3D nanopositioning based on tunneling current sensing

and piezoactuation
Lukasz Ryba

To cite this version:

Lukasz Ryba. 3D nanopositioning based on tunneling current sensing and piezoactuation. Signal and
Image processing. Université Grenoble Alpes, 2015. English. �NNT : 2015GREAT112�. �tel-01267258�

HAL Id: tel-01267258

https://theses.hal.science/tel-01267258v1
Submitted on 9 Feb 2016

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
THÈSE
Pour obtenir le grade de
DOCTEUR DE L’UNIVERSITÉ GRENOBLE ALPES
Spécialité : AUTOMATIQUE - PRODUCTIQUE
Arrêté ministériel : 7 août 2006

Présentée par

« Łukasz RYBA »

Thèse dirigée par « Alina VODA » et

codirigée par « Gildas BESANÇON »

préparée au sein du Laboratoire GIPSA-Lab (Grenoble Images

Parole Signal Automatique)
dans l' École Doctorale EEATS (Electronique,
Electrotechnique, Automatique et Traitement du Signal)

Nanopositionnement 3D à
base de mesure à courant
tunnel et piezo-actionnement
Thèse soutenue publiquement le « 27 Novembre 2015 »,
devant le jury composé de :

M. Skandar BASROUR
Professeur, Université Joseph Fourier, Président et Examinateur
M. S.O.R. MOHEIMANI
Professeur, Université du Texas à Dallas, Rapporteur
M. Stéphane REGNIER
Professeur, Université Pierre et Marie Curie, Rapporteur
M. Micky RAKOTONDRABE
Maître de conférences, Université de Franche-Comté, Examinateur
Mme. Alina VODA
Maître de conférences, Université Joseph Fourier, Directeur de thèse
M. Gildas BESANÇON
Professeur, Institut Polytechnique de Grenoble, Codirecteur de thèse
M. Mohammad Al JANAIDEH
Maître de conférences, Université de Toronto, Invitée
 Contents

List of notations xv

Introduction 1

1 State-of-the-art 5

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.2 Micro-/Nanoscale applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.3 Tools, challenges and associated control issues . . . . . . . . . . . . . . . . . . . 7

1.3.1 Tunneling current phenomenon . . . . . . . . . . . . . . . . . . . . . . . 7

1.3.2 Piezoelectric actuators - modeling and control . . . . . . . . . . . . . . . 11

1.3.3 Sensors and measurement noises . . . . . . . . . . . . . . . . . . . . . . 18

1.4 3D operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

1.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2 System description and modeling 21

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.2 Experimental setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.3 3D System modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2.3.1 Horizontal X and Y axes . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2.3.2 Vertical Z axis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

3 Static hysteresis and creep 35

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

3.2 Prandtl-Ishlinskii approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

3.2.1 Model initialisation, identification and inversion . . . . . . . . . . . . . . 37

i
ii Contents

3.2.2 Hysteresis compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

3.3 Modified Prandtl-Ishlinskii approach . . . . . . . . . . . . . . . . . . . . . . . . 42

3.3.1 Model initialisation, identification and inversion . . . . . . . . . . . . . . 43

3.3.2 Hysteresis compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

3.4 Observer-based approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

3.5 Observer-based approach with MPI model . . . . . . . . . . . . . . . . . . . . . 50

3.6 Numerical inverse-based hysteresis compensation with adaptation . . . . . . . . 53

3.6.1 Adaptive forgetting as a statistical decision problem . . . . . . . . . . . 55

3.6.2 Numerical indirect model inverse . . . . . . . . . . . . . . . . . . . . . . 57

3.6.3 Hysteresis compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

4 Structural vibration and cross-couplings 65

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

4.2 Linear nominal and perturbed model . . . . . . . . . . . . . . . . . . . . . . . . 66

4.2.1 Model identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

4.2.2 Model reduction and model uncertainty . . . . . . . . . . . . . . . . . . 68

4.3 LQG/LTR approach for vibration reduction . . . . . . . . . . . . . . . . . . . . 71

4.4 H∞ approach for vibration reduction . . . . . . . . . . . . . . . . . . . . . . . . 76

4.4.1 General H∞ problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

4.4.2 SISO design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

4.4.3 MIMO design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

5 Tunneling current 91

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

5.2 Nonlinear and linearized model . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

Contents iii

5.3 Digital control using Pole-placement with sensitivity functions shaping . . . . . 94

5.4 Obtaining tunneling current - experimental validation . . . . . . . . . . . . . . 97

5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

6 3D results for nanopositioning and STM applications 101

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

6.2 Nanopositioning in 3D - experimental validation . . . . . . . . . . . . . . . . . . 104

6.3 Surface reconstruction in STM application in simulation . . . . . . . . . . . . . 110

6.3.1 Plant simulation model . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

6.3.2 SISO approach - 3 PID controllers with decoupling compensators . . . . 113

6.3.3 MIMO approach - Linear Quadratic Integral Controller (LQI) . . . . . . 115

6.3.4 Simulation results for surface reconstruction . . . . . . . . . . . . . . . . 117

6.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

7 Cantilever nanopositioning with electrostatic actuation 121

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

7.2 Multi-mode cantilever model analysis . . . . . . . . . . . . . . . . . . . . . . . . 122

7.3 Tunneling current-based cantilever nanopositioning with electrostatic actuation

and piezoelectric tip actuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

7.3.1 H∞ design for tunneling gap control via electrostatically actuated can-
tilever . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

7.3.2 Cantilever positioning via piezoelectrically actuated tip with proximity

force compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

7.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

8 Conclusions and perspectives 143

Bibliography 154
 List of Figures

1.1 3D Scanning tunneling control system. . . . . . . . . . . . . . . . . . . . . . . . 6

1.2 STM/AFM applications examples [Abramovitch et al., 2007]: (a) STM image
of a gold surface. (b) 3D AFM image of blood cells. . . . . . . . . . . . . . . . . 6

1.3 Tunneling effect: (a) Unbiased tip/surface. (b) Surface biased w.r.t. the tip. . . 8

1.4 Electron behaviour in three energy regions during tunneling effect. . . . . . . . 9

1.5 Adverse phenomena of piezoelectric actuators: (a) Hysteresis. (b) Creep. (b)
Structural vibration. (d) Cross-coupling (from X to Y axis). . . . . . . . . . . . 11

1.6 Viscoelastic creep model [Croft et al., 2001]. . . . . . . . . . . . . . . . . . . . . 14

1.7 Control architectures used to control piezoelectric actuator: (a) Feedforward

control. (b) Feedback control. (d) Feedforward with feedback control. . . . . . . 16

2.1 Schema of the experimental setup. . . . . . . . . . . . . . . . . . . . . . . . . . 22

2.2 Experimental setup of 3D nanopositioning platform: (a) General view. (b) The
heart of the platform. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

2.3 Schema of the full 3D nonlinar model. . . . . . . . . . . . . . . . . . . . . . . . 23

2.4 Schema of the experimental setup in vertical (Z) direction. . . . . . . . . . . . . 26

3.1 2D Schema (inverse-based approach). . . . . . . . . . . . . . . . . . . . . . . . . 36

3.2 Hysteretic nonlinearity: (a) Weighted backlash operator. (b) Hysteresis opera-
tor as a superposition of n + 1 backlash operators [Ang et al., 2007], [Rakoton-
drabe et al., 2010]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

3.3 Identification procedure for X direction (PI approach). . . . . . . . . . . . . . . 39

3.4 PI approach for hysteresis compensation (in X (left) and Y (right) direction)
with 16 backlashes: (a), (b) Before identification. (c), (d) After identification.
(e), (f) Inverted model. (g), (h) Compensation. . . . . . . . . . . . . . . . . . . 40

3.5 Reference tracking using PI approach for X (left) and Y (right) directions (sub-
scripts ’uc’ and ’c’ refer to the uncompensated and compensated case, respec-
tively): (a), (b) Triangle of variable amplitude. (c), (d) Step response. . . . . . 41

v
vi List of Figures

3.6 Weighted dead-zone operator (a) and superposition operator consisting of m+1
one-sided dead-zone operators (b) [Ang et al., 2007]. . . . . . . . . . . . . . . . 42

3.7 Identification procedure for X direction (MPI approach). . . . . . . . . . . . . . 44

3.8 MPI approach for hysteresis compensation (in X (left) and Y (right) direction)
with 16 backlashes and 16 dead-zones: (a), (b) Before identification. (c), (d)
After identification. (e), (f) Inverted model. (g), (h) Compensation. . . . . . . . 45

3.9 Reference tracking using MPI approach for X (left) and Y (right) directions:
(a), (b) Triangle of variable amplitude. (c), (d) Step response. . . . . . . . . . . 47

3.10 2D schema (observer-based approach). . . . . . . . . . . . . . . . . . . . . . . . 48

3.11 Hysteresis approximation: (a) Affine approximation. (b) MPI approximation. . 48

3.12 Observer-based approach for hysteresis compensation (in X (left) and Y (right)
direction) (a), (b) Hysteresis reconstruction. (c), (d) Compensation. . . . . . . 49

3.13 Reference tracking using observer-based approach for X (left) and Y (right)
directions: (a), (b) Triangle of variable amplitude. (c), (d) Step response. . . . 50

3.14 2D schema (observer-based approach with MPI model). . . . . . . . . . . . . . 51

3.15 Hysteresis compensation in the X direction and for 1 Hz (left), 20 Hz (middle)

and 50 Hz (right) for: (a)-(c) MPI approach, (d)-(f) DOB approach, (g)-(i)
MPI+DOB approach. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

3.16 Triangular reference tracking for 1 Hz (left), 20 Hz (middle) and 50 Hz (right)

for: (a)-(c) MPI approach, (d)-(f) DOB approach, (g)-(i) MPI+DOB approach. 53

3.17 Step tracking for: (a) MPI approach, (b) DOB approach, (c) MPI+DOB ap-
proach. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

3.18 Disturbance rate of change (d˙x ) for different frequencies: (a) 1 Hz, (b) 10 Hz,
(c) 20 Hz, (d) 50 Hz. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

3.19 The considered model of X axis with adaptive estimator (AE) and numerical
inversion block (N I) for hysteresis compensation. . . . . . . . . . . . . . . . . . 55

3.20 Triangular reference tracking with adaptation using numerical algorithm N I:

(a) only Newton method, (b) Newton combined with bisection. . . . . . . . . . 59

3.21 Experimental results of hysteresis compensation for sinusoidal reference of vari-

able amplitude xr (t) = 5 + 2sin(2πt) + 2sin(2π1.2t + π) for ρ = 0.4 (left) and
ρ = 0.8 (right): (a), (b) Measured and estimated hysteresis. (c), (d) Inverted
model. (e), (f) Compensation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
List of Figures vii

3.22 Experimental results for tracking sinusoidal reference of variable amplitude: (a)
Without compensation. (b), (c), (d) With compensation for ρ = 0.8, 0.4, 0.0,
respectively. (e) Corresponding tracking errors. (f) Corresponding variable
forgetting factor. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

3.23 Experimental results for tracking triangular reference of variable amplitude: (a),
(c), (e) Without any compensation, with MPI and with the proposed adaptive
compensator, respectively. (b), (d), (f) Corresponding tracking errors. . . . . . 62

4.1 Experimental input/output data used for identification: (a) from X to X (b)
from Y to X (c) from X to Y (d) from Y to Y. . . . . . . . . . . . . . . . . . . 67

4.2 Frequency response of the real system and full-order TITO model. . . . . . . . 68

4.3 Step response of the full-order TITO model. . . . . . . . . . . . . . . . . . . . . 68

4.4 Hankel singular values of the components of the full-order TITO model (red
lines point at the number of states of the reduced model): (a) σxxp (b) σxyp (c)
σyxp (d) σyyp . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

4.5 Frequency response of the full- and reduced-order TITO model. . . . . . . . . . 70

4.6 Step response of the full- and reduced-order TITO model. . . . . . . . . . . . . 70

4.7 Feedback system with output multiplicative uncertainty. . . . . . . . . . . . . . 71

4.8 Feedback loop with two SISO LQG/LTR controllers (based only on the diagonal
part of G(s)). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

4.9 LTR with increasing recovery gain qi = [1, 10, 102 , 106 , 108 , 1010 , 1016 ] (LT FP2i (s)
is in dashed line from green to black, the desired LT FP1i (s) is in solid blue line):
a) Nyquist plot for i = x. b) Nyquist plot for i = y. c) Bode plot for i = x. d)
Bode plot for i = y. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

4.10 Robustness w.r.t. model uncertainty: a), b) Recovered LT FP 2 for X and Y

axis, respectively. c), d) Robust stability check for X and Y axis, respectively. . 76

4.11 Performance of SISO LQG/LTR controllers: a), b) Magnitude response of the

system in open-loop and closed-loop (with real and nominal system) for X and
Y axis, respectively. c), d) Step response of the system in open-loop and closed-
loop (with real and nominal system) for X and Y axis, respectively. . . . . . . . 77

4.12 Magnitude plot of the weighting function and relative errors for X (left) and b)
for Y (right) axis: a) |Wx∆ (jω)| ≥ l∆xx (ω) b) |Wy∆ (jω)| ≥ l∆yy (ω). . . . . . . . 81

4.13 Sensitivity functions for the nominal plant and the obtained controller for X
(left) and for Y (right) axis: a), b) Output sensitivity function. c), d) Comple-
mentary sensitivity function. e), f) Control sensitivity function. . . . . . . . . . 82
viii List of Figures

4.14 Performance of SISO H∞ controllers: a), b) Magnitude response of the system

in open-loop and closed-loop (with real and nominal system) for X and Y axis,
respectively. c), d) Step response of the system in open-loop and closed-loop
(with real and nominal system) for X and Y axis, respectively. . . . . . . . . . . 83

4.15 Reduction of SISO H∞ controllers with 17 states: a), b) Frequency response of

nth order controller for X and Y axis, respectively. c), d) Step response of the
closed-loop system with nth order controller and real system for X and Y axis,
respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

4.16 Frequency response of an open-loop and closed-loop perturbed system for dif-
ferent control approaches for structural vibration reduction. . . . . . . . . . . . 86

4.17 Closed-loop controlled TITO system: a) SISO control. b) MIMO control. . . . 87

4.18 Frequency response of an open-loop and closed-loop nominal system for different
control approaches for structural vibration reduction. . . . . . . . . . . . . . . . 87

4.19 Frequency response of an open-loop and closed-loop perturbed system for dif-
ferent control approaches for structural vibration reduction. . . . . . . . . . . . 88

4.20 Step response of an open-loop and closed-loop real system for different control
approaches for structural vibration reduction when only one axis is excited at
the same time. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

4.21 Step response of an open-loop and closed-loop real system for different control
approaches for structural vibration reduction when only one axis is excited at
the same time. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

5.1 Block scheme of nonlinear plant in vertical Z direction. . . . . . . . . . . . . . . 91

5.2 Fitting the model (solid-red) it = gVb e−kd to the experimental data (dashed-blue). 93

5.3 Closed-loop digital control system with RS controller. . . . . . . . . . . . . . . 95

5.4 Performance of the closed-loop system model with RS controller: a) Output

sensitivity function Syp (z −1 ), b) Input (control) sensitivity function Sup (z −1 ).
c) Closed-loop pulse transfer function HCL (z −1 ) (complementary sensitivity
function). d) Step response of the closed-loop system. . . . . . . . . . . . . . . 98

5.5 Closed-loop control system with real plant. . . . . . . . . . . . . . . . . . . . . 98

5.6 Manual approach (view via high-precision camera): (a) The tip and the surface.
(b) Tip deformation after contact with the surface. . . . . . . . . . . . . . . . . 99

5.7 Fine approach and tunneling current control: (a) Fine approach of the tip
with tunneling current stabilization, (b) Tracking the sinusoidal reference of
tunneling current. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
List of Figures ix

6.1 Signals used for the raster (left column) and the spiral (right column) pattern
generation: (a), (b) Reference xr (t). (c), (d) Reference yr (t). (e), (f) Reference
pattern in X-Y plane. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

6.2 Experimental results for 3D operation of the platform and low signal to noise ra-
tio: (a) Reference tunneling current (3D view). (b) Reference tunneling current
(2D view). (c) Tunneling current without hysteresis and creep compensation
(3D view). (d) Tunneling current without hysteresis and creep compensation
(2D view). (e) Tunneling current with hysteresis and creep compensation (3D
view). (f) Tunneling current with hysteresis and creep compensation (2D view).
(g) X-Y trajectory without hysteresis and creep compensation. (h) X-Y trajec-
tory with hysteresis and creep compensation. . . . . . . . . . . . . . . . . . . . 105

6.3 Reference tunneling current: (a) 3D view. (b) 2D view. . . . . . . . . . . . . . . 106

6.4 Tracking spiral trajectory of frequency 2πωs : (a)-(c) 1 Hz. (e)-(g) 10 Hz. (i)-(k)
20 Hz. (m)-(o) 50 Hz. (d), (h), (l), (p) Corresponding tracking errors. . . . . . 107

6.5 Tunneling current obtained using only MPI (left), only DOB (middle) and
combined MPI+DOB (right) approach for the scanning frequencies 2πωs : (a)-
(c) 1 Hz. (d)-(f) 10 Hz. (g)-(i) 20 Hz. (j)-(l) 50 Hz. . . . . . . . . . . . . . . . . 108

6.6 Tunneling current obtained for scanning frequency 1 Hz and MPI+DOB. (a)
Without any compensation for X and Y axes. (b) With compensation only for
X axis. (c) With compensation only for Y axis. (d) With compensation for
both X and Y axes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

6.7 Simulation plant model used for surface variation zs (t) reconstruction. . . . . . 110

6.8 STM-like controlled system with decentralized PID SISO controllers and de-
coupling compensators. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

6.9 STM-like controlled system with centralized LQI MIMO controller. . . . . . . . 116

6.10 Surface reconstruction: (a) Reference surface variations zr (t). (b) Without any
compensation. SISO approach with subsequent compensation of: (c) cross-
couplings, (d) hysteresis, (e) creep and vibrations. (f) MIMO approach. . . . . 119

6.11 Tunneling current: (a) SISO approach, (b) MIMO approach. . . . . . . . . . . . 120

7.1 Schematic showing the cantilever: (a) Subjected to the external force of value
−Fp (t) acting at a distance xf from the base of the cantilever with the the
damping force pd (x, t), assumed constant per unit length. (b) The equivalent
schematic with the total net force Fc (t) = −Fp (t) + Fe (t). . . . . . . . . . . . . 122

7.2 Modal shape functions for 3 modes with singular points (pointed in red): (a)
1st mode. (b) 2nd mode. (c) 3rd mode. . . . . . . . . . . . . . . . . . . . . . . . 124
x List of Figures

7.3 Frequency response of (a) shaped modal and (b) total displacement (sum of
shaped modes) when the applied input force and measured output displacement
are at the distance x = L = 500 µm (i.e. cantilever end). . . . . . . . . . . . . . 124

7.4 Magnitude plot of different shaped modes at singular points, when the applied
force is at xf = 500 µm (cantilever end): (a) x = 1 µm ≈ 0 µm. (b) x = 252 µm.
(c) x = 391.5 µm. (d) x = 434 µm. . . . . . . . . . . . . . . . . . . . . . . . . . 125

7.5 Magnitude plot of the cantilever total displacement at singular points, when
the applied force is at xf = 500 µm (cantilever end): (a) x = 1 µm ≈ 0 µm.
(b) x = 252 µm. (c) x = 391.5 µm. (d) x = 434 µm. . . . . . . . . . . . . . . . 126

7.6 Magnitude plot of different shaped modes at singular points, when the applied
force is at xf = 252 µm: (a) x = 1 µm ≈ 0 µm. (b) x = 252 µm. (c)
x = 391.5 µm. (d) x = 434 µm. . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

7.7 Magnitude plot of the cantilever total displacement at singular points, when
the applied force is at xf = 252 µm: (a) x = 1 µm ≈ 0 µm. (b) x = 252 µm.
(c) x = 391.5 µm. (d) x = 434 µm. . . . . . . . . . . . . . . . . . . . . . . . . . 128

7.8 Input/output data: (a), (c) Input force Fc (t) for SNR = 30 dB and 15 dB,
respectively. (b), (d) Corresponding output displacements zc (x, t) (total and
coming from individual shaped modes), captured at the position x = 150 µm,
when the input force Fc (t) is applied at xf = L = 500 µm. . . . . . . . . . . . . 129

7.9 Output displacements of the cantilever (shaped modal and total) at singular
points, when the applied force is at xf = 252 µm: (a) x = 1 µm ≈ 0 µm. (b)
x = 252 µm. (c) x = 391.5 µm. (d) x = 434 µm. . . . . . . . . . . . . . . . . . 130

7.10 Frequency response of (a) shaped modal and (b) total displacement (sum of
shaped modes) when the applied input force and measured output displacement
are at the distance x = xf = 0.81L = 405 µm. . . . . . . . . . . . . . . . . . . . 131

7.11 Output displacement of the cantilever (shaped modal and total) for x = xf =
0.81L = 405 µm for the input force with: (a) SNR = 30 dB (Fig. 7.8a). (b)
SNR = 15 dB (Fig. 7.8c). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

7.12 Closed-loop control of the vertical subsystem. The tunneling gap d = zt −

zc (xf ) = d0 + zp − zc (xf ) is regulated by the H∞ controller via cantilever
position zc (xf ) = zcp (xf ) + zce (xf ) by keeping the tunneling voltage y3 (and
as a result the tunneling gap) on the constant level of zr . The interaction
(proximity) force Fp between the tip and the cantilever is estimated via F E
subsystem from the tunneling voltage y3 and the control signals uz and ue and
eliminated by adjusting the position of the tunneling tip zp via outer closed-loop
with PID controller, tuned automatically using Matlab & Simulink software. . . 132
List of Figures xi

7.13 The perturbed 3-mode model (in red) and the nominal 1-mode model (in blue)
in case of: (a) x = xf = L = 500 µm. (b) x = xf = 0.81L = 405 µm. . . . . . . 134

7.14 The relative errors l∆ (ω) together with their upper bounds |W∆ (jω)| used for
H∞ design: (a) x = xf = L = 500 µm and 5th order weight W∆ (s). (b) x =
xf = L = 500 µm and 2nd order weight W∆ (s). (c) x = xf = 0.81L = 405 µm
and 5th order weight W∆ (s). (d) x = xf = 0.81L = 405 µm and 2nd order
weight W∆ (s). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

7.15 Sensitivity functions for the nominal plant and the obtained H∞ controller
for the desired bandwidth 5 kHz and 2nd order uncertainty weight W∆ (s): a)
Output sensitivity function. b) Complementary sensitivity function. c) Control
sensitivity function. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

7.16 H∞ control of the linearized model from the input ue to the output y3 (without
the noise) (the open- loop response of the plant has been scaled by its gain in
order to be with the same scale with the closed-loop one): (a) x = xf = L =
500 µm and 5th order weight W∆ (s). (b) x = xf = L = 500 µm and 2nd order
weight W∆ (s). (c) x = xf = 0.81L = 405 µm and 5th order weight W∆ (s). (d)
x = xf = 0.81L = 405 µm and 2nd order weight W∆ (s). . . . . . . . . . . . . . 137

7.17 Bode diagrams of the scaled open-loop (OL) and closed-loop system with H∞
controller (with 5th order weight W∆ (s)) and the perturbed plant model: (a)
x = xf = L = 500 µm. (b) x = xf = 0.81L = 405 µm. . . . . . . . . . . . . . . 138

7.18 Comparison of H∞ control with the PID controller (PIDt is the closed-loop
response with the first shot parameters of PID proposed by Matlab & Simulink
and PIDl refers to the fastest possible PID controller before the stability loss). . 138

7.19 Simulated proximity force for different tip curvature radius R. Here, the inter-
atomic distance ap = 0.166 nm. The inset plot shows the zoom of the proximity
forces in the tunneling voltage measuring range (i.e. 0-10 V, which corresponds
to the tip/surface distance 0.7-1 nm). . . . . . . . . . . . . . . . . . . . . . . . . 139

7.20 The performance of the two closed-loop scheme from Fig. 7.12 when the prox-
imity force Fp is absent: (a) Tunneling voltage y3 . (b) Control signal ue of
the inner loop with H∞ controller. (c) Cantilever and tip displacements. (b)
Control signal uz of the outer loop with PID controller. . . . . . . . . . . . . . 141

7.21 The performance of the two closed-loop scheme from Fig. 7.12 when the prox-
imity force Fp is present: (a) Tip and cantilever displacements without (ztu
and zcu ) and with (ztc and zcc ) proximity force compensation (without the
creep phenomenon). (b) Estimated proximity force Fp (without the creep phe-
nomenon). (c) Tip and cantilever displacements without (ztu and zcu ) and with
(ztc and zcc ) proximity force compensation (with the creep phenomenon). (d)
Estimated proximity force Fp (with the creep phenomenon). . . . . . . . . . . . 142
 List of Tables

3.1 Amount of hysteresis H% using different compensation methods and different

frequencies - experimental result (uncompensated system has H% = 11.37% of
hysteresis) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

5.1 System parameters for the vertical system. . . . . . . . . . . . . . . . . . . . . . 93

6.1 Tracking error-experimental result . . . . . . . . . . . . . . . . . . . . . . . . . 106

6.2 System parameters used for simulation. . . . . . . . . . . . . . . . . . . . . . . . 118

7.1 Numerical data concerning the cantilever. . . . . . . . . . . . . . . . . . . . . . 122

7.2 Parameters computed for the first three modes. . . . . . . . . . . . . . . . . . . 123

xiii
 List of notations

Chapter 1: State-of-the-art
EF Fermi level of metals
EFt Fermi level of the tip
EFs Fermi level of the surface
φ Work function for metals
Vb Potential difference applied between two electrodes
e Electron charge
me Electron mass
E Electron total energy
~ Reduced Planck constant
V (z) Potential barrier
d(t) Length of the potential barrier (tunneling gap)
z(t) Vertical displacement of tunneling tip
ψ(z) Wave function of electron
ψI (z) and ψIII (z) Wave function of electron outside the potential barrier
ψII (z) Wave function of electron inside the potential barrier
k0 Wave number outside the potential barrier
k1 Wave number inside the potential barrier
A1 , A2 and C1 , C2 Amplitude constants for the wave outside the potential barrier
B1 , B2 Amplitude constants for the wave inside the potential barrier
ρLS (z, E) Local density of states (LDOS)
it (t) Tunneling current
g Proportionality constant of tunneling current
k Constant of tunneling current
u(t) Input of piezoelectric actuator
y(t) Output of piezoelectric actuator
γ̂α,β [u(t)] Relay operator of the Preisach model
α and β Thresholds for the relay operator in the Preisach model
µ(α, β) Density function for the realy operator in the Preisach model
Fr [u](t) Play operator of the Prandtl-Ishlinskii model
r Threshold for the play operator in the Prandtl-Ishlinskii model
p(r) Density function for the play operator in the Prandtl-Ishlinskii model
t0 Time instant at which the creep effect occurs
y0 Piezo output at time t0
γ Rate parameter of the nonlinear creep model
kc0 Low frequency elastic behavior of the linear creep model
kci Spring constant of a spring in the linear creep model
cci Damping constant of a damper in the linear creep model

xv
xvi List of notations

Chapter 2: System description and modeling

ux (t) Control signal for the X axis (voltage amplifier input)
uy (t) Control signal for the Y axis (voltage amplifier input)
uz (t) Control signal for the Z axis (voltage amplifier input)
Ve (t) Control signal for the Z axis (electrostatic actuator input)
V Ax (s) Model of the voltage amplifier (X axis)
V Ay (s) Model of the voltage amplifier (Y axis)
V Az (s) Model of the voltage amplifier (Z axis)
Gvx Gain of the voltage amplifier (X axis)
Gvy Gain of the voltage amplifier (Y axis)
Gvz Gain of the voltage amplifier (Z axis)
ωvx Bandwidth of the voltage amplifier (X axis)
ωvy Bandwidth of the voltage amplifier (Y axis)
ωvz Bandwidth of the voltage amplifier (Z axis)
vpx (t) Piezo input voltage (X axis)
vpy (t) Piezo input voltage (Y axis)
vpz (t) Piezo input voltage (Z axis)
dx (t) Disturbance including hysteresis modeling error and creep (X axis)
dy (t) Disturbance including hysteresis modeling error and creep (Y axis)
dz (t) Disturbance including creep phenomenon (Z axis)
Γx [vpx (t)] Nonlinear operator representing static hysteresis model (X axis)
Γy [vpy (t)] Nonlinear operator representing static hysteresis model (Y axis)
N Lx [vpx ](t) Operator representing total piezo nonlinearity (X axis)
N Ly [vpy ](t) Operator representing total piezo nonlinearity (Y axis)
qpx (t) Output of quasi-static piezo nonlinearity (X axis)
qpy (t) Output of quasi-static piezo nonlinearity (Y axis)
qpz (t) Output of quasi-static part of piezo (Z axis)
Dpx (s) Direct piezo linear dynamics (X axis)
Dpy (s) Direct piezo linear dynamics (Y axis)
Dpz (s) Direct piezo linear dynamics (Z axis)
Xy (s) Cross-coupling dynamics (from Y to X axis)
Yx (s) Cross-coupling dynamics (from X to Y axis)
Zx (s) Cross-coupling dynamics (from X to Z axis)
Zy (s) Cross-coupling dynamics (from Y to Z axis)
xx (t) Direct contribution to piezo displacement (X axis)
yy (t) Direct contribution to piezo displacement (Y axis)
zz (t) Direct contribution to piezo displacement (Z axis)
xy (t) Cross-coupling contribution to piezo displacement (from Y to X axis)
yx (t) Cross-coupling contribution to piezo displacement (from X to Y axis)
zx (t) Cross-coupling contribution to piezo displacement (from X to Z axis)
zy (t) Cross-coupling contribution to piezo displacement (from Y to Z axis)
List of notations xvii

xp (t) Total piezo displacement (X axis)

yp (t) Total piezo displacement (Y axis)
zp (t) Total piezo displacement (Z axis)
Gpx Gain of the piezoelectric actuator (X axis)
Gpy Gain of the piezoelectric actuator (Y axis)
Gpz Gain of the piezoelectric actuator (Z axis)
ωpx Bandwidth of the piezoelectric actuator (X axis)
ωpy Bandwidth of the piezoelectric actuator (Y axis)
ωpz Bandwidth of the piezoelectric actuator (Z axis)
ξpx Damping ratio of the piezoelectric actuator (X axis)
ξpy Damping ratio of the piezoelectric actuator (Y axis)
ξpz Damping ratio of the piezoelectric actuator (Z axis)
d0 Initial tunneling gap between the tip and the surface
d(t) Tunneling gap between the tip and the surface
it (t) Tunneling current
g Proportionality constant of tunneling current
k Constant of tunneling current
P L0t Exponential nonlinearity of the tunneling current
CSx (s) Model of the capacitive sensor (X axis)
CSy (s) Model of the capacitive sensor (Y axis)
CSt (s) Model of the current sensor (Z axis)
Gcapx Gain of the capacitive sensor (X axis)
Gcapy Gain of the capacitive sensor (Y axis)
Gt Gain of the current sensor (Z axis)
ωcapx Bandwidth of the capacitive sensor (X axis)
ωcapy Bandwidth of the capacitive sensor (Y axis)
ωt Bandwidth of the current sensor (Z axis)
vx (t) Output of the capacitive sensor (X axis)
vy (t) Output of the capacitive sensor (Y axis)
vz (t) Output of the current sensor (Z axis)
n1 (t) Measurement noise of the capacitive sensor (X axis)
n2 (t) Measurement noise of the capacitive sensor (Y axis)
n3 (t) Measurement noise of the current sensor (Z axis)
y1 (t) Final output of the capacitive sensor (X axis)
y2 (t) Final output of the capacitive sensor (Y axis)
y3 (t) Final output of the current sensor (Z axis)
H Hamaker constant
R Tip radius
ap Interatomic value
Eef∗ Effective contact stiffness (between the tip and the cantilever)
f
Et Elastic modulus of the tip
E Elastic modulus of the surface (cantilever)
vt Poisson ratio of the tip
v Poisson ratio of the surface (cantilever)
EA Quadratic nonlinearity of the electrostatic actuator
Fe (t) Electrostatic force
Fp (t) Proximity force
Fc (t) Total input net force exciting the cantilever
xviii List of notations

CN T (s) Cantilever model

IN Ttc Nonlinear block representing interaction force
zce (x, t) Cantilver displacement (at the position x) due to electrostatic force
zcp (x, t) Cantilver displacement (at the position x) due to proximity force
zc (x, t) Total cantilver displacement (at the position x) due to total force
xf Position x where the force Fc (t) is applied at the cantilever input
L Length of the cantilever
l0 Distance between the unbended cantilever and the electrostatic plate
l(t) Actual distance between the cantilever and the electrostatic plate
ε0 Vacuum permittivity
Ae Common surface between the cantilever and the electrostatic plate
A Cross sectional area of the cantilever
I Area moment of inertia of the cantilever
ρ Density of the cantilever material
ξ Damping ratio for the modes (here the same for each mode)
pd (x, t) Damping force (sum of the structural and electrostatic damping)
mj Modal mass of j th mode
cj Modal damping coefficient of j th mode
kj Modal stiffness coefficient of j th mode
ωj Modal natural frequency of j th mode
Fcj (t) Modal input force of the j th mode
λj Modal wave constant of the j th mode
qj (t) Fundamental deformation of the j th mode (only time-dependent)
φj (x) Modal shape function for the j th mode (only space-dependent)
nc Number of modes
x3D (t) State vector of the nonlinear 3D model
u3D (t) Input vector of the nonlinear 3D model
y3D (t) Output vector of the nonlinear 3D model
d3D (t) Disturbance vector of the nonlinear 3D model
n3D (t) Noise vector of the nonlinear 3D model

Chapter 3: Nonlinear hysteresis and creep

xr (t) Reference value of the capacitive voltage (X axis)
yr (t) Reference value of the capacitive voltage (Y axis)
ux (t) Control signal for the X axis (voltage amplifier input)
uy (t) Control signal for the Y axis (voltage amplifier input)
V Ax (s) Model of the voltage amplifier (X axis)
V Ay (s) Model of the voltage amplifier (Y axis)
Gvx Gain of the voltage amplifier (X axis)
Gvy Gain of the voltage amplifier (Y axis)
vpx (t) Piezo input voltage (X axis)
vpy (t) Piezo input voltage (Y axis)
List of notations xix

dx (t) Disturbance including hysteresis modeling error and creep (X axis)

dy (t) Disturbance including hysteresis modeling error and creep (Y axis)
dˆx (t) Estimate of dx (t)
dˆy (t) Estimate of dy (t)
Γx [vpx (t)] Nonlinear operator representing static hysteresis model (X axis)
Γy [vpy (t)] Nonlinear operator representing static hysteresis model (Y axis)
N Lx [vpx ](t) Operator representing total piezo nonlinearity (X axis)
N Ly [vpy ](t) Operator representing total piezo nonlinearity (Y axis)
Hx [vpx ](t) Prandtl-Ishlinskii (PI) hysteresis operator (X axis)
xpH (t) Output of PI hysteresis operator (X axis)
r xH Vector of thresholds of PI hysteresis operator (X axis)
r 0 xH Vector of thresholds of inverse PI hysteresis operator (X axis)
w xH Vector of weights of PI hysteresis operator (X axis)
w 0 xH Vector of weights of inverse PI hysteresis operator (X axis)
x pH 0 Vector of initial states of PI hysteresis operator (X axis)
v px 0 Vector of initial states of inverse PI hysteresis operator (X axis)
H rxH [vpx , xpH 0 ](t) Vector of backlashes of PI hysteresis operator (X axis)
H rx0 [xpH , v px0 ](t) Vector of backlashes of inverse PI hysteresis operator (X axis)
H
n Number of backlashes of PI hysteresis operator (X axis)
Sx [xpH ](t) Prandtl-Ishlinskii (PI) superposition operator (X axis)
xpS (t) Output of PI superposition operator (X axis)
r xS Vector of thresholds of PI superposition operator (X axis)
r 0 xS Vector of thresholds of inverse PI superposition operator (X axis)
w xS Vector of weights of PI superposition operator (X axis)
w 0 xS Vector of weights of inverse PI superposition operator (X axis)
S rxS [xpH ](t) Vector of dead-zones of PI superposition operator (X axis)
S rx0 [xpS ](t) Vector of dead-zones of inverse PI superposition operator (X axis)
S
m Number of dead-zones of PI superposition operator (X axis)
x̃p (t) Piezo displacement obtained from the measured output y1 (t) (X axis)
wx Unknown MPI model weights (X axis)
Jx (wx ) Loss function for the identification of model weights (X axis)
Fx Hessian matrix - quadratic term in the loss function (X axis)
f x1 Vector - linear term in the loss function (X axis)
fx0 Scalar - absolute term in the loss function (X axis)
l Number of collected data
U H , U 0S Matrices for the constrained problem
uH , u0S Vectors for the constrained problem
qpx (t) Output of quasi-static piezo nonlinearity (X axis)
qpy (t) Output of quasi-static piezo nonlinearity (Y axis)
Dpx (s) Direct piezo linear dynamics (X axis)
Dpy (s) Direct piezo linear dynamics (Y axis)
xp (t) Total piezo displacement (X axis)
yp (t) Total piezo displacement (Y axis)
Gpx Gain of the piezoelectric actuator (X axis)
Gpy Gain of the piezoelectric actuator (Y axis)
xx List of notations

ωpx Bandwidth of the piezoelectric actuator (X axis)

ωpy Bandwidth of the piezoelectric actuator (Y axis)
ξpx Damping ratio of the piezoelectric actuator (X axis)
ξpy Damping ratio of the piezoelectric actuator (Y axis)
CSx (s) Model of the capacitive sensor (X axis)
CSy (s) Model of the capacitive sensor (Y axis)
Gcapx Gain of the capacitive sensor (X axis)
Gcapy Gain of the capacitive sensor (Y axis)
vx (t) Output of the capacitive sensor (X axis)
vy (t) Output of the capacitive sensor (Y axis)
n1 (t) Measurement noise of the capacitive sensor (X axis)
n2 (t) Measurement noise of the capacitive sensor (Y axis)
y1 (t) Final output of the capacitive sensor (X axis)
y2 (t) Final output of the capacitive sensor (Y axis)
Ts Sampling time
fs Sampling frequency
u1 (t) Control input with DOB compensation (X axis)
u2 (t) Control input with DOB compensation (Y axis)
x(t) State vector (X axis)
Ax , Bx , Cx System matrices (X axis)
Bdx Disturbance matrix (X axis)
wx (t) Process noise (X axis)
xe (t) Extended state vector (X axis)
Aex , Bex , Cex Matrices of the extended system (X axis)
Lx Kalman gain matrix (X axis)
Wx Process covariance (X axis)
N1 Noise covariance (X axis)
Px Solution of Algebraic Riccati Equation (ARE)
ŷ1 (t) Estimate of y1 (t)
lH Major hysteresis loop width
LH Range of the reference signal
p Order of polynomial used in adaptive approach
uk Control voltage in adaptive approach
θk Vector of unknown polynomial coefficients in adaptive approach
ymk Model plant output in adaptive approach
yk Real plant output in adaptive approach
ek Mismatch between real plant and its model in adaptive approach
rk Reference signal in adaptive approach
g(uk , θk ) Parametric model of static nonlinearity in adaptive approach
hk Regression vector in adaptive approach
dk Inverse of the unknown variance in adaptive approach
Θk Vector of all uknown parameters (polynomial coefficients and variance)
Dk Set of the observed input/output data up to the instant k
f (Θk |Dk ) Posterior probability density function (pdf)
f0 (Θk+1 ) Alternative for pdf, when the parameters are not changing
f1 (Θk+1 ) Alternative for pdf, when the parameters are uniquely distributed
fA (Θk+1 ) Unknown pdf
fA∗ (Θk+1 ) Decision space of fA (Θk+1 )
List of notations xxi

ϕ0 Probability that fA (Θk+1 ) = f0 (Θk+1 )

ϕ1 Probability fA (Θk+1 ) = f1 (Θk+1 )
that
0
ϕ


ϕ ≡ ϕ0 , ϕ1
D f kfA Kullback-Leibler divergence
Θ∗ Space of Θk
φ(f, ϕ) Cost function in adaptive approach
%0 Penalty coefficient for the success of alternative f0 (Θk+1 )
%1 Penalty coefficient for the success of alternative f1 (Θk+1 )
ρ Positive tuning parameter in adaptive approach
λk|k Variable forgetting factor in adaptive approach
Pk|k Normalized covariance matrix in adaptive approach
Σk|k Least-squares reminder in adaptive approach
vk|k Number of degrees of freedom in adaptive approach
f (uk , θk , rk ) Function of which the roots are looked for in numerical algorithm
u0 Starting point in the Newton’s method
uL Left value for the interval in the Bisection method
uR Right value for the interval in the Bisection method
uB Control input computed by the Newton’s method
uN Control input computed by the Bisection method

Chapter 4: Structural vibration and cross-couplings

r(t) General reference signal
y(t) General plant output
d(t) General disturbance signal
n(t) General measurement noise signal
Gp (s) General perturbed plant
Gxxp (s) Perturbed plant from the input xh to the output y1
Gyyp (s) Perturbed plant from the input yh to the output y2
Gxyp (s) Perturbed plant from the input yh to the output y1
Gyxp (s) Perturbed plant from the input xh to the output y2
G(s) General nominal plant
Gxx (s) Nominal plant from the input xh to the output y1
Gyy (s) Nominal plant from the input yh to the output y2
Gxy (s) Nominal plant from the input yh to the output y1
Gyx (s) Nominal plant from the input xh to the output y2
C(s) General controller
Cx (s) Controller designed on the basis of Gxx (s)
Cy (s) Controller designed on the basis of Gyy (s)
W∆ (s) General uncertainty weight
∆(s) General normalized complex perturbations
l∆ (ω) General relative magnitude of the neglected dynamics
xxii List of notations

σxxp Hankel singular values of Gxxp (s)

σyyp Hankel singular values of Gyyp (s)
σxyp Hankel singular values of Gxyp (s)
σyxp Hankel singular values of Gyxp (s)
xxx State vector of Gxx (s)
xyy State vector of Gyy (s)
xh Control signal (X axis)
yh Control signal (Y axis)
y1 Output signal (X axis)
y2 Output signal (Y axis)
wxx Process noise of Gxx (s)
wyy Process noise of Gyy (s)
n1 Measurement noise (X axis)
n2 Measurement noise (Y axis)
Axx , Bxx , Cxx State matrices of Gxx (s)
Ayy , Byy , Cyy State matrices of Gyy (s)
Gxxa Augmented Gxx (s) (with integral action)
Gyya Augmented Gyy (s) (with integral action)
xxxa State vector of Gxxa (s)
xyya State vector of Gyya (s)
Axxa , Bxxa , Cxxa State matrices of Gxxa (s)
Ayya , Byya , Cyya State matrices of Gyya (s)
Jx (ẋh ) Quadratic cost function for LQG/LTR controller (X axis)
Jy (ẏh ) Quadratic cost function for LQG/LTR controller (Y axis)
Wxxa Process covariance matrix of wxxa
Wyya Process covariance matrix of wyya
N1 Noise covariance matrix of n1
N2 Noise covariance matrix of n2
Kfxxa Kalman Filter gain associated with Gxxa (s)
Kfyya Kalman Filter gain associated with Gyya (s)
Pxxa Solution of Kalman Filter ARE associated with Gxxa (s)
Pyya Solution of Kalman Filter ARE associated with Gyya (s)
ρx Scaling parameter for Kalman Filter associated with Gxxa (s)
ρy Scaling parameter for Kalman Filter associated with Gyya (s)
Qxxa Weighting matrix for the state of Gxxa (s)
Qyya Weighting matrix for the state of Gyya (s)
Rxxa Weighting matrix for the control of Gxxa (s)
Ryya Weighting matrix for the control of Gyya (s)
Kcxxa LQR gain associated with Gxxa (s)
Kcyya LQR gain associated with Gyya (s)
Sxxa Solution of LQR ARE associated with Gxxa (s)
Syya Solution of LQR ARE associated with Gyya (s)
qx Recovery gain for LQG/LTR associated with Gxxa (s)
qy Recovery gain for LQG/LTR associated with Gyya (s)
KxLQG/LT R (s) LQG/LTR controller (without integral action) for the X axis
KyLQG/LT R (s) LQG/LTR controller (without integral action) for the Y axis
LT FP1x (s) Loop Transfer Function (LTF) at point P1x
LT FP1y (s) Loop Transfer Function (LTF) at point P1y
List of notations xxiii

LT FP2x (s) Loop Transfer Function (LTF) at point P2x

LT FP2y (s) Loop Transfer Function (LTF) at point P2y
P (s) General augmented plant (with weighting functions)
Fl (P (s), C(s)) Lower linear fractional transformation
γ∗ Optimal cost in H∞ design
γ Sub-optimal cost in H∞ design
∆φ Phase margin
∆G Gain margin
∆M Module margin
S(s) General output sensitivity function
R(s) General input sensitivity function
T (s) General complementary sensitivity function
WS (s) General weighting function for S(s)
WR (s) General weighting function for R(s)
WT (s) General weighting function for T (s)
W∆ (s) General weighting function for system uncertainty
ω∆ Bandwidth of W∆ (s)
A∆ Lower bound of W∆ (s)
M∆ Upper bound of W∆ (s)
n∆ Order of W∆ (s)
We (s) General weighting function for tracking
ωe Bandwidth of We (s)
Ae Lower bound of We (s)
Me Upper bound of We (s)
ne Order of We (s)
Wu (s) General weighting function for control
ωu Bandwidth of Wu (s)
Au Lower bound of Wu (s)
Mu Upper bound of Wu (s)
nu Order of Wu (s)
γx Sub-optimal cost in H∞ design (X axis)
γy Sub-optimal cost in H∞ design (Y axis)
Sx (s) Output sensitivity function (X axis)
Sy (s) Output sensitivity function (Y axis)
Rx (s) Input sensitivity function (X axis)
Ry (s) Input sensitivity function (Y axis)
Tx (s) Complementary sensitivity function (X axis)
Ty (s) Complementary sensitivity function (Y axis)
WSx (s) Weighting function for Sx (s)
WSy (s) Weighting function for Sy (s)
WRx (s) Weighting function for Rx (s)
WRy (s) Weighting function for Ry (s)
WTx (s) Weighting function for Tx (s)
WTy (s) Weighting function for Ty (s)
Wx∆ (s) Weighting function for system uncertainty (X axis)
Wy∆ (s) Weighting function for system uncertainty (Y axis)
xxiv List of notations

ωx∆ Bandwidth of Wx∆ (s)

ωy∆ Bandwidth of Wy∆ (s)
A x∆ Lower bound of Wx∆ (s)
A y∆ Lower bound of Wy∆ (s)
M x∆ Upper bound of Wx∆ (s)
M y∆ Upper bound of Wy∆ (s)
n x∆ Order of Wx∆ (s)
n y∆ Order of Wy∆ (s)
Wex (s) Weighting function for tracking (X axis)
Wey (s) Weighting function for tracking (Y axis)
ωex Bandwidth of Wex (s)
ωey Bandwidth of Wey (s)
Ae x Lower bound of Wex (s)
Ae y Lower bound of Wey (s)
Mex Upper bound of Wex (s)
Mey Upper bound of Wey (s)
ne x Order of Wex (s)
ne y Order of Wey (s)
Wux (s) Weighting function for control (X axis)
Wuy (s) Weighting function for control (Y axis)
ωux Bandwidth of Wux (s)
ωuy Bandwidth of Wuy (s)
Aux Lower bound of Wux (s)
Auy Lower bound of Wuy (s)
Mux Upper bound of Wux (s)
Muy Upper bound of Wuy (s)
nux Order of Wux (s)
nuy Order of Wuy (s)
S(s) Output sensitivity function for MIMO system
R(s) Input sensitivity function for MIMO system
T (s) Complementary sensitivity function for MIMO system
W S (s) Weighting function for S S (s)
W R (s) Weighting function for RR (s)
W T (s) Weighting function for T T (s)
σ(X(jω)) Minimum singular value of matrix X
σ(X(jω)) Maximum singular value of matrix X
List of notations xxv

Chapter 5: Tunneling current

uz (t) Control signal for the Z axis (voltage amplifier input)
V Az (s) Model of the voltage amplifier (Z axis)
Gvz Gain of the voltage amplifier (Z axis)
ωvz Bandwidth of the voltage amplifier (Z axis)
vpz (t) Piezo input voltage (Z axis)
dz (t) Disturbance including creep (Z axis)
Dpz (s) Direct piezo linear dynamics (Z axis)
zz (t) Direct contribution to piezo displacement (Z axis)
zx (t) Cross-coupling contribution to piezo displacement (from X to Z axis)
zy (t) Cross-coupling contribution to piezo displacement (from Y to Z axis)
zp (t) Total piezo displacement (Z axis)
zs (t) Surface variations
P iezoz (s) Model of the piezoelectric actuator (Z axis)
Gpz Gain of the piezoelectric actuator (Z axis)
ωpz Bandwidth of the piezoelectric actuator (Z axis)
ξpz Damping ratio of the piezoelectric actuator (Z axis)
d0 Initial tunneling gap between the tip and the surface
d(t) Tunneling gap between the tip and the surface
deq Equilibrium gap between the tip and the surface
∆d(t) Gap deviation from its equilibrium deq
it (t) Tunneling current
ieq Equilibrium tunneling current
∆it (t) Tunneling current deviation from its equilibrium ieq
Vb Bias Voltage
g Proportionality constant of tunneling current
k Constant of tunneling current
P L0t Exponential nonlinearity of the tunneling current
e Electron charge
me Electron mass
φ Work function for metals
CSt (s) Model of the current sensor
Gt Gain of the current sensor
ωt Bandwidth of the current sensor
vz (t) Output of the current sensor
n3 (t) Measurement noise of the current sensor
y3 (t) Final output of the current sensor
zr (t) Reference value of the tunneling voltage
Gz (s) Continuous-time plant model (Z axis)
H(z −1 ) Discrete-time plant model (Z axis)
fs Sampling frequency
dl System pure delay expressed in number of sampling periods
xxvi List of notations

B(z −1 ) Polynomial representing numerator of H(z −1 )

A(z −1 ) Polynomial representing denominator of H(z −1 )
R(z −1 ) Polynomial representing numerator of the digital controller
S(z −1 ) Polynomial representing denominator of the digital controller
PD (z −1 ) Polynomial representing desired closed-loop poles
PF (z −1 ) Polynomial representing auxiliary closed-loop poles
P (z −1 ) Polynomial representing closed-loop poles
HR (z −1 ) Polynomial representing fixed part of R(z −1 )
HS (z −1 ) Polynomial representing fixed part of S(z −1 )
R0 (z −1 ) Polynomial representing free part of R(z −1 )
S 0 (z −1 ) Polynomial representing free part of S(z −1 )
nA Degree of polynomial A(z −1 )
nB Degree of polynomial B(z −1 )
nR Degree of polynomial R(z −1 )
nS Degree of polynomial S(z −1 )
nH R Degree of polynomial HR (z −1 )
nH S Degree of polynomial HS (z −1 )
ω0 Desired closed-loop bandwidth
ξ Desired closed-loop damping
HCL (z −1 ) Closed-loop transfer function
Syp (z −1 ) Output sensitivity function
Syr (z −1 ) Complementary sensitivity function
Sup (z −1 ) Input sensitivity function

Chapter 6: 3D results for nanopositioning and STM applications

xr (t) Reference value of the capacitive voltage (X axis)
yr (t) Reference value of the capacitive voltage (Y axis)
zr (t) Reference value of the tunneling voltage
xr0 Starting point (initial value) for xr (t)
yr0 Starting point (initial value) for yr (t)
Ar Amplitude of the triangle reference signal xr (t)
fr Frequency of the triangle reference signal xr (t)
αr Slope of the ramp reference signal yr (t)
rs (t) Instantaneous spiral radius
θs (t) Instantaneous spiral angle
ωs (t) Spiral angular velocity
Ps Spiral pitch
Rs Non-instantaneous spiral radius
Ns Number of crossing points of the spiral signal
ux (t) Control signal for the X axis (voltage amplifier input)
uy (t) Control signal for the Y axis (voltage amplifier input)
uz (t) Control signal for the Z axis (voltage amplifier input)
List of notations xxvii

V Ax (s) Model of the voltage amplifier (X axis)

V Ay (s) Model of the voltage amplifier (Y axis)
V Az (s) Model of the voltage amplifier (Z axis)
Gvx Gain of the voltage amplifier (X axis)
Gvy Gain of the voltage amplifier (Y axis)
Gvz Gain of the voltage amplifier (Z axis)
ωvx Bandwidth of the voltage amplifier (X axis)
ωvy Bandwidth of the voltage amplifier (Y axis)
ωvz Bandwidth of the voltage amplifier (Z axis)
vpx (t) Piezo input voltage (X axis)
vpy (t) Piezo input voltage (Y axis)
vpz (t) Piezo input voltage (Z axis)
Hx [vpx ](t) Nonlinear operator representing static hysteresis PI model (X axis)
Hy [vpy ](t) Nonlinear operator representing static hysteresis PI model (Y axis)
qpx (t) Output of static piezo nonlinearity (X axis)
qpy (t) Output of static piezo nonlinearity (Y axis)
qpz (t) Output of static part of piezo (Z axis)
Gcreepx (s) Creep linear dynamics (X axis)
Gcreepy (s) Creep linear dynamics (Y axis)
Dpx (s) Direct piezo linear dynamics (X axis)
Dpy (s) Direct piezo linear dynamics (Y axis)
Dpz (s) Direct piezo linear dynamics (Z axis)
Gpx Gain of the piezoelectric actuator (X axis)
Gpy Gain of the piezoelectric actuator (Y axis)
Gpz Gain of the piezoelectric actuator (Z axis)
ωpx Bandwidth of the piezoelectric actuator (X axis)
ωpy Bandwidth of the piezoelectric actuator (Y axis)
ωpz Bandwidth of the piezoelectric actuator (Z axis)
ξpx Damping ratio of the piezoelectric actuator (X axis)
ξpy Damping ratio of the piezoelectric actuator (Y axis)
ξpz Damping ratio of the piezoelectric actuator (Z axis)
Zx (s) Cross-coupling dynamics (from X to Z axis)
Zy (s) Cross-coupling dynamics (from Y to Z axis)
Ẑx (s) Estimate of Zx (s)
Ẑy (s) Estimate of Zy (s)
Gzx Gain of the cross-coupling dynamics (from X to Z axis)
Gzy Gain of the cross-coupling dynamics (from Y to Z axis)
Ĝzx Estimate of Gzx
Ĝzy Estimate of Gzy
ωzx Bandwidth of the cross-coupling dynamics (from X to Z axis)
ωzy Bandwidth of the cross-coupling dynamics (from Y to Z axis)
ω̂zx Estimate of ωzx
ω̂zy Estimate of ωzy
xxviii List of notations

xx (t) Direct contribution to piezo displacement (X axis)

yy (t) Direct contribution to piezo displacement (Y axis)
zz (t) Direct contribution to piezo displacement (Z axis)
zx (t) Cross-coupling contribution to piezo displacement (from X to Z axis)
zy (t) Cross-coupling contribution to piezo displacement (from Y to Z axis)
ẑz (t) Estimate of zz (t)
ẑx (t) Estimate of zx (t)
ẑy (t) Estimate of zy (t)
xp (t) Total piezo displacement (X axis)
yp (t) Total piezo displacement (Y axis)
x̂p (t) Estimate of xp (t)
ŷp (t) Estimate of yp (t)
zp (t) Total piezo displacement (Z axis)
zs (t) Surface variations
us (t) Proper (non-compensating) part of control input uz (t)
uc (t) Compensating part of control input uz (t)
ucx (t) Compensating control for cross-coupling (from X to Z axis)
ucy (t) Compensating control for cross-coupling (from Y to Z axis)
Ĝzz (s) Estimate of piezo transfer function
d0 Initial tunneling gap between the tip and the surface
d(t) Tunneling gap between the tip and the surface
ˆ
d(t) Estimate of d(t)
deq Equilibrium gap between the tip and the surface
∆d(t) Gap deviation from its equilibrium deq
it (t) Tunneling current
ieq Equilibrium tunneling current
∆it (t) Tunneling current deviation from its equilibrium ieq
Vb Bias Voltage
g Proportionality constant of tunneling current
k Constant of tunneling current
P L0t Exponential nonlinearity of the tunneling current
CSx (s) Model of the capacitive sensor (X axis)
CSy (s) Model of the capacitive sensor (Y axis)
CSt (s) Model of the current sensor (Z axis)
Gcapx Gain of the capacitive sensor (X axis)
Gcapy Gain of the capacitive sensor (Y axis)
Gt Gain of the current sensor (Z axis)
ωcapx Bandwidth of the capacitive sensor (X axis)
ωcapy Bandwidth of the capacitive sensor (Y axis)
ωt Bandwidth of the current sensor (Z axis)
vx (t) Output of the capacitive sensor (X axis)
vy (t) Output of the capacitive sensor (Y axis)
vz (t) Output of the current sensor (Z axis)
n1 (t) Measurement noise of the capacitive sensor (X axis)
n2 (t) Measurement noise of the capacitive sensor (Y axis)
n3 (t) Measurement noise of the current sensor (Z axis)
List of notations xxix

y1 (t) Final output of the capacitive sensor (X axis)

y2 (t) Final output of the capacitive sensor (Y axis)
y3 (t) Final output of the current sensor (Z axis)
x(t) State vector of the 3D model for surface reconstruction
u(t) Input vector of the 3D model for surface reconstruction
y(t) Output vector of the 3D model for surface reconstruction
xeq Equilibrium state vector
ueq Equilibrium input vector
yeq Equilibrium output vector
∆x(t) Deviation of the state vector from its equilibrium
∆u(t) Deviation of the input vector from its equilibrium
∆y(t) Deviation of the output vector from its equilibrium
∆x̂(t) Estimate of ∆x(t)
e(t) Tracking error vector
r(t) Reference vector
∆xe (t) Extended state vector deviations (with surface variations)
∆x̂e (t) Estimate of ∆xe (t)
A, B, C System matrices
Ae , Be , Ce Matrices of the extended system (for estimation)
Bd Disturbance matrix
n(t) Measurement noise vector
w(t) Process noise vector
N Measurement noise covariance matrix
W Process noise covariance matrix
L Kalman gain matrix
P Solution of the Kalman Filter ARE
xi (t) Integrated tracking error
xc (t) Extended state vector deviations (with an integral action)
Ac , Bc Matrices of the extended system (for control)
Q LQR weighting matrix for the state
R LQR weighting matrix for the control
K LQR gain matrix
S Solution of the LQR ARE

Chapter 7: Cantilever nanopositioning with electrostatic actua-

tion
Fe (t) Electrostatic force
F̂e (t) Estimate of Fe (t)
Fp (t) Proximity force
F̂p (t) Estimate of Fp (t)
Fc (t) Total input net force exciting the cantilever
F̂c (t) Estimate of Fc (t)
xxx List of notations

zce (x, t) Cantilver displacement (at the position x) due to electrostatic force
ẑce (x, t) Estimate of zce (x, t)
zcp (x, t) Cantilver displacement (at the position x) due to proximity force
ẑcp (x, t) Estimate of zcp (x, t)
zc (x, t) Total cantilver displacement (at the position x) due to total force
ẑc (x, t) Estimate of zc (x, t)
zcer Reference for cantilever displacement due to electrostatic force
z cr Reference for total cantilever displacement
A Cross sectional area of the cantilever
I Area moment of inertia of the cantilever
ρ Density of the cantilever material
ξ Damping ratio for the modes (here the same for each mode)
pd (x, t) Damping force (sum of the structural and electrostatic damping)
Q Quality factor
L Length of the cantilever
mj Modal mass of j th mode
cj Modal damping coefficient of j th mode
kj Modal stiffness coefficient of j th mode
ωj Modal natural frequency of j th mode
λj Modal wave constant of the j th mode
qj (t) Fundamental deformation of the j th mode (only time-dependent)
φj (x) Modal shape function for the j th mode (only space-dependent)
CN T (s) Cantilever model
CN T3m (s) Cantilever model (with 3 modes)
CN T1m (s) Cantilever model (with 1 mode)
IN Ttc Nonlinear block representing interaction force
−1
kf = CN TDC Cantilever force constant
xf Position x where the force Fc (t) is applied at the cantilever input
l0 Distance between the unbended cantilever and the electrostatic plate
l(t) Actual distance between the cantilever and the electrostatic plate
ε0 Vacuum permittivity
Ae Common surface between the cantilever and the electrostatic plate
uz (t) Control signal for the Z axis (voltage amplifier input)
Ve (t) Control signal for the Z axis (electrostatic actuator input)
ue (t) Control signal for the Z axis (after exact linearization)
b Proportionality constant of electrostatic actuator
V Az (s) Model of the voltage amplifier (Z axis)
Gvz Gain of the voltage amplifier (Z axis)
vpz (t) Piezo input voltage (Z axis)
dz (t) Disturbance including creep phenomenon (Z axis)
qpz (t) Output of static part of piezo (Z axis)
Dpz (s) Direct piezo linear dynamics (Z axis)
Gpz Gain of the piezoelectric actuator (Z axis)
ωpz Bandwidth of the piezoelectric actuator (Z axis)
ξpz Damping ratio of the piezoelectric actuator (Z axis)
zp (t) Total piezo displacement (Z axis)
zt (t) Tip displacement (Z axis)
ẑt (t) Estimate of zt (t)
List of notations xxxi

d0 Initial tunneling gap between the tip and the surface

d(t) Tunneling gap between the tip and the surface
ˆ
d(t) Estimate of d(t)
it (t) Tunneling current
ieq Equilibrium tunneling current
g Proportionality constant of tunneling current
k Constant of tunneling current
Vb Bias Voltage
CSt (s) Model of the current sensor (Z axis)
Gt Gain of the current sensor (Z axis)
vz (t) Output of the current sensor (Z axis)
n3 (t) Measurement noise of the current sensor (Z axis)
y3 (t) Final output of the current sensor (Z axis)
EA Electrostatic actuator
FE Proximity force estimator
Gue −>y3 p (s) Perturbed linear model for H∞ cantilever control
Gue −>y3 (s) Nominal linear model for H∞ cantilever control
∆(s) Normalized complex perturbation
W∆ (s) Model uncertainty weight
l∆ (ω) Relative magnitude of the neglected dynamics
WSz (s) Weight for the output sensitivity function
WTz (s) Weight for the complementary sensitivity function
WRz (s) Weight for the input sensitivity function
H Hamaker constant
R Tip radius
ap Interatomic value
Eef∗ Effective contact stiffness (between the tip and the cantilever)
f
Et Elastic modulus of the tip
E Elastic modulus of the surface (cantilever)
vt Poisson ratio of the tip
v Poisson ratio of the surface (cantilever)
 Introduction

General context
The recent years show rapidly growing popularity of nanotechnology and related sciences which
can be explained by the continuous need for object miniaturization. One of the most exten-
sively studied branch by the researchers is the Scanning Probe Microscopy (SPM), which was
originated in the early 80s by the invention of the Scanning Tunneling Microscope (STM), for
which its creators Gerd Binnig and Heinrich Rohrer from IBM laboratory, received the Nobel
Prize. This device has the capability of scanning the surface (conductive or semi-conductive)
with the aim of obtaining its morphological (or electrical) properties with an atomic reso-
lution. Together with its successor - Atomic Force Microscope (AFM) (which can be used
with non-conductive materials as well), it has changed many areas like physics, biology and
material sciences, where not only the atomic-resolution measurement but also manipulation
at such small scale is possible. One of the key requirements of these devices is nanoposition-
ing, which is the positioning of a small object with high accuracy and over high bandwidth.
These features are the main aspects that have attracted the researchers working around this
area. However, the measurement noise, disturbances and nonlinearities strongly affect the
positioning at such a small scale. Those problems, which are usually treated by physicists or
electronics specialists are solved here using system theory (modeling, identification, observa-
tion and control).

Thesis objectives and contribution

The main objective of this thesis is to elaborate high performance control strategies for
a three-dimensional (X,Y,Z) model of nanopositioning system and their real-time validation
on an experimental tunneling current-based platform, started in GIPSA-lab a few years ago
[Blanvillain, 2010], [Ahmad, 2011]. The main challenge is to take into account different non-
linear dynamics, disturbances and noises appearing in such 3D system, composed mainly of
a piezoelectrically actuated tunneling tip (like in STM) and a microcantilever (like in AFM)
and to control the whole system with nanoscale accuracy and high bandwidth.

The contribution of this thesis is twofold. First, the development of a new three-dimensional
model of the nanopositioning device is done, based on the tunneling current phenomenon, with
piezoelectrically actuated tunneling tip and electrostatically actuated cantilever, taking into
account the adverse effects of different nature. They are associated mainly with piezoelectric
actuators (like nonlinear hysteresis, creep, structural vibration, or cross-couplings between
the axes) but also with the tunneling current phenomenon (exponential nonlinearity) and
with the electrostatic actuator (quadratic nonlinearity). Such a 3D model, to the authors’
knowledge, does not appear in the literature. The second contribution is the development
of different control approaches for such 3D model. To that end, the horizontal motion in X
and Y directions is first studied. The nonlinear hysteresis and creep exhibited by piezoelectric

1
2 Introduction

actuators are compensated using different methods (operator-based Prandtl-Ishlinskii model,

its modified version, disturbance observer and combination of the two, as well as an adaptive
approach). The effects of piezo vibration and cross-couplings between X and Y axes are re-
duced using robust control methods (SISO LQG/LTR, SISO H∞ and MIMO H∞ ). Next, the
horizontal motion is combined with the vertical motion of the Z axis, based on the tunneling
current phenomenon, controlled by the piezoelectrically actuated tip and the electrostatically
actuated cantilever (by means of pole-placement with sensitivity functions shaping and H∞
design). The results for three applications are presented. First, 3D nanopositioning on the
fixed part of the cantilever is validated experimentally. In the second application, tested in
simulation, the model of STM-like device is used with the aim of surface reconstruction (using
three SISO PID controllers with decoupling method and MIMO LQI control). Finally, the
aim of the last application is the nanopositioning of the multi-mode cantilever model on its
mobile part by means of electrostatic actuation with piezoelectrically actuated tunneling tip
and with proximity force (appearing between the tip and the cantilever) compensation. All
applications are based on the tunneling current phenomenon.

Thesis organization
Chapter 1: In this chapter a state-of-the-art is presented. Among the nanoscale applica-
tions, the focus is mainly on Scanning Tunneling Microscopy. The key components used in
such nanoscale applications, like piezoelectric actuators or tunneling current sensor (in case
of STM) are described in detail, together with the associated challenge of their control, espe-
cially connected with tunneling nonlinearity and with the adverse phenomena of piezoelectric
actuators like hysteresis, creep, vibration and cross-couplings between axes. At the end, an
overview of the nanopositioning control in three dimensions is tackled.

Chapter 2: This chapter is focused on the description of the lab-made experimental nanopo-
sitioning platform based on tunneling current phenomenon, used in this thesis. The full non-
linear system model in three dimensions is developed. It includes the horizontal model (motion
along X and Y axes) with focus on the piezoelectric actuators, driven by voltage amplifiers
and read by capacitive sensors, and the vertical model (motion along the Z axis) of piezoelec-
trically actuated tunneling tip with a high-gain tunneling current sensor and electrostatically
actuated cantilever.

Chapter 3: In this chapter the problem of modeling and compensation of nonlinear hysteresis
and creep present in piezoelectric actuators (in X and Y directions) is considered. Experi-
mental data are used for hysteresis model identification and results for real-time hysteresis
compensation are presented. To that end, several modeling/compensation methods are used,
starting from the classical and modified inverse-based Prandtl-Ishlinskii (PI/MPI) model, dis-
turbance observer (DOB) and the hybrid method, which combines both of them (MPI/DOB).
Finally, a polynomial-based model of hysteresis with adaptively changing coefficients is used
together with Newton-based inversion algorithm. All the results are obtained experimentally
using the nanopositioning platform of GIPSA-lab.
Introduction 3

Chapter 4: Here the 2D horizontal system (X and Y axes) already compensated for hysteresis
is considered, with the purpose of structural vibration and cross-coupling reduction. The two-
input two-output (TITO) high-order linear model is identified and its order is reduced using
balanced truncation method. Such a model together with specified uncertainty, serves as
a nominal model for the controller design. Three kinds of robust controllers are designed
and experimentally tested on the nanopositioning platform: single variable LQG/LTR, single
variable H∞ and multivariable H∞ .

Chapter 5: This chapter is devoted to the vertical system of piezoelectrically actuated metal-
lic tip. The tip is approached to the fixed metallic surface of the cantilever and the tunnel-
ing current phenomenon is observed. Its exponential dependence on the tip/surface gap is
confirmed by fitting the model curve to the experimental data. The model with tunneling
nonlinearity is then linearized around an equilibrium point and a pole placement controller
with sensitivity functions shaping is designed on the basis of this linearized model. Finally,
experimental results for tunneling current stabilization and tracking are shown.

Chapter 6: In this chapter the horizontal motion in X and Y directions is coupled with the
vertical motion in Z direction (not including the cantilever motion). Two kinds of applications
are considered here: the nanopositioning in three dimensions, validated experimentally, and
STM-like application, tested in simulation. In the nanopositioning application the tunneling
tip tracks the reference trajectory along X, Y and Z axes (in the Z axis, the tunneling current
tracks the prespecified reference called "virtual surface"). Here, three kinds of hysteresis
compensation methods (open-loop MPI, closed-loop DOB and a hybrid MPI/DOB) are tested
in horizontal directions in the presence of tunneling current in the vertical direction. In the
STM-like application, the tunneling tip is moved along horizontal X and Y axes, tracking
the prespecified trajectory, while the tunneling current in vertical Z direction is stabilized
at a constant level. To that end, two control approaches are tested with the aim of surface
reconstruction: SISO PID decentralized control with decoupling compensators and MIMO
Linear Quadratic Integral (LQI) centralized control. In the first (more classical) case, the
surface is retrieved from the control signal of the vertical piezo, and in the second one, a
Kalman observer is used for its reconstruction.

Chapter 7: In this chapter, modeling, analysis and simulation of a multi-mode cantilever

model are done. Properties of mode vanishing near the singular points of the associated
modal shape function are used for actuator and sensor location in order to damp the modes.
This allows to reduce the order of the system, as well as to increase its bandwidth. The
application of three-mode cantilever model nanopositioning based on the tunneling current
phenomenon is presented. The control of tunneling gap through electrostatically actuated
cantilever is done using H∞ design and the cantilever positioning over a distance greater than
1 nm is done through piezoelectrically actuated tunneling tip using PID control. A proximity
force estimator is designed for estimation and subsequent compensation of the interaction
force between the tip and the cantilever.

Chapter 8: This chapter concludes the manuscript and gives some perspectives and possible
extensions.
4 Introduction

Publications

This work gave rise to the following publications:

Journal paper:

• Ryba Ł., Voda A., Besançon G., "Experimental comparison of disturbance observer and
inverse-based hysteresis compensation in 3D nanopositioning piezoactuation", Sensors
and Actuators A: Physical 236, pp. 190-205, Elsevier, 2015.

International conference proceedings:

• Ryba Ł., Voda A., Besançon G., "A real-time inverse-based hysteresis compensation
with adaptation", accepted to: 54th IEEE Conference on Decision and Control (CDC
2015), Osaka, Japan, 15-18 December 2015.

• Oubellil R., Ryba Ł., Voda A., Rakotondrabe M., "Experimental model inverse-based
hysteresis compensation on a piezoelectric actuator", 19th International Conference on
System Theory, Control and Computing (ICSTCC 2015), Cheile Gradistei - Fundata
Resort, Romania, 14-16 October 2015.

• Ryba Ł., Voda A., Besançon G., "3DOF nanopositioning control of an experimental tun-
neling current-based platform", 2014 IEEE Conference on Control Applications (CCA),
(part of 2014 IEEE Multi-Conference on Systems and Control (MSC2014)), Antibes,
France, 8-10 October 2014, pp. 1976-1981.

• Ryba Ł., Voda A., Besançon G., "An LQG/LTR approach towards piezoactuator vibra-
tion reduction with observer-based hysteresis compensation", 19th IFAC World Congress
(IFAC 2014), Cape Town, South Africa, 24-29 August 2014, pp. 5623-5628.

• Ryba Ł., Voda A., Besançon G., "Modelling and control of 3D STM-like scanning device
with application to surface reconstruction", 18th IEEE International Conference on
Methods and Models in Automation and Robotics (MMAR 2013), Miedzyzdroje, Poland,
26-29 August 2013, pp. 479-484.
 Chapter 1

State-of-the-art

Contents
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Micro-/Nanoscale applications . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Tools, challenges and associated control issues . . . . . . . . . . . . . . 7
1.3.1 Tunneling current phenomenon . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.2 Piezoelectric actuators - modeling and control . . . . . . . . . . . . . . . . 11
1.3.3 Sensors and measurement noises . . . . . . . . . . . . . . . . . . . . . . . 18
1.4 3D operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

1.1 Introduction

In this first chapter, a short state-of-the-art is given as a starting point to the work on nanopo-
sitioning, considered in this thesis. First, some micro-/nanoscale applications are briefly re-
viewed. Then, an overview of main phenomena (such as tunneling current) and tools (such as
piezoelectric actuators), that will be useful for us, as well as related control issues are given.
Finally, 3D operation, which plays a fundamental role in the present work, is discussed.

1.2 Micro-/Nanoscale applications

The Scanning Tunneling Microscope (STM), since its invention in the beginning of the 1980s
by Gerd Binnig and Heinrich Rohrer [Binnig and Rohrer, 1986], has experienced extensive re-
search and gained wide acceptance in nanotechnology, mainly due to its capability of measuring
topographical/electrical properties of the surfaces with an atomic resolution. The scheme of
an STM-like scanning device is shown in Fig. 1.1. When the conductive tip is approached
to the metallic surface at a distance lower than 1 · 10−9 m and if it is electrically biased w.r.t.
the surface (bias voltage applied in-between), electrons can leave one material and pass to
another one. Such flow of electrons is called then tunneling current and its intensity depends
on the tip/surface gap. Now, by keeping the tunneling current (and as a result this gap)

5
6 Chapter 1. State-of-the-art

Figure 1.1: 3D Scanning tunneling control system.

(a) (b)

Figure 1.2: STM/AFM applications examples [Abramovitch et al., 2007]: (a) STM image of
a gold surface. (b) 3D AFM image of blood cells.

constant using proper control, and by simultaneously moving the tip along the surface hori-
zontally using piezoelectric actuators, the tip is forced to follow the topographical variations
(or more precisely the contour of constant density of states) of the surface. The amount of this
variations (topographical/electrical) can be retrieved via control signal of the vertical piezo
[Abramovitch et al., 2007].

The Atomic Force Microscope (AFM) [Binnig et al., 1986], which unlike STM uses position
sensor instead of tunneling current sensor, became even more popular, since it applies also
to non-conductive surfaces, thus more focus has been devoted to it in the literature. It has
fundamentally changed research in such areas as biology [Zou et al., 2004], chemistry [Jandt
et al., 2000], physics [Tsukada et al., 2000], materials science [Yamanaka et al., 1999] and
micro-/nanorobotics [Xie et al., 2012].
1.3. Tools, challenges and associated control issues 7

Other applications where the nanopositioning plays very important role include Hard Disk
Drives (HDDs) [Hu et al., 1999], where the read/write head is positioned over data tracks or
nanolithography [Gentili et al., 1993]. Fig. 1.2 provides a couple of STM/AFM illustrative
applications.

1.3 Tools, challenges and associated control issues

1.3.1 Tunneling current phenomenon

In metals the electrons fill the energy levels up the the particular top energy known as the
Fermi level EF as shown in Fig. 1.3a. If two conductive materials (here the tip and the surface)
are brought close enough to each other, there is a small vacuum gap between them which is
a potential barrier. The electrons of one metal can leave it, overcome an energy barrier and
travel to the other metal only if an additional amount of energy, so-called work function φ, is
delivered to them. Both, the tip and the surface without any external supply have the same
electrical potential and therefore their Fermi levels are aligned. Biasing the surface w.r.t. the
tip with the potential Vb , as shown in Fig. 1.3b, decreases the Fermi level of the latter by
the energy eVb . Such a difference between the Fermi levels of the surface EFs = EF and of
the tip EFt = EF − eVb results in filled energy states in the former and empty energy states
in the latter. The filled/empty states force the electrons to jump from one material to the
other one and such a flow of electrons is called tunneling current. The direction of tunneling
depends on the polarity of the constant bias voltage Vb (i.e. if the negative electrode of the
supply is connected to the tip, its Fermi level will increase w.r.t. the one of the surface and
the tunneling direction will be reversed).

In classical mechanics, if a particle of total energy E accounters a potential barrier of energy

V > E it cannot pass through it and will be reflected. In case of quantum mechanics, the
wave–particle duality concept states that every elementary particle (in this case the electron)
can be treated not only as a particle but also as a wave and even if its total energy is smaller
than the potential energy of the barrier there is still a non-zero probability that it may pass
this forbidden region and reappear on the other side as shown in Fig. 1.4. Mathematically,
the electron state can be described by referring to the following time-independent Schrödinger
equation in one dimension (z in this case), that describes the stationary states (indeed in
atoms the potential on the orbits does not change with time):
~2 d2
− ψ(z) + V (z)ψ(z) = Eψ(z) (1.1)
2m dz 2
where ψ(z) is the wave function describing the wave-like behavior of the electron, E is its
total energy, which equals to the sum of the potential energy V (z) and the kinetic energy
(described by the first term in equation (1.1)) and ~ = 1.05 · 10−34 J · s is a reduced Planck
constant. Assuming the square potential barrier of length d, described as follows:

V0 , 0 < z < d
V (z) = (1.2)
0, otherwise
8 Chapter 1. State-of-the-art

energy 𝐸
𝐸𝐹 + 𝜙
potential
work
barrier
function (𝜙)
𝐸𝐹𝑠 = 𝐸𝐹 𝐸𝐹𝑡 = 𝐸𝐹
𝐸𝐹
Fermi level (𝐸𝐹 )

0
surface tip
(a)

energy 𝐸
𝐸𝐹 + 𝜙 − 𝑒Vb
potential tunneling
𝐸𝐹𝑠 = 𝐸𝐹 barrier electron
filled empty
} 𝑒Vb }
states states
𝐸𝐹 − 𝑒Vb
𝐸𝐹𝑡

surface bias
voltage (Vb )
0
tip

(b)

Figure 1.3: Tunneling effect: (a) Unbiased tip/surface. (b) Surface biased w.r.t. the tip.

the space z is divided into three regions as shown in Fig. 1.4: region I (z < 0), region III
(z > d), both outside the barrier, with zero potential energy and region II (0 < z < d) inside
the barrier with constant potential energy V0 . For each region one searches the solution of
(1.1).
Outside the barrier (in regions I and III) E − V (z) > 0. From (1.2), V (z) = 0 and the total
energy of electron is equal to its kinetic energy. Equation (1.1) takes the following form:

~2 d2
ψ(z) + Eψ(z) = 0 (1.3)
2me dz 2
This linear ordinary differential equation can be expressed using Laplace operator as follows:
 2 
~ 2
s + E ψ(s) = 0 (1.4)
2me

which has the following solutions:

√
2me E
s1,2 = ±i (1.5)
~
1.3. Tools, challenges and associated control issues 9

energy 𝐸
𝑉0
potential
𝜙 − 𝑒𝑉𝑏 barrier
𝐸𝐹𝑠 = 𝐸𝐹 tunneling
𝑒𝑉𝑏 electron
𝐸𝐹𝑡

II
(𝐸 < 𝑉)
𝑧
0
I (𝐸 > 𝑉) surface III (𝐸 > 𝑉) tip
𝑧=0 𝑧=𝑑

Figure 1.4: Electron behaviour in three energy regions during tunneling effect.

Hence, the final solution for the wave function in regions I and III is a linear combination of
solutions (1.5) as follows:

ψI (z) = A1 es1 z + A2 es2 z = A1 eik0 z + A2 e−ik0 z (1.6)

ψIII (z) = C1 es1 z + C2 es2 z = C1 eik0 z + C2 e−ik0 z (1.7)

√
where k0 = 2m ~
eE
is a wave number outside the barrier. One can notice that the solutions
(1.6) and (1.7) represent oscillating waves.
Inside the barrier (region II) E − V (z) < 0. From (1.2), V (z) = V0 and equation (1.1) takes
the following form:
~2 d2
ψ(z) + (E − V0 ) ψ(z) = 0 (1.8)
2me dz 2
or in Laplace domain:  2 
~ 2
s + E − V0 ψ(s) = 0 (1.9)
2me
The following solutions are obtained:
p p
2me (E − V0 ) 2me (V0 − E)
s1,2 = ±i =∓ (1.10)
~ ~
This gives the following solution for the wave function inside the barrier:

ψII (z) = B1 es1 z + B2 es2 z = B1 e−k1 z + B2 ek1 z (1.11)

√
2m (V −E)
e 0
where k1 = ~ is a wave number inside the barrier. One can notice that the solution
(1.11) represents a combination of an exponentially increasing and exponentially decaying
terms with decaying rate k1 . Only the decaying term is taken into account for further consid-
erations, since it corresponds to the tunneling through the barrier.

Knowing the wave function ψ(z) allows one to calculate the probability density for the
electron to be found at the specific location z. This probability is proportional to the wave
10 Chapter 1. State-of-the-art

function squared |ψ(z)|2 . We are interested in the case when the tunneling occurs, i.e. we
want to assess the probability that the electron at position z = 0 can cross the barrier and be
found at position z = d. Taking only the decaying part of equation (1.11), ψII (z) = B1 e−k1 z
(where B1 = ψII (0)), this probability can be expressed as follows:

P ∝ |ψII (d)|2 = |ψII (0)|2 e−2k1 d (1.12)

In an unbiased case, like in Fig. 1.3a, V0 − E = φ and in case of tunneling effect when two
metals are biased w.r.t. each other, like in Fig. 1.3b, V0 − E = φ − eVb and when the √ bias
voltage Vb is chosen to be small enough (such that eVb << φ), the wave number k1 ≈ 2m ~
eφ

(when the materials for the tip and the sample surface differ electrically, the work function is
assumed to be an average of the work functions corresponding to each material).
The tunneling current can be related to the density of empty states of the tip or filled states
of the surface (see Fig. 1.3b). The higher the number of these states, the higher value of the
tunneling current. Assuming that after biasing EFs = EF is the Fermi level of the sample
surface, the states filled by electrons in the surface lie on the energy levels between EF − eVb
and EF , hence the electrons will flow to the empty states of the tip between the same energy
levels EFt = EF − eVb and EF . The formed tunneling current is proportional to the sum of
the probabilities (1.12) taken over this energy range as follows:
EF √
X 2me φ
it ∝ |ψII (0)|2 e−2 ~
d
(1.13)
EF −eVb

Defining the local density of states (LDOS) near some energy region E in an interval  and at
the position z as:
E
1X
ρLS (z, E) = |ψ(z)|2 (1.14)

E−

and assuming the continuity of the wave function (i.e. ψ(0) = ψI (0) = ψII (0) and ψ(d) =
ψII (d) = ψIII (d)) the tunneling current (1.13) can be rewritten in terms of LDOS (filled) in
an interval  = eVb near the Fermi level of the sample surface EF at the surface location z = 0
as follows:
EF √ √
1 X 2me φ 2me φ
it ∝ eVb |ψII (0)|2 e−2 ~
d
= eVb ρLS (0, EF )e−2 ~
d
(1.15)
eVb
EF −eVb

or in terms of LDOS (empty) in the same interval  = eVb and again near the Fermi level of
the sample surface EF but at the tip location z = d:

it ∝ eVb ρLS (d, EF ) (1.16)

√ √
Defining a constant g = eρLS (0, EF ) and k = −2 2m ~
e
φ, which depends on the metal
electrical properties, the tunneling current equation is given by:

it = gVb e−kd (1.17)

One can see that it depends not only on the distance d, but also on the bias voltage Vb and
the electrical properties of the used surfaces (φ). Since the principle of STM device is based
1.3. Tools, challenges and associated control issues 11

on the tunneling current phenomenon, in order to have a reliable images of LDOS (which may
approximate the surface topography) one has to keep the bias voltage Vb constant and the
scanned surface must be composed of only one material in order to obtain constant effective
work function φ.

1.3.2 Piezoelectric actuators - modeling and control

Due to their high resolution, high stiffness and fast response piezoelectric actuators are widely
used in micro-/nanoscale applications such as in STM [Binnig and Rohrer, 1986] or AFM
[Abramovitch et al., 2007] and other nanopositioning applications [Devasia et al., 2007], [Se-
bastian and Salapaka, 2005]. However, they exhibit some adverse phenomena of nonlinear
hysteresis, creep, structural vibration and cross-couplings between the axes shown in Fig. 1.5.
Overcoming these difficulties can substantially improve positioning precision, range of opera-
tion and increase the bandwidth.

output
output displacement [μm]
displacement [μm]

creep

input
voltage [V]
0.1
100
time [s]

(a) (b)

output
displacement [μm]
output
reference
input output
voltage (X) ≠ 0 piezoelectric displacement (X) ≠ 0
actuator ≠0
input output
voltage (Y) = 0 displacement (Y) ≠ 0
≠0
0 0.02 0.04
time [s]

(c) (d)

Figure 1.5: Adverse phenomena of piezoelectric actuators: (a) Hysteresis. (b) Creep. (b)
Structural vibration. (d) Cross-coupling (from X to Y axis).
12 Chapter 1. State-of-the-art

Hysteresis.
Hysteresis is a nonlinear phenomenon between the applied input voltage and the output dis-
placement of piezoelectric actuator as shown in Fig. 1.5a [Bertotti and Mayergoyz, 2006]. The
displacement depends not only on the current input voltage but also on its past history (mem-
ory effect). Hysteresis is amplitude-dependent and is more visible in response to high input
voltages (in long-range positioning like during imaging large samples over large displacements).
Moreover, recent studies have shown that this effect is frequency-dependent, as well [Janaideh
et al., 2009b]. Uncompensated, it can significantly limit the accuracy (in case of positioning
the error can reach 15% of the piezo travel range or even go beyond 35% for high frequencies
[Gu et al., 2014b]), lead to high image distortions (in case of scanning applications) or even
be the reason of instability in a controlled system [Shan and Leang, 2012].

Many hysteresis models have been studied in the literature and generally they can be
divided into two categories: physics-based models and phenomenological models. As the
name suggests, the models from the first category are developed based on the physical rules.
The most common models in this group are the Jiles-Atherton model [Jiles and Atherton,
1986] (applied to the ferromagnetic hysteresis) and domain wall model [Smith and Ounaie,
2000] derived from the first one. The main drawback of these models is their complexity and
often limited applicability, since the model derived from the physical rules for one kind of
materials may not be valid for the other kind of materials governed by other physical rules.
Therefore, they are mostly replaced by the phenomenological models, which do not require
the physical insight into the modeling but use mathematical equations to describe hysteresis.
Phenomenological models based on differential equations such as Duhem model [Lin and Lin,
2012] Bouc-Wen model [Rakotondrabe, 2011] or Dahl model [Xu and Li, 2010] can be very
easily added to the dynamical state-space models describing the whole piezoactuated system,
though the general solution is quite difficult to obtain. The most popular phenomenological
models for hysteresis based on integral equations are so-called operator-based models, where
the hysteresis output is obtained as an integral of weighted elementary operators. The models
in this group include Preisach model, Krasnosel’skii-Pokrovskii (KP) model and Prandtl-
Ishlinskii (PI) model and the difference between them lies in the elementary operators they
use. The Preisach model uses a relay operator γ̂α,β [u(t)] and its output is a weighted integral
of many of these operators expressed as:
Z Z
y(t) = P [u](t) = µ(α, β)γ̂α,β [u(t)]dαdβ (1.18)
α≥β

with


 0, u(t) ≤ β
1, u(t) ≥ α







 γ̂α,β [u0 ], ∀τ ∈ [t0 , t], u(τ ) ∈ (β, α)
γ̂α,β [u(t)] = 0, if u(τ ) ∈ (β, α), ∃t1 ∈ [t0 , t), s.t. (1.19)
u(t1 ) = α and ∀τ ∈ (t1 , t] u(τ ) ∈ (β, α)








 1, if u(τ ) ∈ (β, α), ∃t1 ∈ [t0 , t), s.t.
u(t1 ) = β and ∀τ ∈ (t1 , t] u(τ ) ∈ (β, α)

1.3. Tools, challenges and associated control issues 13

where α and β are the thresholds such that α ≥ β and µ(α, β) is called the density func-
tion. The Preisach model is the oldest one among the operator-based models and is widely
recognized owing to its general structure, but due to double integrals appearing in (1.18)
the identification procedure becomes quite complicated. Moreover, this model due to dis-
continuous nature of the relay operator is not analytically invertible and thus the numerical
approximations are usually used for implementation of the inverse compensators. On the other
hand, Prandtl-Ishlinskii model uses so-called play operator Fr [u](t) and its output is again a
weighted integral of these operators expressed as:
Z R
y(t) = Π[u](t) = p0 u(t) + p(r)Fr [u](t)dr (1.20)
0

with
y(0) = Fr [u](0) = fr (u(0), 0)
y(t) = Fr [u](t) = fr (u(t), y(ti ))
(1.21)
for ti < t ≤ ti+1 , 0≤i≤N −1
fr (u, y) = max(u − r, min(u + r, y))

where r is the model threshold, p(r) is a density function and p0 > 0 and R is the up-
per integration limit. The Prandtl-Ishlinskii model can be considered as a subclass of the
Preisach model, since it was derived from the latter. The advantage of the PI model w.r.t.
the Preisach model is that it is simpler and it is analytically invertible, hence it can be used
in real-time applications without any numerical approximations. The drawback of this model
is that due to its symmetry, it cannot describe the asymmetric hysteresis loops. Therefore,
several Modified/Generalized Prandtl-Ishlinskii (MPI/GPI) models for asymmetric and sat-
urated behavior have been reported in literature like dead-zone operators cascaded with the
play operators [Kuhnen, 2003], the model with asymmetric backlash as an elementary operator
[Dong and Tan, 2009] or the play operators based upon hyperbolic-tangent envelope functions
[Janaideh et al., 2009a]. All of these models assume that the hysteresis effect does not depend
on the rate of input signals, though as it was mentioned before, in reality it is rate-dependent.
Therefore, several works have emerged which include the Preisach model with rate-dependent
density function [Mayergoyz, 1988] and Prandtl-Ishlinskii model where the rate-dependence
is included in both the threshold and the density function [Janaideh et al., 2008]. In [Ang
et al., 2007], the model used in [Kuhnen, 2003] has been extended to the rate-dependent
Modified Prandtl–Ishlinskii model, where the slope of the hysteresis was modeled (based only
on the experimental data) as a linear function of the rate of the control input. Other hys-
teresis models include polynomial-based models [Sun et al., 2004], [Bashash and Jalili, 2008],
ellipse-based models [Gu and Zhu, 2010], [Gu and Zhu, 2011] and intelligent modeling such as
using neural networks [Chuntao and Yonghong, 2004], [Dong et al., 2008] or fuzzy systems [Li
et al., 2013b]. In [Cruz-Hernandez and Hayward, 2001] the hysteresis is modeled as a phase
lag between input and output.
14 Chapter 1. State-of-the-art

𝑘𝑐1 𝑘𝑐2 𝑘𝑐𝑛 𝑦

𝑘𝑐0
…
𝑢
𝑐𝑐1 𝑐𝑐2 𝑐𝑐𝑛

Figure 1.6: Viscoelastic creep model [Croft et al., 2001].

Creep.
Another adverse phenomenon present in piezoelectric actuators, which also leads to the loss
of precision, is a creep phenomenon (i.e. the piezo can drift when positioned over extended
periods of time [Abramovitch et al., 2007] and slow-speed operations) shown in Fig. 1.5b. Two
most popular models reported in the literature are the nonlinear logarithmic model [Jung and
Gweon, 2000] and the linear viscoelastic model [Croft et al., 2001]. The first model assumes
that the creep response has a logarithmic shape over time and can be represented by the
following equation:   
t
y(t) = y0 1 + γlog10 (1.22)
t0
where y(t) is the response of piezoelectric actuator to a fixed input voltage and the parameters
t0 , y0 and γ are practically identified from the experimental data, t0 refers to the time at which
the creep effect is obvious (after applying the input voltage and after oscillations vanishing
in the output response, from Fig. 1.5b t0 ≈ 0.1 s), y0 is the displacement at time t0 and γ
determines the rate of the logarithmic response. The linear model of creep, as proposed in
[Croft et al., 2001] consists of a series connection of springs and dampers as shown in Fig. 1.6
and can be described in Laplace domain as follows:
n
y(s) 1 X 1
Gcreep (s) = = + (1.23)
u(s) kc0 cci s + kci
i=1

where u(s) and y(s) are the measured response and input voltage of piezoelectric actuator in
Laplace domain, respectively, kc0 models low frequency elastic behavior (but not creep itself)
of the piezo, kci is the spring constant of each spring, cci is the damping constant of each
damper and n is the number of spring-damper elements. The advantage of this model over
its logarithmic counterpart is that, unlike the latter, its output does not become unbounded
for sufficiently small (t → 0) or large (t → ∞) time [Devasia et al., 2007], [Gu et al., 2014b].
Moreover, both y0 and γ depend on t0 and the identification of parameter γ of model (1.22)
is difficult to determine (even for a fixed time t0 ). Since the creep occurs at low frequencies,
it is usually not taken into account during fast operations.

Structural vibration.
In STM/AFM applications high resolution of the scanned images is often required, how-
ever, the scanning procedure can take sometimes several minutes or even fractions of an hour
[Abramovitch et al., 2007]. This time can be shortened when the scanning frequency is in-
creased. However, fast scanning input signals with high frequency components (especially
1.3. Tools, challenges and associated control issues 15

triangular waveforms, which include odd harmonics) may excite lightly-damped resonance
peaks of the positioning device. This resonant nature of piezoelectric actuators causes an-
other adverse phenomenon called structural vibration (see Fig. 1.5c or the inset of Fig. 1.5b),
which leads to high positioning errors or even instability and is the main phenomenon lim-
iting high-speed operation of the nanopositioning devices. Therefore without proper control
the scanning speed is often limited from 1% up to 10% of piezo’s first resonant frequency
[Moheimani, 2008]. The vibration modeling mainly consists of finding a linear model, after
hysteresis phenomenon (or without taking it into account). The structure of the considered
dynamical system is usually quite uncertain or not known a priori, hence only the input
and output data sets are taken into account for the identification procedure, which can use
such techniques/tools as MATLAB System Identification Toolbox [Li et al., 2013a], sub-space
state space technique [Yong et al., 2010b] or using dynamic signal analyzer [Croft and Deva-
sia, 1999]. These methods very often lead to high order models and therefore proper model
reduction is looked either for the obtained model only (before the controller design), for the
controller model only (after controller design), or in both cases.

Cross-couplings.
Another adverse phenomenon exhibited by piezoelectric actuators is the cross-coupling be-
tween different axes, observed as a nonzero displacement in the axis other than the excited
one, as shown in Fig. 1.5d. This effect increases with both the increased frequency and am-
plitude of the input signals. In STM/AFM applications in the case when the raster scan is
used, one horizontal axis serves as the fast one (typically excited with a triangle input) and
the other one as the slow one (excited with a ramp input). In this case the coupling-based
positioning errors from the faster axis to the probing vertical one is more significant than from
the slower one [Tien et al., 2004]. The coupling effect is one of the major limitations for high
speed imaging in STM/AFM devices and can lead not only to large topography distortions
but can even damage the cantilever/tip probe and/or the sample [Shi et al., 2009]. Moreover,
the cross-couplings between the horizontal X and Y axes are present as well [Yong et al., 2010b].

Control of piezoelectric actuators.

A large number of works has been devoted to overcoming the above-mentioned phenomena in
order to improve positioning accuracy, increase range of operation and bandwidth. Generally
they can be classified into open-loop feedforward and closed-loop feedback methods, or com-
bination of the two as shown in Fig. 1.7. In the feedforward control, models of the addressed
phenomena (such as hysteresis, creep or vibration) are used in order to compensate for them,
by cascading their corresponding inverses one after another in the feedforward path of the
system [Croft et al., 2001] while in the feedback control the nonlinear hysteresis or creep are
assumed as disturbances over a linear system. Inverse-based methods, though computationally
intensive, are common solutions in cases where sensors are not available. On the other hand,
closed-loop techniques are accurate and need not model inversion, but the drawback is that
they require (sometimes expensive) sensors for feedback control. The combination of both,
feedforward and feedback control, usually gives better results than feedforward or feedback
alone, since it allows to take advantage of both approaches - the feedforward control ensures
good tracking, while the feedback control robustness w.r.t. modeling errors of the inverse
16 Chapter 1. State-of-the-art

models, parameter variations as well as disturbances of different nature [Butterworth et al.,

2008].

Piezoelectric actuator
𝑟 Feedforward 𝑢 Nonlinear Linear 𝑦
controller Hysteresis Dynamics

(a)

Piezoelectric actuator
𝑟 Feedback 𝑢 Linear 𝑦
Nonlinear
+-
controller Hysteresis Dynamics

(b)

Feedforward-feedback controller Piezoelectric actuator

𝑟 Feedback Feedforwad 𝑢 Linear 𝑦
Nonlinear
+- Dynamics
controller controller Hysteresis

(c)

Figure 1.7: Control architectures used to control piezoelectric actuator: (a) Feedforward con-
trol. (b) Feedback control. (d) Feedforward with feedback control.

There exists a vast amount of works on feedforward compensation of hysteresis phe-

nomenon and they are mainly based on the inverse model such as the operator-based Preisach
model [Hughes and Wen, 1997], [Tan and Bennani, 2008] the Prandtl-Ishlinskii (PI) model
[Kuhnen and Janocha, 2001], [Rakotondrabe et al., 2010] or its modifications/generalizations
for asymmetric hysteresis loops [Ang et al., 2007], [Janaideh et al., 2009a], [Gu et al., 2014a].
Adaptive inverses of Preisach and PI models can be found in [Tan and Baras, 2005] and
[Kuhnen and Janocha, 1999], respectively. For the differential-based models like Duhem or
Bouc-Wen model the inversion is either impossible or extremely difficult to obtain [Gu et al.,
2014b] and to overcome this problem, the feedforward compensator can be obtained like in
[Rakotondrabe, 2011], where so-called multiplicative-inverse structure is used, which does not
require constructing the model inverse explicitly. Moreover, since inverse PI model is also
PI-type, one can model the hysteresis inverse directly, from the experimental input/output
data, by inverting the input/output axes (i.e. by taking the input data as the output and
vice-versa) and fitting the PI model to such data. The hysteresis can be considered as a
nonlinear bounded disturbance and eliminated in the feedback loop as well using PID con-
1.3. Tools, challenges and associated control issues 17

trol [Leang and Devasia, 2007], robust H∞ [Chuang et al., 2011] approach or disturbance
observer (DOB) [Yi et al., 2009], [Ryba et al., 2014b]. In [El-Shaer et al., 2013] a Generalized
Prandtl-Ishlinskii (GPI)-based robust performance enhancement using DOB was studied and
illustrated in simulation. As reported in [Amin-Shahidi and Trumper, 2013] the hysteresis can
be almost completely avoided using charge amplifiers, which are an interesting alternative for
a voltage-based control. However, due to high complexity and cost of the circuit designs of
charge amplifiers, voltage-based control is mainly used.

The creep phenomenon can be compensated by the feedforward control using either inverse
model of (1.22) [Jung et al., 2001] or (1.23) [Croft et al., 2001] or it can be integrated with PI
model and canceled together with hysteresis by its inverse model [Krejci and Kuhnen, 2001].
It can be also assumed as a disturbance and eliminated by the feedback control, mainly PID
control [Leang and Devasia, 2007], or via disturbance observer together with hysteresis [Ryba
et al., 2014b]. In [Sun and Pang, 2006] post-correction of the AFM image has been done to
compensate for the effect of creep and hysteresis.

The open-loop feedforward control using the inverse model of vibration can be found for
example in [Croft et al., 2001]. As pointed in [Clayton et al., 2009], the exact vibration inverse
[Silverman, 1969] is possible when the dynamical model is minimum-phase (i.e. no zeros on
the open right half of the complex plane). In case of non-minimum systems a the standard
inverse would result in an unbounded feedforward input causing an internal instability in the
controlled system. One can inverse only the minimum-phase part and adjust for the overall
system gain or use for example more accurate Fourier-transform approach [Leang and Devasia,
2007] which leads to the optimal inversion w.r.t. cost function based on the trade-off between
the tracking accuracy and the control effort, similar to LQR designs. When the desired ref-
erence trajectory is known a priori Repetitive Control (RC) [Necipoglu et al., 2011] can be
used. In [Wu and Zou, 2007] an iterative control approach (IIC) was used for compensating
both, hysteresis and vibration of the piezo scanner during high-speed, large-range positioning.
The feedforwad approaches for vibration compensation which do not require model inversion,
include input-shaping methods as in [Schitter et al., 2008a] (where the scanning speed has
been increased by three orders of magnitude compared to conventional AFM) or notch filters
as reported in [Leang and Devasia, 2007], where both hysteresis and creep were eliminated
via high-gain feedback and the resonant peak of dynamic response was modified by the notch
filter, improving the stability margin of the system. The advantage of these methods is that
they do not require to model full system dynamics, but only need the values of the resonant
frequencies and their corresponding damping ratios. A simple proportional-integral-derivative
(PID) controller often fails in dealing with highly resonant positioners especially at high fre-
quencies. When charge control is used to reduce hysteresis phenomenon, an active damping
proved to be an effective way to cope with vibration effects such as Integral Resonant Con-
troller (IRC) [Bhikkaji and Moheimani, 2008] or Positive Position Feedback controller (PPF)
[Mahmood and Moheimani, 2009b]. The combination of inversion-based feedforward control
with active damping feedback control and charge actuation as in [Aphale et al., 2008] proved
to increase the tracking bandwidth in the presence of parameter variations such as resonance
frequencies. In [Habib et al., 2012], an LQG control is used as another damping technique
for lateral positioning of the AFM. However, pure LQG controller is not guaranteed to be
18 Chapter 1. State-of-the-art

robust, unlike LQR [Safonov and Athans, 1977]. Unfortunately, not all the states can be
available and the LQR controller has to work with a state observer. To recover nice proper-
ties of LQR design (at least 60◦ phase margin and 6dB gain margin), an LQG/LTR (Loop
Transfer Recovery) procedure was proposed by [Kwakernaak, 1972] and furtherly developed
by [Doyle and Stein, 1979], [Doyle and Stein, 1981]. This technique has been successfully
applied to suppress the vibration in piezo smart composite wing [Munteanu and Ursu, 2008],
at micro-/nanoscale for dual-stage actuators in HDDs [Hu et al., 1999], in vertical direction in
AFM [Yeh et al., 2008] as well as to reduce the vibration of lightly-damped resonant system
equipped with piezoelectric actuator [Ryba et al., 2014b]. In the recent years, a rapid growth
of modern robust control designs has been recorded in nanopositioning due to uncertainties
of different nature (different operating conditions, hysteresis, unmodeled dynamics etc.). In
[Sebastian and Salapaka, 2005] two kinds of robust controllers were tested experimentally: the
Glover–McFarlane loop-shaping and mixed-sensitivity H∞ controllers. In the first one, the
design task of meeting the performance specifications and robustness are separated into two
steps, which appears to be particularly attractive when the controller already exists and one
wants to robustify it. Moreover, the uncertainty does not need to be explicitly characterized
in this design. The second approach offers more degrees of freedom and achieves simultane-
ously performance and robustness and is used when the uncertainty model is available. In
[Halim and Moheimani, 2001] spatial resonant H2 controller applied to piezoelectric laminate
beam has been designed. Both bandwidth and scanning range can be increased significantly
by using a dual actuated system as reported in [Schitter et al., 2008b] or [Kuiper et al., 2010]
for AFM. The idea is to combine a long-range, low-bandwidth actuator with a short-range,
high-bandwidth actuator. The bandwidth can be also increased by using proper reference
scanning inputs like triangular waveform with cutted sharp top [Schitter et al., 2008a] or
other scanning trajectories including archimedean spiral [Mahmood and Moheimani, 2009a],
cycloid [Yong et al., 2010a] or Lissajous [Tuma et al., 2012] curves.

Finally, to deal with cross-coupling phenomena several works have been reported in the lit-
erature. In [Tien et al., 2004] and [Shi et al., 2009], the Inversion-based Iterative Control (IIC)
method is proposed to cancel the cross-coupling phenomena from the horizontal X and Y axes
to the vertical Z axis in AFM application. The control of cross-couplings between horizontal X
and Y axes of nanopositioning stages have been studied in [Mahmood and Moheimani, 2009c],
where both vibration and cross-couplings have been tackled via Positive Position Feedback
(PPF) control. In [Das et al., 2012], the experimental implementation of the multi-variable
resonant controller is used to reduce both vibration and cross-couplings between the axes of
the piezoelectric tube scanner of AFM. In [Yong et al., 2010b] a parallel flexure-based nanopo-
sitioner based on the compliant mechanism is developed to minimize the cross-coupling effects
with H∞ control to furtherly eliminate these adverse phenomena for fast raster scanning.

1.3.3 Sensors and measurement noises

As mentioned previously, the closed-loop techniques provide an accurate performance, since

they assure robustness w.r.t. model uncertainties, disturbances of different nature and system
parameter variations. This would not be possible without a feedback from the measured system
1.4. 3D operation 19

outputs. Both, the speed and positioning accuracy crucially depend on the sensing mechanism
used in the system [Devasia et al., 2007]. A variety of sensors are used in nanopositioning
and differ from their sensing techniques, including inductive [Brinkerhoff and Devasia, 2000],
capacitive [Horowitz et al., 2004], piezoresistive [Pedrak et al., 2003] and optical [Schneir et al.,
1994] ones. The optical sensors such as laser interferometer [Alexander et al., 1989] are often
quite expensive, but they provide high accuracy. On the other hand, cheaper alternatives,
such as strain gauge sensor (which measures the strain of the object, based on change of its
resistance), do exist. Their resolution is much lower than the one of the optical sensor and
they can even bring non-minimum phase behavior into the controlled system, resulting in a
serious bandwidth limitation [Schitter et al., 2002].

The most popular sensor used to measure displacements in the horizontal scanning axes
of STM/AFM is the capacitive sensor. It is based on a simple technique of non-contact
measurement, based on the change in capacitance between the fixed and movable electrode
(which depends on the distance between the two of them) converted by an electronic circuitry
to voltage units [Devasia et al., 2007]. In the vertical direction of STM, the quantum effect
of tunneling current described in section 1.3.1 appears when the distance between the two
conductive materials is less than 1 nm, which is very difficult to maintain. Due to very
small value of tunnel current (order of nano-amperes), a high-gain sensor is required (109
[V/A]) to capture and amplify it in order to be properly controlled. Among the noises of
different nature [Blanvillain, 2010], the measurement noise introduced by this lab-made high-
gain sensor is the main reason of low signal-to noise ratio at such sub-angstrom scale. The
tunneling phenomenon allows to measure the displacements of electrodes with a very high
precision (several angstroms), but limits the measurement range to 1nm, and [Blanvillain,
2010], [Blanvillain et al., 2009] [Blanvillain et al., 2014] proposed a new approach of using the
tunnel current to precisely control the motion of a cantilever beyond the range of 1 nm by
means of electrostatic actuation of the latter. So called pull-in effect has been observed i.e.
when the tunneling tip is approached to the freely vibrating part of the cantilever sufficiently
close, the proximity Casimir and van der Waals forces attract the latter towards the tip. An
electrostatic actuation with proper control was used to generate the electrostatic effect which
counteract these forces [Blanvillain, 2010], [Blanvillain et al., 2008].

1.4 3D operation

The aim of this work is to control the experimental tunneling current-based nanopositioning
platform in three dimensions. In [Merry et al., 2008], [Habib et al., 2012] or [Das et al., 2012]
3D approaches for AFMs are described, but to the author’s knowledge, there is a lack of
publications devoted to 3D control of STMs. While the horizontal control of STM and AFM
is similar, the principle of work in vertical axis is completely different. The most considered
cases are 1D or 2D control. In previous works it was in particular considered the problem of 1D
robust digital control in vertical Z direction in experiments by pole placement with sensitivity
functions shaping [Ahmad et al., 2012b], and H∞ design (see [Ahmad et al., 2010], [Ahmad
et al., 2012a]), as well as in horizontal X direction [Ryba et al., 2014b] using observer-based
20 Chapter 1. State-of-the-art

hysteresis compensation with vibration reduction by LQG/LTR technique. In [Ahmad et al.,

2012a] also 2D extension is proposed taking into account cross-coupling phenomenon but only
in simulation. Most popular approaches use a SISO strategy (three controllers, one for each
axis). Its performance can yet be deteriorated in case of large model uncertainties, sensor noises
and other external disturbances. In this manuscript a comparison of 3D SISO PID and MIMO
LQI control of STM in simulation is presented with the purpose of surface reconstruction
[Ryba et al., 2013]. The nanopositioning in three dimensions is tested in experiments, with
this difference that tunneling current is not kept on a constant level but tracks a prespecified
reference signal, playing the role of "virtual surface". The above mentioned phenomena of
hysteresis, creep, vibration, cross-coupling between the axes and tunneling nonlinearity are
tackled.

1.5 Conclusions

In this chapter the Scanning Probe Microscopy such as STM and AFM has been briefly
presented together with other nanopositioning applications. The importance, challenges and
difficulties arising from the control of such devices at such small scale have been discussed. The
tools (such as piezoelectric or electrostatic actuators and capacitive or current sensors) that
are used to control the physical laws considered in this manuscript have been presented with
the associated modeling and existing control approaches to account for the adverse phenom-
ena of piezoelectric actuators (such as nonlinear hysteresis, creep, vibration and cross-coupling
between the axes) in both horizontal and vertical directions, the exponential tunneling non-
linearity, as well as the measurement noise of the high-gain current sensor. Finally, a brief
look at the control of nanopositioning devices in 3 dimensions has been given. The considered
experimental setup, which is presented in the next chapter, is a micro-/nanopositioning device,
which shares some components with both STM (tunneling current phenomenon) and AFM
(cantilever). However, its principle of work is much more similar to the STM, and therefore
it will be sometimes called as STM-like device.
 Chapter 2

System description and modeling

Contents
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2 Experimental setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3 3D System modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.1 Horizontal X and Y axes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.2 Vertical Z axis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

2.1 Introduction

In this chapter, the experimental setup used throughout the work is briefly described and
novel 3D nonlinear model is introduced including horizontal X and Y axes, followed by vertical
Z axis. The source of its nonlinear behavior is due to: nonlinear hysteresis of piezoelectric
actuators, exponential dependence of tunneling current on the distance between two electrodes,
quadratic nonlinearity of the electrostatic actuator and nonlinear interaction force between the
tip and the surface.

2.2 Experimental setup

The researches in this thesis are carried out on tunneling current-based platform of GIPSA-
lab shown in Fig. 2.1. Three piezoelectric actuators are driven by control signals from 16-bit
D/A converter of the control card, amplified by a high voltage amplifier E-240-100 of gain
15 [V/V] and bandwidth 4 kHz. A platinum/iridium (Pt-Ir) tunneling tip is moved along
metallic surface in X and Y directions by piezoelectric actuator Tritor T-402-00 of gain 235
[nm/V] and bandwidth 630 Hz while much smaller and stiffer piezoelectric actuator of gain
1.2 [nm/V] and bandwidth 120 kHz moves the tip in Z direction as shown in Fig. 2.2. Two
capacitive sensors CS005 with capaNCDT 6500C device (gain 200 [V/mm] and bandwidth
8.5 kHz) are used to measure the displacements along X and Y axes. In the Z direction, the
distance between tip and surface (< 1 nm) is determined by the value of tunneling current
(nA), measured via high gain (109 [V/nA]) sensor. These three analog outputs are converted

21
22 Chapter 2. System description and modeling

Figure 2.1: Schema of the experimental setup.

into digital values by 16-bit A/D converter of the measurement card and used subsequently
by the real-time application working in Target PC with the sampling rate fs of 20 kHz. The
control algorithms and compensators are developed in Matlab & SimulinkTM software on a
development PC, and downloaded via Ethernet interface into a Target PC. Two acquisition
cards (one for control and one for measurement signals) connected with Target PC through
PCI bus are used together with two anti-aliasing Butterworth filters with cutoff frequency of
20 kHz.

Antialiasing Z Piezo (Z)

filters Y To current sensor
X
capaNCDT Tip and surface

Target PC
Current sensor
Voltage amplifier

Development PC

Y Capacitive
X Piezo (XY) Sensors
(a) (b)

Figure 2.2: Experimental setup of 3D nanopositioning platform: (a) General view. (b) The
heart of the platform.
2.3. 3D System modeling 23

2.3 3D System modeling

In this subsection a full 3D nonlinear model of the considered system, shown in Fig. 2.3 as
P lantxyz , is developed. In this scheme, the nonlinear parts are distinguished as dashed orange
blocks. More details about modeling, control or compensation of each specific subsystem are
given in the subsequent chapters.

0000 𝑃𝑖𝑒𝑧𝑜𝑥
𝑁𝐿𝑥 𝑑𝑥 𝑞𝑝𝑥
𝑢𝑥 𝑣𝑝𝑥 𝑥𝑥 𝑥𝑝 𝑦1
𝑣𝑥
𝑉𝐴𝑥 (𝑠) 𝛤𝑥 ++ 𝐷𝑝𝑥 (𝑠) ++ 𝐶𝑆𝑥 (𝑠) ++
𝑥𝑦
𝑛1
𝑌𝑥 (𝑠) 𝑋𝑦 (𝑠)

𝑦𝑥 𝑛2
𝑁𝐿𝑦 𝑑𝑦 𝑞𝑝𝑦
𝑢𝑦 𝑣𝑝𝑦 𝑦𝑦 𝑦𝑝 𝑣𝑦 𝑦2
𝑉𝐴𝑦 (𝑠) 𝛤𝑦 ++ 𝐷𝑝𝑦 (𝑠) ++ 𝐶𝑆𝑦 (𝑠) ++
𝑃𝑖𝑒𝑧𝑜𝑦 𝑃𝑙𝑎𝑛𝑡𝑥𝑦

𝑍𝑥 (𝑠) 𝑍𝑦 (𝑠) 𝑃𝑙𝑎𝑛𝑡𝑥𝑦𝑧

𝑧𝑥 𝑧𝑦 𝑛3
𝑑𝑧 +
𝑧𝑝
𝑢𝑧 𝑣𝑝𝑧 𝑞𝑝𝑧 +
+ 𝑑 𝑖𝑡 𝑣𝑧 𝑦3
𝑉𝐴𝑧 (𝑠) −𝐺𝑝𝑧 ++ 𝐷𝑝𝑧 (𝑠) +
𝑧𝑧 - 𝑃𝐿𝑡 ′ 𝐶𝑆𝑡 (𝑠) ++
𝑃𝑖𝑒𝑧𝑜𝑧 +
𝑃𝐿𝑡
𝑉𝑒 𝐹𝑒 𝐹𝑐
𝐸𝐴 +- 𝐶𝑁𝑇(𝑠) 𝑑0
𝑧𝑐 (𝑥)|𝑥=𝑥𝑓
𝐹𝑝
𝐼𝑁𝑇𝑡𝑐
𝑃𝑙𝑎𝑛𝑡𝑧

Figure 2.3: Schema of the full 3D nonlinar model.

2.3.1 ⏟
Horizontal X and Y axes
𝑧̂ 𝑐𝑒 (𝑥)|𝑥=𝑥𝑓 =−𝑏𝑉𝑒 2
A 2D scanning system of X and Y horizontal axes is shown in Fig. 2.3 as subsystem P lantxy .
The horizontal piezoelectric actuators in X and Y axes are used to move the tunneling tip
along the surface. As pointed in the previous chapter they are widely used in nanopositioning
applications due to their high resolution, high stiffness and fast response. However, they
24 Chapter 2. System description and modeling

exhibit nonlinear hysteresis, creep, vibration and cross-coupling between the axes. Now, taking
into account all the above-mentioned adverse phenomena, the piezoelectric actuator for the X
(resp. Y axis) can be modeled as follows (see also Fig. 2.3):


 qpx (t) = N Lx [vpx ](t) = Γx [vpx (t)] + dx (t)
 

 ẋpx (t) = Apx xpx (t) + Bpx qpx (t)
 Dpx (s) :


x (t) = Cpx xpx (t) + Dpx qpx (t)
P iezox :  x (2.1)
 ẋ xy (t) = Axy xxy (t) + Bxy yy (t)

 Xy (s) :



 xy (t) = Cxy xxy (t) + Dxy yy (t)
xp (t) = xx (t) + xy (t)




 qpy (t) = N Ly [vpy ](t) = Γy [vpy (t)] + dy (t)
 

 ẋpy (t) = Apy xpy (t) + Bpy qpy (t)
 Dpy (s) :


xx (t) = Cpy xpy (t) + Dpy qpy (t)
P iezoy :  (2.2)
 ẋ yx (t) = Ayx xyx (t) + Byx xx (t)

 Yx (s) :



 yx (t) = Cyx xyx (t) + Dyx xx (t)
yp (t) = yy (t) + yx (t)


or in Laplace domain:


 qpx (s) = N Lx [vpx ](s) = Γx [vpx ](s) + dx (s)
xx (s) = Dpx (s)qpx (s)

P iezox : (2.3)
 xy (s) = Xy (s)yy (s)


xp (s) = xx (s) + xy (s)


 qpy (s) = N Ly [vpy ](s) = Γy [vpy ](s) + dy (s)
yy (s) = Dpy (s)qpy (s)

P iezoy : (2.4)
y (s) = Yx (s)xx (s)
 x


yp (s) = yy (s) + yx (s)
where N Lx [vpx ](t) (resp. N Ly [vpy ](t)) is a nonlinear part which consists of the model of static
nonlinear hysteresis Γx [vpx ](t) (resp. Γy [vpy ](t)) (more details on how to modelize hysteresis
are given in chapter 3) and a slowly varying disturbance dx (t) (resp. dy (t)), which gathers the
effect of creep, unmodeled hysteresis and other low-frequency disturbances. This static (or low
frequency) part of piezo is followed by the direct dynamics Dpx (s) (resp. Dpy (s)) and cross-
coupling Xy (s) (resp. Yx (s)), where xx (t) (resp. yy (t)) is a direct contribution to the piezo
displacement xp (t) (resp. yp (t)) and xy (t) (resp. yx (t)) is a contribution due to cross-coupling
phenomenon. The intermediate variables qpx (t) (resp. qpy (t)) link the nonlinear and linear
parts of piezoelectric actuators. The piezo gain Gpx (resp. Gpy ) is hidden in the hysteresis
models Γx [vpx ](t) (resp. Γy [vpy ](t)), thus the direct dynamical parts have unit static gain
(i.e. Dpx (0) ≈ 1 and Dpy (0) ≈ 1). For the quasi-static behavior (e.g. for the design of the
disturbance observer for hysteresis and creep compensation, described in chapter 3), a simple
second order model Dpx (s) = ωpx 2 /(s2 + 2ξ ω s + ω 2 ) is used (where ξ
px px px px is the damping
ratio and ωpx the undamped natural frequency of piezo). However, for the operation at high
frequencies a full high-order black box dynamical model is needed and identified as shown in
chapter 4.
2.3. 3D System modeling 25

The voltage amplifier which feeds piezoelectric actuator in X (resp. Y) axis can be de-
scribed by a first order system as follows:

ẋ1,vx (t) = −ωvx x1,vx (t) + ux (t)

(2.5)
vpx (t) = Gvx ωvx x1,vx (t)

ẋ1,vy (t) = −ωvy x1,vy (t) + uy (t)

(2.6)
vpy (t) = Gvy ωvy x1,vy (t)

or in Laplace domain:
vpx (s) Gvx ωvx
V Ax (s) = = (2.7)
ux (s) s + ωvx

vpy (s) Gvy ωvy

V Ay (s) = = (2.8)
uy (s) s + ωvy

where ux (s) (resp. uy (s)) and vpx (s) (resp. vpy (s)) are the Laplace transforms of the amplifier
input and output, and Gvx , ωvx (resp. Gvy , ωvy ) are its gain and bandwidth, respectively.

Finally, the piezo displacements xp (t), yp (t) are converted into the voltage units by two
capacitive sensors which can be modeled as follows:

ẋ1,capx (t) = −ωcapx x1,capx (t) + xp (t)

(2.9)
vx (t) = Gcapx ωcapx x1,capx (t)

ẋ1,capy (t) = −ωcapy x1,capy (t) + yp (t)

(2.10)
vy (t) = Gcapy ωcapy x1,capy (t)

or in Laplace domain:
vx (s) Gcapx ωcapx
CSx (s) = = (2.11)
xp (s) s + ωcapx

vy (s) Gcapy ωcapy

CSy (s) = = (2.12)
yp (s) s + ωcapy

where vx (s) (resp. vy (s)) is the Laplace transform of the capacitive sensor output vx (t) (resp.
vy (t)), and Gcapx , ωcapx (resp. Gcapy , ωcapy ) are its gain and bandwidth, respectively. The
final readout from the sensors is noised in practice as follows:

y1 (t) = vx (t) + n1 (t) (2.13)

y2 (t) = vy (t) + n2 (t) (2.14)

where n1 (t) and n2 (t) are the noises of the measured outputs y1 (t) and y2 (t), respectively.
26 Chapter 2. System description and modeling

𝑣𝑝𝑧 𝑢𝑧

voltage
current amplifier
sensor
piezo

tunneling
𝑅
𝑦3 tip
𝑉𝑏
𝑖𝑡 surface
𝑧𝑝 (cantilever)
Z -𝐹𝑝 𝑑
𝑑0 𝑧𝑐
Y 𝑧𝑐 (𝐿)
X 𝑙0 𝑙 𝑧𝑐 (𝑥𝑓 )
𝐹𝑒
𝑥𝑓 𝑧𝑐
𝐿 ) 𝑧𝑐 (𝐿)
𝐹𝑝

electrostatic actuator
𝑉𝑒 𝐹𝑒 𝐿

Figure 2.4: Schema of the experimental setup in vertical (Z) direction.

2.3.2 Vertical Z axis

The scheme of the vertical system, referred in Fig. 2.3 as P lantz , is shown in Fig. 2.4. The
heart of the vertical system model is the tunneling phenomenon between the metallic tip and
the metallic surface. As mentioned in the previous chapter, practically, this phenomenon is
possible if the distance d(t) between the sharp tip and the clean surface is smaller than 1 nm,
therefore the aim is to maintain the tunneling gap d(t) within this distance. The model of
the tunneling current used in this manuscript (shown in Fig. 2.3 as a nonlinear block P L0t ) is
expressed as follows:

gVb e−kd(t) , 0 < d(t) ≤ 1 nm


P L0t : it (t) = (2.15)
0, d(t) > 1 nm
2.3. 3D System modeling 27

This distance can be expressed as (see also Fig. 2.3 and Fig. 2.4):

d(t) = d0 + zp (t) − zc (xf , t)) (2.16)

where d0 is a tip/surface initial distance, zp (t) is the actual tip position, zc (xf , t) represents
the cantilever displacement (see Fig. 2.3) at position xf (in chapter 6 in STM-like application,
surface variations zs (t) are used instead of cantilever displacement). Due to the exponential
term in (2.15), even small variations of the tip/surface distance d(t) greatly influence the
tunneling current (decreasing the gap by 1Å (0.1 nm) causes an increase of this current by an
order of magnitude and vice versa), which makes the control quite difficult and challenging
(an open-loop operation is not possible in this case). On the other hand, taking into account
the fact that a typical atomic diameter is of around 0.3 nm, such strong dependency on the
distance makes the value of the tunneling current dominated by the contribution between the
last atoms of the surface and the tip which in turn results in an atomic resolution of the
obtained images.

The piezoelectric actuator in the Z direction is used to approach/withdraw the tip vertically
to/from the scanned surface. Since the displacements in this direction are very small (order of
nanometers), the hysteresis phenomenon is negligible and the static part of piezo is simply its
gain Gpz . However, the creep is still present and it will be considered as a disturbance dz (t).
The model of the piezoactuator can be defined as follows:


 qpz (t) = −Gpz vpz (t) + dz (t)

 ẋ1,pz (t) = x2,pz (t)




2 x
ẋ2,pz (t) = −ωpz 2
D (s) : 1,pz (t) − 2ξpz ωpz x2,pz (t) + ωpz qpz (t)



 pz
 


 zz (t) = x1,pz (t)

P iezoz : ẋzx (t) = Azx xzx (t) + Bzx xx (t) (2.17)
 Zx (s) :
 z (t) = Czx xzx (t) + Dzx xx (t)
 x




 ẋzy (t) = Azy xzy (t) + Bzy yy (t)
Zy (s) :


zy (t) = Czy xzy (t) + Dzy yy (t)





zp (t) = zz (t) + zx (t) + zy (t)

or in Laplace domain:


 qpz (s) = −Gpz vpz (s) + dz (s)



 zz (s) = qpz (s)Dpz (s)
ω2


Dpz (s) = s2 +2ξpz ωpzpz s+ω2

P iezoz : pz (2.18)
 zx (s) = Zx (s)xx (s)



 zy (s) = Zy (s)yy (s)



zp (s) = zz (s) + zx (s) + zy (s)


where zz (s) is a direct contribution to the piezo displacement zp (s) and zx (s), zy (s) are the
contributions due to cross-couplings from X and Y axes, respectively. Dpz (s) is a vibrational
part of piezo, and ξpz , ωpz are piezo damping and bandwidth, respectively. The piezo moves
downwards (in the negative direction), when the input voltage vpz is increased, hence the
minus appears near the piezo gain Gpz in the first equation of (2.18).
28 Chapter 2. System description and modeling

Similarly to the horizontal axes, the voltage amplifier in the vertical axes, which feeds the
vertical piezo can be modeled as the following first order system:

ẋ1,vz (t) = −ωvz x1,vz (t) + uz (t)

(2.19)
vpz (t) = Gvz ωvz x1,vz (t)

or in Laplace domain:
vpz (s) Gvz ωvz
V Az (s) = = (2.20)
uz (s) s + ωvz
where uz (s) and vpz (s) are the Laplace transforms of the amplifier input and output, and Gvz ,
ωvz are its gain and bandwidth, respectively.

Since the value of the tunneling current is very small (order of magnitude of nA), a high-
gain current sensor is used to capture and amplify it. This sensor can be modeled as follows:

ẋ1,t (t) = x2,t (t)

ẋ2,t (t) = −ωt2 x1,t (t) − 2ξt ωt x2,t (t) + ωt2 Gt it (t) (2.21)
vz (t) = x1,t (t)

or in Laplace domain:
vz (s) Gt ωt2
CSt (s) = = 2 (2.22)
it (s) s + 2ξt ωt + ωt2
where it (s) and vz (s) are the Laplace transforms of the tunneling current it (t) and current
sensor output vz (t), respectively and Gt , ξt , ωt are its gain, damping and bandwidth, respec-
tively. The sub-nanometer scale causes high sensitivity of the tunneling current to the external
disturbances and noises. The signal to noise ratio in this direction is much smaller than in
the horizontal ones. The final measured output is expressed as follows:

y3 (t) = vz (t) + n3 (t) (2.23)

where n3 (t) is the noise of the measured output y3 (t).

Proximity and counteracting forces.

Due to small distance between the materials the inter-atomic proximity forces appear and
depending on this distance they can be either attractive or repulsive. This interaction is a
nonlinear function of the distance d between the tip and the surface and is shown in Fig. 2.3
as an internal physical feedback from this distance through the nonlinear function IN Ttc to
the cantilever input as follows García and San Paulo, 2000, Stark et al., 2004:
(
−HR/6d(t)2 , √ d(t) ≥ ap
Fp (t) = IN Ttc (d(t)) = 2 4 ∗ 3/2 (2.24)
−HR/6ap + 3 Eef f R(ap − d(t)) , d(t) < ap

where R is the tip radius, H the Hamaker constant. The value of ap separates the attractive
(d(t) ≥ ap ) and repulsive (d(t) < ap ) zones. In the attractive zone van der Waals forces
attract the tip versus the surface (it is assumed that the tip apex is spherical and the surface
is plate). When the distance tip/surface is decreased to the interatomic value ap , the adhesion
2.3. 3D System modeling 29

force appears in addiction to the pure van der Waals force according to the Dejarguin-Muller-
Toporov (DMT) theory Derjaguin et al., 1975, which models the repulsive zone. This adhesive
force successively decreases the effect of the van der Waals force and when the two forces are
equal, the atoms of the two materials are in mechanical contact and start to repeal each
other. The effective contact stiffness appearing in the adhesive force is given by Eef ∗
f =
2 2 −1
[(1 − vt )/Et + (1 − v )/E] , where vt and vc are the Poisson ratios and Et and E the elastic
moduli of the tip and the cantilever, respectively. During normal STM operation the tip is not
in contact with the surface, hence the range of tip/surface distances coincides with the range
of attractive force only (d(t) ≥ ap ). Due to the attractiveness of the proximity force during
normal STM operation, the minus appears in (2.24). However this force bends the cantilever
upwards (see Fig. 2.4) in the positive direction of the Z axis, hence it is taken with minus sign
in the scheme from Fig. 2.3.

In order to avoid a stuck between the tip and the cantilever, the proximity force should
be counteracted by another external force. To that end, a metalic plate is added and biased
w.r.t the base of the cantilever with the potential Ve (t). This electrostatic actuator EA (see
Fig. 2.3 and Fig. 2.4) bends the cantilever in opposite direction to the proximity force direction
(with minus sign) and together with the cantilever creates a capacitor with the capacitance
expressed as follows:
ε0 A e
C(l(t)) = (2.25)
l(t)
where Ae is the cantilever surface under the electrostatic plate and l(t) is the distance between
the two. The total energy of this capacitor is equal to the work needed to transport all the
charge from one plate to another and is given by:
1
W = C(l(t))Ve (t)2 (2.26)
2
Due to the potential difference between these two conductive materials an electrostatic force
appears which can be expressed as follows:
dW 1d 1 ε0 Ae
Fe (t) = = C(l(t))Ve (t)2 = − Ve (t)2 (2.27)
dl 2 dl 2 l(t)2

where l(t) = l0 + zc (xf , t) and l0 is the distance between the unbended cantilever (at zero
position) and the electrostatic actuator as shown in Fig. 2.4, zc (xf , t) is the total cantilever
displacement at the tip position xf , defined as the superposition of the displacements caused
by the proximity and the counteracted electrostatic forces as follows:

zc (xf , t) = zcp (xf , t) + zce (xf , t) (2.28)

where zcp (xf , t) > 0, since it has the direction of −Fp > 0 and zce (xf , t) < 0, since it has the
direction of Fe < 0. The total force that is acting on the cantilever is expressed as follows:

Fc (t) = −Fp (t) + Fe (t) (2.29)

30 Chapter 2. System description and modeling

Cantilever model.
In this thesis a multi-mode cantilever of length L with forced vibration and damping is adopted
like in Salapaka et al., 1997 to model the subsystem CN T (s) shown in Fig. 2.3. The cantilever
motion is govern by the following partial differential equation:

∂ 4 zc (x, t) ∂ 2 zc (x, t)
EI + pd (x, t) + ρA = −Fp (t) (2.30)
∂x4 ∂t2
The behavior of this model depends on both the position x and time t. For convenience of
notation, v 0 = dx
dv
will stand for the spatial derivative (here w.r.t. the x variable) and v̇ = dv
dt
time derivative of variable v. The interaction force is applied at the position xf from the base
of the cantilever (in this manuscript this force is the proximity force in the positive direction
−Fp (t) (2.24)). The value of the damping force pd (x, t) can be expressed as follows:

∂zc (x, t)
pd (x, t) = ξ − Fe (t) (2.31)
∂t
and is the sum of the natural structural cantilever damping force and the damping force that is
introduced to the system (here −Fe (t) (2.27)). Using (2.31) in (2.30) one obtains the following
equation:
∂ 4 zc (x, t) ∂zc (x, t) ∂ 2 zc (x, t)
EI + ξ + ρA = Fc (t) (2.32)
∂x4 ∂t ∂t2 | {z }
−Fp (t)+Fe (t)

The cantilever displacement can be expressed as a weighed sum of the time-dependent fun-
damental mode deformations qj (t) (corresponding to the j th mode), with spatially-dependent
weighting functions referred as the modal shape functions φj (x), as follows:
∞
X
zc (x, t) = φj (x)qj (t) (2.33)
j=1

where the modal shape function φj (x) is given by:

φj (x) = (sin λj L + sinh λj L)(cos λj x − cosh λj x)

(2.34)
+(cos λj L + cosh λj L)(sinh λj x − sin λj x)

with λj L being j th mode wavelength which satisfies the following equation:

cos λj L cosh λj L + 1 = 0 (2.35)

The modal shape function (2.34) satisfies the following formulas:

Z L
φj (x)2 dx = L(sin λj L + sinh λj L)2 (2.36)
0
Z L
(φ00j (x))2 dx = λ4j L(sin λj L + sinh λj L)2 (2.37)
0

The fundamental deformation qj (t) of j th mode is the solution of the following 2nd order
differential equation:
mj q̈j (t) + cj q̇j (t) + kj qj (t) = Fcj (t) (2.38)
2.3. 3D System modeling 31

where mj , kj , cj and Fcj (t) are respectively the mass, spring constant, damping coefficient
and force corresponding to j th mode. Using (2.36) and (2.37) they can be expressed as follows:
Z L
mj = ρA φj (x)2 dx = ρAL(sin λj L + sinh λj L)2 (2.39)
0
Z L
kj = EI (φ00j (x))2 dx = EIλ4j L(sin λj L + sinh λj L)2 (2.40)
0
Z L
cj = ξL φj (x)2 dx = ξL(sin λj L + sinh λj L)2 (2.41)
0
Fcj (t) = φj (xf )Fc (t) (2.42)

In state-space with standard state variables x2j−1,c (t) = qj (t), x2j,c (t) = q̇j (t) and using (2.42)
in (2.38), the j th mode can be expressed as follows:
 " #  " #
0 1 0

ẋ2j−1,c (t) x2j−1,c (t)
= k c + φj (xf ) Fc (t) (2.43)
ẋ2j,c (t) − mjj − mjj x2j,c (t) mj

Taking first nc modes (j = 1..nc ), the state space representation for the cantilever takes the
following form:
   
0

0 1 ··· 0 0
 
ẋ1,c (t) x1,c (t)
k c1 φ1 (xf )
 ẋ2,c (t)   − m − m
 1
··· 0 0 
x2,c (t)   m1

1 1    
  .. .. .. .. ..
  
.. = ..  ..   
. +  Fc (t)

 .   . . . . 
 . .
    
 ẋ2nc −1,c (t)   0 0 ··· 0 1  x2nc −1,c (t) 0
   
  
k nc cnc φnc (xf )
ẋ2nc ,c (t) 0 0 · · · −m nc
−m n
x2nc ,c (t) mnc
c
| {z } | {z }
Ac Bc
(2.44)

 
x1,c (t)
 x2,c (t) 
  ..

yc (t) = 1 0 · · · 1 0  0 ·Fc
 + |{z} (2.45)
 
.
| {z } 
 x2nc −1,c (t) Dc
Cc

x2nc ,c (t)
with 2nc -dimensional state vector:
   
x1,c (t) q1 (t)

 x2,c (t)  
  q̇1 (t) 

xc (t) =  .. ..
= (2.46)
   
 .   . 

 x2nc −1,c (t)   qnc (t) 
x2nc ,c (t) q̇nc (t)
or in compact way:
ẋc = Ac xc + Bc Fc
(2.47)
yc = Cc xc + Dc Fc
32 Chapter 2. System description and modeling

Finally, inserting qj (t) = x2j−1,c (t) into (2.33), the displacement of the cantilever with nc
modes can be expressed in state coordinates as follows:
nc
X
zc (x, t) = φj (x)x2j−1,c (t) (2.48)
j=1

Especially, when x = xf , one considers the cantilever displacement at the tip position (see
Fig. 2.4) and (2.48) takes the following form:
nc
X
zc (x, t)|x=xf = zc (xf , t) = φj (xf )x2j−1,c (t) (2.49)
j=1

Cantilever displacement can be expressed also in Laplace domain (assuming first nc modes,
zero initial conditions for qj (0) = q̇j (0) = 0, taking the Laplace transform of (2.38) and using
it in (2.33)) as follows:
nc c n
X 1 1 X φj (xf ) 1
zc (x, s) = φj (x) cj kj
Fcj (s) = φj (x) c kj
Fc (s) (2.50)
s2 + mj s + j s+
2 mj
j=1 mj s + mj j=1 mj mj

One can say that the modal shape function φj (x) allows to distribute in space the modal
contributions. It is a convenient way, since only this shaping function depends on position
and the other part (modal) depends only on time.
Finally, taking into account all the equations from this chapter, one can write the following
full nonlinear 3D state-space model which corresponds to the schema in Fig. 2.3:

−ωvx x1,vx (t) + ux (t)

 

  
 Apx xpx (t) + Bpx qpx (t) 

ẋ1,vx (t)  Axy xxy (t) + Bxy yy (t) 
 ẋ (t)    
 px   −ω capx x 1,capx (t) + x x (t) + x y (t) 

 ẋxy (t)  
   | {z } 
   xp (t) 
 ẋ1,capx (t)   
   −ωvy x1,vy (t) + uy (t) 
 ẋ1,vy (t)   

 ẋpy (t)  
  A py x py (t) + B py q py (t) 

A x (t) + B x (t)

yx yx yx x
   
 ẋ (t)  
yx
−ω x (t) + y (t) + y (t)
   
capy 1,capy y x 
 ẋ1,capy (t)  
   | {z } 
= yp (t)

 ẋ1,vz (t)  
  (2.51)
−ω x (t) + u (t)

vz 1,vz z
  
 ẋ1,pz (t)   
x2,pz (t)
  
 ẋ2,pz (t)   
  2 x 2 q (t) 


 ẋzx (t)    −ω pz 1,pz (t) − 2ξ ω x
pz pz 2,pz (t) + ω pz pz 
Azx xzx (t) + Bzx xx (t)
 
 ẋ (t)   
zy  
A x (t) + B y (t)
   
zy zy zy y
 ẋc (t)  
   
 ẋ1,t (t)  
   A x
c c (t) + B F
c c (t) 

x (t)

ẋ2,t (t)  2,t 
}  −ωt2 x1,t (t) − 2ξt ωt x2,t (t) + ωt2 Gt gVb e−kd(t) 
 
| {z
ẋ (t)
| {z }
3D
it (t)
| {z }
f (x3D (t),u3D (t),d3D (t))
2.4. Conclusion 33

   
y1 (t) Gcapx ωcapx x1,capx (t) + n1 (t)
 y2 (t)  =  Gcapy ωcapy x1,capy (t) + n2 (t) 
y3 (t) x1,t + n3 (t) (2.52)
| {z } | {z }
y3D (t) h(x3D (t),n3D (t))

 
u1 (t)    
 u2 (t)  dx (t) n1 (t)
 u3 (t)  ,
u3D (t) =   d3D (t) =  dy (t)  , n3D (t) =  n2 (t)  (2.53)
dz (t) n3 (t)
Ve (t)

where:

qpx (t) = Γx [Gvx ωvx x1,vx (t)] +dx (t) (2.54)

| {z }
vpx (t)

qpy (t) = Γy [Gvy ωvy x1,vy (t)] +dy (t) (2.55)

| {z }
vpy (t)

qpz (t) = −Gpz Gvz ωvz x1,vz (t) +dz (t) (2.56)
| {z }
vpz (t)

xx (t) = Cpx xpx (t) + Dpx qpx (t) (2.57)

yy (t) = Cpy xpy (t) + Dpy qpy (t) (2.58)
xy (t) = Cxy xxy (t) + Dxy yy (t) (2.59)
yx (t) = Cyx xyx (t) + Dyx xx (t) (2.60)
zx (t) = Czx xzx (t) + Dzx xx (t) (2.61)
zy (t) = Czy xzy (t) + Dzy yy (t) (2.62)

nc
X
d(t) = d0 + x1,pz (t) +zx (t) + zy (t) − φj (xf )x2j−1,c (t) (2.63)
| {z }
j=1
zz (t) | {z }
| {z } zc (xf ,t)
zp (t)

HR −ε0 Ae Ve2 (t)

Fc (t) = + (2.64)
6d(t)2 2(l + zc (xf , t))2
| {z } | 0 {z }
−Fp (t) Fe (t)

2.4 Conclusion

In this chapter, the full nonlinear 3D model has been developed for the considered tunneling
current-based platform. First, the horizontal motion in X and Y axes has been considered. The
34 Chapter 2. System description and modeling

model of piezoelectric actuators (driven by voltage amplifier and read by capacitive sensor)
including nonlinear hysteresis, creep, structural vibration and cross-coupling is given. Next,
the model in vertical Z direction includes the exponential behavior of the tunneling nonlin-
earity between the tip (driven by the piezoelectric actuator) and the cantilever (driven by the
electrostatic actuator). The current sensor catches and amplifies the small value of tunneling
current. The next chapters give more insight to the modeling of the single components of the
full nonlinear model as well as to appropriate control and compensation.
 Chapter 3

Static hysteresis and creep

Contents
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.2 Prandtl-Ishlinskii approach . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2.1 Model initialisation, identification and inversion . . . . . . . . . . . . . . . 37
3.2.2 Hysteresis compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3 Modified Prandtl-Ishlinskii approach . . . . . . . . . . . . . . . . . . . . 42
3.3.1 Model initialisation, identification and inversion . . . . . . . . . . . . . . . 43
3.3.2 Hysteresis compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.4 Observer-based approach . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.5 Observer-based approach with MPI model . . . . . . . . . . . . . . . . 50
3.6 Numerical inverse-based hysteresis compensation with adaptation . . 53
3.6.1 Adaptive forgetting as a statistical decision problem . . . . . . . . . . . . 55
3.6.2 Numerical indirect model inverse . . . . . . . . . . . . . . . . . . . . . . . 57
3.6.3 Hysteresis compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

3.1 Introduction

In this section the nonlinear effect of hysteresis is studied. The measured data from the
capacitive sensors of the experimental nanopositioning platform in X and Y directions are
used together with the corresponding control signals in these directions to identify the model
of static nonlinearity. To that end, the Prandtl-Ishlinskii model in its classical (referred
simply as PI) and modified (referred as MPI) version is first fitted to the measured data,
inverted and cascaded with the corresponding axis of the real plant in the feedforward path.
The presentation of this open-loop technique is followed by that of observer-based hysteresis
compensation as proposed in [Yi et al., 2009]. In this method the hysteresis modeling is given
up and this nonlinear effect is assumed as an unknown slowly varying disturbance over piezo
displacement and reconstructed via disturbance observer (DOB). This technique is considered
as a closed-loop method, since it uses the feedback from the measured displacement. However,
differently from [Yi et al., 2009], the disturbance is considered as a new entry of a state vector

35
36 Chapter 3. Static hysteresis and creep

and reconstructed via state observer together with the system state (see also [Ryba et al.,
2014b] and [Ryba et al., 2014a]), hence the direct inversion of system dynamics is avoided.
Next, the open-loop MPI is combined with closed-loop DOB (i.e. the observed disturbance is
now the mismatch between the real hysteresis response and its MPI model), which is shown
to improve the accuracy of modeling and compensation (see [Ryba et al., 2015b]). Finally,
one parametric polynomial model with adaptively changing coefficients is used to modelize
hysteresis and the proper compensation input is determined in real-time using Newton-based
algorithm, which does not require model inversion (see [Ryba et al., 2015a]).

3.2 Prandtl-Ishlinskii approach

In this section the PI approach is used to model and compensate hysteresis effect. In static
case, for identification purpose of a static hysteresis, the piezo dynamics Dpx (s) can be ne-
glected (i.e. qpx (t) = xp (t) in Fig. 3.1). The backlash operator of the Prandtl-Ishlinskii (PI)

𝑣𝑝𝑥 𝑥𝑝𝐻 𝑥𝑝𝑆 ≈ 𝑞𝑝𝑥

𝐻𝑥 𝑆𝑥
𝛤𝑥 ≈ 𝑁𝐿𝑥

𝑥𝑟 𝑣𝑝𝑥 𝑞𝑝𝑥 𝑥𝑝 𝑣𝑥
−1 −1 𝐷𝑝𝑥 (𝑠)
𝐺𝑐𝑎𝑝𝑥 𝛤𝑥−1 𝐺𝑣𝑥 𝑉𝐴𝑥 (𝑠) 𝑁𝐿𝑥 𝐶𝑆𝑥 (𝑠) ++
𝑢𝑥 𝑦1
𝑃𝑖𝑒𝑧𝑜𝑥
𝑛1
𝑛2
𝑦𝑟 𝑣𝑝𝑦 𝑞𝑝𝑦 𝑦𝑝 𝑣𝑦 𝑦2
−1 𝑢𝑦
−1
𝐺𝑐𝑎𝑝𝑦 𝛤𝑦−1 𝐺𝑣𝑦 𝑉𝐴𝑦 (𝑠) 𝑁𝐿𝑦 𝐷𝑝𝑦 (𝑠) 𝐶𝑆𝑦 (𝑠) ++

𝑃𝑖𝑒𝑧𝑜𝑦
𝑃𝑙𝑎𝑛𝑡𝑥𝑦

Figure 3.1: 2D Schema (inverse-based approach).

model is an elementary nonlinear transformation between the input vpx (t) and the output
xpH (t) as follows:
xpH (t) = HxrH [vpx , xpH 0 ](t)
= max{vpx (t) − rxH , (3.1)
min{vpx (t) + rxH , xpH (t − Ts )}},
where rxH is an input threshold for vpx and Ts is the sampling time. The elementary backlash
operator (3.1) multiplied by a weight value wxH forms so called weighted backlash operator
3.2. Prandtl-Ishlinskii approach 37

(see Fig. 3.2a). The linear weighted superposition of n + 1 backlash operators with different
threshold values finally defines the PI hysteresis operator Hx as follows (see Fig. 3.2b):

xpH (t) = Hx [vpx ](t) = wxH T H rxH [vpx , xpH 0 ](t), (3.2)

where wxH is the vector of weights, r xH is the vector of thresholds such that 0 = rxH0 <
rxH1 < ... < rxHn < +∞, xpH 0 is the vector of initial states of the backlash operators and
H rxH [vpx , xpH 0 ](t) is the vector of backlash operators. In its classical form the PI model is
symmetric, it consists of hysteresis operator Hx only and the superposition operator Sx in
Fig. 3.1 is omitted (i.e. Sx [xpH ](t) = xpH (t)). Therefore the PI model Γx [vpx ](t) = Hx [vpx ](t)
approximates static nonlinearity N Lx and xpH (t) ≈ qpx (see Fig. 3.1). The procedure for
hysteresis compensator Γx design consists of three consecutive steps: model initialization,
model identification and model inversion.

(a) (b)

Figure 3.2: Hysteretic nonlinearity: (a) Weighted backlash operator. (b) Hysteresis operator
as a superposition of n + 1 backlash operators [Ang et al., 2007], [Rakotondrabe et al., 2010].

3.2.1 Model initialisation, identification and inversion

In model initialization step, the range of input voltage max{vpx } is divided into n + 1 (often
equal) intervals defining the threshold values r xH as follows:
i
rxHi = max{vpx }, i = 0..n. (3.3)
n+1
and the backlash operators are initialized to zero (deenergized state):

xpH 0i = 0, i = 0..n. (3.4)

Fig. 3.3 shows the identification procedure of the PI approach, which is the second step of the
compensator design. One searches for the optimal weights w∗xH that minimize (in the least
squares sense) the following model error:

Ex [vpx , x̃p ](t) = Hx [vpx ](t) − x̃p (t)

(3.5)
= wTxH H rxH [vpx , xpH 0 ](t) − x̃p (t)
38 Chapter 3. Static hysteresis and creep

with an output displacement x̃p (t) = Gy1capx

(t)
measured at each time instant t. The quadratic
loss function to be minimized is given by:
l
1X
Jx (wxH ) = (Ex [vpx , x̃p ](ti ))2
2 (3.6)
i=1
1
= wTxH F x wxH + wTxH f x1 + fx0
2
where ti = i · Ts and l is a size of the collected experimental data set. The quadratic term
represented by the Hessian matrix F x , the linear term represented by the vector f x1 and the
scalar fx0 in (3.6) are defined as follows:
l
X
Fx = H rxH [vpx , xpH 0 ](ti )H Trx [vpx , xpH 0 ](ti ) (3.7)
H
i=1

l
X
f x1 = x̃p (ti )H rxH [vpx , xpH 0 ](ti ) (3.8)
i=1

l
X 1
fx0 = x̃2p (ti ) (3.9)
2
i=1
To ensure the existence of the model inverse, the slope of the PI model has to be monotonous
function of the input voltage and since this slope is defined as the sum of the weights of
the activated backlashes, these weights wxH have to be constrained by the following convex
polyhedron: 
P x = w xH : U H w xH − u H ≥ 0 (3.10)
where U H ∈ <n+1×n+1 , uH ∈ <n+1 are defined as:
   
1 0 ··· 0 
1 1 · · · 0

 
UH = . , u = (3.11)

.. . . ..  H  .. 
 .. . . . .
1 1 ··· 1 

and  > 0 is a possibly infinitely small value. This condition ensures strong monotonicity
of the hysteresis curve. Note, that the absolute term (3.9) does not depend on wxH , and as
a result it can be neglected during the optimization procedure. In this case the quadratic
optimization problem constrained by (3.10) and (3.11) boils down to:
 
1 T
QPx : min {Jx (wxH )} = min wxH F x wxH + wTxH f x1 (3.12)
wxH ∈Px wxH ∈Px 2

The third step is the inversion of the hysteresis model identified in the previous step. The
inverse model is also PI type as follows:
T
vpx (t) = Γx−1 [xpH ](t) = Hx−1 [xpH ](t) = w0xH H rx0 [xpH , v px0 ](t) (3.13)
H
3.2. Prandtl-Ishlinskii approach 39

𝒘∗𝑥𝐻
𝐻𝑟 𝑥
𝐻0

𝑢𝑥 𝑣𝑝𝑥 𝐻𝑟 𝑥
𝐺𝑣𝑥 𝐻1

.
. 𝑭𝑥
.

𝑦1 −1
𝑥𝑝
𝐺𝑐𝑎𝑝𝑥

Figure 3.3: Identification procedure for X direction (PI approach).

with the transformed vector of thresholds:

i
X
rx0 H i = wxH j (rxH i − rxH j ), i = 0..n, (3.14)
j=0
| {z }
r 0x =ΞH (r xH ,wxH )
H

vector of weights:
1
wx0 H 0 = , i = 1..n
w xH 0
w xH i
wx0 H i = − ! !, (3.15)
i
P i−1
P
w xH 0 + w xH j w xH 0 + w xH j
j=1 j=1
| {z }
w0x =ΦH (wxH )
H

and vector of initial states of the inverted backlash operators:

i
X n
X
vpx0i = wxH j xpH 0i + wxH j xpH 0j , i = 0..n . (3.16)
j=0 j=i+1
| {z }
v px0 =ΘH (xpH 0 ,wxH )

Symbol 0 refers to the parameters of the inverted model (for instance w0xH is a vector of weights
of the inverted model of hysteresis operator).

3.2.2 Hysteresis compensation

In this subsection hysteresis compensation in real time is performed on the two horizontal
piezoactuated axes of the nanopositioning platform. To that end, the PI model is fitted to
the experimental data, inverted and cascaded with the plant in the feedforward path. A
triangular voltage input signal of positive variable amplitude and frequency 0.1 Hz is chosen
40 Chapter 3. Static hysteresis and creep

10 10
model model
measured measured
8 8

6 6
y1 [V]

y2 [V]
4 4

2 2

0 0
0 2 4 6 8 10 0 2 4 6 8 10
ux [V] uy [V]

(a) (b)

10 10
model model
measured measured
8 8

6 6
y1 [V]

y2 [V]

4 4

2 2

0 0
0 2 4 6 8 10 0 2 4 6 8 10
ux [V] uy [V]

(c) (d)

10 10

8 8

6 6
ux [V]

uy [V]

4 4

2 2

0 0
0 2 4 6 8 10 0 2 4 6 8 10
y [V] y [V]
1 2

(e) (f)

7
uncompensated
uncompensated
8 6 compensated
compensated

5
6
y2 [V]
y1 [V]

4
4 3
2
2
1
1 2 3 4 5 6 7 1 2 3 4 5 6 7
xr [V] yr [V]

(g) (h)

Figure 3.4: PI approach for hysteresis compensation (in X (left) and Y (right) direction) with
16 backlashes: (a), (b) Before identification. (c), (d) After identification. (e), (f) Inverted
model. (g), (h) Compensation.
3.2. Prandtl-Ishlinskii approach 41

10 8

8
6

voltage [V]
voltage [V]

xr
6 yr
y1uc
4 y2uc
y1c
4 y2c

2
2

0 0
0 1 2 3 4 5 0 1 2 3 4
time [s] time [s]

(a) (b)

5 5

4 4
voltage [V]

3 voltage [V] 3

2 2
xr xr
y
1 y1uc 1 y2uc
1uc
y1c y2c
1c
0 0
0 10 20 30 40 50 0 10 20 30 40 50
time [s] time [s]

(c) (d)

Figure 3.5: Reference tracking using PI approach for X (left) and Y (right) directions (sub-
scripts ’uc’ and ’c’ refer to the uncompensated and compensated case, respectively): (a), (b)
Triangle of variable amplitude. (c), (d) Step response.

as an identification signal. This frequency has to be low enough in order not to excite piezo
dynamics, and sufficiently high in order to avoid the creep phenomenon. Fitting the model
curve into the measured data is done using the Quadratic Programming algorithm of Matlab
Optimization ToolboxTM . The identification procedure at each iteration modifies the model
weights until the proper tolerance is achieved. The fitting accuracy increases with the number
of backlashes and in the considered case 16 backlashes have been chosen. Fig. 3.4a (resp.
Fig. 3.4b) shows the hysteresis of a measured response of the piezoactuator and the hysteresis
model with the initial weights, while Fig. 3.4c (resp. Fig. 3.4d) shows the results after the
identification procedure for X (resp. Y) directions. The corresponding inverted PI model
shown in Fig. 3.4e (resp. Fig. 3.4f) is cascaded with the plant system as shown in Fig. 3.1.
The effect of hysteresis was successfully cancelled in both directions (see Fig. 3.4g and Fig. 3.4h,
respectively). Since the classical PI model is symmetric and the measured hysteresis curve is
evidently asymmetric (especially for higher voltages), the model does not catches the hysteresis
to the end (despite high accuracy of 16 backlashes). The fitting accuracy can be substantially
improved using the modified version of the PI approach called MPI, which is presented in the
next section. However, even with PI model the tracking performance is quite good as shown
42 Chapter 3. Static hysteresis and creep

(a) (b)

Figure 3.6: Weighted dead-zone operator (a) and superposition operator consisting of m + 1
one-sided dead-zone operators (b) [Ang et al., 2007].

in Fig. 3.5a (resp. Fig. 3.5b). The creep phenomenon is however still present as shown in
Fig. 3.5c and Fig. 3.5d and to compensated its effect, additional modeling should be done, or
it can be eliminated by feedback control.

3.3 Modified Prandtl-Ishlinskii approach

Since the PI model in its classical form cannot catch asymmetric hysteresis loops (due to the
property of symmetry of the elementary play operator it uses), the Modified Prandtl-Ishlinskii
(MPI) model has been elaborated in order to solve this problem. The idea behind this method
lies in introduction of an asymmetric elementary one-sided dead-zone operator (in addition to
the backlash operator) defined as follows:
(
max{xpH (t) − rxS , 0}, rxS > 0
xps (t) = SrxS [xpH ](t) = (3.17)
xpH (t), r xS = 0

where rxS is a threshold for xpH . Next, an elementary one-sided dead-zone operator (3.17) is
multiplied by a weight value wxS as shown in Fig. 3.6a forming a weighted one-sided dead-zone
operator. The PI superposition operator Sx is given by the linear weighted superposition of
many one-sided dead-zone operators with different threshold values (see Fig. 3.6b) as follows:

xps (t) = Sx [xpH ](t) = wxS T S rxS [xpH ](t), (3.18)

where r xS is the vector of thresholds such that 0 = rxS 0 < rxS 1 < ... < rxS m < +∞, wxS
is the vector of weights and S rxS [xpH ](t) is the vector of dead-zone operators. Finally, the
concatenation of PI hysteresis operator Hx defined by (3.2) and PI superposition operator Sx
defined by (3.18) defines MPI operator Γx , which is an approximation of the static nonlinearity
3.3. Modified Prandtl-Ishlinskii approach 43

N Lx (see Fig. 3.1) as follows:

xps (t) = Γx [vpx ](t) = Sx [Hx [vpx ]](t)

(3.19)
= wxS T S rxS [wxH T H rxH [vpx , xpH 0 ]](t)

The procedure for compensator design for MPI method is similar to the PI approach and it
consists of 3 consecutive steps described in the following subsections.

3.3.1 Model initialisation, identification and inversion

In the initialization step, the threshold values r xH and r 0xS are equally distributed over the
range of input vpx and output xps , respectively, as follows:
i
rxH i = max{vpx }, i = 0..n, (3.20)
n+1
i
rx0 S i =max{xps }, i = 0..m, (3.21)
m+1
and it is assumed that the backlash operators are in their deenergized state as follows:

xpH 0i = 0,
i = 0..n. (3.22)
h 0
i
In the identification step, the optimal weights w∗T
x = w xH
∗T wx∗TS
are looked for in the sense
of least-square minimization of the following error model (see Fig. 3.7):

Ex [vpx , x̃p ](t) = Hx [vpx ](t) − Sx−1 [x̃p ](t)

T
= wTxH H rxH [vpx , xpH 0 ](t) − w0 xS S rx0 [x̃p ](t)
S
!

T
 H [v
rxH px , x pH 0 ](t) (3.23)
T
= w xH w xS 0
−S rx0 [x̃p ](t)
| {z } S
w T | {z }
x
Ψx [vpx ,xpH 0 ,x̃p ](t)

The loss function Jx (wx ) to be minimized is given by:

l
1X 1
Jx (wx ) = (Ex [vpx , x̃p ](ti ))2 = wx T Fx wx (3.24)
2 2
i=1

with the Hessian matrix F x defined as follows:

l
X
Fx = Ψx [vpx , xpH 0 , x̃p ](ti )ΨTx [vpx , xpH 0 , x̃p ](ti ) (3.25)
i=1

Again the monotonicity of the hysteresis loading curve is required for the existence of the
model inverse. This can be assured if the weights wxH and w0xS are constrained by the
following convex polyhedron:
       
UH 0 w xH uH 0
Px = wxH , w0xS : − ≥ (3.26)
0 U 0S w0xS u0S 0
44 Chapter 3. Static hysteresis and creep

𝐻𝑟 𝑥 𝒘∗𝑥
𝐻0
𝑢𝑥 𝑣𝑝𝑥
𝐺𝑣𝑥 𝐻𝑟 𝑥 𝑄𝑃𝑥
𝐻1
𝑯𝑟 𝑥 𝑃𝑥
𝐻
.
. 𝐽𝑥 (𝒘𝑥 )
. 𝒘𝑥
𝐻𝑟 𝑥
𝐻𝑛

𝒘𝑥𝐻 𝒘′𝑥𝑆
𝑆𝑟′𝑥 𝑭𝑥
𝑆0

𝜳𝑥
𝑦1 𝑥𝑝
−1
𝐺𝑐𝑎𝑝𝑥 𝑆𝑟′𝑥 −1 ∑
𝑆1 ×
. 𝑇
.
.
𝑺𝑟′𝑥
𝑆

𝑆𝑟′𝑥
𝑆𝑛

Figure 3.7: Identification procedure for X direction (MPI approach).

where U H and uH are defined like in (3.11) and U 0S ∈ <m+1×m+1 , u0S ∈ <m+1 are given by:
   
1 0 ··· 0 
1 1 · · · 0 
U 0S =  . . .
 0
.  , u = . . (3.27)
  
 .. .. . . ..  S  .. 
1 1 ··· 1 

The constrained quadratic optimization problem for finding unknown weights wx can be
formulated as below:
 
1 T
QPx : min {Jx (wx )} = min wx Fx wx (3.28)
wx ∈Px wx ∈Px 2

Using identified weights w0xS and thresholds r 0xS defined by (3.21), the vector of thresholds r xS
and vector of weights wxS of the one-sided dead-zone operators can be computed as follows:
i
X
rxS i = wx0 S j (rx0 S i − rx0 S j ), i = 0..m, (3.29)
j=0
| {z }
r xS =ΞS (r 0x ,w0x )
S S

1
w xS 0 = , i = 1..m
wx0 S 0
wx0 S i (3.30)
w xS i = − ! !,
i i−1
wx0 S 0 + wx0 S j wx0 S 0 + wx0 S j
P P
j=1 j=1
| {z }
wxS =ΦS (w0x )
S
3.3. Modified Prandtl-Ishlinskii approach 45

10 10
model model
measured measured
8 8

6 6
y1 [V]

y2 [V]
4 4

2 2

0 0
0 2 4 6 8 10 0 2 4 6 8 10
ux [V] uy [V]

(a) (b)

10 10
model model
measured measured
8 8

6 6
y1 [V]

y2 [V]

4 4

2 2

0 0
0 2 4 6 8 10 0 2 4 6 8 10
ux [V] uy [V]

(c) (d)

10 10

8 8

6 6
ux [V]

uy [V]

4 4

2 2

0 0
0 2 4 6 8 10 0 2 4 6 8 10
y [V] y [V]
1 2

(e) (f)

without compensation without compensation

8 8
with compensation with compensation
6 6
y1 [V]

y2 [V]

4 4

2 2

0 0
0 2 4 6 0 2 4 6
xr [V] yr [V]

(g) (h)

Figure 3.8: MPI approach for hysteresis compensation (in X (left) and Y (right) direction) with
16 backlashes and 16 dead-zones: (a), (b) Before identification. (c), (d) After identification.
(e), (f) Inverted model. (g), (h) Compensation.
46 Chapter 3. Static hysteresis and creep

The inverse of the MPI model is finally given by:

vpx (t) = Γx−1 [xps ](t) = Hx−1 [Sx−1 [xps ]](t)
T T (3.31)
= w0xH H rx0 [w0xS S rx0 [xps ], v px0 ](t)
H S

with the transformed vector of thresholds r 0xH ,

vector of weights w0xH and vector of initial
states v px0 of the inverted backlash operators as defined in (3.14)-(3.16). The inverse model
Sx−1 of superposition operator Sx has been already determined by r 0xS and w0xS during ini-
tialization and identification steps, respectively. The number of backlashes and dead-zone
operators has been set to 16. The other identification conditions are the same as in case of PI
model.

3.3.2 Hysteresis compensation

The feedforward hysteresis compensator is obtained by cascading the inverse MPI model Γx−1
with the real plant as shown in Fig. 3.1. The compensation results for each axis are given
in Fig. 3.8 and the tracking performance in Fig. 3.9. Comparing these results with the ones
corresponding to the PI model, one can clearly see that MPI model outperforms its classical
PI version, as unlike the latter it can capture the asymmetric hysteresis curves. However
the creep phenomenon is still present as shown in Fig. 3.9c and Fig. 3.9d. Its effect will be
compensated in the next subsections.

3.4 Observer-based approach

Apart from the hysteresis phenomenon, piezoelectric actuators can drift when positioned over
extended periods of time during slow-speed operation. This adverse effect is called the creep
phenomenon. A low-order state-space disturbance observer is designed for hysteresis and creep
reconstruction and subsequent compensation in both horizontal axes as shown in Fig. 3.10.
Both hysteresis and creep are treated as a slowly varying disturbance dx over the linear affine
representation of hysteresis (see Fig. 3.11a). This disturbance is assumed to be bounded as
shown in [Ryba et al., 2015b] and the first equation of (2.3) is expressed as:
N Lx [vpx ] = Gpx vpx + dx (3.32)
Since the bandwidths of voltage amplifier (4 kHz) and capacitive sensor (8.5 kHz) are relatively
high w.r.t. the bandwidth of the piezoelectric actuator (630 Hz), they can be treated as
constant gains (Gvx and Gcapx , respectively). Similarly, the bandwidth of anti-aliasing filters
is around 20 kHz and as a result it can be neglected as well. The following second order model
of the X axis is used to estimate the slowly varying disturbance dx :
2 x = ω 2 (G v + d ) =
ẍp + 2ξpx ωpx ẋp + ωpx p px px px x
| {z }
N Lx [vpx ]
2 (G G u + d ) = ω 2 (u + d ) (3.33)
ωpx px vx x x
| {z } px
| 1 {z x}
N Lx [ux ] N Lx [u1 ]
3.4. Observer-based approach 47

10 8

8
6
voltage [V]

voltage [V]
6 xxrr
y
yy1uc r
1uc 4 y2uc
4 yy1c
1c
y2uc
2
2

0 0
5 6 7 8 9 10 0 1 2 3 4
time [s] time [s]
(a) (b)

5 5

4 4
voltage [V]

voltage [V]
3 3

2 2
xr xr
y
1 y1uc 1 y2uc
1uc
y1c y2c
1c
0 0
0 10 20 30 40 50 0 10 20 30 40 50
time [s] time [s]
(c) (d)

Figure 3.9: Reference tracking using MPI approach for X (left) and Y (right) directions: (a),
(b) Triangle of variable amplitude. (c), (d) Step response.

It includes the gain Gvx of the voltage amplifier V Ax (s), the piezo dynamics Dpx (s) and the
gain Gcapx of the capacitive sensor CSx (s) and can be rewritten in state-space (with state
variables: x1 = xp , x2 = ẋp ) as follows:
    
ẋ1 0 1 x1
= 2
ẋ2 −ωpx −2ξpx ωpx x2
| {z }
  Ax
 
0 0
+ 2 u1 + 2 dx + wx
ωpx ωpx (3.34)
| {z } | {z }
Bx Bdx
 

x1
y1 = Gcapx 0 + n1
| {z } x2
Cx

where ux is the voltage amplifier input, y1 is the capacitive sensor output, wx and n1 are the
process and measurement noises, respectively. The observer dynamics is chosen to be much
faster then the disturbance variations and therefore those variations can be "seen" by the
48 Chapter 3. Static hysteresis and creep

𝑥𝑟
𝐷𝑂𝐵𝑥
−1
𝐺𝑐𝑎𝑝𝑥 𝑑𝑥
𝑑𝑥 𝑞𝑝𝑥
𝑣𝑝𝑥 𝑥𝑝 𝑣
+- −1
𝐺𝑝𝑥 −1
𝐺𝑣𝑥 𝑉𝐴𝑥 (𝑠) 𝐺𝑝𝑥 ++ 𝐷𝑝𝑥 (𝑠) 𝐶𝑆𝑥 (𝑠) 𝑥 +
+
𝑢1 𝑢𝑥 𝑦1
𝑁𝐿𝑥 𝑃𝑖𝑒𝑧𝑜𝑥
𝑛1
𝑛2
𝑢𝑦 𝑑𝑦 𝑞𝑝𝑦
𝑢2 𝑣𝑝𝑦 𝑦𝑝 𝑣𝑦 𝑦2
−1 𝐷𝑝𝑦 (𝑠) 𝐶𝑆𝑦 (𝑠)
+- −1
𝐺𝑝𝑦 𝐺𝑣𝑦 𝑉𝐴𝑦 (𝑠) 𝐺𝑝𝑦 ++ ++
𝑁𝐿𝑦 𝑃𝑖𝑒𝑧𝑜𝑦
−1
𝐺𝑐𝑎𝑝𝑦 𝑃𝑙𝑎𝑛𝑡𝑥𝑦
𝑑𝑦
𝐷𝑂𝐵𝑦
𝑦𝑟

Figure 3.10: 2D schema (observer-based approach).

𝑥𝑟
𝐷𝑂𝐵𝑥
−1
𝐺𝑐𝑎𝑝𝑥
𝑑𝑥
𝑑𝑥
𝑣𝑝𝑥 𝑖𝑝𝑥 𝑥𝑝 𝑣𝑥
−1
𝐺𝑝𝑥 - −1
𝐺𝑣𝑥 𝑉𝐴𝑥 (𝑠) 𝐺𝑝𝑥 𝐷𝑝𝑥 (𝑠) 𝐶𝑆𝑥 (𝑠)
+
𝑢𝑥 ++ ++
𝑦1
𝐻𝑥 𝑃𝑖𝑒𝑧𝑜𝑥
𝑛𝑥
𝑛𝑦
𝑑𝑦
𝑢𝑦 𝑣𝑝𝑦 𝑖𝑝𝑦 𝑦𝑝 𝑣𝑦 𝑦2
−1 𝐺𝑝𝑦 𝐷𝑝𝑦 (𝑠)
−1
𝐺𝑝𝑦 +- 𝐺𝑣𝑦 𝑉𝐴𝑦 (𝑠) ++ 𝐶𝑆𝑦 (𝑠) ++
𝑑𝑦 𝐻𝑦 𝑃𝑖𝑒𝑧𝑜𝑦
−1
𝐺𝑐𝑎𝑝𝑦 𝑃𝑙𝑎𝑛𝑡𝑥𝑦
𝐷𝑂𝐵𝑦 (a) (b)
𝑦𝑟
Figure 3.11: Hysteresis approximation: (a) Affine approximation. (b) MPI approximation.

observer as constant (d˙x ≈ 0) between two consecutive time instants. Next, the state-space
model (3.34) is augmented with the disturbance dynamics (i.e. with the extended state vector
xe = [xT dx ]T ) as follows:

ẋe = Aex xe + Bex u1 + wex

(3.35)
y1 = Ce x x e + n 1
3.4. Observer-based approach 49

T
CxT
    
Ax Bdx Bx
Aex = , Bex = , Cex = (3.36)
0 0 0 0
The steady-state Kalman observer of this system is defined as:

x̂˙ e = Aex x̂e + Bex u1 + Lx (y1 − Cex x̂e ) (3.37)

where Lx = Px CeTx N1−1 is the observer gain matrix and Px is the solution of the following
associated Algebraic Riccati Equation (ARE):

Aex Px + Px ATex − Lx N1 LTx + Wx = 0 (3.38)

with the process Wx and measurement N1 covariance matrices.

Now, using the estimated disturbance dˆx in (3.32) gives the following input/output depen-

8 8
reconstructed reconstructed
measured measured
6 6
y1 [V]

y2 [V]

4 4

2 2

0 0
0 2 4 6 8 10 0 2 4 6 8 10
ux [V] uy [V]

(a) (b)

8 without compensation 8
without compensation
with compensation with compensation
6 6
y2 [V]
y1 [V]

4 4

2 2

0 0
0 2 4 6 0 2 4 6
xr [V] yr [V]

(c) (d)

Figure 3.12: Observer-based approach for hysteresis compensation (in X (left) and Y (right)
direction) (a), (b) Hysteresis reconstruction. (c), (d) Compensation.

dency:
ŷ1 = NˆLxv [ux ] = Gcapx NˆLx [ux ] = Gcapx (Gpx Gvx ux + dˆx ) (3.39)
Fig. 3.12a (resp. Fig. 3.12b) shows a good consistency of the measured and reconstructed
hysteresis between the input ux (resp. uy ) and output y1 (resp. y2 ). To compensate for
50 Chapter 3. Static hysteresis and creep

10 8

8
6
voltage [V]

voltage [V]
6 xr
yr
y1uc 4 y2uc
4 y1c
y2uc
2
2

0 0
5 6 7 8 9 10 0 1 2 3 4
time [s] time [s]
(a) (b)

5 5

4 4
voltage [V]

voltage [V]

3 3

2 2
xr
yr
y1uc
1 1 y2uc
y1c
y2uc
0 0
0 1 2 3 4 0 10 20 30 40 50
time [s] time [s]
(c) (d)

Figure 3.13: Reference tracking using observer-based approach for X (left) and Y (right)
directions: (a), (b) Triangle of variable amplitude. (c), (d) Step response.

the output disturbance dx (resp. dy ), its estimate dˆx (resp. dˆy ) is then substracted from
the input G−1 −1
capx xr (resp. Gcapy yr ) forming new input u1 (resp. u2 ) as shown in Fig. 3.10.
In Fig. 3.12c (resp. Fig. 3.12d) the reference voltage xr (resp. yr ) (see Fig. 3.10) being a
triangle of variable amplitude and frequency 1 Hz was applied and the corresponding output
voltage y1 (resp. y2 ) was measured without (in red) and with (in blue) DOB. The hysteresis
was successfully compensated as shown in Fig. 3.12c (resp. Fig. 3.12d). The corresponding
tracking triangular reference is shown in Fig. 3.13a (resp. Fig. 3.13b). Unlike MPI, DOB
approach cancels also the creep phenomenon as shown in Fig. 3.13c (resp. Fig. 3.13d).

3.5 Observer-based approach with MPI model

In this section a disturbance observer (DOB) is combined with MPI approach as shown in
Fig. 3.14 and the affine hysteresis approximation (3.32) is replaced by MPI model as follows
3.5. Observer-based approach with MPI model 51

𝑥𝑟
𝐷𝑂𝐵𝑥
−1
𝐺𝑐𝑎𝑝𝑥 𝑑𝑥
𝑑𝑥 𝑞𝑝𝑥
𝑣𝑝𝑥 𝑥𝑝 𝑣
+- 𝛤𝑥−1 −1
𝐺𝑣𝑥 𝑉𝐴𝑥 (𝑠) 𝛤𝑥 ++ 𝐷𝑝𝑥 (𝑠) 𝐶𝑆𝑥 (𝑠) 𝑥 +
+
𝑢1 𝑢𝑥 𝑦1
𝑁𝐿𝑥 𝑃𝑖𝑒𝑧𝑜𝑥
𝑛1
𝑛2
𝑢𝑦 𝑑𝑦 𝑞𝑝𝑦
𝑢2 𝑣𝑝𝑦 𝑦𝑝 𝑣𝑦 𝑦2
−1
+- 𝛤𝑦−1 𝐺𝑣𝑦 𝑉𝐴𝑦 (𝑠) 𝛤𝑦 ++ 𝐷𝑝𝑦 (𝑠) 𝐶𝑆𝑦 (𝑠) ++
𝑁𝐿𝑦 𝑃𝑖𝑒𝑧𝑜𝑦
−1
𝐺𝑐𝑎𝑝𝑦 𝑃𝑙𝑎𝑛𝑡𝑥𝑦
𝑑𝑦
𝐷𝑂𝐵𝑦
𝑦𝑟

Figure 3.14: 2D schema (observer-based approach with MPI model).

(see Fig. 3.11(b)):

N Lx [vpx ] = Γx [vpx ] + dx (3.40)
The disturbance dx is combining now the mismatch between the real hysteresis and the
MPI model. This mismatch includes the creep and other external disturbances as well and is
considered as smaller in magnitude (almost everywhere) than in case of the affine hysteresis
approximation. Fig. 3.15 shows comparison of MPI, DOB and hybrid MPI/DOB in terms of
hysteresis compensation in the X axis and for different input frequencies. One can see the effect
of excited dynamics with increasing frequency. Table 3.1 shows the corresponding percentage
compensation errors. They are evaluated by taking the ratio of the major hysteresis loop
width and the full range of the reference (i.e. H% = LlHH × 100%) as shown in Fig. 3.15a.

Table 3.1: Amount of hysteresis H% using different compensation methods and different fre-
quencies - experimental result (uncompensated system has H% = 11.37% of hysteresis)

MPI DOB MPI+DOB

1Hz 1.10% 0.53% 0.26%
20Hz 3.12% 2.12% 0.98%
50Hz 5.01% 6.25% 4.03%

The corresponding tracking performance is shown in Fig. 3.16. For low frequencies DOB
outperforms MPI approach, however for higher frequencies the performance of low-order DOB
worsens and it can be improved by combining it with MPI approach. The creep phenomenon
is compensated only when DOB is used as shown in Fig. 3.17.

The disturbance model is assumed as constant (i.e. d˙x = 0), which is not true. However,
 52 Chapter 3. Static hysteresis and creep

8 without compensation 8 without compensation 8 without compensation

with compensation with compensation with compensation
7

6 6 6
lH

y1 [V]

y1 [V]
y1 [V]

5
4 4 4
3
LH
2 2 2
1
2 4 6 2 4 6 1 2 3 4 5 6 7
xr [V] x [V] xr [V]
r

(a) (b) (c)

8 without compensation 8 without compensation 8 without compensation

with compensation with compensation with compensation

6 6 6
y1 [V]

y1 [V]
y1 [V]

4 4 4

2 2 2

2 4 6 2 4 6 2 4 6
xr [V] xr [V] xr [V]

(d) (e) (f)

8 without compensation 8 without compensation 8 without compensation

with compensation with compensation with compensation
7
6 6 6
y1 [V]

y1 [V]
y1 [V]

5
4 4 4
3
2 2 2
1
1 2 3 4 5 6 7 2 4 6 2 4 6
xr [V] xr [V] xr [V]

(g) (h) (i)

Figure 3.15: Hysteresis compensation in the X direction and for 1 Hz (left), 20 Hz (middle) and
50 Hz (right) for: (a)-(c) MPI approach, (d)-(f) DOB approach, (g)-(i) MPI+DOB approach.

this assumption is less restrictive in the case of DOB based on MPI model (3.40) than in the
case of DOB with the affine model (3.32), which can be seen in Fig. 3.18, where the rate of
change d˙x is smaller in magnitude in case of MPI-based approach. This can be understood
as follows: adding MPI model linearizes the system to some extent and as a result one may
expect the improved performance w.r.t. the DOB alone or MPI alone.
 3.6. Numerical inverse-based hysteresis compensation with adaptation 53

10 10 10

8 8 8
voltage [V]

voltage [V]

voltage [V]
6 x 6 x 6 xr
r r
y1uc y1uc y
1uc
4 4 y 4
y1c 1c y1c

2 2 2

0 0 0
5 6 7 8 9 10 5 5.05 5.1 5.15 5.2 5.25 5 5.05 5.1
time [s] time [s] time [s]

(a) (b) (c)

10 10 10

8 8 8
voltage [V]
voltage [V]

voltage [V]
6 6 xr 6 xr
xr
yuc y1uc
y1uc 4
4 y1c 4
y1c y1c
2
2 2
0
0 5 5.05 5.1 5.15 5.2 5.25 0
5 6 7 8 9 10 time [s] 5 5.02 5.04 5.06 5.08 5.1
time [s] time [s]

(d) (e) (f)

10 10 10

8 8 8
voltage [V]

voltage [V]

voltage [V]

6 6 xr 6 xr
xr
yuc y1uc
y1uc
4 4 4
y1c y1c y1c
2 2 2

0 0 0
5 6 7 8 9 10 5 5.05 5.1 5.15 5.2 5.25 5 5.02 5.04 5.06 5.08 5.1
time [s] time [s] time [s]

(g) (h) (i)

Figure 3.16: Triangular reference tracking for 1 Hz (left), 20 Hz (middle) and 50 Hz (right)
for: (a)-(c) MPI approach, (d)-(f) DOB approach, (g)-(i) MPI+DOB approach.

3.6 Numerical inverse-based hysteresis compensation with adap-

tation

In this section a pth order polynomial with adaptively changing coefficients is chosen to model
static hysteresis [Ryba et al., 2015a]. For brevity, only the equations of motion for the X axis
(including gains Gvx , Gcapx , nonlinear hysteresis Hx [vpx (t)] and assuming Dpx (s) = 1 for low-
frequency behavior) are given and therefore subscript x is omitted. This simple parametric
54 Chapter 3. Static hysteresis and creep

5 5 5

4 4 4
voltage [V]

voltage [V]

voltage [V]
3 3 3

2 2 2
xr xr
xr
y1uc y1uc
1 y1uc 1 1
y1c y1c
y1c
0 0 0
0 10 20 30 40 50 0 1 2 3 4 0 1 2 3 4
time [s] time [s] time [s]

(a) (b) (c)

Figure 3.17: Step tracking for: (a) MPI approach, (b) DOB approach, (c) MPI+DOB ap-
proach.

20
without MPI without MPI
with GPI 500 with MPI
10
d˙x [µm/s]
d˙x [µm/s]

0 0

−10
−500

−20
1 2 3 4 1 1.1 1.2 1.3
time [s] time [s]
(a) (b)

1500
without MPI without MPI
500 with MPI 1000 with MPI

500
d˙x [µm/s]

d˙x [µm/s]

0 0

−500
−500 −1000

−1500
1 1.05 1.1 1.15 1 1.02 1.04 1.06
time [s] time [s]
(c) (d)

Figure 3.18: Disturbance rate of change (d˙x ) for different frequencies: (a) 1 Hz, (b) 10 Hz, (c)
20 Hz, (d) 50 Hz.

model is described as follows:

p
X
ymk = g(uk , θk ) = θik uik , (3.41)
i=0
yk = ymk + ek , (3.42)
3.6. Numerical inverse-based hysteresis compensation with adaptation 55

where θk is a time-varying vector of model parameters (to be adapted) and ek is the mismatch
(including measurement noise) between the real plant output yk and the model output ymk .

𝐴𝐸

𝜃̂𝑘
𝑢𝑘 𝑢 𝑣𝑝 𝑞𝑝 𝑥𝑝 𝑣𝑥 𝑦𝑚𝑘
𝑟𝑘 𝑁𝐼 𝐷𝐴𝐶 𝑉𝐴(𝑠) 𝐻 𝐷𝑝 (𝑠) 𝐶𝑆(𝑠) 𝐴𝐷𝐶 ++
𝑦𝑘
𝑃𝑖𝑒𝑧𝑜
𝑃𝑙𝑎𝑛𝑡 𝑒𝑘

Figure 3.19: The considered model of X axis with adaptive estimator (AE) and numerical
inversion block (N I) for hysteresis compensation.

3.6.1 Adaptive forgetting as a statistical decision problem

This subsection is devoted to the adaptive parameter estimator AE shown in Fig. 3.19. Using
(3.41)-(3.42), the case when the output yk depends on the observations in the following way
is considered:

θk+1 = θk ,
(3.43)
yk = h0k θk + ek , ek ∼ N (0, 1/dk ) dk+1 =dk
.

where θk ∈ Rp is a randomly variable vector of unknown parameters, hk ∈ Rp is the known

regression vector, 0 stands for the transposition operator. The error ek is assumed to be
normally distributed, discrete with unknown variance 1/dk (to be estimated as well). It is
assumed that there is no drift in the evolution of 1/dk . The following posterior probability
density function (pdf) is used to describe the uncertainty of the set of parameters Θk = {θk , dk }
having the observed data Dk = {uγ , yγ }kγ=1 :

−(vk|k +p−2)/2
f Θk Dk ) ∝ 2π/dk
−1
× exp −(θk − θ̂k|k )0 Pk|k
 
(θk − θ̂k|k )dk /2 (3.44)
 
× exp −Σk|k dk /2

In k|k − 1, k indicates the current time, and k − 1 the time of the last recognized output and
symbol ∝ denotes the proportionality. The data learning step (with accordance with Kalman
56 Chapter 3. Static hysteresis and creep

Filter equations) is given as follows:

Kk ≡ Pk|k−1 hk / 1 + h0k Pk|k−1 hk ,

(3.45)
k|k−1 ≡ yk − h0k θk|k−1 , (3.46)
θ̂k|k = θ̂k|k−1 + Kk k|k−1 , (3.47)
Pk|k = (I − Kk h0k )Pk|k−1 , (3.48)
Σk|k−1 + 2k|k−1 / 1 h0k Pk|k−1 hk


Σk|k = + , (3.49)
vk|k = vk|k−1 + 1. (3.50)

where ≡ stands for the equality by definition. In [Karny, 2006] a detailed discussion about
the particular statistics and their prior settings is given.

In (3.43) it is assumed that the set of parameters Θk does not change, which is an unrealistic
assumption. Here, one wants to determine the predictive pdf f Θk+1 Dk ) and compensate for
this unrealistic assumption. Suppose that no explicit time evolution model is available and
only the following pair of alternatives gives the prior information about model parameters:

f0 (Θk+1 ) ≡ f (Θk |Dk ), (3.51)

−(p+2)/2
f1 (Θk+1 ) ∝ (2π/dk+1 ) , (3.52)

where f0 (Θk+1 ) refers to the case when the parameters are not changing within time interval
(k, k + 1) and f1 (Θk+1 ) to the case when all the parameters are uniquely distributed between
(k, k + 1).

restricted to the decision space fA∗ Θk+1

One defines

an
unknown pdf fA Θk+1

≡
{f0 Θk+1 , f1 Θk+1 } equal to f0 Θk+1 with a probability ϕ0 > 0 and to f1 Θk+1 with
a probability
 0 ϕ1 > 0 subject to the constraint ϕ0 + ϕ1 = 1. For the forthcoming

manipulation
ϕ ≡ ϕ0 , ϕ1 . The distance between the reality f Θk+1 ) and fA Θk+1 is assessed using so
called Kullback-Leibler divergence [Karny, 2006] defined as follows:




D f Θk+1 kfA Θk+1 ≡

!
f Θk+1 (3.53)
Z


f Θk+1 ln
dΘ.
Θ∗ fA Θk+1

where for an arbitrary k, Θ∗ is the space of Θk . The following cost function is defined:






φ f Θk+1 , ϕ ≡ ϕ0 D f Θk+1 kf0 Θk+1
 


 (3.54)
−(1 + ρ)%0 + ϕ1 D f Θk+1 kf1 Θk+1 − %1

with the penalties coefficients %0 and %1 for the success of the corresponding alternatives (3.51)
and (3.52), respectively:

%0 ≡ D f Θk |Dk kf Θk |Dk−1 ,




(3.55)
%1 ≡ D f Θk |Dk kf1 Θk .




(3.56)
3.6. Numerical inverse-based hysteresis compensation with adaptation 57

The only tuning parameter in the considered adaptive method is the
scalar ρ ≥ 0. The higher
value of this parameter, the more trust in the alternative f1 Θk+1 . The decision max-min
problem is then defined as follows:



max min φ f Θk+1 , ϕ . (3.57)
ϕ f (Θk+1 )

Substituting ϕ0 = λk+1|k (ϕ1 = 1 − λk+1|k ) allows to write the result of the decision problem
(3.57) as follows:
Pk+1|k = Pk|k /λk+1|k , Σk+1|k = λk+1|k Σk|k ,
(3.58)
θ̂k+1|k = θ̂k|k , vk+1|k = λk+1|k vk|k ,
where the variable forgetting factor λk+1|k is defined by:

ρ+1
λ−1 ln 1 + h0k Pk|k−1 hk


k+1|k =1 + p+1
Σk|k
 
+ vk|k−1 ln
Σk|k−1
h0k Pk|k−1 hk − vk|k 1 − Σk|k−1 Σ−1


k|k
+
1 + h0k Pk|k−1 hk
vk|k−1 vk|k−1
  


+ vk|k−1 − 1 ln + . (3.59)
vk|k vk|k

Equations (3.58) and (3.59) close the time updating step of the sufficient statistics w.r.t.
(3.57).

3.6.2 Numerical indirect model inverse

In this subsection having a parametric model of static nonlinearity defined by (3.41) one
searches an input uk such that the output of the model ymk follows the reference rk . To that
end, the numerical root-finding algorithm is used to solve for the proper uk in the following
equation:
f (uk , θk , rk ) = ymk − rk = g(uk , θk ) − rk = 0. (3.60)

Note that in this way one does not need to find an explicit inverse model g −1 (rk , θki ) (not
necessarily linear in parameters θki ). The Newton’s iteration algorithm is used for solving
(3.60). It is well known that this algorithm does not guarantee the convergence, unlike for
example the bisection method. On the other hand, it has better (quadratic) convergence than
the latter (linear). Therefore a combination of both approaches is used here. The numerical
algorithm needs the starting point u0 , which in the considered case is restricted to the allowable
control voltage interval from 0 to 10 V. Since the Newton’s method solves the problem locally,
(3.60) may have several solutions even in this interval. Therefore, a natural choice for the
starting point value is the one that is sufficiently close to the expected solution (for example
the control voltage in the previous time instant u0 = uk−1 ).
58 Chapter 3. Static hysteresis and creep

First, the algorithm tries to find an interval [uL uR ] around the starting point u0 where
the solution of (3.60) is located, as follows:

uL = uR = u0 , (3.61)
 
uL := uL − a
until f (uL , θk , rk )f (uR , θk , rk ) < 0.
uR := uR + b

where a > 0, b > 0 are possibly small values. Next, the combined Newton/bisection numerical
algorithm is used (starting from u0 ) with an iteration described by the following equations:

f (ui , θk , rk )
ui+1 i
N =u − , (3.62)
fu0 i (ui , θk , rk )
(uL + uR )
ui+1
B = , (3.63)
2
 i+1
uN if ui+1
N ∈ [uL uR ] ,
ui+1 = i+1 (3.64)
uB if ui+1
N ∈/ [uL uR ]
uL = ui+1 i+1
B if f (uR , θk , rk )f (uB , θk , rk ) < 0, (3.65)
uR = ui+1
B if f (uL , θk , rk )f (ui+1
B , θk , rk ) < 0.

where ui+1 i+1

N and uB stand for control input computed by Newton iteration and bisection step,
respectively. The Newton iteration (3.62) is done everytime when possible and the bisection
(3.63) only if the Newton’s iteration gives solution out of the interval (3.64)-(3.65) where
the solution exists. Therefore, the convergence (which is quadratic) is always assured. The
found voltage (after reaching the maximum number of iterations or the required accuracy) is
applied to the piezoelectric actuator as uk . Fig. 3.20a shows tracking triangular trajectory
of variable amplitude in simulation using Newton approach alone while Fig. 3.20b shows
the result of combined Newton with bisection method. One can clearly see that the latter
approach, unlike Newton’s method alone, is robust against solutions that can be found outside
the local neighbourhood of the previous control voltage (which can be seen as sudden jumps
in Fig. 3.20a).

3.6.3 Hysteresis compensation

In this section experimental results for horizontal (X) axis of a GIPSA-lab system, are pre-
sented. These results are obtained for 3rd order time-varying polynomial (i.e. p = 3 in
(3.41)).

The following sinusoidal reference of variable amplitude is applied to the considered system
in the X direction:
xr (t) = 5 + 2sin(2πt) + 2sin(2π1.2t + π) (3.66)

Fig. 3.21a and Fig. 3.21b show the measured and the reconstructed hysteresis curves via adap-
tation algorithm for parameter values ρ = 0.4 and 0.8, respectively. The actual parameters of
the polynomial model (3.41) estimated by the adaptive estimator AE are passed in real-time
3.6. Numerical inverse-based hysteresis compensation with adaptation 59

10 10
y y
r r
8 8

voltage [V]
voltage [V]

6 6

4 4

2 2

0 0
0 2 4 6 8 0 2 4 6 8
time [s] time [s]
(a) (b)

Figure 3.20: Triangular reference tracking with adaptation using numerical algorithm N I: (a)
only Newton method, (b) Newton combined with bisection.

to the numerical inversion algorithm N I (see Fig. 3.19), which searches for the proper com-
pensation input. The corresponding adapted inverse model, shown in Fig. 3.21c and Fig. 3.21d
is applied in cascade with the real plant. The compensation results are shown in Fig. 3.21e
and Fig. 3.21f, respectively. Note that parameter ρ highly influences the parameter estimation
(and as a result the fitting accuracy). Better fit is obtained for ρ = 0.8

Fig. 3.22a-Fig. 3.22f show the corresponding tracking results. Fig. 3.22a shows tracking
without any compensation. In the intervals 20-25 s, 25-30 s and 30-35 s the adaptive compen-
sator is switched between different values of parameter ρ (ρ = 0.8, 0.4 and 0.0, respectively).
The tracking error increases when ρ decreases as shown in Fig. 3.22e. Fig. 3.22f shows time
evolution of the variable forgetting factor λ during this experiment. The estimator AE for-
gets mostly the data at the time instant when the input voltage changes direction (i.e. the
polynomial parameters rapidly change their values from ascending to descending curve). The
forgetting level depends on the input amplitude as well. The higher value of parameter ρ,
the faster adaptivity to changing conditions. Finally, Fig. 3.23a, Fig. 3.23c and Fig. 3.23e
show tracking performance for triangular reference of variable amplitude without any com-
pensation, with well-known MPI approach and with the proposed adaptive compensator (with
ρ = 0.5), respectively. The corresponding tracking errors are given in Fig. 3.23b, Fig. 3.23d
and Fig. 3.23f, respectively. The results clearly show that the proposed compensator adapts
to both changing amplitude as well as to the direction of the input, thus one avoids addi-
tional modeling for different amplitudes and ascending or descending hysteresis curves. Since
the parameters are updated on-line it outperforms non adaptive approaches based on off-line
identification such as MPI approach.
60 Chapter 3. Static hysteresis and creep

10 10
measured measured
estimated estimated
8 8

6 6

y [V]
y [V]

4 4

2 2

0 0
0 2 4 6 8 10 0 2 4 6 8 10
u [V] u [V]
(a) (b)

10 10

8 8

6 6
u [V]

u [V]

4 4

2 2

0 0
0 2 4 6 8 10 0 2 4 6 8 10
r [V] r [V]

(c) (d)

10 10

8 8

6 6
y [V]
y [V]

4 4

2 2

0 0
0 2 4 6 8 10 0 2 4 6 8 10
r [V] r [V]

(e) (f)

Figure 3.21: Experimental results of hysteresis compensation for sinusoidal reference of vari-
able amplitude xr (t) = 5 + 2sin(2πt) + 2sin(2π1.2t + π) for ρ = 0.4 (left) and ρ = 0.8 (right):
(a), (b) Measured and estimated hysteresis. (c), (d) Inverted model. (e), (f) Compensation.
3.6. Numerical inverse-based hysteresis compensation with adaptation 61

10 10
y xyr
8 r 8 yr
voltage [V]

voltage [V]
6 6

4 4

2 2

0 0
15 16 17 18 19 20 20 21 22 23 24 25
time [s] time [s]
(a) (b)

10 10
xyr xyr
8 yr 8 yr
voltage [V]
voltage [V]

6 6

4 4

2 2

0 0
25 26 27 28 29 30 30 31 32 33 34 35
time [s] time [s]
(c) (d)

1 1

0.99
0.5
0.98
e [V]

0 0.97
λ

0.96
−0.5
0.95
ρ = 0.8 ρ = 0.4 ρ = 0.0 ρ = 0.8 ρ = 0.4 ρ = 0.0
−1 0.94
15 20 25 30 35 20 25 30 35
time [s] time [s]
(e) (f)

Figure 3.22: Experimental results for tracking sinusoidal reference of variable amplitude: (a)
Without compensation. (b), (c), (d) With compensation for ρ = 0.8, 0.4, 0.0, respectively. (e)
Corresponding tracking errors. (f) Corresponding variable forgetting factor.
62 Chapter 3. Static hysteresis and creep

10 1
r
y
8
0.5
voltage [V]

e [V]
0
4

−0.5
2

0 −1
4 5 6 7 8 4 5 6 7 8
time [s] time [s]
(a) (b)

10 1
r
y
8
0.5
voltage [V]

6
e [V]

0
4

−0.5
2

0 −1
4 5 6 7 8 4 5 6 7 8
time [s] time [s]
(c) (d)

10 1
r
y
8
0.5
voltage [V]

6
e [V]

0
4

−0.5
2

0 −1
4 5 6 7 8 4 5 6 7 8
time [s] time [s]
(e) (f)

Figure 3.23: Experimental results for tracking triangular reference of variable amplitude: (a),
(c), (e) Without any compensation, with MPI and with the proposed adaptive compensator,
respectively. (b), (d), (f) Corresponding tracking errors.
3.7. Conclusion 63

3.7 Conclusion

In this chapter the nonlinear hysteresis and creep phenomena have been investigated. To that
end, several methods for their subsequent modeling, identification and compensation have been
tested on the horizontal axes of the the experimental platform. The Prandtl-Ishlinskii (PI)
model due to its symmetric nature cannot capture asymmetric hysteresis curves, however it
still gives much better results than without any compensation. The Modified Prandtl-Ishlinskii
(MPI) model extends the PI approach for the asymmetric curves and therefore substantially
increases the accuracy. Unlike MPI, the observer-based approach (DOB) successfully com-
pensates the creep phenomenon in addiction to hysteresis. Further improvement (especially
for higher frequencies) can be achieved by combining the DOB with MPI. A nice alternative
is an adaptive approach, which uses in the considered case a simple polynomial model with
adaptively changing coefficients. The numerical inversion algorithm works in real-time and
one avoids the explicit model inversion.
 Chapter 4

Structural vibration and

cross-couplings

Contents
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2 Linear nominal and perturbed model . . . . . . . . . . . . . . . . . . . 66
4.2.1 Model identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.2.2 Model reduction and model uncertainty . . . . . . . . . . . . . . . . . . . 68
4.3 LQG/LTR approach for vibration reduction . . . . . . . . . . . . . . . 71
4.4 H∞ approach for vibration reduction . . . . . . . . . . . . . . . . . . . . 76
4.4.1 General H∞ problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.4.2 SISO design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.4.3 MIMO design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

4.1 Introduction

This chapter is devoted to the structural vibration and cross-coupling reduction of horizontal
piezoelectric actuators used to move the tunneling tip in X-Y plane along the scanned surface.
Such a horizontal 2D system is already compensated for hysteresis as treated in the previous
chapter. In [Habib et al., 2012], LQG control was used to tackle for the vibration effects while
positioning the lateral axes of the AFM. As mentioned in chapter 1, pure LQG controller is
not guaranteed to be robust w.r.t. the model uncertainties. It is commonly known that full
state feedback LQR controller has at least 60◦ phase margin and 6dB gain margin and as a
result the closed-loop system with this controller is always stable (see [Safonov and Athans,
1977]). However, from the practical point of view not all the states can be available for state
feedback design, either because there is a lack of sensors in the system or some states are
impossible to measure. In this case a state observer is added to the system to estimate its
states on the basis of the measured outputs and the LQR controller is now based not on the
states but on their estimates (output feedback controller). The performance of this observer,
which is based on the nominal model of the system, depends on how much uncertainty is
in the system and since the real system differs from the nominal one, the robust stability

65
66 Chapter 4. Structural vibration and cross-couplings

margins of LQR cannot be more assured. In [Doyle and Stein, 1979], it has been shown that
these margins can be arbitrarily small even for minimum phase, stable systems. One may
try to "speed-up" observer dynamics by playing with its process and measurement covariance
matrices but this may be not sufficient. To overcome this limitation the LQG/LTR (Loop
Transfer Recovery) procedure was proposed by [Kwakernaak, 1972] and furtherly developed
by [Doyle and Stein, 1981]. The idea behind this method lies in LQR loop recovery w.r.t. the
uncertainty on the intput or output of the minimum-phase plant by relaxing the optimality
either of the quadratic feedback controller or the Kalman filter, respectively. This method
has been applied for example in [Munteanu and Ursu, 2008] for vibration control of smart
composite wing of the aircraft or in [Hu et al., 1999] to assure high track density for dual-
stage actuators in HDDs as well as to piezoelectric tuning fork in Z direction of the AFM [Yeh
et al., 2008]. However, due to small displacements (less than 100 nm), the hysteresis effect has
been completely omitted in this work, which was the main motivation for the present work to
experimentally validate the efficiency of this method for vibration reduction in the presence
of nonlinear hysteresis in long-range positioning. Inspired by [Sebastian and Salapaka, 2005],
other robust H∞ controllers based on the mixed-sensitivity design have been designed and
tested as well.

First, the full 2D linear model (together with cross-couplings) is identified and subsequently
reduced in section 4.2. The model uncertainty is given here as well. Next, in sections 4.3 and
4.4, the experimental studies on robust control of the Multi-Input Multi-Output (MIMO) 2D
plant are presented using Single-Input Single-Output (SISO) LQG/LTR and SISO/MIMO
H∞ strategies, respectively, together with some comparison.

4.2 Linear nominal and perturbed model

4.2.1 Model identification

The linear model of the system already compensated for hysteresis with unit static gain, used
as a vibration model, is identified here. The input identification signal is a chirp voltage
of amplitude 0.5 V around 4 V with frequencies linearly growing with time from 0.1 Hz to
Shannon frequency 0.5fs , where fs is the sampling rate of 20 kHz. Firstly, this input is
applied on the X axis of the system as xh , while the Y axis remains on a constant level (i.e.
yh = 4V ). The corresponding response, measured via two capacitive sensors for the X and Y
axes is shown in Fig. 4.1a and Fig. 4.1c, respectively. The same identification signal is applied
subsequently on the Y axis of the system as yh , while the X axis remains on a constant level
(i.e. xh = 4V ) and again the two measurements are recorded (see Fig. 4.1b and Fig. 4.1d).

40th -order models were fitted to the measured responses in order to obtain very good
accordance with the real data as shown in Fig. 4.2. The diagonal sub-figures correspond to
the direct dynamic behavior (i.e. from the input to the corresponding output), while the off-
diagonal sub-figures show the cross-coupling between the two axes in the frequency domain.
These four high-order single-input single-output (SISO) models together create a full-order
4.2. Linear nominal and perturbed model 67

6 6
input xh input yh
output y1 output y1
5 5
voltage [V]

voltage [V]
4 4

3 3

2 2
0 2 4 6 8 10 0 2 4 6 8 10
time [s] time [s]
(a) (b)

6 6
input xh input yh
output y2 output y2
5 5
voltage [V]
voltage [V]

4 4

3 3

2 2
0 2 4 6 8 10 0 2 4 6 8 10
time [s] time [s]
(c) (d)

Figure 4.1: Experimental input/output data used for identification: (a) from X to X (b) from
Y to X (c) from X to Y (d) from Y to Y.

two-input two-output (TITO) model as follows:

 
Gxxp (s) Gxyp (s)
Gp (s) = (4.1)
Gyxp (s) Gyyp (s)

This model will reflect the behavior of the real system as much as possible and will only serve
as a perturbed model for robustness considerations and not for the controller design. The step
response of the full-order TITO model (4.1) is shown in Fig. 4.3. Again, the non-zero response
on the off-diagonal sub-figures shows that the cross-coupling effect is evident. The amount
of coupling in the low frequencies is almost zero (see Fig. 4.2), since the observers designed
in the previous chapter, mainly for hysteresis compensation, cancel any other low frequency
disturbances, including creep and low frequency cross-couplings. However, with increasing
frequencies the coupling-based positioning errors become quite significant.
 68 Chapter 4. Structural vibration and cross-couplings

From: xh From: yh
20 20
real dataxy
0 0 Gxy (s)
p
To: y1

−20 −20
real dataxx
−40 Gxx (s) −40
Magnitude [dB]

p
−60 0 2 4
−60 0 2 4
10 10 10 10 10 10

20 20
real datayx
0 Gyx (s) 0
p
To: y2

−20 −20
real datayy
−40 −40 Gyy (s)
p
−60 0 2 4
−60 0 2 4
10 10 10 10 10 10
frequency [Hz]

Figure 4.2: Frequency response of the real system and full-order TITO model.

From: xh From: yh
1.5 1.5

1 1
To: y1

0.5 0.5

0 0
output [V]

0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02

1.5 1.5

1 1
To: y2

0.5 0.5

0 0
0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02
time [s]

Figure 4.3: Step response of the full-order TITO model.

4.2.2 Model reduction and model uncertainty

The full-order TITO model presented previously is an idealistic case and cannot be used for
the controller design due to its complexity. Therefore, a proper model reduction is looked for
4.2. Linear nominal and perturbed model 69

in this subsection. This reduced-order TITO model is described as follows:

 
Gxx (s) Gxy (s)
G(s) = (4.2)
Gyx (s) Gyy (s)

To that end, the balanced truncation method is used. First the full order model is converted
to its balanced form and next the states which are weakly controllable/observable (with small
corresponding singular values) are neglected. Here the reduction has been performed for all of
the components of Gp (s) separately, for having more degree of freedom. Fig. 4.4 shows Hankel
singular values of each of the component of the full-order model. The states with singular
values less than the prespecified thresholds (pointed by the black lines) are then neglected
resulting in reduced models of components with the number of states pointed by the red line.
The values of thresholds for the singular values for each of the component were chosen to
be 0.35, 0.18, 0.2 and 0.17, respectively, which results in three 10th -order and one 11th -order
models. The magnitude plot in the frequency domain and step responses of both full- and

0.6 1.5
0.5
Singular Values σyx

Singular Values σxy

0.4 1
0.3

0.2 0.5

0.1

0 0
0 10 20 30 40 0 10 20 30 40
Number of States Number of States
(a) (b)

0.6
2
0.5
Singular Values σyx

Singular Values σyy

0.4 1.5

0.3 1
0.2
0.5
0.1

0 0
0 10 20 30 40 0 10 20 30 40
Number of States Number of States
(c) (d)

Figure 4.4: Hankel singular values of the components of the full-order TITO model (red lines
point at the number of states of the reduced model): (a) σxxp (b) σxyp (c) σyxp (d) σyyp .
 70 Chapter 4. Structural vibration and cross-couplings

reduced-order TITO models are shown in Fig. 4.5 and Fig. 4.6, respectively. One can see a
very good accordance of both models despite order reduction.

From: xh From: yh
20 20
Gxy (s)
p
0 0 Gxy(s)
To: y1

−20 −20
Gxx (s)
−40 p −40
Gxx(s)
Magnitude [dB]

−60 0 2 4
−60 0 2 4
10 10 10 10 10 10

20 20
Gyx (s)
p
0 0
Gyx(s)
To: y2

−20 −20
Gyy (s)
−40 −40 p
Gyy(s)
−60 0 2 4
−60 0 2 4
10 10 10 10 10 10
frequency [Hz]

Figure 4.5: Frequency response of the full- and reduced-order TITO model.

From: xh From: yh
1.5 1.5
Gxy (s)
p
1 1
Gxy(s)
To: y1

0.5 0.5
Gxx (s)
p
0 0
Gxx(s)
output [V]

0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02

1.5 1.5
Gyx (s)
p
1 1
Gyx(s)
To: y2

0.5 0.5
Gyy (s)
p
0 0
Gyy(s)

0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02

time [s]

Figure 4.6: Step response of the full- and reduced-order TITO model.

The controlled system has to be robust w.r.t. the model uncertainty. In many practical
situations the actual system can behave slightly differently from the model on the basis of
4.3. LQG/LTR approach for vibration reduction 71

which the controler has been designed. The full-order ideal TITO model defined by (4.1)
reflects the behavior of the real system with a high accuracy for all frequencies and thus it
will be referred here as a perturbed model Gp (s). However, for real-time implementation only
the reduced-order model, referred here as nominal model G(s), will be used. The uncertain
system can be described in an additive or multiplicative way. The additive uncertainty is a
simple way to represent the discrepancy between the perturbed and the nominal model by
taking the difference between their respective transfer functions. In this work, however, the
output multiplicative structure has been chosen to model the uncertain system as follows (see
Fig. 4.7):
Gp (s) = (1 + W∆ (s)∆(s))G(s) (4.3)
where ∆(s) is any stable transfer function, which represents the normalized complex pertur-
bations such that |∆(jω)| ≤ 1, ∀ω and W∆ (s) is a rational transfer function representing the
uncertainty weight chosen as follows:

Gp (jω) − G(jω)
|W∆ (jω)| ≥ |W∆ (jω)∆(jω)| = l∆ (ω) = , ∀ω (4.4)
G(jω)

where l∆ (ω) represents relative magnitude of the neglected dynamics. In other words W∆ (s)
is chosen such that |W∆ (jω)| is an upper bound on the relative errors l∆ (ω).

𝐺𝑝 (𝑠) 𝑊∆ (𝑠) ∆(𝑠) 𝑑(𝑡)

𝑟(𝑡) 𝑦(𝑡)
+- 𝐶(𝑠) 𝐺(𝑠) ++ ++

++

𝑛(𝑡)
Figure 4.7: Feedback system with output multiplicative uncertainty.

4.3 LQG/LTR approach for vibration reduction

On the basis of the obtained model G(s), a Linear Quadratic Gaussian with Loop Transfer
Recovery (LQG/LTR) controller can be designed. Two SISO LTR controllers will be designed
(each for one axis of the considered system). To that end, two components of G(s) are taken
into account, Gxx (s) (for Cx (s) controller design) and Gyy (s) (for Cy (s) controller design).
These models in state-space are given by:

ẋxx = Axx xxx + Bxx xh + wxx ; xxx ∈ R10


Gxx : (4.5)
y1 = Cxx xxx + n1
72 Chapter 4. Structural vibration and cross-couplings

xyy ∈ R11

ẋyy = Ayy xyy + Byy yh + wyy ;
Gyy : (4.6)
y2 = Cyy xyy + n2
with the control inputs xh and yh (the subscript h means that the system is already compen-
sated for hysteresis). wxx , wyy and n1 , n2 are its process and measurement noises, respectively.
To assure zero steady-state tracking error, the forward path of the controlled system must con-
tain an integrator, thus both systems (4.5) and (4.6) are augmented with an integral action
before controller design procedure (see Fig. 4.8). The augmented system for the X axis is
expressed as:
         
ẋxx Axx Bxx xxx 0 0
; xxxa ∈ R11


 = + ẋh +



 ẋ h 0 0 x h 1 w xx
 | {z } | {z } | {z } | {z }
Gxxa : Axxa xxxa Bxxa wxxa (4.7)
 


x xx

 y1 = Cxx 0 + n1



 | {z } xh
Cxxa

Similarly, the augmented system for the Y axis is expressed as:

         
ẋyy Ayy Byy xyy 0 0
= + ẏh + ; xyya ∈ R12






 ẏh 0 0 yh 1 wyy
 | {z } | {z } | {z } | {z }
Gyya : Ayya xyya Byya wyya (4.8)
 


x yy

 y2 = Cyy 0 + n2



 | {z } y h
Cyya

The optimal LQR solutions ẋ∗h = −Kcxxa xxxa and ẏh∗ = −Kcyya xyya minimize the following
cost functions on trajectories of (4.7) and (4.8), respectively:
Z∞
Jx (ẋh ) = (xTxxa Qxxa xxxa + ẋTh Rxxa ẋh )dt (4.9)
0

Z∞
Jy (ẏh ) = (xTyya Qyya xyya + ẏhT Ryya ẏh )dt (4.10)
0

Here, we consider an LTR design on the plant output (the loop can be recovered also on the
plant input, since the problems are dual). From Fig.4.8 one can see that the loop transfer
function for X direction (resp. Y direction) obtained by breaking the LQG loop at point P1x
(resp. P1y ) (i.e. Kalman Filter (KF) loop transfer function) is:

LT FP1x (s) = Cxxa Φxxa (s)Kfxxa ; Φxxa (s) = (sI11×11 − Axxa )−1 (4.11)

LT FP1y (s) = Cyya Φyya (s)Kfyya ; Φyya (s) = (sI12×12 − Ayya )−1 (4.12)
and the loop transfer function obtained by breaking the LQG loop at point P2x (resp. P2y ) is
simply an open loop transfer function Lx (s) (resp. Ly (s)):
1
LT FP2x (s) = Lx (s) = Gxxa (s)KxLQG/LT R (s) = Gxx (s) KxLQG/LT R (s) (4.13)
s
4.3. LQG/LTR approach for vibration reduction 73

1
LT FP2y (s) = Ly (s) = Gyya (s)KyLQG/LT R (s) = Gyy (s) KyLQG/LT R (s) (4.14)
s

Now, the aim of LTR procedure is to make LT FP2x (s) (resp. LT FP2y (s)) to approach to
LT FP1x (s) (resp. LT FP1y (s)) by properly designed respective LQR controller. In other words,
the LQG/LTR design consists of two steps:
1) Loop transfer design - design LQG loop transfer function on the basis of LQR (full-state
feedback) at point P1x (resp. P1y ).
2) Loop transfer recovery (LTR) - approximate the desired loop transfer function obtained
from step 1 with a recovery procedure at point P2x (resp. P2y ) (which is the plant output).

𝐾𝑥 𝐿𝑄𝐺/𝐿𝑇𝑅 (𝑠) 𝐶𝑥 (𝑠)

𝐶𝑥𝑥𝑎
𝑃1𝑥 𝐺𝑥𝑥𝑎 (𝑠)
𝑥̂𝑥𝑥𝑎
𝑥̂̇𝑥𝑥𝑎 = 𝐴𝑥𝑥𝑎 𝑥̂𝑥𝑥𝑎 𝑥ℎ
+- 𝑥̇ ℎ 1
𝑃2𝑥 −𝐾𝑐𝑥𝑥𝑎 𝐺𝑥𝑥 (𝑠)
+𝐵𝑥𝑥𝑎 𝑥̇ ℎ + 𝐾𝑓𝑥𝑥𝑎 𝑒𝑒𝑠𝑡𝑥𝑥𝑎 𝑠 𝑦1

𝐾𝑦 (𝑠) 𝐶𝑦 (𝑠)
𝐿𝑄𝐺/𝐿𝑇𝑅
𝑃2𝑦 𝐶𝑦𝑦𝑎
𝑃1𝑦 𝐺𝑦𝑦𝑎 (𝑠)
𝑥̂𝑦𝑦𝑎
𝑥̂̇𝑦𝑦𝑎 = 𝐴𝑦𝑦𝑎 𝑥̂𝑦𝑦𝑎 𝑦̇ ℎ 𝑦ℎ
+- 1 𝑦2
−𝐾𝑐𝑦𝑦𝑎 𝐺𝑦𝑦 (𝑠)
+𝐵𝑦𝑦𝑎 𝑦̇ ℎ + 𝐾𝑓𝑦𝑦𝑎 𝑒𝑒𝑠𝑡𝑦𝑦𝑎 𝑠

Figure 4.8: Feedback loop with two SISO LQG/LTR controllers (based only on the diagonal
part of G(s)).

Step 1): (the full state design) is done via Kalman Filter Algebraic Riccati Equation (KF
ARE) for X and Y axes as follows:

Axxa Pxxa + Pxxa ATxxa − Kfxxa N1 KfTxxa + Wxxa = 0 (4.15)

Ayya Pyya + Pyya ATyya − Kfyya N2 KfTyya + Wyya = 0 (4.16)

with properly chosen process and measurement covariance matrices:

T T
Wxxa = E[wxxa wxx a
] = Bxxa Bxx a
, N1 = E[n1 nT1 ] = ρx I1×1 , ρx = 1 (4.17)
T T
Wyya = E[wyya wyya
] = Byya Byya
, N2 = E[n2 nT2 ] = ρy I1×1 , ρy = 1 (4.18)
74 Chapter 4. Structural vibration and cross-couplings

Step 2): (LQR loop recovery at plant output) is done via LQR ARE:

ATxxa Sxxa + Sxxa Axxa − KcTxxa Rxxa Kcxxa + Qxxa = 0 (4.19)

ATyya Syya + Syya Ayya − KcTyya Ryya Kcyya + Qyya = 0 (4.20)

with the weighting matrices for the state and the control effort:
T
Qxxa = qx Cxx C ,
a xxa
Rxxa = I1×1 (4.21)
T
Qyya = qy Cyy C ,
a yya
Ryya = I1×1 (4.22)

Once the controllers KxLQG/LT R (s) and KyLQG/LT R (s) are designed for the augmented systems
(4.7) and (4.8), the integrators are finally inferred into the controllers to control the original
systems (see Fig. 4.8) as follows:

1
Cx (s) = KxLQG/LT R (s) (4.23)
s
1
Cy (s) = KyLQG/LT R (s) (4.24)
s
The Nyquist plots of the loop transfer functions LT FP2x (s) (resp. LT FP2y (s)) for X (resp.
Y) axis are shown in Fig.4.9a (resp. Fig.4.9b) in dashed lines (from green to black) and the
respective bode plots for X (resp. Y) axis are shown in Fig.4.9c (resp. Fig.4.9d). One can
observe their convergence to the desired optimal LQR loop transfer function LT FP1x (s) (resp.
LT FP1y (s)), shown in solid blue line, when the recovery gain qx (resp. qy ) is increased. One
can say that the nice properties of LQR have been recovered. Having the open-loop transfer
functions Lxp (s) and Lyp (s) for the perturbed systems Gxxp (s) and Gyyp (s) defined as follows:

Lxp (s) = Gxxp (s)Cx (s) (4.25)

Lyp (s) = Gyyp (s)Cy (s) (4.26)

and the open-loop transfer functions Lx (s) and Ly (s) for the nominal systems Gxx (s) and
Gyy (s) defined by (4.13) and (4.14), the robust stability of the closed-loop systems w.r.t.
model uncertainties is assured by the following conditions (see also Landau):

|Lxp (jω) − Lx (jω)| < |1 + Lx (jω)|, ∀ω (4.27)

|Lyp (jω) − Ly (jω)| < |1 + Ly (jω)|, ∀ω (4.28)

The recovered open-loop transfer functions LT FP 2x , LT FP 2y are shown in Fig. 4.10a and
Fig. 4.10b, respectively and the graphical verification of robust stability conditions (4.27),
(4.28) is shown in Fig. 4.10c and Fig. 4.10d, respectively. The obtained LQG/LTR SISO
controllers (Cx (s) and Cy (s)), are applied to the nominal (Gxx (s) and Gyy (s)) and perturbed
(Gxxp (s) and Gyyp (s)) SISO models, which are the diagonal parts of TITO nominal and
perturbed models (G(s) and Gp (s)). Figs. 4.11a-4.11b show magnitude plot of the perturbed
SISO models in open-loop and of both perturbed and nominal SISO models in closed-loop.
One can see that the resonant peaks were successfully damped for both axes and as a result
4.3. LQG/LTR approach for vibration reduction 75

LQG/LTR recovery (recov. gain qx −−−> 1*1016) LQG/LTR recovery (recov. gain qy −−−> 1*1016)
2
6
4
1
2

IMAG
IMAG

0 0
−2
−1 −4
−6
−2
−1.5 −1 −0.5 0 0.5 1 1.5 −4 −2 0 2
REAL REAL
(a) (b)

LQG/LTR recovery (qx −−−> 1*1016) LQG/LTR recovery (qy −−−> 1*1016)
100 100
Magnitude [dB]
Magnitude [dB]

0 0

−100 −100

−200 0 −200 0 2 4
2 4
10 10 10 10 10 10

0 400
Phase [deg]
Phase [deg]

−100 200

−200 0

−300 −200

−400 0 2 4
−400 0 2 4
10 10 10 10 10 10
Frequency [Hz] Frequency [Hz]

(c) (d)

Figure 4.9: LTR with increasing recovery gain qi = [1, 10, 102 , 106 , 108 , 1010 , 1016 ] (LT FP2i (s)
is in dashed line from green to black, the desired LT FP1i (s) is in solid blue line): a) Nyquist
plot for i = x. b) Nyquist plot for i = y. c) Bode plot for i = x. d) Bode plot for i = y.

the structural vibration has been significantly reduced as shown in step tracking in Figs. 4.11c-
4.11d. However, it is the TITO model and TITO plant that should be controlled and which
contain the model of couplings, omitted in the controller design procedure. The behavior of
the TITO plant in closed-loop, controlled by the two designed LQG/LTR SISO controllers is
given in section 4.4.3 at the end of this chapter together with other control approaches for the
comparison reason.
76 Chapter 4. Structural vibration and cross-couplings

Recovered LTFP2 Recovered LTFP2

y
x
50 50
Lx(s) Ly(s)

Lx (s) Ly (s)
p
p

Magnitude [db]
Magnitude [db]

0 0

−50 0 2 4
−50 0 2 4
10 10 10 10 10 10
Frequency [Hz] Frequency [Hz]
(a) (b)

Robust stability check (X) Robust stability check (Y)

50 50
|L (s)−L (s)| |Ly (s)−Ly(s)|
x x p
p
|1+Lx(s)| |1+Ly(s)|
Magnitude [db]
Magnitude [db]

0 0

−50 0 2 4
−50 0 2 4
10 10 10 10 10 10
Frequency [Hz] Frequency [Hz]

(c) (d)

Figure 4.10: Robustness w.r.t. model uncertainty: a), b) Recovered LT FP 2 for X and Y axis,
respectively. c), d) Robust stability check for X and Y axis, respectively.

4.4 H∞ approach for vibration reduction

4.4.1 General H∞ problem

The H∞ optimal control problem can be formulated as follows: Among all stabilizing con-
trollers C(s) find an optimal one C ∗ (s) that minimizes the the H∞ norm of the lower linear
fractional transformation Fl (P (s), C(s)), i.e.:

γ ∗ = min ||Fl (P (s), C(s))||∞ = ||Fl (P (s), C ∗ (s))||∞ (4.29)

C(s)

where γ ∗ is an optimal cost and P (s) is an augmented plant (plant G(s) with weighting
functions). There is no guarantee that there exists controller which achieves that minimum.
4.4. H∞ approach for vibration reduction 77

20 20

10 10
Magnitude [dB]

Magnitude [dB]
0 0

−10 −10

−20 −20

OL (perturbed) OL (perturbed)
−30 −30
CL (perturbed) CL (perturbed)
CL (nominal) CL (nominal)
−40 0 1 2 3 4
−40 0 1 2 3 4
10 10 10 10 10 10 10 10 10 10
Frequency [Hz] Frequency [Hz]

(a) (b)

1.5
OL (perturbed)
1.5
CL (perturbed) OL (perturbed)
CL (nominal) CL (perturbed)
CL (nominal)
1
output y [V]

1
output y2 [V]
1

0.5 0.5

0 0
0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02
time [s] time [s]

(c) (d)

Figure 4.11: Performance of SISO LQG/LTR controllers: a), b) Magnitude response of the
system in open-loop and closed-loop (with real and nominal system) for X and Y axis, respec-
tively. c), d) Step response of the system in open-loop and closed-loop (with real and nominal
system) for X and Y axis, respectively.

Therefore the optimal problem (4.29) is often formulated as a sub-optimal one as follows:

γ ∗ = inf ||Fl (P (s), C(s))||∞ ≤ ||Fl (P (s), C ∗ (s))||∞ ≤ γ (4.30)

C(s)

where inf stands for infimum and γ is an upper bound of ||Fl (P (s), C ∗ (s))||∞ , which we would
like to make as close as possible to the optimal value γ ∗ .

In this work, the Mixed-sensitivity H∞ control problem is considered. It consists of the

design of the H∞ optimal controller while shaping the sensitivity function S(s), the control
sensitivity function R(s) and the complementary sensitivity function T (s), defined as follows:

S(s) = (I + G(s)C(s))−1
R(s) = (I + G(s)C(s))−1 C(s) (4.31)
−1
T (s) = (I + G(s)C(s)) G(s)C(s)
78 Chapter 4. Structural vibration and cross-couplings

The objective is to minimize the H∞ norm of the system, for a given performance vector.
In the mixed sensitivity problem, this vector is composed of the sensitivity functions (4.31),
multiplied by the respective weights WS (s), WR (s) and WT (s) and the problem is expressed
as follows:

γ ∗ = inf ||N (C(jω))||∞ ≤ γ

C(s)
   
WS (s)S(s) WS (s)S(s) (4.32)
N (C(s)) =  WR (s)R(s) =  WR (s)C(s)S(s) ,
   

WT (s)T (s) WT (s)T (s)

where γ ∗ , γ are the optimal and the obtained cost values, respectively. The weights are used
to shape their respective sensitivity functions according to the desired requirements. The
tracking error from Fig. 4.7 can be expressed as follows:

e(t) = r(t) − y(t) = S(s)(r(t) − d(t)) + T (s)n(t) (4.33)

We want to keep this error small in both the tracking and the disturbance band. However, from
the constraint S(s) + T (s) = 1 we can see that keeping the output sensitivity function S(s)
small will imply complementary sensitivity function T (s) to be close to one, which will cause
propagation of the noise n(t) into the tracking error. Therefore, since usually the disturbance
rejection d(t) and tracking the reference r(t) is in lower frequencies, one tries to keep S(s)
small only in this band of interest. The desired requirements in our case are defined as follows:

1) Robust stability in face of model uncertainty. The controlled system has to be

robustly stable w.r.t. the model uncertainty. From small gain theorem (see Zames 1966),
it is known that the autonomous system described as a feedback interconnection of a stable
systems S1 (s) and S2 (s) is asymptotically stable if the following condition is fulfilled:

||S1 (jω)S2 (jω)||∞ < 1 (4.34)

In our case (for the output multiplicative uncertainty like in Fig.4.7) if we set S1 (s) = T (s) and
S2 (s) = W∆ (s)∆(s) and taking into account that |∆(jω)| ≤ 1, ∀ω we will assure (4.34) when
the following condition is true (see also for instance [Skogestad and Postlethwaite, 2005]):

||T (jω)W∆ (jω)||∞ < 1 (4.35)

We define the following high-pass nth

∆ -order weighting function with lower and upper bounds
determined by A∆ and M∆ , respectively:
1/n
!n∆
s + ω∆ A ∆ ∆
W∆ (s) = 1/n
(4.36)
1/M∆ ∆ s + ω∆

in order to impose an upper bound on the relative magnitude of the neglected dynamics by
properly choosing the bandwidth ω∆ .
4.4. H∞ approach for vibration reduction 79

2) Good robustness margins. Good robustness margins can be assured if ∆G ≥ 2 and

∆φ ≥ 29◦ , where ∆G and ∆φ stand for the gain and phase margins, respectively. Both con-
ditions can be assured if ∆M ≥ 0.5, where ∆M is the modulus margin. These conditions in
terms of constraints on the sensitivity functions can be expressed as follows:

||S(jω)||∞ ≤ 6dB, ∀ω (4.37)

||T (jω)||∞ ≤ 3.5dB, ∀ω (4.38)

3) Reference tracking and disturbance rejection. We define the following low-pass nth e -
order weighting function with lower and upper bounds determined by Ae and Me , respectively:
!ne
1/n
1/Me e + ωe
We (s) = 1/n
(4.39)
s + ωe A e e

in order to impose the desired performance for tracking accuracy (by choosing Ae possibly
small to mitigate an integral action for low frequencies and as a result to eliminate the steady-
state error and other disturbances up to the desired bandwidth ωe ).

4) Control moderation and noise rejection. We define the following high-pass nth u -order
weighting function with lower and upper bounds determined by Au and Mu , respectively:
!nu
1/n
s + ωu Au u
Wu (s) = 1/n
(4.40)
1/Mu u s + ωu

in order to impose the limit on the actuator force and thus eliminate possible saturations in
the controlled system (by choosing properly the value of Au as an inverse of the maximal
control effort within the bandwidth detetmined by ωu , which is used to attenuate the effect
of noise at high frequencies).
By choosing Me = 2 in (4.39) we will immediately assure (4.37). Thus the weight for shaping
the output sensitivity function will take the form:

WS (s) = We (s), Me = 2 (4.41)

By choosing A∆ = 1/1.5 in (4.36) we will immediately assure (4.38). Thus the weight for
shaping the complementary sensitivity function will take the form:

WT (s) = W∆ (s), A∆ = 1/1.5 (4.42)

By choosing properly ωu in (4.40) we will assure the attenuation of the noise influence on the
control input (in our case we want to reject the noise influence for the frequencies beyond 2
kHz). Thus the weight for shaping the control sensitivity function will take the form:

WR (s) = Wu (s), ωu = 2π · 2000 rad/s (2 kHz) (4.43)

More details about the numerical values of the rest of parameters of the weighing functions
are given in the next section.
80 Chapter 4. Structural vibration and cross-couplings

4.4.2 SISO design

In this subsection two SISO decentralized controllers Cx (s) and Cy (s) are designed for lateral
X and Y axes of the considered system, respectively. The controllers are based on the model
of the respective axis only (i.e. in the design procedure only the diagonal parts of G(s) (i.e.
Gxx (s) and Gyy (s)) are used., the cross-coupling models Gxy (s) and Gyx (s) are neglected and
seen as a disturbance acting on the output of the respective axis).

For SISO transfer function H(s) the H∞ norm is defined as follows:

def
||H||∞ = max |H(jω)| ≥ |H(jω)| (4.44)
ω

The Mixed sensitivity problem (4.32) for the case of two SISO decentralized controllers is
given by:
WSx (jω)Sx (jω) WSy (jω)Sy (jω)
WRx (jω)Rx (jω) ≤ γx , WRy (jω)Ry (jω) ≤ γy (4.45)
WTx (jω)Tx (jω) ∞
WTy (jω)Ty (jω) ∞

where γx and γy are the cost values of optimization procedure. This is equivalent to:

||WSx (jω)Sx (jω)||∞ ≤ γx ||WSy (jω)Sy (jω)||∞ ≤ γy

||WRx (jω)Rx (jω)||∞ ≤ γx , ||WRy (jω)Ry (jω)||∞ ≤ γy (4.46)
||WTx (jω)Tx (jω)||∞ ≤ γx ||WTy (jω)Ty (jω)||∞ ≤ γy
Using the Cauchy-Schwartz inequality, if we have the following conditions, conditions (4.46)
will be satisfied as well:
||WSx (jω)||∞ ||Sx (jω)||∞ ≤ γx ||WSy (jω)||∞ ||Sy (jω)||∞ ≤ γy
||WRx (jω)||∞ ||Rx (jω)||∞ ≤ γx , ||WRy (jω)||∞ ||Ry (jω)||∞ ≤ γy (4.47)
||WTx (jω)||∞ ||Tx (jω)||∞ ≤ γx ||WTy (jω)||∞ ||Ty (jω)||∞ ≤ γy
or equivalently:
γx γy
||Sx (jω)||∞ ≤ ||Sy (jω)||∞ ≤
||WSx (jω)||∞ ||WSy (jω)||∞
γx γy
||Rx (jω)||∞ ≤ , ||Ry (jω)||∞ ≤ (4.48)
||WRx (jω)||∞ ||WRy (jω)||∞
γx γy
||Tx (jω)||∞ ≤ ||Ty (jω)||∞ ≤
||WTx (jω)||∞ ||WTy (jω)||∞
Using (4.44), conditions (4.48) can be finally expressed as follows:
γx γy
|Sx (jω)| ≤ |Sy (jω)| ≤
|WSx (jω)| |WSy (jω)|
γx γy
|Rx (jω)| ≤ , |Ry (jω)| ≤ (4.49)
|WRx (jω)| |WRy (jω)|
γx γy
|Tx (jω)| ≤ |Ty (jω)| ≤
|WTx (jω)| |WTy (jω)|
4.4. H∞ approach for vibration reduction 81

The two perturbed systems can be expressed in the output multiplicative structure as follows:
Gxxp (s) = (1 + Wx∆ (jω)∆xx (s))Gxx (s)
(4.50)
Gyyp (s) = (1 + Wy∆ (jω)∆yy (s))Gyy (s)
where ∆xx (s) and ∆yy (s) represent the normalized complex perturbations such that |∆xx (jω)| ≤
1, |∆yy (jω)| ≤ 1, ∀ω and Wx∆ (jω), Wy∆ (jω) are the uncertainty weights for X and Y axes,
respectively. They are chosen as follows:
Gxxp (jω) − Gxx (s)(jω)
|Wx∆ (jω)| ≥ |Wx∆ (jω)∆xx (jω)| = l∆xx (ω) = , ∀ω (4.51)
Gxx (s)(jω)
Gyyp (jω) − Gyy (s)(jω)
|Wy∆ (jω)| ≥ |Wy∆ (jω)∆yy (jω)| = l∆yy (ω) = , ∀ω (4.52)
Gyy (s)(jω)
Since they are the upper bounds on the relative errors (relative neglected dynamics) l∆xx (ω)
and l∆yy (ω) as shown in Fig. 4.12, they can infer them directly to the controller design proce-
dure, so that it can be robust against them. They were chosen like in (4.36) with the following
numerical values for X and Y axes, respectively:
n∆x = 3, A∆x = 1/10, M∆x = 10, w∆x = 2π · 600 rad/s (4.53)
n∆y = 2, A∆y = 1/40, M∆y = 8, w∆y = 2π · 800 rad/s (4.54)
The robust stability condition (4.35) takes the following form:

20 20

0 0
Singular values [dB]

Singular values [dB]

−20 −20

−40 −40

lx(ω) ly(ω)
−60 −60
|Wx (jω)| |Wy (jω)|
∆ ∆
−80 0 2 4
−80 0 2 4
10 10 10 10 10 10
Frequency [Hz] Frequency [Hz]
(a) (b)

Figure 4.12: Magnitude plot of the weighting function and relative errors for X (left) and b)
for Y (right) axis: a) |Wx∆ (jω)| ≥ l∆xx (ω) b) |Wy∆ (jω)| ≥ l∆yy (ω).

||Tx (jω)Wx∆ (jω)||∞ < 1, ||Ty (jω)Wy∆ (jω)||∞ < 1 (4.55)

The weighting functions for the tracking error were chosen like in (4.39) with the following
numerical values for X and Y axes, respectively:
nex = 1, Aex = 1/1000, Mex = 2, wex = 2π · 400 rad/s (4.56)
ney = 1, Aey = 1/1000, Mey = 2, wey = 2π · 600 rad/s (4.57)
82 Chapter 4. Structural vibration and cross-couplings

20 20
|S | |S |
x y
0 |1/WS | 0 |1/WS |
Magnitude [dB]

Magnitude [dB]
x y

−20
−20

−40
−40
−60
−60
−80 −2 0 2 4 −2 0 2 4
10 10 10 10 10 10 10 10
Frequency [Hz] Frequency [Hz]
(a) (b)

10 20

10
0
Magnitude [dB]

Magnitude [dB]

−10 −10
|Tx| −20 |Ty|
−20 |1/WT |
x −30 |1/WT |
y

−30 −2 0 2 4
−40 −2 0 2 4
10 10 10 10 10 10 10 10
Frequency [Hz] Frequency [Hz]
(c) (d)

30 30

20 20
Magnitude [dB]

Magnitude [dB]

10 10

0 0

−10 −10
|R | |Ry|
x
−20 |1/W | −20 |1/WR |
R
x y
−30 −2 0 2 4
−30 −2 0 2 4
10 10 10 10 10 10 10 10
Frequency [Hz] Frequency [Hz]
(e) (f)

Figure 4.13: Sensitivity functions for the nominal plant and the obtained controller for X (left)
and for Y (right) axis: a), b) Output sensitivity function. c), d) Complementary sensitivity
function. e), f) Control sensitivity function.
4.4. H∞ approach for vibration reduction 83

20 20

10 10
Magnitude [dB]

Magnitude [dB]
0 0

−10 −10

−20 −20

OL (perturbed) OL (perturbed)
−30 −30
CL (perturbed) CL (perturbed)
CL (nominal) CL (nominal)
−40 0 1 2 3 4
−40 0 1 2 3 4
10 10 10 10 10 10 10 10 10 10
Frequency [Hz] Frequency [Hz]

(a) (b)

1.4 1.4
OL (perturbed) OL (perturbed)
CL (perturbed) CL (perturbed)
1.2 1.2
CL (nominal) CL (nominal)
1 1
output y [V]

output y2 [V]

0.8 0.8
1

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02
time [s] time [s]

(c) (d)

Figure 4.14: Performance of SISO H∞ controllers: a), b) Magnitude response of the system in
open-loop and closed-loop (with real and nominal system) for X and Y axis, respectively. c),
d) Step response of the system in open-loop and closed-loop (with real and nominal system)
for X and Y axis, respectively.

Finally the weighting functions for control moderation were chosen like in (4.40) with the
following numerical values for X and Y axes, respectively:

nux = 1, Aux = 1/10, Mux = 2, wux = 2π · 2000 rad/s (4.58)

nuy = 1, Auy = 1/10, Muy = 2, wuy = 2π · 2000 rad/s (4.59)

Final weightings for the sensitivity functions were chosen like in (4.41), (4.42) and (4.43),
which means that the numerical values for A∆x and A∆y in (4.53)-(4.54) were reset to the
higher value 1/1.5 in order to assure good robustness margins. Now, using mixed sensitivity
procedure of Matlab two optimal SISO controllers Cx (s) and Cy (s) were found with their
corresponding cost values as follows:

γx = 1.18, for Cx (s)

(4.60)
γy = 1.34, for Cy (s)
84 Chapter 4. Structural vibration and cross-couplings

60 60
Magnitude [dB]

Magnitude [dB]
40 40
20 20
n = 17 n = 17
0 n = 12 0 n = 13
−20 n=8 −20 n=8
n=6 n=6
−40 0 2 4
−40 0 2 4
10 10 10 10 10 10

50 50
Phase [deg]

Phase [deg]
0 0

−50 −50

−100 0 2 4
−100 0 2 4
10 10 10 10 10 10
Frequency [Hz] Frequency [Hz]
(a) (b)

1.5 1.5
output y1 [V]

output y2 [V]

1 1

0.5 n = 17 0.5 n = 17
n = 12 n = 13
n=8 n=8
n=6 n=6
0 0
0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02
time [s] time [s]
(c) (d)

Figure 4.15: Reduction of SISO H∞ controllers with 17 states: a), b) Frequency response of
nth order controller for X and Y axis, respectively. c), d) Step response of the closed-loop
system with nth order controller and real system for X and Y axis, respectively.

The obtained values are only slightly greater than one, meaning that our requirements are
almost met (more precisely they are met w.r.t. the values (4.60), see also equations (4.49)).
Fig. 4.13 shows the sensitivity functions for the obtained controllers Cx (s) and Cy (s) and their
corresponding nominal systems Gxx (s), Gyy (s). One can see that our requirements are more
severe for the Y axis. They can be relaxed for both axes in order to obtain the cost functions
(4.60) less than one by for example decreasing the required bandwidth. Nevertheless, we are
satisfied with the obtained performance (see Fig. 4.14). Figs. 4.14a-4.14b show magnitude plot
4.4. H∞ approach for vibration reduction 85

of the real plant in open-loop and of both real and nominal plant in closed-loop. One can see
that the resonant peaks were successfully damped for both axes and as a result the structural
vibration has been significantly reduced as shown in step tracking in Figs. 4.14c-4.14d.

The H∞ method produces controllers with high order (which is the sum of the order of the
plant, on the basis of which the controller has been computed, and the orders of the weighting
functions). In our case this order is 17th for each axis and it can be reduced using for example
balanced truncation method (like previously for the plant model reduction). The frequency
response of the full- and reduced-order SISO controllers are shown in Fig.4.15a (for the X axis)
and Fig.4.15b (for the Y axis) and their corresponding step reference tracking performance is
shown in Fig. 4.15c and Fig. 4.15d. One can see that significant controller reduction can be
achieved (with almost no performance deterioration for 12th - and 13th -order and with some
performance deterioration for 8th - and 6th -order). We will keep 12th -order for the X axis and
13th -order for the Y axis (the same order as for the obtained LQG/LTR controllers in the
previous sections in order to compare them at the end of this chapter). If lower performance
is required the low-order controllers can be chosen.

4.4.3 MIMO design

The Mixed sensitivity problem is similar as for the SISO case, with this difference that the
measure of the gain of the system at each frequency are the singular values of this system.
They are satisfactory definition of gain (range of gain) for a matrix transfer function (see
[Skogestad and Postlethwaite, 2005]). For MIMO transfer matrix H(s) the H∞ norm is
defined as follows:
def
||H||∞ = max σ̄(H(jω)) ≥ σ̄(H(jω)), (4.61)
ω

where σ̄(H(jω)) denote maximum singular value of H(s).

By multivariable transfer function shaping we understand the shaping of singular values of
appropriately specified transfer functions. For the mixed sensitivity problem (4.32) the MIMO
counterpart is given as follows:

||W S (jω)S(jω)||∞ < γ

||W T (jω)T (jω)||∞ < γ (4.62)
||W R (jω)R(jω)||∞ < γ

where S(s), T (s), R(s) are sensitivity functions matrices and γ is a cost value of the opti-
mization procedure. Using (4.61), the conditions (4.62) can be expressed as follows:

σ̄(S(jω)) < γ σ̄(W −1

S (jω))
σ̄(R(jω)) < γ σ̄(W −1
R (jω)) (4.63)
σ̄(T (jω)) < γ σ̄(W −1
T (jω))

For comparison reason, the same requirements and thus the same weighting functions were
considered as in SISO H∞ approach. The mixed sensitivity procedure gave the cost value
86 Chapter 4. Structural vibration and cross-couplings

20 10

5
0
0
Singular values [dB]

Singular values [dB]

−20 −5

−10
−40 −15
σ(S) σ(T )
−20
−60 σ(S) σ(T )
σ(WS ) −25 σ(WT )
σ(WS ) σ(WT )
−80 −2 0 2 4
−30 −2 0 2 4
10 10 10 10 10 10 10 10
Frequency [Hz] Frequency [Hz]

(a) (b)

30

20
Magnitude [dB]

10

−10
σ(R)
σ(R)
−20 σ(WR−1 )
σ(WR−1 )
−30 −2 0 2 4
10 10 10 10
Frequency [Hz]

(c)

Figure 4.16: Frequency response of an open-loop and closed-loop perturbed system for different
control approaches for structural vibration reduction.

γ = 1.38. This means that the requirements were almost met as seen in Fig.4.16, showing
the singular values of the sensitivity functions and the singular values of their corresponding
weightings. The order of the obtained sub-optimal centralized TITO controller is 54 (which
is the sum of the orders of four subsystems Gxx , Gyy , Gxy , Gyx ) and the used weighting
functions. This order can be finally reduced using balanced truncation method resulting in
30th -order controller.

To compare all of the considered control approaches for structural vibration reduction,
the SISO decentralized controllers obtained via LQG/LTR and H∞ procedures are compared
together with MIMO H∞ centralized controller for control of the TITO plant Gp (s) as shown
in Fig.4.17. Fig. 4.18 and Fig. 4.19 show the frequency response of the system in open-
loop and closed-loop when different controllers are used with the nominal and perturbed
plant, respectively. For the diagonal parts of the TITO model, H∞ SISO approach give
slightly better results than SISO LQG/LTR approach, but both of them do not reduce cross-
couplings (the off-diagonal parts of the model), being however still better than in open-loop.
4.4. H∞ approach for vibration reduction 87

𝑥𝑟 𝑥ℎ 𝑦1 𝑥𝑟 𝑦1
𝑥ℎ
+- 𝐶𝑥 (𝑠) +-
𝐺𝑝 (𝑠) 𝐶(𝑠) 𝐺𝑝 (𝑠)
𝑦𝑟 𝑦ℎ 𝑦𝑟 𝑦ℎ
+- 𝐶𝑦 (𝑠) +-
𝑦2 𝑦2

(a) (b)

Figure 4.17: Closed-loop controlled TITO system: a) SISO control. b) MIMO control.

From: xr From: yr
20 20
0
0
−20
To: y1

open−loop
−40
−20 SISO LTR
SISO H∞ −60
Magnitude [dB]

MIMO H∞
−40 0 2 4
−80 0 2 4
10 10 10 10 10 10
20 20
0
0
−20
To: y2

−40
−20
−60
−80 0 2
4
−40 0 2 4
10 1010 10 10 10
frequency [Hz]
Figure 4.18: Frequency response of an open-loop and closed-loop nominal system for different
control approaches for structural vibration reduction.
88 Chapter 4. Structural vibration and cross-couplings

From: xr From: yr
20 20
0
0
−20
To: y1

open−loop
−40
−20 SISO LTR
Magnitude [dB]

SISO H∞ −60
MIMO H∞
−40 0 2 4
−80 0 2 4
10 10 10 10 10 10

20 20
0
0
−20
To: y2

−40
−20
−60
−80 0 2 4
−40 0 2 4
10 10 10 10 10 10
frequency [Hz]

Figure 4.19: Frequency response of an open-loop and closed-loop perturbed system for different
control approaches for structural vibration reduction.

Moreover, because of coupling phenomena one can see the undesirable peaks even in the
diagonal parts (see the insets in Fig. 4.18 and Fig. 4.19). One can clearly see that H∞ MIMO
approach outperforms other approaches and successfully reduce the resonant peaks for both
direct models and couplings.

Tracking the step reference of 1 V for the real system and the considered approaches is
shown in Fig. 4.20 when one axis (either X or Y) is excited in the same moment. One can see
that all of the approaches successfully reduce the structural vibrations w.r.t. the open-loop
case. The performance of H∞ SISO and H∞ MIMO controllers is comparable for the diagonal
parts and better than in case of SISO LQG/LTR controllers, however only MIMO approach
deals with cross-coupling phenomena. Fig. 4.21 shows the step responses when both X and
Y axes are excited in the same moment (each response is the sum of the responses of the
diagonal part and the corresponding coupling) showing the effect of coupling directly on the
diagonal responses and the superiority of MIMO approach over the other ones.

4.5 Conclusion

The MIMO approach outperforms the SISO approach while controlling the MIMO plant, since
it takes into account the whole model including cross-coupling between the axes. The H∞
control appears to give better results than classical LQG/LTR approach, since by properly
4.5. Conclusion 89

From: xr From: yr
1.5 0.3
open−loop
SISO LTR
0.2 SISO H∞
MIMO H∞
1
0.1
To: y1

0
0.5 open−loop
SISO LTR −0.1
output [V]

SISO H∞
MIMO H∞
0 −0.2
0 0.01 0.02 0 0.01 0.02

0.15 1.5
open−loop
SISO LTR
0.1 SISO H∞
MIMO H∞
1
0.05
To: y2

0
0.5 open−loop
−0.05 SISO LTR
SISO H∞
MIMO H∞
−0.1 0
0 0.01 0.02 0 0.01 0.02
time [s]

Figure 4.20: Step response of an open-loop and closed-loop real system for different control
approaches for structural vibration reduction when only one axis is excited at the same time.

From: x and y From: xr and yr

r r
1.5 1.5

1 1
output [V]

To: y2
To: y1

0.5 open−loop 0.5 open−loop

SISO LTR SISO LTR
SISO H∞ SISO H∞
MIMO H∞ MIMO H∞
0 0
0 0.01 0.02 0 0.01 0.02
time [s]

Figure 4.21: Step response of an open-loop and closed-loop real system for different control
approaches for structural vibration reduction when only one axis is excited at the same time.
90 Chapter 4. Structural vibration and cross-couplings

designed weighting functions, it can assure more gain at low frequencies, while rolling-offs the
higher-frequency part. The classical LQG/LTR has less degrees of freedom in loop-shaping
design - i.e. the loop is shaped by properly chosen either Kalman Filter covariance matrices
(in case of recovery at the plant output) or LQR weightings (in case of recovery at the plant
input) and the recovery part is done automatically. However these design parameters are
the same for all frequencies and if the model is uncertain, choosing a "high-gain" controller
will excite higher frequency part, and thus we have to limit the desired bandwidth. There
exist of course non-classical methods, like frequency-based LQG, where the weighting matrices
are frequency-dependent, but the classical LQG/LTR approach assumes one shot parameters
while H∞ loop-shaping does not. On the other hand the latter method produces controllers of
high order (which is the sum of the order of the plant and order of the weighting functions) and
subsequent model reduction is needed. The existence of weighting functions makes the tuning
procedure more difficult, however more degrees of freedom allow to shape the loop closer to the
desired one. The last issue is whether MIMO control is really worth of consideration. If the
MIMO plant is strongly non-diagonal (i.e. with significant cross-couplings) the approach may
be interesting, provided that a precision plays more important role than the required hardware
resources and if the designed controller will fit to the hardware limitations. If this is not the
case or the cross-couplings are not significant the SISO approach can be more satisfactory.
 Chapter 5

Tunneling current

Contents
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.2 Nonlinear and linearized model . . . . . . . . . . . . . . . . . . . . . . . 92
5.3 Digital control using Pole-placement with sensitivity functions shaping 94
5.4 Obtaining tunneling current - experimental validation . . . . . . . . . 97
5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

5.1 Introduction

In chapter 2 the tunneling phenomenon has been described in details. It has been shown
that this quantum phenomenon takes place when a conductive tip approaches a conductive
surface at a distance lower than 1 nm and when these metals are biased w.r.t. each other with
constant voltage Vb . This chapter is focused on the issues connected with this phenomenon
which is a heart of the experimental platform of GIPSA-lab. The general block scheme of
the nonlinear plant in Z direction is shown in Fig. 5.1. The focus is given to the tunneling
current phenomenon, hence the cantilever model is not taken into account in this section
(its displacement is replaced by the surface variations zs (t)). The cross-couplings from the
horizontal axes are not considered here either. The tunneling tip is approaching the fixed part

𝑧𝑥 𝑧𝑦

𝑑𝑧 + 𝑛𝑧
𝑣𝑝𝑧 𝑧𝑝
𝑢𝑧 +
+ 𝑑 𝑖𝑡 𝑣𝑧 𝑦3
𝑉𝐴𝑧 (𝑠) −𝐺𝑝𝑧 ++ 𝐷𝑝𝑧 (𝑠) +
𝑧𝑧 - 𝑃𝐿𝑡 ′ 𝐶𝑆𝑡 (𝑠) ++
𝑃𝑖𝑒𝑧𝑜𝑧 +
𝑃𝐿𝑡
𝑧𝑠 𝑃𝑙𝑎𝑛𝑡𝑧
𝑑0

Figure 5.1: Block scheme of nonlinear plant in vertical Z direction.

91

⏟
𝑧̂ 𝑐𝑒 (𝑥)|𝑥=𝑥𝑓 =−𝑏𝑉𝑒 2
92 Chapter 5. Tunneling current

of the cantilever (surface). The nonlinear model is first identified and linearized around an
equilibrium point, then a robust pole-placement controller with sensitivity functions shaping
with RS structure is designed using the framework described in details in [Landau and G.,
2006] and applied to our experimental platform (as in [Ryba et al., 2014a]).

5.2 Nonlinear and linearized model

In chapter 2 the full nonlinear model in 3 dimensions has been developed. Taking into account
differential equations for the vertical direction from the input uz (t) to the output y3 (t) from
(2.51), without taking into account the couplings from the horizontal axes (zx (t) and zy (t)),
one obtains the following nonlinear state-space model:


 ẋ1,vz (t) = −ωvz x1,vz (t) + uz (t)

ẋ1,pz (t) = x2,pz (t)





 ẋ2,pz (t) = −ω 2 x1,pz (t) − 2ξpz ωpz x2,pz (t) + ω 2 (−Gpz Gvz ωvz x1,vz (t) + dz (t))

pz pz
(5.1)

 ẋ1,t (t) = x2,t (t)

 ẋ2,t (t) = −ω 2 x1,t (t) − 2ξt ωt x2,t (t) + ω 2 Gt gVb e−k(d0 +x1,pz (t)−zs (t))




 t t

y3 (t) = x1,t + n3 (t)
We can recall that the first equation of (5.1) represents first order model of the voltage amplifier
described by the following transfer function:
vpz (s) Gvz ωvz
V Az (s) = = (5.2)
uz (s) s + ωvz
The effect of hysteresis in the vertical direction is negligible and the creep phenomenon is
considered as a disturbance dz (t), which can be eliminated in the feedback loop. Hence, the
piezoelectric actuator (second and third equations of (5.1)) is expressed as the following linear
system:
2
−Gpz ωpz
zp (s)
P iezoz (s) : = −Gpz Dpz (s) = 2 2
(5.3)
vpz (s) s + 2ξpz ωpz s + ωpz
The last three equations of (5.1) describe the physical law subsystem P L0t (see Fig. 5.1)
together with the current sensor CSt (s). The source of nonlinearity of (5.1) is the exponential
dependency of the tunneling current on the distance d(t) = d0 + zp (t) − zs (t), when this
distance is less than 1 nm and the constant bias voltage Vb is applied between the tip and the
surface:
gVb e−kd(t) if 0 < d(t) ≤ 1 nm

it (t) = (5.4)
0 if d(t) > 1 nm
The small value of tunneling current is finally amplified by the high-gain current sensor:
vz (s) Gt ωt2
CSt (s) = = 2 (5.5)
it (s) s + 2ξt ωt + ωt2
The numerical parameters used for the model described in this chapter are summarized in
Table 5.1. The constant parameter values of k = 1.65 Å−1 and g = 0.0011 were experimentally
5.2. Nonlinear and linearized model 93

Table 5.1: System parameters for the vertical system.

ωvz Voltage amplifier bandwidth 4 kHz

Gvz Voltage amplifier gain 15 V/V
ωpz Piezoactuator bandwidth 120 kHz
Gpz Piezoactuator gain 1.2 nm/V
ξpz Piezoactuator damping 0.9
ωt Current sensor bandwidth 13 kHz
Gt Current sensor gain 109 V/A
d0 Initial distance tip/surface 1 nm
Vb Bias voltage 1.025 V
g Tunneling current constant 1 0.0011
k Tunneling current constant 2 1.65 Å−1
ieq Equilibrium tunneling current 1 nA
deq Equilibrium distance tip/surface 8.4 Å (angstroms)
me Electron mass 9.109 ·10−31 kg
e Electron charge 1.602 ·10−19 C
eV Electronvolt 1.602 ·10−19 J

found by fitting the nonlinear function it (t) = gVb e−kd(t) to the experimental data as shown
in Fig. 5.2. For identification purpose it was assumed that the tunneling current appears

10
measured
model
8
tunnel current it [nA]

0
7 7.5 8 8.5 9 9.5 10
distance d [m] x 10
−10

Figure 5.2: Fitting the model (solid-red) it = gVb e−kd to the experimental data (dashed-blue).

when the distance is less than 1 nm. The data used for identification are the output from the
current sensor y3 (t) and the piezo control voltage uz (t). Since the hysteresis phenomenon is
negligible for such small piezo displacement zp (t), the latter can be deduced from the control
94 Chapter 5. Tunneling current

signal (knowing the piezo gain Gpz ). The tip moves slowly towards the surface and when
the tunneling current appears, the control voltage is set to zero, which corresponds to the
initial distance d0 = 1 nm. Since the obtained data are quite noisy, they are smoothed using
the moving average method before the curve fitting. Fig. 5.2 shows that this fitting is quite
well, which proves the exponential behavior of the tunneling nonlinearity. From chapter 2,
√ √ 1
k = ce φ, where ce = 2 2me /~ = 1.02−1 /(eV) 2 ≈ 2.548 · 1019 . The identified value of k
corresponds to the effective work function tip/surface φ ≈ 4.2 eV, which is in accordance with
the work functions for metals.

The physical law (5.4) is nonlinear and it is linearized around an equilibrium point (deq , ieq )
before linear control design. The first order Taylor expansion of (5.4) gives:
dit
it (t) = ieq + (d(t) − deq ) = ieq − kieq (d(t) − deq ) (5.6)
dd |d(t)=deq
and subsequently:
∆it = it (t) − ieq = −kieq (d(t) − deq ) = −kieq ∆d (5.7)
or in terms of transfer function:
it (s) it (s)
P L0t (s) = = −kieq = = P Lt (s) (5.8)
d(s) zp (s)

Combining (5.2), (5.3), (5.8), and (5.5) altogether gives:

y3 (s)
Gz (s) = = CSt (s)P Lt (s)P iezoz (s)V Az (s)
uz (s)
2 (5.9)
Gt ωt2 Gpz ωpz Gvz ωvz
= 2 2 ki eq 2 2
s + 2ξt ωt + ωt s + 2ξpz ωpz s + ωpz s + ωvz

The bandwidths of the voltage amplifier, the piezoelectric actuator and the current sensor in
the Z axis are equal to 4 kHz, 120 kHz and 13 kHz, respectively. Since, the bandwidth of
the open-loop system is determined by the smallest bandwidth among all of its components
(in this case voltage amplifier) and the desired closed-loop bandwidth is similar to the one
in open-loop (4 kHz), hence in practice both piezoelectric actuator and current sensor can
be taken as static gains. Therefore only the dynamics of the voltage amplifier is taken into
account for the controller design procedure. For that reason (5.9) reduces to:

y3 (s) Gvz ωvz

Gz (s) = = Gt kieq Gpz (5.10)
uz (s) s + ωvz

5.3 Digital control using Pole-placement with sensitivity func-

tions shaping

In this subsection a pole placement with sensitivity functions shaping controller for vertical
motion of the tip is designed. In [Ahmad et al., 2012b] it was experimentally shown that this
method gives better results than simple PI control. The digital controller has an RS structure
5.3. Digital control using Pole-placement with sensitivity functions shaping 95

𝑝 = 𝑑𝑧
𝑟 = 𝑧𝑟 𝑅(𝑧 −1 ) 𝑢 = 𝑢𝑧 𝑧 −𝑑𝑙 𝐵(𝑧 −1 ) 𝑦 = 𝑦3
+- ++
𝑆(𝑧 −1 ) 𝐴(𝑧 −1 )
++
𝑛 = 𝑛3

Figure 5.3: Closed-loop digital control system with RS controller.

as described in [Landau and G., 2006] and is shown in Fig. 5.3. The linear continuous-
time model (5.10) is first discretized with ZOH giving a discrete-time model used for digital
controller design, which can be expressed in the frequency domain as follows:
z −dl B(z −1 )
H(z −1 ) = (5.11)
A(z −1 )

with z −1 variable defined as:

z −1 = e−jω = cos ω − jsin ω, ω = 2π f /fs (5.12)

where f /fs is a normalized frequency (w.r.t. the sampling frequency), dl stands for system
pure delay expressed in number of sampling periods, B(z −1 ) = b1 z −1 +...+bnB z −nB , A(z −1 ) =
1 + a1 z −1 + ... + anA z −nA are polynomials of the numerator and denominator of the plant
model, respectively. The closed-loop transfer function (complementary sensitivity function)
from the reference r(t) = zr (t) to the output y(t) = y3 (t) is given by:

z −dl B(z −1 )R(z −1 )

Syr (z −1 ) = HCL (z −1 ) = (5.13)
A(z −1 )S(z −1 ) + z −dl B(z −1 )R(z −1 )
The aim is to find unknown coefficients of the following controller polynomials:

R(z −1 ) = r0 + r1 z −1 + ... + rnR z −nR (5.14)

S(z −1 ) = 1 + s1 z −1 + ... + snS z −nS (5.15)

assuming that the plant polynomials A(z −1 ) and B(z −1 ) are known. Desired performances
are defined in terms of requirements on the shape of the output sensitivity function (from
p(t) = dz (t) to y(t) = y3 (t)) and of the input (control) sensitivity function (from n(t) = n3 (t)
or p(t) = dz (t) to u(t) = uz (t)) defined respectively as follows:

A(z −1 )S(z −1 )
Syp (z −1 ) = (5.16)
A(z −1 )S(z −1 ) + z −dl B(z −1 )R(z −1 )

A(z −1 )R(z −1 ) R(z −1 )

Sun (z −1 ) = Sup (z −1 ) = − =− Syp (z −1 ) (5.17)
A(z −1 )S(z −1 ) −d −1
+ z l B(z )R(z )−1 S(z −1 )
Note that:
Syr (z −1 ) + Syp (z −1 ) = 1 (5.18)
96 Chapter 5. Tunneling current

The desired closed-loop poles are defined by the polynomial P (z −1 ) given by:

P (z −1 ) = PD (z −1 )PF (z −1 ) (5.19)

where PD (z −1 ) contains the dominant closed-loop poles (the poles that determine the ’main’
dynamics of the closed-loop system) and they are the roots of the denominator of the following
discretized normalized continuous-time second-order system:

ω02
(5.20)
s2 + 2ξω0 s + ω02

with the desired closed-loop bandwidth ω0 and damping ξ. Polynomial PF (z −1 ) defines aux-
iliary closed-loop poles that are faster than the dominant poles (i.e. they are placed at the
frequency higher than ω0 /2π).
Now, the controller polynomials (5.14) and (5.15) can be found by solving the following equa-
tion (so called "Bezout identity"):

P (z −1 ) = A(z −1 )S(z −1 ) + z −dl B(z −1 )R(z −1 ) (5.21)

The right part of (5.21) is the denominator of the closed-loop transfer function (5.13) and the
left part contains the desired-closed loop poles defined by (5.19).
The controller polynomials (5.14) and (5.15) may contain prespecified fixed parts HR (z −1 )
and HS (z −1 ), respectively. Then polynomials R(z −1 ) and S(z −1 ) can be expressed as:

R(z −1 ) = HR (z −1 )R0 (z −1 ) (5.22)

S(z −1 ) = HS (z −1 )S 0 (z −1 ) (5.23)
From (5.12), z −1 = 1 for f = 0, hence by introducing a fixed part HS (z −1 ) = 1 − z −1 (which is
equal to zero at the zero frequency), one assures zero gain for the output sensitivity function
(5.16) at this frequency (since from (5.23) S(1) = 0). This implies zero tracking error at the
steady-state. Similarly, z −1 = −1 for the Shannon frequency f = 0.5fs , hence by introducing
a fixed part HR (z −1 ) = 1 + z −1 , one assures zero gain for the input sensitivity function (5.17)
at this frequency (since from (5.22) R(−1) = 0). This implies that the controller does not
react to signals close to the Shannon frequency, and the operation is like in open-loop. This is
to avoid an actuator stress for high frequencies, where the cascade interconnection actuator-
plant has generally low gain (see [Landau and G., 2006]).
Inserting (5.22) and (5.23) into (5.21) gives a new expression of Bezout equation (5.21) (in-
cluding fixed parts):

P (z −1 ) = A0 (z −1 )S 0 (z −1 ) + z −dl B 0 (z −1 )R0 (z −1 ) (5.24)

where
A0 (z −1 ) = A(z −1 )HS (z −1 ) (5.25)
B 0 (z −1 ) = B(z −1 )HR (z −1 ) (5.26)
and the problem boils down to resolution of (5.24) w.r.t. the polynomials R0 (z −1 ) and S 0 (z −1 ).
If A(z −1 ) and B(z −1 ) do not have common factors, equation (5.24) has an unique solution
5.4. Obtaining tunneling current - experimental validation 97

with minimal degree provided that the following constraints on the degree of polynomials are
fulfilled:
nP = deg(P (z −1 )) ≤ nA + nHS + nB + nHR + dl − 1
nS = deg(S(z −1 )) ≤ nB + nHR + dl − 1 (5.27)
−1
nR = deg(R(z )) ≤ nA + nHS − 1
The details on how to solve (5.24) can be found in [Landau and G., 2006]. Once the poly-
nomials R0 (z −1 ) and S 0 (z −1 ) are found, they are inserted into (5.22) and (5.23), respectively,
giving the final RS controller ready for implementation.

In the considered case, using the parameters from Table 5.1, the model after discretization
with ZOH defined by (5.11), has the following numerical parameters: a1 = −0.2846, b1 =
212.5, dl = 1. The following characteristics on the designed controller were imposed:
- Closed-loop dominant poles PD : Poles placed at frequency ω0 = 2π × 4000 rad/s and with
damping ξ = 0.9, which corresponds to the highly damped desired system without overshoot
and with the closed-loop bandwidth similar to the open-loop one.
- Closed-loop auxiliary poles PF : Double high frequency real poles placed at -0.2 to improve the
closed-loop robustness without affecting significantly the desired closed-loop performances.
- Controller fixed part HS : HS (z −1 ) = 1 − z −1 , an integrator to assure zero steady-state
tracking error and low frequency disturbance rejection.
- Controller fixed part HR : HR (z −1 ) = 1 + z −1 , a real zero added at Shannon frequency 0.5
fs in order to shape the input sensitivity function and to impose an open-loop behavior close
to this frequency.
With the above characteristics, the obtained RS controller for the Z axis has the following
numerical form:
R(z −1 ) = 10−3 × (1.1634 − 0.8331z −1 + 0.3301z −2 )
(5.28)
S(z −1 ) = 1 − 0.6664z −1 − 0.0723z −2 − 0.2613z −3
The closed-loop sensitivity functions of the the model with the designed RS controller are
given in Fig. 5.4. From the output sensitivity function, shown in Fig. 5.4a, one can see that at
low frequencies the gain is almost zero, which is due to the integral action introduced by the
fixed part HS (z −1 ). The closed-loop system has also a good stability margin (||Syp (z −1 )||∞ <
6dB). At the Shannon frequency |Syp (z −1 )| = 1 and from (5.18) it follows that |Syr (z −1 )| =
|HCL (z −1 )| = 0, which in turn implies |Sup (z −1 )| = 0 at this frequency (see Fig. 5.4b and
Fig. 5.4c). This is due to signal blocking at 0.5fs , introduced by the fixed part HR (z −1 ).
Finally, Fig. 5.4d shows the step response of the digital control system. The response is
without oscillations, due to the high desired damping factor ξ = 0.9. The settling time
ts ≈ 0.17 ms and the obtained closed-loop bandwidth ωbCL = 2π/ts ≈ 3.7 kHz, which is close
to the open-loop bandwidth ωbOL = 4 kHz (the difference is due to the auxiliary poles added
to improve system robustness).

5.4 Obtaining tunneling current - experimental validation

In this subsection, the previously designed RS controller is connected with the Z axis of the
experimental platform as shown in Fig. 5.5. The procedure for obtaining tunneling current
98 Chapter 5. Tunneling current

Syp(z−1) Sup(z−1)
10 0

0
−20
−10 6dB
Magnitude (dB)

Magnitude (dB)
−20 −40

−30
−60
−40

−50 −80

−60
0 1 2 3 4
−100 0 1 2 3 4
10 10 10 10 10 10 10 10 10 10
Frequency (Hz) Frequency (Hz)

(a) (b)

HCL(z−1) 1.4
ts
20
1.2

0 1

1
Magnitude (dB)

−20 0.8
y3 [V]

0.6
−40
0.4

−60
0.2
Ts

−80 0 0
10 10
1
10
2
10
3
10
4 0 0.2 0.4 0.6 0.8 1
Frequency (Hz) time [s] −3
x 10

(c) (d)

Figure 5.4: Performance of the closed-loop system model with RS controller: a) Output
sensitivity function Syp (z −1 ), b) Input (control) sensitivity function Sup (z −1 ). c) Closed-loop
pulse transfer function HCL (z −1 ) (complementary sensitivity function). d) Step response of
the closed-loop system.

𝑧𝑟 𝑅 𝑢𝑧 𝑦3
+- 𝑃𝑙𝑎𝑛𝑡𝑧
𝑆

(a)

Figure 5.5: Closed-loop control system with real plant.

consists of two consecutive steps:

- Manual approach - During this step (called also a "coarse approach") the tip is approached
5.4. Obtaining tunneling current - experimental validation 99

(a) (b)

Figure 5.6: Manual approach (view via high-precision camera): (a) The tip and the surface.
(b) Tip deformation after contact with the surface.

10 measured
tunnel voltage y3 [V]

1 measured
tunnel voltage y3 [V]

8 1.5 reference
reference
6
1
2
4
3 0.5
2
0
0 2 4 6 128 130 132 134 136
time [s] time [s]
(a) (b)

Figure 5.7: Fine approach and tunneling current control: (a) Fine approach of the tip with
tunneling current stabilization, (b) Tracking the sinusoidal reference of tunneling current.

to the surface manually using the micrometer screws until both the tip and the surface are
visible through the camera as shown in Fig. 5.6a. Attention must be paid to avoid physical
contact between the tip and the surface so as not to crash/deform the tip apex, as shown
in Fig. 5.6b, since it can have an adverse influence on the quality of the obtained tunneling
current (therefore a sharp tip and a clean surface is usually required).
- Fine approach - This step is done in three stages via proper control as shown in Fig. 5.7a.
100 Chapter 5. Tunneling current

In the first step, the vertical piezo moves the tip closer to the surface until the tunneling
current appears and saturates (1 in the figure). The saturation indicates that the value of the
tunneling current is too high to be measured by the card with the range 0-10V. Therefore,
in the second stage, the piezo "withdraws" the tip away from the surface and compares the
actual value of tunneling voltage with the reference one (imposed by the user) until they are
equal (2 in the figure), and if this occurs, the previously designed RS controller is switched
on and the loop is closed (3 in the figure). Fig. 5.7a shows the case of tunneling current
stabilization around the reference value zr =0.5 V and Fig. 5.7b shows tracking the sinusoidal
reference. Despite quite significant noise at such small scale arising mainly from the high-gain
current sensor the results are satisfactory. To minimize noise effect, a Faraday cage is used
and the presence of the anti-vibration table helps to attenuate external disturbances.

5.5 Conclusion

In this chapter, a vertical dynamical model has been examined and used to design a digital
controller with a polynomial structure in order to control the tunneling current. To that end,
the pole placement with sensitivity functions shaping method has been used. The procedure to
obtain the tunneling current has been described and the results for its stabilization and track-
ing have been presented. The behavior of this current, when positioned in three dimensions
is described in the next chapter.
 Chapter 6

3D results for nanopositioning and

STM applications

Contents
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.2 Nanopositioning in 3D - experimental validation . . . . . . . . . . . . 104
6.3 Surface reconstruction in STM application in simulation . . . . . . . . 110
6.3.1 Plant simulation model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
6.3.2 SISO approach - 3 PID controllers with decoupling compensators . . . . . 113
6.3.3 MIMO approach - Linear Quadratic Integral Controller (LQI) . . . . . . . 115
6.3.4 Simulation results for surface reconstruction . . . . . . . . . . . . . . . . . 117
6.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

6.1 Introduction

In this chapter, two kinds of applications are under consideration. In section 6.2 a nanoposi-
tioning application in three dimensions is validated experimentally, while in section 6.3 STM-
like results (surface reconstruction) are given in simulation. In the first case, the tip is posi-
tioned along X, Y, and Z axes using three piezoelectric actuators. At each time instant the
tip is moved to the (x(t), y(t)) point of the X-Y plane using a prespecified scanning trajectory.
In the vertical direction the piezo tracks the prespecified reference signal, which plays the
role of "virtual surface". In the second case, the STM application is presented in simulation
and two control strategies are settled with a purpose of surface reconstruction. SISO PID
decentralized control is first tested, where unlike in the previous case, the tunneling current is
kept on a constant level and the surface topography is reconstructed from the control signal of
the vertical piezo. Then a multivariable centralized Linear Quadratic Integral (LQI) control
is proposed, where the tunneling current is again kept constant and the surface topography is
reconstructed via Kalman observer. Various adverse phenomena like tunneling nonlinearity,
actuator hysteresis, creep, vibrations and cross-couplings are compensated. Both strategies
are compared at the end and the results are presented in the form of 3D and more common
2D images, similar to the ones obtained from STM/AFM. Each value of the tunneling voltage
y3 (t) from the current sensor has corresponding points in X-Y plane (y1 (t), y2 (t)), expressed

101
102 Chapter 6. 3D results for nanopositioning and STM applications

in volts (or µm) as well (for the horizontal X,Y axes 1V corresponds to the displacement of
5µm, and for the vertical Z axis 1V corresponds to the tunneling current of 1nA).

To move the tip in X-Y plane, two kinds of scanning trajectories are used: raster and
more recent spiral pattern [Mahmood and Moheimani, 2009a], shown in Fig. 6.1. In the raster
pattern, a triangle input voltage is used in one (fast) direction (here X, see Fig. 6.1a) and a
linear ramp in the other (slow) direction (here Y, see Fig. 6.1c) as follows:
  
1 1
(−1)btfr + 2 c
1
xr (t) = xr0 + Ar 2fr t − tfr + (6.1)
fr 2
yr (t) = yr0 + αr t (6.2)

where xr0 and yr0 are the starting points for the signals xr (t) and yr (t), respectively, Ar , fr
are the amplitude and frequency of the triangle signal, respectively, αr defines the slope of the
ramp signal and the symbol bxc represents the floor function of x.

The spiral pattern centered at point (xr0 , yr0 ) is generated by the pair (rs (t), θs (t)) in polar
coordinates, where rs (t) and θs (t) are the instantaneous spiral radius and spiral angle at time
t, respectively. The following differential equation defines the spiral radius time evolution (see
[Mahmood and Moheimani, 2009a]):

drs (t) Ps ωs (t)

= (6.3)
dt 2π
with the angular velocity ωs and the spiral pitch Ps given by:
2Rs
Ps = (6.4)
Ns − 1
where Rs is the non-instantaneous spiral radius and Ns is defined as the number of times
the spiral curve crosses the line y(t) = yr0 (in Fig. 6.1f Ns = 8 and the crossing places are
numbered from 1 to 8). Separating variables and integrating both sides of equation (6.3) with
rs (0) = 0 gives:
Ps
rs (t) = ωs t (6.5)
2π
and the spiral angle at time t is expressed by:

θs (t) = ωs t (6.6)

Since the piezoelectric actuators are moved in X and Y directions, one needs to translate the
polar coordinates (rs (t), θs (t)) to the Cartesian ones as follows:

xr (t) = xr0 + rs (t)cos θs (t) (6.7)

yr (t) = yr0 + rs (t)sin θs (t) (6.8)

where xr0 = xr (0) and yr0 = yr (0) define a center in the Cartesian space of the spiral reference.
This center is chosen such that the whole spiral curve lies in positive quadrant of the Cartesian
space, since only the positive input voltages are used.
6.1. Introduction 103

7 8

6 7

5
reference x [V]

reference xr [V]
6
r

4
5
3
4
2

1 3

0 2
0 1 2 3 4 5 0 0.5 1 1.5 2 2.5 3 3.5
time [s] time [s]

(a) (b)

7 7

6 6
5
reference y [V]

reference yr [V]

5
r

4
4
3
3
2

1 2

0 1
0 1 2 3 4 5 0 0.5 1 1.5 2 2.5 3 3.5
time [s] time [s]

(c) (d)

7 7

6 Ps
6
5 rs(t)
reference y [V]

reference yr [V]

5
r

4 yr Rs θs(t)
4 0
1 2 3 4 5 6 7 8
3
3
2

1 2
xr
0
0 1
0 1 2 3 4 5 6 7 1 2 3 4 5 6 7 8
reference xr [V] reference xr [V]

(e) (f)

Figure 6.1: Signals used for the raster (left column) and the spiral (right column) pattern
generation: (a), (b) Reference xr (t). (c), (d) Reference yr (t). (e), (f) Reference pattern in
X-Y plane.
104 Chapter 6. 3D results for nanopositioning and STM applications

6.2 Nanopositioning in 3D - experimental validation

In this section the positioning of the tip in three dimensions is validated experimentally.
To that end, the 3D STM-like micro/nanopositioning platform developed in GIPSA-lab is
used. In the horizontal scanning directions the hysteresis and creep are eliminated using the
techniques developed in chapter 3. Due to much smaller displacements, the hysteresis effect
in the Z direction is negligible. The performance of open-loop MPI alone, closed-loop DOB
alone (based on the affine hysteresis approximation) and a hybrid approach (DOB, based on
MPI hysteresis approximation) is compared, while simultaneously controlling the tunneling
current in the Z direction, using pole placement controller with sensitivity functions shaping
[Landau and G., 2006], [Ahmad et al., 2012b], designed in chapter 5. The scanning frequency
in the horizontal X-Y axes is not too high in order not to cause coupling-based positioning
errors in the vertical Z axis and not to lose the tunneling current. Future work can be devoted
to identify and compensate these couplings. In real STM, the aim is to scan the surface in the
X-Y plane and to keep the tunneling current constant (as discussed in simulation in the next
section). In this way the surface topography is retrieved via the control signal. Here instead,
the "virtual surface" is imposed by the proper reference tunneling voltage zr (t) tracked while
moving in X-Y plane (nanopositioning application).

First test is done using DOB in both horizontal axes in order to eliminate the hysteresis
and creep phenomenon and the amplitude of the reference is small, which implies low signal
to noise ratio (the noise in the Z direction is reasonably high w.r.t. the amplitude of the
reference tunneling voltage). This reference zr (t) being sinusoidal variations between 0.25 V
and 1.25 V (which corresponds to tunneling current it (t) between 0.25 nA and 1.25 nA or the
distance d(t) between 9.3 Å and 8.3 Å, respectively) is shown in Fig. 6.2a and Fig. 6.2b in 3D
and 2D view, respectively. The raster scan is used in X-Y plane - a triangle signal (6.1) of
Ar =2 µm amplitude and slow frequency fr =0.5 Hz is applied to the piezoactuator in the X
direction and a linear ramp signal (6.2) in the Y direction as shown in Fig. 6.1e. In our case
both forward and backward movements along the X direction produce a line of the scanned
samples, while moving slowly in the Y direction. For the clarity of results the hysteresis and
creep are already compensated for the Y axis. 2D and 3D images of the "virtual surface"
without creep and hysteresis compensation for X axis for 1 µm ×1 µm (0.2 V × 0.2 V) range
are shown in Fig. 6.2c and Fig. 6.2d, respectively and with compensation in Fig. 6.2e and
Fig. 6.2f, respectively. The scanning trajectory without and with compensation is shown in
Fig. 6.2g and Fig. 6.2h, respectively. The difference between the reference image and the image
without compensation can be more pronounced when none of the two axes is compensated.
In X-Y plane, the obtained image with observer-based hysteresis and creep compensation
resembles the reference one better than wihout compensation (compare Fig. 6.2d and Fig. 6.2f
with Fig. 6.2b), however, the vertical resolution can be deteriorated due to the fact that the
observers work in a feedback loop, thus the measurement noise can be propagated to the
system input. This can be however improved as shown later using a low-pass filter on the
observers’ output as well as by properly tuned their corresponding covariance matrices. The
RS controller tracks the reference tunneling voltage well. However, our lab-made system at
ambient conditions has signal-to-noise ratio not as high as commercial STM, especially at this
6.2. Nanopositioning in 3D - experimental validation 105

(a) (b)

(c) (d)

(e) (f)

(g) (h)

Figure 6.2: Experimental results for 3D operation of the platform and low signal to noise
ratio: (a) Reference tunneling current (3D view). (b) Reference tunneling current (2D view).
(c) Tunneling current without hysteresis and creep compensation (3D view). (d) Tunneling
current without hysteresis and creep compensation (2D view). (e) Tunneling current with
hysteresis and creep compensation (3D view). (f) Tunneling current with hysteresis and creep
compensation (2D view). (g) X-Y trajectory without hysteresis and creep compensation. (h)
X-Y trajectory with hysteresis and creep compensation.
106 Chapter 6. 3D results for nanopositioning and STM applications

(a) (b)

Figure 6.3: Reference tunneling current: (a) 3D view. (b) 2D view.

Table 6.1: Tracking error-experimental result

MPI DOB MPI+DOB

rms (µm) max (µm) rms (µm) max (µm) rms (µm) max (µm)
1Hz 0.0588 0.1557 0.0042 0.0137 0.0022 0.0074
10Hz 0.0964 0.2045 0.0361 0.1334 0.0130 0.0590
20Hz 0.1157 0.2390 0.0760 0.2808 0.0274 0.1070
50Hz 0.1065 0.2839 0.1884 0.5086 0.0706 0.2490

sub-nanoscale. The system is sensitive to external disturbances, which can be sometimes seen
as sudden high jumps of the tunneling voltage. Furthermore, an increase of scanning speed
in the horizontal axes can cause coupling-based positioning errors in the vertical axis. This
makes the control of the all system very challenging.

In the second test, to increase signal to noise ratio, the disturbance observers were "well-
tuned" (their covariance matrices were set to give more trust in the model than to the measure-
ment). Moreover, a low pass filter has been added on the observers’ output and the amplitude
of the reference signal was increased (i.e. one may expect to see better the hills and valleys
of the signal when its amplitude is increased w.r.t. the noise). The sinusoidal reference zr (t)
between 1 V and 7 V (which corresponds to tunneling current it (t) between 1 nA and 7 nA or
the distance d(t) between 8.4 Å and 7.3 Å, respectively) is imposed as shown in Fig. 6.3a and
Fig. 6.3b in 3D and 2D view, respectively. The aim of this test is to compare techniques for
hysteresis and creep compensation developed in chapter 3, namely MPI, DOB and MPI/DOB
approaches. The spiral scanning pattern described in section 6.1 is used (see (6.7) and (6.8))
to move the tip in horizontal directions. The tracking of the spiral pattern in X-Y plane for
different scanning frequencies 2πωs is shown in Fig. 6.4. The values of the center of the spiral
reference (xr0 , yr0 ) = (5V, 4V) and the non-instantaneous radius Rs = 3V. For the clarity
of the presented results, the number of curves Ns = 8 (see equation (6.4)) was chosen. The
6.2. Nanopositioning in 3D - experimental validation 107

y2 [V]
3

ref
DOB
1
2 4 6 8
y [V]
1

(a) (b) (c) (d)

7

5
y2 [V]

ref
DOB
1
2 4 6 8
y [V]
1

(e) (f) (g) (h)

7

5
y2 [V]

2 ref
DOB
1
2 4 6 8
y [V]
1

(i) (j) (k) (l)

7

5
y2 [V]

ref
DOB
1
2 4 6 8
y1 [V]

(m) (n) (o) (p)

Figure 6.4: Tracking spiral trajectory of frequency 2πωs : (a)-(c) 1 Hz. (e)-(g) 10 Hz. (i)-(k)
20 Hz. (m)-(o) 50 Hz. (d), (h), (l), (p) Corresponding tracking errors.

corresponding RMS and maximal tracking errors are summarized in Table 6.1.

Comparing Fig. 6.4b with Fig. 6.4a and Fig. 6.4f with Fig. 6.4e, one can see the advantage
of DOB w.r.t. MPI approach for low frequencies. This is due to the fact that DOB works
in closed-loop, while MPI in open-loop. The performance of DOB worsens w.r.t. MPI model
when the input frequency is increased (compare Fig. 6.4j with Fig. 6.4i and Fig. 6.4n with
Fig. 6.4m). This can be explained by the fact that DOBs are based on low (second) order
model and their accuracy for higher frequencies can be improved by increasing this order,
which increases the model complexity. The accuracy can be also improved by combining DOB
108 Chapter 6. 3D results for nanopositioning and STM applications

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

(j) (k) (l)

Figure 6.5: Tunneling current obtained using only MPI (left), only DOB (middle) and com-
bined MPI+DOB (right) approach for the scanning frequencies 2πωs : (a)-(c) 1 Hz. (d)-(f) 10
Hz. (g)-(i) 20 Hz. (j)-(l) 50 Hz.
6.2. Nanopositioning in 3D - experimental validation 109

(a) (b)

(c) (d)

Figure 6.6: Tunneling current obtained for scanning frequency 1 Hz and MPI+DOB. (a)
Without any compensation for X and Y axes. (b) With compensation only for X axis. (c)
With compensation only for Y axis. (d) With compensation for both X and Y axes.

with MPI, which gives the best performance among all considered configurations as shown in
the third column of Fig. 6.4. The last column of Fig. 6.4 shows the corresponding tracking
errors. For correct image presentation, the number of curves has to be dense enough and
has been increased to Ns = 50. The corresponding images of the "virtual surface" in the Z
direction for different scanning frequencies are shown in Fig. 6.5. The advantage of hybrid
MPI+DOB approach over MPI or DOB alone is visible especially for higher frequencies (see
Fig. 6.5j-Fig. 6.5l). The level of deformation/improvement is indicated by dash-dotted lines.
The red and blue arrows point at two hills and two valleys, respectively. The MPI model
linearizes piezoelectric actuator and improves the performance of low-order observer, while the
presence of DOB assures robustness w.r.t. the hysteresis modeling uncertainties and external
disturbances. Next, MPI+DOB approach is used for 1 Hz scanning frequency and the result for
hysteresis and creep compensation is shown in Fig. 6.6. The adverse phenomena of hysteresis
and creep strongly deform the obtained image, when none of the axis is compensated as shown
in Fig.6.6a. Fig. 6.6b -Fig. 6.6d show the compensation results only for X axis, only for Y
110 Chapter 6. 3D results for nanopositioning and STM applications

axis and for both X and Y axes, respectively. One can see that after compensation the details
(hills and valleys) of the surface, generated by the tunneling voltage are equalized in terms
of shape and size and the obtained image from Fig. 6.6d can be compared with the reference
one from Fig. 6.3b.

6.3 Surface reconstruction in STM application in simulation

In the previous section, experimental results for nanopositioning in three dimensions were
given, where the tunneling current tracked the prespecified reference, called "virtual surface".
In this section the STM application is presented in simulation. The tunneling current is kept
on a constant level (constant current mode), while simultaneously scanning in X-Y plane. The
surface topography is retrieved via control signal in the Z direction. Cross-couplings from the
horizontal X and Y axes to the vertical Z axis are modeled and the results for two approaches
are presented: SISO approach (3 PID decentralized controllers together with decoupling com-
pensators) and MIMO approach (a centralized Linear Quadratic Integral Controller (LQI),
which takes into account the whole system model with cross-couplings).

Figure 6.7: Simulation plant model used for surface variation zs (t) reconstruction.

6.3.1 Plant simulation model

The plant model simulated in this subsection, shown in Fig. 6.7, was developed in the early
stage of the work and differs a little from the model considered in chapter 2. For example it does
not take into account the cross-couplings between the horizontal axes (i.e. xy (t) = yx (t) = 0).
As a result xp (t) = xx (t) + xy (t) = xx (t) and yp (t) = yy (t) + yx (t) = yy (t) in the full nonlinear
model block schema from Fig. 2.3. The hysteresis in X (resp. Y) direction is modeled as a
symmetrical nonlinearity Hx [vpx ](t) (resp. Hy [vpy ](t)) using classical Prandtl-Ishlinskii model
described in chapter 3 and is compensated via its inverse model cascaded in a feedforward
6.3. Surface reconstruction in STM application in simulation 111

path of the corresponding axis. The creep phenomenon for X (resp. Y) directions has been
modeled as a linear system Gcreepx (s) (resp. Gcreepy (s)), used only in the simulated plant
model. However, from the control point of view nothing changes and the creep is still assumed
to be a disturbance dx (t) and dy (t), compensated via feedback control (PID or LQI). This
phenomenon in the Z direction has been neglected since the surface variations are changing
rapidly.

Most of the subsystems are modeled as in chapter 2. Here, this model is recalled with
some modifications. Each horizontal motion is driven by the piezoelectric actuator fed by
the voltage amplifier and read by the capacitive sensor. Like before, the voltage amplifier is
described by the following low-pass system:
ẋ1,vx (t) = −ωvx x1,vx (t) + ux (t)
(6.9)
vpx (t) = Gvx ωvx x1,vx (t)
or in Laplace domain:
vpx (s) Gvx ωvx
V Ax (s) = = (6.10)
ux (s) s + ωvx
The piezoelectric actuator exhibits the adverse phenomenon of hysteresis Hx [vpx ](t) and also
tends to drift in response to a slowly varying input. This drift (creep phenomenon) in the
considered simulated plant is described by a 3rd order transfer function as in [Croft et al.,
2001] as follows:
s3 + 0.630s2 + 0.0216 + 6.25 · 10−5
Gcreepx (s) = 3 . (6.11)
s + 0.516s2 + 0.0171 + 4.69 · 10−5
The proper vibration part Dpx (s) of piezo is given by:
2
ωpx
Dpx (s) = 2 2
(6.12)
s + 2ξpx ωpx s + ωpx
This yields a nonlinear model made of a static nonlinearity Hx [vpx ], followed by the linear
dynamics Gcreepx (s)Dpx (s), known as a Hammerstein structure [Giri and Bai, 2010] and the
simulated piezo in the plant model from Fig. 6.7 is given by:

 qpx (t) = Hx [vpx (t)]

xx (s)
P iezox : qpx (s)
= Gcreepx (s)Dpx (s) (6.13)

xp (t) = xx (t)


However, for the controller design, the creep phenomenon is considered as a disturbance dx (t)
and the control model of piezo takes the following form:

 qpx (t) = Hx [vpx (t)] + dx (t)

xx (s) ω2
P iezox : qpx (s) = Dpx (s) = s2 +2ξpx ωpxpx s+ω2 (6.14)
 px
xp (t) = xx (t)


or in state space with state variables xpx,1 (t) = xp (t) and xpx,2 (t) = ẋp (t):
ẋ1,px (t) = x2,px (t)
ẋ2,px (t) = −ωpx2 x 2
1,px (t) − 2ξpx ωpx x2,px (t) + ωpx qpx (t) (6.15)
xp (t) = x1,px (t)
112 Chapter 6. 3D results for nanopositioning and STM applications

In [Croft et al., 2001] an inverse-based feedfoward control approach is used to compensate for
hysteresis, creep and vibration parts of piezo. Here instead, only hysteresis is compensated
using feedforward inverse-based classical Prandtl-Ishlinskii model, and then the feedback con-
trol is used to compensate for the creep and vibration.
Finally, the capacitive sensor is described as in chapter 2 as follows:

ẋ1,capx (t) = −ωcapx x1,capx (t) + xp (t)

(6.16)
vx (t) = Gcapx ωcapx x1,capx (t)
or in Laplace domain:
vx (s) Gcapx ωcapx
CSx (s) = = (6.17)
xp (s) s + ωcapx
The final sensor output is disturbed by measurement noise as follows:

y1 (t) = vx (t) + n1 (t) (6.18)

Identical modeling is done for the Y axis, and is not given here for brevity.

The modeling in the Z direction is the same as in chapter 2 with the difference that the
creep phenomenon dz (t) of piezo is neglected in this case, since it is assumed that the surface
variations are changing fast enough. The voltage amplifier is described as follows:

ẋ1,vz (t) = −ωvz x1,vz (t) + uz (t)

(6.19)
vpz (t) = Gvz ωvz x1,vz (t)
or in Laplace domain:
vpz (s) Gvz ωvz
V Az (s) = = (6.20)
uz (s) s + ωvz
The piezoelectric actuator of the gain Gpz and the bandwidth ωpz in the Z axis together with
cross-couplings from the horizontal axes to the vertical axis (Zx (s) and Zy (s)) is described by
the following linear model (the displacements along this axis are very small and the effect of
hysteresis is negligible):


 qpz (t) = −Gpz vpz (t)

 ẋ1,pz (t) = x2,pz (t)




2 x
ẋ2,pz (t) = −ωpz 2
D (s) : 1,pz (t) − 2ξpz ωpz x2,pz (t) + ωpz qpz (t)



 pz
 


 zz (t) = x1,pz (t)

P iezoz : ẋ1,zx (t) = −ωzx x1,zx (t) + xp (t) (6.21)
Zx (s) :
= −Gzx ωzx xzx,1 (t) + Gzx xp (t)

 z (t)
 x




 ẋ1,zy (t) = −ωzy x1,zy (t) + yp (t)
 Zy (s) :




 zy (t) = −Gzy ωzy x1,zy (t) + Gzy yp (t)

zp (t) = zz (t) + zx (t) + zy (t)

where the cross-coupling subsystems Zx (s) and Zy (s) are the first order high-pass systems
expressed in Laplace domain as follows:
zx (s) Gzy s
Zx (s) = = (6.22)
xp (s) s + ωzy
6.3. Surface reconstruction in STM application in simulation 113

zy (s) Gzy s
Zy (s) = = (6.23)
yp (s) s + ωzy
where Gzx , ωzx (resp. Gzy , ωzy ) are the gain and the bandwidth of Zx (s) (resp. Zy (s)).

The physical tunneling law is the heart of the considered STM-like system and is modeled like
before as the exponential dependency of the tunneling current on the distance tip/surface d(t),
when the latter is less than 1nm and the bias voltage Vb is applied between the two materials
as follows:
gVb e−kd(t) , 0 < d(t) ≤ 1 nm

0
P Lt : it (t) = (6.24)
0, d(t) > 1 nm
where the distance d(t) is given by:

d(t) = d0 + zp (t) − zs (t) (6.25)

with initial distance d0 between the tip and the surface, piezo displacement zp (t) (disturbed
by the cross-coupling phenomena) and surface variations zs (t).
The tunneling current is captured and amplified by the high-gain sensor of the gain Gt and
the bandwidth ωt modeled like before as follows:

ẋ1,t (t) = x2,t (t)

ẋ2,t (t) = −ωt2 x1,t (t) − 2ξt ωt x2,t (t) + ωt2 Gt it (t) (6.26)
vz (t) = x1,t (t)

or in Laplace domain:
vz (s) Gt ωt2
CSt (s) = = 2 (6.27)
it (s) s + 2ξt ωt + ωt2
providing a noisy output:
y3 (t) = vz (t) + n3 (t) (6.28)

6.3.2 SISO approach - 3 PID controllers with decoupling compensators

Fig. 6.8 shows the block schema of the considered SISO controlled system. As mentioned
before, the inverse Prandtl-Ishlinskii model (Hx−1 and Hy−1 ) is used to linearize piezoelectric
actuators in both horizontal directions and such a linearized system is controlled by two PID
controllers, which compensate the creep phenomena as well. The third PID controller in the Z
direction keeps the tip/surface distance on a desired constant level. The cross-couplings Zx (s)
and Zy (s) from the horizontal axes to the vertical one can be quite strong, thus a proper
compensation is looked for in order to obtain correct image of the surface topography.

As in [Shi et al., 2009], for X to Z coupling identification, the Y axis remains unexcited
(i.e. the reference yr is set to 0), which means that there is no influence on the Z axis coming
from the Y axis (i.e. zy (t) ≈ 0). It is assumed also that the scanned surface is chosen to
be "more or less" flat (i.e. zs (t) ≈ 0). The reason behind this is that the surface variations
zs (t), as well as coupling positioning errors zx (t), zy (t) are viewed as external disturbances on
the tip position zp (t) as shown in Fig. 6.7. The coupling from Y to Z axis is identified in the
114 Chapter 6. 3D results for nanopositioning and STM applications

Figure 6.8: STM-like controlled system with decentralized PID SISO controllers and decou-
pling compensators.

similar way.
Assuming both capacitive sensors CSx (s) and CSy (s) as static gains one obtains:
y1 (t) y2 (t)
xp (t) ≈ x̂p (t) = , yp (t) ≈ ŷp (t) = (6.29)
Gcapx Gcapy
The transfer functions describing the identified coupling models are given by:
ẑx (s) Ĝzx s ẑy (s) Ĝzy s
Ẑx (s) = = , Ẑy (s) = = (6.30)
x̂p (s) s + ω̂zx ŷp (s) s + ω̂zy

where Ĝzx , ω̂zx , Ĝzy , ω̂zy are identified parameters of the coupling subsystems. Neglecting
the dynamics of the piezoactuator Dpz (s) and the dynamics of the voltage amplifier V Az (s)
in the Z direction one obtains:
ẑz (s)
Ĝzz (s) = = −Gpz Gvz (6.31)
uz (s)
The control signal in the Z direction uz (t) consists of two parts: the part that is used for
cross-couplings compensation uc (t) and the proper (non-compensating) part us (t) that is used
to adjust the tip/surface distance when the cross-couplings are compensated as follows:

uz (t) = uc (t) + us (t) = ucx (t) + ucy (t) + us (t) (6.32)

where the compensating components for Zx (s) and Zy (s) couplings are given by:

Ẑx (s) Ẑy (s)

ucx (t) = − x̂p (t), ucy (t) = − ŷp (t) (6.33)
Ĝzz (s) Ĝzz (s)
6.3. Surface reconstruction in STM application in simulation 115

After cross-couplings compensation one can assume that zp (t) ≈ zz (t) ≈ −Gpz Gvz us (t). Ne-
glecting the dynamics of the current sensor CSt (s), the estimated distance can be expressed
as follows:  
ˆ = − 1 ln 1 y3 (t)
d(t) (6.34)
k gVb Gt
Finally, the surface is reconstructed from the current sensor output y3 (t) and the proper part
of the control signal us (t) as follows:
 
ˆ 1 1 y3 (t)
ẑs (t) = d0 + zp (t) − d(t) = fs (us (t), y3 (t)) = d0 − Gpz Gvz us (t) + ln (6.35)
k gVb Gt
The results for the surface reconstruction are given in subsection 6.3.4 together with the results
obtained for MIMO design, for comparison.

6.3.3 MIMO approach - Linear Quadratic Integral Controller (LQI)

In this subsection, a Linear Quadratic Integral (LQI) controller is designed. It is Linear

Quadratic Gaussian controller with an integral action (i.e. LQG for tracking). The technique
is a model-based approach, and is based on the model developed in subsection 6.3.1, with
a state vector x(t) ∈ R(15×1) [composed of the following state variables: x1,vx (t), x1,px (t),
x2,px (t), x1,capx (t), x1,vy (t), x1,py (t), x2,py (t), x1,capy (t), x1,vz (t), x1,pz (t), x2,pz (t), x1,zx (t),
x1,zy (t), x1,t (t) and x2,t (t)], an output vector y(t) ∈ R(3×1) made of y1 (t), y2 (t) and y3 (t), and
an input vector u(t) ∈ R(3×1) made of u1 (t), u2 (t) [assuming hysteresis compensation on X
and Y axes as before - see Fig. 6.9] and u3 (t) = uz (t) is considered. The model, still nonlinear
with exponential tunneling nonlinearity (6.24), is then linearized around an equilibrium point
(deq , ieq ) as in chapter 5 as follows:

it (t) − ieq = −kieq (d(t) − deq ) (6.36)

resulting in a final linear model used for the LQI controller design:

∆ẋ(t) = A∆x(t) + B∆u(t) + Bd zs (t) + w(t)

(6.37)
∆y(t) = C∆x(t) + n(t)

where ∆x(t) = x(t) − xeq , ∆u(t) = u(t) − ueq , ∆y(t) = y(t) − yeq are deviations of re-
spective variables around their equilibrium. It is assumed that process noise w(t) and mea-
surement noise n(t) are zero-mean Gaussian noises with covariances W = E[w(t)w(t)T ] and
N = E[n(t)n(t)T ], respectively (E[x] stands for the expected value of x). A classical LQG
controller is an interconnection of a Kalman estimator and a Linear Quadratic Regulator
(LQR), which is used for a stabilization around zero. The purpose is to scan the surface in in
X-Y plane and to keep the tunneling current on a constant non-zero level in the Z direction,
hence the LQG with an integral action (referred here as LQI) is adopted as shown in Fig. 6.9.
The reference vector r(t) is composed of the references for each direction: xr (t), yr (t), zr (t).

Observer Design. The surface variations zs (t) are considered as an unknown disturbance
on the tip position zp (t). The aim of the designed observer is to reconstruct not only the state
116 Chapter 6. 3D results for nanopositioning and STM applications

Figure 6.9: STM-like controlled system with centralized LQI MIMO controller.

x(t) of the system, but also the surface variations. If the observer parameters are chosen in
such a way that its dynamics is sufficiently fast, the surface variations can be assumed con-
stant (i.e. żs (t) ≈ 0) from the controller point of view (i.e. the observer reconstruction speed
is much faster than the surface variations). The covariance matrices W and N are chosen to
guarantee both stability and proper dynamics of this observer.

The state vector for reconstruction is extended into ∆xe (t) = [∆x(t) zs (t)]T and is a part
of the following state space representation:

∆ẋe (t) = Ae ∆xe (t) + Be ∆u(t) + w(t)

(6.38)
∆y(t) = Ce ∆xe (t) + n(t)

with the state space matrices defined as follows:

   
A Bd B

Ae = , Be = , Ce = C 0 (6.39)
0 0 0

The steady-state Kalman observer for such observable system is given by:

∆x̂˙ e (t) = Ae ∆x̂e (t) + Be ∆u(t) + L(∆y(t) − Ce ∆x̂e (t)) (6.40)

where ∆x̂e (t) is an optimal state estimate of (6.38) in terms of minimizing the mean square
error E[∆x̂e (t) − ∆xe (t)] if the observer gain matrix L = P CeT N −1 is properly chosen. The
matrix P is a solution of the following Algebraic Riccati Equation (ARE):

Ae P + P ATe − LN LT + W = 0 (6.41)
6.3. Surface reconstruction in STM application in simulation 117

Controller Design. As mentioned before, the LQG has to be adapted to the tracking
problem in order to ensure that the output y(t) tracks the reference signal r(t), which can be
easily done via extending the classical LQG feedback from the system state estimate ∆x̂(t)
(the state estimate ∆x̂(t) is used and not the extended state estimate ∆x̂e , since zs (t) is not
controllable) with the feedback from integrated error xi (t) such that ∆u(t) = −Kxc (t), where
xc (t) = [∆x̂(t) xi (t)]T . The integrated error xi (t) is given by:
Z Z
xi (t) = e(t)dt = [r(t) − y(t)]dt (6.42)

The following cost function is minimized:

Z ∞
J(u(t)) = [xc (t)T Qxc (t) + ∆u(t)T R∆u(t)]dt (6.43)
0

with the feedback matrix K = R−1 BcT S, where S is a solution of the following ARE:

ATc S + SAc − K T RK + Q = 0 (6.44)

   
A 0 B
Ac = , Bc = (6.45)
−C 0 0
where Q is a positive semi-definite weighting matrix for the state with integral action xc (t)
and R is a positive definite weighting matrix for the control ∆u(t).

6.3.4 Simulation results for surface reconstruction

The final simulation results for surface reconstruction in the STM-like system for both SISO
and MIMO design are presented in this subsection. The numerical values used for simulation
are summarized in Table 6.2. Most of them are with a good accordance with the parameters’
values of the experimental platform under development in GIPSA-lab (see e.g. [Ahmad et al.,
2012b], [Ryba et al., 2013]). The raster scan is used to scan the surface in X-Y plane, while
simultaneously keeping the tunneling current on the constant level (here 0.9 nA). The X axis
is assumed as the fast scanning axis, where a triangle signal of amplitude 1 µm and frequency
100 Hz is applied and the Y axis is the slow axis, where a linear ramp is applied, which moves
the tip to the next line in a raster pattern. It is assumed that the data are collected only
during the forward movement of the tip (rising part of the triangle).

The influence of different adverse phenomena (cross-couplings, hysteresis, creep and vibra-
tion) on the surface reconstruction are investigated here for both SISO and MIMO designs from
the previous subsections. It is assumed that all these phenomena are already compensated for
the Y axis and subsequently eliminated for the X axis. Fig. 6.10a shows the reference surface
topography being the sinusoidal variations of amplitude 5 Å. Due to large cross-coupling from
the horizontal axes to the vertical one, the image of the surface without any compensation can
be completely deteriorated as shown in Fig. 6.10b. Two decoupling compensators designed in
SISO design in subsection 6.3.2 were switched on and the obtained image shown in Fig. 6.10c
118 Chapter 6. 3D results for nanopositioning and STM applications

Table 6.2: System parameters used for simulation.

ωvx , ωvy , ωvz Bandwidth of voltage amplifier 4 kHz

Gvx , Gvy , Gvz Gain of voltage amplifier 15 V/V
ωpx , ωpy , (ωpz ) Bandwidth of piezoactuator 630, (120000) Hz
Gpx , Gpy , (Gpz ) Gain of piezoactuator 25, (1.2) nm/V
ξpx , ξpy , ξpz Damping of piezoactuator 0.9
ωcapx , ωcapy Bandwidth of capacitive sensor 8.5 kHz
Gcapx , Gcapy Gain of capacitive sensor 200 V/mm
ωt Bandwidth of current sensor 13 kHz
Gt Gain of current sensor 109 V/A
d0 Initial distance tip-surface 1 nm
Vb Biased voltage 1.025 V
g Tunneling current constant 0.0011
k Tunneling current constant 1.65 Å−1
ieq Equilibrium tunneling current 0.9 nA
deq Equilibrium distance tip-surface 8.5 Å
ωzx , ωzy Bandwidth of cross-coupling 100 Hz
Gzx , Gzy Cross-coupling gains (maximum positioning error 3%
w.r.t. scanning displacement)

resembles the reference sinusoidal surface variations. The difference is due to other remaining
phenomena of hysteresis, creep and vibrations. Fig. 6.10d shows the image after hysteresis
compensation and Fig. 6.10e after the creep and vibration compensation. These results have
been obtained for the SISO design with three PID controllers and in this classical approach
the surface is reconstructed from the control signal of vertical piezo [Abramovitch et al., 2007].
However, in case of large model uncertainties and measurement noise, the performance of this
approach deteriorates. In case of MIMO design described in subsection 6.3.3, the surface is
reconstructed via the state observer together with the system state. Since the whole model is
taken into account, including the cross-coupling phenomena and the level of noise is decreased
by properly chosen covariance matrices W and V , the reconstructed surface is better than
in SISO case as shown in Fig. 6.10f. The difference can be seen also in tunneling voltage
stabilization at zr =0.9V (corresponding to 9 nA of tunneling current) as shown in Fig. 6.11.

6.4 Conclusion

In this chapter two applications have been investigated: nanopositioning validated experimen-
tally on the 3D lab-made micro-/nanopositioning device, and STM-like surface reconstruction
tested in simulation. In the first application, the piezoactuated tip followed a prespecified
trajectory in X, Y and Z directions. In the horizontal X and Y axes raster and spiral tra-
jectories were used, and in the vertical Z direction the prespecified sinusoidal trajectory has
been imposed playing the role of "virtual surface". First, the amplitude of this surface was
chosen to be small enough w.r.t. the measurement noise (low signal to noise ratio). The ob-
tained images were influenced by noise, however the surface details can be well-distinguished.
6.4. Conclusion 119

(a) (b)

(c) (d)

(e) (f)

Figure 6.10: Surface reconstruction: (a) Reference surface variations zr (t). (b) Without any
compensation. SISO approach with subsequent compensation of: (c) cross-couplings, (d)
hysteresis, (e) creep and vibrations. (f) MIMO approach.
120 Chapter 6. 3D results for nanopositioning and STM applications

(a) (b)

Figure 6.11: Tunneling current: (a) SISO approach, (b) MIMO approach.

The low-pass filters added on the observers’ output as well as increased signal to noise ratio
improved significantly the presented results and the performance of the following compen-
sation methods for hysteresis and creep was compared: Modified Prandtl-Ishlinskii (MPI),
disturbance observer (DOB) and hybrid MPI + DOB. Illustrative results show, that all of the
approaches give significantly better results than without any compensation among which the
hybrid MPI+DOB approach gives the best results.

The second application - STM-like surface reconstruction was tested in simulation. Two
control strategies were analyzed: decentralized SISO control, where three PID controllers were
used (each for one axis of the considered system) and MIMO centralized control using Linear
Quadratic Integral (LQI) controller. In the first approach the surface was reconstructed via
control signal of the vertical piezo, while in the second approach estimated via state observer.
In both approaches the hysteresis was first compensated using classical Prandtl-Ishlinskii
model. Illustrative surface images have been presented for subsequent cross-couplings, hystere-
sis, creep and vibration compensation for SISO approach. The MIMO approach outperforms
its SISO counterpart in terms of both surface reconstruction and tunneling current stabiliza-
tion, since it is based on the whole system model (including cross-coupling phenomena) and
appears more robust to model uncertainties and measurement noises.
 Chapter 7

Cantilever nanopositioning with

electrostatic actuation

Contents
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
7.2 Multi-mode cantilever model analysis . . . . . . . . . . . . . . . . . . . 122
7.3 Tunneling current-based cantilever nanopositioning with electrostatic
actuation and piezoelectric tip actuation . . . . . . . . . . . . . . . . . 132
7.3.1 H∞ design for tunneling gap control via electrostatically actuated cantilever134
7.3.2 Cantilever positioning via piezoelectrically actuated tip with proximity
force compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
7.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

7.1 Introduction

This chapter is devoted to modeling and control of the electrostatically actuated cantilever
based on the tunneling current measurement. The idea is to consider here the (multi-mode)
dynamics of the cantilever. The electrostatic actuator is used to keep the tunneling gap con-
stant, while the piezoelectric actuator adjusts the tip (and as a result positions the cantilever,
since the gap cantilever/tip is already maintained constant). This extends the former study
of [Blanvillain et al., 2014], which considered a one-mode cantilever model without proximity
force compensation. The numerical data (such as cantilever dimensions) used for simulation
have been taken from the cantilever provider from Nano World company and are summa-
rized in Table 7.1 (tc stands for the cantilever thickness in order to avoid confusion with the
time t). The multi-mode model of the cantilever from [Salapaka et al., 1997] is borrowed
for further studies and analysis for n = 3 first modes. Some "singular points" of the modal
shape functions along horizontal x position of the cantilever are emphasized and used for
actuator/sensor location. Near these singular points the model is degenerating (i.e. the cor-
responding shaped modes are vanishing) and as a result the model order can be reduced. The
success in damping the resonances for the piezoelectric actuators in the horizontal axes, shown
in chapter 4, motivated the use of H∞ approach to control the tunneling gap via resonant
cantilever. The proximity force is approximated from the tunneling current measurement and
then compensated via PID controller which actuates the tip via piezoelectric actuator.

121
122 Chapter 7. Cantilever nanopositioning with electrostatic actuation

Table 7.1: Numerical data concerning the cantilever.

w Cantilever width 100 µm

L Cantilever length 500 µm
tc Cantilever thickness 1 µm
E Cantilever Young’s modulus of elasticity 180 GPa
Q the Q factor 33
n Number of first modes used for simulation 3

7.2 Multi-mode cantilever model analysis

In chapter 2 the equations describing the multi-mode model of the cantilever have been pre-
sented. In this section this model is used and simulated for the first 3 modes. Recalling
equation (2.30), the cantilever motion subject to proximity force −Fp (t) and damping force
pd (x, t) (constant per unit length) can be described as follows (see Fig. 7.1a):

∂ 4 zc (x, t) ∂ 2 zc (x, t)
EI + pd (x, t) + ρA = −Fp (t) (7.1)
∂x4 ∂t2
where I = wt3c /12 is the moment of inertia of the cantilever and A = wtc is its cross sectional
area. The force Fp (t) is the interaction force between the tip and the cantilever and is acting

−𝐹𝑝 (𝑡)
𝑧𝑐 (𝑥, 𝑡) 𝐹𝑐 (𝑡) = −𝐹𝑝 (𝑡) + 𝐹𝑒 (𝑡)
𝑧𝑐 (𝑥, 𝑡)
𝑥𝑓 𝑥
𝑥𝑓 𝑥
𝑝𝑑 (𝑥, 𝑡)
𝐿
𝐿
𝑧 𝑧

𝑥 𝑥
(a) (b)

Figure 7.1: Schematic showing the cantilever: (a) Subjected to the external force of value
−Fp (t) acting at a distance xf from the base of the cantilever with the the damping force
pd (x, t), assumed constant per unit length. (b) The equivalent schematic with the total net
force Fc (t) = −Fp (t) + Fe (t).

at the distance xf from the base of the cantilever. The damping force pd (x, t) is the sum of
the natural structural damping force of the cantilever and the external damping force (here
the electrostatic force Fe (t)) as follows:

∂zc (x, t)
pd (x, t) = ξ − Fe (t) (7.2)
∂t
7.2. Multi-mode cantilever model analysis 123

p
with the damping factor ξ = 3.515/Q EIρA/L4 (assumed constant for all modes). Inserting
(7.2) into (7.1) gives (see Fig. 7.1b):

∂ 4 zc (x, t) ∂zc (x, t) ∂ 2 zc (x, t)

EI 4
+ξ + ρA = Fc (t) (7.3)
∂x ∂t ∂t2 | {z }
−Fp (t)+Fe (t)

where Fc (t) = −Fp (t) + Fe (t) is the total net force acting at a distance xf from the base of
the cantilever, which can be distributed over the cantilever length L.

The cantilever displacement is expressed as the sum of the fundamental deformations qj (t)
shaped by the modal shape functions φj (x) (j = 1..3) as follows:
3
X
zc (x, t) = φj (x)qj (t) (7.4)
j=1

where φj (x) is given by (2.34) and qj (t) is the solution of (2.38). The displacement depends
not only on the time t but also on the horizontal position x along the cantilever. However,
as mentioned in chapter 2, the modal shape function φj (x) allows to distribute in space the
modal contributions qj (t) and it only depends on position, while the modal part depends only
on time. Recalling equation (2.50), the cantilever displacement in Laplace domain can be
expressed as follows (assuming zero initial conditions qj (0) = q̇j (0) = 0):
3
X φj (xf ) 1
zc (x, s) = φj (x) cj kj
Fc (s) (7.5)
s2 + mj
j=1 mj s + mj

with the modal parameters defined by (2.39)-(2.42). The computed parameters for the con-
sidered 3-mode model are summarized in Table 7.2. Note that one can rewrite equation (7.5)
as follows:
3
X 1
zc (x, s) = mj cj kj
Fc (s) (7.6)
2
j=1 φj (x)φj (xf ) s + φj (x)φj (xf ) s + φj (x)φj (xf )

Now taking only the first mode of (7.6) one obtains the equivalent single-mode model, at

Table 7.2: Parameters computed for the first three modes.

j λj φj (L) mj kj cj
1 0.3798 × 104 -8.5673 2.1446 × 10−9 2.7311 2.3184 × 10−6
2 0.9392 × 104 107.6972 3.3694 × 10−7 1.6850 × 104 3.6424 × 10−4
3 1.5710 × 104 −2.5803 × 103 1.9389 × 10−4 7.6025 × 107 2.0961 × 10−1

the force location xf and sensor location x, with the following parameters:

m1 c1 k1
msingle = , csingle = , ksingle = (7.7)
φ1 (x)φ1 (xf ) φ1 (x)φ1 (xf ) φ1 (x)φ1 (xf )
124 Chapter 7. Cantilever nanopositioning with electrostatic actuation

0 150

−2 100

−4 50
φ1(x)

φ2(x)
−6 0

−8 −50

−10 −100
0 100 200 300 400 500 0 100 200 300 400 500
x [µm] x [µm]

(a) (b)

2000

1000

0
φ3(x)

−1000

−2000

−3000
0 100 200 300 400 500
x [µm]

(c)

Figure 7.2: Modal shape functions for 3 modes with singular points (pointed in red): (a) 1st
mode. (b) 2nd mode. (c) 3rd mode.

x = 500 µm, x =500 µm (L = 500 µm)

f
100 60
40
Magnitude (dB)

Magnitude (dB)

50
20
0 0
3 mode model
−20
−50 2 mode model
−40 1 mode model
−100 −60
0 0
1st mode
−45 2nd mode
Phase (deg)

Phase (deg)

3rd mode
−90 −90

−135

−180 −180
3 4 5 6 3 4 5 6
10 10 10 10 10 10 10 10
Frequency (Hz) Frequency (Hz)

(a) (b)

Figure 7.3: Frequency response of (a) shaped modal and (b) total displacement (sum of shaped
modes) when the applied input force and measured output displacement are at the distance
x = L = 500 µm (i.e. cantilever end).
7.2. Multi-mode cantilever model analysis 125

Three resonant frequenciesp were found using

p parameters from Table 7.2 according to the
2
formula fj = ωj /(2π) = kj /mj /(2π) = λj EI/(ρA)/(2π) and they are equal to 5.68 kHz,
35.59 kHz and 99.65 kHz. The force constant at the end of cantilever is given by the formula
kf (L) = Ewt3c /(4L3 ) = 0.036 N/m.

x = 1 µm, xf =500 µm (L = 500 µm) x = 252 µm, xf =500 µm (L = 500 µm)

100 100
1st mode 1st mode
80 80
2nd mode 2nd mode
rd rd
60 3 mode 60 3 mode

40 40
Magnitude (dB)

Magnitude (dB)
20 20

0 0

−20 −20

−40 −40

−60 −60

−80 −80

−100 3 4 5 6 7
−100 3 4 5 6 7
10 10 10 10 10 10 10 10 10 10
Frequency (Hz) Frequency (Hz)

(a) (b)

x = 391.5 µm, xf =500 µm (L = 500 µm) x = 434 µm, xf =500 µm (L = 500 µm)
100 100
1st mode 1st mode
80 nd 80 nd
2 mode 2 mode
rd rd
60 3 mode 60 3 mode

40 40
Magnitude (dB)

Magnitude (dB)

20 20

0 0

−20 −20

−40 −40

−60 −60

−80 −80

−100 3 4 5 6 7
−100 3 4 5 6 7
10 10 10 10 10 10 10 10 10 10
Frequency (Hz) Frequency (Hz)

(c) (d)

Figure 7.4: Magnitude plot of different shaped modes at singular points, when the applied
force is at xf = 500 µm (cantilever end): (a) x = 1 µm ≈ 0 µm. (b) x = 252 µm. (c)
x = 391.5 µm. (d) x = 434 µm.

Analyzing the model (7.5), it can be noticed that depending on the position x, the contri-
bution of the j th mode to the total displacement is different and can vanish completely when
its corresponding modal shape function φj (x) is equal to zero. Here, the positions where this
happens are called the singular points. Fig. 7.2 shows the modal shape functions φj (x) for
the considered model. The singular points are pointed by the red points and it appears that
j th mode has j singular points. It can be also noticed that all modes have singular point at
126 Chapter 7. Cantilever nanopositioning with electrostatic actuation

x = 1 µm, xf =500 µm (L = 500 µm) x = 252 µm, xf =500 µm (L = 500 µm)

100 100
3 mode model 3 mode model
80 2 mode model 80 2 mode model
1 mode model 1 mode model
60 60
40 40
Magnitude (dB)

Magnitude (dB)
20 20
0 0
−20 −20
−40 −40
−60 −60
−80 −80
−100 −100
−120 3 4 5 6 7
−120 3 4 5 6 7
10 10 10 10 10 10 10 10 10 10
Frequency (Hz) Frequency (Hz)

(a) (b)

x = 391.5 µm, x =500 µm (L = 500 µm) x = 434 µm, x =500 µm (L = 500 µm)
f f
100 100
3 mode model 3 mode model
80 2 mode model 80 2 mode model
1 mode model 1 mode model
60 60
40 40
Magnitude (dB)

Magnitude (dB)

20 20
0 0
−20 −20
−40 −40
−60 −60
−80 −80
−100 −100
−120 3 4 5 6 7
−120 3 4 5 6 7
10 10 10 10 10 10 10 10 10 10
Frequency (Hz) Frequency (Hz)

(c) (d)

Figure 7.5: Magnitude plot of the cantilever total displacement at singular points, when the
applied force is at xf = 500 µm (cantilever end): (a) x = 1 µm ≈ 0 µm. (b) x = 252 µm. (c)
x = 391.5 µm. (d) x = 434 µm.

x = 0 µm (cantilever fixed part). Fig. 7.3a shows the bode diagram of the shaped modes
individually and Fig. 7.3b the bode diagram of the whole model (sum of the one, two and
three shaped modes) from the input force Fc (t) to the output cantilever displacement zc (x, t),
where both the applied input force and captured output displacement are taken at the end
of cantilever (x = xf = L = 500 µm). Fig. 7.4 shows the frequency response (magnitude
plot) of the individual shaped modes at singular points, when the input force is applied at the
cantilever end (xf = L). It can be clearly seen that all the modes are vanishing at x = 0 µm
(see Fig. 7.4a), the second mode vanishes at x = 391.5 µm (see Fig. 7.4c) and the third one at
x = 252 µm and x = 434 µm (see Fig. 7.4b and Fig. 7.4d, respectively). The corresponding
frequency response for the total displacement is shown in Fig. 7.5. One can see that placing
7.2. Multi-mode cantilever model analysis 127

x = 1 µm, xf =252 µm (L = 500 µm) x = 252 µm, xf =252 µm (L = 500 µm)

100 100
1st mode 1st mode
2nd mode 2nd mode
50 3rd mode 50 3rd mode
Magnitude (dB)

Magnitude (dB)
0 0

−50 −50

−100 −100

−150 3 4 5 6 7
−150 3 4 5 6 7
10 10 10 10 10 10 10 10 10 10
Frequency (Hz) Frequency (Hz)

(a) (b)

x = 391.5 µm, xf =252 µm (L = 500 µm) x = 434 µm, xf =252 µm (L = 500 µm)
100 100
1st mode 1st mode
2nd mode 2nd mode
50 3rd mode 50 3rd mode
Magnitude (dB)

Magnitude (dB)

0 0

−50 −50

−100 −100

−150 3 4 5 6 7
−150 3 4 5 6 7
10 10 10 10 10 10 10 10 10 10
Frequency (Hz) Frequency (Hz)

(c) (d)

Figure 7.6: Magnitude plot of different shaped modes at singular points, when the applied
force is at xf = 252 µm: (a) x = 1 µm ≈ 0 µm. (b) x = 252 µm. (c) x = 391.5 µm. (d)
x = 434 µm.

the displacement sensor (if such exists) at the position near the singular point can damp the
associated mode and in a consequence reduce the model order. These results have been ob-
tained for the fixed position for the applied force (xf = L). What happens when this position
is also placed in one of the singular points (i.e. when φj (xf ) = 0)? From model (7.5) it
appears immediately that the j th mode contribution will vanish as well in this case. Fig. 7.6
shows the frequency response (magnitude plot) of the shaped modes at singular points in such
a case, where the input force is applied at the position xf = 252 µm (one of the singular
points for the third mode). One can notice, that at each plot the third mode is vanishing,
thus its contribution to the total displacement (see Fig. 7.7) is almost zero. Moreover, at the
128 Chapter 7. Cantilever nanopositioning with electrostatic actuation

x = 1 µm, xf =252 µm (L = 500 µm) x = 252 µm, xf =252 µm (L = 500 µm)

100 100
3 mode model 3 mode model
2 mode model 2 mode model
1 mode model 1 mode model
50 50
Magnitude (dB)

Magnitude (dB)
0 0

−50 −50

−100 −100

−150 3 4 5 6 7
−150 3 4 5 6 7
10 10 10 10 10 10 10 10 10 10
Frequency (Hz) Frequency (Hz)

(a) (b)

x = 391.5 µm, x =252 µm (L = 500 µm) x = 434 µm, x =252 µm (L = 500 µm)
f f
100 100
3 mode model 3 mode model
2 mode model 2 mode model
1 mode model 1 mode model
50 50
Magnitude (dB)

Magnitude (dB)

0 0

−50 −50

−100 −100

−150 3 4 5 6 7
−150 3 4 5 6 7
10 10 10 10 10 10 10 10 10 10
Frequency (Hz) Frequency (Hz)

(c) (d)

Figure 7.7: Magnitude plot of the cantilever total displacement at singular points, when the
applied force is at xf = 252 µm: (a) x = 1 µm ≈ 0 µm. (b) x = 252 µm. (c) x = 391.5 µm.
(d) x = 434 µm.

location x = 391.5 µm, the second mode is vanishing as well (see Fig. 7.6c) and as a result the
corresponding model for the total displacement can be reduced in this case to the first mode
only as shown in Fig. 7.7c. From this observation one can conclude that the model order can
be reduced by both sensor and actuator locations.

Fig. 7.8a and Fig. 7.8c show the input force being the step function with random Gaussian
noise with signal to noise ratio SNR=30 dB and SNR=15 dB, respectively while Fig. 7.8b
and Fig. 7.8d show the corresponding cantilever total displacement with modal contributions
(forced at xf = L = 500 µm and measured at x = 150 µm). Fig. 7.9 shows the case
corresponding to Fig. 7.7 where the force (of SNR = 15 dB) from Fig. 7.8c is kept at xf =
7.2. Multi-mode cantilever model analysis 129

−8 −8
x 10 x 10
1.4 8
total zc
7 st
1 mode
1.2
nd
2 mode
6 rd
3 mode
1

displacement [m]
5
force [nN]

0.8 4

0.6 3

2
0.4
1
0.2
0

0 −1
0 0.002 0.004 0.006 0.008 0.01 0 0.002 0.004 0.006 0.008 0.01
time [s] time [s]

(a) (b)

−8 −8
x 10 x 10
1.4 8
total z
c
7 st
1 mode
1.2
2nd mode
6 rd
3 mode
1
displacement [m]

5
force [nN]

0.8 4

0.6 3

2
0.4
1
0.2
0

0 −1
0 0.002 0.004 0.006 0.008 0.01 0 0.002 0.004 0.006 0.008 0.01
time [s] time [s]

(c) (d)

Figure 7.8: Input/output data: (a), (c) Input force Fc (t) for SNR = 30 dB and 15 dB,
respectively. (b), (d) Corresponding output displacements zc (x, t) (total and coming from
individual shaped modes), captured at the position x = 150 µm, when the input force Fc (t)
is applied at xf = L = 500 µm.

L = 252 µm (singular point for the third mode) and is captured at four singular points. In
all cases, the contribution coming from the third mode is zero, since the force is applied at
the singular point of the modal shape function φ3 (xf ). The total displacement in Fig. 7.9a at
x = 1 µm is almost zero and is increasing with the position x. As expected, for x = 391.5 µm
(singular point of the modal shape function φ2 (xf )), the second mode is additionally vanishing
and the total displacement of the full 3-mode model is exacly equal to the displacement of the
1-mode model.

In the case of the considered tunneling current-based platform, the displacement sensor
130 Chapter 7. Cantilever nanopositioning with electrostatic actuation

−13 −8
x 10 x 10
16 7
total zc total zc
14 st st
1 mode 6 1 mode
nd nd
12 2 mode 2 mode
rd 5 rd
3 mode 3 mode
10
displacement [m]

displacement [m]
4
8

6 3

4
2
2
1
0
0
−2

−4
0 0.002 0.004 0.006 0.008 0.01 0 0.002 0.004 0.006 0.008 0.01
time [s] time [s]

(a) (b)

−8 −8
x 10 x 10
14 16
total z total z
c c
12 st
1 mode 14 st
1 mode
2nd mode 2nd mode
rd 12 rd
10 3 mode 3 mode
displacement [m]

displacement [m]

10
8
8
6
6
4
4
2
2

0 0

−2 −2
0 0.002 0.004 0.006 0.008 0.01 0 0.002 0.004 0.006 0.008 0.01
time [s] time [s]

(c) (d)

Figure 7.9: Output displacements of the cantilever (shaped modal and total) at singular
points, when the applied force is at xf = 252 µm: (a) x = 1 µm ≈ 0 µm. (b) x = 252 µm.
(c) x = 391.5 µm. (d) x = 434 µm.

is not present and the only sensor in the vertical direction measures the tunneling current
(or the gap between the tip and the cantilever). The cantilever displacement zc (xf , t) causes
the change in this gap according to (2.16) which in turn affects the value of the current. In
this case, the measurement and the force location is constrained to the same position x = xf
(tip position). According to the previous observations, in order to damp the mode, x = xf
should be located near one of the singular points, corresponding to this mode. Here, x = xf
is set to 0.81L = 405 µm, which is between the singular points for the second and the third
mode, in order to obtain a satisfactory damping of the two. This can be seen in Fig. 7.10
showing frequency response of shaped modal and total displacement. Comparing this figure
with Fig. 7.3, where the x = xf = L, a significant damping can be noticed. Fig. 7.11a and
7.2. Multi-mode cantilever model analysis 131

Fig. 7.11b show the displacement at x = xf = 0.81L = 405 µm in response to the excitation
force from Fig. 7.8a and Fig. 7.8c, respectively. The contributions from the second and third
mode to the total displacement are negligible.

x = 405 µm, x =405 µm (L = 500 µm) x = 405 µm, x =405 µm (L = 500 µm)
f f
100 100
Magnitude (dB)

Magnitude (dB)
50 50

0 0
3 mode model
−50 −50 2 mode model
1 mode model
−100 −100
0 0
1st mode
−45 2nd mode
Phase (deg)

Phase (deg)
3rd mode
−90 −90

−135

−180 −180
3 4 5 6 3 4 5 6
10 10 10 10 10 10 10 10
Frequency (Hz) Frequency (Hz)

(a) (b)

Figure 7.10: Frequency response of (a) shaped modal and (b) total displacement (sum of
shaped modes) when the applied input force and measured output displacement are at the
distance x = xf = 0.81L = 405 µm.

−7 −7
x 10 x 10
4 4
total zc total zc
st
3 1 mode 3 1st mode
displacement [m]

displacement [m]

2nd mode 2nd mode

2 3rd mode 2 3rd mode

1 1

0 0

−1 −1
0 0.002 0.004 0.006 0.008 0.01 0 0.002 0.004 0.006 0.008 0.01
time [s] time [s]
(a) (b)

Figure 7.11: Output displacement of the cantilever (shaped modal and total) for x = xf =
0.81L = 405 µm for the input force with: (a) SNR = 30 dB (Fig. 7.8a). (b) SNR = 15 dB
(Fig. 7.8c).
132 Chapter 7. Cantilever nanopositioning with electrostatic actuation

7.3 Tunneling current-based cantilever nanopositioning with elec-

trostatic actuation and piezoelectric tip actuation

Fig. 7.12 shows the considered controlled vertical system (more details about its control is
given in the following subsections). It consists of two closed-loops, one inside the other one

−𝑧𝑐𝑟
𝑧̂𝑐𝑝 (𝑥)|𝑥=𝑥𝑓 −𝐹̂𝑝 𝐹̂𝑝
++ 𝐶𝑁𝑇𝐷𝐶 −1 𝐹𝐸
−𝑧𝑐𝑒𝑟
𝐿𝑃 𝑑𝑧 𝑞𝑝𝑧
𝑣𝑝𝑧 𝑛3
𝑧
+
- 𝑃𝐼𝐷 𝑉𝐴𝑧 (𝑠) −𝐺𝑝𝑧 ++ 𝐷𝑝𝑧 (𝑠) 𝑝 + 𝑑 𝑖𝑡 𝑣𝑧 𝑦3
𝑢𝑧 - 𝑃𝐿𝑡 ′ 𝐶𝑆𝑡 (𝑠) ++
𝑃𝑖𝑒𝑧𝑜𝑧
+
𝑃𝐿𝑡
𝑉𝑒 𝐹𝑒 𝐹𝑐
𝐸𝐴 +- 𝐶𝑁𝑇(𝑠) 𝑑0
𝑧𝑐 (𝑥)|𝑥=𝑥𝑓
(∗) 𝐹𝑝
𝐼𝑁𝑇𝑡𝑐
𝑃𝑙𝑎𝑛𝑡𝑧
𝑏 −1

𝑠𝑎𝑡(∗) 𝐻∞ +-
−𝑧̂𝑐𝑒 (𝑥)|𝑥=𝑥𝑓 = 𝑏𝑉𝑒 2 𝑢𝑒
𝑧𝑟

Figure 7.12: Closed-loop control of the vertical subsystem. The tunneling gap d = zt −
zc (xf ) = d0 + zp − zc (xf ) is regulated by the H∞ controller via cantilever position zc (xf ) =
zcp (xf )+zce (xf ) by keeping the tunneling voltage y3 (and as a result the tunneling gap) on the
constant level of zr . The interaction (proximity) force Fp between the tip and the cantilever
is estimated via F E subsystem from the tunneling voltage y3 and the control signals uz and
ue and eliminated by adjusting the position of the tunneling tip zp via outer closed-loop with
PID controller, tuned automatically using Matlab & Simulink software.
⏟
(𝑥)| 2
and proximity force estimator𝑧̂ 𝑐𝑒 F 𝑓
=−𝑏𝑉
E.𝑥=𝑥The 𝑒
inner closed-loop consists of the nonlinear physical
0
law P Lt with the current sensor CSt (s) and linear cantilever CN T (s) actuated by the non-
linear electrostatic actuator EA, controlled by H∞ controller which is designed to control
the tunneling gap on the constant level (i.e. tunneling voltage y3 on the desired level zr ).
Recalling equation (2.27), the nonlinear electrostatic actuator can be described by:

1 ε0 A e
Fe (t) = − Ve (t)2 (7.8)
2 l(t)2

where l(t) = l0 + zc (xf , t), l0 is the distance between the unbended cantilever and the electro-
static plate (see Fig. 2.4) and zc (xf , t) is the total cantilever displacement at the tip position
xf , defined as the superposition of the displacements caused by the proximity and the elec-
7.3. Tunneling current-based cantilever nanopositioning with electrostatic
actuation and piezoelectric tip actuation 133
trostatic forces as follows:
zc (xf , t) = zcp (xf , t) + zce (xf , t) (7.9)
Since in static conditions this force can be approximated as Fe (t) ≈ kf zce (xf , t) (where kf
is cantilever effective stiffness (force constant)), one gets the expression which links directly
zce (xf , t) with the input voltage Ve (t) as follows:

1 ε0 Ae 1 ε0 A e
zce (xf , t) ≈ − 2
Ve (t)2 = − Ve (t)2 (7.10)
kf 2l(t) kf 2(l0 + zc (xf , t))2

Finally, since l0 >> zc (xf , t) further approximation of (7.10) gives (see also [Blanvillain et al.,
2014]):
1 ε0 A e
zce (xf , t) ≈ − Ve (t)2 = −bVe (t)2 (7.11)
kf 2l02
The nonlinear electrostatic actuator can be expressed finally as follows:

Fe (t) ≈ kf zce (xf , t) ≈ −kf bVe (t)2 (7.12)

where the cantilever effective force constant kf is equal to the inverse of the cantilever DC
−1
gain kf = CN TDC . In order to linearize the actuator an exact linearization is used as follows:
r
ue
Ve (t) = (7.13)
b

where ue = bVe2 = −ẑce (xf ). Therefore, the control signal ue (with minus sign) serves as an
estimation of the cantilever position coming from the electrostatic force Fe . Since this force
is negative it bends the cantilever downwards (zce (xf ) < 0). The presence of the square root
in (7.13) restricts ue to be positive. One can design the controller with anti-windup design,
however here the positiveness of ue will be assured by restricting the cantilever displacement
coming from the electrostatic force zce to be negative, therefore ue = sat(ue ) = −ẑce almost
ever and the saturation block is added only for the security reason. Recalling (5.8), the
linearization of the nonlinear physical tunneling law P L0t around equilibrium point gives:

it (s) it (s)
P L0t (s) = = −kieq = − = P Lt (s) (7.14)
d(s) zc (s)

Taking into account the two linearizations (7.13) and (7.14), the H∞ linear controller is based
on the following linearized model from ue to y3 :
−1
Gue −>y3 (s) = −CSt (s)kieq CN T (s)CN TDC (7.15)

Once the inner loop is closed, a well tuned PID controls the tip position (and the cantilever
position as well, since the gap is kept constant by the inner closed-loop). This two-loop
scheme enables positioning the cantilever and tip above 1 nm, which is not possible with
only one loop, since the gap must be kept below this distance. The force estimator F E
estimates the proximity force on the basis of the measured tunneling voltage y3 and the control
signals of the piezoelectric and electrostatic actuators (uz and ue , respectively) and will be
described in the last subsection. Since the total cantilever displacement is the sum of the
134 Chapter 7. Cantilever nanopositioning with electrostatic actuation

250 250
Magnitude (dB)

Magnitude (dB)
200 200

150 150

100 100 perturbed model

nominal model
50 50
180 180
90 90
Phase (deg)

Phase (deg)
0 0
−90 −90
−180 perturbed model −180
nominal model
−270 −270
2 3 4 5 3 4 5
10 10 10 10 10 10 10
Frequency (Hz) Frequency (Hz)

(a) (b)

Figure 7.13: The perturbed 3-mode model (in red) and the nominal 1-mode model (in blue)
in case of: (a) x = xf = L = 500 µm. (b) x = xf = 0.81L = 405 µm.

displacement components coming from the electrostatic and proximity forces (i.e. zc (xf ) =
zce (xf ) + zcp (xf ) ≈ k1f (−Fp + Fe ) = k1f Fc in the static case), knowing the estimated cantilever
displacement ẑcp taken from the estimated proximity force F̂p allows to adjust the reference
for the cantilever displacement coming from the electrostatic force (with minus sign) −zcer in
such a way that the total cantilever displacement zc is kept on the reference level zcr . The
low-pas filter LP with cut-off frequency of 70 Hz is added on the reference −zcer in order to
eliminate the influence of its high frequency components on the tunneling current.

7.3.1 H∞ design for tunneling gap control via electrostatically actuated

cantilever

In this section H∞ controller is designed for the two models corresponding to the two positions
highlighted in the previous section (i.e. x = xf = L = 500 µm and x = xf = 0.81L =
405 µm). In both cases the controller for the 3-mode model (called the perturbed model) is
designed on the basis of the 1-mode model (called the nominal model) as shown in Fig. 7.13a
and Fig. 7.13b. The linearized perturbed model from the input ue to the output y3 (see
Fig. 7.12) including the 3-mode model of the cantilever is given by:
−1
Gue −>y3 p (s) = −CSt (s)kieq CN T3m (s)CN T3mDC
(7.16)

where CN T3mDC is the DC gain of the cantilever model CN T3m (s). The considered nominal
model corresponds to model (7.16) but with 1-mode model of the cantilever and is given by:
−1
Gue −>y3 (s) = −CSt (s)kieq CN T1m (s)CN T1m DC
(7.17)

The output multiplicative structure has been chosen to model the uncertain system (7.16) as
7.3. Tunneling current-based cantilever nanopositioning with electrostatic
actuation and piezoelectric tip actuation 135

100 100
l∆(ω) l∆(ω)
80 80
|W∆(jω)| |W (jω)|
∆
60 60
Singular Values (dB)

Singular Values (dB)

40 40

20 20

0 0

−20 −20

−40 −40

−60 −60

−80 −80

−100 3 4 5 6
−100 3 4 5 6
10 10 10 10 10 10 10 10
Frequency (rad/s) Frequency (rad/s)

(a) (b)

100 100
l∆(ω) l∆(ω)
80 80
|W∆(jω)| |W∆(jω)|
60 60
Singular Values (dB)

Singular Values (dB)

40 40

20 20

0 0

−20 −20

−40 −40

−60 −60

−80 −80

−100 3 4 5 6
−100 3 4 5 6
10 10 10 10 10 10 10 10
Frequency (rad/s) Frequency (rad/s)

(c) (d)

Figure 7.14: The relative errors l∆ (ω) together with their upper bounds |W∆ (jω)| used for
H∞ design: (a) x = xf = L = 500 µm and 5th order weight W∆ (s). (b) x = xf = L = 500 µm
and 2nd order weight W∆ (s). (c) x = xf = 0.81L = 405 µm and 5th order weight W∆ (s). (d)
x = xf = 0.81L = 405 µm and 2nd order weight W∆ (s).

follows:
Gue −>y3 p (s) = (1 + W∆ (s)∆(s))Gue −>y3 (s) (7.18)

where ∆(s) is any stable transfer function, which represents the normalized complex pertur-
bations such that |∆(jω)| ≤ 1, ∀ω and W∆ (s) is a rational transfer function representing the
uncertainty weight chosen as follows:

Gue −>y3 p (jω) − Gue −>y3 (jω)

|W∆ (jω)| ≥ |W∆ (jω)∆(jω)| = l∆ (ω) = , ∀ω (7.19)
Gue −>y3 (jω)
136 Chapter 7. Cantilever nanopositioning with electrostatic actuation

20 10
|Sz|
|1/WS |
z 0
0

−10
Magnitude [dB]

Magnitude [dB]
−20

−20

−40
−30

−60 |Tz|
−40
|1/WT |
z

−80 −2 0 2 4
−50 −2 0 2 4
10 10 10 10 10 10 10 10
Frequency [Hz] Frequency [Hz]

(a) (b)

−180

−190

−200
Magnitude [dB]

−210

−220

−230

−240
|R |
z
−250
|1/W |
R
z
−260 −2 0 2 4
10 10 10 10
Frequency [Hz]

(c)

Figure 7.15: Sensitivity functions for the nominal plant and the obtained H∞ controller for
the desired bandwidth 5 kHz and 2nd order uncertainty weight W∆ (s): a) Output sensitivity
function. b) Complementary sensitivity function. c) Control sensitivity function.

where l∆ (ω) represents relative magnitude of the neglected dynamics (i.e. one chooses W∆ (s)
such that |W∆ (jω)| is an upper bound on the relative errors l∆ (ω)). Fig. 7.14 shows the
relative errors l∆ (ω) together with |W∆ (jω)| for the 5th and 2nd order weight W∆ (s) with
cross-over frequency of 5 kHz and for two cases, where x = xf = L = 500 µm and x = xf =
0.81L = 405 µm (for both cases W∆ (s) are exactly the same). Since W∆ (s) will shape the
complementary sensitivity function, to assure good robustness margin, it is already constrained
at low frequencies by the 3.5 dB from below. It can be seen that for the considered bandwidth
5 kHz, the 5th order weight W∆ (s) successfully bounds the relative errors from above for
the two cases. For the 2nd order weight, the robust stability of the model corresponding to
x = xf = L = 500 µm is not achieved, while for the case where x = xf = 0.81L = 405 µm
it is. Moreover, in the second case the bandwidth can be furtherly increased, since there is
still a little space for robust stability. The following weights have been chosen to shape the
7.3. Tunneling current-based cantilever nanopositioning with electrostatic
actuation and piezoelectric tip actuation 137
sensitivity functions:
s + 6.283 × 104
WSz (s) = (7.20)
2s + 62.83
WTz (s) = W∆ (s) (7.21)
WRz (s) = −Gt kieq /10 = −1.65 × 109 (7.22)
Fig. 7.15 shows the sensitivity functions shaped by the associated weights (2nd order uncer-
tainty weight W∆ (s) in this case). The optimal cost function γ = 1.35 has been obtained.
Fig. 7.16 shows the reference tracking corresponding to the two cases and weights. The open-
loop response of the model has been scaled by its gain in order to be with the same scale
with the closed-loop one. The performance achieved is the same apart from the case where
robust stability is not achieved (x = xf = L = 500 µm and 2nd order weight W∆ (s)). One
can conclude that proper location of the actuator and sensor allows to reduce the order of

7
x 10
4 8
OL
OL
3.5 H∞ 6 H∞
3 ref zr
4 ref zr
output y3 [V]

output y3 [V]

2.5
2
2
0
1.5
−2
1

0.5 −4

0 −6
0 0.002 0.004 0.006 0.008 0.01 0 0.002 0.004 0.006 0.008 0.01
time [s] time [s]

(a) (b)

4 4
OL OL
3.5 H∞ 3.5 H∞

3 ref zr 3 ref zr
output y3 [V]

output y3 [V]

2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

0 0
0 0.002 0.004 0.006 0.008 0.01 0 0.002 0.004 0.006 0.008 0.01
time [s] time [s]

(c) (d)

Figure 7.16: H∞ control of the linearized model from the input ue to the output y3 (without
the noise) (the open- loop response of the plant has been scaled by its gain in order to be with
the same scale with the closed-loop one): (a) x = xf = L = 500 µm and 5th order weight
W∆ (s). (b) x = xf = L = 500 µm and 2nd order weight W∆ (s). (c) x = xf = 0.81L = 405 µm
and 5th order weight W∆ (s). (d) x = xf = 0.81L = 405 µm and 2nd order weight W∆ (s).
138 Chapter 7. Cantilever nanopositioning with electrostatic actuation

the controller and/or increase the desired bandwidth. Fig. 7.17a and Fig. 7.17b show the
frequency response (magnitude) of the scaled open-loop and closed-loop system for the two
considered cases. The comparison with the simple PID controller has been shown in Fig. 7.18.

50 50
OL OL
H∞ H∞

0 0
Magnitude (dB)

Magnitude (dB)
−50 −50

−100 −100

−150 2 3 4 5
−150 2 3 4 5
10 10 10 10 10 10 10 10
Frequency (Hz) Frequency (Hz)

(a) (b)

Figure 7.17: Bode diagrams of the scaled open-loop (OL) and closed-loop system with H∞
controller (with 5th order weight W∆ (s)) and the perturbed plant model: (a) x = xf = L =
500 µm. (b) x = xf = 0.81L = 405 µm.

1.8
OL
1.6 PIDt
PIDl
1.4
H∞
1.2 ref zr
output y3 [V]

0.8

0.6

0.4

0.2

0
0 0.005 0.01 0.015 0.02
time [s]

Figure 7.18: Comparison of H∞ control with the PID controller (PIDt is the closed-loop
response with the first shot parameters of PID proposed by Matlab & Simulink and PIDl
refers to the fastest possible PID controller before the stability loss).
7.3. Tunneling current-based cantilever nanopositioning with electrostatic
actuation and piezoelectric tip actuation 139
The PID gains have been tuned using Matlab & Simulink software. The proposed solution
(PIDt ) has quite low bandwidth and can be increased, however this introduces at some point
significant oscillations until the system is on the stability limit (PIDl ). The model-based H∞
controller clearly outperforms PID.

7.3.2 Cantilever positioning via piezoelectrically actuated tip with prox-

imity force compensation

Recalling equation (2.24), the proximity interaction force between the tunneling tip and the
cantilever is expressed as follows:
(
−HR/6d(t)2 , √ d(t) ≥ ap
Fp (t) = IN Ttc (d(t)) = ∗ (7.23)
−HR/6a2p + 34 Eef f R(ap − d(t))3/2 , d(t) < ap

−8
x 10
1
ap tunneling
voltage
0−10V

0
force Fp [N]

−10
x 10
−2
−4
−1 −6
R = 20 nm −8
R = 30 nm −10
R = 40 nm 8 9 10
R = 50 nm −10
x 10
−2
0 0.2 0.4 0.6 0.8 1
distance tip/surface d [m] −9
x 10
Figure 7.19: Simulated proximity force for different tip curvature radius R. Here, the inter-
atomic distance ap = 0.166 nm. The inset plot shows the zoom of the proximity forces in the
tunneling voltage measuring range (i.e. 0-10 V, which corresponds to the tip/surface distance
0.7-1 nm).

Fig. 7.19 shows the behavior of proximity force (7.23) for the following numerical values:
different tip curvature radius R = [20, 30, 40, 50] nm, the Hamaker constant H = 6.4 × 10−20
∗ 2
J, interatomic value ap = 0.166 nm, the effective contact stiffness Eef f = [(1 − vt )/Et + (1 −
v 2 )/E]−1 with the Poisson ratios vt = 0.28 and v = 0.17 and the elastic moduli Et = 129
GPa and E = 180 GPa of the tip and the cantilever, respectively. As mentioned before, the
interatomic value ap separates the region where only attractive van der Waals force is present
140 Chapter 7. Cantilever nanopositioning with electrostatic actuation

from the region where the repulsive adhesive force is added to the van der Waals force. When
the two forces become equal (proximity force Fp = 0) the materials are in mechanical contact
and the total net proximity force starts to be repulsive. In the considered case the measuring
range of the tunneling voltage is constrained to the standard value of the data acquisition cards
between 0 and 10 V, which corresponds to the tip/surface distance below 1 nm and down to
0.7 nm and therefore during normal operation (where the tunneling current can be measured)
only the attractive proximity force should appear (non-contact mode). The proximity force
could be estimated from the measurement of zc (xf ) if the cantilever displacement sensor (like
in AFM) was available as follows (in static case):

F̂p = F̂e − Fc ≈ kf ẑce (xf ) − kf zc (xf ) = kf ue − kf zc (xf ) (7.24)

In the considered case the only available measurement is the tunneling voltage y3 . The prox-
imity force can be then approximated from the measurement y3 and the control signals ue and
uz as proposed below:
F̂p = F̂e − F̂c ≈ kf ue − kf ẑc (xf ) (7.25)
where ẑc (xf ) is given by:
  
ˆ 1 y3
ẑc (xf ) = ẑt − d = (−Gpz Gvz uz + d0 ) − − ln (7.26)
| {z } k Gt gVb
ẑt | {z }
dˆ

Fig. 7.20 shows the performance of the two closed-loop scheme from Fig. 7.12, when the
proximity force is absent. The tunneling current is kept on the constant level of 1 nA (which
corresponds to the tunneling voltage of 1 V) via electrostatic actuator, by the H∞ controller
as shown in Fig. 7.20a. This corresponds to the distance d = 0.84 nm. The corresponding
control signal ue is shown in Fig. 7.20b. Note that ue = −ẑce (i.e. it is the estimation of the zce
taken with minus sign). The displacements zt and zc of the tip and the cantilever, respectively
are shown in Fig. 7.20c. One can see that the PID controller of the outer loop controls the
cantilever position zc around the reference zcr by adjusting the tip position zt via piezoelectric
actuator (it is possible since the gap between the tip and the cantilever is already kept on
the constant level of 0.84 nm by the H∞ controller of the inner loop). The low-pass filter
LP added on the cantilever reference allows to reduce the peaks appearing in the tunneling
current measurement due to the square signal of this reference. The creep phenomenon which
acts as a disturbance dz over the tip position has been successfully canceled (which can be
also seen on the control signal uz of the tip, which drifts as shown in Fig.7.20d).

Fig. 7.21 shows the performance of the two closed-loop scheme in the presence of the
proximity force. First, the case when the creep dz of the tip is not present is considered.
Fig.7.21a shows the tip and the cantilever displacements without (ztu and zcu ) and with
(ztc and zcc ) proximity force compensation. Without compensation the cantilever cannot
be positioned at the reference signal zcr , since the force −Fp attracts it upwards. As a
result the tip moves upwards as well in order to keep the tunneling gap constant. When
the interaction force is compensated, the proper cantilever and tip positioning is possible.
The estimated proximity force is shown in Fig. 7.21b. Notice that Fig. 7.21c and Fig. 7.21d
7.4. Conclusion 141

−9
x 10
2.5
2
y3
zr 2

control signal ue [V]

1.5
tunneling voltage [V]

1.5

1
1

0.5
0.5

0 0
0.5 0.6 0.7 0.8 0.9 1 0 0.5 1 1.5 2
time [s] time [s]

(a) (b)
−9
x 10
0 0.15

−0.5
control signal uz [V]
displacement [m]

−1 0.1

−1.5

−2 zt 0.05
d=deq
zc
−2.5
zc
r
−3 0
0 0.5 1 1.5 2 0 0.5 1 1.5 2
time [s] time [s]

(c) (d)

Figure 7.20: The performance of the two closed-loop scheme from Fig. 7.12 when the proximity
force Fp is absent: (a) Tunneling voltage y3 . (b) Control signal ue of the inner loop with H∞
controller. (c) Cantilever and tip displacements. (b) Control signal uz of the outer loop with
PID controller.

correspond to the case when the creep phenomenon of piezoactuated tip is present. In this
case the proximity force cannot be well estimated, however the results are still much better
than without compensation. The performance could be improved when the creep model of
the piezo was taken into account in the proximity force estimator. It should be stressed that
this compensation depends on the accuracy of the parameters used in the estimator and other
phenomena like creep.

7.4 Conclusion

In this chapter simulation results for the 3-mode cantilever model have been presented. Know-
ing the singular points of the modal shape functions allows to place the actuator/sensor near
142 Chapter 7. Cantilever nanopositioning with electrostatic actuation

−9 −10
x 10 x 10
10
ztu Fp
−2.6 F̂p
8 zcu
ztc

Proximity Force [N]

6
displacement [m]

−2.8
zcc
4 zc
r −3
2

0 −3.2

−2
−3.4
−4
0 0.5 1 1.5 2 0 0.5 1 1.5 2
time [s] time [s]

(a) (b)
−9 −10
x 10 x 10
10
ztu Fp
−2.6 F̂p
8 zcu
ztc
Proximity Force [N]

6
displacement [m]

−2.8
zcc
4 zc
r −3
2

0 −3.2

−2
−3.4
−4
0 0.5 1 1.5 2 0 0.5 1 1.5 2
time [s] time [s]

(c) (d)

Figure 7.21: The performance of the two closed-loop scheme from Fig. 7.12 when the proximity
force Fp is present: (a) Tip and cantilever displacements without (ztu and zcu ) and with
(ztc and zcc ) proximity force compensation (without the creep phenomenon). (b) Estimated
proximity force Fp (without the creep phenomenon). (c) Tip and cantilever displacements
without (ztu and zcu ) and with (ztc and zcc ) proximity force compensation (with the creep
phenomenon). (d) Estimated proximity force Fp (with the creep phenomenon).

them in order to damp the associated modes and as a result to reduce the order and increase
the bandwidth of the system. The application of such a 3-mode cantilever nanopositioning
has been considered on the basis of the tunneling current. The tunneling gap is kept constant
via electrostatically actuated cantilever by H∞ control. The positioning over the distance of
1 nm was possible using a second closed-loop, which controls the cantilever position through
the tip position via piezoelectrically actuated tunneling tip. To that end, a proximity force
estimator has been designed in order to compensate it, on the basis of the measured tunneling
current and two control signals.
 Chapter 8

Conclusions and perspectives

Conclusions:
Summing up, the main goal of this thesis was to develop 3D model for the nanopositioning
system corresponding to the experimental set-up developed in GIPSA-lab, its proper control
with high accuracy over possibly high bandwidth and experimental validation of the proposed
techniques. To that end, the platform has been equipped for 3D operation with capacitive
sensors for both X and Y horizontal displacements in addition to the vertical actuation.

The developed 3D model finally consists of piezoelectrically actuated scanning horizonal X

and Y axes and vertical probing Z axis. The latter is equipped with piezoelectrically actuated
tunneling tip (like in STM) and with electrostatically actuated micro-cantilever (like in AFM).
Such a model is quite complex and due to several nonlinearities (exponential nonlinearity of
tunneling current, hysteresis of piezoelectric actuators, quadratic nonlinearity of electrostatic
actuator) and other adverse phenomena (creep, vibration, cross-couplings between piezo axes,
high measurement noise of tunneling current sensor), it is a real challenge to control it with
nanoscale accuracy. First, hysteresis and creep phenomena exhibited by piezoelectric actua-
tors in horizontal axes have been compensated in real-time using different techniques (classical
Prandtl-Ishlinskii (PI) model, its modified version (MPI), disturbance observer (DOB) and
a combination of MPI with DOB, as well as an adaptive approach). An identification with
subsequent order reduction of such linearized horizontal 2D model allowed to design several
linear controllers (LQG/LTR, SISO H∞ , MIMO H∞ ) for piezo vibration and cross-coupling
reduction between X and Y axes. Their comparative experimental validation has been pre-
sented. The real challenge is a control of the vertical motion (Z axis), based on the tunneling
current phenomenon, due to several requirements (nanoscale operating range, high sensitivity
to measurement noise and external disturbances, sharp tip, clean surface etc.). A two-stage
procedure for approaching the tip to the surface and obtaining tunneling current has been
used (coarse approach done manually and fine approach done automatically). Tunneling cur-
rent proper control using pole placement with sensitivity functions shaping methodology has
been experimentally validated. Finally, illustrative results for three applications have been
presented:

• 3D nanopositioning application (without cantilever motion), validated experimentally.

The tunneling tip tracks prespecified reference signals in X, Y and Z directions (in the
Z direction the reference is called "virtual surface", since it imitates surface variations).
The adverse phenomena of hysteresis, creep, vibrations and cross-couplings in the hor-
izontal axes were eliminated using above mentioned techniques with tunneling current

143
144 Chapter 8. Conclusions and perspectives

control in vertical axis using pole placement with sensitivity functions shaping method-
ology.

• STM-like application, validated in simulations. The tip is moved along the X and Y
axes, while the reference for the Z axis is kept constant. To that end, two control
strategies have been compared: SISO PID decentralized control (3 PIDs, each for one
axis, with decoupling compensators) and one MIMO LQI centralized controller. In the
first approach, the surface topography is retrieved from the control signal of the vertical
piezo (classical approach) and in the second approach it is reconstructed via Kalman
observer (proposed approach). The second approach gave more satisfactory results.

• Multi-mode cantilever positioning application, tested in simulations. A scheme consist-

ing of two closed-loops (an inner closed-loop with electrostatically actuated cantilever
using H∞ controller and an outer closed-loop with piezoelectrically actuated tunneling
tip using PID controller) allowed to position the cantilever on a distance larger than
1 nm. The interaction force appearing between the tip and the cantilever has been
compensated via proximity force estimator.

Perspectives:
Several perspectives/extensions can be deduced in view of the obtained results:

• The multi-mode cantilever model tested in simulations needs to be validated in exper-

iments. To that end, the sampling frequency should be increased in order to catch at
least two modes. However, the acquisition cards in the present configuration limit the
sampling frequency to 30 kHz. High speed data acquisition, based for example on FPGA
technology can be used for that purpose.

• Cross-couplings from the horizontal to the vertical axes should be compensated. How-
ever, it is hard to identify them from the tunneling-current measurement, since it is
affected also by the surface variations (for that purpose the surface should be ideally
flat). For high frequencies, the tip can move outside the operating range as well, resulting
in loss of tunneling current.

• For the second application (STM-like application), in order to be experimentally val-

idated, a proper reference grating (surface) should be used. Here instead, a "virtual
surface" was imposed as the reference in the vertical direction, which only imitated
the real surface (the first application - 3D nanopositioning). For that reason, the first
application has been tested in experiments, while the second one in simulations only.

• Accurate model parameters are needed for the proximity force compensation using the
proposed estimator. Adding for example another sensor which measures the cantilever
displacement could help with exact force estimation.
 Bibliography

Abramovitch, D.Y., Andersson, S.B., Pao, L. Y., and Schitter, G. (2007). “A Tutorial on the
Mechanisms, Dynamics, and Control of Atomic Force Microscopes”. In: 2007 American
Control Conference (ACC 2007), pp. 3488–3502 (cit. on pp. 6, 11, 14, 118).

Ahmad, I. (2011). “Analyse et commande d’un système de mesure à courant tunnel”. PhD
thesis. GIPSA-Lab, l’Université de Grenoble (cit. on p. 1).

Ahmad, I., Voda, A., and Besançon, G. (2010). “Robust H∞ control of a scanning tunneling
microscope under parametric uncertainties”. In: 2010 American Control Conference (ACC
2010), pp. 6555–6560 (cit. on p. 19).

Ahmad, I., Voda, A., and Besancon, G. (2012a). “Experimental Validation of H∞ SISO control
for high performance tunneling current measurement system and MIMO extension”. In:
2012 IEEE International Conference on Control Applications (CCA 2012), pp. 581–586
(cit. on pp. 19, 20).

Ahmad, I., Voda, A., Besancon, G., and Buche, G. (2012b). “Robust digital control approach
for high performance tunneling current measurement system”. In: Control Engineering
Practice 20(7), pp. 643–653 (cit. on pp. 19, 94, 104, 117).

Alexander, S., Hellemans, L., Marti, O., Schneir, J., Elings, V., Hansma, P. K., Longmire, M.,
and Gurley, J. (1989). “An atomic-resolution atomic-force microscope implemented using
an optical lever”. In: Journal of Applied Physics 65(1), pp. 164–167 (cit. on p. 19).

Amin-Shahidi, D. and Trumper, D.L. (2013). “Improved charge amplifier using hybrid hys-
teresis compensation”. In: Review of Scientific Instruments 84(8), p. 085115 (cit. on p. 17).

Ang, W.T., Khosla, P.K., and Riviere, C.N. (2007). “Feedforward Controller With Inverse
Rate-Dependent Model for Piezoelectric Actuators in Trajectory-Tracking Applications”.
In: IEEE Transactions on Mechatronics 12(2), pp. 134–142 (cit. on pp. 13, 16, 37, 42).

Aphale, S.S., Devasia, S., and Moheimani, S.O.R. (2008). “High-bandwidth control of a piezo-
electric nanopositioning stage in the presence of plant uncertainties”. In: Nanotechnology
19(12), pp. 125503–125512 (cit. on p. 17).

Bashash, S. and Jalili, N. (2008). “A Polynomial-Based Linear Mapping Strategy for Feed-
forward Compensation of Hysteresis in Piezoelectric Actuators”. In: Journal of Dynamic
Systems, Measurement, and Control 130(3), p. 031008 (cit. on p. 13).

145
146 Bibliography

Bertotti, G. and Mayergoyz, I.D. (2006). The Science of Hysteresis: Mathematical Modeling
and Applications. New York, NY, USA: Academic Press (cit. on p. 12).

Bhikkaji, B. and Moheimani, S.O.R. (2008). “Integral Resonant Control of a Piezoelectric Tube
Actuator for Fast Nanoscale Positioning”. In: IEEE/ASME Transactions on Mechatronics
13(5), pp. 530–537 (cit. on p. 17).

Binnig, G. and Rohrer, H. (1986). “Scanning Tunneling Microscopy”. In: IBM Journal of
Research and Development 30, pp. 355–369 (cit. on pp. 5, 11).

Binnig, G., Quate, C. F., and Gerber, Ch. (1986). “Atomic force microscope”. In: Physical
Review Letters 56(9), pp. 930–933 (cit. on p. 6).

Blanvillain, S. (2010). “Contrôle nanoscopique du mouvement par courant tunnel: étude et

réalisation”. PhD thesis. GIPSA-Lab, l’Université de Grenoble (cit. on pp. 1, 19).

Blanvillain, S., Voda, A., and Besancon, G. (2008). “Pull-in control during nanometric posi-
tioning by near field position sensing”. In: 47th IEEE Conference on Decision and Control
(CDC 2008), pp. 5194–5199 (cit. on p. 19).

Blanvillain, S., Voda, A., Besancon, G., and Buche, G. (2009). “The tunnel current as a sub-
nanometer motion sensor”. In: 2009 European Control Conference (ECC 2009), pp. 5003–
5008 (cit. on p. 19).

Blanvillain, S., Voda, A., Besancon, G., and Buche, G. (2014). “Subnanometer Positioning and
Drift Compensation With Tunneling Current”. In: IEEE Transactions on Control Systems
Technology 22(1), pp. 180–189 (cit. on pp. 19, 121, 133).

Brinkerhoff, R. and Devasia, S. (2000). “Output Tracking for Actuator Deficient/Redundant

Systems: Multiple Piezoactuator Example”. In: Journal of Guidance, Control and Dynam-
ics 23(2), pp. 370–373 (cit. on p. 19).

Butterworth, J.A., Pao, L.Y., and Abramovitch, D.Y. (2008). “Architectures for Tracking
Control in Atomic Force Microscopes”. In: Proceedings of the 17th IFAC World Congress
(IFAC 2008). Seoul, Korea, pp. 8236–8250 (cit. on p. 16).

Chuang, N., Petersen, I.R., and Pota, H.R. (2011). “Robust H∞ control in fast atomic force
microscopy”. In: 2011 American Control Conference (ACC 2011), pp. 2258–2265 (cit. on
p. 17).

Chuntao, L. and Yonghong, T. (2004). “A neural networks model for hysteresis nonlinearity”.
In: Sensors and Actuators A: Physical 112(1), pp. 49 –54 (cit. on p. 13).
Bibliography 147

Clayton, G.M., Tien, S., Leang, K.K., Zou, Q., and Devasia, S. (2009). “A Review of Feed-
forward Control Approaches in Nanopositioning for High-Speed SPM”. In: Journal of Dy-
namic Systems, Measurement, and Control 131(6), pp. 061101–19 (cit. on p. 17).

Croft, D. and Devasia, S. (1999). “Vibration compensation for high speed scanning tunneling
microscopy”. In: Review of Scientific Instruments 70(12), pp. 4600–4605 (cit. on p. 15).

Croft, D., Shed, G., and Devasia, S. (2001). “Creep, Hysteresis, and Vibration Compensation
for Piezoactuators: Atomic Force Microscopy Application”. In: Journal of Dynamic Systems
Measurement and Control 123 (1) (cit. on pp. 14, 15, 17, 111, 112).

Cruz-Hernandez, J.M. and Hayward, V. (2001). “Phase control approach to hysteresis re-
duction”. In: IEEE Transactions on Control Systems Technology 9(1), pp. 17–26 (cit. on
p. 13).

Das, S.K., Pota, H.R., and Petersen, I.R. (2012). “Multi-variable resonant controller for fast
atomic force microscopy”. In: 2012 2nd Australian Control Conference (AUCC 2012),
pp. 448–453 (cit. on pp. 18, 19).

Derjaguin, B.V, Muller, V.M, and Toporov, Y.P. (1975). “Effect of contact deformations on
the adhesion of particles”. In: Journal of Colloid and Interface Science 53(2), pp. 314 –326
(cit. on p. 29).

Devasia, S., Eleftheriou, E., and Moheimani, S.O.R. (2007). “A Survey of Control Issues in
Nanopositioning”. In: IEEE Transactions on Control Systems Technology 15(5), pp. 802–
823 (cit. on pp. 11, 14, 19).

Dong, R. and Tan, Y. (2009). “A modified Prandtl–Ishlinskii modeling method for hysteresis”.
In: Physica B: Condensed Matter 404(8–11), pp. 1336 –1342 (cit. on p. 13).

Dong, R., Tan, Y., Chen, H., and Xie, Y. (2008). “A neural networks based model for rate-
dependent hysteresis for piezoceramic actuators”. In: Sensors and Actuators A: Physical
143(2), pp. 370 –376 (cit. on p. 13).

Doyle, J.C. and Stein, G. (1979). “Robustness with observers”. In: IEEE Transactions on
Automatic Control, pp. 607–611 (cit. on pp. 18, 66).

Doyle, J.C. and Stein, G. (1981). “Multivariable Feedback Design: Concepts for a Classi-
cal/Modern Synthesis”. In: IEEE Transactions on Automatic Control 26(1), pp. 4–16 (cit.
on pp. 18, 66).

El-Shaer, A.H., Janaideh, M.A., Krejci, P., and Tomizuka, M. (2013). “Robust performance
enhancement using disturbance observers for hysteresis compensation based on generalized
148 Bibliography

Prandtl-Ishlinskii model”. In: Journal of Dynamic Systems, Measurement, and Control

135(5), p. 051008 (cit. on p. 17).

García, R. and San Paulo, A. (2000). “Dynamics of a vibrating tip near or in intermittent
contact with a surface”. In: Physical Review B 61 (20), R13381–R13384 (cit. on p. 28).

Gentili, M., Giovannella, C., and Selci, S. (1993). Nanolithography: A Borderland Between
STM, EB, IB and X-Ray Lithographies. Norwell, MA: Kluwer: NATO ASI Series E: Ap-
plied Science (cit. on p. 7).

Giri, F. and Bai, E.-W. (2010). Block-oriented nonlinear system identification. Springer-Verlag
(cit. on p. 111).

Gu, G.-Y. and Zhu, L.-M. (2010). “High-speed tracking control of piezoelectric actuators using
an ellipse-based hysteresis model”. In: Review of Scientific Instruments 81(8), p. 085104
(cit. on p. 13).

Gu, G.-Y. and Zhu, L.-M. (2011). “Modeling of rate-dependent hysteresis in piezoelectric
actuators using a family of ellipses”. In: Sensors and Actuators A: Physical 165(2), pp. 303
–309 (cit. on p. 13).

Gu, G.-Y., Zhu, L.-M., and Su, C.-Y. (2014a). “Modeling and Compensation of Asymmetric
Hysteresis Nonlinearity for Piezoceramic Actuators With a Modified Prandtl-Ishlinskii
Model”. In: IEEE Transactions on Industrial Electronics 61(3), pp. 1583–1595 (cit. on
p. 16).

Gu, G.-Y., Zhu, L.-M., Su, C.-Y., Ding, H., and Fatikow, S. (2014b). “Modeling and Control of
Piezo-Actuated Nanopositioning Stages: A Survey”. In: IEEE Transactions on Automation
Science and Engineering PP(99), pp. 1–20 (cit. on pp. 12, 14, 16).

Habib, H., Rehman, O.U., Pota, H.R., and Petersen, I.R. (2012). “Internal reference model
based optimal LQG controller for atomic force microscope”. In: 12th International Confer-
ence on Control Automation Robotics Vision (ICARCV 2012), pp. 294–299 (cit. on pp. 17,
19, 65).

Halim, D. and Moheimani, S.O.R. (2001). “Spatial resonant control of flexible structures-
application to a piezoelectric laminate beam”. In: IEEE Transactions on Control Systems
Technology 9(1), pp. 37–53 (cit. on p. 18).

Horowitz, R., Chen, T.-L., Oldham, K., and Li, Y. (2004). Microactuators for Dual-Stage Servo
Systems in Magnetic Disk Files. Berlin, Germany: Springer-Verlag, pp. 951–981 (cit. on
p. 19).
Bibliography 149

Hu, X., Guo, W., Huang, T., and Chen, B.M. (1999). “Discrete-time LQG/LTR dual-stage
controller design and implementation for high track density HDDs”. In: 1999 American
Control Conference (ACC 1999). Vol. 6, pp. 4111–4115 (cit. on pp. 7, 18, 66).

Hughes, D. and Wen, J.T. (1997). “Preisach modeling of piezoceramic and shape memory alloy
hysteresis”. In: Smart Materials and Structures 6(3), p. 287 (cit. on p. 16).

Janaideh, M.A., Su, C.-Y., and Rakheja, S. (2008). “Development of the rate-dependent
Prandtl–Ishlinskii model for smart actuators”. In: Smart Materials and Structures 17(3),
p. 035026 (cit. on p. 13).

Janaideh, M.A., Rakheja, S., and Su, C.-Y. (2009a). “A generalized Prandtl–Ishlinskii model
for characterizing the hysteresis and saturation nonlinearities of smart actuators”. In: Smart
Materials and Structures 18(4), p. 045001 (cit. on pp. 13, 16).

Janaideh, M.A., Rakheja, S., and Su, C.-Y. (2009b). “Experimental characterization and mod-
eling of rate-dependent hysteresis of a piezoceramic actuator”. In: Mechatronics 19(5),
pp. 656 –670 (cit. on p. 12).

Jandt, K. D., Finke, M., and Cacciafesta, P. (2000). “Aspects of the physical chemistry of
polymers, biomaterials and mineralised tissues investigated with atomic force microscopy
(AFM)”. In: Colloids and Surfaces B: Biointerfaces 19(4), pp. 301 –314 (cit. on p. 6).

Jiles, D.C. and Atherton, D.L. (1986). “Theory of ferromagnetic hysteresis”. In: Journal of
Magnetism and Magnetic Materials 61(1–2), pp. 48 –60 (cit. on p. 12).

Jung, H. and Gweon, D.-G. (2000). “Creep characteristics of piezoelectric actuators”. In: Re-
view of Scientific Instruments 71(4), pp. 1896–1900 (cit. on p. 14).

Jung, H., Shim, J.Y., and Gweon, D.-G. (2001). “Tracking control of piezoelectric actuators”.
In: Nanotechnology 12(1), 14–20 (cit. on p. 17).

Karny, M. (2006). Optimized Bayesian Dynamic Advising: Theory and Algorithms. Secaucus,
NJ, USA: Springer-Verlag New York, Inc. (cit. on p. 56).

Krejci, P. and Kuhnen, K. (2001). “Inverse control of systems with hysteresis and creep”. In:
IEE Proceedings - Control Theory and Applications 148(3), pp. 185–192 (cit. on p. 17).

Kuhnen, K. (2003). “Modeling, identification and compensation of complex hysteretic non-

linearities: A modified Prandtl-Ishlinskii approach”. In: European Journal of Control 9(4),
407–418 (cit. on p. 13).
150 Bibliography

Kuhnen, K. and Janocha, H. (1999). “Adaptive inverse control of piezoelectric actuators with
hysteresis operators”. In: 1999 European Control Conference (ECC 1999), pp. 791–796
(cit. on p. 16).

Kuhnen, K. and Janocha, H. (2001). “Inverse feedforward controller for complex hysteretic
nonlinearities in smart-materials systems”. In: Control and and Intelligent Systems 29(3),
pp. 74–83 (cit. on p. 16).

Kuiper, S., Fleming, A., and Schitter, G. (2010). “Dual actuation for high speed atomic force
microscopy”. In: 5th IFAC Symposium on Mechatronic Systems, pp. 220–226 (cit. on p. 18).

Kwakernaak, H. (1972). Linear Optimal Control Systems. John Wiley and Sons, Inc (cit. on
pp. 18, 66).

Landau, I. D. and G., Zito (2006). Digital Control Systems: Design, Identification and Imple-
mentation. Springer-Verlag (cit. on pp. 92, 95–97, 104).

Leang, K.K. and Devasia, S. (2007). “Feedback-Linearized Inverse Feedforward for Creep,
Hysteresis, and Vibration Compensation in AFM Piezoactuators”. In: IEEE Transactions
on Control Systems Technology 15(5), pp. 927–935 (cit. on p. 17).

Li, Ch.-X., Gu, G.-Y., Yang, M.-J., and Zhu, L.-M. (2013a). “Design, analysis and testing of
a parallel-kinematic high-bandwidth XY nanopositioning stage”. In: Review of Scientific
Instruments 84(12), 125111(1–12) (cit. on p. 15).

Li, P., Yan, F., Ge, Ch., Wang, X., Xu, L., Guo, J., and Li, Pe. (2013b). “A simple fuzzy
system for modelling of both rate-independent and rate-dependent hysteresis in piezoelec-
tric actuators”. In: Mechanical Systems and Signal Processing 36(1), pp. 182 –192 (cit. on
p. 13).

Lin, Ch.-J. and Lin, P.-T. (2012). “Tracking control of a biaxial piezo-actuated positioning
stage using generalized Duhem model”. In: Computers and Mathematics with Applications
64(5), pp. 766 –787 (cit. on p. 12).

Mahmood, I.A. and Moheimani, S.O.R. (2009a). “Fast spiral-scan atomic force microscopy”.
In: Nanotechnology 20(36), 365503–365507 (cit. on pp. 18, 102).

Mahmood, I.A. and Moheimani, S.O.R. (2009b). “Improvement of accuracy and speed of a
commercial AFM using positive position feedback control”. In: 2009 American Control
Conference (ACC 2009), pp. 973–978 (cit. on p. 17).

Mahmood, I.A. and Moheimani, S.O.R. (2009c). “Making a commercial atomic force micro-
scope more accurate and faster using positive position feedback control”. In: Review of
Scientific Instruments 80(6), pp. 063705–13 (cit. on p. 18).
Bibliography 151

Mayergoyz, I.D. (1988). “Dynamic Preisach models of hysteresis”. In: IEEE Transactions on
Magnetics 24(6), pp. 2925–2927 (cit. on p. 13).

Merry, R., Uyanik, M., Koops, R., Molengraft, R. van de, Veghel, M. van, and Steinbuch, M.
(2008). “Modeling, identification and control of a metrological Atomic Force Microscope
with a 3DOF stage”. In: 2008 American Control Conference (ACC 2008), pp. 2716–2721
(cit. on p. 19).

Moheimani, S.O.R. (2008). “Accurate and fast nanopositioning with piezoelectric tube scan-
ners: emerging trends and future challenges”. In: Review of Scientific Instruments 79(7),
p. 071101 (cit. on p. 15).

Munteanu, E. and Ursu, I. (2008). “Piezo Smart Composite Wing with LQG/LTR Control”.
In: IEEE International Symposium on Industrial Electronics, (ISIE 2008), pp. 1160–1165
(cit. on pp. 18, 66).

Necipoglu, S., Cebeci, S.A., Basdogan, C., Has, Y.E., and Guvenc, L. (2011). “Repetitive
control of an XYZ piezo-stage for faster nano-scanning: Numerical simulations and exper-
iments”. In: Mechatronics 21(6), pp. 1098 –1107 (cit. on p. 17).

Pedrak, R., Ivanov, T., Ivanova, K., Gotszalk, T., Abedinov, N., Rangelow, I. W., Edinger, K.,
Tomerov, E., Schenkel, T., and Hudek, P. (2003). “Micromachined atomic force microscopy
sensor with integrated piezoresistive sensor and thermal bimorph actuator for high-speed
tapping-mode atomic force microscopy phase-imaging in higher eigenmodes”. In: Journal
of Vacuum Science and Technology B 21(6), pp. 3102–3107 (cit. on p. 19).

Rakotondrabe, M. (2011). “Bouc-Wen Modeling and Inverse Multiplicative Structure to Com-

pensate Hysteresis Nonlinearity in Piezoelectric Actuators”. In: IEEE Transactions on
Automation Science and Engineering 8(2), pp. 428–431 (cit. on pp. 12, 16).

Rakotondrabe, M., Clevy, C., and Lutz, P. (2010). “Complete Open Loop Control of Hysteretic,
Creeped, and Oscillating Piezoelectric Cantilevers”. In: IEEE Transactions on Automation
Science and Engineering 7(3), pp. 440–450 (cit. on pp. 16, 37).

Ryba, L., Voda, A., and Besancon, G. (2013). “Modelling and control of 3D STM-like scanning
device with application to surface reconstruction”. In: Methods and Models in Automation
and Robotics (MMAR), 2013 18th International Conference on, pp. 479–484 (cit. on pp. 20,
117).

Ryba, L., Voda, A., and Besancon, G. (2014a). “3DOF nanopositioning control of an experi-
mental tunneling current-based platform”. In: 2014 IEEE Conference on Control Applica-
tions (CCA 2014, part of MSC 2014), pp. 1976–1981 (cit. on pp. 36, 92).
152 Bibliography

Ryba, L., Voda, A., and Besancon, G. (2014b). “An LQG/LTR approach towards piezoactuator
vibration reduction with observer-based hysteresis compensation”. In: Proceedings of the
19th IFAC World Congress (IFAC 2014), pp. 5623–5628 (cit. on pp. 17–19, 36).

Ryba, L., Dokoupil, J., Voda, A., and Besancon, G. (2015a). “A real-time inverse-based hys-
teresis compensation with adaptation”. In: 54th IEEE Conference on Decision and Control
(CDC 2015) (accepted) (cit. on pp. 36, 53).

Ryba, L., Voda, A., and Besancon, G. (2015b). “Experimental comparison of disturbance
observer and inverse-based hysteresis compensation in 3D nanopositioning piezoactuation”.
In: Sensors and Actuators A: Physical 236, pp. 190–205 (cit. on pp. 36, 46).

Safonov, M.G. and Athans, M. (1977). “Gain and phase margin for multiloop LQG regulators”.
In: IEEE Transactions on Automatic Control 22(2), pp. 173–179 (cit. on pp. 18, 65).

Salapaka, M.V., Bergh, H.S., Lai, J., Majumdar, A., and McFarland, E. (1997). “Multi-mode
noise analysis of cantilevers for scanning probe microscopy”. In: Journal of Applied Physics
81(6), pp. 2480 –2487 (cit. on pp. 30, 121).

Schitter, G., Stark, R.W., and Stemmer, A. (2002). “Sensors for closed-loop piezo control:
strain gauges versus optical sensors”. In: Measurement Science and Technology 13(4), N47–
N48 (cit. on p. 19).

Schitter, G., Thurner, P.J., and Hansma, P.K. (2008a). “Design and input-shaping control of a
novel scanner for high-speed atomic force microscopy”. In: Mechatronics 18(5–6), pp. 282
–288 (cit. on pp. 17, 18).

Schitter, G., Rijkee, W.F., and Phan, N. (2008b). “Dual actuation for high-bandwidth nanopo-
sitioning”. In: 47th IEEE Conference on Decision and Control (CDC 2008), pp. 5176–5181
(cit. on p. 18).

Schneir, J., McWaid, T.H., Alexander, J., and Wilfley, B.P. (1994). “Design of an atomic
force microscope with interferometric position control”. In: Journal of Vacuum Science
and Technology B 12(6), pp. 3561–3566 (cit. on p. 19).

Sebastian, A. and Salapaka, S.M. (2005). “Design methodologies for robust nano-positioning”.
In: IEEE Transactions on Control Systems Technology 13(6), pp. 868–876 (cit. on pp. 11,
18, 66).

Shan, Y. and Leang, K.K. (2012). “Accounting for hysteresis in repetitive control design:
Nanopositioning example”. In: Automatica 48(8), pp. 1751 –1758 (cit. on p. 12).
Bibliography 153

Shi, J., Wu, Y., Su, Ch., and Zou, Q. (2009). “A control approach to cross coupling compen-
sation of piezotube scanners in tapping-mode imaging”. In: vol. 7378, 73781E–73781E–10
(cit. on pp. 15, 18, 113).

Silverman, L.M. (1969). “Inversion of multivariable linear systems”. In: IEEE Transactions on
Automatic Control 14(3), pp. 270–276 (cit. on p. 17).

Skogestad, S. and Postlethwaite, I. (2005). Multivariable Feedback Control: Analysis and De-
sign. John Wiley & Sons (cit. on pp. 78, 85).

Smith, R. and Ounaie, Z. (2000). “A domain wall model for hysteresis in piezoelectric ma-
terials”. In: Journal of Intelligent Material Systems and Structures 11(1), 62–79 (cit. on
p. 12).

Stark, R.W., Schitter, G., Stark, M., Guckenberger, R., and Stemmer, A. (2004). “State-
space model of freely vibrating and surface-coupled cantilever dynamics in atomic force
microscopy”. In: Physical Review B 69(8), pp. 085412–085421 (cit. on p. 28).

Sun, L., Ru, Ch., Rong, W., Chen, L., and Kong, M. (2004). “Tracking control of piezoelec-
tric actuator based on a new mathematical model”. In: Journal of Micromechanics and
Microengineering 14(11), p. 1439 (cit. on p. 13).

Sun, Y. and Pang, J.H.L. (2006). “AFM image reconstruction for deformation measurements
by digital image correlation”. In: Nanotechnology 17(4), 933–939 (cit. on p. 17).

Tan, X. and Baras, J.S. (2005). “Adaptive identification and control of hysteresis in smart
materials”. In: IEEE Transactions on Automatic Control 50(6), pp. 827–839 (cit. on p. 16).

Tan, X. and Bennani, O. (2008). “Fast inverse compensation of Preisach-type hysteresis oper-
ators using field-programmable gate arrays”. In: 2008 American Control Conference (ACC
2008), pp. 2365–2370 (cit. on p. 16).

Tien, S., Zou, Q., and Devasia, S. (2004). “Control of dynamics-coupling effects in piezo-
actuator for high-speed AFM operation”. In: 2004 American Control Conference (ACC
2004). Vol. 4, pp. 3116–3121 (cit. on pp. 15, 18).

Tsukada, M., Kobayashi, N., Brandbyge, M., and Nakanishi, S. (2000). “Physics of artificial
nano-structures on surfaces”. In: Progress in Surface Science 64(3–8), pp. 139 –155 (cit. on
p. 6).

Tuma, T., Lygeros, J., Kartik, V., Sebastian, A., and Pantazi, A. (2012). “High-speed mul-
tiresolution scanning probe microscopy based on Lissajous scan trajectories”. In: Nanotech-
nology 23(18), 185501–185510 (cit. on p. 18).
154 Bibliography

Wu, Y. and Zou, Q. (2007). “Iterative Control Approach to Compensate for Both the Hysteresis
and the Dynamics Effects of Piezo Actuators”. In: IEEE Transactions on Control Systems
Technology 15(5), pp. 936–944 (cit. on p. 17).

Xie, H., Onal, C., Régnier, S., and Sitti, M. (2012). Atomic Force Microscopy Based Nanorobotics.
Springer-Verlag (cit. on p. 6).

Xu, Q. and Li, Y. (2010). “Dahl Model-Based Hysteresis Compensation and Precise Positioning
Control of an XY Parallel Micromanipulator With Piezoelectric Actuation”. In: Journal of
Dynamic Systems, Measurement, and Control 132(4), p. 041011 (cit. on p. 12).

Yamanaka, K., Noguchi, A., Tsuji, T., Koike, T., and Goto, T. (1999). “Quantitative material
characterization by ultrasonic AFM”. In: Surface and Interface Analysis 27(5-6), pp. 600–
606 (cit. on p. 6).

Yeh, T.-J., Wang, Ch.-D., and Wu, T.-Y. (2008). “Modeling and control of an atomic force
microscope using a piezoelectric tuning fork for force sensing”. In: Simulation Modelling
Practice and Theory 16(7), pp. 768 –782 (cit. on pp. 18, 66).

Yi, J., Chang, S., and Shen, Y. (2009). “Disturbance-Observer-Based Hysteresis Compensa-
tion for Piezoelectric Actuators”. In: IEEE/ASME Transactions on Mechatronics 14(4),
pp. 456–464 (cit. on pp. 17, 35).

Yong, Y.K., Moheimani, S.O.R., and Petersen, I.R. (2010a). “High-speed cycloid-scan atomic
force microscopy”. In: Nanotechnology 21(36), 365503–365507 (cit. on p. 18).

Yong, Y.K., Liu, K., and Moheimani, S.O.R. (2010b). “Reducing Cross-Coupling in a Com-
pliant XY Nanopositioner for Fast and Accurate Raster Scanning”. In: IEEE Transactions
on Control Systems Technology 18(5), pp. 1172–1179 (cit. on pp. 15, 18).

Zou, Q., Leang, K., Sadoun, E., Reed, M., and Devasia, S. (2004). “Control Issues in High-
speed AFM for Biological Applications: Collagen Imaging Example”. In: Asian Journal of
Control 6(2), pp. 164–178 (cit. on p. 6).