Sensors and Actuators A 102 (2002) 139150

Effect of cubic nonlinearity on auto-parametrically

amplified resonant MEMS mass sensor
Wenhua Zhang*, Rajashree Baskaran1, Kimberly L. Turner
Department of Mechanical and Environmental Engineering, University of California, Engineering II, R2355,
Santa Barbara, CA 93106-5070, USA
Received 4 April 2002; accepted 9 August 2002

Abstract
Parametric resonance has been well established in many areas of science, including the stability of ships, the forced motion of a swing and
Faraday surface wave patterns on water. We have previously investigated a linear parametrically driven torsional oscillator and along with
other groups have mentioned applications including mass sensing, parametric amplification, and others. Here, we thoroughly investigate the
design of a highly sensitive mass sensor. The device we use to carry out this study is an in-plane parametrically resonant oscillator. We show
that in this configuration, the nonlinearities (electrostatic and mechanical) have a large impact on the dynamic response of the structure. This
result is not unique to this oscillatormany MEMS oscillators display nonlinearities of equal importance (including the very common parallel
plate actuator). We report the effects of nonlinearity on the behavior of parametric resonance of a micro-machined oscillator. A nonlinear
Mathieu equation is used to model this problem. Analytical results show that nonlinearity significantly changes the stability characteristics of
parametric resonance. Experimental frequency response around the first parametric resonance is well validated by theoretical analysis. Unlike
parametric resonance in the linear case, the jumps (very critical for mass sensor application) from large response to zero happen at additional
frequencies other than at the boundary of instability area. The instability area of the first parametric resonance is experimentally mapped.
Some important parameters, such as damping co-efficient, cubic stiffness and linear electrostatic stiffness are extracted from the nonlinear
response of parametric resonance and agree very well with normal methods.
# 2002 Elsevier Science B.V. All rights reserved.
Keywords: Mass sensor; Chemical sensor; Nonlinearity; Parametric resonance; Damping; Mathieu equation; MEMS

1. Introduction
Parametric resonance has been well established in many
areas of science, including the stability of ships [1], the forced
motion of a swing [2] and Faraday surface wave patterns on
water [3]. We have previously investigated a linear parametrically driven torsional oscillator [4] and along with other
groups have mentioned applications including mass sensing
[5], parametric amplification [6], and others [7].
For high-precision chemical detection (MEMS nose) [8],
mass spectrometry [9], small force detection [1012], etc.
cantilever-based MEMS sensors have been utilized. Microscale oscillators are promising due to their small mass and
high sensitivity. Many MEMS based mass sensors track
shifts in resonant frequency due to changing mass. We have
*
Corresponding author. Tel.: 1-805-893-7849; fax: 1-805-893-8651.
E-mail address: [email protected] (W. Zhang).
1
Rajashree Baskaran has the same contribution to this paper as Wenhua
Zhang.

previously introduced the framework for developing a parametrically driven mass sensor with sensitivity of femtogram
(10
15 g) mass variation [5]. The presence of cubic mechanical and electrostatic nonlinearity significantly alters the
dynamic behavior of the system. Some of these nonlinear
effects are beneficial for the use of the oscillator as a mass
sensor and a detailed investigation is warranted.
A simple harmonic oscillator (SHO) usually refers to an
oscillator that can be modeled by a second-order constant
co-efficient differential equation with time as the independent variable. In a SHO with 1 d.f., when excited with an
external periodic force, the response of the oscillator is
restricted to the externally applied frequency. Also, the
response is amplified near the resonant frequency and,
depending on the damping, the response at all frequencies
sufficiently far away from the resonant frequency is minimal. This is not the case when the co-efficients of the
second-order system are periodic in time (referred to as
parametric oscillators). The dynamics of such oscillators is
significantly different from that of the SHO. The response of

0924-4247/02/$ see front matter # 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 9 2 4 - 4 2 4 7 ( 0 2 ) 0 0 2 9 9 - 6

140

W. Zhang et al. / Sensors and Actuators A 102 (2002) 139150

the oscillator is no longer restricted to the frequency of the

externally applied force. In particular, there are excitations of
the natural frequency, when excited at certain integer fractions
or multiples of the resonant frequency. There is plenty of
literature on such systems modeled by Hills equation and
variations including the introductory book by Cartmell [13].
In the case of harmonic oscillators with time-modulated
stiffness, a sharp transition between zero response and a
large auto-parametric response (sub-harmonic resonance)
exists [14]. Since this transition is dependent on system
parameters, including the mass of the vibrating oscillators,
change in mass can be detected with such a system. In this
mass sensor implementation, the minimum detectable mass
change can be expressed as [5]:
dm

k 1
df0
4p2 f03

(1a)

Theoretically, good sensitivity can be achieved even at

atmospheric pressure.
In this paper, we describe the motivation for the need to
study nonlinear effects and present an introduction to the
bulk-micro-machined parametrically actuated mass sensor.
We then present the development of its model in the
electromechanical domain and present an analytical treatment of the resulting nonlinear Mathieu equations. This
analysis is supported with experimental results. The implications of these results on the behavior of the mass sensor are
discussed.

2. Device

(2)

In this section, we describe the electromechanical system

fabricated using the bulk-micro-machining technique
SCREAM [21]. The device we have studied is an oscillator,
which was designed by Adams et al. [22] for the independent
tuning of linear and cubic stiffness terms. A scanning
electron micrograph of the oscillator is shown in Fig. 1.
The device size is about 500 mm  400 mm. It has two sets of
parallel interdigitated comb finger banks on either end of the
backbone and two sets of non-interdigitated comb fingers on
each side. The four folded beams provide elastic recovery
force for the oscillator. The beams, backbone and the fingers
are 2 mm wide and 12 mm deep. The backbone is 515 mm
long and 20 mm wide. Each of the four recovery folded
beams are 200 mm on the long side and 20 mm on the short
side. Either the interdigitated or the non-interdigitated comb
fingers may be used to drive the oscillator. Fig. 2 is a
schematic of these comb fingers. This oscillator is used to

where A is the amplitude of the oscillation, B the bandwidth,

Q the quality factor, T the temperature and k the stiffness of
the oscillator. It should be noted that damping cannot be
avoided in micro or nano scale [19]. In the case of a
parametrically driven oscillator, the sensitivity depends on
the transition between zero and large response and the
transition can be very sharp. At 7 mTorr, for the case of a
parametric torsional mode MEMS oscillator, the transition
was observed with an input frequency shift of 0.001 Hz [4],
which is the limit of the hardware used (Function generator
HP3245A). Because of the sharp transition, with the same
configuration, parametrically driven mass sensor can be two
orders of magnitude more sensitive than harmonic resonator
based mass sensor. In a cantilever mass senor with dimensions as 22:37 mm  2 mm  0:5 mm, the theoretical sensitivity can be 9:65e
17g [20]. If working in the parametric
mode, it can resolve as small a mass change as 3:62e
19 g.
The sharp transition is a reflection of a sub-harmonic
response pitchfork bifurcation in the driving voltage frequencyamplitude parameter space. The occurrence of this
bifurcation will be shown to be independent of the ambient
pressure (modeled as viscous damping in the dynamics). We
have observed this sharp transition at 450 mTorr and higher.

Fig. 1. A scanning electron micrograph of the oscillator. Note the folded

beam springs (S), the two sets of interdigitated comb finger banks (C) on
both ends of backbone (B) and non-interdigitated comb fingers (N) on each
side of backbone (B).

where k is the system stiffness and f0 the natural frequency.

The sensitivity of a simple harmonic resonator based mass
sensor, such as a cantilever sensor, can be represented as
[1517]:



k
1
1
k 1
dm 2 0 2
2
2 3 df0
(1b)
4p f 0 f0
2p f0
The sensitivity of these two cases is of the same order if
the smallest resolvable frequency shift (df0) due to mass
change is the same. However, the sensitivity of a normal
cantilever mass sensor is strongly dependent on pressure.
The minimum detectable frequency change is inversely
proportional to the quality factor [18]:
Df

1 f0 kB TB
A kQ

W. Zhang et al. / Sensors and Actuators A 102 (2002) 139150

Fig. 2. Schematic of (a) interdigitated comb fingers and (b) noninterdigitated comb fingers that are present in the design of the oscillator
shown in Fig. 1.

study the nonlinear effects on the parametric resonance

phenomenon. An optimized design (stiffer springs and lesser
mass) with a reactive coating specific to the mass sensing
application will be used in future versions of the sensor
implementations.
The driving force of the non-interdigitated comb finger is
proportional to displacement, while the driving force of
interdigitated comb fingers is nearly independent of displacement [22]. When a time varying voltage signal is applied
to the non-interdigitated fingers, the effective stiffness of the
system (mechanical and electrostatic) is dependent on time.
At specific frequencies (related to the natural frequency of
the device and the strength of the electrostatic drive) and
amplitudes of the driving voltage, the sub-harmonics parametric resonance can be excited [4].

3. Theoretical analysis
The approach towards modeling the system is as followswe use numerical simulation to estimate the mechanical and electrostatic parameters of the oscillator design
separately. In the mechanical domain, the details of the
spring design predominantly govern the stiffness of the
oscillator. A finite element package, ANSYS [23], was used
to simulate the spring stiffness. In the electrostatic domain,
the configuration of the comb fingers (interdigitated or noninterdigitated) and the dimensions (length of overlap, gap
between fingers) govern the electrostatic force generated.
The strengths of these electrostatic forces were estimated
using a boundary element solver COULOMB [24]. The
parameters obtained from these simulations were then used
in a single degree of freedom model.
To calculate the stiffness of beams, the device was
simplified as shown in Fig. 3, where the dimensions of
the folded beams (see Fig. 3(b)) are exactly the same as the

141

Fig. 3. Simplified models used in estimation of mechanical stiffness using

ANSYS: (a) shows the schematic of the fixedfixed beams and (b) shows
the folded beam configuration.

oscillator design and the backbone is represented as a solid

beam connecting the two sets of folded beams. Additionally,
a fixedfixed beam (see Fig. 3(a)) with dimensions comparable to the folded beam was used to study the effect of the
folding configuration. ANSYS was used to simulate these
two models. The dimensions used in the simulations are as
follows: length of springs, Ls is 200 mm; length of folded
springs, Lf is 20 mm and backbone equivalent length B is
50 mm. The depth of the features was taken as 12 mm and all
the anchor points were assumed fixed in displacement and
slope of displacement. The material of the structures was
assumed to be single crystal silicon with a youngs modulus
of 170 GPa. The results are shown in Fig. 4. At small
displacement, the stiffness is close for these two models.
But when the displacement is large, their stiffness varies
significantly. To get the linear and higher order stiffness, we
fit the curve with polynomial functions. The recovery force
can be expressed as:
Fr x k1 x k3 x3

(3)

where k1 is the linear stiffness and k3 the cubic stiffness.

Fig. 4. The results of mechanical stiffness calculation using ANSYS. The

curves are fitted with polynomials to extract the linear and cubic stiffness
parameters tabulated in Table 1.

142

W. Zhang et al. / Sensors and Actuators A 102 (2002) 139150

Table 1
Simulation results of linear stiffness k1 and cubic stiffness k3

Fixedfixed spring
Folded spring

k1 (mN/mm)

k3 (mN/mm3)

8.14
7.61

5.8
4.4e
2

HASP
COULOMB

The results are shown in Table 1. In the folded beam model,

the cubic stiffness is two orders less than with the fixedfixed
model, while the linear stiffness is comparable. Though the
cubic stiffness is reduced, it is not negligible in effect on
parametric resonance behavior as we show with perturbation
analysis and experimental results. We can see from the simulation results in Fig. 4 that the total recovery force grows very fast
with displacement for the fixedfixed beam configuration.
Electrostatic force of the non-interdigitated comb finger
was calculated using COULOMB. This boundary element
solver can compute the electrostatic forces generated in any
configuration of the comb finger spacing. We use the parametric modeling ability in the software to map the electrostatic forces as a function of small displacements of the
movable comb fingers relative to the fixed fingers, as shown
in Fig. 5. We then define electrostatic stiffness as the
gradient of this forcedisplacement curve. By fitting a
polynomial curve, we get the linear and cubic electrostatic
stiffness, r1 and r3 separately (see Table 2). The electrostatic
force can be expressed as [22]:
3

Fe x
r1 x r3 x V

(4)

We also show the value of the linear electrostatic stiffness

as found by Adams et al. [22] using a combined Finite
Element/Boundary Element package HASP. The two results,
within 15%, show good correlation.
The oscillator can be simplified as a massspring system
with electrostatic force as the driving force. Using Newtons
laws, the motion of the device can be described by the
following equation:
m

d2 x
dx
c Fr x Fe t; x
dt2
dt

Table 2
Linear electrostatic stiffness r1 and cubic electrostatic stiffness r3
calculated using HASP and COULOMB

(5)

r1 (mN/(mm/V2))

r3 (mN/(mm
3 V
2))

3.9e
4

3.65e
4

1.4e
5

1.6e
5

where x is displacement, m the mass, c the damping coefficient, Fr(x) the elastic recovering force and Fe(t, x) the
electrostatic force. Since either interdigitated comb fingers
or non-interdigitated comb fingers can be used to drive the
oscillator and the dynamic response will be different, these
two cases are addressed, respectively.
3.1. Harmonic excitation using interdigitated
comb fingers
When an ac voltage signal is applied on the interdigitated
comb fingers, Fe is only a function of time, t. Fe can be
written as:
Fe FA cosot

(6)

Substituting Fe and Fr into Eq. (5) and normalizing it,

2

d x
dx
a bx d3 x3 f cost
2
dt
dt

(7)

where
a

c
;
mo

k1
;
mo2

d3

k3
;
mo2

FA
mo2

This is a Duffing equation [25]. Considerable work has

been done on Duffing equation in MEMS devices to study
nonlinear effects on harmonic resonance [22,2628]. By
fitting the frequency response curve, cubic stiffness can be
found.
3.2. Parametric resonance excitation using noninterdigitated comb fingers
When the device is driven by non-interdigitated comb
fingers, Fe is a function of displacement x and time t. Since
the force due to electrostatic interactions has a square dependence on the voltage applied, in order to isolate the parametric
effects from harmonic effects, we use a square rooted sinusoidal voltage signal [14]. For the case of square rooted
sinusoidal voltage being applied (V VA 1 cosot1=2 ):
Fe x; t
r1 x r3 x3 VA2 1 cosot

(8)

Substituting Fe and Fr in Eq. (5) and then normalizing

Fig. 5. Numerical results of variation of electrostatic force with
displacement of non-interdigitated comb fingers simulated using COULOMB. Linear electrostatic stiffness r1 and cubic electrostatic stiffness r3
can be estimated by fitting a cubic polynomial. The results with
experimental data comparison are shown in Table 2.

it,
d2 x
dx
a b 2d cos 2tx d3 d03 cos 2tx3 0
dt2
dt
(9)

W. Zhang et al. / Sensors and Actuators A 102 (2002) 139150

where
2c
;
b
mo
4k3 4r3 VA2
;
d3
mo2

4k1 r1 VA2
;
mo2
d03

2r1 VA2
mo2

It should be noted that we are working with first-order

expansion, neglecting O(e2) terms.
The solution to the harmonic oscillator in x0 (Eq. (14)) is
as follows,

4r3 VA2
mo2

x0 AZ cos 12 x BZ sin 12 x

This is a nonlinear Mathieu equation. All the terms in this

equation except b (which is of order 1) is of small order. The
values of the non-dimensional parameters from simulation at
typical voltage amplitude (20 V) and frequency (52 kHz) are
b 1:01;
a 1:50e
3;
d
1:01e
2;
d3 5:10e
3;
d03
8:75e
4
Hence, we can use perturbation methods to study this
system.
By assuming d e, b b0 eb1 , the equation can be
rewritten as a harmonic oscillator with a perturbation (the
terms on the right hand side of Eq. (10)).
d2 x
b x
dt2
0
e
b1 2 cos 2tx g3

g03

dx
cos 2tx
m
dt
3

143

(16)

Unlike the simple harmonic oscillator case, we have the

constants A and B varying in slow-time scale. Using
Eq. (16) to evaluate the right hand side of the Eq. (15) in
x1 and setting the condition for removal of the resonant
terms, yields the slow-flow equations in A and B (the
dynamics of A and B in slow-time).
dA
m
B
3Bg3 2
g 0 B3

B b1
1
A B2
3
dZ
2
2
8
4

(17)

dB
m
A
3Ag3 2
g 0 A3

A
b1 1

A B2
3
dZ
2
2
8
4

(18)

The characteristics of these two equations are schematically shown as A
B in b
d plane (see Fig. 6(a)). The plane
can be divided into three areas. In area I, one center exists at

(10)
where the terms in Eq. (10) have been scaled with respect to
the small term e,
g3

d3
;
e

g03

d03
;
e

a
e

(11)

We know that for a linear Mathieu equation, the first-order

parametric resonance occurs when the driving frequency is
near twice the resonant frequency (here that would correspond to b0 1). So, we use the method of two variable
expansion [28] to analyze the equation when driving near
the first parametric resonance. The idea of the method is that
the expected solution involves two time scales: the time
scale of the periodic motions and a slower time scale that
modulates the amplitude of the periodic motion. Here, we
use the notation that x represents stretched time (ot), and Z
represents slow-time (et) [28]:
x 2t;

Z et

(12)

And assuming,
x x0 ex1 ;

b 1 eb1

(13)

Using two-variable-expansion method, the equation of

motion can be rewritten as two equations:
4

@ 2 x0
x0 0
@x2

(14)

and
4

@ 2 x1
@ 2 x0

b1 2 cos xx0
x1
4
2
@x@Z
@x

g3 g03 cos xx30
2m

@x0
@x

(15)

Fig. 6. Dynamic characteristics of nonlinear Mathieu equation in the bd

plane. b 1 d are the transition curves, which divide bd plane into
area I, II,
III. Note the damping effects on transition curves
p
(b 1 d 1
m2 ) in Fig. 6(a). The number of stable points change in
each areaone center at (0, 0) in area I, two centers at ( a1, 0), and one
saddle at (0, 0) in area II, three centers at ( a2, 0) and (0, 0) and two
saddles at (0, b) in area III. Fig. 3(b) shows how the positions of the
stable (dark trace) and unstable (broken trace) points vary as b and d are
varied quasi-statically.

144

W. Zhang et al. / Sensors and Actuators A 102 (2002) 139150

(0, 0) which means only one stable trivial solution exists in the
area. In area II, there are two centers at ( a1, 0) and one saddle
at (0, 0), corresponding to one stable non-trivial solution and
one unstable trivial solution. In area III, there are two centers at
( a2, 0), two saddles at (0, b) and one center at (0, 0),
corresponding to one stable non-trivial solution, one unstable
non-trivial solution and one stable trivial solution.
It is convenient to study Eqs. (17) and (18) in polar
system. Assuming
A R cos y;

B R sin y

(19)

We get



dR
mR R
g0 R 2




1 3
sin2y
dZ
2
2
4

0

dy
b
3g R2
g
1

1
3
3 R2
cos2y
dZ
2
2
8
4

(20)
(21)

The fixed points (R

, y
) of these coupled equations are
obtained by setting dR=dZ 0 and dy=dZ 0. This
yields:
sin2y


and
R

4m
4 g03 R
2



4
b1 cos2y



3 g3 2=3g03 cos2y

(22)

(23)

Fixed points (centers and saddles) occur when there is a

real value for R
and to facilitate the discussion, we define a
new variable,
g3eff g3

2g03
cos 2y
3

(24)

Inserting g3 d3 =e and g03 d03 =e into Eq. (24)

yields:



2k3
r3
4

cos2y
g3eff

(25)
3
r1 VA2 r1
where g3eff is the effective nonlinearity parameter of the
system, a sum of contributions from cubic mechanical
stiffness,(fixed for a particular beam design) and voltage
dependent cubic electrostatic stiffness. Looking at Eq. (9), it
might be tempting to conclude that the effective contributions of the mechanical and electrostatic cubic stiffness will
be just a linear sum of k3 and r3 VA2 .
The detailed analysis presented above shows that the
effective contribution is subtle and depends on the linear
electrostatic stiffness r1 and the response of the oscillator
itself (the cos(2y
) term). In the current design, the mechanical cubic stiffness k3 is positive and the electrostatic cubic
stiffness r3 is negative. Hence by varying the applied voltage, VA, we can change the sign of the effective cubic
stiffness. Fig. 7 is the variation of g3eff with the input voltage
for the device under study. The effect of the sign of the
effective cubic stiffness is presented in detail in Section 5.
For clarity of discussion, we present the influence of
damping and nonlinear terms on parametric resonance
separately.
3.2.1. Effect of damping (assuming no cubic nonlinearity)
Eqs. (17) and (18) can be simplified as:
dA

12 mB 12 Bb1
1
dZ

(26)

dB

12 mA
12 Ab1 1
dZ

(27)

Fig. 7. The effective nonlinear parameter as a function of the driving voltage amplitude VA for the spring and electrostatic comb finger design of the oscillator
studied in this paper. Note that the sign of effective nonlinear parameter can be changed at large voltage amplitude.

W. Zhang et al. / Sensors and Actuators A 102 (2002) 139150

Eqs. (26) and (27) are called the slow-flow equations.

These are coupled linear first-order differential equation that
can be easily solved. The slow-time dynamics of the variable
A and B distinguish the resonant and non-resonant regions.
The transition of the real part of the eigenvalues from
positive (exponentially growing solution) to negative (exponentially decaying solution), corresponds to the transition
curve from unstable to stable areas in the first parametric
resonance stability plot [28]:
p
p
b 1 e 1
m2 1 e 2
a 2
(28)

145

of R, equilibrium points are

y
0;

p
;
2

p;

3p
2

and
R
2

4
b cos2y

3g3eff 1

(31)

3.2.2. Effect of cubic nonlinearity (assuming no damping)

We look at the slow-flow equations in polar coordinates
here, assuming damping co-efficient m 0.

Let us assume the effective nonlinearity parameter

g3eff < 0. In the case of y
0 and p, R
2

4=3g3eff b1 1, non-trivial solutions require b1 >
1.
When y
p=2 and 3p/2, R
2
4=3g3eff b1
1, nontrivial solutions exist only for b1 > 1. The characteristics of
Eq. (31) are schematically shown in Fig. 6(b). Since b1 1
corresponds to transition curves from stable to unstable areas in
b
d plane, bifurcation occurs when we quasi-statically vary
frequency of the input voltage across the transition curve. The
growth of R
with respect to b in different areas is schematically shown in Fig. 6(b), where solid line represents stable
solution and dashed line represents unstable solution. The
solutions in different cases are reflected in the results presented
in A
B plane (Fig. 6(a)).

dR

12 R sin2y
dZ

(29)

4. Experimental results

dy
b
3g R2 1

1
3eff
cos2y
dZ
2
2
8

(30)

Fig. 6 also shows the effects of damping on parametric

resonance. The presence of damping limits the occurrence of
the parametric resonance to above certain critical input
voltage amplitude. By substituting a 2c=mo, b
4k1 r1 VA2 =mo2 , e d 2r1 VA2 =mo2 , we can find
the equation of boundary curve in VA
f plane and get
the critical input voltage.

We study the equilibrium (fixed) points of the slow-flow

variable (R
, y
) and analyze their stability characteristics.
The equilibrium points are when the right hand side of the
above equations is identically zero. For non-trivial solutions

A multi-dimensional MEMS motion characterization

suite is used to measure the in-plane movement of the device
[29]. The schematic setup is shown in Fig. 8. The device is
placed in a vacuum chamber, where the pressure can be
pumped to as low as 7 mTorr. The Laser Doppler Vibrometer
with built in controllers and sensor heads (Polytec, OFV-

Fig. 8. Schematic of characterization suite used to test the MEMS oscillator. For the results presented in this paper, we used a vacuum chamber with a
pressure of 7 mTorr.

146

W. Zhang et al. / Sensors and Actuators A 102 (2002) 139150

3001, and OFV-511) uses a 633 nm wavelength, 205 mm

cavity length HeliumNeon laser. The principle of measurement is optical interferometry. It is a heterodyne interferometer, meaning the velocity magnitude as well as the
direction is determined by use of an additional Bragg cell
[30]. A voltage source (HP3245A) is used for input voltage
signal generation and the displacement and velocity outputs
from the Vibrometer were recorded and analyzed with a HP
Spectrum Analyzer (HP89470A) and Tektronics Oscilloscope (TDS 420 A).
To obtain natural frequency and quality factor (defined as
the ratio between the frequency of the peak amplitude and the
band width of the frequency response at half power points), an
ac voltage signal (20 V) is applied on the interdigitated
comb fingers. For this design, a driving voltage of 20 Vapplied
to the interdigitated comb fingers results in an amplitude of
motion less than 1 mm, which can be adequately modeled with
a linear stiffness. The position independent actuation of the
interdigitated fingers is used in order to avoid parametric
excitation effects. At 7 mTorr, the natural resonance frequency
is 26.48 kHz and quality factor is 640 (see Fig. 9).
A larger electrical signal (60 V) is applied on the
interdigitated comb finger to study the nonlinear behavior
of the device. The motion of the device under this driving
voltage is greater than 2.5 mm. Its frequency response follows that of a Duffing equation response (see Fig. 10)
[25,27]. By fitting the curve with the solution to the Duffing
equation, cubic stiffness can be obtained [26,27]. At
7 mTorr, k3 is found to be 0.030 mN/mm3, which compares
well with FEM results (see Table 1).
To avoid coupling between harmonic resonance and
parametric resonance, a square root sinusoidal ac voltage
signal is applied to the non-interdigitated comb fingers [14].
The device displays parametric resonance when excited with
a frequency around twice the natural frequency. The timescale growth of the response (Fig. 11(b)) is quite different

Fig. 10. The oscillators frequency response curve when excited with the
interdigitated comb fingers with a high voltage (VA 60 V). This isolates
the effect of mechanical cubic nonlinearity with that of the electrostatic
nonlinearity since the latter arises only when driven by non-interdigitated
comb drives. The experimental data is represented by squares and the
smooth line is the fit using solution to a Duffing equation. The mechanical
cubic stiffness extracted from this curve is k3 0:030 mN/mm3.

Fig. 9. The oscillators frequency response curve when excited with the
interdigitated comb fingers with a small ac voltage (VA 20 V). At small
amplitudes of motion, the oscillator is essentially linear and the linear
stiffness to mass ratio (natural frequency) and the damping co-efficient
(Qthe quality factor) values are extracted from this data.

Fig. 11. Comparison of timeresponse of oscillator velocity when excited

at (a) harmonic resonance (driving voltage 60 V) and (b) first-order
parametric resonance (driving voltage 20 V). The contrasting feature here
is that the latter has an exponential envelope of growth of the sinusoidal
amplitude, while the former has linear growth.

W. Zhang et al. / Sensors and Actuators A 102 (2002) 139150

Fig. 12. Map of the first parametric resonance transition curve in VA
f

plane (driving voltage amplitudedriving frequency) at 7 mTorr. The
squares represent experimental data. The equation of the curve fit is
derived from nonlinear Mathieu Eq. (28). The Q (quality factor) and linear
electrostatic stiffness r1 can be found by fitting this curve.

from that of a harmonic resonance (see Fig. 11(a)). When

excited from a non-resonant point in the input voltage
frequency space to a resonant region, parametric resonance
grows exponentially with time while harmonic resonance
grows linearly, as shown in Fig. 11(a) and (b).
The first parametric resonance instability area is mapped
by sweeping frequency and amplitude of the input voltage
signal (see Fig. 12). By fitting this data at the transition curve
with the analytical result (Eq. (28)) in VA
f format, the
linear electrostatic force co-efficient and Q factor can be
extracted. Linear electrostatic force co-efficient (r1

3:67e4 mN/(mm/V2)) is in good agreement with the
Boundary Element simulations value of r1
3:65e
4 mN/(mm/V2). The quality factor Q (Q 671) matches
with the result obtained from small signal harmonic frequency response (Q 640).
Fig. 13 is the response of the device when excited with
frequencies corresponding to the first parametric resonance
at VA 20 V. The frequency is swept from low to high (up)
and high to low (down). Different step sizes are used in the
sweeping. When sweeping frequency down, the jump from
small amplitude to large amplitude response occurs around
52.7 kHz. This corresponds to the moving quasi-statically
from region III to region II in Fig. 6(b). When sweeping the
frequency up, the jump (from large amplitude response to
zero response) happens at about 53.7 kHz. This corresponds
to moving from region II to region III in Fig. 6(b). In region
III, there are two possible stable states of response of the
system. Initial conditions determine which one of the two
will be preferred. This jumping point depends on the step
size of frequency. The smaller the step size, the larger is the
jumping frequency (points S and R). Changing the frequency quasi-statically introduces a perturbation in the
system, which alters the initial state of the system. We

147

Fig. 13. Experimental data of frequencyamplitude curves as the input

frequency is swept quasi-statically near first-order parametric resonance
(twice the natural frequency and b 1) at VA 20 V. The figure can be
divided into three areas, I, II and III (refer Fig. 6 for the distinction). Point
P is in the right transition curve (b 1 d) from area III to II and Q is in
the left transition curve (b 1
d) from area II to I. Note there are two
experimental responses in area III, corresponding to the two stable
solutions. The points R and S are in region III where the response jumps
from the large amplitude stable response to the trivial solution.

theorize that a larger frequency jump size perturbs the

system enough to drive it out of the basin of attraction of
the large response. Further discussion of this phenomenon is
presented in Section 5.

5. Discussion
According to the analysis of nonlinear Mathieu equation,
first-order parametric resonance as represented in the b
d
plane (which can be translated to VA
f coordinates) may
be categorized into three areas, I, II and III (see Fig. 6(a)).
The corresponding A
B plane is shown in the same figure.
In area I, only one trivial solution exists. As the frequency is
changed quasi-statically, a bifurcation occurs at the left
boundary (b 1
d). In area II, the trivial solution
becomes unstable and simultaneously a stable sub-harmonic
motion is born. This motion grows in amplitude as d
increases. In the right boundary (b 1 d), the unstable
trivial solution becomes stable again and an unstable subharmonic motion is born. The stable sub-harmonic nontrivial solution born in area II also exists in area III.
The experimentally obtained displacements exactly verify
all the characteristics expected from the analysis. Fig. 13
presents typical frequency response in the three zones
described above. In area I, the response is very small, which
corresponds to the stable trivial solution. In area II, a large
response exists, corresponding to the stable non-trivial
solution, while the unstable trivial solution cannot be found.
In area III, one large and one small response can be found
depending on the testing conditions. Analysis predicts that,
depending on the initial displacement and velocity, the final

148

W. Zhang et al. / Sensors and Actuators A 102 (2002) 139150

displacement converges to one of the two stable states.

When sweeping frequency down, the initial value is small
in area III and hence the trivial stable solution is obtained.
But when sweeping the frequency up, in small step sizes, the
response stabilizes at the non-trivial stable solution. For
large step size sweeps, the response jumps to the trivial
stable solution at a lower frequency. In Fig. 13, Point R
shows this jump. It has been experimentally observed that, if
the step size of the frequency sweep is made very small, the
oscillator can be maintained in the large amplitude response
to higher frequency (see RS in Fig. 13). From the theoretical
analysis, it can be seen that the two stable solutions in region
III have their own basins of attraction. This implies that the
final state of the oscillator in this region depends on the
initial condition of the oscillators. We theorize that the step
size causes a perturbation in the initial conditions, thereby
effecting the frequency at which the jump occurs. It is
worthwhile to note here that, the difference in response
between sweeping input frequency in either direction (up or
down) is due to the bifurcations occurring at the boundaries.
The significant impact of this analysis and its experimental verification on the mass sensor is as follows. In order to
be able to utilize the jump (sharp transition from trivial to
non-trivial solution in b
d plane), for the case of negative
effective nonlinear parameter g3eff, we can use only one of
the instability boundaries (b 1 d). The analysis predicts
that the bifurcations will occur in the other instability
boundary (b 1
d) where the system has a positive nonlinear parameter. In this paper, we limit the discussion to the
case of negative effective nonlinear parameter g3eff. Fig. 14
shows a schematic of how the position of the instability
boundary will change with change in mass. The two traces
are simulations of the sharp boundary between a very small
amplitude (trivial) response and a well-defined large amplitude response (Fig. 15), mapped for the reference oscillator
and a sense oscillator with a mass change of 10
14 kg.

Fig. 14. Schematic of the effect of change in mass on the transition

boundary between regions II and III. This is a simulation of the effect for a
mass change of 10
14 kg.

Fig. 15. Detail of the sharp transition from trivial solution to the nontrivial stable high amplitude solution as the drive frequency is changed
quasi-statically from region III to region II.

It should be noted that, by varying the amplitude of the

driving voltage, the effective nonlinear parameter can be
tuned. The transition of this effective nonlinear parameter
from positive to negative will switch the occurrence of the
bifurcations from one side of the instability boundary to the
other. This switch will be sharp and will occur at a welldefined VA and can then be used to estimate the value of the
cubic electrostatic stiffness with a good degree of accuracy
(see Fig. 7). In the present experimental set up, this was not
done as the required input voltage to change the effective cubic
stiffness to be positive was estimated to be more than the break
down voltage of the device oxide electrical insulator layer.
The presence of the cubic stiffness also results in a distinct
advantage compared to that of the linear parametric mass
sensor. In the case of the linear oscillator with parametric
excitation, the transition boundaries separate regions of
trivial solution from regions of exponentially growing
sub-harmonic solutions. The presence of damping changes
the rate of the exponential growth and alters the shape of the
transition curve (from V to U), but does not limit the
amplitude growth. This implies that a linear model cannot
predict the experimentally observed amplitude. In any real
system, the amplitude is finite and limited by factors not
modeled, like the nonlinearity, or physical constraints. Thus,
the advantage of modeling the cubic nonlinearity is that, we
can estimate the final amplitude of the response inside the
tongue.
Due to the cubic nonlinearity, the range of operation of the
mass sensor is reduced, but the sensitivity of the sensor is not
compromised. The above analysis and experimental verification illuminates the effect of the various design parameters
on the sensor performance. This helps in developing good
design rules. The sharp transitions can also be used as a selfcalibrating tool to extract the system parameters with a good
degree of confidence.
Additionally, from the analysis and experimental results,
we note that the effect of the damping in the system is to shift
the instability tongue from one that looks like a V to one
that looks like a U (see Fig. 6). This implies that there is a
minimum input voltage above which the transition takes
place (also see Fig. 12). The damping does not affect the

W. Zhang et al. / Sensors and Actuators A 102 (2002) 139150

sensitivity of the mass sensor, but rather introduces a constraint in the input voltage signal amplitude, unlike the
sensors based on simple harmonic oscillator resonance
shifts. This is a very critical feature of the sensor since
often the quality factor cannot be controlled with precision
or made very large in test situations.
The response of the oscillator is not effected significantly
due to process variations. The depth of the features is a
parameter not easily controlled in the process conditions.
This does not affect the design since the depth is not a design
parameter and does not effect the natural frequency (the
mechanical stiffness and the mass has the same dependence
on the depth) to the first-order. We have published an
analytical study of the effect of parameter variation due
to process conditions on the frequency stability of lateral
resonant oscillators elsewhere [31]. The stress and stress
gradient effects are minimal in these deep etched single
crystal silicon resonators of the design dimensions used in
this work [32]. However, for implementations of such a mass
sensor in other material systems, the ideas of Mehner et al.
[33] could be used to model the stress and stress gradient
effects in conjunction with our modeling approach.

6. Conclusion and future directions

In this paper, we present the modeling, analysis, and
experimental verification of nonlinearity effects on an
auto-parametric amplification based mass sensor. The effect
of the cubic mechanical stiffness and electrostatic stiffness
changes the dynamic behavior of the oscillator response
when excited parametrically in parts of the parameter (frequencyamplitude of input voltage) space. This restricts the
regions of operation of the mass sensor, but does not alter the
sensitivity. The detailed modeling and analysis also serve as
tools for design of the mass sensor. The sharp transition,
which facilitates the high mass change detection, can also be
used to estimate the system parameters with good degree of
accuracy in the reference oscillator.
For use as a mass sensor, some more issues need to be
taken into consideration. Understanding of the various
sources of noises in the system is one main issue [34]. Also,
we are investigating methods to activate the surface of the
oscillator for selective reaction and perhaps increase the
surface area using porous silicon. The effect of temperature
on the frequency response of parametric resonance is
another critical issue. To increase the sensitivity, the design
and fabrication of a new oscillator with higher natural
frequency is under progress.
Acknowledgements
The authors would like to thank Scott Adams for the
device fabrication. This work is supported by NSF CAREER
award #0093994.

149

References
[1] L. Ruby, Applications of the Mathieu equation, Am. J. Phys. 64
(1996) 3944.
[2] J.A. Burns, More on pumping a swing, Am. J. Phys. 38 (1970) 920
921.
[3] J.W.S. Rayleigh, R.B. Lindsay, The Theory of Sound, Dover
Publications, New York, 1st American, 1945, pp.
[4] K.L. Turner, S.A. Miller, P.G. Hartwell, N.C. Macdonald, S.H.
Strogartz, S.G. Adams, Five parametric resonances in a microelectromechanical system, Nature 396 (1998) 149152.
[5] K.L. Turner, W. Zhang, Design and analysis of a dynamic MEM
chemical sensor, in: Proceedings of the 2001 American Control
Conference, Arlington, VA, USA, 2527 June 2001, pp. 12141218.
[6] J.P. Raskin, A.R. Brown, B. Khuri-Yakub, G.M. Rebeiz, A novel
parametric-effect MEMS amplifier, J. Microelectromech. Syst. 9
(2000) 528537.
[7] A. Olkhovets, D.W. Carr, J.M. Parpia, H.G. Craighead, Nondegenerate nanomechanical parametric amplifier, in: Proceedings of
the 14th IEEE International Conference on Micro Electro Mechanical
Systems, Interlaken, Switzerland, 2125 January 2001, pp. 298
300.
[8] M.K. Baller, H.P. Lang, J. Fritz, C. Gerber, J.K. Gimzewski, U.
Drechsler, H. Rothuizen, M. Despont, P. Vettiger, F.M. Battiston, J.P.
Ramseyer, P. Fornaro, E. Meyer, H.J. Guntherodt, A cantilever arraybased artificial nose, in: Proceedings of the International Conference
on Scanning Probe Microscopy, Cantilever Sensors and Nanostructures, Seattle, WA, USA, 30 May1 June 1999, pp. 19.
[9] T. Bachels, R. Schafer, Microfabricated cantilever-based detector for
molecular beam experiments, Rev. Sci. Instrum. 69 (1998) 3794
3797.
[10] T.D. Stowe, K. Yasumura, T.W. Kenny, D. Botkin, K. Wago, D.
Rugar, Attonewton force detection using ultrathin silicon cantilevers,
Appl. Phys. Lett. 71 (1997) 288290.
[11] T. Kenny, Nanometer-scale force sensing with MEMS devices, IEEE
Sens. J. 1 (2001) 148157.
[12] P. Mohanty, D.A. Harrington, M.L. Roukes, Measurement of small
forces in micron-sized resonators, in: Proceedings of the 22nd
International Conference on Low Temperature Physics, Espoo,
Finland, 411 August 1999, pp. 21432144.
[13] M. Cartmell, Introduction to Linear, Parametric, and Nonlinear
Vibrations, Chapman & Hall, London, 1990, pp. xii, 242.
[14] K.L. Turner, P.G. Hartwell, F.M. Bertsch, N.C. Macdonald,
Parametric resonance in a microelectromechanical torsional oscillator, in: ASME International Mechanical Engineering Congress and
Exposition Proceedings of the Microelectromechanical Systems
(MEMS), Anaheim, CA, USA, 1520 November 1998, pp. 335340.
[15] R. Berger, C. Gerber, H.P. Lang, J.K. Gimzewski, Micromechanics: a
toolbox for femtoscale science: towards a laboratory on a tip, in:
Proceedings of the International Conference on Micro- and
Nanofabrication, Glasgow, UK, 2225 September 1996, pp. 373379.
[16] T. Thundat, P.I. Oden, R.J. Warmack, Microcantilever sensors,
Microscale Thermophys. Eng. 1 (1997) 185199.
[17] T. Thundat, R.J. Warmack, G.Y. Chen, D.P. Allison, Thermal and
ambient-induced deflections of scanning force microscope cantilevers, Appl. Phys. Lett. 64 (1994) 28942896.
[18] B. Vu Thien, N. Garcia, A.L. Levanyuk, A mechanical nanosensor in
the gigahertz range: where mechanics meets electronics, Surf. Sci.
301 (1994) L224228.
[19] R. Lifshitz, M.L. Roukes, Thermoelastic damping in micro- and
nanomechanical systems, Phys. Rev. B, Condens. Matter 61 (2000)
56005609.
[20] Z.J. Davis, G. Abadal, O. Kuhn, O. Hansen, F. Grey, A. Boisen,
Fabrication and characterization of nanoresonating devices for mass
detection, J. Vac. Sci. Technol. B, Microelectron. Nanometer Struct.
18 (2000) 612616.

150

W. Zhang et al. / Sensors and Actuators A 102 (2002) 139150

[21] N.C. Macdonald, SCREAM microelectromechanical systems, Microelectron. Eng. 32 (1996) 4973.
[22] S.G. Adams, F.M. Bertsch, K.A. Shaw, N.C. Macdonald, Independent tuning of linear and nonlinear stiffness coefficients (actuators),
J. Microelectromech. Syst. 7 (1998) 172180.
[23] ANSYS, Release 5.7.
[24] COULOMB, Three-Dimensional Electrostatic Field Solver, Version
5.1 (2000).
[25] R. Grimshaw, Nonlinear Ordinary Differential Equations, CRC
Press, Boca Raton, FL, 1993, pp. viii, 328.
[26] S. Evoy, D.W. Carr, L. Sekaric, A. Olkhovets, J.M. Parpia, H.G.
Craighead, Nanofabrication and electrostatic operation of singlecrystal silicon paddle oscillators, J. Appl. Phys. 86 (1999) 6072
6077.
[27] M. Aikele, K. Bauer, W. Ficker, F. Neubauer, U. Prechtel, J. Schalk,
H. Seidel, Resonant accelerometer with self-test, in: Proceedings of
the EUROSENSORS, 14th European Conference, Copenhagen,
Denmark, 2730 August 2000, pp. 161167.
[28] R.H. Rand, Lecture Notes on Nonlinear Vibrations, Version 34a,
Available online at http://www.tam.cornell.edu/randdocs/, 2000.
[29] K.L. Turner, Multi-dimensional MEMS motion characterization
using laser vibrometry, in: Transducers99 The 10th International
conference on solid-state Sensors and Actuators, Digest of Technical
Papers, Sendai, Japan, 710 June 1999, pp. 11441147.
[30] Polytec, Vibrometer Users Manual (1999).
[31] L. Rong, B. Paden, K. Turner, MEMS resonators that are robust to
process-induced feature width variations, in: Proceedings of the 2001
IEEE International Frequency Control Symposium and PDA
Exhibition, Seattle, WA, USA, 68 June 2001, pp. 556563.
[32] M.T.A. Saif, N.C. Macdonald, Planarity of large MEMS, J.
Microelectromech. Syst. 5 (1996) 7997.

[33] J.E. Mehner, L.D. Gabbay, S.D. Senturia, Computer-aided generation

of nonlinear reduced-order dynamic macromodels. Part II. Stressstiffened case, J. Microelectromech. Syst. 9 (2000) 270278.
[34] T.B. Gabrielson, Mechanical-thermal noise in micromachined
acoustic and vibration sensors, IEEE Trans. Electron Devices 40
(1993) 903909.

Biographies
Wenhua Zhang is currently a Ph.D. candidate in the department of
Mechanical and Environmental Engineering at University of California,
Santa Barbara. He received his master degree in plasma processing from
Institute of Mechanics, Chinese Academy of Sciences in 1999. His
research interests are in dynamics and characterization of MEMS.
Rajashree Baskaran is working towards her Ph.D. at University of
California, Santa Barbara. She got her Masters degree in the field of
theoretical and applied mechanics from Cornell University in 2000. Her
research interests are in the field of studying nonlinear dynamics of micro
electromechanical systems.
Kimberly Turner received her BS in mechanical engineering from
Michigan Technological University in 1994, and her Ph.D. from Cornell
University in 1999. She has been an Assistant Prof. of mechanical and
environmental engineering at UC Santa Barbara since July 1999. She
supervises seven graduate students, and three undergraduate researchers.
Dr. Turner is the recipient of a 2001 NSF CAREER award. She is a
member of many professional societies including ASME, IEEE, ASEE,
SEM, and AVS.