ACCEPTED BY THE JOURNAL OF MICROELECTROMECHANICAL SYSTEMS 1

Microfabricated Neuroaccelerometer: Integrating

Sensing and Reservoir Computing in MEMS
Bruno Barazani, Guillaume Dion, Jean-Franois Morissette, Louis Beaudoin, and Julien Sylvestre

Abstract—This study presents the design, fabrication, and test (such as quadrotor drones), with controllers generating control
of a micro accelerometer with intrinsic processing capabilities, signals that are complex functions of the system dynamics
arXiv:2003.10581v1 [cs.ET] 23 Mar 2020

that integrates the functions of sensing and computing in the same and of sensor data (provided by accelerometers or gyroscopes,
MEMS. The device consists of an inertial mass electrostatically
coupled to an oscillating beam through a gap of 8 µm. The for instance), as well as anomaly detectors in preventive
motion of the inertial mass modulates an AC electrostatic field maintenance systems, where specific vibration patterns must
that drives the beam in its non-linear regime. This non-linearity be identified in sensor data (provided by microphones or
is used to implement machine learning in the mechanical domain, accelerometers, for instance), often in the presence of colored
using reservoir computing with delayed feedback to process or non-stationary noise. This complexity results in controllers
the acceleration information provided by the inertial mass. The
device is microfabricated on a silicon-on-insulator substrate which are often much larger and less energy-efficient than the
using conventional MEMS processes. Dynamic characterization sensor devices in the control system.
showed good accelerometer functionalities, with an inertial mass In an attempt to develop more efficient control systems
sensitivity on the order of 100 mV/g from 250 to 1300 Hz through integration, we propose a new class of MEMS de-
and a natural frequency of 1.7 kHz. In order to test the vices, where both the sensory and the computing functions
device computing capabilities, two different machine learning
benchmarks were implemented, with the inputs fed to the device are implemented in the mechanical response of the same
as accelerations. The neuromorphic MEMS accelerometer was device. While sensory functions are implemented using fairly
able to accurately emulate non-linear autoregressive moving conventional MEMS designs, computing functions exploit the
average models and compute the parity of random bit streams. non-linear dynamics of a mechanical resonator in the MEMS,
These results were obtained in a test system with a non-trivial to implement a form of machine learning known as reservoir
transfer function, showing a robustness that is well-suited to
anticipated applications. computing (RC) [1]. The implementation of RC in new sub-
strates has been the target of several recent studies, which were
Index Terms—MEMS accelerometer, MEMS non-linearity, able to emulate RCs in different hardware platforms such as
recurrent neural networks, reservoir computing, neuromorphic
computing. memristors arrays [2], optical systems [3], [4], [5], mechanical
devices [6], and spintronic devices [7]. This is because new
unconventional computing architectures are expected to exceed
I. I NTRODUCTION the density and the energy efficiency of current technology.
Control systems are generally built from three classes of The approach used in this study enables complex computing
devices: sensors convert stimuli into electrical signals, which to be implemented in the MEMS in a trainable manner, by the
are processed by an electronic controller, which itself gener- repetitive presentation to the device of appropriate responses
ates control signals sent to actuators. MEMS technologies are to randomly selected sensory inputs. As the computing process
especially popular to implement sensors for different types is similar to data processing by a neural network, we call this
of stimuli, such as acceleration, pressure (including sound), new class of devices neuromorphic MEMS. We here demon-
spatial orientation and temperature. This popularity stems from strate experimentally a neuromorphic MEMS accelerometer,
low manufacturing costs, as well as from the sensitivity and or neuroaccelerometer for brevity, by training it to perform
energetic efficiency of the MEMS sensors. These benefits are two different machine learning benchmark tasks on signals
all related to the small physical dimensions of the mechanical that are applied as physical accelerations on the device. To the
components of the MEMS, and the resulting fast dynamics and best of our knowledge, this constitutes the first demonstration
low mechanical losses. MEMS sensors thus produce signals of a single physical device that is both a sensor and a
that are rich in information about the state of a system; (neuromorphic) computer.
these signals must be properly processed by a controller The design of the neuroaccelerometer is based on a conven-
device, which often must implement complex control strate- tional suspended proof mass, that is coupled electrostatically
gies. Familiar examples include underactuated robotic systems to a beam clamped at both ends (section II). The motion of the
proof mass, induced by accelerations applied on the device,
B. Barazani, G. Dion, J.-F. Morissette, L. Beaudoin, and J. Sylvestre modulates the amplitude of a pump signal driving the beam
are with Interdisciplinary Institute for Technological Innovation - 3IT at
University of Sherbrooke, 3000 boul. de l’Universit J1K 0A5 Sherbrooke near resonance, thus establishing the coupling between the
2020 IEEE. Personal use of this material is permitted. Permission from sensing and the computing portion of the device. The non-
IEEE must be obtained for all other uses, in any current or future media, linear dynamics of the beam are exploited to implement a
including reprinting/republishing this material for advertising or promotional
purposes, creating new collective works, for resale or redistribution to servers reservoir computer, using a scheme similar to reference [8],
or lists, or reuse of any copyrighted component of this work in other works. where inputs were applied directly as electrical signals on
ACCEPTED BY THE JOURNAL OF MICROELECTROMECHANICAL SYSTEMS 2

a similar beam, to demonstrate experimentally that MEMS to increase the computational power of simple dynamical
resonators could be used to perform complex computations, systems, at the cost of reduced processing speed.
including the classification of spoken words. While, theoret-
ically, networks of mechanically coupled MEMS resonators A. Suspended Inertial Mass
could address much more complicated computing tasks [9], we
The suspended inertial mass consists of a relatively large
initially focus on a single resonator for this first demonstration
central piece connected to a fixed substrate by compliant
of a neuromorphic MEMS sensor. As described in section
springs. The motion of this inertial mass under inertial forces
III, we have used standard microfabrication techniques to
(external accelerations) can be well approximated by a mass-
build a neuroaccelerometer. The mechanical characterization
spring-damper system. The static sensitivity is given by [13]
of the neuroaccelerometer is presented in section IV, where
fairly conventional sensing capabilities are demonstrated. In x m 1
= = 2, (1)
section V, the neuromorphic computing capabilities of the a k ω0
device are demonstrated, with machine learning benchmarks where x is the mass displacement, a is the external accelera-
realized on input data provided as accelerations acting on the tion, m is the mass, k is the elastic constant of the suspension,
device. The benchmarks include a first task (NARMA) with and ω0 is the system resonance frequency. Note that the sen-
requirements that are similar to those of a robotic controller, as sitivity can be increased by reducing the resonance frequency;
well as a second task (Parity) with requirements resembling this, however, also reduces the bandwidth of the sensor. Figure
those of a signal classification controller. Both benchmarks 1 shows the suspended inertial mass composed of a 590 x 410
demand significant non-linear processing and the ability to µm2 central rectangle attached to two symmetric T shaped
memorize data for some period of time. They are also realized structures. Both the rectangular and the T shaped parts are
on actual hardware, including a test platform with limited fully perforated by 10 x 10 µm2 holes to facilitate fabrication.
bandwidth and a response function that is not perfectly linear. The suspension consists of a pair of 2-stage folded accordion
The neuroaccelerometer was able to learn the benchmark springs that allows motion in a direction y. Each accordion
tasks, in spite of these non-ideal characteristics of the test spring possesses 4 longer members connected to the inertial
system, demonstrating at the same time its robustness and its mass and 4 shorter members connected to the anchors, which
usefulness as a device that can be easily adapted to real-world are fixed to the substrate. The elastic constant of the accordion
systems. spring, kAcc , can be estimated using
4Ewt3
II. D ESIGN kAcc = , (2)
L3l + L3s
A necessary property of physical RC is the ability to map where E is the silicon Young’s modulus, t is the thickness of
their input signals into a high-dimensional state, via non-linear the members, w is the width of the members, and Ll and
dynamics [10]. This mapping allows signals that are originally Ls are the lengths of the longer and the shorter members
not linearly separable to be represented in a space where they
can be processed by linear models. In this study, the non-linear
expansion of the input results from the dynamical response of
a clamped-clamped beam oscillating at large amplitudes [11],
[12]. We have shown previously [8] that this dynamical re-
sponse could be exploited to achieve significant neuromorphic
computational capabilities, in a very small and energy efficient
device. In this work, we leverage the mechanical nature of the
clamped-clamped beam computing system by coupling it to a
suspended proof mass that implements the sensing functions
of the neuromorphic MEMS. 5 µm
40 µm
The neuroaccelerometer thus comprises two principal me-
chanical elements: the non-linear oscillating beam, which has
a high natural frequency (section II-B); and a larger suspended
inertial mass with a much lower natural frequency, designed to
be sensitive to external accelerations (section II-A). When in
operation, a pump voltage applied to the inertial mass induces
an electrostatic force over the beam, driving it near resonance
with large displacements, in its non-linear regime. External
accelerations displace the inertial mass, thus modulating the
amplitude of the driving force over the beam and consequently
the beam oscillation amplitude. The displacement of the beam Fig. 1. Inertial mass suspended by a pair of accordion springs. The mass
is measured with piezoresistive strain gauges. The signal from only moves in the vertical direction and its maximum displacement is 5 µm,
the gap it forms with the bumpers. Blue color indicates the features that are
the gauges is digitized, delayed and fed back to the pump fixed to the substrate, such as anchors and electrical traces, while structures
voltage, in a scheme described in section V-A that is useful in red are free to move.
ACCEPTED BY THE JOURNAL OF MICROELECTROMECHANICAL SYSTEMS 3

[µm]
1.4
1.2
1.0
0.8
0.6 1.9 kHz 21.80 kHz 21. 81 kHz
0.4
0.2
0.0
z
y x F = 5 µN

21.81 kHz 21.91 kHz 41.3 kHz

Fig. 2. Surface plot of the total displacement for a force of 5 µN applied Fig. 3. Simulated vibration modes of the suspended inertial mass. The first
in the y direction. The inertial mass displaces approximately 1.5 µm in the 5 modes are in-plane and the last one is out-of-plane.
same direction.

B. Oscillating Beam
The displacement of the oscillating beam can be approxi-
mated by the Duffing equation:
respectively. Considering E = 125 GPa, t = 2 µm, w = 50
ω0
µm, Ll = 410 µm, and Ls = 366 µm, eq. 2 results ÿ + ẏ + ω02 y + βy 3 = F (t), (3)
in kAcc = 1.7 N/m, and since the two accordion springs Q
are associated in parallel, the suspension elastic constant is where y, t, and F are displacement, time, and external force
2kAcc = 3.4 N/m. The 5 µm gap between the bumper and the per unit mass, respectively. Dots indicate derivative with
inertial mass limits the suspension force to a maximum value respect to time. The beam properties, ω0 , Q, and β, correspond
of 17 µN, which corresponds to a maximum acceleration of to the angular natural frequency, the quality factor, and the
60g, assuming g = 9.8 m/s2 and a silicon density of 2328 cubic stiffness parameter, respectively. Note that the β term
kg/m3 . adds to the restoring spring force and introduces the non-
linearity to the equation. If β = 0, eq. 3 is reduced to the forced
Finite element analysis (FEA) was performed through the damped linear oscillator. In the case of a clamped-clamped
solid mechanics interface of COMSOL Multiphysics R [14] to beam (Fig. 4), the value of β can be approximated by [15]
further simulate the static behavior of the mechanical system.  4
E 2π
Figure 2 shows the surface plot of the mass total displacement β= , (4)
when a force of 5 µN is applied in the y direction. The ratio 18ρ l
of the applied force to the simulated displacement gives 3.4 where l is the beam length and ρ is its density. For the beam
N/m, in agreement with the calculated spring constant. The shown in Fig. 4, a reduced effective beam length can be used
application of the same force in the z direction results in in eq. 4 in order to model the influence of piezoresistives
an out-of-plane displacement of 3 nm, corresponding to an gauges on the mode shape of the beam. In this geometry, β
out-of-plane stiffness of 1820 N/m, approximately 535 times is positive, which leads to an increase of the beam stiffness
the in-plane stiffness. This low cross-sensitivity is caused by with displacement (hardening). Eq. 4 further indicates the
the relatively large device thickness of 50 µm used in this geometric nature of this non-linearity as it depends on the
design. The total displacement of the mass was simulated beam length. When oscillating at large amplitudes, short
while varying the force amplitude from 0.05 to 17 µN. The beams stretch significantly more than long beams, which leads
device showed a linear behavior, with a static sensitivity of to the introduction of a larger non-linear restoring force term
82 nm/g. All simulations showed negligible displacement in in eq. 3. A characteristic of Duffing oscillators is the abrupt
the x direction (< 1 nm), indicating sufficiently low rotation change of the oscillation amplitude for small shifts of force
compliance. amplitude or driving frequency near the oscillator natural
frequency.
In addition to the static analysis, the device dynamic re- In addition to the non-linearity, the natural frequency and
sponse was simulated (Fig. 3). The eigenfrequency analysis quality factor also influence the beam computing performance.
showed in-plane natural frequencies principally around 1.9 The combination of a high natural frequency (larger than 105
kHz and 22 kHz. The former corresponds to large displace- Hz) and relatively low quality factor (∼100) leads to a higher
πf0
ments of the inertial mass in the y direction. The latter processing speed τ −1 = N Q [8], where N is the number of
corresponds to different flexural vibration modes of the sus- virtual nodes (see section V-A). The optimization of the beam
pension springs, with negligible displacement of the inertial properties led to a beam length l of 300 µm and a thickness
mass. Furthermore, the simulation showed a first out-of-plane t of 3 µm. The beam width w of 50 µm was defined by the
vibration mode at 41 kHz. Therefore, parasitic motions are fabrication technology. Using eq. 4 with an effective beam
not expected to significantly affect low frequency acceleration length of 280 µm yields an expected β value of 1.0×1024
measurements. (Hz/m)2 , which fits our experimental data well and provides
ACCEPTED BY THE JOURNAL OF MICROELECTROMECHANICAL SYSTEMS 4

Beam
FE
d

Vd = V0cos( t)
y
x
Fig. 4. FEA model of the beam showing its first mode shape. The longitudinal
strain on the piezoresistive gages allows a differential transduction of beam Fig. 5. Schematics of the coupling between the inertial mass (red) and the
oscillations.
beam (gray). The motion of the mass modulates the force on the beam, which
oscillates near its natural frequency and in its non-linear regime. Piezoresistive
gauges (green) measure the variation of the beam oscillations amplitude.
rich enough computing dynamics for the desired application.
Simulations of the defined beam show a natural frequency
f0 = ω0 /(2π) of 484 kHz. The squeeze film effect is expected beam oscillations amplitude, albeit non-linearly. Therefore,
to dominate viscous damping since the gap between the proof the described device is sensitive to external accelerations and
mass and the beam is small compared to the beam width. The produces non-linear outputs.
beam quality factor is estimated using [16] Note that since the natural frequency of the suspended mass
is more than two orders of magnitude smaller than that of the
ρtd3 ω0
Q= , (5) beam, the inertial mass motion is not dynamically amplified
µw2 when the beam is driven at its resonance. Nevertheless, the
where µ is the dynamic viscosity of air and d is the gap mass equilibrium position is shifted due to the constant force
distance of 8 µm. For µ = 1.8x10−5 Pa·s, the quality factor term in eq. 6, εAV02 /(4d2 ). The static deflection of the beam
of the beam is 241. is smaller due to its higher stiffness.

C. Inertial Mass Coupled to the Oscillating Beam III. FABRICATION

The micro-fabricated device enables the electrostatic cou- The device was built on a (100) silicon on insulator (SOI)
pling of the suspended mass to the oscillating beam. The wafer. The SOI was p-doped so as to increase silicon gauge
beam is located next to the inertial mass, forming a gap d factor and thus the sensitivity of the piezoresistive gauges. The
of 8 µm with its T shaped structure, as can be seen in Fig. 5. device layer, of a thickness of 50 µm, and the handle layer,
An electrical signal is applied on the inertial mass through both with a resistivity of (0.015 ± 0.005) Ωcm, were separated
its anchors in order to generate an attractive electrostatic by a buried oxide layer (BOX) of 4 µm. The manufacturing
force on the beam. This driving signal can be written as flow consisted of a sequence of the following processes:
Vd = V0 cos(ωd t), where V0 is the voltage amplitude and ωd photolithography, silicon etching, liberation, and metallization.
is the angular driving frequency. The electrostatic force FE is Firstly, after dicing and proper cleaning, the substrate was
estimated using the infinite parallel plate approximation, with spin-coated with positive photoresist AZ 9245 at 3800 RPM
for 1 minute to obtain a 4.5 µm thick film on the top of
ε0 A 2 2 ε0 AV02 the device layer. After a soft bake of 30 minutes at 110◦ C,
FE ≃ V cos (ω d t) = [1 + cos(2ωd t)] , (6)
2d2 0 4d2 the photoresist was exposed through the photomask to a UV
where ε0 is the dielectric constant of air and A is the capacitor dose of 200 mJ/cm2 and then developed for approximately 5
effective area of 130 µm × 50 µm. From eq. 6, one concludes minutes. Next, the patterned dice underwent a deep reactive-
that the force is proportional to V02 /d2 and that the beam is ion etching step (DRIE), in which the device layer was etched
driven at twice the driving frequency. 50 µm down, at roughly 90◦ , reaching the BOX layer. Then,
In experiments, ωd is adjusted so that 2ωd is approximately in a procedure similar to [17], the dice was fixed to a silicon
the natural frequency of the beam, leading to large oscillation support wafer of 6 inches, which was flipped and placed
amplitudes, which are measured by piezoresistive gauges over a Teflon dish containing 50 ml of HF 49%. A 500 W
attached to the beam extremities, as shown in Fig. 5. V0 is lamp was used to heat the back of the wafer to evaporate
then fine tuned to reach the beam non-linear regime, which the HF to etch the oxide. The temperature at the wafer was
is detected by the shift in the resonance frequency due to the tuned to 40◦ C since lower temperatures may condense the HF
stiffening of the beam. External accelerations stimulate the and higher temperatures would reduce the etch rate. After 3
motion of the inertial mass, which opens or closes the gap etching periods of 4 minutes each, the inertial mass, beam,
d with the beam resulting in a change of the electrostatic and piezoresistive gauges were free to move, while wider
force. As a result, the external accelerations modulate the features (anchors and the electrical traces) were still connected
ACCEPTED BY THE JOURNAL OF MICROELECTROMECHANICAL SYSTEMS 5

1.2 V0 :

Oscillation amplitude signal (V)

138 75
131 69
1.0 125 62
119 56
0.8 112 50
106 44
0.6 100 38
94 31
88 25
0.4 81 19
0.2
0.0
experiment
1.2 numerical 700

Oscillation amplitude signal(V)

Oscillation amplitude (nm)

1.0 600
0.8 500
0.6 400
Fig. 6. SEM image of the MEMS neuromorphic accelerometer built on a 300
SOI wafer. The inset shows the cross section of an anchored feature. At the 0.4
end of fabrication, the wider features (anchors and electrical traces) remain 200
0.2
connected to the substrate by the sacrificial oxide that was not etched. 100
0.0
0
220 225 230 235 240 245 250 255 260
to the handle through the remaining oxide. Figure 6 shows the fd (kHz)
MEMS device after the liberation, the inset showing the cross
section of a feature fixed to the handle via the oxide. Etching Fig. 7. Measured oscillation amplitude signal (top) and comparison with the
oscillation amplitude of the numerical model (bottom) as a function of the
cycles of at most 4 minutes were necessary since longer driving frequency fd = ωd /2π for different values of V0 . As V0 increases,
etching periods resulted in stiction caused by the condensation the shift of the resonance frequency and appearance of jumps indicate the
of the HF vapor. onset of non-linearity. The inset shows the hysteresis at V0 = 138 V for a
sweep increasing (solid line) and decreasing (dashed line) in frequency.
Finally, a laser cut stainless steel hard mask was placed
over the dice so that only the electrical traces and the bonding
pads were exposed, i.e. not covered by the mask. Then, a 5 Sweeping the driving frequency ωd from low to high values
nm thick film of chrome and a 200 nm thick film of gold were and varying V0 produced the curves shown in Fig. 7. For low
evaporated on the device. The BOX kept the device and the values of V0 , the beam behaved as a harmonic oscillator with
handle layer (partially covered with gold) electrically isolated. a natural frequency of 482.2 kHz (twice ωd /2π = 241.1 kHz)
and a quality factor Q of 145 ± 10. The measured natural
IV. C HARACTERIZATION frequency is in good agreement with the simulated value of
A. Doubly Clamped Beam 484 kHz; however, the measured quality factor is significantly
The characterization of the beam dynamical response is smaller than the calculated value of 241. This indicates that
crucial since the beam is the source of non-linearity used by in addition to squeeze film damping, other loss mechanisms
the RC. In order to test the beam, a sinusoidal driving signal of were active in the system. As V0 was increased, the peaks
amplitude V0 was applied on the inertial mass, which was kept of the curves shifted to higher frequencies and, above the
still (no acceleration), functioning as a fixed drive electrode. critical value of V0 = 106 V, an abrupt change (jump) to
The same driving signal was applied on the handle layer in lower displacement amplitudes occurred at the peak frequency.
order to prevent electrostatic out-of-plane forces induced by a This behavior was associated with a hysteretic response. If the
charged substrate (these forces could lead to the pull-in of the driving signal was instead swept from high to low frequencies,
inertial mass onto the substrate). the curve jumped to a higher amplitude at a frequency that
Oscillations of the beam were measured via the piezoresis- was lower than the sweep up peak frequency (Fig. 7, inset).
tive strain gauges in a differential configuration. This allowed This is typical of a non-linear hardening Duffing oscillator
cancellation of the large electrical feed-through of the driving (β > 0). In Duffing oscillators, the steady-state solution, for
signal, which was mainly caused by parasitic capacitance forcing frequencies between the two jumps (up and down), is
present on the device. This capacitance resulted from the thin multivalued with two stable branches and an unstable one.
oxide layer that separated the polarized handle layer and the Figure 8 shows the beam oscillation amplitude signal as a
large wire-bonding pads used for the readout. The ∼1 mV function of V0 for different values of ωd . Similarly to Fig.
displacement signal stemming from the elongation of the strain 7, the beam displacement became more non-linear as the
gauges, which was one order of magnitude smaller than the driving frequency was increased (near ωd = ω0 /2) and for
feed-through signal, was amplified with a total gain of 760 and ωd /2π > 242 kHz an hysteretic jump shifted the curves to
bandpass-filtered around 2ωd (passband width of 80 kHz) to higher (lower) amplitudes for a sweep up (down). Note that
mitigate noise and attenuate any leftover feed-through at ωd . the higher the driving voltage frequency ωd , the larger was the
The oscillating signal was then demodulated by an envelope jump or change of the oscillation amplitude; and therefore, the
detector and sampled by a 16 bit analog-to-digital converter higher was the non-linearity. The square law behavior apparent
(ADC). in these curves, not described by the Duffing equation, is due
ACCEPTED BY THE JOURNAL OF MICROELECTROMECHANICAL SYSTEMS 6

1.0 fd : sweep up
Oscillation amplitude signal (V)
240 kHz sweep down
239 kHz 90
0.8 241 kHz 2.0
242 kHz
243 kHz Accele ation:
0.6 244 kHz

Sensitivity (V/g)
1.5 0.33g 0

Phase (°)
0.4 0.65g
0.98g
0.2 1.0 1.31g
1.62g (90
0.0 1.91g
800 0.5
1.0 experiment
Oscillation amplitude signal (V)

numerical

Oscillation amplitude (nm)

0.8 600 (180
0.0
0 250 500 750 1000 1250 1500 1750
0.6 Vibration f equency (Hz)
400
0.4
200 Fig. 9. Inertial mass sensitivity as a function of the vibration frequency
0.2 for different acceleration amplitudes. The peaks near 1.7 kHz indicate the
resonance frequency of the inertial mass.
0.0 0
20 40 60 80 100 120 140
V0 (V)
difference is likely due to the observed 10-15% reduction of
Fig. 8. Measured oscillation amplitude signal (top) and comparison with the the spring thickness caused by the occurrence of notches in
numerical model (bottom) as a function of the driving voltage amplitude V0 the end of the DRIE. While the features appearing around
for different values of fd . 250 Hz and 1300 Hz are not fully understood, possible causes
include vibration modes in the packaging of the MEMS device,
to the quadratic dependence of the force with respect to voltage rotational modes of the shaker and vibration or displacement
(eq. 6) and also contributed to the non-linearity of the system. of other structures present on the device (multiple neuroac-
Numerically solving equation 3 using equations 4, 5 and celerometers are patterned on a single chip for convenient
6 while accounting for the electrostatically induced change in testing). Despite this non-ideal sensor response, the device was
equilibrium position of the inertial mass for every new value of able to perform non-trivial computations on sensed inputs in
V0 (i.e. d = d(V0 )) produced the dashed curves in the bottom the neuroaccelerometer configuration (section V), which un-
panel of figures 7 and 8. derscores the robustness of such neuromorphic sensor systems.

B. Inertial Mass Response to Accelerations

V. P ROCESSING ACCELERATION S IGNALS WITH THE
The motion of the inertial mass was sensed by measuring N EUROACCELEROMETER
the change of the piezoresistive signal caused by the variation
of the gap distance d (see eq. 6). As discussed previously, the A. Methods
application of external acceleration in the y direction leads to Neural-like processing capabilities were conferred to the
the displacement of the inertial mass, which closes or open neuroaccelerometer by using the non-linear oscillating beam
the gap with the beam. as the single physical node in a delay-coupled RC [18]. The
Acceleration signals were applied on the device using system, schematized in Fig. 10, was essentially the same as
a PID-controlled electrodynamic shaker. The device was that of our previous MEMS RC [8], with the difference that
mounted near a calibrated integrated electronics piezo-electric voltage signal inputs on the fixed electrode of the MEMS RC
(IEPE) accelerometer used for the control loop. For this were replaced by acceleration signals sensed by the suspended
characterization, the beam was operated in its linear regime, inertial mass, which in turn modulated the electrostatic forcing
at ωd /2π = 241.1 kHz and V0 = 75 V, to obtain an accurate of the beam in a gap-closing configuration.
assessment of the inertial mass response. The piezoresistive In this scenario, the sinusoidal electrostatic pump was first
signal was processed as before to produce an electrical signal amplitude-modulated by a temporal mask m(t) of period τ ,
proportional to the beam oscillation amplitude. with mask values updated at a rate θ−1 . Each sample of the
Figure 9 shows the device sensitivity for different sinusoidal mask defined a virtual node of the reservoir, and the mask
acceleration amplitudes, sweeping the shaker vibration fre- values corresponded to input weights. By choosing θ . T ,
quency from 10 Hz to 2 kHz. The graph shows a plateau from where T = 2Q/ω0 = (96 ± 7) µs is the beam ring-down
250 Hz to 1.3 kHz, where the sensitivity was on the order of characteristic time, the nodes were kept from settling to a
100 mV/g, independent of the acceleration amplitude. The first steady-state value before the mask was updated. This led to an
vibration mode of the inertial mass occurred at (1706 ± 5) interconnection scheme where each virtual node was coupled
Hz, with a quality factor of 19 ± 2. The measured natural at least to one of its nearest neighbor: the state of a given node
frequency of the suspended mass showed good agreement with depended on the state of the previous node, a time θ in the
the simulated value of 1865 Hz, less than 10% error. This past. While this mask signal could take different forms, such as
ACCEPTED BY THE JOURNAL OF MICROELECTROMECHANICAL SYSTEMS 7

the weights for a different dimension of y(k), and y(k) is a

Normalize to [0,1] A column vector. The matrix W was computed offline, following
MEMS
MEMS
a training phase where, after discarding the initial transient,

Power ampli er
M reservoir states were accumulated in an M × (N + 1)

Controller

Steel matrix X such that each row contained the reservoir state x(k)
block
for k = 0, 1, .., M − 1, augmented with a constant bias term
(needed to reproduce signals with non-zero mean). Training
Calibrated IEPE the readout using a ridge regression,
accelerometer
−1
+
W = Y′ X⊺ (XX⊺ + Γ) , (9)
40 dB 20 dB ENV ADC
-
where Y′ is the matrix of desired outputs constructed in a
similar way to X, allowed to prevent from over-fitting by in-
troducing a regularization matrix Γ. Choosing Γ = γI, where
DAC delay
I is the (N +1)×(N +1) identity matrix, produced satisfactory
results. The regularization parameter γ was optimized every
time a new readout weight matrix was computed by choosing
Fig. 10. Experimental setup of the neuroaccelerometer. the value which maximized the performance of the RC in the
testing phase for the given task.
The electronic circuitry shown at the bottom of Fig. 10
different random distributions [3], [19], sinusoidal signals [20],
was implemented by a custom analog front end combined
multiple level digital signals [21] and even chaotic signals [22],
with various commercially available instruments. The main
we chose the most commonly used form of a random binary
obstacles to full integration of the control electronics are the
sequence for the sake of simplicity. In this case, the mask was
high voltage drive signal and the delay loop. The former can
composed of only two values and switched randomly between
be dealt with by miniaturizing the device: according to our
them after each interval θ. Values of 0.45 and 0.7 produced
model, reducing all geometrical dimensions of the oscillating
adequate results for both our neuroaccelerometer and fixed
beam by one order of magnitude (except the gap d, which
drive MEMS RC [8].
only needs to be scaled by a factor of 1/5) would allow drive
For both benchmarks presented below, the input samples
voltages below 10 V. The latter, which currently necessitates
(u(t) in Fig. 10) were scaled to an amplitude of 2g (gain
analog-to-digital conversion in order to appropriately delay the
A = 4g), then held for a duration τ and output via a DAC
feedback signal, could eventually be bypassed by coupling
to the shaker PID controller. Since τ was also the RC output
multiple oscillating beams in a scheme first described in [9],
update period and the number of virtual nodes was N = τ /θ
which does not require external delayed feedback.
with θ determined by T , the dynamical properties of the beam
(T ) and of the acceleration input (τ ) dictated the maximum
available number of virtual nodes. As harder tasks typically B. NARMA Benchmark
require a larger reservoir size N , they also demand a faster
The emulation of non-linear autoregressive moving average
reacting beam, i.e. lower Q and higher ω0 .
(NARMA) models is a widespread machine learning bench-
Memory of recent inputs, which is necessary for processing
mark task [23], [24], [25], [26], [3]. This benchmark is relevant
time series such as the acceleration signals used in sections
as such non-linear filters are often used for signal processing
V-B and V-C, was provided to the RC by a delayed feedback
and control applications. An important part of its appeal stems
loop. The latter added the previous virtual node states to the
from its (non-linear) memory requirements: its current output
current mask pattern such that each virtual node was driven by
is a non-linear combination of many past inputs and outputs.
a superposition of the masked pump signal, the inertial mass
By introducing a time-lag parameter n, a generalized version
physical displacement and its previous response, τ seconds
of its input-output relationship can be written as
earlier. The voltage applied on the inertial mass thus took the
form n−1
X
yn (k + 1) = 0.3yn (k) + 0.05yn(k) yn (k − i)
Vd (t) = V0 [m(t) + αx(t − τ ) + 1] cos (ωd t) , (7) i=0
(10)

where x(t) is the beam displacement envelope signal sampled + 1.5u(k)u(k − n + 1) + 0.1,
by the ADC and α is the feedback gain. The output of the where u(k) is typically an i.i.d. uniform random variable [23],
envelope detector was sampled at the end of each interval θ so [25], [27] over [0, 0.5] constituting the input sequence. For
that N samples were collected per period τ , yielding the vector this study, this random sequence was input, after the scaling
x(k) containing the reservoir state at timestep k = (t − t0 )/τ . procedure described in section V-A, on the electrodynamic
Finally, the RC output vector y(k) was constructed from linear shaker PID controller as a voltage signal, resulting in an
combinations of the virtual node states: acceleration signal u′ (k) that was a distorted version of the
original input, as shown in Fig. 11. Indeed, because of the non-
y(k) = Wx⊺ (k). (8)
ideal mechanical response of the shaker, the original uniform
W is a readout weight matrix with each row corresponding to distribution became gaussian-like and samples separated by
ACCEPTED BY THE JOURNAL OF MICROELECTROMECHANICAL SYSTEMS 8

0.50 uk uk
′
Target Prediction
0.3

n=3
Value

0.25 0.1

0.3

n=5
0.00
0 10 20 30 40 50 0.1
k

n = 20 n = 15 n = 10
0.3
1.0
Autocorrelation

0.1
0.5
0.3
0.0 0.1
−40 −30 −20 −10 0 10 20 30 40 0.3
Time lag
0.1
Fig. 11. The uniform samples u(k) used for the PID setpoint (top left) are 0 20 40 60 80 100
altered by the vibration system, resulting in a different distribution (top right) k
of samples u′ (k) which are used to compute the target (eq. 10). The bottom
panel shows the correlogram for the two signals.
Fig. 12. First 100 timesteps of the RC output (eq. 8) and of the target (eq.
10) in the testing phase for the NARMA task with n = 3, 5, 10, 15, 20.

less than 40 timesteps became correlated. In order to exempt 0.8

the RC from also inverting the shaker transfer function, which 0.04
0.6
NRMSE
would constitute a harder task, the actual acceleration sig-

RMSE
nal measured with the reference accelerometer was sampled, 0.03
0.4
rescaled and offset so that its amplitude and mean were similar 0.02
to those of u(k), then used instead of u(k) to compute the 0.2
target (i.e., u′ (k) was used instead of u(k) in eq. 10). The RC 4 6 8 10 12 14 16 18 20
n
was then trained (eq. 9) using the first M = 4000 samples
of u′ (k). For this task, each dimension of y′ contained the
Fig. 13. NRMSE and RMSE of the RC emulation of eq. 10 as a function of
expected NARMAn output for a different value of the time- the time-lag parameter n.
lag parameter n = 2, 3, .., 20:
 
NARMA2 (k)
 NARMA3 (k)  of optimization, the neuroaccelerometer did not reach the
y′ (k) = 

.

(11) performance of some noiseless software RC (NRMSE values
..
 .  close to 0.1 [28], [23] for n = 10, compared to a value of
NARMA20 (k) 0.5 for the neuroaccelerometer), but performed similarly to
other hardware [3] and software [26] RC. Noise inherent to
Following the training phase, the RC was tested by collecting
physical systems hinders their memory capacity [29], making
the next 400 samples of the RC output y(k) and comparing it
processing time series with long memory requirements more
to y′ (k) via the normalized root mean squared error for each
challenging.
dimension:
v h i
u E (y′ (k) − yn (k))2
u
t n
C. Parity Benchmark
NRMSEn = , (12) The parity benchmark allows a straightforward comparison
σ 2 (y′ n (k))
of the neuroaccelerometer computing capabilities with other
where E denotes the expected value over k = systems [8], [9], [30], [31], [32], [33], [34]. It is a non-linear
4000, 4001, .., 4399 and σ 2 (y′ (k)) is the variance of the pattern classification task requiring a memory of past inputs
target. This metric was then used as a criterion for adjusting which makes it well suited for RC evaluation, as a prototype
the RC hyperparameters, yielding optimized values of for non-trivial signal recognition tasks. A random binary
(V0 , fd , α, θ, γ) = (135 V, 245 kHz, 1.2, 50 µs, 10−3 V2 ). stream u(t) was fed to the shaker controller as a voltage signal,
Figure 12 shows the RC output overlaid on the target resulting in a zero-mean, 2g amplitude acceleration waveform
waveform for n = 3, 5, 10, 15, 20. The reproduction of the input to the MEMS. Since the electrodynamic shaker has a
target became worse as the memory requirement of the task maximum travel distance of 13 mm, long sequences of suc-
increased between n = 3 and n = 20, as shown in cessive identical values in the input would have been heavily
Fig. 13 where the testing error is shown against the time- distorted. In order to mitigate this, input bits were flipped when
lag parameter. The root mean squared error (RMSE) did the input stream exceeded 3 successive bits without switching.
not exactly follow the NRMSE due to different variances Figure 14 shows that the acceleration signal was still distorted
for the different dimensions of y′ (k). At its current state with respect to the controller setpoint due to the non-ideal
ACCEPTED BY THE JOURNAL OF MICROELECTROMECHANICAL SYSTEMS 9

the performance metric, which was computed by thresholding

Shaker setpoint Pump envelope
Acceleration signal Beam oscillation envelope the RC output at 0 and comparing the sign of the resulting
bit stream with that of the target. This last step produced the
Signal (a.u.)

success rates shown beside their respective waveforms in Fig.

15, with the half-width of the 95% confidence Agresti-Coull
interval [35] used as the uncertainty. Hyperparameters for Fig.
15 were (V0 , fd , α, θ, γ) = (135 V, 245 kHz, 0.7, 50 µs, 0.005
V2 ). The RC output for n = 1, which is the identity function of
0.1 1 10 the PID setpoint u(t), showed that the neuroaccelerometer was
(t − t0)/τ able to accurately invert the non-linear response of the shaker.
As was the case for the NARMA benchmark, the time traces
Fig. 14. Typical behavior of the different signals involved in computing the became more noisy when the memory requirement of the task
parity function with the RC, showing different timescales on a logarithmic was increased from n = 1 to n = 6. While thresholding
scale.
them yielded lower success rates than it did for our fixed drive
MEMS RC for n ≥ 3 [8], the latter had direct access to the
undistorted input as a voltage signal, making the task easier.
Input

n=1 (99.5±0.2)%
VI. C ONCLUSION
n=2 (100.0±0.1)%
This paper demonstrated the design, fabrication, and val-
idation of a MEMS accelerometer with built-in computing
y(k)

n=3 (97.4±0.5)%
capabilities. The concept of exploiting the non-linearity of a
n=4 (82±1)% clamped-clamped micro beam to emulate a reservoir computer
had been previously simulated [9] and experimentally demon-
n=5 (70±1)%
strated [8]. In this study, an accelerometer was coupled to the
n=6 (63±2)%
computing beam in order to integrate sensing and computing in
a single MEMS. This is the first experimental demonstration
0 20 40 60 80 100

k of such a highly integrated device, which represents a new

class of MEMS devices. Through integration, these devices
Fig. 15. Following excitation of the RC with the input bit stream (black), have the potential to be much smaller, faster, and more
the RC output waveforms for the P1 to P6 tasks (blue) are thresholded and energy efficient than conventional combined control systems
compared to the target waveforms (red) to produce the success rates shown comprising sensors and separate electronic controller units.
on the right.
In addition, the computing functions were implemented via a
neuromorphic system that has many of the beneficial features
of machine learning algorithms implemented in software.
shaker response, especially for signals with a large bandwidth The same neuromorphic MEMS can be trained to implement
such as the step-wise input for the parity task. Showing the many different data processing tasks, such as classification
signals on a logarithmic scale allows the observation of the (as demonstrated with the parity benchmark) and the imple-
different timescales involved in the system. On the shorter mentation of complex non-linear functions (as demonstrated
timescales of order T ∼ θ = 50 µs, the electrostatic pump with the NARMA benchmark). This training characteristic
envelope alternated between the mask values, and the beam could be useful to simplify the design of control systems,
oscillation amplitude was a non-linear function of this signal. to increase the robustness of applications, and to facilitate
The right half of Fig. 14 shows that the beam response was also the adaptation of systems to changing environments (e.g.
modulated by the slowly evolving acceleration signal (through using continuous unsupervised training). Machine learning
the inertial mass displacement, not shown), of characteristic systems, including RC, have been shown to frequently offer
time τ = N θ = 5 ms. powerful generalization capabilities, with errors on validation
Following this excitation pattern, the neuroaccelerometer data that are not much larger than errors on training data.
had to compute the parity (without delay) of order n = 1 Such generalization capabilities could be especially useful to
to n = 6: increase the robustness of control systems.
 
P1 (k) Neuromorphic MEMS could be used for applications with
P2 (k) n−1
Y strict constraints on volume, weight, response time or energy
y′ (k) =  .  , where Pn (k) = u (k − i) , (13)
 
 ..  consumption. These include many autonomous and robotic
i=0 applications, as well as mobile and wearable devices. Neu-
P6 (k)
romorphic MEMS could also be especially useful in dis-
where the input u(k) was a unit-amplitude version of the signal tributed sensor networks (e.g. the Internet of Things), to limit
sent to the PID controller. As for the NARMA benchmark, the the amount of data transmitted to a central processing unit
training sequence lasted 4000 samples, but the testing phase by providing significant computing power “at the edge”, to
was longer at 2000 samples in order to be more precise on transmit only data corresponding to specific patterns. The
ACCEPTED BY THE JOURNAL OF MICROELECTROMECHANICAL SYSTEMS 10

technological relevance of neuromorphic MEMS will likely [15] H. W. C. Postma, I. Kozinsky, A. Husain, and M. L. Roukes, “Dynamic
be enhanced by future work that could focus on integrating range of nanotube- and nanowire-based electromechanical systems,”
Appl. Phys. Lett., vol. 86, no. 22, p. 223105, May 2005. [Online].
neuromorphic computing capabilities to other MEMS sensors, Available: http://aip.scitation.org/doi/10.1063/1.1929098
and on increasing the computational power of the MEMS [16] M. Bao, Analysis and design principles of MEMS devices, 1st ed.
RC. Increasing the number of physical nodes in the reservoir Amsterdam: Elsevier, 2005, oCLC: 254583926.
[17] Y. Fukuta, H. Fujita, and H. Toshiyoshi, “Vapor Hydrofluoric
network (multiple, coupled resonant non-linear structures) Acid Sacrificial Release Technique for Micro Electro Mechanical
could increase the device processing speed, leading to hybrid Systems Using Labware,” Jpn. J. Appl. Phys., vol. 42, no.
networks formed by physical nodes and virtual nodes (through Part 1, No. 6A, pp. 3690–3694, Jun. 2003. [Online]. Available:
http://stacks.iop.org/1347-4065/42/3690
multiplexing in time). Such MEMS could have simpler drive [18] L. Appeltant, M. Soriano, G. Van der Sande, J. Danckaert, S. Massar,
electronics, to eventually reach the full technological benefits J. Dambre, B. Schrauwen, C. Mirasso, and I. Fischer, “Information
of neuromorphic MEMS sensors. processing using a single dynamical node as complex system,” Nature
Communications, vol. 2, p. 468, Sep. 2011. [Online]. Available:
http://www.nature.com/doifinder/10.1038/ncomms1476
ACKNOWLEDGMENT [19] Y. Kuriki, J. Nakayama, K. Takano, and A. Uchida,
“Impact of input mask signals on delay-based photonic
The authors thank the Ministre de lconomie, de la science reservoir computing with semiconductor lasers,” Opt. Express,
et de linnovation du Qubec and the Natural Sciences and vol. 26, no. 5, p. 5777, Mar. 2018. [Online]. Available:
https://www.osapublishing.org/abstract.cfm?URI=oe-26-5-5777
Engineering Research Council of Canada for financial support. [20] F. Duport, A. Akrout, A. Smerieri, M. Haelterman, and S. Massar,
“Analog input layer for optical reservoir computers,” p. 17.
[21] M. C. Soriano, S. Ortn, D. Brunner, L. Larger, C. R.
R EFERENCES Mirasso, I. Fischer, and L. Pesquera, “Optoelectronic reservoir
computing: tackling noise-induced performance degradation,” Opt.
[1] H. Jaeger and H. Haas, “Harnessing nonlinearity: Predicting chaotic
Express, vol. 21, no. 1, p. 12, Jan. 2013. [Online]. Available:
systems and saving energy in wireless communication,” Science, vol.
https://www.osapublishing.org/oe/abstract.cfm?uri=oe-21-1-12
304, no. 5667, pp. 78–80, 2004.
[2] C. Du, F. Cai, M. A. Zidan, W. Ma, S. H. Lee, and W. D. Lu, “Reservoir [22] J. Nakayama, K. Kanno, and A. Uchida, “Laser dynamical reservoir
computing using dynamic memristors for temporal information computing with consistency: an approach of a chaos mask signal,”
processing,” Nat Commun, vol. 8, no. 1, p. 2204, Dec. 2017. [Online]. Optics Express, vol. 24, no. 8, p. 8679, Apr. 2016. [Online]. Available:
Available: http://www.nature.com/articles/s41467-017-02337-y https://www.osapublishing.org/abstract.cfm?URI=oe-24-8-8679
[3] Y. Paquot, F. Duport, A. Smerieri, J. Dambre, B. Schrauwen, [23] H. Jaeger, “Adaptive Nonlinear System Identification with Echo State
M. Haelterman, and S. Massar, “Optoelectronic Reservoir Computing,” Networks,” p. 8.
Scientific Reports, vol. 2, no. 1, Dec. 2012. [Online]. Available: [24] J. Steil, “Backpropagation-decorrelation: online recurrent learning
http://www.nature.com/articles/srep00287 with O(N) complexity,” in 2004 IEEE International Joint
[4] L. Larger, M. C. Soriano, D. Brunner, L. Appeltant, Conference on Neural Networks (IEEE Cat. No.04CH37541), vol. 2.
J. M. Gutierrez, L. Pesquera, C. R. Mirasso, and Budapest, Hungary: IEEE, 2004, pp. 843–848. [Online]. Available:
I. Fischer, “Photonic information processing beyond Turing: http://ieeexplore.ieee.org/document/1380039/
an optoelectronic implementation of reservoir computing,” Opt. [25] A. Atiya and A. Parlos, “New results on recurrent network training:
Express, vol. 20, no. 3, p. 3241, Jan. 2012. [Online]. Available: unifying the algorithms and accelerating convergence,” IEEE Trans.
https://www.osapublishing.org/oe/abstract.cfm?uri=oe-20-3-3241 Neural Netw., vol. 11, no. 3, pp. 697–709, May 2000. [Online].
[5] K. Vandoorne, P. Mechet, T. Van Vaerenbergh, M. Fiers, G. Morthier, Available: http://ieeexplore.ieee.org/document/846741/
D. Verstraeten, B. Schrauwen, J. Dambre, and P. Bienstman, [26] D. Verstraeten, B. Schrauwen, M. DHaene, and D. Stroobandt, “An
“Experimental demonstration of reservoir computing on a silicon experimental unification of reservoir computing methods,” Neural
photonics chip,” Nat Commun, vol. 5, no. 1, p. 3541, May 2014. Networks, vol. 20, no. 3, pp. 391–403, Apr. 2007. [Online]. Available:
[Online]. Available: http://www.nature.com/articles/ncomms4541 https://linkinghub.elsevier.com/retrieve/pii/S089360800700038X
[6] J. Degrave, K. Caluwaerts, J. Dambre, and F. Wyffels, “Developing an [27] T. Kubota, K. Nakajima, and H. Takahashi, “Dynamical Anatomy of
embodied gait on a compliant quadrupedal robot,” in 2015 IEEE/RSJ NARMA10 Benchmark Task,” arXiv:1906.04608 [cs, stat], Jun. 2019,
International Conference on Intelligent Robots and Systems (IROS). arXiv: 1906.04608. [Online]. Available: http://arxiv.org/abs/1906.04608
Hamburg, Germany: IEEE, Sep. 2015, pp. 4486–4491. [Online]. [28] A. A. Ferreira, T. B. Ludermir, and R. R. de Aquino, “An approach to
Available: http://ieeexplore.ieee.org/document/7354014/ reservoir computing design and training,” Expert Systems with Applica-
[7] J. Torrejon, M. Riou, F. A. Araujo, S. Tsunegi, G. Khalsa, D. Querlioz, tions, vol. 40, no. 10, pp. 4172–4182, Aug. 2013. [Online]. Available:
P. Bortolotti, V. Cros, K. Yakushiji, A. Fukushima, H. Kubota, S. Yuasa, https://linkinghub.elsevier.com/retrieve/pii/S0957417413000420
M. D. Stiles, and J. Grollier, “Neuromorphic computing with nanoscale [29] M. C. Soriano, S. Ortn, L. Keuninckx, L. Appeltant, J. Danckaert,
spintronic oscillators,” Nature, vol. 547, no. 7664, pp. 428–431, Jul. L. Pesquera, and G. Van der Sande, “Delay-based reservoir computing:
2017. [Online]. Available: http://www.nature.com/articles/nature23011 noise effects in a combined analog and digital implementation,” IEEE
[8] G. Dion, S. Mejaouri, and J. Sylvestre, “Reservoir computing with transactions on neural networks and learning systems, vol. 26, no. 2,
a single delay-coupled non-linear mechanical oscillator,” Journal of pp. 388–393, 2015.
Applied Physics, vol. 124, no. 15, p. 152132, Oct. 2018. [Online]. [30] P. Phua and D. Ming, “Parallel nonlinear optimization techniques
Available: http://aip.scitation.org/doi/10.1063/1.5038038 for training neural networks,” IEEE Trans. Neural Netw.,
[9] J. C. Coulombe, M. C. A. York, and J. Sylvestre, “Computing vol. 14, no. 6, pp. 1460–1468, Nov. 2003. [Online]. Available:
with networks of nonlinear mechanical oscillators,” PLOS ONE, http://ieeexplore.ieee.org/document/1257409/
vol. 12, no. 6, p. e0178663, Jun. 2017. [Online]. Available: [31] S. P. Adhikari, Changju Yang, Hyongsuk Kim, and L. O. Chua,
http://dx.plos.org/10.1371/journal.pone.0178663 “Memristor Bridge Synapse-Based Neural Network and Its Learning,”
[10] M. Lukoeviius and H. Jaeger, “Reservoir computing approaches to IEEE Transactions on Neural Networks and Learning Systems,
recurrent neural network training,” COMPUTER SCIENCE REVIEW, vol. 23, no. 9, pp. 1426–1435, Sep. 2012. [Online]. Available:
p. 23, 2009. http://ieeexplore.ieee.org/document/6232461/
[11] K. L. Ekinci and M. L. Roukes, “Nanoelectromechanical systems,” Rev. [32] N. Bertschinger and T. Natschlger, “Real-Time Computation at the
Sci. Instrum., p. 12. Edge of Chaos in Recurrent Neural Networks,” Neural Computation,
[12] Harrie A.C. Tilmans, Miko Elwenspoek, and Jan H.J. Fluitman, “Micro vol. 16, no. 7, pp. 1413–1436, Jul. 2004. [Online]. Available:
resonant force gauges,” vol. 30, no. 1, pp. 35–53, 1992. http://www.mitpressjournals.org/doi/10.1162/089976604323057443
[13] N. Yazdi, F. Ayazi, and K. Najafi, “Micromachined Inertial Sensors,” [33] K. Nakajima, T. Li, H. Hauser, and R. Pfeifer, “Exploiting short-term
PROCEEDINGS OF THE IEEE, vol. 86, no. 8, p. 20, 1998. memory in soft body dynamics as a computational resource,” J. R.
[14] “COMSOL Multiphysics Programming Reference Manual.” [Online]. Soc. Interface, vol. 11, no. 100, p. 20140437, Nov. 2014. [Online].
Available: www.comsol.com Available: https://royalsocietypublishing.org/doi/10.1098/rsif.2014.0437
ACCEPTED BY THE JOURNAL OF MICROELECTROMECHANICAL SYSTEMS 11

[34] J. Schumacher, H. Toutounji, and G. Pipa, “An Introduction to

Delay-Coupled Reservoir Computing,” in Artificial Neural Networks,
P. Koprinkova-Hristova, V. Mladenov, and N. K. Kasabov, Eds. Cham:
Springer International Publishing, 2015, vol. 4, pp. 63–90. [Online].
Available: http://link.springer.com/10.1007/978-3-319-09903-3 4
[35] A. Agresti and B. A. Coull, “Approximate Is Better than ”Exact”
for Interval Estimation of Binomial Proportions,” The American
Statistician, vol. 52, no. 2, p. 119, May 1998. [Online]. Available:
https://www.jstor.org/stable/2685469?origin=crossref