366 JEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 31, NO.

3, MARCH 1996

urf ace Micromachined Accelerometers

Bernhard E. Boser and Roger T. Howe, Fellow, IEEE

Abstract- Surface micromachining has enabled the cofabri- This paper concentrates on polysilicon surface microstruc-
cation of thin-film micromechanical structuries and CMOS or tures integrated with CMOS and Bipolar/MOS technologies.
BipolarMOS integrated circuits. Using linear, single-axis ac- These technologies have been used for various accelerometer
celerometers as a motivating example, this paper discusses the
and resonant gyroscopes [5]-[7], as well as micro-resonator
fundamental mechanical as well as the electronic noise floors
for representative capacitive position-sensing interface circuits.oscillators [8]. Other applications are being explored, such as
Operation in vacuum lowers the Brownian noise of a polysiticon micromechanical IF filters for signal processing and spectrum
accelerometer to below 1 pg/&. For improved sensor perfor- analysis [9], [lo]. The prospects for multisensing IC's using
mance, the position of the microstructure should be controlled several sensing elements or possibly arrays are particularly
using electrostatic force-feedback.Both analog and digital closed-interesting.
loop accelerometers are described and contrasted, with the latter
using high-frequency voltage pulses to apply force quanta to the
Accelerometers serve as a vehicle in this paper to investigate
microstructure and achieve a very linear resplonse. the various aspects of surface micromachined sensors. These
devices have evolved quite far, with commercial parts being
available for some time now [ll], [12]. First generation
I. INTRODUCTION devices achieve a noise floor around 10 m g / a over an
VER the past decade, surface micromachining has be- input range of f 5 0 g or more and with shock survival in excess
come established as a versatile solution for a wide of 2000 g. These specifications are compatible with automotive
variety of sensing problems. Only a few additional processing applications such as airbag release. Second generation devices
steps compatible with standard fabrication techniques and ma- achieve an order of magnitude better resolution and are well
terials are required to cofabricate mechanical sensing elements suited for a very large range of needs, including active car
and the associated electronic interface circuits on a single suspension, shock detection and monitoring, computer input
die. 'Surface micromachined sensors are used, for example, devices, toys, and short term navigation. Substantial future
in the automotive market as crash detectors and for dynamic improvements down to a noise level as low as 1 p g / a
vehicle control. Applications as vibration and shock detectors can be expected based on analysis of the fundamental limits
range from monitoring mechanical stress in airplane wings to of this technology. This level of performance is adequate for
recording mechanical shock of fragile shipping goods. The all but the most demanding inertial navigation applications of
technology is expected to have an even greater impact in acceleration sensors.
prospective applications such as head-mounted displays, where We begin with a brief analysis of sensing elements for
the small size and weight, combined with sophisticated on-chip linear acceleration and their implications for the measurement
signal processing capability, are enabling features. The need in system performance. An overview of the several strategies
this application for several different sensors, including linear for cofabrication of polysilicon microstructures and CMOS
and angular accelerometers and gyroscopes, in a very small provides the context for a description of a modular approach
volume, illustrates the advantages of monolithic fabrication of developed at Berkeley. The paper then discusses the electronic
the sensing elements and associated electronics. interface of the sensing element. Special attention is given to
Mechanical structures fabricated in surface micromachining factors limiting the achievable resolution of micromachined
technologies consist of deposited thin films of polysilicon [ 13, sensors. Particularly, the small size and consequent low mass
aluminum [2], silicon nitride 131, [4], and other materials. of the sensing element results in an elevated thermal noise
Integration of surface microstructures with MOS electronics floor. It will be shown how this limitation can be overcome
is relatively straightforward and economically attractive, with by vacuum packaging and embedding the sensing element in
the microstructure typically occupying only a small fraction of an electronic force-feedback loop, resulting in a performance
the die area. By bringing the sense element onto the integrated corresponding to a sensing element that is several orders of
circuit, surface micromachining leverages the experience and magnitudes larger. The various aspects will be demonstrated
sophisticated processes of IC manufacturing and brings about with examples of commercial accelerometer IC's [6], [12],
all the customary advantages of IC solutions: batch fabrication, [13] and research prototypes [5], [7]. A detailed description
high yield, small size, low power, and low cost. Elimination of an experimental device appears also in an earlier issue of
of a separate sense element miniaturizes the package and also this journal [14].
results in improved reliability.
11. SENSWGELEMENTS
Manuscript received October 25, 1995; revised December 19, 1995. This
work was supported by ARPA and the California PATH program. For sensing physical quantities such as acceleration, angular
The authors are with the Electrical Engineering and Computer Science
'Department, University of California Berkeley, Berkeley, CA 94720-1770
rate, or pressure, a mechanical sense element converts the
USA. unknown quantity into a displacement that is then detected
Publisher Item Identifier S 0018-9200(96)02451-1. and converted to an electrical signal. A conceptual diagram of
0018-9200/96$05 .00 0 1996 IEEE

Authorized licensed use limited to: National Science Council. Downloaded on July 6, 2009 at 12:34 from IEEE Xplore. Restrictions apply.
BOSER AND H O W SURFACE MICROMACHINEDACCELEROMETERS 367

displacement
X A proof-mass
F=ma

XQ F=kx

Fig. 2. Cantilever beam with concentrated load.

reference frame
(a)
In micromechanical systems, the choice of the resonant
frequency is also constrained by other considerations. Low
resonant frequency implies a low spring constant k and high
mass m. The mass of typical surface micromachined sensors
is below one microgram. A one micron displacement due to
a 1 g ( l g = 9.8 m/s2) acceleration would require a spring
constant k less than 10 mN/m. Material properties, geometric
current c-) force constraints, and self resonance of the spring set a lower bound
on 5 that is well above this value.
voltage
flux -
C-C

(b)
velocity
displacement
A further complication is that the resonance frequencies
in different directions cannot be chosen independently. For
example, the ratio of resonant frequencies about the y and z
axis for the cantilever beam shown in Fig. 2 with thickness
Fig. 1. Concept of an accelerometer (a) and the equivalent electrical model
(b). t and width w is equal to the ratio t / w [15], [16]. Typical
dimensions for the suspension of a z axis accelerometer are a
thickness of 2 pm and a width of 5 pm, corresponding to a
a simple single-axis linear accelerometer is shown in Fig. l(a). ratio of resonant frequencies of only 2.5. Therefore, in practical
The inertia of the proof mass restrains the motion of this situations, the suspension alone can provide only a portion of
element in the presence of an external force Fext acting on a a typical requirement for rejection of off-axis accelerations.
reference frame to which the proof mass is attached by means Accelerometer suspensions based on folded trusses have
of a spring. The proof mass is further subject to damping from several advantages, including independence of the spring con-
the surrounding gas ambient or from internal dissipation in the stant on residual stress in the film and the capability of low
spring. The differential equation for the displacement x as a spring constants in a small area. Torsional resonant modes are
function of Fe,, is obtained from Newton’s law potentially near the fundamental mode of the sense element,
making careful design essential. Analytical models of such
d2x
m-
dt2
+ b-dx
dt
+ kx = Fext= ma. (1) suspensions [ 161, [ 171 have been developed that provide
insight. Fig. 3 shows the result of a finite element analysis
In this equation k and b are the spring constant and damping simulation [18] of a proof mass suspended with four single-
coefficient, respectively, and linear relations are assumed. folded trusses that reveals a rotational resonance about either
Solving for x using the Laplace transform yields the second- diagonal at less than twice resonance for the desired motion
order transfer function along the z axis. Typical polysilicon surface micromachined
accelerometers have resonance frequencies in the kilohertz
range and a mass of between 0.1 and 1 pg.
This low mass gives rise to another design challenge faced
in high-sensitivity accelerometers. According to the laws of
with resonant frequency wT = and quality factor thermodynamics, the thermal energy of a system in equilib-
Q = w,m/b. Critical or under-damping (Q 5 0.5) of the rium is k ~ T / 2for each energy storage mode, where k13 is
sensor is assumed since this condition minimizes the thermal Boltzmann’s constant. The minuscule mass of the microma-
noise of the sensor, as will be seen later. chined device implies substantial agitation due to this thermal
At frequencies well below resonance, the displacement energy, a process known at the molecular level as Brownian
x M alw:, which is proportional to the acceleration. This motion [19]. The extent of this disturbance can be appreciated
relationship implies a trade-off between sensitivity and band- readily from the equivalent second-order system presented in
width of the sensor: low resonant frequency results in large Fig. 1(b), which is described by the same differential (1) as
displacements and hence, good sensor resolution but restricts the mechanical system shown in Fig. l(a). In the electrical
the bandwidth of the sensor. This trade-off can be eliminated equivalent, currents represent forces, and voltage corresponds
with feedback, as will be discussed later. to velocity. The electrical component values are proportional

Authorized licensed use limited to: National Science Council. Downloaded on July 6, 2009 at 12:34 from IEEE Xplore. Restrictions apply.
368 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 31, NO. 3, MARCH 1996

(b)
Fig. 3. Finite-element analysis of the resonant frequencies of a proof mass suspended with four single-fold trusses The plots illustrated an exaggerated view
of the motion corresponding to the desired first mode (a) at 7.5 lcHz The parasitic second and third modes occur at 12.6 kHz (b).

to the respective mechanical quantities. The current noise

- head-mounted displays demand better performance. Lower
zz/Af = 4 k B T / R from the resistor corresponds to an noise can be achieved by either increasing the mass or in-
equivalent force noise source associated with the mechanical creasing the quality factor Q by reducing the damping of the
damping element. It is in parallel with the external input sensing element.
Iekxt,corresponding to the measured force in the mechanical The first approach offers only limited improvement because
original, and has a white power spectrum. The equivalent of the relatively modest increase of the mass of the sensor
acceleration spectral density is obtained by back-substituting obtained, for example, by substituting a material with higher

/&y- ,,;,,-.
mechanical for electrical quantities and is density such as tungsten for polysilicon. The lateral dimensions
of the proof mass are restricted by economic constraints
V Q G m= (3) that call for minimizing silicon area as well as processing
difficulties. Polysilicon films typically have small gradients
in the residual stress that cause warpage of large-area plates
For a typical micromachined sensor with m = 0.5 pg, w, = which limits lateral dimensions to a few hundred microns in
271.10 kHz, and Q = 0.5, the input referred noise density is typical technologies.
approximately 200 pg/& at room temperature. By contrast, vacuum packaging of the sensor results in a
For many commercial applications (e.g., in the automotive several orders of magnitude decrease of the thermal noise
sector), this noise level is acceptable, but inertial navigation of the sensor and does not require any modification of the
and other precision applications such as tracking systems for sensor fabrication process. Quality factors of 50 000 have

Authorized licensed use limited to: National Science Council. Downloaded on July 6, 2009 at 12:34 from IEEE Xplore. Restrictions apply.
BOSER AND HOWE SURFACE MICROMACHINED ACCELEROMETERS 369

U P wall

N substrate

Fig. 4. Cross-section of a CMOS wafer with added micromechanical structures [16].

been reported in the literature for in situ phosphorus-doped control that was fabricated in the MICS technology [22].In this
polysilicon [SI, [lo], corresponding to a reduction of the device, the proof mass and spring suspension are fabricated
noise floor in above example from 200 pg/& to below 1 in poly level two. The third layer serves as a limit stop
pg/&. By comparison, the ratio of the density of tungsten and electrostatic actuator. The etch holes in the proof mass
to that of polysilicon is less than an order of magnitude. and top actuator are required for complete removal of the
Of course, the elimination of nearly all mechanical damping sacrificial oxides and to further reduce squeeze film damping,
requires that the position of the suspended mass be controlled thus improving the Q of the structure at atmospheric pressure.
electrically. The technology also permits the fabrication of polysilicon
fuses that hold the structure in place during the release process.
111. FABRICATION
TECHNOLOGY They are cut electrically with a current pulse.
Low strain gradient in the polysilicon films is important
Monolithic integration of the proof mass and suspension for to prevent overhanging structures from warping after release.
an accelerometer or other physical sensor with its associated Polysilicon cantilever beams fabricated in the MICS technol-
electronics requires the merging of surface micromachining ogy typically warp less than 0.1 pm over 400 pm in length.
with an IC fabrication process. Polysilicon and aluminum are Stiction of the sense element or its suspension to the limit stops
the most attractive candidates for the structural elements of or to the substrate due to overrange forces poses an additional
micromechanical sensors. Both are already used in standard challenge for the fabrication technology.
IC technology and can be deposited and patterned to very The first commercial integrated surface machining mi-
accurate dimensions. The negligible fatigue [20] and lack of crosensor technology, the BiMEMS process of Analog
memory of polysilicon make it the material of choice for the Devices, Inc. [23] uses a different integration strategy from
fabrication of high performance micromechanical sensors. The the MICS technology. The sacrificial oxide and structural
advantages of aluminum are its low processing temperatures, polysilicon film are deposited and annealed prior to a standard
which makes integration with electronic processing much more aluminum metallization. In BiMEMS, a diffusion is used to
straightforward. connect the polysilicon microstructure to the circuit, rather
In order to integrate polysilicon microstructures with CMOS than the poly level one and gate polysilicon levels used in
processing, the 600 "C deposition and 950 OC rapid-thermal MICS. Finally, BiMEMS is a linear bipolar/MOS technology
annealing temperature needed for the former must be con- that includes thin film resistors, making it very well suited to
sidered carefully. One approach is to replace the conventional implementing analog closed-loop control of the sense element.
aluminum metallization with tungsten, together with a titanium
silicide diffusion barrier in the contacts, to enable the CMOS
to withstand these high post-processing temperatures. Low IV. POSITIONSENSING
pressure chemical vapor deposition (LPCVD) silicon nitride Acceleration sensors translate the external signal in a cor-
is used to isolate the CMOS from the micromachining process responding displacement that can be measured by several
steps, such as the final etching of the sacrificial layer in means. Piezo-resistive strain gauges are used widely in sen-
hydrofluoric acid. The modular integration of CMOS and sors because of the simple interfacing to off-chip electronic
microstructures (MICS) process uses this integration strategy circuits. Ion-implanted piezo-resistors have been used in un-
and is based on a conventional 3 pm CMOS technology [21]. doped polysilicon resonant microbeams [24] and polysilicon
Fig. 4 shows a cross-sectional view of an MICS wafer. The piezo-resistors have also been embedded in silicon nitride
sensing structures can be seen on the right and consist of three membranes for pressure sensing [3]. Position sensors that
layers of 0.5 pm, 1.5 pm, and 1 pm thick phosphorus-doped measure the capacitance between a conducting polysilicon
polysilicon. The second and third layer are deposited on top of proof mass and a fixed electrode, on the other hand, require
1 and 2 pm thick sacrificial phosphosilicate glass layers that no additional processing and can be extremely sensitive, as
are later removed. will be shown later. The negligible temperature coefficient
Fig. 5 shows an scanning electron microscopy (SEM) pic- is another important advantage of a capacitive sensor read-
ture of a section of a multimode test structure for digital out. The low parasitics that are characteristic of monolithic

Authorized licensed use limited to: National Science Council. Downloaded on July 6, 2009 at 12:34 from IEEE Xplore. Restrictions apply.
370 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL 31, NO 3, MARCH 1996

ac

mt

Fig. 5. Detail of a polysilicon micromachined structure [22].

integration are the key to maximizing the performance with anchor suspension,
this technique. Researchers are also investigating alternative
transduction methods such as tunneling tips in surface mi-
crostructures [25]-[27].
Fig. 6 schematically shows quarter-sections of two sensor
designs with sensitivity to accelerations perpendicular and
parallel to the silicon surface. In the z axis structure, the
capacitance between the proof-mass and the substrate serves
as the mechanical to electrical interface [5];the y axis design
uses a comb-like structure [6].The advantage of the first style
interface is the usually larger capacitor area and value, but
the interface is asymmetric when a top electrode is missing.
anchor tether proof mass
Over the finger structure, the parallel plate arrangement has
the advantage of a much larger sense capacitance for a given
area of up to 1 pF, compared to less than 200 fF for typical
y axis sensing elements.
For position measurement, the variation dC/dx of this
capacitance due to displacement must be maximized. For the
z axis structure, a parallel plate approximation is usually
appropriate and for small displacements, the capacitor change
is approximately equal to C / X OMaximizing
. the sensitivity substrate
therefore calls for minimizing the capacitor gap, XO. In the
second structure, a substantial fraction of the total capacitance (b)
is due to fringing fields and does not change substantially Fig 6. Accelerometers with z (a) and y axis (b) sensitivity
in the presence of small displacements [l6], [as], resulting
in a somewhat lower sensitivity of this arrangement for a capacitor with a 1 pm gap and nominal value of 1 pF changes
given electrode spacing. Exact values can be obtained from by a mere 10 aF due to this displacement, an amount that is
numerical simulation [29]. 100000 times smaller than the capacitor itself. Flicker noise,
The challenge in the design of a capacitive position mea- offsets, and parasitic capacitances represent further difficulties.
surement circuit consists in detecting extremely small capac- The circuit shown in Fig. 7 uses chopper stabilization
itance changes in the presence of much larger parasitics. For and bootstrapping to minimize the noise and maximize the I
example, a sensing element with resonant frequency fr = 5 sensitivity. The voltage V, at the midpoint of the divider
kHz experiences a displacement of only 0.1 A in the presence formed by the two sense capacitors and excited with the ac
of a 1 mg constant acceleration signal. A parallel plate signal V, is proportional to the capacitor mismatch and, hence,

Authorized licensed use limited to: National Science Council. Downloaded on July 6, 2009 at 12:34 from IEEE Xplore. Restrictions apply.
BOSER AND HOWE SURFACE MICROMACHINED ACCELEROMETERS 371

+vsy
7 Csl2

-vs

Fig. 7. Sensor interface using a unity-gain buffer.

C&T
Electronic noise as a function of transistor size.

it is important to derive relationships relating this noise to

basic physics and technology parameters.
At least in principle, all noise sources except the input
referred thermal noise 3
from the amplifier can be made
insignificantly small. For either the circuit shown in Figs. 7 or
-vs 8, this noise appears across all capacitors connected to node
V, and produces an equivalent current noise that is added to
Fig. 8. Sensor interface based on correlated double sampling. the signal current i, = V,AC. In an appropriately designed
amplifier, the input transistor is the dominant noise contributor,
and hence
position of the proof mass. The measurement is performed -
3
at a sufficiently high frequency to suppress offset and flicker
noise and the result demodulated. The parasitic capacitance
in node V, must be minimized to avoid signal attenuation.
assuming a differential input. Apparently, the noise can be
Many designs reduce the attenuation by shielding interconnect
reduced by increasing the transconductance gm. Increasing
capacitances and tying the shield to the output of a unity gain
gm, however, requires increasing either the saturation voltage
buffer as shown. Shields are also needed to avoid electrical
of the input devices, which in practice is limited by the
fields between the sensor and surrounding conductors since
the resulting electrostatic forces are indistinguishable from an supply voltage, or enlarging the gate capacitance C,,. This,
external force and, hence, corrupt the acceleration measure- in turn, increases the capacitance at node V, and hence, the
current noise, thus calling for a trade-off between gm and
ment. Special provisions are needed to control the dc potential
C,,. For a given saturation voltage, these two quantities are
at node V,. The solution shown here relies on a resistor R d c
that typically must be in the megohm range to minimize its related as gm M W T C ~by, the cutoff frequency WT of the
technology. Fig. 9 shows the normalized current noise as a
noise contribution.
function of the ratio of C,, to the sum of all other capacitances
An alternative solution presented in Fig. 8 eliminates both
the demodulator and the need for a resistor or similar element
CT = C, + C, connected to the amplifier input. Clearly, the
noise is minimized for C,, = CT, with a small penalty for
to set the dc potential [7], [30]. This circuit is based on
somewhat larger values of C,, as might be required to meet
an amplifier with an auxiliary input A2 with reduced gain
other circuit requirements, such as bandwidth.
and operates in two phases. First, the sense capacitors are
Based on these considerations and assuming that the ampli-
precharged to a constant voltage. At the same time, the offset
fier input capacitance is chosen optimally, the mean squared
and flicker-noise of the amplifier are stored on the holding
error of the capacitance change divided by the sense capaci-
capacitor Ch. During the second phase, the voltage across
tance C, is
the sense capacitors is changed, causing a charge that is
proportional to the mismatch between the two sense capacitors
to flow into node V,. Since the amplifier input is now a
virtual ground, this charge flows onto the integrating capacitor
C, unattenuated by the parasitic C,, producing an output V,, For a parallel plate sense capacitor with gap 20,the corre-
which is proportional to the position of the proof-mass.
sponding input referred acceleration noise floor is
Like Brownian motion of the sensing element, the noise
floor of the position measurement interface sets a limit on
the achievable sensitivity of the sensor. To assess the ultimate (6)
performance achievable with surface micromachined sensors, 3wT cs

Authorized licensed use limited to: National Science Council. Downloaded on July 6, 2009 at 12:34 from IEEE Xplore. Restrictions apply.
312 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 31, NO. 3, MARCH 1996

This expression demonstrates how the sensitivity of an ac-

celerometer that is limited by electronic noise in the position
measurement circuit can be maximized. Upon first inspection,
the biggest improvement appears to result from minimizing the
proof mass 1
position sense compensator
I
resonant frequency. Accelerometers described in the literature Ffb
feature resonant frequencies in the range of 5 kHz [14]
to over 20 kHz [ 113. Unfortunately, substantially lowering force transducer
these frequencies compromises the reliability of the device,
in particular its robustness to mechanical shock, and increases Fig. 10. Accelerometer with analog force-feedback servo loop.
the probability of the proof mass sticking to the substrate.
Better suppression of undesired resonances is needed as well
before the resonant frequency can be reduced substantially. important in accelerometers because of the trade-off between
Lowering the capacitor gap xo also decreases the noise floor sensitivity and bandwidth imposed by the sensing element.
of the sensor but, of course, depends on technology. In z Feedback increases the useful bandwidth by a factor equal
axis structures, the spacing is dictated by the thickness of to the loop gain, which, thanks to the electronic circuitry,
the sacrificial oxide and must be chosen sufficiently large to can be made large. Consequently, the resonant frequency of
prevent the sensing element from touching the substrate due the sensing element can be optimized for sensitivity alone,
to warpage and mechanical excitation. For current polysilicon regardless of the desired sensor bandwidth.
surface micromachining technologies, minimum vertical or Controlling the displacement of the proof mass is equally
lateral gaps are around 1 pm. important, particularly for vacuum-packaged high-Q devices
The remaining parameters in (6) characterize the electronic which can exhibit motion at the resonant frequency that
interface of the sensor. Accordingly, the sense capacitance C, exceeds the small spacings between the electrodes of the sense
should be maximized, a condition that favors z axis over y capacitor. Finally, imperfections, for example, due to nonlinear
axis designs. The parasitic capacitance C, should not exceed or temperature sensitive springs, are attenuated provided that
the value of C, for not significantly degrading the sensor the force-transducer does not introduce similar errors.
performance. High cutoff frequency and, hence, short channel Electrostatic actuation is the simplest means for generating
length also improve sensor sensitivity. Increasing the sense the feedback force in a micromechanical sensor. The electrodes
voltage V, results in an increased signal current and, hence, of a capacitors with a constant voltage V, across are attracted
better signal-to-noise ratio. In practice, V, is often limited with a force
not only by the supply voltage, but also by the maximum
acceptable electrostatic force exerted on the sensor. This F, = d C(x(j)V,2- - __
v:c ~

force can degrade the linearity of the sensor (see Section V- dxo 2 2x0
B). in the case of a parallel plate capacitor with value C and gap
Typical numbers for surface micromachined sensors, C, =
20. The same capacitor can be used for force-feedback and
0.5 pF, C, = C,, WT = 2.ir x 500 MHz, V, = 0.5 V, fr = 5
sense 1111, or separate electrodes can be added [5].
kHz, and xo = 1 pm give a noise floor of 1 p g / a at room Because of the quadratic dependence of the force on voltage,
temperature. This figure is of the same order as the Brownian electrostatic actuators cannot be used directly, but must be
noise of a vacuum-packaged structure. While this performance
combined with some means of linearization. For symmetric
level has not yet been achieved with micromachined sensors,
sensors [Fig. 6(b)], a simple solution is to apply a voltage
these numbers demonstrate the potential of this technology for
Vi + AV across C+ and V, - AV across C-. Since the
very sensitive inertial measurement applications.
two resulting forces are in opposite directions, the quadratic
The 1 p g / z / H z noise floor corresponds position measure-
terms cancel and the net difference A F = ~ V O A V C /is~ O
ment noise of only A/z/Hz. This means that displace- proportional to the controlling voltage AV. In practice, the
ments as small as the classical diameter of an electron can be
linearity of this technique is limited by the matching accuracy
detected in a 10 Hz bandwidth. In an actual circuit, the noise
of C+ and C-. More importantly, this approach cannot be
would be somewhat larger due to additional noise sources
used at all with asymmetric sensors such as the one depicted
neglected in this analysis.
in Fig. 6(b).
A more general and potentially more accurate solution
V. FORCE-FEEDBACK consists in pulse modulating the feedback signal. Fig. 11
Fig, 10 shows an acceleration sensor embedded in a feed- shows a system that employs a clocked comparator to quantize
back loop. A compensator and force-feedback transducer are the feedback force to only two levels [5], [7], [14]. This
added to the open loop sensor consisting of the proof mass and system is equivalent to a sigma-delta modulator as used in
position measurement circuitry. The feedback force opposes A D conversion [31], except that the noise shaping filter
displacements of the proof mass from its nominal position. has been replaced by the mechanical sensor. If all feedback
Compensation is required for stability and will be discussed pulses are kept equal in length, imbalance merely results in
later. an offset and/or gain error, but does not cause distortion.
Feedback improves many important characteristics of a sen- The pulse-density of the one-bit output stream tracks the
sor, including bandwidth, dynamic range, and, in certain cases, input acceleration, which is obtained by low-pass filtering and
linearity and drift. Increasing the bandwidth is particularly decimating the pulse-density code. Because of the inherent

Authorized licensed use limited to: National Science Council. Downloaded on July 6, 2009 at 12:34 from IEEE Xplore. Restrictions apply.
BOSER AND HOWE: SURFACE MICROMACHINED ACCELEROMETERS 373

vo

I ProofmaSS pasitionsense
Hc(z)
compensator comparator
II -
I I---I i I i
Fig. 11. Accelerometer with digital force-feedback loop.

AD conversion, this approach is referred to as “digital”

feedback, while the technique illustrated in Fig. 10 is called
“analog” feedback in this paper. X
The feedback loops will now be analyzed with respect to ~

stability, residual motion of the proof mass, and, for the case
of digital feedback, quantization noise.

A. Stability
The force-feedback accelerometers shown in Figs. 10 and
11 are unstable without compensation because of the 180
(b)
degree phase delay from the sensing element for frequencies
Fig. 12. Stability of accelerometer with digital feedback loop for two dif-
above the resonance. Several solutions are used in accelerom- ferent actuation delays t d .
eter designs and will be described below: over-damped proof
mass [32], limiting the loop bandwidth electronically [ 111,
[331, and compensation with a lead filter 151, [71, 1341. frequency of the feedback signal. The maximum possible rate
The first solution is simple but because of the high Brownian is f s 14,constrained by the clocking and second-order nature
noise floor of an over-damped low-mass sensing element, of the feedback loop, and is achieved when the total phase
practical only for low performance applications or when much delay is less than 180’ at this frequency. This condition is
larger and thus heavier sensing elements than those built with met when the phase lead from the compensator is larger than
surface micromachining technology are used. approximately t d f s x 4 5 O . Here, t d is the delay from sensing
The second approach uses an electronic low-pass filter to the position of the proof mass to applying the feedback
reduce the loop bandwidth to frequencies well below the signal.
resonance of the proof-mass. Obviously, this technique is Simulation results for two loops operating under this con-
applicable only in situations where either a low measurement dition are shown in Fig. 12. In both cases, a compensator
bandwidth or low sensitivity can be tolerated, since in this case H,(x)= 2 - z-l is ased to add a zero at fs/9 and
the sensor bandwidth is actually reduced to a value less than approximately 27’ phase lead at fs/4. In the first case,
the resonant frequency f T of the sensing element. Stability short feedback pulses are issued almost immediately after
demands that the loop gain is lower than the ratio of f r to the position of the proof mass has been measured. The
the sensor bandwidth. A commercial surface micromachined compensator output V, leads V, despite a small lag of V,,
accelerometer implements this technique [ 111. The sensor has suggesting that a smaller amount of lead (e.g., 5 - z - l ) would
a resonant frequency of 24 kHz and 1 kHz signal bandwidth suffice in this case. Simulation confirms that this is indeed the
and a low-frequency loop gain of about 10, thus diminishing case. In the second case, the feedback pulses are elongated
somewhat the benefits of feedback. and delayed by t d = T / 2 causing V, to arrive just in time
A compensation filter requires additional circuitry but to ensure a correct decision by the comparator. The longer
avoids the aforementioned problems. The basic strategy is feedback pulses round the edges of the position signal, but for
to add a left half-plane zero to the loop transfer function in stability, only the delay t d from position measurement to the
order to decrease the phase delay at the unity-gain frequency. midpoint of the feedback pulse is relevant.
The analysis is straightforward for analog feedback, but
complicated by the nonlinearity in the digital case.
The conventional definitions of stability involving the B. Residual Motion of the Proof Mass
boundedness of states and absence of limit cycles are not The sense and feedback capacitors exert a position depen-
useful for sigma-delta modulators which use oscillations as dent force on the proof-mass. Residual motion consequently
a means for A/D conversion. A more appropriate criterion results in an error force that cannot be distinguished from
follows from analyzing the spectrum of the pulse-density the accelerometer input. Assuming a zero external input and
output. For illustration purposes, consider the digital feedback proper compensation of the feedback loop, the feedback signal
loop in Fig. 11 with zero input. Then the comparator output of the accelerometer is pulse-train at one quarter the sampling
and feedback signal switch rapidly between positive and frequency f s and amplitude amaxequal to the input range of
negative values. To minimize the residual motion of the proof the device. This feedback signal causes the proof mass to move
mass, it is important to maximize the rate of this signal: up and down at the frequency f s / 4 . Owing to the second-order
motion is reduced by a factor four for every doubling of the nature of the proof mass, the amplitude of the fundamental

Authorized licensed use limited to: National Science Council. Downloaded on July 6, 2009 at 12:34 from IEEE Xplore. Restrictions apply.
 374 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL 31, NO 3, MARCH 1996

of the motion is Ax = ~ ~ ~ / ( 2 x f ~Assuming / 4 ) ~ . a full-

scale input range amax = 2 g and f s = 10 MHz, Ax M
l0W3A. Noise or an external acceleration signal randomize
the spectrum of the proof mass displacement and give rise to
low-frequency components falling into the signal-band of the
sensor.
Capacitive position sensing and feedback both exert an
electrostatic force on the proof mass. To first order, this force is
proportional to the position of the proof mass. Residual motion 1 -

T = 4/fs
modulates these forces, resulting in a spurious acceleration
of Fig. 13. Block diagram illustrating the origin of a dead-zone in sensors with
digital feedback.
AX
Aa M (asense + amax)-- (8)
20
VI. CONCLUSION
Usually, the acceleration due to the sense capacitor dominates The advantages of monolithic fabrication of micromachined
because of the need to maximize the sense voltage V, in sensors and associated electronics have been discussed. It has
order to minimize the noise floor of the position measurement been demonstrated that the limits of the technology are beyond
circuit. A typical asense= 50 g results in an error Au M a 1 p g / G noise level, which is compatible with a large
8 pg. Simulations and measurements show that this error variety of demanding applications.
is sufficiently random to raise the noise floor of the sensor.
It is most easily suppressed by choosing a sufficiently high
sampling rate fs. ACKNOWLEDGMENT
A sensor with analog feedback and identical loop bandwidth The authors are very indebted to their students who did
will exhibit the same error, however, because of the absence most of the work reported here.
of limit cycles, it will not result in an increased noise floor but
instead in a slightly increased nonlinearity.
REFERENCES

[l] R. T. Howe, “Surface micromachining for microsensors and microactu-

C. Quantization Noise ators,” J. Vacuum Sci. Technol. B , vol. B6, pp. 1809-1813, 1988.
[2] L. Hombeck, “Current status of the digital micromirror device (DMD)
Digital feedback adds quantization noise as an additional for projection television applications,” in IEDM Tech. Dig., Dec. 1993,
error source [35]. Because of the typically high ratio of the pp. 381-384.
sampling frequency f s to the signal bandwidth of interest, [3] S . Sugiyama et al., “Micromachined sensors using polysilicon sacrificial
layer etching technology,” in IEDMTech. Dig., Dec. 1994, pp. 127-130.
this error is usually insignificant. However, the finite resonant [4] B. E. Cole et al., “512 x 512 infrared scene projector array for low-
frequency of the proof mass causes a dead-zone for inputs background simulations,” in Tech. Dig. Solid-State Sensor and Actuator
less than Workshop, June 1994, pp. 7-12.
151 W. Yun, R. T. Howe, and P. R. Gray, “Surface micromachined,
digitally force-balanced accelerometer with integrated CMOS detection
circuitry,” in Proc. IEEE Solid-state Sensor and Actuator Workshop,
(9) June 1992, pp. 126-131.
[6] S. J. Sherman et al., “A low cost monolithic accelerometer: Prod-
uct/technology update,” in IEDM Tech. Dig., Dec. 1992, pp. 501-504.
For am= = 2 g and f s / f r = 1000, ad M 16 pg. Like errors [7] C. Lu, M. Lemkin, and B. E. Boser, “A monolithic surface microma-
due to residual motion of the proof mass, the dead-zone width chined accelerometer with digital output,” in ISSCC Dig. Tech. Papers.
can be decreased by raising the sampling rate, f s . Smaller IEEE Int. Solid-state Circuits Con$, 1995, pp. 160-161.
[8] C. T. Nguyen and R. T. Howe, “CMOS micromechanical resonator
dead-zones exist also for other inputs that are rational fractions oscillator,” in IEDM Tech. Dig., Dec. 1993, pp. 199-202.
of amax 1311. [9] L. Lin, C. T.-C. Nguyen, R. T. Howe, and A. P. Pisano, “Micro
The cause of the dead-zone can be appreciated from the electromechanical filters,” in Proc. Fifth IEEE Micro ,Electromechanical
Systems Workshop, Feb. 1992, pp. 226-231.
block diagram in Fig. 13, where the proof mass is modeled as [ 101 C. T.-C. Nguyen, “Polysilicon microresonators for frequency references
a linear filter H ( s ) and the order of the linear filtering action and signal processing,” Ph.D. dissertation, University of California,
and feedback summing node have been reversed. Assuming Berkeley, Dec. 1994.
[ 111 Analog Devices, “ADXLSO-Monolithic accelerometer with signal con-
for a moment that the input of the accelerometer labeled amin ditioning,” Datasheet, 1993, One Technology Way, Norwood, MA
is zero, the feedback signal is a square wave with amplitude 02062.
amaxand frequency fs/4. To be detected, an input signal must [12] L. Ristic, R. Gutteridge, J. Kung, D. Koury, B. Dunn, and H. Zunino, “A
capacitive type accelerometer with self-test feature based on a double-
disturb this idle pattern of the modulator. This occurs only for pinned polysilicon structure,” in Dig. Transducers ’93, June 1993, pp.
displacements U that are at least equal in amplitude to the idle 8 10-8 13.
channel residual motion w. [13] R. S. Payne, S. Sherman, S. Lewis, and R. T. Howe, “Surface microma-
chining: From vision to reality to vision,” in ISSCC Dig. Tech. Papers.
In oversampled AID converters with electronic noise shap- IEEE Int. Solid-State Circuits Con$, 1995, pp. 164-165.
ing filters, the resonant frequency of the mechanical noise [14] C. Lu, M. Lemkin, and B. Boser, “A monolithic surface micromachined
shaper corresponds to the filter pole frequencies. These are accelerometer with digital output,” IEEE J. Solid-state Circuits, vol. 30,
no. 12, pp. 1367-1373, Dec.-1995.
usually at much lower frequencies; consequently, dead-zones [l5] J. M. Gere and S. P. Timoshenko, Mechanics ofMaterials. Belmont,
are not observed. CA: Wadsworth, 1984.

Authorized licensed use limited to: National Science Council. Downloaded on July 6, 2009 at 12:34 from IEEE Xplore. Restrictions apply.
BOSER AND HOWE: SURFACE MICROMACHINED ACCELEROMETERS 375

[ 161 G. K. Fedder, “Simulation of microelectromechanical systems,” Ph.D. [34] W. Henrion, L. DiSanza, M. Ip, S. Terry, and H. Jerman, “Wide dynamic
dissertation, University of California, Berkeley, Sept. 1994. range direct digital accelerometer,” presented at Proc. IEEE Solid-state
[17] M. W. Judy, “Micromechanisms using sidewall beams, Ph.D. disserta- Sensor and Actuator Workshop, June 1990.
tion, University of California, Berkeley, Apr. 1994. [35] B. E. Boser and B. A. Wooley, “The design of sigma-delta modulation
[18] ABACUS User’s Manual, 5.2 ed., Hibbitt, Karlsson & Sorenson, Inc., analog-to-digital converters,” IEEE J. Solid-state Circuits, vol. 23, no.
1080 Main Street, Pawtucket, RI, 1992. 6, pp. 1298-1308, Dec. 1988.
[ 191 T. B. Gabrielson, “Mechanical-thermal noise in micromachined acoustic
and vibration sensors,” IEEE Trans. Electron Devices, vol. 40, no. 5,
pp. 903-909, May 1993.
[20] D. W. Burns et al., “A digital pressure sensor based on resonant
microbeams,” in Tech. Dig. Solid-State Sensor and Actuator Workshop,
June 1994, pp. 221-228.
1211 J. M. Bustillo, G. K. Fedder, C. T.-C. Nguyen, and R. T. Howe. “Process
technology for the modular integratkn of CMOS and polysilicon Bernhard E. Boser received the Diploma in elec-
microstructures,” Microsystem Technol., vol. 1, pp. 3041, 1994. trical engineering in 1984 from the Swiss Federal
[22] G. K. Fedder and R. T. Howe, “Integrated tested for multi-mode digital Institute of Technology (ETH) in Zurich, Switzer-
control of suspended microstructures,” in Proc. IEEE Solid-State Sensor land, and the M.S. and Ph.D. degrees from Stanford
and Actuator Workshop, June 1994, pp. 145-150. University, Stanford, CA, in 1985 and 1988, respec-
[23] T. Core, W. Tsang, and S. Sherman, “Fabrication technology for an tively.
integrated surface-micromachined sensor.” Solid State Technol.. DD.
39-47, Oct. 1993.
__ From 1988 to 1991 he was a Member of Technical
Staff at AT&T Bell Laboratories, working on VLSI
~. H. Guckel et al., “Polvsilicon resonant microbeam technology
1241 -_for high implementations of artificial neural networks and
performance sensor applications,” in Tech. Dig. IEEE Solid-State Sensor algorithms for automatic learning. Since 1992 he
and Actuator Workshop, June 1992, pp. 153-160. has been an Assistant Professor in the Department
[25] P. M. Zavracky, F. Hartley, N. Sherman, T. Hansen, and K. Warner, of Electrical Engineering and Computer Sciences, University of California,
“A new force balanced accelerometer using tunneling tip position Berkeley. His current research interests are in the areas of integrated circuits
sensing,” in Dig. Int. Con$ Solid-state Sensors and Actuators, 1993, for data conversion and communication, and on the design and fabrication of
pp. 50-51. micromechanical systems.
[26] H. K. Rockstad, T. W. Kenny, J. K. Reynolds, W. J. Kaiser, and T.
B. Gabrielson, “A miniature high-sensitivity broad-band accelerometer
based on electron tunneling transducers,” in Dig. Int. Con$ Solid-state
Sensors and Actuators, 1993, pp. 836-839.
[27] N. C. MacDonald, “Nanomechanisms and tips for microinstruments,”
in Tech. Dia. Seventh Int. C o d on Solid-state Sensors and Actuators.
June 1993, pp. 8-12.
W. C. Tang, T.-C. H. Nguyen, and R. T. Howe, “Laterally driven Roger T. Howe (S’79-M’84-SM’94-F’96) was
polysilicon resonant microstructures,” Sensors and Actuators, vol. 20, horn in Sacramento, CA on April 2, 1957. He
pp: 25-32, 1989. received the B.S. degree in physics from Harvey
Maxwell Solver, Electrostatic Package, 4.20 ed. Ansoft Corp., 4 Station Mudd College, Claremont, CA in 1979, and the
Square, 660 Commerce Court Bldg: Pittsburgh, PA 15219.- M.S. and Ph.D. degrees in electrical engineering
M. Degrauwe, E. Vittoz, and I. Verbauwhede, “A micropower CMOS- from the University of Califomia at Berkeley, in
instrumentation amplifier,” IEEE J. Solid-State Circuits, vol. SC-20, no. 1981 and 1984, respectively.
3, pp. 805-807, June 1985. He was on the faculty of Carnegie-Mellon Uni-
J. C. Candy and G. C. Temes, “Oversampling methods for A/D and versity during the 1984-85 academic year and was
DIA conversion,” in Oversampling Delta-Sigma Data Converters, J. an Assistant Professor at the Massachusetts Institute
C. Candy and G. C. Temes Eds. New York: IEEE Press, pp. 1-29, of Technolow from 1985 to 1987. In 1987. he
“2

1992. joined the Department of Electrical Engineering and Computer Sciences at

T. Smith, 0.Nys, M. Chevroulet, and Y. DeCoulon, “A 15b electro- the University of Califomia at Berkeley, where he is now a Professor, as well
mechanical sigma-delta converter for acceleration measurements,” in as a Director of the Berkeley Sensor & Actuator Center. His research interests
ISSCC Dig. Tech. Papers, IEEE Int. Solid-State Circuits Con$ , 1994, include silicon microsensors and microactuators, micromachining processes,
pp. 160-161. and integrated-circuit design.
Analog Devices, “ADXLOS-Monolithic accelerometer with signal con- Dr. Howe served as CO-General Chairman of the 1990 IEEE Micro Electro
ditioning,” Datasheet, 1995, One Technology Way, Norwood, MA Mechanical Systems Workshop and is General Chairman of the 1996 Solid-
02062. State Sensor and Actuator Workshop at Hilton Head, SC.

Authorized licensed use limited to: National Science Council. Downloaded on July 6, 2009 at 12:34 from IEEE Xplore. Restrictions apply.