You are on page 1of 15

1

Extended Sense Finger MEMS Gyroscope


Guo Jie Chin, Gen Masuda, and William Myers
the sense motion, lateral comb ngers are used for driving structures, while parallel-plate sense capacitors are used to sense the Coriolis-induced sense displacement, which can be on the picometer scale. [4] Ideally, a gyroscopes lateral comb ngers apply a drive force perfectly aligned with the drive axis to resonate the drive mass. In this situation, the sense mass will move along the sense axis only if there is a Coriolis force. However, manufacturing imperfections cause departures from the ideal mass, stiffness and damping distribution and therefore affect the dynamics. Two main sources of such errors are the Quadrature signal and the Coriolis offset. The quadrature signal is due to manufacturing imperfections causing the drive displacement to have a vector component in the sense direction. Like the Coriolis motion, this displacement along the sense direction is at the drive frequency d . There is a 90 phase difference between this error signal and the desired Coriolis output. The magnitude of this error is usually very large compared to the typical Coriolis output, and it is thus a large source of error. The ratio of quadrature force to Coriolis force is expressed in Equation 1. [2]: Fquadrature Fcoriolis = x 2z (1)

AbstractThis paper details the design, analysis, and simulation of a micro machined gyroscope using the MCNC MUMPS process ow. Index TermsMEMS; gyroscope; Cornell.

I. I NTRODUCTION YROSCOPES have found their calling across the globe and have permeated many diverse industries. With uses ranging from Department of Defence ballistics navigation to video game motion sensing, there exists a large market to exploit: especially if you can deliver a quality device that is affordable, small, and easily integrated. MEMS devices have heeded that call and provided marvellously cost effective and sensitive gyros for the better part of a decade. This paper aims to explore the design, simulation, and testing of one such device. The development of this device was reminiscent of George Boxs timeless adage, All models are wrong, some models are useful. Figure 1 depicts a drastically simplied and miniaturized schematic of the gyroscope that will be considered in this paper.

Here, is the fraction of drive displacement that has a component along the sense direction. Coriolis offset arises due to the drive force having a component along the sense direction. Unlike Quadrature signal, however, Coriolis offset is precisely in phase with the desired Coriolis signal and it is thus indistinguishable from Coriolis signal. The ratio of force from Coriolis offset to Coriolis Force [2] is expressed in Equation 2. Fcoriolis offset x = Fcoriolis 2Qx z
Fig. 1. Reduced Gyroscope Design

(2)

II. BACKGROUND The key requirements for the gyroscope are as follows: ARW: 0.1 / Hr Full Scale Range: 500 /sec Bandwidth: 1-100 Hz 2 Maximum Chip Area: 3mm There are multiple families of MEMS gyroscopes, but the focus of this paper will be on the analysis of a frame gyroscope. Frame gyros have the ability to decouple the drive and sense modes, resulting in lower output errors. As resonant motion in the drive direction is typically much larger than

Here is the fraction of drive force that has a component along the sense direction. Thus, both Quadrature signal and Coriolis offset error forces are usually comparable to or even larger than the Coriolis force that is being sensed. It is thus imperative that such forces are not allowed to couple to the sense mass and corrupt the Coriolis force measured by the sense capacitors. III. IDOS AND ISOD S TRUCTURES MEMS frame gyroscopes may be subdivided into InsideDrive-Outside-Sense (IDOS) and Inside-Sense-Outside-Drive (ISOD) gyros. Each conguration has its respective advantages and disadvantages. However, in light of the sources of error mentioned above, we have chosen to adopt an ISOD conguration for our gyroscope instead of the more conventional IDOS

conguration. We believe that this design is less sensitive to Coriolis offset and Quadrature error. Indeed, ISOD frame gyroscopes may have a factor of three lower Quadrature and Coriolis offset and a factor of ve lower noise oor (limited by electronic noise) in comparison with IDOS frame gyroscopes [3]. In an IDOS gyroscope, error forces arising from the drive combs couple directly to the sense structure, causing large output errors. However, ISOD frame gyroscopes, which have drive springs that are much stiffer along the sense axis, avoid this problem. In our ISOD gyroscope, drive force is applied to the outer frame (drive mass), which causes the entire mass to oscillate along the drive axis. In response to a z-axis rotation rate, Coriolis force acts on both the inner and outer masses. However, only the inner sense mass responds to Coriolis force due to stiff drive springs along the sense axis [See Mechanical Design]. Another advantage with ISOD gyros is that reduced damping and an increased lateral pull-in voltage can be expected in the sense mode. For ISOD gyros, damping and the lateral pull-in along the sense axis are caused by the sense combs only, whereas for IDOS gyros, both the sense and drive combs contribute towards it. Lower damping along sense mode increases the Coriolis signal, and thereby, the scale factor for the device. [4] The beauty of the ISOD gyroscope is that the drive and sense motions are guaranteed to be orthogonal. Further, since only the sense mass reacts to the Coriolis force, the associated damping is relatively small. This leads to higher quality factor and sensitivity of the device. In comparison, with an IDOS gyroscope, both the drive and sense mass react to Coriolis force, resulting in higher damping. IV. ISOD F EATURES Unfortunately, with an ISOD Frame gyroscope the sense combs are driven along the drive direction with the outer mass. This drive motion (several orders of magnitude larger than the Coriolis motion) cause changes in the area of overlap between sense capacitors, and hence causes the sense capacitance to uctuate if the sense ngers are implemented in the typical comb fashion. This is an especially devastating effect because the sense mechanism relies on gap closing capacitors to detect very small displacements. We avoided this problem by using an alternative sense comb design adapted from Moorthi Palaiapan [4]. In our design (See Figure 2.) the Coriolis sense motion is detected by the change in capacitance of the parallel sense combs. Two xed sense combs are surrounded by the suspended sense mass. Hence, even when the mass is oscillating in the drive direction, the area of overlap and sense capacitance is kept constant. Further, to reduce unwanted changes in the sense capacitance in the drive direction, a shielding electrode is used as shown in Figure 2. V. FABRICATION AND C OST In order to minimize the cost of the device, we elected to use the existing MCNC MUMPS Process Flow. Although the minimum resolution in the MUMPS process is a relatively

Fig. 2.

Sense Comb

poor 2m, we believe that this is sufcient given the large 3 mm by 3 mm scale of our device. We only utilized the 2m thick Poly1 layer for our structures, with Poly0 serving as a ground layer. As such, the associated cost for our design is simply dened by the MUMPS process, which is $2500/cm2 . By rescaling this to our gyroscope size, the estimated cost is $225 per gyro. VI. M ECHANICAL D ESIGN Each of the four corners of the main mechanical structure is anchored to the substrate via exures. These mechanical spring suspensions support the structure while allowing it to vibrate in the drive direction. The material used for these suspensions is the same structural material used for the proof mass (polysilicon). This exure design is compliant in the direction of oscillation (driving) and stiff in the orthogonal direction (sensing). In our design, the suspensions used for the outer mass are composed of two xed-xed beams connected in series. The effective spring coefcient is derived in Equation 3. kxed-xed =
3

192EI L3

(3)

Here, E = 170Gpa, I = wh , w = 2m and h = 4m. 12 Taking into account that each corner has two of these in series and their are 4 of these corner springs in parallel, our nal kdrive was 1586 N/m. The outer spring suspension can be seen in Figure 3.
Fig. 3. Drive Spring

The inner mass is supported by a simpler tracked beam suspension consisting of 4 tracked breams in parallel. The effective spring coefcient can be found from Equation 4. ktracked = 12EI L3
3

(4)

Here, E = 170Gpa, I = wh , and w = h = 2m. 12 Considering the 4 of these springs in parallel supporting

the proof-mass, our ksense was 1175 N/m. The outer spring suspension can be seen in Figure 4.
Fig. 4. Sense Spring

operating with far too much damping given our masses, and made the decision to move to a vacuum to actuate and sense over greater distances. To nd 0 in a vacuum we rst needed to nd the mean free path for air by Equation 8. P = 5.1 10 5 Torr m (8)

In order to isolate the motion of the inner mass (See IDOS and ISOD Structures), the inner suspensions should be rigid in the drive direction. The axial Stiffness of these inner beams is expressed in Equation 5 AE (5) L Hence, the sense mass displacement in the drive direction is negligible. kaxial = VII. M ECHANICAL R ESONANCE There are two fundamental ways to calibrate the resonant properties of a MEMS framed gyro. The outer mass frame and inner proof mass of the gyro will each exhibit their own unique mechanical resonance depending on their effective mass and the effective spring constant they are coupled to. Well rst direct our attention to the simpler case of the inner mass. Using the density of Si as 2330 kg/m3 and volume of our structure, we calculated the expected mass to be 0.145 nKg. To nd the resonant frequency, we simply plugged this mass and the aforementioned k value into Equation 6. ks 0,s = = 45.3kHz (6) m The outer mass resonance is very similar, except we simply plugged in meff = mi + mo and the normal outer spring constant kd to get 0,d = 45.21 kHz (note the frames mass was 5.15 nKg). If we needed more precision out of this measurement, we would take into account the spring coupling between the inner and outer mass by applying the Rayleigh-Ritz approximation. However, because the suspensions are designed such that the inner and outer masses oscillate orthogonally, we were able to design the inner springs to be extremely stiff in the drive direction, validating our simplication of the model. An important effect to consider for the project was damping, namely slide-lm damping. This was exceptionally important given the extremely large areas involved (and the damping scales with surface area as well soon see). The general expression for slide lm damping is A v (7) d where 0 is the viscosity of the medium, A is the area parallel with the surface, d is the offset from the surface, and v is the velocity of the structure. We can extract the damping coefcient b from this by noting F = bv, so b = 0 A . When d analytically solving for displacements, we realized we were Fdrag = 0

From this we can nd Kn , a measure of the gas rarefaction effect: Kn = where d is the gap to the substrate. Finally, we d can approximate eff by the following the empirical relation given in Equation 9. [1] eff =
0.788 1 + 2Kn + 0.2Kn e
Kn 10

(9)

We ended up applying 100 Torr, which resulted in eff = kg 1.74 108 ms . This value was much higher than we had originally anticipated when we moved the structure into a vacuum, but adding the parameter gave us one more knob to turn when we began ne tuning our device. Finally, we developed transfer functions relating the displacement and force placed on each structure, given by Equation 10. H(j) =
1 k ( o )2 + j Qo

+1

(10)

Here Q = o m . Q is a good parameter to describe the b overall sharpness of resonance, which dramatically increases in vacuum. The following is a plot of both the drive and sense resonance curves:
Fig. 5. Resonance
xf

0.15

0.10

0.05

Hz 44 500 45 000 45 500 46 000

For our implementation we found Qd = 234 and Qs = 173. This is very moderate for MEMS devices operated in vacuum. Gyros operating in the matched conguration need to be cautious moving over to the vacuum design because the extremely sharp resonances can have a dramatic effect on bandwidth. By balancing the devices Q and bandwidth values, we were able to nd a happy medium. The resonance curves well illustrate the reason we decided to design the structure such that the sense and drive resonances were nearly matched: we were able to produce maximal displacement for both the drive and sense structures, which would have very positive effects on our ARW and sensitivity

downstream, while minimizing the drop in bandwidth. This drop in bandwidth precisely when d = s was compensated for by slowly moving off the mode-matched case until the ARW and bandwidth specications were to taste. One can imagine that if s d , you can achieve very high bandwidth because youll be operating in the relatively at kx regime of the sense resonance curve while also struggling with very large ARW because of the decreased sensitivity. One nal consideration for the resonance of the structure was the negative electrostatic spring caused by the bias difference between the proof mass and its adjacent sense ngers. This effect will effectively decrease the mechanical k and thus shift the resonant frequency down, so we needed to analyse this effect and compensate accordingly. The expression for kel is stated in Equation 11. (11) d2 Even with the relatively large proof-mass bias we applied (150V), this effect was on the order of .059 N/m, and because the mechanical k was 1175 N/m, the effect was negligible yet still included in the complete analysis of the system. VIII. ACTUATION There are a few basic schemes used in MEMS devices to actuate structures. Thermal actuation can achieve extremely high forces at frequencies fast enough for use in a gyro, but the excessive forces are unneeded for such an application and not worth the sacrice in displacement range of the structure. Electrostatic actuation provided all the force required of our structure with certain distinct advantages: it could operate at high frequencies, act over large distances linearly, and needed only proper amplier design to implement electronically. Within electrostatic actuation there are two methods used, both leveraging the forces generated from a differential change in parallel place capacitance. Gap closing actuation can achieve very large forces but suffers two prominent drawbacks: a highly non-linear force as the structure resonates and the concern of pull-in, or when the plates get close enough to dynamically shut together, destroying your actuation. The preferred method of electro-statically actuating the gyro is via a comb-drive, which changes the parallel plate capacitance by altering their overlap area rather than their gap distance. As previously stated, this method is ideal because you can produce tremendous displacements highly linearly because the change in capacitive area is constant as the combs oscillate back and forth. The force generated by a comb nger can be seen in Equation 12.
2 2 1 o txV o tV CV 2 = = (12) F = U= 2 2d 2d Here, d = 2m is the gap distance between the combs, t = 2m is the comb thickness, and V is the applied voltage across the gap. This equation is valid only for 1 capacitor. In our nal comb-nger implementation, each comb is actuated from capacitors on both sides, we are using N combs on a side, and we are applying a voltage across the gap of the

from V = Vac + vdc . Further, we are driving our structure in a balanced comb conguration in which we are actuating from one side of the gyro and we have a matching comb drive, called the Drive Velocity, on the other side that is stabilizing the oscillation and ensuring we are operating at precisely the desired frequency. The nal equation relating applied voltages and force is stated in Equation 13 [1]. Fbalanced comb = 4
0 tN

Vdc vac

(13)

kel =

2 2 0 tlVdc

Our design t N = 114 ngers, each spaced d = 2m apart and t = 2m thick. We applied a Vdc = 150V and vac = 8.5V (both important parameters for adjusting ARW). Because we were applying such a high proof-mass DC bias we had to ensure we werent going to encounter any breakdown 1V with the air between combs. For air, Vbr nm , which can be approximated to 2000V for a 2 m gap, so we were operating well underneath this ceiling. To simplify future analysis, well dene the transfer function in Equation 14 relating force applied and ac voltage.
0 tN Vdc A=4 d

(14)

We were able to theoretically get 20m displacement of the drive element and 28nm displacement of the sense element (a good sign for having a high sensitivity). IX. S ENSING Sensing for our gyro was again accomplished via parallel plate capacitor congurations. In this case, we utilized gap closing sensing which allows us to get greater sensitivity compared to a comb implementation and also worked well with our design schematically because it allowed us to have the sensing ngers run the entire length of the mm scale proof mass. One thing to consider was the linearity of the sensing. Because we were using the gap closing conguration, our current sensed was not linear with sense displacement, yet because we displaced the sense element so slightly compared to the 2m gap, we well approximated the sensing via a rst order Taylor expansion. The linearity of the device will be discussed in depth in the FSR section. Equation 15 shows a derivation for the current output: Q = CVdc = = i(t) = Vdc dQ dC = Vdc dt dt (15)

dC dx dC = i(t) = Vdc x dx dt dx

Noting that the change in capacitance is now from the gap closing in, we can write dC d = dx dx
0A

(d xs )

0A

(xs

d)2

0A d2

(16)

Given x is the velocity of the proof-mass and that it is being driven harmonically, we can write it as x = xs where xs is the maximal displacement of the sense mass. Our nal equation for output current is thus

i(t) = Vdc

0 tlxs 0 (xs d)2

Vdc

0 tlxs d2

(17)

Writing the Taylor approximation of the current allows us to write the transfer function relating current output and sense displacement seen in Equation 18. At times in our analysis we utilized the transfer function and at others we used the full description. Generally, we found the approximation to be very good given the low displacements. We also needed to multiply this by the N ngers that will be be sensing. N 0 tl (18) I(j) = Vdc d2 Given our relatively large sense displacements of 28nmand long sensing ngers, we were able to get very large output current on the order of 1.55A. Figure 6 chart depicts the non-linearity of the sensing device well.
Fig. 6. Sense Mechanism Non-Linearity

to their counterparts. This was an important check given the relatively large bias applied across the sense ngers as well as the physical length of the ngers themselves. Equation 19 well describes the voltage at which such pull in occurs, and given our numbers this ended up being 216V. Vpull in = 8 ks d3 27 0 A (19)

The fact that we were operating safely under this value and that we were displacing so slightly compared to the gap allowed us to safely not concern ourselves with the threat of pull in. One additional advantage of our design is that it features differential sensing ngers, i.e., each nger has a sensing capacitor on either side. As we will discuss in the amplication section, this will give us an extra kick in sensitivity at the output of the amplier. X. A MPLIFICATION As discussed in the sensing section, our output from the mechanical system is a small current. Naturally, it makes sense to boost this signal and convert it to a voltage so that the gyro is readily integrable with other devices. The standard approach for amplifying a current to a voltage is using the trans-impedance op-amp conguration, which simply features a resistor dropped from the negative terminal to the output. If you wire the current output to the negative terminal you will get a gain proportional to the bridge resistor, as in Equation 20 Vout = Rf Iout (20)

Figure 7 also shows how sensitive the device is to various Coriolis acceleration and driving frequency.
Fig. 7. Gyro Sensitivity

Obviously, the op-amp golden rules dont always hold, but given we are inputting 1.55A, we can safely set Rf = 105 and get .155V at the output of the trans-impedance amp. However, recalling that we are using differential sensing ngers, well need two of these ampliers, one for the V+ signal and one for the V signal. At the output of the amps, we can sum this differential pair to give us twice the output. Finally, we need to demodulate our signal so we can pick our z value off at DC. By multiplying the output signal with another sinusoid operating at d we can achieve this result. The nal amplier circuit can be seen in Figure 8. XI. BANDWIDTH One of the most important specications of a gyroscope is its bandwidth. The bandwidth roughly describes what amount of dithering the device can undergo while still being able to detect its effects. Consider if the angular rotation rate of the gyro was of the following form: z = cos(in t) (21)

If we were to plug this into our expression for the Coriolis force, we would nd a modulated signal of the form: One nal important effect we checked for our rst order analysis was whether or not our sense ngers would pull in FCoriolis = 2mz v = 2mxd cos(in t) cos(d t) (22)

XII. N OISE AND A NGLE R ANDOM WALK The discussion of bandwidth leads us to yet another critical metric of a gyroscope: its noise. Noise caused by Brownian motion in the mechanical structure or non-ideal operation of signal ampliers can ruin the utility of a gyro if not properly understood, quantied, and optimized. Our design exhibited noise from four distinct sources: the mechanical structure itself, the input current and voltage noise of the amplier, and the resistance of the gain stage of the amplier. To analyse the noise produced by the mechanical structure, it is useful to nd an equivalent circuit model and then nd the models equivalent Johnson-Nyquist noise current. Equation 10 has an analogue in the electrical domain as described by Equation 23.
Fig. 8. Trans-Impendance Amplier Schematic

T (j) =

( 0 )2

jC j + Q0 + 1

(23)

Equation 22 contains the product of two sinusoids, which in the frequency domain will look like a carrier signal at d with sidebands representing the dithering input at = d in and = d + in . Thus, a good way to characterize bandwidth would be to observe how the relevant transfer functions of the system attenuate or boost these sidebands relative to the carrier. From this description it becomes immediately evident why their exists a trade off between large displacements and sensitivity in the sense direction and bandwidth: if we were operating in the kx regime of the sense structure, the curve is much atter so we would expect to have a greater acceptable sideband range compared to when we are operating very near the sharp resonant peak. Figure 9 depicts the relative magnitude of the right (top line) and left (bottom line) sidebands.
dB 3

w0 1 Here, 0 = LC and Q = RC . By pattern matching we can see this is very close to our mechanical system save for a pre-factor induced by the sensing elements displacement to current transfer function (Equation 18). To compensate this, we introduce a factor which is dened in Equation 24.

= Vdc

N 0 tl d2

(24)

As it happens, our value was 1.96104 . Next, we can use to nd the equivalent resistance, capacitance, and inductance of our structure. The mass maps to an inductance, spring constant to a capacitance, and damping to a resistance in the electrical domain. Thus we can dene the following three transforms: Lm Cm m 2 2 k b 2 (25) (26) (27)

Rm
Hz 20 1 40 60 80 100

Further, functions as the turn ratio of a transformer bridging the LRC circuit and current output. The full circuit diagram can be seen in Figure 10.

Fig. 9.

Gyroscope Sideband Attenuation/Boost

To dene the bandwidth of our gyro, we stipulated that neither sideband can fall outside of 1dB of the carrier. Further, we wanted to ensure that the effect within that range was fairly linear. Figure 9 also shows the sum of the two curves (middle line), and as we wanted, this hangs very near 0dB, so we felt good about the linearity of our device in this range. Calculating the 1dB intersection of the bottom line (the one that breaks the 1db limit rst) resulted in a bandwidth of 60Hz.

Fig. 10.

LRC Equivalent Circuit

With the equivalent circuit model extracted, we can extract our equivalent Johnson-Nyquist noise current via Equation 28.

7
iOut A

imech =

4kb T Rm

(28)

0.00003 0.000025 0.00002 0.000015 0.00001 5. 10


6

Similarly, the noise caused by the integrating resistor in the amplier stage can be calculated by this equation. Next, the amplier current noise can be calculated by Equation 29, where iamp noise = .01pA and v amp noise = 20nV are from the LM155 Op Amp datasheet, Rf is the integrating resistance, and f is the gyro bandwidth. iamp = i2 noise amp f
2 vamp noise 1 + 2 f Rf

1000

2000

3000

4000

5000

Omega deg s

(29)
Fig. 11. Gap Closing Non-Linearities

Finally, we can calculate the total current noise in the systems output by summing these noise sources and integrating over the device bandwidth. Equation 30. ieq = f imech + iint + iamp (30)

Our current noise was thus 46.7pA. It was then necessary to establish what kind of z input could cause such a current. If we knew this, we would know what rate of rotation would be indistinguishable from random noise. Running ieq back through Equation 18, we found our z, noise normalized over the root of the bandwidth to be .102 / Hr, which, as it happens, is the ARW of our gyro. Finding the right blend of parameters to produce this ARW with a broad bandwidth proved a challenge. The terms are highly coupled, so often times adjusting your structure for one will cause an opposite effect for the other. This is where having free parameters like pressure, spring length, Vdc , and vac came in so handy. The most effective way to shift weight around between bandwidth and ARW was to adjust exactly how mode-matched the structures were. We ended up nding a sweet spot between perfectly mode matching (and having almost no bandwidth but extremely small ARW) and being dramatically off resonance (and having bandwidth for days but an ARW around 500 / Hr). XIII. F ULL S CALE R ANGE Full scale range can be roughly dened as the range over which you can make a fairly linear mapping between your sensed current and what angular rate caused it. If your mapping becomes non-linear, it will pose a difcult problem for other devices to interpolate the data properly. Unfortunately, there are many sources that can corrupt the linearity of your sensing. The most noticeable for us was the non-linearity of the gap closing sense capacitors. As your input z becomes large, the 1st order approximation in Equation 18 becomes wildly inaccurate. To analyse the actual response, we plotted the output current vs. z for a few different d + in values to get a sense of a) the range of linearity for a given value and b) how the current will vary for sidebands. The result is show in Figure 11 Figure 11 shows a curve operating exactly at d (middle), -5700Hz offset (bottom), and +5700Hz offset (top). The next step in our evaluation was to see how linear these curves were. We dene our full scale range as the over which

the current doesnt deviate from a straight line by more than 10%. What we found, however, is that if this is a good way of dening full scale range, you can effectively double your FSR by applying a slightly devious linear interpolation. Figure 12 well demonstrates the trick.
Difference 20

15

10

100

200

300

400

500

Omega deg s

Fig. 12.

Linear Interpolation of Sensing

The line originating from the origin represents what could be thought of as a typical way to evaluate the data. It matches quite perfectly the slope of the current curve near the origin, and naturally starts to deviate the further we move away. The other curve that bounces off the x-axis however represents a line of slope somewhat less than that of the current curve near the origin, so as you move to the right it starts to approximate it better and better for a time until the functions intersect, afterwards deviating much as expected. If our denition of FSR is good, then this effectively doubles our range. Its worth mentioning however that this might complicate things if you wanted to estimate the condence in a result. In the traditional interpolation its obvious: the farther away from the origin you are the less accurate your measurement will be. With the tweaked interpolation your condence would technically increase to a point and then decrease afterwards. How the end user wants to apply the gyro might predicate what kind of interpolation of the data they want to use. In any event, the FSR from a traditional line is 251 /s and the FSR from the tweaked version is 503 /s. Another consideration for FSR could be hitting the amplier rails or disengaging the comb ngers. In the case of the former, we would need an z far beyond the linear range

of current output to generate a voltage high enough to clip (again, typical voltages will be around .155V). If we have two op-amp rails at 15V this will clearly not be a concern compared to the sensing non-linearity. In regards to the comb ngers disengaging, we designed our structure such that the ngers couldnt disengage. Again, because our design is ISOD, we had to design a sensing scheme such that the drive motion was decoupled from the sense motion. Our sensing scheme not only accomplishes this but removes any FSR issues relating to disengaging. XIV. T ESTING AND I MPLEMENTATION A. SPICE We were able to leverage several available software suites to simulate aspects of our gyroscope. First, we implemented our equivalent circuit model in SPICE to both double check our calculations and to analyse how the device operated with regard to noise specic to the gain stage, which SPICE happens to do very well. We simulated the frequency response of the structure for various ranges of frequencies and Coriolis forces, which was vital considering the nite-element tool COMSOL did not support the rotational motions required to simulate a gyroscope, so SPICE was our only venue to test the sensing mechanisms. Furthermore, SPICE simulation is a convenient alternative even for simple tasks such as verifying resonant frequency, since COMSOL is often limited by the complexity of the structure. A 3mm x 3mm structure with over 350 drive ngers and 100 inner sense ngers is well beyond what COMSOL could compute in a practical amount of time. Another benet to SPICE simulation was its ability to calculate the system noise that arose from internal impedances in the amplier. As noted, being assigned a strict set of specications for bandwidth and ARW demands accurate noise analysis to verify that the structures sense is not over-attenuated by the noise. In SPICE, the noise is computed by taking the RMS sum of all the Brownian noise contributors [5]: Vnoise =
i

range, the noise dominates the output signal. One of the great advantages to operating the gyro above 10 kHz is that it pushes your operation out to a regime of the amplier that is occupied by less noise. B. COSMSOL Simulating our design using COMSOL posed several unanticipated challenges. Due to the large size of our structure, even after we removed the actuating and sensing comb ngers from the design, COMSOL was still unable to simulate.

Fig. 14.

Drive Structure Resonance Curve Simulation

As such, we tested our design using a reduced structure with a proof mass of 1/3 the lateral dimensions (i.e. 1mm X 1mm), but springs of the same size. In order to obtain a more accurate simulation of resonance, we increased the inner mass density to 9 times the actual density of Polysilicon.

4kb T Ri

(31)

Fig. 15.

Sense Structure Resonance Curve Simulation

Fig. 13.

Electrical Equivalent SPICE Simulation

Noting Figure 13, it can be observed that at the resonant frequency the noise level is practically negligible compared to the output voltage (148 mV), while in the lower frequency

We believe that, as the springs and masses were identical to those of the intended design, this model provided good grounding for us to test our structure. Frequency sweeps in COMSOL derived a driving resonance frequency of around 45.4 kHz (expected 45.21 kHz) and a sensing resonance frequency of around 42.4 kHz (expected 45.29 kHz).

We were also able to run parametric sweeps on the structure to nd that the spring coefcient in the driving direction was 1786 N/m (1586 N/m expected) and the spring coefcient in the sensing direction was 818 N/m (1175 N/m expected). The signicant error in these results may be attributed to the fact that we assumed perfect rigidity of our structure. In reality, the frame may deform slightly, leading to different displacements, and hence, different spring coefcients. Figures 14 and 15 show the simulated frequency responses from the drive and sense elements, respectively. Another advantage of COMSOL was it allowed us to get a good 3D representation of the gyroscope to use in reports like this. One can quickly get an intuitive feel for the gyro by observing just a few of the images generated in COMSOL (See Figures 17 and 18).

XVI. D ESIGN & R ESULTS S UMMARY

TABLE I R ESULTS Parameter Bandwidth Angle Random Walk Tweaked FSR Traditional FSR Current Output Voltage Output Drive Resonant Frequency Sense Resonant Frequency Total Drive Displacement Total Sense Displacement Drive Spring Coefcient Sense Spring Coefcient Electric Spring Coefcient Drive Mass Sense Mass Drive Q Sense Q Effective Viscosity Equivalent Inductance Equivalent Resistance Equivalent Capacitance Mechanical Noise Integrator Noise LM155 Current Noise LM155 Voltage Noise Current Noise Floor Value 60Hz .102 / Hr 503 /s 251 /s 1.55A .155V 45.21 kHz 45.29 kHz 20m 28nm 1586 N/m 1175 N/m .059 N/m 5.15 nKg 0.145 nKg 234 173 1.129 105 Kg ms .377H 457 32.7pF 1.96104
A 3.63 1023
2

C. Cadence One nal piece we implemented for the gyro was a .gds le, which physically denes the structures and their respective process layers according to MUMPS. This is the le that can actually be sent off to a foundry for fabrication. Figure 16 shows a snippit of the .gds le.

1.66 1.66 6.66

Hz A2 1025 Hz A2 1030 Hz A2 1028 Hz

46.7pA

Fig. 16.

Snapshot of GDS Implementation File

TABLE II C HOSEN PARAMETERS Parameter Length of Drive Spring Length of Sense Spring Proof Mass DC Bias AC Drive Voltage Integrating Resistance Trans-Impedance Gain Pressure Value 76m 21m 150V 8.5V 105 105 100 Torr

XV. C ONCLUSION Our analysis was certainly rough, but we felt it gave us good rst-order approximations for how this gyroscope could function. In reality, we are greatly simplifying the real world by neglecting most fabrication imperfections, temperature issues, availability of 150V sources etc. All told, however, we feel good about the general philosophy of our gyroscope, and are condent that given a few design iterations and more robust simulation it could hold up relatively well in a real application. As stated in the Design & Results Summary, we managed to make our model t all required specs. Our ARW came in right on spec while being able to maximize bandwidth. Futher, by utilizing the relatively lot cost pricing model for MCNC, we were able to get a reasonably priced, high performance design.

ACKNOWLEDGMENT The authors would like to thank Professor Sunil Bhave for his instruction and guidance through ECE4320 and the nal project. Further, they would like to recognize all the help Siddharth Tallur and David Hutchison have been over the course of the term.

10

Fig. 17.

Full Scale Gyroscope Zoomed In

Fig. 18.

Full Scale Gyroscope

R EFERENCES
[1] C. Acar and A. Shkel, MEMS Vibratory Gyroscopes Structural Approaches to Improve Robustness New York, NY: Springer, 2009, pp. 102-103,116. [2] W. A. Clark, Micromachined Vibratory Rate Gyroscope, Ph.D thesis, UC Berkeley, 1997. [3] M. Palaniapan. Integrated Surface Micromachines Frame Gyroscopes, Ph.D thesis, U.C. Berkeley: 2002. [4] M. Palaniapan, R.T. Howe and J. Yasaitis, Performance Comparison of Integrated Z-axis Frame Microgyroscopes Proceedings of the IEEE MEMS03 Kyoto, Japan: January 2003, pp. 482-485. [5] Orcad Corporation, Orcad PSPICE Users Guide, First Edition Beaverton, OR: 1998

11

A PPENDIX Mathematica Code


ClearAll["Global*"] echo[x_, unit_] := Print[x, " = ", ToExpression[x], " ", unit] (**********************) (*Parameters*) (**********************) (*Physics + Global*) eSi = 170*109.; u0 = 1.78*10-5; (*damping coefficient for air*) e0 = 8.85*10-12; densitySi = 2330; t = 2*10-6.; w = 2*10-6; substrateOffset = 2*10-6.; proofMassDC = 150; kB = 1.3806503*10-23.; temp = 300.; pressure = 100; (*Structure*) spacing = 20*10-6.; smallW = 10*10-6.; bigW = 100*10-6.; outsideY = 68*10-6; dimY = 3*10-3; iW = 14*10-6; cornerWidth = 12*10-6; bigL = dimY - 2*outsideY; bigN = 114; dimX = 2*(6*10-6 + cornerWidth) + bigN*(smallW + iW) + smallW + 2*bigW; smallL = dimY - 2*(bigW + outsideY + 2*10-6); (*Primary Springs*) lPrimary = 76*10-6; wPrimary = 4*10-6; (*Secondary Springs*) lSecondary = 21*10-6; (*Actuating*) vACactuator = 8.5; nActuatorFingers = Floor[(dimX - 2*(lPrimary + 2*10-6))/(8*10-6)];(*on side*) tActuatorFingers = 2*10-6; dActuatorFingers = 2*10-6; (*Sensing*) nSensingFingers = bigN; (*in total*) tSensingFingers = 2*10-6; dSensingFingers = 2*10-6; lSensingFingers = smallL - 2*(smallW + spacing + 4*10-6); (*Gain*) rF = 2*105; (*Noise*) opAmpCurrent = .01*10-12; opAmpVoltage = 20*10-9; (*BW*) bw = 60; (*Max out this spec*)

12

(**********************) (*Calculations*) (**********************) (*Primary Springs*) iPrimary = t*wPrimary3/12; kPrimary = 4*96*eSi*iPrimary/((lPrimary)3); echo["kPrimary", "N/m"] (*Secondary Springs*) iSecondary = t*w3/12; kSecondaryMech = 4*12*eSi*iSecondary/((lSecondary)3); kElectric = bigN*2*e0*tSensingFingers*lSensingFingers* proofMassDC2/(dSensingFingers)2; kSecondary = kSecondaryMech - kElectric; echo["kElectric", "N/m"] echo["kSecondary", "N/m"] (*Outer Mass*) aOuterMass = (2*bigW + smallL + 4*10-6)2 - (smallL + 4*10-6)2; outerMass = densitySi*t*aOuterMass; echo["outerMass", "kg"] (*Inner Mass*) aInnerMass = (bigN + 1)*smallL*smallW + 28*10-6*smallW*bigN + 8*10-6 (smallL - 2*10-6); innerMass = densitySi*t*aInnerMass; echo["innerMass", "kg"] (*Damping from Book*) kN = 5.1*10-5/(pressure*substrateOffset); uEff = u0/(1 + 2*kN + .2*kN.788*Exp[-kN/10]); echo["uEff", "kg/(ms)"] (*Resonant Frequency*) wD0 = Sqrt[kPrimary/(innerMass + outerMass)]; (*Resonant Frequency of entire structure*) wS0 = Sqrt[kSecondary/ innerMass]; (*Resonant Frequency of inner mass*) fD0 = wD0/(2*Pi); fS0 = wS0/(2*Pi); echo["fD0", "Hz"] echo["fS0", "Hz"] b = uEff*(aInnerMass + aOuterMass)/substrateOffset; bSense = uEff*(aInnerMass)/substrateOffset; echo["b", "Ns/m"] echo["bSense", "Ns/m"] qStructure = wD0*(innerMass + outerMass)/b; qSense = wS0*(innerMass)/b; echo["qStructure", ""] echo["qSense", ""] resonanceCurve[ w_] := (1/kPrimary)/(-(w/(wD0))2 + I*w/(qStructure*wD0) + 1) senseCurve[w_] := (1/kSecondary)/(-(w/wS0)2 + I*w/(qSense*wS0) + 1) (*Actuating*) actuatingCurve[w_] := 4*proofMassDC*vACactuator*e0*tActuatorFingers* nActuatorFingers/(dActuatorFingers); echo["actuatingCurve[w]", "N/v"]

13

(*Displacement*) totalDisplacement[w_] := resonanceCurve[w]*actuatingCurve[w]*vACactuator; totalD = Abs[totalDisplacement[wD0]]*106; echo["totalD", "um"] (*Physics*) fCor[w_] := 2*(outerMass + innerMass)*totalDisplacement[w]*w (*Fc/Omega*) (*Sensing*) iOutFSR[w_, xs_] := e0*nSensingFingers*tSensingFingers*lSensingFingers*w* proofMassDC/(xs - dSensingFingers)2*xs; (*i*) iOut[w_] := e0*nSensingFingers*tSensingFingers*lSensingFingers*w* proofMassDC/(dSensingFingers)2 (*i/xs*) (*Sensing \ Displacement/omega*) senseDisplacement[w_] := iOut[w]*senseCurve[w]*fCor[wD0] (*i/Omega*) totalS = Abs[senseCurve[wD0]*fCor[wD0]]*109; echo["totalS", "nm"] totalI = Abs[senseCurve[wD0]*fCor[wD0]*iOut[wD0]]*109; echo["totalI", "nA"] (*Output Current vs. Omega for FSR*) output[w_, omega_] := iOutFSR[w, fCor[w]*senseCurve[w]*omega] (*Bandwidth*) dBRight[fIn_] := -10* Log10[Abs[senseCurve[fD0*2*Pi]]/ Abs[senseCurve[(fD0 + fIn)*2*Pi]]]; dBLeft[fIn_] := -10* Log10[Abs[senseCurve[fD0*2*Pi]]/Abs[senseCurve[(fD0 - fIn)*2*Pi]]]; (*FSR*) outputFSR[omega_, wSide_] := iOutFSR[wD0 + wSide, fCor[wD0]*senseCurve[wD0 + wSide]*omega] percentDifferenceFSR[omega_, wSide_] := Abs[(outputFSR[omega, 0] outputFSR[omega, wSide])/((outputFSR[omega, 0] + outputFSR[omega, wSide])/2)]*100 (*LRC Model*) eta = nSensingFingers*e0*tSensingFingers*(lSensingFingers)* proofMassDC/(dSensingFingers)2; L = innerMass/eta2; Cp = eta2/kSecondary; R = bSense/eta2 ; wModel = 1./Sqrt[L*Cp]; fModel = wModel/(2. Pi); qModel = 1./(wModel*R*Cp); echo["L", "H"] echo["R", "Ohms"] echo["Cp", "F"] echo["eta", ""] LRCModelCurve[w_] := I*w*Cp/(-(w/wModel)2 + I*w/(qModel*wModel) + 1) *eta

14

(*Noise*) ampNoise = 4*kB*temp/(rF); iAmpNoise = opAmpCurrent2/(bw); vAmpNoise = opAmpVoltage2/(bw*rF2)*1.; mechNoise = 4*kB*temp/(R); iEq = Sqrt[bw*(ampNoise + iAmpNoise + vAmpNoise + mechNoise)]; echo["ampNoise", "A2/Hz"] echo["iAmpNoise", "A2/Hz"] echo["vAmpNoise", "A2/Hz"] echo["mechNoise", "A2/Hz"] echo["iEq", "A"] Off[Solve::ifun]; Solve[iEq == Abs[iOutFSR[wD0, senseCurve[wD0]*fCor[wD0]*omega]], omega]; inputRef = Abs[omega /. %[[3]]]; (*inputRef=Abs[iEq/senseDisplacement[wD0]];*) ARW = inputRef/10*360/2 Pi*60 (*deg/roothour*); echo["ARW", "deg/Sqrt(Hr)"] (*Amplifier*) vOut[w_] := Abs[senseCurve[wD0]*fCor[wD0]*iOut[wD0]]*rF; echo["vOut[wD0]", "V"] tweakpercent[omegaD_] := 100*Abs[Abs[outputFSR[omegaD*2*Pi/360, 0]] 1.405*10-6*omegaD*2* Pi/360]/((Abs[outputFSR[omegaD*2*Pi/360, 0]] + 1.405*10-6*omegaD*2*Pi/360)/2) tPercent = tweakpercent[503]; FSRTweaked = 503; classicpercent[omegaD_] := 100*Abs[Abs[outputFSR[omegaD*2*Pi/360, 0]] 1.55*10-6*omegaD*2* Pi/360]/((Abs[outputFSR[omegaD*2*Pi/360, 0]] + 1.55*10-6*omegaD*2*Pi/360)/2) cPercent = classicpercent[251]; FSRClassic = 251;

echo["FSRTweaked", "deg/s"] echo["FSRClassic", "deg/s"]

(*Plots!*) Plot3D[iOutFSR[w, xs], {w, 0, 3*fS0}, {xs, 0, 1*10-3}, AxesLabel -> {"f[Hz]", "xs[m]", "iOut[A]"}] Plot[{Abs[outputFSR[omegaD*2*Pi/360, 0]], Abs[outputFSR[omegaD*2*Pi/360, 100]], Abs[outputFSR[omegaD*2*Pi/360, -100]]}, {omegaD, 0, 5000}, AxesLabel -> {"Omega(deg/s)", "iOut(A)"}] Plot[{100* Abs[Abs[outputFSR[omegaD*2*Pi/360, 0]] 1.405*10-6*omegaD*2* Pi/360]/((Abs[outputFSR[omegaD*2*Pi/360, 0]] + 1.405*10-6*omegaD*2*Pi/360)/2), 100*Abs[Abs[outputFSR[omegaD*2*Pi/360, 0]] 1.55*10-6*omegaD*2* Pi/360]/((Abs[outputFSR[omegaD*2*Pi/360, 0]] + 1.55*10-6*omegaD*2*Pi/360)/2), 10, 0}, {omegaD, 0, 500}, AxesLabel -> {"Omega(deg/s)", "% Difference"}, Filling -> {3 -> {4}}, FillingStyle -> Directive[Opacity[0.5], Red]] Plot3D[Abs[ iOutFSR[2*Pi*f, Abs[fCor[wD0]*senseCurve[2*Pi*f]*omega]]], {f, 0, 3*wD0/(2 Pi)}, {omega, 0, 10}, AxesLabel -> {"Hz", "Omega", "iOut"}, PlotRange -> {0, 10-6}]

15

Plot[{Abs[resonanceCurve[2*Pi*f]], Abs[senseCurve[2*Pi*f]]}, {f, wD0/(2 Pi) - 1000, wD0/(2 Pi) + 1000}, AxesLabel -> {"Hz", "x/f"}, PlotRange -> {0, .17}] Plot[Abs[senseDisplacement[2*Pi*f]], {f, 0, 25000}, PlotLabel -> "Sense Displacement vs. Driving Force Magnitude", AxesLabel -> {"Hz", "m"}] Plot[{dBRight[fIn], dBLeft[fIn], 1, -1, dBRight[fIn] + dBLeft[fIn]}, {fIn, 0, 100}, AxesLabel -> {"Hz", "dB"}, Filling -> {3 -> {4}}, FillingStyle -> Directive[Opacity[0.5], Orange], PlotRange -> {-3, 3}] Plot[Abs[LRCModelCurve[2*Pi*f]], {f, 0, 2.5*wD0/(2*Pi)}, PlotLabel -> "LRC Model Response Magnitude", AxesLabel -> {"Hz", "i/fc(v)"}] Plot[Abs[iOut[2*Pi*f]*senseCurve[2*Pi*f]], {f, 0, 2.5*wD0/(2*Pi)}, PlotLabel -> "LRC Model Response Magnitude", AxesLabel -> {"Hz", "i/fc(v)"}, PlotRange -> {0, 10}]