MEMS Reference Oscillators - RFICrfic.eecs.berkeley.edu/ee242/pdf/Module_7_1_MEMS_XTAL.pdf · LIN...

MEMS Reference Oscillators

EECS 242B Fall 2014

Prof. Ali M. Niknejad

Ali M. Niknejad University of California, Berkeley EECS 242B, Slide:

Why replace XTAL Resonators?

• XTAL resonators have excellent performance in terms of quality factor (Q ~ 100,000), temperature stability (< 1 ppm/C), and good power handling capability (more on this later)

• The only downside is that these devices are bulky and thick, and many emerging applications require much smaller form factors, especially in thickness (flexible electronics is a good example)

• MEMS resonators have also demonstrated high Q and Si integration (very small size) ... are they the solution we seek?

• Wireless communication specs are very difficult: • GSM requires -130 dBc/Hz at 1 kHz from a 13 MHz oscillator

• -150 dBc/Hz for far away offsets2


Business Opportunity

• XTAL oscillators is a $4B market. Even capturing a small chunk of this pie is a lot of money.

• This has propelled many start-ups into this arena (SiTime, SiClocks, Discera) as well as new approaches to the problem (compensated LC oscillators) by companies such as Mobius and Silicon Labs

• Another observation is that many products in the market are programmable oscillators/timing chips that include the PLL in the package.

• As we shall see, a MEMS resonator does not make sense in a stand-alone application (temp stability), but if an all Si MEMS based PLL chip can be realized, it can compete in this segment of the market

3


Series Resonant Oscillator• The motional resistance of

MEMS resonators is quite large (typically koms compared to ohms for XTAL) and depends on the fourth power of gap spacing

• This limits the power handling capability

• Also, in order not to de-Q the tank, an amplifier with low input/output impedance is required. A trans-resistance amplifier is often used

4

2478 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 39, NO. 12, DECEMBER 2004

Fig. 1. General topology for a series-resonant oscillator.

Unfortunately, a recently demonstrated 10-MHz oscillatorusing a variant of the above CC-beam resonator togetherwith an off-the-shelf amplifier exhibits a phase noise of only

80 dBc/Hz at 1-kHz carrier offset, and 116 dBc/Hz atfar-from-carrier offsets [10]—inadequate values caused mainlyby the insufficient power-handling ability of the CC-beammicromechanical resonator device used [22].

This work demonstrates the impact of micromechanicalresonator power handling and on oscillator performanceby assessing the performance of an oscillator comprised ofa custom-designed sustaining transresistance amplifier ICtogether with three different capacitively transduced microme-chanical resonators: a 10-MHz CC-beam similar to that in[10], a 10-MHz wide-width CC-beam with a 5 wider widthfor higher power handling; and a 60-MHz wine glass diskresonator with a 10 higher power handling capability thanthe wide-width CC-beam, a comparable series motional re-sistance, and a 45 higher of 48 000. The combinationof the wine glass micromechanical disk resonator with thecustom-IC CMOS transresistance sustaining amplifier yieldsa 60-MHz reference oscillator that achieves a phase noisedensity of 110 dBc/Hz at 1-kHz offset from the carrier, and

132 dBc/Hz at far-from-carrier offsets, while consumingonly 350 W of amplifier power, which is substantially lowerthan the milliwatts typically needed by the quartz crystal os-cillators presently used in cellular telephones. Dividing downto 10 MHz for fair comparison, these values correspond to

125 dBc/Hz at 1-kHz offset from a 10-MHz carrier, and aneffective 147 dBc/Hz far-from-carrier value. These valuesnearly satisfy or already satisfy (for some renditions) thereference oscillator specifications for wireless handsets. This,together with the potential for single-chip integration via any ofseveral already demonstrated processes [21], [23], [24], [25],makes the micromechanical resonator oscillator of this workan attractive on-chip replacement for quartz crystal referenceoscillators in communications and other applications.

II. OSCILLATOR DESIGN TRADEOFFS: VERSUS

POWER HANDLING

Fig. 1 presents the top-level schematic of the oscillator circuitused in this work, where the micromechanical resonator is rep-resented by an equivalent series LCR circuit to be specified later.As shown, a series resonant configuration is used, employing a

transresistance sustaining amplifier in order to better accommo-date the somewhat large motional resistance , on the orderof several k , exhibited by some of the micromechanical res-onators to be used. Here, the use of a transresistance amplifierwith small input and output resistances effectively imposes asmaller load on the series resonator, allowing it to retain its verylarge . The conditions required for sustained oscillation canbe stated as follows.

1) The gain of the transresistance sustaining amplifiershould be larger than the sum of the micromechanical res-onator’s motional resistance, plus the input and output re-sistances of the sustaining amplifier, and any other sourcesof loss in the feedback loop, i.e.,

(1)

where is the total resistance that can consume powerwithin the oscillation loop. In essence, this criterion statesthat the loop gain must be greaterthan 1. At start-up, a loop gain of 3 or greater is recom-mended to insure oscillation growth, even in the face ofprocess variations.

2) The total phase shift around the closed positive feedbackloop must be 0 . In this series-resonant topology, an idealsituation exists when both the micromechanical resonatorand transresistance sustaining amplifier have 0 phaseshifts from input to output. In practice, there will bea finite amplifier phase shift, which can be minimizedby choosing the amplifier bandwidth to be at least 10greater than the oscillation frequency.

As the oscillation amplitude builds, nonlinearity in either thesustaining amplifier or the resonator tank reduces orraises , respectively, until the loop gain equals unity, atwhich point the amplitude no longer grows and steady-stateoscillation ensues. As will be seen, unlike quartz crystal os-cillators, where the oscillation amplitude usually limits viaamplifier nonlinearity, surface-micromachined micromechan-ical resonator oscillators without automatic level control (ALC)circuitry generally limit via nonlinearity in the micromechan-ical resonator [21].

Since the main function of a reference oscillator is to provideone and only one output frequency , the output sinusoid ofFig. 1 should ideally be a delta function in the frequency do-main, as shown in the figure. Given that an oscillator’s outputcan be fundamentally modeled as noise filtered by an extremelyhigh- filter, where the of the resonator tank is effectivelyamplified in the positive feedback loop to a value often morethan 10 times its stand-alone value [26], the output can bevery close to a delta function if the initial resonator is large.However, since the resonator is not infinite, there will stillbe output power at frequencies adjacent to . If the power at

(and only at ) is considered the desired output, then thisadjacent “sideband” power is considered unwanted noise. If theoscillator amplitude is constant (which is generally true for thecase of hard limiting or ALC), then amplitude noise is largelyremoved, and half of the total original noise power remains asphase noise [26].

Ramp Rx + Ri + Ro = Rtot

LIN et al.: SERIES-RESONANT VHF MICROMECHANICAL RESONATOR REFERENCE OSCILLATORS, IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 39, NO. 12, DECEMBER 2004 !


Zero’th Order Leeson Model

• Using a simple Leeson model, the above expression for phase noise is easily derived.

• The insight is that while MEMS resonators have excellent Q’s, their power handling capability will ultimately limit the performance.

• Typically MEMS resonators amp limit based on the non-linearity of the resonator rather than the electronic non-linearities, limiting the amplitude of the oscillator

5

L fm =2kT (1 + FRamp)

Po·

Rtot

Rx

·1 +

f0

2Ql · fm

2

Ql =Rx

Rx + Ri + RoQ =

Rx

RtotQ



MEMS Resonator Designs

• Clampled-clamped beam and wine disk resonator are very populator. Equivalent circuits calculated from electromechanical properties.

• Structures can be fabricated from polysilicon (typical dimensions are small ~ 10 um)

• Electrostatic transduction is used (which requires large voltages > 10 V).

6


TABLE IRESONATOR DESIGN EQUATION SUMMARY

Fig. 2. Perspective view schematic and equivalent circuit of a CC-beammicromechanical resonator under a one-port bias and excitation scheme.

is the radian resonance frequency, all other variables are spec-ified in Fig. 2, and an approximate form for has been

used (with the more accurate, but cumbersome, form givenin Table I). As the frequency of varies through the beamresonance frequency, the output motional current magnitudetraces out a bandpass biquad frequency spectrum identical tothat exhibited by an LCR circuit, but with a much higher thannormally achievable by room temperature electrical circuits.Fig. 3 presents the SEM and measured frequency characteristic(under vacuum) for an 8- m-wide, 20- m-wide-electrode,10-MHz CC-beam, showing a measured of 3100.

The values of the motional elements in the equivalent circuitof Fig. 2 are governed by the mass and stiffness of the resonator,and by the magnitude of electromechanical coupling at its trans-ducer electrodes. Equations for the elements can be derived bydetermining the effective impedance seen looking into the res-onator port [5], and can be summarized as

(6)



CC-Beam Resonator

• This example uses an 8-μm wide beamwidth and a 20-μm wide electrode.

• Measurements are performed in vacuum.

• Q ~ 3000 for a frequency of 10 MHz

7


TABLE IRESONATOR DESIGN EQUATION SUMMARY

Fig. 2. Perspective view schematic and equivalent circuit of a CC-beammicromechanical resonator under a one-port bias and excitation scheme.

is the radian resonance frequency, all other variables are spec-ified in Fig. 2, and an approximate form for has been

used (with the more accurate, but cumbersome, form givenin Table I). As the frequency of varies through the beamresonance frequency, the output motional current magnitudetraces out a bandpass biquad frequency spectrum identical tothat exhibited by an LCR circuit, but with a much higher thannormally achievable by room temperature electrical circuits.Fig. 3 presents the SEM and measured frequency characteristic(under vacuum) for an 8- m-wide, 20- m-wide-electrode,10-MHz CC-beam, showing a measured of 3100.

The values of the motional elements in the equivalent circuitof Fig. 2 are governed by the mass and stiffness of the resonator,and by the magnitude of electromechanical coupling at its trans-ducer electrodes. Equations for the elements can be derived bydetermining the effective impedance seen looking into the res-onator port [5], and can be summarized as

(6)

LIN et al.: SERIES-RESONANT VHF MICROMECHANICAL RESONATOR REFERENCE OSCILLATORS 2481

Fig. 3. (a) SEM and (b) frequency characteristic (measured under 20-mtorrvacuum) for a fabricated CC-beam micromechanical resonator with an8- m-wide beamwidth and a 20- m-wide electrode.

where and are the effective stiffness and massof the resonator beam, respectively, at its midpoint, both givenin Table I, and is the electromechanical coupling factor. Thecapacitor represents the static overlap capacitance betweenthe input electrode and the structure.

Of the elements in the equivalent circuit, the series motionalresistance is perhaps the most important for oscillatordesign, since it governs the relationship between andat resonance, and thereby directly influences the loop gain ofthe oscillator system. For the CC-beam resonator of Fig. 2,the expression for can be further specified approximately(neglecting beam bending and distributed stiffness [5]) as

(7)

which clearly indicates a strong fourth-power dependence onthe electrode-to-resonator gap spacing, as well as a square-lawdependence on the dc bias voltage .

As already mentioned, the power-handling ability of a mi-cromechanical resonator greatly influences the phase-noise per-formance of any oscillator referenced to it. For the case ofan oscillator application, the power handling limit of a mi-cromechanical resonator is perhaps best specified by the powerrunning through it when it vibrates at a maximum acceptableamplitude , which is set either by ALC, or bythe pull-in limit (for which at resonance). Using(5) and (7), the maximum power that a resonator canhandle can then be expressed approximately by [22]

(8)

where is the effective stiffness of the resonator at its midpoint.Equation (8) asserts that higher power handling can be attainedwith larger values of stiffness and electrode-to-resonator gapspacing .

The basic CC-beam resonator used in this work is 40 mlong, 2 m thick, and 8 m wide, and with these dimensions,can handle a maximum power of only 39.7 dBm ( 0.11 W)when m, V, and nm. In addi-tion, its series motional resistance under these same conditions

Fig. 4. (a) SEM and (b) frequency characteristic (measured under 20-mtorrvacuum) for a fabricated wide-width CC-beam micromechanical resonator,featuring large beam and electrode widths for lower and higherpower-handling ability. Note that the difference in frequency from that ofFig. 3 comes mainly from the larger electrical stiffness caused by a higherdc-bias and a larger electrode-to-beam overlap.

is 8.27 k , which is quite large compared with the 50 nor-mally exhibited by quartz crystals, and which complicates thedesign of the sustaining amplifier for an oscillator application.

B. Wide-Width CC-Beam Resonator

One convenient method for reducing and increasingpower handling is to widen the width of the CC-beam [11].For example, the width of the previous CC-beam can beincreased from 8 to 40 m, while also increasing the elec-trode width from 20 to 32 m, all without appreciablychanging the resonance frequency, which to first order doesnot depend upon and . Equation (7) predicts that anincrease in beamwidth leads to a smaller , mainly dueto a larger electrode-to-resonator overlap area that increaseselectromechanical coupling via the capacitive transducer. This,together with increasing the electrode width to furtherincrease transducer capacitance, comprises the basic strategyused for the wide-width CC-beam to achieve an smallenough to allow the use of a single-stage sustaining amplifier(to be described later).

Increasing also seems to lower the and increase theeffective stiffness of the CC-beam, and according to (7),this reduces the degree to which is lowered. To illustrate,Fig. 4 presents the SEM and measured frequency character-istic (under vacuum) for a 40- m-wide, 32- m-wide-electrode10-MHz CC-beam, showing a of 1036, which is 3 lowerthan exhibited by previous 8- m-wide beams. This reductionin with increasing beamwidth is believed to arise from in-creased energy loss through the anchors to the substrate, causedby increases in CC-beam stiffness and in the size of its anchors.A direct consequence of the increase in stiffness, governed by(T 1.5) and (T 1.9), is that the beam pumps harder on its anchorswhile vibrating, thereby radiating (i.e., losing) more energy percycle into the substrate. The increase in anchor size further ex-acerbates this energy loss mechanism by creating a better match



CC-Beam with Better Power Handling

• To increase power handling of the resonator, a wider beam width is used [~10X in theory].

• The motional resistance is reduced to 340 ohms (Vp = 13V)

8


Fig. 3. (a) SEM and (b) frequency characteristic (measured under 20-mtorrvacuum) for a fabricated CC-beam micromechanical resonator with an8- m-wide beamwidth and a 20- m-wide electrode.

where and are the effective stiffness and massof the resonator beam, respectively, at its midpoint, both givenin Table I, and is the electromechanical coupling factor. Thecapacitor represents the static overlap capacitance betweenthe input electrode and the structure.

Of the elements in the equivalent circuit, the series motionalresistance is perhaps the most important for oscillatordesign, since it governs the relationship between andat resonance, and thereby directly influences the loop gain ofthe oscillator system. For the CC-beam resonator of Fig. 2,the expression for can be further specified approximately(neglecting beam bending and distributed stiffness [5]) as

(7)

which clearly indicates a strong fourth-power dependence onthe electrode-to-resonator gap spacing, as well as a square-lawdependence on the dc bias voltage .

As already mentioned, the power-handling ability of a mi-cromechanical resonator greatly influences the phase-noise per-formance of any oscillator referenced to it. For the case ofan oscillator application, the power handling limit of a mi-cromechanical resonator is perhaps best specified by the powerrunning through it when it vibrates at a maximum acceptableamplitude , which is set either by ALC, or bythe pull-in limit (for which at resonance). Using(5) and (7), the maximum power that a resonator canhandle can then be expressed approximately by [22]

(8)

where is the effective stiffness of the resonator at its midpoint.Equation (8) asserts that higher power handling can be attainedwith larger values of stiffness and electrode-to-resonator gapspacing .

The basic CC-beam resonator used in this work is 40 mlong, 2 m thick, and 8 m wide, and with these dimensions,can handle a maximum power of only 39.7 dBm ( 0.11 W)when m, V, and nm. In addi-tion, its series motional resistance under these same conditions

Fig. 4. (a) SEM and (b) frequency characteristic (measured under 20-mtorrvacuum) for a fabricated wide-width CC-beam micromechanical resonator,featuring large beam and electrode widths for lower and higherpower-handling ability. Note that the difference in frequency from that ofFig. 3 comes mainly from the larger electrical stiffness caused by a higherdc-bias and a larger electrode-to-beam overlap.

is 8.27 k , which is quite large compared with the 50 nor-mally exhibited by quartz crystals, and which complicates thedesign of the sustaining amplifier for an oscillator application.

B. Wide-Width CC-Beam Resonator

One convenient method for reducing and increasingpower handling is to widen the width of the CC-beam [11].For example, the width of the previous CC-beam can beincreased from 8 to 40 m, while also increasing the elec-trode width from 20 to 32 m, all without appreciablychanging the resonance frequency, which to first order doesnot depend upon and . Equation (7) predicts that anincrease in beamwidth leads to a smaller , mainly dueto a larger electrode-to-resonator overlap area that increaseselectromechanical coupling via the capacitive transducer. This,together with increasing the electrode width to furtherincrease transducer capacitance, comprises the basic strategyused for the wide-width CC-beam to achieve an smallenough to allow the use of a single-stage sustaining amplifier(to be described later).

Increasing also seems to lower the and increase theeffective stiffness of the CC-beam, and according to (7),this reduces the degree to which is lowered. To illustrate,Fig. 4 presents the SEM and measured frequency character-istic (under vacuum) for a 40- m-wide, 32- m-wide-electrode10-MHz CC-beam, showing a of 1036, which is 3 lowerthan exhibited by previous 8- m-wide beams. This reductionin with increasing beamwidth is believed to arise from in-creased energy loss through the anchors to the substrate, causedby increases in CC-beam stiffness and in the size of its anchors.A direct consequence of the increase in stiffness, governed by(T 1.5) and (T 1.9), is that the beam pumps harder on its anchorswhile vibrating, thereby radiating (i.e., losing) more energy percycle into the substrate. The increase in anchor size further ex-acerbates this energy loss mechanism by creating a better match



Disk Wineglass Resonator

• Intrinsically better power handling capability from a wine glass resonator.

• The input/output ports are isolated (actuation versus sensing).

9


Fig. 5. Measured plots of (a) and (b) series motional resistance , versusbeamwidth for V and 13 V (where m, m,

are fixed), showing a net decrease in , despite a decrease in.

between the resonator and the substrate, allowing for more effi-cient energy transfer between the two, hence, lower .

Fortunately, this decrease in is not enough to preventdecrease, and in fact, still decreases as beamwidth increases.Fig. 5(a) and (b) presents measured plots of and versusbeamwidth , showing that despite reductions in , the neteffect of beam widening is still a decrease in . In particular,beam widening has reduced the of a CC-beam resonatorfrom the 17.5 k (with V) of the 8- m-wide devicedescribed above [10], to now only 340 (with V),which can be extracted from the height of the frequency spec-trum shown in Fig. 4(b). It should also be noted that althoughthe of the CC-beam resonator drops from 3100 to 1036 as aresult of a wider beamwidth, this is still more than two or-ders of magnitude larger than that achievable by on-chip spiralinductors, and is still suitable for many wireless applications.

In addition to a lower , and perhaps more importantly, theuse of a wide-width CC-beam provides the additional advantageof a larger power handling ability, which comes about due toincreased stiffness . In particular, the stiffness of the 40- m-wide beam used here is 9240 N/m, which is 4.97 larger thanthe 1860 N/m of the previous 8- m-wide CC-beam, and whichaccording to (8), when accounting for decreased , provides anet 9.06 higher oscillator output power. For the same

V, at which s of 3100 and 1700 are exhibited by the 8- m-wide and 40- m-wide CC-beams, respectively, this results in a9.57 dB lower far-from-carrier phase-noise floor for the 40- m-wide beam.

C. Wine Glass Disk Resonator

Even after widening their beam and electrode widths, thepower handling ability of stand-alone CC-beams is still not suf-ficient for some of the more stringent specifications, such as forthe far-from-carrier phase noise in the GSM standard. In addi-tion, for reference oscillator applications, CC-beam microme-chanical resonators have a frequency range practically limitedto less than 30 MHz, since their s drop dramatically as theirbeam lengths shrink. In particular, the increase in anchor losseswith beam stiffness described above become enormous as fre-quencies rise, to the point where the at 70 MHz drops to only300 [28]. One solution to this problem that retains the simplicityof a beam resonator is to support the beam not at its ends, butat nodal locations, and design the supports so that anchor lossesthrough the supports are minimized when the beam vibrates in

Fig. 6. (a) Perspective-view schematic of a micromechanical wine glass-modedisk resonator in a typical two-port bias and excitation configuration. Here,electrodes labeled A are connected to one another, as are electrodes labeledB. (b) Wine-glass mode shape simulated via finite element analysis (usingANSYS). (c) Equivalent LCR circuit model.

a free–free mode shape. Free–free beam micromechanical res-onators have been successfully demonstrated, one with a fre-quency of 92 MHz and a of 7450 [29].

Even better performance, however, can be obtained byabandoning the beam geometry and moving to a disk geom-etry. In particular, radial-mode disk resonators have recentlybeen demonstrated with s 10 000 at frequencies exceeding1.5 GHz, even when operating in air [18]. wine-glass-modedisks have now been demonstrated at 60 MHz with s on theorder of 145 000 [6], which is the highest yet reported for anyon-chip resonator in the frequency range needed for referenceoscillators. This, together with the much larger stiffness (hence,higher power handling) of a wine glass disk relative to thebeam-based counterparts previously described, inspires the useof a wine glass disk for oscillator applications.

As shown in Fig. 6(a), the wine glass disk resonator of thiswork consists of a 3- m-thick disk floating above the substrate,supported by two beams that attach to the disk at quasi-nodalpoints [6], which have negligible displacement compared toother parts of the disk structure when the disk vibrates inits wine glass mode shape, shown in Fig. 6(b). In this modeshape, the disk expands along one axis and contracts in theorthogonal axis. Four electrodes surround the disk with a lateralelectrode-to-disk gap spacing of only 80 nm, made tiny tomaximize capacitive transducer coupling governed by in (6).Opposing electrodes along a given axis are connected in pairsas shown in Fig. 6(a) to affect a drive forcing configurationalong the -axis that compresses and expands the disk with themode shape of Fig. 6(b), while sensing an oppositely directedmotion along the sense axis . In this configuration, if the -and -axis electrodes are identically sized, current enteringwhere the disk compresses (i.e., where in (T 1.16) isnegative), leaves where the disk expands (i.e., where in


(T 1.16) is positive), giving this device a pass-through natureat resonance with a 0 phase shift from the -axis (input)electrode to the -axis (output) electrode.

The two-port nature of this device whereby the input andoutput electrodes are physically distinct from the resonator it-self further allows a bias and excitation configuration devoid ofthe bias tee needed in Fig. 2, hence, much more amenable toon-chip integration. In particular, the applied voltages still con-sist of a dc bias voltage and an ac input signal , but now

can be directly applied to the resonator itself without theneed for a bias tee to separate ac and dc components. Similarto the CC-beam, these voltages result in a force proportional tothe product that drives the resonator into the wine glass vi-bration mode shape when the frequency of matches the wineglass resonance frequency, given by [30]

(9)

where

(10)

and where is Bessel function of first kind of order ,is the resonance frequency, is the disk radius, and , ,

and , are the density, Poisson ratio, and Young’s modulus,respectively, of the disk structural material. Although hidden inthe precision of (9)’s formulation, the resonance frequency ofthis wine glass disk is to first order inversely proportional to itsradius . Once vibrating, the dc biased (by ) time-varyingoutput electrode-to-resonator capacitors generate output cur-rents governed also by (5) with replaced by ,and with a frequency response also characteristic of an LCRcircuit. However, the equivalent circuit for this two-port disk,shown in Fig. 6(c), differs from the one-port circuit of Fig. 2 inthat the electrode-to-resonator capacitor no longer connectsthe input to the output, but is rather now shunted to groundby the dc biased (i.e., ac ground) disk structure. The removalof from the input-to-output feedthrough path is highlyadvantageous for the series resonant oscillator configurationused in this work, since it better isolates the input from theoutput, allowing the majority of the current through the deviceto be filtered by its high bandpass biquad transfer function.

Pursuant to attaining a closed form expression for the se-ries motional resistance , the electrode-to-resonator overlaparea for each of the two ports of the wine glass disk res-onator is , where and are the radius and thickness ofthe disk structure, respectively. Aside from this difference, the

Fig. 7. (a) SEM and (b)–(c) frequency characteristics (measured under20-mtorr vacuum with different dc bias voltages) of a fabricated 60-MHz wineglass disk resonator with two support beams.

approximate expression for takes on a similar form to thatof (7), and can be written as

(11)

where is now the effective stiffness of the disk. For a 3- m-thick, 32- m-radius version of the disk in Fig. 6, dc biased to

V with 80-nm electrode-to-disk gaps and a of48 000, (11) predicts an of 1.46 k , which is quite reason-able for oscillator implementation.

Fig. 7(a) shows the SEM of the fabricated 3- m-thick,64- m-diameter, 60-MHz wine glass disk resonator used inthis work, with a zoom-in view of the 80 nm gap after release.Fig. 7(b) and (c) present frequency characteristics measuredunder a 20-mtorr vacuum, where the device exhibits a of145 780 with V; and 48 000 with V. From theheight of the peak of the frequency spectrum in Fig. 7(c),can be extracted to be about 1.5 k , which is very close to the1.46 k predicted by (11). Although this number is somewhathigher than exhibited by the wide-width CC-beam resonator, itis still quite amenable to many oscillator applications.

The key to achieving improvements on the scale above liesnot only in the use of a wine glass disk resonator, but also in thespecific advances applied to its design. In particular, the wineglass disk of this work differs from that of a previous prototype[6] in that its thickness has been increased from 1.5 to 3 mand gap reduced from 100 to 80 nm in order to increase itspower handling and lower its impedance according to (11). Inaddition, the number of supports used has been reduced from



Sustaining Amplifier Design

• Use feedback amplifier to create positive feedback trans-resistance

• Automatic gain control is used so that the oscillation self-limits through the electronic non-linearity. This reduces the oscillator amplitude but also helps to reduce 1/f noise up-conversion

10



4 to 2, and the support beamwidth has been reduced from 1.5 mto 1 m, all to decrease energy loss from the disk to the substratethrough the anchors, and thus, maximize the device . Evenwith these enhancements, the measured of 1.5 k for the64 m-diameter 60-MHz wine glass disk with V and

is still larger than the 50 normally exhibitedby off-chip quartz crystals, and thus, in an oscillator applicationrequires a sustaining amplifier capable of supporting higher tankimpedance.

With the above modifications, the stiffness of this wine glassdisk resonator becomes N/m, which is 71.5 the9240 N/m of the previous 10-MHz wide-width CC-beam. Ac-counting for differences in , and , (8) predicts a powerhandling capability 10 higher for the wine glass disk. For thesame V, this should result in a 10-dB lower far-from-carrier phase-noise floor.

IV. SUSTAINING AMPLIFIER DESIGN

To complete the oscillator circuit, a sustaining amplifier cir-cuit compatible with the comparatively large motional resis-tance of micromechanical resonators is needed. As mentionedearlier, and as was done with a previous oscillator [21], a transre-sistance amplifier in series with the resonator is a logical choice,since the low input and output resistances and , respec-tively, of such an amplifier impose relatively small loading onthe resonator, allowing the loaded of the system to be veryclose to the large resonator , without sacrificing power transferthrough the loop. Such an amplifier would need to have suffi-cient gain, per item 1) of Section II; would need to provide a 0input-to-output phase shift to accommodate the 0 phase shiftof the resonator when operating at series resonance, per item 2)of Section II; and would need to do all of the above with min-imal noise and power consumption.

Fig. 8 presents the top-level schematic of the oscillator cir-cuit used in this work, where the micromechanical resonatoris represented by its equivalent LCR circuit (which in this caseassumes the wine glass disk resonator of Fig. 6). As shown,a series resonant configuration is indeed used to best accom-modate the medium-range resistance of the micromechanicalresonator tanks to be used. However, the particular sustainingamplifier circuit of Fig. 8 differs from previous two-stage cir-cuits [10], [21] not only in its use of only one gain stage, butalso in that it achieves the needed 0 phase shift for oscil-lation in only a single stage, which improves both its noiseand bandwidth performance. As shown in the coarse oscillatorschematic of Fig. 8, the sustaining circuit is composed of a fullybalanced differential CMOS op amp hooked in shunt-shuntfeedback on one side, with the output taken from the other.By taking the output from the other side of the differentialop amp, an additional 180 phase shift is added on top of the180 shift from the shunt-shunt feedback, resulting in a total0 phase shift from input to output, while preserving a lowoutput resistance (due to symmetry) obtained via shunt-shuntfeedback. In the detailed circuit schematic of Fig. 9, transis-tors – comprise the basic single-stage, differential opamp, while – constitute a common-mode feedbackcircuit that sets its output dc bias point. The MOS resistors

Fig. 8. Top-level circuit schematic of the micromechanical resonatoroscillator of this work. Here, the (wine glass disk) micromechanical resonatoris represented by its equivalent electrical circuit.

Fig. 9. Detailed circuit schematic of the single-stage sustaining transresistanceamplifier of this work, implemented by a fully differential amplifier in aone-sided shunt-shunt feedback configuration.

and provide resistances and andserve as shunt-shunt feedback elements that allow control ofthe transresistance gain via adjustment of their gate voltages.The need for two of them will be covered later in Section Von ALC.

A. Transfer Function

Expressions for the dc transresistance gain, input resistance,and output resistance, of the sustaining amplifier are as follows:

(12)

(13)

(14)

where is the transconductance of , and are theoutput resistance of and , respectively, is MOS re-sistor value implemented by and , assumed to bemuch smaller than the s, and the forms on the far rights as-sume a large amplifier loop gain . (Note thatthis is amplifier loop gain, not oscillator loop gain.) In practice,


Amplifier Details

• Single-stage amplifier is used to maximize bandwidth. Recall that any phase shift through the amplifier causes the oscillation frequency to shift (and phase noise to degrade)

• Common-mode feedback used to set output voltage. Feedback resistance and Amplitude Level Control (ALC) implemented with MOS resistors

11













(12)

(13)

(14)




Design Equations

• These equations are used to trade-off between power and noise in the oscillator. The device size cannot be too large since the bandwidth needs to be about 10X the oscillation frequency.

12













(12)

(13)

(14)



(12)–(14) indicate that depends mainly on , whileand depend mainly on the of the input transistors, sug-gesting that larger input transistor ratios or larger bias cur-rents can further reduce the input and output resistance, hence,reduce loading. The latter approach, however, will increasethe power consumption of the amplifier and might disturb theimpedance balance in the loop.

In addition, the use of larger s can impact the 3-dB band-width of the transresistance amplifier, which as a rule should beat least 10 the oscillation frequency so that its phase shift atthis frequency is minimal. In particular, as detailed in [31], anamplifier phase shift close to 0 allows the micromechanical res-onator to operate at the point of highest slope in its phase versusfrequency curve, which allows it to more effectively suppressfrequency deviations caused by amplifier phase deviations. Thebandwidth of the transresistance amplifier of Fig. 9 is a functionof parasitic capacitance in both the transistors and the microme-chanical resonator, and is best specified by the full transfer func-tion for the amplifier

(15)where

(16)

is the open loop transresistance gain of the base amplifier withfeedback loading, where

(17)

and where is dc voltage gain of the base op amp, and andare the total parasitic capacitance at the input and output

terminals of the amplifier, respectively, including MOS parasiticcapacitance, pad capacitance, and resonator parasitic capaci-tance. Equation (15) has the form of a lowpass biquad transferfunction, with a dc gain given in (12), and aneffective bandwidth given by

(18)

From (18), the bandwidth can be increased by decreasingand , and increasing the gain-bandwidth product of theamplifier. For best stability, the effective bandwidth should bechosen to be at least 10 greater than the oscillation frequency.For a given drain current, this places a limit on how large the

ratio of the input transistors can be made, since growsfaster than with increasing , hence, bandwidth suffers.Bandwidth needs also constrain to a maximum allowablevalue, while loop gain needs set its minimum value.

B. Sustaining Amplifier Noise

In (2), the phase-noise contribution from the sustaining am-plifier is modeled in the noise factor constant , which canbe expanded into

(19)

where

(20)

where represents the input-referred current noise of thesustaining amplifier, , and is the input-referred voltage noise source of the differential op amp, givenby

(21)

where is 2/3 for long-channel devices, and from 2–3 largerfor short-channel devices. In (21), all common mode noisesources are nulled by the common-mode feedback circuit.In addition, flicker noise is neglected since the oscillationfrequency is beyond the flicker noise corner, and (2) representsonly an approximate expression that accounts for noiseand white noise at large offsets. (If (2) attempted to include

noise, then transistor flicker noise would need to beincluded.)

From (21) noise from this sustaining amplifier improves asthe size of the op amp input transistors and/or their drain cur-rents increase—the same design changes needed to decrease theamplifier and , with the same bandwidth-based restric-tions on input transistor s. For a given resonator andoscillation frequency , the optimal sustaining amplifier designthat still meets wireless handset specifications for the referenceoscillator can be found by simultaneous solution of (2), (12),and (18), to obtain the drain current and input transistor s thatminimize the power consumption.

V. ALC

As will be seen, oscillator limiting via resonator nonlin-earity seems to introduce a phase-noise component thatdominates the close-to-carrier phase noise of micromechanicalresonator oscillators [22]. To remedy this, the oscillator IC ofthis work also includes an ALC circuit, shown in Figs. 8–10,that can insure limiting via electronic methods, rather thanby mechanical resonator nonlinearity. This circuit consists ofan envelope detector that effectively measures the oscillationamplitude, followed by a comparator that compares the am-plitude with a reference value , then feeds back a voltageproportional to their difference to the gate of the gain-con-trolling MOS resistor so as to match the oscillationamplitude to . In this scheme, since the resistance ofis very large at the start of oscillation, another MOS resistor

is placed in parallel with and biased to a channelresistance value equal to , which realizes a loop gaingreater than 3, and thereby insures oscillation start-up once


(12)–(14) indicate that depends mainly on , whileand depend mainly on the of the input transistors, sug-gesting that larger input transistor ratios or larger bias cur-rents can further reduce the input and output resistance, hence,reduce loading. The latter approach, however, will increasethe power consumption of the amplifier and might disturb theimpedance balance in the loop.

In addition, the use of larger s can impact the 3-dB band-width of the transresistance amplifier, which as a rule should beat least 10 the oscillation frequency so that its phase shift atthis frequency is minimal. In particular, as detailed in [31], anamplifier phase shift close to 0 allows the micromechanical res-onator to operate at the point of highest slope in its phase versusfrequency curve, which allows it to more effectively suppressfrequency deviations caused by amplifier phase deviations. Thebandwidth of the transresistance amplifier of Fig. 9 is a functionof parasitic capacitance in both the transistors and the microme-chanical resonator, and is best specified by the full transfer func-tion for the amplifier

(15)where

(16)

is the open loop transresistance gain of the base amplifier withfeedback loading, where

(17)

and where is dc voltage gain of the base op amp, and andare the total parasitic capacitance at the input and output

terminals of the amplifier, respectively, including MOS parasiticcapacitance, pad capacitance, and resonator parasitic capaci-tance. Equation (15) has the form of a lowpass biquad transferfunction, with a dc gain given in (12), and aneffective bandwidth given by

(18)

From (18), the bandwidth can be increased by decreasingand , and increasing the gain-bandwidth product of theamplifier. For best stability, the effective bandwidth should bechosen to be at least 10 greater than the oscillation frequency.For a given drain current, this places a limit on how large the

ratio of the input transistors can be made, since growsfaster than with increasing , hence, bandwidth suffers.Bandwidth needs also constrain to a maximum allowablevalue, while loop gain needs set its minimum value.

B. Sustaining Amplifier Noise

In (2), the phase-noise contribution from the sustaining am-plifier is modeled in the noise factor constant , which canbe expanded into

(19)

where

(20)

where represents the input-referred current noise of thesustaining amplifier, , and is the input-referred voltage noise source of the differential op amp, givenby

(21)

where is 2/3 for long-channel devices, and from 2–3 largerfor short-channel devices. In (21), all common mode noisesources are nulled by the common-mode feedback circuit.In addition, flicker noise is neglected since the oscillationfrequency is beyond the flicker noise corner, and (2) representsonly an approximate expression that accounts for noiseand white noise at large offsets. (If (2) attempted to include

noise, then transistor flicker noise would need to beincluded.)

From (21) noise from this sustaining amplifier improves asthe size of the op amp input transistors and/or their drain cur-rents increase—the same design changes needed to decrease theamplifier and , with the same bandwidth-based restric-tions on input transistor s. For a given resonator andoscillation frequency , the optimal sustaining amplifier designthat still meets wireless handset specifications for the referenceoscillator can be found by simultaneous solution of (2), (12),and (18), to obtain the drain current and input transistor s thatminimize the power consumption.

V. ALC

As will be seen, oscillator limiting via resonator nonlin-earity seems to introduce a phase-noise component thatdominates the close-to-carrier phase noise of micromechanicalresonator oscillators [22]. To remedy this, the oscillator IC ofthis work also includes an ALC circuit, shown in Figs. 8–10,that can insure limiting via electronic methods, rather thanby mechanical resonator nonlinearity. This circuit consists ofan envelope detector that effectively measures the oscillationamplitude, followed by a comparator that compares the am-plitude with a reference value , then feeds back a voltageproportional to their difference to the gate of the gain-con-trolling MOS resistor so as to match the oscillationamplitude to . In this scheme, since the resistance ofis very large at the start of oscillation, another MOS resistor

is placed in parallel with and biased to a channelresistance value equal to , which realizes a loop gaingreater than 3, and thereby insures oscillation start-up once



Amplitude Control Loop

• Precision peak-detector used to sense oscillation amplitude. This is done by putting a MOS diode in the feedback path of an inverting op-amp

13


Fig. 10. (a) Top-level and (b) detailed circuit schematics of the ALC circuit.

power is applied. As the amplitude of oscillation grows andthe ALC reduces ’s channel resistance to below thatof , then dictates the total shunt-shunt feedbackresistance of the sustaining amplifier.

The envelope detector in Fig. 10 [32] combines a classicop amp-based precision peak rectifier design using an MOSdiode in its feedback path, with a capacitive peak sampler

, and a bleed current source (implemented by )with values of 1 pF and 0.1 A, respectively, chosen to trackan oscillator output with an assumed maximum peak-to-peakvariation of V/s. In this circuit, when the input voltageis larger than the capacitor voltage, , is ON and forces

to equal to . On the other hand, when ,is OFF, allowing to hold against the bleed current of

, which discharges at the rate of V/s.To attenuate the ripple at the envelope detector output, the

bandwidth of the subsequent comparator is purposely limited byimplementing it via a two-stage op amp compensated heavily togenerate a low frequency dominant pole. A 1-pF compensationcapacitor is sufficient to split the poles to 83 kHz and 109 MHz,which provides more than 61 of phase margin, and more im-portantly, attenuates the 10 MHz ripple by 41.6 dB.

VI. EXPERIMENTAL RESULTS

Standard and wide-width micromechanical CC-beam res-onators, such as in Figs. 3 and 4, with cross sections asshown in Fig. 11(a), were fabricated using a small vertical gappolysilicon surface-micromachining process previously used toachieve HF micromechanical filters [5]. Micromechanical wineglass disk resonators, such as in Fig. 7, with cross sections as inFig. 11(b), were fabricated via a three-polysilicon self-alignedstem process used previously to achieve disk resonators with

Fig. 11. Final cross-section views of the fabricated (a) CC-beam resonator and(b) wine glass disk resonator.

tiny lateral electrode-to-resonator gaps [6], [17]. Table II sum-marizes the three resonator designs used in the oscillators ofthis work.

Fig. 12 presents a photo of the amplifier IC, which was fab-ricated in TSMC’s 0.35- m process. As shown, the transre-sistance amplifier lies in the upper mid section, close to theinput and output terminals of the chip. The envelope detectorand the comparator of the ALC circuit sit adjacent to the am-plifier, with their 1-pF capacitors clearly visible. The total ICchip area is about m m, which together withthe m m required for the CC-beam resonator, the

m m required for the wide-width CC-beam res-onator, or the m m required for the wine glass diskresonator, yields (to the author’s knowledge) the smallest foot-print to date for any high reference oscillator in this frequencyrange. Tables II and III summarize the design and performanceof the overall oscillator circuit.

Interconnections between the IC and MEMS chips weremade via wire bonding, and testing was done under vacuumto preserve the high of the micromechanical resonators.Fig. 13 presents the measured open-loop gain of the 10-MHzwide-width CC-beam oscillator at various input power levels,showing a spring-softening Duffing nonlinearity that likely con-tributes to limiting of the oscillation amplitude when the ALCis not engaged. The open loop gain is measured by breaking theoscillator feedback loop at the input electrode of the CC-beam,applying an ac signal to this electrode, and measuring the outputpower of the transresistance amplifier by directing it througha buffer, then into an HP 4194A Impedance/Gain-Phase An-alyzer capable of both forward and reverse frequency sweepsto correctly extract peak values in Duffing-distorted curves.As shown in Fig. 13, as the input power increases, the peaksbecome smaller, clearly indicating a decrease in open loopgain. Extracting from these peaks and plotting it versus thecorresponding input powers yields the curves of Fig. 14, where

is seen to increase faster as the input power increases. Forthis particular resonator, the value of crosses a value equalto the transresistance amplifier gain (set by k ), at



Measured Spectra and Time-Domain

• These are the measurements without using the ALC

• The oscillation self-limits due to the resonator non-linearity

• Notice the extremely small oscillation amplitudes

• With the ALC, the oscillation amplitude drops to 10mV

14


Fig. 14. Extracted from the peaks shown in Fig. 13 versus theircorresponding input power. The is also derived fromthe curve and compared with for a typical case to illustrategraphical determination of the steady-state oscillation amplitude.

Fig. 15. Measured steady-state Fourier spectra and oscilloscope waveformsfor (a) the 10-MHz 8- m-wide CC-beam resonator oscillator; (b) the 10-MHz40- m-wide CC-beam resonator oscillator; and (c) the 60-MHz wine glass diskresonator oscillator. All data in this figure are for the oscillators with ALCdisengaged.

wine glass disk design when it comes to power handling. Ac-counting for the corresponding series motional resistances andoutput current of each resonator, the corresponding operatingpowers for each measured amplitudeare 0.061 W, 0.405 W, and 3.44 W. Thus, widening a10-MHz CC-beam from 8- m-wide to 40- m-wide providesa 6.64 increase in oscillator steady-state power, which is atleast consistent with factors given in Section III (but not thesame, since the ones in Section III were based on maximumpower handling). On the other hand, replacing the wide-width

Fig. 16. Phase-noise density versus carrier offset frequency plots for (a) the10-MHz 8- m-wide CC-beam resonator oscillator; (b) the 10-MHz 40- m-wideCC-beam resonator oscillator; and (c) the 60-MHz wine glass disk resonatoroscillator. All were measured using an HP E5500 Phase Noise MeasurementSystem. The dotted line is the phase-noise prediction by (2) and (19).

CC-beam entirely by a 60-MHz wine glass resonator yields an8.49 increase, again consistent with Section III.

The practical impact of the progressively larger powerhandlings among the resonators is clearly shown in the phasenoise versus carrier offset frequency plots of Fig. 16(a)–(c) forthe 8- m-wide 10-MHz CC-beam, the 40- m-wide 10-MHzCC-beam, and the 60-MHz wine glass disk, respectively.The far-from-carrier phase-noise levels of 116 dBc/Hz,

120 dBc/Hz, and 132 dBc/Hz, respectively, are close tothe predicted values of 118.1 dBc/Hz, 123.6 dBc/Hz, and

134.5 dBc/ Hz, using (2) and (19). For fair comparison,the value for the 60-MHz wine glass disk oscillator becomes

147 dBc/Hz when divided down to 10 MHz, which practically(if not strictly) satisfies the GSM specification.

The close-to-carrier phase noise, on the other hand, looksnothing like the expectation of (2). In particular, rather thanthe expected component, a larger-than-expectedcomponent is observed that masks the . The observedcomponent is substantially larger than predicted by first-orderexpressions that assume an aliasing mechanism for noise,whereby amplifier noise aliases via micromechanicalresonator transducer nonlinearity into the oscillator passbandand noise is generated via filtering through the resonatortransfer function [22]. Other formulations for noise basedon resonator frequency dependence on the dc bias voltage[22] also do not correctly predict the magnitude of this noisecomponent.



Experimental Results

• Performance close to GSM specs. DC power and area are compelling

• The measured 1/f noise much larger than expected

15


TABLE IIRESONATOR DATA SUMMARY

Fig. 12. Photo of the sustaining transresistance amplifier IC fabricated inTSMCs 0.35- m CMOS process.

which point the loop gain of an oscillator would drop to 0 dB,the oscillation amplitude would stop growing, and steady-stateoscillation would ensue. Although the plot of Fig. 13 seemsto imply that Duffing nonlinearity might be behind motionalresistance increases with amplitude, it is more likely that de-creases in or with amplitude are more responsible[21], since Duffing is a stiffness nonlinearity, and stiffness (likeinductance or capacitance) is a nondissipative property.

Oscillators with the ALC loop of Fig. 10 disengaged weretested first. Fig. 15 presents spectrum analyzer plots and oscil-loscope waveforms for oscillators with ALC disengaged using

TABLE IIIOSCILLATOR DATA SUMMARY

Fig. 13. Measured open-loop gain of the 10-MHz wide-width CC-beamoscillator circuit under increasing input signal amplitudes. These curves weretaken via a network analyzer sweeping down in frequency (i.e., from higher tolower frequency along the -axis).

each of the three resonator designs summarized in Table I.Fig. 16 presents plots of phase-noise density versus offset fromthe carrier frequency for each oscillator, measured by directingthe output signal of the oscillator into an HP E5500 PhaseNoise Measurement System.

A quick comparison of the oscilloscope waveforms ofFig. 15(a)–(c), which shows steady-state oscillation ampli-tudes of 42 mV, 90 mV, and 200 mV, for the 8- m-wide10-MHz CC-beam, the 40- m-wide 10-MHz CC-beam, andthe 60-MHz wine glass disk, respectively, clearly verifies theutility of wide-CC-beam design and the superiority of the


TABLE IIRESONATOR DATA SUMMARY

Fig. 12. Photo of the sustaining transresistance amplifier IC fabricated inTSMCs 0.35- m CMOS process.

which point the loop gain of an oscillator would drop to 0 dB,the oscillation amplitude would stop growing, and steady-stateoscillation would ensue. Although the plot of Fig. 13 seemsto imply that Duffing nonlinearity might be behind motionalresistance increases with amplitude, it is more likely that de-creases in or with amplitude are more responsible[21], since Duffing is a stiffness nonlinearity, and stiffness (likeinductance or capacitance) is a nondissipative property.

Oscillators with the ALC loop of Fig. 10 disengaged weretested first. Fig. 15 presents spectrum analyzer plots and oscil-loscope waveforms for oscillators with ALC disengaged using

TABLE IIIOSCILLATOR DATA SUMMARY

Fig. 13. Measured open-loop gain of the 10-MHz wide-width CC-beamoscillator circuit under increasing input signal amplitudes. These curves weretaken via a network analyzer sweeping down in frequency (i.e., from higher tolower frequency along the -axis).

each of the three resonator designs summarized in Table I.Fig. 16 presents plots of phase-noise density versus offset fromthe carrier frequency for each oscillator, measured by directingthe output signal of the oscillator into an HP E5500 PhaseNoise Measurement System.

A quick comparison of the oscilloscope waveforms ofFig. 15(a)–(c), which shows steady-state oscillation ampli-tudes of 42 mV, 90 mV, and 200 mV, for the 8- m-wide10-MHz CC-beam, the 40- m-wide 10-MHz CC-beam, andthe 60-MHz wine glass disk, respectively, clearly verifies theutility of wide-CC-beam design and the superiority of the



Array-Composite MEMS Wine-Glass Osc

• Increase power handling capability by coupling multiple (N) resonators together.

• This increases power handling capability by N.

16

Y.-W. Lin, S.-S. Li, Z. Ren, and C. T.-C. Nguyen, “Low phase noise array-composite micromechanical wine-glass disk oscillator,” Technical

Digest, IEEE Int. Electron Devices Mtg., Washington, DC, Dec. 5-7, 2005, pp. 287-290.

Low Phase Noise Array-Composite Micromechanical Wine-Glass Disk Oscillator

Yu-Wei Lin, Sheng-Shian Li, Zeying Ren, and Clark T.-C. Nguyen

Center for Wireless Integrated Micro Systems

Department of Electrical Engineering and Computer Science

University of Michigan, Ann Arbor, Michigan 48109-2122, USA

TEL: 734-764-5411, FAX: 734-647-1781, email: [email protected]

ABSTRACT

A reduction in phase noise by 13 dB has been obtained

over a previous 60-MHz surface-micromachined microme-

chanical resonator oscillator by replacing the single resonator

normally used in such oscillators with a mechanically-coupled

array of them to effectively raise the power handling ability of

the frequency selective tank. Specifically, a mechanically-

coupled array of nine 60-MHz wine-glass disk resonators em-

bedded in a positive feedback loop with a custom-designed,

single-stage, zero-phase-shift sustaining amplifier achieves a

phase noise of -123 dBc/Hz at 1 kHz offset and -136 dBc/Hz

at far-from-carrier offsets. When divided down to 10 MHz,

this effectively corresponds to -138 dBc/Hz at 1 kHz offset

and -151 dBc/Hz at far from carrier offset, which represent 13

dB and 4 dB improvements over recently published work on

surface-micromachined resonator oscillators, and also now

beat stringent GSM phase noise requirements by 8 dB and 1

dB, respectively.

I. INTRODUCTION

Among off-chip components in a wireless communication

circuit, the quartz crystal used in the reference oscillator is

perhaps the most difficult to miniaturize, since Q’s > 10,000

and thermal stabilities better than 35 ppm uncompensated over

0-70°C are generally unavailable on-chip. Recently, however,

on-chip vibrating micromechanical resonators based on

MEMS technology have become increasingly attractive as on-

chip frequency selective elements for communication-grade

oscillators and filters, spurred by demonstrations of Q’s >

145,780 at 60 MHz [1], frequency temperature dependencies

of only 18 ppm over 25-105°C at 10 MHz [2], and by a poten-

tial for single-chip integration with transistors [3]. A recent

reference oscillator using an SOI-based vibrating microme-

chanical resonator [4] has, in fact, already satisfied the GSM

specification (-130 dBc/Hz and -150 dBc/Hz at 1 kHz and far-

from-carrier offsets, respectively, from a 13 MHz carrier).

Oscillators employing surface-micromachined resonator tanks,

however, encumbered by a smaller power handling ability, do

not quite yet satisfy GSM specs [1] over all carrier offsets.

Between SOI- and surface-micromachining, the latter is per-

haps more attractive, since it avoids high-aspect ratios and has

a more successful planar integration history [3][5].

Pursuant to attaining GSM reference oscillator phase noise

specs using a surface-micromachined resonator tank, this work

first recognizes that phase noise is inversely proportional to

resonator tank power handling [6]; then proceeds to lower

phase noise by coupling several micromechanical disk resona-

tors into mechanical arrays that then automatically match up

the individual resonator frequencies and combine their outputs

to significantly raise the overall “composite” resonator power

handling. This increase in power handling leads to oscillator

phase noise improvements of up to 13 dB at close-to-carrier

offsets, and yields an overall phase noise plot that handily sur-

passes GSM specs. As will be shown, much of this rather large

improvement actually derives from a removal of the 1/f3 close-

to-carrier noise that before now had greatly limited the per-

formance of micromechanical resonator oscillators.

II. WINE-GLASS DISK ARRAY RESONATOR

Fig. 1 presents the perspective-view schematic of a 3-

resonator version of the wine-glass disk-array resonator used

to raise power handling in this work, together with a typical

two-port bias and excitation scheme, and an equivalent electri-

cal model. Here, three (or more) wine-glass disks, each identi-

cally designed to 60 MHz, are coupled mechanically by 1!m-

wide, half-wavelength coupling beams connecting each adja-

cent resonator to one another at high-velocity locations.

Fig. 1: Perspective-view schematic of a multi (three) wine-glass disk micro-

mechanical resonator array. The electrical equivalent circuit for the resonator

is shown to the bottom right.

Fig. 2: ANSYS simulated mode shapes for a mechanically coupled three

wine-glass disk micromechanical resonator array.

Amplitude

Freq.

1st Mode

2nd

Mode

3rd

Mode

vi

Co

VP

z

r

!

SupportBeam

Coupling Beam

Anchor

WGDisk

OutputElectrode

InputElectrode

io

vo

RL

Rx Lx Cx io

vo

RL Co

WGDisk WGDisk

vi



IV. EXPERIMENTAL RESULTS

Wine-glass disk array resonators were fabricated via a

three-polysilicon self-aligned stem process used previously to

achieve disk resonators [9]. Fig. 5 presents SEM’s of fabri-

cated 60-MHz wine-glass disk arrays with varying numbers of

coupled resonators, each supported by only two support beams.

Fig. 6 presents measured frequency spectra for a stand-alone

wine-glass disk resonator together with resonator arrays using

3, 5, and 9 resonators mechanically coupled with one another.

Although the single resonator achieves the highest Q of

161,000, the array Q’s are still all greater than 115,000.

From the peak heights, Rx (with VP = 7 V) can be extracted

to be 11.73 k!, 6.34 k!, 4.04 k!, and 2.56 k!, for 1, 3, 5,

and 9 resonator arrays, respectively. The measured Rx reduc-

tion factors of 1.85, 2.90, and 4.58, actually fall short of the

expected 3, 5, and 9, respectively (i.e., the number of the reso-

nators in the array). Although differences in Q contribute

somewhat to the lower multiplication factor, the main culprit

here is the need to split the electrodes between resonators in

order to avoid coupling beams located at high velocity points

(c.f., Fig. 1). Since the coupling beams attach at high velocity

points, the electrodes between resonators must be split and

removed at high velocity points where the current would oth-

erwise have been the largest. This greatly reduces the current

contribution from such inner electrodes, thereby reducing the

factor by which the motional resistance is lowered. A more

detailed determination of Rx using velocity integration over

each electrode (e.g., as done in [10]) yields reduction ratios of

1.84, 2.85, and 4.94, for the 3, 5, and 9 resonator arrays, re-

spectively, which now match the measured values.

Note that the square resonators of [7] did not suffer from

the above problem, since their coupling beams did not inter-

fere with their electrodes. One remedy to the problem for disk

resonators is to couple the disks in a third plane above them,

which would keep the coupling beams clear of the electrodes,

but at the penalty of another masking step.

To ascertain how effectively unwanted modes in the me-

chanically coupled array have been suppressed, Fig. 7 presents

the frequency characteristic for a 5-disk array measured over a

wide frequency range (20 MHz). Here, only a single peak cor-

responding to the first filter mode is observed, which verifies

the utility of half-wavelength coupling beam design and elec-

trode placement in eliminating unwanted modes.

Fig. 5: SEM’s of fabricated wine-glass disk resonator-arrays with varying

numbers of mechanically-coupled wine-glass disks.

Fig. 6: Measured frequency characteristic for a fabricated wine-glass disk

resonator-array.

Fig. 7: Measured frequency spectrum verifying no spurious modes around

the desired mode of the resonator array, achieved via proper electrode excita-

tion and half-wavelength coupling beam design.

Oscillator Design Summary

Process TSMC 0.35 "m CMOS

Voltage Supply # 1.65 V

Power Cons. 350 "W

Amplifier Gain 8 k!

Amplifier BW 200 MHz

Integrated Circuit

Layout Area 50 "m $ 50 "m

Process Polysilicon-Based

Surface Micromachining

Radius, R 32 "m

Thickness, h 3 "m

Gap, do 80 nm

Voltage Supply 10 V

Power Cons. ~ 0 W

Motional Resistance, Rx

5.75 k!, 3.11 k!, 1.98 k!, 1.25 k! for n = 1, 3, 5, 9

MEMS Wine-Glass

Disk Resonator

Array

Layout Area n $ 105 "m $ 105 "m

-110

-100

-90

-80

-70

-60

-50

-40

-30

61.73 61.78 61.83 61.88 61.93Frequency (MHz)

Tra

nsm

issio

n (

dB

)

3 Res.Q = 122,500

1 Res.Q = 161,000

5 Res. Q = 119,500

9 Res.Q = 118,900

DataR = 32 "m h = 3 "m

do = 80 nm VP = 7 V

-70

-65

-60

-55

-50

-45

-40

52 57 62 67 72

Frequency (MHz)

Tra

nsm

issio

n (

dB

) 5-Res. Array

VP = 7 V

Selected

Mode

No Spurious

Modes

Table 1. Oscillator Data Summary

3 WGDisk1 WGDisk

5 WGDisk

9 WGDisk

Support Beams Wine-Glass

Disk

Anchor

Input Electrode

Coupling Beam Output

Electrode

R=32"m

Zoom-in View


Design Summary

• Prototype resonator implemented in a 0.35μm CMOS process shows no spurious modes

• Area is still quite resonable compared to a bulky XTAL

17















to be 11.73 k!, 6.34 k!, 4.04 k!, and 2.56 k!, for 1, 3, 5,


































resonator-array.







Power Cons. 350 "W

Amplifier Gain 8 k!


Integrated Circuit




Radius, R 32 "m

Thickness, h 3 "m

Gap, do 80 nm

Voltage Supply 10 V

Power Cons. ~ 0 W


5.75 k!, 3.11 k!, 1.98 k!, 1.25 k! for n = 1, 3, 5, 9

MEMS Wine-Glass

Disk Resonator

Array


-110

-100

-90

-80

-70

-60

-50

-40

-30

61.73 61.78 61.83 61.88 61.93Frequency (MHz)

Tra

nsm

issio

n (

dB

)

3 Res.Q = 122,500

1 Res.Q = 161,000

5 Res. Q = 119,500

9 Res.Q = 118,900

DataR = 32 "m h = 3 "m

do = 80 nm VP = 7 V

-70

-65

-60

-55

-50

-45

-40

52 57 62 67 72

Frequency (MHz)

Tra

nsm

issio

n (

dB

) 5-Res. Array

VP = 7 V

Selected

Mode

No Spurious

Modes


3 WGDisk1 WGDisk

5 WGDisk

9 WGDisk


Disk

Anchor

Input Electrode


Electrode

R=32"m

Zoom-in View















to be 11.73 k!, 6.34 k!, 4.04 k!, and 2.56 k!, for 1, 3, 5,


































resonator-array.







Power Cons. 350 "W

Amplifier Gain 8 k!


Integrated Circuit




Radius, R 32 "m

Thickness, h 3 "m

Gap, do 80 nm

Voltage Supply 10 V

Power Cons. ~ 0 W


5.75 k!, 3.11 k!, 1.98 k!, 1.25 k! for n = 1, 3, 5, 9

MEMS Wine-Glass

Disk Resonator

Array


-110

-100

-90

-80

-70

-60

-50

-40

-30

61.73 61.78 61.83 61.88 61.93Frequency (MHz)

Tra

nsm

issio

n (

dB

)

3 Res.Q = 122,500

1 Res.Q = 161,000

5 Res. Q = 119,500

9 Res.Q = 118,900

DataR = 32 "m h = 3 "m

do = 80 nm VP = 7 V

-70

-65

-60

-55

-50

-45

-40

52 57 62 67 72

Frequency (MHz)

Tra

nsm

issio

n (

dB

) 5-Res. Array

VP = 7 V

Selected

Mode

No Spurious

Modes


3 WGDisk1 WGDisk

5 WGDisk

9 WGDisk


Disk

Anchor

Input Electrode


Electrode

R=32"m

Zoom-in View















to be 11.73 k!, 6.34 k!, 4.04 k!, and 2.56 k!, for 1, 3, 5,


































resonator-array.







Power Cons. 350 "W

Amplifier Gain 8 k!


Integrated Circuit




Radius, R 32 "m

Thickness, h 3 "m

Gap, do 80 nm

Voltage Supply 10 V

Power Cons. ~ 0 W


5.75 k!, 3.11 k!, 1.98 k!, 1.25 k! for n = 1, 3, 5, 9

MEMS Wine-Glass

Disk Resonator

Array


-110

-100

-90

-80

-70

-60

-50

-40

-30

61.73 61.78 61.83 61.88 61.93Frequency (MHz)

Tra

nsm

issio

n (

dB

)

3 Res.Q = 122,500

1 Res.Q = 161,000

5 Res. Q = 119,500

9 Res.Q = 118,900

DataR = 32 "m h = 3 "m

do = 80 nm VP = 7 V

-70

-65

-60

-55

-50

-45

-40

52 57 62 67 72

Frequency (MHz)

Tra

nsm

issio

n (

dB

) 5-Res. Array

VP = 7 V

Selected

Mode

No Spurious

Modes


3 WGDisk1 WGDisk

5 WGDisk

9 WGDisk


Disk

Anchor

Input Electrode


Electrode

R=32"m

Zoom-in View















to be 11.73 k!, 6.34 k!, 4.04 k!, and 2.56 k!, for 1, 3, 5,


































resonator-array.







Power Cons. 350 "W

Amplifier Gain 8 k!


Integrated Circuit




Radius, R 32 "m

Thickness, h 3 "m

Gap, do 80 nm

Voltage Supply 10 V

Power Cons. ~ 0 W


5.75 k!, 3.11 k!, 1.98 k!, 1.25 k! for n = 1, 3, 5, 9

MEMS Wine-Glass

Disk Resonator

Array


-110

-100

-90

-80

-70

-60

-50

-40

-30

61.73 61.78 61.83 61.88 61.93Frequency (MHz)

Tra

nsm

issio

n (

dB

)

3 Res.Q = 122,500

1 Res.Q = 161,000

5 Res. Q = 119,500

9 Res.Q = 118,900

DataR = 32 "m h = 3 "m

do = 80 nm VP = 7 V

-70

-65

-60

-55

-50

-45

-40

52 57 62 67 72

Frequency (MHz)

Tra

nsm

issio

n (

dB

) 5-Res. Array

VP = 7 V

Selected

Mode

No Spurious

Modes


3 WGDisk1 WGDisk

5 WGDisk

9 WGDisk


Disk

Anchor

Input Electrode


Electrode

R=32"m

Zoom-in View


Measured Phase Noise

• Meets GSM specs with comfortable margin

18



For oscillator testing, the IC and MEMS chips were inter-

connected via wire-bonding, and testing was done under vac-

uum to preserve the high Q of the micromechanical resonators

or arrays. Figs. 8-10 present oscillator performance data, start-

ing with the obligatory oscilloscope and spectrum analyzer

waveforms, and culminating in a plot of phase noise density

versus offset from the carrier frequency. The last of these

shows a phase noise of -123 dBc/Hz at 1 kHz offset from the

carrier, and -136 dBc/Hz at far-from-carrier offsets. This far-

from-carrier noise floor is about 4 dB better than that of an

oscillator using a single wine-glass disk resonator, verifying

the utility of resonator array design. In addition, Fig. 10 also

shows that the undesired 1/f3 noise (seen in previous microme-

chanical resonator oscillators [1]) is removed when a coupled

array is utilized, due to its increased power handling ability.

With 1/f3 noise suppressed, an expected 1/f2 dependence com-

monly exhibited by high Q oscillators then remains, improving

the close-to-carrier phase noise at 1 kHz offset from -110

dBc/Hz to -123 dBc/Hz. Dividing down to 10 MHz for fair

comparison, this corresponds to -138 dBc/Hz at 1 kHz offset

and -151 dBc/Hz far from the carrier, which more than satis-

fies (by 8 dB and 1dB) the stringent GSM requirement.

V. CONCLUSIONS

Via use of a mechanically-coupled array approach to boost

the power handling ability of a “composite” micromechanical

resonator, a 60-MHz series resonant oscillator divided down to

10 MHz has been demonstrated with phase noise values of -

138 dBc/Hz at 1 kHz offset and -151 dBc/Hz at far-from-

carrier offsets, both of which now satisfy stringent GSM speci-

fications for communications reference oscillators. This, to-

gether with its low power consumption of only 350 !W, and

its potential for full integration of the transistor sustaining

circuit and MEMS device onto a single silicon chip, makes the

micromechanical resonator-array oscillator of this work an

attractive on-chip replacement for quartz crystal reference

oscillators in communications applications. And all of this

made possible by effectively harnessing the integration advan-

tage of micromechanics, which allows a designer to break the

“minimalist” paradigm that dictates the use of one and only

one quartz crystal in an oscillator, and instead, permits the use

of as many micromechanical resonators as needed, with little

size or cost penalty.

Acknowledgments. This work was supported under DARPA

Grant No. F30602-01-1-0573.

REFERENCES

[1] Y.-W. Lin, S. Lee, S.-S. Li, Y. Xie, Z. Ren, and C. T.-C. Nguyen, “Series-

resonant VHF micromechanical resonator reference oscillators,” IEEE J.

Solid-State Circuits, vol. 39, no. 12, pp. 2477-2491, Dec. 2004.

[2] W.-T. Hsu and C. T.-C. Nguyen, “Stiffness-compensated temperature-

insensitive micromechanical resonators,” Technical Digest, MEMS’02,

Las Vegas, NV, 2002, pp. 731-734.

[3] A. E. Franke, J. M. Heck, T.-J. King, and R. T. Howe, “Polycrystalline

silicon-germanium films for integrated microsystems,” J. Microelectro-

mechanical Systems, vol. 12, no. 2, pp. 160-171, April 2003.

[4] V. Kaajakari, T. Mattila, A. Oja, J. Kiihamäki, and H. Seppä, “Square-

extensional mode single-crystal silicon micromechanical resonator for

low-phase-noise oscillator applications,” IEEE Electron Device Letters,

vol. 25, no. 4, pp. 173-175, April 2004.

[5] T. A. Core, W. K. Tsang, S. J. Sherman, “Fabrication technology for an

integrated surface-micromachined sensor,” Solid State Technology, vol.

36, no. 10, pp. 39-47, Oct. 1993.

[6] W. P. Robins, Phase Noise in Signal Sources. London: Peter Peregrinus,

Ltd., 1982.

[7] M. U. Demirci, M. A. Abdelmoneum, and C. T.-C. Nguyen, “Mechani-

cally corner-coupled square microresonator array for reduced series mo-

tional resistance,” Digest of Technical Papers, Transducers’03, Boston,

MA, 2003, pp. 955-958.

[8] S. Lee and C. T.-C. Nguyen, “Mechanically-coupled micromechanical

resonator arrays for improved phase noise,” Proceedings, IEEE Int. Ultra-

sonics, Ferroelectrics, and Frequency Control 50th Anniv. Joint Conf,

Montreal, Canada, Aug. 2004, pp. 280-286.

[9] J. Wang, Z. Ren, and C. T.-C. Nguyen, “1.156-GHz self-aligned vibrating

micromechanical disk resonator,” IEEE Trans. Ultrason., Ferroelect.,

Freq. Contr., vol. 51, no. 12, pp. 1607-1628, Dec. 2004.

[10] F. D. Bannon III, J. R. Clark, and C. T.-C. Nguyen, “High-Q HF micro-

electromechanical filters,” IEEE J. Solid-State Circuits, vol. 35, no. 4, pp.

512-526, April 2000.

Fig. 8: Measured steady-state oscilloscope waveform for the 60-MHz wine-

glass disk resonator-array oscillator.

Fig. 9: Measured steady-state Fourier spectrum for the 60-MHz wine-glass

disk resonator-array oscillator.

Fig. 10: Phase noise density versus carrier offset frequency plots for the 60-

MHz wine-glass disk resonator-array oscillator, measured using an HP

E5500 Phase Noise Measurement System. The two star symbols show the

GSM specification for close-to-carrier and far-from-carrier offsets.

230 mV

-90

-80

-70

-60

-50

-40

-30

-20

-10

0

61.5 61.7 61.9

Frequency (MHz)

Po

we

r (d

B)

-160

-140

-120

-100

-80

-60

-40

-20

1.E+01 1.E+02 1.E+03 1.E+04 1.E+05

Offset Frequency (Hz)

Ph

ase

Nois

e (

dB

c/H

z)

1/f

3 Noise

1/f

2 Noise

SingleResonator

9-ResonatorArray

FrequencyDivided Down

to 10 MHz

























V. CONCLUSIONS





















Grant No. F30602-01-1-0573.

REFERENCES






Las Vegas, NV, 2002, pp. 731-734.







vol. 25, no. 4, pp. 173-175, April 2004.



36, no. 10, pp. 39-47, Oct. 1993.


Ltd., 1982.




MA, 2003, pp. 955-958.










512-526, April 2000.









230 mV

-90

-80

-70

-60

-50

-40

-30

-20

-10

0

61.5 61.7 61.9

Frequency (MHz)

Po

we

r (d

B)

-160

-140

-120

-100

-80

-60

-40

-20

1.E+01 1.E+02 1.E+03 1.E+04 1.E+05


Ph

ase

Nois

e (

dB

c/H

z)

1/f

3 Noise

1/f

2 Noise

SingleResonator

9-ResonatorArray


to 10 MHz















to be 11.73 k!, 6.34 k!, 4.04 k!, and 2.56 k!, for 1, 3, 5,


































resonator-array.







Power Cons. 350 "W

Amplifier Gain 8 k!


Integrated Circuit




Radius, R 32 "m

Thickness, h 3 "m

Gap, do 80 nm

Voltage Supply 10 V

Power Cons. ~ 0 W


5.75 k!, 3.11 k!, 1.98 k!, 1.25 k! for n = 1, 3, 5, 9

MEMS Wine-Glass

Disk Resonator

Array


-110

-100

-90

-80

-70

-60

-50

-40

-30

61.73 61.78 61.83 61.88 61.93Frequency (MHz)

Tra

nsm

issio

n (

dB

)

3 Res.Q = 122,500

1 Res.Q = 161,000

5 Res. Q = 119,500

9 Res.Q = 118,900

DataR = 32 "m h = 3 "m

do = 80 nm VP = 7 V

-70

-65

-60

-55

-50

-45

-40

52 57 62 67 72

Frequency (MHz)

Tra

nsm

issio

n (

dB

) 5-Res. Array

VP = 7 V

Selected

Mode

No Spurious

Modes


3 WGDisk1 WGDisk

5 WGDisk

9 WGDisk


Disk

Anchor

Input Electrode


Electrode

R=32"m

Zoom-in View

























V. CONCLUSIONS





















Grant No. F30602-01-1-0573.

REFERENCES






Las Vegas, NV, 2002, pp. 731-734.







vol. 25, no. 4, pp. 173-175, April 2004.



36, no. 10, pp. 39-47, Oct. 1993.


Ltd., 1982.




MA, 2003, pp. 955-958.










512-526, April 2000.









230 mV

-90

-80

-70

-60

-50

-40

-30

-20

-10

0

61.5 61.7 61.9

Frequency (MHz)

Po

we

r (d

B)

-160

-140

-120

-100

-80

-60

-40

-20

1.E+01 1.E+02 1.E+03 1.E+04 1.E+05


Ph

ase

Nois

e (

dB

c/H

z)

1/f

3 Noise

1/f

2 Noise

SingleResonator

9-ResonatorArray


to 10 MHz


Phase Noise: Model for Resonator

• The system is non-linear due to the electrostatic mechanism and the mechanical non-linearities

19

kaajakari et al.: analysis of phase noise and micromechanical oscillators 2323

Fig. 1. Schematic representation of noise aliasing in micro-oscillator.A linear resonator would filter out the amplifier low-frequency 1/f-noise present at the resonator input, but nonlinear filtering elementwill result in noise aliasing.

Fig. 2. Mechanical lumped model for the resonator.

where m is the lumped mass, γ is the damping coefficient,Fe is the electrostatic forcing term, and k is the mechanicalspring constant. We also define the natural frequency ω0 =!

k/m and the quality factor Q = ω0m/γ. The resonatordisplacement x due to the force Fe is given by:

x = H(ω)Fe, (2)

where the force-displacement transfer function H(ω) from(1) is:

H(ω) =k−1

1 − ω2/ω20 + iω/Qω0

. (3)

The electrostatic force actuating the resonator is:

Fe =12

∂C

∂x(Udc + uac)

2 , (4)

where Udc is the direct current (DC)-bias voltage over thegap, uac is the alternating current (AC)-excitation voltage,and:

C = ϵ0Ael

d − x, (5)

is the transducer working capacitance that depends on thepermittivity of free space ϵ0, the electrode area Ael, andthe nominal electrode gap d. The current through the elec-trode is:

isig =∂CU

∂t≈ ∂C

∂tUdc + C0

∂uac

∂t, (6)

Fig. 3. The electrical equivalent circuit for MEMS-based oscillator.

where C0 is the capacitance at zero displacement. In (6),the first term is due to the capacitance variations (mo-tional current im), and the second term is the normal AC-current through the capacitance. The electromechanicaltransduction factor is identified as [12]:

η = Udc∂C

∂x≈ Udc

C0

d. (7)

The resulting relation between the motional current im,the mechanical transducer velocity x, the excitation volt-age uac, and the force Fe at the excitation frequency are:

im ≈ ηx,

Fe ≈ ηuac,(8)

where the displacement x is assumed to be small com-pared to the gap d. By substituting (8) into (1), an electri-cal equivalent circuit shown in Fig. 3 can be derived. Thecomponent values are:

Rm =√

km/Qη2 = k/ω0Qη2,

Cm = η2/k,

Lm = m/η2, andC0 = ϵ0Ael/d0.

(9)

The important observation is that, to obtain a smallmotional resistance Rm, a large electromechanical trans-duction factor is needed requiring either a small gap d ora large DC-bias voltage Udc. In practice, the voltage usu-ally is limited by system considerations and, thus, a smallgap, typically less than 1 µm, is needed. Unfortunately, aswill be seen in the following sections, the small gap will re-sult in unwanted nonlinear effects that limit the vibrationamplitude and cause noise aliasing.

B. Nonlinear Electrostatic Spring Force

Due to the inverse relationship between the electrodedisplacement and the parallel plate capacitance, the elec-trostatic coupling introduces nonlinear spring terms. Ad-ditionally, nonlinear effects of mechanical origins are pos-sible, and most fundamentally material nonlinearities setthe limit for the miniaturization [4] [13]. In this paper,however, the gap is assumed small, and therefore, the ca-pacitive nonlinearity dominates. Thus, a linear mechanical






x = H(ω)Fe, (2)


H(ω) =k−1

1 − ω2/ω20 + iω/Qω0

. (3)


Fe =12

∂C

∂x(Udc + uac)

2 , (4)


C = ϵ0Ael

d − x, (5)


isig =∂CU

∂t≈ ∂C

∂tUdc + C0

∂uac

∂t, (6)



η = Udc∂C

∂x≈ Udc

C0

d. (7)


im ≈ ηx,

Fe ≈ ηuac,(8)


Rm =√


Cm = η2/k,


(9)









x = H(ω)Fe, (2)


H(ω) =k−1

1 − ω2/ω20 + iω/Qω0

. (3)


Fe =12

∂C

∂x(Udc + uac)

2 , (4)


C = ϵ0Ael

d − x, (5)


isig =∂CU

∂t≈ ∂C

∂tUdc + C0

∂uac

∂t, (6)



η = Udc∂C

∂x≈ Udc

C0

d. (7)


im ≈ ηx,

Fe ≈ ηuac,(8)


Rm =√


Cm = η2/k,


(9)









x = H(ω)Fe, (2)


H(ω) =k−1

1 − ω2/ω20 + iω/Qω0

. (3)


Fe =12

∂C

∂x(Udc + uac)

2 , (4)


C = ϵ0Ael

d − x, (5)


isig =∂CU

∂t≈ ∂C

∂tUdc + C0

∂uac

∂t, (6)



η = Udc∂C

∂x≈ Udc

C0

d. (7)


im ≈ ηx,

Fe ≈ ηuac,(8)


Rm =√


Cm = η2/k,


(9)









x = H(ω)Fe, (2)


H(ω) =k−1

1 − ω2/ω20 + iω/Qω0

. (3)


Fe =12

∂C

∂x(Udc + uac)

2 , (4)


C = ϵ0Ael

d − x, (5)


isig =∂CU

∂t≈ ∂C

∂tUdc + C0

∂uac

∂t, (6)



η = Udc∂C

∂x≈ Udc

C0

d. (7)


im ≈ ηx,

Fe ≈ ηuac,(8)


Rm =√


Cm = η2/k,


(9)









x = H(ω)Fe, (2)


H(ω) =k−1

1 − ω2/ω20 + iω/Qω0

. (3)


Fe =12

∂C

∂x(Udc + uac)

2 , (4)


C = ϵ0Ael

d − x, (5)


isig =∂CU

∂t≈ ∂C

∂tUdc + C0

∂uac

∂t, (6)



η = Udc∂C

∂x≈ Udc

C0

d. (7)


im ≈ ηx,

Fe ≈ ηuac,(8)


Rm =√


Cm = η2/k,


(9)









x = H(ω)Fe, (2)


H(ω) =k−1

1 − ω2/ω20 + iω/Qω0

. (3)


Fe =12

∂C

∂x(Udc + uac)

2 , (4)


C = ϵ0Ael

d − x, (5)


isig =∂CU

∂t≈ ∂C

∂tUdc + C0

∂uac

∂t, (6)



η = Udc∂C

∂x≈ Udc

C0

d. (7)


im ≈ ηx,

Fe ≈ ηuac,(8)


Rm =√


Cm = η2/k,


(9)









x = H(ω)Fe, (2)


H(ω) =k−1

1 − ω2/ω20 + iω/Qω0

. (3)


Fe =12

∂C

∂x(Udc + uac)

2 , (4)


C = ϵ0Ael

d − x, (5)


isig =∂CU

∂t≈ ∂C

∂tUdc + C0

∂uac

∂t, (6)



η = Udc∂C

∂x≈ Udc

C0

d. (7)


im ≈ ηx,

Fe ≈ ηuac,(8)


Rm =√


Cm = η2/k,


(9)









x = H(ω)Fe, (2)


H(ω) =k−1

1 − ω2/ω20 + iω/Qω0

. (3)


Fe =12

∂C

∂x(Udc + uac)

2 , (4)


C = ϵ0Ael

d − x, (5)


isig =∂CU

∂t≈ ∂C

∂tUdc + C0

∂uac

∂t, (6)



η = Udc∂C

∂x≈ Udc

C0

d. (7)


im ≈ ηx,

Fe ≈ ηuac,(8)


Rm =√


Cm = η2/k,


(9)




kaajakari et al.: analysis of phase noise and micromechanical oscillators: ieee transactions on ultrasonics, ferroelectrics, and frequency control, vol. 52, no. 12, december 2005


Non-Linear Spring Constant

• The second-order correction in the spring constant dominates

• Electrostatic non-linearity limits the drive level at high vibration amplitudes.

• The system can become chaotic at high drive amplitudes. The critical amplitude before a bifurcation is given by

20

2324 ieee transactions on ultrasonics, ferroelectrics, and frequency control, vol. 52, no. 12, december 2005

model is used, and the accurate nonlinear model is usedfor the electromechanical transduction [14].

The nonlinear electrostatic spring constants are ob-tained by a series expansion of the electrostatic force:

F =U2

dc

2∂C

∂x. (10)

Including terms up to the second order gives for theelectrostatic spring:

ke(x) = k0e(1 + k1ex + k2ex2)

k0e = −U2DCC0

d2 , k1e =32d

, and k2e =2d2 . (11)

The linear electrostatic spring k0e is negative, and thuslowers the resonance frequency. Of the nonlinear terms, thesecond-order correction k2e can be shown to dominate [1].

The electrostatic nonlinearity limits the resonator drivelevel as at high-vibration amplitudes; the amplitude-frequency curve is not a single valued function and oscilla-tions may even become chaotic [1] [4]. Therefore, the max-imum usable vibration amplitude can be estimated fromthe largest vibration amplitude before a bifurcation. Thiscritical vibration amplitude can be written as [4], [15]:

xc =2!

3√

3Q|κ|, (12)

where:

κ =3k2ek0e

8k− 5k2

1ek20e

12k2 . (13)

Defining the drive level as the motional current throughthe resonator, the maximum drive level is given by (8) and(12) and can be written as:

imaxm = ηω0xc. (14)

As will be seen in Section IV, the maximum drive levelsets the noise floor obtainable with microresonator-basedoscillator.

The analysis in this section assumed that the nonlinear-ity is due to the capacitive spring effects, but (12) and (13)are valid also for cases in which mechanical nonlinearitiesdominate. As an example, in the comb drive actuated res-onators, the capacitance depends linearly on displacement[16]. In such a case, the nonlinear springs and the max-imum vibration amplitude is estimated from mechanicalnonlinearities [4].

III. Aliasing of Noise in Microresonator

In addition to the nonlinear spring effects, the capaci-tive coupling results in up- and down-conversion of noise.As seen from the trigonometric identity:

2 cos∆ωt · cosω0t = cos (ω0 + ∆ω) t + cos (ω0 − ∆ω) t,(15)

Fig. 4. Schematic representation of noise aliasing. Low-frequencynoise un(∆ω) present at filter input is aliased to carrier side-bandsω0 ±∆ω due to mixing in resonator.

the low-frequency noise signal at ∆ω multiplied with thecarrier signal at ω0 results in additional near-carrier noiseside-bands at ω0±∆ω. As illustrated in Fig. 5, this mixingin the resonator causes aliasing of low-frequency noise tocarrier side-bands.

The aliasing of low-frequency noise can be very detri-mental to the oscillator phase noise performance aslow-frequency 1/f -noise can be considerably larger thanthe thermal noise floor. The typical low-frequency noisesources present at the resonator input are the active sus-taining elements (transistors) in the oscillator circuit thatmay have a significant amount of 1/f -noise [17]. Resonatorbiasing also may be noisy, especially if it is implementedwith a charge pump. Notably mechanical 1/f -noise alsomay be significant if the resonator is scaled to nanome-ter scale [18]. However, the noise up-mixing analysis pre-sented here is not limited to a specific noise source as onlythe magnitude of the noise at resonator input is needed topredict noise aliasing in the resonator.

In this section, the noise up-mixing due to the elec-trostatic coupling is analyzed. Mixing due to the electro-static transduction is compared to mixing due to nonlinearspring effects. As Fig. 4 indicates, the analysis in this sec-tion is an open-loop, up-conversion analysis of differentup-mixing mechanisms. The closed-loop oscillator analy-sis, including the effect of positive feedback, is presentedin Section IV.

A. Mixing Due to Capacitive Current Nonlinearity

Fig. 5(a) illustrates how low-frequency noise un at ∆ωis mixed to a higher frequency due to the time varying gapcapacitance. The capacitance is:

C(x) ≈ C0

"1 +

x0

d

#, (16)

where x0 is the resonator displacement at the excitationfrequency. The current through the resonator due to thevoltage un is then:

in =∂(C(x)un)

∂t≈ C0

dx0un + C0un. (17)

The first term in (17) isresponsible for the noise up-conversion and results in noise current at ω0 ± ∆ω. Us-




F =U2

dc

2∂C

∂x. (10)


ke(x) = k0e(1 + k1ex + k2ex2)

k0e = −U2DCC0

d2 , k1e =32d

, and k2e =2d2 . (11)



xc =2!

3√

3Q|κ|, (12)

where:

κ =3k2ek0e

8k− 5k2

1ek20e

12k2 . (13)














C(x) ≈ C0

"1 +

x0

d

#, (16)


in =∂(C(x)un)

∂t≈ C0

dx0un + C0un. (17)





F =U2

dc

2∂C

∂x. (10)


ke(x) = k0e(1 + k1ex + k2ex2)

k0e = −U2DCC0

d2 , k1e =32d

, and k2e =2d2 . (11)



xc =2!

3√

3Q|κ|, (12)

where:

κ =3k2ek0e

8k− 5k2

1ek20e

12k2 . (13)














C(x) ≈ C0

"1 +

x0

d

#, (16)


in =∂(C(x)un)

∂t≈ C0

dx0un + C0un. (17)





F =U2

dc

2∂C

∂x. (10)


ke(x) = k0e(1 + k1ex + k2ex2)

k0e = −U2DCC0

d2 , k1e =32d

, and k2e =2d2 . (11)



xc =2!

3√

3Q|κ|, (12)

where:

κ =3k2ek0e

8k− 5k2

1ek20e

12k2 . (13)














C(x) ≈ C0

"1 +

x0

d

#, (16)


in =∂(C(x)un)

∂t≈ C0

dx0un + C0un. (17)





F =U2

dc

2∂C

∂x. (10)


ke(x) = k0e(1 + k1ex + k2ex2)

k0e = −U2DCC0

d2 , k1e =32d

, and k2e =2d2 . (11)



xc =2!

3√

3Q|κ|, (12)

where:

κ =3k2ek0e

8k− 5k2

1ek20e

12k2 . (13)














C(x) ≈ C0

"1 +

x0

d

#, (16)


in =∂(C(x)un)

∂t≈ C0

dx0un + C0un. (17)




Noise Aliasing in Resonators

• As we have learned in our phase noise lectures, 1/f noise can alias to the carrier through time-varying and non-linear mechanisms. Since 1/f noise is high for CMOS, this is a major limitation

21




F =U2

dc

2∂C

∂x. (10)


ke(x) = k0e(1 + k1ex + k2ex2)

k0e = −U2DCC0

d2 , k1e =32d

, and k2e =2d2 . (11)



xc =2!

3√

3Q|κ|, (12)

where:

κ =3k2ek0e

8k− 5k2

1ek20e

12k2 . (13)














C(x) ≈ C0

"1 +

x0

d

#, (16)


in =∂(C(x)un)

∂t≈ C0

dx0un + C0un. (17)




Mixing: Capacitive Current Non-Linearity

• This term is usually much smaller (by 10X ~ 100X) than mixing due to the force non-linearity

22




F =U2

dc

2∂C

∂x. (10)


ke(x) = k0e(1 + k1ex + k2ex2)

k0e = −U2DCC0

d2 , k1e =32d

, and k2e =2d2 . (11)



xc =2!

3√

3Q|κ|, (12)

where:

κ =3k2ek0e

8k− 5k2

1ek20e

12k2 . (13)














C(x) ≈ C0

"1 +

x0

d

#, (16)


in =∂(C(x)un)

∂t≈ C0

dx0un + C0un. (17)





F =U2

dc

2∂C

∂x. (10)


ke(x) = k0e(1 + k1ex + k2ex2)

k0e = −U2DCC0

d2 , k1e =32d

, and k2e =2d2 . (11)



xc =2!

3√

3Q|κ|, (12)

where:

κ =3k2ek0e

8k− 5k2

1ek20e

12k2 . (13)














C(x) ≈ C0

"1 +

x0

d

#, (16)


in =∂(C(x)un)

∂t≈ C0

dx0un + C0un. (17)



(a) (b)

(c)

Fig. 5. Different mixing mechanism for the noise voltage un at ∆ω tohigh-frequency noise current. (a) Time-varying capacitor (plate dis-placement x) results in up-converted noise current. (b) Square forcelaw results in mixing of noise and signal voltages, un and uac, respec-tively. (c) Nonlinear spring force results in mixing of low-frequencyand signal frequency vibrations.

ing the relation between the signal voltage and resonatordisplacement given by (2) and (8), the up-converted noisecurrent due to capacitive current mixing can be written as:

icn = 2Γcuacun, (18)

where we have defined the current aliasing factor:

Γc =Qω0η2

2kUdc. (19)

As (15) shows, the current icn given by (18) has equalamplitudes at frequencies ω0 ±∆ω.

B. Mixing Due to Capacitive Force Nonlinearity

The second main up-conversion avenue is illustrated inFig. 5(b). Due to the square force law, the low-frequencynoise voltage un at ∆ω is mixed with the high frequencysignal voltage uac at ω0. The capacitive force is given by:

Fn =U2

2∂C

∂x≈ (Udc + uac + un)2

2C0

d

!1 + 2

x0

d

",(20)

where the first three terms of power series expansion of thecapacitance have been kept. The products uacun and xun

result in up-converted noise at ω0 ± ∆ω. Thus, the forceat ω0 ±∆ω is:

Fn(ω0 ±∆ω) ≈ C0

duacun + 2

C0

d

x0

dUdcun.

(21)

This high-frequency noise force near the resonator reso-nance excites the resonator, and the displacement is givenby (3). Close to the resonance, the noise-induced displace-ment is:

xFn ≈ −jQ

Fn

k, (22)

and the resulting noise current is:

iFn = ηxn = −jηω0xn. (23)

Substituting (21) and (22) to (23) and using x0 =−jQηuac/k leads to:

iFn = 2ΓF uacun, (24)

where we have defined the force aliasing factor:

ΓF ≈ Qω0η2

2kUdc

#1 − j2

QηUdc

kd

$. (25)

Here the first term in brackets is due to the squareforce law [product uacun in (20)], and the second termin brackets is due to the nonlinear capacitance [prod-uct xun in (20)]. If we had kept only the first term ofthe power series expansion of capacitance [linear capaci-tance, C(x) = C0(1 + x/d)], then the force aliasing factorwould be:

ΓF ≈ Qω0η2

2kUdc(linear C), (26)

which is the same as the current aliasing factor givenby (19).

It is of interest to compare the two terms in brackets in(25). Substituting typical microresonator parameters (Ta-ble I) gives (2QηUdc/kd) ≈ 63. Thus, the second term inbrackets in (25) is the dominant term and the noise alias-ing could be significantly reduced with a linear couplingcapacitor.


(a) (b)

(c)





Γc =Qω0η2

2kUdc. (19)




Fn =U2

2∂C


2C0

d

!1 + 2

x0

d

",(20)



Fn(ω0 ±∆ω) ≈ C0

duacun + 2

C0

d

x0

dUdcun.

(21)


xFn ≈ −jQ

Fn

k, (22)






ΓF ≈ Qω0η2

2kUdc

#1 − j2

QηUdc

kd

$. (25)


ΓF ≈ Qω0η2





(a) (b)

(c)





Γc =Qω0η2

2kUdc. (19)




Fn =U2

2∂C


2C0

d

!1 + 2

x0

d

",(20)



Fn(ω0 ±∆ω) ≈ C0

duacun + 2

C0

d

x0

dUdcun.

(21)


xFn ≈ −jQ

Fn

k, (22)






ΓF ≈ Qω0η2

2kUdc

#1 − j2

QηUdc

kd

$. (25)


ΓF ≈ Qω0η2






Mixing: Capacitive Force Non-Linearity

• The form is the same as the capacitance non-linearity, but the magnitude is much higher and dominates for most resonators. A linear coupling capacitor has much reduced noise up-conversion

23


(a) (b)

(c)





Γc =Qω0η2

2kUdc. (19)




Fn =U2

2∂C


2C0

d

!1 + 2

x0

d

",(20)



Fn(ω0 ±∆ω) ≈ C0

duacun + 2

C0

d

x0

dUdcun.

(21)


xFn ≈ −jQ

Fn

k, (22)






ΓF ≈ Qω0η2

2kUdc

#1 − j2

QηUdc

kd

$. (25)


ΓF ≈ Qω0η2





(a) (b)

(c)





Γc =Qω0η2

2kUdc. (19)




Fn =U2

2∂C


2C0

d

!1 + 2

x0

d

",(20)



Fn(ω0 ±∆ω) ≈ C0

duacun + 2

C0

d

x0

dUdcun.

(21)


xFn ≈ −jQ

Fn

k, (22)






ΓF ≈ Qω0η2

2kUdc

#1 − j2

QηUdc

kd

$. (25)


ΓF ≈ Qω0η2





(a) (b)

(c)





Γc =Qω0η2

2kUdc. (19)




Fn =U2

2∂C


2C0

d

!1 + 2

x0

d

",(20)



Fn(ω0 ±∆ω) ≈ C0

duacun + 2

C0

d

x0

dUdcun.

(21)


xFn ≈ −jQ

Fn

k, (22)






ΓF ≈ Qω0η2

2kUdc

#1 − j2

QηUdc

kd

$. (25)


ΓF ≈ Qω0η2





(a) (b)

(c)





Γc =Qω0η2

2kUdc. (19)




Fn =U2

2∂C


2C0

d

!1 + 2

x0

d

",(20)



Fn(ω0 ±∆ω) ≈ C0

duacun + 2

C0

d

x0

dUdcun.

(21)


xFn ≈ −jQ

Fn

k, (22)






ΓF ≈ Qω0η2

2kUdc

#1 − j2

QηUdc

kd

$. (25)


ΓF ≈ Qω0η2





(a) (b)

(c)





Γc =Qω0η2

2kUdc. (19)




Fn =U2

2∂C


2C0

d

!1 + 2

x0

d

",(20)



Fn(ω0 ±∆ω) ≈ C0

duacun + 2

C0

d

x0

dUdcun.

(21)


xFn ≈ −jQ

Fn

k, (22)






ΓF ≈ Qω0η2

2kUdc

#1 − j2

QηUdc

kd

$. (25)


ΓF ≈ Qω0η2






Mixing: Non-Linear Spring Force

• Amplitude of noise at low-frequency is very small due to resonator Q. The noise is up-converted through the spring non-linearity.

• This term is the smallest of the three, about 500X smaller than the capacitance non-linearity.

24


TABLE IResonator Characteristics at Udc = 20 V.

Parameter Symbol Value Units

Resonance frequency f0 13.2 [MHz]Effective spring constant k 16.8 [MN/m]Effective mass m 2.44 [nkg]Quality factor Q 100,000Electrode area Ael 12,000 [µm2]Transducer gap d 0.2 [µm]Motional capacitance Cm 168 [aF]Motional inductance Lm 867 [mH]Motional resistance Rm 718 [Ω]Critical amplitude xc 36.0 [nm]

C. Mixing Due to Nonlinear Spring Force

The up-conversion avenue due to nonlinear spring forceis illustrated in Fig. 5(c). The force due to the noise voltageFn = ηun results in low-frequency resonator vibrations.Because these vibrations are far from the resonance, theamplitude is given by:

xn = H(ω)Fn ≈ ηun

k. (27)

Due to the nonlinear spring effects, these low-frequencyvibrations are multiplied with the vibrations at the signalfrequency. Assuming spring force F = k0x(1 + k1x) andsubstituting x = x0 + xn, the up-converted noise force is:

F kn = 2k0k1x0xn. (28)

The resulting current can be evaluated as in Section III-B. Assuming that the nonlinear spring is dominated bythe capacitive effects given by (11), the up-converted noisecurrent due to nonlinear spring mixing is given by:

ikn = 2Γkuacun, (29)

where we have defined the spring aliasing factor:

Γk = j3Q2ω0η4Udc

2d2k3 . (30)

D. Comparison of Mixing Mechanisms

The ratio of aliasing factors due to the current up-conversion and nonlinear spring mixing given by (19) and(30), respectively, is:

!!!!Γc

Γk

!!!! =1

3Q

"dk

ηUDC

#2

. (31)

Substituting typical microresonator parameters (Ta-ble I) into (31) gives |Γc/Γk| ≈ 500. Thus, we can concludethat the main aliasing mechanism is the nonlinear electro-static transduction (capacitive current and force nonlin-earity) and not the nonlinear spring effects.

For a linear capacitance C(x) = C0(1 + x/d), the alias-ing factors for the current and force up-conversions given

by (19) and (26), respectively, are equal. Substituting typ-ical microresonator parameters (Table I) into (31) gives|Γc| = |ΓF | = 34.8 µA/V2 (linear C). With the inclusionof the nonlinear terms in capacitance, the force aliasingfactor given by (25) is |ΓF | = 2.2 mA/V2. As noted be-fore, this is 63 times larger than for a linear capacitance.Thus, for a parallel plate coupling capacitor, the force up-mixing is the dominant aliasing path.

E. Simulation of Noise Aliasing

To verify the analytical results, the aliasing of a low-frequency signal was simulated with a harmonic balancecircuit simulator [14]. The resonator was excited using ahigh-frequency signal at the resonance frequency and asmall-frequency signal at frequency ∆f . Fig. 6(a) showsthe simulated aliasing factors Γ at different frequencies∆f obtained using the accurate model with all capaci-tive nonlinearities included [14]. At small-frequency off-sets (∆f < f0/2Q) the magnitude of aliasing factor isvery close to the analytical estimate of |Γ| = 2.2 mA/V2

given by:

Γ = ΓF + Γc, (32)

where ΓF and Γc are given by (25) and (19), respectively.Outside the resonator bandwidth (∆f > f0/2Q), the alias-ing is significantly reduced as the motion is not enhancedby the resonator quality factor.

For a linear capacitor, the aliasing shown in Fig. 6(b)is smaller by a factor of 63. Again, the result is close tothe analytical result of Γ = 69.6 µA/V2 given by the sumof (19) and (26). Figs. 6(c) and (d) show the aliasing withonly the force and current nonlinearity included in themodel, respectively. It is shown that, for small-frequencyoffsets (∆f < f0/2Q), both effects are equal and agreewith (19) and (26). At higher offsets, the effect of forcenonlinearity is reduced as the resonator is not excited farfrom the resonance. The aliasing due to current nonlinear-ity does not show this effect as it is due to direct modula-tion of the capacitance.

To further validate the noise aliasing analysis, the alias-ing factors were simulated at different noise levels. Thealiasing factors remained unchanged to noise voltages lessthan 10 mV. At higher noise levels, the oscillation fre-quency was changed as the noise started to be significantin comparison to bias voltage. As the typical noise levelsare less than 1 µV/

√Hz, the first-order mixing analysis

is enough for accurate estimation of noise aliasing in theresonator.

IV. Analytical Phase Noise Model

Here an analytical model for the noise in a closed-loop micromechanical oscillator is developed. The modelis based on well-known Leeson’s model for the phase noise[7] [8] [10] and is expanded to incorporate the 1/f -noisealiasing in microresonators.


(a) (b)

(c)





Γc =Qω0η2

2kUdc. (19)




Fn =U2

2∂C


2C0

d

!1 + 2

x0

d

",(20)



Fn(ω0 ±∆ω) ≈ C0

duacun + 2

C0

d

x0

dUdcun.

(21)


xFn ≈ −jQ

Fn

k, (22)






ΓF ≈ Qω0η2

2kUdc

#1 − j2

QηUdc

kd

$. (25)


ΓF ≈ Qω0η2











k. (27)


F kn = 2k0k1x0xn. (28)




Γk = j3Q2ω0η4Udc

2d2k3 . (30)



!!!!Γc

Γk

!!!! =1

3Q

"dk

ηUDC

#2

. (31)






given by:

Γ = ΓF + Γc, (32)















k. (27)


F kn = 2k0k1x0xn. (28)




Γk = j3Q2ω0η4Udc

2d2k3 . (30)



!!!!Γc

Γk

!!!! =1

3Q

"dk

ηUDC

#2

. (31)






given by:

Γ = ΓF + Γc, (32)















k. (27)


F kn = 2k0k1x0xn. (28)




Γk = j3Q2ω0η4Udc

2d2k3 . (30)



!!!!Γc

Γk

!!!! =1

3Q

"dk

ηUDC

#2

. (31)






given by:

Γ = ΓF + Γc, (32)










FBAR Resonator

• Another “MEMS” technology is the Thin Film Bulk Wave Acoustic Resonators (FBAR)

• It uses a thin layer of Aluminum-Nitride piezoelectric material sandwiched between two metal electrodes

• The FBAR has a small form factor and occupies only about 100µm x 100µm.

25

37

2.4.1 FBAR Resonator

The FBAR resonator [Ruby01] consists of a thin layer of Aluminum-Nitride

piezoelectric material sandwiched between two metal electrodes. The entire structure is

supported by a micro-machined silicon substrate as shown in Fig. 2.10. The metal/air

interfaces serve as excellent reflectors, forming a high Q acoustic resonator. The FBAR

has a small form factor and occupies only about 100!m x 100!m.

Fig. 2.10: (Left) structure (right) photograph of a FBAR resonator.

The FBAR resonator can be modeled using the Modified Butterworth Van Dyke circuit

as shown in Fig. 2.11 [Larson00]. Lm, Cm and Rm are its motional inductance,

capacitance and resistance respectively. Co models the parasitic parallel plate capacitance

between the two electrodes and Cp1 and Cp2 accounts for the electrode to ground

capacitances. Losses in the electrode are given by R0, Rp1 and Rp2.

Fig. 2.11 Circuit model of the FBAR resonator.

Rp

Cp1

Lm Cm Rm

C0 R0

Rp

Cp2

Electrodes

Air

Air

AlN

Si Si

Drive Electrode

Sense

Electrode

100 !m


FBAR Resonance

• Very similar to a XTAL resonator. Has two modes: series and parallel

• Unloaded Q ~ 1000

• This technology will not be integrated directly with CMOS, but there is a potential for advanced packaging or procesing.

26

37

2.4.1 FBAR Resonator

The FBAR resonator [Ruby01] consists of a thin layer of Aluminum-Nitride

piezoelectric material sandwiched between two metal electrodes. The entire structure is

supported by a micro-machined silicon substrate as shown in Fig. 2.10. The metal/air

interfaces serve as excellent reflectors, forming a high Q acoustic resonator. The FBAR

has a small form factor and occupies only about 100!m x 100!m.

Fig. 2.10: (Left) structure (right) photograph of a FBAR resonator.

The FBAR resonator can be modeled using the Modified Butterworth Van Dyke circuit

as shown in Fig. 2.11 [Larson00]. Lm, Cm and Rm are its motional inductance,

capacitance and resistance respectively. Co models the parasitic parallel plate capacitance

between the two electrodes and Cp1 and Cp2 accounts for the electrode to ground

capacitances. Losses in the electrode are given by R0, Rp1 and Rp2.

Fig. 2.11 Circuit model of the FBAR resonator.

Rp

Cp1

Lm Cm Rm

C0 R0

Rp

Cp2

Electrodes

Air

Air

AlN

Si Si

Drive Electrode

Sense

Electrode

100 !m

38

The frequency response of the FBAR resonator is shown in Fig. 2.12. The FBAR

behaves like a capacitor except at its series and parallel resonance. It achieves an

unloaded Q of more than a 1000.

Fig. 2.12 Frequency response of the FBAR resonator.

2.4.2 Advantages of FBAR Resonator

1. High Q factor

The Q factor of the FBAR resonator is more than 1000, which is much higher than the

Q-factor of an on-chip LC resonator. The high Q factor allows implementation of low

loss filters and duplexers to attenuate the out of band blockers and reject the image

signals. In some applications, the bandwidth of these FBAR filters is sufficiently small

for channel filtering, relaxing the linearity requirement of mixers and removing the need

for baseband/IF filters. The high Q FBAR resonator also substantially improves

oscillator’s phase noise and reduces its power consumption. It could also potentially

100M 1G 10G1

10

100

1000

Impe

da

nce

(!

)

Frequency (Hz)

Parallel

resonance

Series

resonance


FBAR Oscillator

• Rm ~ 1 ohm

• gm ~ 7.8 mS used (3X)

• C1=C2=.7pF

• gm/Id ~ 19, Id ~ 205μA

• Start-up behavior shown below:

27 44

performance oscillators [Chee05a].

The schematic of the low power FBAR oscillator is shown in Fig. 3.3. The Pierce

oscillator uses a CMOS inverting amplifier comprising of transistors M1 and M2. The

FBAR is modeled using the modified Butterworth Van Dyke model [Larson00].

Capacitance C1 and C2 include the capacitances due to the FBAR electrodes, transistors,

pad and interconnects. A large resistor Rb is used to bias the gate and drain voltage of the

transistors at Vdd/2 to maximize the allowable voltage swing and minimize its loading on

the FBAR. Transistors M1 and M2 share the same current but their transconductances gm1

and gm2 sum, reducing the current needed for oscillation by half. The transistors are also

designed to operate in the sub-threshold regime to obtain higher current efficiency.

Fig. 3.3: Schematic of an ultra low power FBAR oscillator.

Capacitors C1 and C2 transform the amplifier’s transconductance 21 mmm ggg !" into a

negative resistance 21

2

21

CC

gg mm

#

!$ at frequency #. Thus, a higher negative resistance

C2 C1

M1

M2

Rb

X Y

FBAR

Vdd

C0

Lm Cm Rm

R0

44

performance oscillators [Chee05a].

The schematic of the low power FBAR oscillator is shown in Fig. 3.3. The Pierce

oscillator uses a CMOS inverting amplifier comprising of transistors M1 and M2. The

FBAR is modeled using the modified Butterworth Van Dyke model [Larson00].

Capacitance C1 and C2 include the capacitances due to the FBAR electrodes, transistors,

pad and interconnects. A large resistor Rb is used to bias the gate and drain voltage of the

transistors at Vdd/2 to maximize the allowable voltage swing and minimize its loading on

the FBAR. Transistors M1 and M2 share the same current but their transconductances gm1

and gm2 sum, reducing the current needed for oscillation by half. The transistors are also

designed to operate in the sub-threshold regime to obtain higher current efficiency.

Fig. 3.3: Schematic of an ultra low power FBAR oscillator.

Capacitors C1 and C2 transform the amplifier’s transconductance 21 mmm ggg !" into a

negative resistance 21

2

21

CC

gg mm

#

!$ at frequency #. Thus, a higher negative resistance

C2 C1

M1

M2

Rb

X Y

FBAR

Vdd

C0

Lm Cm Rm

R0

47

Fig. 3.6: Measured startup transients of an FBAR oscillator.

1. Initial power up. When the oscillator is powered up, the supply charges the gate of

the transistors to their biasing voltage through Rb with a time constant !initial ~ RbC1.

With Rb = 60 k" and C1 = 700 fF, the biasing point can be reached in ~ 3*!initial.

(~126 ns). Smaller Rb results in a shorter !initial but increases its loading on the

resonator. With the FBAR impedance equals to ~ 2 k" at parallel resonance, Rb

loads the resonator by ~ 3%.

2. Exponential growth. Once the operating point is reached, the amplifier acquires

sufficient loop gain and the oscillation amplitude builds up exponentially with a

time constant !exp = mXY

m

RZ

L

#$

]Re[. A higher Q resonator has a larger Lm and/or

smaller Rm, which leads to a lower steady state power consumption but longer

startup time for a given Re[ZXY]. For Lm = 147 nH, Rm = 0.9" and if Re[ZXY] is

chosen to be -3* Rm, !exp = 82 ns. With an initial voltage of ~1%V, it takes 12.6!exp &

Oscillator

turns on Exponential

growth

Gain compression

Steady state

oscillation VDD

gating

signal

Oscillator

transient

response


Measured Results on FBAR Osc

• Operate oscillator in “current limited” regime

• Voltage swing ~ 167 mV, Pdc ~ 104 μW

28

49

Fig. 3.7: Die photo of the FBAR oscillator

3.2.4 Measured Results

The oscillator is self-biased with a 430mV supply and dissipates 89µW for sustained

oscillation at 1.882 GHz. The measured zero to peak output voltage swing is 142mV.

The output spectrum of the oscillator is shown in Fig. 3.8. A clean output signal is

obtained and no close-in spurs are observed. Second, third, fourth and fifth harmonics are

measured to be -43.8 dBc, -45.5 dBc, -68.8 dBc and -69.7 dBc respectively.

Fig. 3.8: Output frequency spectrum of FBAR oscillator.

Force electrode

Sense electrode FBAR

CMOS Die

800µm

Bond wires

49

Fig. 3.7: Die photo of the FBAR oscillator

3.2.4 Measured Results

The oscillator is self-biased with a 430mV supply and dissipates 89µW for sustained

oscillation at 1.882 GHz. The measured zero to peak output voltage swing is 142mV.

The output spectrum of the oscillator is shown in Fig. 3.8. A clean output signal is

obtained and no close-in spurs are observed. Second, third, fourth and fifth harmonics are

measured to be -43.8 dBc, -45.5 dBc, -68.8 dBc and -69.7 dBc respectively.

Fig. 3.8: Output frequency spectrum of FBAR oscillator.

Force electrode

Sense electrode FBAR

CMOS Die

800µm

Bond wires

50

The measured phase noise performance is shown in Fig. 3.9. The oscillator achieves a

phase noise of -98dBc/Hz and -120dBc/Hz at 10kHz and 100kHz offsets respectively.

The good phase noise performance is mainly attributed to the high Q FBAR resonator.

Fig. 3.9: Measured phase noise performance of FBAR oscillator.

A better phase noise performance is obtained by operating the oscillator at the edge of

the current limited regime [Ham01]. Fig. 3.10 shows the output voltage swing and

measured phase noise at various power consumptions. The optimal measured phase

noise is -100 dBc/Hz at 10kHz offset and -122 dBc/Hz at 100kHz offset and it occurs

when the output voltage swing is 167mV with the oscillator consuming 104µW. Beyond

this operating point, the oscillator transits into the voltage limited regime which the

transistor’s output resistance decreases and loads the FBAR, resulting in a poorer phase

noise performance.

10k 100k 1M 10M-140

-130

-120

-110

-100

-90

Phase N

ois

e (

dB

c/H

z)

Frequency offset (Hz)

-98 dBc/Hz

-120 dBc/Hz

Instrument’s

noise floor

MEMS Reference Oscillators - RFICrfic.eecs.berkeley.edu/ee242/pdf/Module_7_1_MEMS_XTAL.pdf · LIN...

Documents

Transcript of MEMS Reference Oscillators - RFICrfic.eecs.berkeley.edu/ee242/pdf/Module_7_1_MEMS_XTAL.pdf · LIN...