Available online at www.sciencedirect.

com

Sensors and Actuators A 142 (2008) 306315

Modelling the dynamics of a MEMS resonator: Simulations and experiments

R.M.C. Mestrom a, , R.H.B. Fey a , J.T.M. van Beek b , K.L. Phan b , H. Nijmeijer a
Eindhoven University of Technology, Department of Mechanical Engineering, Dynamics and Control Group, P.O. Box 513, 5600 MB, Eindhoven, The Netherlands b NXP Semiconductors, Research Laboratories, High Tech Campus 4, P.O. Box WAG 02, 5656 AE Eindhoven, The Netherlands Received 29 September 2006; received in revised form 7 February 2007; accepted 10 April 2007 Available online 19 April 2007
a

Abstract Nonlinearities in MEMS silicon resonators are caused by different effects. Depending on the resonator layout, different nonlinearities may be dominant in the resonator response. Based on experimental results, a modelling approach is proposed to arrive at a nonlinear dynamic model that potentially captures the observed behaviour. Apart from the model, consisting of a mechanical and an electrical (measurement) part, the effect of thermal noise is also estimated. With the proposed model, a quantitative match between the simulation and experimental results is established such that a good starting point is achieved for a more thorough modelling procedure. 2007 Elsevier B.V. All rights reserved.
Keywords: MEMS resonators; Nonlinear oscillations; Modelling; Experiments

1. Introduction Micro-electromechanical silicon resonators provide an interesting alternative for quartz crystals as accurate timing devices in oscillators for modern data and communication applications. Their compact size, feasibility of integration with IC technology and low cost are major advantages. However, as the resonators are small in size, they have to be driven close to or even into nonlinear regimes in order to store enough energy for a sufciently good signal to noise ratio [1]. Depending on the specic resonator layout, different nonlinearities may be dominant in the resonator dynamic behaviour. The presence of the nonlinearities is relevant for oscillator performance and has to be incorporated in future resonator design optimisation. In order to determine the inuence of resonator nonlinearities on the performance of oscillators, the dynamic behaviour of resonators has to be understood. Clampedclamped beam resonators have been realised and measured by various research groups, see for instance [2]. However, a combined numerical and experimental analysis of the nonlinear behaviour has not been addressed extensively. In this paper, a heuristic approach

for modelling the dynamics in a clampedclamped beam MEMS resonator will be proposed. With the model, a quantitative match between numerical and experimental investigations of nonlinear dynamics in a clampedclamped beam resonator will be established. The current paper is a much more extensive version of [3]. It contains more details of the modelling of the resonator. Moreover, new experimental and numerical results will be presented. The steady-state dynamic behaviour of the resonator can be described by a dynamic characteristic called an amplitudefrequency curve. The outline for this paper is as follows. First, in Section 2, experimental results of a clampedclamped beam resonator will be discussed. Next, in Section 3, a modelling approach will be proposed to describe the dynamic effects observed in the measurements. Additionally, the numerical analysis will be described briey. Experimental and numerical results will be compared in Section 4. Finally, in Section 5, some conclusions will be drawn. 2. Experiments using a clampedclamped beam resonator An example of a clampedclamped beam resonator is depicted in Fig. 1. Its characteristic vibration shape is shown

Corresponding author. Tel.: +31 40 247 2811; fax: +31 40 246 1418. E-mail address: r.m.c.mestrom@tue.nl (R.M.C. Mestrom).

0924-4247/$ see front matter 2007 Elsevier B.V. All rights reserved. doi:10.1016/j.sna.2007.04.025

R.M.C. Mestrom et al. / Sensors and Actuators A 142 (2008) 306315

307

Fig. 1. Schematic layout of a clampedclamped beam resonator. (a) Characteristic vibration shape and (b) actuation by a dc and an ac voltage component.

in Fig. 1a. Due to out-of-plane vibration, the resonator is often called a exural resonator. The actuation of the resonator is realised by means of a dc (Vdc ) and an ac (Vac ) voltage component, which are applied to the electrodes of the resonator by means of bias tees, see Fig. 1b. During experiments, the resonator output voltage Vout is measured. This quantity is related to the beam motion, as will become clear from Section 3. Resonators are fabricated using silicon-on-insulator (SOI) wafers. The fabrication of the resonator is schematically depicted in Fig. 2. First, aluminum bondpads are dened on the wafer surface, see Fig. 2a. A layer of resist is deposited on top of the wafer and the resonator layout is dened by means of lithography. Next, the resonator layout is etched into the 1.4 m thick SOI layer down to the buried oxide layer by means of deep reactive ion etching (DRIE) (Fig. 2b). The resist is removed (Fig. 2c) and, nally, the resonator is released from the substrate through isotropic etching of the buried oxide layer using an HF wet etch solution (Fig. 2d). A microscope image of the clampedclamped beam resonator can be seen in Fig. 3. Here, the grey material is silicon (Si), thin dark lines are lithography etch gaps and the white, grainy material corresponds to the aluminum (Al) bond pads and electrical lines. Six aluminum bond pads can be distinguished. These are designed in such a way that they t the ground-signal-ground probes that are used during the measurements. The outer four bond pads are connected to ground, such that the beam itself is grounded. The middle two bond pads are used for actuation and measurement purposes.

Fig. 3. Clampedclamped beam resonator (beam dimensions: length, 44 m; width, 4 m; thickness, 1.4 m).

During the experiments, the MEMS resonator is located in a vacuum (pressure p = 4.6 104 mbar). The steady-state dynamic behaviour of the resonator is investigated by constructing a so-called amplitudefrequency plot. For this purpose, the ac excitation frequency f is slowly increased (sweep up) and decreased (sweep down) in steps of 250 Hz around the fundamental frequency of the resonator. At each frequency, after 2 s, the output voltage Vout is measured, which is a measure for the beam motion (see also Section 3.2). An example of a typical amplitudefrequency curve for the clampedclamped beam considered is depicted in Fig. 4. The excitation is given by a dc bias voltage of Vdc = 70 V and an ac excitation amplitude of Vac = 139 mV. From this gure, it can be seen that the resonance frequency of the resonator is approximately 12.875 MHz. However, due to nonlinearities, the resonance peak is bent to the left (lower frequencies). As a result, the steady-state dynamic behaviour of the resonator is found to depend on the sweep direction. Sudden jumps in the response, indicated by black arrows, occur at dif-

Fig. 2. Fabrication of the resonator. (a) Dene Al bond pads, (b) DRIE of top Si layer, (c) remove resist and (d) HF wet etching of SiO2 .

Fig. 4. Example of an amplitudefrequency curve.

308

R.M.C. Mestrom et al. / Sensors and Actuators A 142 (2008) 306315

ferent frequencies. This is called frequency hysteresis and has also been reported in ref. [4]. However, in ref. [4], a longitudinal mode beam resonator was investigated. Frequency hysteresis is a typical nonlinear effect, which may be caused by softening nonlinear behaviour. This will be explained in Section 3. 3. Modelling approach 3.1. Mechanical model The nonlinear phenomena observed in the measurement (Fig. 4) demand for a nonlinear dynamical model. A single degree-of-freedom (1DOF) model, in line with [1,4], that may capture these phenomena is a Dufng-like model. The Duffing equation is a classical nonlinear dynamic equation which can be used to describe the response of a periodically excited mass-spring-damper system with linear viscous damping force and a spring force containing both a linear and a cubic term. From literature, it is known that a Dufng model may describe nonlinear effects like softening and hardening behaviour and sub- and superharmonic resonances [57]. In contrast to a more rst principles-based approach, as for instance found in refs. [810], a more or less heuristic approach is utilised here. The single degree-of-freedom model for the mechanical part will form a suitable starting point for gaining insight in the nonlinear dynamics of the MEMS resonator, without the need for extensive distributed modelling. On the one hand, essential nonlinear dynamics are believed to be captured by a 1DOF model (see also refs. [1,4]) and it will allow for a more straightforward nonlinear dynamic analysis. On the other hand, results will be less accurate than the ones based on a distributed modelling approach. The following model is proposed, which describes the dynamic behaviour of the exural displacement x of the beam resonator: m + b + k(x)x = Fe (x), x x (1)

Taylor expansion of the electrostatic force in x: Fe (x, t) = 1 C0 2 V 2 d0 dc + 4 x x3 + 8 3 + h.o.t. d0 d0

1 C0 2 (2Vdc Vac sin(2ft) + Vac sin2 (2ft)) 2 d0 x x2 x3 + 3 2 + 4 3 + h.o.t. d0 d0 d0 (4)

1+2

reveals that this force introduces softening nonlinear behaviour. In (4), h.o.t. denotes higher-order terms. The nonlinearity in spring stiffness results from a combination of softening silicon material nonlinearities (higher-order elastic effects, see refs. [1,11]) and hardening geometric nonlinearities (mid-plane stretching of a clampedclamped beam, see for instance refs. [5,6]). A starting point for the nonlinear stiffness function, including terms up to fourth order is written as: k(x) = k0 + k1 x + k2 x2 + k3 x3 + k4 x4 , (5)

where k0 is the stiffness parameter in the linear part of the spring force and k1 , k2 , k3 and k4 denote the stiffness parameters in the nonlinear part of the spring force. This approach allows for extension of the stiffness function to arbitrary order. Depending on the specic layout of the clampedclamped silicon beam resonator, k1 to k4 may become positive or negative, resulting in hardening or softening behaviour, respectively. In ref. [4], a softening spring stiffness was observed (only terms k1 and k2 for a longitudinal bulk-mode beam resonator). The higher-order stiffness parameters k3 and k4 have not yet been found in literature related to MEMS resonators. However, as will become clear from Section 4, where the match between simulations and experiments will be established, these terms have to be included, at least in the current model. 3.2. Measurement circuit An electrical circuit is used for the actuation and detection of the beam motion. The electrical circuit representation of Fig. 1b is shown in Fig. 5a. In order to be able to measure the ac component of Vout , dc decoupling is present by means of an additional decoupling capacitor Cdec . This is depicted in Fig. 5a. The resonator output voltage is measured on a 50 resistor (R2 ). In order to relate the output voltage Vout , measured on R2 , to the beam motion, consider Fig. 5b, which contains only the output part of the electrical circuit (grey part in Fig. 5a). C(x) denotes the variable resonator capacitance. The differential equations with the two unknowns v and i for this electrical circuit are determined from node analysis. By using Vout = iR2 , two coupled differential equations for current i and voltage v are obtained: C(x) Cdec dv = dt 1 Vdc C(x) x vi+ + , R1 x R1 = i, (6) (7)

where m, b and k(x) are the effective (lumped) mass, damping and nonlinear stiffness of the system, respectively. Since the beam is excited over the total beam length (see Fig. 1), x is some characteristic displacement measure for the beam, vibrating in the rst mode. Furthermore, x and x denote the rst and second time derivative of x, respectively. The electrostatic force Fe (x,t) is given by: Fe (x, t) = 1 C 0 d0 1 C 0 d0 V 2 (t) V 2, 2 (d0 x)2 1 2 (d0 + x)2 2 (2)

where C0 is the capacitance over the gap when x = 0 and d0 is the corresponding initial gap width. V1 (t) and V2 denote the applied voltages on the electrodes and are written as: V1 (t) = Vdc + Vac sin(2ft), V2 = Vdc , (3)

where Vdc is the so-called bias voltage, and Vac and f are the amplitude and frequency of the ac voltage, respectively. The capacitance on the readout side for arbitrary x is given by C(x) = C0 d0 /(d0 + x).

dv di R2 dt dt

R.M.C. Mestrom et al. / Sensors and Actuators A 142 (2008) 306315

309

or Vac x, gives m + b + x x k0 2 C0 2 V 2 dc d0 x= C0 Vdc Vac sin(2ft). d0 (9)

2 2 Next, the modied stiffness k0,e = k0 2(C0 /d0 )Vdc and the electromechanical coupling coefcient (see ref. [14]) = (C0 /d0 )Vdc are dened. In this way, (9) can be rewritten as

m + b + k0,e x = Vac sin(2ft). x x

(10)

By using the electromechanical coupling coefcient, a relation between the electrical current and the mechanical velocity can be established as im = [14]. Substitution of this relation into x (10) followed by division by results in:
Fig. 5. Measurement conguration for the clampedclamped beam resonator. (a) Vout measurement and (b) output circuit.

m dim k0,e b + 2 im + 2 2 dt

im dt = Vac sin(2ft).

(11)

Eq. (11) can be rewritten as a harmonically excited RLC circuit: Lm dim 1 + R m im + dt Cm im dt = Vac sin(2ft). (12)

where, in (6), the time derivative of the resonator capacitance C(x) is elaborated as dC(x)/dt = (C(x)/x) . x 3.3. Thermal noise In every electrical system, thermal agitation of the electrons inside electrical conductors forms a source of noise. This noise is called JohnsonNyquist or thermal noise [12,13]. In ref. [13], the expression for the root mean square (rms) of the noise voltage vn generated in a resistor is given as vn = 4kB TR f, (8)

in which the motional inductance, motional resistance and motional capacitance (see ref. [14]) are dened as Lm = m , 2 Rm = b = 2 k0,e m 2 Q and Cm = 2 , k0,e (13)

where kB denotes Boltzmanns constant, T the absolute temperature, R the resistor value and f is the bandwidth in which the noise is measured. An estimate for the thermal noise can be found by using the resonator motional resistance Rm for the resistor value R in (8). An expression for Rm will be derived next. Due to presence of other resistors in the measurement circuit, the estimate for vn can be considered to be conservative. In order to determine the noise generated in the circuit with the clampedclamped beam resonator, the resistance value of its equivalent circuit representation can be used in (8). For this purpose, a linear (ki = 0, i = 1, . . ., 4) mechanical model is considered, in which the electrostatic force is expanded in a Taylor series around x = 0. Furthermore, it is assumed that the ac voltage is much smaller than the bias voltage (Vac Vdc ) and that small 2 beam motions occur (x d0 ). Neglecting terms containing Vac

respectively. Note that the quality factor Q is dened in a linear sense: Q = k0 m/b. By (13), the mechanical parameters of the resonator are related to RLC parameters of the equivalent circuit. The motional resistance Rm can be used to obtain a conservative estimate for the thermal noise using (8). Other noise sources or effects might also be present in the system, both in the actuation/sensing of the resonator and in the measurement equipment. However, within the scope of the current modelling approach, individual noise effects cannot be distinguished properly and, therefore, the discussion will be limited to thermal noise. 3.4. Numerical approach

The total simulation model for the clampedclamped beam resonator and the measurement circuit consists of a statespace description of the 1DOF resonator model (1) (using (2), (3) and (5)), together with the equations for the measurement circuit (6)(7). Therefore, dene the state column T T x = [ x x v i ] = [ x1 x2 x3 x4 ] . In this way, the total model can be written in a form that is more suitable for numerical implementation, as x = f (x, t), where x2 1 1 C 0 d0 1 C 0 d0 V 2 (t) V2 bx2 k(x1 ) + 2 1 2 2 m 2 (d0 x1 ) 2 (d0 + x1 ) f (x, t) = (14) 1 d0 + x1 C0 d0 Vdc . x2 x3 + x4 2 C0 d0 R1 R1 (d0 + x1 ) d0 + x1 1 C0 d0 C 0 d0 Vdc x4 x2 x3 + 1 + C 0 d0 R 2 R1 (d0 + x1 )Cdec R1 (d0 + x1 )2

310

R.M.C. Mestrom et al. / Sensors and Actuators A 142 (2008) 306315

In the numerical simulations, this representation of the total model allows for the mechanical part to be calculated rst and and electrical (measurement) part to be calculated afterwards, since the states x3 and x4 do not appear in the rst two elements (rows) of (14). The steady-state nonlinear dynamic behaviour of the total model is investigated. Hereto, numerical collocation and continuation techniques, available in the numerical package AUTO [15], are applied to the mechanical model for determining periodic solutions (resonator vibrations) for varying excitation frequency f. Next the results from the mechanical model are used as input for the measurement circuit part and time integration is applied to calculate the response of the measurement circuit in order to obtain Vout . Results will be discussed in Section 4. 4. Numerical and experimental results For comparing numerical and experimental results, a series of measurements has been performed on the clampedclamped beam resonator. The bias voltage during all measurements has been kept constant to Vdc = 70 V and the ac amplitude Vac has been varied between 20 and 349 mV. First, the measurement of the (almost) linear response at the lowest excitation value will be used to estimate the motional resistance Rm and, thus, the thermal noise in the measurements. The result of the measurement and the numerical t is depicted in Fig. 6. In this gure, the black curves (both solid and dashed) denote experimentally determined peak to peak values of periodic solutions. These curves are rather noisy, since, for low ac excitation values, the peak to peak value of Vout is dominated by noise. The solid grey line denotes simulation results. Here, the response of the numerical model is shifted vertically 3.25 mV resulting in similar peak to peak amplitudes as in the measurement. This shift value can be justied to a large extent by thermal noise, which will be explained below. In Fig. 7, a part of the time history at point A in Fig. 6 can be seen. In this gure, the solid black line denotes the measured time history and the dashed line denotes the (scaled) ac excitation

Fig. 7. Part of time history at point A in Fig. 6 at f = 12.877 MHz. Table 1 Numerical values for some parameters in the total model for Vdc = 70 V and Vac = 20 mV Parameter m k0 b Q f0 f0,e d0 C0 R1 R2 Cbias Value 1.505 103 3.083 109 6.000 103 12.949 106 12.878 106 0.330 106 0.185 1015 1.0 106 50.0 0.082 106 2.330 103 2.273 1013 Unit kg N/m N s/m Hz Hz m F

F kg/m3

Fig. 6. Amplitudefrequency curve, Vdc = 70 V, Vac = 20 mV.

signal. The simulated time history (without the effect of thermal noise) is depicted as a solid grey curve. Initial estimates for some parameter values in the numerical model have been obtained from the physical dimensions and the mass of the resonator. After some ne-tuning of m, k0 and b parameter values for the numerical model have been collected in Table 1. Here, the mechanical linear natural frequency is dened as f0 = (1/2) k0 /m. The actual natural frequency of the resonator is slightly different, due to the effect of the bias voltage: f0,e = (1/2) k0,e /m, see (9). The value f0,e matches very well with the measured value of 12.875 MHz, see Fig. 4. With these values, the motional resistance Rm can be estimated using (13): Rm = 1.998 M . The rms thermal noise voltage results from (8). Measurements are performed using a bandwidth of 250 MHz and the temperature of the resonator is estimated to be 300 K. This yields an rms noise voltage of vn = 2.88 mV. This value corresponds very well to the vertical shift of 3.25 mV applied to the numerical result in Fig. 6 and, therefore, thermal noise is considered to be a plausible explanation for the observed effect. In the time history of the experimental result (Fig. 7), the effect of the noise can clearly be observed. Next, measurements for other Vac values are considered. Some parameters in the numerical model are adjusted such that the simulated amplitudefrequency curves match the measured ones. Parameters that are allowed to change are damping parameter b, the mass m and the nonlinear stiffness parameters k1 to k4 . The resonator effective mass m is allowed to change because for higher excitation values, a slightly different beam vibration

R.M.C. Mestrom et al. / Sensors and Actuators A 142 (2008) 306315

311

Fig. 8. Amplitudefrequency curve, Vdc = 70 V, Vac = 87 mV.

Fig. 10. Amplitudefrequency curve, Vdc = 70 V, Vac = 279 mV.

shape may take place, such that, effectively, more mass of the beam is moving, resulting in a (very) small increase in lumped mass. Results for ac excitation values of Vac = 87 mV and Vac = 279 mV are depicted in Figs. 811. In Figs. 8 and 10, the measured responses are denoted by black curves (solid for sweep up and dashed for sweep down). Furthermore, stable and unstable parts of the numerical amplitudefrequency curves (in grey) are indicated. Solid grey curves correspond to stable periodic solutions, whereas dashed grey curves correspond to unstable ones (not seen in measurements). The transition between stable and unstable periodic solutions is characterised by cyclic fold (cf) bifurcations (see for instance refs. [6,7]). Bifurcations mark the qualitative changes in system behaviour that may occur under variation of system parameters, in this case, under variation of the excitation frequency f. The cyclic fold bifurcations in Figs. 8 and 10 are responsible for sudden jumps in the measured amplitudefrequency behaviour during frequency sweeps (see also Fig. 4). This can be understood by considering the numerical amplitudefrequency curves. Starting from a frequency below the fundamental resonance, the systems response follows the stable low amplitude branch if the excitation frequency is increased until the lower cf point. From there, an increase in frequency makes the system jump to the upper branch of stable solutions. This jump is clearly observed in the experimental curves. An analogous

explanation holds for the sweep down behaviour by considering the upper cf bifurcation point. A good quantitative match is obtained between the numerical and experimental results. The shape of the softening resonance peak can be predicted accurately, as well as the frequencies at which sudden jumps occur. Additionally, the time histories for points B on the stable high-amplitude branches match very well (Figs. 9b and 11b). A clear distinction is present between the time histories at points A (Figs. 9a and 11a), in which thermal noise dominates, and the time histories at points B in which the noise is of signicantly lower order than the periodic signal itself. Furthermore, a discrepancy is present between the measured and simulated peak to peak values for excitation frequencies lower than the fundamental resonance, see Fig. 10. Apart from the thermal noise considerations used for the linear response, other (nonlinear) effects are also present, resulting in an additional offset between the numerical and experimental curves. This offset might result from noise effects that have not been included in the current model, see also Section 3.3. Another plausible explanation might be that a higher vertical shift in the numerical curve should be applied for higher ac excitation values, in order to match with the noise level at frequencies below the resonance frequency. As a result, the quality of the t would decrease for frequencies higher than the resonance frequency. Additionally, some kind of electrical anti-resonance effect at a frequency just above the resonance frequency of the resonator

Fig. 9. Parts of time histories at selected points in Fig. 8 at f = 12.86625 MHz. (a) Time signal at point A and (b) time signal at point B.

312

R.M.C. Mestrom et al. / Sensors and Actuators A 142 (2008) 306315

Fig. 11. Parts of time histories at selected points in Fig. 10 at f = 12.8405 MHz. (a) Time signal at point A and (b) time signal at point B.

might be present, which might justify this explanation. However, in the current modelling approach, this effect has not been included. Some parameters in the nonlinear stiffness function (5) are not needed to match the peak to peak amplitudes of the simulation model to the experimental values. Parameters k1 and k3 of the nonlinear stiffness function are set to zero. This indicates that the mechanical stiffness function is symmetric around x = 0. This is expected, because the bias voltage Vdc is applied to both sides of the beam. In this way, the only parameters that vary between successive measurements are b, m, k2 and k4 , since they can be considered as effective (lumped) parameters. The remaining parameters in (14) are assumed to be constant over changes in Vac . Parameter variations as a function of the ac excitation amplitude Vac are depicted in Fig. 12. Parameters k2 and k4 remain relatively constant after a certain Vac value. Due to limitations in the experimental setup, steps in Vac are not uniform. From this gure, it can be seen that the nonlinear stiffness terms have to be included from a certain Vac value. Due to the very small exural beam displacement x, the nonlinear stiffness parameters terms have large magnitudes: k2 = O(1015 ) N/m3 and k4 = O(1030 ) N/m5 . Furthermore, k2 and

k4 have different signs. The negative sign of k2 indicates softening nonlinear behaviour, which is observed in Fig. 8. However, for large vibration amplitudes, the positive sign of k4 (hardening) partly compensates for the softening effect of k2 , since, near 12.845 MHz, the resonance peak slightly curves upwards, see Fig. 10. In this way, it follows that both k2 and k4 have to be used in the model, which justies the inclusion of higher order stiffness terms (5) in the modelling approach. The damping parameter b is used for ne-tuning the location of the cyclic fold bifurcation points such that they match the frequencies at which jumps in the experiment occur. Finally, it can be seen that the effective lumped mass of the resonator is hardly inuenced by Vac ; it only has to be increased slightly for large excitation values. The parameter changes with the excitation parameter Vac , although they are relatively small, emphasise/conrm that the modelling approach applied here is of heuristic nature. The Dufng-like modelling approach does not give analytical insight in how parameters will change with excitation parameters. A more fundamental modelling based on rst principles will be necessary for this. In order to address the predictive value of the proposed modelling approach, amplitudefrequency curves have been cal-

Fig. 12. Parameter variation as a function of the ac excitation voltage.

R.M.C. Mestrom et al. / Sensors and Actuators A 142 (2008) 306315

313

Fig. 13. Amplitudefrequency curves for various Vac values, based on averaged values for the tted parameters.

Fig. 15. Experimental subharmonic resonance, Vdc = 70 V, Vac = 350 mV.

culated for a range of Vac values using averaged values for the tting parameters b, m, k2 and k4 . This is depicted in Fig. 13, where, for three Vac values, both the experimental and numerical response are depicted. For higher Vac values, the response can be seen to become increasingly nonlinear. The calculated responses for the averaged parameter values qualitatively represents the dynamic behaviour of the resonator for a certain Vac value. However, as indicated before, for the lumped modelling approach, the quantitative predictive nature of the model is limited. In order to investigate the inuence of the ac excitation voltage on the size of the frequency hysteresis interval, the frequencies where the cyclic fold bifurcations occur are depicted versus Vac in Fig. 14. From this gure, it is observed that frequency hysteresis occurs for Vac > 28 mV and that, for increasing Vac , the hysteresis interval rst increases but then decreases again for high Vac values (200 mV).

Another phenomenon that has been observed in the experiments is the presence of a 1/2 subharmonic resonance, which occurs at twice the fundamental resonance frequency, near 25.753 MHz. For a 1/2 subharmonic resonance, the fundamental frequency in the system response is half the excitation frequency. For excitation settings Vdc = 70 V and Vac = 350 mV. the measured subharmonic resonance is depicted in Fig. 15. The subharmonic resonance peak is initiated by two so-called period doubling (pd) bifurcations (see for more background on these and similar types of bifurcations [6,7]) near f = 25.753 MHz, see the inset in Fig. 15. In the 1/2 subharmonic resonance peak (the upper branch of solutions), the resonator vibrates with a fundamental frequency (around 12.875 MHz) that is half the frequency of the excitation signal (around 25.75 MHz). This can also be seen from the time histories at point A, see Figs. 15 and 16, in which the period time of the response (black solid line, labeled experiment) is twice that of the excitation (black dashed line). The numerical model is also capable of predicting these period doubling bifurcations and 1/2 subharmonic solution branches for parameter values of b = 4.625 1010 N s/m, k2 = 8.0 1015 N/m3 and k4 = 1.1 1030 N/m5 . The noisecorrected simulated amplitudefrequency plot is also depicted in Fig. 15, together with the measured one. A good correspondence is present. However, using these parameter values, the match

Fig. 14. Loci of cyclic fold bifurcations near the fundamental resonance.

Fig. 16. Part of time history at point A in Fig. 15 at f = 25.687 MHz. The system response has a base frequency of 12.844 MHz.

314

R.M.C. Mestrom et al. / Sensors and Actuators A 142 (2008) 306315 [11] K.Y. Kim, W. Sachse, Nonlinear elastic equation of state of solids subjected to uniaxial homogeneous loading, J. Mater. Sci. 35 (2000) 31973205. [12] J.B. Johnson, Thermal agitation of electricity in conductors, Phys. Rev. 32 (1928) 110113. [13] H. Nyquist, Thermal agitation of electric charge in conductors, Phys. Rev. 32 (1928) 97109. [14] T. Mattila, J. Kiiham ki, T. Lamminm ki, O. Jaakkola, P. Rantakari, A. a a Oja, H. Sepp , H. Kattelus, I. Tittonen, A 12 MHz micromechanical bulk a acoustic mode oscillator, Sens. Actuators A 101 (2002) 19. [15] E. Doedel, A.R. Champneys, T.F. Fairgrieve, Y.A. Kuznetsov, B. Sandstede, X. Wang, AUTO97: Continuation and Bifurcation Software for Ordinary Differential Equations (with HomCont), Technical Report, Concordia University, 1998.

around the fundamental resonance of the resonator deteriorates, since too much softening behaviour is predicted. From the time history at point A in Fig. 15, depicted in Fig. 16, it is observed that both the experiment and the simulation show a periodic response with half the excitation frequency. However, the difference in shape between the experimental and numerical time history is signicant. An explanation for this has not yet been found. 5. Conclusions The relatively simple dynamic model of the clamped clamped beam resonator utilised here, is able to predict the measured resonator response for various parameter settings qualitatively and in many cases even quantitatively. Characteristic nonlinear dynamic steady-state behaviour is very well predicted by the model. Therefore, it represents a good rst step in the modelling process and a suitable starting point for understanding and predicting the dynamic behaviour of MEMS resonators. Future work will incorporate improvement and extension of the numerical model in order to obtain better predictions. Namely, the small parameter variation for changing ac excitation voltage suggests that a more fundamental (or rst principlesbased) modelling approach is required. Furthermore, in future research, oscillator design aspects and the effect of different resonator layouts will be addressed, since the long-term goal of the project is to derive guidelines for optimal resonator layout and to predict the performance of resonators in oscillator circuits, based on more enhanced numerical models. References
[1] V. Kaajakari, T. Mattila, A. Oja, H. Sepp , Nonlinear limits for singlea crystal silicon microresonators, IEEE ASME J. Microelectromech. Syst. 13 (2004) 715724. [2] T. Mattila, O. Jaakkola, J. Kiiham ki, J. Karttunen, T. Lamminm ki, P. a a Rantakari, A. Oja, H. Sepp , H. Kattelus, I. Tittonen, 14 MHz micromea chanical oscillator, Sens. Actuators A 97/98 (2002) 497502. [3] R.M.C. Mestrom, R.H.B. Fey, J.T.M. van Beek, K.L. Phan, H. Nijmeijer, Nonlinear oscillations in MEMS resonators, in: Proceedings of the Eurosensors XX Conference, G teborg, Sweden, September 1720, 2006, o M2B-P26. [4] V. Kaajakari, T. Mattila, A. Lipsanen, A. Oja, Nonlinear mechanical effects in silicon longitudinal mode beam resonators, Sens. Actuators A 120 (2004) 6470. [5] W. Weaver Jr., S.P. Timoshenko, D.H. Young, Vibration Problems in Engineering, fth ed., Wiley, New York, 1990 (Chapter 2). [6] J.J. Thomsen, Vibrations and Stability; Order and Chaos, second ed., McGraw-Hill Publishing Company, London, 2004 (Chapter 5). [7] R.H.B. Fey, D.H. van Campen, A. de Kraker, Long term structural dynamics of mechanical systems with local nonlinearities, J. Vib. Acoust. 118 (1996) 147153. [8] M.I. Younis, A.H. Nayfeh, A study of the nonlinear response of a resonant microbeam to electric actuation, Nonlinear Dynam. 31 (2003) 91117. [9] E.M. Abdel-Rahman, A.H. Nayfeh, Secondary resonances of electrically actuated resonant microsensors, J. Micromech. Microeng. 13 (2003) 491501. [10] A.H. Nayfeh, M.I. Younis, Dynamics of MEMS resonators under superharmonic and subharmonic excitations, J. Micromech. Microeng. 15 (2005) 18401847.

Biographies
R.M.C. Mestrom (1981) received his MSc degree (cum laude) from the Department of Mechanical Engineering at Eindhoven University of Technology in 2005. Since 2005 he has been a PhD student in the Dynamics and Control group of the Department of Mechanical Engineering at Eindhoven University of Technology. His current interest and research topic are nonlinear dynamics in microsystems. R.H.B. Fey (1964) received his MSc (cum laude) and PhD degrees from the Department of Mechanical Engineering at Eindhoven University of Technology in 1987 and 1992, respectively. In 1992 he received the Shell Prize for his PhD thesis. From 1992 till 2002 he was working as a research scientist at the TNO Center of Mechanical and Maritime Engineering in Delft, a part of the TNO Netherlands Organization for Applied Scientic Research. At TNO, he worked a.o. on state estimation of ships loaded by underwater shock waves, nonlinear dynamics of solar panels, crashworthiness of steel and composite ship hulls, buckling of cylindrical structures, and uid-ship interaction. Since 2002 he works as an assistant professor in the Dynamics and Control group of the Department of Mechanical Engineering at Eindhoven University of Technology. His current research interests are nonlinear dynamics in micro-systems, piecewise linear systems, rotating discs, and base-isolation systems, and dynamic buckling of thin-walled structures. Rob Fey is author and co-author of more than 35 papers. J.T.M. van Beek (1969) received his MSc degree in applied physics and a degree in technological design from the Eindhoven University of Technology. From 1996 to 1998 Joost van Beek held a post-doc position at the Massachusetts Institute of Technology. At MIT he worked in the eld of micro-machining and micro-fabrication of UV blocking lters for space applications. From 1998 till 2006 he was a senior scientist and project leader at Philips Research and was active in the eld of RF Integrated Passive Devices and RF-MEMS. Currently, Joost van Beek is a senior scientist and project leader at NXP Semiconductors and is involved in research on MEMS oscillators and lters. Joost van Beek is co-author of more than 30 papers and holds over 15 patents. K.L. Phan (1971) received the PhD degree from the University of Twente, The Netherlands in 1999. His PhD research work was on magnetic thin lms for recording media. From 1999 to 2000 he was a researcher with the International Training Institute for Material Science in Hanoi, and subsequently from 2001 to 2002 he worked at the University of Twente as a research fellow. In 2002 he joined Philips Research, where he has been involved in various research topics including MRAM, magnetic sensors and recently MEMS resonators. H. Nijmeijer (1955) obtained his MSc degree and PhD degree in Mathematics from the University of Groningen, Groningen, The Netherlands, in 1979 and 1983, respectively. From 1983 until 2000 he was afliated with the Department of Applied Mathematics of the University of Twente, Enschede, The Netherlands. Since 1997, he was also part-time afliated with the Department of Mechanical Engineering of the Eindhoven University of Technology, Eindhoven, The Netherlands. Since 2000, he is a full professor at Eindhoven, and chairs the Dynamics and Control section. He has published a large number of journal and conference papers, and several books, including the classical Nonlinear Dynamical Control Systems (Springer Verlag, 1990, co-author A.J. van der Schaft), with A. Rodriguez, Synchronization of Mechanical Systems

R.M.C. Mestrom et al. / Sensors and Actuators A 142 (2008) 306315 (World Scientic, 2003), with R.I. Leine, Dynamics and Bifurcations of NonSmooth Mechanical Systems (Springer Verlag, 2004) and with A. Pavlov and N. van de Wouw, Uniform Output Regulation of Nonlinear Systems (Birkhauser, 2005). Henk Nijmeijer is editor in chief of the Journal of Applied Mathematics, corresponding editor of the SIAM Journal on Control and Optimization,

315

and board member of the International Journal of Control, Automatica, Journal of Dynamical Control Systems, International Journal of Bifurcation and Chaos, International Journal of Robust and Nonlinear Control, and the Journal of Applied Mathematics and Computer Science. He is a fellow of the IEEE and was awarded in 1990 the IEE Heaviside premium.