Statistical Properties of a Fluctuation-Driven Nanomechanical Duffing Resonator

In this section, we provide details on the measurement of the modal parameters, including the linear spring constant $k_{n}$, the nonlinear Duffing constant $\alpha_{n}$, and the quality factor $Q_{n}$. To extract the experimental value of $k_{n}$, we turn to the equipartition of energy, where $k_{n}=\frac{k_{B}T}{\langle{{z_{n}}}^{2}\rangle}$. We use the same notation as in the main text: $k_{B}$ is the Boltzmann constant, $T=300\rm~K$ is the ambient temperature, $\langle{{z_{n}}}^{2}\rangle$ is the mean-squared amplitude of mode $n$, and is computed as the area under the power spectral density (PSD) measurement in thermal equilibrium (i,e., without external drive). In Fig. S1(a), we show a representative thermal measurement using gray symbols, and a Lorentzian fit to the resonance peak is shown with a black continuous curve.

We measure the Duffing constant $\alpha_{n}$ using both the noise-based drive shown in the main manuscript and the deterministic drive presented below. Using deterministic frequency sweeps near an eigenmode, we follow the standard steps [landau2013course, 23, 16] that lead to the expression for the frequency of the peak $\omega_{p,n}$ as a function of the peak amplitude $z_{p,n}$ and other modal parameters, resulting in the well-studied backbone curve [13, 14, 23]:

Eq. (S1) is parabolic in $z_{p,n}$ and is shown as the solid black line in Fig. S1(b). We note that the $z_{p,n}$ corresponds to the peak value of the displacement. We compute the experimental values of $\alpha_{n}$ by fitting a parabolic function (black continuous line) using measured $z_{p,n}$ values at the largest drives to Eq. (S1). In the main manuscript, the Duffing constant is determined from stochastic excitation, where the response is based on the mean-squared amplitude $\langle{z_{n}}^{2}\rangle$. Accounting for the peak and RMS values in $z_{n}$, we achieve the same values of $\alpha_{n}$ using both methods.

We extract $Q_{n}$ from the ringdown measurements of a mechanical mode driven by a harmonic burst signal in the linear regime close to its resonant frequency $\omega_{n}$. During a ringdown, the time-dependent modal amplitude ${z_{n}}(t)$ decays exponentially as

Here, $\phi$ is an arbitrary phase, and $z_{N}$ is the technical detection noise that arises in a typical broadband measurement. As shown in Fig. S1(c), $z_{N}\approx 30~\rm mV$, and we also show this detection noise as the red-shaded area in Fig. 3(b, inset) of the main manuscript. From Fig. S1(c), we extract the upper envelope of the ringdown from measurement (gray symbols), i.e., $z_{p,n}e^{-{\omega_{n}t}/2{Q_{n}}}+z_{N}$, and find the value of $Q_{n}$ by fitting an exponential decay function with an offset (black continuous line) to the envelope.

Our synthetic noise is a band-limited white noise generated in MATLAB and continuously played on an arbitrary waveform generator (AWG) to stochastically drive a mechanical mode of our doubly clamped beam resonator. The schematic for the electronic drive is shown in Fig. 1(a) of the main manuscript. The fluctuating component of the noise, $v(t)$, is generated in the time domain and consists of $\sim 10^{5}$ decaying sine waves. Each sine wave has a randomly assigned frequency, decay time, starting time point, and phase. To ensure noise quality, we set the total number of time points to $N_{t}=5\times 10^{6}$, limited by the AWG’s memory depth. We define a uniform probability distribution within a specific range for each wave parameter: the frequency is within the noise bandwidth $\Delta f\approx 1~\rm{MHz}$; the decay time is between $10^{-3}$ and $10^{-2}\rm~s$; the starting time point is between $0$ and $N_{t}$; and the starting phase is between $0$ and $2\pi$. Each decaying wave in the time domain is a Lorentzian peak with a finite linewidth in the frequency domain, and a combination of such randomly generated waves results in the band-limited Gaussian noise, shown as the black curve in Fig. 1(c) of the main manuscript.

In the main manuscript, we showed that the PDF of the fluctuating noise component $v(t)$ is Gaussian [Fig. 1(b)]. We have also measured the statistics of the drive noise signal we generated. We add a DC offset to our fluctuating noise signal, $V+v(t)$, such that $V\gg v(t)$ and hence $(V+v(t))^{2}\approx V^{2}+2Vv(t)$. Since the force generated in the electrothermal actuator is proportional to $(V+v(t))^{2}$, we compute the PDF of the drive power $(V+v(t))^{2}$ from measured time traces of $V+v(t)$. The computed PDF is shown in Fig. S2(a) for mode $n=2$, where the $x-$axis is centered at the mean power value $\mu\approx 4\times 10^{-2}~\rm{V^{2}}$ and has units of $\sigma\approx 3.1\times 10^{-3}~\rm{V^{2}}$. The noise power has a Gaussian distribution (continuous curve) even for $\gtrsim\pm 3\sigma$. We conclude that any observed deviation from Gaussianity is due to the nonlinear dynamics — rather than the forcing.

Finally, we further analyzed the drive noise by recording the signal in the time domain for $100~\rm\mu s$ and computing the normalized autocorrelation function, $\langle v(t)\,v(t+\Delta t)\rangle/\langle v^{2}\rangle$, shown in Fig. S2(b). The inset of Fig. S2(b) shows that the correlation time is $\sim 1~\rm\mu s$, in agreement with $1/\Delta f\approx 1~\rm\mu s$, and it takes $\approx 7$ cycles for the autocorrelation function to decay, consistent with $\omega_{4}/2\pi\Delta f\approx 7.6$. Given that the relaxation time of the driven resonator mode is $2\pi Q_{4}/\omega_{4}\approx 170~\rm\mu s$, our force noise can be considered delta-correlated for all practical purposes. We show a typical autocorrelation function of a linear response ($\tilde{\alpha}\sim 10^{-3}$) in Fig. S2(c) for mode $n=3$. Here, the decay envelope is nicely fitted with an exponential decay function (dashed curve), with the relaxation time constant of $\approx 130~\rm\mu s$. In our experiments, we verified the Gaussianity and the same time and frequency domain properties for the noisy drive for all modes under study.

We compute the experimental value of $T_{\text{eff}}$ from the theoretical variance of a Duffing resonator, as shown in Eq. (3) in the main manuscript. To better illustrate the deviation of the variance of a Duffing oscillator from the variance of a harmonic oscillator, we show the dimensionless variance ${k_{n}}\langle{z_{n}}^{2}\rangle/k_{B}T_{\text{eff}}$ for mode $n=4$ in Fig. S3. The theoretical curve for the variance [28] as a function of $\tilde{\alpha}$ is shown in Fig. S3 as a black continuous curve and is expressed as

We find the experimental values of $T_{\text{eff}}$ from Eq. (S3) using the experimental values of $k_{n}$, $\alpha_{n}$, $\omega_{n}$ as described in the main manuscript. We compute the $\langle{{z_{n}}}^{2}\rangle$ value by integrating the area under the PSD curves at different power levels. A representative PSD is shown in the inset of Fig. S3, and the area under the PSD curve is highlighted in gray. The experimental $T_{\text{eff}}$ values are shown as symbols. From equipartition of energy, $k_{n}\langle{{z_{n}}}^{2}\rangle/(k_{B}T_{\text{eff}})=1$ for a harmonic resonator. In our experiments, we reach the dimensionless variance values of $\approx 0.96$ due to nonlinear stiffening [3].

We first measure the stochastic response $z_{n}(t)$ of a noise-driven mode in the time domain at different levels of $\tilde{\alpha}$. For each $z_{n}(t)$, we create a histogram of $z_{n}$ with $\sim 50$ discrete interval bins. We normalize the histogram as $\int_{-\infty}^{\infty}P(z_{n},\tilde{\alpha})\,dz=1$ to get the corresponding PDF curve, such as the ones shown in Figs. 1(b) and (d) of the main manuscript. For the PDF curves shown in Fig. 4(a) in the main manuscript, we plot the dimensionless displacement, $\tilde{z}=z_{n}/\sqrt{k_{B}T_{\text{eff}}/k_{n}}$, on the $x$-axis. Using each experimental $P(z_{n},\tilde{\alpha})\,dz$, we calculate the $m^{\rm{th}}$ order moment of our PDFs, $\langle{z_{n}}^{m}\rangle$, using the following expression

As we construct our PDF curves using a finite data set $z_{n}(t)$ drawn from a nearly Gaussian distribution, the number of standard deviations $n_{\sigma}$ displayed in our PDF is fundamentally limited by the independent sample size $N$ [westfall2013understanding]. This limitation stems from the extremely fast decay rate of the Gaussian function, which has a quadratic exponential form. To reliably capture a displacement amplitude $|\tilde{z}|>n_{\sigma}$, we require at least $5$ expected independent sample in the corresponding probability region, i.e., $NP(|\tilde{z}|>n_{\sigma})\geq 5$. For a Gaussian PDF,

In our experiments, the fluctuation amplitude remains coherent for a time span of a ringdown, i.e., $\tau_{c}\sim 2\pi Q_{n}/\omega_{n}$. As a result, the number of independent response amplitude values is $N\approx t/\tau_{c}$, where $t$ is the finite time of the measured signal. For the PDF curve shown in Fig. 4(a) of the main manuscript, $N\approx 400$. Following Eq. (S5), the maximum number of standard deviations $n_{\sigma}$ is computed as $n_{\sigma}\lesssim\sqrt{2\ln\frac{N}{5}}\approx 2.96$. We calculate the $n_{\sigma}$ values across all modes studied in experiments, and all our PDF curves are accurate up to at least $\approx 2.9|\tilde{z}|$.

We derive the expressions for the theoretical PSD curves shown in Fig. 2(a) in the main manuscript. We start with the Duffing resonator equation with damping for a single mode $n$,

Here, ${z_{n}}$ denotes the out-of-plane displacement at the modal antinode closest to the center, $\omega_{n}$ is the angular eigenfrequency, $\alpha_{n}$ is the nonlinear stiffening coefficient (Duffing constant), $Q_{n}$ is the quality factor, $\xi(t)$ is the stochastic driving force arising from the input noise, and $m_{n}$ is the modal mass.

To obtain the theoretical frequency-response curves for stochastically driven Duffing resonators, we incorporate the method of averages [6, 23] to remove the fast oscillating component of the dynamics and obtain the slowly varying envelope of the resonator amplitude. Using path integral methods [6, 8, 32], one can solve for the so-called correlator, i.e., the envelope of the autocorrelation function of the response amplitude, as follows:

where the overbar denotes complex conjugate, $\psi(t)=\cosh{(at)}+\frac{\omega_{n}(1-2i\zeta)}{2Q_{n}a}\sinh{(at)}$, $a=\frac{\omega_{n}}{2Q_{n}}\sqrt{1-4i\zeta}$, and $\zeta=2Q_{n}({\omega_{p,n}-\omega_{n}})/{\omega_{n}}$. For linear oscillations, $\zeta=0,$ thus the amplitude decay becomes identical with linear ringdown with the initial amplitude ${k_{B}T_{\text{eff}}}/{k_{n}}$ and an envelope of $e^{-\omega_{n}t/2Q_{n}}.$ In our experiments, we achieve the highest value of $\zeta\approx 30$.

Next, we compute the approximate autocorrelation function of a stochastic Duffing resonator by multiplying the correlator from Eq. (S7) with harmonic oscillations at the modal eigenfrequency, i.e., $\langle A\left(t\right)\bar{A}\left(0\right)\rangle e^{-i\omega_{n}t}$. Using the Wiener-Khinchin theorem and the Fourier convention from [paul2006stochastic], the theoretical PSD of the nonlinear stochastic response is computed as

where $\Re$ corresponds to the real component. Finally, we numerically compute the theoretical response PSD curves for all values of $T_{\text{eff}}$ using Eqs. (S7) - (S8).

The dimensionless probability distribution function for the displacement of a Duffing resonator is defined with the standard Boltzmann distribution, as shown in Eq. (7) in the main manuscript. To find the precise probability density function (PDF), we find an analytical form for the normalization constant, also known as the partition function [chandler1987introduction].

We find the partition function $\mathcal{Z}(c_{1},c_{2})$ which normalizes the probability distribution by using the definition of a quartic Gaussian integral [amdeberhan2012evaluation]. This definition is available in mathematical reference tables (see Eq. 3.469.1 in Gradshteyn and Ryzhik [gradshteyn2007table]). This matches the analytical expression we found using Mathematica [mathematica]. Using the definition from Gradshteyn and Ryzhik, the position component of the partition function is

where $c_{1}$ and $c_{2}$ are constants. For our work, we choose $c_{1}=1/2$, $c_{2}=\tilde{\alpha}/4$ based on the nondimensionalization of the nonlinearity to the parameter $\tilde{\alpha}$. This simplifies the partition function to an expression in agreement with Ref. [28]. This reduces the nondimensional PDF to

where parameters are defined in the main manuscript.

To the best of our knowledge, explicit expressions for the higher even-order moments ($m\geq 4$) are not explicitly written, but were theorized by Rahman in a footnote [rahman1996stationary]. We use Mathematica [mathematica] to take a set of appropriate partial derivatives of Eq. (S9) to find the first four even-order moments of the displacement. For example, $\partial\mathcal{Z}(c_{1},c_{2})/\partial c_{1}$ from Eq. (S9) provides the second order moment $\langle\tilde{z}^{2}\rangle$. The computational derivatives matched with the recurrence relations of the modified Bessel function of the second kind, see Eq. (9.6.26) from Ref. [abramowitz1965handbook]. We provide explicit expressions for the even-order moments of a Duffing resonator’s response PDF curves $\langle{{z_{n}}}^{m}\rangle$, for the first four even moments $m=2,4,6,8$. The odd-order moment values vanish due to symmetry in the PDF. The first even moment $m=2$ is the variance of a stiffening Duffing resonator, i.e.,

which is equivalent to the variance expression from Roberts and Spanos [28]. The higher-order moments are expressed as follows:

Using Eqs. (S11) - (S14), we compute the standardized moment values ${\langle{z_{n}}^{m}\rangle}/{\langle{z_{n}}^{2}\rangle^{m/2}}$ as a function of $\tilde{\alpha}$. We show the computed theory curves in Fig. 4(c) of the main manuscript.

The thermodynamic temperature $T$ in our system increases due to Joule heating from the electrothermal noise drive. To estimate $T$, we use the following protocol: we first measure all the mechanical parameters without any heating or drive as described above in Section I.1 and establish baselines. We then apply a known off-resonance electrical voltage comparable to the drive voltages and measure the fluctuations of the mode again. The effect of this electrical voltage is to heat the beam, without applying mechanical force to the mode. A typical measurement is shown in Fig. S4(a). The perturbed eigenfrequency $\omega_{n}$ is immediately available from the downshifted peak; the perturbed spring constant $k_{n}$ is found from $\omega_{n}$ by assuming the mode mass does not change; the quality factor $Q_{n}$ can be found from the linewidth. To find $T$, we assume that the beam is still close to equilibrium since it is not driven by a force noise — even though heat is dissipated in the system. We integrate the area under the curve and find $\langle{z_{n}}^{2}\rangle$. Then, using the perturbed $k_{n}$ values, we estimate the temperature increase from the equipartition of energy, i.e., $T=k_{n}\langle{z_{n}}^{2}\rangle/k_{B}$. The values of $T$ found using this protocol are shown in Fig. S4(b) as a function of the applied rms voltage value.

Using this approach, we estimate the maximum temperature increase of each mode for the maximum applied drive power, as shown in Table S1. Also shown are the values of the mechanical parameters at this maximum temperature. We also confirm that, as a result of Joule heating, the quality factors $Q_{n}$ decrease by a factor $\lesssim 2-3$ compared to the ringdown $Q_{n}$ values shown in Table I of the main manuscript. In this work, the decrease in $Q_{n}$ does not strongly affect the results, such as the fits discussed in Section II.1. We therefore used the unperturbed $Q_{n}$, where needed.

We detect the stochastic amplitude fluctuations using a homodyne Michelson interferometer. Here, the mechanical displacement is transduced into a phase difference between the reference and object laser beams, resulting in a measurable change in optical intensity on the photodetector. We use the following expression to convert the detected optical intensity $D$ to the displacement $z_{n}$ [wagner1990optical, dolleman2017amplitude]

Here, $a_{o}$ and $a_{r}$ are the electric field amplitudes of the object and the reference laser beams, respectively, $d_{r}-d_{o}$ is the path difference between the two laser beams, and is set to $d_{r}-d_{o}=\lambda/4$ for highest sensitivity [wagner1990optical], $\lambda=633~\rm nm$ is the laser wavelength, and $k=2\pi/\lambda$ is the wavenumber. In our experiments, the highest measured peak displacement is ${z_{n}}\approx 45~\rm nm$, which results in $\approx 5\%$ error with respect to the linearized Eq. (S15). To this end, we created a numerical calibration curve to convert the measured $D$ to $z_{n}$ based on Eq. (S15) to correct for the slight nonlinearity in the optical transducer.

We have conducted experiments in which we increased the noise bandwidth to demonstrate its effect on the response. In Fig. S5, we show two response PSD curves, $G_{z,4}({\omega\over 2\pi})$, driven by noise with a bandwidth of $\approx 1.6\rm~MHz$, which is the bandwidth used in our experiments (black trace) and $\approx 3.3\rm~MHz$ (red trace) centered around the peak. The response here is nonlinear, and $T_{\text{eff}}\sim 10^{7}~\rm K$. While a larger bandwidth results in broader tails in the response, the response at the tails is 2-3 orders of magnitude smaller and negligible. When we integrate the areas under these two response curves, we find that the mean-square amplitude $\langle{z_{4}}^{2}\rangle$ increases by $\lesssim 4\%$ for the wider bandwidth.