Main

The FDR expresses the equilibrium thermal fluctuations of an observable, O, in terms of available thermal energy, kBT (where kB is the Boltzmann constant), and the linear response of that observable to an applied field, F (ref. 10). It also relates the time (t) dependence of the fluctuations, found in the autocorrelation function, C(t)=〈δ O(t′+t)δ O(t′)〉t, and the time-dependent susceptibility, χ(t)=O(t)/F, where F is applied at t=0. In equilibrium, C(t) should contain the same information about the dynamics that is found in χ(t), such as the spectrum of relaxation times. This is seen clearly in plotting χ(t) versus C(t) and obtaining a straight line, with a negative slope inversely proportional to temperature10. Deviations from the FDR have been intensively studied theoretically in ageing or driven glassy systems, both on the macroscale8,10 and nanoscale12. Ageing occurs in glassy materials that have been quenched from high to low temperature. During ageing, the response functions depend both on time, t, and on the age of the system since the quench, tw. However, χ(t,tw) and C(t,tw) need not have the same dependence. This can be understood conceptually by viewing ageing dynamics as hopping on a tilted energy landscape11: energy-lowering transitions cause more rapid decorrelation than is possible by thermal energy alone. Prominent glass-transition models15 place fragile structural glasses6,16 in the same universality class as mean-field p-spin models with single-step replica symmetry breaking17. In such models10 and simulations of structural glasses7, the χ(t,tw) versus C(t,tw) asymptotically collapse (with increasing tw) to a single scaling function χ(C). χ(C) has two distinct linear regions, one agreeing with the FDR for short times, ttw<tw, then abruptly bending to a second shallower line in the strong-ageing regime, when ttw>tw. This second linear regime has a slope reduced by a violation factor, X, relative to equilibrium, and is described as having an effective temperature, Teff=T/X.

Thus far, experiments studying the breakdown of the FDR in structural and soft colloidal glasses have focused on the quasi-equilibrium regime with tw>ttw, and have given a range of results that have found strong18,19,20, weak21 or non-existent22 FDR violations. The strong-ageing regime has been difficult to access in these systems, owing largely to instrumental and statistical challenges of measuring thermal noise at the very low frequencies required. Here, using electric force microscopy (EFM) techniques23,24, we probe long-lived nanoscale polarization fluctuations and dielectric responses in polyvinyl acetate (PVAc) films, just below the glass-transition temperature, Tg (304 K), through a.c. detection of local electrostatic forces. This approach allows the probing of much smaller volumes than conventional dielectric spectroscopy. Such volumes produce large polarization fluctuations that can be detected above instrumental background down to very low frequencies. By scanning, we can study, in effect, many samples in parallel, which improves measurement statistics. The time-dependent polarization signal, VP(t), is measured (see the Methods section) to observe spontaneous noise, δ VP, or after application of a tip bias, Vd.c., to produce an experimental susceptibility signal Δχex(t)=ΔVP(t)/Vd.c..

In probing nanometre-scale regions, thermal fluctuations can be seen readily in the polarization as measured locally by VP(t) (refs 23, 24). The noise can also be used to produce spatiotemporal images of the dynamics23,24. In this approach, VP is repeatedly measured along a one-dimensional spatial line (see Fig. 1). These space-time images of the surface polarization are striking, in that they clearly show the spatio-temporal aspects of glassy dynamics; for example, the correlations seen along the time axis are distinctly longer at the lower temperatures. The autocorrelation function C(t)=〈VP(t)VP(t+t′)〉 should relate to the glassy part of the susceptibility, Δχex(t), and should obey23,24 an FDR of the form:

where ceff is an effective tip capacitance. ceff was calculated from a sphere-plane image charge model of tip capacitance for the experiment to be 6.3±1.8×10−18 F. By plotting Δχex(t) versus C(t), equation (1) was verified for several temperatures near Tg in equilibrium. The slopes of the lines were nearly identical, and required ceff=8.45±1×10−18 F to produce correct temperatures, a reasonable agreement given the simplification of the tip geometry.

Figure 1: Spatio-temporal images of thermal fluctuations measured on a single spatial dimension in equilibrium.
figure 1

The grey scale indicates the sample polarization. a,b, Image recorded at 301.5 K (a), showing long-lasting correlations relative to the image at 305.5 K (b). Inset: Measurement set-up showing an EFM cantilever with a conducting tip, biased with a sinusoidal voltage, above a dielectric polymer sample coated on a conducting electrode.

Four micrometre spatiotemporal images such as those in Fig. 1 were used to determine spatially averaged C(t) and Δχexp(t). The instrumental resolution set by the tip radius, tip height, scan speed and filter settings used produced noise that was spatially correlated over 120 nm. Thus, in these scans, 32 independent regions were measured in parallel, and several images were averaged. For C(t), scans with Vd.c.=constant (near 0) were used. Right and left 2 s/line scans were averaged to give a 0.25 Hz sampling rate. In Fig. 2a, normalized C(t) and Δχex(t) are shown for equilibrium at 303.5 K (Tg). The shapes of the two curves are nearly identical and they are fitted by stretched-exponential or Kohlrausch–Williams–Watts (KWW) functions C(t)=C(0)exp[−(t/τ)β] and , where τ is the alpha-relaxation time and β is a stretching exponent. For equilibrium, identical KWW parameters can be used, in this case β=0.53 and τ=75 s, which confirms the FDR is satisfied. The technique produces a slightly reduced β relative to bulk values25.

Figure 2: Response and correlation functions.
figure 2

a,b, Normalized local dielectric response (χ) and correlation (C) functions measured in equilibrium (a) and during ageing (b), all with stretched-exponential (KWW) fits. Equilibrium χ and C curves are fitted with identical KWW parameters: τ=75 s and β=0.53. The ageing curves in b require different stretching exponents. For χ,β=0.40±0.04, and for C,β=0.65±0.04, and both require τ=385 s. Inset: Temperature quench profile with Tg indicated as the start of ageing.

For ageing experiments, the samples were heated to well above Tg to 324 K, held for 1 min, cooled and stabilized at a final temperature, TF, with a cooling rate of 5 K min−1 through Tg (see the temperature profile in Fig. 2b). For TF=302 K, where the equilibrium τ150 s, there were no observable FDR violations. The data and analysis shown here are for TF=298 K, where τ2,500 s. The start of ageing, tw=0, was defined as the dynamic Tg(304 K) crossing. Topographic images showed that thermal drift of the tip position was significant up to tw3 min. Thereafter, spatiotemporal images were collected, and the small drift contribution to C at early tw could be determined and subtracted. Ten or more quenches were used for each correlation measurement (>320 samples) and six or more (>192 samples) for relaxation. For relaxation, Vd.c.=0.2 V was applied at tw to give Δχex(t,tw). Data points at tw and t were pair-wise correlated to calculate C(t,tw), but to improve statistics, C(t,tw) was averaged over 1-min-wide windows of time centred on tw and t.

In Fig. 2b, normalized correlation and relaxation curves are shown for ageing with tw=3 min. The shapes of the two curves are now very different, with Δχex(t,tw) more stretched compared with C(t,tw). A single KWW is no longer an ideal fit, but can be used to fit reasonably well the intermediate to long times, with β=0.64±0.04 for C and β=0.39±0.04 for Δχex with τ=385 s. Ageing of response in structural glasses for shallow quenches can be understood as a slow evolution from higher (fictive) temperature dynamics to the lower temperature equilibrium dynamics, modelled as an evolving relaxation rate (1/τ) (ref. 26) that stretches out the dynamics and gives a reduced β. More surprising is that the correlation function has an increased β, indicating a compressed spectrum of relaxation rates.

Given the different stretching parameter, β, for χ and C during ageing, the FDR clearly can no longer hold. In Fig. 3a, a plot of Δχex(t,tw) versus C(t,tw) for various tw and equilibrium is shown. Whereas the equilibrium curve is linear, as predicted by equation (1), the ageing curves all show a distinct curvature, trending towards horizontal. This is indicative of a failure of the FDR at the smallest values of C, in the strong-ageing regime. Qualitatively, this behaviour is predicted by various mean-field models and simulations7,10,12. Most striking is that there appears to be a near-collapse of all curves onto a global curve, χ(C). Such a collapse is achieved only in the asymptotic tw limit in mean-field spin-glass models10 and with scaled axes in spin-glass experiments13. However in Lennard-Jones structural glass simulations, a similar collapse was found7. The local slope of the global χ(C) curve gives the violation factor: X(C)=−dχ(C)/dC (ref. 10). In simulations, p-spin models and spin-glass experiments, X(C) is a discontinuous step-function with two distinct values: X=1 for ttw<tw and X<1 for ttw>tw (refs 6, 16). The class of models including the Sherrington–Kirkpatrick spin-glass model that show continuous replica symmetry breaking have a continuous X(C) (refs 5, 17). In the present case, if we join all of the data of Fig. 3a, and treat them as a single function, χ(C), we can determine X(C) from the local slope of this curve. In Fig. 3b, the calculated X(C) is shown. This function appears to be continuous: a linear or power-law fit work equally well, with R2=0.80. The power-law fit is shown, giving X(C)C0.57. A step function fits the data rather poorly, with R2=0.35.

Figure 3: Susceptibility versus correlation for ageing and equilibrium.
figure 3

a, Normalized χ(t,tw) plotted versus normalized C(t,tw) for various waiting times after the quench. Equilibrium data for 303 K are also plotted along with their linear fit. b, FDR violation factor, X, of all ageing data in a joined and treated as a single universal curve, plotted on a log–log plot versus C, and fit to a power-law.

We have found that during ageing in a fragile structural glass, the spontaneous dynamics are very different from relaxation dynamics. The spectrum of relaxation times appears stretched in a relaxation experiment, and compressed in a fluctuation measurement, compared with the equilibrium spectrum, requiring a modified form of the FDR. The way in which the equilibrium FDR is modified may tell us something about the way the glassy states are organized6,27 in equilibrium. A single effective temperature during ageing was not found, but the collapse of the data to χ(C), with an apparent continuous FDR violation factor, X(C), are distinctive signatures, similar to a Sherrington–Kirkpatrick class of mean-field spin-glass models, and therefore provide guidance in finding a successful model of fragile structural glasses. With larger data sets, a fixed-t analysis could more accurately determine effective temperatures28 and X(C). Open questions include the temperature dependence of X for deep quenches, and how Teff can be larger than the initial temperature of the system. The role of fragility should be explored, because in a fragile glass, dynamics slow markedly in a small range of temperature, requiring energy landscape changes much larger than kBT.

Methods

A variation of scanning probe microscopy, EFM, is used, which involves oscillating a small silicon cantilever with a sharp metal-coated tip at its resonance frequency, f0, in ultrahigh vacuum. The EFM tip is held a distance z=18±5 nm above a dielectric sample, and biased with voltage, V, relative to a conducting substrate on which the sample is coated (see Fig. 1, inset). The bias provides an electrostatic force between the tip and the sample surface, which shifts the cantilever resonance frequency by δ f(V), which is then measured. The tip electrostatic force can be calculated from the charging energy on the tip capacitance, U=1/2CtipV2 and F=−dU/dz. The bias-induced shift in the cantilever resonance frequency, δ f(V), can be obtained from the force gradient dF/dz, which supplies a fractional change in the cantilever spring constant, k. Sample polarization produces a surface potential and an image charge on the tip, both acting to produce an effective d.c. offset, VP, to the applied bias. The electrostatic component of δ f is therefore:

As the signal is proportional to the force derivative on the sharp tip, it is most sensitive to sample polarization near the surface. For a conical tip with a typical 30 nm tip radius, a region of about 60 nm in diameter and 20 nm in depth below the surface is probed25.

By applying a 200 Hz oscillating bias, Va.c.=V0sin(ω t), with V0=1 V and using a lock-in amplifier to detect the oscillation in δ f2V0VPsinω t, the polarization contribution to the tip–sample interaction can be isolated, and VP can determined. To study dielectric relaxation, a d.c. bias, Vd.c.=0.2 V, is added as a step function to the a.c. bias, the resulting time-dependent response of VP(t) is measured and an experimental susceptibility is defined as Δχex(t)=ΔVP(t)/Vd.c., where ΔVP(t) is the slow portion of the response. Thus, Δχex(t) is proportional to the glassy part of the dielectric susceptibility, Δχ(t). The samples studied here are PVAc, Mw=167,000 (Sigma), which has a bulk glass transition of Tg=308 K. Thin films (1 μm) of PVAc were prepared by dissolving the polymer in toluene and spinning the solution onto Au-coated glass substrates, drying in air and annealing for 24 h in high vacuum at Tg. Samples were mounted in an ultrahigh-vacuum SPM system (RHK SPM 350) for measurements on a chilled-water-cooled stage. Samples are radiatively heated from below and the temperature was measured with a small thermocouple clamped to the sample surface. Previously we showed that the near-surface Tg of PVAc is mildly suppressed by 4 K (ref. 25), as we see in the present experiment.

For the FDR analysis, we use the standard method of plotting fixed-tw curves. Although this has been shown28 to underestimate Teff, an attempt to use the more accurate fixed-t analysis29 produces very poor-quality curves. Both methods produce the two-slope behaviour in Lennard-Jones glasses28.