WorksPhysics and Astronomy
Time-temperature superposition from multi-temperature chirp rheometry of thermorheologically simple materials via physics-informed Gaussian processes
Abstract
Inferring constitutive laws from transient response data is a central inverse problem in soft condensed-matter physics, particularly when the relevant relaxation spectrum extends beyond the time window of a single experiment. Time-temperature superposition (TTS) constrains this problem for thermorheologically simple materials in the linear viscoelastic regime by allowing temperature to rescale the characteristic relaxation time without changing the underlying shape of the relaxation spectrum. We incorporate this constraint into a physics-informed machine-learning framework for time-resolved mechanical spectroscopy via chirp rheometry. We encode compact fractional viscoelastic models as causal convolutional features derived from the Boltzmann superposition principle, and we jointly fit multi-temperature chirp records using sparse variational Gaussian process regression. The model infers horizontal shift factors, constitutive parameters, and an observational-noise scale directly from time-domain stress-strain signals measured in the linear viscoelastic regime, without first extracting frequency-domain spectra or manually shifting separate curves. For synthetic Fractional Maxwell data, the method recovers the generating constitutive law from a prescribed model library and latent shifts over stress-noise levels from 3% to 25%. Applied to a DGEBA-IPDA epoxy dataset spanning the glass-transition regime, the pooled likelihood distinguishes competing fractional rheology models using the Bayesian information criterion. These results recast TTS master-curve construction as probabilistic inference over latent time rescaling and linear viscoelasticity, while also identifying limitations, such as homoscedastic noise and weakly informative spectral regimes.
article_main.pdf
Introduction
Many problems in soft condensed-matter physics require inferring an underlying material description from transient, noisy responses that reveal the relevant dynamics only indirectly. In linear viscoelastic rheology, this inverse problem is particularly acute because complex fluids and soft solids can exhibit relaxation timescales far beyond those accessible in a single experiment[1, 2, 3]. A finite measurement window therefore reveals only part of the relaxation spectrum, even when the stress response is sampled densely in time.
For thermorheologically simple materials within the linear viscoelastic regime, time–temperature superposition (TTS) provides an additional physical constraint. Temperature changes the characteristic relaxation timescale while leaving the shape of the relaxation spectrum invariant. Measurements acquired at different temperatures can then be represented by one latent master response after applying a temperature-dependent horizontal shift factor to the time or frequency axis[4]. The resulting master curve extends the experimentally accessible range, but its construction is usually treated as a separate analysis step rather than as part of constitutive inference.
Conventionally, TTS requires a sequence of operations that are both experimental and analytical. Mechanical spectra are acquired across multiple temperatures and frequency (or time) ranges, reduced to estimates of the storage and loss moduli (, ), and then aligned by horizontal and, where required, vertical shifts. Conventional frequency sweeps can take tens of minutes to acquire each spectrum when the target relaxation window spans many decades. Chirp rheometry addresses this acquisition bottleneck by compressing mechanical spectroscopy into a few seconds[5, 6, 7, 8, 9, 10, 11, 12]. Frameworks such as BOTT[13] and TORCh[14] exploit this speed to acquire multi-temperature datasets suitable for master-curve construction, while open-source packages such as herme–rheo streamline the chirp waveform generation and Fourier analysis of experimental data. However, the corresponding analysis remains largely sequential. Each chirp is generally converted into separate Fourier estimates of and , after which the spectra are shifted and interpreted together. This transformation reduces a densely sampled transient record to a smaller set of frequency-domain estimates and can propagate phase-estimation errors into the inferred moduli [16]. The problem is particularly severe when the viscous contribution is greater than the elastic one (i.e., a loss factor value ), because a small phase error can produce a large relative error in the loss modulus. Treating spectral reduction, curve shifting, and constitutive fitting as independent stages also obscures how uncertainty introduced at one stage influences the next [17, 18].
These limitations motivate a different interpretation of TTS. Rather than viewing master-curve construction as post-processing, we treat it as a latent-variable inverse problem. Multi-temperature chirp responses measured for a thermorheologically simple material share a single constitutive viscoelastic description, while temperature modulates the characteristic timescale via record-specific latent shift factors. The shifts and constitutive parameters can therefore be inferred together from the raw stress–strain histories under a common likelihood. This formulation retains the causal structure of the transient data and allows uncertainty in time rescaling and material parameters to be represented within the same model.
The proposed approach connects two related lines of work while changing the object of inference. Hernández et al. 19, {citet}`Pawlak and Denisiewicz {citep}`20, and Genovese et al. 21 ` formulated multi-temperature calibration as an inverse problem in which fractional-derivative parameters and a Williams–Landel–Ferry (WLF) shift law {citep}`22` are fitted simultaneously, but their analyses begin with frequency-domain complex moduli. Conversely, {citet}`Lennon et al. {citep}`23 used Gaussian process regression to align master curves flexibly without specifying a constitutive viscoelastic description. Here, we combine the statistical framework of Gaussian process regression {citep}24–26 with the physical, compact structure of fractional viscoelasticity by embedding candidate linear viscoelastic constitutive kernels and the TTS constraint directly into the raw time-domain records, rather than applying them after spectral extraction. We then use the Bayesian information criterion (BIC) {citep}27–29` to select the most parsimonious description from a predefined library of constitutive models.
We extend a physics-informed Gaussian-process (GP) framework previously developed for single-chirp inference [30] to multi-temperature chirp rheometry. Candidate Fractional Maxwell and Fractional Zener models [31, 32, 33, 34] define causal convolutional features derived from the Boltzmann superposition principle. We train a sparse variational GP [35, 36] on pooled records from all temperatures, providing a scalable likelihood for shared constitutive parameters, horizontal shift factors, noise scale, and model comparison. We first test parameter and shift recovery on synthetic Fractional Maxwell data with known ground truth. We then analyze a DGEBA-IPDA epoxy dataset spanning the rubbery, glass-transition, and glassy regimes.
Methods
The joint inference workflow is summarized in Figure 1. For each temperature, the measured strain rate is first convolved with the relaxation-kernel components of a candidate linear viscoelastic model, producing a set of physics-informed features. The temperature enters through a latent horizontal shift factor , which rescales the shared reference relaxation time to give the record-specific relaxation time . The dimensionless kernel-shape parameters and constitutive prefactors remain shared across records. The features and measured stresses are then pooled and used to train a single sparse variational Gaussian process (GP) under a common likelihood. The resulting posterior provides estimates of the constitutive parameters, horizontal shift factors, and observational-noise scale. These estimates are subsequently used to construct a reduced angular frequency master curve. Repeating the complete inference procedure for each candidate constitutive model enables model comparison using values of BIC.
Figure 1. Overview of the joint time-domain inference workflow. Multi-temperature chirp measurements provide the strain and stress histories. For each candidate linear viscoelastic model, the measured strain rate is convolved with the constitutive relaxation-kernel components to generate physics-informed features. Temperature-dependent horizontal shift factors rescale the characteristic relaxation time , while the kernel shape and constitutive prefactors are shared across records. The features and measured stresses are then pooled and supplied to a sparse variational Gaussian process. Joint inference yields the constitutive parameters, horizontal shift factors, and observational-noise scale, from which the reduced-frequency master curve is constructed. Candidate constitutive models are compared using the Bayesian information criterion.
Constitutive physics as a feature map
Within the linear viscoelastic regime, the stress response to an imposed strain history is governed by the Boltzmann superposition principle,
where is time, is the relaxation modulus, and is the strain rate. Equation (1) expresses the stress as a causal memory of the preceding deformation history. For a given candidate constitutive model, we write the relaxation modulus as
where the kernels encode the assumed viscoelastic topology. They depend nonlinearly on the kernel parameters , which may include dimensionless shape parameters, such as fractional exponents, and characteristic relaxation times. The coefficients , corresponding to moduli or viscosities, enter the constitutive relation linearly.
Substituting Eq. (2) into Eq. (1) gives
Each is the stress contribution that the th kernel component would generate under the measured strain history. Equation (3) therefore maps the raw deformation record to a feature space whose coordinates are determined by the candidate constitutive law.
For fixed kernel parameters , which govern the form and timescale of the relaxation spectrum rather than its overall magnitude, Eq. (3) is linear in the constitutive prefactors . The constitutive model thus determines the feature coordinates, while a regression layer determines how those features combine to reproduce the measured stress. We implement this layer using a GP with a linear covariance function. This choice preserves the linear constitutive structure while providing a probabilistic treatment of the prefactors and observational noise.
Specifically, we place a GP prior on the map from the feature vector to the observed stress:
where is independent Gaussian observational noise with variance . The covariance function is
where is the signal variance. Because the covariance is linear, the posterior mean remains linear in the physics-informed features. The GP therefore does not replace the constitutive law with an unrestricted nonlinear regression. Instead, it provides a Bayesian regression layer over features generated by that law.
Equations (3)–(5) describe the inference problem for a specified relaxation timescale. We next extend this construction to multiple temperatures by allowing each record to have a temperature-dependent timescale while retaining a common relaxation-kernel shape and common constitutive prefactors.
Latent time–temperature shifts
We consider an number of chirp records,
measured at temperatures , with samples in record . For a thermorheologically simple material, temperature changes the characteristic relaxation time without changing the shape of the linear relaxation spectrum. We therefore define the shared reference-state parameters as , where contains the shared dimensionless kernel-shape parameters and is the characteristic relaxation time at the reference temperature .
The record-specific relaxation time is defined as
The reference condition removes the scale degeneracy between and the horizontal shift factors. No WLF, Arrhenius, or other parametric temperature law is imposed during training; the independent shift factors are inferred directly from the pooled records.
For record , the features are constructed using the shared shape parameters and the record-specific relaxation time :
Temperature is therefore not included as a generic GP input. Instead, it enters through a physically constrained latent variable that rescales the memory kernel. Although this formulation could, in principle, be extended to time–concentration or other forms of superposition, we do not explore such extensions in this work.
The record-specific feature matrices and stress vectors are pooled as
A single GP is trained on this pooled feature space. The model therefore contains three classes of unknown quantities: shared constitutive parameters, record-specific horizontal shift factors, and GP and variational parameters. We optimize these quantities jointly via Adam[37] rather than through separate spectral reduction, curve shifting, and constitutive fitting stages.
Because the pooled datasets can contain hundreds of thousands of time samples, exact GP inference is computationally impractical. We instead maximize a sparse variational evidence lower bound (ELBO) under a Gaussian likelihood[36]. The optimized quantities include , , the independent shift factors, the inducing-point locations, the variational parameters, and the GP hyperparameters. The input features and stress targets are standardized before training. The inducing-point representation reduces the computational cost while retaining a common probabilistic objective for parameter estimation and model comparison. Because the covariance is linear, the required number of inducing points is on the order of the number of input features.
Parameter extraction and master-curve construction
After training, the GP posterior is converted into physically interpretable constitutive parameters. For the linear covariance in Eq. (5), the constitutive prefactor associated with feature is obtained from the posterior sensitivity,
where is the posterior-mean stress. After reversing the feature and target standardization, this derivative equals the corresponding posterior weight in physical units. Credible intervals are obtained by applying the same transformation to samples from the posterior distribution of the weights.
The inferred constitutive parameters define an analytic complex modulus for each candidate model. The response at temperature is placed on the master curve using the reduced angular frequency
An empirical WLF, Arrhenius, or other temperature law may subsequently be fitted to the inferred shift factors, but such a law does not constrain the time-domain likelihood used here.
Candidate models and model comparison
We define a library composed of constitutive models before fitting the GP to the chirp data. For the synthetic benchmark, the candidates are the Fractional Maxwell model (FMM), Fractional Maxwell Gel (FMG), Fractional Maxwell Liquid (FML), and classical Maxwell model[33]. For the experimental epoxy data, the library is extended to the Fractional Zener family and its nested limits. Each candidate defines a different set of basis kernels , shape parameters, and linear prefactors. We repeat the full inference procedure, including optimizing the shift factors, independently for each candidate. Consequently, we do not evaluate a candidate using shifts estimated under a different constitutive model.
The fitted candidates are compared using
where is used as a surrogate for the negative log marginal likelihood[29, 35, 36, 38]. The parameter count includes the fitted constitutive parameters, linear prefactors, independent horizontal shift factors, and GP hyperparameters. Lower BIC values are preferred.
Because the likelihood is represented by a variational lower bound and the transient observations may be temporally correlated, the resulting BIC values are interpreted as comparative model-selection scores under a common fitting procedure rather than as exact Bayes factors. The sensitivity of the rankings to an autocorrelation-based effective sample size is examined separately.
Results and discussion
Synthetic Fractional Maxwell benchmark
We first evaluate whether the proposed framework can recover a shared constitutive law, temperature-dependent shift factors, and an observational-noise scale directly from multi-temperature transient data. We also test whether the trained GP predicts the response to an unseen deformation history and whether the pooled likelihood discriminates among competing constitutive models. For this controlled benchmark, we generate synthetic data using a Fractional Maxwell model (FMM).
The FMM consists of two springpots connected in series. A springpot of fractional order obeys the constitutive relation
where denotes the fractional derivative of order , and is the corresponding quasi-property, with units Pa s. The limiting cases and recover a Hookean spring and a Newtonian dashpot, respectively, whereas describes an intermediate power-law element.
For the two springpots in the FMM, we denote the fractional exponents by and , with $0 \leq \beta \leq \alpha \leq 1, and their quasi-properties by and , with units Pa s and Pa s, respectively. Because the elements are connected in series, they experience the same stress, giving the constitutive equation
In our framework, it is convenient to reparameterize the two quasi-properties in terms of a characteristic relaxation time at the reference temperature state and a characteristic modulus , defined by
Thus, sets the magnitude of the elastic plateau at high frequencies in Pa units, while sets the crossover timescale between the two fractional regimes. With this parameterization, the FMM contains one kernel component, , with prefactor . Its dimensionless shape parameters are , and its relaxation time for record is . Its relaxation modulus is
where is the two-parameter Mittag–Leffler function [39]. The corresponding feature for record is
The independent record-specific parameters are the horizontal shift factors , subject to . The corresponding record-specific relaxation times, , are derived quantities rather than independently fitted parameters. We generate the synthetic data using , , Pa, and s at °C. Because the generating value is , the synthetic material lies at the fractional Maxwell Gel (FMG) boundary of the full FMM. This boundary case provides a stringgent test of both parameter recovery and model discrimination.
The temperature-dependent shifts follow the WLF relation
with and K. We generate nine records at temperatures from 30 to 90 °C. Each record uses the same 0.6–30 rad s optimally windowed chirp (OWCh) [5], sampled at 500 Hz, with 5% additive stress noise in the baseline case (see Supplementary Material, Sec. S1). The strain amplitude increases with temperature, as shown in Figure 2(a), mimicking experimental protocols in which the excitation amplitude is adjusted to maintain a measurable response across different material regimes.

Figure 2. Joint fit of nine synthetic FMM chirps spanning 30–90 °C at °C. (a) Imposed chirp strain at four representative temperatures. (b) Stress (points), posterior-mean prediction of the joint fit (black), and 95% credible band (gray), which are imperceptible in this plot because their width is thinner than the line thickness. (c) Storage and loss moduli obtained by applying the discrete Fourier transform (DFT) separately to each record and plotted against the angular frequency . The low-temperature scatter in reflects the ill-conditioning of phase-sensitive spectral estimation when , rather than a higher imposed noise level. (d) Master curve at °C plotted against the reduced frequency , using the shift factors inferred by the joint fit. Solid and dashed black lines are obtained by applying the same DFT reduction to the GP posterior-mean predictions and are not fitted directly to the spectral points. Inset: recovered , the reference constraint , and the generating WLF law.
The joint fit closely reproduces the pooled stress records, with , RMSE Pa, and [Figure 2(b)]. Because the synthetic stresses contain independent, zero-mean Gaussian noise, an adequate model should leave residuals that are centered near zero and exhibit little temporal correlation. The record-resolved residuals show no systematic drift (Figure S1), and their integrated autocorrelation times remain close to the white-noise limit, , across most temperatures (Figure S2). These diagnostics indicate that the shared constitutive kernel and latent shifts capture the deterministic response without leaving substantial time-correlated structure in the residuals.

Figure S1. Time-domain residuals for the synthetic FMM benchmark at 5% additive stress noise. Residuals are shown separately for each temperature-dependent chirp record. The residuals remain centered around zero over the chirp duration, indicating that the inferred constitutive response and latent horizontal shifts capture the deterministic stress response across temperatures.

Figure S2. Residual autocorrelation analysis for the synthetic FMM benchmark at 5% additive stress noise. The bar plot shows the integrated residual autocorrelation time, , for each temperature-dependent chirp record. These estimates quantify the reduction in independent information caused by temporal correlation in the residuals and are used to define an effective sample size for model-selection diagnostics.
The sparse variational GP returns and Pa, with a 95% credible interval of Pa for . The inferred reference relaxation time differs by approximately 1% from the generating value s, and the fitted remains indistinguishable from its boundary value of zero. We repeat the inference at stress-noise levels from 3% to 25% to determine whether the parameter recovery persists as the observations become less informative. Across this range, remains within 0.3% and within 1.7% of their generating values (Figure S3). The complete fits at 3%, 8%, 15%, and 25% noise are shown in Figure S4–Figure S7. The inferred observational-noise scale also tracks the injected noise in both absolute and relative units (Figure S8). Thus, the likelihood separates the dominant constitutive response from the observational noise over the tested range.

Figure S3. Parameter recovery across synthetic FMM noise levels. Parameters are shown for the generating FMM and for the BIC-winning model selected using the full-sample criterion. (a) Fractional exponent as a function of injected noise level. (b) Exponent . (c) Characteristic time scale at the . (d) Modulus scale . (e) WLF coefficient . (f) WLF coefficient . The comparison shows how increasing observational noise affects recovery of the constitutive and time–temperature superposition parameters.

Figure S4. Synthetic FMM benchmark at 3% additive stress noise. The figure follows the same layout as Fig. 2 of the main text: imposed chirp strains, time-domain stress fits with posterior uncertainty, unshifted frequency-domain spectra, shifted reduced-frequency master curve, and inferred horizontal shift factors.

Figure S7. Synthetic FMM benchmark at 25% additive stress noise. The figure follows the same layout as Figure 2 of the main text.

Figure S8. Recovery of the observational noise level in the synthetic FMM noise benchmark. a GP-inferred homoscedastic noise scale plotted against the injected stress-noise scale in physical units of Pa; the identity line indicates perfect recovery. b The same comparison expressed as a percentage of the noiseless stress scale. c Ratio of inferred to injected noise as a function of injected noise level, showing that the learned Gaussian likelihood accurately tracks the imposed homoscedastic noise across the tested range.
Although the model is trained entirely in the time domain, its inferred response must also remain consistent with the corresponding frequency-domain representation via Fourier analysis. Figure 2(c) shows and obtained by applying the discrete Fourier transform (DFT) separately to the observed, noisy records of the stress and plotting the resulting estimates against the unshifted frequency. After applying the GP-inferred shift factors, the individual spectra collapse onto a common reduced-frequency response [Figure 2(d)]. The solid and dashed black curves are obtained by applying the same DFT reduction to the posterior-mean stress predictions. Their agreement with the shifted raw DFT points shows that the time-domain model reproduces both the frequency-domain response and its superposition. The inferred shift factors also agree closely with the generating WLF values and with independent estimates obtained using a frequency-domain master-curve method [40], apart from a modest discrepancy at the lowest temperature (Figure S9). The low-temperature discrepancy identifies an important limitation of shift-factor inference that we discuss further in the analysis of the epoxy dataset.

Figure S9. Comparison of inferred horizontal shift factors with the mastercurves package. (a) Horizontal shift factor as a function of temperature for the present time-domain GP inference and for the mastercurves construction. (b) Ratio between the logarithmic shift factors obtained by the present method and by mastercurves. (c) Master curve obtained using mastercurves. Agreement between the inferred shifts supports the interpretation of the latent time-rescaling variables as physically meaningful TTS shift factors.
We next evaluate the predictive capability of the trained GP using a broadband chirp strain history that is not included in the training data (Figure S10). The master curve constructed from the nine training chirps spans to rad s, corresponding to 5.3 decades. Using the same sampling interval as the training records, s (i.e., 500 Hz), we generate a test chirp six times longer than an individual training chirp ( samples). This single chirp spans 0.09 to 25.7 rad s, or 2.5 decades, and therefore covers 46% of the logarithmic frequency range obtained by superposing the nine training chirps. From the imposed test strain, we compute the physics-informed features using the constitutive parameters inferred from the joint multi-temperature fit and provide them to the trained GP. The GP then accurately predicts the corresponding stress response without refitting (Figure S10), acting as a surrogate of the learned constitutive stress–strain relationship.

Figure S10. (a) Unseen chirp and (b) the prediction of the stress response using the physics-informed features computed from the constitutive parameters inferred in the joint multi-temperature fit. (c) Discrete Fourier transform of the predicted stress compared to the master curve from Figure 2 in the main manuscript.
Model discrimination in the synthetic benchmark
We compare the BIC values obtained by independently fitting the FMM, FMG, Fractional Maxwell Liquid (FML), and classical Maxwell models to the synthetic data generated at different noise levels. The FMG fixes and thus corresponds to the generating model, whereas the full FMM treats as an additional fitted parameter. Therefore, we aim to test whether the BIC defined in Eq. (11) identifies the FMG model as the most parsimonious description, and we evaluate the effect of increasing the noise level in the selection.
Table 1 separates two qualitatively different model-selection regimes. The FML and classical Maxwell models are rejected at every tested noise level by values of approximately – owing to model misfit. The distinction between the FMM and FMG is more subtle because the models are nested and the generating value lies at the boundary of the full FMM. At every tested noise level, the FMG has the lowest BIC, with the full FMM trailing by –. In these cases, the improvement in the surrogate likelihood from fitting the additional exponent does not offset the corresponding complexity penalty, and the BIC favors the most parsimonious description.
Table 1: BIC-based model comparison for the synthetic Fractional Maxwell benchmark. and BIC are absolute values for the baseline 5% noise case. The remaining columns report BIC = BIC BIC at each noise level, such that BIC = 0 identifies the selected model. All candidates are fitted to the same pooled dataset () at each noise level. Lower values are preferred.
| Model | Circuit | 5% | BIC 5% | BIC 3% | BIC 5% | BIC 8% | BIC 15% | BIC 25% | |
| FMG | ; | 15 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | ||
| FMM | ; | 16 | 7.0 | 7.8 | 6.6 | 8.1 | 10.4 | ||
| FML | ; | 15 | |||||||
| Maxwell | ; | 14 |
Table 1. BIC-based model comparison for the synthetic Fractional Maxwell benchmark. and BIC are absolute values for the baseline 5% noise case. The remaining columns report BIC = BIC BIC at each noise level, such that BIC = 0 identifies the selected model. All candidates are fitted to the same pooled dataset () at each noise level. Lower values are preferred.
Lastly, we investigate the effect of residual autocorrelation. Because the synthetic observations contain independent white noise by construction, an effectivesample-size correction to the BIC in Eq. (11) is not required in this benchmark. We nevertheless recompute the comparison using an autocorrelation-based effective sample size, both in the complexity penalty alone and in the likelihood and penalty terms together, as a robustness check. The selected model remains unchanged at every tested noise level (Figure S11). We further examine the practical relevance of this correction for the epoxy data, whose residuals exhibit temporal correlation.

Figure S11. Effect of effective sample size on BIC model selection across synthetic FMM noise levels. (a) scores computed using the full number of time-domain samples, , in both the likelihood and penalty terms. (b) scores computed using the full likelihood while replacing the penalty sample size by the effective sample size, . (c) Criterion obtained when both the likelihood and penalty terms are evaluated using . The panels report the model-selection decision as a function of injected noise level, showing that the selected model is robust to the effective-sample-size correction.
Experimental DGEBA-IPDA epoxy
We next analyze the multi-temperature chirp dataset for a DGEBA-IPDA epoxy reported by Sheridan et al. 13. The records span 100–180 °C in 5 °C increments and contain pooled time samples across the glass-transition regime. Figure 3(a) presents a representative subset of the chirp records for different temperatures. The authors assembled the complex modulus isothermal linear viscoelastic spectra into a master curve and represented it using a Prony-series generalized Maxwell model. The resulting DGEBA-IPDA master curve exhibited the broad glass transition expected for a heterogeneous epoxy thermoset. These observations motivate a constitutive description that captures both a broad distribution of relaxation times and a low-frequency elastic contribution.

Figure 3. Joint Fractional Zener fit of the DGEBA-IPDA chirp dataset over 100 °C ≤ ≤ 180 °C at = 175 °C. Color denotes temperature using one sequential scale. (a) Imposed strain at four representative temperatures. The amplitude increases by approximately two orders of magnitude at 180 °C, together with a change in axis multiplier, to maintain a measurable torque after softening. (b) Measured stress (colored points), posterior-mean prediction (black), and 95% credible band (gray). (c,d) Storage and loss moduli obtained by applying the DFT separately to each record and plotted against the unshifted angular frequency . The lowest-temperature records lie largely on the glassy plateau, whereas higher-temperature records sample the glass transition. (e) Reduced-frequency master curve using at = 175 °C. Solid and dashed black lines result from applying the same DFT procedure to the posterior-mean time-domain predictions rather than fitting the spectral points directly. Inset: inferred horizontal shift factors; horizontal gray line demarcates the reference temperature state.
Sheridan et al. 13 noted that the arbitrary strain history used in a broadband chirp obscures the direct connection between the measured stress response and a constitutive representation. With our modeling framework, this connection can be made explicit in the time domain using constitutive convolution features, as we demonstrate above for the case of synthetic data. This avoids an intermediate spectral decomposition during inference and allows the constitutive parameters and horizontal shift factors to be estimated jointly from the original time-domain records.
Here, we replace the discrete Prony-series representation used by Sheridan et al. 13 with a compact Fractional Zener model. The model consists of the FMM branch connected in parallel with an equilibrium spring of modulus . The spring does not relax, so its shape kernel is
and the corresponding convolutional feature for record is
Assuming zero initial strain, .
Because the FMM branch and the equilibrium spring are connected in parallel, their relaxation moduli add. Therefore, the Fractional Zener relaxation modulus can be written as
The model therefore has features, shared dimensionless shape parameters , a shared reference relaxation time , record-specific relaxation times , and linear prefactors .
The joint model reproduces the stress histories over the full temperature range, with RMSE = 6158 Pa and (Figure 3b). We verified model convergence by examining the optimization traces of the shape parameters, which approach stable values (Figure S12). The DFT-derived spectra show the expected temperature separation before shifting (Figure 3c,Figure 3d). The low-temperature records lie mainly on the glassy plateau, while for higher temperatures records traverse the softer transition regime. Applying the inferred horizontal shifts produces a reduced-frequency master response spanning approximately 14.2 decades (Figure 3e). The rapid decrease of with temperature seen in the inset of Figure 3e is qualitatively consistent with WLF-type behavior near the glass transition.

Figure S12. Convergence diagnostics for the experimental DGEBA–IPD dataset. The optimization histories of (a) the ELBO and (b) model parameters are used to assess the stability of the joint time-domain inference and to verify that the inferred constitutive response and horizontal shift factors correspond to a converged solution.
The fitted value lies close to the elastic boundary , raising the question of whether the additional flexibility of the general FZM is warranted or whether a more parsimonious model can describe the pooled data adequately. At the same time, the shape of the loss modulus in Figure 3(e) provides a physical reason for retaining . The two fractional exponents allow the model to represent different asymptotic behaviors at short and long times, or equivalently at high and low reduced frequencies, and thereby capture the asymmetric shape of across the relaxation regime. Fixing constrains one of these asymptotic branches and gives a less adequate representation of the observed transition.
To determine whether this additional flexibility is supported statistically, we fit each candidate model independently. The tested library compares the results from Figure 3 with the FZM with fixed at zero, the FMM, the FMG, and the classical Zener model. We then compare the independently optimized fits using the BIC. Table 2 and Figure S13 summarize the fitted shape parameters, relaxation times, prefactors, goodness-of-fit statistics, and model-selection scores.
Table 2: Model comparison for the DGEBA-IPDA chirp dataset over 100–180 °C, using pooled time samples. Here, , and , with . The quantity is the corresponding difference after accounting for the autocorrelation-based effective sample size . Lower values are preferred. The reported values of , , and are scaled by . RMSE is reported in Pa, the relaxation time in ms, and the moduli in MPa. A dagger denotes a parameter fixed at zero, and a dash denotes a parameter that is not applicable.
| Model | RMSE | [ms] | [MPa] | [MPa] | ||||||||
| FZM | 25 | −4.56702 | 0 | 0 | 17447.0 | 6156 | 0.9818 | 0.556 | 0.0298 | 10.740 | 0.2311 | |
| FMM | 24 | −2.74115 | 1.826 | 0.188 | 1936.3 | 7774 | 0.9709 | 0.437 | 0.0200 | 14.217 | – | |
| FZM () | 24 | −2.08187 | 2.485 | 0.154 | 9028.0 | 8452 | 0.9656 | 0.491 | 19.215 | 0.1912 | ||
| FMG | 23 | −1.16919 | 3.398 | 0.191 | 3530.3 | 9496 | 0.9566 | 0.460 | 19.359 | – | ||
| Zener | 23 | 5.05969 | 9.626 | 0.336 | 10413.0 | 0.7888 | – | – | 13.986 | 0.4695 |
Table 2.

Figure S13. (a) BIC model selection for the experimental DGEBA–IPD dataset, computed using the nominal sample size and the effective sample size . (b) Independent samples resulting from each fit. (c) Epochs used for each fit and its budget.
Under the BIC expression in Eq. (11), the general FZM is preferred, followed by the FMM, the nested FZM with , the FMG, and the classical Zener model. This ordering, shown in Table 2, is driven primarily by the likelihood term rather than by the complexity penalty. The candidates differ by only one or two fitted parameters, whereas changes substantially across the models. The same trend is reflected in the goodness-of-fit statistics: decreases from 0.9818 for the general FZM to 0.7888 for the classical Zener model. Thus, under the complete sample-size assumption, the improvement in fit provided by the general FZM more than compensates for its additional parameter.
Because the experimental residuals are temporally correlated, the sample size overstates the number of independent observations. Recalculating the BIC using the autocorrelation-based effective sample size (Figure S14 and Figure S15) substantially reduces the differences among the candidate models but leaves the selection of the general FZM unchanged. The corrected ranking places the nested FZM with second, followed by the FMM, FMG, and classical Zener model. The similar scores of the nested FZM, FMM, and FMG show that their relative ordering is sensitive to the treatment of residual correlation. Nevertheless, the general FZM remains preferred under both calculations. The nonzero value of is therefore supported both physically, through the asymmetric shape of in Figure 3(e), and statistically, through the model-selection scores.

Figure S14. Time-domain residuals for the experimental DGEBA–IPD chirp records. Residuals are shown separately for each temperature-dependent record and provide a diagnostic of measurement variability, instrumental effects, and possible constitutive-model mismatch after joint time-domain inference.

Figure S15. Residual autocorrelation analysis for the experimental DGEBA–IPD dataset. The integrated autocorrelation time, , is reported for each temperature-dependent chirp record. These values quantify temporal dependence in the residuals and determine the corresponding effective number of independent observations used in the model-selection analysis.
Having established the preferred constitutive model, we next examine the inferred horizontal shift factors, shown in the inset of Figure 3(e), by comparing them with two frequency-domain estimates. These are the shifts obtained using the data-driven method of Lennon et al. [23] and those reported by Sheridan et al. [13], who applied pairwise horizontal shifting to minimize the noise-weighted mismatch in and . The three estimates agree broadly at high temperatures and through the glass-transition region but diverge at the lowest temperatures (Figure S16 and Figure S17), demonstrating that master-curve alignment is not uniquely determined over the full temperature range.

Figure S16. (a) Comparison of horizontal shift factors for the experimental DGEBA–IPD dataset. The temperature-dependent shift factors inferred jointly by the time-domain GP framework are compared with those obtained using the mastercurves package. (b) Agreement between the two estimates supports the interpretation of the latent time-rescaling variables as physically meaningful time–temperature superposition factors, while deviations identify temperatures at which the inferred rescaling is method-dependent. (c) Master curve obtained using the mastercurves package.

Figure S17. Frequency-domain master curve for the experimental DGEBA–IPD dataset constructed using the approach from (a) this work, (b) the mastercurves package and (c) the shifts from Sheridan, Zauscher, and Brinson [3]. The shifted storage and loss moduli are plotted against reduced frequency to assess the quality of time–temperature superposition and to provide an independent comparison with the response inferred from the pooled time-domain chirp records.
This pattern is physically and statistically informative. Near the transition, the measured response has substantial slope and curvature, and neighboring temperatures sample overlapping but distinguishable portions of the relaxation spectrum. Horizontal displacement is therefore constrained by identifiable features in the data. At low temperatures, by contrast, the measured window lies largely on the glassy plateau. A large horizontal shift can then produce only a small change over the observed frequency interval, making weakly identifiable. In this regime, the inferred shift depends more strongly on the representation and optimization objective of the superposition method.
Although no estimate can be identified as the uniquely correct shift without independent ground truth, the present method has the advantage of inferring all shift factors jointly from the raw stress–strain records under a common causal constitutive model and likelihood. It therefore avoids phase-sensitive spectral estimation and pairwise curve alignment. This integration also introduces an important limitation: the inference is tied to the prescribed constitutive-model library, and if none of the candidate models adequately represents the material response, no valid solution can be obtained within the framework. Thus, the method provides a more integrated physical and statistical basis for the inferred shifts, but it remains sensitive to constitutive-model misspecification and does not eliminate weak identifiability in featureless plateau regions, such as the low-temperature glassy plateau at high reduced angular frequencies or an equilibrium plateau, when present.
Conclusion
We formulated time-temperature superposition as a joint probabilistic inverse problem using raw multi-temperature chirp records. Fractional relaxation kernels generate causal, physics-informed convolutional features, and a sparse variational GP jointly infers the constitutive parameters, horizontal shift factors, and observational-noise scale. The approach therefore combines constitutive inference and master-curve construction without prior spectral reduction or manual curve alignment.
For the synthetic FMM benchmark, the method recovers the constitutive law, relaxation scale, noise level, and latent shifts over stress-noise levels from 3% to 25%. For the experimental DGEBA-IPDA dataset, the FZM is preferred among the tested models and yields a reduced-frequency response spanning approximately 14.2 decades. Comparisons with frequency-domain methods recover the same dominant shifting trend but reveal weak shift-factor identifiability on the low-temperature glassy plateau. Furthermore, model discrimination is reliable for clearly different constitutive structures and remains sensitive to small likelihood differences for nested fractional models near a parameter boundary.
The present framework is limited to horizontal shifts and assumes homoscedastic, temporally independent Gaussian noise. Because chirps acquired at different temperatures can have different torque amplitudes and signal-to-noise ratios relative to the instrumental resolution, future work should consider a record- or signal-dependent heteroscedastic likelihood. Vertical shifts would require record-specific factors , a reference constraint, and additional physical regularization to distinguish them from the shared constitutive prefactors and horizontal shifts. The feature construction also requires relaxation-modulus kernels with a characteristic timescale that can be rescaled by . Models formulated primarily through creep compliance and retardation times, including common fractional Kelvin–Voigt variants, are therefore not directly compatible without deriving an equivalent relaxation kernel or reformulating the inference in terms of creep compliance.
Overall, the framework shows how physics-informed convolutional features can connect machine learning with rheological inverse problems. Constitutive convolution provides a causal, interpretable representation, while probabilistic regression estimates constitutive parameters, temperature-dependent time rescaling, and observational uncertainty. The master curve is thus one outcome of a broader inference that also provides constitutive parameters, model comparisons, and information about where the response is well constrained or remains sensitive to modeling assumptions.
Supplementary Material
The Supplementary Material is available free of charge at [link to be inserted after acceptance].
The Supplementary Material contains the additional synthetic-noise benchmarks, recovery of the observational-noise scale, residual and autocorrelation diagnostics, effective-sample-size analyses, optimization histories, parameter-recovery trends, and independent comparisons of the inferred shift factors and reduced-frequency master curves. The file also includes results for two additional datasets of thermorheologically simple materials in Figure S18 and S19.

Figure S18. Application of the joint time-domain framework to the photopolymer dataset reported by Sheridan, Zauscher, and Brinson [3] at . Consult the original publication for further details about the material and experimental protocol. The layout follows the same as in Figure 3 of the main manuscript.
Acknowledgments
Funding: I. M. V. thanks the Vilho, Yrjö, and Kalle Väisälä Foundation of the Finnish Academy of Science and Letters and the Finnish Foundation for Technology Promotion for personal funding. The authors acknowledge funding from the FinnCERES flagship (151830423), Business Finland (210129, 211835, 211909, 211989), the Future Makers program, and the Finnish Cultural Foundation. K. M. and M. J. A. acknowledge the financial support provided by the Finnish Ministry of Education and Culture through the Pilot for Doctoral Programmes (Pilot Project: Mathematics of Sensing, Imaging and Modelling). The authors acknowledge funding from the European Union: the ARCHIBIO-FOAM project has received funding from the European Union’s Horizon Europe research and innovation program under grant agreement No 101161052. Views and opinions expressed are, however, those of the authors only and do not necessarily reflect those of the European Union or the European Innovation Council and SMEs Executive Agency (EISMEA). Neither the European Union nor the granting authority can be held responsible for them. The authors acknowledge the computational resources provided by the Aalto Science-IT project, and thank Mohua Das and Prof. Gareth McKinley for their guidance on the numerical simulation of chirps.
Author contributions: I. M. V.: Conceptualization, Methodology, Software, Formal analysis, Visualization, Investigation, Writing—original draft, Writing—review and editing. K. M.: Formal analysis, Validation, Writing—review and editing. J. K.: Supervision, Validation, Writing—review and editing. M. A.: Supervision, Validation, Writing—review and editing, Funding acquisition, Project administration.
Competing interests: The authors declare that they have no competing interests.
Data and materials availability: All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Materials. The data and code that support the findings of this study are openly available in GitHub repository RheoGP (v1.1.0) at https://github.com/mirandil/rheogp and 10.5281/zenodo.23558075, Ref. [41].

Figure 4 (PDF p. 12).
Supplementary Information: Time–temperature superposition from multi-temperature chirp rheometry of thermorheologically simple materials via physics-informed Gaussian processes
Isaac Y. Miranda-Valdez; Kourosh Mobredi; Juha Koivisto; Mikko J. Alava
supplementary.pdf
Supplementary Information: Time–temperature superposition from multi-temperature chirp rheometry of thermorheologically
Numerical simulation of chirp rheometry
We obtained the numerical simulations from the main paper using a strain-controlled, optimally windowed chirp (-OWCh) protocol following Geri et al. [1]. The imposed strain is
where is the nominal strain amplitude, is a dimensionless tapering window, and is the accumulated phase. To probe the frequency interval over a duration , the chirps use a logarithmic sweep,
so that the instantaneous angular frequency increases exponentially from to ; is an initial phase offset, zero unless stated otherwise. To reduce spectral leakage, a chirp is tapered with a Tukey-type window,
where is the taper fraction; the reported data used throughout.
Complementary results for the numerical simulation
Residual diagnostics
We next examined the residual structure of the synthetic FMM benchmark used in the main text. Because joint inference is performed directly in the time domain, residual diagnostics directly check whether the fitted constitutive memory and inferred time-rescaling variables capture the deterministic stress response, leaving only the imposed observational noise.
Figure S1 shows the residuals for the 5% noise case from the main paper, separated by chirp record at every synthetic temperature. The residuals remain centered around zero across temperatures, with no systematic drift over the chirp duration. This indicates that the joint model captures the dominant temperature-dependent viscoelastic response and that the remaining discrepancies are consistent with observational noise.
To quantify possible temporal correlations in the residuals, we computed the residual autocorrelation for each chirp record. Figure S2 summarizes the resulting autocorrelation time scale, , as a function of temperature. The estimated autocorrelation time provides a measure of the number of effectively independent samples in each time-domain record and motivates the use of an effective sample size when comparing models by information criteria.
Synthetic FMM benchmark at additional noise levels
To assess the robustness of the joint time-domain inference procedure, we repeated the synthetic Fractional Maxwell model (FMM) benchmark described in the main text at four additional stress-noise levels. In addition to the baseline 5% case discussed in the main article, we considered additive Gaussian stress noise with amplitudes of 3%, 8%, 15%, and 25% of the noiseless stress scale. For each noise level, the same nine-temperature chirp protocol, WLF shift law, reference temperature, and generating FMM parameters were used:
with and temperatures spanning 30–90 °C. We also kept the imposed chirp frequency range and sampling rate fixed, so changes in the inference results can be attributed directly to increased observational noise.
For every noise level, we trained the same sparse variational Gaussian-process model on the pooled time-domain stress–strain records. We inferred the horizontal shift factors independently for each temperature, subject to the reference constraint , and shared the constitutive parameters across all records. We imposed no WLF form during training; as in the main text, we compared the generating WLF law only after inferring the latent shift factors.
We assessed how the recovered constitutive and time–temperature superposition parameters vary with noise level. Because the BIC variants based on and give the same model-selection decision for these synthetic cases (§ 4.1.2.5), the comparison is reported for the BIC-winning model obtained using the full-sample criterion.
Figure S3 compares the recovered parameters as a function of injected noise level for the generating FMM and for the BIC-winning model. Panels (a)–(d) show the evolution of the constitutive parameters , , , and , respectively. Panels (e) and (f) show the recovered WLF coefficients and . The parameter estimates remain close to the generating values at low and moderate noise levels, while increasing noise broadens the uncertainty and leads to larger deviations, particularly for parameters associated with weaker or nearly nested model components.
Figure S4 shows the joint fit for the 3% noise case. At this lowest noise level, the stress trajectories are reproduced with very small residuals, and the recovered shift factors closely follow the generating WLF curve. The reduced-frequency representation collapses the independently transformed spectra onto the FMM master curve, indicating that the time-domain likelihood resolves both the fractional constitutive parameters and the latent time-rescaling variables with high precision. Figure S5 shows the corresponding result for 8% stress noise. The time-domain predictions remain stable across the full temperature range, despite the visibly increased scatter in the stress response and the frequency-domain moduli obtained from separate DFT reductions. The inferred shifts continue to collapse the spectra onto a single reduced-frequency master curve, indicating that the pooled likelihood leverages the shared constitutive structure to regularize inference across temperatures. Figure S6 reports the 15% noise case. Although the direct spectral estimates become substantially noisier, particularly for the loss modulus, the joint time-domain fit still recovers the dominant fractional-relaxation behavior. The collapse of the shifted spectra remains clear, illustrating that the inference is not driven by fitting each noisy spectrum independently but by

Figure S5. Synthetic FMM benchmark at 8% additive stress noise. The figure follows the same layout as Figure 2 of the main text.

Figure S6. Synthetic FMM benchmark at 15% additive stress noise. The figure follows the same layout as Figure 2 of the main text.
estimating a shared memory law from all raw chirp records simultaneously.
Figure S7 gives the highest-noise case, with 25% additive stress noise. This case provides a stringent test of the latent-variable formulation. Even when the individual DFT-derived spectra are highly scattered, the joint model retains the correct qualitative master-curve structure and estimates shift factors consistent with the underlying trend. The increased uncertainty at this noise level is reflected in the broader posterior bands and in the reduced ability to distinguish nearly nested model variants, as discussed in the BIC analysis in the main text.
Across the four additional noise levels, the same qualitative conclusions hold as in the 5% case presented in the main text. The pooled time-domain likelihood recovers the shared fractional constitutive law, estimates physically meaningful horizontal shift factors without imposing a WLF form, and produces a reduced-frequency master curve consistent with the generating model. Increasing the stress noise mainly broadens the posterior uncertainty and degrades the independent DFT estimates, rather than causing a systematic failure of the time-domain inference. These results support the robustness of the proposed latent time-rescaling formulation over a broad range of experimentally relevant noise amplitudes.
Figure S8 summarizes the recovery of the Gaussian likelihood noise scale across the synthetic FMM noise benchmarks. The inferred homoscedastic noise closely follows the injected noise level, both in absolute stress units and when expressed relative to the noiseless stress scale, indicating that the GP model correctly identifies the observational noise rather than absorbing it into the constitutive response.
Shift-factor comparison with mastercurves
As an external comparison for the inferred horizontal shifts, we also compared the latent shift factors obtained by the present time-domain method with shift factors obtained using the mastercurves package of Lennon, McKinley, and Swan [2]. This comparison tests whether the probabilistic time-domain inference recovers shifts consistent with a well-established method for frequency-domain master-curve construction.
Figure S9 shows the resulting comparison. Panel (a) compares as a function of temperature for the present method and for the mastercurves result. Panel (b) reports the ratio between the logarithmic shifts obtained by the two methods, highlighting deviations in the inferred horizontal rescaling. Panel (c) shows the master curve obtained using mastercurves. The two methods give nearly identical shift factors over most of the temperature range, with only a small discrepancy at the lowest temperature. This close agreement supports the interpretation of the latent variables inferred by the GP model as physically meaningful horizontal shift factors.
Prediction of a broadband test chirp
Figure S10 shows the stress response predicted by the trained GP for the broadband test chirp discussed in the main text, which is not included in the training data.
Effective sample size and BIC model selection
We then investigated how the treatment of sample size affects Bayesian information criterion (BIC) model selection across noise levels. Since chirp time series contain many closely spaced samples, the nominal number of time points, , can overstate the amount of independent information when residuals are temporally correlated. We therefore compared three BIC variants: one using the full number of samples in both the likelihood and penalty terms, one using the full likelihood but replacing the penalty term by an effective sample size, , and one in which both the likelihood contribution and the penalty term are evaluated using .
Figure S11 summarizes the resulting model-selection decisions as a function of injected noise level. Panel (a) shows scores computed using the full number of samples, . Panel (b) shows the corresponding scores when only the complexity penalty is computed using . Panel (c) shows the criterion obtained when both the likelihood and penalty terms are scaled by . Across these variants, the selected model remains unchanged for the tested synthetic-noise cases, indicating that the model-selection outcome is robust to the effective-sample-size correction. This insensitivity is expected for the synthetic benchmark because the injected observational noise is independent and approximately white. Consistent with this, Figure S2 shows that the integrated residual autocorrelation time remains close to its minimum value, with most temperature records corresponding to roughly one effectively independent sample per time step. Thus, the effective sample size is nearly equal to the nominal number of samples, and replacing by produces only a small correction to the BIC penalty.
Complementary results for the DGEBA–IPD dataset
We next applied the joint time-domain inference framework to the experimental DGEBA–IPD dataset. In contrast to the synthetic FMM benchmark, the experimental data have no prescribed constitutive parameters, known horizontal shift factors, or injected noise levels. We therefore assess the inference through optimization convergence, the fitted noise model, time-domain residuals, residual autocorrelation, constitutive-model selection, and comparison with an independent frequency-domain master-curve construction. Together, these diagnostics test whether the inferred constitutive response and latent horizontal shifts provide a stable and internally consistent description of the multi-temperature chirp measurements.
Figure S12 shows the optimization history for the DGEBA–IPD inference. The convergence traces provide a direct check that the variational objective and inferred quantities approach stable values before the fitted response and latent shift factors are interpreted. Establishing convergence is particularly important for the experimental dataset because model adequacy cannot be evaluated by comparison with known generating parameters.
We next compared the candidate constitutive models using the Bayesian information criterion. Figure S13 summarizes the model-selection results for the experimental data. Because the residuals can be temporally correlated, the comparison includes the effect of replacing the nominal number of time-domain samples, , with the effective sample size, , in the information criterion. Reporting BIC differences relative to the best-supported candidate makes it possible to assess both the preferred constitutive description and the sensitivity of the ranking to the treatment of temporal correlation.
Figure S14 shows the time-domain residuals for the individual DGEBA–IPD chirp records. These residuals provide a record-by-record diagnostic of the fitted constitutive response. In contrast to the synthetic case, departures from zero can reflect not only observational noise but also instrumental effects, experimental variability, preprocessing, and deviations of the material response from the selected constitutive model. Their temporal and temperature-dependent structure therefore reveals where systematic discrepancies remain after joint inference.
To quantify temporal dependence in the experimental residuals, we computed the integrated autocorrelation time for each chirp record. Figure S15 reports the resulting autocorrelation time scale, , as a function of temperature. Temporal correlation reduces the amount of independent information contained in the densely sampled time series; accordingly, these estimates provide the basis for defining an effective sample size, , for model-selection diagnostics. Differences among temperatures can reflect changes in measurement conditions, instrumental response, preprocessing, or unmodeled material behavior.
As an external assessment of the latent time-rescaling variables, we compared the horizontal shift factors inferred by the time-domain GP framework with those obtained using the mastercurves package of Lennon, McKinley, and Swan [2]. Figure S16 presents the temperature-dependent shifts obtained by the two methods. The GP estimates are inferred jointly from the raw stress–strain records and the shared constitutive response, whereas the mastercurves estimates are obtained from a frequency-domain superposition procedure. Their comparison therefore tests whether two distinct inference strategies identify a consistent temperature dependence of the material time scale.
Figure S17 shows the corresponding frequency-domain master curve constructed using mastercurves. The storage and loss moduli are plotted against reduced frequency after applying the independently estimated horizontal shifts. The degree of overlap among the temperature-dependent records provides an external assessment of time–temperature superposition and complements the shift-factor comparison in Figure S16. It also provides an independent frequency-domain representation against which the broadband response inferred directly from the time-domain chirps can be evaluated.
Overall, the DGEBA–IPD results provide an experimental counterpart to the synthetic FMM benchmark. Because the true constitutive parameters, horizontal shifts, and observational-noise levels are unknown, validation relies on mutually consistent diagnostic evidence rather than recovery of prescribed generating quantities. The convergence history establishes the stability of the optimization, the noise and residual analyses characterize the unexplained experimental variability, and the autocorrelation analysis quantifies the reduction in effective sample size. The model-selection results identify the constitutive description supported by the pooled time-domain records, while the independent mastercurves comparison assesses the physical consistency of the inferred horizontal shifts and reduced-frequency response. Together, these results demonstrate how the joint time-domain framework can be evaluated when applied to experimental multi-temperature chirp rheometry.
Implementation details
The framework is implemented in Python using GPyTorch[4] for GP inference and PyTorch for automatic differentiation and gradient-based optimization. All experiments were run on a single NVIDIA H200 MIG GPU with 71 GB of VRAM. The learning rates were fixed at 0.001 for the GP hyperparameters and 0.005 for the constitutive parameters. Optimization was performed for a maximum of 20000 epochs, with early stopping triggered when the ELBO did not improve for 500 consecutive epochs. Two inducing points were used for models with one input feature, such as the fractional Maxwell model, whereas three were used for models with two input features, such as the fractional Zener model.
Code and data supporting this study are openly available at https://github.com/mirandi1/rheogp.
Applications to additional experimental datasets
To assess the applicability of the proposed framework beyond the DGEBA–IPD dataset, we considered two additional experimental systems reported in the literature: the photopolymer dataset of Sheridan, Zauscher, and Brinson [3] and the acrylic-based adhesive dataset of Perego, Mills, and Vadillo [5]. Figure S18 shows the application of the proposed approach to the photopolymer dataset using a fractional Zener model composed of two springpots and a spring. The fitted parameters were , , , with a 95% confidence interval of , and with a 95% confidence interval of . The fit yielded RMSE = 6172 Pa, = 0.9485, and BIC = −34216.0. Figure S19 presents the corresponding analysis of the acrylic-based adhesive dataset, which accompanies hermes-rheo, an open-source Python package for rheological data analysis. For this material, we used a fractional Zener model composed of three springpots, with the third fractional order denoted by . The fitted parameters were = 0.6406, = 0.1262, = 0.1808, = 1.00015 × s, = 6.7376 × Pa with a 95% confidence interval of [6.72536 × , 6.76024 × ] Pa, and = 1.05711 × Pa s with a 95% confidence interval of [1.04194 × , 1.06819 × ] Pa s. From , the corresponding springpot prefactor is

Figure S19. Application of the joint time-domain framework to the acrylic-based adhesive dataset reported by Perego, Mills, and Vadillo [5] at . Consult the original publication for further details about the material and experimental protocol. The layout follows the same as in Figure 3 of the main manuscript.
1.845 Pa s. The retardation time is therefore s, yielding the equivalent equilibrium modulus Pa. The fit yielded RMSE = 426.2, = 0.9914, and BIC = .
Acknowledgments
See main text.
References
- [1]Scott Blair, G. W. Elementary rheology; Academic Press: London and New York, 1969.
- [2]Song, J.; Holten-Andersen, N.; McKinley, G. H. Soft Matter 2023, 19, 7885–7906.
- [3]Ricarte, R. G.; Shanbhag, S. Polymer Chemistry 2024, 15, 815–846.
- [4]Ferry, J. D. Viscoelastic Properties of Polymers, 3rd ed.; John Wiley & Sons: New York, 1980.
- [5]Geri, M.; Keshavarz, B.; Divoux, T.; Clasen, C.; Curtis, D. J.; McKinley, G. H. Physical Review X 2018, 8, 041042.
- [6]Das, M.; Vadillo, D. C.; Perego, A.; McKinley, G. H. Journal of Rheology 2026, 70, 361–381.
- [7]Perego, A.; Vadillo, D. C.; Mills, M. J. L.; Das, M.; McKinley, G. H. Rheologica Acta 2025, 64, 391–406.
- [8]Waeterloos, J. L.; McKinley, G. H.; Clasen, C. Rheologica Acta 2025, 64, 633–646.
- [9]Hudson-Kershaw, R. E.; Das, M.; McKinley, G. H.; Curtis, D. J. Journal of Non-Newtonian Fluid Mechanics 2024, 333, 105307.
- [10]Holly, E. E.; Venkataraman, S. K.; Chambon, F.; Winter, H. H. Journal of Non-Newtonian Fluid Mechanics 1988, 27, 17–26.
- [11]Winter, H. H.; Chambon, F. Journal of Rheology 1986, 30, 367–382.
- [12]Ghiringhelli, E.; Roux, D.; Blésès, D.; Galliard, H.; Caton, F. Rheologica Acta 2012, 51, 413–420.
- [13]Sheridan, R. J.; Zauscher, S.; Brinson, L. C. Soft Matter 2024, 20, 7811–7820.
- [14]Vadillo, D. C.; Perego, A.; Whitcomb, K.; Behling, R. E.; Gill, A. M.; Das, M.; McKinley, G. H. Macromolecules 2026, 59, 6893–6905.
- [16]Angerman, P.; Faulhammer, L.; Bauer, L.; Haeberle, J.; Curtis, D. J. Rheologica Acta 2026, 65, 735–744.
- [17]Gergесova, M.; Zupančič, B.; Saprunov, I.; Emri, I. Journal of Rheology 2011, 55, 1–16.
- [18]Maiti, A. Rheologica Acta 2019, 58, 261–271.
- [23]Lennon, K. R.; McKinley, G. H.; Swan, J. W. Data-Centric Engineering 2023, 4.
- [29]Kass, R. E.; Raftery, A. E. Journal of the American Statistical Association 1995, 90, 773–795.
- [30]Miranda-Valdez, I. Y.; Koivisto, J.; Alava, M. J. Direct inference of viscoelastic memory from chirp rheometry via physics-informed Gaussian processes. 2026; https://arxiv.org/abs/2608.16306.
- [31]Bagley, R. L.; Torvik, P. J. Journal of Rheology 1983, 27, 201–210.
- [32]Bonfanti, A.; Kaplan, J. L.; Charras, G.; Kabla, A. Soft Matter 2020, 16, 6002–6020.
- [33]Jaishankar, A.; McKinley, G. H. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences 2013, 469, 20120428.
- [34]Miranda-Valdez, I. Y.; Niinistö, A.; Mäkinen, T.; Lejon, J.; Koivisto, J.; Alava, M. J. Digital Discovery 2025, 4, 1075–1082.
- [35]Titsias, M. Variational Learning of Inducing Variables in Sparse Gaussian Processes. Proceedings of the Twelfth International Conference on Artificial Intelligence and Statistics. Hilton Clearwater Beach Resort, Clearwater Beach, Florida USA, 2009; pp 567–574.
- [36]Hensman, J.; Fusi, N.; Lawrence, N. D. Gaussian Processes for Big Data. Proceedings of the Twenty-Ninth Conference on Uncertainty in Artificial Intelligence. 2013; pp 282–290.arxiv.org/abs/1309.6835
- [37]Kingma, D. P.; Ba, J. Adam: A Method for Stochastic Optimization. 2017; https://arxiv.org/abs/1412.6980.
- [38]Freund, J. B.; Ewoldt, R. H. Journal of Rheology 2015, 59, 667–701.
- [39]Garrappa, R. SIAM Journal on Numerical Analysis 2015, 53, 1350–1369.
- [40]Lennon, K. R.; McKinley, G. H.; Swan, J. W. Proceedings of the National Academy of Sciences 2023, 120.
- [41]Miranda-Valdez, I. RheoGP: physics-informed Gaussian processes for chirp rheometry. 2026; https://doi.org/10.5281/zenodo.23558075.
