Glassy relaxation as fast exponential decay in a hidden space of local excitations

A supercooled liquid is a liquid cooled fast enough that it never crystallizes. As it cools, it slows down by more than ten orders of magnitude, yet a snapshot of its structure looks almost the same as a snapshot of an ordinary liquid. Its relaxation back to equilibrium is also not a simple exponential: correlations die off as a "stretched" exponential, which looks like many timescales mixed together. In a separate body of mathematics, viscoelastic materials have history-dependent stress. These materials provably return to equilibrium exponentially fast once you add hidden "augmented" variables, one per memory mode, so that the enlarged system has no memory and contracts in time. We propose to build that kind of enlarged system for a glass. The hidden variables would not be fitted. Each one would be a population of physically identified local rearrangements, the "excitations" that recent landscape-search algorithms extract directly from glass configurations. The couplings between these variables would be measured from how triggering one excitation shifts the energy barriers of its neighbors.

The payoff is a sharp, falsifiable account of the gap between structure and dynamics. The hidden excitation state decays exponentially at a rate set by the smallest eigenvalue of a generator, which is a matrix assembled entirely from static landscape data. The observed stretched-exponential relaxation is then that exponential decay seen through a projection onto an observable. Two hypotheses can be tested against each other. In the first, the slowdown is explained by independent excitations with a broad spread of barriers, so the generator is diagonal. In the second, elastic interactions between excitations reshape the spectrum, so the off-diagonal terms are required. The proposal tests whether a temperature-transferable, parameter-free generator can predict the α-relaxation time (the main structural relaxation time, τ_α), the stretching exponent, and the activation energy at temperatures it never saw. It also tests which of the two hypotheses is needed to do so.

Core Claim

The work borrows its mathematical template from Resolvent Estimates for Viscoelastic Systems of Extended Maxwell Type. That paper shows that a viscoelastic equation with a memory integral does not generate a semigroup, meaning a well-behaved time-evolution operator. Rewriting it with auxiliary variables produces a contractive semigroup with exponential decay. Anisotropic Extended Burgers Model extends the same construction to richer relaxation tensors. We transfer this template to glasses. Let $\xi(t)\in\mathbb{R}^K$ be the deviations from equilibrium of the occupation of $K$ classes of local excitations, binned by barrier height and local environment. We posit linear dynamics near equilibrium:

$$\dot\xi = -(\Lambda + C)\,\xi,\qquad \Lambda_{kk} = \nu_0\, e^{-E_k/T},$$

Here $E_k$ is the class barrier, $\nu_0$ is an attempt frequency, $T$ is temperature, and $C$ is the coupling matrix encoding how firing an excitation of class $k'$ changes the rate of class $k$ through elastic stress transfer. A measured relaxation function $\phi(t)$, such as the self-intermediate scattering function or the shear-stress autocorrelation, is predicted as $\phi(t) = w^\top e^{-(\Lambda+C)t}\xi_0$, with $w$ the projection of excitation occupations onto the observable.

The asymptotic decay rate is the spectral gap of $\Lambda+C$. Stretching arises because $w$ and $\xi_0$ spread weight over many eigenmodes. The claim has three parts, and each can fail on its own. First, $\Lambda$ and $C$ measured on static configurations at one set of temperatures predict τ_α and the stretching at held-out temperatures with no refitting. Second, the generator is dissipative: its symmetrized form is positive definite under an equilibrium-weighted inner product, which is the glass analogue of the contractivity proved for the Extended Maxwell reduced system. Third, the coupling $C$ is or is not necessary. Comparing the full generator with its diagonal ablation settles whether the decoupling of dynamics from structure is a disorder effect or an interaction effect.

Research Program

The motivating evidence comes from two lines of glass physics that disagree about what grows on cooling.

Local vs. Cooperative: Unraveling Glass Transition Mechanisms with SEER and Theories of the Glass Transition Based on Local Excitations introduce SEER (Systematic Excitation ExtRaction). It extracts roughly 200 local excitations per configuration in polydisperse soft spheres. On cooling, these excitations involve fewer particles while their barriers rise, and the barriers track the activation energy. In those models the surrounding medium enters only through elasto-plastic "kicks". We replace those empirical kicks with a measured coupling matrix and ask what spectrum it produces.

The second line is Structure–dynamics correlation and its link to fragility and dynamic heterogeneity. It reports that microscopic barriers scale linearly with an inverse structural order parameter, while macroscopic barriers scale quadratically. It also reports that the correlation between structure and dynamics is strongest in attractive, enthalpy-dominated systems. This paper constrains the design: any generator must reproduce the linear-to-quadratic amplification, and it must do so for the right reason.

A third paper supplies a closely related but less structured precedent. Relation between local diffusivity and local inherent structures in the Kob-Andersen Lennard-Jones model writes bulk diffusion as an integral of Arrhenius local diffusivities over the distribution of basin depths in the energy landscape. That is effectively our diagonal model, with no couplings and no spectral analysis.

The program has three stages.

Stage one: build the diagonal part. We reimplement SEER, since no public code is listed, on continuously polydisperse soft repulsive spheres (Poly-IPL10, N = 2000, equilibrated with swap Monte Carlo) and on the Kob-Andersen 80:20 Lennard-Jones mixture. Barriers and local descriptors are binned into $K$ classes to form $\Lambda$.

Stage two: measure the couplings. For a sample of excitations $i$, we push the configuration across $i$'s saddle into its final state, re-extract the excitations within a few cage radii, and record the barrier changes $\Delta E_{ij}$. Linearizing the resulting rate changes around equilibrium occupations gives $C$. We then check whether its symmetric part under the Boltzmann-weighted inner product is positive definite. This is the dissipativity test, and it decides whether exponential decay is guaranteed rather than merely observed.

Stage three: analyze the spectrum. We compute the eigenvalue density of $\Lambda + C$, its gap as a function of temperature, and the weights that $w$ places on each mode. Together these predict $\phi(t)$ and its stretching exponent.

We also prepare a secondary analysis of the linear-versus-quadratic barrier tension. A purely diagonal model with roughly Gaussian barrier spread $\sigma$ already gives $\log\tau_\alpha \approx \bar E/T + \sigma^2/(2T^2)$, because averaging $e^{E/T}$ over a Gaussian yields a variance term. That term can mimic quadratic growth if $\sigma$ scales with the order parameter. The full generator can produce quadratic growth by a different route, through the coupling. We report which route the data support, rather than assume the coupling is needed. The contact-geometric convergence results of Expectation Variables on a Para-Contact Metric Manifold assume a well-behaved Markov generator. Our measured generator is exactly such an object, so if dissipativity holds, those exponential-convergence guarantees apply to the excitation expectations. We treat this as a consequence to note, not a core deliverable.

Baselines

Naive reference: a stretched-exponential fit. The do-nothing reference fits a Kohlrausch stretched exponential (or a Vogel-Fulcher-Tammann law for τ_α) to the training temperatures and extrapolates it to held-out temperatures. It uses the same simulation trajectories and has two or three free parameters. Beating it means a lower extrapolation error in log τ_α at the coldest held-out temperature, with no parameters refit there.

Strongest physics competitor: SEER local-barrier prediction. This is the activation-energy prediction of Theories of the Glass Transition Based on Local Excitations, which predicts relaxation from the local barrier distribution and the Debye-Waller relation $E_a/T = 3.7 + 20.1x + 11.5x^2$. It reports local-barrier prediction accurate over five orders of magnitude in relaxation time on Poly-IPL10. No per-temperature error value is reported on Lacuna, so we will recompute it on our own extraction to get a number to beat. It is the strongest competitor because it uses the same excitation data. Our generator must match it on τ_α and also predict what it does not: the shape of $\phi(t)$, the stretching exponent, and the stress autocorrelation.

Diagonal ablation. The key ablation sets $C=0$. This reduces our model to an excitation-level version of the basin-integral picture in Relation between local diffusivity. It uses identical $\Lambda$, binning, and $w$. If it predicts as well as the full model, coupling is unnecessary, and we will say so.

Black-box dissipative closure. OnsagerNet and GFINN closures, from Learning Macroscopic Dynamics from Partial Microscopic Observations, are trained with its codebase MLDS-NUS/Learn-Partial on the same LAMMPS trajectories. They get matched latent dimension $K$ and a matched hyperparameter budget. These are dissipative by construction, but their latent variables are learned rather than physically identified, and they have only been demonstrated on simple Lennard-Jones fluids. Beating them means better extrapolation to held-out temperatures, where a learned closure has no physical reason to transfer.

Per-particle reference point. For particle-level predictions we compare against Geo-GNN, a graph neural network enriched with three-body bond-angle features, from Geometry-enhanced graph neural network for learning the smoothness of glassy dynamics. It reports a Pearson correlation of about 0.798 with dynamic propensity in the caging regime ($S(q,t)=0.8$) on Kob-Andersen with N = 4096, against about 0.515 for a standard graph neural network. We use its 400/400 train/test configuration split. We do not expect to beat it on propensity. The comparison asks whether a few-parameter physical model comes within reach of it on the long-time regime.

Why not just…?

Why not just fit a sum of exponentials (a Prony series) to the relaxation function? Any stretched exponential can be fit this way, so a good fit proves nothing. Our modes and rates come from static landscape data, and the test is prediction at temperatures excluded from all fitting. A Prony series has no way to make that prediction without refitting.

Why not just use independent barriers, a trap-model-style integral over local relaxation rates? This is a real threat. The Gaussian-variance argument above shows it could, in principle, produce both the stretching and the apparent quadratic barrier growth reported in Structure–dynamics correlation. The diagonal ablation is exactly this alternative, run on identical inputs. If it matches the full generator within statistical error at held-out temperatures, the conclusion is that disorder, not interaction, drives the decoupling. That would itself be a clear result about the local-versus-cooperative debate.

Why not just train a dissipative neural closure end to end? The Learn-Partial pipeline would likely fit the training temperatures as well as or better than our model. Its latent variables, however, carry no physical meaning, so nothing ties them to how barriers change with temperature. The held-out temperature experiment shows whether that matters. If the learned closure extrapolates equally well, our model's advantage reduces to interpretability, and we will report it that way.

Why not just predict mobility with a stronger graph neural network? Geo-GNN already reaches about 0.8 correlation, but it outputs a per-particle propensity at one timescale. It does not output a relaxation law, a decay rate, or a guarantee of dissipation, and its own analysis leaves the causal mechanism open. Our generator answers a different question. We use Geo-GNN only as a reference ceiling for the particle-level comparison.

Proposed Experiments

The decisive experiment is temperature transfer on Poly-IPL10. We build $\Lambda$ and $C$ from configurations at the three warmest supercooled temperatures. We then predict τ_α, the stretching exponent, and $E_a$ at the two coldest temperatures, and compare against flash-reheating measurements of $E_a$ and direct molecular dynamics of $\phi(t)$. The full model is judged against all baselines by error in log τ_α and in the stretching exponent. The same protocol on Kob-Andersen across the densities ρ = 1.1 to 1.6 studied in Structure–dynamics correlation tests the enthalpy-versus-entropy prediction. That paper finds structure-dynamics correlation weakest in the purely repulsive WCA truncation, so we predict the measured coupling $C$ should be weakest, and the diagonal model closest to adequate, for WCA. If this ordering comes out the other way, it falsifies the interpretation of $C$ as elastic facilitation.

A second check is spectral. As temperature drops, the gap of $\Lambda+C$ should close faster than the gap of $\Lambda$ alone if coupling matters. It should also scale quadratically in the inverse structural order parameter while the diagonal entries scale linearly. A third check is dissipativity: if the symmetrized generator acquires negative eigenvalues at the lowest temperatures, the near-equilibrium linear picture breaks down there. That result would mark where exponential-decay theory stops applying to glasses, which we would report as a finding rather than hide.

Finite-size effects are checked by repeating coupling measurements at N = 2000 and N = 8000. This matters because the 155-particle study in Relation between local diffusivity could not resolve elastic stress transfer at long range.