TL;DR
Sampling equilibrium configurations of amorphous systems (like glasses and supercooled liquids) is notoriously difficult due to astronomical relaxation timescales and the need for unbiased physical observables. This paper introduces the Equivariant Riemannian Stochastic Interpolant (eRSI), combining Riemannian flow matching on a flat torus with symmetry-preserving graph neural networks to provide tractable likelihoods for exact Boltzmann reweighting.
What Problem Does This Paper Address?
In condensed matter physics, amorphous materials—such as structural glasses and liquid mixtures—lack the crystalline periodicity of lattice solids and the isolated structure of single biomolecules. Simulating their equilibrium states using standard methods (Molecular Dynamics or local Monte Carlo) suffers from severe kinetic slowdowns.
Generative models (like normalizing flows or diffusions) can propose independent candidate states, but doing so for physical sampling imposes two major constraints:
- Periodic Boundary Conditions (PBCs): Particles interact over a flat torus ($\mathbb{T}^d$), meaning particle coordinates wrap around box edges. Standard Euclidean diffusion or flow matching models ignore this, creating severe unphysical particle overlaps at the boundaries.
- Tractable Likelihoods: Estimating unbiased thermodynamic observables (e.g., specific heat, free energy) via importance sampling requires calculating exact configuration likelihoods $q(C)$. Standard diffusion models lack cheap, exact likelihoods, while standard continuous normalizing flows trained via maximum likelihood are computationally expensive due to Jacobian divergence computations during training.
Methods
The authors propose eRSI (Equivariant Riemannian Stochastic Interpolant), which unifies manifold flow matching with group equivariance tailored to amorphous systems:
- Geodesic Interpolation on the Torus: The interpolation follows torus geodesics defined via exponential and logarithmic maps on $\mathcal{M} = [0, L)^d$: $$I_L(t, x_0, x_1) = \exp_{x_0}(t \log_{x_0}(x_1))$$
- Symmetry Group on the Torus ($\mathcal{G}_{\mathcal{C}}$): The target Boltzmann measure is invariant under particle permutations ($S_N$), uniform spatial translations, and hyperoctahedral symmetries (coordinate axis swaps and sign flips, $B_d$). The authors prove that Riemannian stochastic interpolation retains these symmetries when using equivariant vector fields.
- Torus-Equivariant Graph Neural Network: The authors adapt the E($n$)-equivariant GNN (EGNN) architecture so that coordinate differences and distance calculations explicitly use the torus logarithmic map $\log_{X_{(i)}}(X_{(j)})$ and geodesic metric $d_{\mathcal{M}}$, ensuring the learned velocity field $\hat{v}(t, C)$ is rigorously $\mathcal{G}_{\mathcal{C}}$-equivariant.
- Minibatch Optimal Transport: Transport costs during training are minimized by matching base and target configurations using the Hungarian algorithm over particle permutations under the torus pre-metric.
Datasets
The framework was evaluated on two canonical computational physics benchmarks:
- 2D Inverse Power Law (IPL) Binary Mixture: A standard benchmark for glass formers consisting of an equimolar binary mixture with interaction potential $W(r) \propto (\sigma / r)^{12}$ shifted to zero at a cutoff. Tested at $N = 10$ and $N = 44$ particles at number density $\rho = 0.5$ and temperature $T = 0.1$.
- 2D Ternary Kob–Andersen (KA) Mixture: A 3-species system ($N = 44$, composition $5/11, 3/11, 3/11$) interacting via shifted Lennard-Jones-like potentials at temperatures $T = 1.0$ (liquid) and $T = 0.32$ (supercooled glassy regime).
- Datasets of $10^5$ independent equilibrium configurations were generated using long Metropolis–Hastings Monte Carlo simulations (augmented with swap Monte Carlo moves for low-temperature KA).
Key Insights and Ideas
- Symmetries naturally transfer to the optimal velocity field: The authors provide theoretical proofs showing that if the base and target distributions are invariant under the torus symmetry group $\mathcal{G}_{\mathcal{C}}$ and the interpolant is equivariant, the optimal velocity field $v^\star(t, C)$ is guaranteed to be $\mathcal{G}_{\mathcal{C}}$-equivariant.
- Geometry is non-negotiable for importance sampling: Standard Euclidean equivariant flow matching (
eFM) places particles without respecting the periodic box boundaries. When evaluated under periodic boundaries, configurations have near-zero distances across edges, resulting in near-infinite potential energies ($U^\star \to \infty$). These states get assigned zero importance weights, causing importance sampling estimators to break down completely as system size grows. - Reweighting handles imperfect generative models: Even when raw model samples have slight deviations in structural statistics (such as the radial distribution function $g(r)$), having a tractable likelihood allows self-normalized importance sampling (IS) to recover exact equilibrium expectations.
Key Results
- Ablation Comparison: Evaluated against eFM (Euclidean symmetries, no torus geometry) and RSI (torus geometry, but non-equivariant MLP):
- $N = 10$ IPL: RSI completely failed to capture structure ($g(r)$ was flat). eFM generated reasonable samples but exhibited noisy reweighted estimators. eRSI converged rapidly to ground-truth potential energy $U$ and specific heat $c_V$.
- $N = 44$ IPL: eFM completely failed; its energy and specific heat estimators stabilized far away from the ground truth due to boundary overlaps. In contrast, eRSI successfully matched target observables and radial distribution profiles $g(r)$.
- Effective Sample Size (ESS): The ESS of eRSI broke away from the $1/R$ penalty at a much smaller sample threshold ($R \approx 10^3$ for $N=10$) compared to eFM ($R \approx 10^4$), while non-equivariant RSI never broke away from $1/R$ over the tested budget.
- Scaling: Increasing eRSI network capacity (from 22k up to ~580k parameters) and training dataset size ($10^4 \to 10^5$) significantly improved generative quality and reduced the sample budget needed for reweighting.
Limitations and Future Work
- High-Dimensional Scaling & 3D: Experiments were demonstrated on 2D systems with moderate particle counts ($N \le 44$). While extending the architecture to 3D is theoretically straightforward, scaling to larger systems poses sample-efficiency challenges for importance sampling.
- Inference Bottleneck: Computing exact divergence traces via automatic differentiation along ODE trajectories becomes computationally expensive during importance weight evaluation for large particle batches.
- Training Data Requirement: The model relies on large pre-computed MCMC datasets ($10^5$ configurations). The authors propose integrating self-training or adaptive MCMC methods (e.g., sequential tempering or flow-augmented MCMC) to bypass the need for extensive ground-truth training data.
Summary updated