Riemannian Stochastic Interpolants for Amorphous Particle Systems
Louis Grenioux, Leonardo Galliano, Ludovic Berthier, Giulio Biroli, Marylou Gabrié · 2025
Read Riemannian Stochastic Interpolants for Amorphous Particle Systems by Louis Grenioux, Leonardo Galliano, Ludovic Berthier, Giulio Biroli, Marylou Gabrié (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-16607.txt.
Introduction
Louis Grenioux and coauthors build Boltzmann generators for glasses, which are disordered solids with no repeating pattern. Sampling fair layouts of such particle systems at balance is slow, and past generator tools worked for crystals and large molecules but not for glasses.
Their method folds the needed symmetries straight into flow-based samplers that respect curved spaces and wrap-around edges, using graph networks that treat like particles alike. Tests show that built-in symmetry sharpens accuracy a lot. They also expose a deep flaw: small numeric errors pile up when odds are tracked, breaking exact reweighting. That flaw marks a basic limit for flow-based tools in this field.
Catalog record
| CREATOR | Louis Grenioux, Leonardo Galliano, Ludovic Berthier, Giulio Biroli, Marylou Gabrié |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 81 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2512-16607
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Riemannian Stochastic Interpolants for Amorphous Particle Systems Louis Grenioux, Leonardo Galliano, Ludovic Berthier, Giulio Biroli, Marylou Gabrié Abstract Modern generative models hold great promise for accelerating diverse tasks involving the simulation of physical systems, but they must be adapted to the specific constraints of each domain. Significant progress has been made for biomolecules and crystalline materials. Here, we address amorphous materials (glasses), which are disordered particle systems lacking atomic periodicity. Sampling equilibrium configurations of glass-forming materials is a notoriously slow and difficult task. This obstacle could be overcome by developing a generative framework capable of producing equilibrium configurations with well-defined likelihoods. In this work, we address this challenge by leveraging an equivariant Riemannian stochastic interpolation framework which combines Riemannian stochastic interpolant and equivariant flow matching. Our method rigorously incorporates periodic boundary conditions and the symmetries of multi-component particle systems, adapting an equivariant graph neural network to operate directly on the torus. Our numerical experiments on model amorphous systems demonstrate that enforcing geometric and symmetry constraints significantly improves generative performance. 1 Introduction Successes of diffusion models, flow matching, and stochastic interpolants in generative modeling have led to scientific applications, particularly for generating microscopic configurations in physical systems. Such physics systems pose unique challenges: data are often scarce, and substantial prior knowledge exists in the form of invariances. Adapting generative models to these constraints has yielded notable successes for biomolecules ( 23 ; 39 ) and their ligands ( 17 ) , or for crystalline materials ( 47 ; 26 ) . Generative models are often trained to approximate an unknown data distribution and evaluated through visual fidelity or likelihood metrics. By contrast, sampling from a known energy landscape requires methods that (i) efficiently generate configurations with high Boltzmann weight and (ii) provide tractable likelihoods or importance weights so that thermodynamic averages remain unbiased; as already pioneered in diverse areas of physics ( 49 ; 57 ; 2 ) . This distinction motivates different algorithmic and evaluation choices In this work, we focus on sampling amorphous particle systems — disordered arrangements of interacting particles, possibly of different subtypes, that lack the atomic periodicity of crystals. The most prominent examples are structural glasses and glass-forming liquids ( 8 ) . A major obstacle to their theoretical understanding lies in the extremely long equilibration timescales, even at moderate system sizes , which severely limit the applicability of traditional numerical methods, such as molecular dynamics or local Monte Carlo, to sample from the equilibrium Boltzmann distribution ( 9 ) . The state-of-the-art swap Monte Carlo algorithm extends this reach, but remains limited to specific models ( 48 ) . These limitations highlight the potential of generative models as an alternative route to equilibrium sampling in amorphous materials. To take into account symmetries of amorphous particle systems, prior work considered equivariant continuous normalizing flows (CNFs) trained via maximum likelihood ( 33 ) , but this approach is computationally expensive. Diffusion-based models, by contrast, offer a simulation-free and scalable alternative ( 60 ) ; however, while they generate visually realistic amorphous samples, they lack tractable likelihoods and therefore cannot be used to correctly sample the Boltzmann distribution. Motivated by these limitations, we focus on flow matching and, more generally, stochastic interpolants ( 5 ; 3 ; 41 ) , that can (i) be trained using a simulation-free mean-squared loss while they admit a CNF formulation and therefore can (ii) easily incorporate symmetries using an equivariant velocity field ( 37 ; 22 ) and (iii) have a tractable likelihood to unbias the generated samples ( 13 ) . Moreover, recent extensions to Riemannian manifolds ( 12 ; 58 ) make this framework particularly well suited to simulations using periodic boundary conditions (PBC), as required to minimize boundary effects in amorphous materials, a direction also explored for crystalline systems in 47 . Guided by these goals, we make the following contributions: • We introduce the equivariant Riemannian stochastic interpolant (eRSI) framework, which extends Riemannian stochastic interpolants by incorporating invariance constraints tailored to amorphous particle systems; • We prove that the optimal marginal paths and velocity fields induced by our objective respect the full symmetry group relevant to amorphous systems with multiple species on the torus. These theoretical guarantees are original contributions to the flow matching and stochastic interpolant frameworks. • We adapt the graph neural network of 52 to respect the full set of symmetries relevant for amorphous materials; • We apply our framework to a canonical model of metallic glass formers widely used in theoretical and computational studies of amorphous materials, and benchmark it against symmetry- and geometry-agnostic baselines. Our approach yields higher-quality generations, both in individual snapshots and in averaged physical observables (via importance sampling reweighting), while also exhibiting improved scalability with system size. 2 Preliminaries 2.1 Amorphous particle systems Structure of the configuration space. A configuration of N particles is denoted by C\in{\mathcal{C}} and decomposed as C=(s,X) , where s=(s_{1},\ldots,s_{N})\in{\mathcal{S}}^{N} encodes the species of each particle within a finite set {\mathcal{S}} , and X specifies their spatial coordinates over {\mathcal{M}}^{N} where {\color[rgb]{0,0,0}{\mathcal{M}}=[0,L)^{d}} is the d -dimensional flat torus of length L>0 . On the torus, particle coordinates are not unique: for any (s,X)\in{\mathcal{C}} , all configurations of the form (s^{\prime},X^{\prime}) with s^{\prime}=s,\;X^{\prime}=X+kL,\;\;k\in{\mathbb{Z}}^{Nd}, represent the same physical state. This ambiguity is lifted by applying component-wise the modulo operator A\;\%\;L=A-\left\lfloor\frac{A}{L}\right\rfloor L\;,\qquad A\in{\mathbb{R}}\;, mapping all equivalent coordinates back into the fundamental domain of the torus. The flat torus implements periodic boundary conditions (PBC), as particles exiting one side of the domain re-enter from the opposite side. We consider the nearest image distance \mathrm{d}_{{\mathcal{M}}} \mathrm{d}_{{\mathcal{M}}}(X,Y)=\min_{k\in{\mathbb{Z}}^{d}}\mathrm{d}_{E}\left(X,Y+kL\right)\;, (1) which is a distance on {\mathcal{M}} but not on {\mathbb{R}}^{d} since it doesn’t satisfy the triangular equality. Equilibrium distribution. Denoting X_{(i)}\in{\mathcal{M}} the coordinates of the i -th particle, the system is governed by the potential energy \displaystyle\mathrm{U}_{\star}(s,X)\;=\;\sum_{i=1}^{N}\sum_{j<i}^{N}\mathrm{W}\left(s_{i},s_{j},\mathrm{d}_{{\mathcal{M}}}(X_{(i)},X_{(j)})\right)\;, (2) where \mathrm{W}:{\mathcal{S}}\times{\mathcal{S}}\times{\mathbb{R}}^{+}\to{\mathbb{R}} is a pairwise interaction potential. At temperature T , the equilibrium distribution is the Boltzmann measure p_{\star}(\mathrm{d}s,\mathrm{d}\text{vol}_{X})\;=\;\frac{1}{\mathcal{Z}}\exp\left(-\frac{\mathrm{U}_{\star}(s,X)}{k_{\rm B}T}\right)\,\mathrm{d}s\,\mathrm{d}\text{vol}_{X}\;, (3) where k_{\rm B} is the Boltzmann constant, \mathrm{d}\text{vol}_{X} is the volume element over {\mathcal{M}} and the partition function \mathcal{Z}=\int\exp\left(-\mathrm{U}_{\star}(s,X)/k_{\rm B}T\right)\mathrm{d}s\mathrm{d}\text{vol}_{X} ensures the normalization. This joint distribution implements a uniform distribution on the specie of particles and correspond to sampling from a “semi-grand” canonical ensemble in the language of statistical mechanics, where the total number of particles is fixed but the composition is not. We will also be interested in sampling from p_{\star} conditioned on a composition. Our objective is to build a generative model that approximates p_{\star} (or its conditional) to allow efficient sampling of equilibrium configurations. In the following, we examine the symmetries on the multidimensional torus induced by the potential, and introduce Riemannian stochastic interpolants, the generative modeling framework underlying our approach. 2.2 Invariances on the torus space Figure 1: Illustration of invariance group actions on a configuration of the 2D 10-particle IPL model. The system contains two particle species with different effective diameters (see Section 5 ). The symmetrized transformation shown corresponds to a 90^{\circ} counterclockwise rotation – equivalently, an axial symmetry with respect to the diagonal from the bottom-left to the top-right corner. Definition 1 . Let G denote a set of group actions acting on the configuration space {\mathcal{C}} . A probability density q on {\mathcal{C}} is said to be