ATLAS: A Foundation Neural Sampler for Amorphous Materials

Research Question and Approach

Simulating amorphous materials is notoriously difficult because their potential energy surfaces contain numerous metastable basins, making conventional molecular dynamics and Monte Carlo methods inefficient at low temperatures. The authors introduce ATLAS, a neural sampler that learns a diffusion process to generate Boltzmann-distributed atomic configurations directly from a target energy function. Unlike previous generative models that require precomputed structural datasets, ATLAS is trained using a self-consistent, force-informed bootstrapping approach. It utilizes an E(3)-equivariant graph neural network to learn the forward and backward stochastic dynamics, allowing it to generalize across system sizes, temperatures, and chemical compositions.

ATLAS demonstrates exceptional efficiency and accuracy in sampling amorphous systems. In two-dimensional Kob-Andersen glass models, the sampler reproduces structural distributions, free energies, and entropies with less than 0.2% error in the low-temperature glass regime, while requiring over 500 times fewer energy evaluations than parallel tempering Markov chain Monte Carlo (PT-MCMC). The model exhibits strong transferability, as it can be trained on small systems and applied to significantly larger ones without retraining. Furthermore, the authors show that ATLAS can be steered toward specific target properties—such as mechanical bulk moduli or short-range-order parameters—in complex metallic glasses like Cu-Zr and Cr-Co-Ni. By coupling ATLAS with an LLM-guided agent, the researchers successfully performed inverse design in an eight-element chemical space, identifying a Pareto frontier for high-entropy metallic glasses with minimal oracle evaluations.

ATLAS functions as a foundation model for amorphous materials, unifying equilibrium sampling, thermodynamic estimation, and property-guided design. By amortizing the cost of sampling across different thermodynamic and chemical conditions, it drastically reduces the computational burden of materials discovery. This framework enables researchers to explore complex, high-dimensional configuration spaces that were previously inaccessible, providing a scalable tool for the rational design of disordered materials with tailored mechanical and functional properties.