ATLAS: A Foundation Neural Sampler for Amorphous Materials
ATLAS is presented as a diffusion-based sampler that generates Boltzmann-distributed amorphous structures from a target energy function.
The paper says ATLAS uses an equivariant graph neural network and generalizes across system size, temperature, and composition. In Kob-Andersen tests, it matches parallel tempering MCMC distributions and thermodynamic quantities while using more than 500 times fewer energy evaluations. In metallic-glass cases, it recovers short-range-order trends and steers structures toward target order parameters and bulk moduli. The authors also pair it with a language-model agent to search an eight-element glass space, reporting a Pareto frontier within 480 oracle evaluations. ArXiv · AI/CL/LG's note
The paper says ATLAS uses an equivariant graph neural network and generalizes across system size, temperature, and composition. In Kob-Andersen tests, it matches parallel tempering MCMC distributions and thermodynamic quantities while using more than 500 times fewer energy evaluations. In metallic-glass cases, it recovers short-range-order trends and steers structures toward target order parameters and bulk moduli. The authors also pair it with a language-model agent to search an eight-element glass space, reporting a Pareto frontier within 480 oracle evaluations. ArXiv · AI/CL/LG's note
score 5