Speaker
Description
Machine Learning Interatomic Potentials (MLIPs) are undergoing a paradigm shift from narrowly trained, system-specific models to universal frameworks intended to generalize across the periodic table. While early MLIPs focused on accelerating atomistic simulations for a particular material or chemistry, recent developments emphasize capturing complex thermodynamic and kinetic properties in arbitrary material compositions.
This presentation introduces the Graph Atomic Cluster Expansion (GRACE) and explains how its underlying architecture enables accurate, large-scale simulations of complex material compositions. By efficiently managing the combinatorial explosion inherent to multi-component systems, GRACE preserves linear scaling with system size. The completeness of the underlying basis expansion makes GRACE not only a model but a unifying framework: a wide range of existing MLIP architectures can be recovered as special cases, and the same completeness enables principled extensions to interactions involving dynamic charge transfer and magnetism that have historically challenged conventional atomistic models.
Model accuracy is ultimately bounded by training data, and conventional datasets carry strong thermodynamic and chemical bias toward near-equilibrium structures. GRACE counters this with uncertainty-driven, maximum-entropy, chemistry-agnostic sampling, decoupling structural generation from thermodynamic constraints and establishing a broad physical prior. We demonstrate this on stable molecular dynamics simulations at high temperature and pressure containing more than 90 chemical species simultaneously in a single simulation, resolving emergent structures in complex mixtures — such as high-entropy alloys — without a priori chemical assumptions.