Foundational Graph Atomic Cluster Expansion for Atomistic Thermodynamic and Kinetic Simulation

The atomic-scale simulation of technological materials requires a departure from task-specific approximations toward universal foundations capable of generalizing across diverse chemistries. While first-principles methods offer the necessary accuracy, they lack the scalability required for foundational coverage. This presentation introduces the Graph Atomic Cluster Expansion (GRACE), a framework designed to bridge this gap by providing a mathematically complete and computationally efficient description of atomic interactions.
GRACE unifies equivariant message-passing with high-body-order polynomials, enabling the development of foundational parameterizations that span the periodic table. I will demonstrate that GRACE facilitates robust thermodynamic and kinetic simulations for a wide range of materials, including materials with complex, multi-component compositions. Specific applications will be presented, including the computation of phase diagrams and the prediction of diffusion kinetics in complex solids. Finally, I will address the incorporation of magnetism and charge transfer and discuss the systematic generation of comprehensive training datasets for creating transferable and accurate foundational parameterizations.