Speaker
Description
The emergence of foundational machine learning interatomic potentials (MLIPs) represents a transformative shift in atomistic materials simulations. These models aim to provide a scalable bridge between the predictive accuracy of first-principles methods and the temporal and spatial scales required for complex molecular dynamics simulations. Unlike traditional "bespoke" potentials tailored to specific chemical compositions or phases, foundational MLIPs are designed for universal applicability, demonstrating remarkable generalization across diverse chemical spaces and robustness in out-of-distribution environments.
This presentation focuses on the Graph Atomic Cluster Expansion (GRACE), a framework which provides a complete and efficient description of atomic interactions and unifies many current MLIP approaches. We will demonstrate the versatility of GRACE in simulations of thermodynamic, functional and mechanical properties across a broad spectrum of multicomponent systems, ranging from complex alloys to functional ceramics. Furthermore, we will critically evaluate the current issues surrounding foundation models including data quality and integrity, uncertainty quantification and strategies for ensuring reliability when moving into unknown regions of the potential energy surface, and model distillation as a path toward deriving efficient, task-specific models.