Advancing Machine-Learned Interatomic Potentials: Enhancing Accuracy and Robustness in Materials Science Applications
Data Science Seminar
Yangshuai Wang (University of British Columbia)
Abstract
The success of molecular simulation relies heavily on the accuracy and efficiency of force-fields. A novel technique named machine-learned interatomic potentials (MLIPs) has been developing rapidly. The idea is to bridge the significant gap in accuracy and capability between ab initio electronic structure models and classical mechanistic models. The MLIPs are becoming part of the standard toolbox of computational materials science as demonstrated by the increasing number of successful applications leading to new scientific discoveries, including amorphous materials, high-pressure systems, phase diagrams and reaction dynamics of molecules. In this talk, I will provide an overview of the current state-of-the-art MLIPs, highlighting recent applications in materials science. Specifically, I will explore the transition from small-but-accurate models like the Atomic Cluster Expansion (ACE) method to large-and-robust models established using the MACE (Message passing neural networks with ACE) architecture. Additionally, I will provide a brief overview of current challenges and potential pathways for resolution, setting the stage for advancements in AI for Science.