Machine Learning Interatomic Potentials for Simulations of Materials at Extreme Conditions

Molecular dynamics (MD) simulations have become an essential tool for investigating the physical and chemical behavior of materials under extreme conditions. By predicting the behavior of every atom in a system based on mathematical descriptions of interatomic interactions, MD provides a detailed atomic-scale understanding of material dynamics with fine time resolution, which is often unattainable through experiments. However, to enable transformative scientific discoveries, MD requires highly accurate models of how atoms interact with each other, almost as precise as those from quantum-mechanical density functional theory. MD also needs to simulate very large billion-atom systems with such quantum accuracy. To address these challenges, we are developing machine learning interatomic potentials (MLIPs) by (1) efficiently fingerprinting a unique local atomic environment around each atom using state-of-the-art descriptors, (2) employing a large database of experimentally relevant quantum molecular dynamics (QMD) data, (3) training the MLIPs with robust machine learning methodologies, and (4) extensively validating them against QMD and experimental data. The efficient implementation of these MLIPs on GPU-based exascale supercomputers enable us to perform quantum-accurate molecular dynamics simulations at experimental time and length scales, making a direct comparison with laser-driven dynamic compression experiments possible. The development of Machine Learning Informed Physics (MLIP) is a crucial component of each research project we undertake. It enables groundbreaking simulation campaigns at exascale supercomputers.