Aidan Thompson1
Sandia National Labs1
The molecular dynamics method (MD), as implemented in the LAMMPS[1] particle simulation code, is a powerful tool for explicitly sampling the phase space distribution of hundreds or billions of physically and chemically interacting atoms. It provides a wealth of information on how particular microscopic interactions lead to a vast range of emergent behaviors on much larger length and time scales, complementing theory and experiment. The overall usefulness of the method is sensitive to how well the chosen interaction potential approximates the true physical and chemical conditions. Over many decades, this has driven the emergence of many prominent interaction potentials that provifde a good tradeoff between accuracy and cost. At the one extreme of minimal complexity are well-established models for simple fluids (Lennard Jones particles), polymer melts (FENE chains), and metals (EAM). In the case of chemical reacting systems, both organic and inorganic, the ReaxFF reactive potential has proven quite effective. More recently, relatively expensive machine-learning (ML) potentials have been found to approach the accuracy of very expensive quantum methods. These are trained to reproduce the energy and forces of many small configurations of atoms obtained from quantum electronic structure calculations e.g. Density Functional Theory. Examples of ML potentials implemented in LAMMPS include Behler-Parinello, GAP, SNAP[2], the Atomic Cluster Expansion (ACE)[3], and ALEGRO. In this talk, I will give a general overview of LAMMPS capabilities, as well as the those the LAMMPS-integrated FitSNAP software [4] for generating quantum-accurate machine-learning interatomic potentials. I will describe several recent scientific applications of LAMMPS that combine innovative physics models, machine-learning interatomic potentials, and extreme scale computing resources.<br/><br/>[1] Thompson et al., Comp. Phys. Comm., 271:108171, 2022. DOI 10.1016/j.cpc.2021.108171 (URL https://www.lammps.org) [2] Thompson et al., J. Comp. Phys., 285:316, 2015. DOI 10.1016/j.jcp.2014.12.018 [3] Lysogorskiy, npj Comp. Mat. 7:1, 2021. DOI 10.1038/s41524-021-00559-9 [4] A. Rohskopf et al., Journal of Open Source Software 8, 5118 (2023). DOI 10.21105/joss.05118 (URL https://fitsnap.github.io)