Speaker: Ji Qi (UC San Diego and LLNL)
Session Chair: Daniel Schwalbe-Koda (UC Los Angeles)

Abstract: Machine learning interatomic potentials (MLIPs) enable accurate simulations of materials at scales beyond conventional first-principles approaches, and they have played increasingly important roles in understanding and design of materials. However, MLIPs are only as accurate and robust as the data they are trained on. In this seminar, I will present DImensionality-Reduced Encoded Clusters with sTratified (DIRECT) sampling as an approach to select a robust training set of structures from a large and complex configuration space. By applying DIRECT sampling on the Materials Project relaxation trajectories dataset with over one million structures and 89 elements, we develop an improved materials 3-body graph network (M3GNet) universal potential that extrapolate more reliably to unseen structures. We further show that molecular dynamics (MD) simulations with universal potentials such as M3GNet can be used in place of expensive ab initio MD to rapidly create a large configuration space for target materials systems. For demonstration, we combined this scheme with DIRECT sampling to develop a reliable moment tensor potential for titanium hydrides without the need for iterative augmentation of training structures.

In this seminar, I will walk through two Jupiter notebooks to showcase DIRECT sampling with the two example cases demonstrated in our manuscript, so that audience can expect to reproduce our major results with no trouble. Hopefully, DIRECT sampling will serve as a straightforward, efficient, useful plug-in for the robust training of MLIPs across any compositional complexity.