Benchmarking Structural Evolution Methods for Training of Machine Learned Interatomic Potentials

Abstract

When creating training data for machine-learned interatomic potentials (MLIPs), it is common to create initial structures and evolve them using molecular dynamics to sample a larger configuration space. We benchmark two other modalities of evolving structures, contour exploration and dimer-method searches against molecular dynamics for their ability to produce diverse and robust training density functional theory data sets for MLIPs. We also discuss the generation of initial structures which are either from known structures or from random structures in detail to further formalize the structure-sourcing processes in the future. The polymorph-rich zirconium-oxygen composition space is used as a rigorous benchmark system for comparing the performance of MLIPs trained on structures generated from these structural evolution methods. Using Behler-Parrinello neural networks as our machine-learned interatomic potential models, we find that contour exploration and the dimer-method searches are generally superior to molecular dynamics in terms of spatial descriptor diversity and statistical accuracy.