Rational Design of Machine-Learned Interatomic Potentials

The predicative capability of a molecular dynamics simulation is determined by how accurately the chosen interatomic potential (IAP) captures a material’s bonding characteristics. Integrating machine learning techniques into IAP development has led to powerful improvements in that accuracy, while also introducing new levels of complexity to the training process. The choice of model (hyper)parameters, training set structures, validation tests, and whether and how to use tools such as active learning is frequently a material-specific, highly empirical process. To accelerate development and usage of machine-learned IAPs, new approaches to promote reproducibility and rational design of IAP are needed, ideally ones that are consistent with FAIR (Findable, Accessible, Interoperable, Reusable) research practices. In this talk, I will discuss promising developments in this area, amongst them software tools and techniques our group has developed while training machine-learned IAPs for multicomponent, chemically-complex materials used in extreme environments.