Efficient Generation of Materials Data for Machine Learning

The biggest bottleneck to machine learning (ML) for materials science is the generation of training data. In this talk, I will discuss various approaches to efficiently generate and use materials data to develop ML models. For instance, I will demonstrate the use of universal interatomic potentials to pre-generate a large configuration space of structures, as well as a DImensionality Rduced Encoded Clusters with sTratified (DIRECT) sampling approach to create a robust training set for an ML interatomic potential (MLIP). I will also discuss the application of multi-fidelity techniques to maximize the return on scarce, high quality data. While a major focus of this talk will be on MLIP development, I will also highlight the applicability of these techniques to other ML-enabled applications.