Training Data Selection for Accuracy and Transferability of Interatomic Potentials

Abstract

Advances in machine learning (ML) techniques have enabled the development of interatomic potentials that promise both the accuracy of first principles methods and the low-cost, linear scaling, and parallel efficiency of empirical potentials. Despite rapid progress in the last few years, ML-based potentials often struggle to achieve transferability, that is, to provide consistent accuracy across configurations that significantly differ from those used to train the model. In order to truly realize the promise of ML-based interatomic potentials, it is therefore imperative to develop systematic and scalable approaches for the generation of diverse training sets that ensure broad coverage of the space of atomic environments. This work explores a diverse-by-construction approach that leverages the optimization of the entropy of atomic descriptors to create a very large ($>2\cdot10^{5}$ configurations, $>7\cdot10^{6}$ atomic environments) training set for tungsten in an automated manner, i.e., without any human intervention. This dataset is used to train polynomial as well as multiple neural network potentials with different architectures. For comparison, a corresponding family of potentials were also trained on an expert-curated dataset for tungsten. The models trained to entropy-optimized data exhibited vastly superior transferability compared to the expert-curated models. Furthermore, while the models trained with heavy user input (i.e., domain expertise) yield the lowest errors when tested on similar configurations, out-sample predictions are dramatically more robust when the models are trained on a deliberately diverse set of training data. Herein we demonstrate the development of both accurate and transferable ML potentials using automated and data-driven approaches for generating large and diverse training sets.