Optimal Development of Transferable Machine Learning Interatomic Potentials using Active Learning

Automated methods for generating atomistic configurations make possible the creation of vast and diverse datasets where potentials that exhibit consistent accuracy across diverse configurations can be trained. However, automated generation frameworks struggle to create configurations of physical relevance to the potential development task of interest. For that reason, a training dataset that combines configurations generated using automated frameworks and domain expertise is expected to yield better performing potentials. Nevertheless, integrating configurations from these two frameworks is not a trivial task given the fact that data-driven methods often yield a very large number of configurations, which places a severe computational burden for calculating the results, and domain-expertise methods are not capable of scaling in order to generate a vast number of configurations. Consequently, there is a critical need for an enhanced training protocol that can integrate both types of configurations in an optimized and data-driven manner. This work addresses this challenge by establishing an automated protocol to train a potential using an ensemble of neural network based potentials and active learning. The developed protocol trains the potential iteratively by automatically incorporating a fixed set of configurations that maximize the information gained.