AutoPot: New Software Automates Construction of Machine-Learning Potentials for Computational Physics

Researchers from Brigham Young University, Lawrence Livermore National Laboratory, the University of Minnesota, UCLA, and Cross Stream Consulting have published a preprint describing an inverse design strategy for building bespoke interatomic potentials (IPs) using an information-matching (IM) active learning framework. Interatomic potentials allow atomistic simulations at scales inaccessible to first-principles methods, but their accuracy depends heavily on how training data is selected. The IM approach differs from conventional active learning strategies by requiring that selected training data deliver at least as much parameter-space information as needed to achieve user-specified uncertainty targets for particular material properties. Because directly simulating plastic strength is computationally expensive, the team employed an indirect strategy targeting cheaper intermediate quantities of interest that correlate with strength. Results showed that IM enabled precise parameter constraints with minimal training data, yielding accurate predictions for both intermediate and final quantities. However, the authors acknowledge that model error remains a key limitation, and they propose a post hoc uncertainty inflation correction as a practical mitigation. The study illustrates both the potential and the current boundaries of uncertainty-aware active learning for complex materials property prediction.