Building Useful Machine-Learned Interatomic Potentials

Interatomic Potentials have long been used for atomistic modeling where the interesting questions are out of reach by first-principles approaches. Traditional empirical potentials are typically fitted to experimental data. They typically have poor general accuracy but are physically well-behaved. On the other hand, machine-learned interatomic potentials are far more expressive than physically motivated interatomic potentials like Lennard-Jones, Stillinger-Weber, Embedded Atom Potentials, etc., but they are also more likely to be completely wrong outside of the training domain, are more difficult to train reliably, and are computationally expensive. We have developed MLIPs for the Hf-Ni-Ti shape memory alloy. We share cautionary tales, best practices for generating training sets, and demonstrate how community tools make for "easy entry" to realistic thermodynamic modeling with these potentials