Gaussian Process Interatomic Potentials

Gaussian processes (GPs) are a Bayesian regression technique with big upside for machine learning interatomic potentials. Unlike black-box methods like neural networks, GPs are interpretable, uncertainty-aware, and allow for the integration of physics-informed behavior through the use of carefully designed Bayesian priors. We develop state-of-the-art GP models for interatomic interactions, constructed from both quantum level simulations [1] and experimental scattering data [2].

[1] J. Kermode et al. Improved uncertainty quantification for Gaussian process regression based interatomic potentials. arXiv (2022), https://arxiv.org/abs/2206.08744

[2] B.L. Shanks, et al. Transferable Force Fields from Experimental Scattering Data with Machine Learning Assisted Structure Refinement. J. Phys. Chem. Lett. (2022), https://pubs.acs.org/doi/10.1021/acs.jpclett.2c03163