Active Learning of Bayesian Force Fields

Abstract

Simulating matter at the atomistic scale can accelerate drug design and materials discovery, but the most accurate atomistic simulation methods are prohibitively expensive. In the past decade, machine learning (ML) has emerged as a powerful tool for combining the computational efficiency of classical force fields with the accuracy of quantum-mechanical methods such as density functional theory (DFT). However, most modern ML force fields are difficult to train, requiring thousands of expensive DFT calculations and detailed prior knowledge of the material of interest. In this thesis, I present a closed-loop Bayesian inference method that automates the training of many-body ML force fields using structures drawn "on the fly" from molecular dynamics simulations. Our online active learning algorithm uses the internal uncertainty of a Gaussian process (GP) regression model to decide whether to accept the model prediction or to perform a DFT calculation that updates the training set of the model. To enable large-scale molecular dynamics simulations with the resulting force fields, mean predictions of the GP are mapped onto much faster spline-based and parametric models. I discuss applications to superionic diffusion in silver iodide, hydrogen chemisorption on platinum, and martensitic phase transitions in the shape-memory alloy nickel titanium. The method is made available in an open-source software package called FLARE (Fast Learning of Atomistic Rare Events).