Automated training of machine learned potentials with Bayesian active learning

Machine learned interatomic potentials are often manually trained and restricted to single-component and nonreactive systems, severely limiting the practical application of these models. We present an adaptive Bayesian inference method for automating the training of multi-element interatomic potentials using structures drawn “on the fly” from molecular dynamics simulations. Within an online active learning algorithm, the internal uncertainty of a Gaussian process (GP) regression model is used to decide whether to accept the model prediction or to perform a first principles calculation to augment the training set of the model and reoptimize its hyperparameters. Model uncertainties are derived from the variance of the predictive posterior distribution of the GP, which is shown to correlate with true model error on independent test sets. The GP models are based on low-dimensional, explicitly multi-element two- and three-body kernels that can be mapped onto highly efficient cubic spline models suitable for large scale molecular dynamics simulations. Applications to a range of single- and multi-element systems will be discussed, including vacancy and adatom migration in Aluminum, fast-ion diffusion in AgI, and surface segregation in Pd/Ag alloys.