Accelerating atomistic modelling with active learning

Machine learning provides a path toward fast, accurate, and large-scale materials simulation, promising to combine the accuracy of ab initio methods with the computational efficiency of analytical potentials. However, training current state-of-the-art models, which include Neural Network Potentials and Gaussian Approximation Potentials, often requires hundreds of CPU hours and databases containing tens of thousands of chemical environments. Moreover, these potentials are trustworthy only for chemical configurations that fall within the training set and have so far been restricted to single- or few-component systems. In this talk, we present a multi-component on-the-fly learning scheme that refines the machine learned force-field when a new chemical configuration is encountered, opening the door to ML-driven molecular dynamics that can capture complex many-body dynamics spanning previously unprecedented length- and time-scales.