William Witt1
University of Cambridge1
Data-driven (or machine-learned) interatomic potentials have become an increasingly vital tool for the computational materials scientist. Created from the results of small quantum-mechanical calculations, such potentials enable first-principles-accurate simulations on unprecedented length- and time-scales. This talk will summarize recent work towards automated production of Atomic-Cluster-Expansion-type potentials, where the training protocol is adapted from the random structure searching approach to atomistic structure prediction.