Skoltech reports active-learning method for large composite-material simulations

Researchers at Skoltech reported a machine-learning approach for modeling heterogeneous materials that uses active learning on local chemical configurations. The method identifies unreliable energy predictions during a simulation, sends selected fragments for density-functional-theory calculations, then adds them to training data and retrains the potential. In a WC-Co composite case study, the researchers said it handled systems containing tens of thousands of atoms with accuracy comparable to direct DFT calculations and described brittle-to-ductile behavior as cobalt content increased.