Machine Learning to Analyze the Atomic Energy Landscape in Poly-Crystalline Materials

Plastic deformation of poly-crystalline materials occurs at defects such as grain boundaries. At a small scale, the plasticity typically consists of atoms at the defect core shifting between metastable positions (rearranging). Predicting these rearrangements at grain boundaries is challenging due to the structural complexity. We adapt a machine learning technique used successfully on disordered glassy materials to study atomic plasticity in crystalline metals. We first catalog the atomic rearrangements that occur, in large MD simulations of poly-crystalline aluminum and nickel at finite temperature. We train a support vector machine (SVM) to identify the local structure surrounding a particle just before it rearranges. The SVM classifies structures as susceptible to rearrangement with 90% accuracy in cross-validation. For each atom, we calculate a value derived from the SVM that is correlated with the susceptibility to rearrange, called softness, which provides a new view on atomic-scale plasticity in poly-crystals. We obtain a well-defined energy barrier for rearrangements for grain boundary particles and find that the average energy barrier is consistent with published experimental measurements.