Data-driven prediction of dislocation plasticity – DAPREDIS

Metals are structural materials par excellence as they posses a rare combination of strength and ductility. These properties are due to the motion and interaction of dislocations, line defects in the crystal structure. Quantitative prediction of dislocation-enabled plasticity is a “grand challenge” except in a handful of situations, primarily due to their multiscale nature: elementary dislocation properties are controlled by atomic structure, dislocations and other defects form mesoscale microstructures, whilst all mechanisms must be adequately averaged to define the macroscopic response. A key issue is that current multiscale methods transfer information across scales with rigid models that lack measures of data uncertainty or similarity, preventing data-driven analysis and prediction at scale. Similar issues faced construction of molecular dynamics (MD) interatomic potentials from ab initio data; the solution was a descriptor representation of atomic structures. With uncertainty/similarity measures, data-driven methods have transformed MD potential accuracy, error estimation and production speed. DAPREDIS will develop descriptors for dislocation dynamics to revolutionise how information transfer to the mesoscale, and use recent innovations of the team to predict the future of simulation trajectories, to overcome timescale limitations.