Atomistic Dislocation Dynamics Using Bayesian Force Fields

Dislocation dynamics present a difficult simulation task for existing classical and ab initio methods due to the simultaneous accuracy and length-scale requirements for reliable simulation and comparison to experimental data. These limitations ultimately prohibit advanced understanding of plastic deformation (e.g. via edge and screw dislocations) of materials. Here, we develop a Bayesian machine-learned force field (MLFF) from first-principles training data that extends quantum-mechanical accuracy to large length-scale molecular dynamics simulations which permit direct observation of high-temperature and high-stress dislocation dynamics in Cu with atomistic resolution. In concert, a generalizable training protocol is defined for the construction of MLFFs for description of dislocations, which can be trivially employed for other metal and alloy systems of interest. The resulting FLARE MLFF provides excellent predictions of static bulk properties (e.g. bulk modulus and elastic tensor), stacking fault energies, and reliable, dynamic evolution of edge and screw dislocations, as well as cross-slip mechanisms that allow for non-planar movement of screw dislocations across a broad range of temperature and applied stress. Each of these observations from molecular dynamics simulations are then compared to experimental data, including the screw and edge stacking fault widths, edge dislocation mobility coefficient, and screw dislocation cross-slip activation energy, where FLARE provides the best agreement among the methods tested. Such simulations permit increased understanding of material responses to extreme conditions by direct atomistic insight at the mesoscale with near quantum mechanical accuracy.