Predicting grain boundary dislocation structures through multidimensional neural networks and high-throughput phase-field calculations

Abstract

Phase-field modeling has been a useful method for describing crystal defects (such as dislocations and grain boundaries) at the continuous level. However, energy functional based on the crystallography and elasticity, as well as the kinetic governing equations for seeking the equilibrium state, may prevent it from widespread application. Therefore, surrogate models that directly map material properties (i.e., the γ-surfaces) to the equilibrium dislocation configurations are highly necessary from the view of improving both practical usability and computing efficiency. Using output data from phase-field calculation, the current work trains a simple neural network (NN) to investigate the geometrical outline of GB dislocations. Moreover, a deep NN concerning complex input and output is constructed to predict the whole configuration of the dislocation networks and stacking faults. The two NNs build up the relationship between the γ-surface and the GB dislocation configurations, which are validated through phase-field simulation in this work and atomic simulation reported previously. This work provides an important framework that bridges the physics-based model and machine learning techniques through data generation and transmission.