Machine learning for exploration of defects in 2D grain boundaries

Grain boundaries (GBs) in two-dimensional (2D) materials have a profound impact on various material properties, yet computationally predicting their realistic interfaces/structures is a challenge. Topological 2D structures can be naturally transferred into graphs, and herein we combine evolutionary algorithms and graph theory for defective structure search. We benchmarked our method on laterally interfacing graphene, and rank-ordered 128 predicted structures according to their corrected formation energy. From the statistical analysis of primitive rings, a correlation was determined between the ring distribution and the formation energy. Our workflow was further expanded to probe silicene interfaces and the associated

extended defects. For accelerated sampling, we also deploy a graph neural network (GNN) as a surrogate for DFT in our evolutionary search. This talk will also discuss the importance of the distribution and extent of training data needed to build an adequate GNN surrogate.