Accelerating the Discovery of Two-Dimensional Interfaces with Graph Neural Network and Evolutionary Search

Machine learning (ML) methods have attracted considerable attention in the field of materials science due to their flexibility, accuracy, and efficiency compared to traditional simulation approaches for material discovery. Among the various ML methods available, graph neural networks (GNNs) have emerged as a prominent technique for predicting material properties. This is attributed not only to recent advancements in GNNs but also to the inherent correlation between atomic structures and graphs.<br/><br/>One significant challenge in harnessing the remarkable properties of two-dimensional (2D) materials is the occurrence of defects during synthesis, such as grain boundaries (GBs). Predicting the realistic atomic structure of GBs and their impact on material properties is a complex task that surpasses the capabilities of conventional trial-and-error methods, owing to the vast design space involved.<br/><br/>In this abstract, we propose a combination of GNN and evolutionary search methods to address this challenge. We originally developed a search workflow based on genetic algorithms (GA) for exploring 2D GBs applied to graphene and silicene. The workflow involves transforming GBs into graphs and utilizing graph isomorphism checks to identify multiple material states during the GA searches, and multiple novel silicene GBs were predicted. Separately, we developed a GNN architecture capable of predicting the energy of 2D GB structures, which serves as an efficient and accurate surrogate for computationally expensive quantum mechanical simulations. We utilized datasets generated from quantum mechanical simulations for training and testing the GNN model. The high accuracy of the GNN model on the test set demonstrates its ability to provide reliable estimations of quantum mechanical calculations.<br/><br/>In the work presented here, the research combines the GA search with this GNN surrogate, to fully leverage an automated design search. Using this method, we successfully uncovered previously unreported GBs in blue phosphorene. Additionally, the flexibility of the GNN approach allowed us to generalize the method to other 2D materials, thus expediting the process of discovering novel materials.