Machine Learning-Driven Evolutionary Search and Theoretical Analysis of Novel 2D Grain Boundaries

Defects such as grain boundaries (GBs) present during the synthesis of two-dimensional (2D) materials often impede the exploitation of their unique properties. On the other hand, these imperfections, along with lateral heterostructures formed between dissimilar 2D materials, can also be used to tailor material properties for specific applications. Theoretical examination of GBs or interfaces in 2D materials, a precursor to experimental endeavors, demands the prediction of their physically realistic atomic structures—a challenge due to the immense potential design space. This study presents a workflow employing a genetic algorithm (GA) integrated with different machine learning (ML) techniques to predict 2D interfaces. We outline the methodological framework for constructing 2D material interfaces, encompassing supercell construction, GA operations on GB structures, and defining a fitness function. Enhancements to the initial GA approach include the integration of graph isomorphism checks to predict metastable GB phases and the incorporation of a multi-objective GA (MOGA) for simultaneous multiple fitness evaluations. Extending the workflow to novel 2D materials, we developed a graph neural network (GNN) for the efficient and precise prediction of material energies. The GNN, trained on two datasets derived from density functional theory (DFT) and empirical potential energy calculations, demonstrated high accuracy. This refined GA workflow, augmented by the GNN model, was applied to the search for blue phosphorene (BlueP) GBs. Comparing GNN-guided GA outcomes with empirical potential and DFT results validated the GNN's replication of energy hypersurfaces of DFT and empirical potential. Application of our workflow enabled the prediction and DFT validation of multiple previously unreported GB phases for silicene and BlueP. We investigated the formation energy and mechanical properties of silicene GBs, uncovering insights instrumental for experimental research. Moreover, we explored the transition of band structures for BlueP GBs induced by GBs, indicating opportunities for defect engineering. Our findings affirm that the developed workflow expedites the discovery of novel 2D interfaces, deepens our understanding of their properties, and facilitates defect engineering for prospective technological applications.