Graph Neural Network and Tight Binding Approaches for Fast and Accurate Predictions of Defect Energetics

Defects play an important role for predicting materials behavior. While there are have been a lot of efforts for underatnding and harnessing the properties of perfect materials, such applications are limited for defect based materials. Some of the key challenges are: availability of a large enough database of defect properties, and a generalized frameowrk for predicting such behavior. In this work, we generate a vacancy formation energy database for 2D and 3D materials available in the JARVIS-DFT database with 1000 entries using density functional theory. This database consists of thermodynamically stable elemental solids, 2D chalcogenides and 3D binary solids. In order to accelrate such predictions, we train tight-binding and graph neural models that can accurately predict the defect formation behavior. For the graph neural network predictions, we use the atomostic line graph neural network model (ALIGNN) and for the tight-binding model we use the tb3py package, which uses three body interactions as well as the two body interactions. All the database and tools from this work are publicly distributed to enhance reproducibility.