Invariant Graph Neural Network Models for Accelerated Prediction of Equilibrium Structure, Defect Energetics and Electronic Properties of Semiconductor Materials

Accelerated material property predictions are an important component of large scale search and optimization of new semiconductor device materials. We present ongoing work developing and applying invariant graph neural network models for modeling the properties of multicomponent semiconductor materials. We use such models to predict defect energetics in semiconductor materials. Initial results on prediction of equilibrium interface structures and electronic properties in semiconductor heterointerface systems will be discussed as well. Robust and fast uncertainty quantification methods are of particular concern for predictive design of semiconductor heterointerface systems, for which currently-available reference data is limited. We will discuss how we use graph neural network uncertainty quantification for active learning of force fields, out-of-distribution detection for predictive models, and for targeted allocation of theory-based modeling.