Exploration of Deep Learning Models for Accelerated Defect Property Predictions and Device Design of Cubic Semiconductor Crystals

In this work, we present an exploration of deep learning models for predicting defect properties in cubic phase semiconductors. The nature of impurity energy levels strongly influences the performance of semiconductors in a wide range of applications, such as solar cells, field effect transistors, and qubits for quantum computing. In this work, we employ two types of deep learning models, a crystal defect graph neural network and a chemical environment-encoded artificial neural network, to predict defect properties. The models are trained on a data set of charge-dependent defect formation energies obtained from density functional theory computations and descriptors based on elemental properties, defect local environment, and relevant semiconductor properties. We assess the models' performance and showcase their capability in optimizing semiconductor devices, particularly when used in tandem with compositionally constrained thermodynamics and technology computer-aided design models.