Using Machine Learning to Explore Defect Configurations in Cd/Zn-Se/Te Compounds

To address the challenge of accurately and quickly predicting defect properties in semiconductors, we developed a workflow combining high-throughput density functional theory (DFT) computations with machine learning (ML) approaches, specifically active learning (AL) and crystal graph-based neural networks (GNNs). Using a comprehensive DFT dataset of Cd/Zn-S/Se/Te configurations, we trained GNN models on over 6200 crystal structures to predict crystal formation energy (CFE) for bulk and defective structures. Employing the Atomistic Line Graph Neural Network (ALIGNN) framework, we achieved high-accuracy CFE prediction models, which were then applied on 13,000 hypothetical defects in Cd<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">x</inf>Zn<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1−x</inf>Te and CdSe<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">x</inf>Te<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1−x</inf> compounds. New DFT calculations were strategically performed by targeting high-uncertainty configurations to enable iterative retraining of the models. Eventually, the optimized ALIGNN models enabled screening of low-energy defects, dopants, and defect complexes, including group V elements, Cu, Cl, and O in Cd/Zn-Se/Te compounds. ALIGNN models were also combined with optimization algorithms to identify low-energy defect structures in ZnTe (suitable back contact for CdTe), and predictions were verified by DFT. Our approach offers an effective surrogate for expensive DFT and generating libraries of defect structures for CdSeTe solar cells.