Tailoring Semiconductor Defect Properties using Multi-fidelity Graph Neural Networks and Active Learning

To address the challenge of predicting defect properties in semiconductors and designing novel materials with tailored defect behavior, we developed a workflow integrating active learning (AL) and crystal graph-based neural networks (GNNs), trained on high-throughput density functional theory (DFT) data. Our dataset includes a range of binary compounds like CdTe, ZnTe, anion- or cation-site alloys, and various defects, simulated across charge states. We rigorously trained GNN models on >6200 structures, accurately predicting crystal formation energy (CFE) for bulk and defective structures. Later, we ran hybrid HSE+SOC calculations, combining them with PBE data to train a multi-fidelity Atomistic Line Graph Neural Network (ALIGNN) model, achieving <20 meV/atom RMSE for CFE in neutral states. The ALIGNN model, coupled with perturbation-based geometry optimization, enables efficient defect formation energy predictions as a function of Fermi level and chemical potentials. This allows high-throughput screening of defects and dopants, generating libraries of low-energy defect configurations in (Cd,Zn)-(Te,Se,S) for solar cell applications. AL further refines model predictions by iterating new DFT calculations.