Modeling crystal defects using defect-informed neural networks

Machine learning has revolutionized the study of crystalline materials forenabling rapid predictions and discovery. However, most AI-for-Materialsresearch to date has focused on ideal crystals, whereas real-world materialsinevitably contain defects that play a critical role in modern functionaltechnologies. The defects and dopants break geometric symmetry andincrease interaction complexity, posing particular challenges for traditionalML models. Addressing these challenges requires models that areable to capture sparse defect-driven effects in crystals while maintainingadaptability and precision. Here, we introduce Defect-Informed EquivariantGraph Neural Network (DefiNet), a model specifically designedto accurately capture defect-related interactions and geometric configurationsin point-defect structures. Trained on 14,866 defect structures,DefiNet achieves highly accurate structural predictions in a single step,avoiding the time-consuming iterative processes in modern ML relaxationmodels and possible error accumulation from iteration. We furthervalidates DefiNet’s accuracy by using density functional theory (DFT) relaxation on DefiNet-predicted structures. For most defect structures,regardless of defect complexity or system size, only 3 ionic steps arerequired to reach the DFT-level ground state. Finally, comparisonswith scanning transmission electron microscopy (STEM) images confirmDefiNet’s scalability and extrapolation beyond point defects, positioningit as a groundbreaking tool for defect-focused materials research.