Advanced Theory and Simulations Моделирование и симуляция
Prediction of Defect‐Dependent Electronic Properties in 2D TMDC Materials Using Multitask Graph Neural Networks
Аннотация
ABSTRACT Defects strongly influence the electronic and energetic properties of two‐dimensional transition‐metal dichalcogenides (TMDCs), but density functional theory calculations become computationally demanding when exploring large defect configuration spaces. We develop a multitask graph neural network based on Principal Neighborhood Aggregation (PNAConv) to predict the Fermi level, bandgap energy, and formation energy of defective and . The study uses 5933 configurations per material from the 2D Materials Defect database, including vacancy, substitutional, and mixed vacancy–substitutional defects. Four graph representations are evaluated by combining DFT‐relaxed or unrelaxed structures with Simple or Enriched features. Paired‐bootstrap analysis shows that structural relaxation provides more consistent improvements than feature enrichment, particularly for Fermi‐level and bandgap‐energy prediction. PNAConv is further benchmarked against SchNet, Random Forest, and CatBoost using the same fixed data partition. PNAConv achieves the lowest mean MAE for Fermi‐level and bandgap‐energy prediction for both materials, whereas Random Forest achieves the lowest mean formation‐energy MAE. A target‐aligned comparison with prior 2DMD results further places the PNAConv predictions in the same low‐error range as previously reported sparse defect‐centered models.