Physics‐Informed Machine Learning for Accurate and Physically Valid Prediction of Semiconductor Defect Levels

Abstract Accurately predicting semiconductor defect energy levels with strong physical validity remains a major bottleneck in accelerating materials discovery. Routine density function theory (DFT) calculations based on generalized gradient approximations such as Perdeco Burke Ernzerhof (PBE) are computationally cheap yet systematically underestimate defect levels, while hybrid functionals such as Heyd Scuseria Ernzerhof (HSE) improve accuracy but require prohibitive computational cost. A physics informed machine learning model is presented for predicting HSE calculated defect concentrations from PBE defect results and host material descriptors. A multilayer perceptron with one hidden layer achieved strong numerical performance with test R2 equal to 0.9615 and mean absolute error equal to 0.1635 eV, yet produced large physical violations with 73.33 percent of predictions falling outside the host bandgap. Adding a physics informed loss reduced violation rates to 59.44 percent with R2 above 0.94. A sigmoid scaled output multilayer perceptron achieved zero percent violations but with reduced flexibility and R2 equal to 0.5169. SHAP analysis confirmed PBE defect levels and host bandgaps as dominant features. Incorporating physical constraints greatly enhanced scientific validity and supports practical scalable high throughput semiconductor defect sorting.