Sharpness-aware minimization with physics-informed regularizations for predicting semiconductor material properties in molecular dynamics

Abstract

In recent years, the growing adoption of artificial intelligence across diverse scientific fields has significantly increased demand for advanced semiconductor chips, necessitating innovations in semiconductor material design. Accurate prediction of semiconductor material properties is essential for improving chip performance, as these properties directly affect electrical, thermal, and mechanical characteristics. Traditionally, density functional theory has been the gold standard for atomic-scale simulations in material property prediction; however, its high computational cost limits scalability. Molecular dynamics simulations provide a scalable alternative by leveraging the power of machine learning force fields (MLFFs); however, semiconductor systems present unique challenges due to non-equilibrium dynamics, surface defects, and impurities. These factors often result in out-of-distribution (OOD) atomic configurations, which can significantly degrade model performance. To address this challenge, we propose Physics-Informed Sharpness-Aware Minimization (PI-SAM), a novel framework designed to enhance the prediction of semiconductor material properties across diverse datasets and challenging OOD scenarios. Specifically, PI-SAM leverages sharpness-aware minimization to achieve flatter loss minima, improving the model’s generalization. Additionally, it incorporates physics-informed regularizations to enforce energy-force consistency and account for potential energy surface curvature, ensuring alignment with the underlying physical principles governing semiconductor behavior. Experimental results demonstrate that our PI-SAM outperforms competing methods, especially on OOD datasets, underscoring its effectiveness in improving generalization.