Self supervised ensemble learning models for 2D semiconductors bandgap prediction

Effective representation and processing of crystal structures are crucial for the successful application of machine learning techniques in predicting material properties. In this research, we have developed a Self-Supervised Ensemble model (SSE) to predict the Heyd–Scuseria–Ernzerhof (HSE06) bandgap of two-dimensional semiconductors. Specifically, our model is capable of autonomously extract crystalline structural information of materials using an autoencoder, which is then used to refine the preliminary bandgap calculations obtained through the low-cost Perdew–Burke–Ernzerhof (PBE) method. By correcting the PBE bandgap through ensemble learning, we have successfully approximated the HSE06 bandgap with a root mean squared error (RMSE) of 0.372 eV and a mean absolute error (MAE) of 0.262 eV. Furthermore, we validated the model’s performance on three-dimensional materials with diverse and complex structures, demonstrating robust generalization capabilities. Our research lays a foundational framework for the screening and synthesis of two-dimensional semiconductors with significant potential applications.