Prediction of Frequency-Dependent Dielectric Function for Solid Materials: A Multi-Fidelity Machine Learning Approach with Physical Insights

The quest for innovative solid materials, ideally suited for integration into next-generation optoelectronic devices, demands the development of accurate and efficient tools for predicting frequency-dependent optoelectronic properties. Data-driven predictive models can efficiently screen a broad spectrum of materials, alleviating the scalability constraints of first-principles methods like Density Functional Theory (DFT) while retaining their accuracy. In this study, we employ deep graph neural networks (GNNs) to forecast the frequency-dependent complex dielectric function of solid materials. We implement regularizations in the loss function to refine the error distribution and enhance GNN precision regarding physical features such as band gaps, resonant peak locations, and function smoothness. Our dataset encompasses 17,805 crystal structures generated using the OptB88vdW DFT functional and 7,358 refined using the more accurate Tran-Blaha modified Becke Johnson (MBJ) potential. We employ transfer learning to fine-tune our GNNs, enabling us to compute absorption coefficients with MBJ-level accuracy. The synergy of advanced ML techniques, including GNNs, physically informed learning, and transfer learning across multi-fidelity datasets, yields precise predictions of the frequency-dependent dielectric function, enhancing generalizability beyond observational domain and advancing optoelectronic device development.