Integrating transfer learning and uncertainty quantification for data-efficient materials discovery

Abstract A major challenge in machine learning for materials lies in the simultaneous use of large, lower-accuracy datasets and small, high-accuracy datasets for data-efficient screening. Here, we show that the combination of transfer learning and uncertainty quantification enables highly efficient multi-fidelity materials screening. Using band gap prediction as a representative application, a model pre-trained on generalized gradient approximation density functional theory (GGA-DFT) data was fine-tuned on experimental and higher-level DFT datasets. The transfer learning model reduced the MAE by approximately 39% on the experimental dataset (0.833 ± 0.104 eV–0.502 ± 0.063 eV) and by 54% on the strongly constrained and appropriately normed (SCAN) dataset (0.716 ± 0.093 eV–0.331 ± 0.031 eV) relative to the respective target models trained from scratch, and matched or exceeded the best accuracy reported in the literature without requiring feature engineering. Ensemble-based uncertainty estimates guide the screening process and minimize the need for new calculations. In a photovoltaic case study, only six additional meta-GGA calculations are required to identify multiple DFT-validated candidate materials with band gaps near the target value. More broadly, this work establishes a general, data-efficient strategy applicable for accelerating materials discovery by leveraging heterogeneous datasets.