Existing methodologies for materials screening, such as density functional theory (DFT) and experimental measurements, often face significant limitations in terms of computational cost, data availability, and predictive accuracy for certain properties. In particular, machine learning (ML) models, while invaluable for accelerating materials discovery, are frequently constrained by limited and unevenly distributed training data, especially for properties that are challenging to predict with DFT—such as optical and excited-state properties. These limitations hinder the broader applicability of ML-driven approaches in materials design. Here, we address these challenges by developing and applying a transfer-learning framework based on deep learning, which enables more accurate predictions even in data-scarce regimes.
Transfer learning enhances prediction performance on small datasets by leveraging knowledge from large datasets. This study demonstrates that using transfer learning, based on an element-centered fingerprint (ECFP) representation in deep learning, significantly improves band gap predictions. A model pre-trained on the Open Quantum Materials Database (OQMD) was fine-tuned on the Materials Project and an experimental dataset, achieving lower prediction errors. Specifically, we currently achieve a 45% reduction in error as compared to models trained from scratch. These results highlight the potential of transfer learning in materials informatics, enabling accurate predictions while efficiently utilizing available data.
