Accurate and rapid predictions with explainable graph neural networks for small high-fidelity bandgap datasets

Accurate and rapid bandgap prediction is a fundamental task in materials science. We propose graph neural networks with transfer learning to overcome the scarcity of training data for high-fidelity bandgap predictions. We also add a perturbation-based component to our framework to improve explainability. The experimental results show that a framework consisting of graph-level pre-training and standard fine-tuning achieves superior performance on all high-fidelity bandgap prediction tasks and training-set sizes. Furthermore, the framework provides a reliable explanation that considers node features together with the graph structure. We also used the framework to screen 105 potential photovoltaic absorber materials.