Machine Learning Integrating Surface Features and Crystal Similarity for Exploring 2D Materials

The property prediction of two-dimensional (2D) materials has been constrained to specific systems or the highly self-similar Computational 2D Materials Database (C2DB) due to limited data availability and inadequate feature extraction. These challenges impede extrapolated predictions and large-scale exploitation across diverse databases. We developed a model named crystal surface and cluster network (CCSN) that integrates crystal surface and cluster features while leveraging transfer learning from bulk material databases to enhance model performance. Compared to the widely adopted CGCNN model, our approach achieved a 30% reduction in mean absolute error for bandgap prediction on the Materials Cloud 2D crystals database, characterized by low self-similarity and limited data volume. To enable extrapolated predictions, we developed a prediction workflow based on crystal similarity, which selects the most similar database for model training and determines the necessity of applying transfer learning. This method was applied to predict bandgaps of 8,218 crystals without bandgap values in the C2DB database and subsequently validated through our DFT calculations. The proposed prediction workflow, based on the CCSN model, would enable the discovery of 2D materials through rapid property prediction and can be generalized to improve any prediction models dealing with scarce or biased data sets.