Bandgap Prediction in Low-Dimensional Semiconductors via Transfer Learning: Augmenting Small Datasets with Pretrained CGCNN Models

Abstract The bandgap regulation of low-dimensional semiconductor materials is the core challenge of optoelectronic device design, and the computational cost and experimental cycle of traditional methods seriously restrict the R&D efficiency. Aiming at the problems of overfitting and cross-dimensional generalization in existing deep learning models under small sample conditions, this study proposes a collaborative solution that integrates model optimization and transfer learning. By systematically introducing the regularization strategy into the crystal graph convolutional neural network (CGCNN), the bandgap prediction error of the test set is successfully reduced by 25.7%, while maintaining the high-precision feature extraction ability of the training set (R 2 =0.998), which significantly improves the stability of the model. Based on the optimized network architecture, the transfer learning framework of "bulk material pre-training - low-dimensional material fine-tuning" is further constructed, and the knowledge transfer is completed by using the 20377 individual phase data of the Materials Project, and finally the average absolute error of bandgap prediction is 0.15 eV on the two-dimensional semiconductor test set, which is 35% higher than the accuracy of the traditional training method, and the prediction error of the formed energy and total energy is reduced by 22%-30% at the same time. This study not only confirms the adaptability of deep learning in cross-dimensional material computing but also provides a solution with both computational efficiency and generalization capabilities for accurate prediction in small-sample scenarios.