Band gap prediction of perovskite materials based on transfer learning

The band gap is a key physical quantity in material design. First-principles calculations based on density functional theory can approximately predict the band gap, which often require significant computational resources and time. Deep learning models have the advantages of good fitting ability and automatic feature extraction from the data, and are gradually being applied to predict the band gap. In this paper, aiming at the problem of quickly obtaining the band gap value of perovskite materials, a feature fusion neural network model named CGCrabNet is established, and the transfer learning strategy is used to predict the band gap of perovskite materials. CGCrabNet extracts features from both chemical equation and crystal structure of materials, and fits the mapping between features and band gaps. It is an end-to-end neural network model. Based on the pre-training data obtained from the Open Quantum Materials Database (OQMD dataset), the CGCrabNet parameters can be fine-tuned by using only 175 perovskite material data to improve the robustness of the model.The numerical experimental results show that the prediction error of the CGCrabNet model for band gap prediciton based on the OQMD dataset is 0.014eV, which is lower than that obtained from the prediction based on Compositionally restricted attention-based network (CrabNet). The mean absolute error of the model developed in this paper for the prediction of perovskite materials is 0.374eV, which is lower 0.304eV, 0.441eV and 0.194eV than that obtained from random forest regression, support vector machine regression and gradient boosting regression, respectively. The mean absolute error of the test set of CGCrabNet trained only using perovskite data is 0.536 eV, and the mean absolute error of the pre-trained CGCrabNet has decreased by 0.162 eV, which indicates that the transfer learning strategy has significant role in improving the prediction accuracy of small data sets (perovskite material data sets). The difference between the predicted band gap of some perovskite materials such as SrHfO3and RbPaO3 by the model and the band gap calculated by first-principles is less than 0.05eV, which indicates that the CGCrabNet can quickly and accurately predict the properties of new materials and accelerate the development process of new materials.