A New Model Based on Graph Neural Network Feedback for Crystal Property Prediction

The prediction of crystal properties is crucial in crystal design. Currently, most methods employ graph neural networks to model crystal structures and have achieved satisfactory prediction accuracy. However, accurately simulating the complex relationships between atoms and molecules in crystal structures remains challenging. The prediction performance of crystal graph networks also depends on the effective extraction of crystal features, typically requiring a close correlation between the properties to be predicted and the extracted features. In this study, a model retraining method is utilized, where the loss function in the retraining phase is constructed using feedback from a pre-trained crystal graph network model to adjust model parameters and enhance the accuracy of predicting crystal properties. This method enhances the predictive performance of multiple crystal graph neural networks and accelerates the development of high-performance crystals without the need to add various features or incorporate additional neural network layers. The research findings demonstrate the significant effectiveness of the proposed method in enhancing the prediction performance of four models: CGCNN, OGCNN, GATGNN, and ALIGNN. Optimization metrics support this claim.