Predicting Li Transport Activation Energy with Graph Convolutional Neural Network

Exploring activation energy in ionic transport is one of the critical pathways to discovering high-performance inorganic solid electrolytes (ISEs). Although traditional machine learning methods have achieved relatively accurate activation energy predictions, they suffer from issues such as complex descriptor construction and poor generalization. Graph neural network (GNN) has gained widespread usage in accurate material property prediction due to their ability to uncover structure-property relationships latent in materials data in an end-to-end way. However, current graph representation methods and corresponding GNN models have not been widely applied to the ion transport properties of materials. Here, we introduce the interstitial network graph representation method, and design a GNN model to predict activation energy in Li-containing compounds. As a result, the dynamic ion migration process is characterized, enabling the GNN to automatically capture the inherent mechanisms of ion transport. Performance tests demonstrate that interstitial network representation method achieves high prediction accuracy, with a 10% improvement in prediction accuracy compared with the crystal structure representation method. The developed model can be used to screen and design ISEs and in general providing new ideas for applying machine learning in materials science.