An explainable ionic liquid property model by coupling ionic fragment contribution with graph neural network

Abstract Ionic liquids (ILs) have potential applications in various fields due to their unique advantages. Recently, machine learning (ML) presented excellent predicting ability to the property ILs. But the lack of interpretability poses a significant challenge in effectively guiding their design. In this study, we developed an interpretable Attentive Ionic Fragment Contribution (AIFC) model for IL property prediction by combining Ionic Fragment Contribution (IFC) with graph neural networks (GNNs). The AIFC first predefined 99 ionic fragments (IFs), then the IF graph embedding was encoded by training GNN. Furthermore, integrated with an attention‐based method, the proposed model not only shows the better prediction abilities but also gives the sequence of IF importance in all IFs while predicting the target properties, such as CO₂ solubility, viscosity, density, thermal decomposition temperature and melting point. Therefore, the proposed method will be very helpful for the design of functional ionic liquids.