Fig. 1. Machine learning workflow for the discovery of ILs with high conductivity (σ) and wide electrochemical window (ECW).

The permutation of 74 cations and 30 anions forms an IL pool containing 2220 unique ILs. Three open-source platforms, including RDKit, Psi4, and PyG are employed to generate the molecular descriptors for the raw dataset. Unsupervised learning contains boxplots, pair plots, and hierarchical clustering, which are essential analytical methods for investigating the structure and correlations of variables in the dataset. Supervised learning leverages both regression and classification based on SVM, RF, XGBoosting and GCNN. The IL pool will initially be classified as a solid or liquid group. Then the ILs with liquid phase at RT will be further classified based on the σ ≥ 5 mS cm⁻¹ or not. Meanwhile, we also employ regression to predict the absolute σ values of the ILs for reference. Finally, ECW > 4 V is the final screening criterion for the final recommendation list of potential ILs.