Abstract
Blending two ionic liquids (ILs) offers a promising route to discover superior solvents by combining complementary molecular features of each IL to achieve tailored thermophysical properties of mixed solvents. However, an exhaustive exploration of the vast combinatorial and compositional design space of IL-IL pairs through experiments or high-fidelity simulations is prohibitively time-consuming. We propose a quantum chemical and thermodynamic data-driven machine learning (ML) framework for predicting the miscibility of binary IL mixtures based on activity coefficients. Specifically, we train an artificial neural network (ANN) using COSMO-RS-generated mixture activity coefficients for over 4.9 million IL-IL compositions. The approach is further generalized across temperatures by using excess enthalpy. The trained ANN shows high accuracy as a threshold-based binary miscibility classifier. We further explore the similarity across IL mixtures to understand how molecular descriptors and mixture compositions correlate with miscibility. By combining ML-based structural analysis, we go beyond the 150 years old solubility principle of like dissolves like and deconvolute the roles of molecular identity in determining the miscibility of ILs. Finally, we perform miscibility experiments for several important binary IL-IL mixtures. The experimental observations are largely consistent with the miscibility trends predicted by the computational model.