Cross-Learning Transport and Thermodynamic Properties of Supercooled Ionic Liquids via Physics-Informed Multitask Neural Networks.

Predicting coupled transport and thermodynamic properties of supercooled ionic liquids (ILs) is critical for low-temperature energy applications, yet sparse data and the absence of low-temperature phenomenological relations make the prediction of these properties challenging. To address this challenge, we develop a physics-informed multitask neural network to predict viscosity (η), specific heat capacity (Cp), and density (ρ) of pure ILs, IL mixtures, and IL solutions. The η branch is constrained by the Vogel-Fulcher-Tammann (VFT) equation, while Cp and ρ are trained with lower-order physics-guided losses because their phenomenological relations in the supercooled regime are absent. We show that the multitask model captures nonlinear Cp and ρ behavior through cross-learning from η, unlike independently trained single-task models. By implementing a VFT-type correction to the interaction potential energy, we show quantitatively that the temperature dependence of Cp can be explained by the loss of configurational entropy due to kinetic limitations. The multitask model was then used to screen candidates with optimum η, Cp, and ρ for application as low-temperature heat transfer fluids, and the structural features governing these properties were identified by explainable AI algorithms and quantum chemical surface charge density calculations.