A hybrid machine learning approach with Jouyban–Acree mixing rule for predicting dielectric constants of binary liquid mixtures

Accurate prediction of the dielectric constant (ε) of binary liquid mixtures remains challenging due to complex non-ideal intermolecular interactions and limited availability of experimental data. In this work, both direct machine learning prediction and a hybrid physics-informed approach combining machine learning with the Jouyban–Acree mixing rule were investigated for predicting dielectric constants of binary liquid mixtures. Multiple Linear Regression, CatBoost, and Extra Trees models were employed either to predict ε directly or to estimate the effective non-ideal interaction parameter K used in the simplified Jouyban–Acree equation. The models were developed using a dataset comprising 163 binary liquid systems over a range of temperatures and evaluated under two validation protocols involving unseen mixtures and previously unseen components. The results demonstrate that direct machine learning models provide satisfactory predictive performance when dielectric constants of the pure components are unavailable and no additional physicochemical information can be incorporated into the prediction procedure. However, when dielectric constants of the pure liquids are known, the hybrid approach shows significantly improved generalization capability. Under the more rigorous component split validation, direct machine learning models exhibited a substantial decrease in performance, with average absolute relative deviations (AARD) reaching up to 32%. In contrast, the hybrid models maintained high predictive accuracy, achieving AARD values near 8.5%. The proposed approach demonstrates that integration of classical thermodynamic mixing models with machine learning provides an effective and reliable strategy for predicting dielectric constants of complex binary liquid mixtures.