Interpretable Machine-Learning Models for Predicting Shear Viscosity and Thermal Conductivity of Binary Fluid Mixtures

Accurate prediction of fluid transport properties is essential and challenging, especially for non-ideal mixtures and under extreme conditions. In this work, we systematically evaluate several machine learning (ML) regression methods─including linear, nonlinear, and decision-tree-based ensemble models─to predict shear viscosity and thermal conductivity for two representative fluid mixtures, N2/O2 and CO2/H2S, in the gas, liquid, and supercritical thermodynamic regions. These mixtures were selected to analyze the influence of composition and intermolecular interaction strength on the predictive accuracy because the two mixtures have different types of intermolecular interactions. Tree-based ensemble models consistently yielded superior predictive accuracy, with the Light Gradient Boosting Machine (LightGBM) performing particularly well for both properties, achieving low root-mean-square error (RMSE) values across all phases. SHAP (SHapley Additive exPlanations) analysis provided interpretability, showing that composition strongly influenced predictions in the gas phase, while pressure and temperature significantly impacted predictions in liquid and supercritical phases. Specifically, higher pressures and lower temperatures required larger corrections due to enhanced molecular interactions and structural complexities, especially for the CO2/H2S mixture. This study demonstrates the capability of interpretable ML models to accurately predict transport properties and offer physically meaningful insights into fluid mixture behaviors.