Coupling machine learning and group contribution method to predict density and viscosity of ionic liquid-inorganic solvent-water ternary mixtures

Accurate prediction of the physical properties of mixed systems is essential for efficient industrial design and process optimization. This study explores the use of a hybrid approach that combines the group contribution (GC) method with advanced machine learning (ML) algorithms to predict the density and viscosity of ternary mixtures comprising ionic liquids (ILs), inorganic solvents (ICs), and water. The machine learning algorithms employed include artificial neural networks (ANN), XGBoost, and LightGBM. A comprehensive dataset was compiled, encompassing 5738 density data points for 25 classes of ILs and 34 classes of ICs, and 1551 viscosity data points for 12 classes of ILs and 16 classes of ICs. The results indicate that all three ML-GC models significantly outperformed traditional GC models without ML algorithms. Notably, the XGBoost-GC model exhibited the highest prediction accuracy for density, with an R2 of 0.9983, while the ANN-GC model with 7 neurons in the hidden layer delivered the best predictions for viscosity, with an R2 of 0.9967. Additionally, the Shapley Additive Interpretation (SHAP) analysis reveals that the molar fractions of water and IC primarily influence density, whereas the presence of the IC anion NO3 and temperature significantly impact viscosity. Moreover, a practical example is provided to illustrate the real-world applicability of the ML-GC models developed in this work, highlighting their potential to enhance industrial processes through precise property predictions.

混合体系物理性质的精准预测，对于高效工业设计与流程优化至关重要。本研究探索了一种将基团贡献（group contribution, GC）法与先进机器学习（machine learning, ML）算法相结合的混合方案，用于预测由离子液体（ionic liquids, ILs）、无机溶剂（inorganic solvents, ICs）与水构成的三元混合物的密度与黏度。所采用的机器学习算法包括人工神经网络（artificial neural networks, ANN）、XGBoost以及LightGBM。研究人员构建了一套综合数据集，涵盖25类离子液体与34类无机溶剂共5738条密度数据，以及12类离子液体与16类无机溶剂共1551条黏度数据。结果表明，三款基于机器学习-基团贡献（ML-GC）的模型均显著优于未结合机器学习算法的传统基团贡献模型。其中，XGBoost-GC模型在密度预测任务中表现出最高精度，其决定系数（R²）达0.9983；而隐藏层含7个神经元的人工神经网络-基团贡献（ANN-GC）模型则在黏度预测任务中效果最佳，决定系数（R²）为0.9967。此外，夏普利可加性解释（Shapley Additive Interpretation, SHAP）分析显示，水与无机溶剂的摩尔分数是影响密度的主要因素，而无机溶剂阴离子NO₃⁻的存在与温度则对黏度具有显著影响。本研究还提供了一则实用案例，用以说明本文所构建的ML-GC模型的实际应用价值，凸显了其通过精准的物性预测助力工业流程优化的应用潜力。