文献类型：期刊文献

英文题名：Machine Learning-enhanced QSPR Model for Predicting the Viscosity of Ionic Liquids

作者：Zheng, Yichao[1];Ji, Changzheng[1];Shi, Zhaochong[1];Li, Ruixue[2];Shi, Jialin[1];Peng, Changjun[1];Liu, Honglai[1]

机构：[1]East China Univ Sci & Technol, Sch Chem & Mol Engn, Shanghai 200237, Peoples R China;[2]Chengdu Univ Technol, Coll Energy, Coll Modern Shale Gas Ind, Chengdu 610059, Peoples R China

年份：2026

卷号：321

外文期刊名：CHEMICAL ENGINEERING SCIENCE

收录：;EI(收录号：20254819595387);WOS:【SCI-EXPANDED(收录号:WOS:001630505500001)】；

基金：This research was financially sponsored by the National Natural Science Foundation of China (No. 22078086) .

语种：英文

外文关键词：Viscosity; Ionic liquids; COSMO-SAC; Machine learning; QSPR

摘要：Viscosity is a critical transport property of ionic liquids (ILs), essential for their industrial applications. However, predicting IL viscosity remains challenging due to existing imbalanced datasets. Many prediction models rely on random data processing methods, which often masks true generalization performance and results in inflated statistical metrics on the test dataset. In this study, three Quantitative Structure-Property Relationship (QSPR) models were developed by partitioning the dataset based on IL types and compared with models trained on randomly partitioned data. The dataset includes 6,932 experimental viscosity values from 198 distinct ILs. All three models utilized descriptors derived from the conductor-like screening model for segment activity coefficient (COSMO-SAC) method. Models I and II combined regression algorithms with distinct empirical equations to form hybrid models, while Model III was developed based on an artificial neural network (ANN). Model II demonstrated superior generalization capability over Model I, achieving a coefficient of determination (R2) of 0.8298 on the test set. Model III demonstrated the highest prediction accuracy. However, its root mean square error (RMSE) was 0.5942, which is higher than Model II's 0.5647, indicating a greater presence of outliers. The findings of this study indicate that incorporating empirical formulas into regression algorithm-based predictive models can significantly improve both prediction accuracy and generalization performance. Furthermore, the comparative analysis of the two partitioning strategies revealed that random partitioning yielded better statistical metrics on the test set than IL-type partitioning. However, such performance reflects predictive ability only for previously encountered ILs and lacks extrapolative potential for new ILs.