欢迎访问过程工程学报, 今天是

过程工程学报 ›› 2026, Vol. 26 ›› Issue (7): 703-714.DOI: 10.12034/j.issn.1009-606X.225261

• 研究论文 • 上一篇    下一篇

基于可解释性的机器学习模型预测一元胺溶液的黏度

朱颖1， 张子腾1*， 唐望2， 刘天雄2， 高红霞2， 梁志武2   

  1. 1. 中国能源建设集团江苏省电力设计院有限公司，发电工程公司，江苏 南京 211102 2. 湖南大学二氧化碳捕获与封存国际合作中心(iCCS)，化石能源低碳化高效利用湖南省重点实验室，湖南大学化学化工学院，湖南 长沙 410082
  • 收稿日期:2025-10-13 修回日期:2025-12-25 出版日期:2026-07-28 发布日期:2026-07-28
  • 通讯作者: 张子腾 zhangziteng@jspdi.com.cn
  • 基金资助:
    中国能源建设集团江苏省电力设计院有限公司科技项目

Prediction of viscosity of monoamine solutions based on interpretable machine learning models

Ying ZHU1,  Ziteng ZHANG1*,  Wang TANG2,  Tianxiong LIU2,  Hongxia GAO2,  Zhiwu LIANG2   

  1. 1. Power Generation Engineering Company, Jiangsu Electric Power Design Institute Co., Ltd., China Energy Engineering Group, Nanjing, Jiangsu 211102, China 2. Joint International Center for CO2 Capture and Storage (iCCS), Provincial Hunan Key Laboratory for Cost-effective Utilization of Fossil Fuel Aimed at Reducing CO2 Emissions, College of Chemistry and Chemical Engineering, Hunan University, Changsha, Hunan 410082, China
  • Received:2025-10-13 Revised:2025-12-25 Online:2026-07-28 Published:2026-07-28

RichHTML

PDF

摘要： 在胺法二氧化碳(CO2)捕集工艺中，胺基吸收剂的黏度直接影响气液传质效率，而传统实验测定方式存在耗时费力、成本高昂的局限。鉴于此，迅猛发展的机器学习技术为胺基吸收剂物性预测提供了高效工具，本研究聚焦一元胺溶液体系，构建了两种机器学习模型用于预测其黏度。首先利用RDKit工具从胺分子的SMILES标识符中提取分子描述符，继而构建随机森林(RF)与梯度提升决策树(GBDT)两种机器学习模型。通过特征工程进行系统的分子描述符筛选与超参数优化，模型性能评估结果显示，RF与GBDT模型在测试集上的决定系数(R2)分别为0.9939和0.9986，其中GBDT模型表现出更高的预测精度与稳健性。进一步采用SHAP (SHapley Additive exPlanations)方法对GBDT模型进行可解释性分析，结果显示胺质量分数越高，黏度越大；温度越高，黏度越低。从分子结构层面看，分子量(MolWt)对黏度整体呈正向贡献，电子状态描述符MinAbsEStateIndex对黏度呈负贡献，表明电子分布差异会调控分子间相互作用并影响黏度。本研究为胺法CO2捕集工艺中吸收剂的高效筛选提供了重要的理论依据。

关键词: 机器学习, 黏度, 预测, 二氧化碳, 构效关系

Abstract: Viscosity is a critical transport property in amine-based CO2 capture systems, as it directly influences solvent circulation, pumping energy consumption, and gas-liquid mass transfer rates within absorption-desorption columns. Despite its importance, viscosity determination in traditional research and industrial practice largely depends on experimental measurements, which are not only labor-intensive and time-consuming but also impractical for evaluating the rapidly expanding library of candidate absorbents. To address this methodological bottleneck, this study establishes a data-driven prediction framework for accurately estimating the viscosity of monoamine aqueous solutions using machine learning techniques. Molecular descriptors are systematically extracted from SMILES representations via RDKit and subjected to feature engineering to eliminate redundancy and enhance model generalizability. Two ensemble-learning algorithms, random forest (RF) and gradient boosting decision trees (GBDT), are developed and optimized using Bayesian hyperparameter tuning. Benchmarking on independent test datasets reveals that the RF and GBDT models achieve coefficient of determination (R2) of 0.9939 and 0.9986, respectively, with GBDT exhibiting higher robustness toward extrapolative predictions under varying temperature and concentration conditions. To elucidate model decision mechanisms, SHAP analysis is employed, revealing that viscosity increases monotonically with amine mass fraction while decreasing with temperature, aligning with thermodynamic expectations. From the perspective of molecular structure, molecular weight (MolWt) makes a positive contribution to overall viscosity, while the electronic state descriptor MinAbsEStateIndex makes a negative contribution to viscosity, indicating that differences in electron distribution can regulate intermolecular interactions and affect viscosity. Overall, this work bridges computational chemistry and process engineering, offering a scalable pathway toward accelerated solvent discovery and optimization for next-generation carbon capture technologies.

Key words: machine learning, viscosity, prediction, carbon dioxide, structure-property relationship