基于可解释性机器学习的纳滤膜去除有机微污染物研究

Prediction of organic micropollutant rejection by nanofiltration membrane based on interpretable machine learning

纳滤膜对微污染物去除效率受膜特性、微污染物性质和实验条件等因素影响,优化这些影响因素对纳滤工艺的成功应用至关重要,然而实验优化过程不能同时兼顾多因素对去除效能的交互影响.为此,基于线性、非线性和集成学习算法,开发了4种纳滤膜对有机微污染物去除效率的预测模型,并验证了模型的预测性能和可行性.在拟合程度、稳健性和外部预测能力等方面对开发的模型进行了对比分析研究,结果表明,用集成学习算法开发的XGBoost模型能够准确识别影响膜分离过程中的关键因素,在预测纳滤膜对微污染物的去除效率方面表现出强大的潜力(R_(adj)^(2)=0.977,Q_(ext)^(2)=0.877,Q_(LOO)^(2)=0.875).此外,利用SHAP解释方法定量解析了各驱动因素对纳滤膜去除微污染物效率的贡献,证明纳滤膜的截留分子量、膜表面接触角和污染物的尺寸是微污染物去除过程中重要的影响因素.首次将模型解释方法应用于特征变量的选择过程,使集成算法(XGBoost)模型的性能得到进一步优化.所开发的机器学习模型,有利于优化实验设计,为复杂的膜分离系统的建模与应用提供科学依据和技术支持.

The removal efficiency of nanofiltration membrane for micropollutants is affected by diverse factors including membrane characteristics,micropollutant properties and experimental conditions.Optimizing influencing factors is crucial to the successful application of nanofiltration.However,the experimental optimization process cannot simultaneously take into account the interactive effects of multiple factors on the removal efficiency.Therefore,based on linear,nonlinear and ensemble learning algorithms,four prediction models were developed to predict the removal efficiency of organic micropollutants by nanofiltration membranes,and the prediction performance and feasibility of the models were verified.The results showed that the ensemble XGBoost model can accurately identify the influencing factors in the membrane separation process and demonstrated powerful potential for predicting removal efficiency of organic contaminants by nanofiltration membrane(R^(2)_(adj)=0.977,Q^(2)_(ext)=0.877,Q^(2)_(LOO)=0.875).Moreover,based on SHAP interpretation method,the contribution of each driving factor to the removal efficiency of micropollutant by nanofiltration membrane was quantitatively analyzed,and molecular weight cut-off of nanofiltration membrane,surface contact angle of nanofiltration membrane and compound size of micropollutant were demonstrated to be the most important influencing factors in micropollutant removal by the nanofiltration.The model interpretation method is first applied to the selection of feature variables to further optimize the performance of XGBoost model.The proposed machine learning method can facilitate the optimization of experimental design,and provide scientific basis and technical support for the modeling and application of the complex membrane separation systems.