| 引用本文: | 徐达梁,徐杭镔,靳心瑶,刘超,费兆轩,姚杰,张子峰,李圭白,梁恒.基于机器学习的纳滤膜预测筛选模型构建与评估[J].哈尔滨工业大学学报,2024,56(6):8.DOI:10.11918/202402004 |
| XU Daliang,XU Hangbin,JIN Xinyao,LIU Chao,FEI Zhaoxuan,YAO Jie,ZHANG Zifeng,LI Guibai,LIANG Heng.Construction and evaluation of prediction model for nanofiltration membranes based on machine learning[J].Journal of Harbin Institute of Technology,2024,56(6):8.DOI:10.11918/202402004 |
|
| |
|
|
| 本文已被：浏览 1446次 下载 1888次 |
码上扫一扫！ |
|
|
| 基于机器学习的纳滤膜预测筛选模型构建与评估 |
|
徐达梁,徐杭镔,靳心瑶,刘超,费兆轩,姚杰,张子峰,李圭白,梁恒
|
|
(城市水资源与水环境国家重点实验室(哈尔滨工业大学),哈尔滨 150090)
|
|
| 摘要: |
| 纳滤净水技术是应对水资源危机和水质安全保障的核心技术之一。然而,纳滤膜性能长期受渗透性与选择性制约,亟需开发高性能纳滤膜。纳滤膜制备过程涉及水相单体质量分数、水相添加剂质量分数、油相单体质量分数、聚合时间等因素,传统的试误实验法需消耗大量的人力、物力与财力。依据纳滤膜制备参数,构建基于机器学习的纳滤膜预测筛选模型。结果表明,XGBoost机器学习模型可有效预测纳滤膜纯水通量与截留性能,对纯水通量和截留性能的R2评价指标分别为0.84和0.90。采用SHAP值法对XGBoost机器学习模型中的输入参数进行量化分析,发现水相单体质量分数与基膜类型对纯水通量有最高的绝对平均SHAP值,分别为2.77与2.59,而面向纳滤膜截留性能的关键参数绝对平均SHAP值相对接近。单体子结构特征分析结果显示,亲水性子结构特征与支链型子结构特征有助于提升纳滤膜纯水通量,胺基则促进纳滤膜的截留性能。构建的纳滤膜预测筛选模型有助于关键参数的识别与优化,为纳滤膜的开发提供理论与技术指导。 |
| 关键词: 纳滤膜 机器学习 性能预测 XGBoost模型 |
| DOI：10.11918/202402004 |
| 分类号:TU991 |
| 文献标识码:A |
| 基金项目:国家自然科学基金(52300083)；哈尔滨工业大学城市水资源与水环境国家重点实验室开放基金(QA202323) |
|
| Construction and evaluation of prediction model for nanofiltration membranes based on machine learning |
|
XU Daliang,XU Hangbin,JIN Xinyao,LIU Chao,FEI Zhaoxuan,YAO Jie,ZHANG Zifeng,LI Guibai,LIANG Heng
|
|
(State Key Laboratory of Urban Water Resources and Water Environment (Harbin Institute of Technology), Harbin 150090, China)
|
| Abstract: |
| Nanofiltration is one of the core technologies to address water crises and ensure water quality safety and security. However, the performance of nanofiltration membranes has long been limited by the permeance/selectivity trade-off. Thus, developing high-performance nanofiltration membranes is an urgent need. Nanofiltration membrane fabrication involves factors such as aqueous phase monomer concentration, aqueous phase additive concentration, oil phase monomer concentration, polymerization time, etc. Traditional trial-and-error experimental method consumes substantial manpower, material and financial resources. In this study, based on the fabrication parameters of nanofiltration membranes, we constructed a machine learning-based predictive screening model for nanofiltration membranes. The results show that the XGBoost machine learning model effectively predict the permeance and rejection performance of nanofiltration membranes, with R2 evaluation scores of 0.84 and 0.90, respectively, for the permeance and rejection performance. The quantitative analysis of the input parameters in the XGBoost machine learning model using the SHAP value method reveals that the aqueous-phase monomer concentration and the substrate type had high average SHAP values of 2.77 and 2.59 for permeance. The average SHAP values of the key parameters oriented towards the rejection performance were relatively close. The results of the monomer substructure characterization show that the hydrophilic and branched substructure features contribute to the permeance, while the amine group promote the rejection performance. The nanofiltration membrane prediction and screening model can identify and optimize the key parameters, providing theoretical and technical support for developing nanofiltration membranes. |
| Key words: nanofiltration membranes machine learning performance prediction XGBoost model |
|
|
|
|
