Scholarlink

<p>Nanofiltration (NF) membranes, especially polyamide-based thin-film composite (TFC) structures, are essential in water treatment, providing enhanced selectivity and permeability at minimal operating pressures. Nevertheless, performance optimization is impeded by the intricate, varied composition of these membranes and the lack of cohesive, high-quality samples. This study introduces a machine learning system developed using laboratory-scale experimental data from the Open Membrane Database (OMD) to predict two critical performance metrics such as water permeance (A) and salt permeance (B). We developed two regression models utilizing gradient boosting, XGBoost and Decision tree algorithms, employing designed features derived from membrane construction methods, hydraulic and osmotic pressures, feed salinity, concentration polarization, and an estimated active-layer thickness. The models attained robust prediction accuracy, with average R<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(^{2}\)</EquationSource> </InlineEquation> values above 0.84 for A and 0.85 for B during cross-validation. To improve interpretability, SHAP (SHapley Additive exPlanations) analysis was employed to identify primary predictors, indicating that concentration polarization and osmotic pressure were the most significant features. Water permeance estimates corresponded more closely with classical transport theory, but salt permeance displayed greater variance, indicating the complexity of the underlying data and physicochemical heterogeneity. Our findings highlight the effectiveness of ensemble machine learning methods in elucidating nonlinear structure–property correlations and affirm the feasibility of data-driven predictions for membrane performance evaluation. Despite the study’s limitations due to its dependence on literature-derived data with few experimental descriptors, it presents a reproducible and interpretable framework for rational evaluation of NF membranes. Subsequent research ought to incorporate more detailed structural and electrostatic data, with hybrid modeling methodologies, to connect empirical findings with first-principles comprehension.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Predicting water and salt permeance of polymeric desalination membranes using interpretable machine learning

  • 期刊论文

  • 发表日期：

    2026年7月30日

查看全文
  • A. Chahbi,
  • S. Rochd,
  • Y. Ezaier,
  • H. Dari,
  • M. Korchi,
  • R. Moultif,
  • A. Hader,
  • I. Tarras,
  • Y. Boughaleb

摘要

Nanofiltration (NF) membranes, especially polyamide-based thin-film composite (TFC) structures, are essential in water treatment, providing enhanced selectivity and permeability at minimal operating pressures. Nevertheless, performance optimization is impeded by the intricate, varied composition of these membranes and the lack of cohesive, high-quality samples. This study introduces a machine learning system developed using laboratory-scale experimental data from the Open Membrane Database (OMD) to predict two critical performance metrics such as water permeance (A) and salt permeance (B). We developed two regression models utilizing gradient boosting, XGBoost and Decision tree algorithms, employing designed features derived from membrane construction methods, hydraulic and osmotic pressures, feed salinity, concentration polarization, and an estimated active-layer thickness. The models attained robust prediction accuracy, with average R \(^{2}\) values above 0.84 for A and 0.85 for B during cross-validation. To improve interpretability, SHAP (SHapley Additive exPlanations) analysis was employed to identify primary predictors, indicating that concentration polarization and osmotic pressure were the most significant features. Water permeance estimates corresponded more closely with classical transport theory, but salt permeance displayed greater variance, indicating the complexity of the underlying data and physicochemical heterogeneity. Our findings highlight the effectiveness of ensemble machine learning methods in elucidating nonlinear structure–property correlations and affirm the feasibility of data-driven predictions for membrane performance evaluation. Despite the study’s limitations due to its dependence on literature-derived data with few experimental descriptors, it presents a reproducible and interpretable framework for rational evaluation of NF membranes. Subsequent research ought to incorporate more detailed structural and electrostatic data, with hybrid modeling methodologies, to connect empirical findings with first-principles comprehension.