Tree-based ensemble learning algorithms for predicting the performance of organic solvent reverse osmosis membranes

Abstract

Organic solvent reverse osmosis (OSRO) offers a safer and less energy-intensive alternative to conventional thermal separations, but accurately predicting membrane performance remains challenging. In this work, six tree-based machine learning (ML) algorithms were developed to predict membrane performance. A dataset containing 160 experimental data points was collected. Pearson correlation analysis combined with Recursive Feature Elimination with Cross-Validation (RFECV) was employed to screen seven physically meaningful descriptors related to molecular interactions, operating conditions, and material properties. Six tree-based machine learning algorithms, including Decision Tree (DT), Random Forest (RF), Gradient Boosting Regressor (GBR), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LGBM), and Categorical Boosting (CatBoost), were trained to predict permeance and rejection. Model predictive stability was assessed using leave-one-out cross-validation (LOOCV) and 5-fold cross-validation. CatBoost achieved the best overall predictive performance, achieving test-set R² values of 0.954 for permeance and 0.926 for rejection, together with low AAE, RMSE, and MAPE values. Furthermore, SHapley Additive exPlanations (SHAP) analysis indicated that permeance is primarily governed by solvent molar volume and solvent mass fraction, whereas rejection is strongly influenced by the solubility parameter distances between solvent, solute, and membrane, as well as the solute molar volume. These findings provide a data-driven framework for predicting OSRO membrane performance and identifying influential factors for membrane design.