Explainable AI for permeate flux prediction in forward osmosis: SHAP interpretability and theoretical validation for enhanced predictive reliability

Abstract

This study presents an integrated machine learning (ML) and explainable artificial intelligence (XAI) framework for predicting permeate flux in forward osmosis (FO) desalination systems with high accuracy and enhancing the interpretability of model results. Advanced tree-based boosting models-extreme gradient boosting (XGBoost), light gradient-boosting machine (LGBM), and categorical boosting (CatBoost)-were employed, and they achieved notable prediction accuracy, with XGBoost displaying an R2 of 0.9716. Principal component analysis was used to preprocess the FO experimental data to reduce dimensionality, and k-means clustering was applied to uncover underlying patterns before model training. To ensure transparency in model interpretation, Shapley additive explanations (SHAP) was used, which identified osmotic pressure difference and water permeability as the most influential variables affecting permeate flux predictions. The ML models were compared with a physicochemical FO model to validate their reliability, and the ML predictions were found to align with the fundamental principles of the FO process. This dual-validation approach improves the accuracy and reliability of predictions while providing a transparent framework for FO process optimization. Integrating XAI techniques with rigorous theoretical validation offers a solid foundation for future advancements in FO desalination technology, addressing the need for interpretability and practical relevance in predictive models for real-world applications.