Prediction of permeate flux of ceramic membrane-based direct contact membrane distillation process utilizing machine learning approaches

Ceramic membrane-based direct contact membrane distillation (DCMD) has emerged as a highly promising technology for seawater desalination, owing to the exceptional chemical stability, mechanical robustness, thermal resistance, antifouling properties, and longevity of ceramic membranes. However, the intricate interplay of multiple operational parameters in DCMD systems complicates accurate permeate flux prediction, presenting a persistent challenge in process optimization. To address this, we systematically evaluate the predictive capabilities of machine learning (ML) techniques and compare three non-ensemble and six ensemble ML models. Leveraging a curated dataset of 357 experimental data points from published studies, each comprising nine input features and one output feature exclusively for ceramic membranes-based DCMD, we created an ML framework for flux prediction. Performance assessments revealed the superior predictive accuracy of ensemble models over non-ensemble models, with the Extremely Randomized Trees (ERT) model achieving best performance (test set metrics: R² = 0.905, MAE = 2.614, RMSE = 4.588). The conducted SHapley Additive exPlanations (SHAP) analysis identified permeate-side temperature and feed flow rate as the most influential factors governing permeate flux, while Partial Dependence Plots (PDPs) elucidated the nonlinear relationships and interactions among key input parameters. Furthermore, particle swarm optimization (PSO) was employed to fine-tune the top-performing model (ERT), identifying the optimal values for the five most critical features to maximize membrane distillation efficiency. Overall, this study demonstrates the potential of ML in advancing membrane science, while the integration of interpretability analysis with flux prediction allows to accelerate the development of more efficient ceramic membrane-based DCMD for seawater desalination.