Modeling of flat sheet-based direct contact membrane distillation (DCMD) for the robust prediction of permeate flux using single and ensemble interpretable machine learning

Direct contact membrane distillation (DCMD) has emerged as a promising technology for water desalination and treatment while offering advantages such as high energy efficiency and the adequacy to treat high-salinity feeds. However, due to complex interactions among multiple process parameters, the accurate prediction of permeate flux, a critical performance indicator of the DCMD process, remains a challenge that numerous traditional predictive techniques fall short of achieving. Thus, for the first time, this study investigated the effectiveness of various machine learning techniques, including four single and four ensemble models, to predict the permeate flux in the DCMD process. Utilizing a comprehensive dataset of 475 experimental points compiled from the literature, these models were built considering ten key process parameters, with the membrane material (PTFE or PVDF) being the sole categorical input. The performance evaluation demonstrated that the advanced ensemble models consistently outperformed the relatively simpler single models. Among all techniques, the extreme gradient boosting (XGB) model exhibited the most accurate and reliable predictions of permeate flux, confirmed by the lowest error metrics of MAE = 1.94 LMH, MAPE = 9.90 %, and RMSE = 2.59 LMH, along with the highest R2 (97.33 %), on the test dataset. In addition, to overcome the limited interpretability of black-box models, the Unified Shapley Additive Explanation technique was used, and the feed temperature and feed flowrate were identified as the most influential features on the permeate flux. Additionally, a user-friendly web interface was developed for the best predictive model, providing an accessible tool to advance DCMD applications in sustainable water treatment.