Toward Sustainable Desalination: Experimental Evaluation and Predictive Modeling of Membrane Capacitive Deionization

Membrane capacitive deionization (MCDI) is a promising electrochemical technique for water desalination. However, accurately predicting its salt removal efficiency (SRE) remains challenging because desalination performance is governed by complex nonlinear interactions among operating conditions and feed-water characteristics. In this study, a laboratory-scale MCDI system was designed, fabricated, and experimentally evaluated to investigate key operational parameters (applied voltage and feed flow rate) and feed-water characteristics (total dissolved solids, salinity, pH, and temperature) on desalination performance. To address the limited number of experimental observations, 25 experimental samples were augmented to a dataset of 150 samples using a Bayesian network-based DataSynthesizer. Four ensemble machine learning algorithms, namely, random forest (RF), categorical boosting (CatBoost), eXtreme gradient boosting (XGBoost), and gradient boosting regression (GBR), were optimized using Bayesian hyperparameter optimization (Optuna). Among the investigated models, GBR achieved the highest predictive performance with training and testing R² values of 0.999 and 0.958, respectively, outperforming RF, CatBoost, and XGBoost. To improve model interpretability, SHapley Additive exPlanations (SHAP), partial dependence plots (PDP), individual conditional expectation (ICE), and accumulated local effect (ALE) analyses were employed to quantify feature importance and nonlinear relationships. Thus, the proposed framework can serve as an interpretable artificial intelligence tool for accurately predicting SRE and supporting the design, operation, and optimization of MCDI systems.