Modeling and prediction of desalination performance in a scaled-up membrane capacitive deionization system using machine learning and deep learning
Journal
Water Research
Journal Volume
298
Start Page
125839
ISSN
00431354
Date Issued
2026-06-15
Author(s)
Liu, Huei-Cih
Abstract
Membrane capacitive deionization (MCDI) is a promising desalination technology characterized by low energy consumption. However, its practical application often relies heavily on trial-and- error approaches, which are time consuming and inefficient for operation tuning. Since traditional mechanistic models tend to oversimplify complex physicochemical interactions, data-driven modeling plays an essential role in predicting the performance of MCDI system. In this study, a scaled-up MCDI system was operated under varying influent condictions, which served as dataset for the machine learning (ML) and deep learning (DL) models, to establish a robust data-driven modeling farmework. Four ML and DL models, including random forest (RF), extreme gradient boost (XGBoost), multilayer perceptron (MLP), and long-short term memory (LSTM), were developed to predict effluent conductivity (ECout). To ensure a leakage-free assessment, a strictly chronological cycle-based data splitting strategy was adopted. Additionally, hyperparameter tuning was conducted to balanced predictive performance and model stability, while Shapley Additive Explanations (SHAP) were employed to decode the underlying physicochemical logic of the black-box models. Among these models, the RF model achieved superior robustness and stable predictive accuracy, with a coefficient of determination (R2) of 0.79, root mean square error (RMSE) of 36.25 μs/cm, and a mean absolute error (MAE) of 22.92 μs/cm. The SHAP analysis validated the model's interpretability by identifying time, current, and pH as the most influential input features, thereby revealed that current as the driving force for electromigration, and pH as a critical indicator of electrochemical stability. Overall, this study demonstrates the potential of ML/DL models to capture nonlinear dynamics in large-scale MCDI systems, providing a bridge between data-driven predictions and process understanding.
Subjects
Desalination
Machine learning
Scale-up MCDI
SHAP interpretability
Publisher
Elsevier Ltd
Type
journal article