Modeling and prediction of desalination performance in a scaled-up membrane capacitive deionization system using...

Area of Science:

• Water treatment technologies
• Electrochemical processes
• Data-driven modeling

Background:

• Membrane capacitive deionization (MCDI) is a low-energy desalination method.
• Current operational tuning relies on inefficient trial-and-error.
• Mechanistic models struggle with complex MCDI interactions.

Purpose of the Study:

• To establish a robust data-driven modeling framework for scaled-up MCDI systems.
• To predict effluent conductivity (EC_out) using machine learning (ML) and deep learning (DL) models.
• To enhance MCDI process understanding and optimize operations.

Main Methods:

• Developed and compared four ML/DL models: Random Forest (RF), XGBoost, MLP, and LSTM.
• Utilized a chronological, cycle-based data splitting strategy for rigorous model assessment.
• Employed Shapley Additive Explanations (SHAP) for model interpretability and feature importance analysis.

Main Results:

• The RF model demonstrated superior performance with R²=0.79, RMSE=36.25 μs/cm, and MAE=22.92 μs/cm.
• SHAP analysis identified time, current, and pH as key predictors of effluent conductivity.
• Model interpretability confirmed current's role in electromigration and pH's impact on electrochemical stability.

Conclusions:

• ML/DL models effectively capture nonlinear dynamics in large-scale MCDI systems.
• Data-driven modeling offers a viable alternative to traditional methods for MCDI optimization.
• This approach bridges predictive capabilities with fundamental process understanding for improved desalination.