Interpretable machine learning modeling of capacitive deionization for contribution analysis of electrode and process features

Abstract

Capacitive deionization (CDI) is a promising technique used to desalinate water via electrosorption of ions inside the porous structure of two oppositely charged electrodes. Developing a numerical model to predict CDI desalination performance and to understand how electrode and process features jointly contribute to desalination is very crucial for rational CDI system designing. However, the non-linear behavior of CDI and the interconnectivity of the parameters make this a challenging task. In this work, two different machine learning (ML) models of Artificial Neural Network and Random Forest have been implemented to predict the electrosorption capacity of CDI with a reasonable accuracy based on important electrode and process features. Then, based on the established models, the contribution and relative importance of each feature in deionization are determined and validated. The specific surface area of electrodes and the electrolyte salt concentration are defined as the most important electrode and process features, respectively. Oxygen and nitrogen elements of the electrode material are shown to have a suppressing and enhancing impact on deionization, respectively. A nitrogen-rich electrode with a dominant channel-pore fraction is expected to show high deionization capacity according to the established models which is in agreement with previous experimental and theoretical findings. This study shows the strong abilities of ML in predicting the non-linear behavior of the CDI system and in revealing the role of each feature in desalination.