Machine Learning-Guided Prediction of Desalination Capacity and Rate of Porous Carbons for Capacitive Deionization

Nowadays, capacitive deionization (CDI) has emerged as a prominent technology in the desalination field, typically utilizing porous carbons as electrodes. However, the precise significance of electrode properties and operational conditions in shaping desalination performance remains blurry, necessitating numerous time-consuming and resource-intensive CDI experiments. Machine learning (ML) presents an emerging solution, offering the prospect of predicting CDI performance with minimal investment in electrode material synthesis and testing. Herein, four ML models are used for predicting the CDI performance of porous carbons. Among them, the gradient boosting model delivers the best performance on test set with low root mean square error values of 2.13 mg g-1 and 0.073 mg g-1 min-1 for predicting desalination capacity and rate, respectively. Furthermore, SHapley Additive exPlanations is introduced to analyze the significance of electrode properties and operational conditions. It highlights that electrolyte concentration and specific surface area exert a substantially more influential role in determining desalination performance compared to other features. Ultimately, experimental validation employing metal-organic frameworks-derived porous carbons and biomass-derived porous carbons as CDI electrodes is conducted to affirm the prediction accuracy of ML models. This study pioneers ML techniques for predicting CDI performance, offering a compelling strategy for advancing CDI technology. Nine features are selected as input into four machine learning (ML) models, with the aim to predict the capacitive deionization (CDI) performances. Gradient boosting model delivers best performance. Different porous carbons are further prepared to validate the rationality of predicted ML results. The consistence between ML prediction and experimental validation proves the feasibility of ML in CDI field. image