A data-driven machine learning approach to predict desalination capacities of faradic materials for capacitive deionization

Capacitive deionization (CDI) stands out as a promising desalination method owing to its energy efficiency, environmental sustainability, and high performance. Faradic materials, with high theoretical capacities and abundant redox reaction sites, have shown substantial potential for practical CDI applications. However, the exact impact of material properties and operational conditions on desalination capacity of faradic materials remains unclear, requiring numerous time-consuming and inefficient CDI experiments. Machine learning (ML) offers a prospective solution by training large amounts of data to provide a more efficient approach for predicting the desalination capacity of faradic materials with minimal investment in time and resource. This work employed four ML models to predict the desalination capacity of faradic materials and investigate the relationships between various features and desalination capacity. Notably, the gradient boosting decision tree model achieved high prediction accuracy with a mean absolute error of 6.34 mg g⁻¹ and root mean square error of 8.58 mg g⁻¹. SHapley Additive exPlanations analysis identified specific capacitance as the most influential feature for desalination capacity. Finally, Prussian blue analogues and molybdenum disulfide, as representative faradic materials, were prepared and tested for CDI, confirming the effectiveness of ML predictions. This study demonstrates the excellent function of ML in predicting the desalination capacity of faradic materials and highlights its potential to accelerate the development of advanced CDI systems.