Explainable deep learning model for membrane capacitive deionization operated under fouling conditions

Abstract

To avoid fouling problems during operation, membrane capacitive deionization (MCDI) requires proper cleaning processes. In this study, we assessed seven different conditions to investigate the effects of flushing conditions and foulant concentration on the recovery rate of the MCDI salt adsorption capacity. Two representative deep learning models, namely the long short-term memory (LSTM) and temporal fusion transformer (TFT) models, were developed to simulate effluent salt concentrations under fouling conditions. The prediction results obtained using the two models indicated that the TFT model (R2, 0.945–0.993; RMSE, 0.051–0.151) was superior to the LSTM model (R2, 0.631–0.993; RMSE, 0.051–0.740) in terms of performance and applicability. Analyses of the permutation importance and attention weights were performed to evaluate the importance of input variables and the model-training process. The interpretation of the models based on attention scores revealed that the TFT model used the applied voltage and implementation of flushing as important inputs, which contributed to higher prediction accuracy. Thus, the proposed model could be utilized as an interpretable artificial intelligence model in practical applications to improve the efficiency of MCDI operations involving flushing processes.