Deep learning-enhanced multi-modal modeling for electrosorption performance prediction via Nyquist plots

Abstract

Flow-through electrosorption systems are emerging as a highly promising and efficient technology that integrates selective ion recovery with electrochemical redox functions in aqueous environments. Hence, in this study, we developed a novel electrosorption-based multi-modal model to predict the removal efficiency of Cr(VI) and the regeneration efficiency of Cr(III) by incorporating Nyquist plot-based electrochemical characterization of the cathode electrode (i.e., graphitized nanodiamond) at different annealing temperatures along with key process parameters. As a result, the simple Artificial Neural Network (ANN) model without annealing temperature exhibited a low R2 value (i.e., 0.344-0.829) and high error margins (e.g., Mean Absolute Error (MAE): 0.085-0.148, Mean Squarred Error (MSE): 0.015-0.065). After adopting annealing temperature as variable, performance of ANN model was improved (i.e., R2: 0.716-0.814). Additionally, the prediction performance of the multi-modal model, coupled with ANN and Convolutional Neural Network (CNN), was significantly enhanced compared to the single ANN model, achieving notably higher R2 values (i.e., 0.950-0.977). Therefore, the result proved the necessity of CNN application for prediction with high accuracy. Additionally, Shapley Additive exPlanations (SHAP) value analysis highlighted the critical role of electrode durability tests during charging and discharging cycles, emphasizing their necessity within the system. Finally, 2D simulations provided insights into optimizing the process conditions for operating the flow-through redox-assisted electrosorption system.