An artificial neural network model for capacitance prediction of porous carbon-based supercapacitor electrodes

Among energy storage devices, the last decades have witnessed the rapid spread of usage of carbon-based electrodes for electric double-layer capacitors (EDLCs) due to their large surface area, low cost, and high porosity. It is crucial to develop an accurate and efficient forecasting model for electrochemical performance to reduce the time needed for making suitable designs and choosing testing electrode materials. As a result, the use of machine learning (ML) approaches in creating a predicting model for the capacitance of carbon-based supercapacitors looks critical and provides the electrode characteristics' relative relevance. Data extracted from nearly a hundred published experimental research papers to select supercapacitors with certain electrode morphologies such as mesoporous, nanoporous, microporous, and hierarchical porous carbon electrode. The data was examined using machine learning techniques to predict the supercapacitor's specific capacitance (F/g). Electrode material structural qualities and various physicochemical test features such as electrolyte material, pore volume, and specific surface area. Electrochemical test features acquired via electrochemical impedance spectroscopy (EIS) and galvanostatic charge-discharge (GCD) test investigations for the same purpose include: cell configuration, current density, applied potential window, charge-transfer resistance (RCT), and equivalent series resistance (ESR) were used as input features to predict the corresponding capacitance performance. In the present study, Lasso, Support Vector Machine Regression (SVMR), and Artificial Neural Networks (ANN) with different structures were examined to predict the capacitance of the supercapacitor. The exhibition of the ML models measured concerning the root mean square error (RMSE), the correlation between expected yield and yield provided by the system. The developed ANN model with RMSE, MAE, and R values of 30.82, 46.5624 and 0.89537, respectively, provides outcomes for the prediction that are highly accurate compared to other models created for this purpose. According to the analysis of the input features done using the SHAP (SHapley Additive exPlanations) framework, the specific surface area had the biggest impact on the ANN model.