Data-driven machine learning approach for predicting the capacitance of graphene-based supercapacitor electrodes

Graphene-based nanocomposites have shown strong potential as active components of high-capacity supercapacitors electrodes in energy storage systems. Developing an accurate and effective prediction technique for electrochemical performance is essential to decrease the time required for designing and testing electrode materials. In the present study, experimental data from more than two hundred published research papers have been extracted and examined through several machine learning (ML) models to predict the specific capacitance (F/g) of graphene-based electrode structures using various physicochemical features and diverse electrochemical measurements. The physicochemical features used in this work to predict the specific capacitance of the SCs electrode material include: carbon, nitrogen, and oxygen atomic percentages as well as electrode configuration, pore size, pore-volume, specific surface area (SSA), and ID/IG ratio. Electrochemical test features obtained from galvanostatic charge-discharge (GCD) tests and electrochemical impedance spectroscopy (EIS) analyses for the same purpose include: cell configuration, electrolyte ionic conductivity, electrolyte concentration, applied potential window, current density, charge-transfer resistance (RCT), and equivalent series resistance (RS). Four different ML models were developed: k-nearest neighbors' regression (KNN), decision tree regression (DTR), Bayesian ridge regression (BRR), and artificial neural network (ANN). The developed ANN model, with root mean square error (RMSE) and coefficient of determination (R2) values of 60.42 and 0.88, respectively, delivers extremely accurate prediction results compared to the other models developed for this purpose. The SHAP (SHapley Additive exPlanations) framework analysis of the input characteristics revealed that atomic percentages of nitrogen and oxygen doped graphene had the greatest effect on the ANN model.