Machine learning enabled property prediction of carbon-based electrodes for supercapacitors

For supercapacitor electrodes, carbon is one of the most sought materials. Many synthetic strategies have been reported for carbon materials with a wide range of features in terms of surface area, porosity, the extent of the disorder, and the presence of various doping elements. Though there are many reports on the independent impact of these properties on supercapacitive outcomes of carbon materials, the collective effect of these features on the supercapacitive behavior is still not well understood. Therefore, it is imperative to explore new methods to predict the supercapacitive performance of carbon material without attempting experimentation. In this regard, this work reports on the comparative efficacy of five machine-learning (ML) techniques for estimating the capacitance of carbon-based supercapacitors, including Linear Regression, Lasso Regression, XGBoost, Random Forest Regressor, and Artificial Neural Networks (ANN). We gathered data from various research papers and articles to prepare the dataset, followed by Data Cleaning and Feature Selection to select the best parameters for training our Machine Learning model. Additionally, five factors were chosen to determine their effects on capacitance: voltage window, specific surface area (m2/g), pore size (nm), ID/IG ratio, and extent of N-doping. Out of five ML approaches, XGBoost, Random Forest Regressor, and Artificial Neural Networks (ANN) produced respectable prediction results when tested against actual data. More significantly, this project demonstrates the promise of Machine-Learning in selecting suitable materials for energy storage applications.