Multi-scale modelling of supercapacitors : a combined simulation and machine learning approach

Abstract

Electrical double layer capacitors (EDLCs) store and release energy via re versible adsorption/desorption of ions at the electrode–electrolyte interface. Research on EDLCs mainly focus on improving their energy density while maintaining their at tractive properties such as high power density and long cycle life. EDLC performance is a complex function of the properties of its components, as well as the interactions between them. Given the large number of parameter combinations make traditional experiments remain infeasible for parameter optimization. To address this problem, we use molecular dynamics simulation data for a set of room temperature ionic liq uid/nanoporous carbon based EDLCs. By analyzing the charging kinetics and equi librium behavior of EDLCs using a transmission line model, we construct a simple data-driven method that is capable of quantitatively predicting energy density and time-dependent charging profile as a function of electrode micropore size and elec trolyte composition. In particular, linear and ridge regression, elastic networks, lasso, and neural network models are trained to predict gravimetric and volumetric capac itance (CG and CV ), charging time (τM), and electrical resistance (Rl). The elastic network model yields the best performance with a root mean square error of 3.10 F/g (CG), 0.15 s (τM), 1.09 F/cm3 (CV ), and 0.54 Ohm m (Rl). This model is then used to construct diagrams that show the dependence of the above-mentioned performance metrics to electrode pore size and electrolyte composition, and allow designing EDLCs with a set of predetermined performance criteria. This work can be extended to provide a framework that can quantify the effect of key factors on the EDLC performance.