Physics-driven deep learning for inverse extraction of capacitive parameters and faradaic contribution in supercapacitors

Cyclic voltammetry (CV) is a routine characterization technique for electrochemical capacitors, yet recovering the physical parameters underlying a measured curve remains largely dependent on user-defined analysis and interpretation, particularly when capacitive and faradaic contributions coexist and overlap in the voltammetric response. Moreover, kinetic limitations, ohmic drop, and leakage introduce additional distortions that such analyses cannot disentangle. Here, a physics-driven deep learning framework, in which a physical model is used to generate the training data but does not enter the network architecture or loss function, directly extracts these parameters from a single voltammogram. One million synthetic voltammograms, generated from an explicit electrochemical model that describes each of these contributions, were used to train a dual-input convolutional neural network that reads the current–potential curve together with its scan rate. On 50,000 unseen curves, prediction errors were 1.4% for capacitance, 4.6% for series resistance, and 28% for leakage resistance; the faradaic charge fraction, defined over a 0–10% range, was predicted with an error of 0.17 percentage points. Without retraining, the network was also applied to experimental CVs from two supercapacitors, yielding predictions consistent with independent electrochemical measurements. Implemented as a freely accessible web tool, the framework turns a single voltammogram into a rapid, reproducible fingerprint of electrochemical capacitor behaviour.