Elemental‐Descriptor‐Based Machine Learning for Model‐Consistency‐Guided Screening of Binary Metal–Nonmetal Supercapacitor Electrodes

Abstract

The discovery of high‐performance supercapacitor electrodes remains constrained by experimental trial‐and‐error and heterogeneous literature data. A physically informed machine‐learning framework is presented to estimate the specific capacitance of binary metal–nonmetal systems using a compact and interpretable descriptor space. A curated dataset of 60 experimentally reported materials was constructed from more than 300 literature entries through strict completeness and consistency screening. Seven descriptors were considered, including metal first ionization energy, nonmetal electronegativity, specific surface area, crystallite size, electrolyte conductivity, current density, and potential window. Artificial neural network (ANN) and support vector regression (SVR) models were developed to capture nonlinear relationships while maintaining generalization. The models achieved strong predictive performance, with R 2 values of 0.95 for ANN and 0.92 for SVR and average absolute relative errors of approximately 20%. Sensitivity analyses revealed physically consistent relationships between descriptors and capacitance. ANN–SVR agreement was further used as a heuristic model‐consistency indicator to flag extrapolation‐sensitive predictions. The framework therefore enables conservative, trend‐level prioritization of promising binary metal–nonmetal electrodes before experimental validation.