When universal machine learning models fail in capacitance prediction for carbon-based supercapacitors: Leakage-aware validation and electrolyte screening

Electrochimica Acta. 2026. Vol. 579. Article 150159.

Makarov D. M., Kalikin N., Fashchevsky K. A., Budkov Y.

Machine learning (ML) has become a powerful tool for accelerating the development of carbon-based supercapacitors;
however, most published models assume that a single predictive framework can describe both conventional
porous carbons and hybrid carbon electrodes. Here, we test this assumption using a comprehensive
benchmark dataset and an extended literature-curated dataset enriched with synthesis- and electrolyte-related
descriptors. Three ML algorithms (CatBoost, TabM, and TabPFN) were evaluated using conventional random
splitting and a more rigorous grouped validation based on electrode identity. Under random splitting, the best
model achieved an RMSE of 14 F/g, whereas electrode-wise validation increased the RMSE to approximately 54
F/g. Extending the dataset to include hybrid electrodes and literature-curated measurements (8038 measurements
in total) raised the grouped-validation RMSE to approximately 200 F/g, whereas restricting it to conventional
carbons reduced it to approximately 55 F/g. These results show that, within the present descriptor
space, a single universal model does not generalize reliably across conventional and hybrid electrodes because
the latter introduce additional charge-storage mechanisms that are not encoded by descriptors developed for
conventional carbons. This indicates that conventional and hybrid carbon electrodes should be treated as distinct
materials classes unless mechanism-specific descriptors are introduced. Incomplete reporting of experimental
descriptors was identified as an additional source of epistemic uncertainty. The validated conventional-carbon
model was subsequently employed to screen approximately 11000 ionic liquids with ionic conductivities
above 1 mS/cm. Cyano-containing organoborate anions consistently ranked among the best electrolyte components
for all representative pore architectures, while functionalized organic cations, including imidazolium
derivatives, repeatedly appeared among the highest-ranked candidates.

Keywords: двойной электрический слой, supercapacitors, Electric double layer, Суперконденсаторы, deep machine learning, глубокое машинное обучение, Electric capacitance, Удельная емкость

Publication based on the results of: Прогнозирование свойств молекулярных систем: совмещение методов машинного обучения и классических методов моделирования (2026)