Machine learning steers biomass carbon cathodes to stronger zinc-ion supercapacitors

Nitrogen-doped porous carbon materials have emerged as promising cathode candidates for zinc-ion hybrid supercapacitors, devices that pair a carbon cathode with a zinc anode to combine battery-like energy storage with the rapid charge and discharge of a capacitor. A new study in the Journal of Materials Science reports a machine learning framework designed to predict the specific capacity of these cathode materials before they are ever synthesized, and then backs up the predictions with laboratory experiments on three newly made carbons.

The research team built a descriptor-to-performance predictive pipeline, tuning it through grid search and validating it with fivefold cross-validation. Eight features served as model inputs: nitrogen content, defect density expressed as the ID/IG ratio, specific surface area, total pore volume, oxygen content, micropore ratio, average pore diameter, and the applied current density. These descriptors capture the chemical composition, structural disorder, porosity, and operating conditions that together determine how much charge a doped carbon cathode can store.

Three mainstream tree-based regression algorithms were trained and compared: Random Forest, Light Gradient Boosting Machine, and eXtreme Gradient Boosting. XGBoost proved the strongest performer, reaching a coefficient of determination of 0.92, a root mean square error of 10.08 mAh g–1, and a mean absolute error of 5.74 mAh g–1. In practical terms, the model could explain most of the variation in reported specific capacities while its typical prediction error stayed near ten milliampere-hours per gram, a useful margin for screening candidate materials.

Interpreting why the model made its predictions mattered as much as the predictions themselves. The researchers applied SHapley Additive exPlanations, a technique that assigns each input feature a contribution to every individual prediction, together with Partial Dependence Plots, which show how performance shifts as a single descriptor changes while others are held fixed. This combination allowed the team to trace which of the eight features pushed predicted capacity up or down, turning a statistical model into a source of structure–performance insight.

To test whether the framework held up outside the training data, the group synthesized three nitrogen-doped porous carbons and measured the devices built from them. The zinc-ion hybrid supercapacitor assembled with the optimized carbon cathode and a zinc anode delivered a specific capacity of 160.3 mAh g–1 at a current density of 0.1 A g–1. The model’s prediction for that material, 154.4 mAh g–1, landed close to the measured value, a gap of roughly six milliampere-hours per gram that falls within the error range the model itself reported.

The authors frame the result as a demonstration that machine learning can both establish correlations between material descriptors and specific capacity and accurately predict the performance of new NPCM-based devices. Because the descriptors are measurable synthesis and testing parameters rather than abstract quantities, the approach could in principle let researchers screen biomass-derived carbon recipes computationally and reserve laboratory effort for the most promising candidates.

The work sits within the broader push toward energy materials that are inexpensive, sustainable, and safe. Biomass-derived carbons draw their carbon skeleton from renewable precursors, and nitrogen doping modifies the electronic structure of the carbon lattice in ways that influence charge storage. Zinc-ion hybrid supercapacitors, meanwhile, use aqueous electrolytes and zinc metal, avoiding some of the cost and safety concerns of lithium-ion technology. The study’s contribution is a quantitative bridge between these two areas: a validated model that connects synthesis-controllable features to device-level capacity.

Limitations remain. The model was trained on existing datasets of NPCM-based devices, and its accuracy reflects that training distribution; performance on radically different material classes or far outside the tested current densities is not established by this study. The experimental validation covered three synthesized carbons, with the optimized device tested at 0.1 A g–1, so the reported 160.3 mAh g–1 figure applies to those specific conditions. The authors describe the framework as optimized through grid search and fivefold cross-validation, standard practices that reduce but do not eliminate the risk of overfitting.

Even with those caveats, the study offers a concrete example of how machine learning is moving from generic materials discovery into targeted electrochemical applications. By combining an interpretable XGBoost model with SHAP and PDP analysis and closing the loop with experimental synthesis, the researchers showed that a predicted capacity of 154.4 mAh g–1 could be trusted closely enough to guide cathode design, a step toward faster, data-driven development of zinc-ion hybrid supercapacitor electrodes.

Subject of Research: Machine learning-guided prediction and experimental synthesis of biomass-derived nitrogen-doped porous carbon cathodes for zinc-ion hybrid supercapacitors

Article Title: Integrating machine learning with experimental synthesis of biomass-derived N-doped carbon cathodes for zinc-ion hybrid supercapacitors

Article References: Cheng, S., Su, R., Xiao, L., Lin, X., Zang, L., & Yang, C. (2026). Integrating machine learning with experimental synthesis of biomass-derived N-doped carbon cathodes for zinc-ion hybrid supercapacitors. Journal of Materials Science. https://doi.org/10.1007/s10853-026-14019-6

Image Credits: AI Generated

DOI: 10.1007/s10853-026-14019-6

Keywords: zinc-ion hybrid supercapacitors, nitrogen-doped porous carbon, machine learning, XGBoost, SHAP, biomass-derived carbon, specific capacity prediction, cathode materials, energy storage, fivefold cross-validation, Random Forest, LightGBM