Machine learning improves biomass carbon cathodes for stronger zinc-ion supercapacitors

Nitrogen-doped porous carbon materials are gaining attention as cathodes for zinc-ion hybrid supercapacitors, which combine the energy storage of batteries with the quick charge and discharge capabilities of capacitors. A recent study published in the Journal of Materials Science introduces a machine learning framework that predicts the specific capacity of these cathode materials before synthesis, validated by experiments on three newly developed carbons.

The research team developed a predictive pipeline using a descriptor-to-performance model, optimized through grid search and validated with fivefold cross-validation. The model inputs included eight features: nitrogen content, defect density (ID/IG ratio), specific surface area, total pore volume, oxygen content, micropore ratio, average pore diameter, and applied current density. These descriptors encompass the chemical composition, structural disorder, porosity, and operational conditions that influence the charge storage capacity of doped carbon cathodes.

Three tree-based regression algorithms—Random Forest, Light Gradient Boosting Machine, and eXtreme Gradient Boosting (XGBoost)—were trained and compared. XGBoost emerged as the top performer, achieving 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. This indicates that the model could explain most variations in specific capacities, with a typical prediction error of around ten milliampere-hours per gram, making it effective for screening potential materials.

Understanding the rationale behind the model's predictions was as crucial as the predictions themselves. The researchers used SHapley Additive exPlanations (SHAP) to assign contributions to each input feature for individual predictions, alongside Partial Dependence Plots (PDP) to illustrate how performance changes with variations in a single descriptor while others remain constant. This approach provided insights into which features influenced predicted capacity, transforming the statistical model into a tool for understanding structure-performance relationships.

To verify the framework's applicability beyond training data, the team synthesized three nitrogen-doped porous carbons and tested devices made from them. A zinc-ion hybrid supercapacitor using the optimized carbon cathode and a zinc anode achieved a specific capacity of 160.3 mAh g–1 at a current density of 0.1 A g–1. The model predicted a capacity of 154.4 mAh g–1 for this material, closely matching the measured value and falling within the model's error range.

The authors highlight that machine learning can establish correlations between material descriptors and specific capacity, accurately predicting the performance of new nitrogen-doped porous carbon material (NPCM)-based devices. Since the descriptors are measurable synthesis and testing parameters, this method could allow researchers to computationally screen biomass-derived carbon recipes, focusing laboratory efforts on the most promising candidates.

This research contributes to the broader goal of developing energy materials that are cost-effective, sustainable, and safe. Biomass-derived carbons use renewable precursors, and nitrogen doping alters the electronic structure to enhance charge storage. Zinc-ion hybrid supercapacitors, using aqueous electrolytes and zinc metal, offer a safer and less costly alternative to lithium-ion technology. The study provides a quantitative link between synthesis-controllable features and device-level capacity.

However, limitations exist. The model was trained on existing NPCM-based device datasets, and its accuracy is confined to that training distribution. Its performance on different material classes or outside the tested current densities is not established. Experimental validation involved three synthesized carbons tested at 0.1 A g–1, so the reported 160.3 mAh g–1 capacity applies to those specific conditions. The framework was optimized using grid search and fivefold cross-validation, which reduces but does not eliminate overfitting risks.

Despite these limitations, the study exemplifies how machine learning is advancing from general materials discovery to targeted electrochemical applications. By integrating an interpretable XGBoost model with SHAP and PDP analysis and validating with experimental synthesis, the researchers demonstrated that a predicted capacity of 154.4 mAh g–1 can reliably guide cathode design, paving the way for faster, data-driven development of zinc-ion hybrid supercapacitor electrodes.