# Fileset

[ChemPlusChem25_90_e202500342.pdf](https://mdr.nims.go.jp/filesets/7d89d933-1fcc-4e5f-8c67-a3a933267340/download)

## Creator

[Manickam Minakshi](https://orcid.org/0000-0001-6558-8317), Apsana Sharma, Ferdous Sohel, [Almantas Pivrikas](https://orcid.org/0000-0002-7713-2154), [Pragati A. Shinde](https://orcid.org/0000-0003-1730-2374), [Katsuhiko Ariga](https://orcid.org/0000-0002-2445-2955), [Lok Kumar Shrestha](https://orcid.org/0000-0003-2680-6291)

## Rights

[Creative Commons BY-NC-ND Attribution-NonCommercial-NoDerivs 4.0 International](https://creativecommons.org/licenses/by-nc-nd/4.0/)

## Other metadata

[Machine Learning—Guided Design of Biomass‐Based Porous Carbon for Aqueous Symmetric Supercapacitors](https://mdr.nims.go.jp/datasets/8d976e81-304b-4f26-b331-a96a48adb478)

## Fulltext

Machine Learning—Guided Design of Biomass‐Based Porous Carbon for Aqueous Symmetric Supercapacitorswww.chempluschem.orgMachine Learning—Guided Design of Biomass-BasedPorous Carbon for Aqueous Symmetric SupercapacitorsManickam Minakshi,* Apsana Sharma, Ferdous Sohel, Almantas Pivrikas, Pragati A. Shinde,Katsuhiko Ariga, and Lok Kumar Shrestha*Biomass-derived porous carbon electrodes have attracted signifi-cant attention for high-performance supercapacitor applicationsdue to their sustainability, cost-effectiveness, and tunable poros-ity. To accelerate the design and evaluation of these materials, it isessential to develop accurate and efficient strategies for optimiz-ing their physicochemical and electrochemical properties. Herein,a machine learning (ML) approach is employed to analyze exper-imental data from previously reported sources, enabling the pre-diction of specific capacitance (F g�1) based on various materialcharacteristics and processing conditions. The trained ML modelevaluates the influence of factors such as biomass type, electro-lyte, activating agent, and key synthesis parameters, includingactivation and carbonization temperatures and durations, onsupercapacitor performance. Despite growing interest, compre-hensive studies that correlate these variables with performancemetrics remain limited. This work addresses this gap by usingML algorithms to uncover the interrelationships between bio-mass-derived carbon properties, synthesis conditions, and spe-cific capacitance. Herein, it is demonstrated that an optimalcombination of a carbonized honeydew peel to H3PO4 ratio of1:4 and an activation temperature of 500 °C yields a highly porouscarbon material. When used in a symmetric device with 1 MH2SO4 electrolyte, this material, rich in oxygen and phosphorusspecies, achieves a high specific capacitance of 611 F g�1 at a cur-rent density of 1.3 A g�1. Correlation analysis reveals a strong syn-ergy between surface area and pore volume (correlationcoefficient= 0.8473), and the ML-predicted capacitance closelyaligns with experimental results. This ML-assisted frameworkoffers valuable insights into the critical physicochemical and elec-trochemical parameters that govern supercapacitor performance,providing a powerful tool for the rational design of next-genera-tion energy storage materials.1. IntroductionRechargeable battery technologies are well-established andwidely utilized in consumer electronics and electric vehicles.However, they face several limitations, including substantialweight, large volume, high internal resistance, low power density,and limited transient response. In contrast, supercapacitors, alsoknown as ultracapacitors or electrochemical double-layer capaci-tors (EDLCs), have emerged as promising alternatives for energystorage, owing to their rapid charge–discharge capability, highpower density, low weight, compact size, and low internalresistance.[1]Although supercapacitors have been known since the 1950s,their potential has only recently been fully recognized.[2] Amongvarious types, aqueous supercapacitors are particularly attractivedue to their low cost, environmental friendliness, and excellentcycling stability, making them ideal for applications such asregenerative braking systems in electric vehicles. The perfor-mance of supercapacitors is largely governed by the propertiesof the electrode materials, which directly influence capacitance,energy density, and overall device efficiency. Designing and opti-mizing electrode materials is, therefore, critical to advancingsupercapacitor technology.[3] Carbon-based materials are espe-cially well-suited for this purpose, offering high electricalM. Minakshi, A. Sharma, F. Sohel, A. PivrikasCollege of ScienceTechnology, Engineering & MathematicsMurdoch UniversityMurdoch 6150, Western Australia, AustraliaE-mail: minakshi@murdoch.edu.auM. Minakshi, P. A. Shinde, K. Ariga, L. K. ShresthaResearch Center for Materials Nanoarchitectonics (MANA)National Institute for Materials Science (NIMS)Tsukuba 305 0044, JapanE-mail: shrestha.lokkumar@nims.go.jpK. ArigaDepartment of Advanced Materials ScienceGraduate School of Frontier SciencesThe University of Tokyo5-1-5 Kashiwanoha, Kashiwa 277–8561, Chiba, JapanL. K. ShresthaDepartment of Materials ScienceInstitute of Pure and Applied SciencesUniversity of Tsukuba1–1 Tennodai, Tsukuba 305–8573, Ibaraki, JapanSupporting information for this article is available on the WWW under https://doi.org/10.1002/cplu.202500342© 2025 The Author(s). ChemPlusChem published by Wiley-VCH GmbH. This isan open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in anymedium, provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made.ChemPlusChem 2025, 90, e202500342 (1 of 13) © 2025 The Author(s). ChemPlusChem published by Wiley-VCH GmbHChemPlusChemResearch Articledoi.org/10.1002/cplu.202500342http://www.chempluschem.orghttps://orcid.org/0000-0001-6558-8317https://orcid.org/0000-0001-6558-8317https://orcid.org/0000-0002-7713-2154https://orcid.org/0000-0002-7713-2154https://orcid.org/0000-0002-2445-2955https://orcid.org/0000-0002-2445-2955https://orcid.org/0000-0003-2680-6291https://orcid.org/0000-0003-2680-6291mailto:minakshi@murdoch.edu.aumailto:shrestha.lokkumar@nims.go.jphttp://creativecommons.org/licenses/by-nc-nd/4.0/http://creativecommons.org/licenses/by-nc-nd/4.0/http://doi.org/10.1002/cplu.202500342http://crossmark.crossref.org/dialog/?doi=10.1002%2Fcplu.202500342&domain=pdf&date_stamp=2025-08-29conductivity, chemical stability, affordability, and long cycle life.[4]In particular, porous carbon materials derived from biomass havegained attention due to their sustainability and tunable texturalproperties, which are essential for enhancing ion transport andcharge storage in aqueous electrolytes.Biochar, produced through the thermochemical conversion ofbiomass, provides a sustainable and eco-friendly platform for syn-thesizing a wide range of functional carbon materials, includingporous carbon, heteroatom-doped biochar, carbon nanotubes,graphene, and carbon quantum dots, for advanced applica-tions.[5] The use of biomass-derived carbon in energy storagehas gained significant attention due to its cost-effectiveness, easeof production, and environmental sustainability. Meso- andmicroporous carbons can be readily obtained from diverse agri-cultural and biological sources such as plants, fruits, microorgan-isms, and animal residues.[6] Activated biochar-based electrodematerials, prepared via pyrolysis, hydrothermal treatment, chem-ical activation, and heteroatom doping, have emerged as prom-ising candidates for supercapacitor applications.[7] Thesematerials exhibit favorable physical and chemical properties,including well-developed pore structures, abundant surface func-tional groups, and excellent energy storage capabilities.[8] In con-trast, biochar produced through direct carbonization often suffersfrom limited electrochemical performance due to its underdevel-oped porosity, low graphitization, and insufficient surfacefunctionality.[9]The energy storage performance of activated biochar is influ-enced by multiple factors, including the intrinsic properties of thebiomass precursor, activation conditions, pore architecture, andtesting parameters. The complexity of these interdependent var-iables makes it challenging to identify optimal synthesis path-ways through experimental methods alone.[10] For instance, Jiaet al.[11] conducted 20 carbonization experiments on differentparts of corn straw, revealing that moderate carbonization tem-peratures and low ash content positively affect biochar’s energystorage performance. Similarly, Koutcheiko et al.[12] demonstratedthat optimizing activation conditions can enhance capacitance,but the process remains labor-intensive and lacks standardizedguidelines.To address these challenges, data-driven approaches such asmachine learning (ML) offer powerful tools for modeling and pre-dicting the relationships between synthesis parameters and elec-trochemical performance. ML techniques can simulate complexinteractions between input variables and output metrics,enabling efficient optimization of material properties.[13] Lenget al.[14] applied random forest (RF) and gradient boosting regres-sion models to predict biochar yield and surface area, identifyingpyrolysis temperature as a key factor. Yang et al.[15] used ML tocorrelate activation conditions with specific capacitance,highlighting the importance of surface area and activator dosage.Despite these advances, comprehensive optimization of the fullsynthesis chain, from biomass selection to activation and doping,remains elusive. Repeated experimentation to achieve desiredmaterial properties is time-consuming and inefficient.Therefore, further investigation is needed to understand howdifferent biomass types and processing conditions influencethe electrochemical behavior of activated biochar.[16]In this study, we perform a feature importance analysis toidentify the most influential parameters, such as biomass type,activating agent ratio, and activation temperature, on specificcapacitance. We extend our ML model to predict the capacitanceof honeydew peel-derived porous carbon in acid-aqueous elec-trolytes, building on our previous work,[17] which demonstrated asignificant improvement in specific capacitance from 185 to485 F g�1 in single-electrode configurations. A comparative anal-ysis was conducted using four ML models, RF, Support VectorRegression (SVR), Artificial Neural Network (ANN), and Multi-Layer Perceptron (MLP), to predict the capacitance of biomass-derived carbon supercapacitors. Eleven features were categorizedinto material properties (biomass type, surface area, pore size),thermal treatment parameters (activation/carbonization temper-atures and durations), and electrochemical conditions (activatingagent, electrolyte, voltage window, current density). The capaci-tive contributions from pores and heteroatoms were evaluatedbased on activator dosage and synthesis conditions.Overall, ML techniques were employed to predict and analyzethe influence of biomass carbon properties and experimentalparameters on supercapacitor performance. A comprehensivedataset was compiled and validated using honeydew peel-derived porous carbon. Among the models tested, SVR achievedthe highest coefficient of determination (R2= 0.41398) with amean absolute error (MAE= 0.11359). Both SVR and RF modelsdemonstrated strong predictive capabilities, achieving determi-nation coefficients (R2) of 0.41396 and 0.35689, MAEs of0.11359 and 0.11054, and root mean square errors (RMSE) of0.14897 and 0.13734, respectively. The ML classification algo-rithms highlighted key contributors to supercapacitor perfor-mance, with surface area and pore volume exhibiting strongcorrelation ratios of 0.8473 each. The optimal combination ofhoneydew peel and activator ratio, along with suitable synthesistemperature, resulted in functionalized carbon surfaces enrichedwith phosphorus and oxygen, enhancing wettability and pseudo-capacitance for high-performance aqueous supercapacitors.2. Methodology2.1. ML Approach2.1.1. Dataset DescriptionTo evaluate and predict the specific capacitance of biomass-derived carbon-based supercapacitors, ML techniques wereemployed to analyze the influence of various material and pro-cess parameters. Feature importance scoring, commonly used inregression and classification models, was applied to rank the con-tribution of each input variable to the target output (specificcapacitance). This approach enables researchers to identify themost influential features and reduce the dimensionality of thedataset without compromising model performance. To ensuremodel robustness and prevent overfitting, cross-validation wasChemPlusChem 2025, 90, e202500342 (2 of 13) © 2025 The Author(s). ChemPlusChem published by Wiley-VCH GmbHChemPlusChemResearch Articledoi.org/10.1002/cplu.202500342 21926506, 2025, 10, Downloaded from https://chemistry-europe.onlinelibrary.wiley.com/doi/10.1002/cplu.202500342 by National Institute For, Wiley Online Library on [11/10/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttp://doi.org/10.1002/cplu.202500342implemented. This technique partitions the dataset into multiplefolds, training the model on a subset while validating it on theremaining data. The process is repeated across different folds,providing a more reliable estimate of model generalizability tounseen data.[18]2.1.2. Data Collection and PreprocessingThe dataset was curated from previously reported literature (seeTable S1, Supporting Information), with a focus on maintainingconsistency and reducing complexity. To ensure homogeneity,only aqueous electrolytes such as KOH and H2SO4 were included,while organic electrolytes and ionic liquids (e.g., TEABF4) wereexcluded.[19] Since specific capacitance varies with current den-sity, normalization was applied to account for differences acrossstudies (Table 1).2.1.3. Data SegregationVariables were categorized into continuous and categorical typesto guide appropriate preprocessing and model selection.Continuous variables (e.g., surface area, pore size) were normal-ized or scaled to ensure uniform learning, while categorical var-iables (e.g., biomass type, electrolyte) were encoded using one-hot or label encoding techniques.[20] The dataset was furtherstructured into independent variables (features) and a dependentvariable (specific capacitance), as summarized in Table 2. This dis-tinction is critical for supervised learning, where the model learnsto map input features to the target output.[21] The eleven featuresused in the model were grouped into three categories:1) Material Properties: Biomass type, surface area, and pore size;2) Thermal Treatment Parameters: Activation temperature, car-bonization temperature, and activation time, carbonization time;3) Electrochemical Parameters: Activating agent, electrolytetype, voltage window, and current density.2.1.4. Data ProcessingIn this study, MATLAB software was employed for all stages ofdata handling, including preprocessing, model training, evalua-tion, and feature importance analysis aimed at predicting specificcapacitance. Data handling and preprocessing were conductedto import and structure the dataset efficiently. For regressionmodeling, the Support Vector Machine algorithm with aGaussian kernel was implemented using MATLAB’s fitcsvm func-tion, enabling nonlinear mapping of input features to the targetvariable. To enhance interpretability and reduce dataset complex-ity, Principal Component Analysis (PCA) was applied usingMATLAB’s PCA function. This technique facilitated improved visu-alization of feature distributions and relationships. Various plots,including scatter plots, regression surfaces, and feature correla-tion matrices, were generated using MATLAB’s built-in plottingtools to support exploratory data analysis and model interpreta-tion.[22] Feature importance was assessed using permutation fea-ture importance, which quantifies the increase in prediction errorwhen individual features are randomly permuted. This methodprovides insights into the relative contribution of each featureto model accuracy, helping identify the most influential param-eters affecting specific capacitance.[23,24] A schematic overview ofthe ML workflow is presented in Figure 1.2.1.5. Data CleaningIn this study, data normalization was employed as a critical proc-essing step to scale numerical features to a consistent range, typi-cally between 0 and 1. This ensures that variables with differingTable 1. Variables are segregated into categorical and continuous variables.Categorical variable Continuous variableBiomass material; Activation temperature;Activating agent; and Activation time;Electrolyte Carbonization temperature;Carbonization time;Pore volume;Surface area;Current density;Voltage window, andSpecific capacitanceTable 2. Variables are segregated into independent and dependentvariables.Independent Variable Dependent VariableBiomass material; CapacitanceActivating agent; Electrolyte;Activation temperature;Activation time;Carbonization temperature;Carbonization time;Pore volume;Surface area;Current density; andVoltage windowData SetTrain Machine Learning ModelTest ModelPredictionFigure 1. Schematic representation of the ML workflow involved in modelprediction.ChemPlusChem 2025, 90, e202500342 (3 of 13) © 2025 The Author(s). ChemPlusChem published by Wiley-VCH GmbHChemPlusChemResearch Articledoi.org/10.1002/cplu.202500342 21926506, 2025, 10, Downloaded from https://chemistry-europe.onlinelibrary.wiley.com/doi/10.1002/cplu.202500342 by National Institute For, Wiley Online Library on [11/10/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttp://doi.org/10.1002/cplu.202500342units or magnitudes do not disproportionately influence thelearning process of the ML model. Normalization enhances boththe performance and stability of ML algorithms by enabling uni-form treatment of all input features.[25] Additionally, data prepro-cessing was undertaken to prepare raw data for analysis, whichinvolved cleaning (handling missing values, removing noise),transforming (normalizing and encoding categorical variables),and organizing it for optimal model training. This step ensuredthat the data was clean, consistent, and adequately structuredfor effective learning by the ML model.[26] Continuous variablessuch as surface area, pore volume, and temperature were normal-ized to maintain a uniform scale across the dataset. Categoricalvariables were encoded using appropriate techniques to ensurecompatibility with the ML models.To prioritize features with the highest predictive potential,correlation analysis was performed. This helped identify variablesconsistently reported as significant in influencing specific capaci-tance, guiding feature selection and model refinement.[26]2.2. Physical Characterization of Heteroatom-DopedHoneydew PeelThe structural properties of the synthesized honeydew peel-derived activated carbon (HDP-AC), prepared at a synthesis tem-perature of 500 °C, were investigated using X-ray diffraction(Rigaku, Japan) using Cu Kα radiation. To assess the nature of car-bon bonding and surface structure, Raman spectroscopy was per-formed using an NRS-3100 (JASCO, Tokyo, Japan) instrument witha 532.09 nm neon laser excitation wavelength. Each sample wasanalyzed at a depth of 3 μm across at least three different surfacelocations to ensure representative data. The surface morphologyand elemental composition of HDP-AC were examined usingField Emission Scanning Electron Microscopy (FE-SEM) (TESCANCLARA). For high-resolution imaging of internal structures,Transmission Electron Microscopy (TEM) was conducted usinga JEOL 2200FS TEM operated at 200 kV. TEM specimens were pre-pared by dispersing HDP-AC powder in ethanol, followed bygrinding in an agate mortar and pipetting onto a holey carbonfilm supported on a copper mesh grid.2.3. Electrochemical Characterization of Heteroatom-DopedHoneydew PeelElectrochemical characterization of the HDP-AC was performedusing both three-electrode and symmetric two-electrode config-urations. The working electrode was fabricated by mixing HDP-AC(75 wt%), carbon black (15 wt%), and polyvinylidene difluoride(PVDF, 10 wt%) in 0.4 mL of N-Methyl-2-pyrrolidone (NMP) toform a homogeneous slurry. This slurry was coated onto a graph-ite sheet with an active area of 1 cm2, and the mass of activematerial loaded per electrode was �2mg. In three-electrodetests, platinum wire of 10 cm in length and 1mm in diameterdimension and mercury–mercuric oxide (Hg/HgO) served asthe counter and reference electrode, respectively. In single-electrode tests, HDP-AC served as the working electrode. Thecyclic voltammetry (CV) and galvanostatic charge–dischargestudies of the composites were carried out using SP-150, Bio-Logic Science instruments in 1 M H2SO4 electrolyte at room tem-perature. For HDP-AC, the working electrode was cycled between0 and 1.1 V. The frequency range for electrochemical impedancespectroscopy (EIS) was 10 40MHz with a 5mV bias voltage.The specific capacitance of the symmetric supercapacitordevice was computed from the following (Equation 1).Cs ¼i � ΔtΔv �mðFg�1Þ (1)where Cs is the specific capacitance in (F g�1), i is the current in(mA), Δt is the discharging time (s), m is the mass of the activematerial loaded (mg), and ΔV is the potential window (V). Theenergy and power density of the device were calculated usingthe formulas in (Equations 2) and (3).E ¼ 18CsΔv23.6ðWh kg�1Þ (2)P ¼ E� 3600tðWkg�1Þ (3)where E and P denote the energy density (Wh Kg�1) andpower density (W Kg�1) of the device. For the symmetric device,two identical HDP-AC electrodes (each 2 mg) were assembled in atwo-electrode configuration using 1 M H2SO4 as the electrolyte.3. Results3.1. MLMultiple ML models were evaluated using three key performancemetrics: RMSE, MAE, and the Coefficient of Determination (R2).These metrics provided a robust framework for assessing the pre-dictive accuracy of each model and its alignment with experi-mental values. The training dataset comprised electrochemicalperformance data of heteroatom-doped HDP-AC electrodes, syn-thesized with varying activating agent ratios and tested in 1 MH2SO4 electrolyte. This dataset enabled the models to learnthe relationships between synthesis parameters and specificcapacitance, facilitating reliable predictions and feature impor-tance analysis. The proposed ML model was trained on a datasetcomprising 166 data points, as detailed in Table S1, SupportingInformation. To ensure robust model training and validation, thedataset was partitioned into training, validation, and test sets,each representing approximately 20% of the total data. RMSEis assessed, with lower values indicating better model fit.R-squared (R2) values are analyzed to determine how well theinput features explain the variance in the output variable, withhigher values indicating stronger predictive performance.[27]MAE is also examined, measuring the average absolute differencebetween predicted and actual values, providing insight into over-all prediction accuracy. To further ensure generalizability, fourfoldcross-validation was applied. This technique tests the model’sChemPlusChem 2025, 90, e202500342 (4 of 13) © 2025 The Author(s). ChemPlusChem published by Wiley-VCH GmbHChemPlusChemResearch Articledoi.org/10.1002/cplu.202500342 21926506, 2025, 10, Downloaded from https://chemistry-europe.onlinelibrary.wiley.com/doi/10.1002/cplu.202500342 by National Institute For, Wiley Online Library on [11/10/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttp://doi.org/10.1002/cplu.202500342ability to perform on unseen data by iteratively training and vali-dating across different subsets of the dataset. Additionally, per-mutation feature importance was used to identify the mostinfluential input features. This method evaluates the increasein prediction error when individual features are randomly per-muted, offering insights into which synthesis parameters mostsignificantly impact specific capacitance predictions.The SVR model, illustrated in Figure 2a, achieves an RMSE of0.14897, an MAE of 0.11359, and an R2 of 0.41396, indicating amoderate fit between predicted and actual specific capacitancevalues. The corresponding scatter plot shows a close alignment ofdata points with the trend line, suggesting reasonable predictiveaccuracy. The RF model, shown in Figure 2b, yielded an RMSE of0.13734, an MAE of 0.11054, and an R2 of 0.35689. While RF dem-onstrates slightly lower error metrics compared to SVR, its mar-ginally lower R2 suggests that it captures slightly less variance inthe data, despite effectively minimizing prediction errors. Theneural network classifier (NNC), presented in Figure 2c, resultsin an RMSE of 0.15673, an MAE of 0.12398, and an R2 of0.35909. Although the error values are slightly higher, the scatterplot indicates a generally consistent alignment with actual values,reflecting moderate predictive capability. In contrast, the MLPmodel, depicted in Figure 2d, exhibited the poorest performance,with an RMSE of 1.0518, an MAE of 0.70824, and an R2 of only0.00975. The scatter plot reveals a wide dispersion of data pointsaround the trend line, indicating a lack of predictive accuracy andpoor model fit for specific capacitance prediction. Table S2,Supporting Information, presents the evaluation metrics forthe ML models, highlighting that the SVR model achieves a mod-erate to best fit. It is characterized by relatively low error valuesand the highest R2 score among the models, indicating its supe-rior ability to explain variance in the data. The correspondingscatter plot in Figure 2 further supports this, showing a strongcorrelation between predicted and actual values. Figure S1,Supporting Information, provides a comparative overview ofmodel performance, with error bars on RMSE illustrating the vari-ability in training outcomes across models.The NNC or ANN model exhibits moderately higher RMSE andMAE values, along with an R2 score that is lower than SVR butcomparable to RF. This suggests that while NNC captures certaintrends in specific capacitance, its predictive accuracy is limitedrelative to SVR and RF. The observed pattern may reflect theFigure 2. Comparison of the actual specific capacitance with predicted specific capacitance values of various models a) SVR; b) RF; c) NNC; and d) MLP.ChemPlusChem 2025, 90, e202500342 (5 of 13) © 2025 The Author(s). ChemPlusChem published by Wiley-VCH GmbHChemPlusChemResearch Articledoi.org/10.1002/cplu.202500342 21926506, 2025, 10, Downloaded from https://chemistry-europe.onlinelibrary.wiley.com/doi/10.1002/cplu.202500342 by National Institute For, Wiley Online Library on [11/10/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttp://doi.org/10.1002/cplu.202500342model’s insufficient adaptability to the complexity of the dataset,indicating potential underfitting or the need for further hyper-parameter tuning. In contrast, the MLP model shows a significantdeviation from the trend line, with the highest RMSE and MAEvalues and an extremely low R2 score (as shown in Table S2,Supporting Information). This poor performance indicates thatMLP fails to capture the underlying relationships between inputfeatures and specific capacitance. The widespread scatter of datapoints suggests limitations in the model’s current configuration,possibly due to overfitting or inadequate training.3.2. AnalysisEach ML model’s predictive accuracy for specific capacitance wasevaluated based on its alignment with the actual values and keyperformance metrics (RMSE, MAE, and R2) as shown in Figure 2.Among the models, SVR demonstrated a well-balanced fit, char-acterized by relatively low RMSE and MAE values and a moderateR2 score. This indicates that SVR effectively captures the underly-ing trend in specific capacitance and generalizes well to the data-set without significant overfitting or underfitting. The RF modelalso performed strongly, achieving the lowest RMSE and MAE val-ues, which reflect high predictive accuracy. However, its slightlylower R2 compared to SVR suggests that RF explains less of theoverall variance, despite its effectiveness in minimizing predictionerrors.To further explore feature relationships, the Pearson correla-tion matrix in Figure 3 reveals dependencies and potential multi-collinearity within the dataset. A strong positive correlationbetween surface area and pore volume (0.8473) indicates thatthese features are closely linked, with larger pore volumes gen-erally associated with increased surface area, both of which maycollectively influence specific capacitance. Conversely, a notablenegative correlation between activation temperature and activa-tion time (–0.6725) suggests an inverse relationship, likely reflect-ing trade-offs in synthesis conditions. This inverse relationshipreflects an optimization balance in the material preparation,where higher activation temperatures are often offset by lowercarbonization temperatures and vice versa. Such trade-offs arecommon in fabrication processes aimed at achieving desirableporosity and surface functionality. Features such as activatingagents, electrolytes, and voltage windows exhibit weakercorrelations with other variables, suggesting that they contributeindependently to specific capacitance without strong linear rela-tionships. These weaker correlations enhance their predictivevalue, as they capture unique aspects of capacitance variationwithout overlapping effects.The scatter plots in Figure 4 illustrate the relationshipsbetween individual features and specific capacitance in superca-pacitor performance. Figure 4a shows a weak, scattered correla-tion between biomass materials reported in the literature andspecific capacitance, indicating minimal influence from intrinsicbiomass properties. This suggests that most studies focus moreon physicochemical and electrochemical characteristics than onthe raw material itself.Figure 4b reveals a high concentration of voltage windowsdata points within a narrow range, implying a limited impacton capacitance. Figure 4c demonstrates that surface area hasan insignificant effect, as the points are widely spread withouta discernible pattern, whereas Figure 4d indicates a more notice-able trend with current density, where higher current densitiesappear to stabilize capacitance values, suggesting a moderatecorrelation. Figure 4e shows a scattered relationship with porevolume, indicating no strong influence on capacitance.Figure 3. Feature correlation coefficient diagram by SVR model.ChemPlusChem 2025, 90, e202500342 (6 of 13) © 2025 The Author(s). ChemPlusChem published by Wiley-VCH GmbHChemPlusChemResearch Articledoi.org/10.1002/cplu.202500342 21926506, 2025, 10, Downloaded from https://chemistry-europe.onlinelibrary.wiley.com/doi/10.1002/cplu.202500342 by National Institute For, Wiley Online Library on [11/10/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttp://doi.org/10.1002/cplu.202500342Figure 4f reveals that the majority of electrolyte data points clus-ter at a single level, implying minimal effect. Figure 4g displaysthat varying activating agent levels have little influence on capac-itance, while Figure 4h shows that activation temperatureremains largely uncorrelated, with data points clustering aroundzero. Figure 4i suggests that shorter carbonization time may allowfor a broader range of capacitance values, but does not show astrong trend. Finally, Figure 4j indicates a slight influence of acti-vation time, particularly at shorter durations, yet no clear lineartrend is observed. Collectively, these observations suggest thatspecific capacitance is weakly correlated with most individual fea-tures, with current density showing the most consistent relation-ship. This highlights the importance of considering featureinteractions and nonlinear effects in predictive modeling.Figure 4. The correlation between the specific capacitance and materials properties, thermal treatment, and electrochemical factor of a) Biomass material;b) Voltage windows; c) Surface area; d) Current density; e) Pore volume; f ) Electrolyte; g) Activating agent; h) Activation temperature; i) carbonization time;and j) activation time.ChemPlusChem 2025, 90, e202500342 (7 of 13) © 2025 The Author(s). ChemPlusChem published by Wiley-VCH GmbHChemPlusChemResearch Articledoi.org/10.1002/cplu.202500342 21926506, 2025, 10, Downloaded from https://chemistry-europe.onlinelibrary.wiley.com/doi/10.1002/cplu.202500342 by National Institute For, Wiley Online Library on [11/10/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttp://doi.org/10.1002/cplu.202500342Figure 5 presents the permutation feature importance analy-sis for the two most effective ML models in this study, SVR and RF,highlighting the relative influence of each input feature on spe-cific capacitance predictions. In the SVR model (Figure 5a), “bio-mass materials” emerged as the most influential feature, followedby “pore volume (cm³/g)” and “voltage window (V)”. These resultssuggest that the intrinsic properties of the biomass precursor,along with porosity and operating voltage range, play a signifi-cant role in determining specific capacitance. Features such as“carbonization temperature” and “current density (A g�1),”showed moderate importance, indicating their partial contribu-tion to performance outcomes. In contrast, activation time andcarbonization time was among the least impactful features, sug-gesting limited influence on capacitance within the studiedparameter range. Figure 5b presents the permutation featureimportance rankings for the RF model, revealing notable differ-ences compared to the SVR model. In RF, "Current Density (A/g),""Surface Area (m2g�1)," and "Voltage Window (V)" emerge as themost influential features in predicting specific capacitance. While"Biomass Materials" also shows considerable importance, theirinfluence is slightly reduced compared to their dominant rolein the SVR model. Consistent with SVR, “activation time" and "car-bonization time" rank among the least impactful features in theRF model. These results suggest that features such as "currentdensity," "surface area," "pore volume," and "voltage window"are critical to capacitance prediction, although their relativeimportance varies across models. This variation reflects theunderlying differences in model architecture: SVR is a kernel-based model that is sensitive to nonlinear relationships andcan extrapolate trends even in sparse data regions. In contrast,RF is an ensemble of decision trees and excels at capturinginteractions and thresholds. All four models were trained topredict specific capacitance, and their trends were comparedto evaluate how each responds to changes in input features.This comparative analysis highlights the key factors influencingcapacitance and provides valuable insights for optimizing super-capacitor design. According to model predictions, honeydewpeel-derived carbon in a 1 M H2SO4 electrolyte, when doped withphosphorus-rich heteroatoms, significantly enhances specificcapacitance, even at higher current densities. This improvementis attributed to increased electronic conductivity and enhancedelectrode surface wettability, underscoring the importance ofmaterial selection and surface chemistry in supercapacitorperformance.3.3. Synthesis and Activation of Heteroatom-DopedHoneydew PeelSeveral methods have been investigated for synthesizing AC fromhoneydew peel, as well as from peels of related fruits within thesame botanical family, which have shown promise as electrodematerials for supercapacitors.[17,28–32] In this study, we employeda straightforward synthesis approach. The powdered honeydewpeel was first subjected to carbonization in a muffle furnace at300 °C, with a heating rate of 2.5 °C min�1 for 4 h. The resultingcarbonized material was then thoroughly mixed with an opti-mized concentration of phosphoric acid (H3PO4) as the activatingagent. The resulting slurry was oven-dried and subsequently acti-vated in a muffle furnace at 500 °C, maintaining the same heatingrate for 1 h. After activation, the sample was allowed to cool toroom temperature. Phosphoric acid was selected due to its abilityto generate mesopores, thereby enhancing both pore diameterand volume. Additionally, its use simplifies the recovery of thecarbon product, requiring only a water rinse. This efficiency isattributed to the relatively low reaction temperatures and thecross-linking interactions between H3PO4 and the honeydewFigure 5. Permutation feature importance analysis of SVR and RF.ChemPlusChem 2025, 90, e202500342 (8 of 13) © 2025 The Author(s). ChemPlusChem published by Wiley-VCH GmbHChemPlusChemResearch Articledoi.org/10.1002/cplu.202500342 21926506, 2025, 10, Downloaded from https://chemistry-europe.onlinelibrary.wiley.com/doi/10.1002/cplu.202500342 by National Institute For, Wiley Online Library on [11/10/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttp://doi.org/10.1002/cplu.202500342peel, which inhibit the release of volatiles and the formation of tarduring pyrolysis.[33] The resulting heteroatom-doped honeydewpeel carbon, synthesized at an optimized precursor-to-activatingagent ratio of 1:4, presents a promising pathway toward achiev-ing the theoretical capacitance limits of porous carbon materials.3.4. Physicochemical Properties of Heteroatom-Doped(oxygen and Phosphorous-Rich) Honeydew Peel MaterialThe honeydew peel sample was finely ground into a powder toensure a smooth and homogenous surface for X-ray diffractionanalysis. The X-ray diffraction pattern of the AC derived from hon-eydew peel (Figure 6a) exhibited characteristic peaks around 25°and 44° corresponding to the (002) plane of turbostratic carbon(t-carbon)[29] and the (101) plane of graphitized carbon,[17] respec-tively. Complementary Raman spectroscopy (Figure 6b) revealeddistinct peaks at approximately 1350 cm�1 (D-band, indicatingstructural disorder), 1600 cm�1 (G-band, representing graphiticdomains), and around 2700 cm�1, confirming the formation ofAC. The presence of these peaks reflects the intrinsic molecularstructure of amorphous carbon in the honeydew peel. The inten-sity ratio of the D to G bands (ID/IG) was calculated to be 1.05,suggesting a high degree of defects in the graphitic structure,an advantageous feature for energy storage applications. Theactivation process begins with cellulose depolymerization, fol-lowed by dehydration of biopolymers to form aromatic rings, ulti-mately leading to the development of porous carbon.[33,34] Duringthis process, water-soluble KCl, formed during carbonization, mayreact with the honeydew peel under thermal treatment. Chlorineis released as a gas, while potassiummelts and integrates into thecarbon matrix, contributing to the formation of porous graphiticcarbon. This structural transformation is further supported bySEM images (Figure 7a,b), which reveal a compact and porousarchitecture conducive to efficient ion transport.The use of H3PO4 as an activating agent plays a critical role inpore development during the early stages of activation.[34,35] Themacroporous open structure observed in Figure 7b, resultingFigure 6. a) X-ray diffraction pattern and b) Raman spectra of the HDP-AC. G denotes graphitic and D denotes defect bands in the spectra.Figure 7. a,b) FE-SEM images, c,d) and the corresponding EDS spectra con-firming the presence of oxygen and phosphorus species, e,f ) TEM imagesof the honeydew peel-derived porous carbon.ChemPlusChem 2025, 90, e202500342 (9 of 13) © 2025 The Author(s). ChemPlusChem published by Wiley-VCH GmbHChemPlusChemResearch Articledoi.org/10.1002/cplu.202500342 21926506, 2025, 10, Downloaded from https://chemistry-europe.onlinelibrary.wiley.com/doi/10.1002/cplu.202500342 by National Institute For, Wiley Online Library on [11/10/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttp://doi.org/10.1002/cplu.202500342from intense cross-linking, is well-suited for high-rate capabilityand serves as a buffer zone for rapid ion diffusion.Additionally, the cascaded crystal particles visible in Figure 7bare likely phosphorus deposits on the surface of the AC.Energy-dispersive X-ray spectroscopy (EDS) data (Figure 7c,d)confirm the presence of oxygen- and phosphorus-rich species,with the O/P atomic ratio indicating co-doping of these elementsinto the carbon matrix. TEM images (Figure 7e,f ) further validatethe porous nature and amorphous regions of the synthesized car-bon material.3.5. Electrochemical Properties of Heteroatom-DopedHoneydew Peel MaterialThe electrochemical energy storage properties of heteroatom-doped honeydew peel-derived porous carbon were initiallyassessed using CV at scan rates ranging from 5 to 50mV/s,and galvanostatic charge–discharge (CD) measurements at vari-ous current densities. These tests were conducted in a conven-tional three-electrode system with 1 M H2SO4 as the electrolyte.The corresponding results are presented in Figure 8a,b. Both theCV and CD curves exhibited nearly rectangular and triangularshapes, respectively, characteristic of ideal EDLC behavior, indi-cating excellent capacitive performance and high specific capac-itance reaching up to 485 F g�1.The enhanced specific capacitance is attributed to the pres-ence of oxygen- and phosphorus-rich species, which facilitate theformation of P─O─C structures through cross-linking and oxida-tion reactions. This structural evolution, combined with the pres-ervation of micropores as observed in microscopy images,contributes to superior charge storage capability. Encouragedby the excellent performance within a 1 V voltage window,a symmetric supercapacitor device was fabricated using a two-electrode configuration with 1 M H2SO4 as the electrolyte. TheCV and CD profiles of the symmetric device (Figure 9a,b) closelyresemble those of the three-electrode setup, confirming EDLCbehavior driven by ion adsorption on the porous carbonelectrodes.Notably, the area under the CV curve in Figure 9a is signifi-cantly larger than that of the single electrode in Figure 8a, reflect-ing a higher specific capacitance. The symmetric device achieveda specific capacitance of 611 F g�1 at a current density of 1.3 A g�1(Figure 9b) and delivered a maximum energy density of35Wh kg�1 at a power density of 650W kg�1. This performancesurpasses that of most reported carbon-based symmetric capaci-tors (see Table S1, Supporting Information), which typically relyon materials with high surface area and pore volume. In contrast,our material, phosphorus and oxygen co-doped honeydew peel-derived carbon, exhibits a relatively low surface area of 33 m2g�1.Across our study, the samples demonstrated a wide range ofBrunauer-Emmett-Teller surface areas, from 224 down to3.0049m2g�1, with corresponding pore volumes ranging from0.076 to 0.000867 cm3 g�1. These variations are attributed tothe differing concentrations of the activating agent used, fromlow to high levels. However, under these conditions, both theSVR and RF models predicted relatively low specific capacitancevalues, 229 F g�1 and 208 F g�1, respectively. This discrepancymay be explained by several contributing factors[36,37] 1) insuffi-cient surface area and pore volume for effective ion adsorption.2) A limited voltage window restricting energy storage. 3) moder-ate current density affecting ion transport.These predictions suggest that under suboptimal physicalconditions, such as low surface area and pore volume, the MLmodels estimate low to moderate specific capacitance values.This is expected, as these features are strongly correlated withion accessibility and charge storage capacity. However, in realexperimental settings, additional surface properties often playa significant role in enhancing charge storage, leading to muchhigher capacitance values, such as the observed 611 F g�1. Thediscrepancy between model predictions and experimental resultsmay be attributed to phosphorus and oxygen co-doping, whichcan significantly enhance electrochemical performance by intro-ducing pseudocapacitive behavior and improving conductivity.This highlights the critical importance of validating model predic-tions with actual experimental data, especially when material-level modifications are involved.Figure 8. a) CV, and b) charge–discharge (CD) curves for the AC sample derived from the honeydew peel tested in a three-electrode configuration in 1MH2SO4 electrolyte.ChemPlusChem 2025, 90, e202500342 (10 of 13) © 2025 The Author(s). ChemPlusChem published by Wiley-VCH GmbHChemPlusChemResearch Articledoi.org/10.1002/cplu.202500342 21926506, 2025, 10, Downloaded from https://chemistry-europe.onlinelibrary.wiley.com/doi/10.1002/cplu.202500342 by National Institute For, Wiley Online Library on [11/10/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttp://doi.org/10.1002/cplu.202500342The long-term cycling performance of the biomass-derivedsymmetric supercapacitor, shown in Figure 9c, demonstratesexcellent stability, retaining 98.2% of its initial specific capaci-tance after 15,000 charge–discharge cycles at a current densityof 1.3 A g�1. EIS results, presented in Figure 9d, reveal that theNyquist plot remains nearly unchanged after extended cycling,indicating consistent resistance behavior and robust electrodeintegrity.Overall, the Nyquist plot exhibits features characteristic of anideal capacitor: a nearly vertical line at high frequencies, repre-senting capacitive behavior, and a 45° slope at mid frequencies,attributed to the distributed resistance within the porous carbonelectrode. These experimental findings provide valuable insightsinto the electrochemical behavior of the material and support thepredictive capabilities of ML models. The input variables thatmost significantly influence performance include the nature ofthe biomass precursor, the optimized concentration of the acti-vating agent, and the choice of electrolyte.3.6. Discussions and Limitations of the Current StudyThis study highlights the importance of narrowing the scope ofdata collection to minimize variability in the analysis of superca-pacitor performance. Focusing on specific types of biomass mate-rials, particularly those that are commonly utilized or possesssimilar properties (e.g., cellulose content), is essential to reducevariations associated with diverse starting materials.[28]Additionally, outliers, especially those reporting unusually highspecific capacitance values, should be critically evaluated andpotentially excluded if they demonstrate inconsistencies withother relevant parameters. When available, pore size distributiondata should be incorporated with caution, ensuring consistencyin measurement techniques, as pore architecture plays a criticalrole in determining capacitance.[29] Standardizing these variablesenhances the reliability of comparative analyses and improvesthe predictive accuracy of ML models applied to supercapacitorresearch.It is essential to distinguish between carbonization and hydro-thermal carbonization (HTC) when collecting data for ML appli-cations in supercapacitor research. These two processes yieldcarbon materials with distinct physical and chemical characteris-tics that significantly influence key performance metrics such asspecific capacitance, surface area, and pore structure. Forinstance, honeydew peel-derived carbon produced via conven-tional carbonization exhibits notable differences in surface area,pore size distribution, and surface functional groups compared tomaterials synthesized through HTC.[17,34]These discrepancies directly affect the suitability and perfor-mance of the resulting materials in energy storage applications.Moreover, the operational conditions, such as temperature, pres-sure, and reaction medium (dry vs. wet), differ substantiallyFigure 9. a) CV, b) charge–discharge (CD) curves for the AC sample derived from the honeydew peel tested in a two-electrode configuration as a symmet-ric supercapacitor device in 1MH2SO4 electrolyte. c) cycling stability of the device with the inset showing the charge–discharge curves superimposed for ini-tial and final cycles, and the corresponding d) Nyquist plots showing no visible changes in the shape of the curve.ChemPlusChem 2025, 90, e202500342 (11 of 13) © 2025 The Author(s). ChemPlusChem published by Wiley-VCH GmbHChemPlusChemResearch Articledoi.org/10.1002/cplu.202500342 21926506, 2025, 10, Downloaded from https://chemistry-europe.onlinelibrary.wiley.com/doi/10.1002/cplu.202500342 by National Institute For, Wiley Online Library on [11/10/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttp://doi.org/10.1002/cplu.202500342between carbonization and HTC, leading to unique reaction path-ways and end products. These variations influence carbon yield,structural integrity, and purity, all of which contribute to electro-chemical performance differences.Consequently, materials produced via HTC (hydrochar) mayexhibit electrochemical behaviors that diverge from those of con-ventionally carbonized materials (biochar). For ML applications, itis therefore imperative to treat carbonization and HTC as distinctprocesses. By incorporating the synthesis method as a separatefeature in the ML model, the algorithm can more accurately cap-ture the influence of each process on performance outcomes,such as specific capacitance.[30]The choice of activating agents, such as KOH, CaCl2, K2FeO4,MgCl2, H3PO4, and steam, plays a pivotal role in determining thephysicochemical properties of the resulting carbon materials.While all these agents serve to activate carbon, they operatethrough distinct mechanisms, leading to significant variationsin surface area, pore architecture, and chemical composition.Therefore, treating these agents as equivalent during data collec-tion for ML applications is inappropriate and may compromisemodel accuracy. Chemical activation, involving agents likeKOH, H3PO4, CaCl2, and K2FeO4, induces specific reactions withthe carbon precursor, influencing pore formation, surface area,and the incorporation of functional groups.[31,34] Each agent pro-duces unique pore size distributions and chemical functionalities.In contrast, physical activation, typically using steam or MgCl2,relies on gasification and controlled carbon burn-off at elevatedtemperatures to enhance porosity. Some agents, such as K2FeO4,may also exhibit catalytic effects, further modifying the carbonstructure and its electrochemical behavior.[32]The selected activating agent significantly affects perfor-mance metrics such as specific capacitance, energy density,and cycling stability. For example, KOH-ACs are often favoredfor their high surface area and conductivity, while H3PO4-ACsare valued for their stability in aqueous electrolytes. For ML appli-cations, it is essential to encode each activating agent as a distinctcategorical feature. This enables the model to learn the individualand interactive effects of each agent on material properties andperformance. Additionally, recording process parameters, such asactivation temperature, duration, and carbonization conditions,alongside the activating agent, ensures consistency and enhan-ces model reliability.It is also important to acknowledge that the data used inthis study were sourced from various research articles (listed inTable S1), each conducted under different environmental andexperimental conditions. Such diversity may introduce inconsis-tencies in ML outcomes. Furthermore, variations in terminologyacross studies can lead to misinterpretations and compromise theintegrity of the final analysis. The failure to account for differencesin carbonization methods, biomass types,[37] and activation tech-niques may have influenced the appropriateness of the study’sresults. Therefore, future researchers are advised to exercisecaution and rigor in data collection, ensuring that these criticalfactors are carefully considered and consistently documented.4. ConclusionsThis study presents an ML-guided investigation into oxygen andphosphorus-rich heteroatom-doped carbon material derivedfrom biomass for supercapacitor applications. The aqueous sym-metric supercapacitor demonstrated a strong correlationbetween specific capacitance and the nature of biomass precur-sor, a finding that stands out as particularly novel. By integrating adiverse dataset compiled from various published sources, theresearch effectively highlights the complex interrelationshipsamong critical features such as biomass type, surface area, porevolume, and experimental parameters, including activation andcarbonization temperatures.The performance evaluation of four distinct ML models, SVR,RF, ANN, and MLP, demonstrated their predictive capabilities forsupercapacitor performance. Among these, the SVR modelemerged as the most effective, achieving a coefficient of deter-mination (R2) of 0.41398 and a MAE of 0.11359. These findingsunderscore the potential of ML to enhance the understandingand optimization of biomass-derived carbon materials for energystorage.Importantly, the study identified surface area and pore vol-ume as key contributors to specific capacitance. However, in thecase of heteroatom-doped honeydew peel-derived carbon, theinfluence of biomass composition and dopant chemistryappeared to outweigh the effects of surface area and porosity.This insight provides a valuable foundation for future researchaimed at tailoring carbon material properties for enhancedenergy storage performance. The study also emphasizesthe importance of focused data collection and rigorousconsideration of material preparation methods. These recom-mendations are critical for improving the reliability of ML mod-els and advancing the field of supercapacitor research.Experimentally, the symmetric supercapacitor fabricated in thiswork achieved a specific capacitance of 611 F g�1, with anenergy density of 35 Wh kg�1 and a power density of650 W kg�1, performance metrics that surpass many reportedcarbon-based systems. Overall, this work contributes meaning-fully to the growing body of knowledge in energy storage, advo-cating for the continued integration of ML techniques toaccelerate the development of high-performance, sustainablesupercapacitor materials.AcknowledgementsM.M. acknowledges the Winston Churchill Fellowship to learnglobally and inspire locally. This work was partially supportedby the Japan Society for the Promotion of Science (JSPS)KAKENHI (grant Nos. JP20H00392 and JP23H05459).Open access publishing facilitated by Murdoch University, aspart of the Wiley - Murdoch University agreement via the Councilof Australian University Librarians.ChemPlusChem 2025, 90, e202500342 (12 of 13) © 2025 The Author(s). ChemPlusChem published by Wiley-VCH GmbHChemPlusChemResearch Articledoi.org/10.1002/cplu.202500342 21926506, 2025, 10, Downloaded from https://chemistry-europe.onlinelibrary.wiley.com/doi/10.1002/cplu.202500342 by National Institute For, Wiley Online Library on [11/10/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttp://doi.org/10.1002/cplu.202500342Conflict of InterestThe authors declare no conflict of interest.Data Availability StatementThe data that support the findings of this study are available fromthe corresponding author upon reasonable request.Keywords: biomass · carbon · dopant · energy · machinelearning · storage[1] G. G. Prasad, N. Shetty, S. Thakur, Rakshitha, K. B. Bommegowda, in IOPConf. Series: Materials Science and Engineering, IOP Publishing Ltd., 561,2019, p. 012105.[2] S. Butler, Supercapacitors vs. Batteries: What’s the Difference? https://www.howtogeek.com/786195/supercapacitors-vs-batteries-whats-the-difference/#:~:text=Supercapacitors%20have%20been%20around%20since%20the%201950s.2022, (accessed: July 2025).[3] M. A. Dar, S. R. Majid, M. Satgunam, C. Siva, S. Ansari, P. Arularasan,S. R. Ahamed, Int. J. Hydrogen Energy 2024, 70, 10.[4] L. Wang, X. Hu, Chem. Asian J. 2018, 13, 1518,[5] S. Rawat, R. K. Mishra, T. Bhaskar, Chemosphere 2022, 286, 131961.[6] Y. Liu, J. Chen, B. Cui, P. Yin, C. Zhang, C J. Carbon Res. 2018, 4, 53.[7] X. Yuebin, X. Liu, W. Xiong, H. Wang, X. Ji, F. Kong, G. Yang, J. Xu, Ind. CropsProd. 2021, 174, 114184.[8] J. Li, R. Xiao, M. Li, H. Zhang, S. Wu, C. Xia, Fuel Proces. Technol. 2019, 192,239.[9] D. Zhang, P. Yang, Y. Zhang, H. Liu, Y. Xu, J. Wu, P. Li, J. Energy Storage2022, 56, 106134.[10] V. Sawant, R. Deshmukh, C. Awati, J. Energy Chem. 2023, 77, 438.[11] J. Jia, Z. Yao, L. Zhao, T. Xie, Y. Sun, L. Tian, L. Huo, Z. Liu, Biomass Bioenergy2023, 175, 106858.[12] S. Koutcheiko, V. Vorontsov, J. Biobased Mater. Bioenergy 2013, 7, 733.[13] Z. Dong, X. Bai, D. Xu, W. Li, Bioresour. Technol. 2023, 367, 128182.[14] L. Leng, L. Yang, X. Lei, W. Zhang, Z. Ai, Z. Yang, H. Zhan, J. Yang, X. Yuan,H. Peng, H. Li, Biochar 2022, 4, 63.[15] X. Yang, C. Yuan, S. He, D. Jiang, B. Cao, S. Wang, Fuel 2023, 331, 125718.[16] Y. Sun, P. Sun, J. Jia, Z. Liu, L. Zhao, Y. Zhao, W. Niu, Z. Yao, Chem. Eng. J.2024, 485, 149975.[17] M. Minakshi, A. Samayamanthry, J. Whale, R. Aughterson, P. A. Shinde,K. Ariga, L. K. Shrestha, Chem. Asian J. 2024, 19, e202400622.[18] A. Seraj, M. M. Khanaposhtan, R. Daneshfar, M. Naseri, M. Esmaeili,A. Baghban, S. Habibzadeh, S. Eslamian, Handbook of Hydroinformatics,Elsevier 2023, pp. 89–105.[19] A. Jamaluddin, D. Harjunowibowo, S. Budiawanti, N. A. Kurdhi, S. D. T. Lai,S. Ramesh, Energy Rep. 2023, 10, 3125.[20] C. Rudin, C. Chen, Z. Chen, H. Huang, L. Semenova, C. Zhong, StatisticSurveys 2022, 16, 1.[21] I. Tsochantaridis, T. Hofmann, T. Joachims, Y. Altun, in Proceedings of thetwenty-first int. conf. on Machine learning 2004, p. 104.[22] I. Turk, Practical MATLAB, Apress, Berkeley, CA 2019.[23] Z. M. Omer, H. Shareef, Intell. Syst. Appl. 2023, 19, 200261.[24] R. S. Abirami, G. S. Kumar, SN Comput. Sci. 2022, 3, 1.[25] A. E. Eldeen, M. A. Azim, M. Abdelsattar, A. AbdelMoety, J. Energy Storage2024, 100, 113556.[26] S. A. N. Alexandropoulos, A. S. B. Kotsiantis, M. N. Vrahatis, Knowl. Eng. Rev.2019, 34, e1.[27] N. Redell, (Preprint) arXiv: 1908.09718 2019, 09718, https://arxiv.org/abs/1908.09718.[28] Z. Wang, P. Tammela, M. Strømme, L. Nyholm, Adv. Energy Mater. 2017, 7,1700130.[29] T. Wang, R. Pan, M. L. Martins, J. Cui, Z. Huang, B. P. Thapeliya,C.-L. Do-Thanh, M. Zhou, J. Fan, Z. Yang, M. Chi, T. Kobayashi, J. Wu,E. Mamontov, S. Dai, Nat. Commun. 2023, 14, 4607.[30] M. Aghaaminiha, R. Mehrani, S. Sharma, Biomass Convers. Biorefin. 2023,13, 9855.[31] X. Pang, M. Cao, J. Qin, X. Li, X. Yang, J. Porous Mater. 2022,29, 559.[32] D. M. Demir, M. Doguscu, ChemistrySelect 2022, 7, e202104295.[33] M. A. Yahya, Z. Al-Qodah, C. W. Z. Ngah, Renewable Sustainable Energy Rev.2015, 46, 218.[34] M. Fan, Y. Shao, Y. Wang, J. Sun, H. He, Y. Guo, S. Zhang, S. Wang, B. Li,X. Hu, Renew. Energy 2024, 240, 122151.[35] T. Adinaveen, L. J. Kennedy, J. J. Vijaya, G. Sekaran, J. Mater, Cycles WasteManag. 2015, 17, 736.[36] L. Zhu, H. Yin, L. Ju, B. Zhou, B. Hu, F. Hou, Y. Xie, J.-L. Zhan, W. Du, Ceram.Int. 2024, 50, 48609.[37] W. Du, Z. Zhang, L. Du, X. Fan, Z. Shen, X. Ren, Y. Zhao, C. Wei, S. Wei,J. Alloys Compd. 2019, 797, 1031.Manuscript received: May 20, 2025Revised manuscript received: July 13, 2025Version of record online: August 29, 2025ChemPlusChem 2025, 90, e202500342 (13 of 13) © 2025 The Author(s). ChemPlusChem published by Wiley-VCH GmbHChemPlusChemResearch Articledoi.org/10.1002/cplu.202500342 21926506, 2025, 10, Downloaded from https://chemistry-europe.onlinelibrary.wiley.com/doi/10.1002/cplu.202500342 by National Institute For, Wiley Online Library on [11/10/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttps://www.howtogeek.com/786195/supercapacitors-vs-batteries-whats-the-difference/#:~:text=Supercapacitors%20have%20been%20around%20since%20the%201950s.https://www.howtogeek.com/786195/supercapacitors-vs-batteries-whats-the-difference/#:~:text=Supercapacitors%20have%20been%20around%20since%20the%201950s.https://www.howtogeek.com/786195/supercapacitors-vs-batteries-whats-the-difference/#:~:text=Supercapacitors%20have%20been%20around%20since%20the%201950s.https://www.howtogeek.com/786195/supercapacitors-vs-batteries-whats-the-difference/#:~:text=Supercapacitors%20have%20been%20around%20since%20the%201950s.https://arxiv.org/abs/1908.09718https://arxiv.org/abs/1908.09718http://doi.org/10.1002/cplu.202500342 Machine Learning-Guided Design of Biomass-Based Porous Carbon for Aqueous Symmetric Supercapacitors 1. Introduction 2. Methodology 2.1. ML Approach 2.1.1. Dataset Description 2.1.2. Data Collection and Preprocessing 2.1.3. Data Segregation 2.1.4. Data Processing 2.1.5. Data Cleaning 2.2. Physical Characterization of Heteroatom-Doped Honeydew Peel 2.3. Electrochemical Characterization of Heteroatom-Doped Honeydew Peel 3. Results 3.1. ML 3.2. Analysis 3.3. Synthesis and Activation of Heteroatom-Doped Honeydew Peel 3.4. Physicochemical Properties of Heteroatom-Doped (oxygen and Phosphorous-Rich) Honeydew Peel Material 3.5. Electrochemical Properties of Heteroatom-Doped Honeydew Peel Material 3.6. Discussions and Limitations of the Current Study 4. Conclusions