BIBECHANA Vol. 21, No. 2, August 2024, 159-170 ISSN 2091-0762 (Print), 2382-5340 (Online) Journal homepage: http://nepjol.info/index.php/BIBECHANA Publisher: Dept. of Phys., Mahendra Morang A. M. Campus (Tribhuvan University) Biratnagar Comparative study of machine learning based prediction of supercapacitance performance of activated carbon prepared from bio-based materials Kirti Bir Rajguru1, Sujan Bhandari2, Ganesh Kumar Shrestha1, Chhabi Lal Gnawali1,∗, Bhadra Prasad Pokharel1 1Department of Applied Sciences and Chemical Engineering, Pulchowk Campus, Institute of Engineering (IOE), Tribhuvan University, Lalitpur, Nepal 2Central Department of Physics, Tribhuvan University, Kirtipur, Kathmandu, Nepal ∗Corresponding author: Email: chhabig123@ioe.edu.np Abstract The performance of electrochemical double-layer capacitors (EDLCs) is evaluated by the capacitance of activated carbon (AC) electrodes. The capacitance of AC electrodes is in- fluenced by many factors such as precursor type, activation method, pore structure, surface chemistry and electrolytic properties. In this paper, we present a comparative study of ma- chine learning based prediction of surface area, mesopore volume and total pore volume of activated carbon for energy storage applications. The ML models were trained on a dataset of synthetic data that were generated from the limited number of experimental data and which in- cluded the activation temperature, methylene blue number and iodine number of the activated carbon (AC). The best performing ML model was Random Forest and XGboost model. The results of this study can be used to optimize the production of activated carbon and improve its performance in energy storage applications. Keywords Machine Learning; Activated Carbon; Energy storage; Capacitance. Article information Manuscript received: January 31, 2024; Revised: April 18, 2024; Accepted: April 23, 2024 DOI https://doi.org/10.3126/bibechana.v21i2.62465 This work is licensed under the Creative Commons CC BY-NC License. https://creativecommons. org/licenses/by-nc/4.0/ 159 http://nepjol.info/index.php/BIBECHANA chhabig123@ioe.edu.np https://doi.org/10.3126/bibechana.v21i2.62465 https://creativecommons.org/licenses/by-nc/4.0/ https://creativecommons.org/licenses/by-nc/4.0/ Kirti Bir Rajguru et al./ BIBECHANA 21 (2024) 159-170 160 1 Introduction A data based method called the machine learning algorithm has been applied as a substitute tool to deal with a number of real-world problems. Creat- ing models and techniques that enable computers to learn from data and make predictions or judg- ments without being explicitly programmed is a key component of the artificial intelligence field of ma- chine learning [1, 2]. It is utilized in a variety of programs, including recommendation engines, nat- ural language processing, and identifying pictures. There are various types of machine learning such as unsupervised machine learning and supervised machine learning. In supervised machine learning, the algorithm develops predictions or classifications based on input data after learning from labeled training data. Unsupervised machine learning uses an unlabeled data collection as its foundation [3]. Supercapacitors, also known as electrochemical capacitors, are an intermediate form of energy stor- age between batteries and ordinary capacitors. Su- percapacitors can store more energy than regular capacitors due to their substantially higher energy density. Supercapacitors have received a lot of in- terest recently, due to its potential applications in a wide range of fields, such as consumer electron- ics and renewable energy systems. Whereas bat- teries, store energy chemically, supercapacitors do so electrostatically [4, 5]. Supercapacitors rely on the rapid movement of ions to store and release energy quickly. Larger surface area and appropri- ate pore volumes enhance the device’s energy and power density. Optimizing these parameters help in designing energy storage materials that can store more energy, charge and discharge faster, and has improved overall performance and durability in var- ious applications [6–9]. Fossil fuels can lead to energy crisis due to their finite supply and environmental impact. Transi- tioning to renewable energy source is essential to mitigate these issues. Activated carbon is highly porous form of carbon with a large surface area. Activated carbon (AC) is prepared from various carbonaceous material such as peat, wood, etc. In this study, the methylene blue number (MBN), io- dine number and temperature were utilized to pre- dict the surface area, mesopore volume, and total pore volume of produced activated carbon by using machine learning algorithm [10, 11]. A larger sur- face enhances the size of room for interaction be- tween the electrolyte and electrode material. More active places for electrochemical processes to take place can be made possible by this, increasing the amount of energy stored and accelerating the rate of charge/discharge. Larger pores called mesopores allow ions to pass through them and enter and exit the material [12]. A higher mesopore vol- ume ensures better ion accessibility and diffusion within the material, which improves the overall charge/discharge efficiency and, the device’s power performance. Total Pore Volume includes all sizes of pores, from microspores to mesopores. It is a measure of the overall storage capacity for ions [13, 14]. This research employs statistical and ex- perimental data, to assess the most suitable model for the prediction of more accurate dataset. The as- piration is to employ machine learning techniques which enhance comprehension of the interplay be- tween different aspects of activated carbon and sur- face area, mesopore volume and total pore volume. This, in turn, will streamline and enhance the pre- cision of experimental guidance for future experi- ments [15,16]. 1.1 Machine Learning Models There are different types of machine learning mod- els that can be used to predict the electrochemi- cal properties of activated carbon and its features, as they can capture the non-linearity in the data. Among all these models, we selected those that were most suitable for our data type so that we can pre- dict the electrochemical features of the AC with high degree of accuracy. Various studies like that of Ziang, Su, Wang, Donthula has been done in the use of machine learning for prediction of different types of properties of activated carbon [17]. 2 Materials and Methods 2.1 Experimental Terminalia chebula seed stones were used for the preparation of activated carbon(AC). Terminalia seeds were collected from the market and cleaned to remove any impurities. Phosphoric acid was used as an activating agent. Terminalia seeds powder was mixed in a phosphoric acid solution. The im- pregnation process ensured that the seeds powder absorbed the acid solution uniformly. The impreg- nated powder was subjected to tubular furnace, typically in the range of 400°C- 700°C. It helped in the formation of development of pores and the removal of the volatile components. The phospho- ric acid acted as a dehydrating agent and helped the carbon substance (activated carbon) to develop holes. To get rid of any remaining acids or other contaminants, the prepared activated carbon was thoroughly rinsed with distilled water. The synthe- sized activated carbon (AC) was analyzed by deter- mining its iodine number, methylene blue number, surface area, mesopore volume and total pore vol- ume [18–20]. We took eight data for each feature of methylene blue number, iodine number, surface area, mesopore volume, and total pore volume. Kirti Bir Rajguru et al./ BIBECHANA 21 (2024) 159-170 161 2.2 Computational 2.2.1 Data Collection Figure 1: Heatmap of regression values of different input and output features of AC. Figure 2: Box plot featuring range of input and out- put values of features of the AC. Figure 3: Histogram of all the inputs and output features. The data from the experimental method was used to define the range and variance of each of the re- sultant values obtained using formulas for specific capacitance, surface area, mesopore volume, and to- tal pore volume. The synthetic data was generated using the al- gorithm of producing random number but in range and variance obtained from the experimental data. The large number of synthetic data generation is crucial for machine learning as it helps to expand the dataset, making the model more robust and ca- pable of better generalization. This broader dataset can lead to improved training and more accurate predictions. Machine learning algorithms often re- quire a significant amount of diverse data to effec- tively learn patterns and relationships. The details of the generated data can be studied using the heat map, the box plot and histogram shown in figure 1, 2 and 3 respectively. By creating synthetic data that closely resem- bles the original experimental data, we provide the algorithm with more examples which will poten- tially enhance its performance. The large number of data were generated. To generate a normal random distribution value firstly, we limited the range of in- put and out values by using an algorithm function. Finally, the generated value is scaled to fit within the range as mentioned above. These techniques ensured that the synthetic data closely resembles our experimental data [21,22]. 2.2.2 Machine Learning Process Firstly, the data were checked for missing values and missing values were computed using mean. In the following process, firstly the importance of in- put features and their order of importance were noted for each output feature as shown in the fig- ure 3. Methylene blue number was the most im- portant parameter for the prediction of mesopore volume followed by temperature and iodine num- ber respectively. Finally, iodine number, methylene blue number and temperature in the same order are important for prediction of total pore volume. To ensure fair treatment of features, we applied scaling to bring all the features to a similar scale. Random Forest model was used to calculate fea- ture importance. The normalized reduction in the criterion brought by each feature was guided our selection process. A variety of machine learning algorithms were chosen based on their relevance to the task. The dataset was partitioned into training and validation sets in the ratio of 80:20. Subsequently, each algo- rithm underwent training using the training data, enabling them to grasp underlying patterns [23]. Analysis of the results provided insights into al- gorithm performance, guiding the refinement of the approach for improved outcomes in subsequent it- erations [24]. Kirti Bir Rajguru et al./ BIBECHANA 21 (2024) 159-170 162 Figure 4: Feature importance of input variables for output of ‘’Surface Area". Figure 5: Feature importance of input variables for output of Total Pore Volume. Figure 6: Feature importance of input variables for output of ‘ ’Mesopore Volume’ ’. 2.2.3 Proposed Model Selection There are several machine learning algorithms that are suitable for regression tasks. In this paper fol- lowing machine learning algorithms were described. Linear Regression A fundamental approach model connects be- tween input variable and a continuous output using a linear equation. Linear regression is a fundamen- tal statistical method used for modeling the rela- tionship between a dependent variable and one or more independent variables. The rationale behind linear regression is being simple and interpretable and can be used for robust regression and classifica- tion tasks. It assumes a linear relationship between the variables, attempting to find the best-fitting straight line (or hyperplane in higher dimensions) through the data points [25]. If the true relation- ship between the input feature and surface area is not linear, linear regression may not capture it. The goal of linear regression is to find the coeffi- cients that minimize the sum of squared differences between the predicted values and the actual values in the training data. This is often done by using techniques like Ordinary Least Squares (OLS) or gradient descent [26]. Ridge Regression Ridge regression, a regularization method in lin- ear regression, introduces a penalty term based on the L2 norm of coefficients to mitigate overfitting and multi-collinearity [27]. Lasso Regression Similar to Ridge, Lasso adds a penalty, but it tends to create simpler models by encouraging some feature contributions to become zero. Both Ridge and Lasso are used when dealing with multi- collinear data. Both Lasso and Ridge assume lin- earity and independence. Regularization helps to mitigate overfitting, but their performance depends on the true underlying relationship [28]. Decision Tree Regression This method dissects data into hierarchical de- cision structures to predict the target output. A decision tree is a widely used machine learning algorithm for both classification and regression tasks [29]. It is a tree-like model where an in- put is passed through a series of binary decisions based on the values of its features. Each deci- sion node in the tree represents a feature, and the branches from that node represents the possible feature values. The leaf nodes of the tree contain the predicted output or class. When a new input is given, it traverses the tree based on the feature values, following the decisions at each node, and finally arrives at a leaf node that provides the pre- dicted class (classification) or value (Regression). Decision Trees tend to over-fit noisy data and fea- ture importance may not accurately reflect the true Kirti Bir Rajguru et al./ BIBECHANA 21 (2024) 159-170 163 relevant features of data. Random Forest Regression An ensemble technique constructs multiple de- cision trees and averages their predictions, by en- hancing accuracy and avoiding overfitting. Ran- dom Forest is a machine learning algorithm used for both classification and regression tasks. It is an ensemble method that builds multiple decision trees during training and combines their outputs to make predictions. By introducing randomness in tree construction and feature selection, Random Forest reduces overfitting and improves general- ization. It is a powerful and versatile algorithm often used for complex datasets and can provide valuable insights into feature importance. Ran- dom Forests addresses decision tree limitations by averaging predictions across multiple trees. They handle non-linearites better but may still struggle with subtle relationships [30]. Support Vector Regression (SVR) Support Vector Regression is to used determine a regression line that best suits the data, allowing some flexibility around the line. Support Vector Regression (SVR) is a machine learning algorithm that is used for regression tasks. It is an exten- sion of Support Vector Machines (SVM) and is particularly useful while dealing with continuous or numeric target variables. SVR works by finding a hyperplane that best fits the data points which minimizes the error within a certain margin.SVR uses different kernel functions (such as linear, poly- nomial, radial basis function, etc.) to map the input data into a higher dimensional space, which can help to capture complex relationship between variables. SVR can model non-linear relationships effectively. The choice of kernel function impacts the performance [31]. Gradient Boosting Regression An ensemble strategy that sequentially builds weak learners, with each one rectifying the errors of its predecessor. Gradient boosting adapts well to complex patterns but might over fit if not tuned properly. Cat Boost Regression A gradient boosting algorithm proficient in handling categorical features; can be configured to run without displaying training progress. Cat Boost is a gradient boosting machine learning algo- rithm which particularly well suits for categorical features. It stands for "Categorical Boosting." similar to other gradient boosting methods, Cat Boost also builds an ensemble of decision trees to make predictions. Cat Boost also includes features like automatic handling of missing values, built-in cross validation, and the ability to monitor train- ing progress. It has gained popularity for its ease of use, efficiency, and competitive performance on various machine learning tasks [32]. XG Boost Regression Another robust gradient boosting algorithm de- signed to overcome limitations of traditional gradi- ent boosting. It is important to assess dataset size, feature complexity, interpretability, and computa- tional resources while choosing the most suitable algorithm for research paper. 2.2.4 Model Evaluation In this research, a comprehensive evaluation of the model’s performance is carried out to assess its gen- eralizability and stability. The technique of cross- validation is employed, which involves partitioning the dataset into subsets for both training and vali- dation purposes. This allows us to test the model’s ability to perform consistently across different sub- sets of data, ensuring that it can handle a vari- ety of scenarios effectively. To gauge the accuracy and reliability of the model, appropriate evaluation methods are applied. These methods could encom- pass metrics such as Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and possibly more domain specific metrics tailored to the nature of the dataset. These metrics collectively offer insights into the model’s predictive performance and its closeness to the ac- tual experimental data [33] A crucial aspect of this analysis involves a comparison between the actual experimental data and the model’s predicted values for specific characteristics like capacitance, surface area, mesopore volume, micro pore volume, and to- tal pore volume. By contrasting these values, re- searchers can quantify the model’s accuracy in pre- dicting intricate material properties. The degree of agreement between the predicted and actual val- ues will offer insights into the model’s capability to capture the underlying relationships within the data. In the context of Terminalia chebula seeds, the effectiveness of the model in predicting the ma- terial’s characteristics is a focal point of analysis. The findings obtained through rigorous evaluation methods shed light on how well the model can cap- Kirti Bir Rajguru et al./ BIBECHANA 21 (2024) 159-170 164 ture the nuanced features of the activated carbon derived from these seeds. Researchers will interpret the results to discern the strengths and limitations of the model in reproducing the complex properties of the material. This interpretation forms a critical part of the paper, as it guides the reader in under- standing the implications of the study’s outcomes and its broader implications for the field. CatBoost and XGBoost combine ensemble power with cat- egorical feature handling. They perform well but require tuning [34,35]. 2.2.5 Model Evaluation Parameters In the field of predictive modeling and statistical analysis, the selection of appropriate performance metrics is of paramount importance. This choice determines how effectively a model’s predictions are evaluated and compared against actual values. Four commonly employed metrics for this purpose are Mean Squared Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and R-squared [36]. Mean Squared Error Mean Squared Error is a widely utilized metric that quantifies the average squared differences be- tween the model’s predicted values and the actual observed values. By squaring the errors, MSE em- phasizes larger deviations and penalizes them more than smaller errors. This characteristic makes MSE sensitive to outliers and may lead to models that are overly focused on minimizing large errors [37]. The following formula is used to calculate the mean squared error: MSE = 1 n n∑ i (Yi − Ȳ 2) (1) were, MSE= Mean squared error n= number of data points Yi= Observed values Ŷi= Predicted values Root Mean Squared Error Root Mean Square Error is derived from MSE, RMSE is the square root of the average of squared errors. It has an advantage of presenting errors in the same unit as the target variable, making it more interpretable. RMSE offers insights into the spread or dispersion of prediction errors. While it retains the sensitivity to outliers, the square root trans- formation mitigates the influence of extreme val- ues [38, 39]. The formula to calculate root mean square error is given as: RMSE = √ MSE (2) Mean Absolute Error Mean Absolute Error calculates the average of the absolute differences between predicted and ac- tual values. Unlike MSE, it treats all errors equally, making it less sensitive to outliers. This character- istic makes MAE a favorable choice where outliers should not disproportionately affect the evaluation. However, it may lack the emphasis on larger errors that MSE provides. The following formula used to calculate the mean squared error: MAE = ∑n i (Yi − Ȳi) n (3) where, MAE=Mean absolute error yi=Prediction xi= True value n= Total number of data points R-squared R-squared measures the proportion of the vari- ance in the dependent variable that is explained by the independent variables in the model [40]. It of- fers insights into the goodness of fit and the model’s explanatory power. Higher R-squared values indi- cate a larger portion of the variability in the data is accounted by the model. Nevertheless, it is im- portant to note that R-squared can be misleading when applied to complex models, as it tends to in- crease with the addition of more variables, even if they are not meaningful [40]. The following formula used to calculate the R-squared: R2 = 1− SSR SST (4) While selecting the appropriate metric for eval- uating a model’s performance, researcher’s must consider the nature of the data, the goals of the analysis, and the desired balance between sensitiv- ity to errors and robustness to outliers [41]. In many cases, a combination of these metrics can pro- vide a more comprehensive assessment of a model’s strengths and weaknesses, contributing to a more thorough and nuanced interpretation of results. 3 Results and Discussion 3.1 Statistical analysis of datasets Table.1: depicts the entirety of the generated dataset. Within this dataset, the variables exhib- ited mesopore volume size ranging from 0.3 to 0.899 cm3/g and surface area from 800.09 to 1999.70 m2/g. To assess the degree of dispersion for each Kirti Bir Rajguru et al./ BIBECHANA 21 (2024) 159-170 165 variable, we employed mean values and their cor- responding standard deviations(SD). Additionally, the range of each variable was captured through the utilization of minimum and maximum values. Com- plementing these measures, quartile ranges were furnished to provide supplementary insights into the distribution’s central tendency. “SD" typically stands for "standard deviation," which measures the amount of variation or dispersion in a set of values. So, "SD mean" could refers to the stan- dard deviation of the mean, indicating how much the average value varies from the overall average. 3.2 Machine Learning Prediction 3.2.1 Model Evaluation Nine (9) machine learning models were employed based on the testing dataset. Table.2: displays the error values for various models, including support vector re gression, linear regression, ridge regres- sion, decision tree regression, and random forest re- gression, together with surface area, mesopore vol- ume and total pore volume. Table.2: Error values for the three outputs using different machine learn- ing models. 3.2.2 Comparative Analysis with Chemical Methods We compared ML predictions (e.g. using Random Forest, Gradient Boosting, etc.) with traditional methods. ML models outperform traditional meth- ods due to their ability to capture intricate pat- terns. ML models benefit from larger datasets, whereas traditional methods can work with smaller samples. ML implementation involved computa- tional costs, but traditional methods required ex- pensive equipment. ML models handled noisy data better, but traditional methods were robust even with limited data. 3.2.3 Hyperparameter tuning Hyperparameter tuning is the process of selecting the best combination of hyperparameters for a ma- chine learning model which optimizes its perfor- mance in a given task. Hyperparameters are parameters that are not learned by the model during training, but rather set before training begins, such as learning rate, regularization parameter, number of hidden layers, etc. The selection of appropriate hyperparameters can greatly impacts the performance of a machine learning model. For example, using a high learn- ing rate may cause the model to converge quickly but also result in unstable results or overshooting the optimal solution, while a low learning rate may cause the model to converge slowly but also result in more stable and accurate results. Similarly, se- lecting an appropriate value for regularization can help to prevent overfitting, while choosing the opti- mal number of hidden layers or neurons can impact the model’s ability to learn complex patterns in the data. Hyperparameter tuning involves searching through a range of hyperparameters using vari- ous techniques such as grid search, random search and Bayesian optimization to find the combina- tion that results in the performance on a valida- tion set. By optimizing the hyperparameters, the model’s accuracy and generalization ability can be improved [42]. Hyperparameter tuning in random forest is the process of finding the optimal values for the hyper- parameters of the model which maximize its per- formance on a given dataset. Hyperparameters are values set before training the model that control its behavior, such as the number of trees in the forest, maximum depth of each tree, and minimum number of samples required to split a node. Kirti Bir Rajguru et al./ BIBECHANA 21 (2024) 159-170 166 Kirti Bir Rajguru et al./ BIBECHANA 21 (2024) 159-170 167 Hyperparameter tuning in random forest is the process of finding the optimal values for the hyper- parameters of the model that maximize its perfor- mance on a given dataset. Hyperparameters are values set before training the model that control its behavior, such as the number of trees in the forest, the maximum depth of each tree, and the minimum number of samples required to split a node. For each of the model we have calculated the evaluation metrics like MAE, MSE etc. and also the best parameters so that we can get the best pathway for each model. 4 Conclusion Fossil fuels contribute to air pollution, greenhouse gas emissions, and environmental degradation. Su- percapacitors can help to reduce these effects by providing efficient energy storage and recovery, es- pecially in applications like electric vehicles and re- newable energy systems. They can capture and re- lease energy quickly, minimizing the need for con- stant fossil fuel consumption. Using supercapac- itors can lead to cleaner air and reduced carbon emissions, promoting a more sustainable energy Kirti Bir Rajguru et al./ BIBECHANA 21 (2024) 159-170 168 landscape. A model that predicts the surface area, mesopore volume and total pore volume of bio- based activated carbon was created using various machine learning algorithms. The machine uses the input features namely temperature, methylene blue number and iodine number, and the test cir- cumstances. The prediction performance of the Random Forest Regressor and cat Boost Regressor models was superior than that of the ANN, Ridge, MLP, GBR, DTR and SVR models. The multi- faceted correlations between input parameters such iodine amount and Methylene blue number and out- put variables include surface area, mesopore vol- ume, and total pore volume have been studied us- ing machine learning (ML). In addition, ML might be used for generating models for prediction that includes surface area, mesopore volume, and total pore volume. By using surface area, mesopore vol- ume, and total pore volume, machine learning al- gorithms can assist researchers in choosing the best model and in enhancing the efficacy and reliability of the energy storage process. Machine Learning (ML) is used to analyze the effects of the input parameters on the basis of value of surface area, mesopore volume, and total pore volume, as well as to optimize the required out- put parameters. We located Random Forest Re- gressor and XG Boost Regressor models, among nine different ML methods, are the most suitable models. Since, produced activated carbon possesses substantial values for surface area, mesopore vol- ume, and total pore volume, it can be employed as an electrode material for energy storage in future. Acknowledgement The authors would like to express their gratitude to the Department of Applied Sciences and Chemical Engineering (IOE), Pulchowk Campus. References [1] M. Rahimi, M.H. Abbaspour-Fard, and A. Ro- hani. Synergetic effect of n/o functional groups and microstructures of activated carbon on su- percapacitor performance by machine learning. Journal of Power Sources, 521:230968, 2022. [2] C. Wang, W. Jiang, G. Jiang, T. Zhang, K. He, L. Mu, J. Zhu, D. Huang, H. Qian, and X. Lu. Machine learning prediction of the yield and bet area of activated carbon quantitatively re- lating to biomass compositions and operating conditions. Industrial & Engineering Chem- istry Research, 62:11016–11031, 2023. [3] A.G. Saad, A. Emad-Eldeen, W.Z. Tawfik, and A.G. El-Deen. Data-driven machine learn- ing approach for predicting the capacitance of graphene-based supercapacitor electrodes. Journal of Energy Storage, 55:105411, 2022. [4] S. Jha, S. Bandyopadhyay, S. Mehta, M. Yen, T. Chagouri, E. Palmer, and H. Liang. Data-driven predictive electrochemical behav- ior of lignin-based supercapacitors via machine learning. Energy & Fuels, 36:1052–1062, 2021. [5] H. Su, S. Lin, S. Deng, C. Lian, Y. Shang, and H. Liu. Predicting the capacitance of carbon- based electric double layer capacitors by ma- chine learning. Nanoscale Advances, 1:2162– 2166, 2019. [6] J. Abdi, T. Pirhoushyaran, F. Hadavimoghad- dam, S.A. Madani, A. Hemmati-Sarapardeh, and S.H. Esmaeili-Faraj. Modeling of capaci- tance for carbon-based supercapacitors using super learner algorithm. Journal of Energy Storage, 66:107376, 2023. [7] C.L. Gnawali, S. Shahi, S. Manandhar, G.K. Shrestha, M.P. Adhikari, R. Rajbhandari, and B.P. Pokharel. Porous activated car- bon materials from triphala seed stones for high-performance supercapacitor applications. BIBECHANA, 20(1):10–20, 2023. [8] R. Dubey and V. Guruviah. Machine learn- ing enabled performance prediction of biomass- derived electrodes for asymmetric supercapac- itor. In International Conference on Futuris- tic Communication and Network Technologies, volume 995, pages 453–460, 2021. [9] L. Leng, L. Yang, X. Lei, W. Zhang, Z. Ai, Z. Yang, H. Zhan, J. Yang, X. Yuan, and H. Peng. Machine learning predicting and en- gineering the yield, n content, and specific sur- face area of biochar derived from pyrolysis of biomass. Biochar, 4:63, 2022. [10] C.L. Gnawali, L.K. Shrestha, J.P. Hill, R. Ma, K. Ariga, M.P. Adhikari, R. Rajbhandari, and B.P. Pokharel. Nanoporous activated carbon material from terminalia chebula seed for su- percapacitor application. Journal of Carbon Research, 9:109, 2023. [11] H. Thanh, S.E. Taremsari, B. Ranjbar, H. Mashhadi-Moslem, E. Rahimi, M. Rahimi, and A. Elkamel. Hydrogen storage on porous carbon adsorbents: Rediscovery by nature- derived algorithms in random forest machine learning model. Energies, 16:2348, 2023. [12] S. Zhu, J. Li, L. Ma, C. He, E. Liu, F. He, C. Shi, and N. Zhao. Artificial neural net- work enabled capacitance prediction for car- bon based supercapacitors. Materials Letters, 233:294–297, 2018. Kirti Bir Rajguru et al./ BIBECHANA 21 (2024) 159-170 169 [13] J. Wang, Z. Li, S. Yan, X. Yu, Y. Ma, and L. Ma. Modifying the microstructure of algae- based active carbon and modelling supercapac- itors using artificial neural networks. RSC Ad- vances, 9:14797–14808, 2019. [14] X. Yang, C. Yuan, S. He, D. Jiang, B. Cao, and S. Wang. Machine learning prediction of specific capacitance in biomass derived carbon materials: Effects of activation and biochar characteristics. Fuel, 331:125718, 2023. [15] Z. Yang, Y. Lin, X. Gu, and X. Liang. Pre- diction and optimization model of activated carbon double layer capacitors based on im- proved heuristic approach genetic algorithm neural network. Engineering Computations, 35:1625–1638, 2018. [16] M. Rahimi, M.H. Abbaspour-Fard, and A. Ro- hani. Machine learning approaches to rediscov- ery and optimization of hydrogen storage on porous bio-derived carbon. Journal of Cleaner Production, 329:129714, 2021. [17] K. Donthula, N. Thota, S.B. Anne, and M. Kakunuri. Prediction of capacitance using artificial neural networks for carbon nanofiber- based supercapacitors. In Computer Aided Chemical Engineering, volume 52, pages 1015– 1020, 2023. [18] A. Albalasmeh, M.A. Gharaibeh, O. Mo- hawesh, M. Alajlouni, M. Quzaih, M. Masad, and A. El Hanandeh. Characterization and ar- tificial neural networks modelling of methylene blue adsorption of biochar derived from agri- cultural residues: Effect of biomass type, py- rolysis temperature, particle size. Journal of Saudi Chemical Society, 24:811–823, 2020. [19] S. Joshi, R.G. Shrestha, R.R. Pradhananga, K. Ariga, and L.K. Shrestha. High surface area nanoporous activated carbons materials from areca catechu nut with excellent iodine and methylene blue adsorption. C, 8:2, 2021. [20] S. Joshi and B.P. Pokharel. Preparation and characterization of activated carbon from lapsi (choerospondias axillaris) seed stone by chemi- cal activation with potassium hydroxide. Jour- nal of the Institute of Engineering, 9:79–88, 2013. [21] C. L. Gnawali, S. Manandhar, S. Shahi, R. G. Shrestha, M. P. Adhikari, R. Rajbhandari, B. P. Pokharel, R. Ma, K. Ariga, and L. K. Shrestha. Nanoporous carbons materials from terminalia bellirica seed for iodine and methy- lene blue adsorption and high-performance supercapacitor applications. Bulletin of the Chemical Society of Japan, 96:572–581, 2023. [22] L. K. Shrestha, S. Shahi, C. L. Gnawali, M. P. Adhikari, R. Rajbhandari, B. P. Pokharel, R. Ma, R. G. Shrestha, and K. Ariga. Phyllan- thus emblica seed-derived hierarchically porous carbon materials for high-performance super- capacitor applications. Materials, 15:8335, 2022. [23] S. Bhandari, K. B. Rajguru, C. L. Gnawali, and B. P. Pokharel. Influence of precursor type on activated carbon prepared by phosphoric acid-chemical activation for supercapacitor ap- plications. Journal of Nepal Physical Society, 9(3):12–17, 2023. [24] S. Mishra, R. Srivastava, A. Muhammad, A. Amit, E. Chiavazzo, M. Fasano, and P. Asi- nari. The impact of physicochemical features of carbon electrodes on the capacitive perfor- mance of supercapacitors: a machine learning approach. Scientific Reports, 13:6494, 2023. [25] K. Zhang, S. Zhong, and H. Zhang. Predict- ing aqueous adsorption of organic compounds onto biochars, carbon nanotubes, granular ac- tivated carbons, and resins with machine learn- ing. Environmental Science & Technology, 54:7008–7018, 2020. [26] M. Khan, Z. Ullah, O. Mašek, S. R. Naqvi, and M. N. A. Khan. Artificial neural networks for the prediction of biochar yield: a comparative study of metaheuristic algorithms. BioSource Technology, 355:127215, 2022. [27] A. Hai, G. Bharath, M. F. A. Patah, W. M. A. W. Daud, K. Rambu, P. Show, and F. Ba- nat. Machine learning models for the predic- tion of total yield and specific surface area of bio-char derived from agricultural biomass by pyrolysis. Environmental Technology & Inno- vation, 30:103071, 2023. [28] Z. U. Haq, H. Ullah, M. N. A. Khan, S. R. Naqvi, A. Ahad, and N. A. S. Amin. Com- parative study of machine learning methods integrated with genetic algorithm and parti- cle swarm optimization for bio-char yield pre- diction. BioResource Technology, 363:128008, 2022. [29] J. A. Okolie, S. Savage, C. C. Ogbaga, and B. Gunes. Assessing the potential of machine learning methods to study the removal of phar- maceuticals from wastewater using biochar or activated carbon. Total Environment Research Themes, 1:100001, 2022. [30] S. Nanda, S. Ghosh, and T. Thomas. Ma- chine learning aided cyclic stability prediction for supercapacitors. Journal of Power Sources, 546:231975, 2022. Kirti Bir Rajguru et al./ BIBECHANA 21 (2024) 159-170 170 [31] Y. W. Liew, S. K. Arumugasamy, and A. Sel- varajoo. Potential of biochar as soil amend- ment: prediction of elemental ratios from py- rolysis of agriculture biomass using artificial neural network. Water, Air, & Soil Pollution, 233:54, 2022. [32] A. Shafizadeh, H. Shahbeik, S. Rafiee, A. Moradi, M. Shahbaz, M. Madadi, C. Li, W. Peng, M. Tabatabaei, and M. Aghbashlo. Machine learning-based characterization of hy- drochar from biomass: Implications for sus- tainable energy and material production. Fuel, 347:128467, 2023. [33] J. Wang, X. Zhang, Z. Li, Y. Ma, and L. Ma. Recent progress of biomass-derived carbon ma- terials for supercapacitors. Journal of Power Sources, 451:227794, 2020. [34] K. Rahmani, A. H. Mamaghani, Z. Hashisho, D. Crompton, and J. E. Anderson. Prediction of heel build-up on activated carbon using ma- chine learning. Journal of Hazardous Materi- als, 433:128747, 2022. [35] W. Jiang, X. Xing, S. Li, X. Zhang, and W. Wang. Synthesis, characterization and ma- chine learning based performance prediction of straw activated carbon. Journal of Cleaner Production, 212:1210–1223, 2019. [36] S. K. Arumugasamy, A. Selvarajoo, and M. A. Tariq. Artificial neural networks modelling: gasification behaviour of palm fibre biochar. Materials Science for Energy Technologies, 3:868–878, 2020. [37] L. Wu, Z. Xu, Z. Wang, Z. Chen, Z. Huang, C. Peng, X. Pei, X. Li, J. P. Mailoa, C. Y. Hsieh, and T. Wu. Machine learning acceler- ated carbon neutrality research using big data from predictive models to interatomic poten- tials. Science China Technological Sciences, 65(10):2274–2296, 2022. [38] D. Tena-Gag, I. Golcarenarenji, I. Martinez- Alpiste, Q. Wang, and J. M. Alcaraz-Calero. Machine-learning-based carbon dioxide con- centration prediction for hybrid vehicles. Sen- sors, 23(3):1350, 2023. [39] M. Mądziel, A. Jaworski, H. Kuszewski, P. Woś, T. Campisi, and K. Lew. The devel- opment of co2 instantaneous emission model of full hybrid vehicle with the use of machine learning techniques. Energies, 15:142, 2021. [40] J. Seo, B. Yun, J. Park, J. Park, M. Shin, and S. Park. Prediction of instantaneous real-world emissions from diesel light-duty vehicles based on an integrated artificial neural network and vehicle dynamics model. Science of the Total Environment, 786:147359, 2021. [41] S. Ghosh, G. R. Rao, and T. Thomas. Ma- chine learning-based prediction of supercapac- itor performance for a novel electrode material: Cerium oxynitride. Energy Storage Materials, 40:426–438, 2021. [42] Y. Y. Chia, L. H. Lee, S. Shafiabady, and D. Isa. A load predictive energy management system for supercapacitor-battery hybrid en- ergy storage system in solar application using the support vector machine. Applied Energy, 137:588–602, 2015. Introduction Machine Learning Models Materials and Methods Experimental Computational Data Collection Machine learning Process Proposed Model Selection Model Evaluation Model Evaluation Parameters Results and Discussion Statistical analysis of datasets Machine Learning Prediction Model Evaluation Comparative Analysis with Chemical Methods Hyperparameter tuning Conclusion