Unlocking the Full Potential of Heteroatom-Doped Graphene-Based Supercapacitors through Stacking Models and SHAP-Guided Optimization

Abstract

Graphene-based supercapacitors have emerged as a promising candidate for energy storage due to their superior capacitive properties. Heteroatom-doping is a method of improving the capacitive properties of graphene-based electrodes, but the optimal doping conditions and electrochemical properties are not yet fully understood due to the synergistic effects that occur. Many parameters, such as doping content, defects, specific surface area (SA), electrolyte, and more, could affect the capacitance (CAP). In this study, we use machine learning to solve these critical issues. We applied many models, such as Light Gradient Boost Machine, Extreme Gradient Boost, Polynomial Regression, Neural Network, Elastic Net, Lasso Regression, Ridge Regression, Random Forest, Support Vector Machine, K-Nearest Neighbors, Gradient Boost, AdaBoost, and Decision Tree, to find a suitable model for CAP prediction. Moreover, we enhance the prediction result by taking advantage of the top candidate model and creating a stacking concept (called “stacking models”). The SHAP value was used to identify the range of properties that affect CAP, and it was discussed in detail. Our results suggest that high-CAP graphene supercapacitors should have a large SA, with 4–5% nitrogen, 10–15% oxygen, high percentages of sulfur, a defect ratio close to 1, with acid electrolyte, and a low current density. These findings, along with the developed model and code, are expected to serve as a valuable computational tool for future electrochemical research from fundamental to applications.

Liquid Phase Exfoliation of 2D Materials and Its Electrochemical Applications in the Data-Driven Future

“…More importantly, the optimum doping condition can be identified through scattered interpolation-extrapolation which suggest the percentage of heteroatom-doping on graphene as single, double, and triple atoms doping with oxygen, nitrogen, and sulfur as the main atomic species. Deshsorn et al 268 have utilized an ensemble technique so-called "stacking" to combine the benefits of multiple machine learning algorithms from two families which are regression-based and tree-based algorithms. This strategy tends to increase the prediction accuracy and generalization of the final model over the application of the standalone model.…”

Empowering Capacitive Devices: Harnessing Transfer Learning for Enhanced Data-Driven Optimization

“…29 The remarkable success of TL has been shown in fields such as materials informatics, [30][31][32][33][34][35][36][37] and the design of electrochemical systems. [38][39][40][41] In the study by Deshsorn et al, 23 out of 620 rows, only 159 contain complete data, while the remaining 461 samples with missing data can be considered the source domain for pretraining a learning agent to predict the specific capacitance of a capacitor. The target domain and learning task involve the 159 rows with no missing data, which can be used to fine-tune the pre-trained model and mitigate potential bias towards imputed data.…”

“…Notably, activated carbon's high specific surface area, electrical conductivity, and tunable pore structure have significantly improved CDI's salt adsorption efficiency, 19,20 while features such doping with heteroatoms has markedly increased supercapacitor capacitance depending on material type. [21][22][23] These findings underscore the importance of investigating factors such as specific surface area, dopant concentration, current density (or voltage), and electrolyte concentration to further enhance device performance.In the quest to optimize the capacitance (CAP) of supercapacitors (SC) or the salt adsorption capacity (SAC) of capacitive deionization (CDI) devices, machine learning (ML) techniques can be instrumental in determining the ideal range of attributes. However, ML models often require large datasets, and researchers frequently face the challenge of limited or incomplete data, which hampers the development of successful data-driven models.…”

“…To address this, researchers have employed strategies such as imputation with ML or data augmentation. For instance, Deshsorn et al 23 used KNN imputation on a dataset of 513 records, including specific surface area, defect ratio, and dopant percentages, to enhance their stacked ML model, achieving an absolute mean error of 14.3 F/g. Likewise, Saffarimiandoab et al 19 developed an ML model to predict the SAC of CDI devices with an RMSE of 2.271 mg/g using an imputed dataset featuring variables like applied voltage, electrolyte concentration, and electrode pore characteristics.…”

“…Notably, activated carbon's high specific surface area, electrical conductivity, and tunable pore structure have significantly improved CDI's salt adsorption efficiency, 19,20 while features such doping with heteroatoms has markedly increased supercapacitor capacitance depending on material type. [21][22][23] These findings underscore the importance of investigating factors such as specific surface area, dopant concentration, current density (or voltage), and electrolyte concentration to further enhance device performance.…”