The design and optimization of graphene electrode for high-performance energy storage

Research in 2D materials for electrochemical applications has progressed significantly in recent years, causing a large number of literatures to be published. Coupled with machine learning, more and more complex problems are being surfaced each year. To aid in the gathering and simplification of information, data mining was utilized to digest the large quantify of literature information (e.g., abstract, publication date, research origin) and to summarize the trends, which is reported in the literature review section. After which, the author focuses on optimizing and designing supercapacitors and their electrodes as their main field of interest. From the literature review, it was discovered that supercapacitors have limitations in their surface area to store charge. Fortunately, doping hetero-atoms into the structure of graphene has a positive effect on capacitance. However, the quantity and interaction of hetero-atom doping on graphene supercapacitors is still not fully understood. Therefore, data science and machine learning were utilized by the research team to discover the hidden interactions and most optimized quantity of each parameter for a high-capacitance heteroatom doped graphene supercapacitor. Data science techniques that were utilized were Pearson correlation, 2D scatter plots, 3D scatter plots, bar plots, and contour plots. In the side of machine learning, techniques such as Shapley Additive Explanations (SHAP), Feature Permutation Importance (FPI), and Partial Dependence Plots (PDP) were used to find the interactions of variables based on machine learning models such as Gradient Boost (GB), Light Gradient Boost (LGBM), Extreme Gradient Boost (XGB), Polynomial Regression (POLY), Neural Networks (NN), Elastic Net (ELAS), Lasso Regression (LASS), Ridge Regression (RIDGE), Random Forest (RF), Support Vector Machine (SVM), K-nearest Neighbors (KNN), Adaboost (ADA), Decision Tree (DT), and a Stacking Model (STACK) that was developed to utilize all the advantages of different models. The results of the STACK model are trustworthy, as the STACK model is not as easily affected by “noisy” or “falsified” data compared to other models as the research team has tested. Additionally, uncertainty from base model order of the STACK model is reduced due to utilizing novel approaches developed by the research team. Overall, the final STACK model is proven to be robust and certain in prediction accuracy and generalization. It was then trained on hetero-atom doped carbon data to obtain the feature importance. From analysis, it was discovered that surface area, %N, %S, and %B should be high, current density and electrolyte concentration (varies on type of electrolyte) should be low, %O should be in the range of 10% to 20%, and %P should be roughly 2.5%. Additionally, it was observed that doping with fewer elements (%N and %O only) may lead to more beneficial effects than doping with too many elements (5 elements). Lastly, ranking the importance of each feature yields the top 3 candidates as follows: %N >SA >CONC.