Machine learning approach to understanding the ‘synergistic’ pseudocapacitive effects of heteroatom doped graphene

Abstract

In recent years, graphene has been widely utilised as a supercapacitor electrode, and doping heteroatom on graphene is reported to enhance the pseudocapacitance of the electrode materials significantly resulting in a high energy density. However, the relationship and charge storage mechanism of a so-called “synergistic effect” between those doped atoms including oxygen-, nitrogen-, and sulphur-doping on supercapacitor performances remain inscrutable. In this study, a machine learning model – artificial neural network (ANN) is used to predict the capacitance of heteroatom-doped graphene-based supercapacitors and establish the effects of heteroatom-doping. Trained ANN can accurately predict the capacitance of the electrode, drawing the best synthesis conditions for the heteroatom doped graphene. Furthermore, we successfully demonstrate the synergistic effect that arises from co-doping nitrogen, sulphur, and locate the optimised region for N/S-co-doping with high capacitance, and high retention rate. Machine learning methods allow us to consider a much larger space of heteroatom-doping combinations to maximise the supercapacitor performances and provide a useful guideline for co-doping graphene-based supercapacitors.

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

“…Unlike nitrogen, the effect of oxygen is different because most of the doped oxygen species exhibit non-conductive properties such as epoxide, ether, and carbonyl, which can reduce the CAP properties. However, the quinone group is an active oxygen species that exhibits a great redox contribution as it can transform into hydroquinone while accepting a proton; unfortunately, this cannot be controlled. Typically, oxygen in graphene (also refer to reduced graphene oxide) is derived from the oxygen that remains after the thermal or chemical reduction of graphene oxide.…”

“…Knowing the optimal doping condition is crucial to enhance the capacitive properties of graphene-based supercapacitors and can reduce experimental time, this could be done by employing data analysis and machine learning. 18,19 This approach can avoid the random synthesis procedure, which results in a randomly doped content without knowing the final CAP properties. Hence, the understanding of graphene doping can be done within a single click instead of performing a traditional synthesis method, which require almost a week.…”

“…59 In fact, a high R 2 value can be achieved by optimizing the hyperparameters of the applied model; therefore, K-fold cross-validation should be included when performing machine learning. 18 Even though neural networks show excellent prediction in terms of R 2 and RMSE, it is important to note that in some cases, the maximum error of the train, test, and validation set can be found, which can affect the prediction results.…”

“…The contribution in each of the nitrogen species was reported previously in our work using machine learning techniques. 18 In Figure 7a, the SHAP versus percentage of doped nitrogen content is displayed. The positive impact is mostly observed when doping with nitrogen species, which is consistent with experimental findings showing that pristine graphene exhibits poor CAP.…”

“…Doping sulfur in appropriate amounts has been demonstrated to have a significant positive impact on the predicted CAP (see Figure 7c), primarily due to the faradaic process of sulfone. 18 It is noteworthy that the maximum reported doping content for sulfur is approximately 25 percent, although it should not be concluded that doping with pure sulfur is inherently superior to other elements, such as nitrogen. The available data suggest that sulfur doping is frequently accompanied by co-doping with other elements.…”