Carbon 214 (2023) 118342 Contents lists available at ScienceDirect Carbon journal homepage: www.elsevier.com/locate/carbon Machine learning assisted chemical characterization to investigate the temperature-dependent supercapacitance using Co-rGO electrodes Xiaoyu Liu a, 1, Dali Ji a, **, 1, Xiaoheng Jin a, Vanesa Quintano a, b, Rakesh Joshi a, * a School of Materials Science and Engineering, University of New South Wales, Sydney, NSW, 2052, Australia b Catalan Institute of Nanoscience and Nanotechnology (ICN2), CSIC and BIST, Campus UAB, Bellaterra, 08193, Barcelona, Spain A R T I C L E I N F O A B S T R A C T Keywords: Graphene oxide (GO) intercalated with transition metal oxides (TMOs) has been investigated for optimal Graphene oxide supercapacitance performance. However, attaining the best performance requires conducting numerous exper­ Supercapacitance iments to find an optimal material composition. This raises an important question; can resource consumption Random forest associated with extensive experiments be minimized? Here, we combine the machine learning (ML)-based Machine learning random forest (RF) model with experimentally observed X-ray photoelectron spectroscopy (XPS) data to construct the complete chemical analysis dataset of Co(III)/Co(II) ratio for thermally synthesized Co-rGO supercapacitor electrodes. The ML predicted dataset could be further coupled with other experiment results, such as cyclic voltammetry (CV), to establish a precise model for predicting capacitance, with ML coefficient of determination (R2) value of 0.9655 and mean square error value of 6.77. Furthermore, the error between pre­ dicted capacitance and experimental validation is found to be less than 8%. Our work indicates that RF can be used to predict XPS data for the TMO-GO system, thereby reducing experimental resource consumption for materials analysis. Moreover, the RF-predicted result can be further utilized in experimental and computational analysis. 1. Introduction materials presents significant challenges. Due to various synthesis con­ ditions, materials containing different structures and compositions The development of sustainable and renewable energy storage sys­ exceed the scope of exhaustive methods [16,17]. The synergistic in­ tems has gained significant momentum in recent years, leading to the teractions among components contribute to the complexity of the sys­ emergence of non-conventional energy devices such as supercapacitors. tem, limiting the ability to predict results intuitively [18,19]. It is Graphene-based supercapacitors have emerged as promising candidates technically infeasible to rely solely on experiments to characterize the for next-generation energy storage technology because of their high material and predict the capacitance under all experimental conditions, specific area, fast-charging capability, long-life cycle, and low mainte­ considering prediction accuracy, time efficiency, and cost-effectiveness nance cost [1–5]. Among these, the combination of graphene oxide (GO) [16,17]. Hence, utilizing data-driven methods like machine learning with transition metal oxides (TMOs-GO) has garnered significant (ML) to accelerate the prediction process is necessary [20]. Multiple ML attention in recent years [6–9]. Notably, the integration of cobalt oxide studies have exhibited that it can predict capacitance, which still relies (Co oxide) and reduced graphene oxide (rGO) electrodes have demon­ heavily on the availability of extensive experimental data. It requires strated remarkable potential [10]. The exceptional properties of rGO, collecting experimental conditions and capacitance information from including high surface area [11] and facile functionalization [12], available sources in the literature. However, the data collection method coupled with the benefits of Co oxide, such as the introduction of active may pose errors in the analysis [16]. Employing ML not only for pre­ sites [13,14], improved conductivity [15], and widened voltage win­ dicting capacitance but also for assisting material characterization be­ dow, make this hybrid electrode a highly desirable choice. comes particularly crucial to minimise errors in data analysis. Nevertheless, the investigation of these TMOs-GO supercapacitor Recently, ML-assisted material characterization encompasses * Corresponding author. ** Corresponding author. E-mail addresses: dali.ji@unsw.edu.au (D. Ji), r.joshi@unsw.edu.au (R. Joshi). 1 Contributed equally to this work. https://doi.org/10.1016/j.carbon.2023.118342 Received 5 June 2023; Received in revised form 28 July 2023; Accepted 29 July 2023 Available online 30 July 2023 0008-6223/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by- nc-nd/4.0/). X. Liu et al. Carbon 214 (2023) 118342 microscopy enhancement and spectral analysis, with the latter garnering for sample preparation. The output data are corresponding Co(III)/Co significant attention as an important area of research [21,22]. A com­ (II) ratio obtained from Gaussian fitting of experimentally obtained XPS mon approach in ML-assisted characterization is utilizing existing data data. Random forest models (RF) leverage the collective decision- and establishing models for predicting or screening new datasets [23]. making of multiple individual decision trees to provide accurate pre­ Recent studies have predominantly focussed on ML-assisted Raman dictions and handle complex data patterns [29]. It is regarded as a spectroscopy, aiming to predict parameters such as graphene-based versatile and powerful ML algorithm due to its high prediction accuracy, materials’ thickness, twist angle, and defects [24–26]. Notably, these robustness to noise as well as the outliers [30,31]. The training and test studies employed an approach that predominantly relies on random data of the RF-assisted XPS model were collected from the experimental forest models (RF) [24–26]. RF can achieve a high level of predictive XPS data of our samples. The training set and test set have a ratio of 80% performance with a relatively low computational burden. Moreover, the and 20%, respectively. RF model is also more suitable for predicting with small datasets. However, there has been limited exploration of utilizing RF models to 2.2.2. RF-assisted XPS model establishment assist other material’s characterizations. This prompts the question of After data collection, we used Python to perform modelling on the whether these models can be employed in predicting the performance of PyCharm software. All screened data were initially accepted impartially graphene-based materials, such as capacitance. without any initial judgment or bias regarding the data validity. The In this work, we systematically investigated the relationship among descriptors data were broadly categorized into three categories based on specific capacitance, chemical composition and the synthesis tempera­ the (1) mass ratio of doping transition oxide, (2) experimental condi­ ture for Co-rGO. We have used a thermal synthesis method for this study tions including temperature and corresponding Co(III)/Co(II) ratio. [27]. Previously it has been reported that increasing temperature from All the algorithm-specific hyperparameters were tuned via the grid 150 to 500 ◦ C would increase surface area and enhance conductivity due search, which involves splitting the hyperparameter space into user- to a higher C/O ratio in rGO, weakening the capacitance performance defined grids, from where the optimal hyperparameter values are [28]. We combined for the first time the machine learning random forest searched. The estimators were set as 150. We set max depth and min (RF) model with X-ray photoelectron spectroscopy (XPS) data to accu­ split as variable values to optimise the results until the results are rately predict the Co(III)/Co(II) ratio in Co-rGO system synthesized at acceptable. The max depth refers to the maximum depth allowed for different temperatures with a remarkable accuracy of 99.9%. Further­ decision tree branches in a random forest model. The min split refers to more, based on the RF-assisted XPS results and relevant experimental the minimum number of splits allowed during the construction of a data, we established RF model to predict the Co-rGO capacitance. The decision tree. To enable effective hyperparameter tuning and avoid model exhibited a R2 value of 96.5% and an error (between predicted overfitting simultaneously, the k-fold cross-validation method was used. and experiment data) within 8%. Using these two models, we inferred Considering the dataset size, the K value was set to 10 to prevent that the Co(III)/Co(II) ratio of Co in Co-rGO synthesized through ther­ overfitting. Further hyperparameter detail set for this study can be found mal methods consistently increased within the temperature range of in Supplementary SI 2 Sheet S5. 200–600 ◦ C. Our work demonstrates that ML, particularly the RF model, In this work, two performance functions were applied to analyse the can be utilized for predicting XPS results, thereby reducing the con­ accuracy of output, including R2 (R squared), MSE (Mean Square Error) sumption of actual XPS measurements. The ML-assisted XPS result and MAE (Mean Absolute Error). The functions are described below predictions can be incorporated into ML models for predictions of ma­ [32]: terials properties. ∑( (i) )2 ŷ − y(i) 2. Materials and methods 2 i R = 1 − ∑ (i) 2 (1) (ȳ − y(i) ) i 2.1. Preparation of GO/Co electrodes 1 ∑( (i) )2 MSE = y − ŷ(i) (2) Graphene oxide (purchased from NiSiNa Materials Japan Co., Ltd.) m i ethanol solution (2 mg/mL) was mixed with CoCl2-ethanol solution (0.1 ⃒ mg/mL). The CoCl2 amount was appropriately varied in the host GO 1 ∑N ⃒ ⃒ MAE = |yi − ŷi ⃒ (3) solution to achieve the compositions 2 wt%, 4 wt%, 6 wt%, 8 wt%, 10 wt N i=0 ⃒ %, 12 wt%, 14 wt%, 16 wt% in GO/CoCl2 mixture. An appropriate quantity (i.e., 5 μL) of this mixture (with different mass ratio) was drop where ŷ(i) is the predicted value; y(i) is real value; and ȳ(i) indicates the casted on carbon fibre paper followed by heating in muffle furnace for average of real values. duration of 2 h. The samples were named as X-Co-rGO according to heating temperatures (200, 300, 400, 500, 600 ◦ C), where X is the 2.2.3. Other material characterization temperature. X-ray diffraction (XRD, PANalytical Empyrean II thin-film) with Co Kα radiation (λ = 0.154 nm) in the range 2θ = 18–30◦ and 33–64◦ , with 2.2. Characterization a scan speed of 0.11◦ /s and step size of 0.05◦ was used for crystal structure determination. Raman spectroscopy (Renishaw inVia Raman 2.2.1. RF-assisted XPS data collection Microscope, 532 nm excitation laser) was also utilized for structural X-ray photoelectron spectroscopy (XPS, Thermo Scientific, UK, analysis. The surface morphology was studied using scanning electron ESCALAB250Xi) was used to characterize the valency and ratio of Co microscopy (FEI Nova NanoSEM 230 FE-SEM) with energy dispersive X- (III) to Co(II). It applied 100 eV for survey scans and 20 eV for region Ray spectroscopy (EDS) (shown in the supplementary information SI 1). scans. In our experiments the C1s = 284.5 eV for graphite act as a reference. We selected a subset of Co-rGO samples for XPS experiments based on different synthesis conditions. These conditions include 2.3. Electrochemical measurement 200 ◦ C: 2 wt%, 4 wt%, 6 wt%, 8 wt%, 10 wt%, 12 wt%, 14 wt%, 16 wt%; 300 ◦ C: 4 wt%, 8 wt%, 16 wt%; 400 ◦ C: 16 wt%; 500 ◦ C: 16 wt%; 600 ◦ C: All electrochemical measurements were conducted at ambient tem­ 16 wt% (shown in SI 2 Sheet S1). perature on the workstation (WonAtech Electrochemical Workstation Based on the above XPS data, we conducted data collection for the ZIVE SP1) with a three-electrode system. The Ag/AgCl electrode work as RF-assisted XPS model. The input data are the experimental conditions a reference electrode, and the platinum foil (1 cm × 1 cm) acts as the 2 X. Liu et al. Carbon 214 (2023) 118342 counter electrode. The Co-rGO was prepared as a working electrode, and 2.4. RF capacitance model 1.0 M KOH was used as an electrolyte. Cyclic voltammetry (CV) mea­ surements were used to evaluate the electrochemical performance. In 2.4.1. RF capacitance model data collection our experiments, 100 mV/s scan rate was applied between − 1 V and Using the same approach as above, we established a new random +1V to explore the acceptable potential windows. Based on this, the forest model RF2 to predict CO-rGO capacitance. The training and appropriate potential window (− 0.3 to +0.4V) of Co-GO was set and testing data of the RF2 capacitance model was collected from RF1 and conducted CV analysis with different scan rates (100 mV/s, 200 mV/s, CV experiments. Similar to RF1, the training set and test set have a ratio 300 mV/s, 400 mV/s and 500 mV/s). The specific capacitance C (F/g) of 80% and 20%, respectively. The samples were prepared at different was calculated with the equation below and electrochemical results [33, synthesis temperatures which are 200 ◦ C, 300 ◦ C, 400 ◦ C, 500 ◦ C, and 34]: 600 ◦ C. For each temperature, the samples with 8 different concentra­ ∫ V2 tions: 2 wt%, 4 wt%, 6 wt%, 8 wt%, 10 wt%, 12 wt%, 14 wt%, and 16 wt 1 C= IdV (4) % were prepared. The CV analysis was performed on each sample with mν × V V1 different concentrations at 5 different scanning rates: 100 mV/s, 200 mV/s, 300 mV/s, 400 mV/s, and 500 mV/s to obtain the capacitance Where I is current in Amp; m is the effective electrode mass in g; V is the ∫ V2 data. In total, 200 rows of data were obtained from the experiment potential window in V; ν is the scan rate in mV/s; and V1 IdV is the (shown in SI 2 Sheet S3). The inputs include synthesis temperatures, integrated area of the CV curve. mass ratio, scan rate and Co(III)/Co(II) ratios. The output is the corre­ The electrochemical impedance spectroscopy (EIS) was conducted sponding capacitance value for the Co-rGO system. with a potential amplitude of 5 mV in a frequency from 600 kHz to 0.01 Hz. Prior to the EIS measurements, the open circuit potential (OCP) 2.4.2. RF capacitance model establishment measurements were carried out to stabilize the system, and the obtained Based on the above data, we established the capacitance model. The OCP value was set as the bias voltage in the EIS measurements. descriptors data were broadly categorized into five categories based on the (1) capacitance of samples, (2) mass ratio of doping transition oxide, (3) Co(III)/Co(II) obtained from RF1 and experimental conditions, Fig. 1. (a) Workflow of the whole project. (b–d) XPS spectrum of Co 2p form samples prepared at 200, 400, 600 ◦ C. (e) Comparison of actual Co(III)/Co(II) and predicted Co(III)/Co(II). Different colours represent one iteration of the training and testing sets. (A colour version of this figure can be viewed online.) 3 X. Liu et al. Carbon 214 (2023) 118342 which include (4) synthesis temperature and (5) scan rate. dataset 1. We emphasise here that experimental conditions act as inputs, Here, the RF model parameters are similar to RF1, with estimators and Co(III)/Co(II) ratios act as output (shown in SI 2 Sheet S1). The set at 150. No specific values are set for the maximum depth and min­ predicted Co(III)/Co(II) ratio at extensive experimental conditions, imum sample split. Consequently, the 10-fold cross-validation method including all mass ratios (2 wt% to 16 wt%) and temperature intervals was utilized to validate model accuracy as RF1. Similarly, the accuracy (200–600 ◦ C), are provided in SI 2 Sheet S2. of the predicting capacitance model is validated through R2, MSE, and Fig. 1e exhibits the predicted Co(III)/Co(II) ratio obtained from RF1. MAE. Here the X-axis represents the experimental Co(III)/Co(II) ratio, and the Y-axis represents the predicted ratio. The slope of the black line (as 3. Results and discussion typically known in ML-based analysis) is 1 (Fig. 1e) [39]. Here, we can see that the data points are close to the black line, which indicates that Fig. 1 illustrates the two stages of this study. In the first stage, XPS predicted and actual values are in good agreement [39] with R-squared data of samples with 8 different mass ratios at 200 ◦ C, 4 different mass value of 0.9997, MAE value of 0.0047 and MSE value of 0.0000618. ratios at 300 ◦ C, and with only one mass ratio of 16 wt% at 400 ◦ C, According to RF1 model, we can say that with the synthesis temperature 500 ◦ C, 600 ◦ C was analysed to obtain the Co(III)/Co(II) ratio. These increasing the Co(III)/Co(II) ratio increases. This demonstrates the po­ synthesis temperatures and mass ratios were regarded as inputs. The tential of utilizing machine learning to predict other XPS results by corresponding Co(III)/Co(II) ratio worked as the output to train the first conducting a small portion of XPS experiments when a large number of random forest model (RF1). Additionally, we employ other character­ XPS measurements are required. ization techniques such as XRD, Raman, and SEM to analyse for obtaining chemical compositions and morphology information of the 3.1.2. Analysis of chemical compositions and structure synthesized cobalt oxide. Based on the XPS spectra analysis, it has been confirmed that Co In the second stage, we used the predicted Co(III)/Co(II) ratio from exists in two states: divalent and trivalent. Furthermore, we conducted RF1 and corresponding sample mass ratios, scanning rates and synthesis XRD, Raman, and SEM analysis for samples prepared at (200 ◦ C, 400 ◦ C, temperatures as inputs. The corresponding capacitances as the output to and 600 ◦ C) for structural and morphological characteristics and the XPS train a second random forest model RF2. An additional experimental analysis of Co-rGO materials systems. validation was performed to verify the accuracy of the model. Moreover, Fig. 2a – c shows the O1s spectra in different temperatures, and a EIS measurements are utilized to explain the reasons behind capacitance noticeable difference can be seen in the range below 532eV. For Fig. 2a, variations. the OOH Peak (531.4eV) of 200-Co-rGO can be assigned to the OH group in Co(OH)2 [35], whereas for 400-Co-rGO (Fig. 2b), the OOH peak dis­ 3.1. Stage1 ML-assisted XPS spectral analysis appears, and two new peaks- OL (529.8eV) and OV (531.4eV) appear. The OL corresponds to oxide connected with metal, and OV is assigned to 3.1.1. Model of RF-assisted XPS oxygen defects in the oxide [10]. The observed XPS peaks for The chemical composition and state of Co-rGO were analysed by X- 600-Co-rGO and 400-Co-rGO samples locate at a similar position. Thus, ray photoelectron spectroscopy (XPS). The Co 2p XPS spectrum of 200- it could be understood that Co(OH)2 is retained at low temperatures Co-rGO exhibits two characteristic peaks (Co 2p3/2 781.1eV and Co 2p1/ while transforming into cobalt oxide when temperature increases. 2 796.7eV) and two satellite peaks (785.4eV and 801.7eV), which shown XRD results are shown in Fig. 2 d-e. The XRD analysis was divided in Fig. 1b [35]. The high intensity of satellite peaks and their shape into two ranges (18–30◦ , 35–70◦ ) to avoid high-intensity substrate peaks indicate the existence of Co(II) [36]. Furthermore, the binding energy (~33◦ ). The 200-Co-rGO samples exhibit distinct Co(OH)2 peaks at difference between Co2p3/2 and Co 2p1/2 is close to 16eV, suggesting Co 19.8◦ , 38.0◦ , and 44.0◦ . Additionally, the minor peak at ~36.3◦ in­ (II) is the sole component when the temperature at 200 ◦ C [37]. The dicates the presence of a small amount of Co3O4. For the synthesis valence states of Co in the experimentally obtained XPS data were temperature of 400 ◦ C, more peaks of Co3O4 (36.3◦ ,43.6◦ ) were analysed (as shown in Fig. 1b-d and SI 1 Fig. S1) and based on the observed, while only one Co(OH)2 peak remains at 44◦ . Furthermore, experimental data, RF1 model was established. when the synthesis temperature increases to 600 ◦ C, we can only observe The XPS data for the samples 400-Co-rGO (Fig. 1c) clearly shows the Co3O4 peaks in the XRD plots. Thus, it can be concluded that the Co two Co states in the spectrum. We observed two new Co(III) character­ (OH)2, which was initially formed at low temperature, is converted into istic peaks (780.0eV and 794.8eV) and three small new satellite peaks Co3O4 at 600 ◦ C. We observed a weak peak around 28.4◦ in our XRD (787.0eV, 797.3eV and 804.0eV). The fitting peaks at 780.0eV and plots as shown in Fig. 2d. In order to analyse this peak, we conducted 795.2eV are assigned to Co(III), while the other two peaks at 781.7eV additional XRD measurements of pure GO on carbon fibre paper heated and 797eV are attributed to the presence of Co(II) [35]. Compared with at 200 ◦ C (refer to SI 1, Fig. S2) and observed no peak 28.4◦ . On ana­ XPS data for 200-Co-rGO samples, Co peak for the 400-Co-rGO sample lysing further, we noticed that the interplanar spacing (d-value) corre­ has red shifted suggesting the formation of Co(III) [35]. Additionally, for sponding to 28.4◦ closely resembles with d-value of Co2O3 (012) plane the 400-Co-rGO samples, the binding energy difference between (refer to "XRD standard reference material card 04-007-3333" and Co2p3/2 and Co 2p1/2 is close to 15 eV, which indicates the coexistence reference [40]). This suggests that the peak at 28.4◦ might be due to the of Co(II) and Co(III) [37]. presence of residual Co2O3. While the XPS peaks of 600-Co-rGO samples exhibited similar The synthesis and phase transformation processes were also characteristic peaks range and shape as of 400-Co-rGO samples. How­ observed in the Raman spectra. The 200-Co-rGO sample exhibited ever, the intensity of satellites peaks of 600 ◦ C samples are reduced, distinct Co(OH)2 peaks at ~460 cm− 1 and ~640 cm− 1 [41]. In contrast, suggesting Co(III) is predominant [38]. the samples 400-Co-rGO mainly displayed characteristic peaks of Co3O4 To quantify the trend of change in chemical composition we intro­ (185 cm− 1, 506 cm− 1, 690 cm− 1) [42]. Specifically, the samples also duced a Co(III)/Co(II) ratio aiding in XPS analysis. Supplementary SI 2 have a shoulder perk at ~640 cm− 1 corresponding to the residual Co Sheet S2 shows the Co(III)/Co(II) at different synthesis temperatures. (OH)2. Furthermore, this residue was barely visible in the 600-Co-rGO Combining the XPS analysis with the machine learning tool, we samples, which only exhibited characteristic peaks of Co3O4. constructed the RF1 to predict more Co(III)/Co(II) ratios at more Combining the results of XPS, XRD, and Raman, it could be inferred that extensive experimental conditions. The experimental condition of the the chemical compositions experience a transformation from Co(OH)2 to XPS samples, which include 8 different mass ratios at 200 ◦ C, 3 different Co3O4 when the synthesis temperature varies from 200 ◦ C to 600 ◦ C. The mass ratios at 300 ◦ C, 400 ◦ C, and one mass ratio of 16 wt% at 500 and 200-Co-rGO, 400-rGO-Co and 600-rGO-Co were also characterized by 600 ◦ C, was combined with corresponding Co(III)/Co(II) ratio to build SEM, and the images were shown in SI 1 Fig. S3. 4 X. Liu et al. Carbon 214 (2023) 118342 Fig. 2. (a–c) XPS spectrum of O1s of 200 ◦ C, 400 ◦ C, 600 ◦ C samples. XRD pattern of electrode composite heating in different temperatures in the range of (d) 18–30◦ (e) 33–64◦ All peak location reference to XRD standard reference material card 00-045-0031 (Co(OH)2) and 01-074-1656(Co3O4). (f) Raman spectrum analysis of three different temperatures. (A colour version of this figure can be viewed online.) 3.2. Stage 2 capacitance model establishment and validation Based on our observations, the variations in the capacitance per­ formance of the samples were similar under different mass ratios or scan To establish the relationship among synthesis temperatures, Co(III)/ rates. Therefore, for convenience, we only present the capacitance Co(II) ratios, and electrochemical properties, we conducted CV experi­ trends under 16 wt% and 100 mV/s conditions. (Fig. 3a). For more ments and collected data to develop a new RF2 model. These multiple capacitance results, please refer to the Supplementary Information (SI 1 CV scans were conducted in at different synthesis temperature Figs. S4–S8). With the increase in synthesis temperature, the capacitance (200–600 ◦ C) and mass ratio (2 wt%-16 wt%). Compared to the exper­ first decreases (From Fig. 3a) from 160.9 F/g (for 200-Co-rGO) to 16.0F/ imental conditions of XPS provided above, the experimental conditions g (400-Co-rGO) and then increase to 176.2F/g when the temperature is for CV have a larger and more comprehensive range. Additionally, we raised to 600 ◦ C. have used different scan rates (100 mV/s - 500 mV/s) to enlarge the Based on CV experiments and RF1, we conclude that temperature dataset and refine the model. All the CV plots are provided in SI 1 variation affects Co(III)/Co(II) ratio and thus influences capacitance. To Figs. S4–8. quantify the correlation, a second random forest model (RF2) was Fig. 3. (a) Capacitance at different temperatures and 16 wt%. (b) Comparison of actual specific capacitance and predicted specific capacitance. (c) Comparison of actual specific capacitance and predicted specific capacitance for experimental validation data (550 ◦ C). (d) Resistance at different temperatures and 16 wt%. (A colour version of this figure can be viewed online.) 5 X. Liu et al. Carbon 214 (2023) 118342 constructed. Co(III)/Co(II) ratios from RF1 were combined with Dataset Acknowledgements 2 (SI 2 Sheet S3), which include synthesis temperatures (200–600 ◦ C), mass ratios (2 wt%-16 wt%), scan rates (100 mV/s – 500 mV/s), and Dali Ji acknowledges UNSW Tuition Fee Scholarship and Australian their corresponding capacitances to establish the Dataset 3 (SI 2 Sheet Research Council Discovery Project DP180101436. Vanesa Quintano S4). Thereinto, synthesis temperature, mass ratio, and scan rate act as acknowledges the funding from the European Union’s Horizon 2020 input, and their corresponding capacitances are output (Fig. 1a Stage 2). research and innovation programme under the Marie Skłodowska Curie Fig. 3b exhibits the RF2 predicted capacitance, and the black line Grant Agreement No. 101066462. The authors acknowledge the facil­ indicates the best-predicted situation. It can be observed that the data ities and the scientific and technical assistance of Microscopy Australia points are distributed centrally along the line. The R-squared value is at the Electron Microscope Unit (EMU), Spectroscopy Laboratory and 0.9655, MAE is 6.77, and the MSE is 95.00. To verify the accuracy of this Solid State & Elemental Analysis Unit (SSEAU) within the Mark Wain­ RF2 model, we compared the experimental validation results at 550 ◦ C wright Analytical Centre (MWAC) at UNSW Sydney. with RF2 predicted results (Fig. 3c). It is worth pointing out that pre­ dicted values match well with the experimental values with a mean Appendix A. Supplementary data relative error value of less than 8%. As for the relationship between the experimental conditions and capacitance, it can be concluded that the Supplementary data to this article can be found online at https://doi. capacitance correlates to the Co(III)/Co(II) ratio, with the least perfor­ org/10.1016/j.carbon.2023.118342. mance for the samples with a ratio proximity of 0.7, and it improves as the ratio deviates from 0.7. References The EIS measurements were performed and presented in Fig. (3d and S9) to understand the reasons for the change in capacitance. The results [1] J. Yang, S. Gunasekaran, Electrochemically reduced graphene oxide sheets for use in high performance supercapacitors, Carbon N. Y. 51 (2013) 36–44, https://doi. reveal a highly correlated trend between the diffusion of ions or mole­ org/10.1016/j.carbon.2012.08.003. cules (Warburg impedance) and the capacitance. It demonstrates that [2] B. Xu, S. Yue, Z. Sui, X. Zhang, S. Hou, G. Cao, Y. Yang, What is the choice for the Warburg slope of 200-Co-rGO and 600-Co-rGO is significantly higher supercapacitors: graphene or graphene oxide? Energy Environ. Sci. 4 (2011) 2826–2830, https://doi.org/10.1039/c1ee01198g. than that of 400-Co-rGO, suggesting the first two samples offer a larger [3] M. Segal, Selling graphene by the ton, Nat. Nanotechnol. 4 (2009) 612–614, ion transfer rate [43]. This trend could be attributed to the dihydrox­ https://doi.org/10.1038/nnano.2009.279. ylation of Co(OH)2 at 400 ◦ C, resulting formation of a denser mixture of [4] J. Lyu, M. Mayyas, O. Salim, H. Zhu, D. Chu, R.K. Joshi, Electrochemical performance of hydrothermally synthesized rGO based electrodes, Mater. Today Co(OH)2 and Co3O4 [44,45]. The denser oxide increases the resistance at Energy 13 (2019) 277–284, https://doi.org/10.1016/j.mtener.2019.06.006. this temperature compared to other temperatures. Thus, it could be [5] M.F. El-Kady, Y. Shao, R.B. Kaner, Graphene for batteries, supercapacitors and inferred that the ion transfer resistance in 200-Co-rGO and 600-Co-rGO beyond, Nat. Rev. Mater. 1 (2016) 1–14, https://doi.org/10.1038/ natrevmats.2016.33. is lower than that in 400-Co-rGO, contributing to increased capacitance. [6] W. Ma, S. Chen, S. Yang, W. Chen, W. Weng, Y. Cheng, M. Zhu, Flexible all-solid- state asymmetric supercapacitor based on transition metal oxide nanorods/reduced 4. Conclusion graphene oxide hybrid fibers with high energy density, Carbon N. Y. 113 (2017) 151–158, https://doi.org/10.1016/j.carbon.2016.11.051. [7] J. Liu, J. Jiang, C. Cheng, H. Li, J. Zhang, H. Gong, H.J. Fan, Co3O4 nanowire@ We have systematically investigated the effect of temperature on the MnO2 ultrathin nanosheet core/shell arrays: a new class of high-performance capacitance of Co-rGO prepared by the hydrothermal method. We used pseudocapacitive materials, Adv. Mater. 23 (2011) 2076–2081, https://doi.org/ the RF-assisted XPS database to establish a relationship between tem­ 10.1002/adma.201100058. [8] Z. Chen, V. Augustyn, J. Wen, Y. Zhang, M. Shen, B. Dunn, Y. Lu, High- perature and Co(III)/Co(II) ratio. The rise of the Co(III)/Co(II) ratio is performance supercapacitors based on intertwined CNT/V 2O 5 nanowire attributed to an increase in the Co3O4/Co(OH)2 ratio with temperature. nanocomposites, Adv. Mater. 23 (2011) 791–795, https://doi.org/10.1002/ A machine learning model was developed to predict capacitance under adma.201003658. [9] J. Zhang, J. Jiang, H. Li, X.S. Zhao, A high-performance asymmetric supercapacitor extensive experimental conditions. We also observed that the capaci­ fabricated with graphene-based electrodes, Energy Environ. Sci. 4 (2011) tance is less when the Co ratio approximates 0.7, potentially due to 4009–4015, https://doi.org/10.1039/c1ee01354h. increased ion transfer resistance. The study showcases the efficacy of [10] C. Lai, Y. Sun, B. Lin, Synthesis of sandwich-like porous nanostructure of Co3O4- rGO for flexible all-solid-state high-performance asymmetric supercapacitors, machine learning, particularly the random forest model, in supporting Mater. Today Energy 13 (2019) 342–352, https://doi.org/10.1016/j. XPS spectroscopy. This approach offers the dual advantages of reducing mtener.2019.06.008. resource consumption during extensive XPS measurements and utilizing [11] E. Lim, J. Changshin, M.S. Kim, M.H. Kim, J. Chun, H. Kim, J. Park, K.C. Roh, K. Kang, S. Yoon, J. Lee, High-performance sodium-ion hybrid supercapacitor predicted XPS results for further experimental analysis and computa­ based on Nb2O5@Carbon core–shell nanoparticles and reduced graphene oxide tional simulations. nanocomposites, Adv. Funct. Mater. 26 (2016) 3711–3719, https://doi.org/ 10.1002/adfm.201505548. CRediT authorship contribution statement [12] Y.Y. Khine, X. Ren, D. Chu, Y. Nishina, T. Foller, R. Joshi, Surface functionalities of graphene oxide with varying flake size, Ind. Eng. Chem. Res. 61 (2022) 6531–6536, https://doi.org/10.1021/acs.iecr.2c00748. Xiaoyu Liu: Validation, Investigation, Data curation, Writing – [13] X. Jin, T. Foller, X. Wen, M.B. Ghasemian, F. Wang, M. Zhang, H. Bustamante, original draft, Visualization. Dali Ji: Validation, Investigation, Data V. Sahajwalla, P. Kumar, H. Kim, G.H. Lee, K. Kalantar-Zadeh, R. Joshi, Effective separation of CO2 using metal-incorporated rGO membranes, Adv. Mater. 32 curation, Visualization, Validation, Investigation, Data curation, Visu­ (2020) 1–10, https://doi.org/10.1002/adma.201907580. alization, Conceptualization, Methodology, Writing – original draft, [14] D. Ji, Y. Lee, Y. Nishina, K. Kamiya, R. Daiyan, D. Chu, X. Wen, M. Yoshimura, Writing – review & editing. Xiaoheng Jin: Conceptualization, Meth­ P. Kumar, D.V. Andreeva, K.S. Novoselov, G.H. Lee, R. Joshi, T. Foller, Angstrom- confined electrochemical synthesis of sub-unit-cell non-van der waals 2D metal odology, Writing – review & editing. Vanesa Quintano: Investigation, oxides, Adv. Mater. (2023), 2301506, https://doi.org/10.1002/adma.202301506. Writing – review & editing. Rakesh Joshi: Investigation, Funding [15] T. Foller, R. Daiyan, X. Jin, J. Leverett, H. Kim, R. Webster, J.E. Yap, X. Wen, acquisition, Resources, Supervision, Writing – review & editing. A. Rawal, K.K.H. De Silva, M. Yoshimura, H. Bustamante, S.L.Y. Chang, P. Kumar, Y. You, G.H. Lee, R. Amal, R. Joshi, Enhanced graphitic domains of unreduced graphene oxide and the interplay of hydration behaviour and catalytic activity, Declaration of competing interest Mater. Today 50 (2021) 44–54, https://doi.org/10.1016/j.mattod.2021.08.003. [16] A. Chen, X. Zhang, Z. Zhou, Machine learning: accelerating materials development for energy storage and conversion, InfoMat 2 (2020) 553–576, https://doi.org/ The authors declare that they have no known competing financial 10.1002/inf2.12094. interests or personal relationships that could have appeared to influence [17] R. Batra, L. Song, R. Ramprasad, Emerging materials intelligence ecosystems the work reported in this paper. propelled by machine learning, Nat. Rev. Mater. 6 (2021) 655–678, https://doi. org/10.1038/s41578-020-00255-y. 6 X. Liu et al. Carbon 214 (2023) 118342 [18] P. Simon, Y. Gogotsi, Materials for electrochemical capacitors, Nat. Mater. 7 (2008) supercapacitors, Carbon N. Y. 197 (2022) 163–170, https://doi.org/10.1016/j. 845–854, https://doi.org/10.1038/nmat2297. carbon.2022.06.014. [19] B. Selvaratnam, R.T. Koodali, Machine learning in experimental materials [34] Y. Sun, Y. Yuan, X. Geng, C. Han, S. Lu, I. Mitrovic, L. Yang, P. Song, C. Zhao, chemistry, Catal. Today 371 (2021) 77–84, https://doi.org/10.1016/j. Biochar-derived material decorated by MXene/reduced graphene oxide using one- cattod.2020.07.074. step hydrothermal treatment as high-performance supercapacitor electrodes, [20] J. Ren, X. Lin, J. Liu, T. Han, Z. Wang, H. Zhang, J. Li, Engineering early prediction Carbon N. Y. 199 (2022) 224–232, https://doi.org/10.1016/j.carbon.2022.07.058. of supercapacitors’ cycle life using neural networks, Mater. Today Energy 18 [35] P. Gao, Y. Zeng, P. Tang, Z. Wang, J. Yang, A. Hu, J. Liu, Understanding the (2020), 100537, https://doi.org/10.1016/j.mtener.2020.100537. synergistic effects and structural evolution of Co(OH)2 and Co3O4 toward boosting [21] N.M. Ralbovsky, I.K. Lednev, Towards development of a novel universal medical electrochemical charge storage, Adv. Funct. Mater. 32 (2022), 2108644, https:// diagnostic method: Raman spectroscopy and machine learning, Chem. Soc. Rev. 49 doi.org/10.1002/adfm.202108644. (2020) 7428–7453, https://doi.org/10.1039/d0cs01019g. [36] Y. Brik, M. Kacimi, M. Ziyad, Y. Brik, F. Bozon-Verduraz, Titania-supported Cobalt [22] Z. Shen, H. Liu, Y. Shen, J. Hu, L. Chen, C. Nan, Machine learning in energy storage and cobalt-phosphorus catalysts: characterization and performances in ethane materials, Interdiscip. Mater. 1 (2022) 175–195, https://doi.org/10.1002/ oxidative dehydrogenation, J. Catal. 202 (2001) 118–128, https://doi.org/ idm2.12020. 10.1006/jcat.2001.3262. [23] P. Raccuglia, K.C. Elbert, P.D.F. Adler, C. Falk, M.B. Wenny, A. Mollo, M. Zeller, S. [37] T.J. Chuang, C.R. Brundle, D.W. Rice, Interpretation of the x-ray photoemission A. Friedler, J. Schrier, A.J. Norquist, Machine-learning-assisted materials discovery spectra of cobalt oxides and cobalt oxide surfaces, Surf. Sci. 59 (1976) 413–429, using failed experiments, Nature 533 (2016) 73–76, https://doi.org/10.1038/ https://doi.org/10.1016/0039-6028(76)90026-1. nature17439. [38] R. Xu, H.C. Zeng, Mechanistic investigation on self-redox decompositions of cobalt- [24] N. Sheremetyeva, M. Lamparski, C. Daniels, B. Van Troeye, V. Meunier, Machine- hydroxide-nitrate compounds with different nitrate anion configurations in learning models for Raman spectra analysis of twisted bilayer graphene, Carbon N. interlayer space, Chem. Mater. 15 (2003) 2040–2048, https://doi.org/10.1021/ Y. 169 (2020) 455–464, https://doi.org/10.1016/j.carbon.2020.06.077. cm021732o. [25] A. Srivastava, O.N. Srivastava, S. Talapatra, R. Vajtai, P.M. Ajayan, Carbon [39] M. Zhou, A. Gallegos, K. Liu, S. Dai, J. Wu, Insights from machine learning of nanotube filters, Nat. Mater. 3 (2004) 610–614, https://doi.org/10.1038/ carbon electrodes for electric double layer capacitors, Carbon N. Y. 157 (2020) nmat1192. 147–152, https://doi.org/10.1016/j.carbon.2019.08.090. [26] P. Avouris, M. Freitag, V. Perebeinos, Carbon-nanotube photonics and [40] J. Chenavas, J.C. Joubert, M. Marezio, Low-spin → high-spin state transition in optoelectronics, Nat. Photonics 2 (2008) 341–350, https://doi.org/10.1038/ high pressure cobalt sesquioxide, Solid State Commun. 9 (1971) 1057–1060, nphoton.2008.94. https://doi.org/10.1016/0038-1098(71)90462-5. [27] M. Acik, G. Lee, C. Mattevi, M. Chhowalla, K. Cho, Y.J. Chabal, Unusual infrared- [41] W.H. Lee, M.H. Han, Y.J. Ko, B.K. Min, K.H. Chae, H.S. Oh, Electrode absorption mechanism in thermally reduced graphene oxide, Nat. Mater. 9 (2010) reconstruction strategy for oxygen evolution reaction: maintaining Fe-CoOOH 840–845, https://doi.org/10.1038/nmat2858. phase with intermediate-spin state during electrolysis, Nat. Commun. 13 (2022) [28] B. Zhao, P. Liu, Y. Jiang, D. Pan, H. Tao, J. Song, T. Fang, W. Xu, Supercapacitor 605, https://doi.org/10.1038/s41467-022-28260-5. performances of thermally reduced graphene oxide, J. Power Sources 198 (2012) [42] M. Yu, Z. Wang, C. Hou, Z. Wang, C. Liang, C. Zhao, Y. Tong, X. Lu, S. Yang, 423–427, https://doi.org/10.1016/j.jpowsour.2011.09.074. Nitrogen-doped Co3O4 mesoporous nanowire arrays as an additive-free air- [29] L. Rokach, Decision forest: twenty years of research, Inf. Fusion 27 (2016) cathode for flexible solid-state zinc–air batteries, Adv. Mater. 29 (2017), https:// 111–125, https://doi.org/10.1016/j.inffus.2015.06.005. doi.org/10.1002/adma.201602868. [30] Y. Wang, Y. Li, W. Pu, K. Wen, Y.Y. Shugart, M. Xiong, L. Jin, Random bits forest: a [43] C. Alex, S.C. Sarma, S.C. Peter, N.S. John, Competing effect of Co3+Reducibility strong classifier/regressor for big data, Sci. Rep. 6 (2016) 1–8, https://doi.org/ and oxygen-deficient defects toward high oxygen evolution activity in 10.1038/srep30086. Co3O4Systems in alkaline medium, ACS Appl. Energy Mater. 3 (2020) 5439–5447, [31] B. Prasanth, R. Paul, D. Kaliyaperumal, R. Kannan, Y. Venkata Pavan Kumar, https://doi.org/10.1021/acsaem.0c00297. M. Kalyan Chakravarthi, N. Venkatesan, Maximizing regenerative braking energy [44] R. Salomão, I. Medeiros de Morais Dias, L. Fernandes, Porogenesis in the alumina- harnessing in electric vehicles using machine learning techniques, Electron 12 brucite-magnesia-spinel system, InterCeram Int. Ceram. Rev. 69 (2020) 46–53, (2023) 1119, https://doi.org/10.3390/electronics12051119. https://doi.org/10.1007/s42411-020-0096-6. [32] Z. Wang, Z. Sun, H. Yin, X. Liu, J. Wang, H. Zhao, C.H. Pang, T. Wu, S. Li, Z. Yin, X. [45] R. Salomão, C.C. Arruda, M.L.P. Antunes, Synthesis, dehydroxylation and sintering F. Yu, Data-driven materials innovation and applications, Adv. Mater. 34 (2022), of porous Mg(OH)2-MgO clusters, Evolution of Microstructure and Physical 2104113, https://doi.org/10.1002/adma.202104113. Properties (2020), https://doi.org/10.1007/s42411-019-0067-y. [33] F. Tang, D. He, H. Jiang, R. Wang, Z. Li, W. Xue, R. Zhao, The coplanar graphene oxide/graphite heterostructure-based electrodes for electrochemical 7