Artificial Neural Network Enabled Capacitance Prediction for Carbon-Based Supercapacitors PDF

Accepted Manuscript Artificial Neural Network Enabled Capacitance Prediction for Carbon-Based Supercapacitors Shan Zhu, Jiajun Li, Liying Ma, Chunnian He, Enzuo Liu, Fang He, Chunsheng Shi, Naiqin Zhao PII: DOI: Reference: S0167-577X(18)31406-X https://doi.org/10.1016/j.matlet.2018.09.028 MLBLUE 24...

Accepted Manuscript Artificial Neural Network Enabled Capacitance Prediction for Carbon-Based Supercapacitors Shan Zhu, Jiajun Li, Liying Ma, Chunnian He, Enzuo Liu, Fang He, Chunsheng Shi, Naiqin Zhao PII: DOI: Reference: S0167-577X(18)31406-X https://doi.org/10.1016/j.matlet.2018.09.028 MLBLUE 24896 To appear in: Materials Letters Received Date: Accepted Date: 23 July 2018 6 September 2018 Please cite this article as: S. Zhu, J. Li, L. Ma, C. He, E. Liu, F. He, C. Shi, N. Zhao, Artificial Neural Network Enabled Capacitance Prediction for Carbon-Based Supercapacitors, Materials Letters (2018), doi: https://doi.org/ 10.1016/j.matlet.2018.09.028 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. Artificial Neural Network Enabled Capacitance Prediction for CarbonBased Supercapacitors Shan Zhua, Jiajun Lia, Liying Maa, Chunnian Hea,b,c, Enzuo Liua,b, Fang Hea, Chunsheng Shi*a, and Naiqin Zhao*a, b, c a School of Materials Science and Engineering and Tianjin Key Laboratory of Composites and Functional Materials, Tianjin University, Tianjin 300350, China b Collaborative Innovation Center of Chemical Science and Engineering, Tianjin 300350, China c Key Laboratory of Advanced Ceramics and Machining Technology, Ministry of Education, Tianjin, 300350, China *Email: [email protected](C. Shi); [email protected] (N. Zhao) Abstract Carbon is the most widely used electrode for the supercapacitors. This work applies the artificial neural network (ANN) technology to predict the capacitance of carbon-based supercapacitors. For training the ANN model, we extracted data from hundreds of published papers. Moreover, five features were selected to figure out their impact on capacitance, including specific surface area, calculated pore size, ID/IG ratio, N-doping level and voltage window. Then, several carbon-based samples were chosen to evaluate the performance of ANN. As the result, comparing to other machine learning methods, such as linear regression and Lasso, ANN exhibits the best accuracy and adaptability in the capacitance predication. Keywords: Supercapacitors; Carbon materials; Energy storage and conversion; Simulation and modelling; Artificial neural network 1 1. Introduction Supercapacitors enjoy many merits as the typical energy storage applications, such as high power density and long-cycle life [1]. The key component of a supercapacitor is the electrode, and the most widely used material for the electrode is carbon, due to their excellent physical and chemical properties [1,2]. To enhance the performance of supercapacitors, researchers are eager to increase the specific capacitance of carbon-based electrodes. For examples, they improve the capacitance by increasing specific surface area, controlling pore structure and introducing surface functional groups [1,2]. However, there is currently no accepted theory to systematically explain how to increase capacitance; the modification work on the capacitor electrode only relies on several empirical formulas or rough theoretical models. As the result, the reported data for capacitance are very volatile (varying from 10 to 800 F/g), which is depended on the materials and test conditions [1,2]. Therefore, it is urgent to obtain a sophisticated evaluation system for predicting the capacitive properties of carbon-based supercapacitors. To solve this problem, one feasible method is to draw insights from large amount of data. Traditionally, experimental data are often collected, followed by extrapolation predictions. However, this method does not seem to support a large amount of data for performance prediction with multifactor control. Among various data science technologies, artificial neural network (ANN) is an effective estimation technique to analyze the complex and non-linear behaviors [3]. The advantage of ANN is that it can discover relationships between inputs and outputs of a system without a detailed understanding of the mechanisms involved. Considering the complex variables in carbon-based supercapacitors, it is an ideal problem for ANN technology. Considering the data sources, here comes another issue. At present, in the field of ANN-assisted material design, the data come from several large scale project, such as Materials Genome Initiative [4] and Open Quantum Materials Database [5]. In these databases, the properties of materials including electrical, optical, mechanical properties can be obtained. Yet, in the carbon-based supercapacitors, the main materials are determined, i.e. carbon. The factors that really affect the performance turn to 2 the materials' microstructures and surface functional groups. Even more, the testing details of practical application also have their influences. Unfortunately, no database is available to describe the abovementioned data. To this regard, we believe that one plausible way is to directly extract data from the existing papers. For examples, Ghadbeigi et al. provided performance indicators for Li-ion battery electrode materials, whose data are extracted from ~200 articles [6]. Raccuglia et al. applied a similar human-data-retrieval technique to compile ~4,000 reaction conditions for training machine-learned syntheses of vanadium selenite crystals [7]. Therefore, it is a great potential to obtain data from existed reports and to apply ANN for analyzing the carbon-based supercapacitors. 2. Data collection and feature selection To build the database, we collected more than 300 published papers about carbon-based supercapacitors. Then, 681 sets of data were extracted, including the physical and chemical features of carbon material (i.e specific surface area, pore volume, micropore volume, the ratio of I D/IG, doping elements) and the test system (i.e. electrolyte, test voltage window, and corresponding specific capacity). All of these data are available in the CSV file of supporting information. After that, the features selection is conducted according to the existed scientific understanding. In this research, five variables are selected as the features for the ANN method, which are SSA, calculated pore size, ID/IG ratio, N-doping level and voltage window (Fig. 1). As the electrodes of supercapacitors, the most important feature of carbon is their specific surface area (SSA). The relation of capacitance (C) and SSA (A) can be described by C=εε0A/d (ε is the electrolyte dielectric constant, ε0 is the vacuum permittivity, and d the distance between the center of the ion and the carbon surface) [1,2]. It is necessary to increase the SSA in order to enhance the capacitance (Fig. 1a). Also, the accessibility of electrolyte ions is affected by the pore size of electrode [8]. However, there are many kinds of methods for analyzing the pore size of carbon, such as different detection 3 molecules and different mathematical models for fitting the data. All these differences will bring incompatibility among various researches. Herein, we use mathematical calculation (Wheeling equation[9]) to obtain the average pore size: pore size (calculated)= 4000*PV/ SSA (PV is the pore volume). Furthermore, the crystallinity of carbon materials has an effect on the electrical conductivity, which can be investigated by Raman spectroscopy. In Raman results, the D-bands at around 1360 cm-1 represent the existence of the defects and the amorphous carbon; the G-bands reflect the sp2hybridized carbon locate at around 1570 cm-1; and the intensity ratio between the D-band and G-band (ID/IG) characterizes crystallization degree of carbon materials (Fig. 1e) [10]. Besides of the optimized structures, the suitable components play a significant role in carbonaceous materials for capacitive applications. Among various elements, nitrogen is the most common heteratoms that introduced in carbon matrix [11]. The nitrogen-content functional groups greatly increase the capacitive properties of carbon-based materials by introducing additional pseudocapacitance and quantum capacitance [2, 11]. In addition, the material testing system is desired to mention, especially the electrolyte [1,2]. The main consequence introduced by various electrolytes is the voltage window. According to equation E=CV2[1,2], the energy density (E) of the supercapacitor is proportional to the square of voltage window (V). As the result, voltage window is an important indicator for the overall performance of the devices. 4 Fig. 1. The relation between capacitance and (a) specific surface area, (b) pore size (calculated), (c) pore volume, (d) N-doping amount, (e) the ratio of ID/IG and (f) voltage window. 3. ANN Model Based on the database and the selected features, the ANN model was built by combining three-layer: an input layer with a neuron per input, three hidden layer with the 5*5*2 structure, and a single neuron in the output layer (Fig. 2). In our database, 178 carbon samples have all the abovementioned features. Moreover, these sets of data are divided into training set and testing set. In this experiment, the network was trained with the training data, and the trained network was then evaluated using the test data. For each node of the proposed neural network, the “tanh” function is applied as the activation function, and the cost function is the least squares method. For comparison, two other machine learning strategies (linear regression and Lasso) were also applied to deal with the data. The 5 details of these methods were described in supporting information. The ANN model was built by Tensorflow framework; the linear regression and Lasso were achieved by Scikit-learn package in Python. More details are listed in the Supporting Information. Fig. 2. Illustration of ANN model. 4. Results and discussion After 100000 epochs trains, the capacitance can be predicted from the designed ANN model (Fig. S1). Fig. 3a shows the predicted value versus the real data. From the regression analysis, the correlation coefficient (R2) obtained is 0.91, supporting an accurate prediction of this model. Fig. 3. (a) Comparison of the predicted capacitance and the real data in published papers. (b) Comparison between real capacitances and the predictions of three models. 6 To further approve the merits of ANN, we selected data from several materials other than the data set and substituted it into the ANN model (Fig. 3b). The feature parameters of these samples are very different from each other (Table S1). For example, the specific area varies from 20 to 2856 m2 g-1. Also, the test systems, the heteroatom-doping contents and the pore size distribution vary in a wide range. To this regard, the feasibility of this method can be tested in a comprehensive way. Meanwhile, for the purpose of comparison, the same data were applied in both of linear regression and Lasso as well. At first glance, the prediction results of three methods are in the same order of magnitude with the real capacitance. However, ANN shows the best performance in the aspect of accuracy. Especially, in the sample of No.5, the error of ANN is only 3.3 %, while the errors of linear regression and Lasso are almost 90 %. The superiority of ANN derived from its ability to recognize patterns in a series of input and output data without any prior assumptions about their nature and interrelations. As the result, ANN can minimize the error between the calculated output and the experimental known targets during and achieve a better prediction performance. The deviation of the experiment was mainly due to the small number of data points. If the number of samples increases, the accuracy of predictions will increase. 5. Conclusions This work utilized ANN technology to predict the capacitance performance of carbon materials in supercapacitors. The data were extracted from published papers, and five variables (specific surface area, calculated pore size, ID/IG ratio, N-doping level and voltage window) serve as the features. By tested with the real data, ANN achieves an acceptable prediction result, which is more accurate than linear regression and Lasso. More importantly, this work shows the great potential of ANN for assisting researchers in material science and application design. Acknowledgement 7 This work was supported by the National Natural Science Foundation of China (Grant Nos. 51472177, 51772206 and 11474216), and the Science and Technology Support Program of Tianjin (No. 16ZXCLGX00110, and 16ZXCLGX00070). Shan Zhu was supported by China Scholarship Council. Reference [1] P. Simon, Y. Gogotsi, Nat. Mater. 7 (2008) 845. [2] L.L. Zhang, X.S. Zhao, Chem. Soc. Rev. 38 (2009), 2520-2531. [3] S. Soltanali, R. Halladj, S. Tayyebi, A. Rashidi, Mater. Lett. 136 (2014) 138-140. [4] R. Yuan, Z. Liu, P.V. Balachandran, D. Xue, Y. Zhou, X. Ding, J. Sun, D. Xue, T. Lookman, Adv. Mater. 30 (2018) 1702884. [5] P.V. Balachandran, J. Young, T. Lookman, J.M. Rondinelli, Nat. Commun. 8 (2017) 14282. [6] L. Ghadbeigi, J.K. Harada, B.R. Lettiere, T.D. Sparks, Energ. Environ. Sci. 8 (2015) 1640-1650. [7] E. Kim, K. Huang, A. Tomala, S. Matthews, E. Strubell, A. Saunders, A. McCallum, E. Olivetti, Sci Data, 4 (2017) 170127. [8] J. Chmiola, G. Yushin, Y. Gogotsi, C. Portet, P. Simon, P.L. Taberna, Science 313 (2006) 17601763. [9] A. Wheeler, W.G. Frankenburg, V.I. Komarewsky, E.K. Rideal, P.H. Emmett, H.S. Taylor, Advances in Catalysis; Academic Press, New York, 1951. [10] S. Zhu, J. Li, C. He, N. Zhao, E. Liu, C. Shi, M. Zhang, J. Mater. Chem. A 3 (2015) 2226622273. [11] A. Chen , Y. Yu, T. Xing, R. Wang, Y. Li, Y. Li, Mater. Lett. 157 (2015) 30-33. 8     Highlights ANN was used to predict the capacitance of carbon-based supercapacitors. A database of carbon-based electrodes was established by extracting data. Five variables were selected to figure out their impact on the capacitance. ANN exhibited the best accuracy in the capacitance predication. 9 Graphical abstract Specific surface area Calculated pore size ID/IG N-doping Voltage window 10