Prediction of electrocatalyst performance of Pt/C using response surface optimization algorithm‐based machine learning approaches

• Published in: International journal of energy research Vol. 46; no. 15; pp. 21353 - 21372
• Main Authors: Elçiçek, Hüseyin, Özdemir, Oğuz Kaan
• Format: Journal Article
• Language: English
• Published: Chichester, UK John Wiley & Sons, Inc 01.12.2022
• Subjects: Algorithms; Artificial neural networks; Catalysts; Chemical synthesis; Efficiency; Electrocatalysts; Emission analysis; Energy efficiency; Errors; Fuel cells; Fuel technology; Learning algorithms; Machine learning; Mathematical models; Multilayer perceptrons; Neural networks; Optimization; Parameters; Performance evaluation; Performance prediction; pH effects; Prediction models; Process parameters; Pt/C electrocatalysts; random forest; Response surface methodology; Statistical analysis; Statistical methods; Support vector machines; support vector regression; Temperature; Working conditions
• ISSN: 0363-907X; 1099-114X
• DOI: 10.1002/er.8207

Summary Nowadays, fuel cells have attracted a lot of attention because of their unique efficiency, high ‐power density and zero gas emission, and many studies have been conducted to improve their efficiency. The difficulties that occur must be fully grasped and minimized to optimize the energy efficiency and the performance of the fuel cells. To increase the performance of Pt/C catalysts and ensure effective synthesis, precise control of the synthesis conditions is necessary. In the present study, the effect of the synthesis process parameters on the catalyst performance used in fuel cells was comprehensively investigated using statistical methods and machine learning algorithms. The polyol synthesis process was implemented to prepare efficient Pt/C electrocatalysts with reducing synthesis cost and time. The synthesis parameters including duration of reaction, pH and reaction temperature were experimentally studied to determine the optimal working conditions. This study also intended to create an adequate mathematical model with response surface methodology and a prediction model with machine learning algorithms to predict the amount of reduced Pt and the ECSA value depending on the synthesis parameters, and to understand the interaction of parameters. Various ML algorithms that are multilayer perceptron artificial neural network (MLP‐ANN), support vector regression (SVR) and random forest (RF) model were used and each model's performance was evaluated using several performance indicators (R2, mean absolute errors, mean squared error and root mean square errors). The results show that pH is the prominent parameter for both responses. To obtain maximum Pt/C electrocatalyst performance and reduction of Pt, the optimum parameters are determined as pH of 4, reaction temperature of 135°C, and reaction duration of 1 hour. The validation results show a good agreement between predicted and experimental data is obtained with the developed model. Results have obviously shown that this approach can effective in optimizing the electrocatalyst performance with the multiple process parameters. Moreover, it was found that the MLP‐ANN model was outperformed to predict electrocatalyst performance of Pt/C more precisely compared to SVR and RF model. This study aimed to optimize catalyst performance and predict future outcomes with the highest accuracy. The effects of working parameters which are the duration of reaction, pH and reaction temperature, on the Pt/C electrocatalyst performance and reduction of Pt are examined by response surface methodology. In addition, the outputs of Pt/C electrocatalyst performance and reduction of Pt are combined to get a single objective function by applying the grey relational analysis method. In addition, machine learning algorithms that are multilayer perceptron artificial neural network (MLP‐ANN), support vector regression (SVR) and random forest (RF) model were developed to predict the electrocatalyst performance. The optimum parameters are found as pH of 4, reaction temperature of 135°C, and reaction duration of 1 hour. MLP‐ANN model performance was determined to be the best machine learning algorithm compared to SVR and RF, with an R2 of 99.99%. The prediction results show a good agreement between predicted and actual data obtained with the developed model.