In recent year, there are increasing concerns over the greenhouse gases in that many environmental issues are associated with the emission of the greenhouse gases. Methane (CH4) and carbon dioxide (CO2) are two major greenhouse gases. Many techniques have been applied to mitigate these greenhouse gases in the environment, among which the dry reforming of methane [1] is particularly used for the production of syngas, an intermediate for producing fuels. This method provides a possible solution for both global warming and energy shortages. Therefore, many studies have focused on the evaluation of the catalysts in order to promote this reaction as much as possible. This project aims to predict and optimise the catalytic dry reforming of methane using artificial neural network (ANN) algorithms. Based on the experimental data from previous studies, an ANN model will be constructed and developed to investigate the efficiency of different catalyst compositions, which will assist in analysing experimental data and reduce experimental costs in further studies. This report intends to give a general introduction of this project, and the objectives and expected contributions of the project will be introduced in the following parts. A preliminary introduction to the artificial neural network (ANN) algorithm used in the experiment will also be given in this report.
In order to predict the effects of the three catalyst compositions (x, y, z in wt%) on the catalyst performance and determine the optimal combinations, different values of the catalyst compositions obtained by Corthals [1] are used as input data for the ANN model in this project. The ranges of the three input parameters are constrained to 2–10wt% for Ni, 0–5wt% for CeO2 and 0–5wt% for ZrO2, and the optimal combination of the three parameters should achieve the highest activity and selectivity, as well as the lowest stability.
In this project, an ANN model is developed by applying Keras API in Python, and experimental data measured by Corthals et al.[1] are used for training the ANN model(Ni-CeO2-ZrO2-MgAl2O4.csv). Three catalyst components mentioned above are identified as input variables, and the conversions of CH4 (XCH4) and CO2 (XCO2), selectivity of H2 (H2/CO) and coking amount are selected as output variables to quantify the catalyst performance.
The Rectified Linear Unit (ReLU) function is chosen as the activation function: f(x)=max{0,x} The major benefit of the ReLU function is that it keeps the gradient constant, resulting in faster training rate. Another benefit is that it does not activate all the neurons at the same time, and therefore is more computationally efficient. In backward propagation, the magnitude of error between the predicted output against the actual output is identified by a loss function. In this model, the mean square error (MSE) is chosen as the loss function: MSE= 1/n ∑_(i=1)^n▒〖(y_i-a_i)〗^2 The optimization algorithm for backward propagation in this model is Adam, an extension to the classic stochastic gradient descent (SGD) algorithm. This algorithm exhibits high computation efficiency and low memory requirements compared with SGD algorithm. More details of Adam can be found in [2].
References: [1] S. Corthals, et al., "Design of active and stable NiCeO2ZrO2MgAl2O4 dry reforming catalysts", Applied Catalysis B: Environmental, pp. 263-275, April 2011. [2] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization”, arXiv preprint arXiv:1412.6980, 2014.