Kinetic Study of Product Distribution Using Various Data-Driven and Statistical Models for Fischer–Tropsch Synthesis

Abstract

Three modeling techniques, namely, a radial basis function neural network (RBFNN), a comprehensive kinetic with genetic algorithm (CKGA), and a response surface methodology (RSM), were used to study the kinetics of Fischer−Tropsch (FT) synthesis. Using a 29 × 37 (4 independent process parameters as inputs and corresponding 36 responses as outputs) matrix with total 1073 data sets for data training through RBFNN, the established model is capable of predicting hydrocarbon product distribution i.e., the paraffin formation rate (C 2 −C 15 ) and the olefin to paraffin ratio (OPR) within acceptable uncertainties. With additional validation data sets (15 × 36 matrix with total 540 data sets), the uncertainties of using three different models were compared and the outcomes were: RBFNN (±5% uncertainties), RSM (±10% uncertainties), and CKGA (±30% uncertainties), respectively. A new effective strategy for kinetic study of the complex FT synthesis is proposed: RBFNN is used for data matrix generation with a limited number of experimental data sets (due to its fast converge and less computation time), CKGA is used for mechanism selections by the Langmuir−Hinshelwood−Hougen−Watson (LHHW) approach using a genetic algorithm to find out potential reaction pathways, and RSM is used for statistical analysis of the investigated data matrix (generated from RBFNN through central composite design) upon responses and subsequent singular/multiple optimizations. The proposed strategy is a very useful and practical tool in process engineering design and practice for the product distribution during FT synthesis.

Clean Process to Utilize the Potassium-Containing Phosphorous Rock with Simultaneous HCl and KCl Production via the Steam-Mediated Reactions

“…In this work, 17 experimental conditions were used as training data set (Table S1). For neuron network construction, the activation functions in the neuron network architecture are one of the most important ones, , which are used to transform input (using the summed weighted approach) into the activation of the node. In this work, the rectified linear activation function (ReLU) was used in the Tensorflow during supervised machine learning (SML) process to minimize the errors caused by the vanishing gradient .…”