A simple coupled ANNs‐RSM approach in modeling product distribution of Fischer‐Tropsch synthesis using a microchannel reactor with Ru‐promoted Co/Al ₂ O ₃ catalyst

Abstract

A simple approach using hybrid artificial neural networks (ANNs)-response surface methodology (RSM) was developed to model the detailed product distribution using Ru-promoted cobalt-based catalyst with Al 2 O 3 as the support in a microchannel reactor for Fischer-Tropsch (FT) synthesis. Using the independent process parameters for training, the established model is capable of predicting hydrocarbon production distributions, ie, paraffin formation rate (C 2 -C 15 ) and olefin to paraffin ratio (OPR C 2 -C 15 ) within acceptable uncertainties. The ANNs-RSM model and comprehensive mechanistic model using the Langmuir-Hinshelwood-Hougen-Watson (LHHW) approach were compared to identify a few inherent advantages of the proposed model in modeling complex FT synthesis. The proposed ANNs-RSM model shows its appealing merits, ie, faster converge and higher accuracy (less than ±10% uncertainties was achieved from ANNs-RSM model, while LHHW model achieved less than ±15% uncertainties except a few exceptional high errors uncertainties at certain experimental conditions). The statistical significances to the product distributions and hydrocarbon formation rate during FT synthesis could be easily identified and quantitatively analyzed by the established ANNs-RSM model.The future effective alternative of kinetic study of the complex system such as FT synthesis in the microstructured reactor might follow the methods of using empirical reliable approach for process optimization together with mechanistic kinetic study for detailed reaction pathway discrimination.

A Review of Enhancement of Biohydrogen Productions by Chemical Addition Using a Supervised Machine Learning Method

“…The setting for allowable accuracy was 95%. For the ANNs prediction data matrix, the widely used Box-Behnken design (BBD) and the central composite design (CCD) were used to predict the data matrix generation [16]. Once the supervised data learning was complete, the analysis of variation (ANOVA) based on commercial Design Expert ® Version 11 software package (Stat-Ease, Inc., Minneapolis, MN, USA) was used for statistical analysis.…”

A Review of Biohydrogen Productions from Lignocellulosic Precursor via Dark Fermentation: Perspective on Hydrolysate Composition and Electron-Equivalent Balance

“…In this work, artificial neuron networks (ANNs) were established in Python 2.7 (Python Software Foundation). For supervised learning, there are no consensuses in regarding to how many hidden layers should be utilized [32]. In this work, we employed the most widely used feed backward three layers to train data.…”

“…The matrix of hydrolysate (the substrates such as the concentrations of glucose, xylose, and the inhibitors such as acetate, furfural, and aromatic compounds) plays the most critical role in determining the effectiveness of BioH 2 production once the microbial strains are chosen. In order to provide insightful understanding between the matrix of hydrolysate and BioH 2 production (HY and HER), we employed a recently developed novel correlation algorithm using the established artificial neuron networks (ANNs) combined with Box-Behnken design (BBD) design [32]. The schematic diagram of the established ANNs using training data from the references in Table 2 is depicted in Figure 3.…”

Modeling Fischer–Tropsch kinetics and product distribution over a cobalt catalyst

“…Many of the published models attempted to describe product distribution using ideal Anderson‐Schulz‐Flory (ASF) distribution 5,7,9,10,13‐15,22,23 while; recently, Sun et al 30 formulated an artificial neural networks—response surface methodology approach to product distribution modeling. The former approach is of primary interest as the latter approach lacks two critical qualities of physical modeling: generalization and physical interpretation.…”

“…Rytter and Holmen derived an LHHW kinetic expression embodying both positive and negative effects of water on the FT rate. 7 Many of the published models attempted to describe product distribution using ideal Anderson-Schulz-Flory (ASF) distribution 5,7,9,10,[13][14][15]22,23 while; recently, Sun et al 30 formulated an artificial neural networksresponse surface methodology approach to product distribution modeling.…”