AI-assisted modeling and optimization of large-scale ethylene production by catalytic oxidative dehydrogenation of ethane in a shell-and-tube multi-tubular reactor

Abstract

The present study is devoted to the development of a methodology based on artificial intelligence to model and optimize a novel process for efficient ethylene production on a large scale. To this end, a wall-cooled multi-tubular reactor for the catalytic oxidative dehydrogenation of ethane that benefits from MoVTeNbO catalysts was modeled and optimized with different plans of artificial neural networks (ANNs). A mathematical model was then employed to generate a large database (1440 runs at varied process conditions) for this process. Having screened with a relevancy analysis, the dataset was then utilized to analyze different ANN designs based on single-layer perceptron (SLP) and single-layer cascade feed-forward (SLCFF) modes and different training functions. These designs were compared with radial basis function (RBF) as well. The impacts of train function (Broyden-Fletcher-Goldfarb-Shanno Quasi-Newton (trainbfg) and Levenberg-Marquardt (trainlm)) and hidden layer size were also monitored rigorously. Although the RBF ANN failed to describe this system, the SLP and SLCFF plans could represent the outputs precisely. Based on a robust statistical analysis of the errors, as well as kernel density and histogram analyses, it was proven that the calculations of SLP and SLCFF ANN models were reliable. The model composed of SLP-trainbfg with 20 neurons in the hidden layer was selected as the most accurate and fastest one, which was then combined with a multi-objective genetic algorithm method to search for the optimum operating conditions of the system in terms of higher feed conversion and ethylene generation with lower carbon oxide release. The observations signified that the process working under optimum conditions produces a considerable amount of ethylene at moderate temperature and pressure while releasing small quantities of carbon oxides.