Large-scale fluidized-bed CH4 pyrolysis reactor for simultaneous COx-free H2 and carbon production: Multi-objective optimization and artificial intelligence modeling of different process schemes

The present study is devoted to the development, multi-objective optimization, and artificial intelligence modeling of turquoise H₂ and carbon production through the thermal decomposition of CH₄ also known as pyrolysis. With a kinetic model of reaction and deactivation on the Fe/Al₂O₃ catalyst particles, a mathematical model was derived for a fluidized-bed pyrolysis reactor with a perfectly mixing continuous stirred tank reactor assumption, which is then applied to a genetic algorithm optimization procedure to explore the best performance of the two reactor designs (adiabatic and well-heated). The optimization strategy included two plans for the best operating conditions and processing time. In both optimization plans, the well-heated reactor was superior in terms of higher conversions and product yields, as well as more stable catalysts. This was managed due to the instantaneous heating of the reaction area by molten salt flowing in the shell side of the reactor. The mathematical model was in the next section combined with an artificial intelligence computation approach inspired by neural networks. Extended databanks that included 3840 runs at varied operating conditions in each pyrolysis reactor were then analyzed by Pearson approach to determine the effective input variables and construct the input layer of single- and double-layer perceptron neural networks. The impacts of train function and hidden layer(s) size were also investigated rigorously. Although single-layer neural networks failed to describe the systems in question efficiently, the double-layer modes that benefitted from the trainbr and trainbfg functions could represent the outputs (average temperature, conversion, H₂ yield, and carbon yield) of both systems precisely. Statistical parameters, errors analysis, as well as kernel density and histogram analyses, revealed that the calculations of best models can be dependable. Through a comparison between the models’ outputs and the target variables, it was also revealed that the double-layer network can detect even very small alterations in the operating variables.