Machine Learning Model for CFD Simulations of Fluidized Bed Reactors

Abstract

Using detailed chemical kinetic models in CFD simulations of multiphase reactors is challenging. Detailed kinetic models include radical species that span a wide range of time scales, making the resulting system of ODEs stiff. Solving a large, stiff system of ODEs in multiphase CFD simulations puts a severe constraint on the time step, making such simulations impractical even for lab-scale reactors. Moreover, such simulations are difficult to converge. For this reason, most multiphase reactor CFD simulations rely on global kinetics, even when a detailed kinetic scheme is available. This work targets this problem, considering biomass thermochemical conversion at 1073–1273 K in a fluidized bed reactor as an application. To this end, a gated recurrent unit (GRU) based recurrent neural network (RNN) model is developed to predict the reactants and product evolution along the fluidized bed reactor length. Biomass devolatilization and gas-phase chemistries are represented by kinetic schemes comprising 20 species with 24 reactions and 39 species with 118 reactions, respectively. A reactor network model consisting of ideal reactors is used to generate the training data. A comprehensive range of biomass compositions and operating conditions are used, ensuring a wide range of model applicability. The developed machine learning model is assessed against the unseen test data and CFD-DEM simulations of a lab-scale fluidized bed reactor. The computational cost of CFD-DEM simulations is reduced by 10 orders of magnitude using the GRU-based RNN model.