Data-driven multiscale modeling of gas-particle flows

Abstract

Fluidized beds have a wide application in the industry. The gas-particle flows in fluidized beds are inherently unstable with multiscale structures. Insights gained from computational simulations can effectively accelerate the process and technology development of such complex flow systems. For industrial-scale devices, tracking and resolving the vast number of particles is impractical due to computational cost. Therefore the Two-Fluid Model (TFM), which solves continuum equations of motion for both fluid and particle phases, is introduced. High-resolution TFM simulations can resolve various fine structures when the mesh size is small, but it becomes prohibitively expensive for large-scale systems, leading to the development of filtered Two-Fluid Model (fTFM). fTFM requires an accurate filtered drag model to account for the significant sub-grid contribution arising from the inhomogeneities within the filtering volume. This thesis develops a neural-network-based filtered drag model from a dataset constructed by filtering simulation results of a dense fluidized bed. Compared to previous filtered drag models, the new model includes an additional marker,gas-phase pressure gradient, which is identified through theoretical derivation and budget analysis. Validation tests through a prior analysis show high prediction accuracy, and a posterior analysis with coarse grid simulations show good agreement between fine- and coarse-grid simulations. After the filtered drag model development, we continue to tackle two essential challenges. The first challenge concerns the filter size. Industrial devices require large filter sizes. Computationally developed coarse models are commonly based on rather small fine-grid simulations, and extrapolation is carried out for large-scale simulations. Verification on the accuracy of these extrapolations is generally lacking. We approach this problem through systematic verification tests using a cascading framework. We demonstrate that the extrapolation of the neural-network-based model to large grid sizes works satisfactorily. The second question is related to the gas-particle systems to which the filtered drag model can be applied. We analyze and assemble a dataset from different combinations of gas and particle properties, and identify Reynolds number as a suitable additional marker to combine the results from all different cases. The resulting general model for drag correction is applicable to a wide range of gas and particle characteristics.