A Neural Network\nMethod for Characterizing Particle\nClusters in Gas–Solid Fluidized Beds

Cluster quantification in fluidized beds is notoriously\ndifficult\nand heavily depends on empirical correlations with poor generalization\nability. In this study, a simple and general approach was presented\nto characterize local clusters using machine learning. First the adaptive\nOtsu’s threshold was used to identify clusters from the flow\nfield generated by CFD-DEM simulation, then the information regarding\ncluster characteristics and local hydrodynamics were extracted to\ngenerate a data set for machine learning. The Random Forest results\nindicate that pressure drop is the most dominant influence on cluster\nsize, while solids loading ratio and gas velocity influence cluster\nvelocity significantly. The effects of other local hydrodynamic parameters\nsuch as solid concentration and Reynolds number can not be ignored\neither. Accordingly, a Deep Neural Network was trained to predict\ncluster characteristics based on the accessible mesoscale and macroscopic\nflow parameters. Good predictive capability was achieved without a\nmechanistic understanding of cluster dynamics.