Machine learning and physics-informed modeling of cluster coalescence and breakage in gas–solid flows

Abstract

Particle clusters and their dynamic evolution significantly affect the hydrodynamics and performance of fluidization reactors. Population balance models provide a theoretical framework for describing such dynamic behaviors, whereas quantitative descriptions of breakage and coalescence kernels remain unavailable. In this study, we developed a Lagrangian tracking framework combined with comprehensive criteria for both coalescence and breakage, by which the dynamic evolutions of clusters during their growth, coalescence, and breakage were tracked and analyzed from CFD-DEM simulations. Accordingly, deep neural networks were used to quantify, for the first time, the coalescence and breakage kernels with key influencing factors determined by Random Forest-assisted analysis. To further enhance model accuracy and generalization, a physics-informed neural network (PINN) incorporating both volume conservation and priori knowledge on daughter size distribution was developed for breakage modeling. Results demonstrate that PINN outperforms conventional DNNs owing to its consideration of physical constraint and the ability to suppress data noise.