Coupling numerical simulation and artificial intelligence prediction: A computational fluid dynamics–discrete element method and deep learning approach to gas–solid flows.

The article focuses on developing a fluid-dynamics prediction method (FPM) based on deep learning (DL) to accelerate and enhance computational fluid dynamics–discrete element method (CFD–DEM) simulations of gas–solid flows. Using datasets generated from accurate CFD–DEM simulations of a three-dimensional fluidized bed, the FPM employs artificial neural networks (ANN) and UNet architectures to predict gas-phase velocity and pressure by incorporating local, neighboring, and global spatiotemporal flow information, with physics-informed loss functions ensuring physical consistency. Two approaches are proposed: replacing the CFD solver with FPM coupled to DEM (FPM–DEM), achieving about a tenfold speedup in gas-phase prediction and roughly 2.5 times faster overall simulation with comparable accuracy up to 10,000 CFD time steps; and integrating FPM with CFD–DEM (FPM–CFD–DEM) to correct cumulative errors, enabling longer-term stable predictions and doubling simulation speed. The study highlights the potential of AI techniques to improve efficiency in simulating complex multiphase gas–solid systems, while noting challenges in extending the method to complex geometries beyond Cartesian grids.