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

With the development of computing technology and numerical methods, a coupling discrete element method (DEM) with computational fluid dynamics (CFD) is becoming an effective simulation method for deeper understanding of large-scale gas–solid flow systems. Traditionally, the computational cost of the DEM part is usually much higher than the CFD part due to explicitly tracking much more solid particles than the CFD grids. However, with coarse-graining and high-performance computing, the CFD part is actually the wall-time bottleneck in many cases. To mitigate this problem, a deep-learning-based fluid-dynamics prediction method (FPM) is proposed to replace or accelerate the CFD. Accurate CFD–DEM simulation results provide the dataset for deep learning using the artificial neural network (ANN) and UNet architecture, with time-series of local, neighboring, and global information of the gas–solid flow to consider its critical spatiotemporal heterogeneity feature. The FPM can be incorporated into CFD–DEM (FPM–CFD–DEM) or simply replace the CFD (FPM–DEM). Simulations have demonstrated that FPM–DEM can be 10-fold faster than original CFD–DEM under similar accuracy, while FPM–CFD–DEM can only double the speed of CFD–DEM but predict reasonable gas–solid dynamics for much longer time. Future work may exploit remarkable potential of artificial intelligence techniques in developing efficient simulation methods of the gas–solid or other multiphase system.