283896 Efficient Parallelization of Large Scale Fluidized Bed Flow Computations

Wednesday, October 31, 2012: 8:30 AM
Conference C (Omni )
Amit Amritkar, Surya Deb and Danesh Tafti, Mechanical Engineering, Virginia Tech, Blacksburg, VA

In a gasifier, the carbonaceous material undergoes several processes at high temperatures, including pyrolysis, combustion and gasification reactions. Fluidized bed technology is used in gasifiers because the process provides good mixing and promotes uniform heat and mass transfer between the gases and solid particles (e.g., granular material). There is usually a notable difference in the fluidization behavior between various solid fuel particles due to the particle characteristics, gas-particle interactions and especially the reaction kinetics. Research has been conducted on some aspects of the gasification process, but the physics and chemistry of reacting particles in a fluidized bed are less well understood. Therefore, it is important and timely to increase the fundamental understanding of particle mixing in fluidized bed gasifiers to help maximize the benefit of this technology. Current work aims to increase the fundamental understanding of large scale fluidized beds for solid fuel conversion using computational fluid dynamics (CFD). This is achieved using in-house CFD code GenIDLEST which has the capability of Discrete Element Method (DEM) for getting accurate predictions for gasifier flows. Traditionally the CFD simulation of the fluidized beds has been limited to two dimensional (2D) flow simulations. The current work deals with the three-dimensional (3D) flow simulations of an industrial gasifier with O(106) particles. With large number of particles the coupled CFD-DEM computations become very expensive in terms of computational resources. Using the domain decomposition strategy, two different bed configurations are evaluated for communication and computational costs. To study the computational efficiency of the system, strong scaling study of the multiphase flow problem has been performed on the Nautilus system at NICS. Computational costs exceed communication costs for 1.3 million particle case with 256 cores, clearly indicating the limits on the minimum number of particles per core. The scaling of the DEM solver is much better as compared to the fluid solver since the fluid computational workload per core is much smaller as compared to particulate phase.

Extended Abstract: File Not Uploaded