Skip over navigation

Analysis of Gas-Particle Flows through Multi-Scale Simulations

Speaker: Yile Gu
Series: Final Public Oral Examinations
Location: Lapidus Lounge (E-Quad A210)
Date/Time: Tuesday, December 12, 2017, 2:00 p.m. - 3:30 p.m.

Multi-scale structures are inherent in gas-solid flows, which render the modeling efforts challenging.  On one hand, detailed simulations where the fine structures are resolved and particle properties can be directly specified can account for complex flow behaviors, but they are too computationally expensive to apply for larger systems. On the other hand, coarse-grained simulations demand much less computations but they necessitate constitutive models which are often not readily available for given particle properties. The present study focuses on addressing this issue, as it seeks to provide a general framework through which one can obtain the required constitutive models from detailed simulations. 

To demonstrate the viability of this general framework in which closures can be proposed  for different particle properties, we focus on the van der Waals force of interaction between particles. We start with Computational Fluid Dynamics (CFD) - Discrete Element Method (DEM) simulations where the fine structures are resolved and van der Waals force between particles can be directly specified, and obtain closures for stress and drag that are required for coarse-grained simulations. Specifically, we develop a new cohesion model that appropriately accounts for van der Waals force between particles to be used for CFD-DEM simulations.  We then validate this cohesion model and the CFD-DEM approach by showing that it can qualitatively capture experimental results where the addition of small particles to gas fluidization reduces bubble sizes. Based on the DEM and CFD-DEM simulation results, we propose stress models that account for the van der Waals force between particles. Finally, we apply machine learning, specifically neural networks, to obtain a drag model that captures the effects from fine structures and inter-particle cohesion. We show that this novel approach using neural networks, which can be readily applied for other closures other than drag here, can take advantage of the large amount of data generated from simulations, and therefore offer superior modeling performance over traditional approaches.