UBC Theses and Dissertations
Particle-resolved simulation and data-driven modelling of flows laden with polydisperse spheres Cheng, Zihao
Polydisperse particle-laden flows are ubiquitous in both nature and industry across diverse disciplines. An in-depth understanding of the intricate interactions between the dispersed solid phase and carrier fluid phase is essential in terms of designing and optimizing the key components of various industrial processes. Due to the large separation of temporal and spatial scales across different phases in real-world problems, previous studies encountered difficulty establishing a proper two-way coupling with both satisfactory accuracy and efficiency: either the various interactions are coarse-grained and oversimplified by the average drag closures in large-scale models, or the computational costs may soar beyond affordability while fully resolving the boundary of smallest particles in industrial-level simulations. To address this challenge, we develop two deterministic models based on the particle-resolved simulation data to estimate the force and torque fluctuations within polydisperse sphere assembly, which provides a channel for bottom-up knowledge transfer to evaluate the macroscale interphase interactions via the microscale hydrodynamic behaviors. Accordingly, the primary contributions of this dissertation feature the following aspects: First, we develop a highly scalable Immersed Boundary-Lattice Boltzmann method (IB-LBM) that is implemented on the adaptive quadtree/octree grids to model the complex solid-fluid interactions in different contexts such as flows laden with fixed or moving particles. Second, we perform a series of particle-resolved simulations of flow past random arrays of moderately to strongly bidisperse and polydisperse spheres which seemed computationally impossible on uniform grids in the past. We explore the statistical distributions of hydrodynamic forces and torques exerted on individual spheres over a wide range of simulation parameters. Finally, we extend two data-driven methods to predict such force and torque distributions, namely a Microstructure-informed Probability-driven Point-particle (MPP) model and a Physics-Informed Neural Network (PINN). We demonstrate their applicability from different perspectives including prediction and generalization performance, model complexity (i.e., computational efficiency), and interpretability in the form of binary and trinary interactions. We highlight the potential of our PINN model, which appropriately balances accuracy and efficiency, to substitute the conventional average drag closures for solid-fluid coupling in Eulerian-Lagrangian simulations.
Item Citations and Data
Attribution-NonCommercial-NoDerivatives 4.0 International