Speaker
George Biros,The University of Texas at Austin
Abstract
Simulating particulate suspensions governed by Stokesian flows is computationally intensive due to strongly nonlinear fluid-structure interactions, moving interfaces, and multiscale hydrodynamic phenomena. In this work, we introduce VesNet, a novel computational framework that significantly accelerates the simulation of Stokesian particulate suspensions without significantly compromising accuracy. in particular for simulations with large number of deformable particles. Vesnet is trained using a boundary integral equation-based solver. It uses several subnets for approximating self-interactions, background flow interactions, and near-particle lubrication effects. VesNet is not a stand-alone net, as it involves standard algorithmic steps for boundary reparameterization and N-body-type far-field interactions. A GPU-implemented VesNet achieves a speedup of over 100X over a MATLAB multithreaded CPU implementation of the full solver, and about 10X over a Python GPU implementation of the full solver. To access its accuracy, we study the reconstruction of single-vesicle phase diagrams, two-vesicle interactions, and simulations with 1000s of vesicles in Taylor-Green and Poiseulle flow regimes. VesNet is able to accurately reproduce the quantities of interest and enables large scale simulations with modest computational resources.