Title:  High-Performance Multiscale Fluid Simulation: GPU-Based Representations, Solvers, and Adaptive Algorithms

Date: Friday, 17th  April 2026

Time: 10:00 AM - 12:00 PM (Eastern Time)

Location: CODA C0908 Home Park

Teams Link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_NzQyYTVjMzQtZjY1NS00ZDc0LTljOGQtYjY1YzhkZGI0MjQ2%40thread.v2/0?context=%7b%22Tid%22%3a%22482198bb-ae7b-4b25-8b7a-6d7f32faa083%22%2c%22Oid%22%3a%22be1e8809-9fed-468d-b479-4e061696ba91%22%7d

Mengdi Wang

Ph.D. Student

School of Interactive Computing

Georgia Institute of Technology

Committee members

Dr. Bo Zhu (advisor): School of Interactive Computing, Georgia Institute of Technology

Dr. Gregory Turk: School of Interactive Computing, Georgia Institute of Technology

Dr. Sehoon Ha: School of Interactive Computing, Georgia Institute of Technology

Dr. Eftychios Sifakis: School of Computer, Data & Information Sciences, University of Wisconsin-Madison

Dr. Matthew Cong: Senior Research Scientist, NVIDIA Corporation

Abstract

In this thesis, we develop representations, solvers, and transport algorithms for high-performance multiscale fluid simulation, addressing the challenge that important fluid phenomena such as thin films, sub-grid interfaces, and fine vortical structures span spatial scales far smaller than the simulation domain. First, we introduce specialized representations for regimes where standard volumetric methods are insufficient, including a codimensional SPH method for thin fluid films and a sub-grid interface tracking approach on triangle meshes, enabling accurate and efficient simulation with strong geometric fidelity and mass conservation. Next, we present a fully GPU-resident adaptive simulation framework for large-scale three-dimensional flows, combining an efficient octree data structure, a matrix-free algebraic multigrid solver, and a hybrid flow-map transport scheme to capture rich vortical structures on adaptive grids. Finally, we demonstrate that the proposed system achieves effective resolutions of hundreds of millions of cells on a single GPU and produces high-fidelity simulations with significantly enhanced vortical detail. These results show that combining specialized representations with adaptive GPU algorithms provides an effective approach to simulating multiscale fluid phenomena.