Thesis Defence: Physics-Informed Gaseous Fluid Flow Reconstruction From Visual Data

Keyi Wu, supervised by Dr. Shan Du, will defend their thesis titled “Physics-Informed Gaseous Fluid Flow Reconstruction From Visual Data” in partial fulfillment of the requirements for the degree of Master of Science in Computer Science.

An abstract for Keyi Wu’s thesis is included below.

Defences are open to all members of the campus community as well as the general public. Registration is not required for in-person defences.

Abstract

Gaseous fluid flow quantification is crucial for measuring and assessing flow properties. Vision-based gaseous fluid flow reconstruction enables the recovery of density and velocity fields together with their temporal evolution from visual observations, providing a foundation for the quantification of gaseous leakage, emissions, and evaporation. However, existing reconstruction methods often lack sufficient physical constraints, resulting in limited physical plausibility.

This thesis systematically reviews and classifies vision-based gaseous fluid flow quantification methods, gaseous fluid flow image and video datasets, and vision-based gaseous fluid flow reconstruction approaches. Based on the identified limitations of current techniques, a physics-informed gaseous fluid flow reconstruction method, PhysGasFluid, is proposed. It integrates smoothed particle hydrodynamics, position based dynamics, position based fluids, and 3D Gaussian splatting. Specifically, to mitigate tensile instability, a Cubic Spline kernel is employed to generate artificial pressure. A Lagrangian baroclinic turbulence model is incorporated to capture vortex forces and reproduce rotational motion. In addition, a divergence-free loss is introduced to preserve instantaneous volume and enforce incompressibility at the velocity level, complementing the position-level density constraints. PhysGasFluid is evaluated on the ScalarFlow, FluidNexus-Smoke, and FluidNexus-Ball datasets through the tasks of novel view synthesis and resimulation.

Experimental results demonstrate that PhysGasFluid consistently achieves superior visual consistency and physical plausibility, yielding improvements in image-based metrics and incompressibility measures relative to the state-of-the-art baseline. These findings demonstrate the effectiveness of integrating fluid-dynamics principles into the reconstruction process and highlight the potential of physics-informed methods for recovering physically plausible gaseous fluid flows from visual data. Future work includes the development of realistic datasets, enhanced environmental modeling, and more efficient reconstruction algorithms.