Physics-informed gaseous fluid flow reconstruction from visual data

Gaseous fluid flow quantification is an essential task for measuring and assessing quantitative flow properties. Vision-based gaseous fluid flow reconstruction enables the understanding and demonstration of density and velocity fields and volumetric evolution over time, providing an important foundation for gaseous fluid flow quantification. However, purely visual data-driven methods are prone to producing unrealistic reconstruction results with limited physical plausibility. This thesis systematically analyzes and classifies vision-based gaseous fluid flow quantification methods, gaseous fluid flow image and video datasets, and vision-based gaseous fluid flow reconstruction methods. Based on the identified limitations, a novel 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, a cubic spline kernel is adopted to generate artificial pressure and mitigate tensile instability; a Lagrangian baroclinic turbulence model is employed to capture vortex force effects and induce rotational motion; and a divergence-free loss is introduced to preserve instantaneous volume and enforce incompressibility at the velocity level. PhysGasFluid is evaluated on the ScalarFlow, FluidNexus-Smoke, and FluidNexus-Ball datasets through the novel view synthesis and re-simulation tasks. Experimental results show that PhysGasFluid consistently outperforms the state-of-the-art baseline across all the image-based (i.e., MSE, PSNR, SSIM, and LPIPS), constant-density, and divergence-free metrics. These findings demonstrate that PhysGasFluid more effectively recovers physically plausible and visually consistent gaseous fluid flow. Future work includes the diversification of publicly available reconstruction datasets, the development of weather- and obstacle-aware reconstruction models, and the improvement of computational efficiency for reconstruction.