Efficient and detailed simulation of fluids for 3D computer graphics

Abstract

Physical simulations are often used in the field of computer graphics to replicate the motion of complex materials such as liquids and snow. The resolution and visual quality of these simulations are generally limited by the available memory and computation times. In this thesis, we introduce three novel approaches that aim to provide visually detailed simulations with reduced memory requirements and computation times. The first approach uses an explicit mesh surface that follows a particle-based simulation of a liquid. It uses a detail-preserving projection to prevent drifting of the surface from the particles. Additionally, we introduced a new topology matching stage which ensures that the surface topology remains consistent with the particles. This approach makes it possible to have a surface with finer visual details on top of a simulation with fewer particles, thus faster to compute. This approach has been tested with both smoothed-particle hydrodynamics (SPH) and fluid implicit particle (FLIP) simulations with success. Furthermore, it can be run after the whole simulation has been computed, making it ideal for the typical iterative process of visual effects studios. The second approach targets the simulation of a layer of snow on the ground that interacts with animated characters and objects. It decomposes the snow volume into three components: the base layer (untouched snow), dynamic snow, and finer particles of snow. This decomposition allows using a lightweight volumetric representation for the base layer, and a more efficient Semi-Lagrangian fluid simulation for the finer particles. As the characters interact with the snow, it is transferred from the base layer to the dynamic snow, which relies on a memory and computation expensive particle-based simulation. In turn, as dynamic snow particles are shoved in the air, density is added in the finer snow simulation. This decomposition allows us to simulate much larger volumes of snow with more details by focusing memory and computation times where it is needed the most. The third approach efficiently simulates and visualizes small-scale liquids in real-time using a 2.5D simulation of columns of liquid. It introduces a novel surface construction algorithm that better handles overhangs and changes in topology, as well as a new physically-based model for viscosity that is both stable and fast. This model relies on assumptions that enable simplifications of the underlying physical equations. While these assumptions lower the model’s accuracy, our results showed that we can replicate behaviors close to that of much more computationally expensive 3D offline simulations that is based on the full viscosity model.

Date	6 Sept 2025
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Eric Paquette (Supervisor)