The physical and visual quality of particle-based solvers like SPH are defined by the number of particles that are used to discretize the fluid. Generally, the more particles that are used, the smaller the damping artifacts and the more small-scale details like splashes, spray, and surface waves can be reproduced. However, doubling the resolution of a simulation increases the particle number by a factor of 8. This increases the computational cost notably since it depends linearly on the number of particles.
To cope with the increasing demand for more detailed flow structures, we propose a new level-of-detail technique that follows the idea to allocate computing resources to regions where complex flow behavior emerges. Instead of recursively subdividing particles which results in particles of different sizes that interact with each other, we rather use a hierarchy of different resolution levels. While the number of resolution levels is not fundamentally limited with our method, we focus on two scales only.
We use two distinct but coupled simulations; one simulation that computes the whole fluid with a coarse resolution L (blue), and a high-resolution level H that simulates a subset of the fluid with small particles (yellow). Boundary conditions (red) are defined, and feedback forces from H onto L are included. The particles of both simulations can then be merged for the final rendering. An overview of our method is shown in Figure 1.
Our method can stably handle dynamically changing and complex high-resolution regions. We present several examples in the paper using different sampling criteria. Figure 2 shows an example where the area around cylindrical obstacles is computed with quadrupled resolution (factor 64 smaller particles) to get small-scaled splashes and individual droplets at impact locations. In this particular example, our method reduces the total particle number and hence the overall computational cost by a factor of 6.7 compared to the single-scale reference simulation.
The result is visually compared to the single-scale reference simulation in Figure 3, showing that our method produces very similar flow details.