Adaptive Particle Splitting Based on Turbulence Energy for Fluid Simulations on GPUs

Arno in Wolde Lübke (1151205)


Physics-based computer animation for fluids has not only applications in the movie, advertisement or video game industries but also might be found useful for VR training simulations in the near future. Unfortunately, the underlying model of fluids is complex and computational expensive to solve. However, as plausibility outweighs physical accuracy inthe above mentioned applications, simplifications to the model should be made. This is one reason for the growing popularity of particle systems as opposed to grid structures in the field of fluid animation. Particle systems are popular as they allow computations only where necessary, simplify the governing equations greatly and trivially conserve mass.

However, even with the use of a particle system, the computational requirements of fluids are immense. Real-time applications are restricted to only a few thousand particles and design cycles for offline applications are still considerable. To simulate more particles in shorter time, graphics processing units (GPUs), which posses multiple processing units for parallel computations, have been used to speed up the computation of fluids. Furthermore, multi-resolution methods that simulate fluid areas at different scales depending on certain criteria, have been proposed and succeeded in preserving visual detail while significantly reducing the computational cost.

To our knowledge, only a few efforts have been made to combine GPUs and multi-resolution SPH fluids. Furthermore, existing implementations produced no considerable performance gains if the fluid was very turbulent. This talk discusses an easy-to-implement multi-scale simulation method for graphics processing units that splits particles in visually important regions of the fluid into smaller, high-resolution particles. To reduce the computational overhead introduced by the additional simulation scale, a way for splitting only those particles found within turbulent surface regions that are visible to the camera are is discussed. Furthermore, a method that allows for a stable transition between the two simulation scales is presented.