Applying spatially and temporally adaptive techniques for faster DEM-based snow simulation

Abstract: Background. Physically-based snow simulation is computationally expensive and not yet applicable to real-time applications. Some of the prime factors for this cost are the complex physics, the large number of particles, and the small time step required for a high-quality and stable simulation.Simplified methods, such as height maps, are used instead to emulate snow accumulation. A way of improving performance is finding ways of doing less computations. In the field of computer graphics, adaptive methods have been developed to focus computation to where it is most needed. These works will serve as inspiration for this thesis. Objectives. This thesis aims to reduce the total particle workload of an existing Discrete Element Method (DEM) application, thereby improving performance. The aim consists of the following objectives. Integrate a spatial method, thereby lessening the total number of particles through particle merging and splitting, and implement a temporal method, thereby lessening the workload by freezing certain particles in time. The performance of both these techniques will then be tested and analyzed in multiple scenarios. Methods. Spatially and temporally adaptive methods were implemented in an existing snow simulator. The methods were both measured and compared using quantitative tests in three different scenes with varying particle counts. Results. Performance tests show that both the spatial and temporal adaptivity reduce the execution time compared to the base method. The improvements from temporal adaptivity are consistently around 1.25x while the spatial adaptivity shows a larger range of improvements between 1.23x and 2.86x. Combining both adaptive techniques provides an improvement of up to 3.58x. Conclusions. Both spatially and temporally adaptive techniques are viable ways to improve the performance of a DEM-based snow simulation. The current implementation has some issues with performance overhead and with the visual results while using spatial adaptivity, but there is a lot of potential for the future.