Remo Gygax: GPU accelerated numerical simulations and laboratory experiments on viscous avalanches

Remo Gygax
Director: Prof. Y. Podladchikov
Jury: Prof. M. Jaboyedoff, Prof. F. Herman, L. Räss, C. Longchamp

Numerical simulations are an eective tool in natural risk analysis. They are useful to determine the propagation and the runout distance of gravity driven movements such as rock avalanches. A numerical model based on fluid mechanics is proposed to simulate such events. In absence of an accepted mathematical explanation of the interactions inside a rock avalanche bulk, a numerical viscous framework is established. Several studies have shown that the numerical implementation of physical processes of viscous flow produces a good t to actual observations of rock avalanche propagation. The use of a continuum approach enables to implement the equation of motion and the mass balance equation of viscous flow on a fully staggered grid. These equations are solved using a finite difference scheme.

The anticipated numerical results on viscous ow are well documented. Nevertheless, analogue laboratory experiments with glycerine on an incline are conducted to find the dependencies between the major parameters. The fingering of the viscous ow was successfully reproduced, at the same time in the laboratory as well as in the numerical implementation. Nowadays, High Performance Computing hardware is available at reasonable costs. Using parallel GPU computing with the Nvidia C-CUDA programming architecture allows to accelerate the simulation. The difference in performance is compared to the widely used scientific modelling software MATLAB. We obtain a nearly 158 times faster simulation while running the developed code on the parallel High Performance Computing architecture.