Propagating Input Parameter Uncertainty in Geophysical Mass Flow Modeling
Abstract
Computational methods used in modeling geophysical mass flows assume that parameters (friction angles) and initial data (volume/starting location) are known exactly, when often they are only known in a probabilistic sense. For example, if the bed friction angle is known to lie between 15 and 25 degrees, with insufficient data to say which values within the range are more likely than others, proper treatment requires the input for bed friction to be a uniform probability distribution function (PDF) over that range. During the generation of hazard maps from numerical models, geologists need a tool to quantify the effect of this input uncertainty on specific output quantities of interest. Monte Carlo (MC) sampling is prohibitively expensive for sufficient accuracy; for three significant figures of accuracy approximately 106 MC runs are required, for 20 minute single processor runs on 64 processors this would take 217 days. The relative expense of first principle based models like the Savage-Hutter type avalanche model's governing PDEs (which require the solution of a set of nonlinear hyperbolic equations) rules out using the currently popular Stochastic Galerkin method of uncertainty propagation. In this contribution, we discuss the use of an intelligent sampling methodology to propagate the input uncertainty through the TITAN geophysical flow solver. The stochastic output is displayed as a map of the depth that the flow will not exceed at each (X,Y) point for an arbitrary confidence level.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2005
- Bibcode:
- 2005AGUFM.V31D0646D
- Keywords:
-
- 3245 Probabilistic forecasting (3238);
- 8414 Eruption mechanisms and flow emplacement;
- 8488 Volcanic hazards and risks