Surrogate Models and Uncertainty Quantification for Hazard Map Construction
Abstract
Computer models of hazardous phenomena, such as floods, hurricanes, and avalanches, are very expensive to run, and each run produces an enormous amount data. For example, a flood model output consists of water depth and velocity at every point in a large grid, at every instant of time. Furthermore, these models often require speci[|#12#|]fication of several parameters that may not be well characterized, and initial and boundary data that is likewise poorly specifi[|#12#|]ed. These inputs include for instance digital elevation models to represent terrain. Given the high dimension of uncertain features, a study of hazard risk using Monte Carlo type procedures for such models will be prohibitively expensive even with high end supercomputers. We present here systematic procedures that allow the construction of hazard maps based on surrogate models of the physical phenomena created by using a small ensemble of O(100) large scale computer simulations. We introduce the idea of hierarchical global and local (in parameter space) surrogates combined systematically to allow us to quantify uncertainty in the simulation outputs. This characterization may then be used as an easily computable indicator of potential hazard. We present the construction of hazard maps for large scale volcanic flows as an example. Hazard analysis of potential flows at Mammoth Mountain, CA using input uncertainties in digital elevation model, frictional resistance to flow and initiation location and volume (8 dimensional input).
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2010
- Bibcode:
- 2010AGUFMNH51C1250P
- Keywords:
-
- 0520 COMPUTATIONAL GEOPHYSICS / Data analysis: algorithms and implementation;
- 0550 COMPUTATIONAL GEOPHYSICS / Model verification and validation;
- 0560 COMPUTATIONAL GEOPHYSICS / Numerical solutions;
- 1932 INFORMATICS / High-performance computing