Computational Strategies in Uncertainty Quantification for Hazard Mapping

There are many hazards associated with volcanic activities. Amongst them are Pyroclastic flows; a mixture of rock fragments, debris and hot gases that flow down the slope of actives volcanoes at high velocities. These flows have proven to be devastating, and at the same time more than 500 millions people in the world live within potential exposure to such a hazard. A few approaches have been used to try to mitigate the impact of volcanic hazard in general. These include remote sensing technology and developing hazard maps – a graphic representation of safe and risky zones for a given volcanic area. In this dissertation, we develop a workflow for fast creation of accurate hazard maps. We apply this workflow on the case of the Long Valley volcanic region in northern California (USA). We have also made a couple of contributions that, while pertinent to the problem at hand, also have merit in a wide range of applications. First, we develop a Hierarchical Bayesian model that combines data on Pyroclastic flow behavior from various volcanic sites into a ”global” dataset and reduces predictive uncertainty at volcanoes with sparse data. Of particular interest to us is the uncertainty in key input variables for computer simulations of Pyroclastic flows. Secondly, we develop a learn- ing algorithm for experimental resource allocation in the case where multiple objectives need to be achieved simultaneously. This algorithm allows us to compute probability of hazard for multiple locations at the same time, and vastly reduce the time it takes to create hazard maps. These two contributions form the basis of a tool for geo-scientists to rapidly assess risk spatially at a moment notice, and provide hazard maps that can be used as a teaching tool for communities at risk