Design of Experiments: A Sequential Approach for Probabilistic Tsunami Prediction in North Cascadia.

Abstract:

Experimental design plays a key role in scientific studies, especially in cases where the experiments become computationally demanding and expensive to run, as for example the numerical modelling of tsunami hazard. To build statistical surrogates of the computational response for probabilistic predictions, the initial selection of experiments drives the study in a paramount way. Often such a process is random or based on ad hoc knowledge of the topic in question, which sometimes may increase bias susceptibility. In this study, we implement a design algorithm that maximizes the computational information gain over the multidimensional input space and adaptively selects the succeeding set of experiments. The objective is to build Multi-Output Gaussian Process surrogates and use them for probabilistic high-resolution tsunami hazard prediction. We use as a case study for our experiments a potential full-margin rupture along the Cascadia subduction zone. The subduction zone has given rise to large earthquakes in the past and tsunamis that travelled transoceanic distances. The largest and most recent, known megathrust earthquake occurred in 1700. As a source for the tsunami model we use the crustal deformation along the subduction zone which is defined by a set of shape parameters. The multidimensional input space of the deformation parameters is explored to run the training experiments for the statistical emulation. We model the tsunami hazard using a GPU-accelerated nonlinear shallow water equation solver. We focus the high-resolution experiments around Victoria in Vancouver Island and use the outputs at the cell-centres of the computational grid for the emulation. After the surrogates are built, we utilise them to assess the probabilistic tsunami hazard in the region for a set of 2,000 potential scenarios.