Statistics and Prob Seminar: Serge Guillas

Serge Guillas (University College, London)

Uncertainty Quantification for tsunami and climate modeling

In this talk, we first present a first attempt to calibrate a gravity-wave parametrization in the climate model Community Earth System Model (CESM), using dimension reduction through spherical harmonics over the globe. The Bayesian calibration allows us to tune four parameters by comparing variations of zonal wind globally. Second, we present various strategies for the design of experiment and emulation of simulators having uncertain inputs and internal parameters, with applications to tsunami wave modelling. For a simple landslide-generated tsunami model, a fast surrogate is provided by the outer product emulator. It can enable either real-time warnings according to uncertain speed, position and shape of the landslide, or full uncertainty quantification for hazard assessment. We then show somenew realistic simulations and the corresponding emulation for earthquake-generated tsunamis in Cascadia (Western Canada and USA), using VOLNA. VOLNA is a solver of nonlinear shallow water equations on unstructured meshes that is now accelerated on the GPU system Emerald. Strategies for hazard assessment are also discussed. The propagation of uncertainties in both the uplift and in the bathymetry to maximum wave heights are illustrated. We finally propose to use bivariate splines of various degrees and smoothness instead of finite elements in the representation the spatial predictions of the bathymetry as the solution of a Stochastic Partial Differential Equations with INLA (INLA-SPDE). We show that in many situations, this new approach is both effective and computationally efficient.