Séminaire
A reliable quantification of the risk associated with extreme climatic events is crucial for policymakers, civil protection agencies and insurance companies. Studying extremes on a robust statistical basis with complex numerical climate models is however computationally challenging, since extreme events are rare, and thus very long simulations are needed to sample a significant number of them. I will discuss how the problem of sampling extremes in climate models can be tackled using rare event algorithms. Rare event algorithms are numerical tools developed in the past decades in mathematics and statistical physics, dedicated to the reduction of the computational effort required to sample rare events in dynamical systems. Typically they are designed as genetic algorithms, in which a set of cloning rules are applied to an ensemble simulation in order to focus the computational effort on the trajectories leading to the events of interest. I will present a rare event algorithm developed in the context of large deviation theory, and I will show how it can be used to sample very efficiently extreme European heat waves in simulations with the climate model Plasim. This allows to characterise the statistics of heat waves with return times up to millions of years, with computational costs three orders of magnitude smaller than with direct sampling. This allows to sample a large number of trajectories leading to very rare events, which can be used to study their characteristic dynamics, and also to observe ultra rare events that would have never been observed in a normal simulation. I will then discuss how these techniques can be applied to study a wide range of different processes with complex climate models.
bruno.deremble@ens.fr