DATE2022-05-16 18:09:24
IDABSTRACT20220516180924-202
CONTACTsamuel.somot@meteo.fr
PRESENTATIONORAL
INVITED0
IDSESSION5
TITLERegional Climate Model emulator based on deep learning: concept and evaluation of a novel hybrid downscaling approach to study the Mediterranean climate change at fine scale
AUTHORSSamuel Somot (1) ,Antoine Doury (1)
AFFILIATIONS
  1. 1) Cnrm, Université De Toulouse, Météo-france, Cnrs, Toulouse (France)
ABSTRACTDelivering reliable regional and local climate change information for the next decades that is both at fine scale and taking into account all sources of uncertainty is currently an unsolvable problem with currently available dynamical climate models. In particular, current Regional Climate Model (RCM) ensembles strongly undersample the uncertainty range in future projections. To tackle this issue, we propose a completely new approach called “RCM-emulators”. The RCM emulators belong to the family of the hybrid downscaling approaches, combining the physical basis of the dynamical downscaling approach (typically RCM) with the flexibility and the low computational cost of the empirical statistical downscaling (ESD). More specifically, we train a deep neural network to learn the downscaling function of a RCM within existing simulations covering a large range of climate states (past and future). The RCM emulator can then be used to emulate, statistically and at a very low cost, fine scale climate information for century-long time series of daily temperature and precipitation maps. Learning from a given RCP-GCM-RCM triplet, the emulator can emulate the RCM behavior for other socio-economic scenarios, other members of the same GCM or other driving GCMs. We present here the chosen approach and its detailed evaluation thanks to existing simulations not used during the training phase. The RCM-emulator’s evaluation covers temperature and precipitation over a Mediterranean climate area for the mean climate state, the climate change signal and the extreme events.
PAGE61
STATE1