RCM-Emulators for precipitation at daily scale

Delivering reliable regional or local climate change projections for the next decades that are both at fine scale and taking into account all sources of uncertainty is currently an unsolvable problem with dynamical models such as RCMs or CPRCMs. Indeed, it requires building large ensembles to capture the various sources of uncertainty (scenario choice, model choice, natural variability). However these fine scale models are very expensive to run, and the different CORDEX exercises highlight the complexity of completing large SCENARIO-GCM-MEMBER-RCM matrices. 

 

In order to tackle this issue we propose a novel hybrid downscaling method called RCM-Emulator. It aims at combining the physical basis of the dynamical downscaling approach (i.e. RCM) with the low computational cost of the empirical statistical downscaling (ESD) using recent machine learning techniques. The idea is to use existing RCM simulations to learn the transfer function from low resolution field to high resolution surface variables such as temperature or precipitation. The emulator developed here is based on a neural network architecture called UNet and is calibrated following a perfect model framework. The training dataset is a EURO-CORDEX simulation performed with the CNRM-ALADIN RCM at 12km, covering historical and scenario simulations. After presenting the concept of RCM-Emulator and the methodology used to learn the downscaling function, we will evaluate its ability to reproduce the high resolution daily precipitation over Europe in perfect conditions and finally apply it to GCM simulation outputs in order to study its transferability. If successful, this novel tool combined with specifically designed RCM simulations will allow to fill up the 4D-matrices over all CORDEX domains.