Statistical Emulators for Regional Climate Models: Preliminary results

Statistical Emulators for Regional Climate Models: Preliminary results

Predicting some robust information on climate at some local geographical scale is of primary importance to assess the impact of the future climate change. But even more important is to quantify the whole range of uncertainties around the evolution of the climate that translates (i) the imperfections of the climate models, (ii) the natural variability variability and (iii) the uncertainties about the future human emissions of greenhouse gases. One of the nowadays tools used to produce future simulations at the local scale is the Regional Climate Models (RCM): they correspond to high resolution climate models used to downscale over a specific region the information simulated by a Global Climate Model (GCM) scenario simulation.

To cover the full range of uncertainties one should ideally force each RCM with every GCM under different emission scenarios and make several members. It comes down to filling up a huge 4D-matrix [Scenario, GCM, RCM, members]. However regarding the increasing number of climate models (regional and global) and the increasing cost of the RCMs due to their increased complexity and resolution, filling up such matrix becomes unrealistic.

To address this issue we propose a novel approach to merge statistical and dynamical downscaling techniques. The principle relies on three phases. Firstly, some RCM simulations are performed using the classical dynamical downscaling approach. Then, following the statistical downscaling principle, a statistical model is trained to learn the relationship between the large scale information given by the GCM and the local one produced by the RCM, using the runs previously performed. We call this statistical model an emulator. Finally this emulator allows to downscale more GCMs simulation, at a very reasonable cost in order to get a robust ensemble.

In this preliminary work we focus on emulating the surface temperature at the daily scale by testing different machine learning methods (RandomForest, Boosting, Neural Network) sometimes coupled with an a-priori signal decomposition. We train and test the emulator with simulations from the ALADIN RCM forced by the CNRM-CM5 GCM over the period 1950-2100. The different methods are discriminated over hidden simulations using skill scores measuring the match between the emulated series and the pseudo-reality RCM series. Day-to-day scores such as correlation or RMSE are used as well as statistical scores to control on the distribution of the predicted series.