Robust estimation of regional climate change: construction of an hybrid approach between deep neural networks and climate models

Summary

An essential challenge for the climate science community is to provide trustful information about the local impacts of global warming. Climate models are the main tool to study climate evolution according to human activities and greenhouse gas emissions scenarios. They are a numerical representation of the Earth System. Global climate models (GCM) produce worldwide simulations at too low resolution to correctly represent extreme meteorological events that strongly impact our societies. Today we use dedicated regional climate models (RCM) that transform a global low-resolution simulation into a high-resolution one over an area of interest. Nevertheless, the high resolution of those models implies a (much) higher cost that strongly limits the number of those climate simulations and, thus, the necessary exploration of the different sources of uncertainties.

The present work aims to propose a strategy to recreate, at low-cost, high resolution simulations from low-resolution ones. The RCM-emulator introduced here aims to estimate the downscaling function included in a RCM using the recent development of neural networks. This study introduces the concept of RCM-emulator and presents a framework to build, train, and evaluate it. The main result of this study is that the RCM emulator is a credible approach to take up this challenge. Indeed it shows an excellent ability to create realistic high-resolution temperature and precipitation fields, consistent with the low-resolution simulation it downscales. We also study the applicability of the tool the various low-resolution simulations. Moreover, this work also highlights the decisive advantage of using RCM simulations to learn this relationship as it allows to explore future climates and poorly known regions.

The conclusions of this study open the door to further development and various promising applications. Indeed, the RCM-Emulator makes possible the production of robust messages about the local impacts of climate change. Moreover, another significant result of this work is that the emulator performance relies strongly on the calibration set. It is then essential to design the best simulation set to have the most robust emulator implying maybe to revisit the way of choosing which simulation to make with a RCM.