Regional climate model emulator based on deep learning: concept and first evaluation of a novel hybrid downscaling approach

Abstract

Providing reliable information on climate change at local scale remains a challenge of first importance for impact studies and policymakers. Here, we propose a novel hybrid downscaling method combining the strengths of both empirical statistical downscaling methods and Regional Climate Models (RCMs). In the longer term, the final aim of this tool is to enlarge the high-resolution RCM simulation ensembles at low cost to explore better the various sources of projection uncertainty at local scale. Using a neural network, we build a statistical RCM-emulator by estimating the downscaling function included in the RCM. This framework allows us to learn the relationship between large-scale predictors and a local surface variable of interest over the RCM domain in present and future climate. The RCM-emulator developed in this study is trained to produce daily maps of the near-surface temperature at the RCM resolution (12 km). The emulator demonstrates an excellent ability to reproduce the complex spatial structure and daily variability simulated by the RCM, particularly how the RCM refines the low-resolution climate patterns. Training in future climate appears to be a key feature of our emulator. Moreover, there is a substantial computational benefit of running the emulator rather than the RCM, since training the emulator takes about 2 h on GPU, and the prediction takes less than a minute. However, further work is needed to improve the reproduction of some temperature extremes, the climate change intensity and extend the proposed methodology to different regions, GCMs, RCMs, and variables of interest.

A simple hybrid statistical–dynamical downscaling method for emulating regional climate models over Western Europe. Evaluation, application, and role of added value?

“…More complex methods based on state-of-the-art machine learning algorithm could be used (e.g. Doury et al 2022).…”

“…Hybrid statistical-dynamical approaches for downscaling are not new (e.g., Najac et al 2009, Walton et al 2015 but have yet to be widely exploited. The development of machine learning may open the way to the emulation of RCMs, as exempli ed by the pioneering work of Doury et al (2022). Here, we explore a simple way to emulate RCM results, based on the constructed analogues approach (van den Dool 1994), already used in other contexts (Deser et al 2016, Terray 2021, including statistical downscaling (Maurer and Hidalgo 2008, Maurer et al 2010, Werner and Cannon 2016.…”

A perfect model study on the reliability of the added small-scale information in regional climate change projections

“…Since the large-scale change is often dominated by internal variability, which is essentially unpredictable, this is of limited practical user relevance. But, the information can be used for benchmarking statistical emulators of high resolution climate information, such as used to fill in missing RCM simulations in a GCM-RCM matrix (Christensen and Kjellström 2021;Doury et al 2022). At least the statistical emulator should perform better than present-day climatology plus the coarse resolution change derived from the low resolution model.…”

Using Explainability to Inform Statistical Downscaling Based on Deep Learning Beyond Standard Validation Approaches

“…Finally, due to the sensitivity of deep learning algorithms to discrepancies in the scale of features, ERA-Interim and EC-Earth predictors are scaled. There are several ways to perform this scaling, in (Doury et al, 2022) authors rely on the daily spatial mean and standard deviation to scale predictors. To avoid the loss of temporal information due to this spatial scaling, they add as additional predictors the daily spatial mean and standard deviation time series of each predictor.…”

“…The UNET model is a popular CNN out-of-the-self configuration which has been widely used in image recognition problems (Ronneberger et al, 2015). In the context of climate downscaling, UNETs have been used in different studies in Europe (Doury et al, 2022;Quesada-Chacón et al, 2022). This model is composed of two different blocks: encoder and decoder (see Figure 2).…”