ECCE Seminar Series (05.10.2023)

Can deep-learning models downscale rainfall extremes in future climates?

Downscaling future climate projections with deep learning can be orders of magnitude more computationally efficient than using regional climate models (RCMs). Our study focuses on using deep learning models to emulate high-resolution precipitation from the Conformal-Cubic Atmospheric Model (CCAM) RCM at a resolution of 12km. Our study focuses on addressing two key issues with RCM emulators. Firstly, regression-based emulators often “regress to the mean” and smooth high frequency variability and extremes. To overcome this issue, we use a generative machine learning approach known as conditional Generative Adversarial Networks (cGANs). Here, the cGAN learns to generate a realization of high-resolution CCAM precipitation, conditioned on a specific boundary condition (large-scale circulation fields) from a GCM (~150km). Secondly, emulators can struggle to reproduce trends present within an RCM (e.g., trends in extreme rainfall), and thus poorly extrapolate to future climates. This issue extrapolation can be partly linked to the fact the RCM and GCM fields are often inconsistent, making it challenging to train a model to learn such a mapping. To overcome this issue, we simplify model training into two-stages, which improves the emulator performance. Training a model in two stages is like first learning the basics of shooting and dribbling in basketball, and then further refining your skills on something more complex such as three-point shooting. First, we pre-train our cGAN, to map from coarsened RCM fields (at the GCM resolution) to high-resolution RCM precipitation (12km). This is a much simpler problem for an emulator to learn as there are no inconsistencies between coarse and high-resolution RCM fields. Then, we fine-tune our emulator on GCM fields as inputs, thereby “sharing” knowledge from two separate problems. We show that through training CNN in two-stages, we obtain better out-of-sample performance and improved reproduction of trends in extreme rainfall.