Graph Neural Networks for High-Resolution Climate Projections

Graph neural networks (GNNs) have recently emerged as a powerful deep learning framework in scientific applications. This thesis explores their potential in the field of climate science, presenting a novel application of GNNs to climate projections and the challenging task of high-resolution precipitation downscaling. Climate projections are numerical simulations that estimate the future climate evolution under different greenhouse gas emission scenarios. They are typically produced by global climate models (GCMs), which provide essential large-scale information but operate at coarse spatial resolutions, insufficient to capture local phenomena such as intense precipitation. Dynamical downscaling through convection-permitting regional climate models (CP-RCMs) is an effective method to bridge this gap and achieve kilometre-scale resolution. However, CP-RCMs become prohibitively expensive when long climate projections are required or many simulations are needed to estimate the uncertainty of the climate projections. Deep learning models have recently been introduced as an efficient complementary method to emulate the downscaling function between GCMs and RCMs. However, the precipitation phenomenon has been little addressed, particularly in the case of sub-daily data, with multiple open research questions. To address these challenges, this work introduces a new RCM emulator based on GNNs, named GNN4CD (Graph Neural Networks for Climate Downscaling). The proposed model is designed to learn the complex, non-linear relationships between coarser-scale climate drivers (~25km) and local precipitation at very high temporal (1h) and spatial resolution (3km). The GNN-based model offers flexibility for operating on irregular grids and non-rectangular domains, allowing for spatial transferability to regions distinct from those used in training. A novel hybrid imperfect framework enables the emulator to learn from historical reanalysis and observation data, then extrapolate to future climates using climate simulation predictors during inference. The GNN4CD emulator proved effective in estimating high-resolution precipitation in a much shorter time compared to traditional dynamical downscaling. It produced accurate estimates for both the historical period and future projections, capturing the effect of climate change on the precipitation signal. Moreover, it demonstrated potential for spatial transferability and generalisation across different domains and scenarios.