AI-assisted climate downscaling for rapid assessment of ERA5 reanalysis across different geographical domains

State-of-the-art General Circulation Models (GCMs) typically operate at a coarse spatial resolution, posing challenges in accurately assessing regional climate changes and their impacts. This limitation is particularly evident in representing regional-scale topography and meteorological processes, including extreme weather events. Traditional dynamical downscaling methods address these issues but are computationally intensive, while statistical approaches, though efficient, often compromise spatial consistency. This study introduces an innovative application of Conditional Generative Adversarial Networks (cGANs) for climate downscaling to address these challenges. GANs consist of two interconnected components: a generative and discriminative models. The generative model is based on ERA5 climate reanalysis data (Hersbach et al., 2020, ~31 km resolution) and learns to produce high-resolution data. The discriminator uses the VHR-REA_IT dataset (Raffa et al., 2021, ~2.2 km resolution) to distinguish between real and generated data by the GAN (ERA5-DownGAN). This study pioneers the use of cGANs to downscale ERA5 reanalysis data to high horizontal resolution (~2.2 km) for both temperature and precipitation fields. The training phase (1990-2000) allows the cGAN to learn the high-resolution patterns, while the testing phase (2001-2005) evaluates its performance against VHR-REA_IT. The cGAN accurately reproduces patterns and value ranges for both fields, exhibiting a slight tendency toward cooler values. Furthermore, the cGAN downscaling model maintains strong consistency across all percentile classes (from the 1st to the 99th) for temperature, and in nearly all classes for total precipitation, with a tendency to generate outliers in the precipitation fields for the extreme classes (98th-99th percentiles). Additionally, the GAN model developed was validated in collaboration with NCAR through an added case study centered on the U.S. territory. This research demonstrates the significant potential of GANs to address the spatial limitations of traditional climate models, offering a powerful method for high-resolution climate data generation and contributing valuable insights into regional climate dynamics.