A new conditional generative adversarial neural network approach for statistical downscaling of the ERA5 reanalysis over the Italian Peninsula

State-of-the-art General Circulation Models (GCMs) typically operate at a coarse spatial resolution, requiring a refinement to assess regional climate changes and their impacts. This weakness is mainly known for representing regional-scale topography and meteorological processes, particularly those responsible for extreme events. Dynamical downscaling methods are computationally demanding. In contrast, though computationally efficient, statistical approaches often sacrifice spatial consistency. To address these limitations, this work introduces an innovative and robust Conditional Generative Adversarial Neural Network (cGAN) architecture for statistical downscaling, discussing the methodology, advantages, and contributions to refining predictions at a finer scale. By leveraging a generator-discriminator architecture, the cGAN developed permits to downscale ERA5 reanalysis at the local scale to obtain a new high-resolution dataset (∼2.2 km), ERA5-DownGAN. The results obtained show the cGAN's architecture presented accurately reproduces the patterns, value range, and extreme values generated by dynamical models for the 2-m temperature over the Italian Peninsula. • Developed an innovative cGAN model for statistical downscaling of ERA5 reanalysis. • Produced a high-resolution dataset (∼2.2 km) for Italy using AI-based downscaling. • The cGAN successfully replicates 2m-temperature patterns, value ranges, and extreme values. • Demonstrated reduced computational costs compared to dynamical models.