Can AI Models Perform Dynamical Downscaling? Training Deep Generative Models to Downscale ERA5 Variables to High-Resolution COSMO-CLM Dynamical Reanalyses over Italy

One of the most prominent applications of Deep Learning (DL) in Earth System modeling are downscaling techniques which allow to generate high-resolution fields from coarse-scale numerical models simulations. Indeed, a robust DL downscaling model can potentially generate multiple high-resolution climate scenarios from global climate model runs, saving the timely and resourceful applications of regional/local climate models. Additionally, specific DL models can generate uncertainty information and provide an ensemble-like pool of climate scenarios, hardly achievable using traditional numerical simulations due to their high computational requirements. In this contribution, we present the application of the Latent Diffusion Cast (LDCast, Leinonen et al., 2023) to perform the downscaling of ERA5 (Hersbach et al., 2018) data over Italy up to a resolution of 2km. The target high-resolution data used as ground truth come from the Italian high-resolution dynamical reanalyses obtained with COSMO-CLM (Raffa et al., 2021). The goal of the study is to show that recent advancements in generative modeling can provide comparable results with numerical dynamical downscaling models, such as the COSMO-CLM model, given the same input data (i.e., ERA5 data), preserving the realism of fine-scale features and flow characteristics. The training and testing database is composed of hourly data from 2000 to 2020 (~184000 timestamps), and the target variables of the study are 2-m temperature, horizontal wind components, and precipitation. For each target variable, a selection of predictand variables from ERA5 is used as input to the DL models (e.g., 850hPa temperature, specific humidity, and wind). The LDCast model is compared with multiple baselines, both DL-based (e.g., UNET, GANs) and statistically based methods. Preliminary results are presented, highlighting the improvements introduced with this architecture with respect to the baselines. The results are evaluated by different quantitative verification scores, both continuous (e.g., RMSE, MAE, etc.) and discreet (e.g., CRSS). Uncertainty information provided with the ensemble-based generative approach is also qualitatively analyzed (e.g., CRPS, rank histogram).