Deep learning for high-resolution climate projections: a Latent Diffusion Model emulating dynamical downscaling over Italy

Global Climate Models (GCMs) provide critical insights into future climate variability, yet their coarse spatial resolution limits their utility for regional and local-scale impact assessments. AI-driven downscaling techniques have emerged in the last few years as a cost-effective and viable alternative to traditional methods to enhance the spatial resolution of climate projections. Nevertheless, establishing their reliability in unseen climate states remains a priority. This study applies and evaluates a deep generative Latent Diffusion Model, leveraging a residual approach (LDM_res, Tomasi et al., 2025) to downscale GCM outputs (~1°) to high-resolution (~4 km) 6-hourly precipitation and 2-m minimum and maximum temperature fields.

The LDM is developed as an emulator of the COSMO-CLM dynamical model, trained on VHR-REA_IT data (Raffa et al., 2021 - a dynamical downscaling of ERA5). By using aggregated ERA5 data as low-resolution predictors (along with high-resolution static data), the LDM_res model is required to learn to mimic the computationally expensive physics of dynamical downscaling. The model, trained over the past 40 years, is subsequently applied to generate high-resolution climate projections based on the input from four selected CMIP6 GCMs across four different emission scenarios. This modeling chain establishes a hybrid ML-Physics-based system to provide impact assessors with cost-effective, high-resolution climate information.

A central challenge addressed in this work is the evaluation of the model's out-of-distribution generalization—specifically its ability to perform in unseen future climate states and under predictor configurations characteristic of CMIP6 projections. We evaluate the emulator's reliability by comparing its outputs against VHR-PRO_IT, a "twin" dataset of VHR-REA_IT produced using COSMO_CLM to dynamically downscale projections (Raffa et al., 2023), providing a rigorous test of the ML system’s reliability in out-of-domain scenarios.

Furthermore, we compare the LDM_res against traditional statistical (e.g., quantile mapping) and dynamical approaches. Comparative results over the Italian peninsula indicate that while the LDM preserves large-scale seasonal signals from CMIP6 models, it significantly enhances spatial realism and local variability in topographically complex areas. Unlike purely statistical methods, the hybrid ML approach demonstrates superior ability to represent fine-scale heterogeneity in mountainous and coastal regions while maintaining consistency with the original signal.