Can AI be enabled to perform dynamical downscaling? A latent diffusion model to mimic kilometer-scale COSMO5.0_CLM9 simulations

Abstract

Downscaling based on deep learning (DL) is a key application in Earth system modeling, enabling the generation of high-resolution fields from coarse numerical simulations at reduced computational costs compared to traditional regional models. Additionally, generative DL models can potentially provide uncertainty quantification through ensemble-like scenario generation, a task prohibitive for conventional numerical approaches. In this study, we apply a latent diffusion model (LDM) to demonstrate that recent advancements in generative modeling enable DL to deliver results comparable to those of numerical dynamical models, given the same input data, preserving the realism of fine-scale features and flow characteristics at reduced computational costs. We apply our LDM to downscale ERA5 data over Italy up to a resolution of 2 km. The high-resolution target data consist of 2 m temperature and 10 m horizontal wind components from a dynamical downscaling performed with COSMO-CLM. A selection of predictors from ERA5 is used as input, and a residual approach against a reference U-Net is leveraged in applying the LDM. The performance of the generative LDM is compared with reference baselines of increasing complexity: a quadratic interpolation of ERA5, a U-Net, and a generative adversarial network (GAN) built on the same reference U-Net. Results highlight the improvements introduced by the LDM architecture combined with the residual approach, outperforming all the baselines in terms of spatial error, frequency distributions, and power spectra. These findings point out the potential of LDMs as cost-effective, robust alternatives for downscaling applications (e.g., downscaling of climate projections), where computational resources are limited but high-resolution data are critical.