Assessing the physical consistency of high-resolution meteorological variables from a Downscaling Latent Diffusion Model

In recent decades, there has been a growing demand for high-resolution datasets of meteorological variables. Generally, this has been achieved through dynamical downscaling using numerical models. However, to date, km-scale resolution has only been feasible for limited regions or for a single simulation, preventing uncertainty estimation. Recently, deep learning downscaling algorithms have emerged as an alternative to overcome these limitations due to their computational efficiency and ability to reconstruct fine-scale features and flow characteristics. However, while traditional methods rely on physical equations to model the evolution of the atmospheric state, most machine learning algorithms are not constrained by physical laws, and it remains unclear whether the physical consistency of the output fields is preserved. In this study, we address this gap by assessing the physical consistency of results from a Latent Diffusion Model (LDM) trained to emulate the dynamical downscaling from ERA5 reanalysis to CERRA over the Mediterranean region, with a focus on the Italian peninsula. The LDM has already been tested for generating high-fidelity ground variables such as 2m temperature and 10m wind components [1], and precipitation [2] emulating the COSMO-CLM model. Here, we instead leverage the LDM model to downscale dynamical variables at different levels to test its ability to reconstruct a wider configuration of the atmospheric state. The focus variables are mean sea level pressure, geopotential height, specific humidity, and meridional and zonal wind components at distinct pressure levels, which are used as both predictors and target variables. We will present preliminary results showing a first evaluation of the model results, exploring different physical diagnostic constraints, with a specific focus on mass conservation, the thermodynamic relationships between temperature and moisture, and the ratio between geostrophic and ageostrophic wind. By establishing a benchmark for the physical reliability of DL-downscaling techniques in regional climate applications, our work aims to enhance the credibility of such methods and facilitate their broader adoption within the atmospheric sciences. Moving forward, future work will extend this framework to evaluate the model for local-area forecasting applications.