Physical consistency of high-resolution meteorological fields from deep learning-based downscaling

High-resolution meteorological fields are essential for assessing localized impacts of climate change. In recent years, deep learning (DL)-based downscaling techniques have emerged as a computationally efficient and sustainable alternative to dynamical downscaling, demonstrating strong skill in reconstructing fine-scale features and complex flow characteristics. 

Despite these advances, the spatio-temporal physical consistency of DL-downscaled fields remains a critical concern. Because most machine learning (ML) approaches are not explicitly constrained by physics-based equations, their ability to respect fundamental atmospheric balances is still  debated. Similar concerns apply to global data-driven forecasting models, which have revolutionized medium-range weather prediction in recent years but often operate as "black boxes" [1]. 

While the physical integrity of global ML forecasting models has started to receive attention, high-resolution downscaling applications remain largely unexplored from this perspective. In this study, we address this gap by assessing the physical consistency of a Latent Diffusion Model (LDM) - based on the architecture developed by Tomasi et al. 2025 [2] - trained to downscale ERA5 [3] over the Italian peninsula using CERRA [4] as the target dataset. 

Moving beyond standard statistical error metrics, we evaluate the model using a suite of physical diagnostic constraints,  with particular emphasis on mass conservation and thermodynamic relationships between temperature and moisture. Our results provide a benchmark for the physical reliability of DL-downscaling techniques in regional climate applications, thereby enhancing their credibility and facilitating their broader integration within the atmospheric sciences.

[1] Hakim, G. J., and S. Masanam, 2024: Dynamical Tests of a Deep Learning Weather Prediction Model. Artif. Intell. Earth Syst., 3, e230090, https://doi.org/10.1175/AIES-D-23-0090.1.

[2] Tomasi, E., Franch, G., and Cristoforetti, M.: Can AI be enabled to perform dynamical downscaling? A latent diffusion model to mimic kilometer-scale COSMO5.0_CLM9 simulations, Geosci. Model Dev., 18, 2051–2078, https://doi.org/10.5194/gmd-18-2051-2025, 2025.

[3] Hersbach H, Bell B, Berrisford P, et al. The ERA5 global reanalysis. Q J R Meteorol Soc. 2020; 146: 1999–2049. https://doi.org/10.1002/qj.3803

[4] Ridal, M., Bazile, E., Le Moigne, P., Randriamampianina, R., Schimanke, S., Andrae, U., et al. (2024) CERRA, the Copernicus European Regional Reanalysis system. Quarterly Journal of the Royal Meteorological Society, 150(763), 3385–3411. https://doi.org/10.1002/qj.4764