Reproduce “Can AI be enabled to dynamical downscaling? A Latent Diffusion Model to mimic km-scale COSMO5.0\_CLM9 simulations” on OpenResearch

About this paper

Elena Tomasi, Gabriele Franch, Marco Cristoforetti

Researchers at Fondazione Bruno Kessler applied a Latent Diffusion Model with a residual approach (LDM_res) to emulate high-resolution dynamical downscaling of meteorological fields. This model generated realistic 2-m temperature and 10-m wind speed data, accurately reconstructing fine-scale features and reducing bias compared to traditional deep learning models, at a significantly faster inference speed than numerical simulations.

Key insights

• Advanced generative AI, particularly Latent Diffusion Models, can effectively emulate physically-based numerical dynamical downscaling processes, extending beyond simpler super-resolution tasks.
• The proposed residual training approach significantly enhances the LDM's performance, enabling it to focus on and accurately reconstruct fine-scale, locally-driven atmospheric variability.
• LDMs offer robust and realistic generation of high-resolution atmospheric fields, overcoming the instability and mode collapse issues often observed in Generative Adversarial Networks.

Summary by alphaXiv.