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

Tomasi, Elena; Franch, Gabriele; Cristoforetti, Marco

doi:https://doi.org/10.5194/gmd-18-2051-2025

Articles | Volume 18, issue 6

https://doi.org/10.5194/gmd-18-2051-2025

© Author(s) 2025. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/gmd-18-2051-2025

© Author(s) 2025. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 18, issue 6

Development and technical paper

|

01 Apr 2025

Development and technical paper |

| 01 Apr 2025

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

Elena Tomasi, Gabriele Franch, and Marco Cristoforetti

Data sets

Sample dataset for the models trained and tested in the paper 'Can AI be enabled to dynamical downscaling? Training a Latent Diffusion Model to mimic km-scale COSMO-CLM downscaling of ERA5 over Italy' Elena Tomasi, Gabriele Franch, and Marco Cristoforetti https://doi.org/10.5281/zenodo.12934521

2000–2002 Dataset [1/7] for the models trained and tested in the paper 'Can AI be enabled to dynamical downscaling? Training a Latent Diffusion Model to mimic km-scale COSMO-CLM downscaling of ERA5 over Italy' Elena Tomasi, Gabriele Franch, and Marco Cristoforetti https://doi.org/10.5281/zenodo.12944960

2003–2005 Dataset [2/7] for the models trained and tested in the paper 'Can AI be enabled to dynamical downscaling? Training a Latent Diffusion Model to mimic km-scale COSMO-CLM downscaling of ERA5 over Italy' Elena Tomasi, Gabriele Franch, and Marco Cristoforetti https://doi.org/10.5281/zenodo.12945014

2006–2008 Dataset [3/7] for the models trained and tested in the paper 'Can AI be enabled to dynamical downscaling? Training a Latent Diffusion Model to mimic km-scale COSMO-CLM downscaling of ERA5 over Italy' Elena Tomasi, Gabriele Franch, and Marco Cristoforetti https://doi.org/10.5281/zenodo.12945028

2009–2011 Dataset [4/7] for the models trained and tested in the paper 'Can AI be enabled to dynamical downscaling? Training a Latent Diffusion Model to mimic km-scale COSMO-CLM downscaling of ERA5 over Italy' Elena Tomasi, Gabriele Franch, and Marco Cristoforetti https://doi.org/10.5281/zenodo.12945040

2012–2014 Dataset [5/7] for the models trained and tested in the paper 'Can AI be enabled to dynamical downscaling? Training a Latent Diffusion Model to mimic km-scale COSMO-CLM downscaling of ERA5 over Italy' Elena Tomasi, Gabriele Franch, and Marco Cristoforetti https://doi.org/10.5281/zenodo.12945050

2015–2017 Dataset [6/7] for the models trained and tested in the paper 'Can AI be enabled to dynamical downscaling? Training a Latent Diffusion Model to mimic km-scale COSMO-CLM downscaling of ERA5 over Italy' Elena Tomasi, Gabriele Franch, and Marco Cristoforetti https://doi.org/10.5281/zenodo.12945058

2018–2020 Dataset [7/7] for the models trained and tested in the paper 'Can AI be enabled to dynamical downscaling? Training a Latent Diffusion Model to mimic km-scale COSMO-CLM downscaling of ERA5 over Italy' Elena Tomasi, Gabriele Franch, and Marco Cristoforetti https://doi.org/10.5281/zenodo.12945066

Pretrained models presented in the paper 'Can AI be enabled to dynamical downscaling? Training a Latent Diffusion Model to mimic km-scale COSMO-CLM downscaling of ERA5 over Italy' Elena Tomasi, Gabriele Franch, and Marco Cristoforetti https://doi.org/10.5281/zenodo.12941117

Model code and software

LDM_res v1.0 Gabriele Franch, Elena Tomasi, and Marco Cristoforetti https://doi.org/10.5281/zenodo.13356322

Short summary

High-resolution weather data are crucial for many applications, typically generated via resource-intensive numerical models through dynamical downscaling. We developed an AI model using latent diffusion models (LDMs) to mimic this process, increasing weather data resolution over Italy from 25 to 2 km. LDM outperforms other methods, accurately capturing local patterns and extreme events. This approach offers a cost-effective alternative, with potential disruptive application in climate sciences.