Contribution of physical latent knowledge to the emulation of an atmospheric physics model: a study based on the LMDZ Atmospheric General Circulation Model

Crossouard, Ségolène; Thao, Soulivanh; Dubos, Thomas; Kageyama, Masa; Vrac, Mathieu; Meurdesoif, Yann

doi:10.5194/gmd-19-5907-2026

Articles | Volume 19, issue 13

https://doi.org/10.5194/gmd-19-5907-2026

© Author(s) 2026. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/gmd-19-5907-2026

© Author(s) 2026. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 19, issue 13

Development and technical paper

|

07 Jul 2026

Development and technical paper |

| 07 Jul 2026

Contribution of physical latent knowledge to the emulation of an atmospheric physics model: a study based on the LMDZ Atmospheric General Circulation Model

Ségolène Crossouard, Soulivanh Thao, Thomas Dubos, Masa Kageyama, Mathieu Vrac, and Yann Meurdesoif

Supplement

https://doi.org/10.5194/gmd-19-5907-2026-supplement

Data sets

Replication Code and Data for: Crossouard et al. 2025 "Contribution of physical latent knowledge to the emulation of an atmospheric physics model: a study based on the LMDZ Atmospheric General Circulation Model" Ségolène Crossouard et al. https://doi.org/10.7910/DVN/3UFU9J

Model code and software

Replication Code and Data for: Crossouard et al. 2025 "Contribution of physical latent knowledge to the emulation of an atmospheric physics model: a study based on the LMDZ Atmospheric General Circulation Model" Ségolène Crossouard et al. https://doi.org/10.7910/DVN/3UFU9J

Short summary

Current atmospheric models are limited by the computational time required for physical processes, known as physical parameterizations. To address this, we developed neural network-based emulators to replace these parameterizations in the IPSL climate model, using a simplified aquaplanet setup and a realistic configuration. We found that incorporating some physical knowledge, such as latent variables, into the learning process can improve predictions.