Atmospheric Radiation Parameterization by Neural Ordinary Differential Equations and Related Models

Nikolay Yavich, Stefan Maria Ailuro, Chesnokov Alexey Aleksandrovich, Alexander Gavrikov, Evgeny Burnaev, Vladimir Vanovskiy

Keywords: regression models, neural ODEs, numerical weather prediction, neural networks, radiative transfer

TL;DR: We investigated the use of neural ordinary differential equations for emulation of atmospheric radiation transfer and designed a fast and accurate emulator.

Abstract:

Radiation parameterization schemes are crucial components of weather and climate models, however, they are known to be computationally intensive. Alternatively, they can be emulated with machine learning (ML) regression models. Mainly vertical energy propagation motivates the usage of ML models featuring sequential data processing. We investigate these and related models for radiation parameterization using atmospheric data modeled within an Arctic region. We observe that Neural ODE performs best in predicting both the long- and short-wave heating rates. Furthermore, we substitute the architecture with its discrete form to boost its efficiency while preserving competitive performance. The practical applicability of the models is studied for different model sizes. Finally, we link the trained neural network to the operational weather forecast model and assessed its performance versus the conventional radiation parameterization.
We receive a speedup of 26.5 times of the radiation steps
without significant loss of accuracy. The proposed parameterization
emulator dramatically reduces the computational burden and the carbon footprint of weather forecasting.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 12026