Revisiting Machine Learning Approaches for Short- and Longwave Radiation Inference in Weather and Climate Models

Aug 4, 2026

Upload by Quinn

This paper explores machine learning (ML) parameterizations for radiative transfer in the ICOsahedral Nonhydrostatic weather and climate model (ICON) and evaluates achieved speed-ups when ICON runs on GPUs. Five ML architectures—including an MLP, U-Net, BiLSTM, vision transformer, and a random forest baseline—are coupled to ICON using OpenACC-enabled GPU support and a PyTorch-Fortran coupler. The most accurate, stable configuration uses physics-informed normalization and heating-rate penalization during training plus postprocessing, yielding results comparable to ecRad over multi-week simulations.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages24

Reading Time1 h 0 min

AccessFree to download

Introduction

Radiative transfer in Earth system models

ecRad and ICON radiation parameterization

Machine learning models for radiation inference

Coupling and GPU implementation

Training strategies and stability assessment

Results and comparison with physics-based parameterizations

RESEARCH ARTICLE 10.1029/2025MS004956 Key Points: • The ICOsahedral Nonhydrostatic weather and climate model (ICON) is coupled to a neural network radiative transfer parameterization, and both run in tandem on graphics processing units • The ICON model with the radiation emulator is stable for several weeks, with a negligible difference compared to the physical parameterization • Physics‐informed normalization and heating rates penalization during training improve all tested machine learning architectures Correspondence to: G. Bertoli, [gb2956@columbia.edu](gb2956@columbia.edu) Citation: Bertoli, G., Mohebi, S., Ozdemir, F., Jucker, J., Rüdisühli, S., Perez‐Cruz, F., et al. (2025) . Revisiting machine learning approaches for short‐ and longwave radiation inference in weather and climate models. Journal of Advances in Modeling Earth Systems, 17, e2025MS004956 . [https://doi.org/10.1029/2025MS004956](https://doi.org/10.1029/2025MS004956) Received 17 JAN 2025 Accepted 17 AUG 2025 © 2025 The Author(s) . Journal of Advances in Modeling Earth Systems published by Wiley Periodicals LLC on behalf of American Geophysical Union. This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. Revisiting Machine Learning Approaches for Short‐ and Longwave Radiation Inference in Weather and Climate Models Guillaume Bertoli1,2 , Salman Mohebi3, Firat Ozdemir3 , Jonas Jucker4, Stefan Rüdisühli1,5 , Fernando Perez‐Cruz3,6,7, Mathieu Salzmann3, and Sebastian Schemm8 1Institute for Atmospheric and Climate Science, ETH Zurich, Zurich, Switzerland, 2Learning the Earth with Artificial Intelligence and Physics Center, Columbia University, New York, NY, USA, 3Swiss Data Science Center, ETH Zurich and EPFL, Zurich, Switzerland, 4Center for Climate Systems Modeling, ETH Zurich, Zurich, Switzerland, 5Now at Meteomatics, St. Gallen, Switzerland, 6Computer Science Department, ETH Zurich, Zurich, Switzerland, 7Now at Bank for International Settlements (BIS), Basel, Switzerland, 8Department of Applied Mathematics and Theoretical Physics, Cambridge University, Cambridge, UK Abstract This paper explores Machine Learning (ML) parameterizations for radiative transfer in the ICOsahedral Nonhydrostatic weather and climate model (ICON) and investigates the achieved ML model speed‐up with ICON running on graphics processing units (GPUs) . Five ML models, with varying complexity and size, are coupled to ICON; more specifically, a multilayer perceptron (MLP), a Unet model, a bidirectional recurrent neural network with long short‐term memory (BiLSTM), a vision transformer (ViT), and a random forest (RF) as a baseline. The ML parameterizations are coupled to the ICON code that includes OpenACC compiler directives to enable GPU support. The coupling is done with the PyTorch‐Fortran coupler developed at NVIDIA. The most accurate model is the BiLSTM with a physics‐informed normalization strategy, a penalty for the heating rates during training, a Gaussian smoothing as postprocessing and a simplified computation of the fluxes at the upper levels to ensure stability of the ICON model top. The presented setup enables stable aquaplanet simulations with ICON for several weeks at a resolution of about 80 km and compares well with the physics‐based default radiative transfer parameterization, ecRad. Our results indicate that the compute requirements of the ML models that can ensure the stability of ICON are comparable to GPU optimized classical physics parameterizations in terms of memory consumption and computational speed. Plain Language Summary Machine Learning (ML) methods could drastically accelerate existing parts of weather and climate models. This research explores machine learning methods to replace the radiation solver responsible for computing the solar and terrestrial radiative fluxes in the IC

FAQs

Which ICON radiation component is targeted for ML acceleration?

The ML parameterizations replace the radiation solver that computes solar and terrestrial radiative fluxes in ICON.

What ML models are evaluated in the study?

Five models are coupled to ICON: an MLP, U-Net, BiLSTM, vision transformer (ViT), and a random forest (RF) baseline.

What training strategy improves both accuracy and stability?

Physics-informed normalization and penalization of heating rates during training, along with postprocessing and simplified upper-level flux computation, improve performance and stability.

Aug 4, 2026

Upload by Quinn