Bertoli et al. (2025) Revisiting Machine Learning Approaches for Short‐ and Longwave Radiation Inference in Weather and Climate Models

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Journal of Advances in Modeling Earth Systems
• Year: 2025
• Date: 2025-09-01
• Authors: Guillaume Bertoli, Salman Mohebi, Fırat Özdemir, Jonas Jucker, Stefan Rüdisühli, Fernando Pérez‐Cruz, Mathieu Salzmann, Sebastian Schemm
• DOI: 10.1029/2025ms004956

Research Groups

• NVIDIA (developers of the PyTorch-Fortran coupler)
• ICON model development team

Short Summary

This study evaluates several machine learning (ML) architectures as parameterizations for radiative transfer within the ICON weather and climate model on GPUs, finding that a physics-informed BiLSTM model achieves stability and performance comparable to classical physics-based schemes.

Objective

• To explore the effectiveness of various ML parameterizations for radiative transfer in the ICON model and investigate the computational speed-up and stability achieved when running on graphics processing units (GPUs).

Study Configuration

• Spatial Scale: Global aquaplanet simulations at a resolution of approximately $80\text{ km}$.
• Temporal Scale: Simulations spanning several weeks.

Methodology and Data

• Models used:
  • Climate Model: ICON (Icosahedral Nonhydrostatic).
  • ML Architectures: Multilayer Perceptron (MLP), Unet, Bidirectional Recurrent Neural Network with Long Short-Term Memory (BiLSTM), Vision Transformer (ViT), and Random Forest (RF).
  • Baseline: ecRad (physics-based radiative transfer parameterization).
  • Infrastructure: OpenACC compiler directives for GPU support and the PyTorch-Fortran coupler.
• Data sources: Not specified in the provided text.

Main Results

• The BiLSTM model was identified as the most accurate, provided it utilized a physics-informed normalization strategy, a heating rate penalty during training, Gaussian smoothing for post-processing, and simplified upper-level flux computations.
• ML models capable of maintaining ICON stability exhibited memory consumption and computational speed comparable to GPU-optimized classical physics parameterizations.
• The proposed setup successfully enabled stable aquaplanet simulations for several weeks at $80\text{ km}$ resolution.

Contributions

• Demonstrates a viable pipeline for coupling diverse ML architectures to the ICON model using a PyTorch-Fortran coupler on GPUs.
• Identifies the specific physics-informed constraints and post-processing steps necessary to ensure the numerical stability of ML-based radiative transfer in a global climate model.

Funding

• Not specified in the provided text.

Citation

@article{Bertoli2025Revisiting,
  author = {Bertoli, Guillaume and Mohebi, Salman and Özdemir, Fırat and Jucker, Jonas and Rüdisühli, Stefan and Pérez‐Cruz, Fernando and Salzmann, Mathieu and Schemm, Sebastian},
  title = {Revisiting Machine Learning Approaches for Short‐ and Longwave Radiation Inference in Weather and Climate Models},
  journal = {Journal of Advances in Modeling Earth Systems},
  year = {2025},
  doi = {10.1029/2025ms004956},
  url = {https://doi.org/10.1029/2025ms004956}
}