FTorch: A library to couple PyTorch ML models with Fortran climate models
AI for Climate & Nature Community Day
University of Bologna
ICCS - University of Cambridge
ICCS - University of Cambridge
ICCS - University of Cambridge
ICCS - University of Cambridge
University of Kent
ICCS - University of Cambridge
2025-05-01
Motivation : Hybrid modeling
Climate components diagram courtesy of Jack Atkinson.
Globe grid with box by Caltech under Fair use.
Neural Net by 3Blue1Brown under fair dealing.
Facilitating hybrid modeling with FTorch
Many weather and climate models run on code written in Fortran
Today’s AI research happens primarily in Python.
- Language interoperation
- Computationally efficient
- Focus on ease of use and familiarity
Atkinson et al., (2025)
Mathematical Bridge by cmglee used under CC BY-SA 3.0
PyTorch, the PyTorch logo and any related marks are trademarks of The Linux Foundation.”
Collaborations: VESRI and beyond
- DataWave CAM-GW: Neural net parameterizations of gravity waves in CAM
- MiMA-ML: Found online behavior diverges from offline for nearly-identical offline gravity waves models. Mansfield and Sheshadri (2024)
- ICON: ML convection parameterizations work best when causal relations are eliminated. Heuer et al. (2023)
- GloSea6: Replaced BiCGStab solver with ML to speed up forecasting. Park and Chung (2025)
The ICCS received support from Schmidt Sciences
This project also received funding from a C2D3-Accelerate grant
Book a Code Clinic
