AI4NWP
Introduction – AI-based NWP emulators:
Announcement: NEW: Our Real-time visualization page for purely AI-based weather models is now available: https://aiweather.cira.colostate.edu/
Fully AI-based models are emerging for global weather prediction.
These models include:
- Feb 2022: FourCastNet – by NVIDIA (US) (paper, code)
- Nov 2022: Pangu-Weather – by Huawei Cloud (China) (paper, code)
- Dec 2022: GraphCast – by Google (US) (paper, code)
- Jan 2023: ClimaX – by Microsoft (US) (paper, code)
- April 2023: FengWu – from China (paper)
- June 2023: FuXi – from China (paper)
- June 2023: Alibaba SwinRDM – from China (paper)
- June/July 2023: FourCastNet Version 2 (based on spherical harmonics) – from NVIDIA (US) (paper, Tech blog)
The code (and model weights) for some of these models is now available, allowing the community to run them locally and evaluate them. These models include:
- FourCastNet (NVIDIA, US),
- Pangu-weather (Huawei, China),
- GraphCast (Google, US), and
- FourCastNet Version 2 (NVIDIA, US) – soon to come.
- Thank you to ECMWF for developing a standard interface to install these models: https://github.com/ecmwf-lab/ai-models
These new models raise many questions, including:
- Can we trust these models? How exactly should they be evaluated?
- Do these models add value for forecasters? If so, for which applications?
- How should these models be integrated into NOAA applications?
At CIRA and NOAA-GSL we seek to address some of these questions. Our current focus:
- Develop a research agenda to evaluate AI-based global forecasting models.
- Run several of the main models locally, and display them in real-time on a publicly available website. Website has been set up and currently being tested internally. Link will be posted here.
- Consider outputs of the models in CIRA’s daily weather briefings to learn about their strengths and weaknesses in forecasting applications.
- Identify AI-specific weaknesses of these models.
- Evaluate how well existing AI models predict track and intensity of tropical cyclones (TCs)
Recent/upcoming presentations:
- 5th NOAA AI workshop
Sept 18-21, 2023
NOAA AI workshop schedule: https://www.noaa.gov/ai/events/5th-noaa-ai-workshop-2023
Invited presentations in Session 3 (Sept 20, 2023).
Link to Slides - 104th AMS Annual Meeting
Jan 28 – Feb 1, 2024, Baltimore MD & online
Three abstracts submitted.
https://annual.ametsoc.org/index.cfm/2024/
Publications:
- Short discussion in Washington Post article: How Big Tech AI models nailed forecast for Hurricane Lee a week in advance, Washington Post, Sept 21, 2023. https://wapo.st/3rvkOpP
- I. Ebert-Uphoff and K. Hilburn, The outlook for AI weather prediction, Nature, Vol. 619, 20 July 2023, p. 473-474. https://rdcu.be/df8LB.
Figure from Nature article:
CIRA/GSL scientists involved in this project:
Leads:
- Imme Ebert-Uphoff (CIRA lead)
- Jebb Stewart (NOAA-GSL lead)
Running models locally and visualization:
- Jacob T. Radford (CIRA & NOAA/GSL): works on getting models to run locally and developing all visualizations.
- Robert T. DeMaria (CIRA): works on getting models to run locally.
Evaluation of TC properties:
- Mark DeMaria (CIRA)
- Robert T. DeMaria (CIRA)
- Kate D. Musgrave (CIRA)
- Galina Chirokova (CIRA)
Introduced models to CIRA weather briefings:
- Kyle A. Hilburn (CIRA)
Other members involved in developing the research agenda (in alphabetical order):
- Jason Apke (CIRA)
- Randy Chase (CIRA)
- Jeffrey D. Duda (NOAA/GSL)
- Isidora Jankov (NOAA/GSL)
- Christina E. Kumler (CIRES & NOAA/GSL)
- Ryan A. Lagerquist (CIRA & NOAA/GSL)
- Yoonjin Lee (CIRA)
- David D. Turner (NOAA/GSL)
- Lander Ver Hoef (CIRA)
- Matt S. Wandishin (NOAA/GSL)
- Chuck White (CIRA)
Primary contacts:
Imme Ebert-Uphoff (CIRA; iebert@colostate.edu) and Jebb Stewart (NOAA-GSL; jebb.q.stewart@noaa.gov)
