NOAA deploys new generation of AI-driven global weather models

NOAA announced a groundbreaking new suite of operational global weather forecasting models powered by artificial intelligence (AI), significantly improving the speed, efficiency and accuracy of forecasts. This model allows forecasters to quickly provide more accurate guidance while using a fraction of their computational resources.

“NOAA's strategic application of AI is a major step forward in the innovation of American weather modeling,” said NOAA Administrator Dr. Neil Jacobs. “These AI models reflect NOAA's new paradigm of improving large-scale weather and tropical track accuracy and rapidly delivering predictive products to meteorologists and the public at low cost by significantly reducing computational costs.”

The new suite of AI weather models includes three different applications:

• AIGFS (Artificial Intelligence Global Forecasting System): A weather forecast model that implements AI to provide improved weather forecasts faster and more efficiently than traditional models (using up to 99.7% less computing resources).
• AIGEFS (Artificial Intelligence Global Ensemble Forecast System): An AI-based ensemble system that provides different forecast outcomes to meteorologists and decision makers. Initial results show improved performance over traditional GEFS, extending predictive skill by an additional 18 to 24 hours.
• HGEFS (Hybrid-GEFS): A pioneering hybrid “grand ensemble” that combines the new AI-based AIGEFS (above) with NOAA's flagship ensemble model, the Global Ensemble Forecast System. Initial testing shows that the model is a first-of-its-kind approach for operational weather centers, consistently outperforming both AI-only and physics-only ensemble systems.

Learn more about the new AI operating model

AIGFS — A new AI-based system that uses a variety of data sources to generate weather forecasts comparable to those produced by traditional weather forecasting systems such as GFS.

• Performance: Demonstrates improved predictive skill over traditional GFS for many large-scale features. In particular, it has been demonstrated that tracking errors for tropical cyclones are significantly reduced with longer lead times.
• Efficiency: AIGFS' most innovative feature. A single 16-day forecast uses only 0.3% of the production GFS's computing resources and takes approximately 40 minutes. This reduced latency allows forecasters to obtain critical data more quickly than from traditional GFS.

• Room for future improvements: Although track predictions have been improved, v1.0 shows a decline in tropical cyclone intensity predictions, which we plan to address in future versions.

Aigehus — An AI-based 31-member ensemble similar to GEFS that offers a broader range of possibilities for weather forecasters and decision makers rather than a single predictive model solution.

• Performance: Prediction skill is comparable to operational GEFS.
• Efficiency: Requires only 9% of GEFS' computing resources in operation.
• Room for future improvements: Developers continue to improve the ensemble's ability to produce different prediction results.

HGEFS — The most innovative application in the new suite. HGEFS is a 62-member “grand ensemble” created by combining 31 members of the physical GEFS and 31 members of the AI-based AIGEFS.

• Performance: By combining two different modeling systems, one physics-based and one AI-based, HGEFS creates a larger, more robust ensemble that more effectively represents prediction uncertainty. As a result, HGEFS consistently outperforms both GEFS and AIGEFS on most key validation metrics.
• NOAA First: To our knowledge, NOAA is the first organization in the world to implement such a hybrid physics and AI ensemble system.
• Areas for Future Improvement: NOAA continues to work to improve HGEFS hurricane intensity predictions.

NOAA and industry-wide efforts

This initial model suite is the result of Project EAGLE, a joint initiative of NOAA's National Weather Service, Ocean and Atmospheric Research Institute, NOAA's National Environmental Prediction Center's Environmental Modeling Center, and the Earth Prediction Innovation Center.

“With Project EAGLE and the Earth Prediction Innovation Center, NOAA scientists continue to collaborate with members of academia and the private sector to further advance prediction technology,” Jacobs added.

The team leveraged Google DeepMind's GraphCast model as an initial foundation and used NOAA's proprietary Global Data Assimilation System analysis to fine-tune the model. This additional training with NOAA data improved the performance of the Google model, especially when using GFS-based initial conditions.