# AIGFS by NOAA Canonical page: https://technooptimist.io/ai/aigfs Frontier: [AI](https://technooptimist.io/ai.md) · Field: Weather models · Company: [NOAA](https://technooptimist.io/companies/noaa.md) Stage: Deploy (4 of 5: Research → Proto → Pilot → Deploy → Scaling) Updated: last update 2026-07-27, facts checked 2026-10-11 Goal: Give US forecasters global forecasts as good as or better than the physics GFS at a sliver of the computing cost, so more and faster runs fit on the same machines. Status: Operational since Dec 2025; v1.1 (July 2026) improved hurricane intensity and long-range blur. Next milestone: Improve AI ensemble spread and hybrid intensity forecasts, target 2026 ## Milestones | Status | Milestone | Date | Slip | | --- | --- | --- | --- | | Done | AIGFS, AIGEFS and hybrid HGEFS go operational | 17 Dec 2025 | | | Done | AIGFS v1.1: better hurricane intensity, less long-range blur | 27 Jul 2026 | | | Current | Improve AI ensemble spread and hybrid intensity forecasts | 2026 | | ## Upcoming events - **2026**: Improve AI ensemble spread and hybrid intensity forecasts ## Most important updates - **27 Jul 2026**: NOAA upgrades AIGFS to v1.1 to sharpen hurricane intensity and rainfall forecasts. The National Weather Service switched AIGFS to version 1.1 from the 12:00 UTC run on 27 July 2026. ([our story](https://technooptimist.io/stories/noaa-upgrades-aigfs-to-v1-1-to-sharpen-hurricane-intensity-and-rainfall-c559bd53.md) · [original source](https://www.weather.gov/media/notification/pdf_2026/scn26-68_AIGFS_v1.1.pdf)) - **17 Dec 2025**: NOAA puts AI global weather models into operation, led by a retrained GraphCast. NOAA launched three operational AI forecast systems: AIGFS (one best-guess forecast), AIGEFS (a 31-member AI ensemble) and HGEFS (a 62-member hybrid of AI and physics runs). ([our story](https://technooptimist.io/stories/noaa-puts-ai-global-weather-models-into-operation-led-by-a-retrained-cb104e08.md) · [original source](https://www.noaa.gov/news-release/noaa-deploys-new-generation-of-ai-driven-global-weather-models)) ## Current obstacles - **Peak winds in hurricanes**: Training to minimise average error rewards smooth forecasts, so storms come out too weak. v1.1 changed the loss to keep sharp detail. - **Blurring at long lead times**: Forecasts get fuzzier days out. v1.1 trains the model on its own 72-hour forecasts to reduce the blur. ## Physics limits - **Chaos caps how far ahead weather is knowable**: Tiny errors in today's state double every few days, so beyond about two weeks no model, AI or physics, can forecast daily weather. - **AI learns only weather it has seen**: Trained on past decades, an AI model has few examples of record-breaking extremes and may underplay them in a warming climate. ## How it works ### Learn the weather: Trained on past atmospheres The network learned how the global atmosphere changes from one step to the next by studying decades of past weather states. ### Start point: Today's analysis in, forecast out Each run starts from NOAA's best estimate of the current atmosphere and steps forward repeatedly to 16 days, in about 40 minutes. ### Ensemble: Many runs, a range of outcomes 31 slightly different starting states give 31 forecasts; how much they disagree tells forecasters how sure to be. ### Hybrid: Pool AI with physics HGEFS adds the 31 AI runs to the 31 physics runs; mixing two kinds of errors gives a better combined picture. ## Spec sheet | Spec | AIGFS | Provenance | | --- | --- | --- | | Compute vs physics GFS (16 days) | 0.3% | reported | | 16-day forecast run time | About 40 minutes | reported | | AI ensemble members (AIGEFS) | 31 | reported | | AIGEFS compute vs physics GEFS | About 9% | reported | | Hybrid ensemble members (HGEFS) | 62 (31 AI + 31 physics) | reported | | AIGEFS extra useful lead time | 18 to 24 hours (early results) | reported | | Base model | GraphCast (Google DeepMind), fine-tuned | reported | Provenance: reported = stated by the company; estimated = our estimate; sample = a sample figure. ## Papers and demos - **Dec 2025** (code): [Project EAGLE change log: AIGFS v1.0 built on GraphCast](https://epic.noaa.gov/ai/eagle-change-log/). Documents the base model, training data and the loss changes made to improve hurricane intensity. ## About NOAA The US National Oceanic and Atmospheric Administration runs the National Weather Service and its forecast models. Since December 2025 it also runs AI forecast models alongside its physics-based GFS. - Runs AIGFS, an AI global model built by retraining DeepMind's GraphCast on NOAA's own weather analyses. - A 16-day AIGFS forecast uses 0.3% of the computing of the physics GFS and finishes in about 40 minutes. - HGEFS blends 31 AI and 31 physics forecasts; NOAA says no other forecaster runs such a hybrid ensemble. - Headquarters: US - Status: government - Website: https://www.noaa.gov - People: Neil Jacobs (NOAA Administrator) - Partners, customers, investors: Google DeepMind (supplier); Earth Prediction Innovation Center (EPIC) (partner); WindBorne Systems (supplier) Company page: https://technooptimist.io/companies/noaa.md ## Related programs Everything in AI: https://technooptimist.io/ai.md --- The Techno Optimist tracks every frontier of technology, program by program: https://technooptimist.io/ (a Markdown version of any page: add .md to its URL; site map for agents: https://technooptimist.io/llms.txt). Data API (JSON, realtime stream): https://technooptimist.io/developers. People read 5 pages free, then Reader $10/month; these Markdown pages are open to agents.