NOAA has deployed its first operational AI-driven weather forecasting system, a suite of three models that use a fraction of the computing resources required by traditional physics-based approaches while delivering comparable or better accuracy.
The systems went live January 5, 2026, marking a shift from experimental AI weather prediction to actual operational use at the National Weather Service.
The Three Models
AIGFS (Artificial Intelligence Global Forecast System) is the core AI model. A single 16-day forecast uses only 0.3% of the computing resources required by the traditional Global Forecast System and completes in approximately 40 minutes. The model shows improved forecast skill over the traditional GFS for many large-scale weather features and demonstrates significantly reduced tropical cyclone track errors at longer lead times.
AIGEFS (Artificial Intelligence Global Ensemble Forecast System) generates multiple forecast scenarios to capture uncertainty. Its forecast skill matches the operational GEFS while requiring only 9% of the computing resources.
HGEFS (Hybrid Global Ensemble Forecast System) combines 31 members from the physics-based GEFS with 31 members from the AI-based AIGEFS into a 62-member “grand ensemble.” This hybrid approach consistently outperforms both parent systems across most verification metrics.
How It Works
The AI models learn atmospheric patterns and behaviors from decades of historical weather data. Rather than solving physics equations from first principles like traditional numerical weather prediction, they recognize patterns and extrapolate forward.
“These AI models reflect a new paradigm for NOAA in providing improved accuracy for large-scale weather and tropical tracks, and faster delivery of forecast products to meteorologists and the public at a lower cost through drastically reduced computational expenses,” said Neil Jacobs, NOAA administrator.
The models extend forecast life by potentially 18-24 hours compared to traditional systems—a meaningful improvement for emergency preparedness.
The 91-99% Efficiency Claim
The energy reduction numbers are striking: AI programs require 91-99% less computing power than traditional models. However, these percentages exclude the energy-intensive training phase. Training a foundational AI weather model requires significant one-time computational investment, though operational forecasts afterward cost far less.
This tradeoff may prove favorable at scale. Training costs are amortized across thousands of daily forecasts over years of operational use.
What This Means
For forecasters, faster model runs mean more frequent forecast updates and quicker response to developing weather systems. For the public, this translates to earlier warnings for hurricanes, severe thunderstorms, and other hazardous weather.
The reduced computational cost could also enable higher-resolution regional models or more ensemble members—both of which improve forecast confidence.
The Fine Print
NOAA acknowledges the systems need refinement. Hurricane forecasts in particular require ongoing work, as does the diversity of variable outcomes produced by the ensemble system.
The AI models are not replacing traditional physics-based systems—they run alongside them, using the GFS as a framework and information source. Meteorologists can compare outputs and apply judgment.
There’s also a known limitation: some AI weather models produce cold-biased mean temperatures, resembling climates from 15 to 20 years earlier than their prediction period. This suggests limited training exposure to recent extreme heat events. Whether NOAA’s operational models share this bias is unclear from available documentation.
The systems emerged from Project EAGLE (Experimental AI Global and Limited-area Ensemble), a multi-year collaboration across NOAA’s Office of Oceanic and Atmospheric Research, the National Weather Service, academia, and industry partners.