NOAA puts three AI global weather models into operation
NOAA has activated AIGFS, AIGEFS and HGEFS, three AI-based global weather forecasting systems. The new generation speeds up forecasts and drastically reduces supercomputing demands, though significant limitations remain.
The U.S. National Oceanic and Atmospheric Administration (NOAA) has today put three artificial intelligence-based global weather forecasting models into operation. It is a significant step because AI is moving beyond the lab and into the infrastructure that guides public warnings, planning and weather forecasts.
The agency has deployed AIGFS, AIGEFS and HGEFS, systems that complement its traditional numerical models. Those models solve physical equations to simulate the atmosphere and form the foundation of modern forecasting, but they require vast supercomputing resources. The new models learn patterns from historical and recent weather data to produce forecasts much faster.
A 16-day forecast in about 40 minutes
The biggest change comes with AIGFS, short for Artificial Intelligence Global Forecast System. According to NOAA, a 16-day forecast with this system takes approximately 40 minutes and uses just 0.3% of the computing resources required by the conventional operational GFS model. In other words, it cuts computing consumption by as much as 99.7%.
Speed alone does not guarantee a better forecast. In weather forecasting, however, having a useful simulation sooner allows specialists to compare more scenarios, update products and respond with more lead time when severe weather is approaching. NOAA says AIGFS improves GFS forecast skill across many large-scale patterns and significantly reduces tropical cyclone track errors at longer lead times.
There is one important caveat: the first version produces worse tropical cyclone intensity forecasts. Knowing where a storm may go and estimating how strong it will be are separate challenges. The latter remains particularly difficult because it depends on fast, local processes, such as interactions among the ocean, clouds and wind. NOAA identifies that limitation as an area for improvement in future versions.
Not a single forecast, but many possible futures
The second system, AIGEFS, applies AI to ensemble forecasting. A weather ensemble does not provide one definitive answer. Instead, it runs multiple simulations with small variations to estimate which scenarios are most likely and how much uncertainty remains.
That information is essential for decisions such as whether a potential cyclone track warrants additional monitoring even if it is not yet the leading scenario. Emergency managers, power companies, airlines and farmers need more than an indication of the expected weather; they also need to understand the range of plausible outcomes.
NOAA’s early evaluations indicate that AIGEFS outperforms the traditional GEFS and extends forecast skill by 18 to 24 hours. In practice, that margin can amount to nearly an extra day of useful guidance on the evolution of large-scale weather systems.
The hybrid model avoids choosing between physics and AI
The third piece is HGEFS, or Hybrid-GEFS. Rather than replacing the physics-based model with an AI model, it combines AIGEFS members with those from the conventional GEFS in a large hybrid ensemble. NOAA says initial testing shows consistent performance above that of systems based exclusively on AI or exclusively on physics.
The approach is significant. Physics-based models explicitly incorporate known atmospheric laws, while AI models stand out for their speed and their ability to extract patterns from massive volumes of data. A hybrid system aims to harness both strengths: AI’s efficiency and the robustness of physical simulation.
For meteorologists, these products are additional guidance, not an automatic replacement for their work. Operational forecasting requires comparing models with satellite, radar and ground-station observations, as well as assessing local effects that a global model may not resolve in sufficient detail.
AI enters a critical public service
NOAA considers this deployment the first operational adoption of global AI models within a national weather agency’s forecasting infrastructure. Its significance goes beyond reducing computing costs. If quality is maintained or improved using a fraction of the resources, weather centers can devote more capacity to running larger ensembles, testing new configurations and producing guidance with less delay.
The real test begins now, as the models face full seasons and all kinds of weather events. The agency will have to show that the improvements seen in testing hold up in operation and address weaknesses such as cyclone intensity forecasts. But the deployment of AIGFS, AIGEFS and HGEFS has already established a clear direction: weather forecasting is not heading toward a choice between physics-based supercomputing and AI, but toward combining them wherever each adds the most value.