Note that I am not a researcher in machine learning, artificial intelligence, or a meteorologist. All opinions expressed here are those of a layman in this field.

For anybody who has been paying attention, you will have noticed that AI and ML (machine learning) have been improving at an incredible rate over the last decade. If you have spent any time chatting with one of the large language models (LLMs) and exploring their knowledge, it’s hard not to be astounded at their capabilities. Another striking fact is that the LLMs and other ML models we are experiencing now are the worst we will ever encounter—the technology will only get better.

Several teams have been applying ML techniques, and their ability to capture patterns and their evolution, to Numerical Weather Prediction (NWP). For example, Google published a paper describing GraphCast in November 2023. In early December 2024, Google published another paper on GenCast, which applies ML to ensemble weather prediction. These papers caught the attention of several LuckGrib customers who asked if they could gain access to the models.

Since mid-2024, ECMWF has been providing their own ML-based model, AIFS. In mid-December 2025 NOAA NWS made several AI/ML-based models available, including AIGFS and HyGEFS.

Of note is this quote from a paper discussing Hybrid GEFS:

… Recently, data-driven machine learning weather prediction (MLWP) models not only produce superior forecasts compared to traditional numerical weather prediction (NWP) models but also run with magnitudes smaller computing resources and are significantly faster than traditional NWP.

Deterministic vs Machine Learning

For the vast majority of Numerical Weather Prediction models produced to date—what are called deterministic models—the process consists of two major parts:

1. Capturing the initial atmospheric conditions as accurately as possible (data assimilation of observations from stations, buoys, satellites, aircraft, balloons, etc.).
2. Modeling the physical laws that govern how temperature, motion, moisture, radiation, and composition interact to evolve the atmosphere forward in time.

The physics in these models is necessarily an approximation, yet the simulations have become remarkably skillful.

A simplistic view of running a deterministic model is to start with the initial conditions, apply the physics equations for a short time step (often seconds), and repeat—outputting forecast fields every hour or so.

Machine learning models take a fundamentally different approach. They are not (primarily) physics-based. Experts sometimes liken ML weather models to sophisticated compression: during training on vast reanalysis datasets, the model learns statistical patterns of atmospheric evolution and encodes that knowledge in its weights. In effect, it has learned the effective physics of weather systems, including local influences from terrain and coastlines.

Training is extremely compute-intensive, but once trained, inference (running the model forward) is dramatically faster and cheaper than traditional NWP. For example, while ECMWF IFS or GFS can take many hours on supercomputers, AIFS or AIGFS can produce a global forecast in minutes on far more modest hardware.

ML models are still in their early phases

AIFS, AIGFS, and AIGEFS are all operational but still early in their development cycle. They already produce high-quality forecasts—often outperforming their physics-based counterparts in many metrics—yet they have limitations.

• GFS provides hourly output; AIGFS and AIFS currently step every 6 hours. This matters for rapidly evolving systems (hurricanes, sharp fronts).
• Traditional models produce hundreds of parameters; the current AI models offer only a subset. Parameters such as wind gusts, simulated radar, visibility, cloud cover fractions, wave/swell fields, and many diagnostic quantities (vorticity, divergence, etc.) are not yet available.

It is not yet clear how (or how easily) ML models will expand to include these derived fields, though research is active. In physics-based models, many of these parameters emerge naturally from the full 3D atmospheric state. Reproducing them in a purely data-driven framework may require new techniques.

All the same, this is an exciting time in numerical weather prediction. The speed and efficiency gains open fascinating possibilities for more frequent runs, higher resolution, or entirely new applications. I look forward to watching, and offering, these new models.