seasonmap field guide

How AI weather models actually work

Since 2023, machine-learning weather models trained on reanalysis have matched or exceeded physics-based models like the GFS and ECMWF IFS on headline medium-range accuracy scores. This page covers how these models are trained and run, and the systematic bias to account for when reading their forecasts.

Physics models integrate; AI models learn a transition function

A traditional model like the GFS or the Euro is a physics engine: it divides the atmosphere into billions of cells and integrates the fluid-dynamics and thermodynamics equations forward, time-step by tiny time-step, on a supercomputer. Every forecast is computed from first principles, every time.

An AI model is trained differently. It learns from decades of reanalysis — the best available reconstruction of global weather since 1979 — to approximate one function: given the full 3-D state of the atmosphere now, what does it look like six hours from now? There are no equations of motion inside; the trained model is a few hundred million weights encoding that six-hour transition as observed in the historical record.

A forecast is produced by feeding the model the current analysis and applying that six-hour step autoregressively — each output becomes the next input, repeated sixty-plus times to reach 16 days.

Inference cost

Training takes weeks on large compute clusters, but that cost is paid once. Running the trained model is matrix multiplication: a global 10-day forecast takes about a minute on a single GPU, or under an hour on an ordinary CPU, versus the supercomputer-hours required by physics models. This cost difference is why a small operation can run its own global model daily.

Current models

• Pangu-Weather (Huawei, 2023) — the first to match physics-model accuracy on headline scores, published in Nature; a 3-D transformer with openly released weights.
• GraphCast (Google DeepMind, 2023) — a graph neural network on an icosahedral mesh, published in Science; the lineage behind NOAA's experimental AI-GFS runs.
• AIFS (ECMWF, 2024–25) — ECMWF's own data-driven model, now run operationally alongside the physics-based IFS.

Systematic smoothing bias

A model trained to minimize average error converges on a specific behavior: when uncertain, it predicts something smooth. A physics model will commit to a 962 mb bomb cyclone and may be wrong about the details; an AI model will tend to render the same storm at roughly 975 mb — correct on placement and timing, blunted on peak intensity. The same effect appears in precipitation bands and wind maxima, wherever the true field has a sharp gradient.

The practical read: AI models carry strong skill on pattern — jet position, ridge/trough placement, storm track — and are conservative on amplitude, i.e. how extreme a feature gets. Taking pattern from AI output and intensity from physics models and ensembles is a workflow that accounts for this bias.