seasonmap field guide

How AI weather models actually work

In 2023, models with no physics equations inside started beating fifty years of supercomputer meteorology on headline scores. Here's what's actually happening in there, why they run a million times cheaper — and the one systematic flaw you need to price in when you read them.

Physics models solve; AI models remember

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 does something different. It's trained on decades of reanalysis — the best possible reconstruction of every day of global weather since 1979 — until it learns one function extremely well: given the full 3-D state of the atmosphere now, what does it look like six hours from now? No equations of motion anywhere. Just a few hundred million learned weights encoding how weather has always evolved.

To make a forecast, you hand it the current analysis and apply that six-hour jump autoregressively — output becomes input, sixty-plus times in a row — and you've got a 16-day forecast. That's the whole trick.

Why they're absurdly fast

Training takes weeks on serious hardware, but that cost is paid once. Running the trained model is just matrix multiplication: a global 10-day forecast takes about a minute on a single GPU, or under an hour on an ordinary CPU — versus supercomputer-hours for physics models. That's not an incremental improvement; it's the reason a one-person weather site can operationally run its own global model. (Ours is below.)

The family tree

Pangu-Weather (Huawei, 2023) — the first to beat the physics benchmark on headline scores in a Nature paper; a 3-D transformer with openly published weights.
GraphCast (Google DeepMind, 2023) — a graph neural network on an icosahedral mesh; the lineage behind NOAA's experimental AI-GFS runs.
AIFS (ECMWF, 2024–25) — the physics-model world champion building its own challenger, now run operationally alongside the physics IFS.

The flaw: they hedge

Train a model to minimize average error and it learns a rational habit: when uncertain, predict something smooth. A physics model will commit to a 962 mb bomb cyclone and be wrong about the details; an AI model will paint the same storm as 975 mb — right place, right timing, blunted peak. The same blurring shows up in precipitation bands, wind maxima, anything sharp and consequential.

So read them accordingly: AI models are excellent for where the pattern goes — jet position, ridge/trough placement, storm track — and habitually conservative on how extreme it gets. Pattern from the AI models, amplitude from physics models and ensembles, is a genuinely good workflow.

Ours, in public: seasonmap runs Pangu-Weather operationally as CAESAR-OP — a daily rollout out to 16 days from fresh ECMWF open-data analyses, with click-anywhere soundings, an adaptive bias correction trained on our own verification history, and a perturbed-member ensemble arm (CAESAR-ENS) publishing mean and spread. Every run is scored against the RTMA analysis on the public board, next to the models it's trying to beat. When the smoothness tax shows up in our RMSE, you'll see it.

Watch our AI model run live →