~/aixsci
weekly updates

models/GraphCast

GraphCast

GraphCast is a weather forecasting model that treats the globe as a web of connected points and learns, from records of past weather, how one atmospheric state turns into the next.

What it is

Traditional forecasts solve equations of fluid motion on a grid of boxes covering the planet. GraphCast instead learns the step from examples. The atmosphere is described by numbers at many points on the Earth's surface and at different heights: pressure, temperature, wind, moisture. Those points are joined into a mesh, a network of links between neighbours, so each point has a set of connected companions near and far. The model is a graph neural network: software that passes messages along those links, letting every point adjust its own numbers in the light of what its neighbours report. Repeating this a few times lets information travel across the mesh.

The model takes two snapshots of the atmosphere, the current one and the one just before, and outputs the snapshot a short interval later. Feeding that output back in produces the step after, and so on, building a forecast days ahead. It is trained by being shown long archives of reconstructed past weather and adjusted until its predicted next state matches what actually happened. Nothing in it encodes the physics of air and water directly; the patterns are drawn out of the record. Related designs keep a physics solver at the core and use learned pieces only for the parts the solver cannot resolve.

Why scientists use it

The appeal is cost and speed. Solving the equations of the atmosphere at fine resolution takes very large computers and considerable time, which limits how many forecasts can be run and how many alternative futures can be explored. A trained network does its arithmetic once per step and runs quickly on a single modern processor, so forecasts can be produced often, and many slightly different starting points can be tried to gauge uncertainty.

The mesh matters because weather is not purely local. Links that skip across long distances let distant conditions influence one another in a single pass, which is awkward to arrange when a model can only talk to immediate neighbours. The same thinking appears beyond weather, wherever quantities sit at connected points: molecules, power networks, river systems. In each case the model stands in for an explicit calculation of how the system evolves, trading an account of the mechanism for a learned shortcut between one state and the next.

In the record

Limits

A learned model knows only the weather it was trained on. Conditions outside that record, including a changed climate or very rare extremes, are where it is least trustworthy, and it tends to smooth sharp features rather than reproduce their full strength. Nothing guarantees that mass, energy or momentum are conserved, so forecasts can drift as steps accumulate. It also offers no account of why a pattern develops, which makes errors hard to diagnose. And it still depends on physics-based reconstructions of the past for its training and its starting conditions.

Written by AIxSci: the background from general knowledge, the examples from checked records.

Model family

Every article

1articles
0structural-biology
0materials-chemistry
0astronomy
1climate-weather