~/aixsci
weekly updates

models/Pangu-Weather

Pangu-Weather

Pangu-Weather is a weather forecasting model that learns from past atmospheric data and treats the air as a three-dimensional body, so height is handled as a real dimension rather than a pile of separate flat maps.

What it is

Most computer forecasts solve equations of physics on a grid. A learned forecast model does something different: it is shown a vast archive of past atmospheric states and trained to predict what the atmosphere looked like a few hours later. Pangu-Weather does this with a transformer, a kind of neural network that works by letting every part of its input compare itself with every other part, weighing which neighbours matter. Because it learns a direct step from one atmospheric state to the next, a forecast is made by applying that step again and again, rather than by simulating the physics in detail.

Its particular idea is geometry. The atmosphere is not flat. Pressure levels, the stacked layers used to describe height in the air, are handled together with latitude and longitude as one three-dimensional volume. The model also builds in the fact that the Earth is a sphere and that a given place on it has fixed features, such as land or sea. So a pattern can be learned across height as well as across the map.

Why scientists use it

Scientists use learned weather models because, once trained, running them is cheap. A physics simulation of the whole atmosphere needs a large machine for every forecast. A trained network makes its prediction by passing data through fixed arithmetic, so many forecasts can be produced quickly, and ensembles, many slightly different forecasts run together to show the spread of possible weather, become easier to afford.

The three-dimensional treatment matters because weather is vertical. Air rises, sinks and overturns, and conditions high up shape conditions at the surface. A model that stacks independent flat maps must learn vertical relations indirectly. One that treats height as a dimension can represent them directly. The same reasoning appears well beyond weather: medical scans, ocean data and simulations of materials are all volumes, and models that respect that geometry tend to need less data to learn the same structure.

In the record

Limits

A learned model only knows the atmosphere it was trained on. Weather unlike anything in the archive, or a shifting climate, may be forecast poorly, and the model cannot explain itself in terms of physical laws. Applying a short learned step repeatedly tends to blur fine detail and drift as the forecast runs on. It also depends on carefully prepared, consistent input data: fed observations processed differently from its training archive, its skill can fall away without warning.

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