~/aixsci
weekly updates

models/GenCast

GenCast

GenCast is a weather forecasting system built with machine learning that does not give one prediction, but generates many plausible versions of the coming weather, so forecasters can see the range of what might happen.

What it is

Most computer weather forecasts work by solving equations for how air, heat and moisture move. GenCast instead belongs to a family of machine learning models that learn patterns from past weather records. It is a diffusion model: it starts from random noise, which is a field of meaningless values, and then repeatedly refines that noise until it looks like a believable map of the atmosphere a short time ahead. Because the starting noise is different each time, running it again gives a different but equally believable forecast. Chain those short steps together and you get a forecast that reaches days into the future.

Running the model many times produces what forecasters call an ensemble: a collection of possible futures rather than a single answer. If most members of the ensemble show a storm, confidence is high. If they disagree, the weather is genuinely uncertain. The model learns from archives of past atmospheric states, which record things like temperature, pressure, wind and humidity at many heights above the ground and across a grid covering the world. Learning from those records is how it comes to produce fields that behave like real weather rather than noise.

Why scientists use it

Weather is sensitive to tiny differences. Two atmospheres that look nearly identical today can differ sharply in a week, so a single forecast hides the most useful information: how likely each outcome is. Traditional forecast centres build ensembles by nudging the starting conditions and running the heavy physics calculations many times over, which is costly. A model that learns to generate plausible states directly can produce each ensemble member far more cheaply, so more members become affordable.

The approach also shows how one idea travels. Diffusion models, the method of turning noise into structure step by step, were developed for generating images. Here the same machinery generates maps of the atmosphere. What it stands in for is the expensive repeated simulation: the model is not reasoning about physics, it is reproducing the statistics of weather it has seen, including how uncertain weather tends to be.

In the record

Limits

A model that learns from the past can only generate weather resembling what it has seen, so genuinely unusual events may be smoothed away or underestimated. It carries any biases in the archive it learned from. The spread of its ensemble looks like uncertainty but is a learned statistical pattern, not a physical calculation, so it can be confidently wrong. And because no equations are being solved, there is no physical account of why a particular forecast came out as it did.

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