Pangu-Weather is the first learned model to beat operational forecasting

Bi and colleagues at Huawei Cloud published Pangu-Weather in Nature on 5 July 2023. Trained on 39 years of recorded atmospheric data, it was the first machine-learned model reported to beat the operational numerical forecast on accuracy as well as speed, producing a global forecast in about 1.4 seconds on one graphics processor.

Background

Numerical weather prediction integrates the equations of atmospheric physics forward on a grid, and the European Centre’s system is the standard everyone else is measured against.

Machine-learned models had been closing on it. Nvidia’s FourCastNet had shown in 2022 that a learned model could be fast and roughly competitive, and the Pangu team names it as the work they built on. What none of them had done was win.

The authors identify two reasons. Earlier models worked in two dimensions and handled the uneven vertical structure of the atmosphere badly. And they accumulated error, because a long forecast was made by running a short one repeatedly.

What happened

Pangu-Weather was published in Nature on 5 July 2023, by Bi and colleagues at Huawei Cloud, having appeared as a preprint the previous November.

Their answer to the first problem is an architecture that treats the atmosphere as genuinely three dimensional, and adjusts for the fact that grid points are further apart at the equator than near the poles.

Their answer to the second is to train separate models for different forecast lengths, at one hour, three hours, six hours and a day, and then combine them so that a seven-day forecast needs few steps rather than many. Fewer steps means less accumulated error.

Trained on 39 years of recorded weather, it beat the operational forecast on accuracy for horizons from one hour to seven days, and produced a global forecast in about 1.4 seconds on a single graphics processor.

What followed

The European Centre began serving Pangu-Weather forecasts on its own website from late July 2023, which is a notable thing for an institution to do with a model that had just beaten its own.

GraphCast followed in November and beat Pangu in turn, and is more often named as the moment learned forecasting overtook the physics. On the published dates, it was the second to do it.

Why it mattered

Machine-learned forecasting had been close for a year and always slightly behind. This is the paper that crossed, four months before GraphCast, and the ordering matters: GraphCast is usually named as the moment the physics was beaten, and it was the second model to do it.

Sources

• Pangu-Weather: A 3D High-Resolution Model for Fast and Accurate Global Weather Forecast. arxiv.org. Primary source
• Nature publishes paper about Pangu Weather AI Model. huawei.com. Official

Cite this page

AI Achievements. (2023). Pangu-Weather is the first learned model to beat operational forecasting. Retrieved 2026-08-29, from https://achievements.ai/milestone/pangu-weather-3d-neural-networks

@misc{achievements_pangu_weather_3d_neural_networks,
  title  = {Pangu-Weather is the first learned model to beat operational forecasting},
  author = {{AI Achievements}},
  year   = {2023},
  url    = {https://achievements.ai/milestone/pangu-weather-3d-neural-networks}
}