Pangu-Weather is published in Nature

Background

Forecasts for the coming days come from numerical weather prediction. The atmosphere is represented as a grid of variables — temperature, pressure, wind, humidity — and a supercomputer advances the physical equations step by step. The method rests on physics and has improved for decades, but every forecast costs a large amount of computing, and running a global medium-range forecast repeatedly is expensive.

Machine learning takes a different route. Rather than solving the equations, a model is trained on archives of past weather and learns to map the current state of the atmosphere to a future one. Such archives are usually reanalysis: a record in which past observations have been passed through a fixed forecasting system, so that the whole history is internally consistent and can serve as a common yardstick. Alongside medium-range forecasting, there is nowcasting, which covers the very short range, before a full numerical forecast cycle could be completed.

What happened

On 5 July 2023, Nature published Pangu-Weather together with its peer review ^[1]. The paper's title refers to 3D neural networks, and it presented a deep-learning model for global medium-range weather forecasting.

The same issue of Nature carried a second weather paper, NowcastNet, a deep-learning model for nowcasting extreme precipitation, in volume 619, pages 526 to 532 ^[2]. The two papers appeared on the same day and framed AI as an alternative to computationally expensive numerical models ^[1][2].

Key details

The claim made for Pangu-Weather concerned skill, not only speed. The paper reported better deterministic forecasts than ECMWF's operational system across every variable it tested ^[1]. Deterministic here means a single forecast for a given starting state, as opposed to an ensemble of many runs from slightly perturbed initial conditions.

The comparison target was the Integrated Forecasting System of the European Centre for Medium-Range Weather Forecasts ^[1]. When the model was initialised with reanalysis data, its tracking of tropical cyclones was also better than ECMWF-HRES ^[1].

The evaluation used reanalysis data as the common reference for comparing AI and numerical forecasts ^[1]. That choice matters for reading the result: the learned model and the physics-based system were scored against the same archived record, so within that reference the comparison is like for like.

Why it matters

Two things made the publication notable. First, the result was peer reviewed and appeared in a general-science journal, and the same issue carried a second AI weather paper ^[1][2]. Together the two covered different timescales — global medium-range forecasting in one case, nowcasting of extreme precipitation in the other — and both presented machine learning as an alternative to expensive numerical models ^[1][2]. Second, the benchmark was an operational forecasting system rather than another research model, and the reported advantage covered every variable that was tested ^[1].

What the papers did not settle is how such models fit into an operational chain. A learned forecast is only as good as the record on which it was trained and scored, and the use of reanalysis as the common reference is part of that question ^[1]. Framing AI as an alternative to numerical models also leaves open how the two are combined: numerical models carry physical constraints, learned models offer speed, and the balance between them was not fixed by these papers.