Pangu-Weather

What Pangu-Weather is

Pangu-Weather is a machine-learning model that produces medium-range global weather forecasts. Its paper, "Accurate medium-range global weather forecasting with 3D neural networks", appeared in Nature on 5 July 2023, in volume 619, pages 533 to 538.^[1] Kaifeng Bi is the first author, and Lingxi Xie, Heng Zhang and Xin Chen are among the co-authors.^[1]

Medium-range forecasting means predicting the state of the global atmosphere several days ahead. It is the range in which national weather services issue their main global products, and it is the range Pangu-Weather was built and tested for.

How it works

Weather data is not a flat image. At every point on the globe the atmosphere carries several variables at once — temperature, wind, pressure and others — and each of those variables is defined at many different heights. Pangu-Weather applies three-dimensional deep networks equipped with Earth-specific priors to this data, so the model treats the atmosphere as a volume rather than as a stack of unrelated pictures.^[1] The Earth-specific priors are assumptions about the geometry of the Earth and its atmosphere built into the network, instead of priors borrowed from ordinary image processing.^[1]

A medium-range forecast has to be rolled forward many steps, and small errors made at one step compound at the next. Pangu-Weather uses a hierarchical temporal aggregation strategy to reduce this error accumulation in medium-range forecasts: rather than chaining only single short steps, it combines forecasts made over different time intervals.^[1] The model was trained on 39 years of global data.^[1]

Results

The authors evaluated the model on reanalysis data — a globally complete record of the atmosphere produced by combining a weather model with historical observations — and compared it against the operational integrated forecasting system of ECMWF, the European Centre for Medium-Range Weather Forecasts. On that data it produced stronger deterministic forecasts than the operational system in all tested variables.^[1] A deterministic forecast is a single best estimate of the future state; an ensemble forecast, by contrast, runs a set of forecasts from slightly perturbed starting conditions so that the spread of the members expresses uncertainty. The paper reports that the model handled both ensemble forecasting and extreme weather events.^[1]

Forecasting a tropical cyclone is a particularly demanding test, because a small error in the predicted track grows quickly. Initialised with reanalysis data, Pangu-Weather tracked tropical cyclones more accurately than ECMWF-HRES.^[1]

Where it fits in the development of forecasting

The authors frame their work against two existing options. Numerical weather prediction, which advances the physical equations of the atmosphere on large computers, they describe as the most accurate method available today, but computationally expensive.^[1] Earlier AI weather methods, by contrast, were much faster but significantly less accurate than numerical models.^[1] Pangu-Weather's contribution sits in the second gap: it is a learned model whose forecasts, on the variables tested, were stronger than those of an operational global system, while the paper also covers extreme events and ensembles — the parts of forecasting that warnings and day-to-day decisions actually depend on.^[1]