Report: 100,000 Computing Cards Trained China's New AI Weather Model

China's National Meteorological Centre has put an AI ensemble forecasting system into operational use, according to reports published on October 17, moving it beyond research into national weather operations. ^[1]

The system, Fengqing Ensemble V1.0, recently passed an operational review, according to the Hong Kong newspaper Wen Wei Po. Developed under the centre's lead with universities and research institutes, it produces global forecasts for the short and medium term. ^[1]

An ensemble generates several possible forecasts rather than a single prediction. That gives forecasters a way to judge the range of weather that could develop and how much confidence to place in a particular outcome. Bringing such a system into national operations is a practical step beyond demonstrating its performance in a research test.

Wen Wei Po puts Fengqing's effective global forecast range, the period over which its predictions retain useful information, at up to 11 days. It also says the system outperforms the European Centre for Medium-Range Weather Forecasts' latest AI ensemble system on several measures. ^[1]

The 11-day figure is not a guarantee of accurate weather predictions at a particular address. The operational review and underlying performance results have not been published with these accounts, which leave unspecified what the 11-day measure covers, its accuracy threshold and test period, whether testing was independent, which European system was compared, and how Fengqing enters routine forecasting. ^[1][2]

What 11 days means

A forecast's useful range depends on what it predicts and how its success is measured. Tracking the broad atmospheric pattern across the globe is a different task from predicting rain at one location. A model can retain useful information about the first while uncertainty grows sharply around the second.

The same applies to probability forecasts. An ensemble can be useful even when its members disagree, because that disagreement helps describe the uncertainty. If several possible outcomes remain plausible, forecasters need to know that before issuing confident guidance.

The comparison with the European centre concerns several core forecasting indicators, the accuracy of probability forecasts and control of spread error, according to Wen Wei Po. ^[1] Spread describes how far apart an ensemble's predictions are. A useful ensemble should neither bunch its forecasts too tightly around an uncertain outcome nor scatter them so widely that they offer little guidance.

That is why judging an ensemble takes more than checking whether its central prediction was right. Its account of uncertainty must also be credible. A system that repeatedly expresses high confidence before getting the weather wrong can be less useful than one that signals the risk clearly.

AI joins the forecasting machinery

Fengqing combines a Vision Transformer architecture, a type of AI model, with algorithms intended to impose meteorological physics constraints. The stated aim is to represent atmospheric uncertainty and improve probability forecasts for severe weather. ^[1][2]

AI weather models learn patterns from past atmospheric data. Traditional numerical weather prediction instead uses equations describing atmospheric physics to calculate how conditions change. Both approaches seek to turn a picture of the atmosphere now into useful information about what comes next.

They can also work together. Physics-based models and AI systems are not mutually exclusive, and the reported launch should not be read as the retirement of conventional forecasting. The technical design described for Fengqing itself combines a learned model with physical constraints. ^[1][2]

The computing effort behind it was substantial. Fengqing was trained and repeatedly refined on a cluster described as containing 100,000 computing cards, hardware fitted with specialised processors that do the heavy arithmetic for AI training, at a core node of the National Supercomputing Center in Tianjin, according to the reports. ^[1][2]

That number describes the infrastructure used in development. It is not a measure of forecast accuracy.

From a model to a weather service

The operational review is the central development. A research benchmark asks how a model performs under a particular test. Operational use puts its output into the setting where a national forecasting agency works, with the demands that entails for producing usable guidance.

Passing that review does not by itself demonstrate better public warnings. The value will lie in the forecasts and decisions the system helps produce, especially when the atmosphere allows several dangerous outcomes rather than one obvious answer.

The reports identify disaster prevention, transport, energy scheduling, ecological protection and public welfare as intended areas of support. ^[1][2] Those uses make uncertainty part of the product: decisions depend on both the expected weather and the chance that conditions will be worse.

For Fengqing, the next development step is further work on accuracy and forecast lead time, according to the reported plans of its development team. ^[1][2]