Huawei develops a weather forecast model 10,000 times faster than other models

(Nature) magazine, one of the most important scientific references in the world, published a research paper authored by researchers from Huawei on the innovative artificial intelligence model for weather forecasts (Pangu Weather) developed by the Huawei Cloud business unit specialized in cloud computing. The paper dealt with how Huawei developed a global and accurate AI system to predict the weather around the world based on deep learning capabilities and data analysis collected over a period of 43 years.

The new model, which the magazine considered revolutionary, has higher accuracy than traditional numerical forecast methods, and improves forecast speed by up to 10,000-fold, reducing global weather forecasting time to just a few seconds.

The paper, entitled “Accurate Medium-Range Global Weather Forecasts Using Three-Dimensional Neural Networks,” presents a wide range of independent experiments that prove the efficiency of the new model. Huawei’s “Bango Weather” model challenges previous assumptions that AI weather forecasts are less accurate than Numerical conventional predictions.

This is the first time that researchers at a Chinese technology company have exclusively authored a research paper in the journal Nature, according to the journal archives.

The rapid development of computer capabilities over the past 30 years has led to a significant improvement in the accuracy of the traditional numerical weather forecasting system, which has contributed to avoiding many major disaster risks and providing predictions about climate changes. However, the prediction process did not reach the level of speed required.

Therefore, the researchers sought to find out how to use deep learning methods to reach the accuracy required to predict the weather, but systems that rely on artificial intelligence for medium and long-term forecasts remained less than the accuracy of numerical forecasts.

Artificial intelligence was unable to predict severe and unusual weather conditions such as hurricanes, as artificial intelligence models for weather forecasts sometimes lacked accuracy and speed, despite some positive results. This is due to two reasons, the first: the adoption of artificial intelligence models for weather forecasting on networks 2D neurons, so they can't process disparate 3D aerial data. Second, the medium-term weather forecast had cumulative errors when the model was used multiple times.

The number of hurricanes that occur annually around the world is estimated at about 80, most of which cause heavy material losses. At the level of China only, hurricanes in 2022 resulted in direct economic losses of 5.42 billion yuan, according to data from the Chinese Ministry of Emergencies. The earlier warnings are sent, the easier and more effective it will be to take precautionary measures and reduce losses.

In May 2023, Cyclone Mawar captured the attention of the entire world and was named the strongest tropical cyclone of the year so far. According to the China Meteorological Administration, the Pangu Weather model accurately predicted the path of Typhoon Mawar five days before it changed course in the eastern waters of the Taiwan islands.

The accuracy of the (Banjo Weather) model has emerged in scientific experiments as a revolutionary alternative to traditional numerical forecasting methods that require time ranging from one hour to 7 days. With a 10,000-fold increase in its forecasting speed, it can predict even the smallest weather details such as humidity, wind speed, temperature and sea level pressure within a few seconds.

The new model uses a terrestrial 3D architecture transducer to process irregular and complex 3D atmospheric data. It was trained using a temporal hierarchical aggregation strategy on different forecast periods across 1 hour, 3 hour, 6 hour, and 24 hour intervals, which resulted in a reduction in the frequency of forecasting the weather at a given time, as well as in reducing false forecasts.