Thirty seconds, and we'll have the prediction of global weather for the next 10 days.

This achievement comes from "FengWu," a global weather forecast model based on artificial intelligence (AI) released in April this year.

Facing escalating extreme weather events, the demand for more accurate and efficient forecasting systems has been increasing, and AI has been one of the solutions that scientists and industry experts have figured out to count on.

Medium-range weather forecasting

The name of the AI weather forecasting system "FengWu" was derived from the ancient Chinese anemometer used from the Han Dynasty which has been considered as the earliest prototype for measuring wind speed and orientation, according to Ouyang Wanli, leading scientist at Shanghai AI Laboratory, one of its developers.

Just as the expectations in its name, the system, based on a deep learning architecture, is capable of generating high-precision global meteorological forecast results for the next 10 days in 30 seconds.

Global medium-range weather forecast is one of the most important and difficult tasks in meteorological and climate forecasting.

"AI can use past meteorological factors, such as temperature, to predict future weather and get better results," Ouyang told the Global Times.

Currently, HRES, the most advanced physical model on such forecast, has an effective forecast time of 8.5 days, while FengWu extends the forecast range to 10.75 days based on reanalysis data with the error reduced by 19.4 percent.

The lab is now exploring the use of the model to forecast extreme weather such as typhoons with preliminary results achieved.

Bai Lei, a scientist at the lab, told Global Times that the model made an accurate track forecast of super typhoon Mawar which took place in May.

Bai said more data are constantly used to improve the effectiveness of the algorithm and targeted modifies are made to further improve its ability for extreme weather forecasting.

"The resolution can be improved. For example, we can forecast the weather in a district, and in the future we want to narrow it down to a street, and we are working to making it more accurate and precise," said Bai.

With its excellent performance and accuracy, the AI model is expected to complement the traditional physical models in digitalization of weather forecast and support industries such as agriculture and forestry.

Taking the power industry as an example, the FengWu model can assist in accurate prediction of wind speed and sunlight intensity, which is crucial for wind power and solar power generation, said Ouyang.

More recently, China's tech giant Huawei launched its latest AI model Pangu-Weather in early July, which "challenges the previously held assumptions that the accuracy of AI weather forecast is inferior to traditional numerical forecasts," said the company.

The model, trained on 39  years of global data, can forecast global weather for a 7-day lead time with higher accuracy than the Integrated Forecast System, one of the best global numerical weather prediction system, while being more than 10,000 times faster at the same spatial resolution, according to a report published on the academic journal Nature.

The model also shows a better performance in extreme weather forecasts and ensemble forecasts, said the report.

Nowcasting of extreme precipitation

Nowcasting refers to very short-range weather forecast over a period up to 6 hours ahead, which provides detailed information about the weather in real time.

Nowcasting with "high resolution, long lead times and local details" is greatly needed to forecast extreme precipitation, a considerable contributor to meteorological disasters, and mitigate its socioeconomic effects, said a Nature report on July 5.

The report introduced NowcastNet, a nowcasting model integrating physical rules and deep learning for extreme precipitation.

The model, jointly developed by researchers from UC Berkeley, Tsinghua University and the Chinese Meteorological Administration, can make high-resolution prediction up to 3 hours in advance over regions of 2,048 km × 2,048 km based on radar observations from the U.S. and China, said the report.

In the study, 62 meteorologists evaluated the model and ranked it first in 71 percent of cases against some other leading nowcasting models.

"NowcastNet provides skillful forecasts at light-to-heavy rain rates, particularly for extreme-precipitation events accompanied by advective or convective processes that were previously considered intractable," the report said.