What will the weather be like tomorrow? AI predicts

The Earth we live on is experiencing one of the hottest summers in human history.

On July 3, 2023, the global average temperature recorded at 2 meters above the earth's surface exceeded 17 degrees Celsius for the first time, the highest temperature ever recorded . In addition to human activities, the strong "El Niño" is also one of the culprits for the intensification of global high temperatures this summer.

Figure | Global average temperature 2 meters above the earth's surface (Source: University of Maine)

High temperatures will not only cause business and production to stop in some areas, causing economic losses to the local area, but will also endanger our health and even cause death, such as heat stroke caused by prolonged exposure to high temperature environments.

In addition to abnormally high temperatures, the frequent extreme weather events in recent years, such as tsunamis, typhoons (hurricanes), floods, hail, etc., will also have immeasurable negative impacts on human economic life.

The World Bank estimates that globally, improved weather forecasting and early warning systems could not only save tens of thousands of lives, but also bring economic benefits of US$162 billion each year; in addition, over the past 50 years, more than 34% of recorded disasters, 22% of related deaths (1.01 million people) and 57% of related economic losses (US$2.84 trillion) were also related to extreme precipitation events.

Photo: The island after being devastated by a catastrophic hurricane.

Therefore, how to timely and accurately predict short-term and future weather conditions has become one of the important topics that scientists are trying to solve.

Artificial intelligence (AI) is expected to be a faster and cheaper alternative to improve extreme weather forecasts .

So what is the potential of AI to assist with weather forecasting? And how well are forecasts currently done?

Today, two research papers on "AI Weather Forecasting" published in Nature mentioned two AI-based weather forecasting methods , one of which can predict global weather patterns a week in advance, and the other can predict short-term weather, such as extreme precipitation events.

According to the paper, these two AI-based methods are as accurate as existing methods and can even predict weather phenomena that were previously difficult to predict.

However, there is still some uncertainty and controversy about whether and when these AI models can become the main tools for meteorologists to make weather forecasts. The main considerations are the operating costs of these AI models after commercialization and whether they can win people's trust .

What will the weather be like tomorrow? AI predicts

Weather forecasts play an important role in helping save lives and reduce property losses, especially as extreme weather events become more frequent.

Currently, the most accurate forecasting system is numerical weather forecasting, which mainly relies on physical equations, but has high requirements on computing power and needs support from methods such as high-performance computing. Moreover, it is usually slow, with a single simulation taking several hours, even on a supercomputer with hundreds of nodes.

In recent years, some AI-based methods have shown the potential to significantly improve the speed of weather forecasting, producing a 24-hour forecast in a few seconds, but the accuracy is generally not as good as numerical weather forecasting.

In a paper titled "Accurate medium-range global weather forecasting with 3D neural networks", a research team led by Tian Qi, chief scientist of Huawei Cloud AI, proposed an AI-based weather forecasting system - Pangu-Weather. Its main technical contributions include designing the 3DEST architecture and applying a hierarchical time aggregation strategy for medium-term forecasting.

Figure｜Network training and reasoning strategies. (Source: Nature)

It is reported that this AI model uses 39 years of global reanalysis weather data as training data, and its prediction accuracy is comparable to that of the world's best numerical weather forecast system IFS; more importantly, the former is more than 10,000 times faster than the latter at the same spatial resolution.

However, as mentioned in the paper, Pangu meteorology also has certain limitations.

For example, the model was trained and tested on reanalysis data, while actual weather forecasting systems use observational data; weather variables such as precipitation were not studied, and ignoring these factors may cause the model to lack some capabilities (such as using precipitation data to accurately predict small-scale extreme weather events such as tornadoes); the method may produce smoother forecast results, increasing the risk of underestimating the scope of extreme weather events; and there are problems such as time inconsistency.

Despite this, Pangu Meteorology still demonstrates the potential of large pre-trained models in weather forecasting. In the future, the research team hopes to further iterate the model by integrating more vertical levels and atmospheric variables, integrating the time dimension and training 4D deep networks, and using deeper and wider networks.

Extreme precipitation is an important factor leading to meteorological disasters, and therefore accurate forecasts with high resolution, adequate lead time, and local details are highly needed to mitigate their socioeconomic impacts.

In meteorological forecasting, there is a very short-term weather forecast, namely the **“nowcast”** from the current time to the next 6 hours, which is mainly used to provide detailed information on the immediate weather and is very important for risk prevention and crisis management of extreme precipitation events.

However, current approaches suffer from blurring, dissipation, and intensity or position errors, physics-based numerical methods have difficulty capturing key chaotic dynamics (such as convection initiation), and data-driven learning methods fail to obey intrinsic physical laws (such as convection conservation).

In another paper, a joint research team led by Michael Jordan, a leading figure in the field of machine learning and professor at the University of California, Berkeley, and Wang Jianmin, a professor at Tsinghua University, proposed a model called NowcastNet that combines physical laws and deep learning for real-time precipitation forecasting .

Figure｜NowcastNet's design architecture. (Source: Nature)

It is reported that NowcastNet performs well in nowcasting. Based on radar observation data, it can make high-resolution precipitation forecasts for an area of ​​2048 km × 2048 km three hours in advance . In an assessment of extreme precipitation prediction capabilities, the model surpassed other methods in about 70% of predictions. It performed particularly well in extreme precipitation events that were previously difficult to predict.

Should we trust AI?

In a News & Views article published alongside the paper, Colorado State University research professor Imme Ebert-Uphoff and research assistant Kyle Hilburn argue that these approaches are so promising that they could even spark a paradigm shift in which generative AI-based models could completely replace numerical weather forecasting .

However, they also caution that these AI models also present some potential risks. For example, when operating under conditions that have never been seen before, the behavior of AI systems is often unpredictable.

Therefore, avoiding these risks requires the involvement of meteorologists who learn to design, evaluate and interpret such systems .

In this regard, Professor Russ Schumacher, also at Colorado State University, said, " People's concerns about AI are usually that it is a black box . You just input some numbers and get some numbers out, but you don't know how the process works in between."

In addition, although meteorologists have not yet fully embraced this "AI magic," they are already learning how to feed the atmospheric observation data used to run traditional models into AI models in real time, as well as gaining more experience using these AI models.

In the future, it may still take some time to build trust in the “black box” model, but the practice of using AI models in weather forecasting will continue to move forward in this AI wave.

What do you think about AI weather forecasting?

Reference Links:

https://www.nature.com/articles/s41586-023-06185-3

https://www.nature.com/articles/s41586-023-06184-4

https://www.nature.com/articles/d41586-023-02084-9

https://www.newscientist.com/article/2381069-earth-has-just-experienced-the-hottest-day-we-have-ever-seen/

https://www.washingtonpost.com/weather/2023/07/04/ai-weather-forecasts-hurricanes-tornadoes/