Hailstorm Center collects data, large models support extreme weather forecasts, "Storm Chasers" are on stage

Content at a glance: Extreme weather such as thunderstorms, hail, and tornadoes are always unpredictable and hard to avoid. However, Australian researchers risked their lives to go deep into hailstorms to collect data, just to make weather forecasts more accurate. With the help of supercomputers and AI, can humans catch up with the storm and make the chaotic weather system less unpredictable? Let's watch "Storm Chasers" starring storm chasers.

Keywords: large model, storm chaser, extreme weather

Author | Xuecai

Editor | Sanyang

This article was first published on HyperAI WeChat public platform~

In the well-known adventure disaster film "Twister" released in 1996, the protagonist brought detection equipment into the center of the tornado in order to conduct in-depth research on tornadoes and realize real-time data recording.

Inspired by this movie, Australian meteorologists Joshua Soderholm and Julian Brimelow began their own storm-chasing journey and successfully brought a small weather sensor ice probe (hailsonde) into a hail storm to collect meteorological data in the hailstorm, thus revolutionizing the research method of extreme weather.

Figure 1: Still from the movie "Tornado"

The ice probes are similar in shape to hailstones and weigh about 24 grams. They are attached to balloons and released into the hailstorm together. After entering the center of the hailstorm, the two separated and the ice probes experienced the trajectory of the hailstones in the hailstorm just like the hailstones, and recorded the growth conditions of the hailstones moving in the hailstorm. In addition, the ice probes also recorded significant ice growth and rotated half a circle with the supercell mesocyclone.

Figure 2: Trajectory of two ice probes in a hailstorm

"It started as an extracurricular project to see if we could use existing technology to create a device like the one in the movie. We had to solve a lot of engineering problems during the production process to ensure that the ice probe could survive the extreme conditions of a hailstorm," said Joshua Soderholm.

Figure 3: The structure of the ice probe, with 3D printed parts, batteries and other electronic devices encapsulated in a polystyrene shell

"Collecting data from the center of a hailstorm is like chasing the white whale of meteorology - dangerous but fascinating. Data collected from the center of a hailstorm will improve our ability to model hailstorms and provide direct evidence of how hail behaves in hailstorms. But it's not as easy as it sounds - you have to be in the right place at the right time, and you have to encounter the right hailstorm."

After a few days of bad luck, they hit a supercell and successfully put two ice probes into the hailstorm. A supercell is a single-cell strong thunderstorm system with a horizontal scale of more than ten kilometers and a life span of tens of minutes to hours. It is larger, more persistent, and has more severe weather than ordinary mature single-cell thunderstorms. After the ice probe was captured by the supercell, it separated from the balloon and then floated with the hailstorm like hail, and was finally carried to an area 7 kilometers away by winds of more than 120 kilometers per hour.

Figure 4: Joshua Soderholm is launching an ice probe

Unpredictable weather systems

Weather forecasting requires human participation. Even with the use of supercomputers, satellite data and radar data, it is still difficult to make accurate predictions about weather systems. In 1961, American meteorologist Edward Norton Lorenz tried to use a computer program to predict future weather.

After getting the result, he used the output value of the program in the middle step as the input value of the next step and ran the program again. However, because the input value only retained 3 decimal places, and the program was run with 6 floating point numbers, this one thousandth deviation made the output value of the program completely different from the result obtained last time.

Based on this, he proposed the concept of chaotic system. The meteorological system is a typical chaotic system. It is not completely random, but it can easily change dramatically due to changes in certain factors. In other words, the meteorological system is a very sensitive system.

The "butterfly effect" is an exaggerated but typical example. A butterfly flapping its wings twice in the tropical rainforest of South America may bring a tornado to the United States. The source of all this is that the butterfly disturbs the initial variables of the system.

Figure 5: Butterfly Effect

Therefore, it is difficult to achieve completely accurate weather forecasts. The existing meteorological forecasting method, namely numerical weather forecasting (NWP), first divides the forecast area into grids, and then uses supercomputers to solve partial differential equations through numerical simulation to obtain the results.

This method is very time-consuming. Even if a supercomputer with hundreds of nodes is used, it takes several hours to predict the weather for the next 10 days. At the same time, due to the limitation of grid resolution, some small-scale meteorological processes will be parameterized by approximate functions, which will bring errors to meteorological forecasts.

Because of this, it is difficult for NWP to be perfect for small-scale extreme weather and medium- and long-term weather forecasts. After the formation of Typhoon Dusurui this year, different institutions used supercomputers to predict the path based on different models, and the structures were very different. Even the predictions made by the same model are constantly revised as meteorological conditions change, and relatively accurate predictions can only be made before the typhoon lands.

The subsequent Typhoon No. 6 Kanu also had a unique movement. It suddenly turned across the Pacific Ocean, then began to wander, and finally hit Japan directly, leaving even supercomputers confused.

Figure 6: The Global Ensemble Forecast System (GEFS) forecasts the path of Typhoon Kanu No. 6 in 2023.

At the same time, since the weather forecasts made by various agencies vary greatly, weather forecasts still require the participation of forecasters. Forecasters will combine all weather forecast results and combine local climate characteristics, terrain conditions, personal experience, etc. to make the final weather forecast, but it still cannot be guaranteed to be completely correct. There is no way, the weather system is so unpredictable.

Figure 7: Path of Typhoon Wayne No. 16 in 1986

Extreme weather chaser

Small-scale supercells are a loophole in medium- and long-term weather forecasts. Supercells are characterized by rapid formation and difficulty in prediction, and are prone to extreme weather such as thunderstorms, hail, heavy rainfall or tornadoes .

On the evening of August 16, 2021, Haidian District, Beijing, encountered a supercell and heavy rain. The water under the Hanhe Road Railway Bridge rose to 1.75 meters in 30 minutes, resulting in the death of 2 people. On the afternoon of August 13, 2023, Dafeng District, Yancheng City, Jiangsu Province encountered a tornado, killing 2 people and injuring 15 people. The formation of this tornado was also related to the supercell.

Figure 8: Tornado in Dafeng District, Yancheng

However, spectacular meteorological landscapes such as thunderstorms, hail, and tornadoes can feast the eyes of explorers, and therefore attract many storm chasers like Soderholm. Whenever a typhoon approaches or a supercell is about to form nearby, storm chasers will make full preparations and rush towards the storm.

At the same time, as the first witnesses of extreme weather, storm chasers can also collect first-hand information on extreme weather, provide valuable materials for meteorological research, enrich the database of existing computing models and AI models, and make important contributions to the development of meteorology.

A large meteorological model comparable to NWP

As early as 2021, Alibaba Cloud revealed that the DAMO Academy and the National Meteorological Center jointly developed an AI algorithm for weather forecasting and successfully predicted multiple severe convective weather events. In September of the same year, Deepmind published an article in Nature, using a deep generative model to make real-time forecasts of rainfall.

At the beginning of this year, Deepmind officially launched GraphCast, which can predict the global weather for the next 10 days with a resolution of 0.25° within one minute. In April, Nanjing University of Information Science and Technology and Shanghai Artificial Intelligence Laboratory jointly developed the "Fengwu" weather forecast model, which has a further reduction in error compared to GraphCast.

Subsequently, Huawei launched the "Pangu" meteorological model. Due to the introduction of a three-dimensional neural network in the model, the prediction accuracy of "Pangu" exceeded the most accurate NWP prediction system for the first time. Recently, Tsinghua University and Fudan University have successively released the "NowCastNet" and "Fuxi" models. The former is very useful in short-term extreme weather forecasts, while the latter extends the forecast time to 15 days.

Figure 9: The path forecasts of the Pangu model and ECMWF for Typhoon Connie No. 25 (Figure a) and Typhoon Yutu No. 26 (Figure b) in 2018.

Red: Pangu model predictions

Blue: ECMWF forecast

Black: Actual situation

It can be seen that the weather forecasting model is approaching and even surpassing the traditional NWP analysis model in terms of prediction accuracy and prediction time. At the same time, compared with NWP, the weather forecasting of the AI ​​model requires lower equipment conditions and takes less time. Using only a Google TPU v4, GraphCast can predict future weather within minutes.

However, the existing AI big models can only predict future weather by learning from past weather data. Therefore, in extreme weather and sudden weather scenarios, big models also need the assistance of other algorithms, and even more so the participation of people . At this time, the weather data provided by the storm chasers active in the center of the storm is even more important for the optimization of the AI ​​big model. Humans and big models working together will surely make an excellent "Storm Chaser".

Reference Links:

[1] https://phys.org/news/2023-08-movie-inspired-technology-successfully-hail-eye.html

[2] http://m.nmc.cn/ty/

[3] http://henan.china.com.cn/tech/2021-06/22/content_41599891.htm

[4] https://arxiv.org/abs/2212.12794

[5] https://www.nature.com/articles/s41586-023-06185-3/figures/4

[6] Jiang Yanru, Analysis of Typical Weather Processes[M]. Beijing: Meteorological Press, 2016.

This article was first published on HyperAI WeChat public platform~