- Статьи
- Science and technology
- It was accepted as weather data: the "pocket" AI model predicts weather anomalies for the month ahead
It was accepted as weather data: the "pocket" AI model predicts weather anomalies for the month ahead
Scientists have created an AI program for predicting climate processes based on the principle of neural network video generators. The neural network builds possible scenarios based on known weather data, which significantly speeds up and simplifies calculations. It can be run even on a regular laptop: it generates a detailed forecast for eight days and predicts climatic deviations for up to a month. For example, she could have predicted an abnormal snowfall in Moscow. Such data is especially important for industries sensitive to weather risks, from rescue services to agriculture and insurance. At the same time, experts emphasize that new approaches should be used together with traditional methods: neural networks provide high—speed calculations, while classical models provide the necessary accuracy and reliability of forecasts.
How a neural network calculates climate processes
The Marchuk program is based on algorithms similar to those used in video generation, the developers explained. Based on known weather data, it predicts possible atmospheric conditions, just as neural networks "complete" the following frames in the generated videos. This approach makes it possible to significantly speed up and simplify the calculation of weather scenarios compared to traditional methods.
— The developed model has 276 million parameters (about 600 times less than large language models such as GPT-3 chat). It's possible to run it even on a laptop with a good graphics card. The program can make a detailed forecast of up to eight days in 7.5 minutes, and also predict anomalies that will occur in 15-30 days," Konstantin Sobolev, one of the developers and head of the Generative AI for Video group at the FusionBrain laboratory at the AIRI Institute, told Izvestia.
According to the expert, from the point of view of mathematics, weather data are 4D tensors - sets of numbers, each of which corresponds to certain coordinates, time, and physical parameters (temperature, wind, humidity, etc.). A similar structure works in video generation programs, but instead of physical quantities, color channels appear in them.
The Marchuk program is named after mathematician Gury Marchuk, the founder of numerical modeling of the atmosphere and climatic processes. His developments made it possible to use computers for weather forecasting and laid the foundation for many hydrodynamic models. From 1986 to 1991, the scientist headed the USSR Academy of Sciences.
— To make a high-quality forecast, Marchuk just needs to get weather data for the last day. More precisely, four "snapshots" of the state of the atmosphere with an interval of six hours. Based on them, the neural network can predict the weather for a period of one to eight days in advance," Konstantin Sobolev explained.
Such data, for example, is necessary for farmers to understand when to sow and harvest crops. They are also important for power engineers, who need to determine the operating modes of thermal power plants in advance. In general, the development can be in demand by anyone whose operational activities depend on weather conditions: rescue services, utilities, insurance companies and banks.
— The model can simulate the probability distribution of weather scenarios. This is useful if we want to assess the likelihood of a natural disaster in a few weeks or months," Kirill Sobolev said.
To achieve this effect, the classical methods of weather forecasting resort to tricks. As a result, they don't work very well, he added.
How the program can predict weather disasters
— Marchuk, in particular, could predict the abnormal snowfall that took place in Moscow on April 27. The publicly available version currently provides a forecast for five days. This is due to the limited amount of archived data. But when integrated with operational weather forecast models, this time is shortened, Marchuk begins to "feel" weather events in 15 days," said Ivan Oseledets, Director General of the AIRI Institute.
First, global data is loaded into the model for one day, after which the program generates a forecast for up to 30 days in advance. Several such scenarios are calculated, which allows us to obtain a set of possible scenarios for the development of the weather situation. Further, based on these statistics, it is possible to estimate the probability of certain events, the expert explained.
In addition to the neural network, the developers have created an AI meteorologist, a system for automatically generating text forecasts. It translates the numerical data of the model into coherent explanations accessible to the average user, said Ivan Osedelets. The system has been tested in locations with different climates, such as Cork, Ireland, Manila, Philippines, Chennai, India, and Da Nang, Vietnam.
— Weather forecast is one of the most demanded digital services. Many people use it every day without thinking about the complex infrastructure behind the usual numbers and visual icons in applications," Peter Vytovtov, head of the machine learning and forecast quality group at Yandex Weather, told Izvestia.
In general, the approach proposed by the developers fits into broader trends that can be observed at the world's leading conferences, where more and more attention is being paid to combining machine learning with physically based models.
— AI-based models can give an impetus to the further development of domestic predictive schemes. The Marchuk model created by the scientists of the AIRI Institute has been transformed into a functioning technological forecasting line, and at the current stage it can be concluded that a promising scientific and technological direction has been created in Russia,— commented Roman Vilfand, Honored Meteorologist of the Russian Federation.
At the same time, according to the procedures and standards of the Russian Hydrometeorological Service, in order to make a decision on the introduction of a new model, it is necessary to conduct independent tests and compare its forecasts with the results of existing hydrodynamic models used in the department, he noted.
— Modern approaches in the field of weather forecasting often formulate a problem in a probabilistic formulation. In this sense, the presented model is very relevant," says Mikhail Krinitsky, head of the Laboratory of Machine Learning in Earth Sciences at the Moscow Institute of Physics and Technology and a senior researcher at the P. P. Shirshov Institute of Oceanology.
However, like other tools, neural networks can be trusted only after comprehensive verification, he said. In particular, questions arise as to whether standard metrics are sufficient for analyzing atmospheric phenomena with their rich multiscale dynamics. It is also interesting how the model will perform in difficult regions. For example, in the Arctic.
According to the expert, neural network algorithms can be enhanced by expanding the set of factors taken into account, increasing spatial resolution and integrating physical parameters into machine learning processes. At the same time, AI models can be used more effectively in conjunction with classical approaches: neural networks will speed up calculations, and traditional methods will ensure the necessary accuracy.
Переведено сервисом «Яндекс Переводчик»
