Caught by the whirlwinds: Russian neural network will make ultra-accurate forecasts of storms in the Arctic

Scientists have created a neural network that allows for more accurate prediction of Arctic storms. It identifies errors in global weather models that tend to smooth out small—scale vortices and temperature anomalies, which are key factors in sudden bad weather. The implementation of the development can improve the safety of navigation on the Northern Sea Route, as well as air traffic and resource extraction in the region. The experts interviewed by Izvestia confirmed the relevance of the solution and noted that it could be improved if data from the Arctic-M and Meteor-M satellites, as well as specialized models of the ocean and ice conditions, were included in the training sample.

How a neural network detects extreme weather events

Russian experts have created a neural network that more accurately detects sudden and severe storms, including polar cyclones and the Novaya Zemlya bor — one of the most dangerous weather events typical of the Arctic region.

As the developers explained, the program is based on an algorithm that "penalizes" global weather models for smoothing out small eddies and temperature anomalies — the key causes of sudden storms. This averaging of data is a common problem with artificial intelligence—based models, which causes extreme but important weather events to be lost. In machine learning, a "penalty" is a mathematical mechanism that causes a model to have less confidence in certain decisions or actions.

— In the Arctic, the safety of navigation, aviation, resource extraction and other activities depends on the quality of forecasts. At the same time, there are few weather stations in the region, and global weather models often have too low a resolution or pronounced systematic errors," Mikhail Krinitsky, senior researcher at the P.P. Shirshov Institute of Oceanology of the Russian Academy of Sciences and head of the Machine Learning laboratory in Earth Sciences at MIPT, told Izvestia.

There are also neural network models, but they are also trained on arrays of data with low spatial resolution and are predisposed to averaging data, therefore they smooth out small vortices and temperature anomalies, he added. In addition, systematic temperature errors lead to incorrect forecasts of the state of the ice cover.

It was accepted as weather data: the "pocket" AI model predicts weather anomalies for the month ahead

How artificial intelligence can help prepare for natural disasters

To solve this problem, a special BERTUNet program was developed. Its feature is that it corrects large-scale forecast errors, but at the same time intentionally preserves small vortex structures without smoothing them out so as not to lose critically important local weather anomalies.

The work was attended by employees of the P.P. Shirshov Institute of Oceanology of the Russian Academy of Sciences, the Moscow Institute of Physics and Technology, the Skolkovo Institute of Science and Technology and the AIRI Institute.

How the program reduced the errors of global models

According to the researchers, the new neural network was trained on several types of data. On the one hand, the ERA5 archive was used — the world's largest database of meteorological data with a resolution of 0.25 degrees (the size of the geographical grid cell, at the equator it corresponds to approximately 28 km × 28 km). On the other hand, a more detailed Weather Research and Forecasting model with a step of 6 km was used. Additionally, measurements of satellites and ground-based weather stations on land and at sea were included in the training sample. In general, the study covered observations in the waters of the Kara and Barents Seas over 4.5 years.

— The results showed that the initial temperature error at the surface in some cases was almost 5 degrees. The new neural network lowered it to 2.1 degrees. The inaccuracy in the indication of wind speed decreased by about 20%. At the same time, spectral analysis showed that the energy of small atmospheric vortices in the corrected forecast remained at the level of the initial high resolution, whereas conventional correction methods completely suppressed these structures," commented Viktor Golikov, PhD student at the Institute of Physics and Technology of the Russian Academy of Sciences, a research engineer at the Skoltech Center for Artificial Intelligence.

According to scientists, the development is primarily focused on ensuring safety on the Northern Sea Route. In particular, it allows us to obtain more accurate forecasts of wind at the sea surface, which is necessary for optimal navigation of vessels in difficult ice and wave conditions in the Arctic.

As Gismeteo commented to Izvestia, the Arctic is a white spot on the weather map, since geostationary satellites do not see above the 70th parallel, and polar orbiters provide only fragments.

— The surface is mottled — ice, water, islands, mountains. Due to the melting of ice and evaporation, temperatures can be 6-8 degrees higher than normal, which is offset by cold collapses in Europe and Russia. Most models overestimate the ice content and underestimate the liquid water in the atmosphere. And the characteristic phenomena — polar mesocyclones, boron of Novaya Zemlya — require high resolution from the model," said Leonid Starkov, the portal's leading meteorologist.

He clarified that meteorologists are currently compensating for the lack of observational data by expanding the observation network and combining sources. Such hybrid techniques use the maximum of the entire arsenal — the ERA5 archive, satellite instruments, research vessels and weather stations on water and land. The authors of the new development are just teaching the neural network to coordinate the forecast with several sources at once.

What are you grinning at: The "UZI" of the Arctic ice ridges will make navigation safer

How "digital twins" of frozen ridges will help to understand their internal structure

According to the expert, BERTUNet can be improved by expanding the set of factors affecting the weather. In particular, it is important to take into account the parameters of sea ice — cohesion, thickness, age and other characteristics. It is also important to connect data from the Russian Arctic-M and Meteor-M polar satellites, and implement approaches combining data from different models, including ocean and ice systems, in order to improve forecast accuracy.

— The use of machine learning in weather forecasting tasks is actively developing. These models can work faster than classical approaches, and also help to find complex dependencies in data. The proposed solution is focused on application in the Arctic region. However, such tools may be relevant for other areas, since high—quality forecasting of small-scale phenomena is useful in various fields of human activity," added Peter Vytovtov, head of the machine learning and forecasting quality group at Yandex Weather.

I would like to see the development of the development in two directions. Firstly, the analysis and improvement of the model for cases of strong winds, and secondly, the addition to the model of the possibility of adjusting precipitation forecasts as one of the most important indicators, he added.