Institute of Oceanology and MIPT: Neural network for forecasting extreme winds in the Arctic

Main article: Neural networks (neural networks)

2025: Creating a neural network to predict extreme winds in the Arctic

Scientists from the Moscow Institute of Physics and Technology (MIPT) and the Institute of Oceanology named after P.P. Shirshov RAS created a neural network to accelerate and improve the modeling of extreme weather events in the Arctic. It details data from global meteorological services, showing dangerous vortices and storms with the accuracy of complex physical models, but 50 times faster. MIPT announced this on December 3, 2025.

Existing global meteorological models, such as the European Center for Medium-Term Forecasts (ECMWF) global forecast or the American GFS, have low spatial resolution and do not "see well" relatively small but extremely dangerous atmospheric vortices, which on the ground appear as sudden hurricane winds and strong waves in the ocean. The traditional way to get a better picture in certain regions today is to launch a high-detail hydrodynamic model (for example, WRF). It usually takes colossal calculations and time.

𝓲

Фото: MIPT

The Russian researchers trained artificial intelligence on data from a high-precision physical model of WRF. As a result neuronet , I learned from the "blurred," insufficiently detailed data of the global GFS forecast to generate a "clear" and detailed picture of winds in the Arctic almost instantly.

The main quality of our development is speed, provided that there are minimal quality losses. The neural network produces a highly detailed forecast for the Barents and Kara Seas more than 50 times faster than resource-intensive physical models. In addition, it shows storms that others do not reproduce: the number of vortex structures in the neural network data almost coincides with the WRF reference model (a difference of less than 3%), while the global ERA5 forecast underestimates the number of such phenomena by about half, "said Mikhail Krinitsky, one of the authors of the study, head of the machine learning laboratory in Earth Sciences MIPT.

The effectiveness of the method was also clearly demonstrated on the example of an extreme weather event - the New Earth bora, observed in February 2022. This is a strong cold wind, rapidly collapsing from the mountains of Novaya Zemlya and posing a serious danger to shipping. While global reanalysis data could ERA5 show this phenomenon implicitly, weakly and only in the southern part of the archipelago, the neural network model, like the WRF reference model, much more clearly reproduced the structure, direction and strength of the boron throughout the western coast of Novaya Zemlya.

𝓲

Фото: MIPT

As the tests showed, the results of the wind speed modeling were significantly closer to the data of real weather stations than the data of ERA5 or GFS. In addition, when wind data calculated by the neural network were used to model wind waves in the ocean, the calculated wave heights matched buoy measurement data with an accuracy comparable to that obtained from the WRF reference model.

The development opens the way to more reliable and significantly cheaper computational and prompt forecasting of extreme weather events to ensure the safety of navigation, the operation of ports and oil and gas platforms in the Arctic.