Disentangling complex weather with machine learning analysis

Weather in the mid-latitude regions, between the tropics and the poles, is notoriously difficult to predict. While tracking short-term movements, such as cyclones and anticyclones, can provide reliable forecasts up to a few weeks in advance, longer predictions are much more difficult and unreliable. Central to the problem is the atmospheric circulation system, which is nonlinear and high-dimensional, and therefore very difficult to model.

Mukhin et al. developed a method to disentangle these long-term dynamics using a machine learning approach. Their technique identified important weather patterns, which are stationary, recurrent states of the atmospheric system.

“Based on several non-linear machine learning methods for data analysis, we could both reliably identify weather patterns and obtain a set of variables to represent their dynamics,” said author Dmitry Mukhin. “In addition, our approach includes the study of the dynamic properties of these patterns, such as predictability and persistence, using a recently developed recurrence analysis approach.”

Using their method, the researchers identified key weather patterns affecting long-term winter trends in the Northern Hemisphere. Future research could predict the severity of winters in North America and Europe based on these patterns and their dynamic properties.

The authors plan to continue to develop their model while incorporating elements from tropical regions. They encourage others to view their method as a template for long-term forecasting of complex weather events.

“We believe that the approaches to the study of nonlinear systems developed in our collaboration can be useful for a deeper understanding of atmospheric variability on long-term scales,” said Mukhin.

Source: “Revealing recurrent regimes of mid-latitude atmospheric variability using novel machine learning method,” by Dmitry Mukhin, Abdel Hannachi, Tobias Braun, and Norbert Marwan, Chaos (2022). The article can be accessed at https://doi.org/10.1063/5.0109889 .