Machine learning reveals clues to improved weather forecasting in our atmosphere

New mathematical analysis of atmospheric conditions in the Atlantic and Pacific have revealed how to interpret and make sense of key aspects of the variability of the weather.

Experts at the University of Leicester (UK), SISSA (Italy) and Vrije Universiteit Amsterdam (the Netherlands) have used statistical mechanics and machine learning techniques to comb through atmospheric data such as pressure and temperature fields and identify patterns that indicate specific modes that reflect variations in weather.

Their conclusions have been published in the journal npj Climate and Atmospheric Science and could help scientists better understand how to improve weather prediction and how to predict the impact of climate change.

A particular challenge in making weather forecasts is that anticipating atmospheric processes becomes more difficult over longer timescales. Forecasting beyond a week becomes very challenging, in part because large scale atmospheric features associated with slow variations in the atmosphere become prominent. 

The most relevant features are known as blockings, which are persistent anomalies extending over thousands of kilometres that can last for up to a few weeks. Blockings are associated with persistent high-pressure systems, which disrupt the typical west-to-east flow of the mid-latitude jet, the approximately zonal wind blowing in the upper troposphere of the mid-latitudes.

Blocking events can span several days to a few weeks and are often associated with weather phenomena such as heat waves, cold spells, extensive droughts leading to wildfires, and floods. It is also often proposed that climate change will also manifest itself as modification in the nature and frequency of occurrence of such weather patterns.

To find a new approach to tackle this problem, the scientists applied algorithms developed for studying molecular dynamics, used in condensed matter physics, to atmospheric dynamics. They have been able to provide rigorous foundations to understanding atmospheric dynamics as an alternation of weather patterns. In particular, it provides a clearer picture of those patterns corresponding to blocked flows and zonal flow. Using this technique, they could identify not only when the atmosphere was experiencing a ‘blocking’ or a ‘flow’, but explain how the transition between such conditions take place.

Additionally, the method used in the study allows to better understand the interplay of planetary-scale patterns such as the Pacific-North American teleconnection pattern (PNA) and the North Atlantic Oscillation (NAO). 

Professor Valerio Lucarini from the University of Leicester School of Computing and Mathematical Sciences, said: “It is exciting that we have been able to identify much studied and extremely relevant features of the atmosphere as sort of ‘quasi-particles’, using algorithms originally developed for studying problems in molecular dynamics. This is an important step for advancing our understanding of the fundamental mathematical and physical properties of the climate system and our ability to anticipate weather and climate risk.” 

• Unsupervised detection of large-scale weather patterns in the northern hemisphere via Markov State Modelling: from blockings to teleconnections is published in npj Climate and Atmospheric Science, DOI: 10.1038/s41612-024-00659-5