Semester 2, Week 7: Dr Ioana Colfescu, University of St Andrews

The North Atlantic Oscillation — a large-scale seesaw in atmospheric pressure between the Azores and Iceland — is the single most important driver of winter weather variability across Europe and the North Atlantic. When it is strongly positive, mild and wet westerly winds dominate Western Europe; when it is strongly negative, cold Arctic air spills southward, bringing severe winters to the British Isles and Scandinavia. Despite decades of research, predicting the NAO weeks to months in advance remains one of the hardest problems in climate science.

One promising avenue is to use sea surface temperatures as predictors. The ocean warms and cools slowly compared to the atmosphere, meaning that its current state carries memory of past conditions — and potentially information about future atmospheric behaviour. But reading that signal is difficult: it is weak, geographically complex, and — crucially — may not be stationary in time. As the climate warms, the ocean is changing in ways that could alter, strengthen, or entirely reshape the statistical relationship between sea surface temperatures and the NAO.

In this study we ask two questions. First, how well can machine learning models of increasing sophistication — from simple linear regression all the way to deep convolutional neural networks — predict the winter NAO from observed sea surface temperature patterns? Second, and more importantly, has that predictability changed between the early observational record (1950–1969) and the most recent decades (2004–2023), a period marked by accelerated warming, dramatic Arctic sea ice loss, and intensifying ocean circulation changes?