Deep learning illuminates atmospheric blocking events of past, future

"This study set out to extract a paleoweather signal from paleoclimate records using a deep learning model that infers atmospheric blocking frequency from surface temperature," said Karamperidou. "This is a unique study and the first attempt to reconstruct a long record of blocking frequencies based on their relationship with surface temperature, which is complex and unknown. Machine learning methods can be very powerful for such tasks."

Training the deep learning model

Karamperidou developed a specialized deep learning model, which she trained using historical data and large ensembles of climate model simulations. The model was then capable of inferring the frequency of blocking events from anomalies in seasonal temperature reconstructions over the Last Millennium. These past temperature reconstructions are relatively well-constrained by extensive networks of tree-ring records sensitive to temperature during the growing season.

"This approach demonstrates that deep learning models are powerful tools to overcome the long-standing problem of extracting paleoweather from paleoclimate," said Karamperidou. "This approach can also be used for the instrumental period of climate history, which began in the 18th century when routine weather measurements were made, since we only have reliable data to identify blocking since the 1940s, or possibly only the satellite era (post-1979)."

Frequency of future blocking events

There isn't yet a scientific consensus regarding how climate change will change the frequency of blocking events. These strong, persistent mid-latitude high-pressure systems can have significant impacts for Hawai'i, where flooding has accompanied persistent North Pacific blocks, and also worldwide, for example, in the Pacific Northwest and Europe, where summertime blocking can bring extreme heat waves.

So understanding changes in the frequency of these events, especially as they relate to the other big players for climate, such as El Niño and the long-term patterns of sea surface temperatures in the tropical Pacific, is very important for Hawai'i. This study allowed Karamperidou to relate blocking frequencies in the mid- and high-latitudes to tropical Pacific climate variability in the long context of the last millennium, which is essential for climate model validation and to narrow down uncertainties in future climate projections of blocking.

Open research and transparency

Karamperidou worked with two UH Manoa students to create a unique web-interface to explore the deep learning model and the resulting reconstructions. She highlighted that sharing results and methods in this way is important for Open Research best practices and transparency, especially as the application of machine learning and artificial intelligence expands rapidly into many aspects of daily life. The web-interface is hosted on Jetstream-2, an NSF-supported cloud computing system whose regional partners include the University of Hawai'i Information Technology Services -- Cyberinfrastructure and the Hawai'i Data Science Institute.

In the future, Karamperidou plans to explore a range of features and architectural enhancements of the deep learning model to expand its applications for climate phenomena and variables directly related to high socioeconomic impacts.