Extracting paleoweather from paleoclimate through a deep learning reconstruction of Last Millennium atmospheric blocking

Projected changes in atmospheric blocking and associated extreme weather are marked by considerable uncertainties. While paleoclimate records could help reduce these uncertainties, their low temporal resolution makes extracting synoptic-scale signals challenging. Here, a deep learning model is developed to infer summertime blocking frequency from tree-ring-based gridded reconstructions of Northern Hemisphere surface temperature over the Last Millennium. The model, despite not directly incorporating paleoclimate proxies or their locations, is implicitly constrained by them. The reconstructions highlight the tropical Pacific’s strong influence on blocking variability at interannual-to-centennial time scales. A weakened tropical Pacific zonal temperature gradient during the Little Ice Age correlates with a hemispherically reduced -yet more variable interannually- blocking frequency and altered regional patterns. This deep learning approach offers a pathway for extracting paleoweather signals from paleoclimate records that enables improved understanding of blocking response to external forcing and constraining of model projections of blocking under climate change scenarios. The tropical Pacific’s weakened zonal temperature gradient during the Little Ice Age significantly impacted blocking variability, resulting in reduced blocking frequency and altered regional patterns, according to a deep learning model for extracting paleoweather signals from paleoclimate records in the Northern Hemisphere.