Artificial Intelligence reconstructs information about past climate extremes

Our understanding of long-term changes and variations in regional climate extremes is limited due to missing observational data – the further we go back in time, the sparser the observational record becomes. This lack of long-term data complete also undermines a more complete understanding of the underlying processes driving these extreme events. In our research, published in the journal Nature Communications, we use Artificial Intelligence (AI) to reconstruct missing information about observed warm and cold temperature extremes in the early 20th century. The Deep Neural Network is trained on a large ensemble of historical climate model simulations, to learn the typical spatial relationships of temperature extremes across the European region. We scrutinised and evaluated the fidelity of the method in different data set-ups by testing the ability to predict withheld but known data, and this evaluation showed that the AI-based reconstruction outperforms other statistical methods to reconstruct that missing data. Interestingly, the reconstructed climate extremes data enabled us to identify both warm and cold extremes in the early 20th century, for which no measurement-based data existed but whose occurrence could be verified based on anecdotal contemporary reports and other proxies like mortality data. This novel AI-reconstructed long-term climate extremes dataset therefore provides an important basis for a robust quantitative understanding of long-term changes in regional climate extremes, and to study the physical processes that were causing these historical climate extremes.