Modelling the Earth System - From Tipping Elements to Reconstructions

Abstract

Climate change can trigger climate tipping points, which are among the major threats to human society. Tipping points are thresholds beyond which a system undergoes abrupt, often irreversible, changes even if the external forcing is brought to a halt. Several large-scale elements in the Earth system are considered tipping elements with global consequences once critical thresholds are crossed and self-reinforcing changes are triggered. However, there is a large uncertainty as to whether some Earth system components should be considered tipping elements. The precise values of the critical thresholds remain uncertain, and it is unclear whether these can be temporarily exceeded without triggering a tipping point. Moreover, incomplete historical records complicate the inference of past dynamics of these components and current reconstruction methods introduce biases into higher-order statistics that are used to assess their stability. On the other hand, with the increasing availability of data and advancements in computational power, deep learning (DL) offers new advances in climate science, ranging from reconstructions to hybrid climate models. This thesis presents an in-depth study of two distinct tipping elements: the Greenland ice sheet (GrIS) and the coupled system of the South American Monsoon and the Amazon rainforest (SAMS). Furthermore, we introduce a novel deep learning-based method to reconstruct spatiotemporal climate fields. By combining model- and observation-based analyses, we show that the SAMS is approaching a critical transition in response to deforestation, potentially leading to a large-scale reduction in precipitation rates in large parts of South America. We associate the critical transition with a weakening of the oceanic moisture inflow due to forest degradation. Subsequently, we use two independent ice-sheet models and show for the first time that the GrIS's critical threshold can be temporarily exceeded without prompting a transition to an alternative state. Timely reversal of surface temperatures can prevent a complete retreat of the ice sheet due to the slow timescale of the ice loss. Lastly, we present a new deep learning-based reconstruction method. The model learns the underlying spatial relationships from climate model output and can inpaint observation-based datasets. Our method outperforms previous reconstruction methods and can realistically reconstruct known historical events, highlighting the potential of DL.