Quantifying earth system interactions during the last glacial period

Abstract

The earth system is a complex dynamical system, which makes it challenging to infer causal interactions and mechanism. Understanding causal interactions during periods of rapid climate change in the past is crucial for building models that can predict future climate response to forcing. The last ice age is a time interval of great interest because it was characterized by rapid climate change events known as Dansgaard-Oeschger events. However, climate models have so far struggled to reproduce such rapid paleoclimate changes, and traditional correlation-based analyses of paleoclimate time series do not provide sufficient insights into the complex causal networks of the climate system. In this thesis I use recently developed methods for causal network reconstruction to test the empirical claim that there are causal connections in the observed paleoclimate records from the last glacial period. Specifically, I use the data-driven Optimal Causation Entropy causal network reconstruction algorithm, combined with the recently proposed Chatterjee and Azadkia-Chatterjee coefficients for measuring statistical associations, to obtain quantitative constraints on the directionality and strength of dynamical coupling between the observed climate system components. My results suggest atmospheric CO2 as a primary driver of the climate system during the last ice age. Specifically, atmospheric CO2 drives Antarctic annual-mean temperature and Patagonian dust production. Furthermore, annual-mean Antarctic temperature and atmospheric CH4 emerge as important secondary drivers. Specifically, Antarctic annual-mean temperature drives Antarctic sea-ice extent and has a smaller role in Patagonian dust production. Atmospheric CH4 contains causally relevant signals on both sea-ice extent and Patagonian dust production. Like any statistical approach, the causal network reconstruction has its limitations, and paleoclimate proxy data come with additional challenges due to indirect measurements and chronological uncertainty. As such, my results should be considered perliminary pending further sensitivity tests. As I hope this thesis shows, however, causal network reconstruction from paleoclimate records can provide an alternative perspective that has the potential to bridge the data-driven and the model-driven climate research communities.