Dependency structures of climate variability patterns through causal discovery algorithms

Kurzfassung

Understanding and accurately projecting climate change and its impacts is critical for guiding global efforts in mitigation and adaptation. Earth system models (ESMs), such as those participating in the Coupled Model Intercomparison Project (CMIP), represent the physical, chemical, and biological processes and feedbacks of the Earth’s climate and are the primary tools to project climate change. However, the inherent complexity of these models leads to significant uncertainties in future climate projections. This dissertation aims to address these uncertainties by developing and applying advanced causal discovery algorithms and statistical techniques to uncover the complex interactions driving the patterns of climate variability. The dissertation is structured around three major contributions focused on uncertainty quantification: causal effect estimation, time series causal discovery, and their applications to reduce the uncertainty of climate model projections. Regarding the first contribution, this thesis develops a framework for establishing uncertainty bounds for long-term causal effects in spatio-temporal systems. This involves deriving explicit formulas for quantifying long-term effects in spatially aggregated vector autoregressive (SAVAR) models, which provide a deeper understanding of the system’s response to external perturbations. The asymptotic properties of these estimators are also derived, offering possibilities to quantify uncertainties of the estimates under suitable conditions. To validate the theoretical findings, several numerical experiments are conducted, comparing different methods for estimating the effects of an external perturbation on the system. Among the methods tested, the approach using causal discovery consistently outperforms others, underscoring that causal discovery is a powerful tool to investigate complex systems. Second, recognizing the importance of accurately capturing temporal dependencies in complex systems, this dissertation further develops advanced methodologies to enhance time series causal discovery. In particular, it introduces a novel bootstrap approach for time series causal discovery that preserves temporal dependencies and lag structures. The original time series data is bootstrapped using the temporal-preserving approach, and a causal discovery algorithm is run independently on each bootstrap sample to produce the same number of graphs. These graphs are then aggregated by majority voting into an output graph, with the frequency of each edge type in the bootstrap ensemble providing confidence measures for the causal graph links. While it increases computational demands, the approach enhances the performance of time series causal discovery algorithms. Its versatility allows it to be integrated with various causal discovery methods, making it a valuable tool for a wide range of real-world applications, including applications in climate science. Third, the dissertation develops methods with the Earth System Model Evaluation Tool (ESMValTool), a community diagnostic and performance metrics tool for routine evaluation of Earth system models, to first derive unweighted multi-model precipitation and temperature projections. This study reveals large uncertainties in these unweighted projections, particularly regarding precipitation over land, highlighting the challenge of accurately projecting future precipitation under climate change. Building on this, this thesis then develops an approach that reduces uncertainties in climate model projections for precipitation over land with causal discovery. By performing causal discovery on sea level pressure (SLP) components from Coupled Model Intercomparison Project Phase 6 (CMIP6) climate models, the thesis identifies dynamical patterns of SLP that influence precipitation variability. It then introduces a novel multi-model weighting scheme, termed causal weighting. This weighting scheme puts more weight on models that exhibit higher performance in reproducing causal fingerprints of observed dynamical SLP patterns and on models with more distinct patterns to consider interdependencies in the ensemble. This weighting scheme is successfully applied to CMIP6 and reduces the uncertainty range of precipitation projections over land, demonstrating its effectiveness. Overall, the thesis advances the state-of-the-art in causal discovery methods and effect estimation, as well as their applications to climate science. The findings provide a methodological foundation for improving climate model projections, particularly in areas where uncertainties have profound implications, ultimately supporting more effective decision-making in a world that experiences rapid climate change.