Understanding dependency structures in the major modes of climate variability with causal discovery

Abstract

Accurate climate change projections depend on the ability of climatemodels to correctly represent major modes of climate variability and large-scale teleconnections. However, the evaluation of modes of variability is to date largely based on means, climatologies, or spectral properties which cannot always reveal whether aclimate model correctly simulates the dynamical dependencies between different climatological processes or lagged long-distance teleconnections. The goal of this Master thesis is to construct causal interdependency networks for the major modes of varia-bility from modelsparticipating in the Coupled Model Intercomparison Project Phase 6 (CMIP6).This is done by applyingan existingnovel algorithm forcausal discovery to construct causal net-works for the major modes of variability from models and reanalysisdata and are compared, thus going beyond simple descriptive statistics of the model output. In addition, changes in dependency structures are analysed for different future scenarios. The Tigramite python package uses time series as input data and reconstructs acausal graph. The latter is a special kind of graphical model that translates the conditional dependency structures of the selected components at different time lags. The input time series data are generated using the NCAR Climate Variability Diagnostics Package (CVDP) within the Earth System Model Evaluation Tool (ESMValTool). The ESMValToolis a community diagnostics and performance metrics tool for the evaluation of Earth System Models (ESMs). For the resulting time series of the major modes of climate variability, a graphical model is constructed that encodes the link strength by which climate variability indices andcertain atmospheric or ocean variables (sea surface temperature, sea level pressure, precipita-tion) are interconnected to modes of climatevariability. The graphical model also provides thetime lagby whichthe detected causality is effective. First, the methodisused for CMIP6 datatoreconstruct two climateexamples that have previously been applied to reanalyses data.Theyil-lustrate teleconnections between surface-air temperatures (TAS) overEastern and Western Europe, in addition to the interplay between sea level pressure(PSL)anomalies over the Western Pacific and TAS anomalies over both the Central and Eastern Pacific.Performance of recreating the same results as the NCEP-NCAR reanalysis datasetis found to vary from one model to the other. None-theless,themulti-model mean (averaging 6 CMIP6 models) reproducesall links created by the reference datasetwith similar time lags.Second, the algorithm is used to analyse the interplay between El Niño-Southern Oscillation (ENSO) and four other major modes of climate variability: the Pacific DecadalOscillation (PDO), the Pacific North American pattern (PNA), the Indian Ocean Dipole(IOD), and the North Atlantic Oscillation (NAO). The idea is to study the basin-wide teleconnections involving these modes over the Pacific Ocean, the Indian Ocean, and the North Atlantic Ocean. Here,a regime-orientedanalysis is appliedto showhow the basin-wide ENSO connections decay when PDO and ENSO are out of phase.Based on comparison with the reference reanalysis dataset, most of the models reproduce the well-established strong connectionbetween PDO and PNA, and the negatively correlatedPNA─NAOconnectionfor historical and pre-industrial controlsimulations. Results also show thatthe causal network changes according to different future scenarios.Not only connections arefound to weakenbut also the overallnumber of connections is found to drop significantly when moving from a moderate radiative forcing scenario to a high radiative forcing scenario.