Multi-proxy Evaluation of Abrupt Climate Transition Predictability in Paleoclimate Records

The Dansgaard-Oeschger (D-O) events represent iconic tipping points in the Earth's climate system. However, objectively identifying these transitions and extracting reliable early warning signals (EWS) from high-resolution but noisy paleoclimate archives remains a significant challenge. In this study, we implement a systematic framework to evaluate and compare multiple computational methods for identifying abrupt climate shifts in paleoclimate records. To address the non-stationarity and proxy-specific noise inherent in different records, we employ an adaptive signal decomposition technique. This allows for the extraction of high-frequency dynamical features to quantify indicators of critical slowing down, specifically temporal autocorrelation within sliding windows. Results indicate that the deep learning-based framework exhibits superior robustness in capturing transient waveforms across different proxy types compared to conventional linear or state-space models. Notably, we observe significant discrepancies in transition timing and EWS strength between the different records. High-frequency atmospheric components demonstrate a more pronounced loss of resilience prior to major D-O transitions, suggesting that atmospheric reorganization may serve as a highly sensitive precursor to large-scale climate reorganization. Our findings highlight the potential of combining machine learning with advanced signal processing to diagnose the proximity of climate thresholds. This integrated framework provides a robust basis for assessing the stability of the coupled ice-ocean-atmosphere system and offers new insights into the predictability of abrupt climate changes during the last glacial period.