New Framework Detects Geographic Tipping Points with Causal Networks.

Detecting localized geographic tipping points in complex systems such as ecosystems, climate subsystems, or ice sheets presents significant challenges for early warning systems. Traditional spatial indicators often fall short due to issues like spatial dilution, reliance on Euclidean assumptions, and susceptibility to correlated noise. This research introduces a novel framework, SpatioTemporal Causal Network Diagnostics (ST-CND), to overcome these limitations. ST-CND conceptualizes a geographic field as a dynamic, time-evolving directed causal network. Its core workflow involves three key steps: first, it infers causal links between spatial nodes using transfer entropy, replacing fixed spatial neighborhoods with data-driven information flow topology. Second, it estimates local recovery rates within identified subnetworks using dynamic mode decomposition. Finally, it pinpoints the most vulnerable subnetworks by combining signals of high internal fluctuation, strong internal synchronization, and low external coupling, effectively filtering out false alarms from spatially correlated noise. Validated on synthetic bifurcations and real-world sea-surface temperature benchmarks, including the Indo-Pacific SST and North Atlantic AMOC, ST-CND provides localized and interpretable warnings. For the AMOC task, it achieved an AUROC of 0.783, outperforming existing baselines, demonstrating its potential for robust spatial early warning in Earth system science.