Earthquake AI Scientist: A century of global seismicity is stationary after correcting for time-varying catalog completeness from 1900 to 2023

Keywords: Earthquake, Agent

TL;DR: Earthquake AI Scientist

Abstract: Reported global earthquake counts rise through the twentieth century, but whether this reflects a physical increase in seismicity or evolving detectability has remained unresolved. We compile a 1900-2023 global catalogue (37,331 significant events; minimum M ~ 5.5) and explicitly reconstruct a time-varying magnitude of completeness, Mc(t), across instrumentation eras. Gutenberg-Richter parameters are estimated strictly above Mc(t) with bootstrap uncertainty, and completeness-consistent rate tests are conducted globally and by tectonic regime. Once Mc(t) is enforced, the apparent secular rise in counts disappears; decadal b-values are stable within uncertainty; and large-earthquake rates are statistically stationary at robust thresholds (notably M >= 7.5-8.0). We further translate this baseline into decadal forecasts: for 2025-2034 we predict N10(M >= 7.5) = 42.3 events on average (90% interval [32, 53]) and N10(M >= 8.0) = 8.3 ([4, 13]), implying about 56% annual probability of at least one M >= 8.0 event. These results resolve the long-standing debate over secular changes in global seismicity and establish a completeness-aware, uncertainty-quantified foundation for probabilistic hazard models, risk pricing and performance benchmarking of global monitoring networks. Beyond seismology, the framework generalises to observation-limited records (e.g., volcanic unrest, landslides, epidemiological surveillance, biodiversity monitoring), providing a template for bias-corrected trend detection and forecasting.

Submission Number: 30