Answers are grounded in this paper and cite their sources.
Automated catalogs combine ordinary picking variability, moderate unmodeled travel-time effects, and gross errors or misassociations. Our Bayesian location framework accounts for these effects through Student-t residuals and latent inlier–outlier indicators—binary pick assignments inferred jointly with source parameters. Event-specific residual scales accommodate heterogeneous data quality; pick-anomaly probabilities and hypocentral uncertainty support phase- and event-level screening. A neural surrogate predicts travel times from a 3D velocity model, and a station-informed search selects a low-misfit starting solution for posterior sampling. With approximately 15% contaminated phases, the joint model yielded mean horizontal, depth, and origin-time errors of 2.992 km, 4.676 km, and 0.285 s on 2998 common synthetic events, below those of a Student-t-only variant with the indicators removed (an ablation test), NLLoc, and the tested hypocenter inversion with Stein variational inference (HypoSVI) configuration. Retaining half the events with the lowest posterior-width scores reduced horizontal and depth errors to 1.229 and 1.722 km, supporting empirical event ranking. For associated-phase data from the 2022 Luding sequence, the 7000 lowest-uncertainty events achieved China Earthquake Networks Center (CENC)-relative recall of 0.855, the fraction of reference events recovered by one-to-one matching within 1 s and 5 km. For associated inputs, CENC-relative recall was similar to that of HypoSVI, and indicator ablation showed little additional gain on these inputs.
No takes yet. Share an insight, caveat, or question.
A 2026 study studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: