Robust Bayesian Location and Event-Level Quality Control
Manuscript-specific reproducibility archive for BSSA-D-26-00196. This release uses the completed, independently initialized run 46, not the superseded shared-initializer/pilot-overlap HypoSVI experiments.
Start with REPRODUCE.md, ARTIFACT_INDEX.md, and the real-data policy.
Current evidence
The common synthetic input has 3,206 events and 86,827 P/S observations. Per event, round(0.15 × phase count) arrivals receive independent Gaussian gross errors with an 8 s standard deviation. All methods retain the same contaminated picks. The common numerical set has 2,998 events and 86,411 rows; 208 single-station P+S inputs count as failures in the all-input summary.
| Method | Mean horizontal error (km) | Mean depth error (km) | Mean time error (s) |
|---|---|---|---|
| Student-t + indicator | 2.992 | 4.676 | 0.285 |
| Student-t only | 3.161 | 5.239 | 0.316 |
| NLLoc EDT | 7.561 | 8.472 | 0.698 |
| HypoSVI run 46 | 5.884 | 5.406 | 0.419 |
NLLoc uses fresh 32-worker EDT results and reported P/S uncertainties of 0.3/0.4 s, as does HypoSVI. These are distinct from the ordinary generative noise standard deviations of 0.2/0.3 s.
HypoSVI uses independently trained official EikoNet P/S networks on the same 3-D physical model. Raw initializer candidates use input information only; HypoSVI ranks/refines them with its own EikoNet, not the proposed surrogate. Eight configurations and two seeds on 60 disjoint validation base events select 300 particles, 350 epochs, and step size 0.5. The test freezes this choice with seed 82001. Validation includes clean/contaminated pairs but does not establish any cross-method clean-data advantage. The formal run completed on NVIDIA TITAN RTX CUDA; its archived 503-check audit passed.
The proposed surrogate is supervised using labeled source–receiver travel times, without reading a velocity field or enforcing an eikonal loss. The shared velocity model generated controlled training/test labels offline and defines the comparator physical reference.
Scope of QC and real-data claims
Posterior widths correlate with synthetic error and define lower-error retained subsets. Nominal intervals remain undercovered, and chains do not explore every remote mode. The additional saved-chain check applies the actual QC score, with explicit diagnostic thinning, and supports approximate within-selected-basin repeatability rather than calibrated error probabilities.
Real results use 10,487 direct windows / 224,644 valid phase rows and 10,950 associated events / 176,243 valid rows. All three locators receive the same respective input. HypoSVI uses published defaults (150 particles, 175 epochs, step size 1.0), without reference-catalog tuning. At K=7,000, associated-input recall differences between the proposed method and HypoSVI are statistically unresolved for CENC, Liu et al., and their derived consensus; this does not establish equivalence. The consensus is not a third independent catalog. Real results measure reference-relative agreement, not absolute location error or true completeness.
Public and restricted material
Public: synthetic observations and truth; 3-D velocity structure; trained supervised and EikoNet models; pinned comparator source; source-only tests; synthetic posterior diagnostics/particles; configurations; aggregate results.
Not public: real earthquake/station coordinates, identifiers, origin times, phase files, station tables, posterior samples, event-level differences, or restricted-input fingerprints. No file with a .pha extension is included. Real-data rendered media and manuscript PDFs/maps are also excluded from this release. Authorized real-input reproduction requires the original permitted data and saves all event-level products in an ignored private area. These materials exist but are not redistributed.
Obtain and cite the Luding comparison data through Liu et al. (2023), Auto-Building of Aftershock Sequence and Seismicity Analysis for the 2022 Luding, Sichuan, Ms 6.8 Earthquake, Acta Geophysica Sinica 66(5), 1976, https://doi.org/10.6038/cjg2023Q0841. CENC inputs remain restricted.
Computational provenance — SI only
The manuscript contains no performance section or timing table. SI Table S7 preserves native location-stage timings on one workstation. NLLoc uses 32 CPU workers; both neural methods use the same TITAN RTX.
Direct/associated times are 446.5/390.2 s for NLLoc, 824.5/699.5 s for all three proposed chains, and 19,526.4/15,328.1 s for HypoSVI. NLLoc is fastest in this implementation-level record. Timer boundaries differ slightly and exclude training/table construction, initialization, specified I/O, and downstream analysis. These are not end-to-end, hardware-independent, or equal-posterior-quality speed claims.
Integrity
- HypoSVI upstream: 4d449575d3266507741e844e4f1a39935c65f987.
- EikoNet upstream: 59dd11f59e4f327188ebd609d0a0666d76fd983f.
- NLLoc upstream: 062a4bee41d0f94dac404f09e139b7c8b3348a52 (7.1.04).
- Run-specific synthetic hashes and selection attestations are retained.
- MANIFEST.sha256 verifies this release's file bytes.
- tools/privacy_audit.py is a mandatory current-tree and history gate.
Do not merge an old clone's history into the cleaned public release. An old checkout can retain withdrawn restricted derivatives in its local history; archive it privately and use a fresh clone of the cleaned release.