A catalogue-verified multiverse audit of machine-learning seismic susceptibility modelling in western Yunnan, China

Abstract. Machine-learning (ML) classifiers are increasingly applied to regional seismic susceptibility map- ping, frequently reporting high discrimination metrics that imply a reliable link between surficial environmental proxies and earthquake locations. We argue that such claims are often conditioned on undocumented analytical choices rather than on a stable physical signal. Using the active Western Yunnan fault belt (97.0°–103.5° E, 22.0°–28.5° N) as a test bed, we execute a transparent, reproducible methodological audit of a complete ML susceptibility workflow. The audit proceeds through four linked stages: catalogue-provenance verification, leakage-aware predictor screening, predictor-coverage quality assurance (QA), and a multiverse evaluation of model behaviour across model families, spatial cross-validation (CV) designs, background-sampling strategies and target definitions. We find that a legacy working inventory contained 173 of 214 records (80.8 %) that could not be verified against the formal China Earthquake Networks Center (CENC) bulletin, and we replace it with a formally verified catalogue spanning 2009–2023. Naïve spatial joining silently discarded the majority of mainshocks; coverage reconstruction recovered the matched sample from 123 to 330 of 331 events (99.7 %). Across twelve defensible analytical branches, the mean spatial area under the receiver-operating-characteristic curve (AUC) ranged from 0.55 to 0.79, with the highest values attached to the least stable configurations. Independent spatial point-process intensity models confirmed that the full surficial predictor stack provided little measurable incremental gain (∆D² = +0.002) over a simple distance-to-fault baseline. We conclude that, under the tested non-circular surficial predictors, the apparent skill of regional ML susceptibility models is highly conditional on analytical specification. We provide an auditable reporting checklist and a predictor roadmap that prioritises deep geodetic and tectonic covariates for future work.