When does array moveout help borehole phase picking? A leave-one-site-out, confound-free benchmark of array versus per-trace deep learning

ISAO KUROSAWA

Microseismic monitoring of fluid-injection operations — enhanced geothermal systems, hydraulic fracturing and geological CO2 storage — is central to managing induced seismicity and imaging reservoir processes (Ellsworth 2013; Grigoli et al. 2017), and is increasingly performed on dense borehole and fibre-optic (DAS) arrays (Karrenbach et al. 2019; Lellouch et al. 2019; Lindsey & Martin 2021). It depends on detecting and timing P and S arrivals in continuous borehole records, often at sampling rates an order of magnitude higher than the regional networks on which most deep-learning (DL) pickers were trained. Established surface-trained models such as GPD (Ross et al. 2018), PhaseNet (Zhu & Beroza 2019) and EQTransformer (Mousavi et al. 2020), distributed through frameworks such as SeisBench (Woollam et al. 2022), have transformed catalogue completeness, but their behaviour on high-frequency downhole arrays is only beginning to be characterized. Systematic cross-domain benchmarking shows that such pickers transfer within a data domain yet degrade on recordings with different characteristics (Münchmeyer et al. 2022), and induced-seismicity applications often require in-domain retraining (Park et al. 2020; Chai et al. 2020). Lim et al. (2025) benchmarked four surface-trained DL pickers on the Preston New Road (PNR-1z) borehole dataset and found that, although the models detect induced seismicity, they do not transfer to the high-frequency borehole regime without fine-tuning. That study frames the problem we address here: rather than asking how well surface-trained models transfer, we ask how well a borehole-trained model generalizes to an unseen borehole site, and whether using the array — specifically the across-station moveout — helps or hurts that generalization.
Two design philosophies are in tension. An array-level picker can, in principle, exploit the coherent sweep of an arrival across depth-ordered stations (moveout) as an additional discriminative cue, much as array methods and analysts do (Rost & Thomas 2002). A per-trace picker ignores this cue and treats each station independently. The array cue is attractive, but it couples the model to the geometry of the training arrays; if a new site has a different aperture, station spacing or source–array configuration, a moveout-dependent detector may extrapolate poorly. Whether the net effect across realistic sites is positive, negative or neutral is an empirical question that, to our knowledge, has not been tested in a confound-free way.
We make three contributions. (i) We evaluate a velocity-free, waveform-direct picker under a strict zero-shot LOSO protocol on the eight-site AMBER benchmark, reporting event-bootstrap confidence intervals (Efron & Tibshirani 1994) so that per-site differences are interpretable rather than anecdotal. (ii) We isolate the contribution of moveout with a confound-free ablation — the same network, data and training schedule, differing only in whether the station axis is available — avoiding the sampling-rate, architecture and training-domain confounds that make off-the-shelf comparisons inconclusive. (iii) We quantify when moveout helps by relating the array advantage to a data-driven measure of each site’s across-station moveout, and we diagnose the dominant failure mode as out-of-distribution geometry rather than intrinsic difficulty. The result is a cautionary, operationally actionable finding: across sites there is no systematic benefit to array moveout, the per-trace picker is more robust in the worst case, and we recommend it as the safer default for unattended monitoring.