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

Deep-learning phase pickers are increasingly deployed on borehole microseismic arrays, yet two questions central to operational monitoring remain poorly quantified: how well such models generalize to a site they were not trained on, and whether the cross-station moveout that an array provides actually improves that generalization. We give the first confound-free answer using the AMBER benchmark of eight three-component downhole arrays. A velocity-free, waveform-direct picker is trained under a strict leave-one-site-out (LOSO) protocol — seven sites for training and zero-shot evaluation on the eighth with no fine-tuning — and the contribution of moveout is isolated by an ablation in which the identical two-dimensional U-Net is trained either on the full array (the station axis is convolved, so moveout is available) or per trace (single-station inputs, moveout removed), so that the only difference is the presence of moveout. Across the eight held-out sites the two configurations are statistically indistinguishable in median F1 (0.885 versus 0.878; event-bootstrap 95 per cent confidence intervals overlap at the majority of sites): array moveout confers no systematicgeneralization benefit. The effect is instead strongly site-dependent and, at one site, catastrophic. Where a new array’s median P moveout exceeds the largest value seen in training (forge_19, 70 ms versus a 62 ms training maximum) the array P detector collapses (F1 0.18) while the per-trace model recovers it (0.76); probability diagnostics show a detection failure, not a timing error, and a station-shuffle control attributes the collapse to non-transferable cross-station coupling. Because unattended monitoring is governed by worst-case rather than average performance, we recommend the velocity-free per-trace picker as the safer default and treat array moveout as a conditional, in-distribution enhancement guarded by an explicit out-of-distribution check.