Pick Aggregator: A Multimodel Ensemble of Machine Learning Pickers to Generate an Onshore–Offshore Seismic Catalog for Puerto Rico and the Virgin Islands

Abstract Offshore seismic swarms north of Puerto Rico and the Virgin Islands remain poorly characterized due to limited azimuthal coverage from the inland Puerto Rico Seismic Network (network code PR). Here, we integrate data from six U.S. Geological Survey (USGS) ocean-bottom seismometers (ZZ network, mid-2015 to mid-2016) with the PR network using a novel multimodel ensemble approach for machine learning (ML) phase picking, named pick aggregator. ML pickers have automated rapid seismic phase detection, but their performance degrades when applied to data that differ significantly from training datasets. Onshore–offshore experiments with sparse station coverage, complex instrumentation, and lower signal-to-noise ratios provide a challenge. The pick aggregator performs multidomain parallel computations on the SeisBench platform to catalog seismic phases across eight ML model combinations pretrained on ocean and land data. We then aggregate seismic picks identified by at least two models and associate them using a Gaussian Mixture Model Associator. The 1-yr ensemble ML catalog includes 10,948 events, composed of 55,704 P wave and 33,899 S wave picks, with mean probabilities of 0.65 and 0.64, respectively. The ZZ network substantially improves offshore azimuthal coverage and provides thousands of near-source seismic phases (25% of the P wave and 28% of the S wave picks). The relocated-ML catalog with >5.3 M added differential times includes 8025 events (115.8% increase compared with the relocated USGS catalog using the same stations). The relocated-ML catalog reveals an intense seismic swarm in the north-northwest direction southeast of the Main ridge at 20–40 km depth that aligns with the fast shear-wave polarization of previous studies, suggesting an incipient slab tear near 65.1° W. Seismicity depth variations along the strike indicate a steeper western segment with a wider depth range and a retreating eastern segment, supporting a differential rollback mechanism. Our multimodel ensemble approach mitigates inconsistencies among ML pickers pretrained on different datasets, enabling superior performance and robust phase detection in diverse tectonic environments with varying signal characteristics.