Source-linked AI summary
PhaseLink: A Deep Learning Approach to Seismic Phase Association
Zachary E. Ross, Yisong Yue, Men-Andrin Meier, Egill Hauksson, Thomas H. Heaton
TL;DR
Seismic phase association must link detections from different sensors to common earthquakes despite unknown event counts and temporal overlap. PhaseLink uses recurrent neural networks to learn this association without a hypocenter grid, training on synthetic pick sequences from a simple 1D model. It achieves strong performance on dense southern California sequences and is reported to associate events separated by only seconds, while remaining limited by its demonstrated local-network scope and lack of hypocenter determination during association.
Problem
Phase association must determine how many earthquakes occurred and which picks belong to each event when sequences overlap and the number of clusters is unknown.
Method
PhaseLink uses recurrent neural networks to associate time-ordered picks without searching a hypocenter grid, training on synthetic sequences generated with a simple 1D velocity model.
Results
PhaseLink achieved outstanding performance on an active southern California aftershock sequence containing events separated by only seconds in origin time.
Takeaways & Limitations
The method can be trained entirely on synthetic arrival-time data and applied to networks lacking large labeled training datasets.
Takeaways & Limitations
The demonstrated method targets local seismic networks, and applying it regionally or teleseismically may require additional modifications.
Abstract
from arXiv · showhide
Seismic phase association is a fundamental task in seismology that pertains to linking together phase detections on different sensors that originate from a common earthquake. It is widely employed to detect earthquakes on permanent and temporary seismic networks, and underlies most seismicity catalogs produced around the world. This task can be challenging because the number of sources is unknown, events frequently overlap in time, or can occur simultaneously in different parts of a network. We present PhaseLink, a framework based on recent advances in deep learning for grid-free earthquake phase association. Our approach learns to link phases together that share a common origin, and is trained entirely on tens of millions of synthetic sequences of P- and S-wave arrival times generated using a simple 1D velocity model. Our approach is simple to implement for any tectonic regime, suitable for real-time processing, and can naturally incorporate errors in arrival time picks. Rather than tuning a set of ad hoc hyperparameters to improve performance, PhaseLink can be improved by simply adding examples of problematic cases to the training dataset. We demonstrate the state-of-the-art performance of PhaseLink on a challenging recent sequence from southern California, and synthesized sequences from Japan designed to test the point at which the method fails. For the examined datasets, PhaseLink can precisely associate P- and S-picks to events that are separated by ~12 seconds in origin time. This approach is expected to improve the resolution of seismicity catalogs, add stability to real-time seismic monitoring, and streamline automated processing of large seismic datasets.
1 Introduction
Seismic phase association evolved from manual grouping to automated algorithms that must distinguish earthquake phases from unrelated transients and handle increasingly dense sequences. PhaseLink addresses these challenges with a grid-free deep-learning approach trained on synthetic data.
- Motivation: Automated phase association must determine whether detected signals belong to earthquakes, not merely group seismic phases.STA/LTA detects impulsive transients broadly, so association examines arrival-time patterns across stations for earthquake-consistent combinations.
- Limitations of prior methods: Existing associators continuously search gridded regions and typically require many sensitive hyperparameters and ad hoc stabilization rules.Modern variants may incorporate pick uncertainties, machine learning, or multi-scale detection, but retain the grid-search principle.
- Motivation: Smaller, more frequent events produce denser sequences and substantially larger data volumes that require reliable processing of complicated scenarios.The challenge is especially acute when events occur close together in time and signals approach the noise level.
- PhaseLink: PhaseLink uses recurrent neural networks for grid-free association of temporally overlapping earthquake phases.The method learns temporal and contextual relationships in sequential data rather than searching over a hypocenter grid.
- PhaseLink: PhaseLink is trained entirely on synthesized data and can be adapted to tectonic regimes through an appropriate 1D velocity model while incorporating pick-time errors.The authors also describe the approach as suitable for real-time processing and provide source code for public release.
2 Background on Recurrent Neural Networks
Recurrent neural networks extend neural networks with memory that passes information across sequence elements. This sequential context makes them suitable for phase association, including real-time processing as picks arrive one at a time.
- RNN fundamentals: Recurrent neural networks pass information between successive sequence elements through an internal memory state controlled by learned gates.The gates determine what information is retained during processing.
- RNN variants: Long short-term memory networks use three gates to reduce problems with diminishing information propagation over long sequences.Gated recurrent units use two gates, requiring fewer parameters and enabling faster training.
- RNN variants: Bidirectional recurrent layers process sequences in opposite directions so predictions can use contextual information from both sides.This is useful when information later in a sequence helps interpret an earlier element.
- Application to phase association: RNNs fit phase association because seismic picks form time-ordered sequences and picks arrive incrementally in real-time networks.Their sequential processing supports association as each new phase becomes available.
3 PhaseLink Framework
PhaseLink frames earthquake phase association as supervised clustering of picks, using an RNN to classify links within overlapping windows and aggregate them into earthquake clusters. Synthetic training data and explicit cluster aggregation allow the framework to address overlapping events, false picks, and varied arrival-time conditions.
- PhaseLink treats association as supervised clustering: given N picks, it determines the number of earthquakes and assigns picks to each event.
- The method converts the pick stream into overlapping fixed-length binary classification tasks asking whether each pick matches the first, or root, pick's earthquake.
- A stacked bidirectional GRU RNN receives fixed-length pick sequences and outputs binary links to the root pick.
- Using sliding windows of 500 picks, PhaseLink produces root-relative predictions and aggregates them into clusters representing detected earthquakes.
- Each pick uses five features: station latitude, station longitude, time relative to the root, phase type, and a padding indicator; picks beyond 120 s are discarded and the remainder padded.
- Clusters nucleate from predicted links, merge when they share enough picks, and are retained or discarded according to cluster-size rules.
- The completed association assigns picks to events but does not determine hypocenters, unlike associators that jointly solve event location.
- Synthetic 1D velocity-model data provide scalable training, arrival-time errors, and transfer across tectonic regimes; validation accuracy reaches 99.92%.
4 Results
Controlled tests evaluate PhaseLink on real and synthetic sequences, including overlapping events and comparisons with a conventional grid associator. The method maintains high precision, recovers missed picks through clustering, and remains effective for events separated by roughly 10–12 seconds.
- Controlled tests use known ground truth for every pick, enabling phase-level assessment and direct comparison with existing grid association methods.
- The clustering stage recovers many neural-network misses because each pick needs correct association in only one sub-sequence.
- 14 earthquakes within roughly six minutes are associated successfully, including events spaced only 5–10 seconds apart in origin time.
- For all nmin values, event precision exceeds 0.996, while recall decreases from 0.956 to 0.891 as higher thresholds discard small clusters.
- Against dbgrassoc, PhaseLink has nearly identical event precision (>0.996) but higher event recall (0.956 vs 0.919) and phase recall (0.955 vs 0.904).PhaseLink detects 63 more events and correctly associates 7,482 phases without sacrificing precision.
- In Japanese stress tests, reasonable performance ends when events are approximately 10–12 seconds apart, with precision 0.603 and recall 0.197 at 4-second average spacing.
5 Discussion
PhaseLink addresses dense-sequence phase association without hypocenter grids, using deep neural networks to group phases by seismic propagation patterns. The discussion highlights synthetic-data training, explicit handling of picking errors, strong stress-test performance, and scope for large-scale automated processing.
- Method development: Persistent failure cases can be addressed by adding examples to the training dataset rather than tuning hyperparameters or modifying the algorithm.This is presented as a fundamental difference from grid-based association methods.
- Method: PhaseLink avoids grid searches and earthquake location during association by grouping phases with arrival-time patterns consistent with seismic wave propagation.The grouped phases can then be passed to a location algorithm to determine a hypocenter.
- Training data: Synthetic training data makes PhaseLink applicable to networks lacking abundant labeled phase data, provided a sufficiently accurate velocity model is available.The authors suggest that a 1D layered model is probably adequate in most cases.
- Training data: Picking errors can be incorporated during training, including errors drawn from U ∼[−0.5, 0.5] s.The authors state that this can also help the network handle arrival-time perturbations caused by 3D structure.
- Applications: PhaseLink could improve automated processing of large seismic datasets and support seismic monitoring, microseismic monitoring, and earthquake early warning.The early-warning application is described as potential and depends on fully using recurrent neural networks’ long-term memory.
6 Conclusions
The paper develops a deep-learning method for seismic phase association and reports strong performance on an active southern California aftershock sequence. It can use synthetic arrival-time data and does not require grids.
- PhaseLink achieves outstanding performance on an extremely active southern California aftershock sequence with events only seconds apart in origin time.
- PhaseLink can be trained solely on synthetic seismic phase arrival-time data, enabling application to networks without large labeled training datasets.
- PhaseLink does not use grids for seismic phase association.