Abstract
Seismic phase association is the task of grouping phase arrival picks across a seismic network into subsets with common origins. Building on recent successes in this area with machine learning tools, we introduce a neural mixture model association algorithm (Neuma), which incorporates physics-informed neural networks and mixture models to address this challenging problem. Our formulation assumes explicitly that a dataset contains real phase picks from earthquakes and noise picks resulting from phase picking mistakes and fake picks. The problem statement is then to assign each observation to either an earthquake or noise. We iteratively update a set of hypocenters and magnitudes while determining the optimal class assignment for each pick. We show that by using a physics-informed Eikonal solver as the forward model, we can impose stringent quality control on surviving picks while maintaining high recall. We evaluate the performance of Neuma against several baseline algorithms on a series of challenging synthetic datasets and the 2019 Ridgecrest, California sequence. Neuma outperforms the baselines in precision and recall for each of the synthetic datasets. Furthermore, it detects an additional 3285 more earthquakes than the best baseline on the Ridgecrest dataset (13.5%), while substantially improving the hypocenters.
Introduction
In seismology, phase association is the critical "connect-the-dots" step required to turn raw sensor data into an earthquake catalog. When an earthquake occurs, sensors across a network record the arrival times of various seismic waves, primarily P and S phases. The challenge lies in the sheer volume of data: a large network produces thousands of "picks," many of which are false detections (noise) or originate from multiple overlapping earthquakes occurring nearly simultaneously.
Existing association algorithms often struggle with three main issues: velocity model inaccuracies due to simplified homogeneous earth assumptions, sensitivity to noise picks that are often forced into real earthquake clusters, and the difficulty of detecting small events during intense aftershock sequences where signals are buried under the "moveout" of larger events.
The Neuma Framework
To address these challenges, we introduce Neuma (Neural mixture model association). Neuma treats association as a probabilistic clustering problem solved via the Expectation-Maximization (EM) algorithm.
The core innovations of Neuma include:
- EikoNet Integration: Instead of relying on lookup tables or constant-velocity assumptions, Neuma utilizes EikoNet—a physics-informed neural network (PINN) that solves the Eikonal equation. This allows for rapid, differentiable, and highly accurate travel-time calculations even in complex 3D media.
- Explicit Noise Modeling: Neuma includes a dedicated latent variable for noise. This allows the algorithm to explicitly label a pick as "noise" rather than forcing it into an earthquake cluster, leading to cleaner catalogs.
- Robust Laplace Distributions: We utilize Laplace distributions to model arrival time residuals. Because Laplace distributions have heavier tails than standard Gaussians, the algorithm is significantly more robust against the large outliers common in automated phase picking.
- Warmup Scheme: To prevent the EM algorithm from converging on local minima—where valid but distant picks might be discarded as noise—we implemented a warmup phase that temporarily disables noise filtering until hypocenters stabilize.
Experimental Evaluation
We evaluated Neuma using both synthetic and real-world datasets. The synthetic tests (D1–D4) involved four 24-hour periods with increasing difficulty, simulating 1,080 earthquakes with decreasing inter-event times and over 57,000 "garbage" picks.
Key Results
- Superior Recall: On the most challenging synthetic dataset (D4), Neuma achieved a recall of 0.947, vastly outperforming baselines like PhaseLink (0.228) and GaMMA (0.743).
- Ridgecrest Sequence Performance: When applied to the 2019 Ridgecrest, California sequence (July 4–9), Neuma detected 27,748 earthquakes—a 13.5% increase (3,283 more events) over the best baseline.
- Improved Localization: Neuma significantly refined hypocenter accuracy. Compared to the GaMMA baseline, it reduced systematic depth bias by nearly 5 km per event.
- Selectivity: Despite detecting more events, Neuma associated an average of two fewer picks per event than GaMMA, demonstrating its superior ability to exclude noise and maintain high-confidence associations.
Discussion and Conclusion
Neuma demonstrates that combining physics-informed neural networks with robust mixture modeling provides a powerful solution for seismic phase association. By treating noise as a first-class citizen and leveraging differentiable physics solvers, we can achieve higher sensitivity and accuracy than traditional methods.
While Neuma currently relies on initial data "chunking" via DBSCAN, which requires some hyperparameter tuning, its ability to handle complex velocity structures and high-noise environments marks a significant step forward. Future work will focus on leveraging Neuma’s full 3D velocity handling capabilities to further refine earthquake locations in geologically complex regions.
Summary updated