Abstract
Phase association—the process of grouping seismic wave arrivals by their originating earthquake—is a fundamental task in seismic data processing. It is particularly challenging for high-rate microseismicity, where signals overlap and velocity models are often inaccurate. We present Harpa, a high-rate seismic phase association method that leverages travel-time neural fields and optimal transport. By interpreting arrival time data as probability measures and using an optimal transport loss, Harpa avoids the combinatorial explosion of traditional matching. Numerical experiments show that Harpa efficiently associates high-rate seismicity clouds over complex, unknown wave speeds and maintains high accuracy where traditional methods fail.
Introduction: The "Seismic Scramble"
Modern deep learning "pickers" have revolutionized seismology by detecting thousands of previously invisible microearthquakes. However, transforming these individual "picks" (wave arrivals) into a coherent earthquake catalog requires phase association. As the frequency of detected events increases, the signals from different earthquakes begin to overlap and "scramble," creating a complex combinatorial problem. Traditional algorithms often rely on rigid grid searches and simplistic 1D velocity models, which break down under heavy "traffic" or complex 3D geology. Harpa addresses these challenges by reimagining association as a continuous optimization problem.
Methods: Neural Fields and Optimal Transport
Harpa departs from traditional "one-by-one" matching through several key innovations:
- Optimal Transport (Wasserstein Distance): Instead of matching individual picks, Harpa treats arrival sequences at each station as probability distributions. It uses the Wasserstein-2 distance to minimize the discrepancy between observed arrivals and those predicted by a model, making it robust to scrambled arrival orders.
- Travel Time Neural Fields: Harpa utilizes Implicit Neural Representations (INRs)—specifically SIREN-based networks—to create a continuous, differentiable map of travel times across a wide family of possible wave speeds.
- Generative Latent Space: To handle unknown geology, Harpa searches a low-dimensional latent space of a pre-trained autoencoder to find the wave-speed model that best fits the data during the association process.
- Stochastic Gradient Langevin Dynamics (SGLD): To navigate the non-convex landscape of earthquake source localization, Harpa employs SGLD, adding controlled noise to help the algorithm find the global optimum.
Results: Performance at High Rates
The effectiveness of Harpa was evaluated against state-of-the-art (SOTA) tools like GaMMA, PyOcto, and REAL using synthetic benchmarks (SEG/EAGE overthrust model) and real-world data from the 2014 Iquique (Chile) and 2019 Ridgecrest (California) sequences.
As shown in the figure above, while all methods perform well at low event rates (~0.3 events per minute), the performance of traditional associators degrades sharply as the rate increases. At 30 events per minute (epm)—a rate characteristic of intense aftershock sequences—Harpa maintains near 100% accuracy in both pick assignment and event recovery. Furthermore, Harpa's computational runtime is relatively insensitive to event frequency, and its "unknown wave speed" mode has successfully identified deep geological heterogeneities in real-world datasets.
Conclusion and Future Work
Harpa enables the analysis of high-rate microseismicity that was previously inaccessible by combining physics-informed neural fields with the robustness of optimal transport. Future work will focus on incorporating a wider diversity of seismic phases and improving noise tolerance. Ultimately, Harpa paves the way for real-time, high-resolution imaging of the Earth's interior using the very microearthquakes it associates.
Summary updated