SPIDER: Scalable Probabilistic Inference for Differential Earthquake Relocation

Zachary E. Ross, Jon Wilding, Kamyar Azizzadenesheli, Akiko KatoarXiv 2025 · DOI · arXiv

Abstract

Seismicity catalogs are larger than ever due to an explosion of techniques for enhanced earthquake detection and an abundance of high-quality datasets. Bayesian inference is an appealing framework for locating earthquakes due to its ability to propagate and quantify uncertainty into the inversion results, but traditional methods do not scale well to high-dimensional parameter spaces, making them unsuitable for double-difference relocation where the number of parameters can reach the millions. Here we introduce SPIDER, a scalable Bayesian inference framework for double-difference hypocenter relocation. SPIDER uses a physics-informed neural network Eikonal solver together with a highly efficient sampler called Stochastic Gradient Langevin Dynamics to generate posterior samples jointly for entire seismicity catalogs. We show that traditional double-difference relocation formulations neglect residual correlation between observations with common events, which biases uncertainty estimates. Our formulation is designed to whiten this residual correlation, and is readily parallelized over multiple GPUs for enhanced computational efficiency. We demonstrate the capabilities of SPIDER on a rigorous synthetic seismicity catalog and three real data catalogs from California and Japan. We introduce several ways to analyze high-dimensional posterior distributions to aid in scientific interpretation and evaluation.

Introduction

Determining the precise location of an earthquake—its hypocenter—is fundamental to understanding fault structures and seismic hazards. As seismic monitoring networks expand and detection algorithms improve, the volume of data has grown exponentially. While Bayesian frameworks offer a robust way to quantify the uncertainty of these locations, they have historically struggled with the high-dimensional parameter spaces required for large-scale catalogs.

SPIDER (Scalable Probabilistic Inference for Differential Earthquake Relocation) addresses this gap. By integrating physics-informed neural networks (PINNs) with advanced stochastic sampling techniques, SPIDER enables rigorous Bayesian relocation for catalogs containing hundreds of thousands of events, providing well-calibrated uncertainty estimates that were previously computationally out of reach.

The Challenges of Double-Difference Relocation

The current gold standard for high-precision relocation is Double-Difference (DD) relocation. This method utilizes the difference in travel times between pairs of earthquakes recorded at the same station to cancel out errors caused by unknown subsurface velocity structures. However, two primary obstacles have limited its Bayesian application:

  1. Computational Scalability: Modern catalogs involve millions of parameters. Traditional Markov Chain Monte Carlo (MCMC) methods, which sample the probability distribution of these parameters, become prohibitively slow at this scale.
  2. Uncertainty Overconfidence: Standard DD formulations often assume that errors between earthquake pairs are independent. In reality, when multiple pairs share a common earthquake, their errors are correlated. Ignoring this "shared-event residual correlation" leads to uncertainty estimates that are significantly narrower than they should be, resulting in false confidence in hypocenter precision.

The SPIDER Framework

SPIDER overcomes these hurdles through a combination of three core innovations:

Stochastic Gradient Langevin Dynamics (SGLD)

To achieve scalability, SPIDER employs SGLD, an MCMC variant that operates on mini-batches of data. Similar to deep learning optimization, SGLD uses gradients to guide the model toward high-probability regions. By injecting controlled noise into the process, it allows the system to sample the full posterior distribution rather than just finding a single point estimate. This allows the algorithm to naturally transition from an initial optimization phase to a comprehensive sampling phase.

Differentiable Forward Modeling with EikoNet

Gradient-based sampling requires a differentiable method to calculate travel times. SPIDER utilizes EikoNet, a physics-informed neural network that solves the Eikonal equation. EikoNet replaces traditional, static lookup tables with a fast, differentiable function, allowing for the seamless backpropagation of gradients required by SGLD.

Residual Whitening

To address the issue of overconfidence, the authors introduced a new likelihood formulation that explicitly models and "whitens" the correlation between observations sharing common events. This mathematical correction ensures that the resulting uncertainty bounds are geophysically accurate and statistically well-calibrated.

Experimental Validation

The framework was rigorously tested across four distinct datasets to evaluate its precision and reliability:

  • Synthetic Ridgecrest: Using a benchmark of 1,000 earthquakes with known ground-truth locations, the authors confirmed that SPIDER’s posterior was "well-calibrated." The true locations fell within the predicted 95% confidence interval approximately 95% of the time, whereas traditional methods were 20–30% too narrow.
  • Cahuilla Swarm (California): In a compact cluster of ~22,700 events, SPIDER was used to study fine-scale fault structures, identifying a fault thickness of approximately 75 meters at depth.
  • Ridgecrest Sequence (California): This massive real-world test involved ~326,000 events and 30 million differential time observations. SPIDER successfully processed the entire catalog jointly on GPUs, a task that usually requires segmenting data into smaller, disjoint chunks.
  • Noto Sequence (Japan): Testing ~72,700 events in a region with sparse station coverage, SPIDER demonstrated its robustness in complex, swarm-like sequences.

Scientific Interpretation of "Fuzzy" Seismicity

One of the primary advantages of a Bayesian approach is the ability to visualize uncertainty. Rather than representing earthquakes as static points, SPIDER allows researchers to use Kernel Density Estimation (KDE) on posterior samples. This produces a probabilistic map of seismicity, helping researchers distinguish between real structural features (like the width of a fault zone) and blurring caused by location uncertainty. Furthermore, SPIDER showed a superior ability to maintain the continuity of seismic features, avoiding the artificial "shredding" of fault lines into disjoint clusters often seen in baseline methods like GrowClust.

Conclusion and Future Outlook

SPIDER represents a significant step forward in seismic data analysis, providing a scalable, GPU-accelerated path to rigorous uncertainty quantification. While the current implementation relies on differentiable velocity models and is optimized for cross-correlation data, it sets the stage for more reliable probabilistic seismic catalogs. Future work will focus on extending the "whitening" logic to raw phase picks and further automating hyperparameter tuning to make the framework more accessible to the broader seismological community.

Summary updated

Abstract

Seismicity catalogs are larger than ever due to an explosion of techniques for enhanced earthquake detection and an abundance of high-quality datasets. Bayesian inference is an appealing framework for locating earthquakes due to its ability to propagate and quantify uncertainty into the inversion results, but traditional methods do not scale well to high-dimensional parameter spaces, making them unsuitable for double-difference relocation where the number of parameters can reach the millions. Here we introduce SPIDER, a scalable Bayesian inference framework for double-difference hypocenter relocation. SPIDER uses a physics-informed neural network Eikonal solver together with a highly efficient sampler called Stochastic Gradient Langevin Dynamics to generate posterior samples jointly for entire seismicity catalogs. We show that traditional double-difference relocation formulations neglect residual correlation between observations with common events, which biases uncertainty estimates. Our formulation is designed to whiten this residual correlation, and is readily parallelized over multiple GPUs for enhanced computational efficiency. We demonstrate the capabilities of SPIDER on a rigorous synthetic seismicity catalog and three real data catalogs from California and Japan. We introduce several ways to analyze high-dimensional posterior distributions to aid in scientific interpretation and evaluation.
BibTeX
@misc{ross2025spider,
  author = {Zachary E. Ross and Jon Wilding and Kamyar Azizzadenesheli and Akiko Kato},
  title = {{SPIDER: Scalable Probabilistic Inference for Differential Earthquake Relocation}},
  year = {2025},
  eprint = {2508.12117},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.12117},
  howpublished = {arXiv},
}
Figure