Accelerating Time-Reversal Imaging with Neural Operators for Real-time Earthquake Locations

Hongyu Sun, Yang Yan, Kamyar Azizzadenesheli, Robert W. Clayton, Zachary E. RossarXiv 2022 · DOI · arXiv

Abstract

Earthquake hypocenters form the basis for a wide array of seismological analyses. Pick-based earthquake location workflows rely on the accuracy of phase pickers and may be biased when dealing with complex earthquake sequences in heterogeneous media. Time-reversal imaging (TRI) of passive seismic sources with the cross-correlation imaging condition has potential for earthquake location with high accuracy and high resolution, but carries a large computational cost. Here we present an alternative deep-learning approach for earthquake location by combining the benefits of neural operators for wave propagation and TRI with multi-station waveform recordings. A U-shaped neural operator (UNO) is trained to propagate seismic waves with various source time functions and thus can predict a backpropagated wavefield for each station in negligible time. These wavefields can either be stacked or correlated to locate earthquakes from the resulting source images. We demonstrate the method with the 2D acoustic wave equation on both synthetic and field data. The results show that our method can efficiently obtain high resolution and high accuracy correlation-based TRI of earthquake sources. Moreover, our approach is adaptable to the number and geometry of seismic stations, which opens new strategies for real-time earthquake location and monitoring with dense seismic networks.

Introduction

Locating the hypocenter of an earthquake is a fundamental task in seismology. Traditional workflows typically rely on "picking" arrival times of P and S waves at various stations. However, these methods can struggle in noisy environments, during complex earthquake sequences where events overlap, or in regions with significant geological heterogeneity that causes complex wave interactions.

Time-Reversal Imaging (TRI) offers a robust alternative. By recording seismic signals, reversing them in time, and numerically "re-broadcasting" them back into the Earth model from the sensor locations, waves naturally refocus at the original source. While TRI is physically accurate and high-resolution, its computational cost is prohibitive for real-time use. Standard TRI requires a full physics simulation for every station in the network. For modern dense networks, this process is too slow for immediate monitoring.

Methods: U-shaped Neural Operators

To overcome the computational bottleneck of TRI, this research introduces a deep-learning framework centered on U-shaped Neural Operators (UNO). Unlike traditional neural networks that map discrete inputs to labels, Neural Operators learn mappings between infinite-dimensional function spaces, making them particularly suited for solving partial differential equations (PDEs) like the wave equation.

The UNO architecture maps a source time function (a) to a backpropagated wavefield (u) through a series of operator layers (G).

The UNO is trained to learn the mapping from a source time function (the recorded signal at a station) to the resulting wavefield within a specific velocity model. Key advantages of this approach include:

  • Mesh-Invariance: The model can be trained on a coarse grid and evaluated on a finer grid for higher resolution without retraining.
  • Cross-Correlation Imaging: Once the UNO predicts backpropagated wavefields for all stations, a cross-correlation imaging condition is applied. This suppresses artifacts and "ghost" images more effectively than simple stacking, resulting in a sharper focus on the earthquake source.
  • Efficiency: The heavy computational load is shifted to the offline training phase. Once trained, the UNO can predict wavefields in milliseconds.

Datasets and Training

The model was developed and validated using both synthetic and real-world data:

  • Synthetic Data: 2,000 random source time functions were generated using Gaussian random fields. Ground truth wavefields were produced using a traditional Spectral Element Method (SEM) solver for a 2D velocity model of Oklahoma.
  • Field Data: The model was tested on real-world recordings from the 2016 IRIS community wavefield experiment in Oklahoma, specifically targeting an M2 earthquake recorded by a dense node network.
Examples of (a) synthetic source time functions (waveforms) and (b) their corresponding frequency spectra used for training.

Key Results

The UNO-based TRI approach demonstrates significant improvements in both speed and accuracy:

  • Massive Speedup: In a test dataset of 45 simulations, the traditional SEM solver required 3 to 9 hours. In contrast, the UNO achieved comparable results in just 25 to 28 seconds.
  • High Accuracy and Resolution: The method successfully resolved two synthetic earthquakes occurring just 1.45 seconds apart—a scenario that often confuses traditional picking algorithms. Correlation-based imaging with 64 stations produced clean source images with minimal artifacts.
  • Field Validation: The model accurately located a real M2 earthquake in Oklahoma, matching results from computationally intensive research methods while remaining adaptable to varying station geometries.

Discussion and Future Work

While the current 2D acoustic implementation shows great promise, several avenues for future research remain. Transitioning to 3D elastic wave equations is necessary for broader real-world application but will require more memory-efficient operators. Additionally, the method's performance is tied to the accuracy of the underlying velocity model and the frequency range of the training data. Future work will focus on expanding the model's generalizability to diverse geological structures and higher frequency bands.

Summary updated

Abstract

Earthquake hypocenters form the basis for a wide array of seismological analyses. Pick-based earthquake location workflows rely on the accuracy of phase pickers and may be biased when dealing with complex earthquake sequences in heterogeneous media. Time-reversal imaging of passive seismic sources with the cross-correlation imaging condition has potential for earthquake location with high accuracy and high resolution, but carries a large computational cost. Here we present an alternative deep-learning approach for earthquake location by combining the benefits of neural operators for wave propagation and time reversal imaging with multi-station waveform recordings. A U-shaped neural operator is trained to propagate seismic waves with various source time functions and thus can predict a backpropagated wavefield for each station in negligible time. These wavefields can either be stacked or correlated to locate earthquakes from the resulting source images. Compared with other waveform-based deep-learning location methods, time reversal imaging accounts for physical laws of wave propagation and is expected to achieve accurate earthquake location. We demonstrate the method with the 2D acoustic wave equation on both synthetic and field data. The results show that our method can efficiently obtain high resolution and high accuracy correlation-based time reversal imaging of earthquake sources. Moreover, our approach is adaptable to the number and geometry of seismic stations, which opens new strategies for real-time earthquake location and monitoring with dense seismic networks.
BibTeX
@misc{sun2022accelerating,
  author = {Hongyu Sun and Yang Yan and Kamyar Azizzadenesheli and Robert W. Clayton and Zachary E. Ross},
  title = {{Accelerating Time-Reversal Imaging with Neural Operators for Real-time Earthquake Locations}},
  year = {2022},
  eprint = {2210.06636},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2210.06636},
}

Loading related content...

Figure