Microseismic event detection in large heterogeneous velocity models using Bayesian multimodal nested sampling

Saptarshi Das, M. P. Hobson, Farhan Feroz, Xi Chen, Suhas Phadke, Jeroen Goudswaard, Detlef HohlData-Centric 2021 · DOI

Abstract

In passive seismic and microseismic monitoring, identifying and characterizing events within a high-noise background is a significant challenge. Traditional geophysical inversion methods often yield high false-detection rates and require thousands of computationally expensive forward elastic-wave propagation simulations. This paper proposes a novel framework that utilizes an ensemble of Gaussian process surrogate meta-models to accelerate the generation of template seismograms. By employing the Multimodal Nested Sampling (MultiNest) algorithm, the method navigates complex, multimodal likelihood landscapes to perform 5D spatio-temporal-amplitude inference. Crucially, the approach calculates Bayesian evidence to permit rigorous hypothesis testing, effectively discriminating between true seismic events and environmental noise.

Introduction: The Challenge of Microseismic Monitoring

Monitoring "microseisms"—small-magnitude seismic events often induced by human activity—is vital for the safety and efficiency of operations like carbon capture and storage or hydraulic fracturing. However, these signals are frequently buried under significant background noise, making detection difficult.

Existing methods face two primary hurdles. First, the computational cost of "forward modeling" (simulating wave travel through complex 3D geology) is prohibitive when performed thousands of times during optimization. Second, the mathematical landscape of multi-event detection is "multimodal," meaning it contains many local maxima. Standard algorithms often fail by getting stuck on incorrect peaks or mistaking loud noise for a real signal.

Methodology: A Three-Stage Bayesian Pipeline

To address these challenges, the researchers developed a robust pipeline combining machine learning with advanced Bayesian sampling.

Figure 1: Flowchart of the proposed methodology, illustrating the integration of 3D velocity models, Gaussian process surrogate modeling, and MultiNest sampling for likelihood evaluation.

1. Surrogate Modeling (The Emulator)

Instead of running heavy physics-based simulations for every iteration, the authors trained an ensemble of Gaussian Process (GP) regression models. This "proxy" model learns the relationship between event locations and receiver responses, allowing for the near-instantaneous generation of template seismograms. This significantly reduces the computational overhead of the inference process.

2. MultiNest (The Search Engine)

The framework employs the Multimodal Nested Sampling (MultiNest) algorithm. Unlike standard search methods that seek a single global maximum, MultiNest is designed to map entire complex landscapes, including multiple peaks and curving degeneracies. It calculates the "Bayesian Evidence" (marginal likelihood), providing a mathematical basis to compare the hypothesis that an event occurred against the null hypothesis of pure noise.

3. DBSCAN (The Filter)

Following the sampling phase, the researchers use Density-Based Spatial Clustering of Applications with Noise (DBSCAN) to group the posterior samples. This step is critical for separating distinct, simultaneous seismic events and filtering out random noise samples that do not form coherent spatial clusters.

Experimental Setup and Datasets

The study validated the method using a highly realistic synthetic dataset based on a large 3D marine heterogeneous velocity model consisting of millions of voxels. The GP emulators were trained on 4,000 independent GPU-based wave propagation simulations.

Figure 2: Synthetic seismic traces showing three superimposed events buried in Gaussian noise with a negative Signal-to-Noise Ratio (SNR = -1.6455 dB).

The test scenario involved three superimposed microseismic events with varying amplitudes and origin times. These were buried in Gaussian noise with a negative Signal-to-Noise Ratio (SNR = -1.6455 dB), a condition where the noise is significantly stronger than the signal itself.

Key Results and Insights

  • Successful Multi-Detection: The system successfully identified all three overlapping events despite the high-noise environment.
  • Hypothesis Testing via Evidence: By calculating Bayesian Evidence, the authors demonstrated that the hypothesis of "at least one event" was statistically favored over "no event" by a massive margin (a posterior odds ratio of approximately 4168). This provides a rigorous method for rejecting false positives.
  • Uncertainty Quantification: The method provides a full posterior distribution rather than a single coordinate. This revealed that deeper events naturally carry higher uncertainty regarding their origin time, as the signal travel time to the surface is longer.
  • Resolution Sensitivity: The study found that increasing the number of "live points" in the MultiNest sampler (from 300 to 500) was essential for reliably capturing the weakest of the three events without it being lost to spurious clusters.

Conclusion and Future Work

This Bayesian framework offers a powerful tool for microseismic monitoring in complex environments. By replacing expensive simulations with GP emulators and utilizing MultiNest for global inference, the method achieves both speed and statistical rigor. Future work will focus on incorporating the velocity model as a variable to be solved (tomography) and scaling the approach to manage long-duration recordings with larger receiver arrays using sliding time windows.

Summary updated

Abstract

Abstract In passive seismic and microseismic monitoring, identifying and characterizing events in a strong noisy background is a challenging task. Most of the established methods for geophysical inversion are likely to yield many false event detections. The most advanced of these schemes require thousands of computationally demanding forward elastic-wave propagation simulations. Here we train and use an ensemble of Gaussian process surrogate meta-models, or proxy emulators, to accelerate the generation of accurate template seismograms from random microseismic event locations. In the presence of multiple microseismic events occurring at different spatial locations with arbitrary amplitude and origin time, and in the presence of noise, an inference algorithm needs to navigate an objective function or likelihood landscape of highly complex shape, perhaps with multiple modes and narrow curving degeneracies. This is a challenging computational task even for state-of-the-art Bayesian sampling algorithms. In this paper, we propose a novel method for detecting multiple microseismic events in a strong noise background using Bayesian inference, in particular, the Multimodal Nested Sampling (MultiNest) algorithm. The method not only provides the posterior samples for the 5D spatio-temporal-amplitude inference for the real microseismic events, by inverting the seismic traces in multiple surface receivers, but also computes the Bayesian evidence or the marginal likelihood that permits hypothesis testing for discriminating true vs. false event detection.
BibTeX
@inproceedings{das2021microseismic,
  author = {Saptarshi Das and M. P. Hobson and Farhan Feroz and Xi Chen and Suhas Phadke and Jeroen Goudswaard and Detlef Hohl},
  title = {{Microseismic event detection in large heterogeneous velocity models using Bayesian multimodal nested sampling}},
  booktitle = {Data-Centric},
  year = {2021},
  doi = {10.1017/dce.2021.1},
}
Figure