GP-ETAS: semiparametric Bayesian inference for the spatio-temporal epidemic type aftershock sequence model

Christian Molkenthin, Christian Donner, Sebastian Reich, Gert Zöller, Sebastian Hainzl, Matthias Holschneider, Manfred OpperStat. Comput 2022 · DOI

TL;DR

This paper introduces GP-ETAS, a novel Bayesian framework for modeling earthquake aftershock sequences. By leveraging Gaussian Processes for the background seismicity and a sophisticated data augmentation technique, GP-ETAS offers more robust uncertainty quantification and improved predictive power compared to traditional methods.

What Problem Does This Paper Address?

Earthquake aftershock sequences are a classic example of self-exciting point processes, where past events trigger future ones. The Epidemic Type Aftershock Sequence (ETAS) model is widely used to describe this phenomenon. However, traditional ETAS models often rely on point estimates from maximum likelihood estimation (MLE), which can be insufficient for accurately quantifying uncertainties, especially in seismic hazard assessment. Furthermore, estimating the spatially varying background seismicity rate (μ(x)) non-parametrically and performing full Bayesian inference for the spatio-temporal ETAS model presents significant computational challenges due to intractable integrals and non-Gaussian likelihoods. This paper aims to bridge this gap by developing a flexible, fully Bayesian approach that naturally incorporates prior knowledge and provides comprehensive uncertainty estimates.

Methods

The core innovation is the GP-ETAS model, which combines a Gaussian Process (GP) prior for the non-parametric modeling of the spatially varying background seismicity rate (μ(x)) with a classical parametric form for the triggering function (φ). To overcome the computational hurdles of Bayesian inference for this complex model, the authors employ a data augmentation strategy. This involves introducing auxiliary variables:

  • Latent branching structure (Z): This decouples the background and triggering components in the likelihood.
  • Latent Poisson process (Π): This helps to unbiasedly estimate intractable integrals in the likelihood.
  • Pólya-gamma random variables (w): These are used to represent the logistic sigmoid function in a way that leads to conditional conjugacy with the GP prior, enabling efficient sampling.

These augmentations allow for a Gibbs sampling scheme to efficiently explore the posterior distribution of the model parameters and the background intensity.

Datasets

  • Synthetic Data: Two synthetic datasets were generated to evaluate the model's performance under controlled conditions, simulating different spatial distributions of background seismicity (large area sources vs. small fault-type areas).
  • Real Data: The L'Aquila earthquake sequence in Italy (April 2009 onwards) was used for a case study. This dataset comprises N = 2189 events over a spatial domain of [12°E, 15°E] × [41°N, 44°N] with magnitudes 3.0 < m < 6.5. The data was split into training and testing sets.

Key Insights and Ideas

  • GP for Background Seismicity: Using a Gaussian Process for the background seismicity rate allows for flexible, non-parametric modeling without requiring pre-defined functional forms, which is crucial for capturing complex spatial patterns.
  • Data Augmentation for Tractability: The clever use of data augmentation (latent branching structure, latent Poisson process, Pólya-gamma variables) is key to making the full Bayesian inference computationally feasible, transforming a doubly intractable problem into one solvable by Gibbs sampling.
  • Improved Uncertainty Quantification: The Bayesian framework naturally provides posterior distributions for all parameters, offering a richer understanding of uncertainty compared to point estimates from MLE. This is particularly valuable for seismic hazard assessment.
  • Outperformance on Synthetic and Real Data: The GP-ETAS model consistently outperforms traditional ETAS models (ETAS-classical and ETAS-Silverman) in terms of predictive power (test likelihood) and accuracy in estimating the background intensity.

Key Results

  • Synthetic Data: GP-ETAS recovered the underlying model parameters and background intensity more accurately than standard ETAS models, demonstrating superior performance in both test likelihood and L2 norm error for the background intensity.
  • L'Aquila Case Study: GP-ETAS achieved a slightly better test likelihood compared to ETAS-classical and ETAS-Silverman. The model also provided more stable estimates for the background intensity, avoiding the overshooting observed in ETAS-classical in regions with many aftershocks. The uncertainty bands provided by GP-ETAS for parameters like the upper bound of the background intensity (λ) were also more conservative and informative.
  • Computational Cost: While GP-ETAS is computationally more intensive than MLE-based methods (O((NDU)³) complexity due to GP matrix inversions), it remains feasible for datasets of moderate size (ND ≈ 10⁴ events) and offers significant advantages in terms of inference quality.

Limitations and Future Work

  • Computational Complexity: The current implementation of GP-ETAS has a cubic complexity in the number of data points, which can be prohibitive for very large datasets.
  • Future Work: The authors suggest several avenues for future research:
  • Time-dependent background rate: Extending the GP to model time-varying background seismicity.
  • Bayesian non-parametric triggering function: Applying non-parametric methods to the triggering function as well.
  • Model approximations: Developing approximations (e.g., variational inference, sparse GPs, MAP estimation) to scale the model to larger datasets.
  • Geology-informed priors: Incorporating geological information (e.g., fault locations) into the GP prior for more informed modeling.
  • Alternative sampling methods: Investigating gradient-free affine invariant sampling methods.

Summary updated

BibTeX
@article{molkenthin2022gp,
  author = {Christian Molkenthin and Christian Donner and Sebastian Reich and Gert Zöller and Sebastian Hainzl and Matthias Holschneider and Manfred Opper},
  title = {{GP-ETAS: semiparametric Bayesian inference for the spatio-temporal epidemic type aftershock sequence model}},
  journal = {Stat. Comput},
  year = {2022},
  doi = {10.1007/s11222-022-10085-3},
}
Figure