A neural encoder for earthquake rate forecasting

Oleg Zlydenko, Gal Elidan, Avinatan Hassidim, Doron Kukliansky, Yossi Matias, Brendan J. Meade, A. Yu. Molchanov, Sella Nevo, Yohai Bar‐SinaiScientific 2023 · DOI

Abstract

Forecasting the timing of earthquakes is a long-standing challenge. Moreover, it is still debated how to formulate this problem in a useful manner, or to compare the predictive power of different models. Here, we develop a versatile neural encoder of earthquake catalogs, and apply it to the fundamental problem of earthquake rate prediction, in the spatio-temporal point process framework. The epidemic type aftershock sequence model (ETAS) effectively learns a small number of parameters to constrain the assumed functional forms for the space and time correlations of earthquake sequences (e.g., Omori-Utsu law). Here we introduce learned spatial and temporal embeddings for point process earthquake forecasting models that capture complex correlation structures. We demonstrate the generality of this neural representation as compared with ETAS model using train-test data splits and how it enables the incorporation additional geophysical information. In rate prediction tasks, the generalized model shows >4% improvement in information gain per earthquake and the simultaneous learning of anisotropic spatial structures analogous to fault traces. The trained network can be also used to perform short-term prediction tasks, showing similar improvement while providing a 1000-fold reduction in run-time.

Introduction

Predicting when and where earthquakes will strike remains one of the most difficult challenges in geophysics. For decades, the Epidemic Type Aftershock Sequence (ETAS) model has been the industry standard, treating earthquakes as a self-exciting process where one event increases the probability of subsequent aftershocks. However, ETAS is limited by rigid mathematical assumptions—such as symmetrical spatial kernels—and struggles to integrate auxiliary data like micro-seismicity below the catalog completeness threshold.

This research introduces FERN (Forecasting Earthquake Rates with Neural networks), a deep learning framework that replaces human-engineered formulas with flexible neural embeddings. By learning complex relationships directly from raw data, FERN improves forecasting accuracy and discovers geological structures without prior instruction.

Methodology

FERN utilizes an encoder-decoder neural network architecture designed for spatio-temporal point processes. Instead of relying on fixed mathematical laws, the system employs three distinct neural encoders to represent the "tectonic state":

  1. Recent Earthquakes Encoder: A neural generalization of the ETAS response function that learns how recent events trigger future ones.
  2. Long-term Seismicity Encoder: Captures historical patterns across varying timescales, from minutes to years.
  3. Location Encoder: Learns a unique geographic "fingerprint" for specific grid cells, replacing the static background rate parameters used in traditional models.
The FERN architecture: Inputs include spatiotemporal information and earthquake catalogs, processed through three specialized encoders to feed into decoders for number and rate prediction.

The model was trained and validated using the Japan Meteorological Agency (JMA) earthquake catalog, specifically focusing on three sub-regions near the Japan subduction zone. The study utilized data from 1979–2003 for training and validation, with testing performed on the 2004–2011 period.

Key Results and Insights

Improved Predictive Performance

FERN demonstrated a significant performance boost over the standard ETAS model. Across the test regions, FERN showed a >4% improvement in information gain per earthquake. In specific areas like "Region C," the FERN+ variant (which includes sub-threshold micro-seismicity) achieved a 94% Area Under the Curve (AUC), significantly outperforming the ETAS model's 91%.

ROC curves for Region C comparing Poisson, ETAS, FERN, and FERN+ models. FERN+ shows the highest predictive power with an AUC of 0.94.

Emergent Geology and Efficiency

One of the most striking findings was the model's ability to learn anisotropy. While traditional models assume seismic influence spreads in simple circles, FERN’s learned embeddings automatically aligned with actual geological fault traces. The model effectively "discovered" the shape of the faults just by analyzing the earthquake catalog.

Furthermore, FERN provides a 1,000-fold reduction in run-time for short-term forecasting. While traditional ETAS forecasting requires tens of thousands of stochastic simulations, FERN generates predictions in a single forward pass of the network.

Discussion and Limitations

Despite its success, the authors note that FERN's performance can vary by region. In "Region A," the improvement over ETAS was marginal, likely due to shifts in seismic statistics between the training and testing periods. Additionally, the current model does not yet provide explicit uncertainty estimates, which are critical for real-world hazard assessment. Future work will focus on integrating richer geophysical data, such as tectonic stress and focal mechanisms, to further refine forecasting accuracy.

Summary updated

Abstract

Forecasting the timing of earthquakes is a long-standing challenge. Moreover, it is still debated how to formulate this problem in a useful manner, or to compare the predictive power of different models. Here, we develop a versatile neural encoder of earthquake catalogs, and apply it to the fundamental problem of earthquake rate prediction, in the spatio-temporal point process framework. The epidemic type aftershock sequence model (ETAS) effectively learns a small number of parameters to constrain the assumed functional forms for the space and time correlations of earthquake sequences (e.g., Omori-Utsu law). Here we introduce learned spatial and temporal embeddings for point process earthquake forecasting models that capture complex correlation structures. We demonstrate the generality of this neural representation as compared with ETAS model using train-test data splits and how it enables the incorporation additional geophysical information. In rate prediction tasks, the generalized model shows [Formula: see text] improvement in information gain per earthquake and the simultaneous learning of anisotropic spatial structures analogous to fault traces. The trained network can be also used to perform short-term prediction tasks, showing similar improvement while providing a 1000-fold reduction in run-time.
BibTeX
@inproceedings{zlydenko2023neural,
  author = {Oleg Zlydenko and Gal Elidan and Avinatan Hassidim and Doron Kukliansky and Yossi Matias and Brendan J. Meade and A. Yu. Molchanov and Sella Nevo and Yohai Bar‐Sinai},
  title = {{A neural encoder for earthquake rate forecasting}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-38033-9},
}
Figure