Do earthquakes "know" how big they will be? a neural-net aided study

Neri Berman, Oleg Zlydenko, Oren Gilon, Yossi Matias, Yohai Bar‐SinaiarXiv 2024 · DOI · arXiv

TL;DR

For decades, seismologists have operated under the assumption that the magnitude of an earthquake is essentially a random draw from a fixed distribution, independent of seismic history. This paper introduces MAGNET, a neural network that challenges this "separability assumption" by successfully extracting predictive signals from past earthquake catalogs to forecast the magnitude of future events, significantly outperforming standard statistical benchmarks across three major global seismic regions.

What Problem Does This Paper Address?

Predicting the size of an upcoming earthquake is one of the "holy grails" of geophysics. Currently, most forecasting models rely on the Gutenberg-Richter (GR) law, which states that earthquake magnitudes follow a predictable power-law distribution over large areas and long times. However, these models typically assume separability: the idea that the magnitude of a specific event is statistically independent of when or where it happens. In this view, earthquakes don't "know" how big they will be until the rupture is already underway.

The authors tackle a fundamental debate: Is earthquake magnitude truly stochastic and independent of history, or are there detectable "preparatory phases" in the seismic record that hint at the scale of an impending rupture?

Methods

The authors developed MAGNET (MAGnitude Neural EsTimation), a geophysically-inspired deep learning architecture.

  • The Input: The model takes a "hypocentral catalog"—a list of past earthquake locations, times, and magnitudes.
  • The Architecture: It uses a dual-encoder system. One LSTM (Long Short-Term Memory) network encodes the sequence of recent earthquakes, while another processes the local seismicity rate (energy release).
  • The Query: Unlike models that try to predict if an earthquake will happen, MAGNET is given a "query"—the specific time and location of a future event—and asked to predict its magnitude.
  • The Output: Instead of a single number, MAGNET outputs a Probability Density Function (PDF). Specifically, it uses a mixture of two "stretched Kumaraswamy distributions." This is a key innovation because this mathematical family can smoothly interpolate between a standard power-law (representing uncertainty) and a highly localized "confident" prediction.

Datasets

The model was evaluated on three high-quality, distinct earthquake catalogs to ensure the findings weren't region-specific:

  1. Southern California: The Hauksson Catalog.
  2. New Zealand: The GeoNet Catalog.
  3. Japan: The JMA (Japan Meteorological Agency) Catalog.

These regions offer diverse tectonic settings and varying levels of data completeness, providing a robust testing ground for the model's generalizability.

Key Insights and Ideas

  • Breaking Separability: The most significant "aha moment" is the demonstration that seismic history does contain information about future magnitudes. By outperforming the GR benchmark, MAGNET proves that the "separability assumption" is an oversimplification of crustal physics.
  • Accounting for Artifacts: The authors meticulously addressed potential "spurious" gains. They tested if the model was simply exploiting Short-Time Incompleteness (the fact that small quakes are often missed in the chaotic aftermath of a large one) or Spatial Heterogeneity (local variations in fault properties). Even after conditioning the results to remove these effects, MAGNET maintained a significant advantage.
  • The Preparatory Phase: The results suggest that the state of a fault system evolves in a way that influences the eventual size of a rupture, supporting the theory of a detectable "preparatory phase" prior to major events.

Key Results

  • Likelihood Gains: MAGNET consistently outperformed stationary and moving-window GR benchmarks. For instance, it improved the likelihood of estimating the true magnitude of the Ridgecrest mainshock (California) by a factor of 13 and the Kaikōura event (New Zealand) by a factor of 4 compared to naive predictors.
  • Information Gain: The model showed an average information gain of >0.1 nats per earthquake. While this sounds small, in the context of earthquake forecasting—where gains are notoriously hard to find—this is a substantial margin.
  • Qualitative Shifts: For certain events, MAGNET’s predicted PDF differed qualitatively from the GR benchmark, showing a "confident" peak around higher magnitudes rather than a simple decaying power law.

Limitations and Future Work

  • The "Known Query" Constraint: Currently, MAGNET requires the time and location of the event to be provided. It does not yet solve the problem of predicting when or where an earthquake will nucleate.
  • Black Box Nature: While the model performs well, the specific physical features it identifies in the seismic history (e.g., specific patterns of small quakes) are not yet fully transparent. The authors suggest future work using "explainability" analyses to bridge this gap.
  • Integration: A promising next step is integrating MAGNET’s history-dependent magnitude PDF into existing rate-forecasting models like ETAS (Epidemic-Type Aftershock Sequence) to create more accurate, holistic hazard assessment tools.

Summary updated

Abstract

Earthquake occurrence is notoriously difficult to predict. While some aspects of their spatiotemporal statistics can be relatively well captured by point-process models, very little is known regarding the magnitude of future events, and it is deeply debated whether it is possible to predict the magnitude of an earthquake before it starts. This is due both to the lack of information about fault conditions and to the inherent complexity of rupture dynamics. Consequently, even state of the art forecasting models typically assume no knowledge about the magnitude of future events besides the time-independent Gutenberg Richter (GR) distribution, which describes the marginal distribution over large regions and long times. This approach implicitly assumes that earthquake magnitudes are independent of previous seismicity and are identically distributed. In this work we challenge this view by showing that information about the magnitude of an upcoming earthquake can be directly extracted from the seismic history. We present MAGNET - MAGnitude Neural EsTimation model, an open-source, geophysically-inspired neural-network model for probabilistic forecasting of future magnitudes from cataloged properties: hypocenter locations, occurrence times and magnitudes of past earthquakes. Our history-dependent model outperforms stationary and quasi-stationary state of the art GR-based benchmarks, in real catalogs in Southern California, Japan and New-Zealand. This demonstrates that earthquake catalogs contain information about the magnitude of future earthquakes, prior to their occurrence. We conclude by proposing methods to apply the model in characterization of the preparatory phase of earthquakes, and in operational hazard alert and earthquake forecasting systems.
BibTeX
@misc{berman2024do,
  author = {Neri Berman and Oleg Zlydenko and Oren Gilon and Yossi Matias and Yohai Bar‐Sinai},
  title = {{Do earthquakes "know" how big they will be? a neural-net aided study}},
  year = {2024},
  eprint = {2408.02129},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2408.02129},
}
Figure