A neural encoder for earthquake rate forecasting.

Area of Science:

• Geophysics
• Computational Seismology
• Machine Learning

Background:

• Earthquake timing prediction is a significant challenge in seismology.
• Comparing predictive models for earthquake forecasting remains difficult.
• Existing models like the Epidemic Type Aftershock Sequence (ETAS) model use limited parameters for spatio-temporal correlations.

Purpose of the Study:

• To develop a versatile neural encoder for earthquake catalogs.
• To apply this encoder to earthquake rate prediction within a spatio-temporal point process framework.
• To introduce learned spatial and temporal embeddings for enhanced forecasting models.

Main Methods:

• Developed a neural encoder for earthquake catalog data.
• Applied the encoder to spatio-temporal point process earthquake forecasting.
• Introduced learned embeddings to capture complex correlation structures.
• Incorporated additional geophysical information into the model.

Main Results:

• The generalized neural model demonstrated a [Formula: see text] improvement in information gain per earthquake compared to ETAS.
• The model simultaneously learned anisotropic spatial structures, similar to fault traces.
• Short-term prediction tasks showed similar accuracy improvements.
• Achieved a 1000-fold reduction in computational run-time for the trained network.

Conclusions:

• The neural encoder offers a more general and powerful approach to earthquake forecasting than traditional methods.
• Learned embeddings effectively capture complex spatio-temporal earthquake dynamics.
• The model significantly improves prediction efficiency and accuracy, enabling better short-term forecasting and analysis of seismic patterns.