Earthquake Forecasting Based on Data Assimilation: Sequential Monte Carlo Methods for Renewal Processes

TL;DR

This paper introduces a novel data assimilation approach using sequential Monte Carlo methods for earthquake forecasting with renewal process models, demonstrating improved prediction accuracy over traditional methods.

Contribution

It presents the first fully implementable data assimilation method for earthquake forecasts based on point-process models, validated through extensive simulations.

Findings

Forecasts using the OSIR particle filter outperform benchmark models.

Data assimilation significantly improves earthquake forecast accuracy.

Model parameters can be effectively estimated using marginal data likelihood.

Abstract

In meteorology, engineering and computer sciences, data assimilation is routinely employed as the optimal way to combine noisy observations with prior model information for obtaining better estimates of a state, and thus better forecasts, than can be achieved by ignoring data uncertainties. Earthquake forecasting, too, suffers from measurement errors and partial model information and may thus gain significantly from data assimilation. We present perhaps the first fully implementable data assimilation method for earthquake forecasts generated by a point-process model of seismicity. We test the method on a synthetic and pedagogical example of a renewal process observed in noise, which is relevant to the seismic gap hypothesis, models of characteristic earthquakes and to recurrence statistics of large quakes inferred from paleoseismic data records. To address the non-Gaussian statistics of earthquakes, we use sequential Monte Carlo methods, a set of flexible simulation-based methods for recursively estimating arbitrary posterior distributions. We perform extensive numerical simulations to demonstrate the feasibility and benefits of forecasting earthquakes based on data assimilation. In particular, we show that the forecasts based on the Optimal Sampling Importance Resampling (OSIR) particle filter are significantly better than those of a benchmark forecast that ignores uncertainties in the observed event times. We use the marginal data likelihood, a measure of the explanatory power of a model in the presence of data errors, to estimate parameters and compare models.