Estimating correlation lengths of megathrust events: A Bayesian model selection approach.

The correlation lengths of seismic ruptures are a necessary input for the generation of stochastic slip models, which are routinely employed in Probabilistic Tsunami Hazard Assessment (PTHA). Typically, the correlation lengths and Hurst exponent of Von Karman auto-correlation function (VK-AFC) have been computed from published coseismic slip models, by finding the best-fitting VK-AFC parameters. Since spatial correlations (e.g., Laplacian smoothing) are commonly enforced in slip models, the results of this methodology can be biased. In our work, we propose a novel methodology that allows characterizing correlations lengths and Hurst exponent which better describe an earthquake, based directly on seismic data, instead of its coseismic slip model. To this end, we compute an approximate expression for the Bayesian evidence for an ensemble of prior models, comprising several combinations of VK-AFC parameters. Applying Bayesian model selection, we can obtain the model which is more consistent with the data, by simply selecting the model which maximizes the evidence. We apply this methodology to both synthetic and measured teleseismic data. Our results illustrate the usefulness of the method and are consistent with published models.