Probabilistic Tsunami-Source Estimation with Parsimonious Rupture Kinematics

Abstract:

This work develops a self-parametrized Bayesian inversion to infer the spatiotemporal evolution of tsunami sources (the initial sea-surface displacement) due to megathrust earthquakes. The approach provides uncertainty estimates about the spatial and temporal evolution of rupture and does not depend on assumptions about the megathrust fault or seafloor deformation. In particular, we study the effect of spatial and temporal parametrization complexity of the source with trans-dimensional parametrizations. The inversion is based on a trans-dimensional self-parametrization of the sea surface in both space and time and accounts for model-selection ambiguity associated with the spatiotemporal discretization. The spatial displacement pattern is parametrized by Voronoi nodes that include parameters for latitude, longitude, displacement magnitude, and rupture velocity. Hence, the sea surface is parametrized by a self-adapting irregular grid which matches the local resolving power of the data and provides parsimonious solutions for complex source characteristics. Rupture onset is causally constrained by solving the Eikonal equation on the rupture velocity field. The source-time function of the rupture at each subfault is parametrized as a parsimonious triangle sequence, such that scaling parameters for triangles are only included when required by the data. Regularization is not required, rather Bayesian sampling is applied to quantify the posterior probability for the spatiotemporal model. Predictions are based on Greens-function libraries computed by the tsunami propagation model JAGURS for cases that include linear dispersion effects. The method is examined with both simulated and observed tsunami-waveform data recorded on high-quality sensors (ocean-bottom pressure sensors, GPS gauges, and DART buoys). The observed waveforms are from the great 2011 Tohoku (Japan) earthquake and appear sensitive to rupture velocity but resolution of the source-time function is limited.