Kinematic Inversion of Physically Plausible Earthquake Source Models Obtained from Dynamic Rupture Simulations

One approach to investigate earthquake source processes is to produce kinematic source models from inversion of seismic records and geodetic data. The setup of the inversion requires a variety of assumptions and constraints to restrict the range of possible models. Here, we evaluate to what extent physically plausible earthquake scenarios are reliably restituted in spite of these restrictions. We study which characteristics of ruptures, such as rupture velocity, slip distribution, stress drop, rise time, and slip function, can be reliably determined from the inversion of near-field seismic and geodetic data. Using spontaneous dynamic rupture simulations, we generate five earthquake scenarios, each of which has different characteristics of the source process. Then we conduct a blind test by modeling the synthetic near-source data using a standard inversion scheme that optimizes the fit to the observations while searching for solutions with minimum roughness. The inversion procedure assumes a rupture front propagating away from the hypocenter with variable rupture velocity and a simple cosine slip-time function. Our results show that, overall, slip distribution and stress drop are reasonably well determined even for input models with relatively complex histories (such as a subshear rupture transitioning to supershear speeds). Depth-averaged rupture velocities are also reasonably well resolved although their estimate progressively deteriorates away from the hypocenter. The local rise time and slip function are not well resolved, but there is some sensitivity to the rupture pulse width, which can be used to differentiate between pulse-like and crack-like ruptures. Our test for understanding the inaccuracies in Green's functions shows that random 3D perturbations of 5% standard deviation do not lead to significant degradation of the estimation of earthquake source parameters. As remedies to the current limitations, we propose smoothing slip function parameters and using more complicated inversion schemes only if data necessitates them.