Dynamic earthquake source inversion with Generative Adversarial Network priors

Dynamic source inversion of earthquakes consists of inferring frictional parameters and initial stress on a fault consistent with recorded seismological and geodetic data and with dynamic earthquake rupture models. In a Bayesian inversion approach, the nonlinear relationship between model parameters and data requires a computationally demanding Monte Carlo (MC) approach. As the computational cost of the MC method grows exponentially with the number of parameters, dynamic inversion of large earthquakes, involving hundreds to thousands parameters, is hindered by slow convergence and sampling issues. We introduce a novel multi-stage approach for dynamic source inversion. We divide the earthquake source into a hierarchical set of temporal and spatial stages. As each stage involves only a limited number of independent model parameters, their inversion converges faster. Stages are interdependent: the inversion results of an earlier stage are a prior for the next stage inversion. We use Wasserstein Generative Adversarial Networks to transfer the prior information between inversion stages. As proof-of-concept, we apply a two-stage version of our dynamic source inversion approach to a simulated earthquake scenario generated by dynamic rupture modeling. Compared to direct MC inversion, the two-stage approach achieves substantial improvements in relevant performance metrics, including integrated autocorrelation time, and a large increase in stability across several independent runs. Further application of the two-stage Bayesian inversion method will allow for expanded dynamic modeling studies of large earthquakes, paving the way towards a better understanding of earthquake physics.