Dynamic earthquake source inversion with GAN priors, with application to the 2023 Mw 7.8 Kahramanmaraş, Turkey earthquake

Dynamic source inversion of earthquakes consists of inferring frictional parameters and initial stress on a fault consistent with co-seismic seismological and geodetic data and dynamic earthquake rupture models. In a Bayesian inversion approach, the nonlinear relationship between model parameters and data (e.g. seismograms) 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 a large earthquake, involving hundreds to thousands parameters, shows problems with convergence and sampling. We introduce a novel multi-stage approach to dynamic inversions. We divide the earthquake rupture into several successive temporal (e.g. 0-10 s, 10-20 s, …) and spatial stages (e.g., 100 km, 200 km, …). As each stage requires only a limited number of independent model parameters, their inversion converges relatively fast. Stages are interdependent: earlier stage inversion results are a prior for a later stage inversion. Our main advancement is the use of Generative Adversarial Networks (GAN) to transfer the prior information between inversion stages, inspired by Patel and Oberai (2019). GAN are a class of machine learning algorithms originally used for generating images similar to the training dataset. Their unsupervised training is based on a contest between a generator that generates new samples and a critic that discriminates between training and generator’s images. The resulting generator should generate synthetic images/samples with noise in a low-dimensional latent space as an input. We train GANs on samples of dynamic parameters from an earlier stage of the inversion and use the GAN to suggest the dynamic parameters in a later stage of inversion. We show a proof of concept dynamic inversion of a synthetic benchmark, comparing performance of direct MC dynamic inversion with parallel tempering with our GAN approach. We efficiently handle large ruptures by adopting a 2.5D approximation that solves for source properties averaged across the rupture depth. The 2.5D modeling approach accounts for the 3D effect of the finite rupture depth while keeping the computational cost the same as in 2D dynamic rupture simulations. Additionally we show current results on the dynamic inversion of 2023 Mw 7.8 Kahramanmaraş, Turkey, earthquake.