Inversion of Fault Geometric Parameters Based on Mixture Density Networks: A Case Study of the 2013 Ms7.0 Lushan Earthquake in China

Abstract

Mixture density networks (MDNs) are a class of neural network algorithms in the Bayesian framework that can invert for parameters and estimate the parameter uncertainties at the same time. We test MDNs by inverting the fault geometric parameters in a synthetic test with Okada’s model and demonstrate the efficiency and robustness of the model. To further describe the MDNs algorithm, we invert regional static GPS displacement data for the 2013 Ms7.0 Lushan Earthquake in China to obtain the fault geometric parameters. The inverted geometric parameters, which are in the form of distributions, agree well with the published solutions for this event once the 99.7% confidence interval are taken into account. The 99.7% confidence intervals of the strike and dip angles are [ $$187.5^\circ$$ , $$220.4^\circ$$ ] and [ $$20.76^\circ$$ , $$49.76^\circ$$ ], respectively. To demonstrate the applicability of the inverted geometric parameters, we invert the slip distribution of the Lushan earthquake with the inverted parameters and GPS displacements. From the slip distributions considering the distributions of the parameters, we obtain the 99.7% confidence intervals of the magnitude Mw, maximum slip and depth of maximum slip, which are [6.51, 6.65], [0.30, 0.74] m and [9.43, 13.61] km, respectively.