Probabilistic seismic source inversion incorporating uncertainty in 3D earth models: case studies for southeast Australia and the Korean peninsula

Seismic sources of small to moderate magnitudes can be represented by moment tensors (MTs), a 3x3 matrix representing a dynamically equivalent force system, under the point source approximation. The inference of MTs from observed seismic waveforms, known as moment tensor inversion, is strongly subjected to uncertainties in the observation process, known as data noise, and biases in the forward modeling, representing the theoretical noise. It is reasonable to assume that the theoretical noise is dominated by the imperfect knowledge of the Earth's structures, specifically, a 3D Earth model. In this study, we investigate the MT inversion in the Bayesian framework using the affine-invariant ensemble samplers to explore the parameter space effectively. We consider the combined treatment of the two source uncertainties and their influence on the inverted MT solutions. Empirical data noise covariance matrices are constructed for pre-event ambient noise records. The Monte-Carlo method invoking a small perturbation of the 3D Earth model is used to estimate the structural uncertainty covariance matrices. The feasibility and advantages of the newly developed method are showcased through a series of designed synthetic experiments. Our new approach is then applied to gain insights into selected tectonic earthquakes (e.g., the 2021 Mw 5.9 Melbourne earthquake in SE Australia), the events in the Korean peninsula (e.g., the 2017 Mw 5.5 Pohang earthquake, and the DPRK nuclear tests), where high-quality 3D Earth models are available.