Bayesian Posterior Exploration and Predictive Analysis of Finite-Fault Earthquake Models

Abstract

Although earthquake rupture process is inherently complex, recent inversion techniques enable data-constrained physical models that image the spatiotemporal evolution of fault slip. Among these methods, Bayesian inversion framework yields an ensemble of plausible models to reproduce data, in contrast with conventional methods that produce a single best-fit solution. These posterior model ensembles are valuable for exploring relationships among physical parameters, generating downstream physical observables, and quantifying interlinked uncertainty structures. However, large ensembles (from thousands to millions) of high-dimensional models pose computational challenges. Here, we analyze inferred rupture properties and predictive observables of five Bayesian kinematic finite-fault models for large megathrust earthquakes in Japan, Chile, Ecuador and Nepal: 2011 Mw9.0 Tohoku, 2014 Mw8.1 Iquique, 2015 Mw8.3 Illapel, 2016 Mw7.8 Pedernales, and 2015 Mw7.8 Gorkha. We compare joint probability distributions of kinematic rupture parameters, including slip, rupture speed, rise time, and average slip rate, and assess relationships locally and fault-wise across events. We use posterior model ensembles to compute static surface deformation and on-fault stress changes, dynamic ground motion, and their spatiotemporal covariances. We find patch-level correlations are robust between some parameters given expected tradeoffs, but intra-event correlations between most parameters are variable across events. This result conflicts with predicted relations, e.g., between peak slip rate and rupture speed, established by dynamic rupture simulations, suggesting limitations in adopted physics and/or inversion approaches for heterogeneous rupture scenarios. Our static forward simulations reveal characteristic patterns in the predicted deformation and stress changes across the explored megathrust settings. The predicted surface deformation exhibits spatial variations in amplitude and direction largely inherited from source properties, whereas on-fault static stress is more heterogeneous in space. The variability in uncertainty is influenced by station coverage, source characteristics, and the data-model error structure. Despite source heterogeneity, the event-average static elastic strain drop is tightly constrained within ~100–300 microstrain, while static stress drop spans ~5–25 MPa. In the dynamic simulations, displacement waveforms stabilize earlier and faster in smaller earthquakes. Predicted peak ground displacements (PGD) exhibit near-field discrepancies with empirical scaling laws due to finite-source and directivity effects. Our findings can inform physics-based earthquake modeling and improve quantification of rupture complexity and hazard impacts.