A Bayesian Framework for Global Seismic Hazard Model Uncertainty

Thesis Abstract

This study addresses the persistent mismatch between global seismic hazard assessments and observed ground-shaking outcomes by developing a Bayesian framework that explicitly quantifies and propagates model uncertainty across heterogeneous seismotectonic regions. The aim is to synthesize regional and global seismicity information into a coherent probabilistic framework that integrates epistemic and aleatoric uncertainties in hazard curves, and to evaluate how this framework modifies hazard estimates under varying data availability and tectonic regimes. Specific objectives include (1) formulating a hierarchical Bayesian model that couples finite-fault and stochastic renewal process representations with regional seismogenic source models; (2) implementing priors informed by physics-based crustal deformation theories and by empirical global catalogs (e.g., ISC-GEM, USGS PAGER) to constrain regional rupture probabilities; (3) developing a data assimilation scheme to update hazard estimates as new events are recorded, and (4) performing a comprehensive validation against historical strong-motion records and near-real-time ground-motion prediction scenarios. The methodological approach adopts a mixed-methods research design anchored in quantitative probabilistic modeling and synthetic-perturbation analysis. The population comprises seismic source zones and tectonically active regions across the globe, with a stratified sample that includes high, moderate, and low historical activity. Data collection instruments comprise global earthquake catalogs spanning 1900–2024, instrumentally recorded ground motions from accelerometric networks, historical macroseismic intensity datasets, and geodetic strain-rate measurements. Model specification proceeds within a hierarchical Bayesian framework, integrating Poisson renewal process components for rupture occurrence, a lognormal distribution for ground-motion prediction residuals, and a Gaussian process prior over spatially varying seismicity parameters. The analysis employs Markov Chain Monte Carlo methods, including Hamiltonian Monte Carlo for efficient posterior sampling, and a Bayesian model selection protocol using Bayes factors to compare alternative source-zone representations. Validity and reliability are addressed through cross-validation with withheld catalogs, posterior predictive checks, and sensitivity analyses to prior choices. The study further introduces a modular uncertainty-quantification pipeline enabling real-time updating as new seismic data become available. Expected findings include (i) quantification of epistemic uncertainty reductions achieved by incorporating physics-informed priors and multi-resolution source representations, (ii) identification of region-specific contributions to global hazard uncertainty, particularly in intellectually underserved plates-and-blocks settings, (iii) demonstration that posterior hazard curves exhibit reduced misfit to observed strong-motion intensities relative to traditional probabilistic seismic hazard analysis (PSHA) approaches, and (iv) guidance on the conditions under which the Bayesian framework yields robust hazard estimates with limited data. The research anticipates discovering that integrating geophysical priors and data streams yields tighter credible intervals for hazard metrics while preserving sensitivity to extreme events, thereby enhancing risk-informed decision-making. Theoretical implications extend to refining the treatment of epistemic versus aleatory uncertainty in PSHA, and practically, the framework offers a scalable tool for regional hazard assessment and near-real-time hazard updates. Contributions to knowledge include a novel, globally applicable Bayesian framework for seismic hazard that coherently fuses heterogeneous data sources and tectonic knowledge, a transparent approach to uncertainty decomposition in hazard curves, and a validation protocol bridging historical and instrumentally recorded data across diverse tectonic contexts. Methodological innovations encompass a modular hierarchical model, a data assimilation pathway for ongoing updating, and an explicit treatment of model structure uncertainty via Bayesian model averaging. The study concludes that Bayesian hierarchical modeling improves interpretability and reliability of seismic hazard estimates, particularly in data-sparse regions, and recommends adoption of this framework by national hazard agencies and international observatories to augment decision-support tools for seismic risk mitigation. Policy implications include improved design spectral risk bounds for building codes, enhanced probabilistic risk assessments for critical infrastructure, and a framework for communicating hazard uncertainty to stakeholders under unfolding seismic sequences.