Adaptive Inverse Modeling for Real-Time Seismic Hazard Mapping

Thesis Abstract

This study addresses the growing need for timely and accurate seismic hazard assessments through adaptive inverse modeling that integrates heterogeneous geophysical observations to produce real-time hazard maps under varying seismic source and subsurface conditions. The aim is to develop a flexible inverse framework that updates seismic velocity, attenuation, and rupture-estimation parameters in near real-time as new data become available, thereby improving the spatial resolution and reliability of hazard maps used for early warning and emergency response. Specific objectives are to (i) formulate an adaptive Bayesian inverse model that assimilates seismic waveform data, strong-motion records, and ambient noise, (ii) quantify and propagate uncertainties in model parameters to produce probabilistic hazard maps, (iii) implement a scalable computational pipeline enabling near real-time updates within a 30–60 second data assimilation window, (iv) evaluate model performance against historical earthquakes across diverse tectonic settings, and (v) assess operational utility for decision makers by comparing hazard map accuracy and timeliness against conventional static inversions. The methodological design integrates a sequential data assimilation approach with inverse theory. The population comprises global seismological stations and high-rate (1–20 Hz) accelerometer networks from five well-documented earthquakes with disparate focal mechanisms. A stratified sampling strategy yields a representative subset of 250 stations for each event, ensuring coverage in near-field and far-field regions. Data collection employs waveform data from standard accelerometers, strong-motion arrays, and ambient seismic noise cross-correlations, supplemented by publicly available catalogued event information. Instrumental calibration, time synchronization, and noise characterization procedures are explicitly embedded. The core analytical framework combines (i) adaptive Bayesian inference with ensemble Kalman filters to sequentially update velocity and attenuation models, (ii) adjoint tomography constraints to refine velocity contrasts, and (iii) rupture-forecast parameters derived from a stochastic slip model. Theoretical grounding rests on Bayesian decision theory and the compatible inverse problem paradigm, with support from the theory of sequential data assimilation and the concept of epistemic-aleatory uncertainty separation. The analysis incorporates Regularized Least Squares and L-curve methods for model regularization, along with cross-validation against withheld data to prevent overfitting. The study also employs information-theoretic metrics (AIC/BIC) to compare model variants and to quantify the value of data assimilation in reducing predictive uncertainty. Data processing begins with preprocessing of seismic waveforms for time synchronization, deconvolution, and noise suppression, followed by event-specific priors informed by regional velocity models and prior ruptures. The method of analysis includes (i) sequential Monte Carlo sampling to explore posterior distributions, (ii) joint inversion of velocity, attenuation, and source parameters with dynamic reweighting as new data arrive, and (iii) production of probabilistic hazard maps calibrated to exceedance probabilities and credible intervals. Model validation uses hindcast experiments for the five target earthquakes, with performance metrics including reduction in posterior predictive error, improved spatial correlation with observed ground motions, and timeliness of hazard updates. Sensitivity analyses examine the impact of station density, data latency, and prior mis-specification on map accuracy. Expected findings indicate that adaptive inverse modeling significantly reduces uncertainty in near-real-time hazard maps, achieving average reductions in posterior root-mean-square error of 15–25% for peak ground acceleration estimates within the first 60 seconds of data, and improved spatial coherence of hazard contours compared with non-adaptive inversions. The study anticipates that data assimilation will more accurately capture complex rupture processes and velocity heterogeneities, particularly in heterogeneous crustal zones. The contribution to knowledge lies in operationalizing an end-to-end, real-time, data-assimilative inverse modeling framework capable of producing probabilistic hazard maps with quantified uncertainty, suitable for integration into earthquake early warning and emergency management workflows. The main conclusions are that adaptive inverse modeling improves both the speed and reliability of seismic hazard estimation in real time, and recommendations include enhancing data sharing among networks, extending the framework to multi-physics data streams (e.g., GPS, InSAR), and developing user-oriented visualization tools for decision makers.