A new generation of earthquake source models- Past, present and future

点击翻译按钮获取中文摘要

This project focuses on theoretical and computational developments needed for the next generation of seismic models. This effort concentrates quantifying the impact of uncertainties in models and to rigorously incorporate uncertainties into the algorithms behind the models. The last decade has seen considerable improvements in modeling capability and a substantial expansion of geophysical observations including data from dense seismic networks, networks of permanent GPS installations and spatially synoptic geodetic imaging data from orbiting radar and optical satellites. Despite these efforts, perhaps the biggest obstacle to significant progress in observational earthquake source modeling arises from imperfect predictions of observations due to uncertainties in Earth structure - the impact of which are generally overlooked. Indeed, for large earthquakes, our ability to measure ground motions far exceeds our ability to model them. The expected benefits of this approach include improving the sharpness of our images of seismic and aseismic fault slip processes -- thereby directly impacting both our models for fault mechanics and inferences of seismic hazard. The algorithms developed here are directly applicable to other problems such as models of volcano deformation. All tools developed here will be documented and openly available to the geophysical community as open source. This project develops the concept of the misfit covariance as used in inversions for the distribution of slip on subsurface faults. The misfit covariance is a combination of covariances in observations (often assumed independent) and covariances associated with inaccurate model predictions (often entirely ignored). Yet, this prediction error can dwarf the observation error and can induce important covariances between observations. These covariances dictate the relative weighting between disparate data types as well as the information content found in observations from dense networks. The proposed approach to estimating the full model prediction error relies on developing computationally tractable methods for estimating the sensitivity of geodetic and seismic observations to perturbations in assumed material properties. This approach exploits recent advances in Bayesian earthquake source modeling and new massively parallel computational approaches using GPUs.