Seismo Lab Seminar

When it comes to earthquake ground motions, engineers have access to two testbeds to evaluate possible solutions that withstand the passage of time and the extreme loads of natural hazards: observations, which sample the ground truth but are spatiotemporally limited and do not occur on demand; and simulators, with unprecedented capabilities that nonetheless reflect the input accuracy of their constituent models and in their most high-fidelity versions, require supercomputers and expert users. Recent advancements in machine learning show promising potential to combine the best of both worlds: the between- and within-event variability from observations that encapsulate the physics and variability not mapped in the simulator input models; and the spatial coverage and scaling constraints of rare scenarios from simulations for the scenarios that matter in design -- the catastrophic events that have yet to occur. In this talk, I will present new algorithms that our team has developed to overcome challenges in learning ground motion ensembles and generating new content, and applications in generating time-series for shallow crustal events from seismic networks in Japan; and wavefields from simulated ground motions in the San Francisco Bay Area. Our current work focuses on optimally fusing regularly spaced simulation results and sparse observations into wavefields that realistically depict between- and within-event variability and high frequency statistics; and can generate event-level ground motions at a fraction of time and computational expense of cutting-edge computer simulators.