 |
Center for Nonlinear Studies |
The Cascadia Subduction Zone (CSZ) extends over 1000 km from northern California to northern Vancouver Island. The CSZ is capable of unleashing a magnitude 9 earthquake with resulting 30-meter-high tsunamis. While 43 earthquakes have occurred on the Cascadia fault over the last 10,000 years, it has been eerily silent since 1700 and is considered overdue for a major earthquake.
Plans are underway to deploy a network of seafloor-mounted pressure sensors in the CSZ. The sensors will record acoustic waves in the ocean that are generated when the seafloor is suddenly uplifted during an earthquake. Our goal is to use these observed pressure transients, along with a coupled acoustic-gravity wave propagation forward model, to infer the seafloor motion, and then forward propagate the resulting tsunamis toward coastal regions. Since destructive tsunami waves can arrive onshore in as little as 10 minutes, the inverse solution and subsequent tsunami forecast must be carried out in seconds to be useful for early warning. Not only do we want to predict the mean of the tsunami wave heights, but we also want to equip this digital twin with uncertainty estimates in a fully Bayesian framework.
However, a single forward wave propagation simulation takes an hour on 512 A100 GPUs. To infer the billion parameters representing the uncertain spatio-temporal seafloor motion, state-of-the-art inversion algorithms need hundreds of thousands of such forward simulations. We show that by exploiting the time shift-invariance and linearity of the parameter-to-observable map, transforming the inverse operator from the parameter space to the data space, and devising novel parallel Bayesian inference algorithms that map well onto GPUs, we induce a fast offline-online decomposition that allows the seafloor motion inversion and subsequent tsunami forecast to be carried out exactly in a fraction of a second. Finally, we present fast algorithms for the optimal experimental design problem of optimizing the locations of the seafloor pressure sensors to maximize information gain from the data.
This work is joint with Stefan Henneking (UT Austin), Sreeram Venkat (UT Austin), and Alice Gabriel (Scripps/UCSD).
Bio: Omar Ghattas is Professor of Mechanical Engineering at The University of Texas at Austin and holds the Cockrell Chair in Engineering. He is also Principal Faculty and Director of the OPTIMUS (OPTimization, Inverse problems, Machine learning, and Uncertainty for complex Systems) Center in the Oden Institute for Computational Engineering & Sciences, and a member of the faculty in the Computational Science, Engineering, and Mathematics graduate program. He holds courtesy appointments in Earth & Planetary Sciences and Biomedical Engineering. Before moving to UT Austin in 2005, he spent 16 years on the faculty of Carnegie Mellon University. His current research focuses on theory and algorithms for large-scale Bayesian inversion, stochastic optimal control/design, and digital twins for complex engineered and natural systems. He is a three-time winner of the ACM Gordon Bell Prize, a recipient of the SIAM Geosciences Career Prize and the SIAM Babuska Prize, and a Fellow of SIAM and USACM. He holds BSE (civil and environmental engineering) and MS and PhD (computational mechanics) degrees from Duke University.
Host: Sara Calandrini (scalandrini@lanl.gov)