Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Explore a groundbreaking conference talk from the MFEM Workshop 2025 that demonstrates how to achieve real-time Bayesian inference for tsunami early warning systems using extreme-scale computing. Discover Stefan Henneking's innovative approach to creating a Bayesian inversion-based digital twin that processes acoustic pressure data from seafloor sensors and employs 3D coupled acoustic-gravity wave equations to infer earthquake-induced spatiotemporal seafloor motion in real time. Learn how this system forecasts tsunami propagation toward coastlines with quantified uncertainties, specifically targeting the Cascadia subduction zone with one billion parameters. Understand the computational challenges involved, where computing the posterior mean alone would traditionally require 50 years on a 512 GPU machine, and see how the research team overcame this by exploiting shift invariance of the parameter-to-observable map and developing novel parallel algorithms. Examine the fast offline-online decomposition strategy where the offline component requires just one adjoint wave propagation per sensor, scaled to the full El Capitan system using 43,520 GPUs with 92% weak parallel efficiency through MFEM. Witness how the online component exactly solves the Bayesian inverse and forecasting problems in just 0.2 seconds on a modest GPU system, achieving a remarkable ten-billion-fold speedup for real-time tsunami warning applications.
