Overview

• Researchers ran a full-physics Bayesian inverse simulation with over a billion parameters on El Capitan, using 43,500 AMD MI300A APUs and 55.5 trillion degrees of freedom to precompute tsunami wave responses.
• The resulting digital twin uses seafloor pressure sensor data and uncertainty-aware models to infer tsunami heights in under a second during unfolding events.
• Findings appear in an arXiv preprint and earned a spot as a finalist for the 2025 ACM Gordon Bell Prize for its record-breaking computational scale.
• Team members say seconds-scale forecasts could offer critical lead time for coastal evacuations in fast-moving scenarios like a Cascadia Subduction Zone rupture.
• The system remains a research demonstration pending expanded seafloor sensor networks, operational validation, integration with existing warning infrastructure, and further open exascale research as El Capitan transitions to classified NNSA missions.