Overview

• LLNL researchers executed a one-time offline precomputation on El Capitan using over 43,500 AMD Instinct MI300A APUs and the MFEM library to link earthquake-induced seafloor motion to tsunami waves.
• The digital twin solves a billion-parameter Bayesian inverse problem in under 0.2 seconds, delivering an estimated 10-billion-fold speedup over existing forecasting methods.
• Real-time forecasts infer seafloor motion from pressure sensor data and compute full 3D acoustic-gravity wave propagation with uncertainty quantification.
• Transitioning from demonstration to operational early-warning depends on deploying denser seafloor sensor networks and integrating the precomputed libraries into tsunami warning centers.
• El Capitan has since moved to classified NNSA tasks, so future large-scale computations will rely on porting the offline results to smaller GPU clusters.