When a powerful offshore earthquake strikes, coastal communities may have mere minutes to prepare for the impacts of a devastating tsunami. Providing an accurate warning in that narrow window presents an enormous scientific challenge: sophisticated physics-based simulations can take too long, while faster forecasting methods often sacrifice accuracy.
In the latest episode of the Big Ideas Lab podcast, researchers explain how they used Lawrence Livermore National Laboratory's (LLNL's) exascale El Capitan supercomputer, advanced algorithms and digital twin technology to overcome that tradeoff, demonstrating an approach that could someday deliver highly accurate tsunami forecasts in less than a minute. Listen on Apple or Spotify.
The episode features LLNL computational mathematician Tzanio Kolev and Livermore Computing Chief Technology Officer Bronis R. de Supinski, who discuss the work and its implications for scientific computing and tsunami early-warning systems.
The research focuses on the Cascadia Subduction Zone, a 700-mile-long fault extending from Northern California to British Columbia, Canada. A major earthquake could generate a tsunami reaching parts of the Pacific Northwest coastline in less than 10 minutes, leaving little time to assess the threat and issue warnings.
Traditionally, scientists have relied on simplified models that produce rapid but less accurate predictions or high-fidelity simulations that capture complex earthquake and tsunami physics but require enormous computational resources. LLNL researchers, collaborating with scientists at the University of Texas at Austin and the University of California, San Diego, developed a different approach. Using El Capitan's exascale capabilities (about 1.8 quintillion calculations per second in real-world runs), the team performed the most computationally demanding calculations in advance, enabling much faster predictions when new sensor data becomes available.
Using LLNL's open-source MFEM finite-element library, the researchers developed detailed simulations of the seafloor and ocean. On El Capitan, they solved a problem involving more than 55 trillion degrees of freedom across more than 43,000 graphics processing units, achieving unprecedented resolution. The achievement earned the team the 2025 Association for Computing Machinery Gordon Bell Prize, the most prestigious honor in high-performance computing.
The researchers combined the simulations with a digital twin, a computational representation of a physical system. In the proposed forecasting system, hundreds of seafloor sensors would detect pressure changes during an earthquake. Mathematical techniques called inverse methods would use those measurements and precomputed simulations to reconstruct seafloor movement and predict tsunami arrival times and wave heights.
In simulated earthquake scenarios, the team demonstrated forecasts in seconds with accuracy comparable to computationally intensive physics-based simulations. Although the proposed sensor network does not yet exist, the results point toward faster, more reliable tsunami warnings.
The project also illustrates El Capitan's broader impact on scientific computing. Calculations that once required large portions of supercomputers to work for weeks or months on a single problem are becoming more routine, enabling researchers to tackle increasingly complex problems. As de Supinski explains, combining physics simulations, large datasets and computational models is transforming research across LLNL's scientific and national security missions, from stockpile modernization to materials science.
Listen to the latest episode of Big Ideas Lab to learn how LLNL researchers are advancing tsunami forecasting and pushing the boundaries of supercomputing - on Apple or Spotify.
