Scientists at Lawrence Livermore National Laboratory (LLNL) have helped develop an advanced, real-time tsunami forecasting system — powered by El Capitan, the world’s fastest supercomputer — that could dramatically improve early warning capabilities for coastal communities near earthquake zones.
The exascale El Capitan, which has a theoretical peak performance of 2.79 quintillion calculations per second, was developed with funding from the Advanced Simulation and Computing (ASC) program at the National Nuclear Security Administration (NNSA). As described in a preprint paper selected as a finalist for the 2025 ACM Gordon Bell Prize, researchers at LLNL harnessed the machine’s full computing power in a one-time, offline precomputation step, prior to the system’s transition to classified national-security work. The goal: to generate an immense library of physics-based simulations, linking earthquake-induced seafloor motion to resulting tsunami waves.
The project used more than 43,500 AMD Instinct MI300A Accelerated Processing Units (APUs) to solve extreme-scale acoustic-gravity wave propagation problems, producing a rich dataset that enables real-time tsunami forecasting on much smaller systems. By front-loading the intensive computation work on El Capitan, the team was able to solve an extremely high-fidelity Bayesian inverse problem that makes it possible to generate rapid predictions in seconds — during an actual tsunami — using modest GPU clusters. Researchers said this capability could fundamentally transform the future of early warning systems and save lives.
Developed in partnership with the Oden Institute at the University of Texas at Austin (UT Austin) and the Scripps Institution of Oceanography at the University of California, San Diego (UC San Diego), the resulting tsunami “digital twin” models the effects of seafloor earthquake motion using real-time pressure sensor data and advanced physics-based simulations. This dynamic, data-driven system can infer the earthquake’s impact on the ocean floor and forecast the tsunami's behavior in real time — complete with uncertainty quantification.
“This is the first digital twin with this level of complexity that runs in real time,” said LLNL computational mathematician Tzanio Kolev, co-author on the paper. “It combines extreme-scale forward simulation with advanced statistical methods to extract physics-based predictions from sensor data at unprecedented speed.”
By leveraging LLNL’s Hewlett Packard Enterprise/AMD exascale supercomputer El Capitan in the offline precomputation step, the team was able to solve a billion-parameter Bayesian inverse problem in less than 0.2 seconds, accurately predicting tsunami wave heights at an astounding 10-billion-fold speedup over existing methods.
Researchers said the capability could radically improve emergency response and save lives, forming the backbone of next-generation early warning systems. For near-shore events like a future magnitude 8.0 or larger earthquake along the Cascadia Subduction Zone in the U.S.’s Pacific Northwest, the first destructive waves could reach the coast within 10 minutes — leaving little time for evacuation, researchers said.
Conventional tsunami warning systems often rely on seismic and geodetic data to infer earthquake magnitude and location but typically use simplistic models that fail to capture the complexity of fault ruptures, which can lead to false alarms or dangerously late warnings. The team’s approach instead uses data from seafloor pressure sensors and solves a full-physics model of acoustic-gravity wave propagation in the ocean in record time.
As seafloor sensor networks, including distributed acoustic sensing, become more widespread along earthquake-prone coasts and computational infrastructure continues to improve, the team sees a clear path to deploying the approach in future tsunami warning systems, resulting in faster, smarter and more reliable emergency alerts.
“This framework represents a paradigm shift in how we think about early warning systems,” said senior author of the study Omar Ghattas, professor of mechanical engineering and principal faculty in the Oden Institute at UT-Austin. “For the first time, we can combine real-time sensor data with full-physics modeling and uncertainty quantification — fast enough to make decisions before a tsunami reaches the shore. It opens the door to truly predictive, physics-informed emergency response systems across a range of natural hazards.”
At the heart of the system is MFEM, LLNL’s open-source finite element library, which enables scalable, GPU-accelerated simulations of many physical phenomena, including acoustic-gravity wave propagation in the ocean. Running these simulations on 43,520 APUs of El Capitan, MFEM performed the most compute-intensive phase — solving the acoustic-gravity wave equations to precompute mappings between ocean-floor motion and sensor data — using a staggering 55.5 trillion degrees of freedom, shattering the previous record for the largest unstructured mesh finite element simulation.
“MFEM’s high-order methods and GPU readiness, developed under the ASC program at LLNL and the Department of Energy’s (DOE) Exascale Computing Project, made it possible to scale to the full machine,” Kolev said. “This was really a first-of-its-kind demonstration of how we can use that power not just for raw performance, but also for mission-relevant, time-critical decisions in many MFEM-based applications.”
According to Kolev, once the massive precomputations are complete, the online steps of inferring the seafloor motion and forecasting the tsunami wave heights in real time can be performed on far more modest GPU clusters since the underlying algorithms are designed to map well onto GPUs.
“This work is important because it shows that we can solve an inverse problem of enormous size — not for 10 or 15 variables, but for millions, or even billions of variables, very quickly,” said Kolev. “In the past, you'd either have a fast model that's not accurate, or a full-physics model that takes hours or days. Now we’re showing that we can do both — accurate and fast — using principled mathematics and modern computing.”
Kolev said the Bayesian inversion framework isn’t a “one-off capability” limited to tsunamis and could be applied to a wide range of complex systems — from real-time wildfire tracking and subsurface contaminant tracking to space weather forecasting and even intelligence applications where fast, data-driven decisions are needed.
Collaborators included Veselin Dobrev and John Camier of LLNL; Omar Ghattas, Stefan Henneking, Milinda Fernando and Sreeram Venkat of UT-Austin; and Alice-Agnes Gabriel of UC San Diego.
