Big Ideas Lab - Tsunami Forecasting

Episode Date: October 27, 2026

Scientists at Lawrence Livermore National Laboratory are proving that it is possible to design a tsunami warning system that can deliver both accuracy and speed. The multi-pronged solution rides on the enormous exascale computing power of LLNL’s El Capitan supercomputer along with physics-based modeling and digital twin technologies in a collaborative effort. Guests featured (in order of appearance):Tzanio Kolev, Computational Mathematician, LLNLBronis R. de Supinski, Chief Technology Officer of Livermore Computing, LLNL--Big Ideas Lab is a Mission.org original series.Executive Produced by Levi Hanusch.Sound Design, Music Edit and Mix by Matthew Powell.Script by Levi Hanusch.Story Editing by Levi Hanusch.Audio Engineering and Editing by Matthew Powell.Narrated by Matthew Powell.Video Production by Levi Hanusch.Brought to you in partnership with Lawrence Livermore National Laboratory. Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.

Roughly 20 miles off the Pacific Northwest Coast, two tectonic plates meet along the Cascadia subduction zone. The moment they shift, a massive earthquake begins. And an invisible clock starts. In this particular case, it's less than 10 minutes. The Z-floor lurches upward, lifting the enormous column of water above it. Propeled by the Earth itself, the resulting wave travels upwards of 500 miles per hour across the open ocean. A tsunami. There is no stopping what nature has set in motion.

Only time and decisions. Life and death choices that rely on complex predictions. There were basically two ways to predict tsunamis. We had something accurate but very slow. and something fast, but not accurate. A problem that can't be solved after the clock starts. Because once it does, times already up. That could generate 30 feet and more of water hitting the shore,

and that could be an enormous amount of economical value and lives and danger to the coastal communities. What if we didn't have to choose between speed and accuracy? Using a simulated network of sensors on the sea floor, a digital twin, and one of the world's most powerful supercomputers to run physics-based simulations, scientists at Lawrence Livermore are demonstrating a future where a large-scale computation performed in advance can help produce an accurate prediction in seconds when an earthquake occurs. We can't change the clock, but we can move the hardest part. of the problem before it ever begins. Welcome to the Big Ideas Lab, your exploration inside Lawrence Livermore National Laboratory.

Hear untold stories, meet boundary-pushing pioneers, and get unparalleled access inside the gates. From national security challenges to computing revolutions, discover the innovations that are shaping tomorrow today. The waves we see break on the beach are surface waves. They're mostly wind-generated, driven by energy near the ocean surface. A tsunami starts deeper. Its energy reaches far below the surface, moving an enormous volume of water with it. It's the difference between dumping a bucket of water and unleashing an entire reservoir. When an earthquake happens, there is shaking on the seafloor

which displaces the water above it, and it creates vertical column of water that starts moving towards the shore. Sanyo Kolev is a computational mathematician at Lawrence Livermore, working with a multidisciplinary team on tsunami forecasting methods. Tsunami is a big wave that in this case is generated by an earthquake. So we're focusing on a particular mode of tsunami generation and on a particular area, which is called the Cascadia Subduction Zone. The Cascadius abduction zone runs offshore all the way from Northern California to British Columbia.

There are two tectonic plates that meet there, and because they're rubbing against each other, that has the potential to produce very big earthquakes. The volumes of water involved, the sheer speed at which they travel, and the distance they move inland are the ingredients for a potent disaster. The mission is to provide accurate information as close as possible to when the clock begins. We had methods for that, but they were very inaccurate. There were a lot of false positives. You couldn't really rely on them. At the other end of the spectrum, there were methods that we know are accurate based on partial differential equations, on advanced math.

But to do a simulation at the scales involved, we're talking about hundreds of miles in vertical direction and tens of miles in horizontal direction, would take probably years. You want to give a warning to people in essentially in real time, maybe a minute after the earthquake happened, but certainly not years after that.

An accurate forecast that takes years can't help in the minutes that matter, but a fast forecast with poor accuracy can trigger false warnings. Computer scientists and mathematicians at Lawrence Livermore set out to find another way, Collaborating with geoscientists at the University of California, San Diego, and digital twin experts from the University of Texas, Austin, scientists on the M-FEM team, a Lawrence Livermore team led by Sanyo that develops an open-source library of finite element methods, created a computational forecasting approach designed to show how a real-world tsunami warning network could finally deliver both speed and accuracy together. The solution starts by addressing the most immovable part of the equation.

Time. Instead of waiting for an earthquake and then trying to solve an enormous physics problem in real time, they could do much of that work in advance and build a system designed to use it when the real event came. At Livermore, Sonio and the M-Fem team developed the open-source finite element software used to turn those equations into computations that can run across a supercomputer. And finite elements are a mathematical technique that allows us to take real-world processes and translate them into things that the computers understand. Basically, take a process like a tsunami and digitize it, represented in a computer in a way where we can, model what will happen if a earthquake generates a tsunami. It's a kind of approach where we break the big problem into small individual pieces.

Those small individual pieces are the blocks in finite element modeling. You can think of it kind of like Lego blocks. We combine to build a big representation of the device or the process that we are studying. In this case, they're a representation of the ocean and sea floor. The smaller and more numerous the elements become, the more detail the model can capture. But each one also adds computation. And inside the individual elements, we perform some computations, and then we combine them together to get the full solution. Computations involving such physics-based modeling are laborious and time-intensive,

and they require extraordinary computational scale. If you want to understand a large area by hundreds of miles by tens of miles wide ocean, and you want to understand individual behavior at points, it is mind-boggling amount of precision and understanding of phenomena. The more precisely scientists represent an area that large, the more computation the model demands. So how far can you push the resolution? And what kind of machine could even handle it?

At the top of an Olympic downhill course, everything narrows. No attention is held in reserve. For the next few minutes, everything is committed to getting one athlete down one mountain as fast as possible. Everything available is pushed toward one extraordinary performance. Supercomputing has its own version of that. They call it a hero run. Bronis R. D. Sapinski is the chief technology officer for Livermore Computing at Lawrence Livermore National Laboratory. A hero run is a run on the largest computers available to you that takes all of the capability that's available.

They're the kind of runs required to model the ocean and sea floor at the level of detail the tsunami team needed. In order to perform the calculation has to use that capability over a very extended period of time, at least a month, if not three or four months. But advances in high-performance computing are changing the definition, compressing calculations and producing simulations that once consumed large portions of supercomputers for months in vastly shorter timeframes. At Lawrence Livermore, that machine is El Capitan, one of the fastest supercomputers in the world.

With our previous system, we were able to do some pretty complex 3D models. But El Capitan is such a step forward in terms of the computing capability that is being delivered to the program, that we've been able to take those 3D model runs and take them from being hero runs to being much more commonplace. The supercomputer takes what used to be an Olympic event and turns it into something scientific. can repeat regularly. So now people are able to accomplish in a week's time and even in sometimes a day's time what was taking months in the past. L-CAP can turn yesterday's extraordinary calculation into something scientists can do again and again.

But the tsunami team still found a way to give the machine a hero run of its own. A problem with more than 55 trillion degrees of freedom. That is 55 followed by 12 zeros. The team adapted and optimized M-FEM for El Capitan's GPUs. Then scaled the calculation across more than 43,000 of them. The result was a 55.5 trillion degree of freedom simulation. The largest unstructured mesh finite element simulation reported at the time. A degree of freedom is kind of like an individual point where we know something.

So the more degrees of freedom you have, the higher the resolution, the better the image, the sharper the detail. It's like a pixel on your screen. And we're talking about trillions of pixels. So this is millions of 4K TVs, if you want to think about it this way. Except this image is in three dimensions. And it's moving. Changing over time as an earthquake moves the seafloor and its wave moves across the ocean.

It's a scale of modeling that was previously out of reach. there was a time when millions of degrees of freedom was a very hard problem. Then billions of degrees of freedom, that's a thousand times more. Over the year, it's become a symbol of a very, very difficult hero calculation, big problem to solve. Now with Al Capway in the trillions, so that's million millions, right? That's a ridiculous amount of resolution. The tsunami problem needed El Capitan. But El Capi Tan also needed problems like the tsunami model.

It was one of the earliest real applications we ran. The supercomputer was still relatively new, and real applications could stress it in ways benchmark tests couldn't. There's this process of running things more and more aggressively until you finally feel like, well, the system really is ready to run, and then you put real applications on there, and real applications end up doing things that none of your tests, do. Running something that is doing a fairly complex calculation like the tsunami modeling is a good way to take it so it's working that much more reliably. We really pushed it all the way. We used

almost all the GPUs, we used almost all the memory. We filled it up with this enormous problem, and it held up and it gave us really impressive results. All of that heavy computational work happens before the earthquake. Once the clock starts again, the model needs to understand. what the earthquake is doing in near real time. There are no sensors right now. This is a work that is a mathematical and computer science that shows what is possible if we have sensors. An array of sensors on the ocean floor

could gather the information necessary to understand how the seafloor is moving. Below the ocean floor, sensors track seismic activity and wave movement in real time. That data feeds into an advanced, early warning system. There is a precedent for that in Japan with Fukushima. That earthquake was close to shore that generated the tsunami that led to the

Fukushima disaster. After the earthquake, they installed a network of sensors. It's called SNET on the floor. And this is the same idea here in the Cascadia region. For this area, the team proposed a system of 600 sensors to detect changes in pressure on the seafloor. They are still quite far apart. This is a very large region.

The sensors provide only pieces of the full picture. You know what happened on the seafloor just at some points where the sensors are, but you want to infer what happened everywhere. In order to understand what happened, scientists need to reconstruct how the entire seafloor moved, using only measurements from the individual sensors. To do that, they use a technique called an inverse method, which allows them to estimate what conditions created the data they observed. Instead of starting with the full motion of the seafloor and calculating what the sensors would detect,

the problem runs in the opposite direction. Start with what the sensors detected. Then use a model built on the detailed simulations already run on El Capitan to infer how the rest of the seafloor moved. We combine this with a digital twin. It's a representation of a real phenomena on a computer that can absorb information from the real world and produce actionable items in the digital realm. In this case, it's a twin of the Cascadia subduction zone. The twin is not simply a digital copy of the seafloor.

It represents the process connecting an earthquake in Cascadia to the tsunami it could send towards shore. A digital twin takes information from real sensors on the seafloor, which are activated during earthquake. It produces predictions on where the tsunami will hit and how high the waves will be and gives you the prediction that then you can act on. That combination is what allows the model to be fast without sacrificing the accuracy of the full physics-based solution. The digital twin ultimately is giving answer of the same quality as if we were to solve the full problem in a traditional way. It's not like it's giving an approximate answer or an answer that is at a course resolution less accurate.

You can evaluate things very quickly in the digital twin. We can then evaluate with much more confidence and higher fidelity in the physics-based model. The digital twin ties the live measurements from the earthquake to the detailed physics calculated in advance. It lets the system reconstruct the larger motion of the seafloor and turn it into the forecast scientist's need when the clock starts again. This time, the system doesn't have to start from zero. The idea is that when an earthquake happens, this sensors sense the pressure differences on the floor.

They give that data to a local station. That is the input to the digital twin. We crunch that data very quickly. We combine it with pre-computed stored simulations on LCAP. And then we spit out, okay, in this particular... point on the shore. This is when a big wave will come. This is how big it's going to be. That's the idea of the system. And we've demonstrated that it works in simulated scenarios. The problem that once forced an impossible choice between speed and accuracy can now produce

an answer while there's still time to use it. We were able to combine real time with accuracy. And so we are able to actually get prediction after a earthquake in a matter of tens of seconds. That is of the same accuracy as if we were to solve the big, complicated, fully mathematical problem that we know describes that phenomenon well. That was never possible before this work. The result was more than a faster tsunami calculation. Lawrence Livermore scientists had taken a physics problem that could require years to solve at full scale and demonstrated a way to produce a prediction with the same accuracy in under a minute. being able to now have a real time, early warning system for something that before would have taken years.

I think the combination of all of this was just really amazing. The sensor network itself does not yet exist. But the computational demonstration was strong enough to win the 2025 Gordon Bell Prize, the highest honor in high-performance computing. What made it stand out, I think, is the fact that we apply high-performance computing to a real life, easy to understand the impact of simulation. It combined advanced math with advanced computer science and the world's largest machine and the biggest finite element problem ever solved.

El Capi Tan was central to making it possible. El Cap is just an amazing machine. It certainly changed the game in what's possible in terms of simulations. The fact that El Cap had this super-powerful GPUs, allowed us to do simulations that are super complex, at super high resolution that we couldn't even dream before. The tsunami project did more than put a new supercomputer to the test. It demonstrated what's possible when calculations that once represented extraordinary hero runs become something scientists can begin building entire workflows around.

There are many other places in the world. They have similar challenges. These same ideas could be applied to. The tsunami workflow is one example of a broader shift in scientific computing. Across Livermore's scientific and national security missions, researchers are increasingly combining high-fidelity physics simulations, large datasets, and SI to explore complex problems more quickly. That kind of design cycle of large data models, large physics simulations informing them,

then using those to generate models that you can use to explore that design space and then go back and rerun simulations with those inputs and see, okay, yes, it is working the way the model predicts. That's already true in the national security space in tsunami modeling, protein folding, and in weather prediction five years from now. Nobody will think twice about it. And that's what excites me about where L. Capital, is driving those kinds of workflows. The tsunami project gives that future a tangible proof point.

We can't slow the earthquake. We can't stop the wave or add a single minute to the clock. But scientists at Lawrence Livermore National Laboratory have shown we can change how much time has to be spent waiting on an answer. And how reliable that answer can be. Thank you for tuning in to Big Ideas Lab. Lab. If you loved what you heard, please let us know by leaving a rating and review. And if you haven't already, don't forget to hit the follow or subscribe button in your podcast app to keep up with our latest episode. Thanks for listening.