Big Ideas Lab - AI at the Lab

Episode Date: May 20, 2025

Artificial intelligence is transforming how science is conducted at Lawrence Livermore National Laboratory. From accelerating drug discovery to optimizing complex experiments, AI is helping researchers work faster, smarter and with greater precision. Today we’ll explore how scientists are using Cognitive Simulation - an AI-driven approach that combines physics, data, and machine learning - to model the real world.-- Big Ideas Lab is a Mission.org original series. Executive Produced by Levi Hanusch and Lacey Peace.Sound Design, Music Edit and Mix by Daniel Brunelle. Story Editing by Daniel Brunelle. Audio Engineering and Editing by Matthew Powell. Narrated by Matthew Powell. Video Production by Levi Hanusch. Guests featured in this episode (in order of appearance): Brian Spears, Director of the AI Innovation Incubator, LLNLKelli Humbird, Group Leader, Computational Physics, LLNLBrought to you in partnership with Lawrence Livermore National Laboratory.

This week we took a giant step forward with the release of Chat GPT 4.0. Chat GPT has been held as a game changer. AI is at our fingertips. Hey, I'm Chat GPT, your AI assistant built by OpenAI. I can help with writing. But how is it driving the next wave of scientific discovery? While artificial intelligence can feel a little unnerving, in the world of science, it's ushering in a golden age of knowledge. The things we focus on are pretty different from what a lot of these other companies that do AI machine learning focus on.

Most of us ask AI questions just for fun or out of curiosity. What should I make for dinner with these ingredients? Garlic butter pasta. Can you give me movie recommendations? Here are three popular movie recommendations. Can you create a poem in the style of Walt Whitman? Certainly. I sing the earth that thrums beneath concrete.

The wires tangle like... But the questions scientists at Lawrence Livermore are asking AI are reshaping our world. Hey, AI Chatbot and Chat chat GPT, help me understand what happens if I put a high pressure load on this material. What would happen if I drove a shock wave that's so strong that it ionizes the material? That is, it's so strong a pressure wave

that it rips the electrons off of the atoms and causes radiation to propagate through the material. And we've already seen the results. A nuclear fusion reaction that produced more energy than was used to create it. Able to recreate the temperatures and pressures close to what exists in the core of stars. Artificial intelligence helped Livermore scientists predict and optimize the experiment that achieved fusion ignition, the same process that powers the stars. Today on The Big Ideas Lab, we explore how artificial intelligence at Lawrence Livermore is reshaping science with real-world impact and what comes next.

Welcome to The Big Ideas Lab, your weekly 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. While most people use AI as a smart personal assistant, at Lawrence Livermore, it's a way to speed up the scientific process using reasoning models that generate, refine, and test ideas faster than humans. We have some unique capabilities at Lawrence Livermore.

We have the highest powered computers in the world for science with machines like El Capitan. We also have the world's foremost experimental facilities like the laser at NIF and incredible production capabilities for advanced manufacturing. Those tools are fantastic. What they really need is a capability to be steered at high rate, to have hypotheses and then have those winnowed down by doing a high-performance simulation. It's huge amounts of work to go do each one of these experiments. So for every idea that we have, we can do 10 experiments and across a set of ideas, we have maybe 100 or 200 of these experiments, but they are incredibly high precision.

Here's where AI comes in. After eight decades of development of those computing and experimental capabilities, onto the stage have come companies like OpenAI and Anthropic. And they produced AI tools that we can call reasoning models. Those reasoning models can help us understand math and science and produce hypotheses based on our data. AI essentially provides an accelerated feedback loop where every hypothesis feeds a simulation.

Every simulation leads to better data. And that data helps refine the next set of ideas. It's an immeasurable acceleration of your capabilities because we don't really know right now how much time we're spending on ideas that we wish we didn't, until we push those all the way through the production chain of thinking about pushing information from idea to simulation to experiment. And so we used AI tools to go do that, so we could take our simulation capability and get it dialed in and perfectly honed so it would tell us what we should expect in the experiment. That same feedback loop, where AI narrows thousands of possibilities into just a handful

of promising experiments, is precisely what played out in one of the lab's most historic achievements. Fusion Ignition. most historic achievements, fusion ignition. Artificial intelligence helped scientists at the lab identify which experiments were most likely to succeed in achieving fusion ignition. It processed massive amounts of simulation data, identified patterns, and guided decisions that led to ignition. You wake up one morning and your tool tells you, you're more likely than not to ignite, and it's pretty exciting.

That tool was part of an approach to scientific modeling known as cognitive simulation, an AI-driven system that can learn from both experiments and simulations to make real-time predictions. It's pretty mind-blowing that we use this to get fusion ignition for the first time in human history. time predictions. piece of nuclear fuel and got more energy out than what we put in with the laser. That explosion was a carefully engineered fusion experiment using a powerful laser to compress a tiny capsule of hydrogen fuel under extreme heat and pressure until the atoms fused.

That's really creating a little star about the diameter of your hair for about a hundred trillions of a second. The exciting part was we had modeling and simulation tools that told us that it looked like this was going to happen. We had experimental tools that told us, yeah, the data is indicating that if you go in this particular direction, this might happen. And then we used this COGSIM piece and the COGSIM piece said, looking at all of the simulations and the data we have from the past, I've got a capability to analyze new designs. Cognitive simulation, or COGSIM, combines physics-based models, experimental data, and

AI that learns from both. It's built on decades of research and provides a foundation to intelligently evaluate experimental scenarios that have never before been tested. We showed those tools a new design, and it said you've got a greater than 50% chance of igniting, that is getting more energy out than what we put in with a laser. Greater than 50% is not overwhelming confidence, but for the prior six decades that number had been tiny. So 15%, 5%, next to nothing.

This marked a paradigm shift. For the first time ever, the predictive models indicated a significant chance of success. A prediction that successful fusion ignition confirmed. While achieving fusion ignition with the help of AI was a major milestone, scientists at Lawrence Livermore are applying AI in many other areas. The common goal is to understand how a physical system will behave, whether it's a fusion reaction, a new material, or even a drug compound, before running a single real-world experiment. This allows them to test ideas virtually, make adjustments, and predict outcomes in advance. Cognitive simulation is our Livermore brand, our story for the way that we couple AI to physical science.

Really it is the combination of our simulation and experimental capabilities coupled with deep neural networks and AI. So you can imagine it this way. We can take deep neural networks, these AI tools, we can train them on our simulation models, then they have a picture of what the world should look like, but those models are always imperfect. They approximate the real world, then we incorporate experimental data,

and that gives our models not only an understanding of the way the world should be, but a picture of the way it actually is, so that this cognitive simulation model that knows both is actually elevated. AI becomes a powerful predictive tool when it understands both what should happen and what actually happens.

It's got a picture of what the world ought to be like and what it is actually like, so it can make accurate and precise predictions for what we will actually see in the laboratory the next time we do an experiment. At Lawrence Livermore, scientists are using this approach to accelerate research in areas like national security, advanced materials, in areas like national security, advanced materials, and drug development. Experiments in these fields can be expensive, time-consuming, or even impossible to run in the real world. Kelly Humbert is a design physicist at the lab. What's interesting about this approach is we have the ability to incorporate new experimental data as we acquire it. We have these models that train

on these large simulation data sets. We've gotten really good at leveraging our high-performance computing resources to make massive data sets. We can train machine learning models on that data, and then we can modify these machine learning models using the experimental data. Kelly's team uses CogSIM to find faster,

more accurate answers in complex scientific systems. The way I like to visualize it is to think of it as a map. Our simulations give us a map of what they think the design space for ICF looks like. We know that there's a mountaintop where the yield is high and there's a cliff where you can fall off and get low yield if you wander too close to the edge. And our simulations give us an idea of where they think that mountain is. Our experimental maps suggests maybe the

mountain isn't quite where they said it was or maybe it's not quite as big as they said it was. And this cognitive simulation technique lets us make the experimental map in a pretty data efficient way. While AI is a powerful tool, scientists at the lab are clear about its limits with constant testing, challenge, and validation to ensure it delivers accurate results.

I think we might be the largest machine learning skeptics out there, and AI skeptics, you know, we're the ones who embrace it for our jobs, but we challenge these models and these ideas too because we want to make sure that we're doing our jobs to the best of our abilities. There are two things that I am principally concerned about. The first is that AI can give you amazingly correct answers, but it's not guaranteed to give you answers. So you can get the wrong answer.

The second thing that I'm worried about is take a system that can give the right answer, but is not guaranteed to, and then try to turn it into high consequence actions for making actual things that go to help people, that go into systems that they're gonna use and operate. The goal is to use AI to support better decisions, faster modeling, and smarter experiments without losing the scientific rigor or human judgment to make meaningful decisions.

I spent years learning how to run high-performance computers in order to execute simulations to answer physics questions. I am looking forward to the world where all of that capability that I learned for years of driving these simulations is offloaded to an AI system. Because I never really wanted to be the world's best computer jockey of executing those simulations. Having the right attitude of, these are tools that can help you do your job more efficiently or faster,

but they're not replacing the final human analysis of the decisions we're about to make or of the experiment we're about to field. These technologies are just letting us get to the suite of possible answers a lot more efficiently and taking into account a lot more information than we can hold in our brains at any given time. I think as long as the scientists are approaching these things as tools to help them in their work, not tools to replace them in their work, and continue to be skeptical and really hard

on these models, we'll ensure that the answers that we're using from them are ones that everyone feels good about. While achieving fusion ignition was a monumental milestone, the applications of AI at Lawrence Livermore National Labs extend beyond energy research. One of the most impactful areas is in healthcare, particularly in accelerating drug discovery, a process that traditionally takes years. The first things that we've done with AI

are incorporate them into our scientific method operations. The flagship case is probably in our bio-resilience science where there are AI tools that are helping us produce candidates for new drugs. Finding new treatments for diseases like cancer is a complex, time-consuming challenge. It can take years of research that costs billions of dollars. The lab is partnering with Bridge Bio-Oncology Therapeutics and the Frederick National Laboratory

for Cancer Research, using its AI-driven drug discovery platform to develop a novel medication targeting genetic mutations linked to nearly 30% of all cancers. Together, we're showing that when scientific ingenuity and cutting-edge technology meets with novel public-private partnerships, possibilities are endless. That partnership has already resulted in the development of three new cancer drugs currently working their way through the FDA approval process. Recently, Lawrence Livermore researchers also published a paper on their successful use of a different AI-based platform to preemptively optimize existing antibodies to neutralize

a wide range of potential variants of SARS-CoV-2, the virus that causes COVID-19. The work marks a promising step in using AI to counter evolving viruses and protect against future pandemics. So the process that used to take years to go into making a medication, the discovery of the molecule part can now be weeks. And then there are future bottlenecks that we're looking at. You have to make and manufacture those molecules. You have to put those through clinical trials and make sure that they're safe and effective. And then you can go back and use that. So the lab has already in its first hit with AI, taking the initial phase of molecular discovery and shorten that down by a tremendous amount. And now that program on our biocide is turning its attention to,

how do you then accelerate with AI the production capability of making that molecule faster, of making it manufacturable. That acceleration, going from concept to solution in a fraction of the time, is one of the most transformative strengths AI brings to the lab. AI is going to show up in absolutely everything that we do. It doesn't really matter what part you're making. You could be making brake pads, you could be making jet turbine engines, you could be making parts for the nuclear stockpile. The time from idea to execution is now 10 times smaller so you can actually go do it and you can outrun any of the things that are burdening

your manufacturing system. And so what you can see is AI is leaving your laptop and going out into the real world and doing things of consequence. At Lawrence Livermore, artificial intelligence is accelerating the scientific method. By pairing AI with physics, biology, and engineering, researchers are solving complex problems faster, testing ideas more efficiently, and pushing science forward with greater precision. From recreating the power of the stars to advancing medicine, AI is reshaping science. Many years from now it would be very cool to have a model that is a domain expert in fusion,

for example, and can store a lot of data right at the tip of its memory where we can't as humans necessarily do that. I think it's possible that these models can help us come up with hypotheses to problems that we haven't solved yet. It'll be really cool to get to a place where we might have scientific assistance based on these AI models that can just

help us think through really large quantities of data more efficiently than we can just as humans. So I think it still feels a little bit like a pipe dream, but I think we'll get there in the next few years based on the trajectory of progress. The first time I used a model and I had that moment of grief, like, oh no, the thing I plan to do with my career for the next four years has been done. Not completely, but has been done well. My reaction to that, for me, after just a few seconds really was, oh my God, I can now do the next five years of stuff that I was planning today. And so the story there is for all of us as humans on the planet to understand what we

uniquely bring to the world that we're producing, the capabilities that we bring, and differentiate between those things we're doing that we're okay being offloaded to another thing, like a large language model, and identifying what is the special sauce that we bring uniquely. As AI becomes more integrated into science, the tools may change. But the curiosity, creativity, and critical thinking behind discovery remain deeply human. Researchers are using AI to answer big questions, ask better ones, and get to the answers faster. Thank you for tuning in to Big Ideas 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.