Researchers at Lawrence Livermore National Laboratory (LLNL) have reached a milestone in combining AI with fusion target design by deploying AI agents on two of the world's most powerful supercomputers to automate and accelerate inertial confinement fusion (ICF) experiments.
Part of an AI framework called the Multi-Agent Design Assistant (MADA), LLNL scientists and collaborators are merging large language models (LLMs) with state-of-the-art simulation tools to interpret natural language prompts from human designers and using the platform to generate full physics simulation decks for LLNL's next-generation 3D multiphysics code, MARBL. One of MARBL's key strengths is enabling the design and analysis of mission-relevant, high-energy-density experiments, including ICF.
In ICF experiments at LLNL's National Ignition Facility (NIF), fusion energy is produced when the facility's 192 laser beams converge on a tiny target of deuterium and tritium, creating a fusion chain reaction. The MADA team is using the exascale El Capitan - the world's fastest supercomputer at 2.79 exaFLOPs peak - and its smaller sibling Tuolumne, to test the AI system. The framework incorporates what the team calls an Inverse Design Agent (IDA) to engineer new ICF targets.
Jon Belof, a physicist and principal investigator at LLNL, said the project's origins date back to 2019.
"At the time, we were really interested in the question of what happens if you combine AI with shockwave physics. It was kind of a strange idea - at least a lot of people thought it was," Belof said. "Interestingly enough, that wound up being the simple part. As large language models have advanced, the notion of having semi-autonomous AI systems working alongside the human for ICF design seemed like a natural next step."
But as AI rapidly accelerates in capability and accuracy, the MADA team - which includes National Nuclear Security Administration (NNSA) Tri-Lab collaborators at Los Alamos (LANL) and Sandia National Laboratories - has since turned that "strange idea" into a sophisticated AI-driven design workflow that is getting results.
In a recent demonstration, an open-source LLM - fine-tuned on internal documentation for MARBL - successfully took a hand-drawn capsule diagram and a natural language request from a human designer, then produced a complete simulation deck and ran thousands of simulations to explore variations in ICF capsule geometry to come up with a novel target design.
The AI-driven design paradigm arrives at a critical time for fusion research. Following LLNL's historic ignition achievement at NIF in Dec. 2022, the Lab has its sights set on developing a robust ignition platform that would open new possibilities for national security applications.
Belof said tools like MADA - which drastically compresses design cycles and explores vast design spaces - could play a key role in identifying the optimal conditions for scaling up fusion yields. By pairing human insight with AI-driven exploration, LLNL hopes to navigate the complex physics of high-gain implosions faster and more efficiently than ever before.
"In principle, AI agents offer a way for us to pursue not only 3-4 distinct ICF design concepts at once - but hundreds or possibly thousands," Belof explained. "Rather than the human running ensembles of simulations, they will be able to run ensembles of ideas. This concept could be massively transformative in nature."
The heart of the MADA system is its AI "agents" - autonomous software entities composed of two key components: an LLM that can understand and respond to human language, and a specialized "tooling" interface - an executable function that enables the agent to perform domain-specific tasks. For MADA, the agent's tooling can generate structured simulation input files and launch them on high-performance computing (HPC) systems.
"We are putting AI in the driver's seat of a supercomputer, which is something that has never been done before," Belof said.
Supporting the Inverse Design Agent is another vital component: the Job Management Agent (JMA). While the IDA handles design generation, the JMA drives execution of large-scale simulation workflows across LLNL's supercomputers, interacting with the Flux scheduler and workflow management tools like Merlin.
The JMA ensures jobs are properly queued, resources are allocated, and simulation outputs are efficiently harvested and returned for downstream analysis. Together, these agents operate in a coordinated fashion - the IDA proposing simulation strategies, and the JMA managing the execution pipeline - forming a seamless loop between AI planning and HPC execution.
"The Job Management Agent brings AI and HPC together to coordinate agents that handle resource management and workflow optimization at massive scales, giving us a critical advantage as we push toward a robust ignition platform for fusion energy," said JMA Team Lead Giselle Fernandez.
This iterative workflow enables an unprecedented level of interactivity between designers and their simulations. Instead of manually coding and launching individual jobs - a process that could take days or weeks - researchers can now explore thousands of design variations in parallel, simply by having a conversation with an AI agent.
"The agent can then take a capsule diagram and a plain-language prompt like, 'Explore the effect of changing a certain part of the geometry and translate that into a valid simulation deck for MARBL," Belof explained. "It then runs that deck, collects results, and can even build a training dataset to power a surrogate model."
The MADA approach leveraged HPC to run massive ensembles - typically tens of thousands of ICF simulations in a single study - across LLNL's Tuolumne, the world's 12th fastest supercomputer. The output of these simulations is then used to train a machine-learning model known as PROFESSOR, which can generate instant feedback for designers exploring new capsule geometries.
"Once trained, the PROFESSOR model generates implosion time histories - radius as a function of time - that change instantaneously when the human designer changes the input geometry," Belof said. "It's a powerful new tool to ICF designers that is made possible with AI/machine learning plus HPC."
By enabling natural language interaction, image interpretation and rapid simulation-to-model pipelines, the MADA project demonstrates how AI can be embedded directly into high-stakes scientific workflows. The result is a new stage of national security design work - one that replaces slow, manual iteration with collaborative AI augmentation.
The implications could extend far beyond ICF. As more exascale-class systems like El Capitan come online, MADA offers a blueprint for how AI agents could act as digital collaborators in domains ranging from materials discovery to weapons certification.
"It's really about enhancing human productivity through AI, in a transformative way," Belof said. "And I think this project shows that we're just beginning to tap what's possible. AI tools have the potential for allowing us to best allocate resources and help understand tradeoffs that will be needed for the next generation of enhanced fusion facilities."
NNSA's Advanced Simulation & Computing program is funding the work. Other LLNL MADA team members include deputy principal investigator Charles Jekel, MARBL Project Lead Rob Rieben, and researchers Will Schill, Meir Shachar and Dane Sterbentz. Nathan Brown of Sandia and Ismael Djibrilla Boureima of LANL also contributed to the work.
