Researchers at Lawrence Livermore National Laboratory (LLNL) have achieved a milestone in integrating AI with fusion target design, automating and accelerating inertial confinement fusion (ICF) experiments by deploying AI agents on two of the world's most powerful supercomputers.

As part of the Multi-Agent Design Assistant (MADA) AI framework, LLNL scientists and collaborators combined large language models (LLMs) with advanced simulation tools to interpret natural language prompts from human designers and generate complete physical simulation platforms for LLNL's next-generation 3D multi-physics code MARBL. MARBL is capable of designing and analyzing mission-relevant high-energy-density experiments, including ICF.

In ICF experiments conducted at LLNL's National Ignition Facility (NIF), 192 laser beams converge on a tiny target containing deuterium and tritium, triggering a fusion chain reaction to produce fusion energy. The MADA team tested the AI system on the exascale supercomputer El Capitan (the world's fastest, with a peak speed of 2.79 exaFLOPs) and the smaller Tuolumne supercomputer. The framework incorporates an Inverse Design Agent (IDA) for designing new ICF targets.

LLNL physicist and principal investigator Jon Belof stated that the project originated in 2019, when the team became interested in combining AI with shock wave physics. With advances in large language models, the idea of a semi-autonomous AI system collaborating with humans on ICF design became the natural next step. Now, the MADA team—including collaborators from Los Alamos and Sandia National Laboratories under the National Nuclear Security Administration's tri-lab effort—has transformed this idea into a sophisticated AI-driven design workflow with tangible results.

In a recent demonstration, an open-source LLM fine-tuned on MARBL internal documentation successfully took hand-drawn capsule sketches and natural language requests from human designers, generated complete simulation platforms, and ran thousands of simulations to explore variations in ICF capsule geometry, yielding novel target designs.

The emergence of AI-driven design paradigms comes at a critical moment in fusion research. Following LLNL's historic ignition achievement at NIF in December 2022, the laboratory is focused on developing robust ignition platforms to open new possibilities for national security applications.

Belof noted that tools like MADA can dramatically compress design cycles, explore vast design spaces, and play a key role in identifying optimal conditions for increasing fusion yield. By combining human insight with AI-driven exploration, LLNL aims to probe the complex physics of high-gain implosions faster and more efficiently. In principle, AI agents can explore hundreds or even thousands of different ICF design concepts simultaneously, bringing transformative change.

At the core of the MADA system are its AI "agents," composed of LLMs capable of understanding and responding to human language and "tool" interfaces that perform domain-specific tasks. For MADA, agent tools can generate structured simulation input files and launch them on high-performance computing (HPC) systems.

Another important component supporting the Inverse Design Agent is the Job Management Agent (JMA). The IDA handles design generation, while the JMA drives large-scale simulation workflows on LLNL supercomputers, interacting with the Flux scheduler and workflow management tools like Merlin. The JMA ensures jobs are properly queued, resources allocated, and simulation outputs efficiently collected for downstream analysis. These agents operate in a coordinated manner, forming a seamless loop between AI planning and HPC execution.

JMA team lead Giselle Fernández stated that the Job Management Agent, combining AI and HPC, provides a key advantage in advancing robust fusion energy ignition platforms.

This iterative workflow enables an unprecedented level of interactivity between designers and simulations. Researchers no longer need to manually code and launch individual jobs; they can now converse with AI agents to explore thousands of design variants in parallel.

The MADA approach leverages HPC to run large-scale ensembles on LLNL's Tuolumne supercomputer, typically performing tens of thousands of ICF simulations in a single study. These simulation outputs are used to train a machine learning model called PROFESSOR, providing instant feedback to designers exploring new capsule geometries. Once trained, the PROFESSOR model generates implosion timelines that change instantaneously as human designers alter input geometries—a powerful new tool for ICF designers.

By enabling natural language interaction, image interpretation, and rapid simulation-to-model workflows, the MADA project demonstrates how AI can be directly embedded in high-stakes scientific workflows, ushering national security design work into a new era—augmented by collaborative AI rather than slow manual iteration.

Its impact could extend far beyond ICF. As more exascale systems come online, MADA provides a blueprint for AI agents acting as digital collaborators in fields from materials discovery to weapons certification. Belof said this elevates human productivity in transformative ways through AI, noting the project shows they are only beginning to tap the possibilities, with AI tools having the potential to help optimally allocate resources and understand trade-offs needed for next-generation enhanced fusion facilities.

The work was funded by the National Nuclear Security Administration's Advanced Simulation and Computing program. Other LLNL MADA team members include deputy principal investigator Charles Jekel, MARBL project lead Rob Rieben, and researchers Will Hill, Mel Shacham, and Dan Steben. Nathan Brown from Sandia National Laboratories and Ismael Gibrail Boureima from Los Alamos National Laboratory also contributed.