Scientists at Lawrence Livermore National Laboratory have achieved a breakthrough in fusion research by deploying artificial intelligence agents on the world’s most powerful supercomputers to automate and accelerate the design of targets for inertial confinement fusion experiments. The Multi-Agent Design Assistant (MADA) system uses large language models combined with state-of-the-art simulation tools to interpret natural language prompts from human designers, marking the first time AI has been put “in the driver’s seat of a supercomputer.”
This development comes as fusion research enters a new phase following LLNL’s historic ignition achievement at the National Ignition Facility in December 2022. The researchers have since hit that milestone seven additional times, most recently in April when they nearly tripled the energy produced in the original breakthrough.
The MADA framework operates through two specialized AI components running on the exascale El Capitan supercomputer with 2.79 exaFLOPs peak performance and its smaller companion Tuolumne. The Inverse Design Agent translates hand-drawn capsule diagrams and natural language requests into complete simulation decks, while the Job Management Agent manages large-scale simulation workflows across LLNL’s supercomputing infrastructure using tools like the Flux scheduler and Merlin workflow management.
In recent demonstrations, an open-source large language model fine-tuned on MARBL documentation successfully generated thousands of simulations exploring variations in fusion capsule geometry from simple sketches. The system provides instant feedback through PROFESSOR, a machine-learning model that generates implosion time histories changing instantaneously as designers modify input parameters. According to physicist Jon Belof, this approach could enable researchers to pursue “hundreds or possibly thousands” of distinct design concepts simultaneously rather than the traditional 3-4 concepts, allowing humans to “run ensembles of ideas” instead of just simulations.
This advancement is part of a wider trend of AI integration across the fusion energy sector. Microsoft Research held its inaugural Fusion Summit in March 2025, bringing together researchers to explore how AI can accelerate fusion research and bring this energy to the grid sooner. At North America’s largest fusion facility, DIII-D, researchers are already applying AI to active plasma control to avoid disruptive instabilities.
Meanwhile, Lawrence Livermore has partnered with Amazon Web Services to develop AI-driven troubleshooting and reliability systems for the National Ignition Facility, deploying generative AI capabilities across 22 years of operational history to help staff resolve issues rapidly.
These AI advances arrive at a critical time for fusion research, as the technology could play a key role in identifying optimal conditions for scaling up fusion yields and navigating the complex physics of high-gain implosions faster than ever before. The implications extend beyond fusion, offering a blueprint for how AI agents could act as digital collaborators in domains ranging from materials discovery to weapons certification.
