AI agent cuts particle accelerator experiment setup time by 100x at Berkeley Lab

Researchers at Lawrence Berkeley National Laboratory have shown that an autonomous AI agent can prepare and run complex physics experiments at a major particle accelerator, reducing setup time by two orders of magnitude and pointing to a new role for artificial intelligence in large-scale scientific infrastructure.

The system, known as the Accelerator Assistant, has been deployed at the Advanced Light Source, a synchrotron facility in Berkeley that supports around 1,700 experiments each year. According to the research team, the agent was able to autonomously prepare and execute a multistage physics experiment, cutting preparation time by a factor of 100 compared with conventional workflows.

“This is something that can save you serious time. In the paper, we say two orders of magnitude for such a prompt,” said Thorsten Hellert, a staff scientist in the Accelerator Technology and Applied Physics Division at Berkeley Lab and lead author of the research.

The Advanced Light Source accelerates electrons to near the speed of light around a 200-yard circular ring, producing intense ultraviolet and X-ray beams that are distributed across 40 beamlines. Its control system manages more than 230,000 process variables, and even short interruptions can halt experiments across the facility. “It’s really important for such a machine to be up, and when we go down, there are 40 beamlines that do X-ray experiments, and they are waiting,” Hellert said.

To cope with that complexity, the Accelerator Assistant is driven by a large language model running on an NVIDIA H100 GPU. Requests are routed through multiple frontier models, including Gemini, Claude and ChatGPT, while the agent writes Python code and operates either autonomously or with a human operator supervising decisions.

Operators interact with the system through a command line interface or an Open WebUI. The team said the agent uses Osprey to apply agent-based AI safely, with inference running locally via Ollama on an H100 node inside the control room network, or externally through the CBorg gateway. Tight integration with EPICS, the standard control system for accelerators, and Jupyter Notebook allows the agent to generate and execute Python code that analyses data, visualises results and directly interfaces with accelerator hardware.

The work demonstrates that AI agents can move beyond analysis and into active control roles in highly sensitive scientific environments, provided safeguards and human oversight are in place. In this case, the researchers showed the agent could manage complex, multistep experimental procedures that would normally require extensive preparation by expert operators.

Berkeley Lab said the approach is now being expanded across US particle accelerator facilities under the Department of Energy’s Genesys mission. The team has also begun collaborations with international projects, including ITER and the Extremely Large Telescope, suggesting the same agent-based techniques could be applied to other large, data-intensive scientific instruments.

If those efforts succeed, AI agents like the Accelerator Assistant could become a standard layer in future research facilities, helping scientists spend less time configuring machines and more time interpreting results.

The Recap

• AI agent deployed to support experiments at ALS accelerator.
• System uses NVIDIA H100 and routes requests through external models.
• Team reports setup time reduced by one hundredfold.