In the rolling hills of Berkeley, California, an AI agent is supporting high-stakes physics experiments at the Advanced Light Source (ALS) particle accelerator.
Researchers at the Lawrence Berkeley National Laboratory ALS facility recently deployed the Accelerator Assistant, a large language model (LLM)-driven system to keep X-ray research on track.
The Accelerator Assistant — powered by an NVIDIA H100 GPU harnessing CUDA for accelerated inference — taps into institutional knowledge data from the ALS support team and routes requests through Gemini, Claude or ChatGPT. It writes Python and solves problems, either autonomously or with a human in the loop.
This is no small task. The ALS particle accelerator sends electrons traveling near the speed of light in a 200-yard circular path, emitting ultraviolet and X-ray light, which is directed through 40 beamlines for 1,700 scientific experiments per year. Scientists worldwide use this process to study materials science, biology, chemistry, physics and environmental science.
At the ALS, beam interruptions can last minutes, hours or days, depending on the complexity, halting concurrent scientific experiments in process. And much can go wrong: the ALS control system has more than 230,000 process variables.
“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,” said Thorsten Hellert, staff scientist from the Accelerator Technology and Applied Physics Division at Berkeley Lab and lead author of a research paper on the groundbreaking work.
Until now, facility staff troubleshooting issues have had to quickly identify the areas, retrieve data and gather the right personnel for analysis under intense time pressure to get the system back up and running.
“The novel approach offers a blueprint for securely and transparently applying large language model-driven systems to particle accelerators, nuclear and fusion reactor facilities, and other complex scientific infrastructures,” said Hellert.
The research team demonstrated that the Accelerator Assistant can autonomously prepare and run a multistage physics experiment, cutting setup time and reducing efforts by 100x.
The ALS operators interact with the system through either a command line interface or Open WebUI, which enables interaction with various LLMs and is accessible from control room stations, as well as remotely. Under the hood, the system uses Osprey, a framework developed at Berkeley Lab to apply agent-based AI safely in complex control systems.
Each user is authenticated and the framework maintains personalized context and memory across sessions, and multiple sessions can be managed simultaneously. This allows users to organize distinct tasks or experiments into separate threads. These inputs are routed through the Accelerator Assistant, which makes connections to the database of more than 230,000 process variables, a historical database archive service and Jupyter Notebook-based execution environments.
“We try to engineer the context of every language model call with whatever prior knowledge we have from this execution up to this point,” said Hellert.
Inference is done either locally — using Ollama, which is an open-source tool for running LLMs with a personal computer, on an H100 GPU node located within the control room network — or externally with the CBorg gateway, which is a lab-managed interface that routes requests to external tools such as ChatGPT, Claude or Gemini.
The hybrid architecture balances secure, low-latency, on-premises inference with access to the latest foundation models. Integration with EPICS (Experimental Physics and Industrial Control System) enables operator-standard safety constraints for direct interaction with accelerator hardware. EPICS is a distributed control system used in large-scale scientific facilities such as particle accelerators. Engineers can write Python code in Jupyter Notebook that can communicate with it.
Basically, conversational input is turned into a clear natural language task description for objectives without redundancy. External knowledge such as personalized memory tied to users, documentation and accelerator databases are integrated to assist with terminology and context.
“It’s a large facility with a lot of specialized expertise,” said Hellert. “Much of that knowledge is scattered across teams, so even finding something simple — like the address of a temperature sensor in one part of the machine — can take time.”
Using the Accelerator Assistant, engineers can start with a simple prompt describing their goal. Behind the scenes, the system draws on carefully prepared examples and keywords from accelerator operations to guide the LLM’s reasoning.
“Each prompt is engineered with relevant context from our facility, so the model already knows what kind of task it’s dealing with,” said Hellert.
Each agent is an expert in that field, he said.
Key considerations
- Investor positioning can change fast
- Volatility remains possible near catalysts
- Macro rates and liquidity can dominate flows
Reference reading
- https://blogs.nvidia.com/blog/ai-copilot-berkeley-x-ray-particle-accelerator/#content
- https://www.nvidia.com/en-us/
- https://blogs.nvidia.com/?s=
- NVIDIA Rubin Platform, Open Models, Autonomous Driving: NVIDIA Presents Blueprint for the Future at CES
- Japan Science and Technology Agency Develops NVIDIA-Powered Moonshot Robot for Elderly Care
- Admins finally get the power to uninstall Microsoft Copilot on Windows 11 Pro, Enterprise, and EDU versions — devices must meet specific conditions to allow the
- Jensen Huang discusses the economics of inference, power delivery, and more at CES 2026 press Q&A session — 'You sell a chip one time, but when you build softwa
- SanDisk to double price of 3D NAND for enterprise SSDs in Q1 2026 — hyperscalers to pay top dollar for storage as AI continues to roll
Informational only. No financial advice. Do your own research.