Within the rolling hills of Berkeley, California, an AI agent is supporting high-stakes physics experiments on the Superior Mild Supply (ALS) particle accelerator.
Researchers on the Lawrence Berkeley Nationwide Laboratory ALS facility just lately deployed the Accelerator Assistant, a big language mannequin (LLM)-driven system to maintain X-ray analysis on monitor.
The Accelerator Assistant — powered by an NVIDIA H100 GPU harnessing CUDA for accelerated inference — faucets into institutional data information from the ALS help staff and routes requests by Gemini, Claude or ChatGPT. It writes Python and solves issues, both autonomously or with a human within the loop.
That is no small process. The ALS particle accelerator sends electrons touring close to the pace of sunshine in a 200-yard round path, emitting ultraviolet and X-ray mild, which is directed by 40 beamlines for 1,700 scientific experiments per 12 months. Scientists worldwide use this course of to review supplies science, biology, chemistry, physics and environmental science.
On the ALS, beam interruptions can final minutes, hours or days, relying on the complexity, halting concurrent scientific experiments in course of. And far can go unsuitable: the ALS management system has greater than 230,000 course of variables.
“It’s actually essential for such a machine to be up, and once we go down, there are 40 beamlines that do X-ray experiments, and they’re ready,” mentioned Thorsten Hellert, workers scientist from the Accelerator Expertise and Utilized Physics Division at Berkeley Lab and lead writer of a analysis paper on the groundbreaking work.
Till now, facility workers troubleshooting points have needed to rapidly determine the areas, retrieve information and collect the appropriate personnel for evaluation underneath intense time stress to get the system again up and working.
“The novel method provides a blueprint for securely and transparently making use of massive language model-driven programs to particle accelerators, nuclear and fusion reactor services, and different complicated scientific infrastructures,” mentioned Hellert.
The analysis staff demonstrated that the Accelerator Assistant can autonomously put together and run a multistage physics experiment, slicing setup time and lowering efforts by 100x.
Making use of Context Engineering Prompts to Accelerator Assistant
The ALS operators work together with the system by both a command line interface or Open WebUI, which allows interplay with varied LLMs and is accessible from management room stations, in addition to remotely. Below the hood, the system makes use of Osprey, a framework developed at Berkeley Lab to use agent-based AI safely in complicated management programs.
Every consumer is authenticated and the framework maintains customized context and reminiscence throughout classes, and a number of classes may be managed concurrently. This enables customers to arrange distinct duties or experiments into separate threads. These inputs are routed by the Accelerator Assistant, which makes connections to the database of greater than 230,000 course of variables, a historic database archive service and Jupyter Pocket book-based execution environments.
“We attempt to engineer the context of each language mannequin name with no matter prior data we’ve from this execution up so far,” mentioned Hellert.
Inference is completed both domestically — utilizing Ollama, which is an open-source device for working LLMs with a private laptop, on an H100 GPU node positioned inside the management room community — or externally with the CBorg gateway, which is a lab-managed interface that routes requests to exterior instruments equivalent to ChatGPT, Claude or Gemini.
The hybrid structure balances safe, low-latency, on-premises inference with entry to the newest basis fashions. Integration with EPICS (Experimental Physics and Industrial Management System) allows operator-standard security constraints for direct interplay with accelerator {hardware}. EPICS is a distributed management system utilized in large-scale scientific services equivalent to particle accelerators. Engineers can write Python code in Jupyter Pocket book that may talk with it.
Principally, conversational enter is changed into a transparent pure language process description for targets with out redundancy. Exterior data equivalent to customized reminiscence tied to customers, documentation and accelerator databases are built-in to help with terminology and context.
“It’s a big facility with a whole lot of specialised experience,” mentioned Hellert. “A lot of that data is scattered throughout groups, so even discovering one thing easy — just like the handle of a temperature sensor in a single a part of the machine — can take time.”
Tapping Accelerator Assistant to Help Engineers, Fusion Vitality Growth
Utilizing the Accelerator Assistant, engineers can begin with a easy immediate describing their aim. Behind the scenes, the system attracts on rigorously ready examples and key phrases from accelerator operations to information the LLM’s reasoning.
“Every immediate is engineered with related context from our facility, so the mannequin already is aware of what sort of process it’s coping with,” mentioned Hellert.
Every agent is an skilled in that area, he mentioned.
As soon as the duty is outlined, the agent brings collectively its specialised capabilities — equivalent to discovering course of variables or navigating the management system — and might routinely generate and run Python scripts to research information, visualize outcomes or work together safely with the accelerator itself.
“That is one thing that may prevent critical time — within the paper, we are saying two orders of magnitude for such a immediate,” mentioned Hellert.
Wanting forward, Hellert goals to have the ALS engineers put collectively a wiki that paperwork the numerous processes that go on to help the experiments. These paperwork might assist the brokers run the services autonomously — with a human within the loop to approve the plan of action.
“On these high-stakes scientific experiments, even when it’s only a TEM microscope or one thing that may value $1 million, a human within the loop may be crucial,” mentioned Hellert.
The work has already expanded past ALS as a part of the DOE’s Genesys mission, with the framework being deployed throughout U.S. particle accelerator services. Subsequent up, Hellert simply started collaborating with engineers on the ITER fusion reactor — the world’s largest — in France for implementing the framework to be used within the fusion reactor facility. He additionally has a collaboration within the works with the Extraordinarily Massive Telescope ELT, in northern Chile.
Benefiting Humanity: Scientific Impression of Experiments Supported by ALS
Past optimizing the accelerator and different industrial operations, the work on the ALS straight allows scientific breakthroughs with international impression. The ability’s steady X-ray beams underpin analysis in well being, local weather resilience and planetary science.
Through the COVID-19 pandemic, ALS researchers helped characterize a uncommon antibody that would neutralize SARS-CoV-2. Structural biology experiments at Beamline 4.2.2 revealed how six molecular loops of the antibody latch onto and disable the viral spike protein. The findings supported the fast improvement of a therapeutic that remained efficient by a number of variants.
ALS science additionally contributes to climate-focused analysis. Steel-organic frameworks (MOFs) — a category of porous supplies able to capturing water or carbon dioxide from air — have been extensively studied throughout a number of ALS beamlines. These experiments supported foundational work that in the end led to the 2025 Nobel Prize in Chemistry, recognizing the transformative potential of MOFs for sustainable water harvesting and carbon administration.
In planetary science, ALS measurements of samples returned from NASA’s OSIRIS-REx mission helped hint the chemical historical past of asteroid Bennu. X-ray analyses supplied proof that such asteroids carried water and molecular precursors of life to early Earth, deepening our understanding of the origins of the planet’s liveable circumstances.
