Tuning a particle accelerator is no easy task. To achieve the necessary beam, physicists spend hours adjusting parameters, searching for optimal values for dozens of variables. It is like trying to tune a musical instrument with a hundred strings, where each one affects the sound of the others.
Berkeley Lab has simplified this process with the help of AI. Their Advanced Light Source — a major X-ray facility for research — now employs an AI copilot that helps scientists tune equipment and conduct experiments faster.
Why Accelerators Need a Helper
The Advanced Light Source is a synchrotron, a facility the size of a football field that accelerates electrons to nearly the speed of light. When these electrons pass through special magnets, they emit bright X-ray light. This light is used to study materials at the atomic level — from new batteries to proteins.
The issue is that every experiment requires its own setup. Scientists have to manually adjust beam parameters, and this takes a lot of time. It is especially challenging for beginners — those just starting to work with such facilities.
The Berkeley team created a system that automates this process. The AI copilot analyzes data in real-time and suggests which parameters to change to achieve the desired result faster.
What the System Can Do
The AI copilot is built on large language models — the same technology behind chatbots like ChatGPT. However, here, the model was trained to work not with text, but with physical accelerator data.
The system can perform several tasks. First, it translates scientists' requests into command language for the equipment. You can write in plain language: «Increase beam energy by 10%», and the copilot will determine which parameters need to be adjusted.
Second, the copilot automatically optimizes settings. If you need a beam with specific characteristics, the system will select the necessary values itself. Previously, this took hours; now it takes minutes.
Third, the system helps diagnose problems. If something goes wrong, the AI analyzes the data and suggests where to look for the cause.
How It Works in Practice
At the core of the copilot are NVIDIA GPU accelerators that process data in real-time. The accelerator generates a vast amount of information: sensor readings, beam parameters, and diagnostic data. All of this needs to be analyzed quickly to make a decision.
Data collected over years of accelerator operation was used to train the system. The AI learned how experienced operators adjust the equipment and learned to mimic their actions. But not just copy — the model can adapt to new situations and offer solutions that are not in the training data.
An interesting detail: the copilot works as an advisor, not as an autopilot. The final decision is always made by a human. The system offers options and explains its logic but does not interfere directly with accelerator control. This is crucial for safety — a particle accelerator is, after all, a complex device where a mistake can be costly.
Who This Helps
The primary goal of the project is to make accelerators more accessible. Currently, working effectively with such facilities requires extensive experience. New users have to spend a long time understanding the nuances of tuning, and this takes time away from the research itself.
With the AI copilot, the barrier to entry is lowered. Scientists can focus on their experiments rather than the technical details of managing the accelerator. This is especially beneficial for small research groups that do not have their own accelerator specialists.
Furthermore, the system speeds up work even for experienced operators. Routine tasks are automated, freeing up time for more complex experiments. According to the Berkeley team's estimates, the copilot can reduce tuning time by 30–50%.
What's Next
Currently, the system is operating in test mode on one beamline at the Advanced Light Source. The next step is to expand it to other sections of the accelerator and add new functions. For example, the team plans to teach the copilot to predict when equipment will need maintenance.
Interestingly, a similar approach can be applied to other types of scientific equipment. Any complex facility where there are many parameters and precise tuning is needed could potentially get its own AI assistant. We're talking not only about accelerators but also telescopes, reactors, and large experimental installations.
The project demonstrates how AI is changing the very organization of scientific work. Previously, interacting with equipment required specialized knowledge and experience. Now, using language models, one can communicate with technology almost like with a colleague — in everyday language, without needing to remember all the technical details.
Of course, questions remain. How reliable is the system in non-standard situations? How will it behave during a serious failure? Will it really be able to replace years of human experience, or will it remain just a convenient tool? Answers will emerge as the copilot works in real-world conditions.
But it's already clear: AI doesn't just help analyze data after the experiment. It's becoming an active participant in the research process itself, making complex equipment more accessible and efficient.