The Multi-Office Accelerator Team (MOAT) Seed project
The MOAT Seed project introduces the Osprey agentic AI platform to unify fragmented accelerator facility knowledge and automate complex tasks like troubleshooting, analysis, and design across the U.S. Department of Energy complex, thereby reducing reliance on custom scripting and enabling the transfer of shared AI capabilities between national laboratories.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Particle accelerators are among the most complex machines ever built by humans. They are vast, intricate networks of magnets, vacuum chambers, and electronic sensors that guide beams of subatomic particles at nearly the speed of light. These machines serve as powerful microscopes for scientists, allowing them to probe the fundamental building blocks of matter, create new materials, and develop advanced medical treatments. However, keeping these massive engines running smoothly is a monumental task. The machines generate millions of data points every second, and their behavior changes constantly as conditions shift. For decades, the people who operate them have relied on a combination of deep expertise, manual data checks, and custom-written computer scripts to diagnose problems, tune the beam, and keep experiments running. This process is slow, labor-intensive, and often limited to what a single expert can manage at one specific facility.
The challenge is not just the sheer volume of information, but the difficulty of connecting it. A problem might be visible in a logbook entry written by an operator hours ago, hidden in a stream of archived sensor data, or revealed only by a complex simulation that takes hours to run. Traditionally, a human operator had to act as the bridge, manually pulling data from different sources, writing code to analyze it, and then deciding what to do next. This paper describes a new approach to this problem, developed by a team of scientists from seven major national laboratories in the United States. They have built a shared software platform called Osprey, designed to act as an intelligent assistant that can read natural language, access all these different data sources, and perform complex analyses automatically. Instead of replacing human experts, the system is designed to handle the heavy lifting of data retrieval and calculation, allowing the experts to focus on the bigger picture.
The core of this work is the concept of an "agentic" artificial intelligence. Unlike a simple chatbot that answers questions, an agent is a software program that can plan a sequence of actions to achieve a goal. In this project, the researchers connected these advanced reasoning models to the actual control systems of particle accelerators. The system was tested across a wide variety of facilities, including electron storage rings, proton accelerators, and laser-plasma devices. The goal was to see if a single, shared software framework could learn to speak the language of different machines, retrieve historical records, analyze live data, and even write the code needed to fix a problem or tune a beam.
The results show that this approach works. In one demonstration at the Jefferson Lab, an operator noticed a strange signal on a beam monitor that did not match the actual movement of the particle beam. Instead of spending hours manually inspecting the data, the agent was asked to investigate. It automatically retrieved the signal, broke it down into different parts, identified the oscillating sections, and calculated the frequency of the disturbance. It then presented a clear, annotated graph and a summary of the findings to the operator in a fraction of the time it would have taken a human to do the same work manually. In another instance at the Advanced Photon Source, the system analyzed thousands of archived data points to find patterns in equipment failures. It successfully identified a specific power supply that was behaving differently from the others, a subtle issue that might have been missed during a routine check.
The system also proved capable of handling more complex, multi-step tasks. At the Advanced Light Source, researchers asked the agent to analyze why a specific part of the machine was behaving erratically. The agent did not just look at the numbers; it also searched the electronic logbooks, which are like digital diaries where operators write down what they were doing and what they observed. By combining the raw data with the human notes, the agent discovered that a vacuum alarm had triggered a chain reaction, causing the beam to jump. It explained the entire sequence of events, linking the alarm to the beam movement, something that would have required a human to cross-reference two different systems manually.
Beyond just looking at data, the agent was tested on its ability to control the machine. At the Jefferson Lab, the system was asked to stabilize a wobbling beam signal. It wrote its own control code, identified the correct magnets to adjust, and began making real-time corrections to keep the beam steady. It adjusted its own settings as it went, learning what worked and what did not, all without a human writing a single line of the control script. Similarly, at the SLAC National Accelerator Laboratory, the agent helped tune a laser-driven accelerator. It analyzed past performance data, selected the best algorithms for the job, and configured a complex optimization routine. This process, which usually takes an expert a significant amount of time to set up, was completed in minutes, and the resulting beam quality was significantly improved.
The project also explored how these agents could help design new accelerators. In one test, the system was given a task to design a short section of a particle beamline from scratch. The goal was to transport a beam from one part of a machine to another while keeping it focused and stable. The agent was given the rules of physics and the specifications of the machine, and it was free to decide how many magnets to use and where to place them. It wrote its own simulation scripts, ran thousands of tests, and iterated on its design until it found a solution that matched the performance of a design created by human experts. This entire process took only about fifteen minutes, a task that would typically take days of manual work and simulation.
The researchers emphasize that this is not about replacing the human operators or physicists. The system is designed to work under strict safety rules, where a human must approve any changes to the machine before they are made. The agent acts as a powerful tool that can handle the tedious, repetitive, and data-heavy parts of the job. It can search through years of records in seconds, spot patterns that are invisible to the human eye, and run complex calculations that would otherwise be too slow to be practical. By connecting these capabilities across different laboratories, the team showed that a solution developed at one facility can be adapted and used at another, breaking down the barriers that have kept these tools isolated in the past.
The paper concludes that this shared approach represents a significant shift in how accelerator science is conducted. It moves the field away from isolated, facility-specific experiments toward a collaborative ecosystem where tools and knowledge are shared. The initial results suggest that with this kind of assistance, the time required to tune a machine could be reduced dramatically, and the quality of the beams could be improved. While there are still challenges to overcome, particularly in ensuring the reliability and safety of these autonomous systems, the work demonstrates a clear path forward. It shows that by combining advanced artificial intelligence with the deep expertise of human scientists, it is possible to unlock new levels of performance and efficiency in the world's most complex machines.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.