DOE Expands AI Assistant That Runs Particle Accelerators to 16 Institutions

If you've ever tried to get two labs to share research software, you know the usual outcome: everyone nods, then goes back to their own stack. That's basically been the norm in particle accelerator facilities for decades, since every machine is a custom build with its own quirks, instruments, and failure modes. So it's worth pausing on what the Multi-Office Accelerator Team (MOAT) just pulled off: an AI assistant that actually plugs into a new facility and works, without months of bespoke integration.

That assistant is called Osprey, and the Department of Energy just handed it a serious vote of confidence. DOE's Genesis Mission, a national initiative pushing AI ("Super Intelligence," in DOE's branding) into scientific discovery, announced Phase II funding for MOAT-Core, led by Lawrence Berkeley National Laboratory. The award scales Osprey's deployment to additional facilities and pushes its scope beyond day-to-day operations into accelerator design and digital twin simulation.

For context: Osprey first went live at Berkeley Lab's Advanced Light Source (ALS), a synchrotron light source used for materials and chemistry research. MOAT then generalized it into a multi-facility platform, expanding to eight facilities across seven National Labs. Phase II grows the collaboration to 16 institutions, including Argonne, Brookhaven, Fermi, Thomas Jefferson, Oak Ridge, Pacific Northwest, and SLAC National Laboratories, plus Cornell, Michigan State's Facility for Rare Isotope Beams, Old Dominion University, the University of Chicago, and industry partners.

"The Genesis Mission offers a unique opportunity to leverage the rapidly expanding power of agentic assistants and revolutionize the way that we design and operate particle accelerators," said Jean-Luc Vay, MOAT-Core lead and head of the Advanced Modeling Program in Berkeley Lab's Accelerator Technology & Applied Physics Division.

What Osprey Actually Does

Think of Osprey as a natural-language layer sitting on top of accelerator control systems. Operators can use it to set up experiments, tune beam parameters, troubleshoot hardware faults, and pull together the massive data streams that come off these machines during a run. That's not trivial. Accelerators generate huge volumes of telemetry, and sorting signal from noise during a live operational issue usually eats up an expert's attention for hours.

The harder engineering problem, according to Thorsten Hellert, an ATAP staff scientist on the ALS Accelerator Physics Program, wasn't the AI itself. It was making the thing portable.

"It's a big software engineering task to develop a platform that works at multiple facilities, where there are different constraints and different instruments," Hellert said. "Because accelerators are custom machines, sharing software is not the default in our community. But this project got us focused on building one thing together at a scale that has never happened before, and there are a lot of benefits from it. You can plug Osprey into a new facility and it works."

That portability is the real headline here, even more than the funding number. A platform that generalizes across heterogeneous, custom-built scientific hardware is a genuinely hard distributed systems problem, not just a model fine-tuning exercise. Hellert thinks Osprey could eventually scale to dozens of facilities worldwide.

In Phase II, MOAT-Core is pushing Osprey past operations and into design work. That means connecting it to digital twins, which are simulation models that mirror a physical accelerator closely enough to test changes virtually before touching the real machine. The pitch is straightforward: feed the system everything it's learned from operating real accelerators, then use that knowledge to run faster, more accurate simulations for planning future experiments and designing next-generation machines.

That matters because accelerator design right now is brutally slow. Engineers choose from a huge combinatorial space of magnets, components, strengths, lengths, and positions, trying to find configurations that actually work. "There are an infinite number of possibilities, and only a handful of those are actually good solutions," Hellert said. "The question of how to find them is a mathematically very challenging problem, but these coding agents and agentic assistants are made for speedrunning these kinds of workflows." That's a combinatorial optimization problem dressed up in physics, and it's exactly the kind of search-heavy task where agentic systems, ones that can iterate, test, and self-correct across many candidate solutions, tend to earn their keep.

The stakes aren't abstract. Large accelerators and light sources make up roughly half of DOE's User Facilities, supporting research across physics, chemistry, biology, and materials science. Osprey is slated to help run and optimize ALS-U, the major upgrade currently underway at the Advanced Light Source, which will let researchers study materials at atomic precision for applications in microelectronics, energy storage, pharmaceuticals, and quantum computing. Hellert also flagged a commercial angle: smaller accelerators used in industry, like those powering EUV lithography for chipmaking, could benefit from the same tooling developed at the National Labs.

MOAT-Core isn't an isolated bet. Berkeley Lab researchers are also partners on five other Phase II-funded projects announced the same day, spanning extreme-environment hardware design (AXESS), AI for high-performance computing (AI4HPC), error correction for scientific quantum computing (ASQC), lattice QCD, and subsurface reasoning for geoscience applications (MAESTRO). Two new Phase I awards add accelerator-focused generative AI and autonomous catalysis discovery to the portfolio. Altogether, Berkeley Lab now leads 13 Phase I Genesis Mission projects and partners on more than 30 others, a sign of how broadly DOE is spreading its AI-for-science bets rather than concentrating them in one shot.

Key Takeaways

• Osprey, MOAT's AI accelerator assistant, started at Berkeley Lab's ALS and now runs across eight facilities at seven National Labs.
• Phase II Genesis Mission funding expands the collaboration to 16 institutions, including academic and industry partners, and extends Osprey from operations into accelerator design and digital twin simulation.
• The core engineering win isn't a smarter model, it's portability: a single platform that plugs into heterogeneous, custom-built accelerator hardware without facility-specific rebuilds.
• Agentic AI is being aimed at the combinatorial search problem of accelerator component design, a space with effectively infinite configurations and very few good solutions.
• Osprey will help optimize ALS-U, the Advanced Light Source's major upgrade, and could eventually extend to industrial accelerators used in chip lithography.