Models & Labs · United States

Berkeley Lab: Artificial Intelligence for Particle Accelerators Gets Another Boost

Credited to HPCwire · hpcwire.com · 13h ago

Oct. 8, 2026 — The U.S. Department of Energy (DOE) today announced significant funding to scale up promising projects that use Artificial Intelligence (AI), commonly referred to as artificial intelligence, to accelerate discovery. Lawrence Berkeley National Laboratory (Berkeley Lab) will lead the Multi-Office Accelerator Team Core Project (MOAT-Core), which applies AI to better design and run the particle accelerators that advance research in physics, chemistry, biology, and materials science.

The AI accelerator assistant Osprey was first deployed at Berkeley Lab’s Advanced Light Source. MOAT transformed Osprey into a multi-facility platform and expanded it to eight facilities at seven National Labs. Photo credit: Marilyn Sargent/Berkeley Lab.

MOAT-Core is part of the Genesis Mission, a national initiative to mobilize government, industry, academia, nonprofits, and international partners to advance AI for science and technology. Using the DOE-built American Science and Security Platform, partners work together on shared infrastructure that connects researchers with data, compute, and AI tools to accelerate breakthroughs. Partners are focusing on critical challenges in scientific discovery, energy, national security, health, space, and more.

“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, lead of MOAT-Core and head of the Advanced Modeling Program in Berkeley Lab’s Accelerator Technology & Applied Physics (ATAP) Division.

The goal of the Phase II Genesis Mission awards is to expand the impact of projects that have already shown potential for AI advantage and demonstrated a trajectory toward a transformative scientific capability. MOAT’s award grows the project to 16 collaborating institutions and will span multiple years.

As a prototype, MOAT showed that their AI “accelerator assistant,” Osprey, could adapt to work at different facilities and optimize how particle accelerators run. Accelerator operators can use natural language to set up experiments, fine-tune settings for the particle beam, troubleshoot equipment issues, and record and synthesize incredible amounts of data.

“It’s a big software engineering task to develop a platform that works at multiple facilities, where there are different constraints and different instruments,” said Thorsten Hellert, an ATAP staff scientist and part of the Accelerator Physics Program at the Advanced Light Source (ALS), where Osprey was first deployed. “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. I think Osprey will eventually be a platform that we use at dozens of facilities worldwide.”

In the next phase, MOAT-Core will deploy Osprey’s capabilities at additional accelerators across the DOE complex and expand the tool to connect to digital twins and assist in accelerator design. By applying all of the information the AI learns at real machines, researchers will be able to quickly run more accurate simulations, plan future experiments, and design the next generation of particle accelerators — a process that currently takes years of computations and testing.

“You’re choosing from magnets and components of different types, and strengths, and lengths, and positions,” Hellert said. “There are an infinite number of possibilities, and only a handful of those are actually good solutions. 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.”

Osprey is designed to improve the performance of large accelerators and light sources, which account for half of DOE’s User Facilities and cover a broad range of science. For example, the AI tool will help run and optimize the ALS-U, the state-of-the-art upgrade underway at the Advanced Light Source. Researchers will use the enhanced ALS to study molecules and materials with atomic precision and to develop new technologies in microelectronics, energy storage, pharmaceuticals, quantum computing, and more.

AI tools like Osprey could also be important for running smaller accelerators with growing uses in industry, such as making microchips with EUV lithography. “There’s a lot of benefit to industry from these kinds of tools that are developed at the National Labs,” Hellert said.

The MOAT-Core collaboration includes: Argonne, Berkeley, Brookhaven, Fermi, Thomas Jefferson, Oak Ridge, Pacific Northwest, and SLAC National Laboratories; Cornell University; Michigan State University, Facility for Rare Isotope Beams; Old Dominion University; the University of Chicago; and industry partners.

Berkeley Lab researchers are partners on five additional projects, pending final agreements, that received Phase II Genesis Mission funding awards today: Accelerating eXtreme Environment Specs-to-Silicon (AXESS), Artificial Intelligence for High Performance Computing (AI4HPC), Application-Aware Error Correction Codesign for Scientific Quantum Computing (ASQC), Lattice QCD at the Intelligence Frontier, and MAESTRO (Multi-agent AI Expert for Subsurface Reasoning and Optimization).

DOE also announced additional Phase I Genesis Mission awards. Berkeley Lab researchers will partner on two projects, pending final agreements: From Beam Loss to Beam Intelligence: Adaptive Generative AI for Autonomous Accelerator Facilities, and Closed-Loop Autonomous Discovery of Mechanistic Activity-Selectivity-Stability Rules in Catalysis. Berkeley Lab continues to lead 13 previously announced Phase I Genesis Mission projects and collaborates as a partner on more than 30 others.

Models & LabsUnited States

The Fallout From OpenAI’s Revenue Surprise

Credited to The New York Times$ · nytimes.com4h ago

Useful

News that the artificial intelligence giant’s sales weren’t growing as fast as investors had expected added to concerns about the A.I. boom.

View the article · $The New York Times

Front Method

Models & Labs · United States

Microsoft and Anthropic play invoice tennis with startup's $17,600 Claude bill

Credited to The Register · theregister.com · 6h ago

Photograph · The Register

Vegalabs says $21,168 in unused Azure credits expired as separate Marketplace charges piled up

A Norwegian startup says it ran up a $17,600 bill using Claude in Microsoft Foundry after assuming its Azure credits would cover the cost.

We reported a similar case in March, involving third-party model charges that weren't covered by Azure sponsorship credits.

This time, the customer says Microsoft and Anthropic each directed it to the other to resolve the charges.

Vegalabs AS says it received $25,060 in Azure credits through Microsoft for Startups and deployed Claude in August, expecting its analysis workloads to use that balance.

However, Microsoft's sponsorship rules exclude Anthropic models purchased through Azure Marketplace. Vegalabs says Microsoft attempted to collect $16,500 from its card, with the invoiced total subsequently reaching $17,600 before tax.

Vegalabs says it monitored the sponsorship portal, which continued to show available credits while separate Marketplace charges accumulated. Microsoft's Claude deployment guide says credit-only sponsored subscriptions aren't supported, but accounts with a card on file will have that card charged instead.

The company says it deleted the deployment within an hour of discovering the charges. Its card issuer declined the attempted payment over suspected fraud, and $21,168 in unused sponsorship credits expired on September 8.

Vegalabs acknowledges it did not read the credit exclusions or configure a budget alert. It disputes neither the usage nor the documented rule, but says the deployment interface failed to make the separate billing clear. Its request for a waiver then bounced between Microsoft and Anthropic support.

The Register has seen communications from Microsoft directing the user to Anthropic. "Marketplace transactions are processed through a separate billing pipeline associated with the third-party publisher. As a result, Azure Support does not have the ability to directly waive, refund, or adjust these Marketplace charges on the customer's behalf."

"If Anthropic approves the request, they can initiate the appropriate refund authorization through Microsoft."

Anthropic support, however, said Microsoft did not require its authorization to process a refund and directed Vegalabs back to Microsoft.

Vegalabs says four of the seven support replies it received were AI-generated, including both responses from Anthropic. It says those responses incorrectly described its account as an "Azure Student Sponsorship."

We asked Microsoft on October 6 if it had any comment on the situation, but beyond an acknowledgment and a request to let us know what Anthropic had to say, the company had nothing further to share. We also asked Anthropic for its take on the situation, but the company did not respond.

Microsoft's rules distinguish between credit-covered services and separately billed models offered through the same platform. Vegalabs' experience suggests that distinction needs to be clearer where customers deploy them.

For now, the company has a bill and two support teams directing it elsewhere. ®