‘AI for Science’ Initiative at DOE Gains Traction

The Department of Energy has begun to flesh out its aspirations to tailor artificial intelligence models to scientific applications. The push is part of a DOE-wide initiative called Frontiers in AI for Science, Security, and Technology (FASST), formally announced last month.

“Imagine we had a basic science AI foundational model like ChatGPT for English — but it speaks physics and chemistry,” said Deputy Energy Secretary David Turk at an AI expo in Washington, D.C.

DOE so far has released few details about the initiative, but one of its architects elaborated on its aims in a presentation last week to the department’s Advanced Scientific Computing Advisory Committee. Rick Stevens, head of the computing directorate at Argonne National Lab, offered a “sketch of a plan” for the initiative, which he envisioned could involve adding thousands of staff to DOE’s ranks.

American companies currently have the global edge in commercial AI, Stevens said, but the federal government is “proportionately under-investing in noncommercial, nondefense AI.” Stevens also argued that leading AI companies lack the resources to train their consumer-facing models on the extremely large amounts of scientific data DOE generates and are therefore ill-equipped to solve such scientific problems.

“The private sector is not going to do the lifting for science. We talked to them — they’re interested in it; they can’t afford to do it,” he said.

The global race to be on the leading edge of AI development is also front of mind. “There’s massive competition between what I call the western democracies including Japan [and other] friends; the semi-aligned petrostates, which we can’t figure out exactly what they’re doing but they’re going to put a ton of money into AI; and then the adversarial Sino-Russian players,” he said.

FASST will focus on developing AI foundation models to accelerate work across DOE mission areas. These models would be different from the more general-purpose AI dominating the commercial space because they would be trained to support a specific field. Stevens gave 11 examples of potential foundation models that together could span DOE’s portfolio, from particle physics to electric grid infrastructure. He suggested DOE might also produce a higher-level AI model that could combine output from multiple foundation models for cross-disciplinary research.

Stevens noted the AI executive order issued in late 2023 gives DOE a mandate to pursue FASST’s objectives but only using existing appropriations. He did not provide a cost estimate for the initiative but said it would “fit” within the $32 billion per year budget for non-defense AI research proposed in the bipartisan AI roadmap released last month by Senate Majority Leader Chuck Schumer (D-NY). So far DOE has requested $455 million for AI-related projects in fiscal year 2025, a 53% increase over the amount it spent during fiscal year 2023.

Stevens estimated that DOE would need to add “on the order of 2,000 more staff” to realize the full potential of the FASST initiative, with each foundation model requiring around 100 people to develop and evaluate. He also pointed out that DOE would need to reorganize its vast troves of scientific data into workable training formats.

Building the FASST models could “completely consume” both of the exascale computers operated by the DOE Office of Science, Stevens said, suggesting that it would need to build more supercomputers if those resources are to remain available for other research. He also estimated DOE would need to build the capacity to support “something like 200,000” researchers using the AI models each day once they are built, based on the current staff size of the national lab system and its user base.

As to why DOE ought to lead the initiative, Stevens argued the department is uniquely suited to develop tailored AI models because it has a large and interdisciplinary scientific workforce, sophisticated supercomputers, and an incredible amount of training data at its facilities.

“We have all the ingredients to make huge progress in AI, we just don’t have the mandate or didn’t have the mandate until recently,” Stevens said.