Editor’s note: This article is adapted from an 11 August post on FYI, which reports on federal science policy. Both FYI and Physics Today are published by the American Institute of Physics.
As lawmakers scramble to respond to the emergence of artificial intelligence, the Department of Energy is making the case for itself to assume a leading role in developing AI tools for fundamental research, energy technology development, and national security.
Although DOE itself has not taken an official position on a potential AI initiative, staff from DOE national labs have presented an initial vision in the form of a 200-page report that was published this summer. DOE staff have drawn from the report in congressional briefings, including a Senate-wide briefing last month organized by Senate Majority Leader Chuck Schumer (D-NY), who is preparing to advance major legislation on AI. The initiative proposed in the report is designed in part as a follow-on to DOE’s $1.8 billion Exascale Computing Project (ECP), which is about to conclude following the deployment last spring of the first US exascale computer at Oak Ridge National Laboratory and the planned deployment of additional machines at Argonne and Lawrence Livermore National Laboratories.
The report, titled “Advanced Research Directions on AI for Science, Energy, and Security,” paints a picture of AI as a potential way to dramatically accelerate the pace of R&D and enable previously impossible scientific capabilities. It also casts AI as a strategically essential technology that will have multiplicative effects and determine the leaders and followers in other technological arenas. “Establishing leadership in AI and in the underlying capabilities, including high-performance computing, will be intimately tied to the nation’s future and its role in the global order,” the report argues.
AI research potential
The national labs’ report proposes six main areas where AI could be applied across DOE’s mission: foundation models for scientific discovery and synthesis; surrogate models for scientific computing; property inference and inverse design; design, prediction, and control of complex engineered systems; autonomous discovery; and software engineering. It also details “grand challenges” associated with each area.
AI foundation models are general-purpose models that can be applied to a wide range of tasks, including the language models (such as ChatGPT) and image generators that have recently grabbed public attention. The report authors imagine creating new foundation models that are dedicated to scientific and national security problems, though they acknowledge the task is daunting enough to require a “moonshot” level of effort. They envision foundation models synthesizing research relevant to, for instance, how clouds affect Earth’s climate or how vortices evolve in fusion plasmas.
“Regardless of the specific problem being studied, a frequent challenge is the vast amount of existing knowledge that could potentially be relevant to its solution—a quantity that typically far exceeds the cognitive capacity of any one individual or even team,” the report says. “The recent and considerable successes achieved with large language models suggest that a transformative solution may be on the horizon.”
The report describes surrogate models as “simpler yet faithful” representations of complex, real-life systems, with the models themselves trained on the outputs of other computational models. The DOE authors recommend launching pilot programs to develop surrogate models of plasma turbulence and Earth’s oceans, among other systems. DOE’s Energy Exascale Earth System Model is flagged as an existing project that could be made much less computationally demanding by applying surrogate models.
AI could also be used to infer the properties of specific materials or to identify materials that might meet certain criteria, according to the report. The authors further envision using AI to help control complex machines and experiments, such as DOE’s various user facilities. For example, AI could be used to tune particle accelerators in real time or calibrate the diagnostic systems of laser experiments and nuclear reactors. Such efforts are currently time-consuming, with some relying on massive conventional computing systems.
The report authors also imagine using AI-controlled robots to help automate the process of scientific discovery, such as by autonomously manufacturing specialized materials and using AI-generated designs as a starting point for new nuclear weapons systems.
Among the overarching challenges identified is that many AI tools suffer from an inability to determine how they arrive at particular conclusions. “The ‘black box’ nature of AI models confounds our ability to validate the results, hindering adoption,” the report states.
Funding obstacles
Among the report’s lead organizers is Rick Stevens, the head of Argonne’s Computing, Environment, and Life Sciences Directorate. In the Senate-wide briefing, Stevens described a recent meeting of the advisory committee for DOE’s Advanced Scientific Computing Research (ASCR) program, which is co-leading the ECP with the National Nuclear Security Administration. “The pitch that we’re making is that only DOE and NNSA can really advance this responsible codesign of AI R&D with a strong focus on science, energy, and national security,” he said. He added, however, that the department “is probably not the right organization to be the lead on monitoring and regulation.”
Stevens explained how DOE could leverage its experience organizing large, interdisciplinary teams and running exascale computers to spearhead an even larger project focused on developing AI tools relevant across the entire department. “This is not going to be a small initiative. This could be several times or more the scale of what we have been doing with ECP,” he said.
Asked about the prospects for launching such an effort given the recent budget caps set by Congress, Stevens expressed hope that it could receive funding as early as fiscal year 2024 through a special appropriation, perhaps via a follow-on to the CHIPS and Science Act. Schumer has said he wants to develop a bill that would include direct funding for technology areas beyond semiconductors.
Under the spending caps, the House and Senate have proposed to cut ASCR’s regular annual budget by 5%, to just over $1 billion for FY 2024. “The numbers are what they are, and it remains to be seen what we end up with at the end of the day,” acting ASCR head Ceren Susut tells FYI. “I think what we can do from our point of view is to prepare for a big initiative and also respond to the congressional staffers’ questions about what we can do, what are the concerns about AI, what are the opportunities about AI, and how DOE can take a leadership role.”
Susut notes that DOE is in the early stages of planning for such an initiative and that the AI report is a major input that will be supplemented with additional ideas from stakeholders. “The case that we’re trying to make is that we have a unique role to play here,” she says.
Global competition in supercomputing
The DOE report authors emphasize that the current international race to develop AI technology is closely connected to competition to develop ever-faster supercomputers. “Progress in designing and deploying supercomputers in China, Japan, Europe, and other nations has resulted in a competitive AI position that cannot be ignored,” the report states.
That sentiment is echoed in a separate report recently published by the advisory committee for ASCR titled “Can the United States Maintain Its Leadership in High-Performance Computing?” The committee expresses deep concern about the outlook for supercomputer development in the US. Uncertainty about what will come after the ECP is “generating much anxiety” among staff members working on the project, the report states, and there is a risk that many will leave the national labs to work in the private sector.
The committee concludes that US leadership in high-performance computing (HPC) “has eroded” despite the impending completion of the three exascale machines produced by ECP. The committee writes that “it seems clear that China has at least matched US HPC capabilities” and that China reportedly plans to deploy 10 exascale machines by 2025.
