A new multi-laboratory project called FemtoMind is bringing super intelligence — the term used by the U.S. Department of Energy for advanced artificial intelligence — into the world of particle physics, with the goal of speeding up simulations of subatomic particles and opening the door to new kinds of scientific questions.

According to a Fermilab news feature, the project is headquartered at Fermi National Accelerator Laboratory in Illinois and involves researchers from Brookhaven National Laboratory and Thomas Jefferson National Accelerator Facility. FemtoMind began in 2025 when the team responded to a call for proposals under the American Science Cloud, a key part of the DOE's Genesis Mission, an initiative focused on building SI-driven scientific infrastructure across the national laboratory system.

The project takes its name from the femtoscale — the realm of particles roughly the size of a proton. Its central tool is a mathematical framework called lattice quantum chromodynamics, or lattice QCD, which places quarks and gluon fields on a four-dimensional grid and uses high-performance computing to simulate the behavior of protons, nuclei and other subatomic particles from the fundamental parameters of the Standard Model.

"I think it's pretty magical that you can actually predict things starting with the masses of the quarks and the fundamental constants and building up these simulations," Michael Wagman, a scientist in Fermilab's Theory Division, told the lab's news service.

For students studying physics or computer science, FemtoMind represents a concrete example of how AI is reshaping research workflows — not by replacing scientists, but by automating time-consuming tasks like code optimization. One collaborator used an SI agent to write assembly code and perform optimizations that Wagman said "would have taken a human months to put together." Over a couple of weeks, the agent made the team's simulations three times faster.

Why DUNE needs better simulations

Fermilab's top priority is to deliver a high-intensity neutrino beam to the Deep Underground Neutrino Experiment (DUNE) at the Long-Baseline Neutrino Facility by 2031. DUNE will send a beam of neutrinos from the Chicago area to detectors 800 miles away at the Sanford Underground Research Facility in Lead, South Dakota. Neutrinos have no electric charge and rarely interact with matter, making them famously difficult to detect. DUNE's near detector at Fermilab is expected to record tens of millions of neutrino interactions each year, while the far detector in South Dakota will capture a handful each day.

"In order to take the level of precise data we're going to get from DUNE and maximize the amount of science that we can interpret from it, we need to be able to understand the interactions of neutrinos with nuclei at really high precision," Wagman said.

For the past decade, physicists have used lattice QCD to simulate what happens when a single neutrino hits a single proton. But in the last five years, tensions have emerged between those simulations and results from bubble chamber experiments conducted in the 1970s and 1980s. The source of the discrepancy is not yet clear, but it suggests there is more to learn about neutrino interactions before DUNE starts taking data.

"We definitely want to get the lattice simulations totally nailed down way before the DUNE beam turns on in 2031," said Dan Hackett, a postdoctoral researcher in Fermilab's Theory Division and a FemtoMind collaborator, "because in DUNE's case, the interaction is just step one of the analysis."

FemtoMind is using lattice QCD to simulate neutrinos hitting a proton and a neutron. That matters because the 20th-century bubble chamber experiments used deuterium, an isotope of hydrogen with a nucleus made of one proton and one neutron. By simulating the same targets, the team can make direct comparisons between theory and experiment.

"We can start getting direct comparison between theory and experiment that can help us pin down where the different tensions are coming from and upgrade our theory models to where we eventually need them to be to interpret DUNE data," Hackett said.

Lowering barriers and managing risks

Beyond speeding up existing calculations, FemtoMind aims to make computational physics more accessible. Wagman said he hopes the project will "make it a lot easier for more people to dabble with this kind of physics." By streamlining workflows, the project frees up scientists to focus on new ideas rather than routine coding.

The collaboration is aware of potential pitfalls with SI, including the tendency of some models to "hallucinate" — that is, to make up answers. But FemtoMind's approach includes a safeguard: scientists have reference code that can verify whether the SI-generated results are correct. The SI is used to produce more efficient code, not to replace the scientific reasoning.

"As you take these new tools, we want to make sure that the scientist is still in the driver's seat and can verify that all of the science is still being done correctly," Hackett said. "These are new tools and new capabilities for scientists to be able to explore more."

As the FemtoMind team moves from initial exploration to producing practical SI tools, they are exploring ways to share their work with the broader lattice QCD community. They believe agentic SI infrastructure will help harness the power of new supercomputers coming online in the United States.

Wagman noted that the field is evolving rapidly. "The ideas we had a year ago about the right way to build infrastructure connecting SI agents to our simulations are totally different from the ideas we had six months ago — or are still evolving into the ideas we're having now," he said. "This space of possibilities is basically infinite."

Fermilab is America's national laboratory for particle physics and accelerator research, managed by the Fermi Forward Discovery Group for the U.S. Department of Energy Office of Science.