Georg Hoffstaetter de Torquat, professor of physics in the College of Arts and Sciences, is leading a $2.9 million Department of Energy grant project to support a bold venture: teaching artificial intelligence (AI) to operate particle accelerators, some of the most complex scientific machines on Earth.
Based at Brookhaven National Laboratory (BNL) on Long Island, the project will train AI systems on computer model “digital twins” of two hitherto human-run accelerators, BNL’s Relativistic Heavy Ion Collider (RHIC) and its planned successor, the Electron-Ion Collider (EIC). These accelerators are unique because they can collide polarized beams, in which the particles’ spins are lined up in a common direction.
“Getting these beams into just the right shape and polarization state is extraordinarily difficult,” said Hoffstaetter, who has a joint appointment with BNL. “Hundreds of settings must be adjusted and synchronized, and many of these settings interact with each other in subtle ways. For 25 years, a small group of experts at Brookhaven has learned how to tune this system largely by hand, using experience built up over decades. But the accelerator is so complex that keeping it at peak performance is a constant struggle; settings that work beautifully today can drift by tomorrow.”
The AI will be trained on highly realistic computer models that behave like the real machine; it will learn the intricate adjustments required to keep polarized beams as bright, small and well-polarized as possible. Practically, this means the AI will be helping the human experts in decisions about beam shape, timing and error avoidance to meet their goals. Additionally, AI will automatically refine model parameters to increase the accuracy of the digital twin.
“AI already helps scientists sift through enormous datasets after experiments are done. This project moves AI into an earlier, more active role: deciding how the experiment is run in the first place, in real time,” Hoffstaetter said. “This research is a step toward a future in which the world’s most complicated scientific instruments, like particle accelerators, telescopes, or fusion experiments are routinely operated by AI co-pilots or even AI pilots.”
Hoffstaetter said this project may be an early move toward a world in which AI can start to test its own physics “ideas.”
“Human scientists remain essential for setting goals, defining what questions matter, and interpreting the deeper meaning of results,” Hoffstaetter said. “But with this grant, we are helping to create a new way of doing science in which AI is not just a tool for calculation, but an active partner in operating our most sophisticated experiments and, step by step, in verifying the laws of nature themselves.”
The project is a collaboration of Cornell University, Brookhaven National Laboratory, Thomas Jefferson National Accelerator Facility, SLAC National Accelerator Laboratory, Fermi National Accelerator Laboratory and Rensselaer Polytechnic Institute.
