A powerful machine learning algorithm that nudges gold atoms closer together
image: An image of ion collisions captured by Brookhaven National Laboratory's STAR detector. The multicolored lines represent quarks and gluons released during the collision. With GPTune, ion collisions would produce more subatomic particles (more colored lines). These additional quarks and gluons allow scientists to peer deeper into the fundamental behavior of nature.
Credit: Brookhaven National Laboratory
At Brookhaven National Laboratory’s (BNL’s) Relativistic Heavy Ion Collider (RHIC), billions of gold ions race through magnets at nearly the speed of light. Thousands of times per second, they collide head-on, breaking into smaller particles that reveal fundamental secrets of nature and the Universe’s origins, including moments after the Big Bang.
RHIC, the first machine to collide heavy ions—atoms stripped of their electrons—uses a 2.4-mile circular collider to guide ion beams traveling in opposite directions. Aligning these beams for maximum collisions requires tuning nine injector knobs to adjust properties like size, shape, and intensity, a task as challenging as juggling nine bowling pins. Manually controlling the beams demands years of expertise.
To help operators fine-tune beams and generate more powerful collisions, researchers from BNL, Lawrence Berkeley National Laboratory (Berkeley Lab), and Michigan State developed a machine learning algorithm to increase beam intensity—how many ions fit into a beam pack. This is akin to focusing a flashlight for brighter, more direct light. A paper describing this work was presented at the 15th International Particle Accelerator Conference in Nashville, Tennessee.
“Colliders are big, complex machines,” said Ji Qiang, a senior scientist at Berkeley Lab. “Our algorithm addresses uncertainties to better control beams traveling at nearly the speed of light.”
Learn by Doing
A machine learning algorithm is a mathematical expression that processes data in order to recognize patterns in that dataset. The more data that goes into a machine learning algorithm, the more precise the algorithm’s mathematical description of a phenomenon becomes.
At RHIC, the intensity of the ion beams is represented by a mathematical function so complex that researchers have not yet developed an equation that exactly describes it.
“The function describing the beam is not known, so we are trying to learn the function using data collection and an algorithm,” said Sherry Li, a senior scientist in Berkeley Lab’s Applied Math and Computational Research Division (AMCR) and on the research team. “Machine learning techniques fill in this knowledge gap.”
The research team used their machine learning software, GPTune, to process the nine beam control parameters set for the RHIC’s Electron Beam Ion Source (EBIS). EBIS prepares beams for experiments by removing electrons from atoms—i.e., ionizing the atoms before they enter the RHIC. Leveraging data collected at FC96, a Faraday cup measuring location near the end of the system, GPTune applied advanced statistical methods to predict and improve beam performance with each adjustment.
During the first attempt to optimize EBIS’s beam intensity, most of the beam intensity values were below the initial level during sampling and the early optimization phase. However, after evaluating 45 control parameter configurations suggested by GPTune, a beam’s intensity started to show signs of improvement.
“This was the most exciting moment of the experiment,” BNL researcher and team member Xiaofeng Gu said. “The beam intensity eventually emerged from these low points and surpassed the initial value.”
After about 25 additional control parameter evaluations, the algorithm increased the average beam intensity by 22%.
At XF14, a current transformer measuring location in the beamline, GPTune was applied to optimize 10 control parameters in the extraction beamline. Here, the beam’s average intensity increased by 43%, a more significant improvement.
Finally, the team evaluated the combined improvement in beam intensity when optimizing both injection and extraction control parameters. The algorithm achieved a 68%–71% increase in total beam intensity at the extraction location and a 22%–24% increase at the booster injection location.
Going forward, the research team plans to implement GPTune at other beamlines at the RHIC. Doing so will ultimately increase the detector’s total luminosity. Furthermore, the research team can now apply the algorithm and approach to subjects outside of particle colliders.
“The successful optimization of Brookhaven’s ion beam also revealed to us the versatility of the GPTune algorithm,” Li said. “We are excited to apply GPTune now to broader scientific challenges, including instruments for many different fields of study.”
This research project for EBIS control parameters optimization is funded through the DOE’s Nuclear Physics office. The GPTune development was initially funded through the Department of Energy’s Exascale Computing Project (ECP), as part of the xSDK4ECP project. It is now funded through the DOE ASCR’s Applied Mathematics program. The team includes researchers from Berkeley Lab (J. Qiang, X.S. Li, Y. Liu), BNL (X. Gu, B. Coe, T. Kanesue, M. Okamura), and Michigan State University (Y. Hao).
This story was written by David Krause.