The Machine Learning Lunch Series unites potential collaborators across Jefferson Lab, further progressing projects that apply this powerful tool to nuclear and accelerator physics problems.
In the fall of 2018, the scientific computing group at the U.S. Department of Energy’s Thomas Jefferson National Accelerator Facility held a workshop focused on machine learning. The room was packed. So packed that one attendee, Chris Tennant, a Jefferson Lab staff scientist, realized he didn’t recognize many of the faces in the crowd—which he wanted to change.
Tennant, who is an accelerator physicist at Jefferson Lab, had been studying machine learning on his own. But after seeing that many of his colleagues were also interested in the subject, he wanted to find a way to connect them.
“After that workshop, I thought it was important to get a machine learning community together in a more informal, frequent way,” Tennant said. He decided to start a weekly Machine Learning Lunch Series.
Machine learning is a branch of artificial intelligence in which algorithms perform a specific task without explicit instructions. Machine learning algorithms take sample data, called training data, and build a mathematical model that uses pattern identification and inference to make predictions or decisions instead of being told exactly what to do. While there’s no dedicated machine learning group at Jefferson Lab, a handful of staff are working on projects that apply this powerful tool to problems in their fields of expertise, including nuclear and accelerator physics problems.
“I want the lunch series to foster collaboration among those studying machine learning on their own,” Tennant said. The first lunch was held on Feb. 13, 2019, and about 10 people showed up. By the fourth lunch, there were 20 attendees from across the lab, including experimentalists, theorists, and scientific computing group members. The lunches are every Wednesday, noon to 1 p.m. in CEBAF Center room F324-325. Any Jefferson Lab staff members interested in machine learning may attend.
The meetings are not intended to provide detailed tutorials about machine learning, but rather serves as a space for open discussion.
“Because machine learning is a popular and growing field, there are lots of free online resources about it,” Tennant said. “The lunch series is more of a time for active participation instead of passive listening.”
One attendee, Thomas Britton, a postdoc in Jefferson Lab’s Hall D, manages data quality monitoring for the Gluonic Excitations Experiment, or GlueX, which seeks to learn more about the strong force—the force that binds all matter together. GlueX generates particles called mesons, which are made up of smaller particles called quarks. Quarks are held together by the strong force. Studying mesons made by GlueX reveals more about the strong force.
When GlueX is running, Britton must check the quality of the data it produces daily. Some days, there are already hundreds of plots of data produced by 5 a.m. It’s up to him to examine them all.
“Sometimes I can’t make it through all the plots,” Britton said. “I think this job could be automated: A machine would be better and more thorough.”
He’s currently working on a project that uses machine learning to examine GlueX data quality and decided to attend the lunch series to hear what others around the lab are up to with machine learning. He gave a presentation about his work at one of the lunches, sharing some of the dos and don’ts he learned while working on his own project. There, he learned that another attendee had done work on a system similar to what Britton is trying to build to monitor data quality.
“He will be a good source of information when the time comes,” Britton said. “I believe that there may be a lot of overlap between various projects, and when someone runs into a problem, someone else at the lab may have already battled that demon.”
Since he started attending the lunch series, Britton has made more progress.
“Thomas has done some amazing stuff in a short amount of time,” Tennant said. “It’s remarkable to me how quickly things have taken off since the lunch series started. Projects in the early stages weeks ago are already showing impressive results.”
Tennant has also made quick progress on his own machine learning project. Like Britton, Tennant wants to automate a job that currently requires human effort and expertise. The project is aimed at reducing the unscheduled downtime of the Continuous Electron Beam Accelerator Facility, a DOE Office of Science User Facility and Jefferson Lab’s main accelerator.
The CEBAF accelerator uses superconducting radiofrequency cavities to accelerate particles. These cavities can turn off, or trip, due to a multitude of factors, including mechanical vibrations or local heating. When the cavities trip, the particle beam turns off, and the physicists using the accelerator stop receiving experimental data.
Usually, when a trip occurs, a subject matter expert is required to determine the type of fault. The goal of Tennant’s project is to train a machine learning program to do what the expert does. The machine learning model would recognize patterns in signals from cavity data that indicate what type of trip occurred. Using a machine to determine what caused the trip would be faster than a human and would provide important feedback to the accelerator operators.
“So far, I have encouraging results,” he said.
For those who can’t make it to the weekly lunches, the series is supplemented by a machine learning Slack channel, known as the #ml channel on the on the JLab 12 GeV Slack workspace. These different avenues of engagement are intended to develop a community that will progress machine learning at Jefferson Lab.
Or, as Britton puts it: “If the lab wants to explore machine learning, we should not work like hermits.”
By Chris Patrick
Contact: Kandice Carter, Jefferson Lab Communications Office, 757-269-7263, kcarter@jlab.org