A new project aims to use machine learning to improve up-time of particle accelerators
NEWPORT NEWS, VA – More than 1,600 nuclear physicists worldwide depend on the Continuous Electron Beam Accelerator Facility for their research. Located at the Department of Energy’s Thomas Jefferson National Accelerator Facility in Newport News, Va., CEBAF is a DOE User Facility that is scheduled to conduct research for limited periods each year, so it must perform at its best during each scheduled run.
But glitches in any one of CEBAF’s tens of thousands of components can cause the particle accelerator to temporarily fault and interrupt beam delivery, sometimes by mere seconds but other times by many hours. Now, accelerator scientists are turning to machine learning in hopes that they can more quickly recover CEBAF from faults and one day even prevent them.
Anna Shabalina is a Jefferson Lab staff member and principal investigator on the project, which has been funded by the Laboratory Directed Research & Development program for the fiscal year 2020. The program provides the resources for Jefferson Lab personnel to make rapid and significant contributions to critical science and technology problems of mission relevance to the lab and the DOE.
Shabalina says her team is specifically concerned with the types of faults that most often bring CEBAF grinding to a halt: those that concern the superconducting radiofrequency acceleration cavities.
“Machine learning is quickly gaining popularity, particularly for optimizing, automating and speeding up data analysis,” Shabalina says. “This is exactly what is needed to reduce the workload for SRF cavity fault classification.”
SRF cavities are the backbone of CEBAF. They configure electromagnetic fields to add energy to the electrons as they travel through the CEBAF accelerator. If an SRF cavity faults, the cavity is turned off, disrupting the electron beam and potentially requiring a reconfiguration that limits the energy of the electrons that are being accelerated for experiments.
Shabalina and her team plan to use a recently deployed data acquisition system that records data from individual cavities. The system records 17 parameters from a cavity that faults; it also records the 17 parameters from a cavity if one of its near neighbors faults.
At present, system experts visually inspect each data set by hand to identify the type of fault and which component caused it. The information is a valuable tool that helps CEBAF operators for how to mitigate the fault.
“Each cavity fault leaves a unique signature in the data,” Shabalina says. “Machine learning is particularly well suited for finding patterns, even in noisy data.”
The team plans to work off of this strength of machine learning to build a model that recognizes the various types of faults. When shown enough input signals and corresponding fault types, the model is expected to be able to identify the fault patterns in CEBAF’s complex signals. The next step would then be to run the model during CEBAF operations so that it can classify in real time the different kinds of faults that cause the machine to automatically trip off.
“We plan to develop machine learning models to identify the type of the fault and the cavity causing instability. This will give operators the ability to apply pointed measures to quickly bring the cavities back online for researchers,” Shabalina explains.
If successful, the project would also open the possibility of extending the model to identify precursors to cavity trips, so that operators would have an early warning system of possible faults and can take action to prevent them from ever occurring.
Contact: Kandice Carter, Jefferson Lab Communications Office, 757-269-7263, firstname.lastname@example.org