June 1, 2021
Staff Accomplishment

PNNL Researchers Use Machine Learning to Study Beam Outages

Data science tools will help to improve energy efficiency at Fermilab

Jan Strube PNNL

PNNL physicist Jan Strube pairs advanced computing with high energy physics to optimize detection capabilities.

(Image composite by Cortland Johnson | Pacific Northwest National Laboratory)

Have you ever stopped at a traffic light and wondered if you should turn off your engine to save gas and reduce carbon dioxide emissions? If you could predict how long the light would stay red, it would help you determine if you should turn off your car.

At Fermi National Accelerator Laboratory (Fermilab), the U.S. Department of Energy (DOE) Office of Science’s particle physics and accelerator facility located near Chicago, Illinois, they are experiencing a similar situation on a much larger scale. Researchers at Pacific Northwest National Laboratory (PNNL) are collaborating with scientists at Fermilab to forecast time and duration of accelerator beam outages using data science.

The project is one of 14 national laboratory-led efforts awarded $37 million over three years for “Data, Artificial Intelligence, and Machine Learning (ML) at DOE Scientific User Facilities.”

Jan Strube, a physicist at PNNL, is leading the project funded by DOE’s Office of Science, High Energy Physics program to improve the operational efficiency of Fermilab's particle accelerators. Strube also holds a joint appointment at the University of Oregon’s Institute for Fundamental Science. “Fermilab is working with PNNL on this issue because of our expertise in machine learning and data analytics,” said Strube.

PNNL researchers are partnering with Fermilab colleagues to implement best practices for data collection and apply modern ML tools to extract information from the Fermilab accelerator complex. Through their research, they aim to understand and forecast beam outages, optimize accelerator performance, and improve energy efficiency.

Advanced controls infrastructure

Fermilab operates a network of sophisticated scientific machines called particle accelerators that physicists use to conduct particle physics experiments. The accelerator systems vary in age, constituent technology, and mode of operation and are managed by an advanced controls infrastructure.

Accelerator operators monitor the systems using more than 200,000 controllable devices, automated alarms with inter-operating software, and independent safety systems. During normal operations, 15,000 alarms and additional status indicators are triggered each day. During machine study periods and in some phases of operations, information rates from accelerator alarms can be significantly higher.

Potential for enormous energy savings

The power consumption of Fermilab’s accelerator complex is typically between 30 to 40 megawatts during normal operations. In comparison, 1 megawatt is enough power about 800 U.S. homes.

During times without beam (for example, due to glitches or unscheduled maintenance), the power consumption could be reduced significantly if the time and duration of the outage could be accurately forecasted. Depending on the duration, systems not in use could be powered down, go into standby mode, or simply stay on, thereby increasing energy efficiency.

As PNNL researchers develop new methods to predict accelerator system downtime and adjust for power outages, they are demonstrating their capabilities that are relevant to high-energy particle physics programs. “I am confident our work will result in energy savings and hopeful that it will point out ways to increase energy efficiency in the future, paving the way to a more sustainable scientific infrastructure,” said Strube.

Other collaborators on this new project include PNNL computer scientist Vinay Amatya, PNNL data scientist Milan Jain, and William Pellico and his team at Fermilab.