NEWPORT NEWS, Va. – Maintaining high-power particle accelerators at peak performance requires sophisticated and precise control systems. For example, the main research machine at the U.S. Department of Energy’s Thomas Jefferson National Accelerator Facility is equipped with hundreds of finely tuned components that accelerate electrons to 99.999 percent of the speed of light.
Electrons receive this boost from radio frequencies within a series of resonant structures known as cavities. Cavities become superconducting at temperatures lower than deep space. These cavities form the backbone of the Continuous Electron Beam Accelerator Facility (CEBAF) at the Jefferson Institute. This facility is a unique DOE Office of Science user facility that supports the research of more than 1,650 nuclear physicists around the world. CEBAF also holds the distinction of being the world’s first large-scale installation and application of this superconducting radio frequency (SRF) technology.
However, changes in the properties of the cavity can increase the thermal load, resulting in a sudden loss of the zero resistance state. This will force the system to shut down through self-protection mechanisms. Such incidents are called system trips.
An accelerator trip causes a delay in the experiment. To address this, the Jefferson Lab’s team of data scientists and accelerator experts, in collaboration with university and DOE national laboratory partners, developed new machine learning (ML) techniques to transparently model the physics of CEBAF. Their goal is to provide operators with a reliable tool to predict how each cavity will behave during accelerator operation and tuning, staying one step ahead of potential failures and minimizing disruption to experiments.
“CEBAF is a fast-moving system,” said Kishan Rajput, a data scientist at the Jefferson Institute. “It’s very dynamic and operates on a microsecond scale. So it requires very fast models, primarily neural networks, running on fast computers.”
many pieces in play
Each of CEBAF’s 418 cavities can exhibit different thermal behavior and trip susceptibilities, making it difficult to build a unified model for the entire system.
To address this, researchers conducted two complementary studies in which they evaluated the performance of various ML models by simulating a representative section of CEBAF consisting of approximately 200 SRF cavities. Based on these results, the authors introduced a framework that captures the behavior of CEBAF within this approximately 200-dimensional space using an explainable ML model constrained by the underlying physical principles.
The team focused on an ML technique known as reinforcement learning (RL). RL can be compared to teaching a computer how to play chess.
“Instead of giving the ML agent a bunch of chess games that humans have played, we’re giving it the rules of chess,” said Armen Kasparian, a data scientist at the Jefferson Institute. “We tell them to play the game millions of times and learn from that experience.”
Deep differentiable reinforcement learning (DDRL) is a version that optimizes neural networks by backpropagating errors. Without differentiability, sampling high-dimensional spaces is like playing chess blindfolded. However, DDRL significantly speeds up model training, allowing scientists to solve complex problems quickly.
“We modified traditional RL to take advantage of the power of differentiability, and the results were surprising,” Rajput said. “Simply put, DDRL beat all other algorithms.”
board management
The CEBAF cavity is made from ultra-pure niobium. They are housed within a large steel structure known as a cryomomodule, which maintains an ultra-high vacuum and is submerged in liquid helium to cool the system to just a few degrees Fahrenheit above absolute zero (about 2 degrees Kelvin).
In supercooled conditions, the SRF cavity becomes very sensitive to temperature. Small changes in energy can increase the local heat load and eventually trip the system. When operating hundreds of such cavities, there is an optimal trade-off between the total amount of heat generated and the number of trips over a given period of time. This trade-off can be visualized graphically. In a graph, the relationship between these two competing goals forms a so-called Pareto front.
The research team compared the mathematical form of the accelerator control constraints to the established physics of the northern section of CEBAF, which houses about half of the accelerator’s SRF cavity. They found that by introducing the concept of explainability into physics-based constraint requirements, RL algorithms can reliably model high-dimensional environments.
Explainability is a major topic in the AI debate. Essentially turning the “black box” of AI into a gray box, explainability describes how an ML model makes decisions. It will also go a long way in improving the transparency and trustworthiness of AI.
“When you ask a neural network a question, it’s hard to explain why you get a certain answer,” said Jonathan Coren, an assistant professor at Old Dominion University and a collaborator at the Jefferson Institute. “To gain more insight, we let the model build a mathematical equation about what it sees and use that to influence its predictions. This equation is more transparent than a neural network.
“If it matches what we know about the underlying physics, we may have more confidence in the model’s subsequent decisions. But if the equation is ‘wrong’, that’s a red flag to re-evaluate the decision, the model, and even the physics of the problem.”
The RL algorithm performed well in this unique and highly parameterized chess game. With this approach, CEBAF operators may one day be able to avoid the time-consuming task of comparing the equations of the RL model to the governing physics of the accelerator and react in advance of system trips.
“This is a test of our proposed hypothesis and conclusions,” Koren said. “So differentiability is the secret sauce that allows these RL algorithms to solve these very difficult high-dimensional problems.”
This research was supported in part by the DOE Office of Science’s Advanced Scientific Computing Research Program and the Office of Nuclear Physics. The Hampton Roads Biomedical Research Consortium provided additional funding through the Advanced Computing Collaborative Institute for Environmental Research (ACES) between the Jefferson Institute and ODU. DOE’s SLAC National Accelerator Laboratory also contributed through DOE’s Institute-Directed Research and Development Program.
Read more
Harnessing the power of gradient-based simulation for multi-objective optimization of particle accelerators
Explainable physics-based constraints in reinforcement learning for accelerator optimization
Advanced science and technology computing research
Nuclear Physics (NP) Homepage
ODU Jefferson Lab establishes joint advanced computing institute for environmental research
SLAC National Accelerator Laboratory
contact: Matt Cahill, Jefferson Laboratory Communications Office, cahill@jlab.org
Jefferson Science Associates, LLC manages and operates the Thomas Jefferson National Accelerator Facility (Jefferson Lab) for the U.S. Department of Energy’s Office of Science. JSA is a wholly owned subsidiary of Southeastern Universities Research Association, Inc. (SURA).
The DOE Office of Science is the largest supporter of basic research in the physical sciences in the United States, working to address some of the most pressing challenges of our time. For more information, visit https://energy.gov/science.
