BYLINE: Amber Rose
Recent advancements at the ATLAS user facility are enhancing operations and efficiency to help unlock new insights into the universe’s fundamental forces.
Newswise — In the quest to solve fundamental mysteries about the universe, nuclear physics stands at the forefront, probing the very building blocks of matter and the forces that govern their interactions. At the U.S. Department of Energy’s (DOE) Argonne National Laboratory, a prominent fixture in this exploration is the Argonne Tandem Linac Accelerator System (ATLAS), a DOE Office of Science user facility.
Here, scientists explore our fundamental understanding of the atomic nucleus. What forces bind protons and neutrons together at the heart of atoms? How do these interactions give rise to the various elements that make up our universe? What secrets do rare and exotic nuclei whisper about the origin of matter itself?
“With AI, we’re not just improving efficiency; we’re redefining what’s possible in beam tuning. This technology allows us to explore new dimensions of precision and reliability that were previously out of reach.” — Daniel Santiago, Argonne physicist
These questions drive research at ATLAS, where recent advancements in artificial intelligence (AI) are revolutionizing operations and enabling groundbreaking research. By integrating cutting-edge AI technologies, ATLAS is not only enhancing the efficiency of its operations but also clearing the path for new avenues of discovery that could shed light on the intricate details of the particles that shape our universe.
More time experimenting, less time tuning
A linear accelerator, or linac, uses electromagnetic fields to accelerate charged particles, such as electrons or protons, to very high speeds along a straight path toward a target — a material or object that the accelerated particles collide with during experiments, initiating various nuclear reactions. The process begins with a source that produces these charged particles, which are then injected into an accelerator. ATLAS is capable of producing beams from all naturally occurring elements, from hydrogen to uranium, and can accelerate them up to 20% of the speed of light.
These beams consist of either stable or radioactive particles, depending on the research goals. Stable beams, whose particles do not decay over time, have been widely used in nuclear physics research and have significantly contributed to our current understanding of nuclear properties. To push further into new frontiers and gain deeper insights, researchers are increasingly using radioactive beams. These beams consist of unstable particles that decay quickly. Although working with such beams can be incredibly challenging, they provide scientists unique opportunities to study rare and exotic nuclei that cannot be accessed with stable beams.
As they travel from the source to the target, the accelerated beams are steered through numerous electromagnetic elements, which, much like lenses, bend and focus light. Ultimately, these beams collide with the target, causing the target’s atomic nuclei to become excited, break apart or even fuse with incoming particles. Analyzing the resulting particles and radiation emitted from these interactions gives scientists valuable insights into the structure and behavior or atomic nuclei and nuclear reactions.
Currently, many of the electromagnetic elements must be manually adjusted to optimize beam transmission. Imagine tuning an old-fashioned radio to get the clearest signal possible. But instead of adjusting one dial, there are over 100 dials to adjust, each affecting the signal in a different way. From past experience tuning this radio, you know approximately where each dial should be, but the signal can also change unpredictably, requiring constant fine-tuning. The goal is to get the strongest, clearest signal with minimal static, which means adjusting each dial individually to minimize interference and maximize clarity. This process is clearly time-consuming and requires a significant amount of manual effort.
At ATLAS, this tuning process can take a day or two to achieve optimal conditions, eating into precious experimental time. Recently, scientists have implemented several new AI and machine-learning programs to help automate the beam-delivery process with minimal human intervention. These upgrades are expected to enhance operational efficiency and beam availability, further supporting critical scientific research and applications.
The success of ATLAS’s operations hinges on the expertise and dedication of its accelerator operator team. These skilled professionals are responsible for the meticulous tuning of the accelerator, ensuring that the beams are precisely configured for researchers’ experiments. The introduction of AI and machine-learning tools at ATLAS will empower them with advanced technologies that enhance their capabilities.
Improving operations and redefining the possible
One of these new improvements at ATLAS is aimed at a unique source of radioactive ions called nuCARIBU, which generates beams of radioactive isotopes from neutron induced fission — splitting the target nuclei into smaller parts.
The “nuCARIBU-matic” project aims to automate components of the radioactive beam-delivery process using machine learning and hardware improvements.
“We believed we could enhance the efficiency and robustness of our operations,” said Physicist Daniel Santiago. “By developing an automated beam-tuning process, we also aim to tackle workforce challenges, given the limited number of operators available for tuning and optimization.”
The production and delivery of radioactive ion beams requires specialized knowledge and skills relying on a small group of expert operators. Automating the process helps alleviate the burden on human operators by speeding up the process, improving consistency and delivery to users.
Santiago and postdoctoral appointee Sergio Lopez-Caceres used Bayesian optimization, a statistical method, to find the best settings for beam transport. This method determines how to adjust many settings simultaneously to achieve optimal beam performance with minimal losses. Santiago and Lopez-Caceres were able to optimize eight to 10 electrostatic elements at a time and apply this approach to multiple sections of the beamline, including an evacuated pipe in which the radioactive ions travel.
“Humans are good for many things, but we’re not great at exploring an eight-dimensional parameter space,” Lopez-Caceres said. “The program isn’t magic; it’s just sophisticated math and computing. But it’s something beyond what a human expert can do, and that’s really exciting.”
Recently, the nuCARIBU team conducted a live test of the system, successfully optimizing over 60 beamline elements with minimal human intervention and achieving beam transmission comparable to that of human experts.
“With AI, we’re not just improving efficiency; we’re redefining what’s possible in beam tuning,” Santiago said. “This technology allows us to explore new dimensions of precision and reliability that were previously out of reach.”
The team is now working on expanding the machine-learning program to additional sections of the nuCARIBU beamline, with the goal of fully automating the transportation of radioactive beams from the source to users’ experimental stations.
Seeking an elusive digital twin for safe testing and experimentation
On the main superconducting linac beamline, which is mostly used for stable ion beams, Accelerator Physicist Brahim Mustapha is leading efforts to enhance beam-tuning efficiency by integrating machine learning into daily operations.
The first step was to establish a reliable data-collection process to ensure all relevant information, such as beam parameters and machine settings, are accurately captured and linked. For this, ATLAS researchers first developed a comprehensive data collection system and used machine-learning techniques to optimize settings. From this solid data foundation, the team then built models that could optimize beam parameters by starting from nonoptimal conditions and refining them through iterative processes to reach desired outcomes.
Like the nuCARIBU-matic project, the team used Bayesian optimization and also included reinforcement learning — in which models learn by trial and error to achieve a goal — to develop models that can quickly and accurately fine-tune beam settings. While these methods already existed, they were adapted to fit the unique challenges of beam tuning at ATLAS.
In addition to developing a comprehensive data collection system, Mustapha and his team have a goal of creating a virtual machine model, also called a “digital twin,” of the accelerator.
“A digital twin mimics the behavior of the actual machine, allowing it to be used both online, in real-time alongside the accelerator, and offline, independently for testing and development,” Mustapha explained.
A digital twin serves as a safe environment to test and optimize beam settings without the risk of damaging elements of the machine itself. It is a crucial tool for exploring new tuning strategies and refining existing ones without compromising safety.
“But to do that, you need the model to be as realistic as possible, like an exact copy of the machine itself, which is not an easy task,” Mustapha noted.
Creating an accurate digital twin is challenging due to the complexity of a full-scale accelerator system like ATLAS. Many factors, such as the misalignment of elements, errors in settings and changes over time, can affect the accuracy of the model. Despite these challenges, Mustapha and his team aim to create a model that behaves as close to the real machine as possible.
Although still in the development phase, this transition promises to reduce the time required for beam tuning, increase the availability of beam time for experiments and improve the overall quality of the beams delivered to researchers.
Perfecting the recipe for consistent and reliable results
Postdoctoral appointee Khushi Bhatt’s focus is on optimizing the production and delivery of radioactive beams by the Argonne In-Flight Radioactive Ion Separator (RAISOR). RAISOR is designed to produce and separate radioactive ion beams, but its beams tend to be short-lived, presenting a unique challenge for both production and handling.
RAISOR starts by using a stable ion beam to bombard a target, creating a mix of reaction products, including the desired radioactive isotopes — called the secondary beam — and other unwanted particles. It then uses magnetic and electric fields to isolate the desired isotopes, much like picking out the M&M’s from a trail mix. Once separated, these isotopes are directed to users’ experimental stations.
To improve RAISOR’s performance, Bhatt has developed two diagnostic beam stations that provide real-time data on beam properties. This data is essential for training machine-learning models that seek to automate the tuning process and improve beam quality and transmission on RAISOR.
“The goal, and this is where machine learning comes in, is to increase the purity of the desired isotope and complete the tuning and transmission as quickly as possible,” said Bhatt. “Currently, this takes two to three days. If we can reduce it to one or one and a half days, all that other time can be dedicated to doing more experiments and collecting more data.”
Each diagnostic station is equipped with detectors that measure different parameters of the beam, such as its shape and position, ensuring the beam is correctly aligned and focused as it travels through the accelerator. The stations record correlated data, capturing both the changes made to the beamline settings by operators and the resulting effects on the beam.
Bhatt explains this process like making a new pasta dish for the first time. You experiment with different ingredients and techniques, and once you succeed, you want to know exactly what you did so you can replicate it. The diagnostic stations are like the final dish, showing the outcome of the beam adjustments.
In beam optimization, understanding the specific changes that lead to the desired outcome is essential. A feedback loop, using machine-learning algorithms, provides real-time data on beam performance and allows the system to adjust the settings as needed. Just as you taste-test your pasta sauce and add salt if it’s bland, the feedback loop ensures the “recipe” for optimal beam quality is followed and refined, leading to more consistent and reliable experimental results.
Advancing autonomy: Achieving unmatched speed and precision
The integration of AI at ATLAS is not just about improving current operations; it’s about envisioning a future where fully autonomous operation is possible. While this ultimate goal may still be just over the horizon, the progress made so far is a significant step toward realizing that vision.
“If we use machine learning to optimize our process, we can save time,” Bhatt said. “This means more time for experiments, leading to more discoveries and innovations through better data collection.”
As ATLAS continues to push the boundaries of nuclear physics research, the role of AI will undoubtedly expand, offering new opportunities for innovation and discovery. The collaboration between human expertise and machine intelligence is setting the stage for a new era in scientific exploration, one where the mysteries of the atomic nucleus are unraveled with unprecedented precision and speed.
This work was funded by the DOE Office of Science for Nuclear Physics and the Laboratory Directed Research and Development program at Argonne.
Argonne Tandem Linac Accelerator System
This material is based upon work supported by the U.S. Department of Energy (DOE), Office of Science, Office of Nuclear Physics, under contract number DE‐AC02‐06CH11357. This research used resources of the Argonne Tandem Linac Accelerator System (ATLAS), a DOE Office of Science User Facility.
Argonne National Laboratory seeks solutions to pressing national problems in science and technology by conducting leading-edge basic and applied research in virtually every scientific discipline. Argonne is managed by UChicago Argonne, LLC for the U.S. Department of Energy’s Office of Science.
The U.S. Department of Energy’s Office of Science is the single largest supporter of basic research in the physical sciences in the United States and is working to address some of the most pressing challenges of our time. For more information, visit https://energy.gov/science.
