Humans in the loop: How accelerator operators are partnering with AI

In the accelerator control room at the Department of Energy’s SLAC National Accelerator Laboratory, operators work long shifts to keep the world’s most powerful X-ray free-electron laser running smoothly. For 8 to 12 hours at a time, they monitor rows of computer screens that track the performance of the Linac Coherent Light Source (LCLS). It’s these accelerator operators’ job to make quick tweaks so that the machine delivers precise beams for boundary-pushing experiments ranging from protein imaging to quantum materials research.

While advancements in algorithms and automation have transformed many aspects of this work, the human side – how operators make decisions, sometimes under stress – remains less well understood. 

A team of researchers at SLAC is working to change that.

“There’s a lot of focus on improving the machines,” says collaborator Roussel Rahman, “but the human side is much harder to study.”

The team includes Rahman, Jane Shtalenkova and Wan-Lin Hu. Rahman, a cognitive scientist with a background in mechanical engineering, studies how people learn and master complex tasks in real-world settings using mathematical models. Hu, an expert in human-centered design and human-machine interaction, brings insights from her previous work improving LCLS user interfaces. Shtalenkova, an accelerator operations machine specialist at SLAC with over six years of control room experience, provides firsthand knowledge of how operational expertise develops.

Together, they form a multi-disciplinary team united in their mission to optimize human-machine collaboration. By studying how operators think and act under pressure, they aim to refine control room tools, improve training and pave the way for smarter cooperation between humans and machines.

This effort is part of a broader push at SLAC to build the workforce of the future – training accelerator scientists and engineers, as well as creating new opportunities in accelerator research and development to engage staff and attract talent.

At the same time, the lab is building new tools and technologies to keep up with the growing flood of scientific data. By combining algorithms, artificial intelligence and machine learning, these efforts are helping teams manage complex experiments more efficiently.

Making decisions in an ever-changing environment

When operating an accelerator, conditions change constantly. No two shifts are exactly alike, and operators must rely on experience and judgment to interpret the system’s behavior and decide what to do next.

“The solution that worked yesterday may not work today,” Shtalenkova says.

To study this process, the researchers analyzed 14 years of electronic logs, text entries that document what happened during each shift and what actions were taken. Using language processing models, the team identified patterns in how operators tackled complex tasks, such as tuning the beam for optimal performance. They found clear differences between novice and expert operators: Experts leveraged their experience and domain knowledge to prioritize and sequence subtasks more efficiently.

To complement this analysis, the team also conducted a study with current operators, recording screen activity, machine data and operator actions while tuning sessions were underway, then talked with the operators afterward to understand the reasoning behind their decisions.

“It’s not just what the operators are clicking,” Hu says. “It’s about the data they’re seeing on the screens, who they’re coordinating with and how they’re adapting on the fly.”

This part of the study also helped the team identify which parts of the interface were most important – and where design improvements could reduce cognitive load or streamline workflows.

Strengthening human-AI collaboration

While automation has advanced, certain tasks in the control room remain firmly in human hands – especially those involving rare or unexpected situations. AI systems often struggle with edge cases where data is limited, but humans often manage them with practical, experience-based reasoning. In the control room, it’s clear that people make the system work. 

“It’s really difficult to find the optimal approach,” Hu says, “but humans have very simple, robust and effective ways of dealing with uncertain situations.”

The SLAC research team is highlighting this expertise and showing how it can be harnessed. Rather than replacing humans, they envision a future where AI supports them. “We’re interested in building human-AI teams that are stronger together than either one alone,” Rahman says.

The team’s approach could extend to other high-stakes environments, such as air traffic control, power grid management or healthcare – any setting where skilled humans must navigate complex systems in real time. As the team prepares to publish their findings, they’re already planning next steps: refining their models, exploring AI teammates and sharing their insights with the broader scientific community.

For Shtalenkova, the work is deeply personal. “I remember two years into my career, I turned around and thought, ‘How does my brain even know how to do this?’” she says. “When I first walked into the job, it was just a flood of unfamiliar signals and data. But over time, my brain developed the connections to understand what I'm seeing and react appropriately.”

Now, she’s using the team’s findings to improve SLAC’s operator training program. Previously, training relied heavily on hands-on experience and informal knowledge sharing. With new data-driven insights, the team hopes to offer clearer guidance for navigating complex tasks.

“I’ve seen firsthand how impactful operators’ strategic decisions are to the success of the laboratory’s vision,” Shtalenkova says. “This research makes me hopeful. It reminds us that people are still essential in some of the most advanced places in science.”

For questions or comments, contact SLAC Strategic Communications & External Affairs at communications@slac.stanford.edu.