Few systems match the complexity of nuclear power generators. Safe operations depend on continuous monitoring of many conditions simultaneously, and on regular maintenance. It all makes nuclear a prime candidate for improvement with artificial intelligence.

That’s exactly what engineer Rick Vilim, department manager of Plant Analysis, Control and Sensors at the Argonne National Laboratory, aims to do.

Vilim said computer power has grown to the point it can enable a decades-old goal of automating many of the feedback-and-manual adjustment processes typically done by human operators watching dials and looking at graphs. Beyond that, Argonne analysts have constructed a digital twin of commercial reactors now in service. The twin operates in addition to a mock physical plant that uses electrically generated heat and 1,000 sensors to simulate a nuclear reactor.

“Now that we have the plant in virtual form, we can design controllers for it virtually,” Vilim said, referring to the digital twin. “We can automate things and create algorithms that previously couldn’t really be implemented because the computer power wasn’t there.”

This development occurs in the context of an Argonne project to simplify nuclear power plant operation, even as the next generation of generators takes shape.

“The plants on the drawing board today are much simpler and basically operate somewhat by themselves, if you will,” Vilim said. “They’re self-regulating.” He noted that one new-generation plant, which uses new materials and cooling mechanisms, is already under construction in Wyoming under the Energy Department’s Advanced Reactor Demonstration Program.

“One of the things Argonne has been working over on the past 20 years,” Vilim said, “is just trying to simplify the principles behind the operation of nuclear power plant so fewer individuals are required to operate it.”

Remote sensing and diagnostics are also becoming part of nuclear power operations, Vilim said.

“Not so much from the standpoint of active control,” he said. “There will always be a control room on site with operators.” Rather, the remote monitoring centers – which can handle multiple plants – will function to gauge the health of a plant’s components. Here again, remote monitoring goes back decades, but “our current work is basically trying to move the needle on some of these algorithms, what they can do in terms of predictive capability and control,” Vilim said.

Physics, not hallucinations

In a famous movie scene, actor Henry Fonda plays a nuclear plant manager, panicked on seeing a dangerous reading. He taps on the mechanical meter and it unsticks, finally showing an accurate reading. It doesn’t solve the problem, but at least two measurements agree.

As plant instrumentation has become digital, that particular anomaly – which Vilim said used to occur occasionally in real life – has largely disappeared. The contemporary challenge is taking manual operations of controls out of human hands, or augments human judgment. Argonne’s approach, Vilim said, takes numerical values produced by instrumentation and feeding them to an algorithm. The algorithm in turn invokes a more precise response.

“It can do operations in a more precise manner than an operator would,” Vilim said, and it can entertain much more complexity in trying to control the machine than the operator could.”

Vilim added, “The operator is limited by human bandwidth. You can only perceive so many things at once. There’s real advantage in these algorithms.” That advantage, he said, comes from training the algorithms with historical data that incorporates physics controlling nuclear operations.

“So we have the same intuitive understanding in the algorithm [as a human operator], but it’s even better because it actually can assign numerical values to things,” Vilim said.

Argonne research and development in AI involves using industry-supplied data from operations, seeing what they produce when processed by the algorithms, and comparing outputs with the real world results. This technique applies to a variety of tasks.

“One is monitoring, just determine the status of equipment,” Vilim said. “Another is control, managing the facility, so process variables stay where they’re supposed to and you meet the demand or the objective, which is to produce electricity.”

Key to training and evaluating AI is that the industry data sets arrive “blind,” Vilim said.

“They’re blind in the sense that we don’t know what is going on with the data set, physically in the plant” that generated the data, although the plant operators know. This arrangement keeps Argonne’s evaluations objective.

“We run our algorithms in the data, and we tell them what we found,” Vilim said, “and they’ll say, ‘okay, that’s what we saw’” or not. In once case, a plant’s managers handed Argonne a set of data from completely normal operations.

“We came back and said we don’t see anything,” he said. “And they said, ‘Well, that’s right, we wanted to see if your algorithm could handle this.’”

Yet in another instance, the monitoring algorithm, working on 17 hours’ worth of archived data from a plant startup, found something operators missed as it was occurring. A heat transfer component had degraded slightly in a process known as fouling. Vilim said that the variance between human observed and algorithm calculated results showed the greater precision possible with AI.

“That was pretty gratifying to see the algorithm can actually pick that up,” Vilim said. “It’s much more sensitive.”

As it has in other domains, artificial intelligence has started producing predictive, and not just look-back, analytics in nuclear power operations.

“We know that the utilities have a real interest in trying to better schedule maintenance,” Vilim said. “But you don’t want to be maintaining something doesn’t need to be maintained.” AI work therefore centers on predicting how equipment will wear.

“We’re aiming to be able to provide the nuclear operator with a measure of how a particular component is progressing through service life,” Vilim said. The idea is to provide alerts when a piece of gear has, say, six more months of useful life.

Vilim said Argonne is trying out generative AI, using its nuclear plant digital twin. He said that by constraining the generative training to data produced by real-world plan operations, the generative algorithms won’t produce the “hallucinations” that dog other applications.

“Since we have the physics and we know where the plant operates, we don’t hallucinate,” Vilim said. “The digital twin brings that. It’s basically an incarnation of the physics, and we use that to put guardrails on our algorithms.”

Learning Objectives:

• Argonne National Laboratory mission and overview
• Automation trends
• Industry analysis