Artificial Intelligence Has Entered the Nuclear Industry and Its Early Benefits Are Just the Tip of the Iceberg

A collaboration between startup Atomic Canyon and the Oak Ridge National Laboratory allowed construction of a sentence-embedding model using 53 million pages of Nuclear Regulatory Commission documents. Understanding nuclear terminology opens doors for artificial intelligence to revolutionize many processes within the industry. The nuclear power sector is regarded by many onlookers as a slow-moving industry. Strict regulatory frameworks mean extensive testing, documentation, and approval processes are necessary for any modifications to nuclear facilities or procedures. Furthermore, nuclear plants represent massive investments, which makes owners and operators err on the side of caution when it comes to making changes to proven designs. Major upgrades must be thoroughly validated given the complexity of nuclear systems and the potential consequences of failure, and the studies conducted as part of the process can take years. The industry relies heavily on experienced personnel, who are trained to follow detailed procedures, which means they may be resistant to adopting new approaches. So, in the end, things often do change slowly. However, change does happen, and technology has worked its way into the nuclear industry as the development of advanced reactor designs and small modular reactors have accelerated innovation in the field. Recently, artificial intelligence (AI) even found its way into the nuclear conversation as a startup called Atomic Canyon collaborated with the Department of Energy’s Oak Ridge National Laboratory (ORNL) to develop an advanced AI model capable of understanding complex nuclear terminology.

Training AI

“The first thing you need to build artificial intelligence is a data set—you need access to information,” Trey Lauderdale, founder and CEO of Atomic Canyon told POWER. Lauderdale is not a nuclear expert, but he has founded and supported multiple companies over the past 15 years with a focus on utilizing technology to improve processes, which is what Atomic Canyon’s AI platform is designed to do for nuclear power plants, manufacturers of next-generation reactors, and government and national laboratories. “One thing we quickly realized about nuclear power is there is a tremendous amount of data. The Nuclear Regulatory Commission—the NRC—actually has a database called ADAMS [which stands for Agencywide Documents Access and Management System], where there’s all sorts of public information that’s available that can be viewed by anyone, and it’s all available on their website,” Lauderdale said. As Atomic Canyon’s team started building AI models and experimenting with the ADAMS dataset, its experts quickly discovered a problem: all of the AI models that are generally available would get confused when they ran into “nuclear words.” Lauderdale explained, “The nuclear vernacular is very complex. It has all sorts of acronyms and words that these AI models haven’t seen enough examples of. So, what ends up happening is the AI hallucinates. That’s AI speak for: ‘It makes stuff up.’ As you can imagine, in an industry like nuclear, making stuff up is very, very bad.” Lauderdale’s team realized they didn’t necessarily need to create a new large language model (LLM) to solve the problem, they just needed to build sentence-embedding models for AI applications so nuclear terminology could be understood. “To do that, you need access to a lot of what’s referred to as GPUs—graphical processing units,” Lauderdale said. A typical start-up might raise millions of dollars and buy a bunch of GPUs to do a project like this, but Atomic Canyon had a better option: work with the government. ORNL is home to Frontier (Figure 1), a supercomputer that was touted as the world’s fastest when it debuted in May 2022 and has maintained that title through the most recent rankings in May 2024. “It was quickly discovered that this was a project that was worthwhile of the world’s fastest supercomputer—the ability to go train AI models on nuclear terminology and then have an output which is basically a more advanced search application that could be used to help find documents,” Lauderdale said.

The results were astounding. Within just six months, the team developed an advanced AI model capable of understanding complex nuclear terminology. This specialized open-source AI model has set new benchmarks for accuracy, efficiency, and speed in AI search. Developed to be open source, the model is available to ORNL, the nuclear national lab complex, independent researchers, and nuclear institutions. It will also be integrated into Neutron, Atomic Canyon’s AI search platform. The open-source aspect was important to Lauderdale. “I would argue a big ethos of the nuclear industry is inherently open source in nature,” he said. In many industries, there is intense competition, and companies are reluctant to share information that might jeopardize their competitive advantage. But Lauderdale said the nuclear industry is just the opposite. “The statement I’ve heard over and over is: ‘An accident at any plant is an accident at every plant.’ As a result, you have so much sharing of data—whether it’s with the NRC, with INPO [Institute of Nuclear Power Operations], which is an organization that provides quality metrics to all these nuclear power plants—there is an ethos of openness, transparency, and sharing information back and forth,” he explained. “Transitioning to the technology and the code that we’ve built, we’ve used government resources to build this. There was the NRC data. There was Oak Ridge. And it’s our view that enabling artificial intelligence to understand nuclear terminology is so foundational to any AI application that’s going to be built, we want to open source this code, which is a way of saying, we want to make sure any party, even our competitors—people that might build competitive applications—are able to leverage this tool set as they build their own apps. And all we ask in exchange is that as they make improvements—as they add different features—that they commit back to the project,” he said.

The AI Revolution

Yet, Lauderdale suggested this is just the beginning. “Everyone hears there’s this big AI revolution that’s occurring, and we’re all as a society starting to realize the energy demands for artificial intelligence are going to be astronomical, to the point that Three Mile Island is going to be reopened for a Microsoft data center,” Lauderdale said. “We are at the tip of the iceberg, because all the hyperscalers are talking about 10x, 20x, crazy growth, and that is going to require reliable energy that’s available 24/7, that’s safe, and ideally not emitting carbon. So, nuclear is that path.” However, the path is not necessarily straightforward or easy. “It’s our view that not only is AI going to need nuclear, nuclear is actually going to need AI,” Lauderdale predicted. “I think there’s the opportunity to start applying artificial intelligence in very safe manners, in foundational manners.” While Lauderdale doesn’t think AI is ready to start running nuclear power plants on its own yet, he does believe it can provide benefits to operators and developers. “The ability to have AI help people search and find documents is a very worthwhile cause,” he said. “It’s where we started. And then from there, once you have that foundation, you build layer after layer and more advanced applications.” “Where the early wins can really occur is in the licensing process,” Tom Evans, leader of the High-Performance Computing (HPC) Methods for Nuclear Applications group in the Nuclear Energy and Fuel Cycles division at ORNL (Figure 2), told POWER. “The knowledge-based barriers to understanding and navigating the licensing process are large, and if you can do something to reduce those barriers significantly, you’ve already, just out of the gate, made a massive, massive improvement to the entire process and to the prospects for actually being able to deploy nuclear power more economically,” he said.

Yet, Evans suggested there are multiple vectors through which AI can play a role. One area is in complex design analysis. Evans said ORNL often helps the NRC and other industry stakeholders by doing this type of analysis for them using supercomputer models. He explained that analysts usually start by searching through previous analysis to find something similar that they can use as a reference. From there, they adjust inputs to account for the differences in the new design and then run simulations. Evans noted, however, that just identifying the most applicable past analysis to use as a starting point can take a great deal of time. This is where AI could improve the process. Rather than having an analyst search through reams of data, the AI tool, which has literally been trained on 53 million pages from the NRC’s ADAMS database, can quickly provide the most relevant files. This would save the analyst a lot of effort. Atomic Canyon has larger ambitions too. Lauderdale said his team is in discussions with several companies in the nuclear field, who have proprietary datasets they’d like to incorporate AI into. He said his group could install an enterprise version of the software that would ingest the proprietary data, allowing users to search their droves of internal data. “That’s kind of the next iteration of where we’re going,” said Lauderdale. Beyond his own company’s aspirations, Lauderdale suggested the nuclear industry isn’t as slow moving as some people may believe, and, in fact, could end up being a leader in the AI revolution. “One of the key ingredients you need to build fantastic world-class AI models is data. And the more data you have, the better the models you can build,” he said. “Because the nuclear power space has been documenting so much information, I actually think this space has the opportunity to become really a thought leader and an innovator in the artificial intelligence realm, and that’s what gets me really excited.”

—Aaron Larson is POWER’s executive editor.