A California startup is deploying what it says is the first commercial installation of generative AI at a US nuclear power plant, but don’t get too excited (or worried) about what it’s going to be doing quite yet – it’s a pretty run-of-the-mill use in an enterprise environment.
Atomic Canyon and Pacific Gas and Electric (PG&E) today announced plans to deploy the former’s Neutron Enterprise AI platform at the Diablo Canyon Power Plant in California, the only nuclear power plant left in the state. Rather than do something far-fetched, or potentially dangerous, Neutron Enterprise will instead be turning a full Nvidia AI stack installed at the plant to speed up document search and retrieval.
That’s not to say such work isn’t important in the nuclear power industry, mind you. As a PG&E spokesperson Suzanne Hosn told The Register, something as simple as pouring vast amounts of technical documentation and regulatory requirements into an AI system capable of retrieval-augmented generation (RAG) has huge potential to save plant operators and engineers time.
"The ability to quickly locate and retrieve specific documents is not just a matter of convenience – it’s an imperative," Hosn told us.
Different groups at a nuclear power plant – engineering, operations, and regulatory officials, for example – all have to maintain and be ready to comb documents for references regarding various analyses, reports, and evaluations, and all those documents are rarely centralized, Hosn said. Regardless, everyone still has to search through documents to find what they need – the perfect task for RAG-capable AI trained on nuclear regulations and the history of Diablo Canyon’s operations.
As one example, Atomic Canyon explained to The Register what would need to be reviewed to replace a section of concrete in a safety-related structure. To do so appropriately, engineers would have to review the original design basis, study its complete maintenance and modification history, ensure compliance with new design standards without deviating too greatly from the original design basis, verify changes will meet Nuclear Regulatory Commission (NRC) standards, and document everything they do.
"Missing a critical historical document could lead to decisions that compromise the structure’s ability to perform its safety function or maintain regulatory compliance," an Atomic Canyon spokesperson told us.
It can take weeks to find the right documents in order to begin such operations, Atomic Canyon added.
"Diablo Canyon is positioned to significantly enhance its data management capabilities, potentially setting a new standard for efficiency and timelines," Hosn said. "We’ll enhance the expertise and focus of our people by uncovering and connecting vast data sets and reducing required but tedious administrative execution."
And this is just the start, she said.
"In the future, we’ll add in additional areas, such as design, engineering, and maintenance. It will be a phased approach," Hosn noted. Atomic Canyon confirmed plans to expand into other areas, including supporting operational decision making, depending on the results of the document retrieval rollout.
New AI industry, same AI task – for now
Development of Neutron Enterprise, which Atomic Canyon describes as "the first open source sentence embeddings solution tailored specifically for nuclear data," was done in collaboration with Oak Ridge National Laboratory (ORNL) in Tennessee. The models, dubbed Fermi (not to be confused with ORNL’s own FERMI, or Fusion Energy Reactor Models Integrator), are tailored for the nuclear industry, but are the same old optical character recognition-powered RAG systems used in various other industries.
Atomic Canyon used ORNL’s Frontier supercomputer to train its Fermi AI models on Nuclear Regulatory documents, the company said in a September press release announcing its work with the laboratory. According to the company, its AI "returns the correct search result within the top ten results about 98 percent of the time, and within the top five results approximately 93 percent of the time" based on "a new industry-first evaluation benchmark," though what it’s comparing itself to is unclear.
To operate on site, Diablo Canyon will be getting its very own full-stack Nvidia AI solution, including enterprise hardware, Nvidia Triton Inference Server software, and an unknown quantity of Nvidia Hopper GPUs.
That said, AI isn’t exactly new in the nuclear industry – but the generative element is. According to International Atomic Energy Agency (IAEA), there’s a good reason for a lack of generative AI in the nuclear industry – it’s too unpredictable.
"While generative AI can help with administrative tasks, as in other industries, its use in operating nuclear power plants is not yet possible due to its novelty and opaqueness," the IAEA said in a September 2023 bulletin. "It is not yet entirely understood how artificial networks function and come to conclusions."
This isn’t a new issue in AI either – concerns about a lack of explainability have been cropping up for years now, and with systems as sensitive as nuclear power opening the door to generative systems, explainability is more important than ever. The IAEA noted that efforts are under way to develop such systems, but until we get to that point, it’s not likely generative AI will have much role in the nuclear industry beyond helping operators find the right policies and regulations in less time.
While waiting for private companies to develop their own solutions to the explainability hurdle in nuclear power AI, the IAEA is taking its own steps to "build confidence" in AI applications for the nuclear industry including reactor design, plant operations, and training by partnering with Purdue University’s Center for Science of Information (CSI).
"Without reliable quantification, the nuclear community’s ability to realize the potential of AI will be diminished and this will negatively impact its ability to remain competitive in the energy market," said CSI nuclear engineering professor Hany Abdel-Khalik.
Atomic Canyon told us that explainability concerns are a key reason for it starting with document retrieval before advancing into more delicate tasks.
"Using RAG architecture means every Neutron Enterprise response is grounded in specific, verifiable source documents from a controlled knowledge base of nuclear industry documentation," Atomic Canyon said. "This provides clear evidence chains and reduces the risk of hallucination or incorrect information."
Atomic Canyon said that, as it expands Neutron Enterprise’s capabilities, it intends to stick with a RAG model that’s explainable rather than having to add such capabilities after the fact, but the company acknowledged explainability remains a problem for AI development across the entire industry.
"Our focus on nuclear-specific solutions and foundation in document retrieval positions us to address these challenges systematically as we expand Neutron Enterprise’s capabilities," Atomic Canyon said.
Atomic Canyon said the Nvidia GPUs and other hardware are being deployed at Diablo Canyon, with plans to get the whole system up and running by early next year. The company is also in talks with "potential customers across the nuclear energy sector," but it’s entirely possible those additional revenue sources, like the expansion of Neutron Enterprise’s capabilities, may depend on how well the system performs in its first commercial rollout.
Time might be limited at Diablo Canyon, though. The plant was scheduled to shut down next year, but received a temporary extension to keep operating while its owners try to convince the NRC to give it 20 more years of operation. State officials only want Diablo Canyon operating through 2030 to help ease California into more renewable energy generation, meaning the plant’s future beyond the next five years is uncertain, AI-enabled efficiencies or no. ®
{Categories} _Category: Implications{/Categories}
{URL}https://www.theregister.com/2024/11/13/nuclear_plant_generative_ai/{/URL}
{Author}unknown{/Author}
{Image}https://regmedia.co.uk/2024/03/25/nuclear_shutterstock.jpg{/Image}
{Keywords}{/Keywords}
{Source}Implications{/Source}
{Thumb}{/Thumb}