EPRI explores how AI can address the nuclear power industry’s welder shortage
Nuclear power is grabbing headlines. For example, early in 2025, French President Emanuel Macron and India’s Prime Minister Narendra Modi announced plans to jointly develop small modular reactors (SMR). In 2024, Constellation Energy and Microsoft declared their intention to restart a closed reactor at Three Mile Island to generate carbon-free energy for data centers the tech giant needs to power artificial intelligence (AI). “Nuclear energy is making a strong comeback, with rising investment, new technology advances and supportive policies in over 40 countries,” said the International Energy Agency (IEA) in a new report, The Path to a New Era for Nuclear Energy.
While the IEA forecasts global nuclear capacity growth will reach between 650 gigawatts to over 1000 gigawatts by 2050, there remain significant policy and financial challenges. One little-mentioned challenge—a hurdle both for new construction and maintenance on existing nuclear power plants—is a shortage of skilled welders to build, replace, reinforce, or repair piping systems, reactor vessels, valves, and other critical infrastructure. In the U.S., for example, the American Welding Society estimates there will be a shortage of 400,000 welders across the economy.
“Experienced welders are retiring, and industries are struggling to replace those roles with the same quality of welders,” said Jon Tatman, an EPRI principal team lead whose research focuses on developing welding solutions for the electric power industry. “We don’t expect that trend to stop anytime soon.” The labor shortage doesn’t just impact welders. There also aren’t enough inspectors to ensure that welds are correctly done, a cause of potential project delays even when welders are available and work quickly and efficiently.
While efforts to encourage and train more young people to pursue welding careers are important and necessary, improved recruitment and retention won’t be enough to tackle the labor supply and demand gap. Advances in artificial intelligence (AI) and sensor technology have the potential to both address the welder shortage and significantly improve the precision, quality, and safety of welds. EPRI has been collaborating with the Fraunhofer Institute and the welding equipment supplier Liburdi Dimetrics to develop and commercialize AI-driven adaptive welding capabilities.
How Sensors, Cameras, and AI Can Mimic a Skilled Welder
The current approach to mechanized welding at nuclear projects combines both automation and the guidance of a skilled welder. At a very basic level, a welder will observe the speed, position, voltage, and other aspects of the wire making the weld to ensure its quality. “You have an operator watching the weld and adjusting it to ensure nothing goes off course. They’re watching it remotely, so even though it’s mostly automated, you still require an operator to sit there and watch,” said Nick Mohr, an EPRI program manager and welding engineer.
EPRI and its partners have explored how a combination of cameras, sensors, and algorithms can potentially eliminate the need for a skilled welder to continuously observe a weld and sometimes adjust it to ensure its quality. The welding system uses two algorithms, both developed by Germany’s Fraunhofer Institute. The first analyzes images of a weld in real-time and adjusts the wire to ensure a precise and safe weld. The algorithm examines a host of features, like the weld pool (the melted metal), the wire, and the groove, to determine whether any adjustments are needed.
“The camera continuously monitors the weld in real-time and collects data, which the algorithm processes in real-time,” Tatman said. “It identifies different key areas of the weld, like the weld pool, the welding wire, and the groove geometry. By analyzing different aspects of the weld, the algorithm effectively acts like a skilled welder, making decisions and adjustments necessary to maintain optimal welding conditions.”
The algorithm’s ability to adjust depends on its training. EPRI and Liburdi Dimetrics worked with AI experts at Fraunhofer to provide images and data to instruct the algorithm about what good and bad welds look like. “You really need welding expertise upfront to guide the AI development,” Tatman said. “Without that baseline of knowledge built into the algorithm, it won’t be able to make the right adjustments. The cameras and the sensors are like the welder’s eyes and ears, and the algorithm is like the brain. All of that must be working well to automate these welds.”
One algorithm continuously monitors and adjusts weld settings in real time, fine-tuning the welding wire’s entry position into the weld pool to maintain optimal conditions. Meanwhile, a second, future-focused algorithm predicts the optimal weld parameters for adding new layers of metal. This predictive capability ensures process consistency and preemptively prevents potential defect formation.
Adaptive welding also has the potential to eliminate the need for post-weld inspections. If algorithms become sufficiently precise, real-time monitoring and inspections could follow along with the automated welds and immediately raise potential issues to address and correct. “If you can inspect as you’re welding, you could save a lot of time on the backend in terms of your post-inspection,” Mohr said. “We haven’t done that yet, but that is a potential future research effort.”
Exploring Adaptive Welding Applications
While future research to improve adaptive welding is necessary, it is not purely a laboratory technology. Liburdi Dimetrics, a welding vendor, has already licensed the technology from EPRI and begun to incorporate the algorithms into their welding systems. Liburdi Dimetrics has also identified about 10 applications where adaptive welding can be used.
EPRI and Liburdi Dimetrics have already worked with the Fluor and the Central Plateau Cleanup Company to leverage and directly apply this developed adaptive welding technology. The application involves a dry storage canister seal welding application to contain highly radioactive nuclear waste, a scenario where it’s important to keep welding personnel away from dangerous radiation levels. “You can’t have an operator nearby to monitor what’s happening,” Tatman said. “The driver of fully automating the process is safety concerns.” A similar system could be used for commercial nuclear spent fuel canister welding, and another possible future application is for the welding needed at a plant’s spent fuel pool liner.
The specific applications identified can also guide future algorithm development. For instance, algorithms could be trained to automatically shut down a weld if it recognizes a condition that indicates a safety concern or a faulty weld.
Regardless of the potential applications identified for adaptive welding, its implementation into real-world scenarios will inevitably require building trust among operators. “For the first few times in the field, there are going to be people watching. Mechanized welding always has an emergency stop,” Mohr said. “You’re going to have an expert welder there looking at it and using adaptive welding won’t save time and money at first. But as trust in the model builds, I think you’re going to get to the point where the welder sets this up, gets it started, and walks away.”
As much as adaptive welding has the potential to improve safety, precision, and efficiency, one thing it won’t do is fully replace the need for actual human welders. Indeed, even the use of adaptive welding requires a skilled welder to set up and operate the system. In the case of a circular storage container, where the same weld is necessary repeatedly, a welder will still be needed to initiate the automated welds.
“There is always going to be a need for manual welders,” Mohr said. “In many cases, it’s just going to be quicker to send a welder to do a job rather than getting all the equipment needed for adaptive welding into the field and setup. What adaptive welding provides is another tool to improve welding consistency and address the labor shortage.”
EPRI Technical Experts:
Jon Tatman, Nick Mohr
For more information, contact techexpert@eprijournal.com.
