Artificial intelligence watchdog uses virtual eyes to change nuclear safety, humans cannot walk

Inside the nuclear reactor, the surge in temperature and radiation make human inspections impossible, and new digital guardians are standing. Researchers have developed artificial intelligence that can predict potential device failures 1,400 times faster than traditional methods, which could revolutionize the security of one of our most powerful but demanding energy sources.

Breakthrough researchers from the University of Illinois Urbana-Champaign, who have trained AI to act as “virtual sensors” in places that are too hostile to physical measurement devices. The technology, published in Nature’s NPJ Material Degradation, could change the way we monitor the most critical and inaccessible parts of nuclear power plants.

“Our research introduces a new approach to ensuring the safety of nuclear systems by using advanced machine learning techniques to monitor critical conditions in real time,” said Syed Bahauddin Alam, assistant professor in the Department of Nuclear, Plasma and Radiological Engineering. “Traditionally, measuring certain parameters within nuclear reactors is very challenging because they are often in difficult to reach or extremely strict environments.”

This innovation addresses the ongoing challenge in nuclear power generation: how to monitor components that continue to encounter extreme conditions. Current physical sensors cannot be placed anywhere they need, leaving potential blind spots in the security monitoring system.

Alam’s team developed what they call the DeepOnet, which, once trained, can immediately predict complete thermal and hydraulic conditions throughout the reactor system. AI is specifically aimed at monitoring the “hot legs” of pressurized water reactors, a key component in which water flows overheated.

“Thinking of it as having a virtual map that illustrates how the reactor is operating without having to put physical instruments in a blob of risk is what keeps us feedbacking,” Alam explained. “This not only speeds up the surveillance process, but also makes it more accurate and reliable. By doing so, we can detect potential problems before they become serious, thus improving safety and efficiency.”

The study leverages NCSA’s Delta supercomputer to train complex AI models using its advanced NVIDIA A100 GPU. This computational muscle is crucial for teaching AI to predict using the speed and accuracy required for real-world applications.

“I’m very grateful to Illinois Computing for funding this study,” Alam said. “Working with Dr. Diab Abueidda and Dr. Seid Koric from NCSA is critical to our success. Through the campus-funded program, we leverage Delta’s state-of-the-art supercomputing resources, including compute nodes using NVIDIA A100 GPUs to train and test our models.”

Collaboration between nuclear engineers and AI experts shows how interdisciplinary approaches address challenges that were previously insurmountable. By combining domain expertise with cutting-edge computing, the team created a system that not only collects data, but also can actively interpret it.

“This collaboration embodies the synergies that emerge when advanced AI approaches, high-performance computing resources and domain expertise,” said Abueidda, a research scientist at NCSA. “By uniting our expertise, we accelerate research while improving the accuracy and reliability of critical security measures.”

As nuclear power continues to be evaluated as a carbon-free energy alternative, innovations like this AI watchdog can help address concerns about security and maintenance while optimizing plant operations. Researchers are now exploring the extended applications of their virtual sensing technology in other complex energy systems.