AI Predicts Nuclear Reactor Failures 1,400 Times Faster With Virtual Sensors

Monitoring nuclear reactors just got a major upgrade! AI-powered virtual sensors now deliver lightning-fast predictions of critical failures, helping engineers detect risks before they escalate. This groundbreaking tool offers real-time insights no physical sensor ever could.

Research: Virtual sensing-enabled digital twin framework for real-time monitoring of nuclear systems leveraging deep neural operators. Image Credit: IndustryAndTravel / Shutterstock

Whether it’s for your vehicle or your home, from small-scale uses to the largest, the debate over the most efficient and cost-effective fuels continues. Currently, there’s no shortage of options either.

Nuclear power provides an alternative to more conventional energy options but requires rigorous systems monitoring and safety procedures. Machine learning could make monitoring key elements of nuclear systems easier and respond faster to issues.

Syed Bahauddin Alam, an assistant professor in the Department of Nuclear, Plasma & Radiological Engineering (NPRE) in the Grainger College of Engineering at the University of Illinois Urbana-Champaign, and his team worked with artificial-intelligence and machine-learning experts through Illinois Computes to develop a novel method for real-time monitoring of nuclear energy systems that can infer predictions about 1,400 times faster than traditional Computational Fluid Dynamics (CFD) simulations. NCSA research assistants and NPRE graduate students Kazuma Kobayashi and Farid Ahmed assisted in the development. 

Published in Nature’s npj Materials Degradation, Alam’s research introduces machine learning-driven virtual sensors based on deep-learning operator-surrogate models as a complement to physical sensors in monitoring critical degradation indicators. Traditional physical sensors face limitations, particularly in measuring critical parameters in hard-to-reach or harsh environments, often resulting in incomplete data coverage. Moreover, traditional physics-based numerical modeling methods like CFD are still too slow to provide real-time predictions in nuclear power facilities.

Instead, the novel Deep Operator Neural Networks (DeepONet), when properly trained on graphics processing units (GPUs), can instantly and accurately predict complete multiphysics solutions on the entire domain. DeepONet functions as real-time virtual sensors and addresses these limitations of physical sensors or classical modeling predictions, specifically by predicting key thermal-hydraulic parameters in the hot leg of a pressurized water reactor.

Because components are continuously subjected to extreme temperatures, pressures, and radiation, proper monitoring and inspection of in-service elements of nuclear reactors is essential for long-term safety and efficiency. AI isn’t replacing human oversight but creating new ways to monitor and predict the potential failure of system elements.

“Our research introduces a new way to keep nuclear systems safe by using advanced machine-learning techniques to monitor critical conditions in real-time,” Alam said. “Traditionally, measuring certain parameters inside nuclear reactors has been incredibly challenging because they’re often in hard-to-reach or extremely harsh environments. Our approach leverages virtual sensors powered by algorithms to predict crucial thermal and flow conditions without needing physical sensors everywhere.

“Think of it like having a virtual map of how the reactor is operating, giving us constant feedback without having to place physical instruments in risky spots. This not only speeds up the monitoring process but also makes it significantly more accurate and reliable. By doing this, we can detect potential issues before they become serious, enhancing both safety and efficiency.”

Through the campus-funded Illinois Computes program, Alam utilized allocations on NCSA’s Delta, performing computations for data generation on central processing unit (CPU) nodes and training and evaluation tasks on a computational node with  NVIDIA A100 GPUs. He collaborated with NCSA’s experts in AI-driven scientific computing and high-performance computing.

Massive thanks to Illinois Computes for funding this research. Partnering with Dr. Diab Abueidda and Dr. Seid Koric from NCSA was essential to our success. Through the campus-funded program, we leveraged Delta’s state-of-the-art supercomputing resources, including a computational node with NVIDIA A100 GPUs, to train and test our models efficiently. The NCSA technical staff provided invaluable support throughout the entire process, demonstrating the tremendous impact of combining AI with high-performance computing to advance nuclear safety. We will continue to work on unleashing the power of AI in complex energy systems, pushing the boundaries of what is possible to enhance safety, efficiency and reliability.

Syed Bahauddin Alam, assistant professor in the Department of Nuclear, Plasma & Radiological Engineering

“In this Illinois Computes project, we have fully utilized the unique high-performance computing resources and multidisciplinary expertise at NCSA and the Grainger College of Engineering to advance translational and transformative engineering research in Illinois,” said Seid Koric, senior technical associate director for Research Consulting at NCSA and research professor at the Department of Mechanical Science and Engineering.

“This collaboration exemplifies the synergy that emerges when advanced AI methods, high-performance computing resources, and domain expertise converge,” said Abueidda, a research scientist at NCSA. “Working alongside Dr. Alam’s team and NCSA’s AI and HPC experts, we leveraged the U.S. National Science Foundation-funded Delta’s cutting-edge capabilities to push the boundaries of real-time monitoring and predictive analysis in nuclear systems. By uniting our specialized skill sets, we have accelerated research while enhancing the accuracy and reliability of critical safety measures.

“We look forward to continuing this interdisciplinary approach to drive transformative solutions for complex energy systems. Ultimately, these breakthroughs highlight the promise of computational science in addressing the pressing challenges of nuclear energy.”