AI Deep Learning Tool for High-Fidelity Physics Simulations of Nuclear Reactor Systems

Nuclear reactors provide operators with hundreds of data channels that indicate the state of the reactor. During an accident, the flood of information from these channels can be overwhelming, making it difficult for operators to determine the correct actions to reduce the accident impact. Reactor accidents negatively impact public perception of nuclear power and the commercial nuclear industry as a whole.
Objective: This project will generate an Artificial Intelligence (AI) system that will have the capability to aggregate the available data from the reactor instrumentation and correctly identify the reactor condition. During an accident, the AI tool will be able to provide operators with information about the accident and the data that was used in that determination. This tool will help operators to respond correctly to reactor accidents and will also help inform better placement of instrumentation.
Phase I Effort: Development and training of the AI system will commence in Phase I for a prototypical reactor event. Initially, the AI algorithm will be trained to recognize a range of loss of coolant accident (LOCA) events for a generic Westinghouse three-loop reactor unit. At this stage, the cases used to train the AI will be based on simulated data from thousands of reactor system simulations using an analysis code such as TRACE. Once trained, the AI system will be validated against its ability to correctly distinguish a LOCA event from non-LOCA events or normal reactor operations. A series of challenge scenarios will then be presented that include missing information simulating sensor failures in a reactor system. Output from the AI tool will include a confidence level regarding the event detection, along with an importance weighting to help identify the sensors with the most predictive value.
Commercial Applications: The AI tool developed in Phases I, II, and III will be extremely valuable to nuclear utilities and will help operators to run nuclear plants more safely. Using deep learning neural networks to aggregate large volumes data and recognize patterns that reveal the potential for an accident prior to any excursion would help ensure the safe and reliable operation of reactor systems. An improved public perception of the safety of nuclear power would lead to increased likelihood of future nuclear plants being built, therefore reducing the overall climate impacts from fossil fuels.