In the realm of energy research, predicting critical heat flux (CHF) is a crucial task for ensuring the safety and efficiency of nuclear reactors. A team of researchers from the University of Pittsburgh, including Aidan Furlong, Robert Salko, Xingang Zhao, and Xu Wu, has been exploring the use of machine learning (ML) models to improve CHF predictions. Their latest study, published in the journal Nuclear Engineering and Design, investigates the application of these models to rod bundle geometries, which are more representative of actual reactor cores.
Previous research has shown that ML models can outperform conventional methods like empirical correlations or lookup tables (LUTs) in predicting CHF for isolated subchannels. However, rod bundles present additional complexities, such as crossflow between channels, pressure losses due to spacer grids, and the influence of unheated conductors. The researchers aimed to determine if ML models trained on tube-based CHF data could be generalized to predict CHF in rod bundles.
The team implemented three ML-based approaches in the CTF subchannel code: a purely data-driven deep neural network (DNN) and two hybrid bias-correction models. They then used these models to predict CHF location and magnitude in the Combustion Engineering 5-by-5 bundle CHF test series. To evaluate the performance of the ML models, they compared the results with three baseline models: the W-3 correlation, the Bowring correlation, and the Groeneveld LUT.
The results showed that, on average, all three ML-based approaches produced more accurate predictions of CHF magnitude and location than the baseline models. Among the ML models, the hybrid LUT model exhibited the most favorable performance metrics. This suggests that ML models trained on tube-based CHF data can indeed be generalized to predict CHF in rod bundles, offering a promising avenue for improving the safety and efficiency of nuclear reactors.
The practical applications of this research for the energy sector are significant. More accurate CHF predictions can lead to better-informed decisions about reactor design and operation, ultimately enhancing the safety and performance of nuclear power plants. As the researchers continue to refine these models, they could become an invaluable tool for the energy industry.
This article is based on research available at arXiv.