RAPID Enabled Physics-Based Neural Networks for Predicting 3-D Fission Distributions in JSI TRIGA Reactor

Abstract

The current methods for high-fidelity simulations of nuclear reactor systems are complex and computationally expensive. To reduce computation time, artificial intelligence (AI) and machine learning (ML) are being considered. Despite showing promise for solving various neutronics problems, the limited availability of high-fidelity data constrains ML applications to simpler problems or systems. This paper utilizes the RAPID code system for its effectiveness at rapidly producing large quantities of high-fidelity data. This has enabled the development of physics-based neural networks (NN) to predict 3-D fission distributions as a function of CR positions for the JSI TRIGA Mark-II research reactor. We developed a NN architecture that contains two hidden layers, 4400 neurons per hidden layer, with Leaky ReLU activation functions. This model was capable of predicting more than 99% of the fission values in the fuel elements within ±0.5% rel. diff. The model also predicted about 98% of the fission values in the fuel followers within ±10% rel. diff. It was determined that errors in the fuel follower predictions did not significantly impact calculated power peaking factors, which fell within the range of -0.39% to 0.91% rel. diff. Hyperparameter tuning and its effect on model performance is also discussed, with some comparisons to simpler ML models developed in a previous study.