Physics-Informed Machine Learning Methodologies Using RAPID for Predicting Eigenvalue and 3-D Fission Distribution in JSI TRIGA Mark-II Research Reactor

Abstract

The most common methodologies for high-fidelity simulations of nuclear reactors are very slow and require significant computer resources. Machine learning (ML) enables computers the ability to learn from data, allowing well-trained models to produce results quickly and accurately. However, the challenges of machine learning involve proper algorithm selection and, more importantly, the quantity/quality of the data. The RAPID code system enables very fast and accurate simulations of nuclear reactor systems. RAPID's algorithms have been both computationally and experimentally validated using the JSI TRIGA Mark-II research reactor. Accordingly, RAPID was used to generate a complete physics-informed dataset for training ML models to predict the system eigenvalue (keff) and 3-D fission neutron distributions as a function of various control rod configurations. A total of 157,324 high-fidelity simulations were performed to generate training and testing datasets, which contain keff and 3-D fission distributions. The ML models analyzed include linear regression, polynomial regression, k-nearest neighbors (kNN), random forest (RF), and neural networks (NN). Results show that kNN regression is both fast and accurate for calculating keff, achieving an RMSE score of 26.37 pcm in under one second. For predicting 3-D fission distributions, ML models were evaluated across core and control rod fuel follower (CR-FF) regions. The NN model accurately predicted ~99 % of core fission values within ±0.5 % rel. diff. and ~98 % of the CR-FF fission values within ±10 % rel. diff. in under 10 seconds. Power peaking factors calculated from the NN model's predicted fission values fell within -0.39 % to 0.91 % rel. diff., demonstrating that larger errors in the CR-FF regions had minimal impact.