Low-fidelity and High-fidelity Data Processing and Machine Learning Model Development for BWR Core Simulation

The integration of low-fidelity (LF) and high-fidelity (HF) data in nuclear reactor simulations has been a significant advancement in the field of nuclear engineering. In a recent study, researchers have developed a machine learning (ML) model that combines LF data from the Purdue Advanced Reactor Core Simulator (PARCS) with HF data obtained from Serpent Monte Carlo simulations to predict reactor parameters with higher accuracy.

The LF model, based on the PARCS code, consists of a quarter symmetry model of a boiling water reactor (BWR) core with 560 fuel bundles encircled by reflectors. The HF data, generated using Serpent simulations, provide higher resolution and detailed geometry structures to reproduce PARCS solutions under the same core conditions. The data collected from both LF and HF simulations were processed and normalized for training, validation, and testing of the ML model.

The ML model, named BWR-ComodoNet, is based on a 3D–2D convolutional neural network architecture that processes spatial data in their actual dimensions. The model takes input features such as core flow rate, control rod pattern, and nodal exposure to predict errors in the core eigenvalue (\(k_{eff}\)) and nodal power distribution. The predicted errors are then added to the LF data to obtain the predicted HF parameters.

The optimization of the ML model was conducted using Bayesian Optimization to minimize validation loss, with hyperparameter tuning over 500 trials. The model architecture includes ReLU activation functions, dropout layers for regularization, early stopping, and a learning rate schedule to improve generalizability and prevent overfitting.

The study demonstrates the potential of integrating LF and HF data with ML models to enhance the accuracy of predicting reactor parameters. By combining the strengths of both simulation approaches, researchers can achieve more reliable and efficient nuclear reactor simulations, leading to advancements in reactor design, safety, and performance optimization.