Preliminary development of machine learning-based error correction model for low-fidelity reactor physics simulation

Abstract

Better prediction capability in reactor simulation procedures can result in better fuel planning, increased safety, and compliance with the Technical Specifications. Motivated by this necessity in the nuclear industry, we develop a method to improve the current reactor core simulation process using a machine learning approach. With a well-trained machine learning model, it is possible to predict the errors of the low-fidelity diffusion-based core simulator without a significant increase in complexity and computational cost. For the machine learning models, we have tested two different models based on Deep Neural Network and Extreme Gradient Boosting trained on high-fidelity Monte Carlo reactor simulation data. The proposed method has been verified in this work on simple 2x2 boiling water reactor color sets. We collected large data points that include different variations of assembly configuration, burnup, void fraction, and control blade insertion in both low-fidelity and high-fidelity data. The developed models can accurately predict errors in eigenvalue and assembly power. Utilizing the predicted errors, the machine learning-aided simulation results in a significant improvement over the conventional reactor simulation approach.