Integrating core physics and machine learning for improved parameter prediction in boiling water reactor operations.

Area of Science:

• Nuclear Engineering
• Computational Physics
• Machine Learning Applications

Background:

• Accurate simulation of Boiling Water Reactor (BWR) operations is critical for nuclear fuel management and safety adherence.
• High-fidelity simulations offer precision but are computationally prohibitive for real-time applications.
• Existing low-fidelity methods require refinement to meet operational accuracy demands.

Purpose of the Study:

• To develop and validate a novel machine learning (ML) based method for enhancing the accuracy of BWR operation simulations.
• To reduce computational burdens associated with high-fidelity nuclear reactor simulations.
• To improve the prediction of key nuclear reactor parameters like core eigenvalue and power distribution.

Main Methods:

• Integration of a machine learning model with conventional two-step (lattice physics followed by nodal diffusion) BWR simulation techniques.
• Training a neural network on the discrepancies between high-fidelity and low-fidelity simulation results.
• Focusing ML model development on error correction for regular BWR operational scenarios.

Main Results:

• The ML-based error correction reduced nodal power error in low-fidelity simulations to approximately 1% on average.
• Core eigenvalue prediction accuracy was improved to under 100 pcm.
• The enhanced simulation method demonstrated effectiveness under normal variations in control rod patterns and core flow rates.

Conclusions:

• The proposed ML-integrated approach offers a computationally feasible pathway to highly accurate BWR operation simulations.
• This methodology provides a promising solution for enhancing nuclear reactor operation and management strategies.
• The study highlights the potential of targeted ML applications in refining established simulation workflows for nuclear engineering.