| Summary: | Accurately predicting the behavior of nuclear fuel performance is essential for the safe and economic operation of nuclear reactors. Computer codes of different fidelities have been developed over past decades to simulate the behavior of nuclear fuels, such as the multi-dimensional, parallel, finite element-based code BISON, and the NRC-auditing code FRAPCON. Multiple areas of research remain to be addressed in fuel performance while physics-based approaches often reach their limits. Studies to be presented in this thesis therefore revolve around applying data-driven methods to address these issues.
First, discrepancies always exist between code predictions and real-world responses, thus uncertainties must be quantified for the code predictions for benefit of decision making, operation safety and design optimization. Systematic validation and verification are performed for BISON first, followed by a holistic sensitivity analysis (SA) framework built upon a complete set of uncertain input parameters. The number of uncertain input parameters can be effectively reduced based on the obtained qualitative importance ranking, benefiting the subsequent uncertainty quantification (UQ). To enhance the predictability, a novel Bayesian inference framework is introduced to efficiently calibrate the expensive high fidelity tools, possibly without resorting to approximate surrogate methods. The calibrated prediction aligns better with experimental observations, and is subject to significantly reduced uncertainty.
Second, while full-core monitoring of fuel behaviors can provide the most realistic assessment of safety margins, its computational cost for use in design and operation optimization is prohibitive. Machine learning (ML) methods were used to construct fast-running full-core surrogates, which achieves a runtime acceleration of more than 10,000 (1,000) times compared to FRAPCON for the standard (high burnup) PWR cores, allowing for direct coupling of full-core fuel response into core design optimization in the future. Then for purpose of full-core PCI monitoring which requires BISON as the high-fidelity simulation tool, a physics-informed multi-fidelity ML framework is introduced to significantly reduce the number of necessary code runs. Finally, deep learning models are trained to predict the spatiotemporal distribution of the cladding hoop stress. The proposed data-driven methods for the selected applications enlightens the nuclear community on practical pathways to realize meaningful improvements in fuel performance assessment.
|
|---|