UC Berkeley

Accelerated fuel qualification for advanced nuclear systems requires mechanistic understanding coupled with rigorous uncertainty quantification to compress traditional timelines. This dissertation investigates whether mechanistic models grounded in first-principles physics, systematically validated against experimental data across multiple operating regimes, and calibrated through Bayesian inference can substantially reduce epistemic uncertainty in fuel performance predictions while establishing defensible operational boundaries. The work applies this hypothesis to uranium mononitride (UN) fuel for micro-reactor deployment, addressing the challenge that historical UN irradiation data exhibit significant scatter and gaps at high burnup and elevated temperatures relevant to modern micro-reactor designs.A mechanistic model for fission gas release and swelling model was investigated in the BISON fuel performance code. The model explicitly captures temperature-dependent diffusion mechanisms through decomposition of effective diffusivity into thermal equilibrium, defect-assisted, and thermal components. Multi-population bubble tracking differentiates nucleation pathways, and grain boundary coalescence provides mechanistic basis for understanding saturation transitions and release. Validation against various historical irradiation test cases revealed regime-dependent model performance with robust agreement in low-temperature, low-burnup conditions and interpretable limitations at regime boundaries.The dissertation establishes that mechanistic modeling with rigorous uncertainty quantification may successfully enable compressed accelerated fuel qualification timelines. Broader methodological contributions include regime-stratified validation identifying mechanistic regime boundaries based on physics rather than arbitrary thresholds, surrogate-accelerated Bayesian inference achieving significant computational speedup while maintaining statistical rigor through hyperparameter optimization and empirical calibration inflation factors, and mechanistically guided prior specification improving parameter convergence by 30 to 50% while maintaining Bayesian validity. These approaches are generalizable to other high-dimensional inverse uncertainty quantification problems for expensive mechanistic models.The surrogate-accelerated Bayesian Markov chain Monte Carlo (MCMC), using latin hypercube sampling (LHS)-sampled input parameters to train Gaussian process emulators, achieved robust parameter within the primary regime for low temperature and low burnup. Parameter uncertainty was reduced by 80 to 90% through Bayesian inference, with posterior distributions full convergence (Gelman-Rubin R < 1.01). Forward propagation of calibrated posteriors produced predictions with 86.4% experimental coverage, 0.30 vol% mean absolute error for swelling, and 0.05% mean absolute error for fission gas release. Comparison of prior (uninformed) versus posterior predictions quantified information gain, with swelling prediction accuracy improving 78.6% and fission gas release accuracy improving 92.6%.