Content area
Abstract
Pebble bed reactors (PBRs), boasting superior safety features and online fueling, present unique challenges for fuel characterization and core management. This dissertation addresses these challenges at two levels: the measurement of discharged fuel pebbles, and the use of those measurements to predict highly delayed reactivity effects. The work focuses primarily on the Kairos Power generic high-temperature fluoride salt-cooled reactor (gFHR) benchmark and integrates high-fidelity depletion and transport modeling, optical design, machine learning (ML), and detector response simulation to develop new approaches for PBR monitoring and operation.
Chapter 2 explores the use of bent crystal diffraction (BCD) spectrometry for rapid, accurate burnup measurement of freshly discharged pebbles. The high activity and short measurement times inherent to PBR operation make conventional gamma spectroscopy difficult. BCD spectrometers, which leverage coherent scattering in highly perfect or mosaic crystals to act as narrow-band energy filters, are widely used in astrophysics and nuclear physics but have not yet been applied to continuously fueled reactors. Using Serpent 2 depletion and SHADOW 3 ray tracing, viable spectrometer configurations using mosaic silicon were designed and the corresponding pebble spectra were simulated. ML regression models trained on simulated measurements predicted burnup, residence time, radial path, fluence, and plutonium content with high accuracy (i.e. R2 0.995 for burnup) under conservative conditions for fuel decay time (1.5 days) and measurement periods (20 seconds).
Chapter 3 investigates core-level prediction using measurable pebble properties. Due to the high temperature and dynamic fuel bed, in-core instrumentation is impractical, while the consequences of changes to fuel insertion and operation conditions may take months to manifest. A Long Short-Term Memory (LSTM) recurrent neural network was developed to leverage operational history and batches of simulated pebble measurements to predict reactivity and principal components of flux and power meshes. Training data were generated with PEARLSim, a PBR simulator performing zone-wise depletion. On unseen data, the model achieved R2 values of 0.932 for reactivity, 0.925 for the first flux component, and 0.979 for the first power component. Forecasting capability was also implemented. The LSTM was integrated with a steering algorithm to guide PEARLSim through the transition core quickly while maintaining criticality, demonstrating the model’s ability to improve with subsequent simulations.
Chapter 4 extends the application of BCD spectrometry to molten salt reactors (MSRs), which share the challenge of measuring fuel with extremely high activity levels but with even shorter lived nuclides. Unit cell depletion models provided representative compositions and spectra, from which BCD designs were optimized for safeguards and burnup relevant isotopes. The results demonstrate that BCD spectrometry can isolate gammas from 239Np and enable confident measurement of plutonium production rates with a high signal-to-noise ratio. Clean burnup measurements using the 661 keV peak, however, were shown to be less feasible for MSRs than PBRs.
Chapter 5 outlines a pathway for accelerated pebble-wise depletion modeling using ML models trained on a combination of pebble features and interpolated values from coarse flux and power meshes. While not fully implemented, this approach could enable computationally efficient simulation of BCD measurements across diverse operating states, further improving the generality of regression models for pebble history prediction.
This dissertation demonstrates that bent crystal spectroscopy, combined with machine learning models and time series analysis methods, offers a powerful new toolkit for fuel measurement, safeguards, and operational optimization in continuously fueled advanced reactors. These methods are particularly valuable for safe operation during non-equilibrium core states and could be extended to a variety of next-generation designs.
