Transfer learning for the development and analysis of spent fuel libraries for a Lead-Cooled Small Modular Reactor

The increasing demand for accurate and efficient nuclear fuel cycle analysis in advanced reactor systems has highlighted the limitations of traditional high-fidelity simulations due to their computational cost.
This thesis proposes a hybrid modeling approach that integrates deep learning and transfer learning to predict the isotopic composition of spent fuel in Lead-cooled Fast Reactors (LFR), using the ALFRED demonstrator as a reference system. A comprehensive dataset was generated using Monte Carlo based
burnup calculations for both pin-cell and full-core configurations. A neural network (NN) was developed and trained on pin-cell data to estimate nuclide inventories across a wide range of operational parameters. To extend the model's predictive capability to core-level scenarios, a transfer learning
strategy was applied by fine-tuning the pre-trained model using a limited number of full-core simulations. Results demonstrate that the proposed framework significantly improves prediction accuracy. Particularly for minor actinides and long-lived fission products, even under data-constrained
conditions. This approach offers a promising pathway to accelerate reactor fuel cycle analyses and supports the development of real-time digital tools for next-generation nuclear systems.