Transfer Learning for Multifidelity Surrogate Modeling of Fuel Irradiation in Lead-Cooled Fast Reactors

Pin cell depletion models are widely used in fuel cycle analyses but can miss core-level spectral and leakage effects that are particularly important in fast reactors, leading to biased nuclide buildup and discharge inventories. Conversely, full-core Monte Carlo depletion provides higher fidelity but is too expensive for large parametric studies. This work proposes a data-efficient multifidelity surrogate based on transfer learning (TL) to bridge pin cell and full-core physics for small modular reactor–lead-cooled fast reactor spent fuel inventory prediction.A feedforward neural network is first pretrained on a large Serpent 2 pin cell data set (nine inputs, 151 nuclide mass density outputs) and then fine-tuned using a limited full-core ALFRED Serpent 2 data set generated with a sequential experimental design to ensure robust space-filling coverage.Fivefold cross-validation shows that TL achieves full-core test losses on the order of 10−6 with only a few thousand full-core samples and outperforms both training from scratch on the high-fidelity data set and direct application of the pin cell model to full-core conditions. Nuclidewise evaluations at discharge indicate systematic error reductions for important actinides and fission products, and benchmarking against the ALFRED equilibrium core depletion calculations shows consistent improvements in batchwise inventory agreement.Finally, the surrogate is integrated into the ANICCA fuel cycle irradiation module to showcase its performance in the context of fuel cycle analysis.