Surrogate-assisted prediction of radionuclide inventories for radiochemical characterization of activated CANDU reactor materials

Abstract

Accurate prediction of nuclide-specific activity is important for radiological characterization of activated structures during nuclear power plant decommissioning. However, repeated calculations become computationally demanding when neutron flux, impurity concentration, and cooling time vary. In addition, a surrogate model that directly predicts final activity must simultaneously represent irradiation-dependent nuclide formation and post-irradiation decay, which can reduce accuracy for decay-chain-dominated radionuclides. This study evaluates three surrogate approaches for activated CANDU structural materials. The three-variable response-surface surrogate (3RSS) directly predicts final activity. The response surface–decay model (RSDM) predicts end-of-irradiation (EOI) activity using a quadratic response surface and subsequently applies physical decay calculations. The hybrid surrogate–decay model (HSDM) replaces the response surface with Gaussian process regression while retaining the same decay module. Performance was evaluated using 2,000 independent Sobol verification cases. The RSDM and HSDM generally outperformed the 3RSS, demonstrating the benefit of separating EOI prediction from cooling-time calculations. The HSDM achieved the lowest errors for most radionuclides and showed lower sensitivity to limited training data. These results indicate that the HSDM provides an accurate and computationally efficient approximation of MCNP–ORIGEN calculations within the predefined input domain.