A novel irradiation module for ANICCA fuel cycle code based on multi-task learning

Abstract

This paper introduces an alternative approach to irradiation modelling within the context of nuclear fuel cycle codes like ANICCA, the fuel cycle code developed by the Belgian Nuclear Research Centre (SCK CEN). The focus lies on upgrading the irradiation module by substituting the CRAM-based depletion calculations for a more flexible and innovative approach based on Multi-Task Learning (MTL). Utilizing data generated with SERPENT2 Monte Carlo simulations, two MTL neural networks are developed to surrogate irradiation processes of Uranium Oxide (UOX) and Mixed Oxide (MOX) fuels, respectively. MTL enables simultaneous learning of the evolution of the inventories for different observables - i.e., transuranium elements, fission products, minor actinides and fertile materials - offering improved predictive capabilities compared to non-MTL neural networks, as demonstrated in the cross-validation tests. Hyperparameters of the models were found using Bayesian optimization, resulting in superior model performance compared to the old CRAM-based model. Verification against SERPENT2, used as a reference code for comparison, demonstrates the stability and accuracy of the MTL-based models, outperforming the original CRAM method for the majority of predicted isotopes.