Speaker
Description
Accurate fission product yield (FPY) data are essential for reactor design and safety studies. Existing nuclear data libraries provide FPY only at limited neutron energies, leaving large gaps in the intermediate region that affect predictions for accelerator-driven systems (ADS) and advanced reactors. We developed a physics-informed machine-learning model using a Bayesian Neural Network (BNN) combined with a nuclear shell-structure factor and optimized by the Widely Applicable Information Criterion (WAIC). This approach [1,2] reproduces both global and fine structures of FPY distributions while maintaining physical consistency. The predicted energy dependence agrees well with recent experimental data [3], confirming the model’s reliability. Independent yields were used to calculate the production of delayed-neutron precursors and the energy-dependent delayed neutron yield (DNY). For key minor actinides such as 241Am, reliable DNY values were obtained for the first time, improving the understanding of reactivity control and safety margins in subcritical systems. The proposed framework demonstrates that integrating physical insight into machine learning can provide accurate and continuous nuclear data, enhancing the predictive capability of reactor simulations for next-generation nuclear systems.
[1] J. Chen et al., J. Nucl. Sci. Technol. 61, 1509–1520 (2024).
[2] J. Chen et al.. PRL submitted.
[3] A. Tonchev et al., Phys. Rev. C 111, 054620 (2025).
