A Bayesian Network-Based Framework for Predicting Fission Charge Yields

Abstract

Predicting neutron-induced fission charge yields is a long-standing challenge due to the complex, non-equilibrium nature of the fission process and the limited availability of experimental data across a broad range of neutron energies. This work presents a Bayesian machine learning framework that integrates Gaussian Process Regression (GPR), Bayesian Neural Networks (BNNs), and Mixture Density Network (MDN) outputs to address these challenges. The GPR component augments sparse datasets with synthetic samples and associated uncertainties, while BNN layers capture model uncertainty through variational inference. The MDN output models the multimodal nature of charge yields and accounts for data-driven (aleatoric) uncertainty. The proposed model accurately reproduces known features of fission yields, such as odd-even staggering and energy dependence, and generalizes well to isotopes not included in the training set. Its probabilistic predictions are consistent with experimental observations and semi-empirical models, offering a robust tool for fission yield modeling.