Study of the average fission-neutron multiplicity using machine learning

摘要: The average fission-neutron multiplicity ¯ν sets the neutron balance of fission systems and is a key
input to evaluated nuclear data libraries. We train four machine-learning models, a random forest
(RF), a Gaussian process regression (GPR), a Bayesian neural network (BNN), and a Kolmogorov–
Arnold network (KAN), on the JENDL-5 evaluated ¯ν(E) data for the total (MT452), prompt
(MT456), and delayed (MT455) multiplicities (1342 points for 88 nuclides and 1065 points for 78
nuclides). On a random 80/20 split all four models reproduce the total and prompt multiplicities
with test R2 ≃ 0.987–0.998, an interpolation benchmark because five of the six features are constant
per nuclide. When the test set contains only nuclides unseen in training, the KAN retains R2 = 0.983
on the total and prompt multiplicities, above the RF (0.95), the BNN (0.91), and the GPR (0.91);
the BNN is the most accurate on the delayed multiplicity (R2 = 0.746, relative RMSE 42–54%).
SHAP attribution for the RF ranks the incident energy, the mass number, and the fission barrier
first, consistent with statistical-model systematics. Separate INCL–ABLA++ calculations for seven
actinides follow the evaluated curves across the library range and extend the comparison to 1 GeV,
where the BNN stays closest to the cascade curve. The models thereby give a fast estimate of ¯ν for
nuclides and energies the library does not cover, relevant to applications such as nuclear reactors
and accelerator-driven systems.