Predictions of neutron-capture reaction cross sections using a Bayesian neural network approach combined with a physically motivated empirical formula*

Abstract Neutron-capture reaction cross sections are studied using the Bayesian neural network (BNN) approach in combination with a recently proposed empirical formula for the cross sections. In addition to the proton and neutron numbers, four physical quantities are found to be important for improving the predictive performance of the BNN approach: the pairing effect-related variable δ , shell effect-related variables and , and theoretical neutron-capture reaction cross sections. The BNN approach more effectively describes the Maxwellian-averaged ( ) cross sections (MACS) at = 30 keV than the TENDL-astro 2023 theoretical library calculated using TALYS code based on the Hauser–Feshbach statistical model. The root-mean-square deviation of the BNN approach with respect to the natural logarithm of the experimental MACS data from the Karlsruhe Astrophysical Database of Nucleosynthesis in Stars is reduced to 0.1373, compared with the value of 0.2545 for TENDL-astro 2023. The BNN predictions align well with MACS trends predicted by TENDL-astro 2023 when extrapolated to the unknown region, though there are quantitative deviations between them.