Predictions for the ( n,2n ) reaction cross section based on a Bayesian neural network approach

Nuclear $(n,2n)$ reaction cross sections are studied based on the Bayesian neural network (BNN) approach. Three physical quantities besides the proton and neutron numbers are proposed to improve the performance of the BNN approach. These three physical quantities are the incident neutron energy with respect to the reaction threshold, the physical quantity related to the odd-even effect, and the theoretical $(n,2n)$ reaction cross section, and they are included as the inputs to the neural network. The BNN approach has better performance in the description of the $(n,2n)$ reaction cross sections than the theoretical library TENDL-2021 calculated by the talys code based on the Hauser-Feshbach statistical model, especially for heavy nuclei. The root-mean-square deviation of the BNN approach with respect to the evaluation data is reduced to 0.10 barns compared to 0.25 barns of TENDL-2021. The extrapolation ability of the BNN approach is verified with the $(n,2n)$ cross section data that are not used to train the neural network. Furthermore, it is found that the BNN approach still well describes the trend of the $(n,2n)$ cross sections with the incident neutron energy predicted by TENDL-2021 even when extrapolated to the unknown region.