Neural network-based systematics of giant dipole resonance parameters and impact on (γ,n) reaction cross sections

Abstract

Reliable giant dipole resonance (GDR) parameterization is essential for photonuclear reaction modeling in the (γ,n) reaction energy region, where dipole absorption dominates. This work develops a staged artificial neural network (ANN) scheme to infer the GDR centroid energy, width, and peak cross section from physically motivated nuclide descriptors compiled from evaluated nuclear-structure information provided by the National Nuclear Data Center (NNDC), using Reference Input Parameter Library (RIPL-3) values as learning targets. To obtain performance estimates that reflect genuine transfer to unseen nuclei and to avoid leakage from repeated entries, the dataset was partitioned nucleus-wise, assigning all records of a given nuclide to a single subset. Parameter inference was organized in three sequential stages to preserve the established coupling among centroid energy, width, and peak strength. A SHAP-based interpretability analysis was performed at each stage to quantify the relative importance of the input descriptors. The resulting ANN-derived parameter sets were implemented in TALYS and compared with the nucleus-specific GDR parameters internally available in the built-in TALYS structure/gamma/gdr/ data files under otherwise identical reaction-model conditions. Their impact on calculated (γ,n) reaction excitation functions was evaluated against experimental photoneutron reaction cross sections from the EXFOR library via the JANIS interface, quantifying agreement with experiment achieved by the neural-network-based GDR systematics. The even-even 112–124Sn isotopic chain was examined as a representative case of isotopic continuity. Overall, implementation of the ANN-derived GDR parameter triplets in TALYS generally improves agreement between calculated (γ,n) reaction excitation functions and EXFOR data relative to the built-in nucleus-specific TALYS GDR values where available.