Evaluation of the Local–Global Graph Framework for Interpretable Gamma-Ray Spectrum Identification in Nuclear Security

Accurate identification of gamma-emitting isotopes in measured spectra is essential for nuclear security, safeguards, and non-proliferation. However, reliable radionuclide identification using NaI(Tl) detectors remains challenging under low-count statistics, background interference, shielding, and mixed-source conditions. In this study, we adapt and evaluate an interpretable, geometry-driven approach based on the previously developed Local–Global (LG) Graph methodology. In the LG-Graph framework, each spectral peak is modeled as a triangle defined by its apex and adjacent minima. Features such as energy, area, full width at half maximum (FWHM), and slope are extracted from each triangle. These peak-level features form Local Graphs, which are sequentially connected to form a Global Graph that captures the ordered structure of the spectrum. A library of LG-Graph signatures enables shape- and sequence-based matching for isotope identification. When evaluated against a multiple linear regression (MLR) baseline using single- and mixed-source measured spectra, LG-Graph produced higher precision and fewer false-positive isotope assignments, whereas MLR maintained higher recall. The LG representation also provides peak-level traceability by allowing individual isotope detections to be related to the measured peak features that contributed to the matching score. These results demonstrate a trade-off between selectivity and sensitivity under the decision criteria used in this study and support further evaluation of LG-Graph using larger, more balanced measured datasets relevant to nuclear-security applications.