DeepNSI: Element identification in experimental photoneutron spectra for illicit material detection

Abstract

We present DeepNSI (Deep Neutron Spectrum Identification), a deep learning framework for identifying elemental composition from photon-induced neutron spectra in realistic inspection scenarios. Targeted toward the detection of illicit materials, DeepNSI consists of an ensemble of element-specific convolutional neural networks trained on a hybrid dataset of simulated and experimental photoneutron spectra. Special emphasis is placed on detecting light elements such as nitrogen and oxygen, which are key signatures of explosives and chemical threats. The model incorporates Monte Carlo Dropout to provide predictive uncertainty and employs a post-processing step based on non-negative least squares (NNLS) to reconstruct the experimental spectrum from reference components. Evaluation on real data—including organic compounds and complex configurations involving shielding materials—demonstrates robust element identification, with uncertainty estimates supporting decision confidence. Although NNLS coefficients are influenced by nuclear cross sections and cannot be interpreted as direct concentrations, their trends across samples reflect meaningful compositional differences. These results establish DeepNSI as a reliable tool for interpretable, machine-learning-based elemental analysis in photon interrogation systems coupled with photoneutron spectrometry.