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

Area of Science:

• Nuclear Physics and Engineering
• Materials Science
• Artificial Intelligence

Background:

• Photon interrogation systems are crucial for security applications, but accurate elemental analysis remains challenging.
• Detecting light elements like nitrogen and oxygen is vital for identifying explosives and chemical threats.
• Existing methods often lack the precision and interpretability needed for complex inspection scenarios.

Purpose of the Study:

• To develop a robust deep learning framework, DeepNSI, for elemental composition identification from photon-induced neutron spectra.
• To enhance the detection capabilities for light elements, crucial for security applications.
• To provide interpretable results with uncertainty quantification for reliable decision-making.

Main Methods:

• Developed DeepNSI, an ensemble of element-specific convolutional neural networks.
• Trained models on a hybrid dataset of simulated and experimental photoneutron spectra.
• Incorporated Monte Carlo Dropout for predictive uncertainty and Non-Negative Least Squares (NNLS) for spectral reconstruction.

Main Results:

• Demonstrated robust elemental identification on real-world data, including organic compounds and shielded materials.
• Achieved reliable detection of light elements, key signatures for explosives and chemical threats.
• NNLS trends reflected meaningful compositional differences, supporting interpretable analysis.

Conclusions:

• DeepNSI offers a reliable, machine-learning-based approach for elemental analysis in photon interrogation systems.
• The framework provides interpretable results with uncertainty estimates, enhancing confidence in security applications.
• DeepNSI advances the field of photoneutron spectrometry for material inspection and threat detection.