Machine Learning Spectroscopy for High-Resolution Radiation Detection

Machine Learning-Enhanced Spectroscopy Technology for High-Resolution Radiation Detection Using Low-Cost Detectors Transforms low-energy resolution gamma- and x-ray detector data into high-resolution spectra—reducing cost, size, and cooling

requirements without sacrificing performance. Technology Summary This INL technology enables high-energy-resolution radiation spectroscopy using low-cost, room-temperature detectors such as sodium iodide (NaI) scintillators. Traditionally, researchers and engineers rely on high-purity germanium (HPGe) detectors, lanthanum bromide (LaBr3) or similar for applications requiring fine energy discrimination; however, these systems are expensive, fragile, or require cryogenic cooling. The presented approach applies a compact convolutional neural network (CNN) architecture to reconstruct high-energy-resolution spectra from low-resolution measurements. Using four convolution-max pooling layer pairs (128–16 filters) followed by dense layers, the model captures spectral features typically only visible with HPGe detectors. The network contains roughly 1.6 million parameters (6.2 MB total), enabling fast, portable deployment in embedded or field devices. The technology offers a new analytical pathway for radiation spectroscopy—maintaining data fidelity while reducing total system cost, weight, and operational complexity. Problem Addressed High cost and complexity of high-energy-resolution detectors: HPGe systems provide excellent energy resolution (~0.2%) but are 10×–100× more expensive than scintillation-based systems. Limited operational flexibility: HPGe detectors require cryogenic cooling and are unsuitable for mobile or high-radiation environments. Low detection efficiency and count-rate performance: HPGe detectors have lower detection efficiency per detector volume and cannot handle high count rates without peak deformation or detector dead time, leading to data degradation. Restricted deployment scenarios: Field, space-based, and confined monitoring applications require detectors that are robust, efficient, and thermally independent. Solution Data-driven energy resolution enhancement: Employs a convolutional neural network to reconstruct high-resolution spectra from low-resolution detector inputs. Compact, deployable model: 1.6M-parameter neural network (6.2 MB) allows rapid inference on low-power devices. Detector-agnostic implementation: Can be adapted for gamma, x-ray, neutron, or charged-particle spectroscopy. Scalable to various hardware: Applicable to NaI, CsI, or plastic scintillators, enabling energy peak discrimination comparable to HPGe without cryogenic operation. Key Advantages Cost Reduction: Enables ≥10× lower system cost and maintenance by replacing HPGe with NaI or other inexpensive detectors. Operational Simplicity: Eliminates need for liquid nitrogen or cryogenic cooling systems. Higher Throughput: Supports higher count rates with minimal peak deformation. Improved Deployability: Suitable for remote, field, and mobile environments where HPGe is impractical. Cross-Technology Applicability: Adaptable for gamma-ray, x-ray, and neutron detection systems. Market Applications Nuclear materials monitoring and safeguards – real-time isotope discrimination without cryogenic infrastructure. Space-based radiation detection – lightweight, low-power alternative to HPGe for satellite payloads. Industrial quality control and non-destructive testing – improved spectral resolution using existing NaI-based systems. Medical and environmental radiation monitoring – portable spectrometers with enhanced fidelity for imaging and dosimetry. Homeland security and defense – deployable gamma-ray detection for special nuclear material tracking. This notice is not a solicitation for funding or a commitment by DOE/INL to procure services. Rather, it is intended solely to notify industry of an INL technology available for licensing and commercialization. Javier Martinez Contracting (No Street Address

2) Idaho Falls, USA to this opportunity.