Energy-Efficient Radioisotope Identification Using Spiking Neuromorphic Systems

View at publisher Summarise with VIRAT
Authors: Felix Wang, Mark Plagge, Yang Ho, Michael Krygier, Craig Michael Vineyard, Aaron Hill
Cited by: 0
References: 61

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This work presents the development and benchmarking of energy-efficient spiking Radioisotope Identification (RIID) algorithms on neuromorphic hardware, with applications to high-performance, ultra-low-power nuclear material detection. We develop a template-based matching algorithm based on sparse coding and the Locally Competitive Algorithm (LCA) that approximates state-of-the-art RIID methods such as those found in the widely known Gamma Detector Response and Analysis Software (GADRAS). We additionally evaluate simple machine learning (ML) models for RIID which are trained through the Python RadioIsotope IDentification (PyRIID) software. These algorithms are implemented as spiking neural networks (SNNs) and deployed on neuromorphic hardware platforms, including Intel Loihi 2, SpiNNaker2, and the BrainChip Akida accelerator. To quantify performance, we evaluate these algorithms against both synthetic and real-world datasets as well as compare against traditional RIID algorithms through GADRAS. The synthetic dataset was generated through PyRIID and consists of 25 isotopes across 124 shielding configurations, and the real-world dataset consists of roughly 200 test samples spanning 7 different isotopes of interest (133Ba,226Ra,235U,232Th,238U,239Pu,131I). We also measure energy efficiency of these algorithms through power profiling on the neuromorphic hardware. Significantly, our neuromorphic implementations consumed less than 1 mJ per inference, representing dramatic orders of magnitude reduction in energy consumption compared to traditional ARM-based GADRAS deployments. Our results also demonstrated the robustness of our developed LCA-based approach compared to a multi-layer perceptron ML model, which excelled on synthetic data and known noise models but failed to generalize to challenging real-world data. In contrast to the GADRAS baseline performance (F1 score) of 0.623, the ML-based model only achieved 0.466, whereas the LCA-based implementation was more closely matched at 0.579. Together, these findings underscore the potential of spiking neuromorphic hardware as a transformative technology for advancing RIID implementations, offering robust, energy-efficient, and field-deployable solutions for real-world applications such as real-time nuclear material detection.

Fields of Study & Subjects

Advanced Memory and Neural Computing Radiation Detection and Scintillator Technologies Neural Networks and Applications

Publication details

Journal / source
IEEE Transactions on Nuclear Science
Volume
73
Issue
5
Pages
1888–1901
Publication date
25 March 2026
Language
EN