Site-Directed Research and Development logo, green and blue with orange writing

Deploying Isotope Identification Machine Learning Models

Site-Directed Research and Development logo, green and blue with orange writing

Home / Mission / Site-Directed Research and Development / Deploying Isotope Identification Machine Learning Models

Project #: 22-142 | Year 1 of 1

Garrett Dean,a Ronald Guise,b Krikor Hovasapian,b Christopher Joines,b Michael Bassb

aRemote Sensing Laboratory-Andrews; bRemote Sensing Laboratory-Nellis

Executive Summary

This project sought to develop and evaluate edge compatible artificial intelligence (AI) models, specifically to detect anomalies and identify isotopes for radiation detectors in order to reduce the reliance on human subject matter experts (SMEs). Exploring resource constrained models, quantization, pruning, and feature selection of real-world datasets, this project evaluated the effectiveness of these AI models in remote and edge platforms.

Description

Previous AI radiation detection research has studied various machine learning models but had neglected edge AI applications. This SDRD project attempted to implement and study multiple edge AI architectures on different hierarchical levels of radiation telemetry to support low power, miniaturized identification and sense making AI in alignment with the NNSS S&T thrust area of Enabling Technologies for Autonomous Systems and Sensing. The project intended to study edge AI through two approaches: isotope identification and classifying whether further adjudication was required. Ultimately, only basic, inaccurate, isotopic identification and anomaly identification, not at the edge level, was achieved.

Basic neural networks were created in Python using existing packages such as scikit-learn (Pedregosa et al. 2011) and PyTorch (Paske et al. 2019). These were created, trained, and run on traditional computer hardware to gain a fundamental understanding of the machine learning algorithm creation process with the intention of porting into edge systems at a later time. The overall results were disappointing, likely due to a combination of a small training data set and user error in selecting the relevant parameters—such as number of layers, layer size, solver type, etc. The creation of the small training and test data set was achieved using a combination of real and synthetic data. The real data was taken from the Remote Sensing Laboratory’s (RSL) Search Management Center database. Synthetic data was created using GADRAS (Gamma Detector Response and Analysis Software) to determine the ideal response, including scatter and continuum in the absence of background, and summed to known background spectra. This set contained examples of six nuclides (more are likely contained within, but further investigation is needed): Cs-137, Co-60, I-131, Tc-99m, Tl-201, F-18 (511 keV), and numerous variations on what composes a background spectrum. Overall, the project was highly informative, and the learning process was enlightening and useful, even if the results were less than spectacular.

Figure 1. Layout and schedule of originally-planned progression among controller, hardware, and software connections. Multichannel analyzers (MCAs) were physically connected to detectors, such as sodium iodide (NaI) crystals, and recorded radiation spectra. Data was then wirelessly transmitted as follows: one or more MCAs to controller (RSL-Nellis’s [RSLN] Gemini); controller to intermediary (RSLN’s Android application Gamut); or intermediary to cloud, either directly or indirectly through a communication relay (RSLN’s Multi-Path Comms Device [MPCD] [Essex 2006]). Once telemetered, an SME reviewed them.
Figure 1. Layout and schedule of originally-planned progression among controller, hardware, and software connections. Multichannel analyzers (MCAs) were physically connected to detectors, such as sodium iodide (NaI) crystals, and recorded radiation spectra. Data was then wirelessly transmitted as follows: one or more MCAs to controller (RSL-Nellis’s [RSLN] Gemini); controller to intermediary (RSLN’s Android application Gamut); or intermediary to cloud, either directly or indirectly through a communication relay (RSLN’s Multi-Path Comms Device [MPCD] [Essex 2006]). Once telemetered, an SME reviewed them.

Conclusion

The initial thrust to extend edge AI capabilities for radiation anomaly detection and identification was shifted to the development of machine learning on traditional architecture and self-education about such algorithms, with the aim of returning to the original objectives once the fundamentals were better understood by the inexpert PI. While some progress was made in this endeavor, primarily on a training set, no notable progress or achievements occurred, nor were algorithms developed. The project was unable to achieve the initial laid out objectives. However, this project has spurred collaboration with STL on a similar project and may be revisited at a later point in time.

Mission Benefit

Today, deploying hundreds to thousands of detectors, as would occur in a major disaster and already occurs for major public events such as the Super Bowl or inaugurations, yields plateaued radiation awareness, overloads SMEs, and produces terabytes of data that cripple telemetry infrastructures and is too vast for post-processing. This presents a necessity to innovate and move radiation AI to edge devices to reduce these bottlenecks and proportionately increase radiation awareness with detector deployments. Edge AI isotope identification and anomaly classification touches on 1) Consequence Management—reduction of cost by minimizing agent deployments and redundant measurements, 2) Search—reduction in skill barriers for operators and human error, and 3) Aerial Measuring System—decrease in human flights through radiation plumes, costly multiple flights, and down time. A complete edge AI pipeline would accelerate deployment of future edge AI that identify system failures, automate equipment initialization, and prevent time spent in hazardous environments.

Publications, Technology Abstracts, Presentations/Posters

This work was presented at the NNSS FY22 SDRD annual project review.

References

Essex, J. 2006. “Multipath Communications Device.” InFY 2005 Site-Directed Research and Development, 231-237. Las Vegas NV: National Security Technologies, LLC.

Paszke, A. et al., 2019. “PyTorch: An Imperative Style, High-Performance Deep Learning Library.” In Advances in Neural Information Processing Systems 32. Curran Associates, Inc., 8024–8035. http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf.

Pedregosa, F., et al. 2011.Scikit-learn: Machine Learning in Python. JMLR 12 (85): 2825-2830. https://jmlr.csail.mit.edu/papers/v12/pedregosa11a.html.

This work was done by Mission Support and Test Services, LLC, under Contract No. DE-NA0003624 with the U.S. Department of Energy. DOE/NV/03624–1616.

Return to Enabling Technologies for Autonomous Systems and Sensing
Go to SDRD Annual Report Index
Return to SDRD Homepage