Developing Support Vector Machine Prediction Capabilities of Uranium Enrichment Based on Gamma-Gamma Coincidence Signature

Introduction: Nuclear safeguards objectives necessitate robust means of spent nuclear fuel and special nuclear materials assay. Gamma-gamma coincidence is a nuclear analytical technique that is capable of performing nondestructive assay of nuclear materials with low detection limits. However, the resulting spectra are often complex and difficult to analyze. In particular, spent nuclear fuel contains an ensemble of fission products, many of which contribute to the coincidence signature of the sample. Furthermore, the energy resolution of cerium-doped lanthanum bromide (LaBr3:Ce) detectors is insufficient to discriminate between many spectral features in dense datasets, resulting in many overlapping features. This adds further difficulty to the analyst searching for features which are indicative of sample enrichment. The main challenge is therefore determining which features of a coincidence spectrum are most useful for sample identification and/or classification. This work develops an automated machine learning capability to analyze coincidence spectra and identify useful features.

Methods: Three 20 mg samples of uranium of varying enrichment (0.7%, 3.0%, and 63%) were irradiated in UT Austin’s TRIGA reactor for one hour at 500 kW. Measurements were then conducted with coincident LaBr3:Ce detectors of each sample once every few days for a month post-irradiation. Twelve one-hour coincident measurements were made for each of the three uranium samples at decay times of: 4, 6, 8, 11, 13, 15, 18, 20, 22, 25, 27, and 29 days.

A set of binary classification support vector machine algorithms was trained with supervised learning on gamma-gamma coincidence data from irradiated uranium samples of natural, low, and high-enriched uranium. The resulting model are capable of classifying uranium samples into the correct enrichment regime (natural, low, or high) based on their gamma-gamma coincidence spectra.

Results: The results of this work are a dataset of coincidence spectra representing three enrichment regimes across a thirty-day range of decay times and SVM models which have been developed with this dataset. Currently, the binary classification SVM of natural versus high-enriched uranium is capable of accurately classifying 12 out of 13 testing samples (which are separate from the samples used to train the SVM).

This represents a 92% classification accuracy. Strategies for improving the accuracy of the SVM models are currently being pursued. These include the acquisition of additional training data and varying of the parameters input to Scikit-learn in the development of the algorithms.

Conclusions: This eliminates the need for a spectral analyst to parse through the data and manually determine which spectral features are indicative of enrichment. The basic functions used by the model to make classification decisions can also inform the analyst of less obvious spectral features that are indicative of enrichment. Ongoing work is being performed to improve classification accuracy and eventually reframe the experiment as a regression problem.