Application of Artificial Neural Networks to Elemental Assay Data for Nuclear Forensics Analysis
Abstract
An Artificial Neural Network (ANN) is applied to elemental assay data of microscopic, actinide bearing particles obtained using energy dispersive x-ray spectroscopy via a scanning electron microscope (SEM-EDS) and Electron ProbeMicro Analysis (EPMA). This technique provides a non-destructive assessment of the composition of particles that is suitable for nuclear forensics applications. A moment transformation was applied to the data before the ANN was used to compare and group like-particles together using a Siamese network and triplet loss function. A moment transformation provided a noticeable increase in accuracy across all runs. Models using triplet loss had nearly perfect precision when two observations were the same, and provided a preliminary means to compare unknown samples to a database of known samples. Adjusting the hyper parameters could further increase the performance of the models.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 2021
- Accession Number
- AD1145746
Entities
People
- Jason G. Seik
Organizations
- Air Force Institute of Technology