Uniform Manifold Approximation and Projection as Anomaly Detection Applied to Masking Scenarios of Plutonium

Unsupervised learning strategies have proven successful for detecting anomalies in various applications. UMAP, the uniform manifold approximation and projection technique, is a transformation that maps complex input data to a reduced dimensional map. It is applied to real-world spectral data of Plutonium, measured with Sodium Iodide (NaI:Tl), Lanthanum Bromide (LaBr3:Ce,Sr) and Cadmium Zinc Telluride (CZT) detectors. Through the manifold mapping, similar spectral structures are clustered while anomalies are highlighted by a different position in the reduced dimensional subspace. In addition, the simpler t-Student Neighbour Embedding algorithm is used on the same data for comparison with UMAP. The paper shows further geometric strategies to provide reproducible results with the mapping and physics-informed learning algorithms that implement existing UMAP and t-SNE techniques in a forward prediction of the spectroscopic scenario. Data from real-world tests involving Plutonium and medical nuclides like Iodine I-131 is tested and evaluated with this forward prediction. Also naturally occurring masks like K-40 are considered and discussed.