Machine learning classification for irradiated material gamma ray spectra

This work presents a preliminary evaluation of Machine Learning (ML) methods for automated gamma-ray spectroscopy classification across two levels of spectral complexity and three levels of detector resolution. Samples analyzed include both single isotope sources and materials irradiated at the University of Texas at Austin Research Reactor. Sample collection was conducted with Sodium Iodide (NaI(Tl)), Cadmium Zinc Telluride (CZT), and High Purity Germanium (HPGe) detectors. A collection of ML techniques was used and compared to classify spectra according to their correct identity; the same classifiers were first applied to the single-isotope sources before being applied to the complex, temporally evolving irradiated samples. Activated samples included natural uranium oxide (U 3 O 8 ), and NIST-certified coal, soil, and tomato leaves irradiated at the UT Austin TRIGA reactor. A five-step preprocessing pipeline consisting of counts-per-second normalization, interpolation onto a common energy grid, Log 1p transformation, total-area normalization, and Poisson noise augmentation was applied to place all spectra on a physically meaningful, detector-agnostic representation. Principal Component Analysis (PCA) reduced the 3001-dimensional feature vectors to between 4 and 13 components prior to classification. Three classifiers were evaluated under Leave-One-Out Cross-Validation (LOO-CV): Random Forest (RF), Support Vector Machine with RBF kernel (SVM), and k-Nearest Neighbors (KNN). For single-isotope sources and complex irradiated samples, all three classifiers achieved no observed failures (accuracy 1.000, Matthews Correlation Coefficient 1.000) on all three detectors. Clopper-Pearson 95% confidence intervals and per-class prediction confidence analysis indicate that observed accuracy and prediction reliability diverge for low-statistics measurements, most notably for short-count CZT spectra. These results establish a proof-of-concept preprocessing and classification baseline for experimental gamma spectroscopy datasets with variable acquisition times, detector types, and post-irradiation histories, and provide a framework for extension into more complex identification.