Machine Learning Classification for Irradiated Material Gamma Ray Spectra

Technical Summary: Machine Learning Classification for Irradiated Material Gamma Ray Spectra

Problem Statement
Accurate and timely identification of irradiated nuclear materials is a critical challenge in nuclear security, environmental monitoring, and safeguards verification. Traditional gamma-ray analysis relies on peak-search methods requiring expert intervention, accurate efficiency calibration, and significant time, often failing to leverage the full spectral shape of the data. While recent Machine Learning (ML) advances show promise for automated spectroscopy, existing literature predominantly focuses on single-isotope identification or relies on simulated datasets that lack the complexity of real experimental measurements. Furthermore, the classification of thermally neutron-activated environmental matrices, where spectral characteristics evolve continuously with post-irradiation cooling time, remains an unaddressed gap in ML spectroscopy literature. This study addresses the need for a robust, generalizable ML pipeline capable of handling variable acquisition times, detector resolutions, and dynamic spectral evolution in experimental datasets.

Methodology
The study evaluates a complete ML classification pipeline applied to experimental gamma-ray spectra collected from two facilities: Georgia Institute of Technology and the University of Texas at Austin (UT Austin). The dataset includes single-isotope sources (133Ba, 60Co, 137Cs, 152Eu, 22Na) and NIST-certified environmental matrices (coal, soil, tomato leaves, natural uranium oxide) irradiated at the UT Austin TRIGA reactor. Measurements were acquired simultaneously using three detector types: Sodium Iodide (NaI(Tl)), Cadmium Zinc Telluride (CZT), and High Purity Germanium (HPGe).

The methodology employs a rigorous five-step preprocessing pipeline designed to create a detector-agnostic, physically meaningful representation of the spectra:

1. Counts-per-second (CPS) Normalization: Converts raw counts to count rates to account for variable acquisition times.
2. Interpolation: Resamples spectra onto a common 3001-point energy grid (0–3000 keV) to unify channel counts across detectors.
3. Total-Area Normalization: Divides by the integrated area to focus on spectral shape rather than absolute intensity.
4. Log1p Transformation: Compresses the dynamic range between photopeaks and the Compton continuum, handling zero-count channels without producing negative infinity.
5. Poisson Noise Augmentation: Applied to raw integer counts prior to preprocessing to increase the effective training set size while preserving correct counting statistics.

Dimensionality reduction is performed using Principal Component Analysis (PCA), retaining enough components to explain 95% of the variance (resulting in 4 components for single isotopes and 4–13 for irradiated samples). Three classifiers are evaluated under a Leave-One-Out Cross-Validation (LOO-CV) strategy: Random Forest (RF), Support Vector Machine (SVM) with an RBF kernel, and k-Nearest Neighbors (KNN).

Key Contributions
This work distinguishes itself from existing literature through three primary contributions:

1. Experimental Validation on Complex Matrices: Unlike prior studies relying on simulations or single-isotope tasks, this work evaluates classification on fully experimental spectra of neutron-activated environmental matrices with no simulated data in the pipeline.
2. Temporal Evolution Handling: It addresses the classification of spectra that change continuously with post-irradiation cooling time, a dynamic not previously tackled in ML spectroscopy literature.
3. Cross-Detector and Physically Motivated Augmentation: The study enables direct comparison of CZT, HPGe, and NaI detectors using identical pipelines. Furthermore, it applies Poisson noise augmentation to raw counts rather than processed features, preserving the physical statistics of the measurement process.

Results

• Single-Isotope Sources: All three classifiers (RF, SVM, KNN) achieved perfect observed accuracy (1.000) across all detector types, validating the pipeline's ability to distinguish well-resolved photopeaks.
• Complex Irradiated Samples:
  • NaI and CZT Detectors: All three classifiers achieved perfect accuracy (1.000) and Matthews Correlation Coefficient (MCC) of 1.000 on the complex irradiated samples.
  • HPGe Detector: RF and KNN achieved perfect accuracy (1.000). SVM achieved 0.974 accuracy, with a single misclassification of a tomato leaf spectrum acquired near 2,000 minutes post-irradiation. This error occurred because the short-lived 24Na component decayed, causing the spectral shape to resemble background; the SVM's margin-based boundary was less tolerant of this drift than the ensemble (RF) or neighbor-based (KNN) methods.
• Statistical Reliability: Clopper-Pearson 95% confidence intervals indicate that while observed accuracy is high, the true population accuracy could plausibly be lower (e.g., lower bounds around 0.86–0.90) due to the small sample size (5–8 spectra per class).
• Prediction Confidence: Analysis of prediction confidence revealed a divergence between accuracy and reliability. For instance, CZT measurements of uranium oxide with short counting times (5 minutes) were classified correctly but with low mean confidence (0.642), suggesting that while the prediction was correct, it was unstable and susceptible to perturbation.

Significance and Claims
The authors present this work as a validated proof-of-concept rather than a definitive performance benchmark. The study establishes a generalizable preprocessing and classification baseline for experimental gamma spectroscopy datasets characterized by variable channel counts, acquisition times, and post-irradiation histories.

The paper modestly claims that the results demonstrate the viability of the pipeline for proof-of-concept purposes. The convergence of three independent algorithms (RF, SVM, KNN) on similar results supports the claim that the observed performance reflects genuine separation of sample classes rather than artifacts of the small dataset. The authors explicitly note that the pipeline was trained and tested within detector types independently; cross-detector generalization (training on one detector, testing on another) was not evaluated and remains an open question. The work serves as a foundational step toward more complex identification tasks, such as post-detonation fission product spectra, where measurement windows are shorter and identification is critical.