Material Identification using Multi-Modal Intrinsic Radiation and Radiography

Khoa Nguyen; Brendt Wohlberg; Oleg Korobkin; Marc Klasky

physics.ins-detMar 27, 2026

Material Identification using Multi-Modal Intrinsic Radiation and Radiography

Authors: Khoa Nguyen, Brendt Wohlberg, Oleg Korobkin, Marc Klasky

Organizations: Los Alamos National Laboratory, Los Alamos, NM 87545, USA

Abstract

We investigate multi-modal material identification for special nuclear material (SNM) configurations using a combination of X-ray radiography, high-resolution γ-ray spectroscopy, and neutron multiplicity measurements. We consider a Beryllium Reflected Plutonium sphere (BeRP) ball surrounded by one or two concentric shielding shells of unknown composition whose radii are assumed known from radiography. High-purity germanium (HPGe) spectra are reduced to net counts in selected Pu-239 photo-peaks, while neutron multiplicity information is summarized by Feynman variances Y2 and Y3 computed from factorial moments of the neutron counting statistics. Using synthetic data generated with the Gamma Detector Response and Analysis Software (GADRAS) for a range of shielding materials and thicknesses, we cast the material identification problem as a supervised multi-class classification task over all admissible shell-material combinations. We demonstrate that a random forest classifier trained on combined gamma and neutron features achieves almost perfect identification accuracy for single-shell cases, and substantial performance gains for more challenging double-shell configurations relative to gamma-only classification. Alternative statistical and machine-learning formulations for this multi-class problem are examined along with examination of the impact of model-mismatch between the forward model and the test cases as given by variations in the statistical noise. Opportunities for extending the approach to more complex geometries and experimental data are also discussed.

Figures & tables

Panel of Figure 1 — Figure 1: Feynman Variance $Y_{2}$ and $Y_{3}$ for BeRP ball versus shielding shell thickness

Panel of Figure 2 — Figure 2: General problem formulation.

Method	# Noisy samples	Train Acc.	Test Acc.
RF	0 (clean data)	1.000	0.995
	2	1.000	0.993
	10	1.000	0.999
	100	1.000	1.000
	1000	1.000	1.000
XGBoost	0 (clean data)	0.048	0.048

Table 1: Single-shell material identification performance of different ML approaches, including random forest (RF), XGBoost, K-nearest neighbors (KNNs), support vector machine (SVM), and multilayer perceptrons (MLPs). We consider a 60s time window, a 0.5 inch single shell configuration, Gamma+

Y_{2}

Y_{3}

as feature and different number of noisy samples per training simulations.

N=10{,}000

independent noise realizations per feature set and thickness are generated for testing. Boldface indicates the best (highest) test accuracy within each method.

Thickness(in)	Features	Train Acc.	Test Acc.
0.5	Gamma	1.000	0.979
	Gamma + $Y_{2}$	1.000	0.995
	Gamma + $Y_{2}$ + $Y_{3}$	1.000	0.995
1.0	Gamma	1.000	0.979
	Gamma + $Y_{2}$	1.000	0.999
	Gamma + $Y_{2}$ + $Y_{3}$	1.000	1.000

Table 2: Single-shell material identification performance versus shell thickness for three feature sets (Gamma only, Gamma+

Y_{2}

, Gamma+

Y_{2}

Y_{3}

) using 60s time window, clean data for training, and

N=10{,}000

independent noise realizations per feature set and thickness for testing. Boldface indicates the best (highest) test accuracy within each thickness (ties bolded).

Time window	Train Acc.	Test Acc.
3600s	1.000	1.000
60s	1.000	0.995
15s	1.000	0.979

Table 3: Single-shell material identification performance using random forest and diffenrent time windows. We consider the 0.5 inch shell configuration, Gamma+

Y_{2}

Y_{3}

as feature and clean data for training.

N=10{,}000

independent noise realizations per feature set and thickness are generated for testing. Boldface indicates the best (highest) test accuracy (ties bolded).

# Noisy samples	Train Acc.	Test Acc.
0 (clean data)	1.000	0.995
1	1.000	0.980
2	1.000	0.993
10	1.000	0.999
100	1.000	1.000
1000	1.000	1.000

Table 4: Single-shell material identification performance using random forest and 60s time window while considering the 0.5 inch shell configuration, Gamma+

Y_{2}

Y_{3}

as feature and either clean data or different number of noisy samples per training simulations (note that when using noisy samples, the clean data are not included in the training).

N=10{,}000

independent noise realizations per feature set and thickness are generated for testing. Boldface indicates the best (highest) test accuracy (ties bolded).

Figure 3: Classification errors by material and thickness in single shell configuration using Gamma+ $Y_{2}$ + $Y_{3}$ as feature.

Panel of Figure 4 — Figure 4: Normalized confusion matrices for single-shell material identification at two shell thicknesses (rows) and three feature sets (columns): Gamma only (left), Gamma + $Y_{2}$ (middle), and Gamma + $Y_{2}$ + $Y_{3}$ (right), using 60s time window, clean data for training, and $N=10{,}000$ independent noise realizations per feature set and thickness for testing. Each panel shows the classification performance across all candidate shell materials under noisy measurement conditions. We normalize these matrices so that the sum of classification errors for each reference material is 1 (if they are non-zeros).

Time window	Thicknesses (in)	Features	Train Acc.	Test Acc.
60s	0.5 / 0.5	Gamma	1.000	0.671
		Gamma + $Y_{2}$	1.000	0.815
		Gamma + $Y_{2}$ + $Y_{3}$	1.000	0.870
	3.0 / 3.0	Gamma	1.000	0.687
		Gamma + $Y_{2}$	1.000	0.808
		Gamma + $Y_{2}$ + $Y_{3}$	1.000	0.840

Table 5: Double-shell material identification performance for representative shell-thickness combinations (inner/outer) using either 60s or 1800s time window, and clean data for training.

N=10{,}000

independent noise realizations per feature set and thickness are generated for testing. Boldface indicates the best (highest) test accuracy within each thickness combination (ties bolded).

Time window	Train Acc.	Test Acc.
3600s	1.000	0.992
1800s	1.000	0.990
600s	1.000	0.968
60s	1.000	0.870
15s	1.000	0.808

Table 6: Double-shell material identification performance using random forest and diffenrent time windows. We consider the 0.5/0.5 inch shell configuration, Gamma+

Y_{2}

Y_{3}

as feature and clean data for training.

N=10{,}000

independent noise realizations per feature set and thickness are generated for testing. Boldface indicates the best (highest) test accuracy (ties bolded).

Time window	# Noisy samples	Train Acc.	Test Acc.
60s	0 (clean data)	1.000	0.870
	1	1.000	0.731
	2	1.000	0.875
	10	1.000	0.983
	100	1.000	0.993
	1000	1.000	0.995

Table 7: Double-shell material identification performance using random forest with 60s or 1800s time window while considering the 0.5/0.5 inch shell configuration, Gamma+

Y_{2}

Y_{3}

as feature and either clean data or different number of noisy samples per training simulations (note that when using noisy samples, the clean data are not included in the training).

N=10{,}000

independent noise realizations per feature set and thickness are generated for testing. Boldface indicates the best (highest) test accuracy (ties bolded).

Time window	Noise level	Features	Train Acc.	Test Acc.
60s	$\sigma_{2}=0.02,\sigma_{3}=0.2$	Gamma	1.000	0.671
		Gamma + $Y_{2}$	1.000	0.815
		Gamma + $Y_{2}$ + $Y_{3}$	1.000	0.870
	$\sigma_{2}=0.04,\sigma_{3}=0.4$	Gamma	1.000	0.644
		Gamma + $Y_{2}$	1.000	0.805
		Gamma + $Y_{2}$ + $Y_{3}$	1.000	0.847

Table 8: Double-shell material identification performance for 0.5/0.5 inch shell configuration, with different level of noise using (Gamma only, Gamma+

Y_{2}

, Gamma+

Y_{2}

Y_{3}

) using either 60s or 1800s time window, clean data for training, and

N=10{,}000

independent noise realizations per feature set and thickness for testing. Boldface indicates the best (highest) test accuracy within each thickness (ties bolded).

Thickness (in)	Features	Entropy
$0.5/0.5$	Gamma	3.337
	Gamma + $Y_{2}$	2.622
	Gamma + $Y_{2}$ + $Y_{3}$	2.362
$3.0/3.0$	Gamma	3.313
	Gamma + $Y_{2}$	2.672
	Gamma + $Y_{2}$ + $Y_{3}$	2.547

Table 9: Double-shell material identification performance with entropy using (Gamma only, Gamma+

Y_{2}

, Gamma+

Y_{2}

Y_{3}

) using 1800s time window, clean data for training, and

N=10{,}000

independent noise realizations per feature set and thickness for testing. Boldface indicates the best (highest) test accuracy within each thickness (ties bolded).

Panel of Figure 5 — Figure 5: Misclassification fractions for representative two-shell configurations (rows) under three feature sets (columns): Gamma only (left), Gamma + $Y_{2}$ (middle), and Gamma + $Y_{2}$ + $Y_{3}$ (right) using 1800s time window and clean data for training. $N=10{,}000$ independent noise realizations per feature set and thickness are generated for testing. Each panel summarizes the fraction of test cases that are misclassified for the specified shell-thickness configuration (inner/outer).

Panel of Figure 6 — Figure 6: Misclassification fractions for asymmetric two-shell configurations (rows) under three feature sets (columns): Gamma only (left), Gamma + $Y_{2}$ (middle), and Gamma + $Y_{2}$ + $Y_{3}$ (right) using 1800s time window and clean data for training. $N=10{,}000$ independent noise realizations per feature set and thickness are generated for testing. The reduction in misclassification fraction when adding neutron multiplicity features highlights the complementary information provided by neutron transport relative to gamma attenuation in layered shielding.

Panel of Figure 7 — Figure 7: Misclassification fractions for representative two-shell configurations (rows) under three configuration of training data (columns): clean data (left), 10 noisy samples per simulation (middle), and 100 noisy samples per simulation (right) using 6.0/3.0 in shells configuration, Gamma+ $Y_{2}$ + $Y_{3}$ as feature, and 60 or 1800s time window. $N=10{,}000$ independent noise realizations per feature set and thickness are generated for testing. Each panel summarizes the fraction of test cases that are misclassified for the specified shell-thickness and time window configuration (inner/outer).

Panel of Figure 8 — Figure 8: Fraction of classification errors for (a) only shell 1 is misidentified, (b) only shell 2 is misidentified, and (c) both shells are misidentified, by material and thickness in double shell configuration with 0.5/0.5 thickness, using 1800s time window and Gamma+ $Y_{2}$ + $Y_{3}$ as feature. We normalize these errors so that for each material pair, the sum of errors in (a), (b), and (c) is $1$ (if they are non-zeros).

Panel of Figure 9 — Figure 9: Normalized confusion metric with 0.5/0.5 shell thickness using 1800s time window, Gamma+ $Y_{2}$ + $Y_{3}$ as feature, and clean data for training. $N=10{,}000$ independent noise realizations per feature set and thickness are generated for testing. We normalize these matrices so that the sum of classification errors for each reference material is 1 (if they are non-zeros).

Multimodal Classification

Multimodal classification combines images, text, audio, signals, metadata, or other inputs to assign labels more accurately and reliably.

See all 199 papers

Peak-Based Nuclide Identification in HPGe $γ$ -Spectrometry with Machine Learning and SHAP
Jun 12, 2026Samuel Emmons, Kelly Truax, Maurice Lonsway +2Classification
Pacific Northwest National Laboratory, Richland, 99354, WA, USA
Computer vision-based neural networks for radioisotope identification in urban environments
Jun 30, 2026Masen Bachleda, Peter LalorConvolutional Neural Networks
Pacific Northwest National Laboratory, Richland, WA 99354 USA
STAR-XPC: Set Transformer for Estimating X-ray Transmission in Polymer Composites
Oct 20, 2026Can Erer, H. Birkan Yilmaz, Nebahat AralNeural Surrogate Modeling Materials Science
Department of Computer Engineering, Bo˘gaziçi University, Bebek, 34342, Istanbul, Turkiye · Independent Researcher, Istanbul, Turkiye

Jun 12, 2026J/K move · Enter open · S save

Peak-Based Nuclide Identification in HPGe $γ$ -Spectrometry with Machine Learning and SHAP

Samuel Emmons, Kelly Truax, Maurice Lonsway, Bruce Pierson+1

Pacific Northwest National Laboratory, Richland, 99354, WA, USA

Classification

Open paper AskPDF arXiv

High-purity germanium gamma spectra often require time-consuming analyses from subject matter experts. Photopeaks within these spectra are carefully fitted and numerical methods are employed to assist with nuclide identification (NID) and quantification. Amending the list of nuclides identified by analysis software can be nontrivial. When many samples need to be analyzed, it is therefore challenging to make timely and correct decisions. Supervised machine-learning-based NID can serve as an expert-informed, automated tool to improve the initial set of radionuclides suggested to an analyst and more effectively drive subsequent quantification. To that end, we implemented machine learning models that map photopeaks carefully fitted by analysts to NID results for experimental spectra containing various isotopic combinations drawn from a set of 65 isotopes. The best model achieved an F1 score of 0.97, markedly surpassing the F1 score of 0.84 achieved by traditional software when compared using a nuclide library comprising the same 65 isotopes assessed by the models. Finally, we illustrated the most important input features for model predictions using Shapley Additive Explanations. These explanations revealed that the models use physically relevant photopeaks when making predictions for the isotopes in our nuclide library.

Abstract

Multimodal Classification

Explore similar work

Peak-Based Nuclide Identification in HPGe γγγ-Spectrometry with Machine Learning and SHAP

Computer vision-based neural networks for radioisotope identification in urban environments

STAR-XPC: Set Transformer for Estimating X-ray Transmission in Polymer Composites

Peak-Based Nuclide Identification in HPGe $γ$ -Spectrometry with Machine Learning and SHAP