cs.LGMar 5, 2026

Unsupervised domain adaptation for radioisotope identification in gamma spectroscopy

Authors: Peter Lalor, Ayush Panigrahy, Alex Hagen

Organizations: Pacific Northwest National Laboratory, Richland, WA 99352 USA · University of Washington, Seattle, WA 98195 USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Training machine learning models for radioisotope identification using gamma spectroscopy remains an elusive challenge for many practical applications, largely stemming from the difficulty of acquiring and labeling large, diverse experimental datasets. Simulations can mitigate this challenge, but the accuracy of models trained on simulated data can deteriorate substantially when deployed to an out-of-distribution operational environment. In this study, we demonstrate that unsupervised domain adaptation (UDA) can improve the ability of a model trained on synthetic data to generalize to a new testing domain, provided unlabeled data from the target domain is available. Conventional supervised techniques are unable to utilize this data because the absence of isotope labels precludes defining a supervised classification loss. We compare a range of different UDA techniques, finding that feature alignment strategies, particularly via maximum mean discrepancy (MMD) minimization or domain-adversarial training, yield the most consistent improvement to testing scores. For instance, using a custom transformer-based neural network, we achieve a testing accuracy of 0.904±0.0220.904 \pm 0.022 on an experimental LaBr3_3 test set after performing unsupervised feature alignment via MMD minimization, compared to 0.754±0.0140.754 \pm 0.014 before alignment. Overall, our results highlight the potential of using UDA to adapt a radioisotope classifier trained on synthetic data for real-world deployment.

Figures & tables

ScenarioDomainAcquisition ModeIsotopic CompositionNumber of IsotopesSize
sim-to-sim (HPGe)SourceGADRASMixed551.3×1061.3\times 10^{6}
TargetGeant4Mixed551.3×1061.3\times 10^{6}
sim-to-real (LaBr 3 )SourceGADRASSingle321.4×1061.4\times 10^{6}
TargetExperimentSingle3215,091
sim-to-real (NaI(Tl))SourceGADRASSingle321.4×1061.4\times 10^{6}
TargetExperimentSingle3210,440
Table 1: Summary of the datasets used in this study.
ScenarioAUROC (features)AUROC (inputs)MMD 2R\mathrm{R}
sim-to-sim (HPGe)[0.993, 0.996][0.957, 0.967][2.74, 2.93][1.36, 1.40]
sim-to-real (LaBr 3 )[0.982, 0.988][0.852, 0.874][2.26, 2.55][1.87, 1.93]
sim-to-real (NaI(Tl))[0.983, 0.989][0.780, 0.808][1.30, 2.53][1.93, 1.99]
Table 2: 95% CI for each domain separability metric, indicating a statistically significant domain gap between source and target domains for all domain adaptation scenarios.
Table 3: Evaluation metric scores on the target-domain test dataset for seven different UDA techniques. The “Source-only” row indicates models that were only trained on source-domain data with no domain adaptation step, serving as a baseline. The “Train on Target” row indicates models that were trained on labeled target-domain data, serving as a theoretical best-case scenario. Cell entries are calculated as the mean and sample standard deviation across 10 random trials. The best performing method for each architecture (excluding “Train on Target”) is highlighted in bold.
Figure 1: UMAP visualizations of the feature extractor outputs for source spectra (‘ o ’ markers) and target spectra (‘ + ’ markers) using a TBNN-LinEmb architecture. In the sim-to-sim scenario (top row), color indicates domain (blue = source, red = target). In the sim-to-real scenarios (middle and bottom rows), color indicates class, which is possible because the sim-to-real data is single-label.
Table 4: Comparing the effects of UDA on a variety of diagnostic metrics ( ↑\uparrow higher is better, ↓\downarrow lower is better). Source-only and UDA TBNN-LinEmb models are shown for all three domain adaptation scenarios. Entries for the best model are bolded when significant ( p<0.01p<0.01 ). Overall, UDA provides improvements using a wide variety of metrics beyond classification accuracy in the sim-to-real scenarios.
Figure 2: SHAP explanations comparing source-only (left panels) and DAN models (right panels) across three experimental example spectra. For 40 K (top row), the source-only model highlights the LaBr 3 detector’s intrinsic 32 keV X-ray peak, whereas the DAN model instead identifies the correct 1460 keV line. For 233 U (middle row), the source-only model relies primarily on peaks associated with visible 232 U-chain contamination, while the DAN model identifies the more subtle 93 keV feature of 233 U. For 239 Pu (bottom row), the source-only model predicts 124 Sb by emphasizing the 603 keV peak (likely due to trace contamination or activation-related background), whereas the DAN model highlights the characteristic 239 Pu peaks around 345, 375, and 413 keV.
Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

ArchitectureParameterSearch SpaceBest run
MLPNum Dense Layers{1,…,4}\{1,\ldots,4\}2
 Dense1 Hidden Units[512, 8192]2k[512,\,8192]_{2^{k}}8192
 Dense2 Hidden Units[256, 4096]2k[256,\,4096]_{2^{k}}4096
CNNNum Convolutional Layers{1,2}\{1,2\}1
 Conv Filters[16, 256]2k[16,\,256]_{2^{k}}16
 Conv Kernel Size{3,5,7,9}\{3,5,7,9\}5
Appendix
Table 5: Source-only architecture hyperparameter optimization summary for MLP, CNN, TBNN (Li et al. ), TBNN-LinEmb, and TBNN-NonlinEmb architectures. The table details the search spaces and best-performing hyperparameter values obtained via TPE Bayesian optimization using the source dataset from the sim-to-sim scenario. The best-performing hyperparameters are used for all models across different scenarios in this study. See the search space notation paragraph above for interval and sampling conventions.
   Best run
   sim-to-simsim-to-realsim-to-real
ArchitectureParameterSearch SpaceHPGeLaBr 3NaI(Tl)
MLPLearning rate[10−5, 2×10−3]log⁡[10^{-5},\,2{\times}10^{-3}]_{\log}4.60e-51.80e-31.32e-3
 Batch size[32, 512]2k[32,\,512]_{2^{k}}256512256
 Weight decay[10−7, 10−3]log⁡[10^{-7},\,10^{-3}]_{\log}1.54e-51.96e-61.21e-4
 Dropout[0.0, 0.4][0.0,\,0.4]0.2300.3940.224
Appendix
Table 6: Source-only training hyperparameter optimization summary for MLP, CNN, TBNN (Li et al. ), TBNN-LinEmb, and TBNN-NonlinEmb architectures. The table details the search spaces and best-performing hyperparameter values obtained via TPE Bayesian optimization using the source dataset. See the search space notation paragraph for interval and sampling conventions.
   Best run
   sim-to-simsim-to-realsim-to-real
ArchitectureParameterSearch SpaceHPGeLaBr 3NaI(Tl)
MLPFE learning rate †[10−8, 10−2]log⁡[10^{-8},\,10^{-2}]_{\log}5.87e-71.67e-42.18e-4
 DD learning rate †[10−8, 10−2]log⁡[10^{-8},\,10^{-2}]_{\log}9.29e-63.87e-65.03e-6
 Batch size[64, 512]2k[64,\,512]_{2^{k}}12812864
 Weight decay[10−7, 10−3]log⁡[10^{-7},\,10^{-3}]_{\log}6.89e-62.35e-68.33e-4
Appendix
Table 7: ADDA hyperparameter optimization summary for MLP, CNN, TBNN (Li et al. ), TBNN-LinEmb, and TBNN-NonlinEmb architectures. The table details the search spaces and best-performing hyperparameter values obtained via TPE Bayesian optimization. See the search space notation paragraph for interval and sampling conventions.
   Best run
   sim-to-simsim-to-realsim-to-real
ArchitectureParameterSearch SpaceHPGeLaBr 3NaI(Tl)
MLPLearning rate[10−7, 10−3]log⁡[10^{-7},\,10^{-3}]_{\log}5.80e-69.72e-46.84e-4
 Batch size[64, 512]2k[64,\,512]_{2^{k}}256256128
 Weight decay[10−7, 10−3]log⁡[10^{-7},\,10^{-3}]_{\log}1.92e-58.15e-56.24e-6
 Dropout[0.0, 0.4][0.0,\,0.4]0.02781.10e-33.71e-3
Appendix
Table 8: Same as Table 7 , but using DAN.
   Best run
   sim-to-simsim-to-realsim-to-real
ArchitectureParameterSearch SpaceHPGeLaBr 3NaI(Tl)
MLPFE learning rate †[10−8, 10−2]log⁡[10^{-8},\,10^{-2}]_{\log}1.31e-61.05e-33.69e-4
 DD learning rate †[10−8, 10−2]log⁡[10^{-8},\,10^{-2}]_{\log}1.06e-51.52e-41.17e-4
 Batch size[64, 512]2k[64,\,512]_{2^{k}}256512512
 Weight decay[10−7, 10−3]log⁡[10^{-7},\,10^{-3}]_{\log}1.26e-72.06e-41.12e-6
Appendix
Table 9: Same as Table 7 , but using DANN.
   Best run
   sim-to-simsim-to-realsim-to-real
ArchitectureParameterSearch SpaceHPGeLaBr 3NaI(Tl)
MLPLearning rate[10−7, 10−3]log⁡[10^{-7},\,10^{-3}]_{\log}1.34e-48.95e-49.41e-4
 Batch size[64, 512]2k[64,\,512]_{2^{k}}512512128
 Weight decay[10−7, 10−3]log⁡[10^{-7},\,10^{-3}]_{\log}1.37e-61.19e-66.11e-7
 Dropout[0.0, 0.4][0.0,\,0.4]0.2978.24e-30.0839
Appendix
Table 10: Same as Table 7 , but using DeepCORAL.
   Best run
   sim-to-simsim-to-realsim-to-real
ArchitectureParameterSearch SpaceHPGeLaBr 3NaI(Tl)
MLPLearning rate[10−7, 10−3]log⁡[10^{-7},\,10^{-3}]_{\log}9.25e-47.15e-47.24e-4
 Batch size[64, 512]2k[64,\,512]_{2^{k}}512512256
 Weight decay[10−7, 10−3]log⁡[10^{-7},\,10^{-3}]_{\log}8.33e-65.23e-69.17e-4
 Dropout[0.0, 0.4][0.0,\,0.4]0.2070.1770.216
Appendix
Table 11: Same as Table 7 , but using DeepJDOT.
   Best run
   sim-to-simsim-to-realsim-to-real
ArchitectureParameterSearch SpaceHPGeLaBr 3NaI(Tl)
MLPLearning rate[10−7, 10−3]log⁡[10^{-7},\,10^{-3}]_{\log}9.06e-45.36e-48.24e-4
 Batch size[64, 512]2k[64,\,512]_{2^{k}}5126464
 Weight decay[10−7, 10−3]log⁡[10^{-7},\,10^{-3}]_{\log}7.45e-76.91e-66.83e-6
 Dropout[0.0, 0.4][0.0,\,0.4]0.02290.1830.121
Appendix
Table 12: Same as Table 7 , but using Mean Teacher.
   Best run
   sim-to-simsim-to-realsim-to-real
ArchitectureParameterSearch SpaceHPGeLaBr 3NaI(Tl)
MLPLearning rate[10−7, 10−3]log⁡[10^{-7},\,10^{-3}]_{\log}5.85e-43.21e-46.21e-4
 Batch size[64, 512]2k[64,\,512]_{2^{k}}128128128
 Weight decay[10−7, 10−3]log⁡[10^{-7},\,10^{-3}]_{\log}2.00e-51.99e-45.92e-7
 Dropout[0.0, 0.4][0.0,\,0.4]0.1240.1010.311
Appendix
Table 13: Same as Table 7 , but using SimCLR.
Table 14: Calculated pp -values from a one-sided Wilcoxon signed-rank test, testing the null hypothesis that UDA provides no improvement over the source-only baseline model. pp -values below a significance threshold of 0.010.01 are bolded, indicating statistically significant improvement.
Figure 3: UMAP visualizations of the feature extractor outputs for source spectra (‘ o ’ markers) and target spectra (‘ + ’ markers) for all UDA methods for the sim-to-sim domain adaptation scenario. Color indicates domain (blue = source, red = target).
Figure 4: Same as Fig. 3 , but for the sim-to-real LaBr 3 domain adaptation scenario. Color indicates class.
Figure 5: Same as Fig. 3 , but for the sim-to-real NaI(Tl) domain adaptation scenario. Color indicates class.
More in this topic

Synthetic-to-Real Domain Adaptation

Synthetic-to-real domain adaptation transfers models trained on simulated or generated data to real-world settings by reducing distribution shifts.

See all 137 papers

Explore similar work

CardsList

  1. 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

  2. Joint Distribution Alignment for Universal Domain Adaptation

    Aug 25, 2026Shizhe Li, Hongshan Pu, Mengying Xie +2Domain AdaptationDistribution Matching

    School of Software Engineering, South China University of Technology, Guangzhou 510006, China · School of Future Technology, South China University of Technology, Guangzhou 511442, China · College of Computer Science, Chongqing University, Chongqing 400044, China

  3. Pseudo-Labeling for Unsupervised Domain Adaptation with Kernel GLMs

    Mar 19, 2026Nathan Weill, Kaizheng WangUnsupervised Domain AdaptationCovariate Shift

    Department of IEOR, Columbia University. · Department of IEOR and Data Science Institute, Columbia University.