Speaker
Description
In conventional energy spectrometry that uses Gas Electron Multiplier (GEM) detectors, a serious issue is severe spectral dispersion. This dispersion blurs the measured energy deposition profiles, making them less distinct. At the same time, a dominant low energy pedestal overlap appears, were signals from different events merge together at lower energies. Because of these limitations traditional analysis methods become ineffective. Simple energy windowing techniques and principal component analysis (PCA) based methods fail to provide reliable results. Consequently, these conventional approaches cannot achieve robust discrimination between different nuclides when the radiation sources have diverse or varying spectral characteristics.
To overcome this, we introduce a generalized classification methodology based on 1D CNN trained on an Augmented Log Z Score spectral fingerprint. Raw, variable length event lists are converted into fixed length (2048 bin) spectra, followed by a logarithmic transformation for variance stabilization and global Z score standardization. This preprocessing ensures that the CNN learns important intrinsic spectral shape independent of source activity, counting time, or specific radionuclide identity. The 1D CNN architecture consists of three convolutional blocks with aggressive regularization to extract subtle, hierarchical features necessary for fine discrimination. Evaluated on unseen test data under trained operational conditions, the model achieves very high overall generalization accuracy of >98% and cent per cent accuracy for pure, single source isotopes. Furthermore, the model demonstrates successful interpolation, accurately mapping samples measured at untrained integration times to their nearest trained spectral neighbors. However, extrapolation tests involving out of distribution geometric conditions cause prediction confidence to collapse (46–60%), revealing a “timing compression bias” and a breakdown in mixture isolation for unknown source configurations. These results establish the Augmented Log Z Score 1D CNN as a high fidelity tool for nuclear spectroscopy. Simultaneously defining the critical requirement for geometric diversity in training data to enable robust analysis across unknown radiation sources and environments.
Keywords: GEM detector, 1D CNN, nuclear spectroscopy, nuclide identification, spectral fingerprint, machine learning
Acknowledgements: This work was supported by the project “Implementation of the Activities of the Lithuanian Particle Physics Consortium” (Project No. SV5240067), funded by the state budget of the Republic of Lithuania, and by the project “Development and Applications of Microstructured Gaseous Electron Detectors” (Project No. S-CERN-24-8), funded by the Research Council of Lithuania under the programme Prospective R&D Projects for Lithuania's Membership at CERN.
| Name of the speaker | Muhammad usman uraf Ishafaq |
|---|---|
| Eligible for the Georges Charpak Young Scientist Award. | yes |