A comparative study of machine learning and conventional methods for determining the dead layer thickness of an HPGe detector
N.D. Thong, N.H.K. Vi, L.N.D. Uyen, N.V. Thiem, T.T. Thanh, V.T. Minh, P.L. Ho, C.T. Tai, C.V. Tao
Radiation Physics and Chemistry 249(2026)114162
Abstract:
The dead layer of p-type HPGe detectors grows progressively due to lithium diffusion, degrading detection efficiency at low gamma-ray energies. This study compares four regression approaches for dead layer estimation— conventional G4-scan interpolation, single-source Linear Regression (LR), single-source Random Forest (RF), multi-source LR and multi-source RF combining 241Am and 109Cd — applied to an ORTEC GEM50P4-83 detector at two epochs separated by eight years. A single Geant4 simulation campaign (N = 107 events/run, 1301 points) trained all models, with GUM-compliant uncertainties throughout. All four ML predictions agree with the G4-scan reference (1.323 ± 0.019 mm) within 0.006 mm for the 2018 dataset. When Beer–Lambert linearity is confirmed (R2 > 0.998) and features are restricted to 𝑙𝑛(𝜖), Linear Regression achieves a crossvalidated MAE of 0.009 mm, outperforming all tree-based benchmarks (MAE = 0.010 mm), consistent with the Gauss–Markov theorem. A quantitative threshold analysis shows that multi-source LR reduces total uncertainty by 27% when 𝛿𝜖Cd ∕𝜖Cd < 2.04% — a condition satisfied by the present measurements. Dead layer growth rates of 0.208 mm/year (2015–2018) and 0.009 mm/year (2018–2026) are consistent with nonlinear lithium diffusion deceleration.
More detail: https://doi.org/10.1016/j.radphyschem.2026.114162
