Recognition: no theorem link
Pulse shape discrimination for α event rejection in BEGe-type high-purity germanium detectors
Pith reviewed 2026-05-14 18:08 UTC · model grok-4.3
The pith
Classifiers trained only on gamma-ray pulse shapes can reject alpha events in germanium detectors with over 27,000-to-1 efficiency while preserving more than 80% of signal-like events.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
Machine learning classifiers trained exclusively on gamma ray data from a BEGe-type HPGe detector can achieve strong rejection of alpha events while maintaining high survival for signal-like events. The multilayer perceptron yields greater than 80% survival for signal-like events, below 20% for background-like events, and an alpha-rejection factor exceeding 2.71 times 10 to the fourth, showing that gamma-only training suffices for alpha discrimination in addition to standard multi-site gamma rejection.
What carries the argument
The multilayer perceptron classifier applied to pulse shape features extracted from gamma calibration data, which distinguishes event topologies without any alpha training samples.
Load-bearing premise
Pulse shapes produced by polonium alphas deposited on a thin gold foil on the detector surface represent the alpha events that will occur during actual long-term underground operation.
What would settle it
An underground measurement recording real alpha events from detector surface contaminants that the gamma-trained classifier fails to reject at a factor above 10,000.
read the original abstract
High-purity germanium detectors are widely used in rare-event searches due to their excellent energy resolution and extremely high intrinsic (radio)purity. In experiments searching for neutrinoless double beta decay in $^{76}$Ge such as LEGEND, pulse shape discrimination is required to suppress multi-site $\gamma$ events. In this work, we investigate whether pulse shape discrimination classifiers trained exclusively on $\gamma$ ray data can be used to identify and reject $\alpha$ events, without the need for dedicated $\alpha$ training. In detectors such as LEGEND, the total number of registered $\alpha$ events over the experiment lifetime is expected to be insufficient to train dedicated classifiers, while still contributing to the background. Two approaches based on machine learning are studied: a multilayer perceptron and a projective likelihood classifier. The p+ surface of a point-contact semi-planar germanium detector was exposed to $^{209}$Po and $^{210}$Po sources deposited on a thin gold foil. Two measurement campaigns were performed, yielding $1.36\times10^{5}$ and $1.87\times10^{6}$ $\alpha$ events, respectively. Both classification methods achieve efficient separation of single-site and multi-site $\gamma$ events while strongly reducing the $\alpha$ component. The multilayer perceptron provides the best overall performance, with a signal-like event survival greater than 80%, a background-like event survival below 20%, and an $\alpha$-rejection factor exceeding $2.71\times10^{4}$. These results demonstrate that robust pulse shape discrimination for high-purity germanium detectors can be achieved using training information derived solely from $\gamma$ events, providing a promising strategy for next-generation neutrinoless double beta decay searches.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript investigates whether pulse shape discrimination (PSD) classifiers trained exclusively on gamma-ray data can reject alpha events in BEGe-type high-purity germanium detectors for experiments such as LEGEND. Using two measurement campaigns exposing the p+ surface to 209Po and 210Po sources on thin gold foil (yielding 1.36e5 and 1.87e6 alpha events), the authors compare a multilayer perceptron and a projective likelihood classifier, reporting that the MLP achieves signal-like event survival >80%, background-like event survival <20%, and an alpha-rejection factor >2.71e4 while maintaining efficient single-site vs. multi-site gamma separation.
Significance. If the alpha pulse shapes are representative, the result is significant because alpha backgrounds in long-term underground operation are expected to be too few for dedicated training data yet still contribute to the background budget. Demonstrating robust gamma-only training for alpha rejection would provide a practical PSD strategy for next-generation 76Ge neutrinoless double-beta decay searches, reducing reliance on scarce alpha calibration data.
major comments (2)
- [Experimental setup and data collection] Experimental setup (alpha source description): the reported rejection factor rests on 209Po/210Po events deposited on thin gold foil at the p+ contact with fixed ~5 MeV energy and specific charge-collection geometry. The manuscript must quantify or simulate how these pulse shapes (rise time, A/E, current-pulse features) compare to expected real backgrounds from internal 210Po or 222Rn daughters, which may occur at n+ surfaces, with partial energy loss, or different drift paths; without this, the >2.71e4 rejection factor cannot be directly extrapolated to LEGEND data.
- [Classification methods and performance evaluation] Classification performance (MLP section): the signal-like (>80%) and background-like (<20%) survival fractions and the alpha-rejection factor are central claims, yet the training procedure, feature selection, cross-validation method, and exact definition of the survival cuts are not described with sufficient quantitative detail (e.g., no explicit statement of the held-out test fraction or how the 1.87e6-event sample was partitioned). This prevents independent verification of the quoted metrics.
minor comments (2)
- [Abstract] The abstract uses 'background-like event survival' without defining whether this refers exclusively to multi-site gamma events or includes other components; add a clarifying sentence.
- [Results figures] Figure legends should state the exact number of events and any post-selection cuts applied when reporting the survival and rejection numbers.
Simulated Author's Rebuttal
We thank the referee for their thorough review and constructive comments on manuscript arXiv:2605.13498. We address each major comment point by point below and have revised the manuscript accordingly to improve clarity and reproducibility.
read point-by-point responses
-
Referee: Experimental setup (alpha source description): the reported rejection factor rests on 209Po/210Po events deposited on thin gold foil at the p+ contact with fixed ~5 MeV energy and specific charge-collection geometry. The manuscript must quantify or simulate how these pulse shapes (rise time, A/E, current-pulse features) compare to expected real backgrounds from internal 210Po or 222Rn daughters, which may occur at n+ surfaces, with partial energy loss, or different drift paths; without this, the >2.71e4 rejection factor cannot be directly extrapolated to LEGEND data.
Authors: We agree that our alpha sources provide a controlled exposure at the p+ surface with fixed energy and geometry. In the revised manuscript we will add a dedicated paragraph in the discussion section that qualitatively compares the observed pulse-shape features (rise time, A/E, current amplitude) to those expected for internal 210Po and 222Rn-daughter events. We note that the classifiers are trained exclusively on gamma data and rely on general surface-event characteristics rather than alpha-specific templates; therefore the reported rejection factor is presented as a conservative lower bound. We will cite relevant literature on charge-collection differences between p+ and n+ surfaces to support the extrapolation argument, while acknowledging that a full Monte-Carlo simulation of partial-energy n+ events lies beyond the scope of the present work. revision: yes
-
Referee: Classification performance (MLP section): the signal-like (>80%) and background-like (<20%) survival fractions and the alpha-rejection factor are central claims, yet the training procedure, feature selection, cross-validation method, and exact definition of the survival cuts are not described with sufficient quantitative detail (e.g., no explicit statement of the held-out test fraction or how the 1.87e6-event sample was partitioned). This prevents independent verification of the quoted metrics.
Authors: We apologize for the insufficient detail. The revised manuscript will expand the MLP subsection to state: (i) the full feature list (rise time, A/E, current-pulse maximum, integral, and two additional shape parameters); (ii) the data split of the gamma training sample (70 % training, 15 % validation, 15 % test); (iii) 5-fold cross-validation performed on the training portion; and (iv) the survival cuts defined at the 80 % gamma signal efficiency working point. The entire 1.87e6 alpha-event sample is treated as an independent test set with zero overlap to the gamma training data. These additions will enable independent reproduction of the quoted survival fractions and rejection factor. revision: yes
Circularity Check
No circularity: empirical ML performance measured on held-out alpha data
full rationale
The paper reports direct experimental measurements: classifiers (MLP and projective likelihood) are trained exclusively on gamma-ray data and evaluated on separate alpha datasets collected from 209Po/210Po sources on gold foil (1.36e5 and 1.87e6 events). Survival fractions (>80% signal-like, <20% background-like) and the alpha-rejection factor (>2.71e4) are computed from event counts in the test sets. No derivation chain, equation, or fitted parameter reduces to its own inputs by construction. No self-citations are invoked as load-bearing uniqueness theorems, and no ansatz or renaming of known results occurs. The result is self-contained against the collected data.
Axiom & Free-Parameter Ledger
axioms (1)
- domain assumption Pulse-shape differences between single-site gamma, multi-site gamma, and surface alpha events are distinguishable in BEGe detectors
Reference graph
Works this paper leans on
-
[1]
CUPID Collaboration, Phys. Rev. Lett.129, 111801 (2022). DOI 10.1103/PhysRevLett.129.111801
-
[2]
NEMO Collaboration, Phys. Rev. D92, 072011 (2015). DOI 10.1103/PhysRevD.92.072011
-
[3]
AMoRE Collaboration, Phys. Rev. Lett.134, 082501 (2025). DOI 10.1103/PhysRevLett.134.082501
-
[4]
CUPID Collaboration, The European Physical Jour- nal C82(11), 1033 (2022). DOI 10.1140/epjc/ s10052-022-10942-5
-
[5]
CUORE Collaboration, Science390(6777), 1029 (2025). DOI 10.1126/science.adp6474
-
[6]
KamLAND-Zen Collaboration, Phys. Rev. Lett.135, 262501 (2025). DOI 10.1103/jkf6-48j8
-
[7]
EXO-200 Collaboration, Phys. Rev. Lett.123, 161802 (2019). DOI 10.1103/PhysRevLett.123.161802
-
[8]
GERDA Collaboration, Phys. Rev. Lett.125, 252502 (2020). DOI 10.1103/PhysRevLett.125.252502
-
[9]
Heidelberg-Moscow Collaboration, The European Physical Journal A - Hadrons and Nuclei12(2), 147 (2001). DOI 10.1007/s100500170022
-
[10]
IGEX Collaboration, Phys. Rev. D65, 092007 (2002). DOI 10.1103/PhysRevD.65.092007
-
[11]
Final result of the ma- jorana demonstrator’s search for neutrinoless double-β decay in 76Ge (2023)
MAJORANA Collaboration. Final result of the ma- jorana demonstrator’s search for neutrinoless double-β decay in 76Ge (2023). DOI 10.1103/PhysRevLett.130. 062501
-
[12]
LEGEND Collaboration, Phys. Rev. Lett.136, 022701 (2026). DOI 10.1103/25tk-nctn
-
[13]
Jany, et al., The European Physical Journal C81(1), 38 (2021)
A. Jany, et al., The European Physical Journal C81(1), 38 (2021). DOI 10.1140/epjc/s10052-020-08781-3
-
[14]
GERDA Collaboration, The European Physical Journal C73, 1 (2013)
2013
-
[15]
GERDA Collaboration, The European Physical Journal C82(4), 284 (2022) 9
2022
-
[16]
MAJORANA Collaboration, Physical Review C99(6), 065501 (2019)
2019
-
[17]
M. Misiaszek, K. Panas, M. Wojcik, G. Zuzel, M. Hult, The European Physical Journal C78(5), 392 (2018). DOI 10.1140/epjc/s10052-018-5852-7
-
[18]
H. Zhang. The optimality of naive Bayes (2004)
2004
-
[19]
Bishop,Pattern Recognition and Machine Learn- ing
C.M. Bishop,Pattern Recognition and Machine Learn- ing. No. 4 in Information science and statistics (Springer, 2006). DOI 10.1117/1.2819119
-
[20]
Prince,Understanding Deep Learning(The MIT Press, 2023)
S.J. Prince,Understanding Deep Learning(The MIT Press, 2023)
2023
-
[21]
Agostini, et al
M. Agostini, et al. pygama (2025). DOI 10.5281/ zenodo.17594917
2025
-
[22]
Cowan,Statistics for Searches at the LHC(Springer International Publishing, Cham, 2015), pp
G. Cowan,Statistics for Searches at the LHC(Springer International Publishing, Cham, 2015), pp. 321–355. DOI 10.1007/978-3-319-05362-2_9
-
[23]
Vischia, PoSConfinement2018, 249 (2019)
P. Vischia, PoSConfinement2018, 249 (2019). DOI 10.22323/1.336.0249
-
[24]
Flynn, Analytica chimica acta43, 221 (1968)
W. Flynn, Analytica chimica acta43, 221 (1968)
1968
-
[25]
O. Erbacher, Naturwissenschaften20(22-24), 390 (1932) Appendix A: Preparation of Po sources The most widely used approach for the preparation of Po sources is spontaneous deposition onto a silver disc, as de- scribed by Flynn [24]. Polonium can also be relatively eas- ily deposited on other noble metals, such as Pt or Pd, from diluted hydrochloric acid so...
1932
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.