Searches for neutrinoless double-beta decay, a radioactive process that would show neutrinos are their own antiparticles, look for a handful of events per year inside germanium crystals. Most of what these detectors record is background, mainly gamma rays. Signal and gamma background can be told apart by the shape of the electrical pulse each event produces. A signal deposits its energy at one point in the crystal. A gamma ray usually scatters several times and deposits energy at several points. Neural networks now separate these two classes as well as the standard hand-built cut. They have one serious weakness: there are no labeled signal events to train on. Labels come from a calibration source, and the "signal-like" calibration events sit near the crystal surface, while real decays happen uniformly through the bulk. A classifier can learn shortcuts tied to position, such as how long the charge takes to drift to the contact. Those shortcuts are correct on calibration data and wrong on the events that matter. The result is an uncertainty in signal acceptance. For the standard cut this is ±3.5%, and it feeds directly into the final physics result.
We propose to attack this uncertainty with noise injection used deliberately. During training, the classifier sees random perturbations only along the directions physics says should not matter: where in the crystal the event occurred, the detector's real electronic noise, and the long pulse tail. The directions that carry the "one site versus many sites" information are left intact. The claim is that targeted noise makes the classifier's acceptance flat across event position. That flatness should shrink the gap between calibration and bulk acceptance without losing background rejection. Generic white Gaussian noise of the kind used elsewhere in machine learning should fail to do this, because it blurs the fast rising edge that carries the discriminating information.
Background and the Gap
The detectors in question are high-purity germanium (HPGe) crystals with a small point-like readout contact. Each event is recorded as a digitized charge waveform with three parts: a flat baseline, a fast rise as charge drifts to the contact, and a slow exponential decay set by the readout electronics. Multisite event discrimination for the majorana demonstrator established the standard discriminant, $A/E$, which is the peak current divided by the energy.
That cut is tuned on a thorium-228 calibration source. The source supplies two kinds of reference events:
- Signal proxy. The double-escape peak at 1592 keV contains single-site events.
- Background proxies. The full-energy peak and the single-escape peak contain mostly multi-site events.
The cut retains (90 ± 3.5)% of single-site events and cuts single-escape events to about 6%. The ±3.5% is attributed explicitly to the difference in position between calibration events and bulk events. This paper motivates the proposal, because it identifies the uncertainty we target.
Deep learning based pulse shape discrimination for germanium detectors supplies the architecture we reuse. It compresses 256-sample waveforms into a 7-dimensional code with a convolutional autoencoder, then classifies that code with a 141-parameter network. It reaches 90% background rejection at 90% signal acceptance and catches multi-site events that $A/E$ misses. It trains on clean, unperturbed pulses and never measures whether its acceptance depends on position. That omission is the gap.
CycleGAN-driven transfer learning for electronics response emulation in high-purity germanium detectors (CPU-Net) translates simulated pulses into realistic ones. It improves the overlap of drift-time distributions between simulation and data, for example from 5.4% to 22.5% IoU on double-escape events. It also reports unstable training and leaves open whether its output helps a downstream classifier. We use it in two ways: as a tool to build realistic simulated bulk-decay test sets, and as a competing source of training augmentation.
On the noise side, three papers supply the mechanism and one constrains it:
- Noise as kernel smoothing. Kernel Regression and Backpropagation Training With Noise shows that training with input noise behaves like kernel smoothing, with the noise distribution acting as the kernel. This suggests that the shape of the noise, not just its size, decides which directions the classifier is smoothed along. That is the formal hook for using anisotropic, targeted noise.
- Noise and label contamination. Feature Noise Boosts DNN Generalization Under Label Noise shows that isotropic feature noise limits memorization of corrupted labels. This is relevant because the double-escape peak sits on a Compton continuum that contaminates its labels. The work is only on images, and it notes that the right noise level must be tuned.
- Noise on irrelevant spans. Explaining the Efficacy of Counterfactually Augmented Data shows that noising non-causal spans moves a classifier toward causal features. Noising causal spans destroys out-of-domain accuracy, dropping it by about 28.7% on Yelp.
- The constraint. From Data Quality to Model Quality: an Exploratory Study on Deep Learning shows that badly structured perturbations badly damage performance. Together with the span-noising result, this tells us not to noise the rising edge.
Core Claim
A neural pulse-shape classifier trained on calibration data with noise confined to nuisance directions will have signal acceptance that is nearly independent of event position. Its predicted bulk-decay acceptance will therefore match its calibration acceptance more closely than either the $A/E$ cut or an unregularized network, at equal background rejection. Isotropic noise of the same total power will not achieve this.
Research Program
The network follows Holl et al.: an autoencoder that compresses each pulse to a short code, plus a small classifier. We add three noise sources, each with a physical justification.
Electronic noise overlay. We add real baseline traces recorded in the same run and channel to the training waveforms. This makes the noise color and amplitude realistic rather than white.
Span noise. We replace the pre-trigger baseline and the late tail with resampled segments from other events. The tail is the RC-decay region, and it encodes the electronics and slow surface charge rather than the event's topology. This follows the logic of Kaushik et al.
Nuisance-subspace latent noise. This is the central ingredient. Within the double-escape sample, every event is single-site, so variation in the autoencoder code among these events is, to first order, position and electronics rather than topology. We estimate that variation and inject noise in the code along it:
$$ \tilde z = z + \epsilon,\qquad \epsilon \sim \mathcal{N}\!\left(0,\ \lambda^2\, P\,\Sigma_{\text{DEP}}\,P\right). $$
The symbols are:
- $z$ is the autoencoder code of a training event.
- $\Sigma_{\text{DEP}}$ is the covariance of codes among double-escape events, after sideband subtraction of the underlying continuum.
- $P$ projects onto the top directions of $\Sigma_{\text{DEP}}$ that correlate with measured drift time. Drift time is the time from event start to the current peak and serves as our position proxy.
- $\lambda$ is a single scale, tuned on calibration data alone.
In the language of Koistinen and Holmström, this replaces a round smoothing kernel with one elongated along position. The classifier is smoothed where it should be invariant and stays sharp where single-site and multi-site pulses differ.
Safeguard. A projection that also captures topology would erase signal information. To check for this, we measure the classifier's separation of double-escape from single-escape events with and without the noise. A drop in rejection beyond the statistical error flags leakage.
Label contamination. The continuum under the double-escape peak is handled by sideband-weighted labels for every method compared, so that label cleaning does not confound the comparison.
Baselines
The do-nothing reference is the Holl et al. architecture trained on the same calibration events with no noise. This is the most direct test of whether noise of any kind helps. It must reach roughly the published 90% background rejection at 90% signal acceptance before any comparison is trusted.
The physics reference is a drift-time-corrected $A/E$ cut tuned as in Alvis et al. Its reported operating point is (90 ± 3.5)% single-site retention with the single-escape peak reduced to about 6%. It is the strongest competitor on the actual question, because it is transparent and its systematic uncertainty is already characterized. Beating it means two things together: an equal or lower single-escape survival fraction at 90% double-escape acceptance, and a smaller calibration-to-bulk acceptance shift.
The strongest machine-learning competitor for transfer is the same classifier trained on CPU-Net-translated simulated pulses. Those pulses can include simulated bulk single-site events, so this route attacks the position bias directly. It gets the same classifier, the same calibration data, and the same hyperparameter budget.
The generic regularization baseline is isotropic Gaussian feature noise in the style of Zeng et al. It is applied both to waveforms and to codes, at matched total noise power and with the same tuning budget for its scale.
The ablations add the three noise sources one at a time. The key ablation is the electronic overlay plus span noise without the latent nuisance noise. This isolates whether the targeted latent term, rather than realistic noise in general, is responsible for any gain.
Fairness. For every method, the noise scale is tuned only on calibration data, using a fixed number of configurations, never on the bulk test sets.
Why not just…?
Why not just reweight calibration events so their drift-time distribution matches the bulk? Importance weighting is the obvious statistical fix, and it is a real threat. It cannot create events in regions the calibration sample barely covers. The deep interior near the contact is exactly where external-source double-escape events are rare. It also inflates variance where weights are large. We include reweighted training as an additional baseline. If it closes the calibration-to-bulk gap as well as nuisance noise does, we will report that the cheaper method suffices. We expect it to win where drift-time coverage overlaps and to fail on the short-drift tail. Our binned acceptance curves show this directly.
Why not just train on CPU-Net-translated simulations of bulk decays? This is the principled route, and we test it head-on. The CPU-Net authors report run-to-run instability and inherited simulation flaws, such as overly narrow peak distributions. Both would show up as acceptance biases that noise injection does not have. The comparison is decided by acceptance agreement on the realistic simulated bulk test set and on data. The two approaches may also combine. Training on translated pulses with nuisance noise added is one of our arms.
Why not just use isotropic Gaussian noise, which already has theoretical backing? The evidence on structured perturbations, from Kaushik et al. and He et al., suggests isotropic noise strong enough to blur position information will also blur the rising edge. The rising edge is where single-site and multi-site pulses differ. The matched-power isotropic baseline tests this. If it matches the targeted noise, the anisotropy idea is wrong.
Why not just keep $A/E$? It is transparent. However, Holl et al. show it misses multi-site pulses whose secondary structure a learned model catches. Its ±3.5% uncertainty is exactly what we aim to reduce. If the network cannot beat it on both rejection and uncertainty, the network is not worth deploying.
Proposed Experiments
Data. We use thorium-228 calibration data from point-contact detectors, the MAJORANA and LEGEND workflows described in the file. This includes the inverted-coaxial detector and the Geant4/siggen simulations used by CPU-Net. Access depends on collaboration cooperation.
Calibration benchmark. We report single-escape and full-energy survival at 90% double-escape acceptance, with statistical errors.
First decisive check: position flatness. We plot double-escape acceptance in bins of drift time and summarize it by the slope and the spread across bins. Nuisance noise should flatten this curve relative to the unregularized network and $A/E$, at unchanged rejection.
Second decisive check: controlled bulk test. We simulate single-site decays uniformly through the bulk and translate them to data-like pulses with CPU-Net's emulator. We also simulate double-escape events with identical processing. The difference in acceptance between the two is our estimate of the calibration-to-bulk systematic. We compare it against the ±3.5% reported for $A/E$. Because CPU-Net may itself carry biases, we repeat this with raw siggen pulses plus overlaid real baselines and check that the ranking of methods is the same.
Third check: robustness to electronics. We train on one run and test on runs with different noise floors, measuring rejection drift.
On data. Where the collaboration has sufficient exposure, acceptance of the two-neutrino double-beta continuum gives a genuine bulk single-site check. This decay is common, occurs throughout the bulk, and is single-site. No reported result was found on Lacuna for this measurement with neural pulse classifiers.
Success criterion. A visibly smaller calibration-to-bulk shift than both the unregularized network and $A/E$, with single-escape survival no worse than about 6%. If the targeted latent noise does not beat both the overlay-only ablation and drift-time reweighting on the shift metric, we will conclude that the targeting adds nothing.
References
- Multisite event discrimination for the majorana demonstrator
- Deep learning based pulse shape discrimination for germanium detectors
- CycleGAN-driven transfer learning for electronics response emulation in high-purity germanium detectors
- Kernel Regression and Backpropagation Training With Noise
- Feature Noise Boosts DNN Generalization under Label Noise
- Explaining the Efficacy of Counterfactually Augmented Data
- From Data Quality to Model Quality: An Exploratory Study on Deep Learning