Searches for neutrinoless double-beta decay, a hypothetical nuclear decay that would show neutrinos are their own antiparticles, are limited by counting statistics. They expect at most a handful of signal events against a handful of background events. In that regime two numbers set the reach of the experiment: the fraction of background a selection cut removes, and how precisely the cut's signal acceptance is known. Germanium experiments make this cut using the shape of the current pulse each event produces. Today that cut is a single hand-built ratio, $A/E$, the peak current divided by the energy. It is calibrated with radioactive sources that sit outside the crystal, and the mismatch between those calibration events and decays spread uniformly through the crystal shows up as an acceptance uncertainty of 3.5% in the MAJORANA analysis (Multisite event discrimination for the majorana demonstrator).
We propose a learned frequency-domain filter for these pulses that adapts to only one thing: the electronic noise spectrum of each detector and run, measured from signal-free baseline traces. It does not adapt to the content of the event being classified. Learnable spectral filters can separate structure from noise more cleanly than fixed filters do. Keeping the filter blind to event content means it cannot sculpt the signal it is supposed to count, so its acceptance can be predicted from calibration data. The deliverable is a pulse-shape cut that should reject more multi-site background than $A/E$ at equal acceptance and whose acceptance stays stable across noise conditions. We will judge it by its effect on median half-life sensitivity, not by classifier accuracy.
Core Claim
The obvious way to combine these two directions is to take an instance-adaptive spectral filter from vision and apply it to detector waveforms. Instance-adaptive here means the filter coefficients are predicted from each input. Examples are the per-pixel hyperparameter network in An Integrated Framework of Frequency-Domain Denoising with Learnable Parameters in Variational Autoencoders and the soft gate between frequency bands in HarmoniAD. We argue this is the wrong kind of adaptivity when the goal is a counting measurement. If the filter depends on the event, the cut's acceptance depends on event features such as interaction depth. Those features are distributed differently in calibration events, which sit near the surface, and in the signal events, which are uniform through the bulk. The calibration-to-bulk systematic then grows instead of shrinking. The file also notes that freely adapting spectral choices multiply hidden trial factors, the extra chances to find a fluctuation that inflate false discoveries.
Our claim is that the adaptivity should depend on the measured noise power spectrum $\hat S_n(k)$ at frequency index $k$. That quantity can be estimated without any physics events. The filter acts on the full complex spectrum, so phase is kept:
$$Y(k) = H_\theta\big(k;\, \hat S_n\big)\, X(k),$$
Here $X(k)$ is the discrete Fourier transform of the baseline-subtracted, energy-normalized current pulse. $H_\theta$ is a small network that outputs a complex gain per frequency and is initialized at the classical Wiener filter, the textbook noise-optimal linear filter. $Y(k)$ is passed to a compact classifier.
Phase matters because the timing of the staggered charge arrivals that mark a multi-site event lives largely in phase, and Double Chooz's spectral discriminator discarded it by using power alone. Training adds a penalty on how much the acceptance of single-site calibration events varies across noise conditions. This makes acceptance stability a training target rather than something checked after the fact.
The new contribution is not "a neural network on pulse spectra." It is a filter whose adaptivity is restricted to an observable that carries no signal, plus an evaluation that converts background rejection and acceptance uncertainty into half-life sensitivity.
Research Program
The method comes from three sources. Double Chooz showed that moving pulse discrimination into the frequency domain helps a great deal: the figure of merit was $Q = 3.56$ against $1.42$ for the best time-domain method (Novel event classification based on spectral analysis of scintillation waveforms in Double Chooz). Its discriminator, however, was a fixed linear sum of power-spectrum bins. The learnable FFT-VAE showed that learned frequency parameters beat a fixed FFT filter on images. It was tested only under Gaussian noise and needed an extra regularizer to keep its predicted parameters stable. Initializing at Wiener and conditioning on a measured noise spectrum is our answer to that instability. Holl et al. (Deep learning based pulse shape discrimination for germanium detectors) showed that a time-domain autoencoder with a 141-parameter classifier matches $A/E$. They also found multi-site pulses whose high first current peak fools $A/E$, which motivates looking at the later structure in the pulse. They evaluated a single detector setup and did not study drift in the noise spectrum. That drift is the condition our method is designed for.
We will build four things. The first is a waveform pipeline: energy normalization, alignment, DFT, and a noise-spectrum estimate per channel per run, taken from pretrigger baselines and forced-trigger traces. The second is the filter $H_\theta$. It is a residual on top of Wiener with a few hundred parameters, conditioned on a low-dimensional summary of $\hat S_n$. The third is a classifier trained on $^{228}$Th calibration data. Double-escape-peak events at 1592 keV serve as the single-site proxy and full-energy-peak events as the multi-site proxy, following both Holl et al. and MAJORANA. The fourth is a sensitivity calculator. It takes the background surviving the cut, the acceptance, and the acceptance uncertainty and computes median 90% CL half-life sensitivity with Poisson toy experiments. When background dominates, sensitivity grows roughly as $\varepsilon/\sqrt{B}$, because the excess must exceed background fluctuations of size $\sqrt{B}$. An acceptance systematic enters the toys as a nuisance parameter.
The data will come from two kinds of sources. Real calibration and baseline data from point-contact germanium detectors cover several noise conditions, for example different channels, runs, and deliberately altered shaping or grounding on a test stand. Simulated pulses with known interaction positions, overlaid with recorded noise traces, are the only way to measure the calibration-to-bulk acceptance shift directly. That is because the true single-site/multi-site label and the depth are known in simulation.
Baselines
The naive reference is no pulse-shape cut, energy selection only. It sets the background level that every cut divides.
The physics baseline is the MAJORANA $AvsE$ cut (Multisite event discrimination for the majorana demonstrator), recalibrated per run as in that analysis. It reports $(90 \pm 3.5)\%$ single-site retention, about 6% multi-site survival in the $^{208}$Tl single escape peak, and a factor-of-3 reduction in background index in the region of interest. It is the strongest competitor on stability because it has run for years with only a scalar to calibrate.
The strongest learned baseline is the Holl et al. time-domain autoencoder plus classifier (Deep learning based pulse shape discrimination for germanium detectors), which reports 90% double-escape-peak acceptance at 90% full-energy-peak rejection. We will reimplement it at the same 256-sample input length. We will also train a second version with the same recorded-noise augmentation our method sees, so that any gain cannot be credited to augmentation.
The frequency-domain baseline is a Double Chooz-style linear power-spectrum discriminator (Novel event classification based on spectral analysis of scintillation waveforms in Double Chooz), refit on germanium pulses. No reported germanium result was found on Lacuna for it.
Three ablations isolate the key ingredient. In the fixed-Wiener ablation, the noise-conditioned residual is removed. In the content-conditioned ablation, the filter is predicted from the event itself, in the style of the FFT-VAE network. In the power-only ablation, phase is dropped.
Every method gets the same calibration events, the same train/validation/test split by run, the same hyperparameter search budget, and an operating point fixed at 90% double-escape-peak acceptance on validation runs. Beating the baselines means two things at once on held-out runs and detectors. Multi-site survival must be lower than both $AvsE$ and the augmented Holl model. The spread in single-site acceptance must be no larger than $AvsE$, and the simulated calibration-to-bulk acceptance shift must be smaller than $AvsE$'s. Together these must yield a higher median half-life sensitivity in the toy experiments.
Why not just…?
Why not just whiten the pulses with the measured noise spectrum and feed them to an existing CNN? This is the most serious threat. If whitening or a fixed Wiener filter followed by the Holl classifier matches our method, then the learned residual adds nothing. In that case the honest result is "noise-matched preprocessing is what matters." The fixed-Wiener ablation and the whitened-Holl baseline test this directly. We will report it as a positive finding for the simpler pipeline if they tie. Our reason to expect a gap is limited but concrete. Wiener filtering is optimal for estimating a known signal shape under stationary Gaussian noise. Germanium noise includes non-stationary pickup and microphonics. The discrimination task also depends on phase structure that a gain-only Wiener filter does not reweight with the classification goal in mind.
Why not just train the time-domain network on many noise conditions? Augmentation can teach invariance implicitly, and convolutions can in principle learn any filter. We include the augmented Holl model for exactly this reason. The difference we test is extrapolation. Our filter is given the noise spectrum explicitly, so it should handle a held-out run with a new spectrum. An augmented network has to have seen something similar. The held-out-detector split will show whether this matters.
Why not just keep $A/E$ and recalibrate it every run? It works, it is stable, and its systematics are understood. A learned cut is worth adopting only if it improves sensitivity after its own systematic is charged against it. Better rejection alone is not enough. The sensitivity toys make that comparison explicit. A gain in rejection that comes with a larger acceptance uncertainty would count as a loss.
Why not use a content-adaptive filter, as in vision? The FFT-VAE and HarmoniAD gates gain power by reacting to each input. In a counting measurement that same reaction couples acceptance to event depth and topology. The content-conditioned ablation, evaluated on simulated bulk-versus-surface single-site events, will show whether this coupling is real. If it turns out negligible, the restriction we impose is unnecessary, and we will say so.
Decisive Checks
The first test uses held-out runs and detectors. At 90% double-escape-peak acceptance we compare multi-site survival in the $^{208}$Tl single escape peak and full-energy peak against $AvsE$'s roughly 6%. We also compare the run-to-run standard deviation of single-site acceptance.
The second test uses simulation overlaid with real noise. We measure acceptance for uniformly distributed bulk single-site events minus acceptance for source-like near-surface single-site events. This is the quantity behind MAJORANA's 3.5% uncertainty. Lower is better.
The third test checks for sculpting. On background-only Compton-continuum data and toys, the acceptance as a function of energy must be smooth through $Q_{\beta\beta}$, and the cut must create no localized excess. Any trained filter that produces a spurious bump fails regardless of its rejection.
The fourth test combines the measured background, acceptance, and acceptance uncertainty in Poisson toy experiments with a nuisance parameter. It reports median 90% CL half-life sensitivity relative to $AvsE$ and to the Holl model. EXO-200's 25% sensitivity gain from a spatial network (Search for Neutrinoless Double-β Decay with the Complete EXO-200 Dataset) gives a sense of what a meaningful improvement from analysis alone looks like. That result comes from a different detector type, and EXO-200 itself names mismatch between simulation and data as a dominant systematic. That is why our acceptance-stability penalty is trained on real calibration data rather than on simulation.
The proposal succeeds if the noise-conditioned filter beats whitened and augmented baselines on held-out-detector sensitivity. It fails informatively if fixed Wiener filtering plus an existing classifier captures the whole gain.
References
- Multisite event discrimination for the majorana demonstrator
- An Integrated Framework of Frequency-Domain Denoising with Learnable Parameters in Variational Autoencoders
- HarmoniAD: Harmonizing Local Structures and Global Semantics for Anomaly Detection
- Novel event classification based on spectral analysis of scintillation waveforms in Double Chooz
- Deep learning based pulse shape discrimination for germanium detectors
- Search for Neutrinoless Double-<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mi>β</mml:mi></mml:mrow></mml:math>Decay with the Complete EXO-200 Dataset