Counting Charge Clusters: Variable-Dimension Waveform Models for Germanium Background Rejection

Searches for neutrinoless double-beta decay in germanium detectors must keep a signal that deposits all its energy at one point, while rejecting gamma-ray backgrounds that scatter and deposit energy at several points. Today the main tool is a single number read off each pulse: the peak current divided by the energy, called $A/E$. Neural alternatives treat each pulse as a fixed 256-sample vector and compress it with a reconstruction loss. We propose a different kind of model. It asks directly how many separate charge clusters produced a pulse, allowing zero, one, two, or more. For each cluster that exists, it infers that cluster's arrival time and charge. Clusters that do not exist get no parameters at all.

The idea comes from speech synthesis. Pitch exists during voiced sounds and simply does not exist during unvoiced ones. The multi-space distribution, or MSD (a probability model whose sample space is a union of pieces with different dimensions), was built so that systems would stop filling the absent pitch with invented values. Germanium pulses have the same structure, because a secondary charge cluster is either present with a continuous time and charge, or absent. We claim that a discriminant built on the posterior probability of "exactly one cluster" will retain more single-site signal than $A/E$ or a reconstruction-loss autoencoder at the same background rejection. We also claim this model can be checked against a computable statistical limit on how close two clusters can be and still be told apart. Models that average over this ambiguity, rather than represent it, cannot be checked that way.

Core Claim

A multi-site pulse is the sum of a variable number of current pulses from charge clusters. Each cluster has a drift-time offset $t_k$ and charge fraction $q_k$, and each passes through the detector's electronics. Write $\theta_{1:K}$ for the cluster parameters and $K\in\{0,1,2,\dots,K_{\max}\}$ for the cluster count. The model is

$$p(x) = \sum_{K} w_K \int p(x \mid K,\theta_{1:K})\, p(\theta_{1:K}\mid K)\, d\theta_{1:K},$$

where $x$ is the digitized waveform, $w_K$ is the prior weight of each subspace, and the $K=0$ subspace is baseline-only noise with no continuous parameters. This is an MSD with more than two subspaces. The proposed signal discriminant is $P(K=1\mid x)$.

This differs from padding with zero-charge slots. In a fixed $K_{\max}$-slot model, an absent cluster is a continuous parameter pushed to zero. A regression or likelihood fit then places probability mass on tiny phantom clusters. A multi-space distribution (MSD) approach to speech recognition of tonal languages shows the analogous bias in pitch modeling, and Improve the Implementation of Pitch Features for Mandarin Digit String Recognition Task documents the related mismatch caused by imputed values. Both are direct analogues of what we expect for germanium pulses, though they were measured on speech.

The sensitivity limit enters as a quantitative ceiling. Two clusters separated by drift-time difference $\Delta t$ can only be distinguished if $\Delta t$ is large compared with the Cramér–Rao bound, the smallest variance any unbiased estimator can reach:

$$\mathrm{Var}(\widehat{\Delta t}) \ge \mathcal{I}(\Delta t)^{-1}.$$

Here $\mathcal{I}$ is the Fisher information, meaning how sharply the waveform likelihood changes with $\Delta t$ under the detector's noise and electronics response. Predicting the single-site and multi-site event discrimination power of dual-phase time projection chambers applied exactly this analysis to spatial separation in xenon detectors and found that maximum-likelihood and CNN estimators approach the bound. That paper supplies our evaluation method. We carry it into the time domain for germanium pulses and ask a sharper question: does a discriminant's multi-site detection rate as a function of $\Delta t$ track the bound-implied curve, or does it fall short because training averaged over the ambiguity?

Research Program

The forward model uses standard germanium pulse simulation. Geant4 energy depositions are passed through siggen-style charge drift, as in CycleGAN-driven transfer learning for electronics response emulation in high-purity germanium detectors, which gives every simulated pulse a ground-truth cluster count and cluster times. Deposits are merged into clusters by a spatial merging radius, and the merging radius is reported as a sensitivity parameter. Electronics effects are handled by a small parametric response: the RC tail constant, rise-time filtering, overshoot, and a noise spectrum estimated from pre-trigger baselines. We fit these parameters to real calibration pulses. We also test whether passing simulated pulses through a CPU-Net emulator changes the conclusions.

Inference is amortized and trained on simulation. One network head predicts the subspace weights $P(K\mid x)$. For each $K$, a mixture-density head predicts the posterior over $\theta_{1:K}$. We use mixture-density heads because Rich Prosody Diversity Modelling with Phone-Level Mixture Density Network showed that mixture outputs keep multimodal targets that single-Gaussian and mean-squared-error heads flatten. Training maximizes the likelihood of the true $(K,\theta)$, so the model never regresses a mean waveform.

On real thorium-228 calibration data there are no per-event labels. We therefore run expectation-maximization over the subspace weights and electronics parameters, the same branching inference that MSD hidden Markov models use, as in A one-step tone recognition approach using MSD-HMM for continuous speech. This produces a useful by-product: an estimate of how many events in each calibration peak are truly single-site. The double-escape peak (DEP) proxy for signal is not pure, and neither is the full-energy peak (FEP) proxy for background. Purely discriminative classifiers trained on these proxy labels inherit that contamination as a ceiling.

The analysis has two parts. In simulation, we compute the Fisher information for the two-cluster model as a function of $\Delta t$, charge ratio, and signal-to-noise ratio. From it we derive the best achievable two-versus-one detection curve, using a likelihood-ratio test on the true simulator as an oracle. On real data, we compare discriminants at fixed operating points using the calibration peaks.

Baselines

The naive reference is the classical $A/E$ cut as used by Multisite event discrimination for the majorana demonstrator. That paper reports $(90 \pm 3.5)\%$ single-site retention, a reduction of the thallium-208 single escape peak (SEP) to about 6% of its original count, and a factor of 3 suppression of the background index. We will tune $A/E$ on our own data with the standard energy-dependent correction, using the same events and the same 90% DEP acceptance point. That way the comparison does not rely on a number transferred from a different detector.

The strongest existing learned method is the autoencoder plus small classifier of Deep learning based pulse shape discrimination for germanium detectors. It compresses 256-sample pulses to 7 latent dimensions with a reconstruction loss and reports 90% FEP rejection at 90% DEP acceptance, matching $A/E$. We reimplement it on identical splits, give it the same hyperparameter search budget and seeds as our model, and also give it access to the simulated pretraining data. Beating it means a statistically significant gain in DEP acceptance at 90% FEP rejection, or lower SEP survival at 90% DEP acceptance. Significance is judged against the binomial uncertainty of the calibration samples. With roughly 3,000 DEP events, as in the CPU-Net dataset, the acceptance uncertainty is about $\sqrt{0.9\cdot 0.1/3000}\approx 0.5\%$. A gain must therefore be at least about 1.5 percentage points to be trustworthy, and we aggregate over detectors to tighten this.

A strong supervised CNN trained directly on DEP-versus-FEP labels, with the same budget, is included as a direct test of whether the generative structure matters at all.

The key ablation isolates the MSD ingredient. It keeps the same network, simulator, and training but replaces the variable-dimension space with $K_{\max}$ always-present slots, where absent clusters are represented by charge driven to zero. The discriminant becomes one minus the summed charge in slots 2 through $K_{\max}$. A second ablation replaces the likelihood heads with mean-squared-error regression of $\theta$, to isolate the effect of over-smoothing.

Why not just…?

Why not just train a CNN classifier on DEP versus FEP? This is the most serious competitor, and it may match us at the standard operating point. Its weakness is that the training labels are contaminated, because FEP contains genuine single-site events and DEP contains some multi-site ones. A classifier learns the proxy rather than the physics. We test this in simulation, where true labels exist. Both methods are evaluated against true cluster counts as contamination is varied. If the CNN matches us on truth labels even at realistic contamination, our advantage on headline numbers disappears. What we would still contribute is calibrated cluster-time posteriors and peak-purity estimates, and we will say so plainly.

Why not just zero-pad absent clusters with a presence flag? This is the fixed-slot ablation above. Evidence from pitch modeling says imputed values create bias and pass errors downstream (Robust F0 modeling for Mandarin speech recognition in noise). Whether the effect matters for germanium pulses is an empirical question. The test looks at single-site pulses at low signal-to-noise ratio, where phantom secondary clusters with small charge should appear in the padded model and inflate false multi-site calls.

Why not just use a diffusion model of waveforms and a likelihood ratio? Diffusion models avoid mean collapse, as shown in Beyond Oversmoothing: Evaluating DDPM and MSE for Scalable Speech Synthesis in ASR. That paper reports a WERR of 1.46, against an MSE plateau around 2.4–2.5. But a diffusion model gives no explicit cluster count and no posterior over $\Delta t$. That makes comparison with the Fisher-information limit indirect, and it requires per-class density models trained on contaminated proxy classes. We include a conditional diffusion likelihood-ratio baseline in simulation only. If it matches our detection curve, the case for the variable-dimension structure rests on interpretability alone.

Why not just keep $A/E$? $A/E$ is robust and well understood. However, Deep learning based pulse shape discrimination for germanium detectors found multi-site pulses with high current peaks that fool it. In xenon detectors, Search for Neutrinoless Double-β Decay with the Complete EXO-200 Dataset raised signal efficiency from 80.8% to $96.4\pm3.0\%$ by using full topology. The comparison against the Fisher-information limit shows directly how much separation information $A/E$ throws away as a function of $\Delta t$.

Proposed Experiments

In simulation, we plot multi-site detection probability against $\Delta t$ at fixed single-site false-positive rate for every method. Each curve is compared with the oracle likelihood-ratio curve and with the Fisher-information resolution scale. The decisive simulation result is that the MSD posterior tracks the oracle curve close to the bound, while the mean-squared-error and padded ablations fall measurably short in the region near the bound.

On real thorium-228 calibration data we report three things for every method: DEP acceptance at 90% FEP rejection, SEP survival at 90% DEP acceptance, and the variance across seeds. Seed variance matters because CPU-Net reports multi-peaked results across training runs. We also report the expectation-maximization estimates of single-site purity for DEP and FEP and cross-check them against simulation predictions.

Three outcomes would refute the proposal. The first is that the padded ablation matches the MSD model, which would mean the variable-dimension structure adds nothing. The second is that the supervised CNN matches us on truth labels under realistic contamination. The third is that the electronics domain gap, measured by the drift-time and peak-current overlap metrics of CPU-Net, is larger than any discrimination gain, in which case simulation-trained posteriors cannot be trusted on data.