Verifying Each Germanium Pulse Against Its Own Noise

Searches for neutrinoless double-beta decay in germanium detectors look for a handful of events per year. Before any physics cut is applied, the data must be "cleaned": electronic noise, cross-talk, and pileup (two pulses overlapping in one recorded trace) must be removed without throwing away real events. The current best automated cleaner for LEGEND learns one fixed decision boundary from a few thousand hand-labelled waveforms. At high event rates it still lets about 15% of anomalous events through. Computer vision faced a similar problem and found a different answer. Its trackers build a small model of the specific object, and of the background immediately around it, while they run, and they check each hypothesis against that local model rather than against a model of the whole class.

We propose to bring that habit to germanium waveforms. Each channel keeps a single-pulse template that is updated online, but only from trusted events. Each event's own pre-trigger baseline, the stretch of trace recorded just before the pulse, supplies a noise model for that event alone. A trace is then accepted only if one pulse from this channel, under this event's noise, explains it. Put in two sentences: instead of asking "does this waveform look like the physics class?", we ask "is this waveform fully explained by one pulse from this detector as it behaves today, given the noise measured a few microseconds earlier?" We expect this to catch pileup and slow drifts that a static classifier misses, and we design the experiments so that a well-fed static classifier gets every chance to prove us wrong.

Core Claim

The leakage problem in waveform cleaning comes largely from judging each waveform in isolation, without its own temporal context and without the detector's current state. A hypothesize-and-verify test fixes both, provided the test is specialized per channel and per event at runtime. The test is a residual likelihood ratio. Let $x$ be the recorded trace and $s_c$ the current template for channel $c$. Fit amplitude $a$ and start time $t_0$. Estimate a noise autocovariance $\Sigma_e$ from event $e$'s own pre-trigger window, shrunk toward the channel's running average because one window is short. The statistic is

$$\Lambda_e = \min_{a,t_0} (x - a\,s_c(\cdot - t_0))^\top \Sigma_e^{-1} (x - a\,s_c(\cdot - t_0)) \;-\; \min_{a_1,a_2,t_1,t_2} (x - a_1 s_c(\cdot - t_1) - a_2 s_c(\cdot - t_2))^\top \Sigma_e^{-1} (\cdots).$$

The second term uses the same residual form with a two-pulse fit, with amplitudes $a_1, a_2$ and start times $t_1, t_2$. A large $\Lambda_e$ means a second pulse explains the trace much better than one pulse does. Traces where even the two-pulse residual is improbable under $\Sigma_e$ are flagged as non-physical noise.

The specific claim is narrow. At the physics-event sacrifice reported for the static cleaner, this test should reduce high-rate anomalous leakage. It should also hold that performance under drift in detector response, where a frozen classifier degrades.

Relation to Prior Work

The cleaner we compete with is León et al.'s LEGEND data cleaning. It compresses waveforms with a wavelet transform, clusters them with affinity propagation (an exemplar-finding method whose memory cost grows with the square of the number of events, which caps training at about 10,000 events), has experts label the clusters, and trains a support vector machine. That paper motivates this proposal in two ways. Its high-rate leakage is the gap we target. It also states that changing conditions force re-clustering and re-labelling, which is the burden online adaptation would remove. It classifies each waveform without the baseline that precedes it; we use that baseline as the event's local background.

The mechanism comes from vision. Ramanan's segmentation-based verification rejects false detections by building a colour model from a window's interior and its surrounding ring on the fly. Our pre-trigger window plays the role of the ring. Real-Time MDNet freezes shared features and retrains only a small target-specific head during tracking. Likewise, exemplar-based LDA tracking and Vondrick and Ramanan's instance-specific annotation models update instance models online. Our analogue is a fixed fitting procedure with a per-channel template that changes. The adaptive Gaussian-mixture work of Ladicky et al. and Woyczyk et al. supports re-estimating noise statistics per instance rather than globally.

The same literature also limits the design. Self-trained instance models drift toward whatever they mistakenly accept. We therefore gate template updates on events from calibration-source runs and in known energy lines, rather than on the cleaner's own accepted events. This follows the anchor idea in The truth about cats and dogs, where a reliably detected part seeds the appearance model for the rest of the body.

On the physics side, Arnquist et al.'s delayed charge recovery shows that a one-dimensional cut on a pulse's tail rejects surface alphas with 0.2% bulk loss. The same paper notes that surface charge drifts over time. Its success shows that physically motivated pulse-shape statistics are hard to beat, so we compete on a different kind of event: overlaps and noise rather than surface alphas. CATHODE in CMS Open Data also contrasts data against a learned background, but its background density is global over the dataset; ours is per event. The MAGIC random forest study warns that models trained on simulation degrade on real data. That supports building our templates from measured events rather than simulated ones.

Research Program

The system has three parts. The first is a per-channel template store, initialized from calibration data and updated by an exponential moving average. Only calibration-line events whose single-pulse residual is well within the noise may update it. The second is a per-event noise estimator. It computes a short-lag autocovariance from the pre-trigger window and shrinks it toward the channel's running estimate, with the shrinkage weight set by the window length. The third is the verification statistic $\Lambda_e$, together with a goodness-of-fit flag for non-physical traces.

The whole pipeline is a few small least-squares fits per event, with no pairwise matrix. It therefore avoids the memory ceiling of affinity propagation. The new anomaly classes that affinity propagation discovers can still be surfaced, as clusters of high-residual events.

Ablations target each ingredient separately:

  1. Template frozen at initialization versus updated online.
  2. Per-event noise versus a single channel-wide noise estimate.
  3. Gated updates versus ungated self-training, to measure drift directly.

Baselines

The naive reference is a fixed set of hand-tuned cuts on baseline RMS, rise time, and a simple second-peak search in the current pulse. This is the kind of manual tuning León et al. set out to replace. It is run on the same data with thresholds tuned on the same labelled split. Any learned method must clearly improve on it, or none of this matters.

The strongest existing method is the wavelet, affinity-propagation, and support-vector pipeline of León et al.. On the Full Chain Test stand high-rate data it reports a maximum physics sacrifice of 0.024±0.004% and a maximum anomalous leakage of 15.034±0.224%. On low-rate data it reports 0.000±0.021% sacrifice and 0.000±0.330% leakage. Low-rate data therefore offers little headroom, and the contest is on high-rate data. We will re-implement this pipeline and confirm we reproduce its numbers before comparing. It will receive the same labelled exemplars and the same expert labelling budget, and it may retrain whenever our method's templates update. Beating it means lower high-rate leakage, outside the stated uncertainty, at sacrifice no worse than 0.024%.

We add a stronger static competitor that is not in the file. It is a convolutional classifier on the raw trace plus the pre-trigger window, trained on the same labels with a matched tuning budget. It isolates whether gains come from online specialization or simply from seeing the context.

The key-ingredient ablation is our own method with templates and noise frozen at day zero. Any difference between it and the full method, measured under drift, is the value of runtime specialization.

Why not just…?

Why not just append the pre-trigger baseline as extra features to the static classifier? This is the real threat. A flexible static model given the baseline window may learn the same comparison. The convolutional baseline with context is built to test exactly this. If it matches our leakage on stationary data, the per-event noise component adds nothing there. The proposal would then reduce to its drift claim, and we would report that plainly. We expect a gap under drift, because a static model has no way to learn what the detector's response will look like next year.

Why not just retrain the support vector machine periodically? That is the remedy León et al. themselves describe, and with enough labelling effort it would track drift. The cost is repeated expert re-labelling and the 10,000-event clustering cap. We will compare at an equal labelling budget: retraining at the same cadence as our template updates, using exemplars drawn from the same calibration events. If periodic retraining matches us at zero extra human effort, our advantage shrinks to cost and latency, which we would then quantify.

Why not just fit a two-pulse model with a fixed template, as pileup rejection often does? A fixed template and global noise estimate is precisely ablations 1 and 2 combined. If it performs as well as the full method, the specialization is unnecessary. The comparison is therefore built into the program rather than argued away.

Proposed Experiments and Decisive Checks

The primary testbed is the Full Chain Test stand data, which includes both high-rate detector data and quiet dummy-board noise. We need access to these data through the LEGEND collaboration; this is a dependency, not a given. Labels follow León et al.'s expert-assigned classes. Metrics are anomalous leakage and physics sacrifice, both with binomial uncertainties, and leakage reported at matched sacrifice.

Drift is tested two ways. First, in held-out data we inject controlled shifts in gain, shaping time, and noise colour. Second, wherever the data span multiple runs, we train on early runs and test on later ones. We report leakage as a function of shift size for every method.

Three outcomes would be decisive.

  1. If leakage falls clearly below 15.03% at no more than 0.024% sacrifice, and the context-fed static classifier does not close the gap, the core claim holds.
  2. If our method's advantage over frozen baselines grows with injected drift while ungated self-training degrades, both the online update and the anchor gating are justified.
  3. If the context-fed static classifier and periodic retraining both match us, the claim fails. We would then publish the method as a memory-light substitute for affinity-propagation clustering only.

A secondary check applies the same per-event baseline noise model to the delayed-charge-recovery tail slope on Arnquist et al.'s scanner data. The target there is to reduce the reported 0.2% bulk loss without losing the full alpha rejection that paper reports.