Calibration-Conditioned Waveform Encoders for Germanium Detector Arrays

Experiments that search for extremely rare nuclear decays, such as neutrinoless double-beta decay, decide whether each recorded event is "signal-like" by looking at the shape of a microsecond-long electrical pulse from a germanium crystal. Current machine-learning approaches compress each pulse into one fixed vector and then train a tiny classifier on top. The classifier is trained on calibration data, meaning events from a radioactive source that stand in for real signal and background. The compression itself never sees which detector the pulse came from or what that detector's calibration looks like. That matters because a tonne-scale array contains many crystals with different shapes, electric fields and noise. A pulse feature that separates signal from background in one crystal can be useless in another.

We propose to make the pulse encoder take an extra input: a small set of calibration pulses from the same detector and run. This set acts as a "task description" that rescales the encoder's internal features, the way few-shot image classifiers adapt to a new set of classes from a handful of examples. The aim is one encoder, trained across many simulated and real detectors, that reaches per-detector-tuned discrimination on a new crystal from a few hundred calibration events. It should also do better on the rare background subclasses that a fixed compression throws away.

Core Claim

The claim is that the right context for a germanium pulse-shape classifier is that detector's own calibration data, not hand-picked operating variables. When a learned encoder is conditioned on this context, two things should improve together. First, it should need fewer calibration events to reach a target background rejection on a detector it has never seen. Second, it should reject more of the rare multi-site background subclass that fools the standard current-amplitude cut.

The motivating evidence comes from both directions. On the physics side, Deep learning based pulse shape discrimination for germanium detectors trains an unsupervised convolutional autoencoder on 256-sample pulses and compresses each into a fixed 7-dimensional vector. A 141-parameter classifier on that vector reaches 90% rejection of full-energy-peak background at 90% acceptance of double-escape-peak signal proxies. This exactly matches the physics-derived amplitude-over-energy ($A/E$) cut. Along the way, the autoencoder exposed a subclass of multi-site events with sharp current peaks that fool $A/E$. The compression was tuned for reconstruction, not discrimination, and was never asked which detector it was looking at.

On the representation side, TADAM: Task dependent adaptive metric for improved few-shot learning shows that a support-set summary fed through feature-wise scale-and-shift layers (FiLM) improves 5-way 5-shot mini-Imagenet accuracy from 68.2% for prototype averaging to 76.7%. Long-Tailed Learning Requires Feature Learning proves that fixed feature maps fail on rare categories: 0.7% versus 99.8% for learned features on their synthetic task. That proof uses discrete symbolic data, so we treat it as motivation for watching the tail rather than as a guarantee for continuous waveforms.

The closest physics precedent for conditioning is Model-Based Deep Learning Algorithm for Pulse Shape Discrimination in High Event Rates. Its weighting network turns hand-crafted pileup variables (spacing to neighboring pulses, energy ratios) into a vector that reweights features for neutron/gamma separation in CLYC scintillators. It reaches about 98.7% accuracy at 10 Mcps. We differ in three ways. Our context is a set of calibration waveforms rather than a few scalars. Our target is detector-to-detector variation in an ultra-low-background array rather than pileup. And we test the conditioning by deliberately feeding the wrong detector's calibration.

Research Program

The model has two parts. The first is a one-dimensional convolutional encoder over the raw current pulse. Its intermediate layers are modulated by per-channel scale $\gamma$ and shift $\beta$. The second is a set encoder that produces $(\gamma, \beta)$ from a calibration context $C_d$ for detector $d$. That context contains a few hundred double-escape-peak pulses (signal proxy) and full-energy-peak pulses (background proxy), plus unlabeled continuum pulses. The set encoder pools per-pulse embeddings separately by label, in the spirit of TADAM's prototype averaging. It also cross-attends between the unlabeled pulses and the two labeled pools, borrowing the inter-sample attention of Cross Attention Network for Few-shot Classification. The output is a signal-likeness score per event.

Training is episodic. Each episode samples one detector and run, draws a context set and a disjoint query set from that detector, and scores the query set. Across episodes the model sees many geometries (point-contact and coaxial), impurity profiles, noise levels and electronics responses. Most of this variety comes from pulse simulation. Real calibration runs from a smaller number of detectors are mixed in. TADAM needed an auxiliary classification loss with an annealing schedule to avoid collapse, so we start from a pretrained, unconditioned encoder. We switch on the modulation with an initialization at $\gamma=1, \beta=0$ rather than training the conditioning from scratch.

Two analyses accompany the build. The first is a tail analysis: we isolate the $A/E$-fooling multi-site subclass identified by Holl et al. and report rejection on it separately. The second is a sculpting analysis: we check that signal-proxy acceptance stays flat across energy and drift-time bins. A data-driven filter in a counting experiment must not invent or remove structure, a risk emphasized by the same-sign control channel in Isolating Unisolated Upsilons with Anomaly Detection in CMS Open Data.

Baselines

The naive reference is the physics cut. Per-detector $A/E$ is calibrated on the same $^{228}$Th data and the same context events the model receives. It defines the floor, at the 90% background rejection / 90% signal acceptance operating point reported in Deep learning based pulse shape discrimination for germanium detectors.

The first learned baseline is that paper's own pipeline. It uses a static autoencoder bottleneck plus a small classifier, with the classifier retrained per detector on the context events only. Its reported number is the same 90%/90%. Beating it means higher full-energy-peak rejection at 90% double-escape-peak acceptance on held-out detectors, at equal calibration statistics.

The strongest competitor is a supervised convolutional classifier with the same architecture and parameter count as ours, but no conditioning. It is pretrained on the full multi-detector pool and then fine-tuned on each new detector's context set. Hyperparameters, the number of fine-tuning steps and early stopping are tuned on the same validation detectors and with the same budget as our model. This is the honest test, because it gets the same data and the same feature learning.

The second strongest competitor is a weighting network in the style of Model-Based Deep Learning Algorithm for Pulse Shape Discrimination in High Event Rates. It conditions only on scalar metadata: detector type, bias voltage, baseline noise level and drift time. For this method the file reports about 98.7% neutron/gamma accuracy at 10 Mcps on CLYC. We will reimplement it on that pileup task to confirm our reimplementation, then run it on the germanium task.

The key ablation feeds our model a context set drawn from a different detector. If performance does not drop, the conditioning is not doing work and the claim fails.

For alpha surface events specifically, the reference is delayed charge recovery combined with $A/E$, from α-event characterization and rejection in point-contact HPGe detectors. It rejects 100% of surface alphas beyond 3 mm from the detector center within statistics, at 0.2% bulk sacrifice. We do not expect to beat this. It serves as a check that conditioning does not degrade the slow-tail information that cut uses.

Why not just…?

Why not just fine-tune a pretrained classifier on each detector's calibration data? This is the real threat. With enough calibration statistics it may match the conditioned model. Our claim is limited to the low-calibration regime and to the tail subclass. We will sweep context size from about 50 to several thousand events per class and report the crossover point. If fine-tuning matches us at every size, we will report the conditioning as unnecessary. There is a practical argument for conditioning: per-detector weights multiply across a large array and must be re-validated after every recalibration, whereas one conditioned model only swaps its context set. But we will not lean on that argument if accuracy is a tie.

Why not just append detector metadata as extra inputs to one supervised network? Metadata such as bias voltage and detector type misses detector-specific field irregularities and charge-trapping behavior. Those show up only in the pulses themselves, which is the reason Deep learning based pulse shape discrimination for germanium detectors gives for difficulty in coaxial detectors. The weighting-network baseline is exactly this alternative in its strongest published form. Comparing it with our model on held-out geometries answers the question directly.

Why not just widen the autoencoder bottleneck so less is discarded? A larger reconstruction bottleneck still spends capacity on the dominant pulse features. Holl et al. show that the informative multi-site structure is a small secondary feature. We include a 32-dimensional static autoencoder as an extra baseline. If it closes the gap on the tail subclass, the problem is compression size, not lack of conditioning, and we would revise the claim.

Proposed Experiments

The primary testbed is $^{228}$Th calibration data in germanium, using double-escape-peak and full-energy-peak events as in Holl et al. Real calibration runs come from collaboration detectors, with LEGEND Full Chain Test stand data as in Machine learning-powered data cleaning for LEGEND where access permits. These are supplemented by simulated pulses over a grid of geometries. Evaluation always uses leave-detector-out splits, with an entire geometry family held out for the hardest test.

The metrics are four. The first is full-energy-peak rejection at 90% double-escape-peak acceptance. The second is rejection on the $A/E$-fooling subclass. The third is the number of calibration events needed to reach 90% rejection. The fourth is the maximum deviation of signal-proxy acceptance across energy and drift-time bins.

The secondary testbed is the CLYC pileup setting. Here the weighting network's reported 98.7% at 10 Mcps is the reference. Context sets will be drawn from the same rate regime, which asks whether waveform-set conditioning adds anything beyond hand-crafted pileup scalars.

The decisive checks are three. The wrong-detector context must hurt performance, and the right context must help most on held-out geometries. At small calibration budgets the conditioned model must beat the fine-tuned supervised baseline, and the size of that budget must be stated. And none of the gains may come with energy-dependent acceptance. For COCO-FACET-style analogies in physics, no reported result found on Lacuna for germanium tail-subclass rejection beyond the qualitative observation in Holl et al. That number will be established here, not compared against.