Typing Every Pulse: Energy-Decorrelated, Open-Set Event Classification for Germanium Rare-Event Searches

Experiments that look for extremely rare nuclear decays, such as neutrinoless double-beta decay in germanium detectors, record millions of electrical pulses. Almost all of them are not the signal. Some are instrumentation junk: electronic glitches, baseline jumps, and two pulses overlapping in time (pileup). Others are real radiation that looks almost like the signal: gamma rays that deposit energy at several points in the crystal, and alpha particles that hit the detector surface. Today these are removed by separate, sequential filters. A data-cleaning stage flags non-physical pulses, and hand-built pulse-shape cuts then reject each physical background with its own single number. We propose one learned classifier that assigns every pulse a type. The types are signal-like, multi-site gamma, surface alpha, pileup, glitch, and a catch-all "unknown" that gets split into new types as they appear. The classifier is trained under one extra constraint: within each type, its decision must carry no information about the pulse's energy.

That constraint is the point, not a decoration. A rare-event search counts events in a narrow energy window, and it learns how well its cuts work from calibration peaks at other energies. If a classifier's acceptance drifts with energy, two things go wrong. The efficiency measured on calibration no longer holds at the search energy, and the surviving background spectrum can be bent into shapes that look like a signal. Collider physicists call this "mass sculpting," and they solved it only for a binary anomaly score. Our claim is that fine-grained typing plus per-type energy decorrelation lets each background component be counted and calibrated separately in the search window. It also lets a genuinely new background source surface as a new cluster rather than as unexplained excess. A binary clean/dirty filter cannot give either result.

What exists and what is missing

Three pieces of prior work each supply part of this, and none combines them.

Machine learning-powered data cleaning for LEGEND is the closest system and our main baseline. It compresses germanium waveforms with a wavelet transform and groups them with affinity propagation, a clustering method that picks representative "exemplar" pulses. Experts label the exemplars, and a support vector machine is trained on them. It separates non-physical artifacts from physics pulses. It does not type among physical interactions. It cannot scale past roughly ten thousand training events because affinity propagation needs quadratic memory. It reports 15.034±0.224% anomalous-event leakage on high-rate data, at a maximum physics sacrifice of 0.024±0.004%. That leakage figure is the gap we most directly target.

Robust Anomaly Detection for Particle Physics Using Multi-Background Representation Learning supplies the method we borrow. It trains a classifier across several known backgrounds so that its internal features carry structure. It then removes dependence on the search variable with Nuisance Randomized Distillation (NuRD), which combines importance reweighting with an adversary that tries to predict jet mass from the features. It scores anomalies by maximum logit or Mahalanobis distance. On jet images it reports AUROC 0.914 and a sculpting measure (Jensen–Shannon divergence of the background mass shape) of 0.118, against 0.881 and 0.255 for a background-only autoencoder. This result also motivates our constraint, since it shows sculpting is real and fixable. Its limits are that it outputs a single out-of-distribution score, it has never been applied to time-series pulses, and it gives no type to what it flags.

Trident supplies the open-set discovery mechanism. It keeps a separate one-class model per traffic type and sets rejection thresholds with extreme value theory, a statistical model of the tails of each type's score distribution. It clusters rejected samples into new types without retraining existing ones, and reports about 90% adjusted mutual information on unknown attack classes versus below 60% for DBSCAN. It works only on network features and has no physics invariance constraint.

The broader anomaly-detection literature constrains our design. Evaluation of Machine Learning Algorithms for Anomaly Detection and LIVABLE document that multi-class anomaly classifiers degrade on long-tailed class distributions. Rare types therefore need care, which is why we keep a density-based rejection option rather than relying on softmax. The failure of most methods on the multi-mode Black Box 3 in The LHC Olympics 2020 is the collider-side evidence that lumping heterogeneous anomalies into one score loses them.

Core Claim

A single waveform encoder can meet two conditions at once. First, it separates instrumentation artifacts and physical background types in one shared representation. Second, it is trained so that, given the true type, its representation is statistically independent of deposited energy. Writing $z$ for the encoder output, $y$ for the event type and $E$ for energy, the target is

$$ I(z; E \mid y) = 0, $$

where $I$ is mutual information. This is exactly the conditional independence NuRD's reweighting-plus-adversary enforces, now applied with a multi-class label instead of a background/anomaly label.

We make three predictions under this constraint. The first is lower leakage than the cleaning-then-cuts chain at matched signal sacrifice. The second is per-type acceptance that is flat across calibration energies, so efficiencies transfer to the search window. The third is that pulses rejected by every known type can be clustered into physically coherent new types.

Research Program

The encoder is a one-dimensional convolutional network on raw waveforms, with energy supplied only to the adversary and never to the encoder. Class structure comes from three label sources of different quality.

Instrumentation classes (glitch, baseline shift, pileup) come from the expert-labeled exemplars of the LEGEND data-cleaning work on the Full Chain Test stand data. They are expanded by propagating exemplar labels to neighbors, and pileup is augmented by overlaying real single pulses at controlled time offsets.

Physical classes come from calibration data of the kind used in CycleGAN-driven transfer learning for electronics response emulation on the V06643A detector. Its double-escape peak is rich in signal-like single-site events, and the single-escape and full-energy regions are rich in multi-site events. Because these peaks are mixtures, we train with mixture-aware weak supervision in the spirit of the classification-without-labels methods reviewed in the LHC Olympics paper, rather than treating peak membership as a clean label.

Simulated pulses passed through that paper's learned electronics emulator give a third, fully labeled but imperfect source. The emulator's reported mode instability becomes a known systematic that we vary. Surface alphas use the characterization of α-event characterization and rejection in point-contact HPGe detectors where suitable data are available.

On top of the encoder we fit one Gaussian per known type in feature space, as in NuRD's Mahalanobis scoring, and set per-type rejection thresholds with extreme value tail fits, as in Trident. Events rejected by every type go to a buffer. The buffer is clustered with a scalable method such as mini-batch k-means or a streaming variant, which removes the affinity-propagation memory ceiling. A cluster is registered as a new type once it is stable across resampling, and registration adds a Gaussian without retraining the encoder. The deliverable for a physics analysis is a type-resolved count in the search window, with per-type efficiencies measured on calibration and a stated uncertainty on their energy transfer.

Baselines

The naive reference is no machine-learned typing. This is the existing hand-built chain: expert cleaning cuts followed by the standard single-number pulse-shape cuts for multi-site events and surface alphas (the delayed-charge-recovery cut for alphas). It sets the floor any proposal must clear. The file contains no reported leakage-at-fixed-sacrifice for this chain on these datasets (no reported result found on Lacuna for the standard germanium pulse-shape-cut chain on FCT or V06643A data). We therefore run it ourselves with cut values tuned on the same calibration split.

The strongest existing system on this exact data is the affinity-propagation-plus-support-vector-machine cleaner of Machine learning-powered data cleaning for LEGEND. It reports 15.034±0.224% leakage at 0.024±0.004% sacrifice on high-rate data, and 0.000±0.330% leakage on low-rate data. We rerun it on identical train and test splits with its published pipeline. For cleaning, beating it means lower leakage at equal or lower sacrifice on the high-rate set, outside the combined uncertainties. Low-rate data already sits at zero leakage and serves only as a no-regression check. Because it does not type physical backgrounds, we chain it with the hand-built cuts to make an end-to-end competitor.

The strongest learned competitor is a NuRD-style encoder applied to waveforms following Robust Anomaly Detection for Particle Physics Using Multi-Background Representation Learning. It is trained binary, as known backgrounds against everything else, with a single out-of-distribution score. It uses the same encoder architecture, data and tuning budget as our method, meaning the same number of hyperparameter trials and the same GPU hours. On its native benchmark it reports AUROC 0.914 and sculpting divergence 0.118. We reproduce that on the hls4ml jet dataset as a sanity check that our multi-class variant does not lose detection power or gain sculpting there.

A Trident-style ensemble of per-type one-class autoencoders, as in Trident, tests whether a shared encoder is needed at all.

Two ablations isolate the key ingredients. The first drops the energy adversary and reweighting, leaving an unconstrained multi-class classifier with Mahalanobis rejection; it isolates the decorrelation. The second collapses the types to binary while keeping decorrelation; it isolates the typing.

Why not just…?

Why not just keep the hand-built cuts, which physicists understand and trust? Their interpretability is real, and we keep them as the reference. However, the chain is sequential, and whatever cleaning misses reaches the physics cuts, which were not designed for glitches. The cleaning paper motivates its own work by noting that the manual cuts had to be re-tuned when run conditions shifted. The experiment settles this: if our method does not beat the chain in leakage at matched sacrifice, the extra complexity is not justified.

Why not just train an ordinary multi-class network and use its softmax or Mahalanobis score for unknowns? This is the most serious threat, because it is our method minus one ingredient. It will probably classify about as well. Our bet is that it fails on energy transfer: without the constraint, it will learn energy-correlated cues, since low-energy pulses are noisier. Its acceptance for signal-like events will then differ between calibration and search energies, and it will distort the background spectrum. The first ablation tests exactly this. If the unconstrained network turns out to be flat in energy anyway, we will report that the decorrelation is unnecessary for germanium pulses. The contribution would then shrink to the unified typing and the evaluation protocol.

Why not just fix affinity propagation's memory problem with a scalable clustering method and keep the support vector machine? That would address the scaling complaint. It would still rely on hand-picked wavelet features, still need a human to label every new cluster before the classifier can use it, and still have no mechanism tying acceptance to energy. We include a mini-batch clustering variant of the existing pipeline as an extra baseline, so that any gain cannot be attributed merely to more training data.

Why not just use Trident as is? Separate one-class models avoid forgetting, but each sees only its own type, so none learns the boundary between signal-like and multi-site pulses. That boundary is the hardest one here. Trident's baseline row will show whether a shared discriminative encoder buys accuracy on physical types.

Proposed Experiments

All germanium experiments use the Full Chain Test stand data (high-rate detector-model data and low-rate dummy-board data) and the V06643A thorium calibration data, with time-ordered splits so that test data come from later runs. The jet experiment uses the hls4ml jet dataset.

The first decisive check is leakage versus sacrifice. We trace full curves on held-out labeled data and report the high-rate operating point matched to the 0.024% sacrifice of the existing cleaner, with bootstrap uncertainties.

The second is energy transfer. For each type we measure acceptance separately in each calibration peak and in energy sidebands. We report the largest acceptance difference across energies, together with the Jensen–Shannon divergence between the background spectrum before and after selection, the same sculpting measure NuRD uses. Our method should match the unconstrained ablation's classification quality while cutting both energy measures substantially. If it cannot, the core claim fails.

The third is discovery. We hold out one entire type from training, repeating the test with pileup, surface alpha and one glitch class in turn. We then measure how cleanly the rejected buffer clusters it, by adjusted mutual information against the true labels, compared with DBSCAN and with the scalable-clustering baseline. Trident's network-traffic figures give context but are not a like-for-like target.

The fourth is sim-to-real robustness. We retrain the physical heads using emulator outputs from different training seeds of the electronics emulator and report the spread in per-type efficiency. This spread is the systematic uncertainty an analysis would inherit.

The last is the cross-domain sanity check on jets. Adding typed heads to NuRD should keep AUROC near 0.914 and sculpting divergence near 0.118. A clear drop would signal that multi-class boundaries and decorrelation genuinely conflict, as the tension highlighted between the two literatures would predict.