Rare-event searches, such as the hunt for neutrinoless double-beta decay in germanium crystals, run for five to ten years. They separate a handful of signal-like events from background by looking at the shape of each electrical pulse. This step is called pulse shape discrimination: a rule, or a small neural network, that reads a digitized waveform and decides whether the energy was deposited at one point (signal-like) or at several (gamma background). Every such rule assumes that the detector's response stays fixed. Real hardware does not stay fixed. Preamplifier decay times shift, charge trapping in the crystal grows, and surface layers change. A discriminator tuned on one calibration week can slowly start rejecting real signal or admitting background. The usual fix is to insert a radioactive calibration source. That fix is costly in exactly this field, because every insertion adds dead time and contamination risk to an experiment whose target is fewer than one background count per tonne-year (The MAJORANA Project).
We propose to track the detector's response without labels and without sources. The tracking uses two streams the detector already produces. The first is injected electronic test pulses and signal-free baseline traces, which measure the readout chain. The second is an anchor population: a class of physics events whose true signal/background makeup is known in advance and does not change over time. In enriched germanium, the ordinary two-neutrino double-beta decay continuum is such a population. Its rate is constant on any human timescale, and it is intrinsically single-site, so any drift in how those pulses look must come from the instrument, not from the physics. A small set encoder turns each time window's pulser traces and anchor events into a "drift state." The pulse-shape classifier is then conditioned on that state.
Core Claim
Conditioning the classifier on a drift state estimated from fixed-composition anchor populations can hold signal acceptance and background rejection steady across long deployments. We claim it does better than three cheaper options: periodic recalibration of scalar cuts, retraining the classifier to be blind to drift, and aligning the overall data distribution. The last option is the subtle one. It also adapts without labels, but it cannot tell instrument drift apart from a real change in background (for example, more radon arriving), so it silently absorbs the very physics the experiment exists to measure. Anchoring the drift estimate to populations of known composition is what separates "the detector changed" from "the world changed." That separation is the new ingredient.
Prior Work and What It Leaves Open
Three germanium-specific lines of work supply the methods we reuse, and all three assume the detector does not change.
Deep learning based pulse shape discrimination for germanium detectors compresses 256-sample waveforms into a 7-dimensional code with an autoencoder, then classifies them with a 141-parameter network. It matches the standard current-to-energy ratio cut (called $A/E$) at 90% background rejection with 90% signal-proxy acceptance, but it was trained and tested on a single static calibration run. We reuse its encoder design and add drift conditioning.
Ad-hoc Pulse Shape Simulation using Cyclic Positional U-Net learns to translate idealized simulated pulses into realistic ones. It improves tail-slope histogram overlap with real data from 0% to 83.2%, but it maps to one frozen snapshot of the hardware. We use it as a starting point for building a testbed with controlled drift.
Machine learning-powered data cleaning for LEGEND clusters waveforms with affinity propagation and then trains an SVM on human-labeled clusters. It reports 0.024% physics sacrifice and 15.034% anomalous leakage on high-rate test-stand data. It is limited to roughly 10,000 events by quadratic memory, and it needs a human relabeling pass whenever conditions change. Our method is meant to run continuously without that relabeling step.
The motivating evidence that drift matters, and that temporal context helps, comes from outside physics. Using context to make gas classifiers robust to sensor drift tracks 36 months of metal-oxide gas-sensor aging with a recurrent "context" pathway. It reaches 88.2% mean accuracy, against 86.9% for a static network and 76.4% for an SVM ensemble, and it shows that simply making the static network bigger does not help. Its context stream, however, consists of labeled past samples, which a rare-event experiment never has. Follow-ups in the same line, Using Context to Adapt to Sensor Drift and the blind-calibration methods Blind Drift Calibration of Sensor Networks using Sparse Bayesian Learning and A Deep Learning Approach for Blind Drift Calibration of Sensor Networks, recover gain and offset drift without references. They do not touch event-by-event classification and do not separate response drift from changes in what is being measured.
The materials literature constrains how we model drift. It shows that degradation is continuous and that it shows up under operation, not in storage. Hai‐Ting Lu et al. separate shelf stability from operational stability. Bin Liu et al. document progressive surface oxidation. This motivates a drift state that changes smoothly in time rather than in discrete jumps.
Finally, α-event characterization and rejection in point-contact HPGe detectors rejects surface alphas through slow charge recovery in the pulse tail. That creates a concrete hazard for a drift-blind discriminator: growing surface dead layers or charge trapping also slow the tail of genuine bulk events.
Research Program
The method has three parts.
A forward model of drift. We start from real calibration waveforms and apply physically parameterized perturbations $\theta$. These cover preamplifier decay-constant shifts, exponential charge-trapping loss, noise-spectrum changes, and a slow-collection surface layer of variable thickness. We optionally refine the result with a translator in the style of the Cyclic Positional U-Net, conditioned on $\theta$. This yields training data where the drift is known exactly.
A drift encoder. A permutation-invariant set encoder $g$ takes all pulser traces, baseline traces, and anchor-window events $S_t$ from time window $t$ and outputs a drift estimate $\hat z_t = g(S_t)$. It is trained to recover $\theta$.
A drift-conditioned classifier. The classifier $f(x, \hat z_t)$, built on a Holl-style autoencoder code, is trained on drifted calibration proxies. These are double-escape-peak events as the single-site proxy and full-energy-peak events as the multi-site proxy. The training objective is
$$\min_{f,g}\ \mathbb{E}_{\theta}\,\mathbb{E}_{(x,y)\sim p_\theta}\big[\ell(f(x, g(S_\theta)), y)\big] + \lambda\,\mathbb{E}_\theta\,\|g(S_\theta)-\theta\|^2 ,$$
Here $p_\theta$ is the distribution of proxy waveforms $x$ with labels $y$ under drift $\theta$, $S_\theta$ is a window of unlabeled pulser and anchor events under the same drift, $\ell$ is cross-entropy, and $\lambda$ weights the auxiliary drift-recovery term.
At deployment, only unlabeled windows are needed. Because the drift estimate pools $n$ anchor events, its error should shrink roughly as $1/\sqrt{n}$, as for any average of independent samples. This sets the trade-off between window length and how fast the system can follow drift, and we will measure that trade-off directly.
A side product addresses the surface-alpha hazard. Once drift is modeled explicitly, the tail-slowing caused by dead-layer growth enters through $\hat z_t$ rather than being read as alpha-like morphology. This lets us test whether slow-charge-recovery rejection can be kept from drifting into rejecting bulk signal.
Baselines
Frozen discriminator. The naive reference point is a discriminator trained once and never touched. We use both a fixed $A/E$ threshold and the Holl autoencoder classifier frozen at time zero. Its reported operating point is 90% full-energy-peak rejection at 90% double-escape-peak acceptance (Deep learning based pulse shape discrimination for germanium detectors). Holl et al. never evaluated it under drift.
Periodically recalibrated $A/E$. This is the strongest existing practice and the honest competitor. The $A/E$ cut is renormalized at every calibration run, and the readout is corrected using the decay constant measured from test pulses. We found no reported result on Lacuna for how stable recalibrated $A/E$ acceptance is between calibrations. We will therefore measure it ourselves, at several calibration intervals, on identical data. Beating it means one of two things. Either we hold signal-proxy acceptance within a fixed tolerance band at matched rejection over a longer interval than recalibrated $A/E$ achieves at the same interval. Or we match its stability while using no calibration source between anchor points.
Drift-randomized classifier. This classifier is trained on the same drift forward model but without conditioning, so it learns to ignore drift. It receives the same simulated data, architecture, and hyperparameter search budget as our method.
Marginal-alignment adaptation. This baseline adapts without labels by re-estimating normalization statistics, or by aligning feature distributions, on each unlabeled window. It is the most dangerous competitor, because it also adapts without labels.
Ablations. These isolate the key ingredient. We compare drift states estimated from pulser and baseline traces only, from anchor events only, and from both. We also include an oracle that is given the true $\theta$.
For the data-cleaning task on test-stand data, the reference is the affinity-propagation-plus-SVM pipeline at its reported 0.024% sacrifice and 15.034% high-rate leakage (Machine learning-powered data cleaning for LEGEND). We run it on the same windows under its 10,000-event limit.
Why not just…?
Why not just recalibrate $A/E$ more often? This is the real threat. If drift is slow relative to a monthly calibration, recalibrated $A/E$ may already be good enough, and our contribution then shrinks to "fewer source insertions." We accept that outcome as a valid result. Source insertion carries the dead-time and contamination costs that tonne-scale backgrounds cannot absorb, as documented by The MAJORANA Project and by radiopurity concerns in cryogenic shields (Measurement of the 85Kr specific activity in the GERDA liquid argon). Recalibrated $A/E$ also inherits the scalar cut's known blind spots: Holl et al. show it misses multi-site pulses that have a sharp current peak. The calibration-interval sweep settles this question directly.
Why not just train a classifier that is invariant to drift? Invariance discards information whenever drift and class signal share a direction in waveform space. Charge trapping and surface layers both slow pulse tails, which is exactly the feature used to reject alphas. An invariant model must give up some of that feature at every drift level. Comparing the drift-randomized baseline against the oracle-conditioned model measures how much is lost. If that gap is negligible, conditioning is unnecessary and we will say so.
Why not just align the unlabeled data distribution window by window? This would plausibly match our method if background composition never changed. But that never-changing composition is precisely what an experiment cannot assume. Radon ingress or a change in potassium-42 activity (Electromagnetic backgrounds and potassium-42 activity in the DEAP-3600 dark matter detector) shifts the overall data distribution, and alignment would "correct" that shift away. The decisive experiment below tests exactly this case.
Proposed Experiments
Controlled drift testbed. We build it from $^{228}$Th waveform sets of the kind used by Holl et al. and Li et al. Drift is injected along smooth trajectories over simulated multi-year timelines. We report signal-proxy acceptance at fixed background rejection for every window, together with its spread over time, plus error in recovering $\theta$.
Decisive check. We inject response drift and a genuine increase in the multi-site background fraction at the same time. Success means our method keeps signal acceptance flat and reports the increased background rate correctly. Marginal alignment is expected to under-report the background increase. If our method also absorbs the change, the anchoring idea fails.
Window-size sweep. We vary the number of anchor events per window to check the expected $1/\sqrt{n}$ behavior of drift-estimate error. This tells us what event rates make the method practical.
Real long-duration data. We apply the method to multi-month LEGEND-style test-stand data like that used by León et al., comparing against the cleaning pipeline's sacrifice and leakage figures. Long-term archives of calibration runs from an enriched-detector array, where the two-neutrino continuum anchor actually exists, require collaboration access. We treat this as the necessary final validation, not an assumption.
Out-of-domain check on gas sensors. On the 36-month Vergara gas-sensor benchmark, we replace Warner et al.'s labeled context stream with unlabeled windows. We ask how much of their gain over 86.9% (to 88.2%) survives without labels. This tests the label-free half of the claim outside physics, though that benchmark has no fixed-composition anchor population.
References
- The MAJORANA Project
- Deep learning based pulse shape discrimination for germanium detectors
- Ad-hoc Pulse Shape Simulation using Cyclic Positional U-Net
- Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines
- Using context to make gas classifiers robust to sensor drift
- Using Context to Adapt to Sensor Drift
- Blind Drift Calibration of Sensor Networks using Sparse Bayesian Learning
- A Deep Learning Approach for Blind Drift Calibration of Sensor Networks
- Nonenzymatic hydrogen peroxide electrochemical sensor based on carbon-coated SnO2 supported Pt nanoparticles
- Surrounding Sensitive Electronic Properties of Bi2Te3 Nanoplates—Potential Sensing Applications of Topological Insulators
- $$\alpha $$-event characterization and rejection in point-contact HPGe detectors
- Measurement of the 85Kr specific activity in the GERDA liquid argon
- Electromagnetic backgrounds and potassium-42 activity in the DEAP-3600 dark matter detector