Decay Clocks as Labels: Federated Training of Germanium Background Classifiers on Real Cosmogenic Events

Germanium detectors that search for extremely rare nuclear decays have one hard internal background. While the crystals are made and shipped at the Earth's surface, cosmic rays turn some of their atoms into radioactive isotopes. Once the crystal is underground, those isotopes decay and produce events that look like the signal. Experiments reject these events by pulse shape discrimination: a classifier reads the recorded electrical pulse and decides whether the energy was deposited at one point (signal-like, "single-site") or at several points (background-like, "multi-site"). Every such classifier today, from the classic $A/E$ cut to neural networks, is trained on stand-ins. The usual stand-in is thorium calibration-source peaks, which approximate single-site and multi-site events. None is trained on the cosmogenic events it must actually reject. This proposal uses a fact the physics already supplies. Each detector's cosmogenic background fades on a known exponential clock, which starts when that crystal went underground. The clock start and the approximate initial dose are recorded in material-tracking databases. So the fraction of cosmogenic events in a given energy window is known as a function of time, up to a few fitted constants. That is enough to train a classifier from label proportions on real background data.

The second ingredient, federated learning, makes the method work across detectors whose owners cannot pool raw waveforms. In federated learning, each site trains a model locally and shares only parameter updates with a coordinating server, never the raw data. The link here is not decorative. Label-proportion learning from one detector's decay curve is fragile, because anything that drifts with time looks like decay. Different detectors, though, started their clocks at different dates and received different exposure doses. The spread of exposure histories across sites is what separates a true decay signature from calendar-time drift. The core claim is that a shared pulse-shape encoder, trained federatedly with per-detector adaptation layers and a decay-clock proportion loss, will reject real cosmogenic multi-site backgrounds better than calibration-trained classifiers. It is also expected to lose little relative to an idealized model trained on pooled data.

What Exists and What Is Missing

On the classifier side, Deep learning based pulse shape discrimination for germanium detectors has three parts. It compresses 256-sample waveforms into a 7-dimensional code with a convolutional autoencoder (a network trained to reconstruct its input through a narrow bottleneck). It attaches a 141-parameter classifier to that code. It reaches 90% multi-site rejection at 90% single-site acceptance on thorium calibration peaks, matching $A/E$ while catching some pulses that fool $A/E$. We reuse its architecture as the shared encoder. It is trained on one detector with calibration labels only, and we change both of those choices. A feasibility study of multi-electrode high-purity germanium detector compares LightGBM (gradient-boosted trees) with a fully connected network on 180,000 simulated waveforms. LightGBM reaches AUC 0.69 against 0.61 for the network. That result constrains us: small tree models can beat neural nets on this task, so any neural federated method must be compared against a strong tree baseline. Machine learning-powered data cleaning for LEGEND builds a semi-supervised noise filter on a single test stand. Its authors note that detector electronics and noise differ from stand to stand. That observation motivates our per-detector adaptation layers.

On the activation side, The Majorana Parts Tracking Database records each part's surface exposure as "sea-level equivalent" hours. It found 17.6 sea-level-equivalent days accumulated in one 5-day shipment. It treats that record as bookkeeping and never feeds it into data analysis. We use it as the prior on each detector's initial cosmogenic dose and clock start. Cosmogenic production of ³⁷Ar in the context of the LUX-ZEPLIN experiment shows how a cosmogenic background decays away underground, becoming comparable to other backgrounds after about 150 days. It also reports that production-rate models carry a factor-of-2-to-3 uncertainty. That constrains the design: the exposure record can serve only as a loose prior, and the initial activity must be fitted, not trusted.

On the federated side, FedHealth shares lower feature layers across clients and adapts upper layers locally with correlation alignment, which matches each client's feature covariance to the shared one. It reports a 5.3% absolute accuracy gain over a non-personalized baseline on phone sensor data. We borrow the split between shared and local layers. LEAF supplies the evaluation habit we adopt: report the spread of per-client performance (10th, 50th and 90th percentiles) and communication cost, not just the average.

The research file lists no paper that applies federated learning to cosmogenic backgrounds. More importantly for novelty, it lists no work that trains a pulse-shape classifier on real in-situ cosmogenic events through their decay curve, whether centrally or federatedly. The decay-clock supervision is the new part. Federation is how it is deployed, and exposure heterogeneity across sites is what makes it identifiable.

Core Mechanism

For detector $k$, bin its events in a background-dominated energy window (chosen around the double-beta decay energy, using calibration-independent energy selection) into time bins $t$ measured from the date the crystal went underground. The expected cosmogenic fraction in that bin is

$$\pi_k(t) = \frac{A_k e^{-\lambda t}}{A_k e^{-\lambda t} + B_k},$$

where $\lambda$ is the known decay constant of the relevant cosmogenic isotope, $A_k$ is the detector's initial cosmogenic rate in the window, and $B_k$ is the rate of steady (non-decaying) events. $A_k$ gets a prior from the tracking database, widened by a factor of three to reflect the production-model uncertainty; $B_k$ is fitted freely. The classifier $f_\theta$ sees only the waveform, never the time stamp. It outputs a probability that the event is cosmogenic multi-site. It is trained so that its average output in each bin matches the predicted fraction, alongside the ordinary calibration loss $\mathcal{L}_{\text{cal}}$:

$$\mathcal{L}_k = \sum_t n_{k,t}\big(\bar f_{\theta}(k,t) - \pi_k(t)\big)^2 + \alpha\,\mathcal{L}_{\text{cal},k},$$

where $n_{k,t}$ is the number of events in the bin, $\bar f_\theta(k,t)$ is the mean classifier output over them, and $\alpha$ sets the calibration term's weight. The network is split in two. The encoder parameters are averaged across sites with federated averaging (FedAvg). Each detector keeps local parameters: a small input layer that absorbs its electronics response, and its own $A_k, B_k$. The decay clocks of different detectors start on different calendar dates. A shared encoder therefore cannot explain all sites' proportion curves by tracking calendar-time effects such as gain drift or seasonal noise. It can explain them only by learning the waveform features of the decaying population.

Research Program

The work proceeds in three stages. The first builds a federated simulation benchmark, since the file identifies no public real-waveform dataset. Start from the simulated waveforms of the multi-electrode study (A feasibility study of multi-electrode high-purity germanium detector) and from simulated point-contact and coaxial geometries. Create 8 to 20 synthetic detector clients. Each gets its own electronics transfer function, noise spectrum and slow gain drift. Each also gets an exposure history drawn from realistic tracking-database profiles, with initial doses perturbed by the factor-2-to-3 model uncertainty. Background events are drawn as a decaying cosmogenic multi-site population plus steady populations, with statistics varied from generous down to realistic low-background counts. Simulated ground truth lets us score rejection of the actual cosmogenic class.

The second stage studies identifiability and the method's failure modes on this benchmark. We measure how much exposure (detector-days) and how much spread in clock start dates are needed before the proportion loss improves on calibration-only training. We also measure how strongly correlated drift, where gain drift is aligned with decay time, can fool it.

The third stage is a real-data pilot inside one collaboration's computing environment. There the pooled model is also allowed, so the cost of federation can be measured directly rather than argued. A cross-collaboration federated deployment follows only if the pilot shows a gain.

Baselines

The naive reference point is the $A/E$ cut, tuned per detector on thorium calibration data as in standard practice. Deep learning based pulse shape discrimination for germanium detectors reports that it achieves 90% multi-site rejection at 90% single-site acceptance on calibration peaks. We evaluate it on the benchmark's true cosmogenic class at the same 90% signal acceptance.

The strongest existing method is the autoencoder plus 141-parameter classifier of the same paper, trained per detector on that detector's calibration data. On calibration peaks it reports the same 90%/90% operating point. The file reports no number for it on true cosmogenic events, and producing that number is part of our evaluation. Because trees beat neural nets on simulated multi-electrode data (LightGBM AUC 0.69 against 0.61 in the feasibility study), we also run per-detector LightGBM on the same features.

The upper reference is centralized pooled training with the same encoder, per-detector input layers and proportion loss. It is the same model with federation removed. Every method gets the same waveforms, the same hyperparameter search budget (a fixed number of trials on a held-out calibration split) and comparable GPU-hours. Federated runs are additionally reported with bytes communicated.

Two ablations isolate the key ingredients. The first removes the proportion loss, giving a federated model trained on calibration only (FedHealth-style). This isolates what the decay-clock supervision adds. The second trains the proportion loss on a single detector without federation. This isolates whether cross-site clock diversity is what defeats drift confounding. Beating the baselines means higher true-cosmogenic rejection at 90% single-site acceptance than per-detector calibration training, with a gap to pooled training of at most a few percentage points. This must hold at the 10th percentile of detectors, not only on average.

Why not just…?

Why not just pool the waveforms and train centrally? Inside a single collaboration this is usually possible, which makes it a real threat to the federation part of the claim. It is not a threat to the proportion-loss part. We therefore report pooled training as the ceiling. If the federated model comes within a few points of it, the conclusion is that federation lets decay-clock supervision cross collaboration boundaries at little cost. If the gap is large, the honest recommendation is to pool wherever governance allows. The file offers only general arguments about data sovereignty and blinding; it gives no evidence that pooling is actually forbidden in these experiments.

Why not just train each detector on its own thorium calibration data? Calibration peaks are stand-ins, and the concern is that cosmogenic multi-site events differ from them in how they are distributed inside the crystal and in their energy. No paper in the file measures this mismatch. The comparison against the second ablation and the per-detector baselines on the benchmark's true cosmogenic class tests it directly. If calibration-trained models already reject true cosmogenic events as well as the proportion-trained ones, the central premise fails and we would say so.

Why not just use the tracking-database exposure to subtract the predicted cosmogenic rate statistically, with no new classifier? Subtraction corrects the expected count but rejects no individual events, so it does not improve the signal-to-background ratio. It also inherits the factor-2-to-3 production uncertainty reported in the ³⁷Ar study. Our method fits $A_k$ from the time shape of the data and uses the database only as a prior. We will compare how well the two approaches recover $A_k$ on the benchmark.

Why not just pretrain on simulation and fine-tune locally? This is cheap and could capture most of the benefit of a shared encoder. We include it as an extra baseline under the same tuning budget. It still relies on calibration labels at each site and does nothing about the proxy mismatch.

Decisive Checks

On the simulation benchmark, the decisive result is higher true-cosmogenic rejection at 90% single-site acceptance than the per-detector calibration-trained network and LightGBM, at realistic event statistics. A second requirement is that the single-detector proportion ablation degrades clearly when gain drift is aligned with decay time, while the federated version does not. This would show that clock diversity across sites is the working ingredient.

On real data, there is no ground truth, so the check is falsifiable in a different way. The fraction of window events the classifier rejects should fall over time at the known decay rate. It should also rank detectors by their recorded exposure, even though time was never an input. It should do so without lowering acceptance on thorium single-site proxy events.

The main risks are too few background events near the decay energy, and drift that happens to track exposure. If the benchmark shows that the required exposure exceeds what real detectors accumulate before their cosmogenic background has decayed away, the idea fails. Stage two is designed to show that early.