Reading the Relaxation Spectrum of Delayed Charge in Germanium Detectors

Germanium detectors searching for neutrinoless double-beta decay, a nuclear decay that would show neutrinos are their own antiparticles, are limited by a few background events per year. One of the most damaging comes from alpha particles that strike the detector's passivated surface. On that surface, some of the liberated charge is trapped and released slowly, so the event's energy is under-measured and can land in the signal window. The standard defense, Delayed Charge Recovery (DCR), measures one number: the slope of the waveform's trailing edge. That one slope implicitly treats charge release as a single relaxation process.

This note proposes to replace the slope with the full distribution of release rates hidden in the tail. The tail is modeled as a superposition of exponential decays, and the weight on each decay rate is recovered by regularized inversion. The approach borrows two things from the study of non-Fermi liquids, metals whose electrons do not behave as long-lived independent particles. The first is a physical expectation. Electrons coupled with random strengths to a bath of two-level systems (defects that flip between two configurations) relax with a broad, non-exponential spectrum, not a single rate. The second is a numerical tool. Recovering such a spectrum from a noisy decay is the same ill-posed inverse-Laplace problem that quantum Monte Carlo practitioners face when they extract real-frequency transport from imaginary-time data. The bet is practical: a few rate-spectrum features should separate surface alphas from bulk events in regimes where the single slope is weak, at lower bulk-signal loss.

Core Claim

The trailing edge of a germanium waveform after the main charge collection can be written as

$$f(t) = \int g(\Gamma)\, e^{-\Gamma t}\, d\Gamma ,$$

where $f(t)$ is the baseline-subtracted tail, $\Gamma$ is a release rate, and $g(\Gamma)$ is the spectrum of release rates. DCR effectively compresses $g$ into one number.

Our claim is that $g$ carries additional, discriminating structure for surface alphas, especially for two classes of event:

  • events near the point contact, where weighting-potential effects distort the tail;
  • low-amplitude, heavily degraded events, which are exactly those that leak into the region of interest.

A low-dimensional summary of $g$ should therefore beat DCR at equal bulk sacrifice. Example summaries are the log-rate centroid, the spectral width, and the weight at the slowest resolvable rates.

A secondary, falsifiable physics question is whether $g$ is narrow, as a single-trap picture would predict, or broad and roughly flat in $\log\Gamma$, as random-coupling two-level-system models predict.

Relation to Prior Work

The proposal is motivated by Arnquist et al.. That work scanned a collimated alpha source over a MAJORANA-type detector with the TUBE scanner and introduced DCR. DCR rejects all surface alphas more than 3 mm from the center within statistical uncertainty, loses 0.2% of bulk events, and cuts the region-of-interest background by an order of magnitude. The same paper reports long-term surface charge drift that needs empirical correction. That drift is a hint that the release process is not a fixed single rate. We keep their physical observable and change how it is summarized.

The physical expectation of broad relaxation comes from the solvable models of Tulipman et al. and Bashan et al.. Those papers couple itinerant electrons to two-level systems with random couplings and obtain non-quasiparticle, non-exponential dynamics in bulk metals. They compute bulk transport, not single-event charge release at a semiconductor surface. We use them only as a motivation for expecting a broad $g(\Gamma)$, not as a model we claim applies quantitatively to germanium oxide layers.

The method side borrows from the analytic-continuation problem described in the non-Fermi liquid literature. There, the imaginary-time Green's function is a Fredholm integral of a spectral function, inverted with entropy-regularized fits. The tail inversion above is the same mathematical problem, with the same noise sensitivity.

Quantum Loop Topography, extended in Driskell et al., supplies the design lesson. A shallow classifier with 10 hidden neurons, fed physically chosen correlation features, mapped the non-Fermi liquid fan from three training points. QLT does this on simulated Green's functions. We apply the same principle, "physics-shaped features plus a tiny model," to recorded detector waveforms, where it has not been tried.

Research Program

Inversion method. We will build a tail-inversion pipeline. Each waveform's post-rise segment is fit on a fixed log-spaced grid of rates with a nonnegativity constraint. Two regularizers will be compared: Tikhonov smoothing and a maximum-entropy prior. Regularization strength is set by cross-validation on bulk calibration events only, so the choice cannot be tuned to alphas.

Resolvable range. The rates we can resolve are bounded. The fastest is set by digitizer sampling and the slowest by the recorded window. Because the millisecond release in Arnquist et al. exceeds typical windows, we expect to see only the fast end of $g$. We will report the resolvable decade range explicitly rather than over-interpreting the slowest bins.

Classifier. From each recovered spectrum we take three to five summary features. These feed a classifier with one hidden layer of about 10 units, deliberately matched in size to the QLT setting.

Physics analysis. Using scanner position as ground truth, we will map how the spectrum's centroid and width vary with distance from the point contact and with alpha energy degradation. Where data at different bias voltages or temperatures exist, we will check whether the width changes the way thermally activated single traps predict.

Baselines

Energy only. The naive reference point applies no pulse-shape cut, only the energy window. It sets the raw alpha leakage that every method must reduce.

DCR. The strongest existing method is DCR as defined in Arnquist et al.. Its reported performance is 100% surface-alpha rejection beyond 3 mm from the center at 0.2% bulk sacrifice. We will recompute DCR on exactly the same waveforms and the same tail window, and fix its threshold to give 0.2% bulk sacrifice on held-out calibration data. All methods are compared at that sacrifice.

Beating DCR means one of two outcomes:

  • lower alpha leakage at matched 0.2% sacrifice in the near-contact positions (within 3 mm) and among low-energy degraded alphas;
  • equal leakage at a clearly lower bulk sacrifice.

The file contains no reported DCR leakage figure for the near-contact region (no reported result found on Lacuna for near-contact leakage), so we must measure it ourselves.

Stretched exponential. The simplest rival that captures non-exponential tails is a two-parameter stretched-exponential fit to the tail, fed into the same small classifier. It is described further below.

Wavelet features with a support-vector machine. The León et al. pipeline was built for data cleaning. It reports 0.024±0.004% physics sacrifice with 15.034±0.224% anomalous leakage on high-rate LEGEND test-stand data. We will retrain it for the alpha task as a generic-feature learned baseline.

One-dimensional convolutional network. A convolutional network on the raw tail segment is given the same tuning budget, a fixed number of hyperparameter trials, as every learned model here.

Ablation. The key ingredient is regularized inversion. To isolate it, we replace the spectrum features with an unregularized nonnegative fit on the same grid, which keeps the representation but drops the stabilization. We also replace them with randomly projected tail samples of equal dimension.

Why not just…?

Why not just keep DCR? Beyond 3 mm DCR already reaches 100% rejection within statistics, so there is little to gain there. The open space is near the contact and at low degraded energies, where a single slope conflates tail shape with amplitude. If the spectrum features do not improve leakage there at matched sacrifice, the proposal fails and DCR stands. That is a clean outcome either way.

Why not just fit a stretched exponential? This is the real threat. A stretched exponential is exactly what a broad rate spectrum looks like when summarized by two numbers, and it is cheaper and more stable than an inversion. If it matches the spectrum features, we will adopt it as the practical discriminator. The inversion would then be reduced to a diagnostic for interpreting drift. The spectrum earns its place only if alpha populations differ in spectral shape beyond centroid and width, for example a two-component structure near the contact. The head-to-head comparison above is designed to decide this.

Why not just train a deep network on the raw waveform? Zhang et al. found that convolutional networks needed far more data and were more noise-sensitive than physics-chosen features in their setting. Scanner alpha datasets are small, so we expect the same pattern here. This remains an empirical question. If the network wins at equal training size, we report that. We also measure each model's degradation as training data shrinks, since that is where we predict the small model's advantage.

Why not just call it classical distributed traps? A broad spread of classical trap depths can produce the same broad $g(\Gamma)$ as random coupling to two-level systems. Nothing in this note's discrimination claim depends on which picture is right. We keep the many-body interpretation as a secondary hypothesis. It is testable only through how the spectrum width depends on temperature or bias, and we will report it as unresolved if the data cannot separate the two pictures.

Decisive Experiments

Primary dataset. The primary dataset is a TUBE-style alpha scan of a point-contact detector, as in Arnquist et al., paired with thorium-228 calibration for bulk events. Access to that scan data or an equivalent new scan is the main dependency. Low-rate LEGEND test-stand data from León et al. serve as a check that spectrum features do not misfire on electronic transients.

Primary metric. Alpha leakage at 0.2% bulk sacrifice, binned by beam radius and by degraded energy.

Secondary metrics. Leakage across a sweep of sacrifice values, and stability of the spectrum features under the long-term surface drift that DCR must correct for.

Decisive check. In the near-contact bins, the spectrum classifier must beat both DCR and the stretched exponential at matched sacrifice, by more than the bootstrap uncertainty.

Physics outcome. Separately, we will report whether bulk and surface spectra are narrow or span a decade or more of resolvable rates. Either answer is informative.