A Discriminability-Uniform Space for Germanium Pulse Shapes

Germanium detectors hunting for neutrinoless double-beta decay decide event by event whether a recorded current pulse came from one compact charge cloud, which looks like signal, or from several separated deposits, which looks like gamma background. Today that decision rests on a handful of hand-tuned scalar cuts, recalibrated detector by detector. They break down exactly where the detector physics changes. In low-impurity detectors, weak-field corners stretch single-site charge clouds so that they look multi-site. The standard cut also carries a 3.5% efficiency uncertainty, because calibration events sit near the detector edges while real decays fill the bulk. Color science faced the same structural problem decades ago. Distances between raw RGB values mean different things in different parts of the color cube. The fix was to fit a new coordinate system to measured just-noticeable differences, so that one unit of distance means one discriminable step everywhere.

We propose to build the analogous space for detector pulses. We will learn an embedding of digitized waveforms in which Euclidean distance counts "just-discriminable differences" in charge-cloud topology. The unit of discriminability is set by the noise-limited precision with which topology can be recovered at each interaction location, and distance along nuisance directions such as drift time or depth is suppressed. The claim is that a single threshold in this space transfers across detector regions and across detectors with different impurity profiles. In the regions where physics makes pulses hard to tell apart, the space compresses honestly, so the classifier can mark events as "undecided" instead of misclassifying them.

Core Claim

Current pulse-shape discrimination has a fragmentation problem that resembles the one logicians describe for non-monotonic formalisms. Separate empirical dialects coexist: the amplitude-to-energy ratio A/E, the delayed-charge-recovery tail slope, and per-detector energy corrections. Each is calibrated independently and combined by sequential hard thresholds. In logic, the remedy has been a shared semantic space in which the dialects become special cases. Nonmonotonic Databases and Epistemic Queries argues for this, and approximation fixpoint theory reconstructs many semantics as fixpoints over one lattice of intervals, as shown for fuzzy programs in Approximation Fixpoint Theory as a Unifying Framework for Fuzzy Logic Programming Semantics.

We take two things from that line of work. The first is the aim of one shared space in which cuts from different detectors are directly comparable. The second is its central device: a decision is an interval ordered by precision, not a single point. We deliberately do not build a logic-programming system for pulses. A full fixpoint machinery over three or four cut variables would add formality without adding discriminating power.

The mechanism comes from perceptual color mapping. Color matching functions for a perceptually uniform RGB space treats the design of a uniform color space as an optimization over transformations, scored against psychophysical discrimination-threshold data. We replace human thresholds with physical ones. Let $x$ be a 256-sample waveform, $s$ a scalar topology variable describing how far the charge is spread along the drift direction, and $r$ the interaction location. Let $\sigma_s(r)$ be the smallest uncertainty in $s$ achievable from a pulse at location $r$ under the measured electronic noise. This is the Cramér–Rao bound, computed from simulated pulse derivatives. We then train an embedding $f$ on simulated pairs of pulses $(x_i, x_j)$ so that

$$\mathcal{L}(f)=\sum_{(i,j)}\Big(\|f(x_i)-f(x_j)\| - \big|s_i-s_j\big|\,/\,\bar\sigma_s(r_i,r_j)\Big)^2 ,$$

where $\bar\sigma_s(r_i,r_j)$ is the average bound at the two locations. Pairs that differ only in nuisance variables such as depth, drift path, or preamplifier phase get a target distance of zero. In the resulting space, one unit is one noise-limited discriminable step of topology, wherever the event occurred. Weak-field corners have large $\sigma_s$, so both signal-like and background-like pulses from those corners land closer together, near the decision boundary.

The decision itself is three-valued, borrowing the interval idea. An event is accepted if it lies more than one unit inside the single-site region, rejected if it lies more than one unit outside, and otherwise undecided with its interval reported. On the Relation Between Approximation Fixpoint Theory and Justification Theory shows that such interval-valued conclusions can carry explicit justifications. Our analogue is modest: for each event we report which topology feature and which location-dependent $\sigma_s$ placed it where it is.

Relation to Prior Work

Two papers set the physics baselines and motivate the proposal. Multisite event discrimination for the majorana demonstrator defines the energy-corrected A/E cut and identifies calibration-versus-bulk spatial mismatch as the source of its 3.5% systematic. α-event characterization and rejection in point-contact HPGe detectors shows that the delayed-charge-recovery tail slope rejects surface alphas almost perfectly. We therefore leave alpha rejection to that cut and target only the single-site versus multi-site decision.

Two further papers constrain the design:

Metric work from color and imaging supplies alternatives we test, not methods we adopt. Multiscale Sliced Wasserstein Distances as Perceptual Color Difference Measures builds a multiscale transport distance for images. We use a one-dimensional version on current pulses as a hand-designed competitor. Fuzzy color model and clustering algorithm for color clustering problem and Layered Logic Classifiers: Exploring the 'And' and 'Or' Relations combine graded memberships or logical connectives over fixed features. Their operators are heuristic, and they do not fix the underlying geometry, so they cannot make one threshold mean the same thing everywhere.

Research Program

The first component is a simulated pulse library with ground truth. We will use a standard detector field-and-drift simulation for point-contact germanium geometries at two impurity profiles that mimic PONaMa-I and PONaMa-II. The library will contain single-site and multi-site deposits sampled uniformly through the volume and superposed on noise traces recorded from real baseline data. For each location we compute $\sigma_s(r)$ numerically. This yields a discriminability map, the pulse-space counterpart of an ellipse chart of color-discrimination thresholds, and it is an interpretable scientific product in its own right.

The second component is the embedding. It is a small one-dimensional convolutional encoder of roughly the size of Holl et al.'s, trained with the loss above and then fine-tuned only through the decision threshold on real calibration data.

The third component is the three-valued decision layer and its per-event justification record.

The analysis questions are:

  • whether uniformity measured in simulation survives on real data, checked through agreement of distances between real calibration peaks;
  • whether one threshold transfers across detectors;
  • whether the undecided set concentrates in the weak-field corners rather than spreading at random.

Baselines

The naive reference point is a nearest-neighbor classifier on raw, baseline-subtracted waveforms using plain Euclidean distance. It shows how much any learned or physics-based geometry adds. We run it on the same pulses and with the same threshold procedure as everything else.

The strongest physics baseline is the energy-corrected A/E cut of Multisite event discrimination for the majorana demonstrator. It reports $(90\pm3.5)\%$ single-site retention, reduces the Tl-208 single-escape peak to about 6% of its original count, and suppresses the region-of-interest background by a factor of 3. We re-derive its energy correction on exactly our calibration data. To beat it, we must reject a larger fraction of single-escape-peak events at 90% double-escape-peak acceptance, and shrink the calibration-to-bulk efficiency gap below 3.5% on the simulated uniform-bulk test.

In low-impurity detectors, the relevant reference is the multi-site survival that Characterization of high purity germanium point contact detectors with low net impurity concentration reports for its standard cut: $35.2\pm0.2\%$ in PONaMa-II against $9.6\pm0.9\%$ in PONaMa-I. Beating it means closing most of that gap at matched signal acceptance, counting accepted and undecided separately so that abstention cannot masquerade as rejection.

The strongest learned baseline is the autoencoder plus 141-parameter classifier of Deep learning based pulse shape discrimination for germanium detectors. It reports 90% rejection of full-energy-peak events at 90% acceptance of double-escape-peak events. We retrain it on our pulses with the same encoder size, training set, and hyperparameter search budget as our model.

Two ablations isolate the key ingredient:

  • Location-blind scaling. We keep the full training pipeline but replace $\sigma_s(r)$ with a constant.
  • No nuisance pairs. We drop the zero-target nuisance pairs and train with class labels only, using a standard supervised contrastive loss.

Why not just…?

Why not just whiten the autoencoder latents with a Mahalanobis metric? Whitening equalizes noise variance, but it does so globally. The discriminability map is location-dependent, and Mertens et al. show that the weak-field corners are exactly where a global scale is wrong. This is nevertheless a real threat, since much of the benefit could come from simple noise normalization. We include whitened Holl latents as a baseline. If they match our model on transfer and on low-impurity multi-site survival, the location-dependent unit is not earning its place, and we will say so.

Why not just add drift time or position estimates as extra features to A/E? A position-dependent correction to A/E is the obvious physicist's fix. It could plausibly recover much of the low-impurity loss. It stays one scalar, though, so it cannot express the secondary-peak multi-site pulses that Holl et al. found fooling A/E, and it gives no principled "undecided" region. We implement it as an additional baseline, fitting a two-dimensional A/E-versus-drift-time correction on the same data. The comparison on low-impurity multi-site survival settles whether geometry matters beyond position correction.

Why not just use conformal prediction on the Holl classifier to get abstentions? Conformal prediction gives calibrated abstention under exchangeability. Calibration events and bulk decays are not exchangeable, since the 3.5% systematic arises precisely from their differing spatial distributions. We therefore expect conformal sets calibrated on the peaks to be miscalibrated in the bulk. We test this directly by measuring coverage on simulated uniform-bulk events for both methods.

Why not just build a fuzzy or logic-programming combination of the existing cuts? Combining A/E and delayed charge recovery through graded rules changes how evidence is aggregated. It does not change what the evidence measures. If the input features already confuse diffused single-site pulses with multi-site pulses, no connective separates them. Our ablation with location-blind scaling plays the role of "better aggregation without better geometry."

Proposed Experiments

The first experiment is a uniformity check, the analogue of scoring a color space against threshold data. On held-out simulated pairs we compute the ratio of embedding distance to target discriminability. We report its coefficient of variation, binned by detector region and impurity profile, for our space, the Holl latent space, whitened latents, and raw waveforms.

The second experiment is transfer of the calibration-to-bulk systematic. We set the threshold at 90% acceptance on edge-weighted simulated double-escape-peak events and measure acceptance on uniformly distributed single-site events. The gap is compared with the 3.5% figure.

The third experiment is real calibration data. The MAJORANA Th-228 double-escape, single-escape, and full-energy peaks provide the three peak populations, and Holl et al.'s 256-sample evaluation split provides the waveforms. Access depends on collaboration release. On these data we report single-escape and full-energy-peak rejection at 90% double-escape-peak acceptance.

The fourth experiment is low-impurity robustness, using PONaMa-II-like simulations and, where available, the PONaMa scan data. We report multi-site survival against the 35.2% reference, with undecided events counted separately.

The decisive checks are two. First, the undecided set should lie disproportionately within the weak-field corners of the discriminability map. If it is spatially random, the three-valued layer is cosmetic. Second, location-dependent scaling should beat both the whitened-latent baseline and the drift-corrected A/E baseline on the second and fourth experiments. If neither holds, the pulse-space analogy with perceptual uniformity does not transfer, and the discriminability map would remain the main contribution.