Experiments searching for neutrinoless double-beta decay use germanium crystals and must reject background gamma rays that mimic the signal. They do this by looking at the shape of each recorded electrical pulse. A true decay deposits its energy at one point, while a background gamma usually scatters several times. The standard cut, and the neural networks proposed to replace it, are tuned on calibration events from an external radioactive source. Those events interact mostly near the crystal surface, whereas real decays happen uniformly throughout the bulk. The cut's signal efficiency therefore carries a systematic uncertainty of about 3.5% that comes from this position mismatch (Multisite event discrimination for the majorana demonstrator). For a search that counts a handful of events per tonne-year, that uncertainty matters as much as raw rejection power.
We propose to use contrastive regularization, a training loss that pulls chosen pairs of examples together in a network's internal representation and pushes others apart, for a specific purpose. The loss is not aimed at separating signal from background, which ordinary supervised training already does. It is aimed at making the representation ignore where in the crystal an event happened while keeping how many sites it had. The positive pairs come from simulation: the same energy-deposition pattern is placed at two different locations in the crystal. The claim is that a discriminator built on such a representation matches existing methods on calibration rejection. It should also have an acceptance that stays flat across the crystal volume, so the calibration-derived efficiency transfers to bulk signal events with a smaller systematic.
What exists and what is different
The closest prior work is Deep learning based pulse shape discrimination for germanium detectors. It compresses 256-sample germanium pulses into a 7-dimensional code with an autoencoder, a network trained to reconstruct its input through a narrow bottleneck, using plain reconstruction error. A 141-parameter classifier trained on calibration peaks then sits on top of that code. The system reaches 90% rejection of 2614 keV full-energy-peak events at 90% acceptance of 1592 keV double-escape-peak events, which is parity with the physics cut. It also catches some multi-site events that fool that cut. The bottleneck has no geometric constraint, and the evaluation never asks whether acceptance depends on interaction position.
The obvious extension would add a supervised contrastive loss that uses the calibration labels, pulling single-site peak events together and pushing multi-site ones apart. We think that extension is weak for two reasons. First, a skeptic would correctly note that an end-to-end supervised classifier optimizes the same separation directly. Second, the calibration "labels" are proxies. The double-escape peak sits on a Compton continuum of multi-site events, and all of the calibration data share the same surface-biased position distribution. A label-driven contrastive loss would sharpen exactly the bias we want to remove. This matches the representation-learning literature's own warning that supervised contrastive forces misbehave when labels come from coarse physical proxies.
Contrastive regularization has been brought near physics and near "signal versus noise" before, but for different jobs. Pixel Detector Background Generation using Generative Adversarial Networks at Belle II uses contrastive conditioning inside a generator's discriminator so that it synthesizes realistic, sensor-specific background hitmaps. It does not use contrastive learning for event selection. It also reports singular-value explosions that forced early stopping, which is a stability warning we take seriously. A Diffusion Model with Contrastive Learning for ICU False Arrhythmia Alarm Reduction contrasts true arrhythmias against artifacts in a diffusion model's latent space, but its contrast is label-driven and aims at separation rather than invariance to a nuisance variable. Our proposal differs from both in what defines a positive pair. The positive pair is a physical intervention that changes only position, and the loss is judged by stability of efficiency rather than by accuracy.
Two further results shape our expectations. Phantom Embeddings and MAPLE show that latent-space regularizers give real but modest accuracy gains on standard benchmarks: 93.50% to 94.91% on CIFAR-10 for the former. MAPLE is also sensitive to clustering hyperparameters and relabeling instability. This constrains the proposal. We should not expect contrastive structure to move headline rejection much, which is why the claim is about robustness. The pileup work in Model-Based Deep Learning Algorithm for Pulse Shape Discrimination in High Event Rates reports 1.7 times fewer misclassifications than a purely data-driven network by building physics into the model. That motivates encoding the physics, here a known nuisance variable, into training rather than hoping the network discovers it.
Core Claim
A representation trained so that relocated copies of the same deposition pattern map to nearby codes will support a single-site versus multi-site classifier whose acceptance varies less with interaction position than either the standard cut or an unconstrained network. It will do this at equal rejection of calibration multi-site events. The key ingredient is the pairing by position intervention. Neither contrast by class label nor simply training on more position-diverse data should reproduce the effect.
Research Program
We would build an encoder in the style of Holl et al., a one-dimensional convolutional network with a small bottleneck, trained on three signals at once. Reconstruction error on unlabeled real pulses keeps the code faithful to measured data rather than to simulation. A contrastive term on simulated pairs enforces position invariance. A small classifier head is trained on real calibration peaks, with sideband-subtracted weights so that continuum events under the double-escape peak are not counted as single-site. The training objective for the encoder $f$ is
$$ \mathcal{L} = \mathcal{L}_{\text{rec}}(x_{\text{real}}) + \lambda \, \mathcal{L}_{\text{NCE}}\big(f(s(D, \mathbf{r}_1)), f(s(D, \mathbf{r}_2))\big), $$
where the terms are defined as follows. $\mathcal{L}_{\text{rec}}$ is the reconstruction error on real unlabeled pulses $x_{\text{real}}$. $s(D, \mathbf{r})$ is the simulated pulse from a deposition pattern $D$ (a set of energy deposits with fixed relative geometry) placed at crystal position $\mathbf{r}$. $\mathcal{L}_{\text{NCE}}$ is the standard contrastive loss that treats the two placements of the same $D$ as a positive pair and every other pattern in the batch as a negative. $\lambda$ weights the two terms.
Negatives include single-site patterns at other positions. The loss therefore does not collapse all single-site events together; it removes only the position direction. Positions $\mathbf{r}_1, \mathbf{r}_2$ are sampled to span the surface-to-bulk range. Relocating the same pattern means the same particle track, so the loss does not need to trust any class label.
The sim-to-real gap is the main technical risk. Simulated pulses will differ from measured ones in electronics response and charge-trapping details. We address this by keeping the reconstruction and classification terms on real data only. We also report every result both on held-out simulation, where true positions are known, and on real data, where they are not. We will monitor the largest singular values of each layer throughout training, given the instabilities reported at Belle II.
Baselines
The naive reference point is the current-amplitude-over-energy cut from Multisite event discrimination for the majorana demonstrator. That paper reports $(90 \pm 3.5)\%$ single-site retention, with the 2103 keV single-escape peak suppressed to about 6% and a factor-of-3 background reduction in the region of interest. We will apply it to the same detector data with its published energy correction, tuned on the same double-escape peak. Beating it means an acceptance spread across position bins in simulation that is clearly smaller than the spread of the cut, at matched rejection.
The strongest existing learned method is the Holl et al. autoencoder plus classifier. It reports 90% full-energy-peak rejection at 90% double-escape-peak acceptance. We will reimplement it with the same bottleneck size, training data, and hyperparameter search budget as our model. Its position dependence has not been reported; no reported result found on Lacuna for position-resolved acceptance of any learned germanium discriminator. We will therefore measure it ourselves. Beating it means equal rejection at 90% acceptance, within statistical error, together with a smaller acceptance spread and closer agreement between calibration acceptance and the bulk-event proxy described below.
Because an end-to-end supervised network is the natural skeptic's choice, we include one. It is a convolutional classifier of the same size, trained directly on weighted calibration labels, and given the same tuning budget.
Four ablations isolate the key ingredient. The first uses our full architecture but trains on the simulated position-diverse pulses with class labels and no contrastive term. It tests whether the data diversity, not the pairing, does the work. The second replaces position-paired positives with label-defined positives, which is the obvious supervised-contrastive idea. The third is a domain-adversarial variant in which a second head tries to predict position from the code and the encoder is trained to defeat it. The fourth is importance reweighting of simulated calibration events to the bulk position distribution. All variants share the data, architecture, compute, and number of tuning trials.
Why not just…?
Why not just train a supervised classifier end-to-end on the calibration peaks? On calibration rejection it may well match or beat us, and we do not claim otherwise. Its training data, however, come entirely from the surface-biased calibration distribution, so nothing in its objective discourages it from relying on position-correlated pulse features such as drift time. The comparison that settles this is acceptance versus simulated radius and depth at a fixed rejection. If the supervised network turns out to be as flat as ours, the contrastive term is unnecessary and we will say so.
Why not just add label-supervised contrastive loss to the Holl bottleneck? That is the obvious combination. It inherits both flaws of the calibration labels: continuum contamination under the peak, and a single shared position distribution. It can sharpen class separation without touching position dependence. The second ablation tests this directly. We expect it to help rejection slightly, in line with the modest gains seen for latent regularizers elsewhere, and not to help flatness.
Why not just reweight or adversarially decorrelate against position? This is the real threat. Reweighting corrects the average efficiency if the simulation's position distribution is right. Adversarial decorrelation removes position information that a second head can detect, but such training is often unstable in small networks. Our pairing enforces invariance on each pattern rather than on the population. That should matter where multi-site geometry and position interact, for example near the point contact. We cannot argue this away. The third and fourth ablations will be compared on the same flatness and bulk-proxy metrics. If either matches us, it is the simpler method and should be preferred.
Why not just improve the position correction of the scalar cut? Better corrections of the amplitude-over-energy ratio exist in spirit. Still, the cut compresses the pulse to one number and misses multi-site events with an early sharp spike, which Holl et al. showed a learned code catches. A corrected cut remains a baseline, but it cannot recover information it discards.
Proposed Experiments
The experiments use two data sources. The first is thorium-228 calibration data from point-contact germanium detectors of the kind used in the MAJORANA and LEGEND analyses. Access through a collaboration or a public release is a dependency we cannot resolve from the literature. The second is a pulse-shape simulation of the same detector geometry, used to generate paired deposits and to provide ground-truth positions.
The first decisive check is on simulation. We bin held-out single-site events by radius and height and report the maximum deviation of acceptance across bins at the working point that gives 90% acceptance on simulated calibration-like events. Multi-site events get the same treatment for rejection. The target is a clearly smaller spread than the scalar cut, the Holl baseline, and the no-contrast ablation, with statistical uncertainties from bootstrap resampling.
The second check is on real data, where position is unknown. Low-energy data in enriched detectors are dominated by two-neutrino double-beta decay, a bulk-distributed and largely single-site population. A position-invariant discriminator should give acceptance there that agrees with its calibration acceptance after accounting for the expected multi-site fraction. The discrepancy between the two is our real-data estimate of the position systematic, compared across all methods.
The third check guards against losing ground on the main task. We report full-energy-peak and single-escape-peak rejection at 90% double-escape-peak acceptance, which must stay within error of the Holl result of 90%.
Finally, we measure robustness to the simulation itself. We retrain with deliberately perturbed simulation parameters, such as impurity profile and electronics shaping. The question is whether the gain survives a realistic sim-to-real mismatch or disappears, which would mean the invariance learned was specific to the simulator rather than physical.
References
- Multisite event discrimination for the majorana demonstrator
- Deep learning based pulse shape discrimination for germanium detectors
- Pixel Detector Background Generation using Generative Adversarial Networks at Belle II
- A Diffusion Model with Contrastive Learning for ICU False Arrhythmia Alarm Reduction
- Phantom Embeddings: Using Embeddings Space for Model Regularization in Deep Neural Networks
- Gaussian Latent Representations for Uncertainty Estimation using Mahalanobis Distance in Deep Classifiers
- Model-Based Deep Learning Algorithm for Pulse Shape Discrimination in High Event Rates