Composing the Unseen Signal: Attribute-Based Zero-Shot Recognition of Neutrinoless Double-Beta Decay

Experiments searching for neutrinoless double-beta decay, a hypothetical nuclear decay that would prove the neutrino is its own antiparticle, must pick out at most a few signal events per tonne-year from a much larger flood of radioactive background. Modern searches increasingly hand this job to neural networks, and those networks share an awkward property: they learn what "signal" looks like from simulation only, because no real signal event has ever been recorded. Whatever the simulation gets wrong about the detector is therefore built into the classifier. Zero-shot learning addresses a structurally identical problem. There, a model recognizes a category it has never seen by describing that category through attributes ("has a curved beak") that it learned to detect on other, seen categories.

We propose to treat the decay signal as a true zero-shot class. In a gas time projection chamber, which records each event as a 3D track of ionization, the signal is fully described by a short list of physical attributes: one connected track, a dense energy "blob" at both ends, no separate satellite deposits, a position inside the bulk of the gas, and a total energy at the decay's Q-value. Every one of these attributes occurs in some class of real, recordable events. The two-neutrino decay produces two-electron tracks at lower energies. The thallium-208 double-escape peak produces electron–positron tracks with two blobs. Gamma photopeaks produce one-blob tracks with scattered satellites. We will train attribute detectors on these real seen classes and recognize the signal by composing the attributes. The question is whether this removes the simulated signal from the classifier and, with it, the simulation-to-data systematic.

Core Claim

A discriminator whose signal definition comes from physically meaningful attributes learned on real data will hold its efficiency under detector mismatch better than a classifier trained end to end on simulated signal. This should matter most when no single real proxy class matches the signal in energy, position, and topology at the same time. The new element is not a better network but a different source of supervision. Signal is never a training label. It enters only as an attribute vector fixed by physics. The hard technical problem is the semantic gap specific to this setting: attributes are learned at energies and positions where seen classes live (two-neutrino events below the Q-value, double-escape tracks at 1593 keV near calibration sources), and they must transfer to 2458 keV in the bulk.

Relation to Prior Work

Three detector papers motivate the gap. The complete EXO-200 analysis (Search for Neutrinoless Double-Beta Decay with the Complete EXO-200 Dataset) turns wire signals into images and trains a deep network to separate signal from background. It raised signal efficiency from 80.8% to 96.4±3.0% and sensitivity by about 25%. It is still a closed-set classifier trained on simulated signal, and its file summary names detector-modeling systematics as a dominant limitation. PandaX-III's topological study (Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment) supplies our attribute vocabulary: track energy ratio, blob charge, end-track twist, and length. It applies these as independent one-dimensional cuts and reaches about seven orders of magnitude of gamma suppression at above 20% signal efficiency. We reuse those variables as attribute definitions. Unlike that work, we learn their detectors jointly from real classes rather than thresholding simulated distributions. In germanium detectors, Holl et al. (Deep learning based pulse shape discrimination for germanium detectors) already train on a real proxy, the double-escape peak versus the full-energy peak. They reach 90% background rejection at 90% signal acceptance, but through a 7-dimensional latent code with no physical meaning. That work constrains us: it shows a single real proxy can work well when it is close to the signal. Our claim must therefore be tested exactly where proxies and signal differ.

The method comes from zero-shot vision. DAZLE (Fine-Grained Generalized Zero-Shot Learning via Dense Attribute-Based Attention) grounds each attribute through its own spatial attention map and adds a calibration loss against seen-class overconfidence. We adapt the per-attribute attention to point clouds of ionization, so that the "end blob" attribute attends to track endpoints rather than to the whole event. Sample-graph methods (Graph Knows Unknowns: Reformulate Zero-Shot Learning as Sample-Level Graph Recognition) depend on keypoint detectors that do not exist for tracks, so we do not use them. For the energy-and-position shift we borrow the idea from Zero-Shot Domain Adaptation with a Physics Prior that a physics-derived invariant front end can align features across domains without target data. Instead of reflection invariants, we normalize charge-cloud width by the drift-depth-dependent diffusion and normalize along-track coordinates by the energy-dependent electron range. That paper also warns, through its stated trade-off, that hard invariants can remove useful information, so the invariant layer is an ablation, not an assumption. For events that match no known attribute pattern, Informed Democracy (Informed Democracy: Voting-based Novelty Detection for Action Recognition) gates novel inputs with Monte-Carlo dropout, which requires many stochastic forward passes. We replace this with a single-pass residual: the distance between an event's predicted attribute vector and the nearest known class signature. The LEGEND data-cleaning work (Machine learning-powered data cleaning for LEGEND) still needs human experts to label anomaly clusters. Our residual could flag such clusters automatically, though we treat this as a secondary outcome.

Research Program

The model encodes a track, given as 3D charge hits or 2D strip projections, with a point-cloud encoder. It optionally passes through the diffusion- and range-normalizing front end first. One attention head per attribute then produces an attribute score. The attributes are connectedness, the blob charge at each end, satellite energy fraction, end twist, normalized length, drift-depth bulk score, and energy. Each score is conditioned on the reconstructed energy through feature-wise modulation, so that "long for its energy" is learned rather than "long."

Supervision for the attributes comes only from real seen classes. Each class carries its own physics-defined attribute signature, and some entries are left unknown where physics does not fix them. Two-neutrino events are two-blob, connected, and in the bulk. Double-escape events are two-blob, connected, and near the source. Thallium photopeak and bismuth-214 line events are single-blob and usually carry satellites. Surface events are identified by diffusion width. The loss is a masked attribute loss plus a class-compatibility loss over seen classes, with DAZLE's calibration term holding probability mass back for the unseen signature.

At test time an event's signal score is its compatibility with the signal signature at the Q-value. The attribute residual gives an open-set flag. A deliberate design choice is that seen classes differ from signal in at least one attribute value or in the energy and position region. This is what turns the transfer into a test, not something that holds by construction.

Baselines

The naive reference is the PandaX-III cut chain from Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment: sequential one-dimensional cuts on energy ratio, blob charge, twist, and length, with thresholds set on nominal simulation. Its reported operating point is about seven orders of magnitude of gamma suppression at above 20% signal efficiency. Under our perturbed-detector conditions we expect it to lose efficiency. Beating it means higher signal efficiency at equal background rejection on the "real" test domain.

The strongest existing method is an end-to-end supervised network in the style of EXO-200, trained on simulated signal versus simulated backgrounds. EXO-200 reported 96.4±3.0% signal efficiency and about a 25% sensitivity gain over its earlier analyses (Search for Neutrinoless Double-Beta Decay with the Complete EXO-200 Dataset). That figure is from a liquid xenon detector, not our setup, so it is context rather than a target. We will run the same encoder backbone, the same total event count, and an equal hyperparameter search budget. We will also give it the same unlabeled real-domain events through a standard domain-adversarial adaptation variant, so that it has every resource our method has except attribute supervision on real classes.

The real-proxy baseline is a classifier trained directly on the real double-escape peak versus the real photopeak, following Deep learning based pulse shape discrimination for germanium detectors. That paper reports 90% background rejection at 90% signal acceptance in germanium. No reported result was found on Lacuna for any attribute-based or zero-shot discriminator on double-beta-decay data.

Three ablations isolate the key ingredients. The first is gradient-boosted trees on the hand-computed PandaX-III variables, with signal-compatibility rules fit on the same real seen classes. If this matches the full model, learned attribute grounding adds nothing beyond data-driven thresholds. The second removes energy conditioning. The third removes the physics-normalizing front end. Beating the strongest baseline means a statistically significant gain in the figure of merit $\epsilon/\sqrt{b}$ (signal efficiency over the square root of surviving background) on the mismatched test domain. Our method must also lose less than 2 percentage points of efficiency relative to its own matched-domain value.

Why not just…?

Why not just train on simulation and adapt to real data? This is the most plausible competitor, and it is the second baseline. Domain adaptation aligns feature distributions using unlabeled real data, but its signal decision boundary is still learned from simulated signal tracks. If the simulation misplaces blob density, adaptation can align the bulk distributions while leaving that boundary wrong. The closed-loop study below measures this directly, because there the true "real" signal is known.

Why not just use the hand-crafted topological variables with a multivariate classifier? These variables already are attributes, and PandaX-III left multivariate combination open. This is a real threat. We test it as the first ablation, and we will report honestly if it ties. A tie would still be a useful result: the gain then comes from using real seen classes to define signal, not from attention. That would argue for the cheaper method.

Why not just train on the double-escape peak, which is already a two-blob real proxy? In germanium detectors this works well, matching the standard pulse-shape method. In a gas detector, double-escape tracks sit at 1593 keV rather than 2458 keV, cluster near calibration sources, and include a positron. Our hypothesis is that composition beats the proxy exactly where these differences matter: at the Q-value, in the bulk, under position-dependent diffusion. If the proxy baseline is within statistical error everywhere, we will conclude that the zero-shot framing is unnecessary for this detector.

Proposed Experiments

The first experiment is a closed-loop study in which truth is known. Using the REST-based PandaX-III simulation (PandaX-III: Searching for neutrinoless double beta decay with high pressure 136Xe gas time projection chambers), we generate a nominal "simulation" domain and several "real" domains. These have perturbed transverse and longitudinal diffusion, strip gain non-uniformity, readout noise, and modified electron multiple scattering, and we also vary the 3 mm strip versus 3D pixel readouts that the topology paper compared. Every model trains with nominal simulation plus real-domain seen classes only, and is scored on real-domain signal. The decisive check is the curve of signal efficiency at fixed background rejection against mismatch size. Our claim predicts a flatter curve than both the simulation-trained network and the proxy classifier.

The second experiment is leave-one-class-out on real classes, the standard zero-shot protocol from Fine-Grained Generalized Zero-Shot Learning via Dense Attribute-Based Attention. Each real class with a known signature, such as the double-escape peak or bulk two-neutrino events in a high energy band, is withheld in turn and recognized only from its attribute vector. Real calibration classes stand in for the unseen signal here, so this tests transfer without relying on simulation at all. We report the harmonic mean of seen and unseen accuracy, plus efficiency for the withheld class at fixed background rejection.

The third experiment extrapolates in energy. Attribute detectors are trained on two-neutrino events below 1.8 MeV and tested on simulated two-electron tracks at 2.0, 2.2, and 2.46 MeV. This separates the energy-conditioning and range-normalization ablations.

The fourth experiment tests the open-set residual. Out-of-catalogue backgrounds are injected into the simulation: surface alphas, pile-up, and electronic transients. We report the area under the receiver-operating curve for flagging them, against Monte-Carlo-dropout gating at matched compute.

If access to germanium calibration waveforms like those of Deep learning based pulse shape discrimination for germanium detectors is available, we will repeat the second experiment with single-site, bulk, and surface attributes. That repeat checks whether the conclusion carries across detector types.

The proposal fails cleanly in two cases. If the hand-feature ablation or the proxy classifier matches the full model under mismatch, attribute composition is not worth its complexity. If the closed-loop efficiency curve is no flatter than the adapted simulation-trained baseline, attribute supervision on real classes does not reduce the simulation systematic.