Two-Endpoint Point Sets for Robust Topological Discrimination in Xenon Gas TPCs

Neutrinoless double-beta decay is a hypothetical radioactive decay in which a nucleus emits two electrons and no neutrinos. Seeing it would show that the neutrino is its own antiparticle. In a high-pressure xenon gas detector, the decay appears as one wiggly ionization track with a dense clump of charge, or "blob," at each end, because both electrons slow down and stop there. The dominant background is a gamma ray that knocks out a single electron with the same total energy. That electron makes a similar wiggly track but has only one blob. Today this distinction is made either with a handful of hand-tuned cuts or, in other detectors, with a neural network that reads the whole event image and outputs "signal" or "background." We propose a third option. The network first predicts a small, explicit geometric object: a deformable chain of points that follows the primary track and ends in exactly two endpoint points. Each endpoint carries a measured charge. Every energy deposit is also accounted for as primary track, detached satellite, or unexplained. The final decision is made from these few physical quantities, not from the raw image.

The bet is not that this wins on clean simulation; a strong end-to-end classifier probably ties it there. The bet is that it holds up when simulation and reality disagree, and when the detector sees background shapes the simulation never contained. Those are the two failure modes that decide whether a machine-learned cut can be trusted in a search whose signal rate is below one count per year. The endpoint representation targets the first failure mode. The "unexplained charge" channel targets the second. It borrows the open-world detection idea that unlabeled things should not be forced into a closed set of classes.

Core Claim

The PandaX-III design is a 200 kg, 10-bar xenon gas time projection chamber (TPC) read out through strips in two projections, XZ and YZ. Its best published discrimination comes from four hand-built topological variables: track energy ratio, blob charge, end-track twist and track length. Together they suppress background by about seven orders of magnitude while keeping signal efficiency only "above 20%" (Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment). Topology alone gives a rejection factor of roughly 30–35 (PandaX-III: Searching for neutrinoless double beta decay with high pressure 136Xe gas time projection chambers). Those authors list multivariate and machine-learning analysis as future work. This paper motivates the proposal: the physics is clear, but the cuts are sequential one-dimensional thresholds that throw away correlated structure and most of the signal.

The EXO-200 liquid xenon experiment shows what learning can buy. A network reading 2D wire images raised signal efficiency from 80.8% to 96.4±3.0% and half-life sensitivity by about 25%. The same paper flags simulation-versus-data shape mismatch as a systematic (Search for Neutrinoless Double-β Decay with the Complete EXO-200 Dataset). This paper motivates the proposal and also constrains it: learned discrimination helps, but a whole-image classifier's efficiency is hard to validate on real data.

Our claim is that an intermediate representation fixes the validation problem. That representation is two endpoints with their blob charges, plus an accounting of all charge. Blob charge, endpoint separation and unexplained-charge fraction are quantities a physicist can check on calibration events in real data. A whole-image classifier's score is not.

The geometric part of the design adapts point-set tracking. RPT replaced bounding boxes with nine learned points that shift toward an object's extremities. This avoids the background contamination that rectangles cause for bending or elongated targets, a failure documented in VOT2014 and Pose-Configurable Generic Tracking of Elongated Objects. We reuse the offset-point mechanism but change what the points mean. They form an ordered chain, the first and last points are the electron endpoints, and endpoints must share drift coordinate Z across the two strip projections. That last constraint is specific to the stripped readout and gives a way to pair up the two 2D views.

The research file notes that point sets struggle with disconnected clusters such as Compton satellites. We handle this outside the chain: detached deposits are explained by a separate satellite head rather than bent into the track.

Research Program

The backbone is a submanifold sparse convolutional network, meaning a network that computes only at detector cells that actually recorded charge. Following SparseSSNet, we use it because dense convolutions smear sparse tracks and blur the endpoints that carry the signal. SparseSSNet reached 99.6% shower-versus-track accuracy against 95.9% for a dense network, using 1 GB rather than 6 GB of memory.

The file points out that purely submanifold features cannot reach into neighboring empty cells, which matters for measuring blob spread and diffusion. We therefore read out endpoint features by pooling charge within a learned radius around each predicted endpoint, instead of taking features only at occupied sites.

The network has three heads on each projection. The chain head predicts K ordered points, initialized along the principal axis and refined by learned offsets. It is trained against the simulated true electron path with a Chamfer-type loss. It also receives an endpoint loss that matches the two terminal points to the true electron stopping points, allowing either ordering. Cross-projection Z agreement enters as a penalty. The charge head assigns each hit to primary chain, satellite, or unexplained. The first two classes are supervised from simulation truth. "Unexplained" is never supervised as a positive class. Instead, we follow the objectness-decoupling logic of OW-DETR, which learns whether something is an object separately from which class it belongs to. We also follow the image-adaptive background modeling of D2Zero. Charge that no head explains with high confidence is left unexplained rather than pushed into a label. This matters because the known failure of closed-world training, documented in Zero-Shot Object Detection, has a dangerous mirror image here: an unmodeled background topology gets squeezed into whichever class it most resembles, possibly "signal."

The final discriminant is a small monotone model on six inputs: the two endpoint blob charges, endpoint separation, chain length, twist near each endpoint, and unexplained-charge fraction. It is trained on simulation.

We choose offset refinement over generative point diffusion (DiffusionTrack) because an event is a single static frame and iterative denoising adds cost without an obvious benefit. We keep diffusion as an optional variant only if the offset chain fails on strongly curled tracks.

The figure of merit is the standard background-limited half-life sensitivity. It follows from requiring a signal count comparable to the square root of the expected background count:

$$T_{1/2} \propto \varepsilon \sqrt{\frac{M t}{B \, \Delta E}}$$

Here $\varepsilon$ is signal efficiency, $M t$ is exposure, $B$ is background index per unit energy, mass and time, and $\Delta E$ is the energy window. We compare methods on $\varepsilon/\sqrt{B}$ at matched operating points. Doubling efficiency at fixed background is worth as much as cutting background fourfold.

Baselines

The naive reference point is energy selection alone in the region around the 2458 keV decay energy, with no topology. This fixes the scale of what topology adds.

The first real competitor is the published REST topological cut set (Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment). Its reported operating point is roughly seven orders of magnitude of total background reduction at signal efficiency above 20%, with a topology-only rejection factor of about 30–35 (PandaX-III). We rerun it on our simulated samples and reproduce the reported point before comparing anything. We also give it a fair modern upgrade: a gradient-boosted classifier on the same four variables, tuned with the same budget as our model. This checks whether "use ML" merely means "combine the four variables better."

The strongest competitor is an end-to-end sparse classifier: the same submanifold backbone, the same data, the same augmentations and the same hyperparameter search budget, but with a single signal-versus-background output. It is the natural PandaX-III analogue of the EXO-200 network, which reported 96.4±3.0% signal efficiency and about 25% sensitivity gain on its own detector (EXO-200). That number comes from a different medium and readout and is not directly comparable. No reported result was found on Lacuna for a learned classifier on PandaX-III stripped-readout topology.

A dense 2D convolutional classifier on the two projections is included to test the sparsity argument locally.

Two ablations isolate the key ingredients. The first keeps our trained backbone and heads but feeds the backbone's pooled features straight to the final classifier, bypassing the six physical quantities. This tests whether the bottleneck matters. The second removes the unexplained channel by forcing every hit into primary or satellite.

Beating the strongest baseline means one of two things. The first is matching its $\varepsilon/\sqrt{B}$ on nominal simulation to within statistical error while losing clearly less under the shifts below. The second is exceeding it outright at the 20% and 40% efficiency operating points.

Why not just…?

Why not just train the sparse end-to-end classifier and stop there? On nominal simulation it may well match or beat us, and we expect that. Its weakness is that its score has no data-checkable intermediate, so a simulation-data mismatch shifts efficiency invisibly, which is the systematic EXO-200 names. The stress tests below decide this directly. If the end-to-end classifier degrades no more than ours under shifted diffusion, gain, noise and dead strips, the structured bottleneck is not justified and we would report that.

Why not just feed the four REST variables to a boosted classifier? This removes the sequential-threshold loss cheaply, and it may recover much of the efficiency gap. It cannot fix errors in the variables themselves, such as mis-identified endpoints when a track curls back on itself in one projection or crosses a dead strip. Measuring endpoint-localization error per method on simulation truth shows whether better geometry, rather than better combination, drives any gain.

Why not just harden the end-to-end classifier with augmentation or domain adaptation? This is the most serious threat. So the end-to-end baseline receives exactly the perturbation augmentations our model sees during training. The held-out shifts in the test are drawn from ranges and background species excluded from all training. If augmentation alone closes the gap, the geometric part of the proposal adds nothing beyond interpretability, and we would say so.

Proposed Experiments

All training and test samples come from the REST and Geant4-based PandaX-III simulation chain described in PandaX-III. It produces 10-bar xenon with 1% trimethylamine quencher, 3 mm strip readout, and signal plus bismuth-214 and thallium-208 gamma backgrounds from detector components.

The first experiment is the nominal comparison. For every method we report background rejection at fixed signal efficiency, full efficiency-versus-rejection curves, and $\varepsilon/\sqrt{B}$.

The second experiment is the decisive robustness check. We train on nominal settings, then test on held-out perturbations: transverse diffusion scaled by ±20%, strip gain nonuniformity, added electronic noise, and randomly masked dead strips. For each perturbation we report the change in signal efficiency at a fixed nominal threshold. The prediction is that our efficiency shift is smaller and, crucially, predictable from the shift in endpoint blob-charge distributions. Those distributions can be monitored on calibration events in data; end-to-end scores cannot.

The third experiment tests the open-set claim. We inject background species absent from training, namely surface uranium-chain events near the readout plane and alpha tracks. We then measure the fraction accepted as signal with and without the unexplained channel. The known Z-position ambiguity of this readout makes surface events a real threat (Topological background discrimination).

The fourth experiment is a data-like validation proxy. We simulate thallium-208 double-escape events, which are genuine two-blob electron-positron tracks near 1592 keV, alongside single-electron Compton events at matched energy. We check whether per-endpoint blob charge measured on these calibration-like samples predicts the signal efficiency at the decay energy.

The idea fails if the end-to-end classifier, given the same augmentations, shows equal or smaller efficiency shifts in the second experiment and equal open-set rejection in the third. In that case the right recommendation to the experiment is the simpler model.