Experiments searching for neutrinoless double-beta decay, a hypothetical nuclear decay that would show neutrinos are their own antiparticles, look for a signal occurring less than once per tonne-year. That signal sits under a flood of gamma-ray background. In a gas time projection chamber (TPC), a detector that records the ionization trail a charged particle leaves in gas, the signal has a recognizable shape: one continuous track with a dense energy "blob" at each end. A gamma-ray electron usually leaves only one blob. Reading this shape precisely requires 3D information. Full 3D pixel readouts, however, bring channel counts, heat loads and radioactive material that the experiment cannot afford. PandaX-III therefore reads out two orthogonal sets of strips, and each event arrives as two 2D pictures, X versus drift time and Y versus drift time. Drift time is the time the ionization takes to drift to the readout, so it stands in for depth. Pairing hits across the two pictures is ambiguous, and the handcrafted cuts used today do not resolve that ambiguity.
This note proposes a learned "pseudo-3D" reader for these strip data. Pseudo-3D (or 2.5D) models are the family of medical-imaging methods that give a cheap 2D network a short window of neighboring slices instead of a full 3D volume. Here the slices are drift-time slices. At each short window of drift time, the model lets the X-strip and Y-strip signals from that same window exchange information, so it can learn which X hits belong with which Y hits. The claim is concrete and has direct consequences for detector design. PandaX-III found that its 3 mm strip readout, analyzed with handcrafted cuts, matches only a coarse 10 mm 3D pixel readout. We hypothesize that most of the information lost by projecting onto strips is recoverable. If so, a drift-aligned cross-view model on 3 mm strips should approach the discrimination of a fine 3D pixel readout analyzed by a comparably strong 3D network.
Core Claim and What Is New
The motivating evidence comes from PandaX-III itself. Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment builds four handcrafted features: track energy ratio, blob charge, end-track twist and track length. Together with energy and diffusion cuts, these reach roughly $10^{-7}$ background survival at above 20% signal efficiency. Track shape alone contributes a rejection factor of about 30–35. The same paper shows that 3 mm strips match 10 mm 3D pixels, and it explicitly leaves multivariate and machine-learning analysis to future work. That paper and the detector overview PandaX-III: Searching for neutrinoless double beta decay with high pressure 136Xe gas time projection chambers establish the gap: no learned model has been applied to the strip projections. They also leave open the question of how much of the strip-versus-pixel difference is a readout limit and how much is an analysis limit.
The closest learned work treats detector views separately. In Vertex-finding and reconstruction of contained two-track neutrino events in the MicroBooNE detector, a 2D segmentation network labels each of three wire-plane images independently. A separate stochastic "walking" algorithm then assembles the 3D track. That pipeline loses efficiency where tracks overlap in projection. We keep the cheap 2D representation but move the cross-view matching inside the network, where it is aligned by drift time. KamNet: An integrated spatiotemporal deep neural network for rare event searches in KamLAND-Zen* shows that learned models beat conventional analysis in a double-beta decay search. Its events, however, are near-point-like light flashes seen by photomultipliers on a sphere, not extended tracks projected onto strips, so it motivates the direction without supplying the method.
The method borrows from medical 2.5D work. Hierarchical Encoder-Decoder With Soft Label-Decomposition for Mitochondria Segmentation in EM Images uses a 5-slice input to predict the center slice. EfficientCellSeg maps neighboring slices to color channels and pairs a large encoder with a small decoder. Both show that a short window of neighboring slices recovers most of what a 3D network provides, at a fraction of the cost. There is one important difference. In those papers, all slices share the same in-plane coordinates. In a strip TPC, the two views share only the drift axis. Stacking the two views as color channels would therefore overlay unrelated coordinates, which makes no geometric sense. That mismatch is why the key ingredient here is attention across views within a drift window, not channel stacking.
Research Program
The input to the model is two charge maps, X-strip versus drift time and Y-strip versus drift time, with drift time binned consistently in both. Each view is encoded by a small shared 2D convolutional encoder that sees a window of $\pm k$ drift bins, which is the pseudo-3D element. After encoding, the features of each view are pooled into one token per drift bin and strip. Tokens from one view attend to tokens from the other view only within the same drift window. This restriction encodes the physical constraint that charge arriving at the same time comes from the same depth.
The model has two outputs. The first is an event-level signal-versus-background score. The second is an auxiliary per-drift-bin prediction of the true 3D charge centroid and of whether that bin lies in a Bragg-peak blob, where a particle deposits most energy just before stopping. Simulation provides the true 3D deposits, so this auxiliary target is free. It gives the cross-view attention a direct incentive to solve the pairing problem rather than memorize shortcuts.
Training data will be Monte Carlo events produced with the REST simulation chain described in the PandaX-III work: 10 bar xenon-136 with 1% trimethylamine, and 3 mm strip readout. Signal is neutrinoless double-beta decay. Backgrounds are bismuth-214 and thallium-208 gamma events that deposit energy in the region near the xenon-136 decay energy (2458 keV). The same deposits will also be digitized as 3D pixels at 1, 3 and 10 mm pitch, so the readout comparison uses identical events. The analysis restricts to a fixed energy window, so the network cannot win by using energy alone. Diffusion, the spreading of the charge cloud with drift distance, is simulated as in PandaX-III. The diffusion-width depth proxy is kept as a separate, fixed cut applied identically to every method, so that improvements are attributable to topology.
Baselines
The naive reference point is the published handcrafted topological analysis. It uses PandaX-III's four features with their cut values, re-run on our simulation so that its shape-only rejection factor can be checked against the reported ~30–35. It is also reported at the reported operating point of $10^{-7}$ survival with more than 20% signal efficiency for the full chain. Reproducing these numbers validates the simulation before anything else is compared.
The second baseline is a gradient-boosted decision tree trained on those four features plus total energy and blob count. This is exactly the multivariate step the PandaX-III authors anticipated, and it is the cheapest way to exploit correlations between features. It gets the same training events and a hyperparameter search of equal wall-clock budget.
The third baseline follows the independent-views pattern of the MicroBooNE segmentation network. The same 2D encoder processes each strip view separately, and the two views are merged only by concatenating event-level features at the end. This baseline has identical parameters and training budget. It is also the ablation that isolates the key ingredient: the only difference is drift-aligned cross-view attention.
The fourth baseline is the strongest competitor. It is a sparse 3D convolutional network that sees only the active voxels, run on the 3D pixel simulations at 10 mm and at 3 mm. On true 3D data, it is the best available learned reader. The 10 mm pixel version tells us whether we beat the readout that handcrafted strips were found equivalent to. The 3 mm pixel version is the ceiling. It is tuned with the same budget.
The fifth baseline also uses a sparse 3D network, but applies it to "ghost voxels" built from the strip data. At each drift bin, every X hit is paired with every Y hit, and each pair is weighted by the product of the two charges. This produces a 3D volume containing both real hits and false pairings. The network must learn which pairings are real, a job our model does with attention.
No reported result was found on Lacuna for any learned model on the PandaX-III strip benchmark. The only published reference numbers are the handcrafted figures above. Beating a baseline means lower background survival at matched signal efficiency in the energy window. The significance is judged by bootstrapping over independent simulated test sets of at least $10^7$ background events after preselection, since the operating point sits in the far tail.
Why not just…?
Why not just train a BDT on the handcrafted features? It is cheap, it can be interpreted, and it may capture much of the correlation gain the PandaX-III authors expected. It cannot, however, recover information the features never measured, such as which of two overlapping endpoints is the true blob when the views disagree. If the BDT closes most of the gap to the 3 mm pixel ceiling, the learned reader is not justified, and we will say so. The comparison is built in.
Why not just run a 2D network on each view and fuse late? This design is simpler, and it is how the MicroBooNE network handles its views. Its known weakness is overlap in projection, where MicroBooNE's efficiency drops. Our late-fusion ablation measures exactly this. If cross-view attention gives no gain over late fusion, the pairing hypothesis is wrong.
Why not just build ghost voxels and use a sparse 3D network? This is the real threat. It is a legitimate way to give a 3D network the strip data, and it may perform as well. Our argument against it is weaker than a hard proof. The number of ghost voxels grows with the product of the hit counts in the two views at each drift bin, and the charge-sharing constraint is only implicit in the weights. We will compare accuracy, and also compute and latency, which matter for the real-time filtering direction raised in Intelligent experiments through real-time AI. If ghost voxels match us on both, the conclusion becomes "learned readers recover the projection loss" regardless of architecture. That conclusion would still settle the detector-design question.
Proposed Experiments and Decisive Checks
The central experiment plots shape-only rejection at fixed signal efficiency for each method, against pixel pitch for the 3D references, with the strip-based methods drawn as horizontal lines. The claim is supported if the cross-view model on 3 mm strips clearly exceeds the sparse 3D network on 10 mm pixels. It is strongly supported if it reaches the 3 mm pixel ceiling within bootstrap error. The claim fails if the model sits no better than the BDT or the late-fusion ablation.
A second check tests where the gains come from. We bin events by projected opening angle and by the overlap between the two blobs in each view. The model should gain most exactly where it resolves pairings that the handcrafted cuts and late fusion get wrong. The auxiliary 3D-centroid error, split by the same bins, directly tests whether the attention learned the pairing.
Because a rare-event limit cannot rest on a model that is overconfident about simulation, the third check tests robustness. We perturb diffusion, gain and strip noise in the test simulation away from their training values. We then report how far the background survival moves compared with the BDT. A model that gains in rejection but becomes much less stable than the BDT under these shifts would not be usable in a real analysis. That trade-off, not the headline rejection number, is the result that should decide adoption.
References
- Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment
- PandaX-III: Searching for neutrinoless double beta decay with high pressure 136Xe gas time projection chambers
- Vertex-finding and reconstruction of contained two-track neutrino events in the MicroBooNE detector
- KamNet: An integrated spatiotemporal deep neural network for rare event searches in KamLAND-Zen*
- Hierarchical Encoder-Decoder With Soft Label-Decomposition for Mitochondria Segmentation in EM Images
- EfficientCellSeg: Efficient Volumetric Cell Segmentation Using Context Aware Pseudocoloring
- Intelligent experiments through real-time AI: Fast Data Processing and Autonomous Detector Control for sPHENIX and future EIC detectors