Experiments that look for neutrinoless double-beta decay, a hypothesized nuclear decay so rare that its half-life would exceed $10^{25}$ years, separate a handful of signal events from gamma-ray backgrounds mainly by the shape of the ionization track. In a high-pressure xenon gas time projection chamber (TPC), which is a gas volume where ionization electrons drift to a segmented readout plane that records the track in 3D, a signal event leaves one continuous track with a dense energy "blob" at each end. A gamma background usually leaves a single blob or several disconnected pieces. How well this shape can be seen depends on how finely the readout is segmented and on how much the electron cloud spreads out ("diffuses") as it drifts. That spread is not a fixed property of the detector. It grows with drift distance, so every event is effectively recorded at its own resolution.
This proposal treats readout resolution as a known, per-event input to a learned signal-versus-background classifier, rather than as a nuisance the classifier should become invariant to. We train one classifier on a single fine-grained simulation that is re-digitized on the fly through a physical readout model covering pitch, strip-versus-pixel projection, and depth-dependent diffusion. The network is told the effective resolution of each event, which can be estimated from the drift-depth proxy that PandaX-III already uses. We then ask two questions. Does this single classifier match retrained specialists at each configuration and beat them on drift depths and pitches it never saw? And does a learned classifier change the detector-design curve that so far has been drawn only with hand-built topological cuts?
Core Claim
Resolution loss in a track detector erases specific discriminating information. When the blobs smear into the track at coarse pitch or long drift, a good classifier should shift its weight toward features that survive, such as track length and the fraction of energy in the main track. Methods that enforce invariance to blur force the model onto the features available at the worst resolution, and so throw away information at good resolution. A classifier conditioned on the measured per-event resolution can use the best features each event supports. We claim this conditioned classifier will hold rejection at fixed signal efficiency across the full drift length, where both invariance-style training and fixed-resolution training lose ground at large depths. We also claim that the significance-versus-pitch curve it produces will differ measurably from the classical-cut curve that motivated PandaX-III's choice of readout.
Grounding in Prior Work
The application and the benchmark come from the PandaX-III program. The baseline detector paper, PandaX-III: Searching for neutrinoless double beta decay with high pressure 136Xe gas time projection chambers, specifies 10 bar enriched xenon with a trimethylamine quencher added to limit diffusion. It projects a shape-only rejection factor of about 30–35. Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment uses the REST simulation framework to evaluate four hand-built track features: track energy ratio, blob charge, end-track twist, and track length. It reports roughly seven orders of magnitude of background suppression at above 20% signal efficiency. It also finds that a 2D 3 mm strip readout matches the significance of a 10 mm 3D pixel readout. Separately, it shows that pulse width, which grows with drift depth, rejects surface background to a factor of 0.15 while keeping 75% of signal.
This work both motivates and supplies our method. It shows that resolution controls discrimination. It also gives us a per-event depth estimate, which we reuse as the conditioning signal. However, it evaluates each readout separately, uses only classical features, and never asks whether a classifier tuned at one resolution survives at another. Its design conclusion rests on those classical features, which its own discussion notes miss multi-dimensional correlations.
Two learned-discriminator results show that learned models can beat classical reconstruction in this field, and both are tied to a single detector geometry. KamNet uses a symmetry-respecting spherical network in KamLAND-Zen to reach 74.0% rejection of carbon-10 background at 90% signal efficiency, versus 61.5% for a conventional CNN. The MicroBooNE vertexing work, Vertex-finding and reconstruction of contained two-track neutrino events in the MicroBooNE detector, treats wire readouts as images. It reports that the pipeline fails on short low-energy tracks that span too few wires and on high-energy tracks whose small opening angles merge into a single blob. That failure pattern is the one we expect for blob features at coarse effective resolution.
On the vision side, AerialFormer keeps a high-resolution convolutional stem so that tiny objects survive downsampling. We borrow that design choice for the network that reads charge maps, but AerialFormer has no notion of a known, varying resolution. Deep Diffusion-Invariant Wasserstein Distributional Classification makes predictions stationary under diffusion and downsampling perturbations. It reaches 93.0% versus 73–75% for ResNet/DenseNet on perturbed CIFAR-10. It is the clearest representative of the invariance approach we argue against for this setting. When the blur is real physics rather than a corrupting nuisance, the true optimal classifier changes with blur, so full invariance is the wrong target. Its restriction to roughly 128-dimensional Gaussian outputs is a further practical limit. StepSPT adapts to low-resolution target domains with style prompts and reports drops on low-resolution EuroSAT. Its authors acknowledge it has no explicit foreground/background mechanism. This constrains us: unsupervised adaptation alone should not be expected to recover blob information.
Research Program
The first component is a readout operator. We simulate signal and background events once in REST at the finest granularity available (1 mm 3D pixels, minimal diffusion). We then write a re-digitization function that takes true energy depositions plus a configuration (pitch, strip or pixel projection, drift depth, diffusion coefficients, electronic noise level) and returns readout hits. This makes the cost of sampling a new resolution a re-digitization rather than a new Geant4 run. It is also exactly what lets each event in training carry a different effective resolution.
The second component is the classifier. It is a sparse 3D convolutional network over charge maps (with a 2D two-view variant for strips) that keeps a full-resolution stem in the spirit of AerialFormer. It is conditioned on a resolution vector made of pitch, readout type, and the estimated drift depth. Conditioning enters through feature-wise scaling of intermediate layers. At test time the drift depth is never taken from simulation truth: it is estimated from the pulse-width proxy, so errors in that estimate propagate honestly.
The third component is analysis. We measure the gradient attribution placed on blob regions versus track body as effective resolution worsens, to test whether the conditioned model actually reallocates its attention as hypothesized. We then recompute the significance-versus-readout curve with the learned classifier.
Baselines
The naive reference point is the classical topological cut chain of the PandaX-III topology study, rerun by us in REST on identical event samples and configurations. Its reported operating point is about seven orders of magnitude of background suppression at above 20% signal efficiency, and the PandaX-III detector paper reports 30–35 from shape alone. We expect to reproduce these within statistical error before comparing anything. A boosted decision tree on the same four classical features plus pulse width serves as the multivariate classical baseline. It tests whether gains come from learned spatial features or merely from combining existing ones.
The strongest competitor is a set of specialist networks with the identical architecture. Each specialist is trained only on data digitized at one fixed configuration and depth bin, with the same total number of training events and the same hyperparameter search budget as the conditioned model. We tune over learning rate, width, and augmentation on a held-out validation split common to all. On configurations seen in training, these specialists are effectively an oracle. Our model counts as competitive if it is within statistical error of them there. It counts as beating them if it shows higher rejection at fixed signal efficiency on held-out drift depths and on pitches withheld from training (for example 5 mm and 7 mm pixels when training on 1, 3, and 10 mm).
Two ablations isolate the key ingredient. First, the same network trained on randomized configurations but given no resolution input tests whether randomization alone is sufficient. Second, the same network trained with an added penalty pushing its features to be invariant across re-digitizations of the same event is our stand-in for the invariance approach of DeepWDC. No reported result found on Lacuna for any deep-learning classifier on the PandaX-III REST benchmark, so all learned baselines will be produced by us.
Why not just…?
Why not just retrain a network for each readout configuration? This is the most serious alternative. For a finite list of discrete pitches it may match our model, and we treat it as the oracle to match rather than to beat. It cannot handle the fact that resolution varies continuously from event to event with drift depth inside one detector. Depth-binned specialists fragment the training data, and their boundaries become sources of error. They also say nothing about configurations not yet simulated. The held-out-depth and held-out-pitch experiments settle this. If specialists binned finely in depth match us everywhere, the conditioning contribution is void and the project reduces to the design-curve study.
Why not just randomize the readout during training without telling the network the resolution? Randomization is our own training scheme minus one input, so this is the ablation that matters most. It could genuinely work if the network can infer resolution from the charge map itself, for instance from track width. We will report it head to head. If it closes the gap, we will say so and conclude that explicit conditioning is unnecessary. The design-curve and attribution results would still stand on their own.
Why not just upsample or deblur coarse readouts and apply the fine-resolution classifier? Interpolation cannot restore an end-track scattering angle or a blob's charge density that coarse pitch has averaged away. Smoothing instead spreads blob charge into the track and erodes the very two-blob boundary the analysis depends on. We test this directly: we pass bicubic-upsampled and learned-super-resolved coarse events to the 1 mm specialist and compare rejection against the conditioned model.
Why not just keep the classical cuts, which already give seven orders of magnitude? Classical cuts are the right baseline and remain the trusted analysis. The topology study itself notes that they miss correlations among features. In KamLAND-Zen, a learned model added rejection beyond a conventional pipeline (74.0% vs 61.5% at 90% efficiency in KamNet). The open question is whether a comparable margin exists in TPC topology, and whether it holds at coarse effective resolution, where cut thresholds tuned for one depth drift out of calibration.
Proposed Experiments and Decisive Checks
All experiments use REST simulations of PandaX-III with 136Xe two-blob signal events at the Q-value and gamma backgrounds from detector materials, including surface uranium-238. Readouts span 1–10 mm 3D pixels and 3 mm 2D strips over the full drift length, with an electronic noise and dead-channel model added to address the topology study's acknowledged lack of realistic electronics. The primary metric is background rejection at fixed signal efficiency, reported at the classical operating point and at 75% efficiency. We also report the significance figure of merit used in the topology study and the surface-background factor at matched signal retention.
The first decisive check is drift-depth robustness. We plot rejection versus depth for the classical cuts, specialists, unconditioned randomization, the invariance-penalized model, and the conditioned model. The hypothesis fails if the conditioned curve does not lie above the others in the deepest third of the drift volume. The second check is held-out pitch generalization. The third is a sensitivity test: we deliberately perturb the depth estimate to quantify how much the gain depends on accurate conditioning.
The final check recomputes the strip-versus-pixel comparison with the best learned classifier. If the 3 mm strip readout no longer matches 10 mm pixels, or if finer pixels gain much more than classical features suggested, that is a concrete design-relevant result for the tonne-scale phase. Because the study is entirely simulation-based, we will state plainly that real-data validation, once detector runs exist, is the step that could overturn any of these conclusions.
References
- PandaX-III: Searching for neutrinoless double beta decay with high pressure 136Xe gas time projection chambers
- Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment
- KamNet: An Integrated Spatiotemporal Deep Neural Network for Rare Event Search in KamLAND-Zen
- Vertex-finding and reconstruction of contained two-track neutrino events in the MicroBooNE detector
- AerialFormer: Multi-resolution Transformer for Aerial Image Segmentation
- Deep Diffusion-Invariant Wasserstein Distributional Classification
- Step-Wise Distribution-Aligned Style Prompt Tuning for Source-Free Cross-Domain Few-Shot Learning