Searches for neutrinoless double-beta decay must pick out a few true decays from a far larger number of gamma-ray events. In a high-pressure xenon gas detector, the two kinds of event leave different ionization tracks. A true decay emits two electrons from one point, so its track has a dense charge "blob" at each end. A gamma background usually produces one electron, so its track has a single blob. The current discriminator for PandaX-III uses four hand-built track measurements and keeps only slightly more than 20% of signal. Two detector effects blur the evidence. The first is diffusion: the ionization cloud spreads by an amount that grows with drift distance, so blobs smear more the farther the event sits from the readout. The second is the missing start time: without an external trigger, the event's absolute depth is unknown. We propose to handle both with a tool from point-cloud registration, which is the problem of finding the pose, and sometimes the scale, that aligns two 3D shapes.
The tool is the Fourier-basis "functional" representation of Scale-Invariant Fast Functional Registration (FLS). FLS turns a point set into Fourier coefficients in a single averaging pass. It solves for pose and scale separately, and it down-weights high frequencies so that alignment proceeds from coarse to fine. Our observation is that the nuisance factors in a detector track have the same structure that FLS was built to factor out. An unknown position is a translation. An unknown orientation is a rotation. Diffusion is a Gaussian blur, and a Gaussian blur is diagonal in the Fourier basis, meaning it scales each coefficient independently by a known function of one width parameter. We would encode each track functionally, register it to a canonical frame, estimate its diffusion width the way FLS estimates scale, undo the blur within the frequency band where noise allows, and classify on the corrected coefficients. The diffusion width is also a direct proxy for drift depth, so the same estimate serves the surface-background rejection that Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment currently performs with a separate pulse-width cut.
Core Claim
Diffusion width is depth-dependent and is currently left out of the discriminator. Deep events are blurred more, and that blurring erodes exactly the end-blob contrast that separates signal from background. Hand cuts and generic learned classifiers trained over all depths average this effect away. If diffusion is modeled explicitly as a per-event parameter in a Fourier representation, track discrimination should become roughly uniform in depth. Pooled signal efficiency at fixed gamma rejection should rise above the four-cut baseline. Performance should also degrade less when the true diffusion coefficient differs from the simulated one. The registration literature supplies the method. The rare-event problem supplies a nuisance structure (translation, rotation about the drift axis, and a one-parameter blur) that this method fits unusually well.
Mechanism
PandaX-III's baseline readout uses 3 mm Micromegas strips, a gas-avalanche charge readout. It records two projections of each event: X against drift time and Y against drift time. A projection of a 3D charge density, meaning its integral along one axis, has Fourier coefficients equal to a slice of the 3D coefficients (the projection-slice theorem). The X–Z strip view therefore gives the 3D coefficients with $k_y=0$, and the Y–Z view gives those with $k_x=0$. The two views share the $k_z$ axis, which gives a natural consistency check between them.
For a projection with transverse and longitudinal diffusion widths $\sigma_T$ and $\sigma_L$, the observed coefficient at frequency $(k_x,k_z)$ is
$$c^{\mathrm{obs}}_{k} = c^{\mathrm{true}}_{k}\,\exp\!\left(-\tfrac12\sigma_T^2 k_x^2 - \tfrac12\sigma_L^2 k_z^2\right) + \eta_k,$$
where $\eta_k$ is readout noise. Both widths grow as the square root of drift length times the gas diffusion constant, so one depth parameter sets both. We estimate that parameter per event by maximum likelihood over a short band of low frequencies. Track shape at low frequency is dominated by overall length and extent rather than by blob detail, which reduces confusion between blur and intrinsic shape. We then apply the inverse factor only where the estimated signal-to-noise ratio exceeds one. Above that band we keep FLS's frequency down-weighting, which here plays the role of a regularized deconvolution rather than a coarse-to-fine schedule.
Translation is removed by centering on the charge centroid. Centering absorbs the unknown start time, since a shift in depth only changes coefficient phases. Rotation about the drift axis is removed by aligning to the principal axes. Principal axes have sign ambiguities, which are the 180° flips that Scale-Invariant Fast Functional Registration names as its main failure mode. We do not try to resolve the flips. We average the classifier output over the four sign choices, which is cheap and makes the output exactly invariant to them.
The classifier is a small multilayer perceptron, a standard feed-forward network, on a fixed-length coefficient vector. Its cost is on the order of the 0.76 million floating-point operations reported in Efficient Learning on Point Clouds with Basis Point Sets, so it could also run in trigger-level filtering of the kind described in Intelligent experiments through real-time AI.
Relation to Prior Work
Two PandaX-III papers motivate the work and define the target. PandaX-III: Searching for neutrinoless double beta decay with high pressure 136Xe gas time projection chambers specifies the detector: 10 bar enriched xenon with 1% trimethylamine (TMA), a quenching gas added to keep diffusion low. It reports a topology-only rejection factor of 30–35 and assumes uniform drift. Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment applies four sequential cuts: energy fragmentation, blob charge, end-track twist, and length. It uses pulse width to reject surface events and leaves multivariate methods to future work. We differ from both by treating diffusion as an estimated per-event parameter shared by the topology and depth tasks, rather than as fixed smearing in topology and a separate cut for depth.
Within registration, our method comes from FLS. We reuse its averaging encoding and frequency weighting and replace its scale estimate with a diffusion-width estimate. Efficient Learning on Point Clouds with Basis Point Sets and PointNetLK serve as competing encoders. PointNetLK also supplies a design constraint: its finding that max-pooled descriptors give non-smooth objectives is why we avoid max pooling in the encoder.
Vertex-finding and reconstruction of contained two-track neutrino events in the MicroBooNE detector is a cautionary example. Its 2D segmentation plus local stepwise 3D tracking reaches about 69% efficiency and breaks down for short tracks and small opening angles. Those are the same regimes in which our global encoding could also fail, so we report results by track length.
The pulse-shape successes in germanium detectors, in α-event characterization and rejection in point-contact HPGe detectors, and in liquid argon, in First Results from the DEAP-3600 Dark Matter Search with Argon at SNOLAB, show that a well-chosen physical scalar can beat elaborate machinery. We take that seriously: the diffusion estimate is meant to be such a scalar, not a black box.
Research Program
We will generate signal and gamma-background events (from the trace radioisotopes bismuth-214 and thallium-208) in the REST simulation framework of Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment. Truth depth will be retained for every event, and both the 3 mm strip readout and a 3D pixel readout will be simulated. We will then build the encoder and diffusion estimator.
A first check comes before any classifier: the depth estimate must be validated against truth and compared with the paper's pulse-width estimator. A second simulation set will vary the diffusion constant, the TMA fraction, the strip gain uniformity, and the dead-strip patterns. Training will use only nominal conditions. This set tests what a real deployment would face.
Baselines
The naive reference applies only the energy region-of-interest window and a single-connected-cluster requirement. It shows how much of the gain comes from topology at all.
The physics baseline is the four sequential cuts of Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment. That paper reports roughly 10^7 gamma suppression at signal efficiency above 20%, and surface uranium-238 background reduced by a factor of 0.15 at 75% signal retention. PandaX-III: Searching for neutrinoless double beta decay with high pressure 136Xe gas time projection chambers separately reports a topology-only rejection factor of 30–35. These figures may include different preselections, so we will re-implement the cuts on our own simulation and reproduce them before comparing. Beating this baseline means higher signal efficiency at equal background acceptance, after energy cuts, at the 10^-7 operating point quoted for the full chain.
The strongest learned competitors, and the real threats, are three. The first is a gradient-boosted decision tree on the four variables plus pulse width, which is the multivariate analysis the paper defers. The second is a two-channel 2D convolutional network on the raw strip projections, trained with translation and rotation augmentation. The third is a point network or basis-point-set encoder on reconstructed hits. We found no reported result on Lacuna for any machine-learning classifier on PandaX-III topology. Each competitor gets the same training events and the same hyperparameter search budget, measured in trials and GPU-hours. Each is selected on the same validation set.
Three ablations isolate the key ingredient. The first turns diffusion correction off, keeping canonicalization and frequency weighting. The second keeps the diffusion estimate as an input feature but does not deblur. The third replaces sign-symmetrization with arbitrary sign choices. The claim fails if the first ablation matches the full method.
Why not just…?
Why not just feed pulse width into a boosted tree on the four existing variables? This is the most serious alternative. Diffusion depth would then reach the classifier as a feature, and a tree can learn depth-dependent thresholds. The difference is that the tree still sees blob charge and twist as already-blurred scalars. Our method corrects the full shape before any summary is computed. The second ablation (diffusion as a feature, no deblurring) is the matched comparison. If it closes the gap, the honest conclusion is that a better-cut analysis suffices, and we would report that.
Why not just train a convolutional network on the two strip projections with augmentation? Augmentation can teach invariance to position and rotation, and a network can learn that depth predicts blur. However, it must learn both from data, and it gets no guarantee outside the training distribution. The shifted-diffusion test set addresses this directly. If the network matches us at nominal conditions but degrades more under shifted diffusion, the explicit model is justified. If it matches us everywhere, our contribution reduces to interpretability and lower cost.
Why not just use a standard registration matcher or a point network on hits? The file reports that the correspondence-based registration method TEASER recovered 0% of alignments under heavy noise and partial overlap, against 56% for FLS combined with Iterative Closest Point, in Scale-Invariant Fast Functional Registration. In addition, tracks have no fixed template to match against. PointNet costs about 440 million floating-point operations per event, against 0.76 million for a basis-point-set multilayer perceptron, according to Efficient Learning on Point Clouds with Basis Point Sets. The point-network baseline tests whether that cost buys accuracy here.
Decisive Checks
The primary metric is signal efficiency at fixed background acceptance, swept toward 10^-7. Because the extreme tail is statistics-limited, we will estimate it with large simulated background sets and bootstrap intervals. We will also report the sensitivity figure of merit: efficiency divided by the square root of the background count.
The decisive plot shows efficiency against truth drift depth. Our claim predicts a flat curve for the full method and a falling curve for the four-cut baseline, the boosted tree, and the no-deblur ablation. Two secondary checks follow. The first is surface-background rejection from the diffusion estimate, compared with pulse width at 75% signal retention. The second is the size of the gap under shifted diffusion constant and dead strips. Comparing strip and pixel readout tests whether the projection-slice treatment narrows the strip-versus-10 mm-pixel equivalence that the PandaX-III paper reports.
A failure would also be informative. If the depth curves are flat for every method, then diffusion does not limit topology at TMA-quenched pressures, and the open problem lies elsewhere.
References
- Scale-Invariant Fast Functional Registration
- Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment
- Efficient Learning on Point Clouds with Basis Point Sets
- Intelligent experiments through real-time AI: Fast Data Processing and Autonomous Detector Control for sPHENIX and future EIC detectors
- PandaX-III: Searching for neutrinoless double beta decay with high pressure 136Xe gas time projection chambers
- PointNetLK: Robust & Efficient Point Cloud Registration Using PointNet
- Vertex-finding and reconstruction of contained two-track neutrino events in the MicroBooNE detector
- $$\alpha $$-event characterization and rejection in point-contact HPGe detectors
- First Results from the DEAP-3600 Dark Matter Search with Argon at SNOLAB