Experiments that search for extremely rare decays often record particle tracks in a gas-filled chamber whose readout plane is cut into thousands of thin strips. Every detector loses some of these strips over a long run, and a dead strip leaves a gap in the recorded track. That gap does two kinds of damage. It can break one continuous track into two pieces, so a genuine signal looks like background and is thrown away. It can also delete ionization charge, so the event's measured energy falls below its true value and may leave the narrow energy window where the signal is expected. We propose to repair these gaps using a physical fact that generic inpainting ignores. In a two-dimensional strip readout, each ionization cloud is recorded twice, once on the X strips and once on the Y strips, at the same drift time. The charge missing from a dead X strip is therefore still present, in total, in the Y projection at that moment. A small graph network then only has to decide how that known amount of charge is spread across the dead strips. It never has to decide how much charge there was.
The practical stakes are concrete. The PandaX-III design shows that four track-shape quantities can suppress gamma backgrounds by about seven orders of magnitude while keeping signal efficiency above 20% (Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment). Those quantities are the fraction of energy in the main track, the charge in the dense "blobs" where stopping electrons end their paths, the scattering twist near track ends, and the track length. That result assumes every strip works. Sensitivity in these searches grows with exposure, and backgrounds must stay below about one count per tonne-year (The MAJORANA Project). Losing signal to readout gaps, or inventing blob charge in background events, would erode exactly the margin the design depends on.
Core Claim
Our claim is that dead-strip repair for topological discrimination should treat the conjugate projection as a hard constraint on the total charge, and use learned graph propagation only to distribute that charge. Existing graph imputation methods reconstruct missing values by smoothing over neighbors, and smoothing is precisely what flattens a blob. Under the constraint, the repair cannot remove blob charge from an event. It also cannot add charge that the other projection did not record.
Formally, let $q_X(x,t)$ and $q_Y(y,t)$ be the charges on X strip $x$ and Y strip $y$ in drift-time bin $t$. Let $D$ be the set of dead X strips, and let $r$ be the calibrated ratio of charge shared between X and Y strips. The charge missing from the dead strips in bin $t$ is
$$\Delta(t) = r\sum_{y} q_Y(y,t) - \sum_{x \notin D} q_X(x,t).$$
We impute each dead strip as
$$\hat q_X(x,t) = \max(\Delta(t),0)\, w_x(t), \qquad \sum_{x\in D} w_x(t)=1,$$
where the weights $w_x(t)$ come from a softmax over a graph network's outputs. The network runs on a graph whose nodes are (strip, time-bin) cells, and its edges link neighboring strips and neighboring times. When a dead region exists in both projections at the same drift time, the constraint is only partly available. The network then falls back to ordinary neighborhood reconstruction, and it reports a per-event uncertainty on the repaired charge.
Relation to Prior Work
Several papers motivate the gap. PandaX-III (Topological background discrimination) defines the discriminators and shows that the 2D 3 mm strip readout matches a 10 mm pixel readout. Its simulations assume a fully live readout, which is the assumption we relax.
Two lines of graph imputation supply the propagation machinery but constrain its design. The first is spatiotemporal sensor imputation. GRIN (Filling the G_ap_s) places message passing inside a recurrent unit and can reconstruct entirely missing sensors, improving mean absolute error by 20–40% over non-graph baselines on traffic and air-quality data. We reuse its idea of predicting a masked node purely from its neighbors. The second is feature propagation on static graphs. The hypergraph method of Self-supervised Guided Hypergraph Feature Propagation minimizes Dirichlet energy, a penalty on differences between connected nodes. It keeps 79.41% Cora accuracy at 99% missing features, but it is a low-pass filter by construction.
Propagate and Inject documents the failure mode that worries us: at extreme missingness, propagated features collapse toward a low-variance mean. Its remedy injects random synthetic values. That remedy is unsafe here, because random values could manufacture a fake blob in a background event. Our constraint is a physically justified way to avoid the same collapse. The adversarial graph autoencoder of Missing data imputation with adversarially-trained graph convolutional networks matches the distribution of imputed values, but it has no notion of conservation.
Incomplete multi-view clustering transfers structure from a complete view to guide an incomplete one (Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding, High-Confident Local Structure Guided Consensus Graph Learning). Our use of the Y projection is the detector analogue of that idea. The difference is that the transfer here is an exact charge budget, not a learned affinity.
Particle-physics graph networks (Graph Neural Networks for Particle Tracking and Reconstruction) reach 95.9% edge efficiency on TrackML, but they assume every sensor that should fire did fire. The LEGEND cleaning pipeline (Machine learning-powered data cleaning for LEGEND) represents the rejection philosophy that is standard in this field: corrupted events are discarded, not repaired. We test whether repair can do better without adding background.
Research Program
We will build a simulation study in the REST-based PandaX-III setup. It will contain signal events and the 214Bi and 208Tl gamma backgrounds described in the PandaX-III paper. Each event will be overlaid with dead-strip masks drawn from three patterns: isolated random strips, contiguous banks such as a failed readout chip, and regions dead in both projections at once. We will sweep the dead fraction from small to severe.
The graph network is trained self-supervised, in the style of GRIN. We mask live strips in simulated events and train the network to predict the distribution weights $w_x(t)$ against the true charge profile. No physics labels are used, so the imputer never learns to favor signal. After repair, the unchanged PandaX-III reconstruction and cuts run on the output.
The same masking protocol applied to real calibration data, by deliberately hiding working strips, gives a check that transfers from simulation to a real detector. It uses no labels and no new hardware.
Baselines
The upper reference is the fully live readout. For it, PandaX-III reports gamma suppression of about seven orders of magnitude at signal efficiency above 20% (Topological background discrimination). We found no reported result on Lacuna for any dead-strip configuration. We will therefore regenerate the live-readout reference ourselves, with identical cuts, so that every method is compared against the same simulation.
The do-nothing reference zeroes dead strips and runs the standard chain. The conservative physics reference vetoes any event whose track crosses a dead strip. This option introduces no bias but pays in exposure.
The simple engineering baseline has two parts. It relaxes the hit-clustering distance so that one-strip gaps are bridged, and it fills dead strips by linear interpolation from their neighbors.
The generic graph imputer runs GRIN's spatial decoder and Dirichlet-energy feature propagation on the same (strip, time) graph, with the same parameter budget and hyperparameter search as our model. We also run a variant that adds variance injection in the style of Propagate and Inject.
The key ablation is constraint-only repair. It computes $\Delta(t)$ and spreads it uniformly or by linear interpolation, with no learning. It isolates how much of any gain comes from the charge budget and how much from the learned distribution.
The strongest competitor is an end-to-end graph classifier on the raw projections, trained with dead-strip augmentation and given an energy-regression head. It receives the same simulated data and the same compute budget.
Beating a baseline means one of two things. Either we get higher signal efficiency at matched background suppression, measured from high-statistics background samples, or we get a smaller energy bias and less spill of signal out of the energy window, with no increase in background leakage.
Why not just…?
Why not just veto events that touch dead strips? A veto is unbiased, but PandaX-III tracks extend over many strips, so even a modest dead fraction removes a large share of candidates. The signal efficiency starts only a little above 20%. The dead-fraction sweep will measure this exposure loss directly against repair.
Why not just bridge gaps in clustering and interpolate? This is a real threat for isolated single dead strips when the track is continuous. It restores track continuity, which is what the energy-fraction cut needs. It cannot recover charge lost when a blob or a track end falls on a dead bank. It also risks merging separate Compton scatter sites into one track. If this baseline matches ours on isolated-strip masks, we will report that. The case for our method then rests on contiguous banks, blob charge, and energy recovery.
Why not just train a robust classifier end to end? This may match us on classification alone. It does not, however, return repaired charge that the existing interpretable cuts and the energy-window analysis can use. It also offers no per-event audit of what was filled in. If its energy head matches our energy recovery as well, the honest conclusion is that imputation adds interpretability rather than sensitivity, and we would state that.
Why not just use the constraint without learning? For isolated dead strips this may capture most of the gain. The ablation exists to show this. We expect learning to matter where the deficit must be spread across a contiguous bank, and where both projections are dead at the same drift time.
Decisive Checks
The first check is background safety. For background events, the distribution of blob charge after repair must match the live-readout distribution. Leakage at the full cut chain must also match within statistical uncertainty. This requires the large background samples the PandaX-III analysis already uses, and any excess kills the method.
The second check is signal recovery. We will plot signal efficiency at fixed suppression against dead fraction for every baseline and every mask pattern. We will also measure the fraction of signal that stays inside the energy window.
The third check is sim-to-real transfer. On real data with working strips masked on purpose, the repaired charge should match the hidden truth with the error level seen in simulation.
The method succeeds if, on contiguous-bank masks, it recovers most of the gap between the do-nothing reference and the live-readout reference, and it does so without any measurable rise in background.
References
- Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment
- The MAJORANA Project
- Filling the G_ap_s: Multivariate Time Series Imputation by Graph Neural Networks
- Self-Supervised Guided Hypergraph Feature Propagation for Semi-Supervised Classification with Missing Node Features
- Propagate and Inject: Revisiting Propagation-Based Feature Imputation for Graphs with Partially Observed Features
- Missing data imputation with adversarially-trained graph convolutional networks
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding
- High-Confident Local Structure Guided Consensus Graph Learning For Incomplete Multi-view Clustering
- Graph Neural Networks for Particle Tracking and Reconstruction
- Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines