Searches for very rare physics processes, such as neutrinoless double-beta decay or dark matter, separate a handful of signal events from backgrounds that are millions of times more common. They do this with classifiers built on a fixed list of inputs: five hand-made track-shape variables, a fixed-length compressed waveform, or a fixed set of kinematic quantities. Real detectors do not hold still. Readout strips die, veto panels are added, and readout pitch changes between prototype and final design. Each such change breaks the classifier's input format. The collaboration then has to retrain and, more expensively, re-validate that the new classifier has not created a fake bump in the energy spectrum.
We propose a discriminator that reads an event as an unordered set of readout tokens. Each token carries its measured value and a physical description of the sensor that produced it, namely position, pitch, readout type, and calibration, rather than a column index or a column name. A learned per-token weight decides how much each sensor contributes, and those weights renormalize automatically over whichever sensors are actually present. The weights are trained under an explicit penalty that keeps the final score statistically independent of event energy on background. This targets the failure that makes physicists distrust adaptive models: an instance-dependent weighting can quietly learn to favor background events at a particular energy and carve a false peak into the spectrum, a problem usually called sculpting.
Core Claim
Two things can hold at the same time for a set-based discriminator. First, if its tokens are described by physical metadata instead of position in a vector, it can be trained on one readout configuration and applied with no retraining to degraded or modified configurations, while keeping most of the background rejection of a classifier retrained specifically for each configuration. Second, if its adaptive per-sensor weights are constrained to carry no information about energy on background, it achieves this without sculpting the background energy spectrum in the region of interest.
The first property is new to rare-event physics. The second is new to adaptive feature fusion, where weights are normally left free to amplify any combination that lowers the training loss. The proposal stands or falls on whether both survive against a fixed-length masked network trained with random channel dropout, which is the simple competitor most likely to match it.
Where This Sits
The three directions play different roles here. Signal-to-background discrimination supplies the application and the success criterion. The fixed-dimension input problem supplies the stress test. Adaptive weighting supplies the mechanism, and it is also where the risk lies.
On the physics side, the motivating evidence comes from three analyses. Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment uses five sequential cuts on reconstructed xenon gas tracks. These are track energy ratio, blob charge at the track ends, end-track twist, track length, and pulse diffusion width. Together they reach roughly seven orders of magnitude of gamma suppression at above 20% signal efficiency. Those authors also compare a 2D 3 mm strip readout with a 3D pixel readout and say multivariate methods are the natural next step. Machine learning-powered data cleaning for LEGEND compresses germanium detector waveforms to a fixed wavelet vector before clustering and classifying them with a support vector machine. CATHODE, applied in Isolating Unisolated Upsilons with Anomaly Detection in CMS Open Data, builds a background template with a density model on fixed kinematic vectors and raises a dimuon signal from 1.6σ to 6.4σ. All three are tied to one input layout. We keep their physics observables as reference points but change how inputs enter the model.
The same CMS analysis also constrains the design. Its reliance on sidebands orthogonal to the resonant mass, and on a same-sign control channel that shows no signal, is exactly the kind of check that a free adaptive weighting could fail. We adopt that check as a pass/fail test rather than treating it as optional.
On the input-format side, two lines of work supply the method. TransTab shows that representing each feature as a self-describing token lets one model ingest changing column sets. It uses a gated transformer and contrastive training over column subsets. LBC does the same through language prompts to a large language model. Both depend on human-readable column names, and both lose their advantage when names are opaque. Detector channels are opaque by nature: "strip 124" means nothing to a language model.
Our departure is to describe a sensor by its physics instead of by a name. Its position and pitch tell the model where it sits and how fine it is. Its readout type and gain tell the model how to read its value. Multimodal Inference with Incremental Tabular Attributes adds new attributes to frozen encoders through a disentangled encoder. However, it only handles additions, and it assumes moderate class balance. Our setting is dominated by removals (dead channels) and by background-to-signal ratios of $10^5$ or worse.
Benchmark evidence motivates why the shift matters. TabFSBench finds that removing 20% of the most relevant features catastrophically degrades boosted trees and tabular deep networks even with mean imputation. Realistic Evaluation of TabPFN v2.5 in Open Environments notes that even a strong tabular foundation model has a fixed input width.
On the fusion side, FFA-Net and ComBo show the value of data-dependent weighting over static summing or concatenation. FFA-Net does this with channel and pixel attention for image dehazing. ComBo does it with a small transformer over tokens from several frozen vision backbones, plus a sparsity penalty that prunes backbones. Both need aligned grids, and neither constrains what the weights may depend on. We reuse their gating idea on sets of arbitrary size and add the energy-independence constraint.
Research Program
An event becomes a set $S$ of tokens, one for each active readout element. Each token concatenates three things: the element's measured quantity, a fixed descriptor of the element, and a few local features. The measured quantity is charge, pulse width, or arrival time, expressed as a fraction of the event total so that absolute energy scale is not directly visible. The descriptor holds 3D position, pitch, readout type, and gain. The local features are computed from the element's neighbors in the event.
A small attention encoder passes information among tokens. A pooling step then forms the event representation:
$$z = \sum_{i \in S} \frac{\exp\big(g(t_i)\big)}{\sum_{j \in S} \exp\big(g(t_j)\big)}\, \phi(t_i),$$
Here $t_i$ is the encoded token, $g$ is a learned scalar gate, and $\phi$ is a learned value map. Because the normalization runs only over sensors that are present, a dead strip simply vanishes from the sum, and an added veto panel simply joins it. No padding is needed.
A classifier head maps $z$ to a score $s$. Training minimizes
$$\mathcal{L} = \mathrm{BCE}(s, y) + \lambda\, \mathrm{dCor}^2\big(s, E \mid y = \text{background}\big),$$
The first term is cross-entropy on signal label $y$. The second is the squared distance correlation, a dependence measure that is zero only under independence, between score and reconstructed energy $E$, computed on background events in each batch. The weight $\lambda$ trades rejection against flatness.
Using energy fractions makes energy harder to read. The penalty is what actually enforces independence, since total charge can still be inferred from track length and token count.
Training uses PandaX-III-style simulated $0\nu\beta\beta$ and gamma background events in the REST framework at the full 3D pixel configuration. During training, random subsets of tokens are removed. We will also study a variant that freezes the encoder and fine-tunes only the gate $g$ on a few hundred events from a new configuration. This cheap adaptation can be compared against both zero-shot transfer and full retraining.
Baselines
The naive reference point is the published REST sequential cuts, applied unchanged to degraded readouts with their nominal thresholds. Physicists would actually do this before anyone retrains. Its value on the nominal configuration is reported: approximately seven orders of magnitude of background reduction at signal efficiency above 20%, with the diffusion cut alone reducing surface uranium background to a factor of 0.15 at 75% acceptance (PandaX-III topology paper).
The strongest existing competitor is a boosted decision tree on the five REST observables, retrained separately for every test configuration on that configuration's own simulation. This oracle captures the true workflow "just re-simulate and retrain." Because simulation is cheap relative to detector time, it is the bar that matters.
It gets the same simulated events, the same signal and background counts, and the same hyperparameter search budget as our model (equal tuning trials, measured in GPU or CPU hours). The PandaX-III paper reports only cut-based results. We found no reported multivariate result on Lacuna for this suite, so the oracle's numbers will be produced by us.
The fixed-width competitor is a network over the full maximum-size readout image. Missing elements are set to zero and paired with an explicit presence mask, and the network is trained with the same random channel removal as our model and given matched parameter count.
Two ablations isolate the key ingredients. The first replaces the adaptive gate with uniform mean pooling over present tokens, a plain Deep Sets-style model. The second removes the energy-independence penalty by setting $\lambda = 0$. A third ablation swaps physical descriptors for learned per-channel identifiers, which tests whether the physics description, rather than set structure alone, carries transfer to unseen pitches.
We will declare a win on the deployment question if two conditions hold together. Under dead-strip and pitch-change shifts, our zero-shot model must reach at least 90% of the oracle's background rejection at the oracle's signal efficiency. It must also beat the masked fixed-width network by more than seed-to-seed spread. On the physics question, the background energy spectrum after selection must be statistically indistinguishable from the pre-selection shape across the region of interest, while the $\lambda = 0$ ablation must not pass that test.
Why not just…?
Why not just re-simulate and retrain a classifier for each configuration? This is the most serious alternative, and we treat it as the oracle rather than dismissing it. Our argument is practical. The number of dead-channel patterns grows combinatorially and changes run by run. Each retrained classifier also needs its own blinded sculpting and control-region validation, which costs far more than the training itself. The experiment will show how close zero-shot transfer and gate-only fine-tuning come to the oracle. If the gap is large, the honest conclusion is that the set model is a fast interim tool and not a replacement.
Why not just zero-fill missing channels, add a mask, and train with random dropout? For dead channels alone, this may work as well. TabFSBench's negative result concerns imputation without dropout training, so it does not settle the question. The masked fixed-width network is therefore a full baseline. Its structural limit is that it cannot accept sensors it was never sized for. These include a new veto panel, a different pitch, or the 2D-strip versus 3D-pixel change. Our incremental and pitch-change tests target exactly those cases. If it matches us on removals, we will narrow the claim to additions and geometry changes.
Why not just pool all present sensors equally? Uniform pooling also handles variable sets, and it cannot sculpt through instance-dependent weights. However, PandaX-III's blob-charge observable shows that the discriminating information sits in a few endpoint elements, which equal pooling dilutes. FFA-Net and ComBo report large gains from adaptive over static combination in vision, though not in physics. The uniform-pooling ablation measures whether the gate earns its sculpting risk. If it does not, we drop the gate.
Proposed Experiments
The primary testbed is the PandaX-III REST simulation suite. We train on the 3D pixel readout and test on six shifts. The first is the 2D 3 mm strip readout from the original paper. The next three are random dead-strip fractions of 5%, 10%, and 20%, where the 20% level mirrors the TabFSBench threshold. The fifth is contiguous dead regions simulating a failed readout module. The sixth is coarser and finer pitches that never appear in training. An incremental test adds a simulated scintillation-light channel that is absent at training time.
Metrics are background rejection at fixed signal efficiency and signal efficiency at the cut-based rejection. Background counts are converted to projected counts per year in the region of interest using the paper's normalization. Sculpting is measured as a chi-square and a Kolmogorov distance between pre- and post-selection background energy spectra in the region of interest, using statistics at least ten times larger than the expected signal sample.
A secondary sculpting stress test uses the 2016 CMS Open Data DoubleMu dataset. CATHODE's boosted-tree stage is replaced by our gated classifier, with and without the penalty conditioned on dimuon mass. The density-model stage is kept identical, and the published 6.4σ significance serves as the reference. The decisive check is the same-sign dimuon control channel, where CATHODE reports no significant signal. Any fake peak there from the $\lambda = 0$ variant, alongside its absence with the penalty, would be direct evidence that the constraint is necessary.
The idea succeeds if four results appear together on the primary testbed. The constrained gated model must stay within 10% of the oracle's rejection on unseen readouts. It must beat the masked fixed-width network on pitch changes and additions. It must beat uniform pooling on the nominal configuration. And it must show a flat background energy spectrum where the unconstrained version does not.
References
- Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment
- Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines
- Isolating Unisolated Upsilons with Anomaly Detection in CMS Open Data
- TransTab: Learning Transferable Tabular Transformers Across Tables
- LBC: Language-Based-Classifier for Out-Of-Variable Generalization
- Multimodal Inference with Incremental Tabular Attributes
- TabFSBench: Tabular Benchmark for Feature Shifts in Open Environments
- Realistic Evaluation of TabPFN v2.5 in Open Environments
- FFA-Net: Feature Fusion Attention Network for Single Image Dehazing
- Fantastic Features and Where to Find Them: A Probing Method to combine Features from Multiple Foundation Models