When Detector Faults Cluster: Correlation-Aware Nuisance Models for Anomaly Searches and Unfolding

Physics analyses usually treat detector inefficiency as a rate: some fraction of channels is dead, miscalibrated, or noisy. That rate is propagated as a systematic uncertainty, meaning an error that more data will not shrink. Hardware reliability work shows that the rate is not enough. When the same number of faults bunch together in space, because they share a controller, a bonding region, or a thermal hotspot, the system behaves much worse than when the faults are scattered. We propose to carry this lesson into precision inference. Our claim is that the systematic error a detector fault induces in a machine-learned analysis depends on how the faults are arranged in space, not only on how many there are. Standard nuisance models assume independent channels, and so they will under-report uncertainty. In the most sensitive analyses they will also create spurious "discoveries."

We test this on two analyses where the stakes are high and public benchmarks exist. The first is a model-independent new-physics search built on contrastive embeddings. The second is machine-learned unfolding with frequentist error bars. We then build and test one remedy. Nuisance variations are drawn from a spatially clustered fault prior borrowed from hardware reliability models, and these draws are used both to train the background model and to span the uncertainty basis.

Core Claim

Consider $N$ readout channels. Each fails with probability $p$, and failures come in clusters of typical size $m$ with within-cluster correlation $\rho$. The variance of the number of failed channels in any analysis region is the standard design-effect expression:

$$\mathrm{Var}(K) = N p (1-p)\,[1 + (m-1)\rho].$$

It follows from summing correlated Bernoulli variables. An independent-channel nuisance model sets $\rho = 0$ and so understates this variance by the factor $1+(m-1)\rho$. That factor can be large, since a shared front-end controller can take out dozens of adjacent cells together.

For a density-sensitive anomaly test, the problem is worse than extra variance. A contiguous dead or biased patch in the angular coordinates of the detector carves a localized, coherent hole or bump into the event distribution. That is exactly the kind of structure such tests are built to flag. We therefore predict two things. First, at matched fault rate, clustered faults will produce spurious significance and interval under-coverage much larger than scattered faults. Second, a background model augmented with independent random dropout will not remove the effect, while one augmented with a clustered prior will. The second prediction is the part that would be new. If independent dropout suffices, the spatial-correlation ingredient is unnecessary, and the proposal fails on its own terms.

Grounding in Prior Work

Two hardware studies motivate the claim and supply the fault models. Cost-Effective TSV Grouping for Yield Improvement of 3D-ICs builds yield models for vertical interconnects in stacked chips under independent versus clustered defects, controlled by a clustering factor $\alpha$. Under strong clustering ($\alpha=2$, $p=0.01$), naive one-to-one redundancy reaches only 7% yield. We reuse its clustering model as a generative prior over fault maps.

Error Analysis of Programmable Metasurfaces for Beam Steering studies arrays of tunable reflective cells. It separates fault types (stuck-at, out-of-state, deterministic, biased) from spatial arrangements (independent, clustered, aligned) and finds that deterministic clustered faults are far more damaging than biased ones, redirecting the beam to specular reflection at about 33% error. Fault Tolerance in Programmable Metasurfaces: The Beam Steering Case traces clustering to failures of the shared controllers that drive blocks of cells. We adopt this two-axis taxonomy for detectors. Faults that silently shift a region's response are the physics analogue of "deterministic" errors, and they are the ones a simple dead-channel mask cannot see. Reduce retrains neural networks around permanent accelerator faults. It acknowledges geographic clustering but evaluates only uniform random injection. This shows how entrenched the independence default is, even among researchers who know it is wrong.

On the inference side, Solving Simulation Systematics in and with AI/ML documents that networks trained on simulation learn unmodeled detector quirks. The hardware effects it treats are global parameters or diffuse noise, not discrete clustered dropouts, so it motivates our concern but does not address our case. Anomaly-preserving contrastive neural embeddings compresses collider events into low-dimensional contrastive embeddings and feeds them to the New Physics Learning Machine goodness-of-fit test. It reaches 12.1σ on a blind black-box dataset, against 1.4σ for a variational autoencoder and 0σ on raw features. That sensitivity is precisely why we expect clustered faults to be dangerous. The paper does not study instrumental artifacts.

OmniFold and Moment Unfolding remove detector distortion without binning. They assume smooth detector response, and OmniFold omits systematic propagation from its demonstration. Frequentist Uncertainties on Neural Density Ratios with wifi Ensembles fits linear weights over frozen networks, which act as basis functions. It propagates the resulting covariance to physics parameters and reaches nominal 68% coverage at ensemble size 16 in a Gaussian toy. Its stated weakness, bias when the true ratio lies outside the span of the basis, is where our method plugs in. Finally, Maximum-likelihood Regression with Systematic Errors … II models systematics as a scalar intrinsic variance and shows that a quasar absorption feature falls from $p<0.01$ to $p=0.181$ at 10% systematic error. We use it as a reminder that ignoring structured systematics inflates significance, and we contrast our spatially structured prior with its unstructured scalar treatment.

Research Program

We first build a fault emulator that acts on the angular grid of pseudorapidity and azimuth, the standard detector coordinates. Fault maps are drawn from three priors at matched mean rate: independent cells, Zhao-style clusters with tunable $\alpha$, and block failures that mimic a shared controller. Each faulted region is assigned one of three behaviors: dead (hits removed), biased (energy scaled by a drawn factor), or stuck (energy replaced by a fixed value). For the unfolding benchmark we inject faults at the calorimeter and tracking stage of the Delphes simulation, so they act on jet constituents. For the black-box anomaly benchmark we apply them at object level, deleting or rescaling physics objects whose direction falls in a faulted region.

Second, we measure the damage. We train the contrastive encoder and the goodness-of-fit reference on pristine simulation. We then evaluate on background-only pseudo-data carrying each fault prior, and on signal-injected pseudo-data, sweeping fault rate and cluster size. For unfolding, we run OmniFold and Moment Unfolding on faulted pseudo-data and record the bias of the unfolded moments.

Third, we build the remedy, a correlation-aware nuisance model with two parts. The contrastive encoder is trained with augmentations drawn from the clustered fault prior, so that plausible fault patterns become invariances rather than features. For uncertainty, we train each member of a wifi ensemble under a different sampled fault map. The span of the basis then includes fault-induced distortions, and the sandwich covariance absorbs them instead of leaving them as unquantified misspecification. The same fault-map ensemble supplies nuisance variations for the anomaly test.

Baselines

The naive reference point is to do nothing. We train the Metzger et al. supervised SimCLR Transformer with the goodness-of-fit test, and OmniFold and Moment Unfolding, on pristine simulation, and apply them unchanged to faulted pseudo-data. The ideal is zero spurious significance on background-only data and 68% coverage for 68% intervals. The size of the departure defines the problem.

The strongest existing detector for the anomaly task is that same contrastive pipeline, with its reported 12.1σ black-box significance as the published reference. We will rerun it ourselves on identical data splits and compute budget, and report our own reproduced figure alongside it. For controlling instrumental false positives under clustered faults, we found no reported result on Lacuna, so the baselines below are ones we construct.

The practitioner's standard fix is masking: identify regions with anomalously low occupancy and exclude them from both reference and data. The conventional systematic treatment is a single smooth global efficiency nuisance, analogous to a scalar intrinsic variance. For unfolding uncertainty, the strongest method is a wifi ensemble with a pristine basis, which the file reports reaching nominal coverage in its toy setting.

The key ablation replaces our clustered-prior augmentation and basis with independent random dropout at the identical mean rate. Every method gets the same number of training runs, the same encoder architecture and the same hyperparameter search budget.

Beating the baselines means three things. On background-only data carrying held-out clustered and block fault maps, our method must keep spurious global significance below 2σ, whereas the naive and independent-dropout versions exceed it. It must retain at least 80% of the reproduced signal significance on signal-injected data. And its unfolded-moment intervals must cover within a few percent of 68% where the pristine-basis wifi ensemble under-covers.

Why not just…?

Why not just mask the dead channels? Masking handles faults that show up as silence and stay fixed. The metasurface results show that the most damaging clustered failures are not silent: deterministic and biased cells keep responding with the wrong value. A detector block with a drifting gain looks alive. We include masking as a baseline and expect it to succeed on dead faults and fail on biased or stuck ones. If it also succeeds on those, the remedy is unnecessary for static faults, and we will say so.

Why not just augment with random dropout? This is the real threat, because it is cheap and widely used. The design-effect equation predicts that independent augmentation underestimates variance by $1+(m-1)\rho$. It also never shows the network a coherent regional hole, so it cannot teach invariance to one. The matched-rate ablation settles the question directly. If independent dropout controls false positives as well as the clustered prior, our central claim is refuted.

Why not use data-driven translation or sideband ratios, as in Viren et al. and Daya Bay? Image translation fits simulation to data, but Viren et al. concede it introduces second-order systematics without an error budget. Daya Bay cancels detector effects with near/far ratios between identical detectors. That works when the fault affects the signal and control regions equally. A spatial patch that correlates with event kinematics breaks this equality. It is a genuine alternative when that correlation is weak, so we will run a sideband-ratio variant and report the regime in which it matches our method.

Decisive Checks

Three results would confirm the claim. At matched fault rate, clustered and block faults must inflate spurious significance and under-coverage far more than scattered faults, with the gap growing with cluster size roughly as the design-effect factor predicts. Independent-dropout augmentation must fail to close that gap while clustered-prior augmentation closes it. The fault-spanned wifi ensemble must restore nominal coverage on moments unfolded from faulted pseudo-data without losing substantial signal sensitivity.

A null result on the first check would itself matter to experiments. It would show that fault rate alone is a sufficient systematic descriptor for these methods.