Non-Redundant Discovery of Leaking Backgrounds in Germanium Detector Waveforms

Experiments that search for neutrinoless double-beta decay must reject nearly every background event. This decay is a hypothetical radioactive process whose observation would show that neutrinos are their own antiparticles. Once the remaining background in the signal window drops below about one count per tonne-year, sensitivity grows linearly with exposure instead of with its square root (The MAJORANA Project). What blocks that regime is not the bulk of the background, which simple pulse-shape cuts already remove. The blocker is the few small populations of background events that look enough like signal to pass those cuts. Today physicists find these populations by hand. In the clearest recent example, analysts read post-hoc feature attributions from a boosted decision tree and from them devised a new cut on events with unusually high current amplitude (Interpretable boosted-decision-tree analysis for the Majorana Demonstrator).

We propose to automate that discovery step. The method takes waveforms that survive the standard cuts and reports a short sequence of compact, human-readable event populations, each stated as a simple rule on pulse features. After each report, the reference model of "what we already expect" is updated so that the next report must be genuinely new. The reference model has to absorb detector-to-detector and drift-in-time calibration differences. Otherwise the most "surprising" findings would simply be one crystal behaving slightly differently from another. To handle this, we estimate those calibration differences as hidden variables by expectation-maximization (EM), alternating between inferring the hidden quantities and refitting the parameters. The fit uses the known calibration peaks as anchors.

Core Claim

Subjective interestingness scores a pattern by how surprising it is relative to an explicit probability model of current knowledge, then folds each accepted pattern into that model. This approach removes redundancy in tile and cluster mining (Maximum entropy models and subjective interestingness: an application to tiles in binary databases; Subjectively Interesting Alternative Clusters). On multi-detector physics data, however, it fails in a predictable way: the dominant "surprises" are calibration nuisances rather than physics. Our claim is that coupling the maximum-entropy background (the least-committal distribution matching known statistics) to an EM estimate of per-detector, time-varying nuisance transforms turns the method into a usable background-discovery tool. Concretely, a small number of reported patterns should recover the leaking populations that physicists found by hand. When those patterns are converted into cuts, they should match or exceed supervised multi-site rejection without being trained on calibration labels as signal proxies.

The model for this is simple. Each event $i$ recorded on detector $d(i)$ at time $t_i$ has a feature vector $x_i$ that includes energy, A/E (peak current divided by energy, the standard multi-site discriminator), delayed charge recovery (the slope of the waveform's tail, which flags surface alphas), drift time, tail slope and current-peak count. We write

$$x_i = A_{d(i)}\, z_i + b_{d(i)}(t_i) + \varepsilon_i ,$$

where $z_i$ is the event's position in a shared, detector-independent feature space, $A_d$ is a per-detector distortion near the identity, $b_d(t)$ is an offset that drifts slowly over time, and $\varepsilon_i$ is noise. The maximum-entropy background lives on $z$. It is constrained by global moments and by every pattern accepted so far. A candidate pattern is a subset $S$ selected by a short conjunctive rule on $z$, and it is scored as

$$\mathrm{score}(S) = \frac{-\log P_{\text{bg}}(\text{statistics of } S)}{L(S)} ,$$

that is, its self-information under the current background divided by its description length $L(S)$ (the number of rule terms), as in De Bie's tile framework.

Research Program

The work has three coupled pieces.

The first piece is the nuisance estimator. The offsets $b_d(t)$ follow a random walk, so the E-step is a Rauch–Tung–Striebel smoother over each detector's event stream, and the M-step updates $A_d$ and the noise. This is the machinery of Learning linear state-space models with sparse system matrices, which uses Student's-t priors to break the rotation ambiguity of hidden states and reports 95% topology recovery against 0% for classical methods. We reuse its prior to keep each $A_d - I$ sparse, so only a few features distort per detector. The essential identifiability device is anchoring. Events from the thorium-228 double-escape peak (DEP, mostly single-site) and full-energy peaks are physically the same populations on every detector, so their means and covariances in $z$ are tied across detectors. Physics differences in how many events each detector sees cannot then be absorbed as calibration.

The second piece is the background and search loop. The background is a Gaussian maximum-entropy model on $z$, constrained by its accepted patterns. This extends the alternative-clusters construction, which used spectral relaxation to find clusters; we instead use rule search over interpretable features so that every output is directly a candidate cut. After a pattern is accepted, the background is re-projected, and one or two EM sweeps re-estimate the nuisances under the new constraints. This prevents a real population from being absorbed into a detector's offset.

The third piece is the stopping rule. We stop when the score of the best remaining pattern falls to the level achieved on events drawn from the fitted background itself. This plays the role of the sharp drop in information gain seen in the alternative-clusters work.

All development uses the Majorana Demonstrator Data Release for AI/ML Applications: about 3.19 million thorium-228 calibration waveforms with energy and four pulse-shape labels including A/E and delayed charge recovery, split 75/20/5 with a blind set. The file does not say whether detector identifiers and timestamps are included. If they are absent, detector membership becomes an additional latent variable in the E-step, a mixture over calibration states. That variant is harder, and we will report it separately rather than hide it.

Baselines

The naive reference point is the standard physics cuts alone, A/E and delayed charge recovery (Multisite event discrimination for the majorana demonstrator), with no discovery step. The data release reports that roughly 1.2 million of 3.2 million events pass them. Anything we add must reduce leakage beyond these cuts while keeping DEP acceptance at the same level, nominally 90%.

The strongest supervised competitor is the multi-site boosted decision tree, which the file reports as rejecting 8.6% more single-escape-peak events than the standard A/E cut, with 164 surviving events in the background estimation window against 153 for the standard analysis (Interpretable boosted-decision-tree analysis for the Majorana Demonstrator). The convolutional-autoencoder classifier is a second supervised reference, reported at 90% background rejection and 90% signal acceptance, matching A/E (Deep learning based pulse shape discrimination for germanium detectors). We will retrain a LightGBM tree on the same released splits and features with an equal hyperparameter budget. To beat it, the cuts derived from our reported patterns must reject at least as many single-escape-peak events at matched DEP acceptance, and they must do so using no more than five or so rule terms. A tie in rejection, achieved without signal-proxy labels and with readable cuts, is itself a meaningful result. A loss of more than a few percent would mean discovery does not substitute for supervision.

The strongest unsupervised competitor for the discovery task itself is DEMUD. DEMUD repeatedly picks the point least explained by a subspace of already-seen data and then adds it to that subspace, and it is the redundancy-aware ranker in Domain-Agnostic Outlier Ranking Algorithms. Alongside it we run Isolation Forest and PCA reconstruction error, which DORA reports as the best methods on planetary images (precision 0.56) and galaxy data (0.42) respectively. All of these run on the identical feature matrix of post-cut survivors.

Two ablations isolate our key ingredients. The first removes the nuisance model entirely: we run maximum-entropy search on raw $x$. The second replaces EM with per-detector z-scoring of features. For both, no reported result found on Lacuna; they are new measurements.

Why not just…?

Why not just run DEMUD on the survivors? This is the most serious threat, because DEMUD already deflates iteratively. It differs from our method in three ways. It ranks individual events rather than populations, it has no notion of detector nuisance, and it has no principled stopping point. Whether these differences matter is an empirical question. The planted-population experiment below scores both methods on how many distinct planted populations appear in the first $K$ reports. If DEMUD on nuisance-corrected features matches us, we will report that the EM nuisance model is the real contribution and the maximum-entropy scoring is not.

Why not just standardize features per detector and cluster? Per-detector standardization is cheap, but it erases real composition differences. A detector closer to the thorium source sees proportionally more multi-site events, and standardizing its features shifts its genuine physics toward the array average. Anchoring on calibration peaks is designed to avoid exactly this. The z-scoring ablation tests the argument directly, with planted populations injected at unequal rates across detectors. If z-scoring recovers them equally well, the EM component is unnecessary and we will drop it.

Why not just train a better supervised classifier? Supervised training on DEP and single-escape-peak proxies learns to separate known classes. Calibration events, however, are spatially biased relative to bulk signal events, and a population that is absent or mislabeled in the proxies cannot be learned. The tree analysis already showed that the leaking high-A/E population had to be found by reading the model afterward (Interpretable boosted-decision-tree analysis for the Majorana Demonstrator). Our method targets that discovery step and is complementary to classification. The comparison above checks that its outputs are at least competitive as cuts.

Proposed Experiments

The first decisive check uses planted populations with known ground truth. We construct synthetic multi-site events by summing two time-shifted, energy-scaled DEP waveforms from the release. We keep those that pass the standard A/E cut, which typically means sub-millimetre-to-few-millimetre separations near the point contact. We then inject three such families at different rates into different detectors. On the actual waveforms, we also apply per-detector gain and rise-time drifts of realistic size. The metrics are the number of planted families recovered among the first $K$ reports, the mean Jaccard overlap between reported patterns as a measure of redundancy, and the fraction of reports that are pure calibration artifacts. We expect the no-nuisance ablation to spend its first reports on the drifts, and the full method to report the planted families.

The second check is rediscovery on real data. Without using the tree's attributions, does the loop report a high-A/E population with drift-time dependence comparable to the one the boosted-decision-tree analysis found by hand, and at what rank?

The third check is physics utility. We turn the top patterns into cuts and measure single-escape-peak rejection at fixed DEP acceptance against the retrained tree. We also measure the energy-histogram intersection-over-union metric defined by the data release on the blind 5% split; no reported IoU value for a baseline was found on Lacuna. Finally, we check that the stopping rule halts at a similar number of patterns across random restarts and data subsamples, since an unstable count would mean the reported populations are not trustworthy.