Where Does Pulse-Shape Discrimination Break? Separating Shape from Position in Germanium Detectors

Germanium detectors with a tiny "point" electrode are among the most sensitive instruments for rare-decay and dark-matter searches. They are sensitive because the small contact has very little capacitance, so electronic noise stays low. Their signal is useful because the shape of each electrical pulse reveals whether the energy was deposited at one spot (a single-site event, like the hoped-for neutrinoless double-beta decay) or scattered across several spots (a multi-site event, typical of gamma-ray background). Classifying these pulse shapes is well developed. A one-number cut, the ratio of peak current to energy ($A/E$), and small neural networks both reject about 90% of background while keeping about 90% of signal-like events. The weak point is elsewhere. Every such classifier is tuned on calibration events from an external thorium source, and those events sit in different parts of the crystal than real bulk decays would. In the MAJORANA Demonstrator this mismatch leaves a 3.5% systematic uncertainty on signal acceptance that no one has removed (Multisite event discrimination for the majorana demonstrator).

We propose to find out where in a classifier's decision this calibration-to-bulk gap lives, and then to remove it. Split each pulse into two physically defined views. The first is the shape of the current pulse, aligned at its peak, which carries the event's spatial pattern (one site or several). The second is the position information, mainly drift time and rise time, which says where in the crystal the charge started. Sample-level partial information decomposition (a way to split what two inputs jointly tell you about a label into parts that are shared, unique to one input, or available only when both are seen together) then shows how much of each event's classification depends on shape and position jointly. Our hypothesis is that the acceptance gap is carried almost entirely by this joint, shape-times-position information. Calibration events are spatially biased, so classifiers learn position-dependent thresholds that do not carry over to uniform bulk decays. If the hypothesis holds, the fix is structural. Build the classifier as separated specialists: a shape specialist that cannot see position, and a checking module that rejects decisions which only make sense given position. This borrows the "context firewall" and "judge" ideas from multi-agent task decomposition, applied here to neural modules rather than to language models.

Core Claim

The claim is testable and specific. For a pulse-shape classifier $f$ trained on calibration data, define the acceptance gap $\Delta\varepsilon = \varepsilon_{\text{cal}} - \varepsilon_{\text{bulk}}$. Here $\varepsilon_{\text{cal}}$ is the fraction of calibration single-site events (the 1592 keV double-escape peak) the classifier keeps, and $\varepsilon_{\text{bulk}}$ is the fraction of uniformly distributed bulk single-site events it keeps. For each event, the information the two views $S_{\text{shape}}$ and $S_{\text{pos}}$ carry about the topology label $Y$ splits as

$$I(S_{\text{shape}}, S_{\text{pos}}; Y) = R + U_{\text{shape}} + U_{\text{pos}} + Syn,$$

where $R$ is information both views carry, each $U$ is information only one view carries, and $Syn$ is information present only when both are seen together. We predict three things. First, events with high per-event $Syn$ account for most of $\Delta\varepsilon$. Second, a classifier forbidden from using $Syn$ (and $U_{\text{pos}}$) keeps nearly all its background rejection, because the physics of a single-site event is a property of pulse shape and not of position. Third, the residual $\Delta\varepsilon$ of such a classifier falls well below the 3.5% systematic now attributed to the calibration mismatch.

This is different from the closest existing work. Deep learning based pulse shape discrimination for germanium detectors trains an autoencoder plus a 141-parameter classifier on raw 256-sample waveforms. It matches $A/E$ (90% background rejection at 90% signal acceptance) and catches multi-site events that $A/E$ misses. It treats the waveform as one undivided input and does not ask whether what it learns transfers from calibration to bulk. Efficient Quantification of Multimodal Interaction at Sample Level supplies the measurement tool, called LSMI. LSMI estimates shared, unique and joint information per sample, quickly enough for large datasets: about 11 minutes on UCF-101 versus over 13 hours for batch estimators. It also reports that models lean on joint-information samples under distribution shift. That work covers video, audio and text with two inputs and has never been used to diagnose a physics systematic. We use LSMI unchanged as a measuring instrument. Our two inputs are physically defined views of a single detector signal, which fits its two-source limitation rather than fighting it.

Why the literature points both ways, and why that matters

Two cited results pull in opposite directions, and this proposal is built to settle which one applies here. LSMI reports that joint ("synergistic") information is what models rely on to generalize out of distribution, which would suggest keeping it. Synergistic information supports modality integration and flexible learning in neural networks solving multiple tasks uses the same kind of decomposition on trained networks. It finds that joint-information units are fragile hubs whose removal collapses performance, and that noise during training pushes networks toward redundant backup pathways. In germanium pulse-shape discrimination the "distribution shift" is a known, physical change in where events sit in the crystal. The topology label should not depend on position at all. Joint shape-and-position information is therefore, by physics, information that should not be needed. Our prediction sides with the fragility result for this specific kind of shift, and the experiments can falsify it. The broader lesson concerns when joint information helps transfer and when it signals a shortcut. That lesson would transfer outside physics.

Research Program

We first build a labeled testbed in simulation. This addresses the core obstacle that true neutrinoless double-beta decays have never been observed, so no real signal labels exist (Deep learning based pulse shape discrimination for germanium detectors). We use Geant4 energy depositions from (i) an external thorium-228 source and (ii) uniform bulk single-site and multi-site decays. These are propagated through a weighting-potential pulse model for a p-type point-contact crystal (Ramo–Shockley induced current, steeply rising near the contact) with realistic preamplifier noise. Each event carries true position and true topology. We also include the effects known to distort pulses near the outer surface and over time: slow, energy-degraded pulses from the lithium-diffused transition layer, with full-charge-collection depths drawn around the measured 0.77 ± 0.12 mm and 0.933 ± 0.094 mm (Characteristics of Signals Originating Near the Lithium-Diffused N+ Contact of High Purity Germanium P-Type Point Contact Detectors). We also include charge trapping of the kind corrected empirically in Charge trapping correction and energy performance of the Majorana Demonstrator.

Second, we define the two views carefully, because the result depends on them. The position view uses drift time, the 10–90% rise time and the tail slope. The shape view is the current pulse cropped to a fixed window around its peak and normalized in amplitude. In simulation we verify that position cannot be regressed from the shape view beyond a small, measured floor. Because physics does not separate the two views perfectly (pulses near the contact are slightly sharper), we report that floor rather than assume it is zero.

Third, we train a set of classifiers on calibration-peak labels only, as real experiments must. These are $A/E$, the autoencoder classifier, and an unrestricted convolutional network seeing the full waveform. For each one, we run LSMI on its learned representation to get per-event $R$, $U$ and $Syn$. We then test whether $Syn$ predicts which bulk events are wrongly accepted or rejected relative to calibration.

Fourth, we build the firewalled design. A shape specialist sees only the shape view. A position specialist sees only the position view and is used solely as a judge. During training, the judge is trained to predict the shape specialist's errors from position, and the shape specialist is penalized whenever the judge succeeds. This is an adversarial independence constraint, the neural counterpart of the independent tester that catches the coder's blind spots in Self-collaboration Code Generation via ChatGPT. At inference time, events whose LSMI profile still shows high joint information are flagged for review rather than silently classified.

Fifth, as a bounded and clearly secondary test, we try a language-model agent loop on the review step only. When flagged events cluster, a planner proposes physical explanations (a thicker transition layer, a trapping gradient). An executor runs pulse-simulation sweeps. A judge compares simulated and observed distributions. This mirrors the Thinker–Judge–Executor split of MACM: Utilizing a Multi-Agent System for Condition Mining in Solving Complex Mathematical Problems. MACM itself reports weak spatial reasoning, so the planner never reasons about geometry; all field and position computation goes through code. This automates the manual simulation-versus-scan matching done in the transition-layer study. It stays in the proposal only if it beats a scripted search, as described below.

Baselines

The do-nothing reference is no pulse-shape cut at all. It keeps 100% of signal with no calibration gap and no background suppression. It anchors the trade-off and reminds us that a gap of zero is trivial unless rejection is retained.

The physics standard is the $A/E$ cut with quadratic energy normalization from Multisite event discrimination for the majorana demonstrator. It reports (90 ± 3.5)% single-site retention, reduces the 2103 keV single-escape peak to about 6% of its original counts, and suppresses the background index in the region of interest by a factor of 3. We also run a stronger variant that physicists would reach for first: $A/E$ normalized as a function of drift time as well as energy. This is the obvious hand-built way to remove position dependence, and it is the most direct competitor to our firewall. Both are tuned on the same simulated calibration set with the same cut-optimization procedure.

The strongest learned method is the autoencoder-plus-classifier of Deep learning based pulse shape discrimination for germanium detectors. It reports 90% rejection of 2614 keV full-energy-peak events at 90% acceptance of 1592 keV double-escape-peak events, on the same 256-sample waveform format we use. We reimplement it and give it the same training data, the same hyperparameter-search budget in GPU hours, and the same early-stopping rule as our models. We found no reported result on Lacuna for the calibration-to-bulk acceptance gap of any learned germanium classifier, so we measure that gap ourselves for every baseline.

A distribution-matching baseline reweights calibration events by drift time so their position distribution matches uniform bulk decays, then retrains the unrestricted network. Its tuning budget equals ours.

Two ablations isolate the key ingredient. The first removes the adversarial judge from the firewalled model but keeps the shape-only input. The second removes the input restriction but keeps the judge. If the shape-only input alone closes the gap, the judge is unnecessary. If the restriction alone fails, joint information is leaking through features we assumed were pure shape.

Beating the baselines means reaching, at equal 90% calibration signal acceptance, a calibration-to-bulk gap below 1% in simulation, with residual single-escape-peak survival no worse than the roughly 6% of $A/E$. It also means beating drift-time-normalized $A/E$ and reweighting on the gap at matched background rejection.

Why not just…?

Why not just keep the $A/E$ cut? It is simple, interpretable and stable, and its performance numbers are the standard we must match. Its known failures are twofold. It misses multi-site events that happen to have a sharp current peak (Deep learning based pulse shape discrimination for germanium detectors), and it carries the 3.5% acceptance systematic (Multisite event discrimination for the majorana demonstrator). The drift-time-normalized variant is a real threat to this proposal. If it closes the gap as well as the firewalled network, the right conclusion is that the problem was one missing correction term. The decomposition then serves as a diagnostic confirming why, which we would report as the result.

Why not just train a bigger convolutional network on the whole waveform? Capacity does not address the problem. The network is trained on calibration events, and any position dependence it learns from their spatial bias is rewarded during training. Our third step measures exactly this. If the unrestricted network's gap is already small and uncorrelated with joint information, our hypothesis is wrong and we will say so.

Why not just reweight calibration events to the bulk position distribution? Reweighting corrects the mix of positions only where calibration events exist. External-source events are sparse deep in the crystal, so the large weights needed there make the estimates noisy. This alternative may well work, which is why it is a full baseline. The comparison that settles it is the gap and its uncertainty as a function of depth. We expect reweighting to match us near the surface and degrade in the deep bulk. If it does not degrade, reweighting wins.

Why not just use a single language model with step-by-step reasoning for the review step, or a scripted fit? A scripted grid search over transition-layer depth and trapping strength is the honest competitor to the agent loop, more so than a single prompted model. The test injects known perturbations into simulated "data", such as moving the full-charge-collection depth from 0.77 to 0.93 mm or adding a trapping gradient. We then compare how often each method recovers the injected cause under the same simulation budget. MACM's gains over single-model prompting (54.68% to 76.73% on the hardest MATH tier) were on symbolic problems, so they do not transfer by default. If the agent loop does not beat the scripted search on these known injections, we drop it.

Proposed Experiments

The decisive experiment is the attribution test in simulation. For each baseline classifier we rank bulk events by LSMI joint information. We then ask what fraction of the acceptance gap comes from the top tenth of that ranking, compared against the same fraction under rankings by shared information, by classifier confidence, and at random. A concentration well above the confidence ranking supports the core claim. No concentration falsifies it, and in that case the firewall has no motivation.

The second test is the firewalled model against all baselines. We report background survival at the 2614 keV full-energy and 2103 keV single-escape peaks at 90% calibration acceptance, and the calibration-to-bulk gap with uncertainty broken down by drift time. We add transition-layer slow pulses and trapping as stress conditions to check that removing joint information does not open a new weakness near the surface, which is the fragility warning from Proca, Rosas et al. read in reverse.

The third test moves to real data without signal labels. We compare acceptance on the 1592 keV calibration peak with acceptance on the low-energy continuum dominated by two-neutrino double-beta decay. This is a uniform bulk population of mostly single-site events, the measurement pursued in Progress Toward A 2νββ Measurement For The Majorana Demonstrator. We use it only as a bulk check on agreement, not as a training label. We expect the firewalled model's two acceptances to agree more closely than those of the unrestricted network.

The final test is the injected-perturbation recovery comparison between the agent loop and the scripted search, which decides whether the multi-agent review stays in the method.