Field-Conditioned Pulse Shape Discrimination Across Germanium Detectors

Searches for extremely rare processes such as neutrinoless double-beta decay run arrays of high-purity germanium crystals. To tell a candidate signal from background radioactivity, they look at the shape of each electrical pulse. A decay deposits its energy at one point and produces a clean, single-step pulse. A gamma ray that scatters several times produces a pulse with kinks. An alpha particle hitting the crystal surface produces a pulse with a slow tail. How a given deposit turns into a pulse is set by physics that is fully specified and cheap to compute for each crystal. The internal electric field comes from the electrostatic Poisson equation. The "weighting potential" comes from the Shockley–Ramo theorem: it says how much current a moving charge induces on the readout contact at each point in the crystal.

Despite this, today's learned pulse classifiers see only the waveform. They are retrained or retuned for every crystal, and they break when crystal geometry changes, for example when moving from point-contact to inverted-coaxial designs. We propose a different setup. The classifier receives each detector's computed field and weighting-potential maps alongside the raw waveform, so the classification rule depends on the detector's own physics rather than being memorized for one crystal. The model is pretrained on simulated pulses from thousands of synthetic crystal geometries, adapted to real detectors using only a handful of calibration events, and trained under a penalty that stops it from distorting the background energy spectrum. If this works, a newly built or unusual crystal could get a competitive discriminator from its geometry and a few hundred calibration pulses.

Core Claim

Supplying the solved field maps as an input works better than either of the two usual options: describing the detector with a few scalar geometry numbers, or training a separate model per detector. The advantage should be largest on detector geometries held out during training and when calibration data is scarce. We also claim that this conditioning can be combined with nuisance decorrelation, meaning a penalty that makes the classifier's output uninformative about energy and drift time. The goal is that the background continuum in the region where the signal would appear keeps its shape after the cut.

Where This Sits in Prior Work

Work this proposal improves on. Two lines show that learned discriminators beat hand-built physics cuts. Both remain tied to individual detectors or to engineered features.

  • The convolutional-autoencoder-plus-small-network approach in Deep learning based pulse shape discrimination for germanium detectors reads raw 256-sample pulses. It matches the standard amplitude-over-energy ($A/E$) cut and catches multi-site pulses with subtle secondary bumps. However, it is trained and tested on a single detector channel and receives no geometry information.
  • The gradient-boosted trees in Interpretable boosted-decision-tree analysis for the Majorana Demonstrator are trained on hand-computed summary features pooled across point-contact (PPC) and inverted-coaxial (ICPC) detectors. Shapley-value analysis shows they implicitly learn drift-time and field corrections. They still need oversampling to handle detector imbalance and cannot read waveforms.

The hand-built baseline is α-event characterization and rejection in point-contact HPGe detectors. It introduced the delayed-charge-recovery tail cut for surface alphas and sets a very high bar for that class of background.

Work that supplies the method. We borrow ideas from three neural-operator papers, which learn mappings between functions rather than fixed-size vectors.

  • PDEformer shows that feeding the governing equation as an explicit input lets one model generalize across equations. It has only been demonstrated on one-dimensional benchmark equations. We feed the solved fields rather than the symbolic equation, because for detectors the equation is fixed and what varies is the geometry and the distribution of impurities in the crystal.
  • Transolver groups mesh points into a small set of learned "physics slices" so that attention costs scale linearly with mesh size on arbitrary geometries.
  • Universal Physics Transformers compresses any discretization into a fixed set of latent tokens.

We use a Transolver-style encoder to turn each detector's field mesh into a fixed set of tokens. The waveform encoder then attends to those tokens. Neither operator paper addresses classification or instrument response.

Closest bridging work.

  • FM4NPP pretrains a large state-space model on detector hits and reports a 2.9× gain over training from scratch when labels are scarce. It is tied to one cylindrical detector and to processed hit coordinates, not waveforms or field physics.
  • RESuM uses learned surrogates for rare-background design optimization in LEGEND. It works offline on shielding geometry and is not an event-level filter.

Work that constrains the design. Robust Anomaly Detection for Particle Physics Using Multi-Background Representation Learning shows two things. First, training to separate several known backgrounds yields better representations than a single-background autoencoder. Second, adversarial decorrelation (their NuRD method) cuts mass sculpting from a Jensen–Shannon divergence of 0.255 to 0.118. We adopt both lessons. We train on multiple background classes (multi-site gamma, surface alpha, near-contact events), and we penalize dependence of the classifier score on energy and drift time.

Research Program

Simulation corpus. We first build a simulation corpus across a broad family of germanium geometries. The family covers PPC, ICPC, and broad-energy designs, with varied radius, height, contact size, surface-layer thickness, and impurity gradient. For each geometry we solve the electric field and the weighting potential on a mesh. We then generate pulses from simulated energy deposits (single-site, multi-site, surface) with a GEANT4-plus-field-solver pipeline of the kind described in the research file. Each pulse is passed through randomized electronics response, noise, and charge-trapping models. This randomization is deliberate, because the sim-to-real gap in this field comes largely from unmodeled electronics and trapping.

Model. The model has three parts:

  • a field encoder that maps the detector's field and weighting-potential meshes to a fixed set of tokens;
  • a waveform encoder that handles variable-length pulses;
  • cross-attention from waveform to field tokens, followed by a classification head over background classes.

Training objective. Training minimizes

$$ \mathcal{L} = \mathcal{L}_{\text{cls}} - \lambda \, \mathcal{L}_{\text{adv}}(s; E, t_{\text{drift}}), $$

where:

  • $\mathcal{L}_{\text{cls}}$ is the cross-entropy over event classes;
  • $\mathcal{L}_{\text{adv}}$ is the loss of a critic network that tries to predict energy $E$ and drift time $t_{\text{drift}}$ from the classifier score $s$;
  • $\lambda$ sets how strongly we push the score to carry no information about those nuisances.

Adaptation to real detectors. For each real detector we compute its field maps from its nominal geometry. We then fine-tune either only a small adapter or nothing at all, using $N$ calibration events from thorium-228 runs. The calibration provides three classes:

  • double-escape-peak events as a single-site proxy;
  • single-escape and full-energy-peak events as multi-site background;
  • alpha-source data where available.

Baselines

Naive reference: standard cuts with per-detector tuning. The first reference point is the standard $A/E$ cut plus the delayed-charge-recovery cut, tuned per detector exactly as experiments do now. It costs nothing new. For alphas it is near-perfect: α-event characterization and rejection in point-contact HPGe detectors reports 100% rejection of surface alphas at radii above 3 mm with only 0.2% bulk loss. We therefore do not expect to beat it on alphas. Instead we require parity on alphas and look for gains on multi-site rejection.

Strongest existing learned methods: per-detector trained models. These are the autoencoder-plus-network model of Deep learning based pulse shape discrimination for germanium detectors and the boosted trees of Interpretable boosted-decision-tree analysis for the Majorana Demonstrator. The reported numbers are as follows:

  • The autoencoder model achieves 90% full-energy-peak rejection at 90% double-escape-peak acceptance.
  • The boosted trees reject 8.6% more single-escape-peak events than $A/E$, with 164 events surviving in the background estimation window versus 153 for the standard analysis.

These models are the real threat because each detector already receives calibration data. We will train both on the same calibration events available to our model at every $N$, with an equal hyperparameter-search budget measured in trials. The autoencoder will also get unsupervised pretraining on the same unlabeled real pulses we use.

Beating these baselines means two things on held-out detectors:

  • higher multi-site rejection at a fixed 90% acceptance of double-escape-peak events, at matched $N$;
  • reaching the full-data per-detector performance with at least an order of magnitude fewer calibration events.

All of this must hold with no worse energy sculpting.

Ablations that isolate the key ingredient.

  • Scalar conditioning: the same network and pretraining, but conditioned on a vector of geometry scalars instead of field maps.
  • No conditioning: the same network with no detector information at all.
  • No decorrelation: our model with $\lambda=0$.

Why not just…?

Why not just retrain a small model per detector on its calibration run? Every detector does get calibrated, so this is the strongest alternative, and it may win when calibration statistics are plentiful. Our claim is limited to two regimes:

  • detectors early in deployment or with low calibration statistics;
  • geometries unlike the existing fleet.

The sample-efficiency curves across $N$ settle this directly. If per-detector training matches us at $N$ of a few hundred events, the transfer claim fails, and we will report it that way.

Why not just feed geometry scalars (radius, height, contact size) as extra inputs? This is much simpler, and it could capture most of the benefit for regular crystals. It is the central ablation. Field maps should matter mainly where scalars miss structure: impurity gradients and the complex weighting potentials of inverted-coaxial detectors. Our prediction is therefore specific. The two models should be close within the PPC family and should separate on held-out ICPC detectors. If they do not separate, the field-operator machinery is unjustified and we will drop it.

Why not just train a one-class anomaly detector on background? Robust Anomaly Detection for Particle Physics Using Multi-Background Representation Learning found that single-background autoencoders were both weaker (area under the curve of 0.881 versus 0.914) and sculpted the search variable much more. In germanium the main background classes are well characterized by calibration, so multi-class training is natural. We will still include a one-class autoencoder trained on pulses from the energy continuum as a check.

Why not just train on simulation and fine-tune, without field input? This is our "no conditioning" ablation. It tests whether pretraining across many geometries helps only because of data volume, or because the model has learned to read the field.

Proposed Experiments

Synthetic transfer. In simulation, we hold out entire geometry families (train on PPC and broad-energy designs, test on ICPC) and measure multi-site rejection at 90% single-site acceptance as a function of $N$. Simulation also lets us check something the real double-escape-peak proxy cannot. These events sit in a particular energy and spatial distribution, whereas true signal is spread uniformly through the bulk. We will measure the acceptance of uniformly distributed simulated single-site events separately from double-escape-peak acceptance.

Real data. On MAJORANA calibration data of the kind used by the boosted-tree analysis (access via the collaboration or any public release), we run leave-one-detector-out and leave-one-type-out evaluations. We report:

  • single- and full-energy-peak rejection;
  • alpha rejection;
  • survival counts in the background estimation window;
  • the Jensen–Shannon divergence between the background energy spectrum in the continuum window before and after the cut.

That divergence is the sculpting metric, following Robust Anomaly Detection for Particle Physics Using Multi-Background Representation Learning. No reported sculpting number for germanium pulse shape discrimination was found on Lacuna, so our baselines' values will be measured, not quoted.

Field-swap test. At test time we give each detector another detector's field maps. If the model actually uses the field, performance should drop in proportion to how different the two detectors' weighting potentials are. If performance does not drop, the conditioning is decorative.

Outcome. The project succeeds if it shows three things together: a held-out-geometry gap over scalar conditioning, a sample-efficiency gain over per-detector training, and no added energy sculpting. Each of these alone is falsifiable with the data above.