Searches for neutrinoless double-beta decay run arrays of dozens to hundreds of germanium crystals deep underground. They need background rates below about one count per tonne-year (The MAJORANA Project). A key tool is pulse shape discrimination: classifying each recorded electrical pulse by its shape. The signal deposits its energy at essentially one point in the crystal, while most gamma-ray backgrounds scatter at several points. The standard single-number test, the ratio of peak current to energy (called $A/E$), keeps about 90% of single-site events and removes about 90% of multi-site ones (Multisite event discrimination for the majorana demonstrator). A small neural network matches that performance and also catches multi-site events that fool $A/E$ (Deep learning based pulse shape discrimination for germanium detectors). That network, however, was trained for one detector. Every crystal in a real array differs in impurity profile, crystal orientation, dead layer and electronics, and the only labeled data come from external calibration sources. Those sources illuminate a spatial region different from where the signal occurs.
Speech recognition faced the same structure: one shared model, many "speakers," and very little data per speaker. It found a workable answer by giving each speaker a tiny adaptation vector, sometimes a physically meaningful one. We propose to treat each germanium detector as a speaker. A shared pulse-shape network carries a handful of per-detector physical parameters, which we call a "detector code": electronics time constants, a drift-time stretch, and gain and noise levels. The network is meta-learned, meaning it is trained across many detectors specifically so that each detector's code can be fitted from a few dozen unlabeled waveforms recorded during normal data-taking. The shared classifier then works correctly on that detector without any detector-specific labels.
Core Claim
The main claim is that the adaptation step should be unsupervised and in situ, and that meta-learning is what makes this safe. An unsupervised inner loop fits the detector code by reconstructing the detector's own physics-run waveforms. An outer loop trains the shared weights so that the code obtained this way yields correct single-site versus multi-site decisions on labeled calibration data. Formally, with $d$ indexing detectors, $\theta$ the shared network weights, $\phi$ the small detector code with learned starting point $\phi_0$, $U_d$ a small set of unlabeled physics-run waveforms from detector $d$, $C_d$ its labeled calibration events, and $\alpha$ an inner step size:
$$\phi_d' = \phi_0 - \alpha \nabla_{\phi}\, \mathcal{L}_{\text{rec}}(\theta, \phi_0; U_d), \qquad \min_{\theta,\,\phi_0} \sum_d \mathcal{L}_{\text{cls}}(\theta, \phi_d'; C_d).$$
Here $\mathcal{L}_{\text{rec}}$ is waveform reconstruction error and $\mathcal{L}_{\text{cls}}$ is classification loss. The inner step may be repeated a few times.
This design gives three properties that no single prior piece provides. First, the data that set each detector's code are spatially distributed like the signal. They are bulk events recorded between calibrations, not surface-heavy calibration peaks. The file identifies that spatial mismatch as the source of the $\pm 3.5\%$ acceptance systematic (Multisite event discrimination for the majorana demonstrator). Second, the code is a few interpretable numbers, so a physicist can check it against independent measurements. Third, the outer loop explicitly trains the network so that "the code that best reconstructs pulses" is also "the code that classifies correctly." A conventionally trained model with a reconstruction-fitted code has no such guarantee.
The constraint that makes this a genuinely low-resource problem is the event rate. In a low-background array, the number of bulk events a single detector records between weekly calibrations is small. Our rough estimate from the known double-beta rate is tens, not thousands. The file itself gives no such rate.
Relation to Prior Work
Three works supply the method, and each lacks one ingredient we need.
The speaker-code approach of Speaker Code Based Speaker Adaptive Training Using Model Agnostic Meta-Learning meta-learns an initial speaker code that adapts in 2 gradient steps instead of 12. It reached 13.0% word error rate on Hub5, a 5.8% relative gain. Its inner loop uses labeled adaptation utterances, and its code is an abstract vector. Ours uses unlabeled waveforms and a physical code.
Acoustic Model Adaptation from Raw Waveforms with Sincnet shows that adapting only 80 interpretable filter cutoffs moves an adult speech model to children's speech, cutting word error rate from 68.19% to 31.65%. It also shows that the learned shifts reproduce vocal-tract-length normalization. This is our precedent for physically constrained adaptation parameters. SincNet, however, starts from fixed mel-scale cutoffs and adapts with supervised labels.
Conditional layer normalization in AdaSpeech: Adaptive Text to Speech for Custom Voice supplies the mechanism for letting a small code modulate a network's internal activations. It does so with about 5K parameters per speaker, versus 14.1M for full decoder fine-tuning.
Two further works constrain the design. Meta-Learning with Latent Embedding Optimization shows that inner-loop gradient steps in high-dimensional weight space overfit tiny support sets, and that adapting in a low-dimensional space avoids this. That finding motivates keeping $\phi$ to under about ten dimensions. It also makes the second-order meta-gradient cheap, which addresses the training-cost concern raised for Speaker Code Based Speaker Adaptive Training Using Model Agnostic Meta-Learning.
On the physics side, the motivating evidence is threefold. The two-stage model of Deep learning based pulse shape discrimination for germanium detectors (a 256-sample to 7-dimensional autoencoder followed by a 141-parameter head) has no mechanism for transfer between detectors. Per-detector $A/E$ tuning carries a $\pm 3.5\%$ efficiency uncertainty. And channel-by-channel variation is listed as the scaling problem for tonne-scale arrays. The file also raises a warning that could contradict our approach: few-shot adaptation on a spatially biased support set could worsen efficiency systematics. We take this seriously. It is precisely why our support set is physics-run data rather than calibration data.
Research Program
The model has three components. An encoder receives the raw current pulse, with the detector code modulating its normalization layers. It produces a low-dimensional latent description of the event, following the autoencoder design of Deep learning based pulse shape discrimination for germanium detectors. A decoder produces an idealized induced-current pulse from that latent description. That pulse is then passed through a differentiable detector-response layer whose only parameters are $\phi$: preamplifier rise and decay constants, a drift-time scale factor, gain, and noise amplitude. A small classifier head, comparable in size to the published 141-parameter head, reads the latent description.
Because $\phi$ sits in an explicit forward model of the electronics and charge drift, fitting it by reconstruction is physically meaningful. The classifier sees pulses that have effectively been "normalized" to a common detector.
Meta-training treats each detector in the training set as a task. For each task, a support set of unlabeled physics-run or low-energy waveforms drives a few inner steps on $\phi$. A query set of labeled calibration events then drives the outer update on $\theta$ and $\phi_0$. Following the standard calibration practice in Multisite event discrimination for the majorana demonstrator, single-site labels come from the $^{208}$Tl double-escape peak and multi-site labels from full-energy peaks. Since $\phi$ is tiny, exact second-order meta-gradients are affordable. We will still report a first-order variant. At deployment, adapting a detector means running a few gradient steps on under ten numbers, which costs almost nothing. The classifier itself stays small enough for the lightweight inference pipelines these experiments require.
The data needed are digitized waveforms from a multi-detector point-contact array, covering both calibration and background runs, such as the MAJORANA DEMONSTRATOR data. The file describes 28 tested detectors with energy resolutions of 1.6 to 2.2 keV and dead layers of 0.1 to 0.2 mm (Testing the Ge Detectors for the MAJORANA DEMONSTRATOR). If collaboration data are not accessible, we develop the method first on pulses simulated with the MAGE/GEANT4 chain, drawing electronics and drift parameters per virtual detector. We then treat real data as the decisive test. The file notes that simulation fails to reproduce dead layers and electronics perfectly, so simulation-only results would not settle the question.
Baselines
The naive reference is per-detector $A/E$ with its standard energy correction, tuned on each detector's own calibration peaks exactly as in Multisite event discrimination for the majorana demonstrator. The reported numbers are $(90 \pm 3.5)\%$ single-site retention, $^{208}$Tl single-escape peak reduced to about 6%, and a factor-3 reduction of background in the region of interest. It is the established method, and it uses more per-detector labeled data than our approach does. Beating it means one of two things. Either we achieve lower single-escape-peak survival at the same 90% double-escape-peak acceptance on held-out detectors, or we achieve equal rejection with a smaller detector-to-detector spread in acceptance.
The strongest learned competitor is the Holl et al. autoencoder plus classifier trained separately for each detector (Deep learning based pulse shape discrimination for germanium detectors). It reports 90% background rejection at 90% signal acceptance, which matches $A/E$. We will run it with the same architecture size, the same calibration statistics per detector, and the same hyperparameter search budget as our model. Its natural extension, a pooled network trained on all training detectors and then applied or head-fine-tuned on each held-out detector, is also included.
The ablation that isolates the key ingredient is the same architecture and detector code, trained jointly across detectors without meta-learning. At test time, the code is fitted by the same reconstruction loss from the same unlabeled support set, starting from zero or from the average code. This is the analogue of the zero-initialized speaker code in Speaker Code Based Speaker Adaptive Training Using Model Agnostic Meta-Learning.
A second ablation replaces the physical code with an unconstrained learned vector of equal size, adapted in the style of Meta-Learning with Latent Embedding Optimization. This measures what, if anything, physical interpretability costs in accuracy.
All methods receive identical detector splits, identical labeled calibration events, and matched compute and tuning budgets. Evaluation uses leave-detectors-out cross-validation, so no method ever sees labels from the detectors it is scored on. The exception is the per-detector baselines, which are allowed those labels by construction, and that advantage makes them harder to beat.
Why not just…?
Why not just keep tuning $A/E$ per detector? It is cheap, transparent, and good. It is the reference we must beat, not a straw man. The evidence against relying on it alone is that it has a known blind spot and a known cost. Holl et al. show it misses multi-site events whose first current spike looks single-site. Its tuning inherits the calibration-source spatial bias behind the $\pm 3.5\%$ uncertainty. If our method cannot beat per-detector $A/E$ on held-out detectors, the proposal fails, and we will say so.
Why not just train one network on pooled data from all detectors, or fine-tune its small head per detector? This is a real threat. Pooling may already average out most channel variation. Head-only fine-tuning on calibration peaks is simple. Two points argue against it being sufficient. Head fine-tuning uses surface-biased labeled events, which is the very bias we want to avoid. And pooling offers no way to track detector drift between calibrations. The pooled and head-fine-tuned baselines test this directly. If they match our method in both rejection and acceptance spread, the per-detector code adds nothing.
Why not just fit the detector code by reconstruction without any meta-learning? This is the most serious alternative. The experiment is designed to settle it with the joint-training ablation above. Our hypothesis is that without the outer loop, the reconstruction-optimal code does not align with the classification-optimal code. The result would be good reconstruction with drifting classifier acceptance. If the ablation matches the meta-learned model, we will drop the bilevel training. What remains, a physical code fitted without labels, would still be the contribution.
Why not just simulate each detector in full, fitting its impurity profile and fields, and train on the simulation? Full per-detector simulation is the physicist's principled answer, but it is expensive. The file notes that GEANT4/MAGE simulation does not faithfully capture dead layers and electronic transfer functions. Our response layer is a deliberately partial version of this idea: it models only the electronics and drift-time parameters that can be identified from pulses. We will compare the fitted codes against independently measured electronics constants to check whether this partial model is adequate.
Proposed Experiments and Decisive Checks
The primary experiment is leave-detectors-out evaluation on a multi-detector point-contact array. The main metrics are multi-site rejection (full-energy-peak and single-escape-peak survival) at fixed 90% double-escape-peak acceptance, and the standard deviation of single-site acceptance across held-out detectors at a single global cut. Both are reported as a function of unlabeled support-set size, swept from about ten to about a thousand waveforms.
The spatial-bias check compares acceptance measured on the calibration double-escape peak with acceptance on a bulk-dominated, mostly single-site event population from background runs, such as the two-neutrino double-beta continuum in enriched detectors. The prediction is that the gap between these two acceptances is smaller for our method than for calibration-tuned baselines. This is the most direct test of whether in-situ adaptation reduces the systematic that motivated the work.
The interpretability check correlates the fitted time constants and gains with values measured independently from pulser or electronics data. It also tracks the codes across successive calibration periods to see whether code drift predicts drift in $A/E$ behavior.
A secondary, cheaper test applies the recipe to the speech setting it came from. We would meta-learn initial SincNet cutoffs across adult speakers and adapt them to children's speech with unlabeled audio. The comparison point is the reported supervised result of 31.65% word error rate after one epoch (Acoustic Model Adaptation from Raw Waveforms with Sincnet). No reported result was found on Lacuna for unsupervised or meta-initialized SincNet adaptation.
The proposal succeeds if three conditions hold on held-out detectors. First, it beats per-detector $A/E$ in single-escape-peak survival at matched acceptance, or matches it with visibly smaller acceptance spread. Second, the meta-learned model outperforms the jointly trained code ablation. Third, the fitted codes agree with independent electronics measurements.
References
- The MAJORANA Project
- Multisite event discrimination for the majorana demonstrator
- Deep learning based pulse shape discrimination for germanium detectors
- Speaker Code Based Speaker Adaptive Training Using Model Agnostic Meta-Learning
- Acoustic Model Adaptation from Raw Waveforms with SincNet
- AdaSpeech: Adaptive Text to Speech for Custom Voice
- Meta-Learning with Latent Embedding Optimization
- Testing the Ge Detectors for the MAJORANA DEMONSTRATOR