Germanium detectors search for neutrinoless double-beta decay, a hypothetical radioactive decay whose half-life would exceed $10^{25}$ years. They reject most of their background by looking at the shape of each electrical pulse. A decay deposits its energy at one point in the crystal and gives one sharp current spike. A gamma ray that scatters several times gives a smeared, multi-peaked pulse. The standard cut compares peak current to total energy. It works well, but its signal efficiency has to be measured with an external thorium calibration source, and events from that source do not sit where real decays do. In the MAJORANA DEMONSTRATOR this mismatch is the dominant systematic uncertainty, ±3.5% on a 90% signal acceptance (Multisite event discrimination for the majorana demonstrator). That number enters every half-life limit directly. It will matter more at tonne scale, where statistical errors shrink.
We propose to close that gap with a tool from condensed-matter physics: effective Hamiltonian modeling, meaning a compact band-structure model whose allowed terms are fixed by the crystal's symmetries. Electrons in germanium sit in four ellipsoidal energy valleys along the crystal's body diagonals. Holes live in a warped heavy/light-hole band. Both effects make charge drift depend on direction relative to the crystal axes, and that direction dependence is what shapes pulses differently at different positions in the detector. Our claim is specific. If the drift law is written only in the forms that cubic symmetry allows, a pulse simulator needs a handful of parameters. Those parameters can be pinned down from calibration waveforms, and the simulator can then predict signal acceptance for decays spread uniformly through the bulk, with an uncertainty that can be checked against data it never saw. The effective-Hamiltonian direction supplies the method. Signal-to-background discrimination supplies the problem and the benchmark.
Core Claim and What Is New
Existing germanium pulse-shape work treats the detector response in one of two ways. It either calibrates a scalar cut empirically, using a quadratic correction of the amplitude-to-energy ratio against energy tuned on the thorium double-escape peak (Multisite event discrimination for the majorana demonstrator), or it learns directly from calibration waveforms. Two examples of the second approach are an autoencoder with a 141-parameter classifier (Deep learning based pulse shape discrimination for germanium detectors) and wavelet features with clustering and support vector machines for data cleaning (Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines). Both routes inherit the calibration source's spatial bias, because neither has a model of how pulse shape varies with position.
The research file suggested that current simulators treat drift as fully isotropic. That is too strong, since the germanium community already uses empirical, crystal-axis-dependent drift-velocity tables. What is new here is different. We write the drift law as a symmetry-constrained expansion derived by the theory of invariants, the same technique that fixes the form of the 4×4 Dirac Hamiltonian in Na₃Bi (Dirac semimetal and topological phase transitions in A_3Bi (A=Na, K, Rb)) and the 8-band Kane model in Cd₃As₂ (Three-dimensional Dirac semimetal and quantum transport in Cd_3As_2). We fit it jointly with the electronics response to calibration waveforms, and we propagate its posterior into the acceptance of the cut at the decay energy.
For a carrier in field $\mathbf{E}$, written in the crystal frame, cubic symmetry allows a drift velocity of the form
$$\mathbf{v}(\mathbf{E}) = \mu_1(|\mathbf{E}|)\,\mathbf{E} + \mu_3(|\mathbf{E}|)\,\frac{(E_x^3,\,E_y^3,\,E_z^3)}{|\mathbf{E}|^2} + \dots$$
Here $\mu_1$ is the isotropic mobility and $\mu_3$ is the lowest-order cubic anisotropy. Both are smooth functions of field strength. The omitted terms are higher cubic invariants. A second-rank mobility tensor must be isotropic in a cubic crystal, so all anisotropy comes from the field-dependent $\mu_3$ term. This matches the known physics that direction dependence in germanium appears at high fields as electrons repopulate valleys. We will build the model at two levels. The first is this phenomenological invariant expansion. The second is a microscopic four-valley model with ellipsoidal masses and field-dependent valley occupation, which fixes the ratio $\mu_3/\mu_1$ from a few band parameters.
One observable follows directly. Rise time should vary with a fourfold period in azimuth around the detector's symmetry axis, with an amplitude set by $\mu_3$. This signature gives a direct check that the fitted parameters are physical.
We also address a tempting variant head on and reject it. The file's first gap proposes adding Berry-curvature anomalous velocity to germanium pulse formation, drawing on the semiclassical expansion near Weyl nodes in Asymptotic Analysis for Bloch Electrons with Weyl Nodes. Germanium has both inversion and time-reversal symmetry, and in such a crystal the Berry curvature of each band vanishes. Such a term would be zero by symmetry, and the same symmetry reasoning that rules it out is what we use constructively.
Research Program
The first component is a differentiable pulse simulator for p-type point-contact detectors. It chains four stages: an electrostatic field and weighting-potential solve for the measured geometry and impurity profile, a charge cloud of finite size, drift under the symmetry-constrained velocity law for electrons and holes, and a parameterized electronics response. The crystal orientation angle is a fitted parameter. The drift law adds a small number of parameters: $\mu_1$ and $\mu_3$ for each carrier at a few field knots, or the four-valley band parameters. Impurity-profile uncertainty and electronics terms are treated as nuisance parameters.
The second component is inference. We fit these parameters to calibration waveforms: double-escape-peak events (single-site), full-energy and single-escape-peak events (multi-site), and position-tagged events where available. We then sample the posterior, simulate decays uniformly throughout the active volume at the decay energy, apply the unchanged experimental cut, and report acceptance with an uncertainty that comes from the posterior rather than from a fixed engineering allowance.
The third component is a classifier. It is trained on simulated uniform single-site events, with fourfold crystal-symmetry augmentation, and then fine-tuned on calibration data. Its purpose is to test whether a model that knows the geometry also improves rejection, not just the error bar.
The more speculative intersection in the file is using topological semimetals such as Cd₃As₂ or EuB₆ as active target materials. We deliberately do not pursue it. There is no benchmark, no detector data and no reported discrimination figure to test it against. Germanium, by contrast, has a live systematic with a published size.
Baselines
The naive reference point takes the acceptance measured on the double-escape peak as the acceptance at the decay energy, with no position correction. It carries the full ±3.5% systematic reported for the amplitude-versus-energy cut (Multisite event discrimination for the majorana demonstrator), which is the number we aim to shrink.
The standard empirical method is that same cut with its quadratic energy correction. On the thorium calibration it retains (90 ± 3.5)% of single-site events and reduces single-escape-peak events to about 6% of their original count (Multisite event discrimination for the majorana demonstrator). We rerun it on exactly the same calibration runs. Our method does not replace the cut. It replaces the way the cut's acceptance is estimated.
The strongest learned discriminator is the autoencoder-plus-classifier pipeline. It reports 90% rejection on the full-energy peak at 90% acceptance on the double-escape peak (Deep learning based pulse shape discrimination for germanium detectors). We retrain it on identical data splits, with the same hyperparameter search budget we give our classifier. It counts as beaten only if, at matched acceptance, our classifier rejects more multi-site events and its acceptance transfers with smaller held-out error.
The ablation that isolates the key ingredient keeps the same simulator but removes the symmetry constraint. The drift velocity is replaced by an unconstrained per-direction table, or by a small neural velocity field, given at least as many parameters and the same fitting budget. A second ablation forces $\mu_3 = 0$, which is isotropic drift. Our proposal wins only if the constrained model predicts held-out acceptance better than the unconstrained one. Otherwise the symmetry adds nothing beyond parameter count.
Why not just…?
Why not just keep the empirical amplitude-versus-energy correction and accept ±3.5%? Because that uncertainty comes from where calibration events sit in the crystal, not from counting statistics (Multisite event discrimination for the majorana demonstrator). More calibration data from the same external source cannot reduce it. Only a model that can move from the calibration's spatial distribution to the decay's uniform distribution can address it, and our held-out-position test measures whether ours does.
Why not just measure acceptance in situ on two-neutrino double-beta decay events, which are internal and uniformly distributed? This is the most serious competitor, and it may partly win. Its limits are that those events sit well below the decay energy, so the cut's energy dependence must still be extrapolated, and that other backgrounds contaminate the spectrum. We treat it as an independent check rather than something to argue away. If the model's acceptance agrees with the two-neutrino measurement within its posterior band, and the model also predicts the energy dependence the two-neutrino data cannot reach, the approaches combine. If the two-neutrino measurement alone already matches the model's precision at the decay energy, the model's value reduces to cross-checking, and we will report that.
Why not just train a deep network on calibration waveforms, or apply domain adaptation? A network trained on double-escape-peak events learns whatever position cues that surface-biased sample carries, so its acceptance for bulk events is as uncertain as the scalar cut's. The file also shows that data-driven cleaning can leak anomalies at high rates: 15.034 ± 0.224% in pileup-dominated data (Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines). Domain adaptation needs unlabeled target-domain events, and the only such events are the two-neutrino data discussed above. The direct test is the learned baseline's held-out acceptance error compared with ours.
Why not just fit a flexible empirical drift table, which the community already uses? The table may predict pulses equally well. The question is whether it transfers and whether its parameters are identifiable from calibration alone. The unconstrained-drift ablation answers this directly. If it matches the constrained model on held-out acceptance, we will conclude that the symmetry constraint only reduces parameter count and does not improve transfer.
Proposed Experiments
We use MAJORANA DEMONSTRATOR-style thorium calibration data spanning 100 keV to 3 MeV, with the double-escape peak at 1592 keV as the single-site benchmark and the single-escape and full-energy peaks as multi-site benchmarks (Multisite event discrimination for the majorana demonstrator; Deep learning based pulse shape discrimination for germanium detectors). We add the TUBE collimated surface-scanning data, which provides events at known positions (alpha-event characterization and rejection in point-contact HPGe detectors). Access to these datasets depends on collaboration data; the file lists no public release.
The first decisive check is a held-out-position test. We fit the model on calibration runs with the source at one set of positions, predict the cut's acceptance and pulse-shape distributions for runs at other positions, and compare with measurement. Success means a prediction error well below the 3.5% systematic, together with posterior intervals whose coverage is correct across held-out runs.
The second check is the physical signature. The fitted $\mu_3$ must reproduce the fourfold azimuthal variation of rise time in scanned events. A fit that matches acceptance while getting this pattern wrong would indicate a model tuned to one number rather than to the physics.
The third check compares the model's acceptance with the two-neutrino in-situ measurement at overlapping energies. The fourth reports rejection at matched 90% acceptance against the learned baseline.
The surface-alpha cut, which combines delayed charge recovery with the amplitude ratio, already rejects essentially all surface alphas more than 3 mm from the detector center, at 0.2% bulk loss (alpha-event characterization and rejection in point-contact HPGe detectors). It serves as a control. Applying our simulator must not degrade it.
References
- Multisite event discrimination for the majorana demonstrator
- Deep learning based pulse shape discrimination for germanium detectors
- Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines
- Dirac semimetal and topological phase transitions in<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mi>A</mml:mi><mml:mn>3</mml:mn></mml:msub></mml:math>Bi (<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mi>A</mml:mi><mml:mo>=</mml:mo><mml:mtext>Na</mml:mtext></mml:mrow></mml:math>, K, Rb)
- Three-dimensional Dirac semimetal and quantum transport in Cd<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mrow/><mml:mn>3</mml:mn></mml:msub></mml:math>As<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mrow/><mml:mn>2</mml:mn></mml:msub></mml:math>
- Asymptotic Analysis for Bloch Electrons with Weyl Nodes
- $$\alpha $$-event characterization and rejection in point-contact HPGe detectors