Germanium detectors that search for neutrinoless double-beta decay, a hypothetical nuclear process with half-life above $10^{25}$ years, must throw away background events without losing signal. One of the hardest backgrounds is alpha particles that strike the detector's passivated surface, the thin insulating layer that coats part of the crystal. Charge created there is briefly caught in surface "traps" and released slowly, over microseconds to milliseconds, which leaves a characteristic slow tail on the recorded pulse. Experiments reject these events with hand-built tail cuts. The simulators used to train and validate those cuts cannot produce the tail at all, because they model only ideal charge drift in the electric field. This proposal builds the missing piece from physics. We use ab initio molecular dynamics (atomistic simulation in which the forces between atoms come from quantum-mechanical electronic-structure calculations at every step) to generate realistic disordered passivation-layer structures. From those structures we compute the spread of trap energies. That spread becomes a small, physically parameterized trap-and-release stage inserted into the standard pulse simulator.
The claim is narrow on purpose. Quantum calculations cannot predict absolute release times: release rates depend exponentially on trap depth, and density-functional errors of a few hundredths of an electron-volt shift rates by orders of magnitude. They can plausibly predict the shape of the trap-depth distribution, meaning how broad it is, whether it has one or several families, and how it depends on surface chemistry. We therefore let the atomistic simulation fix the shape and calibrate a single global energy offset to data. The resulting simulator should reproduce surface-event tails across source positions it was not calibrated on, and should do so better than either a black-box neural translator or an unconstrained empirical trap model with more free parameters. If it does, it provides labeled surface-alpha waveforms for training discriminators and a physically bounded way to quote systematic uncertainty on surface backgrounds.
Where This Sits
The motivating evidence comes from germanium pulse-shape work. α-event characterization and rejection in point-contact HPGe detectors scanned a collimated alpha source across the passivated surface of a MAJORANA-style detector. It showed that trapped charge produces a slow tail, and introduced a Delayed Charge Recovery (DCR) parameter measuring that tail's slope. Combined with the amplitude-over-energy cut $A/E$, DCR rejected all alpha events at source positions more than 3 mm from the detector center, at a cost of 0.2% of bulk events. Their trapping model is purely empirical, drawn from the scans themselves. Ad-hoc Pulse Shape Simulation using Cyclic Positional U-Net quantified how far the ideal electrostatic simulator siggen is from real Broad Energy Germanium pulses. Measured as histogram overlap (HIoU), the tail-slope match was 0%. The authors trained a CycleGAN-style translator (CPU-Net) that raised tail-slope HIoU to 83.2% and maximum-amplitude HIoU to 36.8%. They also report non-physical artifacts and trouble with surface events and outliers, and they argue that first-principles modeling of these effects is intractable. Deep learning based pulse shape discrimination for germanium detectors matched $A/E$ at 90% background rejection and 90% signal acceptance using an autoencoder plus a tiny classifier. It noted that calibration labels (double-escape versus full-energy peaks) are impure and that clean simulated labels are needed. Together these three papers define the gap: the tail is physically important, the simulators miss it, and the learned patch fails exactly on surface events.
Three atomistic papers supply method and constraints. Silicon Liquid Structure and Crystal Nucleation from Ab Initio Deep Metadynamics trained a neural-network potential on density-functional data to 2.1 meV/atom RMSE and used it to sample disordered and liquid silicon at scale. We reuse that recipe for disordered germanium surfaces. The same paper constrains us, because even the SCAN functional overestimated silicon's melting point by about 10% (1855 K versus 1685 K). That is why we refuse to trust absolute energies. Hydrogen Oxidation Reaction at the Ni/YSZ Interface: An Ab Initio Study shows that AIMD in CP2K can resolve charge-transfer bottlenecks at a buried interface. It is a precedent for interface work, not for detector physics. Predicting in silico electron ionization mass spectra using quantum chemistry is the closest "quantum dynamics to observable signal" pipeline in the file: it runs Born-Oppenheimer dynamics to predict mass spectra. It warns that cheap semi-empirical methods degrade on oxygen chemistry (similarity 520 versus 692 for hydrocarbons), which matters because germanium passivation layers may contain oxygen. It also warns that cost grows quickly with system size. No paper in the file connects atomistic simulation to radiation-detector pulse shapes.
We deliberately did not pick the other visible gap, simulating the 30–100 ns intermediate component of liquid-argon scintillation described in The liquid-argon scintillation pulseshape in DEAP-3600. That component involves excited-state excimer kinetics and electron-ion recombination over tens of nanoseconds across whole ionization tracks. Ground-state AIMD cannot reach it honestly. The germanium surface problem, by contrast, is a ground-state defect-level problem that density-functional methods are actually built for.
Core Claim and Model
A surface event deposits charge $Q$, of which a fraction $f$ is trapped. Trapped carriers are released thermally at the detector temperature $T$. The delayed charge collected by time $t$ is then
$$ Q_{\text{del}}(t) = f\,Q \int g(E)\,\Big[1 - \exp\!\big(-\nu\, t\, e^{-(E+\Delta)/k_B T}\big)\Big]\, dE , $$
where the terms are defined as follows. $g(E)$ is the normalized distribution of trap depths, which the atomistic calculations supply. $\nu$ is an attempt frequency of order a phonon frequency. $\Delta$ is the single global energy offset calibrated to data. $k_B$ is Boltzmann's constant. The rate form is ordinary thermally activated escape. It explains why the tail is a superposition of exponentials and why the absolute depths cannot be trusted: at 77 K, $k_B T \approx 6.6$ meV, so a 0.05 eV error changes rates by roughly a factor of $e^{7.5}$. This delayed-charge stage sits between siggen's ideal drift signal and a standard electronics response (an RC decay and a bandwidth filter). The electronics response is fit once on bulk calibration events.
The atomistic contribution is specifically $g(E)$, which includes its width, any multimodality, and its broadening by thermal motion at operating temperature. Nothing else in the chain is learned from quantum mechanics.
Research Program
The first stage builds interface structures. We generate disordered passivation layers on crystalline germanium by melt-and-quench AIMD in CP2K. We match the layer's composition to the process the detector manufacturer actually uses; we will obtain this rather than assume it, running an amorphous-germanium variant and an oxygen-containing variant if the process is uncertain. Because direct AIMD is limited to a few hundred atoms and picoseconds, we train a MACE equivariant interatomic potential on the AIMD frames. We then use it to produce many larger, independently quenched interface cells and to run long dynamics at 77 K. We validate the potential against held-out AIMD energies and forces, and against radial distribution functions, before trusting it.
The second stage computes trap levels. On a few hundred sampled snapshots, we compute localized in-gap states and their charge-transition levels with a hybrid functional. Hybrid rather than semilocal functionals are necessary because semilocal functionals close germanium's small band gap. Snapshots taken along the 77 K trajectory give the dynamic broadening of each level. The output is an empirical $g(E)$ with bootstrap uncertainty, reported separately for each surface chemistry. Before committing to the pulse stage, we check whether $g(E)$ is stable across functionals up to a rigid shift. That stability is the property the whole method relies on, and if it fails we stop and report it.
The third stage adds the delayed-charge layer to siggen. We fit $\Delta$, $f$, and the electronics response on a calibration subset of surface-scan positions. We then simulate surface-alpha and bulk waveforms with exact labels. Finally, we train the small autoencoder-plus-classifier architecture of Deep learning based pulse shape discrimination for germanium detectors on those labels and test it on real data.
Baselines
The naive reference is ideal siggen followed by the fitted electronics response with no trapping. It reproduces the 0% tail-slope HIoU reported by Ad-hoc Pulse Shape Simulation using Cyclic Positional U-Net and anchors the scale.
The strongest simulation competitor is CPU-Net itself, with reported HIoU of 83.2% on tail slope and 36.8% on maximum amplitude on BEGe thorium-228 data. We will retrain it with its published architecture on our data. We will give it the same real waveforms we use for calibration, including the surface-scan events it previously lacked, and a matched hyperparameter search budget. That way any advantage we show is not simply a data advantage. Beating it means equal or higher HIoU on bulk events, clearly higher HIoU for DCR on surface events at held-out scan positions, and no waveforms outside the physical envelope. We define "outside the envelope" operationally as tails that are non-monotone, or delayed charge exceeding the deposited charge.
The strongest discrimination competitor is the $A/E$ plus DCR cut of α-event characterization and rejection in point-contact HPGe detectors. It is reported to reject 100% of alphas beyond 3 mm from center with 0.2% bulk loss. We will apply it unchanged to the same test events. Because it is already at its ceiling in that region, a simulation-trained classifier "beats" it only by improving rejection at positions within 3 mm of center, where the file gives no reported figure, or by lowering bulk loss at equal rejection. Matching it everywhere while also supplying labeled simulations and uncertainty bands counts as parity, not a win, and we will say so.
The key ablation isolates the atomistic ingredient. We replace the AIMD-derived $g(E)$ with a flexible empirical distribution, a mixture of two to three Gaussians in $E$ with all parameters fit to the same calibration positions. This model has more free parameters than our one-offset version. If it extrapolates to held-out positions and detectors as well as the physics-shaped version, the quantum calculations added nothing.
Why not just…?
Why not just fit an empirical trap model to the surface scans? This is the real threat, and it is exactly the ablation above. A few-parameter release-time distribution fit directly to scan data may reproduce DCR at the fitted positions just as well. Our bet is about extrapolation. The empirical fit has no reason to hold at unscanned radii, on a second detector with the same passivation process, or at a different operating bias. The physics-shaped model changes only one number there. If the held-out comparison shows no gap, we will report that the atomistic stage is unnecessary for this detector class.
Why not just train CPU-Net with surface-alpha data included? Its authors attribute its surface-event failures partly to missing physics, not only to missing data. Adding scan data is a fair fix, so our CPU-Net baseline gets it. What remains is that an unconstrained translator can still produce non-physical waveforms in rare regions, and it gives no parametric handle for systematic uncertainty. We measure the first directly with the envelope-violation rate. The second is a qualitative advantage that we will not oversell.
Why not just keep A/E plus DCR, which already reaches 100%? For deployed rejection at large radii, nothing here will beat it. The value lies elsewhere. The first is the near-center region. The second is replacing impure calibration labels, which Deep learning based pulse shape discrimination for germanium detectors identified as the bottleneck for learned discriminators. The third is estimating how many surface events leak into the signal region at positions calibration never covers.
Decisive Checks
The first check comes before any detector data is used. The shape of $g(E)$ must survive a change of functional up to a rigid shift, and it must survive a change of passivation model within the manufacturer's plausible range. If it does not, the core premise fails.
The second check is held-out extrapolation on the surface-scan dataset of α-event characterization and rejection in point-contact HPGe detectors. We calibrate on alternating radii and test on the rest, and, if collaboration access permits, on a second detector. Metrics are DCR and $A/E$ HIoU in the same convention as CPU-Net, plus the envelope-violation rate. The physics-shaped model must beat the free empirical mixture and retrained CPU-Net at held-out positions.
The third check validates on bulk events. Using the BEGe thorium-228 data, the added trap stage must not degrade bulk-event HIoU relative to the fitted electronics response alone. Here a classifier trained only on simulated labels must reach the 90%/90% operating point of Deep learning based pulse shape discrimination for germanium detectors on real double-escape versus full-energy events. It must also be reported against $A/E$ plus DCR for surface rejection versus bulk loss by radius.
A positive result on the second check, with parity on the third, would be enough to justify extending the method to other passivation chemistries. A null result on the second check would settle, with evidence rather than argument, the dispute between CPU-Net's authors and the experimental collaborations over whether first-principles modeling pays off here.
References
- $$\alpha $$-event characterization and rejection in point-contact HPGe detectors
- Ad-hoc Pulse Shape Simulation using Cyclic Positional U-Net
- Deep learning based pulse shape discrimination for germanium detectors
- Silicon Liquid Structure and Crystal Nucleation from Ab Initio Deep Metadynamics
- Hydrogen Oxidation Reaction at the Ni/YSZ Interface: An Ab Initio Study
- Predicting in silico electron ionization mass spectra using quantum chemistry
- The liquid-argon scintillation pulseshape in DEAP-3600