Position-Resolved Reduced Models of Germanium Pulses for Bulk Signal Acceptance

Searches for neutrinoless double-beta decay in germanium detectors keep a candidate event only if its electrical pulse looks like a single, localized energy deposit. The standard test is the ratio of peak current to energy, called $A/E$. The fraction of true signal that survives this cut is calibrated with an external radioactive source, and that source mostly lights up the outer shell of the crystal. Real decays happen uniformly throughout the bulk. In the MAJORANA DEMONSTRATOR this mismatch is the main reason the signal acceptance carries a ±3.5% systematic uncertainty (Alvis et al.). Removing it requires a pulse simulator that is faithful to the real hardware at interior positions where no calibration data exist.

We propose to build that simulator as a non-intrusive reduced-order model, meaning a small dynamical model learned only from the inputs and outputs of an existing solver and of the detector, without access to the solver's internals. The model has two parts with fixed roles. The first is a compressed map from interaction position to the ideal induced current, learned from queries to the first-principles field and drift code siggen. The second is a learned operator with memory that converts ideal current into the waveform the digitizer actually records. It is shared across positions and fitted to unlabeled calibration pulses. The bet is that this split lets hardware effects learned near the surface carry over to the bulk, something an unstructured translation network cannot do because it never sees positions at all.

Core Claim

The reality-gap corrections used today are not tied to interaction position. Li et al. use CPU-Net, an adversarial network that translates idealized siggen pulses into realistic ones. It learns the translation only from the event population the calibration source provides. The empirical $A/E$ correction of Alvis et al. has the same limitation. Neither can say what the correction should be deep in the crystal.

Our claim is that most of the gap comes from the readout chain and slow charge release. Both are much less position-dependent than the drift physics that siggen already handles. If that is true, a model that keeps position as an explicit parameter in the drift part and learns a shared response part will predict bulk single-site acceptance more accurately than calibration-based extrapolation, and with an uncertainty one can actually compute.

What is new is not the use of a surrogate. It is the factorization, together with a procedure for fitting it to data in which positions are unknown. The test that settles the claim is extrapolation in position, not waveform realism on the calibration set.

Research Program

The drift part. We query siggen as a black box on a dense grid of interaction positions, covering radius, height and azimuth (azimuth matters because charge mobility depends on the crystal axis). This yields ideal induced-current pulses. Across positions these pulses differ mainly by a shift in arrival time and a change in rise shape. That is a transport-type variation, and linear compression handles it badly. Plain proper orthogonal decomposition (a principal-component basis of the simulated pulses) needs many modes to represent it, which is the Kolmogorov barrier described for advection problems.

We therefore compress with the quadratic manifold of Geelen et al., who reconstruct the state from both linear and pairwise-product terms of a reduced coordinate. They report that a quadratic model with $r=30$ matched a linear one with $r=60$ on transport problems. We then interpolate the reduced coordinates over position, in the spirit of the parametric surrogates of Constantine et al.. The result is a pulse generator for any position that is cheap and differentiable. As a simpler control we also try aligning pulses by drift time and then compressing linearly.

The response part. The recorded waveform $y_k$ at time step $k$ is a single-electrode observation of a much larger hidden system. That system includes the preamplifier, RC discharge, and populations of trapped charge. Uy and Peherstorfer show that when only part of a system is observed, a reduced model needs memory terms. Their delay-embedded operator inference matched, using 20% of the state, the accuracy that a memoryless model reached with 99%. We adopt their form:

$$\hat{x}_{k+1}=\sum_{l=0}^{L}\left(\hat{A}_l\hat{x}_{k-l}+\hat{B}_l u_{k-l}\right)+\tau(t_d)\,\hat{C}\,\hat{x}_k,\qquad y_k=\hat{c}^{\top}\hat{x}_k .$$

Here $\hat{x}_k$ is a small latent state of dimension $r$, $u_k$ is the ideal current from the drift part, $L$ is the number of lag steps, and $\hat{A}_l,\hat{B}_l$ are shared linear operators learned by least squares. The last term carries a learned matrix $\hat{C}$ scaled by a low-order function $\tau(t_d)$ of the drift time $t_d$. It is the only place position enters the response, and it is meant to absorb trapping that grows with path length. Surface slow release, which produces delayed charge recovery (DCR), should appear as long-lag structure in $\hat{A}_l$.

One honest observation: without the latent state and without $\tau$, this model reduces to classical ARX system identification. That fact defines one of our ablations rather than being hidden.

Fitting without positions. Calibration pulses come with energies but no positions. We fit by alternating two steps. First, for fixed operators, we infer each event's latent position by minimizing reconstruction error through the cheap drift part. Second, for fixed positions, we refit the operators by regularized least squares, as in operator inference. Charge induction is linear in the deposited charge, so a multi-site event is modeled as a sum of two single-site drift pulses with unknown energy split, passed through the same response. This lets full-energy-peak events, which are mostly multi-site, constrain the response operators as well, instead of serving only as background labels.

Using the model. With the model in hand, we generate pulses for decays placed uniformly through the bulk, add noise sampled from baseline traces, and apply the experiment's actual $A/E$ cut. Repeating the fit under bootstrap resampling and over choices of $r$ and $L$ gives a spread on the acceptance, which we report as its uncertainty.

Baselines

The naive reference is the acceptance read directly from the calibration double-escape peak, with no position correction. This is in effect what the ±3.5% systematic is assigned to cover. The reported figure is single-site acceptance of $(90\pm3.5)\%$ (Alvis et al.). On synthetic data with known truth we measure its bias directly. Any method that does not shrink that bias is not worth deploying.

The strongest existing simulator is CPU-Net (Li et al.). It is the only published system that closes the siggen-to-data gap for these detectors. Its reported histogram-intersection-over-union (HIoU) is 83.2% on the tail-slope parameter and 36.8% on peak current, starting from 0% and 12.4% for raw siggen. We retrain it on the same 65,536 siggen pulses and the same BEGe thorium calibration set, with the same validation split and a comparable hyperparameter search budget.

Beating it on realism means exceeding 36.8% HIoU on peak current, the quantity $A/E$ depends on, without falling below 83.2% on tail slope. Beating it on the actual question means a smaller error in predicted bulk acceptance on the hidden-truth benchmark described below.

The strongest discriminator is the autoencoder classifier of Holl et al.. It compresses 256-sample pulses to 7 latent features and reaches 90% full-energy-peak rejection at 90% double-escape-peak acceptance. We include it to check that pulses from our simulator, used as training labels, do not lose discrimination relative to calibration labels. Matching its 90/90 point while removing the position bias counts as success.

Classical pulse fitting. A skeptic's direct competitor is siggen plus a hand-specified electronics transfer function and a single trapping constant, fitted to data. We run it with the same position-inference loop and the same optimizer budget.

Ablations. Four ablations isolate the key ingredients: removing memory ($L=0$), removing the drift-time-dependent trapping term, collapsing the latent state to the ARX limit, and replacing the quadratic manifold with linear compression.

Why not just…?

Why not just keep $A/E$ and widen the systematic? The scalar cut works well and is cheap. A ±3.5% acceptance uncertainty propagates directly into the half-life limit, and tonne-scale experiments aiming below one count per tonne-year (Elliott et al.) will be limited by it. We do not replace $A/E$. We supply its bulk efficiency, so the online cut remains as cheap as it is now.

Why not just condition CPU-Net or another deep generator on position? Calibration data carry no positions, so there is nothing to condition on. A generator trained only on peripheral events has no grounds for its behavior in the interior. Its documented artifacts and weak peak-current fidelity (Li et al.) are worse precisely for the quantity that matters here. The hidden-truth experiment trains every method only on peripheral events, which tests this directly.

Why not just fit a parametric electronics model to siggen by hand? This is the real threat. If the hardware gap is a simple linear filter plus one trapping constant, the hand model will tie with ours. Our contribution would then shrink to fast, differentiable position inference. We will report that outcome as it stands. Our model should win only when slow surface release and other effects not foreseen by the hand model matter. Two checks probe this: the DCR-bearing alpha data from the TUBE scanner (Arnquist et al.), and injected trapping in the synthetic test.

Why not just run siggen densely and skip reduction? Dense siggen runs give positions but not realism. On their own they do nothing about the gap. Reduction is what makes per-event position inference and the alternating fit tractable, and the speed gain will be measured rather than assumed.

Proposed Experiments

The decisive experiment is a hidden-truth benchmark. We corrupt siggen output with a known hardware response: a preamplifier filter, RC decay, trapping that scales with drift time, and slow surface release. Training data are drawn only from a calibration-like source distribution concentrated near the crystal surface. Every method then predicts single-site $A/E$ acceptance for uniformly distributed bulk events, and we compare against the true value.

The hypothesis is supported if our model's acceptance error is well below the naive bias and below that of CPU-Net and the hand-fitted model, and if its bootstrap interval covers the truth. If the hand model ties, the factorization's value is computational only.

The second experiment uses the BEGe thorium calibration data. We report HIoU for tail slope and peak current against the CPU-Net figures, and double-escape versus full-energy-peak discrimination against Holl et al.'s 90/90 point.

The third experiment uses TUBE alpha-scan data, which have known surface positions. We test whether learned long-lag memory reproduces DCR and whether a cut based on it reaches the reported full alpha rejection with 0.2% bulk loss (Arnquist et al.). These data are also the one real-data check of position inference.

We will treat the method as failed if inferred positions on the known-position scan data are inconsistent, or if acceptance predictions shift by more than the systematic they aim to replace as $r$ and $L$ vary.