Searches for neutrinoless double-beta decay need germanium detectors that throw away background without throwing away signal. The signal deposits its energy at a single point in the crystal. The dominant gamma-ray background scatters at several points a few centimeters apart. The standard way to tell them apart, a single number called $A/E$ (peak current divided by energy), is tuned on events from an external radioactive source. That source illuminates the detector's outer region, while real signal events are spread uniformly through the bulk. The resulting mismatch is one reason the signal acceptance carries a ±3.5% systematic uncertainty in the MAJORANA DEMONSTRATOR (Multisite event discrimination for the majorana demonstrator). Machine-learning replacements so far still train their final classifier on the same source peaks, so they inherit the same bias.
We propose to remove labels from the discrimination step. Each recorded current pulse is explained as a sparse sum of single-site pulse shapes drawn from a library that covers positions throughout the crystal volume. The network that finds this explanation is trained only by minimizing the explanation's own error plus a sparsity cost, never by matching labels. The discriminant is how much better the best two-site explanation fits than the best one-site explanation. This borrows the core move of "unsupervised objective-based training" from wireless engineering, where networks are trained directly on the physical quantity they should maximize. Here that quantity is a sparse-reconstruction likelihood built from detector physics. Because the pulse library spans the whole crystal by construction, the cut does not depend on where calibration events happen to fall.
Core Claim
A deconvolution objective whose building blocks are physically simulated single-site pulses gives a single-site versus multi-site statistic with three properties. First, it needs no calibration labels to train. Second, it can be checked against bulk-uniform event populations. Third, it can be corrected for simulation mismatch using unlabeled data. We claim it can match the rejection of $A/E$ and of autoencoder-based methods on standard calibration peaks. We also claim it can shrink the difference between how many source-calibration events and how many bulk-uniform single-site events survive the cut, which is the quantity behind the ±3.5% systematic. The novelty is not using a neural network for pulse-shape discrimination, which has been done. It is that the training signal is a physics objective rather than labels or generic reconstruction error, and that the hidden representation is a physical list of deposits (time, position, energy) rather than an abstract feature vector.
Mechanism
Let $y$ be a digitized current pulse. Let $D_\phi$ be a dictionary whose columns are unit-energy single-site pulses for a grid of crystal positions and start times. Let $x \ge 0$ be the vector of energies placed on those columns. The objective is
$$F_\phi(x; y) = \tfrac{1}{2\sigma^2}\|y - D_\phi x\|_2^2 + \lambda \|x\|_1,$$
where $\sigma$ is the measured electronic noise level and $\lambda$ sets the price of adding a site. Two things make this tractable. Fixing positions to a grid makes the problem convex in $x$ for a given $\phi$. The dictionary is $D_\phi = D_{\text{sim}} + \Delta_\phi$: a fixed library from field and charge-drift simulation, plus a small global correction shared by all events.
The solver is iterative shrinkage-thresholding, a simple loop of gradient steps followed by zeroing small entries. We unroll it into a fixed number of network layers and learn only one step size per layer, following Step-LISTA from Learning step sizes for unfolded sparse coding.. That paper trains such unrolled solvers directly on the sparse-coding cost, without ground-truth codes. It shows that learned step sizes can exceed the usual conservative $1/L$ limit, where $L$ is the largest eigenvalue of $D^\top D$, once the solution becomes sparse. It also shows that this minimal parameterization is the right one asymptotically. Ablin et al. supply the method; they only tested it on synthetic Gaussian dictionaries and image patches.
Training minimizes $\mathbb{E}_y[F_\phi(x^{(T)}(y); y)]$ over the step sizes and over $\phi$, on unlabeled physics and calibration waveforms. Here $x^{(T)}$ is the output of the last unrolled layer. The discriminant is
$$\Delta(y) = \min_{x:\,\|x\|_0 \le 1} F_\phi(x;y) - F_\phi(x^{(T)}(y);y).$$
The first term is a one-site fit, computed exactly by scanning the dictionary columns. $\Delta$ is near zero when one site suffices and large when a second site pays for its sparsity cost. Under a Gaussian noise model, $\Delta$ is a penalized log-likelihood ratio between the multi-site and single-site hypotheses. That is why we expect it to use more of the waveform than $A/E$.
The correction $\Delta_\phi$ carries a real risk. A flexible correction could learn to make multi-site pulses look single-site. We guard against this in two ways. The correction is shared across all events, so it cannot fit per-event structure. It is also constrained to a low-dimensional, smooth family of perturbations to the simulated response, with an explicit penalty on its size. The ablations below test whether this is enough.
This design builds on, and departs from, the direct-objective networks in wireless engineering. Unsupervised Learning for Passive Beamforming trains a network on negative received signal-to-noise ratio, with a special output layer that enforces the hardware constraint. Deep Learning for Channel Sensing and Hybrid Precoding in TDD Massive MIMO OFDM Systems shows that a loss derived from the task's own rate expression beats a loss based on intermediate estimation error. That second result is our motivating evidence that reconstruction mean-squared error is the wrong surrogate. Two features have no counterpart in that line of work. Our "constraint layer" is non-negativity together with a physically gridded dictionary. Our objective is also a hypothesis-test statistic rather than a throughput.
Research Program
The work has four parts.
The first builds simulated single-site pulse libraries for a p-type point-contact germanium geometry, the detector type used in MAJORANA and LEGEND. It then generates synthetic event sets with known site counts, positions and energy splits. These include asymmetric splits, where one deposit is much smaller than the other. A dual-phase xenon study found that splits of this kind defeat multi-site rejection, because the dim site is hidden by the bright one (Predicting the single-site and multi-site event discrimination power of dual-phase time projection chambers). That study also provides a template we adopt: compare learned methods against a Cramér-Rao-style lower bound on how well two sites can be separated, here computed in drift time rather than transverse position.
The second part trains the unrolled solver with and without the learned correction, and measures how often $\Delta$ fires on single-site events as a function of depth in the crystal.
The third part applies the trained model to thorium-228 calibration data of the kind used by Multisite event discrimination for the majorana demonstrator and Deep learning based pulse shape discrimination for germanium detectors. The model is never trained on peak labels; the peaks are used only for evaluation. Three peaks matter. The double-escape peak at 1592 keV is nearly pure single-site. The single-escape peak at 2103 keV is mostly multi-site. The full-energy peak at 2614 keV serves as the multi-site sample in Holl et al.
The fourth part tests robustness to non-physical pulses using the LEGEND-200 Full Chain Test waveforms of Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines. The test is whether the residual error of the deconvolution flags electronic artifacts as a by-product.
Baselines
The naive reference point is no pulse-shape cut at all. It gives 100% acceptance and no suppression, and every cut's suppression is quoted relative to it.
The incumbent is the energy-normalized $A/E$ cut of Multisite event discrimination for the majorana demonstrator. It is the physics-motivated standard with a decade of deployment. It reports (90 ± 3.5)% single-site acceptance on the double-escape peak, reduces the multi-site single-escape peak to about 6% of its counts, and suppresses the background in the region of interest by a factor of 3. We will reimplement it on the same waveforms, including its quadratic energy correction, with its threshold set to 90% double-escape acceptance. Beating it means lower single-escape-peak survival at matched 90% double-escape acceptance. On simulation, it also means a smaller gap between double-escape acceptance and bulk-uniform single-site acceptance.
The strongest learned competitor is the two-stage method of Deep learning based pulse shape discrimination for germanium detectors. It trains a convolutional autoencoder to compress each 256-sample pulse to 7 numbers, then trains a 141-parameter classifier on double-escape versus full-energy events. It reports 90% full-energy rejection at 90% double-escape acceptance. We retrain it on identical waveforms. Its autoencoder gets the same unlabeled set and training-step budget as our model, and its classifier gets an equal hyperparameter search. We additionally give it a full-waveform supervised classifier variant, so that our comparison is not only against a 7-number bottleneck. Matching it on the peaks while showing a smaller bulk-versus-calibration acceptance gap counts as winning on the question that matters, because a supervised model is trained on exactly the population whose spatial bias we are trying to avoid.
The key ablation is per-event solving of the same objective with the same simulated dictionary, run to convergence with FISTA (an accelerated version of the shrinkage-thresholding loop), with no learned step sizes and no correction. This isolates what learning adds. A second ablation removes only the correction $\Delta_\phi$, which isolates the effect of correcting simulation mismatch.
Why not just…?
Why not just keep $A/E$ and retune it on simulated bulk events? That would reduce the calibration bias if the simulation were exact. But $A/E$ compresses the pulse to one number, and Holl et al. report multi-site events whose sharp first spike passes the cut while later sub-peaks reveal the second site (Deep learning based pulse shape discrimination for germanium detectors). No retuning recovers information the scalar has thrown away. We will run the retuned-$A/E$ variant anyway. If its single-escape survival and bulk acceptance gap match ours, the waveform-level objective adds little and we will say so.
Why not just train a supervised full-waveform classifier on double-escape versus full-energy peaks? On the peaks themselves it will likely be hard to beat, because those are the very distributions it is fitted to. Its weakness is the one the file documents for every source-trained cut: the training population sits near the detector surface while the signal sits in the bulk. The decisive measurement is therefore the acceptance gap on simulated bulk-uniform single-site events, not peak accuracy. If the supervised model shows no larger gap than ours, the label-free argument collapses to a convenience argument.
Why not just fit each event with the simulated library and skip the network? This is the most serious threat, because it optimizes the same objective. Event rates near the region of interest are tiny, so speed alone does not justify a network there. Learning earns its place only if one of two things happens. Either the correction $\Delta_\phi$, which cannot be learned per event, measurably reduces single-site misidentification caused by simulation mismatch. Or amortization becomes necessary to train the correction jointly over millions of calibration pulses. The FISTA ablation decides this directly. If it ties the full model, we will present the contribution as the self-corrected dictionary plus the $\Delta$ statistic, with the unrolled solver as an optional accelerator.
Proposed Experiments
On simulation, we will report four things: single-site acceptance as a function of interaction depth; the gap between double-escape and bulk-uniform acceptance at a fixed 90% double-escape acceptance; two-site detection probability as a function of site separation and energy split, compared against the computed separation bound; and how these quantities degrade when the true detector response is deliberately perturbed away from the simulated library. On calibration data, we will report single-escape and full-energy survival at 90% double-escape acceptance for all baselines, and the stability of the threshold across energies from 1 to 2.6 MeV.
The decisive checks are three. First, the full model must not exceed the incumbent's ~6% single-escape survival at matched acceptance. Second, on perturbed-response simulations, its bulk-versus-calibration acceptance gap must be smaller than that of the supervised baselines. Third, the correction ablation must show that $\Delta_\phi$ lowers single-site misidentification without raising multi-site survival. A rise in multi-site survival would be the signature of the correction absorbing real multi-site structure.
No reported result was found in the file for unsupervised, objective-trained discrimination on germanium waveforms, so the per-event FISTA ablation also serves as the first reference number for this objective.
References
- Multisite event discrimination for the majorana demonstrator
- Learning step sizes for unfolded sparse coding.
- Unsupervised Learning for Passive Beamforming
- Deep Learning for Channel Sensing and Hybrid Precoding in TDD Massive MIMO OFDM Systems
- Predicting the single-site and multi-site event discrimination power of dual-phase time projection chambers
- 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