Searches for extremely rare events, such as neutrinoless double-beta decay in germanium crystals, depend on telling a true signal pulse from a background pulse by its shape. A tiny neural network of 141 parameters already does this as well as the hand-built physics cut. A separate line of work stacks non-silicon materials (spintronic junctions, ferroelectrics, phase-change films) under image sensors so that a sensor can compute in place, without converting every pixel to digital. The natural combination is to run the pulse-shape classifier inside the detector on such hardware. That hardware is low-precision and noisy, and it brings heavy elements such as lead and bismuth that can be radioactive into the most radiopure environment in experimental physics. No one has checked whether the discrimination survives the move, or whether the move is worth anything.
This note proposes a co-design study that answers both questions with one accounting rule. We emulate the documented non-idealities of candidate in-sensor devices on real germanium pulse-shape tasks. We find the lowest weight precision and the highest switching error at which a classifier still holds the published operating point. We then score every hardware option by the net background it leaves: the background it removes minus the background its own materials add. The deliverable is a precision floor for analog waveform discrimination, a training recipe that lowers that floor, and a go/no-go map saying where heterogeneous in-sensor computing beats plain silicon or offline software for rare-event detectors.
Core Claim
We make two claims, and we expect at least one to be uncomfortable for the integration agenda. First, temporal pulse-shape discrimination is more fragile to analog quantization and stochastic activation than the image tasks on which in-sensor hardware has been validated. The discriminating information sits in small features of the waveform tail and secondary current peaks, not in coarse spatial structure. Training that injects device noise should recover most of the lost discrimination down to roughly 4-bit weights.
Second, for low-rate rare-event detectors, the justification for exotic in-sensor computing cannot be bandwidth. It has to show up as a better signal efficiency per square root of residual background once the device's own radioactivity is counted. Where it does not, the right answer is a conventional silicon circuit or offline software, and we will say so.
Context and What Is Missing
The discrimination side supplies the tasks and the targets. Deep learning based pulse shape discrimination for germanium detectors compresses 256-sample waveforms into a 7-number summary with an autoencoder, a network trained to reconstruct its own input through a narrow bottleneck. A 141-parameter classifier on that summary reaches 90% rejection of multi-site gamma background at 90% acceptance of single-site signal. It matches the standard amplitude-over-energy ratio (A/E) and catches multi-site pulses that fool A/E. The proposal reuses this method, but the paper ran only in floating-point software.
α-event characterization and rejection in point-contact HPGe detectors shows how far closed-form physics goes. Its delayed charge recovery parameter measures slow charge release in the millisecond tail. Combined with A/E, it rejects all surface alpha events within statistics while losing 0.2% of bulk events. That result constrains the proposal: any learned in-sensor discriminator has to beat a cut that is already close to perfect on that background. Machine learning-powered data cleaning for LEGEND provides a second task, rejecting non-physical events in electronics data that mimics the LEGEND-200 readout, including high-rate data with pileup. That work is limited to about 10,000 training events and leaks 15.034% of anomalous events in the high-rate set.
The integration side supplies the hardware constraints. Voltage controlled magnetic tunnel junction based ADC less processing-in-pixel stacks voltage-controlled magnetic tunnel junctions, magnetic elements that flip state when an analog sum crosses a threshold, under a 22 nm image sensor. The junctions replace analog-to-digital converters. Their raw switching error is about 7%, pushed below 0.1% only with eight-fold redundant voting. The ferroelectric array in In-sensor image memorization, low-level processing, and high-level computing by using above-bandgap photovoltages stores 4-bit weights (16 levels) but is slow and high-voltage in bulk ceramic form. Waveform Driven Plasticity in BiFeO3 Memristive Devices shows bismuth ferrite devices responding natively to time-overlapping voltage waveforms, but only on isolated devices.
All three were validated on images or synthetic spikes at room temperature, never on detector pulses. That gap motivates the study. The radiopurity demands in The MAJORANA Project, a target below 1 count per tonne-year in the signal region, constrain it. The lead in PLZT and the bismuth in bismuth ferrite are exactly the kinds of elements that carry long-lived radioactive isotopes.
Research Program
The first component is a device-emulation layer for waveform classifiers. Each candidate hardware target becomes a set of perturbations applied during inference.
The ferroelectric target gets weights snapped to 16 levels, with level-to-level programming spread. The spintronic target gets binary threshold activations that flip with a configurable probability, from 7% raw down to the post-voting rate, with redundancy charged as area and energy. The memristive target gets multiplicative conductance noise and drift. A conventional silicon fixed-point target uses the same bit widths but has no stochasticity.
Cryogenic behaviour is unknown, because all these devices were characterized at 300 K. We therefore treat threshold shift and level drift as swept parameters rather than pretending to know them.
The second component is the discriminators. We take the Holl two-stage autoencoder-plus-classifier and two smaller temporal models that map naturally onto analog summation. The first is a bank of a few learned causal filters followed by a threshold, essentially a learned generalization of A/E's peak-current-over-energy. The second is a tail integrator generalizing delayed charge recovery.
We train each model three ways: in floating point, with post-training quantization, and with noise-injected training through the emulator. The third uses a sparsity regularizer of the kind used to protect binarized activations in the spintronic paper.
The third component is the net-background figure of merit. In a background-dominated counting search, sensitivity scales with signal efficiency over the square root of background, because the background fluctuation is Poisson. We therefore score each configuration by
$$ F = \frac{\varepsilon}{\sqrt{(1-r)\,B_0 + B_{\text{dev}}}}, $$
where $\varepsilon$ is signal acceptance, $r$ is the fraction of the original background $B_0$ that is rejected, and $B_{\text{dev}}$ is the background added by the in-sensor computing layer.
$B_{\text{dev}}$ is the film mass times its assayed activity times the probability that a decay deposits energy in the signal window. The last factor comes from a simple geometric simulation. We have no assay numbers, so $B_{\text{dev}}$ will initially be reported as a break-even activity: the largest contamination per gram at which a given device still improves $F$. That number is directly actionable for anyone who later assays these films.
Baselines
The naive reference is no pulse-shape cut at all: full signal acceptance and zero rejection. Every $F$ value is normalized to it.
The strongest physics baseline is A/E combined with delayed charge recovery, as in α-event characterization and rejection in point-contact HPGe detectors. It rejects 100% of surface alphas within statistics at 0.2% bulk loss, and it needs only a peak detector and a tail integrator. We will run it on the same waveforms with its published tuning.
The strongest learned baseline is the floating-point Holl classifier, reported at 90% background rejection at 90% signal acceptance on the thorium-228 double-escape versus full-energy peak task. We retrain it on our data split with the same hyperparameter search budget given to every variant. Beating it in floating point is not the goal. The goal is to match its 90/90 point under device constraints where plain post-training quantization falls below it.
The decisive ablation is the same architecture under the silicon fixed-point target at equal bit width. This isolates whether stochastic analog devices add anything beyond low precision. For the data-cleaning task, the baseline is the LEGEND wavelet, clustering and support-vector pipeline, with its reported 0.024±0.004% sacrifice and 15.034±0.224% high-rate leakage. Beating it means lower leakage at no higher sacrifice on the same high-rate test data.
Why not just…?
Why not just digitize everything and run the classifier offline or on an FPGA? For germanium double-beta-decay searches this is the main threat, and it may win. Physics rates are tiny, and the FPGA route in Intelligent experiments through real-time AI already meets sub-10 µs latency at far higher rates. Bandwidth is not a credible argument here. The study tests whether moving computation into the detector buys anything in $F$, for example by cutting readout cabling mass near the crystal, or during high-rate calibration runs with pileup. If it does not, offline software is the recommendation.
Why not just use the analog A/E and charge-recovery cuts, which are already near perfect? For surface alphas they essentially are. Our learned variants must therefore show gains specifically on the multi-site population that Holl et al. found fools A/E, and on high-rate pileup. If the noise-trained filter bank cannot beat analog A/E on that subset at equal acceptance, the learned in-sensor discriminator has no reason to exist.
Why not just build the 141-parameter network in ordinary silicon CMOS? This is plausible, and it is why the silicon fixed-point target is an ablation rather than an afterthought. Silicon carries none of the lead or bismuth radiopurity risk and none of the cryogenic uncertainty. Exotic devices earn a place only if, at equal $F$, they reach a regime that silicon cannot, such as non-volatile weights without refresh or no converter at all. We expect silicon to win in many cells of the map, and reporting that is a result.
Proposed Experiments
The first experiment is the precision floor. On the thorium-228 double-escape versus full-energy peak task, we sweep weight levels from 64 down to 4 and activation flip probability from 0 to 10%. We record background rejection at fixed 90% signal acceptance and the area under the receiver operating curve for all three training regimes. Calibration waveforms from a point-contact germanium detector are the target data. If the collaboration data is unavailable, we use the TUBE-style alpha scans and Shockley-Ramo pulse simulation as a stand-in, and report results separately. The decisive check is whether noise-injected training holds 90/90 at 16 levels with 7% raw flips, a regime where post-training quantization fails. If it does not, the spintronic target is excluded without redundancy.
The second experiment repeats the sweep on the high-rate and low-rate LEGEND-style electronics data, where pileup makes temporal structure matter most. It tests whether the tail-integrator model lowers anomalous leakage below 15.034% at no added sacrifice.
The third experiment sweeps the cryogenic threshold drift parameter to find the drift tolerance of each model. That gives device groups a concrete spec for measurements at 77 K.
The fourth experiment computes the break-even contamination for each hardware target from $F$. That yields the final map of where heterogeneous in-sensor discrimination beats silicon, analog physics cuts, and offline software.
References
- Deep learning based pulse shape discrimination for germanium detectors
- $$\alpha $$-event characterization and rejection in point-contact HPGe detectors
- Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines
- Voltage controlled magnetic tunnel junction based ADC less processing-in-pixel for extreme edge intelligence
- In-sensor image memorization, low-level processing, and high-level computing by using above-bandgap photovoltages
- Waveform Driven Plasticity in BiFeO3 Memristive Devices: Model and Implementation
- The MAJORANA Project
- Intelligent experiments through real-time AI: Fast Data Processing and Autonomous Detector Control for sPHENIX and future EIC detectors