Spending Digitizer Power Only When Pulses Are Hard to Tell Apart

Radiation detectors that tell neutrons from gamma rays do it by pulse shape discrimination: they look at the shape of each digitized electrical pulse, for example how much light arrives early versus late. How well this works depends on how finely the pulse is digitized, meaning how fast the electronics sample it and how many bits each sample carries. Fine digitization costs power. On one CLYC scintillator testbed, a 500-million-samples-per-second converter draws 858 mW, against 101 mW at 100 million samples per second (Optimization of energy resolution and Pulse Shape Discrimination for a CLYC detector with integrated digitizers). Today that setting is fixed at design time. The instrument is then either over-provisioned for quiet periods or under-provisioned for bursts of high radiation. In those bursts, pulses overlap ("pileup"), and simple discriminators fall below 90% accuracy (Model-Based Deep Learning Algorithm for Pulse Shape Discrimination in High Event Rates).

We propose to treat digitizer configuration as a small, deliberately restricted action space controlled online. Every few milliseconds, a controller may move the acquisition settings at most one rung up or down a short ladder of (sampling rate, bit depth, integration gate) configurations. It is rewarded for meeting a target misclassification rate at the lowest energy. The controller does not decide from count rate alone. It decides from how much of the recent traffic sat near the classifier's decision boundary, which we call the "ambiguity mass." The central bet is that the need for fine digitization tracks this ambiguity, not raw flux. A controller that watches ambiguity should therefore beat both static settings and the obvious rate-triggered switch.

Core Claim

The CLYC digitizer study establishes an "equivalence principle": for quantization noise, one extra effective bit is worth a fourfold increase in sampling rate. Discrimination quality then follows the product $F_s \times 4^{\text{ENOB}}$, where $F_s$ is the sampling frequency and ENOB the effective number of bits. A figure of merit above 3.0 (gamma rejection better than $10^{-12}$) requires a product of at least $2.98 \times 10^7$ MHz, and there is almost no gain beyond $3 \times 10^8$ MHz (Optimization of energy resolution and Pulse Shape Discrimination for a CLYC detector with integrated digitizers). That result was obtained on isolated pulses. Our first claim is that the equivalence principle breaks under pileup. Separating overlapping pulses is a timing problem, and extra bits cannot replace temporal resolution there. So the right configuration depends on the overlap structure of the current traffic, not only on a single fixed product.

Our second claim follows from the first. If the cheap setting is enough for isolated pulses and the expensive one matters only for ambiguous or overlapping ones, then a controller that escalates on observed ambiguity will reach a given error target at substantially lower average power. It should also beat one that escalates on count rate, because equal rates can carry very different ambiguity. A neutron-rich field and a gamma-dominated one at the same rate are an example.

The "restricted action space" thread enters as the controller's design, not as a metaphor. TempoRL showed that reinforcement learning on physical pulse-shaping hardware converges safely when each step may change a parameter by at most 10% of its range (TempoRL: laser pulse temporal shape optimization with Deep Reinforcement Learning). We adopt the same kind of bound, as a one-rung-per-interval ladder, for a different reason: converters, clocks and gate logic need settling time, and a jumpy schedule loses data. TempoRL shapes a laser pulse; we decide how to observe pulses. The bound is the same kind of object, but it serves a different purpose.

Research Program

The environment is built from real waveforms. We would record CLYC pulses at the top configuration (500 MSPS, 16 bits) on a testbed like the ZYNQ-based board of the CLYC study, using $^{137}$Cs and $^{60}$Co for gamma rays and moderated $^{252}$Cf for mixed neutron–gamma fields. Lower configurations are emulated by decimating and requantizing these recordings, as that study's parameter scan effectively did. Pileup is created by superposing recorded single pulses at Poisson arrival times, at rates from 500 kcps to 10 Mcps. This mirrors the regime of the model-based pileup work, but uses measured rather than simulated pulse shapes, which addresses that work's own simulation-to-real caveat.

A flux schedule then drives each episode. These are time-varying traces of rate and neutron fraction, with quiet stretches, bursts and drifts. Power per configuration is taken from bench measurements, anchored to the reported 101 mW and 858 mW endpoints. Settling time is also measured on the board.

For each rung of the ladder we train a per-pulse classifier with the same architecture and tuning budget. We use the model-based design that reweights physics features (rise time, template correlation) with a context network (Model-Based Deep Learning Algorithm for Pulse Shape Discrimination in High Event Rates). We also try a compact autoencoder-plus-tiny-classifier variant, whose 141-parameter classification head is small enough for an FPGA (Deep learning based pulse shape discrimination for germanium detectors). The controller sees per-interval summaries of the classifier output:

  • count rate;
  • a histogram of estimated pileup multiplicity;
  • the fraction of pulses whose classifier probability lies within a margin of 0.5 (the ambiguity mass);
  • the running neutron fraction.

Its actions are up, down or stay on the ladder. The objective is a constrained one:

$$\min_\pi \; \mathbb{E}\Big[\sum_t P(a_t)\Big] \quad \text{s.t.} \quad \mathbb{E}\big[\epsilon_t\big] \le \epsilon^\ast,$$

Here $P(a_t)$ is the measured power of the configuration in interval $t$, $\epsilon_t$ is the misclassification rate in that interval, and $\epsilon^\ast$ is the target. We solve it with a Lagrangian penalty.

Because the emulated environment knows the future flux, we can compute the exact best ladder schedule by dynamic programming, under the same one-rung constraint. This oracle schedule plays the role that "Oracle Reveal" plays in Reveal-and-Click. That benchmark found that uncovering hidden targets lifted Qwen3-VL-8B from 30.8% to 81.9%. This showed that failures came from not acquiring information, not from misreading it. Our analogue splits the total error into two parts. One is information never acquired, which is the gap between cheap and full-resolution classifiers. The other is the scheduling loss of a causal controller relative to the oracle.

We train the controller in two stages: behavioral cloning of the oracle schedule, then reinforcement learning. Following GUI Exploration Lab, we keep the cloning stage short, about one epoch. That paper found that longer supervised training memorized fixed paths and left little room for the later reinforcement-learning stage to improve the policy. We test on flux schedules whose statistics differ from training.

Baselines

The naive reference points are the two static extremes, evaluated with each one's own retrained classifier on identical waveform streams and flux schedules. "Always high" (500 MSPS, 16-bit) bounds achievable accuracy at maximum power. "Always low" (100 MSPS) bounds power at minimum accuracy. Every adaptive method must lie strictly inside the region these two define in power–error space. We have no reported number for the always-high setting under pileup. For pileup accuracy, the strongest reported reference is the model-based deep classifier at about 98.7% accuracy at 10 Mcps with three or more overlapping pulses. That paper reports up to 8 times fewer errors than normalized cross-correlation and 1.7 times fewer than a purely data-driven network, and charge integration and cross-correlation fall below 90% (Model-Based Deep Learning Algorithm for Pulse Shape Discrimination in High Event Rates). We reimplement it as the per-rung classifier, so every controller gets the same, strongest discriminator.

The strongest competing controller is a count-rate hysteresis rule. It steps up the ladder when rate exceeds a threshold and down when rate falls below a lower one, with thresholds tuned by grid search on the training schedules. We give it the same search budget as the learned controller's hyperparameter search. It is the strongest competitor because pileup probability is a known function of rate, and any instrument engineer would build this rule first. No reported result for adaptive digitizer scheduling exists in the file, so this baseline has no external number. Beating it means at least 20% lower average energy at matched error target $\epsilon^\ast$ on held-out schedules, with no more constraint violations.

Two ablations isolate the key ingredients. An ambiguity hysteresis rule uses the same two-threshold logic as the rate rule, driven by ambiguity mass instead of rate. It separates "is ambiguity the right signal" from "does learning help." An unconstrained controller may jump to any rung in one step, and we measure its loss from settling-time dead windows. This tests whether the TempoRL-style bound costs anything or protects anything. The dynamic-programming oracle gives the ceiling.

Why not just…?

Why not just run the 500 MSPS converter all the time? For accuracy this cannot be beaten, and it is the right answer when power is free. The CLYC study shows it is not free: 858 versus 101 mW. That gap matters for space instruments such as the CXPD Cubesat pathfinder and for heat-limited cryostats. Beyond a product of $3 \times 10^8$ MHz it also buys almost nothing on isolated pulses. Our power–error curves show directly how much accuracy, if any, the always-high setting retains over the adaptive controllers at each flux level.

Why not just run the cheap converter with the best classifier? This is a real threat. Under the equivalence principle, a low-rate converter with good effective bit depth may already clear the $2.98 \times 10^7$ MHz threshold on isolated pulses. If it also holds up under pileup, adaptive control is pointless. Experiment one tests exactly this. If the per-rung classifier at 100 MSPS matches 500 MSPS accuracy at every pileup rate, we report the negative result and the proposal collapses to "use the cheap converter."

Why not just switch on count rate? Rate predicts pileup, so this simple rule may capture most of the benefit. We include it as the main controller baseline, and we deliberately build test schedules where rate and ambiguity come apart: same rate, different neutron fraction or pulse-shape spread. If the rate rule matches our controller even there, ambiguity sensing adds nothing and we say so. If the ambiguity hysteresis rule matches the learned controller, the contribution is the signal, not the reinforcement learning, and we would recommend the simple rule.

Why not just use Bayesian optimization to pick settings? It returns one fixed configuration per calibration. That cannot follow flux that changes on millisecond-to-second timescales. Its proposed jumps also ignore settling cost, which is the same failure TempoRL reports for unconstrained black-box optimizers on hardware. The unconstrained-jump ablation quantifies the settling cost.

Proposed Experiments

The first decisive check tests whether the equivalence principle survives pileup. Using full-resolution recordings, we compare pairs of configurations with equal $F_s \times 4^{\text{ENOB}}$ but different sampling rates, at 0.5, 2, 5 and 10 Mcps. For each pair we report misclassification, figure of merit, and gamma rejection at fixed neutron acceptance. If equal-product pairs diverge as rate rises, with the higher sampling rate winning, temporal resolution has value that bits cannot buy. That justifies a sampling-rate ladder. If they do not diverge, the ladder reduces to bit depth and the case for fast switching weakens.

The second check is the controller comparison on held-out flux schedules. We plot energy against interval-level error for the static settings, rate hysteresis, ambiguity hysteresis, the learned controller and the oracle. The headline is energy saved at matched $\epsilon^\ast$ relative to rate hysteresis, together with the fraction of the static-to-oracle gap each method closes. We also report results separately on the rate–ambiguity decoupled schedules.

The third check deploys the best controller on the physical board with live $^{252}$Cf and gamma sources. We vary flux by moving the sources, measure true power draw and settling losses, and confirm that the emulated savings hold on hardware.

A successful outcome is a measured, hardware-validated power reduction at fixed discrimination quality, attributable to ambiguity sensing under a bounded action ladder. A clean negative on experiment one, where cheap digitization plus a strong classifier suffices under pileup, would itself be a useful design rule for low-power instruments.