Knowing When to Stop Pulsing: Overshoot-Aware Programming of Phase-Change Photonic Weights

Some optical chips store numbers, such as neural-network weights or phase settings, in phase-change materials. These are glassy compounds like antimony trisulfide that sit on a light-carrying waveguide and change their refractive index when a small on-chip heater melts or anneals them. The appeal is that the state persists with zero standing power. The cost is that every heat pulse wears the material: devices survive only thousands of switching events. Fine-grained levels are also reached by trains of pulses whose step size is noisy. Chen et al. (2023) report 0.50 ± 0.16 dB per step, a spread of about one third of the mean step. Partial switching runs only one way, so a pulse that overshoots the target cannot be undone by a small opposite pulse. Recovery takes a full melt-and-reset and a fresh ramp, which spends endurance and time. Near the target, then, the next pulse is a gamble.

Machine translation post-editing has already faced the same problem in another setting. When the input is already good, an eager editor makes things worse. Systems that decide whether and where to edit beat systems that edit everything (Findings of the WMT 2022 Shared Task on Automatic Post-Editing). We propose to carry that lesson of restraint into the controller that programs phase-change photonic weights. Before each corrective pulse, a cheap estimator predicts whether the pulse will help the network's output enough to justify its risk of overshoot. Cells that are good enough are locked and never touched again.

Core Claim

Existing selective-programming methods ask which cells to program carefully. They assume each correction can itself be corrected cheaply. SWIM ranks weights in resistive and ferroelectric memory arrays by a curvature-based measure of how much the network loss depends on each one, and spends repeated write-and-read-back cycles only on the sensitive ones. That measure is the diagonal of the Hessian, the matrix of second derivatives of the loss. ELight reorders and regularizes weights offline so that fewer writes are needed when a photonic tensor core is reused across layers. Neither method models the defining asymmetry of phase-change photonics: an overshoot costs a reset.

Our claim is that, for stochastic one-directional pulse programming, the stopping rule matters as much as the selection rule. The stopping rule should compare the expected loss improvement from one more pulse against the pulse's wear and the probability-weighted cost of an overshoot reset. Concretely, for cell $i$ with current transmission error $e_i$, sensitivity $h_i$ (SWIM-style Hessian diagonal), and a random step $s$ drawn from that device's measured step distribution, the controller fires another pulse only if

$$h_i\,\mathbb{E}_s\big[e_i^2-(e_i-s)^2\big] \;>\; c_{\text{pulse}} + P_s\big(e_i - s < -\tau_i\big)\,c_{\text{reset}},$$

Here $c_{\text{pulse}}$ and $c_{\text{reset}}$ are the endurance and time costs of a pulse and of a full reset-and-ramp, and $\tau_i$ is how far past the target the cell may land before a reset becomes necessary. If the inequality fails, the cell is locked. This mirrors how QE-assisted constrained decoding freezes words tagged as correct so that later search cannot damage them. The prediction is that this rule yields fewer resets and fewer total pulses than SWIM-style selection with a fixed tolerance, at equal accuracy. The advantage should be largest exactly where step noise is comparable to the remaining error.

Research Program

The work has three parts.

Characterizing the step distribution. We first measure each cell's pulse-step distribution and how it depends on the cell's current level. This uses the device class of Chen et al. (2023): Sb₂S₃ on silicon PIN-diode heaters, with 5-bit operation. From these measurements we build a calibrated stochastic simulator. The reported ±0.16 dB spread is the anchor. We also fit level dependence, drift after programming, and thermal disturbance of neighboring locked cells. The simulator lets us test the method at network scale even before large arrays exist. Every claim is then re-checked on whatever physical devices are available.

The controller. This is the "hardware quality estimator." It reads a monitor photodiode tap after each pulse and updates a small Bayesian posterior over that cell's step distribution, so early pulses on a cell inform later ones. It then applies the decision rule above. Sensitivities come from SWIM's single-pass second-derivative approximation.

Borrowed post-editing lessons. We take three findings from post-editing as explicit design constraints. The first is error propagation: a falsely "OK" quality tag locks in an error. In our setting, monitor noise can falsely lock a mis-set cell, so we budget a final sparse audit of locked high-sensitivity cells. The second is the "do nothing" reference point: post-editing systems needed more than 25,000 human-edited examples before they reliably beat leaving the translation untouched (Can Automatic Post-Editing Improve NMT?). The analogue here is a single open-loop write. The third is that a separate gate placed in front of an editor underperformed a model trained jointly to edit and to estimate quality (Quality Estimation-Assisted Automatic Post-Editing). That motivates deciding the stop inside the programming loop rather than as a separate pre-filter.

An optional variant tests the tension between analog and segmented heaters. It pairs coarse deterministic binary segments, as in Chen et al. (2024), with a single analog fine-trim segment governed by the controller.

Baselines

The naive reference point is open-loop programming: each weight is written with its nominal pulse count and never verified. This sets the floor on accuracy and the minimum pulse count. A second reference is exhaustive write-verify: every cell is pulsed until it is within a fixed tolerance, and resetting after any overshoot. This bounds the achievable accuracy and the worst-case wear.

The strongest existing competitor is SWIM ported to photonics. It uses Hessian-ranked selection with standard write-verify on the selected cells and a fixed tolerance deadband. It is the strongest because it is the only method in the file that already chooses where to spend correction effort based on output sensitivity. Its reported result for ResNet-18 on CIFAR-10 is that only 10% of write cycles keep the accuracy drop below 0.5%, against 50% for magnitude-based selection and 70% for random selection (SWIM).

The second strong competitor is ELight, which reports a 22.28× reduction in total writes on 5-bit VGG8 (ELight). ELight acts before programming, so it is compatible with every controller. We therefore run every controller both with and without ELight's reordering.

All methods share the same simulator seeds, device draws, sensitivity estimates, and monitor noise. Each gets the same tuning budget, a grid search over its tolerance or cost parameters on a held-out validation split. Beating SWIM means one of two outcomes at the 0.5% accuracy-drop threshold. Either we use significantly fewer total pulses including resets, or we use equal pulses with at least a 2× reduction in reset events, since resets dominate endurance.

The key ablation is SWIM plus a sensitivity-scaled deadband. Its tolerance shrinks with $h_i$, but it ignores step noise and overshoot cost. This isolates whether modeling irreversibility adds anything beyond sensitivity weighting.

Why not just…?

Why not just use SWIM with a tighter or sensitivity-scaled tolerance? This is the real threat. If step noise were small relative to tolerances, a deadband would capture nearly all the benefit. Our expectation rests on the measured step spread being about a third of the mean step. At that spread, a deadband rule fires pulses that overshoot often. The sensitivity-scaled-deadband ablation settles the question directly. We also sweep the simulated step spread. If the gap vanishes across the measured range, we will report that the stopping model is unnecessary and that SWIM transfers as-is.

Why not just use segmented deterministic heaters so there is nothing stochastic to stop? Chen et al. (2024) get 8 repeatable levels from 1:2:4 binary segments. However, each added bit needs another contact and drive line, and fabrication phase errors still need analog trimming. Our hybrid arm compares an all-segmented design against coarse segments plus one controlled analog segment, at matched pad count.

Why not just use volatile thermo-optic heaters with continuous feedback? These heaters draw static power and cause thermal crosstalk between neighbors. They are also long, while phase-change sections are about 30 µm (Chen et al. 2023). We do not re-litigate the choice of technology. The proposal targets designs that have already committed to non-volatility.

Why not just reduce writes offline with ELight? ELight reduces how many writes are nominally needed, but it cannot see the stochastic outcome of any individual write. The two approaches are complementary, and the ELight-plus-controller arm measures how much each contributes.

Proposed Experiments

The network-scale benchmarks reuse the published settings: ResNet-18 on CIFAR-10 following SWIM, and 5-bit VGG8 on CIFAR-10 and CIFAR-100 following ELight. Weights are mapped to simulated Sb₂S₃ cells. We report accuracy drop, total pulses, reset count, projected device lifetime against the reported endurance of more than 1,600 and more than 5,000 cycles, and wall-clock programming time.

The device-scale benchmark is the post-fabrication trimming of a Mach-Zehnder interferometer from Chen et al. (2023), which reached a 24 dB extinction ratio with manually chosen pulse counts. The question is whether an automated, restraint-aware controller reaches 24 dB with fewer resets and pulses than automated fixed-tolerance write-verify. We found no reported overshoot or reset rates on Lacuna for any of these trimming procedures, so we will measure them ourselves.

Three checks decide the outcome. First, the reset count at matched accuracy is the claimed win; if the cost-aware rule does not cut resets relative to the sensitivity-scaled deadband, the core claim fails. Second, the step-spread sweep shows whether any benefit tracks step noise, as the decision rule predicts. Third, a falsely-locked-cell audit measures how often monitor noise locks a cell that is actually wrong, which is the photonic analogue of a quality estimator tagging an error as correct.