Optical computing chips often send light down many waveguides and recombine it at one point. For the result to be correct, the light on every path has to arrive with the right phase. Fabrication shifts each waveguide's width and thickness by a few nanometres, and that is enough to scramble those phases. Phase-change materials (PCMs) offer a way to fix this. A PCM is a thin film that can be switched by a short heat pulse between a glassy state and a crystalline state, and it holds whichever state it is left in without using power. Placed on a waveguide, it acts as a permanent, reprogrammable phase trimmer. The catch is cost. The best foundry-integrated PCM, antimony trisulfide on a 300-mm line, shifts phase by about 0.012π per micron and loses about 0.0044 dB per micron (Deterministic quasi-continuous tuning…). A trimmer with a full 2π range is therefore roughly 167 µm long and costs about 0.73 dB on every arm it sits on. If designers put a full-range trimmer on every path, the light lost in the trimmers can exceed the light lost to the phase errors they correct.
This proposal treats trimmer placement as a design-time synthesis problem, borrowing directly from synchronous path balancing in superconducting logic. In those circuits, delay elements are inserted only where needed, shared across paths, and sized to the expected timing variation. Here, the decisions are which waveguide segments receive PCM trims, how long each trim is, and how many binary-weighted segments it has. The goal is to maximize the fraction of fabricated chips that can be fully corrected with one programming pass, at minimum added loss. Two physical facts enter the yield model. The first is that fabrication errors are spatially correlated, so nearby arms drift together and partly cancel. The second is that a trim segment only switches reliably if its electrical heat pulse arrives intact. That is not guaranteed once pulses are short enough for the drive wiring to behave as a lossy transmission line, and once PIN-heater doping varies from device to device.
Core Claim
In coherent photonic networks where many input-to-output paths share waveguide segments, a variation-aware trim-placement solver should reach the same correction yield as uniform per-arm trimming with substantially less total PCM length, and therefore less insertion loss. It does this by exploiting two things. Sharing places trims on common trunk segments that serve many paths, exactly as selective flip-flop insertion does in superconducting logic. Statistical sizing gives each segment only the range its correlated error distribution requires.
The proposal also addresses an apparent contradiction in the literature. Digital pipelines gain yield from deliberate stage imbalance (Statistical Modeling of Pipeline Delay…), while coherent optics needs near-zero phase mismatch. The claim is that the two are compatible. Imbalance belongs in the trim ranges and in the nominal design offsets, not in the final programmed phases. Spreading where correction capacity sits avoids the "everything becomes critical at once" failure mode that pipeline paper identified.
Research Program
The target class is a coherent fan-out and fan-in network. Light from a source splits through a tree, passes per-path weight elements, and recombines through a combining tree. Every source-to-combiner path must match a reference phase within a tolerance $\epsilon$. Equal-length trees, and meshes built from them, are the natural testbeds because their shared trunks create the sharing opportunity.
We model the network as a directed acyclic graph with edge set $E$. Each edge $e$ may carry a trim of length $L_e$ split into binary-weighted segments, following the 1:2:4 scheme of Deterministic quasi-continuous tuning…. That scheme yields $2^N$ discrete levels from $N$ control lines, so each trim's programmable phase $\theta_e(k_e)$ is quantized. The vector $\xi$ collects all random fabrication parameters: correlated waveguide width and thickness fields, per-segment PCM film variation, and PIN doping variation. The design problem is
$$\min_{\{L_e, N_e\}} \sum_{e\in E} \alpha L_e + \beta N_e \quad \text{s.t.} \quad \Pr_\xi\Big[\exists\, k:\ \max_{p} \big|\phi_p(\xi) + \textstyle\sum_{e\in p}\theta_e(k_e) - \phi_{\text{ref}}\big| \le \epsilon \ \wedge\ \text{all used segments switch}\Big] \ge Y.$$
Here $\phi_p(\xi)$ is the uncorrected phase of path $p$, $Y$ is the target yield, $\alpha$ converts trim length to insertion loss (0.0044 dB/µm from the foundry data), and $\beta$ charges for control lines and routing.
The solver follows the clock-follow-data approach of Delay Balancing With Clock-Follow-Data: an iterative relaxation over placement variables with rounding. The chance constraint is replaced by a sample-average approximation over variation samples. Because that paper reports that rounding thresholds can trap the solver in poor local minima, we will also solve small instances exactly as mixed-integer programs and report the optimality gap.
Evaluating $\phi_p(\xi)$ for thousands of samples is the expensive step. We will build a polynomial-chaos surrogate of edge phase as a function of local width and thickness, using the stochastic-testing collocation of Stochastic Testing Method…. That paper reports speedups over Monte Carlo of up to 1700× in DC sweeps. Its known weakness is that cost grows sharply as the number of random variables rises. We handle this by building the surrogate per edge on a few principal components of the correlated variation field, rather than on all variables at once. Final yields are always checked by plain Monte Carlo.
The second physical ingredient is whether a segment actually switches, which we model at the drive level. Each segment's reset pulse has to deliver enough energy to melt the film and then cool fast enough to freeze it. The PIN-heater switch in Nonvolatile Electrically Reconfigurable Integrated Photonic Switch… uses 6.6 V for 100 ns. Pushing toward sub-nanosecond pulses, to cut heat diffusion into neighbouring cells, puts the pulse spectrum into the multi-gigahertz range. At those frequencies the compact-model work on 2.5D/3D interconnect (Compact models and model standard…) shows that distributed RLCG models with frequency-dependent loss are needed, because lumped RC models lose fidelity. We will model each control line plus its PIN load as a distributed RLCG line inside the differentiable electronic-photonic co-simulator SPIPE, and add a simple transient heat model for the segment. From this we extract a switching-success probability as a function of line length, pulse width, and doping variation. That probability enters the yield constraint and can push the solver toward longer pulses, shorter lines, or fewer and larger segments.
We are honest about the limits of this second ingredient. It is the most speculative part of the proposal, so it is isolated by ablation. If it never changes the optimal placement, we will report that.
Baselines
The naive reference is no trimming: the nominal symmetric layout, with yield measured directly under the variation model. It shows how much of the problem fabrication alone creates.
The practical industrial default is uniform full-range trimming. It puts one 2π-range segmented PCM trim on every leaf-to-combiner arm, the obvious analogue of full path balancing. It is the strongest simple competitor because it always achieves near-maximal correction. Its cost is predictable: about 0.73 dB per arm from the foundry-reported loss and phase-per-length figures. We run it on identical networks, identical variation samples, and identical tolerance $\epsilon$. Our method beats it if it matches its yield to within one percentage point with at least 30% less total PCM length, or reaches higher yield at equal loss.
A stronger statistical baseline is uniform per-arm trimming sized to $k\sigma$ of each arm's marginal phase error. It is statistically informed but neither shared nor correlation-aware. We tune $k$ on held-out samples with the same sample budget our solver uses. This baseline separates the benefit of sizing from the benefit of placement.
The strongest existing method for keeping photonic accelerators accurate under variation is runtime remediation, DOCTOR. It uses sparse calibration and remaps tiles of volatile thermo-optic phase shifters, and it reports restoring CIFAR/Fashion-MNIST accuracy from about 57% to over 90% with 0.14–5% cycle overhead. Its public code lets us run it on top of the uniform-trim baseline and on top of our placement, using the same vision benchmarks and the same variation draws. This answers whether design-time placement reduces what runtime calibration must recover.
We found no reported result on Lacuna for trim-placement yield or loss in photonic fan-in networks. The comparison numbers for that question will therefore come entirely from runs inside this project.
Three ablations isolate the key ingredients: the solver with sharing disabled (trims only on leaf arms), the solver with an independent rather than spatially correlated variation model, and the solver without the drive switching-reliability term.
Why not just…?
Why not just use thermo-optic heaters with closed-loop feedback? Feedback-tuned heaters correct phase reliably, but they draw power for as long as the correction is held, and they heat their neighbours. That heat drift is exactly what DOCTOR has to track at runtime (DOCTOR). We will report static tuning power for this alternative on the same networks, using measured per-arm correction phases. For chips that must hold a fixed phase configuration for long periods, that power is the decisive cost. For chips that are reprogrammed constantly, thermo-optic tuning may simply win, and we will say so.
Why not just put a full-range PCM trim on every arm and calibrate after fabrication? This is the real threat, and it is the main baseline above. It works, and post-fabrication calibration is unavoidable in either case. The question is purely how much loss and how many control lines it costs. If correlated variation turns out to be so strong that every arm needs nearly full range anyway, sharing and sizing will save little, and the proposal's claim fails. The correlation-length sweep described below is designed to expose exactly that regime.
Why not just make layouts compact and symmetric so errors cancel? Compact symmetric layouts raise the correlation between paired arms, which already reduces mismatch. We do not replace this practice; we take it as the starting layout. The solver's correlation-aware variant should then exploit the remaining structure. The ablation with an independent-variation model shows how much of the gain comes from understanding correlation rather than from symmetry alone.
Why not just run Monte Carlo in SPICE or a full-wave solver? Plain Monte Carlo is our final judge of yield. It is too slow to sit inside a placement search, however. The surrogate exists only to make the search tractable, and every reported yield is re-verified by direct sampling.
Proposed Experiments
Solver validation. Before any photonic work, we check that the relaxation-and-rounding solver is implemented correctly. We run it on the ISCAS-85 circuits that Delay Balancing With Clock-Follow-Data used, under its superconducting delay model. The target is a reproduction of its reported average 1.48× area-delay improvement over full path balancing. This only validates the solver machinery; it says nothing yet about photonics.
Main photonic experiment. Binary fan-out/fan-in trees and mesh networks of 16 to 256 paths are generated in SPIPE. Variation is modeled as a correlated Gaussian width and thickness field, with correlation length swept from much shorter to much longer than the network's footprint. A non-Gaussian heavy-tailed variant is included, because Gaussian max-approximations are known to break down (Statistical Modeling of Pipeline Delay…). Each method is optimized on one set of samples and evaluated on 10,000 fresh Monte Carlo chips. The metrics are correction yield at tolerance $\epsilon$, total trim length and resulting insertion loss, control-line count, and the solver's optimality gap against exact solutions on small instances.
Task-level check. The corrected networks are mapped into DOCTOR's photonic accelerator simulator. We report CIFAR-10 and Fashion-MNIST accuracy with and without DOCTOR running on top, to test whether loss saved at design time translates into accuracy or into reduced calibration effort.
Drive-parasitics study. We sweep reset pulse width from 100 ns down to sub-nanosecond values and line length across realistic array sizes. The test is whether including the drive-reliability term changes the chosen placement and the yield verified by Monte Carlo.
Decisive outcome. The central claim survives if the shared, correlation-aware solver delivers uniform-trim yield at meaningfully lower loss across a substantial band of correlation lengths. It fails if the savings vanish everywhere except in contrived settings, and that outcome would itself be useful guidance to photonic designers.
References
- Deterministic quasi-continuous tuning of phase-change material integrated on a high-volume 300-mm silicon photonics platform
- Statistical Modeling of Pipeline Delay and Design of Pipeline under Process Variation to Enhance Yield in sub-100nm Technologies
- Delay Balancing With Clock-Follow-Data: Optimizing Area Delay Tradeoffs for Robust Rapid Single Flux Quantum Circuits
- Stochastic Testing Method for Transistor-Level Uncertainty Quantification Based on Generalized Polynomial Chaos
- Nonvolatile Electrically Reconfigurable Integrated Photonic Switch Enabled by a Silicon PIN Diode Heater
- Compact models and model standard for 2.5D and 3D integration
- SPIPE: Differentiable SPICE-Level Co-Simulation Program for Integrated Photonics and Electronics
- DOCTOR: Dynamic On-Chip Temporal Variation Remediation Toward Self-Corrected Photonic Tensor Accelerators