Turn a component visual into an editable parametric program.

PixCell (pixels → PCells) provides a visual-to-executable interface for photonic component creation. A multimodal agent interprets the representation and uses primitive symbols to construct optimizable parametric programs, allowing process-specific retargeting.

Research contracts

Registry

PixCell’s next research directions are issued as claimable research contracts: bounded bounties that researchers can take on with their agents. Each contract fixes five elements before work begins: a question evidence can settle, acceptance tests, the evidence bundle that must be retained, a verdict procedure that recomputes rather than trusts, and a claim boundary stating what success does not establish. All four are published as funded bounties on the VVUQ4AI marketplace. Settlement is issuer-graded with independent recomputation, and the claimant never writes the verdict. A claim remains a request until assignment is confirmed; the full terms are versioned in the public registry.

VVUQ4AI contract model
01

Shape-aware visual verification

What deterministic metric reflects structural agreement without rewarding area-filling shortcuts?

Success criteria

A frozen suite covers calibrated perturbations, topology and port errors, and adversarial block constructions; the metric is reproducible and improves on IoU and the current boundary baseline under blinded judgments and a locked optimization comparison.

T1 a pure deterministic metric on calibrated rasters, range zero to one. T2 a frozen suite of calibrated perturbations, topology and port errors, and adversarial block constructions. T3 strictly better ordering than IoU and the boundary baseline under blinded judgments and a locked optimization comparison. T4 every adversarial construction scores below its honest counterpart. T5 one rerun script reproduces all raw scores.

Full terms Claim

Benchmark

Coding-agent benchmark. At left, usable visual score and median cost per target for 26 OpenAI and Anthropic configurations. At right, a 26 by 8 matrix of independently recomputed IoU values for targets F1 through F8, with source violations marked.

Recreate

A visual becomes measurable, editable geometry.

A published photonic component is decomposed into primitive geometry and recomposed as an editable parametric cell.
Representation before optimization. PixCell asks the agent to express the component with a small geometric vocabulary, producing a program whose dimensions remain named and editable.
Selected reconstructions of an adiabatic coupler, ring-MZI hybrid, and sub-wavelength grating. Lower and higher-scoring iterative outputs appear above one gate-passing multimodal-agent output for each target.
Selected reconstruction records. Maroon is overlap, pink is candidate-only material, and black is missed target material. These examples show geometry; the complete benchmark above supports the quantitative comparison.

Each cell receives only the silhouette, footprint, and primitive contract and may revise for up to six rounds; pre-built device cells and raw polygon emission are rejected. Every retained geometry and score is independently rebuilt and recomputed by the verifier.

Retarget

Retargetability depends on the program.

An MZI recovered from a calibrated silhouette is retuned for an 8.0 nm free-spectral range on simplified 220 nm SOI, 400 nm SiN, and 400 nm TFLN stack models. The fixed geometry misses every stack; regenerated programs satisfy the analytic FSR and footprint gates on all three, with FSRs from 7.9995 to 8.0007 nm. A generic library MZI reaches the spectral target but exceeds every footprint.

Full-wave calculations define the evaluator's boundary. The analytic MZI fringe spacing agrees with the tested FDTD result to within 2.9%, while the directional-coupler approximation does not. The F1 source and reconstructed fixture differ in geometry, domain, and material model, so their comparison is qualitative.

A blind splitter reconstruction, an independently normalized full-wave field, and sidewall-angle sensitivity across three stack models.
Function and fabrication sensitivity. F1 is reconstructed, retargeted to SOI, and simulated; the source/simulation comparison is qualitative. Sidewall sweeps expose stack-dependent coupling and symmetry-protected balance.

Each retuned MZI passes only on its optimization stack; all six off-diagonal evaluations miss the ±2% interval by 9.3% to 100.7%. The editable program, not one geometry, is the transferable artifact.

Across the separate fixed-representation study, reconstructed programs pass 21 of 30 stack-by-model cases and library PCells pass 22 of 30. Neither representation class dominates.

A three-by-three cross-stack matrix whose three diagonal entries pass and six off-diagonal entries fail.
Cross-stack evaluation. Each row geometry is optimized for one stack and evaluated on all three. Only the diagonal entries satisfy the 8.0 nm ±2% target.

Dataset & model

The same executable representation becomes synthetic training data and a deterministic reward.

Ten examples from the PixCell Dataset arranged in five levels. L0 shows individual primitives, L1 operations, L2 local compositions, L3 structured geometries, and L4 complete photonic components.
The released L0–L4 representation curriculum. Two model-input examples per level progress from primitives and operations to compositions, structured geometries, and complete components. Headers report physical footprints; panels use maximum-visibility rendering rather than a shared visual scale.
Released dataset

PixCell Dataset v2.0.0

Verified image, footprint, and primitive-only program pairs. The compact core spans L0–L4; the depth configuration adds controlled parameter settings and a fixed validation split.

Core
738 rows
Depth
4,560 rows
Levels
L0–L4
Open dataset Released model

PixCell LoRA Run B

A rank-32 adaptation of Qwen3.6-35B-A3B trained for 150 GRPO updates using only executable, verifier-derived geometric reward—without supervised demonstrations or critic feedback.

Direct mean
0.228 IoU
Best of 8
0.467 IoU
Revise loop
0.491 IoU
Open model
Run B training across 150 GRPO updates from L0 through L4, alongside a heatmap of mean IoU on a fixed 80-task probe at the base checkpoint and after each curriculum stage.
Training across the released representation curriculum. The left panel records batch IoU and shaped reward through five 30-step stages. The right panel holds the same 80-task parameter probe fixed across checkpoints; the marked post-L3 probe is retained from a known contention window.

The raw base model produces no executable program in 64 benchmark attempts. The trained policy produces 39 of 64, with mean IoU 0.228 and mean best-of-eight IoU 0.467; three number-free revision rounds bring the independent deployment draw to mean champion IoU 0.491. This is one trained lineage evaluated through repeated samples and revisions, not a claim of parity with frontier coding agents.

Citation