PaNO: Readout-Aligned Simulation for Photonic Devices

Aug 2025 - present · Yuzhe Ma’s group, Microelectronics Thrust, HKUST(GZ) / Guanzhong’s group, SYSU · First author · NeurIPS 2026 under review

Problem. Photonic design runs FDTD/FDFD solvers over and over, at ten-plus seconds per device. Neural surrogates cut this to milliseconds, but they are judged by global field error while designers actually care about local port readouts: powers, splitting ratios, phases, coupling. A field can look accurate overall while the output-plane region drifts, so port power goes wrong and candidate devices get mis-ranked.

Insight. Port readouts are localized integrals over the output plane, produced by modal interference accumulated along the propagation axis. A surrogate must preserve the whole field-to-readout chain, not just dense-field accuracy.

Method.

• Introduced a Field/Mediator/Readout evaluation: dense field reconstruction, propagation and output diagnostics, and localized port readouts, all computed from the predicted complex field with no separate scalar head.
• Designed PaNO, a propagation-aligned operator: anisotropic front end, learned transverse modal tokens, a selective state-space scan along the propagation axis, cross-mode coupling, plus an output-aware reverse-residual branch (PaNO-R2).

Result.

• PaNO cut port-power error from 0.2018 to 0.0739 despite slightly worse field error; PaNO-R2 reached the best field error 0.1471 and port power 0.0551, down 72.7% from NeurOLight, with output profile down 72.5%.
• One forward pass takes 6.19 ms, about three orders of magnitude faster than generating a reference field.
• First-author paper under review at NeurIPS 2026.

Takeaway. Evaluation comes before architecture. Naming the field-to-readout mismatch made the next step obvious: align the objective to the readout chain first, then organize the network’s latent space along that chain.