Title: Can High Global Accuracy in Neural Fields Deceive Photonic Designers? The Case for Port-Centric Evaluation

Abstract:

While neural field surrogates offer significant speedups for photonic design workflows, a critical pitfall exists: a model may demonstrate high global field accuracy yet still fail to correctly rank candidate devices when the final selection relies on localized output-port metrics. This discrepancy is particularly pronounced in MMI splitters and couplers dominated by propagation effects. In these components, key performance indicators such as port power, splitting ratios, phase shifts, and coupling coefficients are driven by accumulated modal interference and output-window aggregation, rather than simple average field similarity.

To address this "field-to-design" mismatch, we employ a Field/Mediator/Readout framework that disentangles dense complex-field errors from propagation-profile and output-window errors prior to port aggregation. To ensure the surrogate aligns with this specific chain of operations, we introduce PaNO (Propagation-aligned Neural Operator). PaNO maintains a full-field prediction interface but structures its latent states around local boundary structures, transverse modal content, axial propagation dynamics, and cross-mode interactions. Additionally, we assess PaNO-R2, a variant incorporating output-aware feedback to refine residual field components near the port region.

In benchmarks using a 15-wavelength tunable $3{\times}3$ MMI with 4,608 held-out fields, PaNO reduced the port-power error of the NeurOLight baseline from 0.2018 to 0.0739, despite exhibiting a slightly higher cMAE. These results underscore that global field accuracy alone does not guarantee fidelity in design-relevant readouts. PaNO-R2 achieved superior performance across cMAE, propagation-profile error, output-profile error, and port-power error, cutting NeurOLight’s port-power and output-profile errors by 72.7% and 72.5%, respectively.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC