Will Accurate Fields Mislead Photonic Design? FromGlobal Accuracy to Port Readout

This paper introduces PaNO, a propagation-aligned neural operator that prioritizes output-port readout fidelity over global field accuracy to prevent neural field surrogates from misguiding photonic device design, particularly in propagation-dominated structures like MMI splitters.

The Gist

The Big Problem: "The Blurry Photo vs. The Sharp Receipt"

Imagine you are a photographer trying to design a new type of camera lens. You have a super-smart AI assistant that can predict what the final photo will look like.

Usually, we judge if the AI is good by looking at the whole picture. If the AI's photo looks 99% similar to the real photo in terms of colors and shapes, we say, "Great job!"

But here is the catch: In the world of photonics (light-based chips), the designer doesn't care about the whole picture. They only care about tiny, specific spots on the edge of the photo (the "ports"). These spots determine how much light goes into a fiber optic cable, how fast data moves, or how the light splits.

The paper argues that an AI can take a "perfect" photo of the whole room but still get the tiny spots completely wrong. It's like a weather forecast that predicts the temperature for the entire city perfectly, but gets the temperature in your specific backyard wrong. If you are planning a picnic in that backyard, the "global" forecast is useless to you.

The Specific Case: The "Light Highway" (MMI Splitters)

The authors tested this on a device called an MMI splitter. Think of this as a highway where cars (light waves) enter, merge, and then split into different lanes.

• The Physics: The cars don't just drive straight; they bounce off walls and interfere with each other (like waves in a pond) as they travel down the road.
• The Result: Where the cars end up at the exit depends on exactly how they interfered along the entire journey.
• The Failure: Old AI models (like NeurOLight) could predict the general "traffic flow" well. But because they didn't pay close enough attention to the specific way the waves interfered, they predicted the cars would end up in the wrong lanes at the exit. This caused the "port power" (the amount of light in the right lane) to be wrong, even though the overall picture looked fine.

The Solution: PaNO (The "Smart Navigator")

The authors built a new AI called PaNO (Propagation-Aligned Neural Operator). Instead of just looking at the image like a standard photo editor, PaNO thinks like a traffic engineer.

1. It understands the journey: Instead of just guessing the final image, PaNO breaks the light down into "modes" (like different types of cars) and tracks how they travel step-by-step down the highway.
2. It respects the physics: It knows that light travels in a specific direction and that waves interact with each other. It simulates this "flow" rather than just guessing the pattern.
3. The "R2" Upgrade: They also made a version called PaNO-R2. This is like having a second pair of eyes that specifically looks at the exit ramp to catch any tiny mistakes the main system missed and corrects them.

The Results: Better at the Job, Even if the Photo is "Blurrier"

The paper ran a massive test with 4,608 different scenarios. Here is what they found:

• The Old Way (NeurOLight): It had a very "sharp" overall picture (low global error), but it often got the exit lane wrong. The light ended up in the wrong port.
• The New Way (PaNO): It had a slightly "blurrier" overall picture (slightly higher global error), BUT it got the exit lanes exactly right. The light went to the correct ports.
• The Winner (PaNO-R2): This version got the best of both worlds. It had the sharpest overall picture and the most accurate exit lanes.

The Key Takeaway:
In designing these light chips, global accuracy is not enough. You can have a model that looks perfect on paper but fails in the real world because it misses the tiny details at the exit. The authors proved that you need to train and test AI specifically on how it handles the journey of the light and the final exit, not just the final image.

Summary Analogy

• Old AI: A painter who copies a landscape perfectly, but paints the wrong door on the house. If you need to enter the house, the painting is useless.
• New AI (PaNO): A painter who understands how the house was built. The painting might have a slightly different shade of blue on the sky, but the door is in the exact right place, and the path leads exactly where it needs to.

The paper concludes that for designing light-based chips, we must stop judging AI just by how "pretty" the whole picture is, and start judging it by whether it gets the critical exit points right.

Technical Summary

Technical Summary: Will Accurate Fields Mislead Photonic Design? From Global Accuracy to Port Readout

1. Problem Statement

Neural field surrogates are increasingly used to accelerate photonic design loops by predicting complex optical fields, thereby avoiding expensive full-wave electromagnetic simulations (e.g., FDFD/FDTD). However, a critical mismatch exists between global field accuracy and localized device readouts.

In propagation-dominated devices like Multi-Mode Interference (MMI) splitters and couplers, design decisions rely on localized output metrics: port power, splitting ratios, relative phases, and coupling behavior. These metrics are derived from the coherent accumulation of modal interference and output-window aggregation along the propagation axis. A surrogate model can achieve high global accuracy (low dense-field error metrics like cMAE) while still misrepresenting the localized intensity profile at the output ports. This "field-to-design" mismatch can lead to the mis-ranking of candidate devices in inverse-design loops or parameter sweeps, even if the overall field reconstruction appears visually accurate.

The paper identifies that dense-field error metrics average over the entire computational window, whereas port quantities are local functionals of the output intensity envelope. Consequently, a model may minimize global error while failing to preserve the specific propagation mediators (modal phases and interference patterns) required for accurate port readouts.

2. Methodology

The Field/Mediator/Readout Diagnostic View
To address this, the authors propose a three-tiered evaluation framework:

1. Field Metrics: Measure dense complex-field reconstruction (e.g., cMAE).
2. Mediator Metrics: Measure propagation-profile consistency and output-window envelope behavior before port aggregation.
3. Readout Metrics: Measure localized device quantities (port power, phase, coupling).
   This decomposition ensures that improvements in global field fidelity do not come at the expense of the intermediate physical quantities that determine device performance.

PaNO: Propagation-Aligned Neural Operator
The authors introduce PaNO, a neural operator designed to align with the physics of MMI propagation while maintaining a full-field prediction interface (no separate scalar port head). Its architecture incorporates specific inductive biases:

• Multi-Scale Anisotropic Stem (MSAS): Uses depthwise convolutions with different kernel sizes along the propagation axis ($w$) and transverse axis ($y$) to respect the elongated nature of interference envelopes and the sharpness of material boundaries.
• Learned Transverse Modal Decomposition: Instead of scanning raw image columns, the model projects transverse slices into learned "modal tokens," exposing the modal organization inherent in the device physics.
• Selective State-Space Propagation: Modal tokens are propagated along the axial direction using a selective state-space model (SSM). This mimics the directed transport of energy and phase accumulation without acting as a traditional PDE solver.
• Controlled Cross-Mode Coupling: A lightweight residual MLP reintroduces cross-mode interactions (essential for coherent intensity calculation) before decoding the full field.

PaNO-R2: Output-Aware Feedback
A variant, PaNO-R2, adds a reverse-residual branch. This branch processes the input features in a reverse axial order to capture output-side discontinuities, weak reflections, or high-frequency residuals that a purely forward-propagating backbone might miss. It produces a spatial residual correction fused with the main output.

3. Key Contributions

1. Empirical Identification of Proxy Mismatch: The paper demonstrates that minimizing dense-field cMAE does not guarantee accurate port readouts, particularly in propagation-dominated devices where output profiles depend on accumulated modal interference.
2. Diagnostic Framework: It formulates the Field/Mediator/Readout view, providing a protocol to evaluate surrogates based on their ability to preserve the chain from full-field prediction to localized device function.
3. PaNO Architecture: The proposal of a propagation-aligned neural operator that encodes local boundary structures, learns transverse modal tokens, and utilizes directed state-space propagation to preserve self-imaging envelopes.
4. Validation and Trade-offs: Through extensive experiments on a 15-wavelength tunable 3×3 MMI benchmark, the paper validates that aligning with propagation physics improves readout fidelity, even when global cMAE is slightly compromised.

4. Experimental Results

The study was conducted on a 3×3 MMI benchmark with 4608 held-out complex-field cases across 15 wavelengths (1.530–1.565 µm).

• Performance vs. Baselines: Compared to NeurOLight (the primary baseline) and other neural operators (FNO, UNet), PaNO achieved significantly lower port-power error (0.0739 vs. 0.2018 for NeurOLight) and better propagation/output-profile errors, despite having a slightly higher cMAE (0.1822 vs. 0.1750).
• PaNO-R2 Superiority: PaNO-R2 achieved the best performance across almost all metrics, including cMAE (0.1471), port-power error (0.0551), and output-profile error, reducing NeurOLight's port-power and output-profile errors by 72.7% and 72.5%, respectively.
• Correlation Analysis: Diagnostic analysis revealed that active-region cMAE has a weak correlation with port-power error (Spearman $\rho \approx 0.21$–$0.28$). In contrast, output-profile error showed a much stronger correlation with port-power error ($\rho \approx 0.47$–$0.76$), confirming that mediator metrics are better predictors of readout failure.
• Generalization: In target-domain adaptation tasks (wavelength transfer and refractive-index shifts), PaNO-R2 consistently outperformed baselines, suggesting that propagation-aligned parameterizations generalize well when the device topology remains fixed but physical parameters shift.
• Efficiency: Inference time remains in the millisecond range (~6.19 ms on RTX 5090), offering a speedup of roughly three orders of magnitude over reference field generation.

5. Significance and Claims

The paper concludes that for photonic devices with localized readouts, global field accuracy alone is insufficient for design-relevant fidelity. The authors claim that surrogates must be evaluated and designed around the full Field/Mediator/Readout chain.

The significance lies in shifting the design objective of neural surrogates from pure image reconstruction to physics-aligned propagation. By preserving the intermediate modal and propagation structures, models like PaNO ensure that the predicted fields yield correct device-level metrics. The authors modestly note that their findings are currently bounded to frequency-domain 2D $H_z$ MMI devices with fixed localized ports, and that phase-sensitive readouts remain a challenge. They position this work as a step toward broader applications in vectorial simulations and other photonic components, emphasizing that the "Field/Mediator/Readout" diagnostic protocol is a necessary tool for reliable photonic AI design.