Why Neural Field Surrogates Are Misleading in Photonic Designs

In the rush to embrace neural field surrogates for photonic design, the allure of speed and efficiency often glosses over a critical issue: accuracy isn't everything. Neural surrogates, while adept at handling global field errors, can stumble when the stakes hinge on localized output-port readouts. In simple terms, they might look great on paper but fail where it counts.

The Problem with MMI Splitters

Take, for example, MMI (Multimode Interference) splitters and couplers. These devices rely heavily on accumulated modal interference and output-window aggregation. This means that accurate global field predictions won't necessarily translate to accurate device performance. The surrogate might tell you one thing, but your output ports might say another.

Enter PaNO, a new approach that aligns these surrogates with what really matters. PaNO, or Propagation-Aligned Neural Operator, organizes data around local boundary structure, transverse modal content, and cross-mode interaction. Sounds complex? it's. But it’s an attempt to bridge the gap between global accuracy and local necessity.

PaNO-R2: The Feedback Variant

PaNO-R2 takes things a step further by introducing output-aware feedback, particularly useful for residual field components near the port region. A benchmark test on a 15-wavelength tunable $3{\times}3$ MMI showed striking results. With 4608 held-out fields, PaNO slashed NeurOLight's port-power error from 0.2018 to 0.0739, despite a slightly higher global field error.

Here's a question: What good is a neural surrogate if it can't deliver where it matters most? These results show that focusing solely on global field accuracy won’t cut it. The key is integrating an awareness of what truly affects device performance.

The Real Takeaway

PaNO and its feedback variant, PaNO-R2, prove that focusing on local specifics, like port-power and output-profile errors, matters significantly more than we thought. They reduced NeurOLight's port-power and output-profile errors by approximately 72.7% and 72.5%, respectively.

So, are neural field surrogates the future of photonic design? Not yet. The key lies in refining these models to consider localized outputs as much as global predictions. Until then, the gap between the lab and real-world application remains wide.