Imagine an architect designing a skyscraper on a digital drafting table. Every beam she places must bear a certain load, every corridor must be wide enough for emergency egress, every window must meet the fire code. Now imagine that her software never allows her to draw a single violation — every line she sketches is, by construction, automatically compliant. She never needs to call a building inspector after the fact, because the palette itself contains only code‑legal geometries.

This is essentially what a team at Koç University has achieved for the design of nanophotonic components. In a preprint (arXiv:2602.03142), Bahrem Serhat Danis, Emir Salih Magden, and their colleagues introduce a framework that bakes fabrication rules directly into the design space itself — not as a penalty to be avoided, but as a constraint that shapes the very set of possible answers. The result is a pipeline that churns out devices ready to be built on a silicon foundry floor, without sacrificing the performance that modern inverse design can offer.

Inverse design flips the engineer’s traditional workflow. Instead of starting from a known geometry and simulating its optical response, you specify the response you want — say, a 50/50 split of incoming light across two output ports — and let an algorithm sculpt the permittivity distribution of a silicon‑on‑insulator chip until it produces that behavior. The approach has yielded spectacularly compact and efficient devices, but it has a recurring headache: the algorithm, left to its own devices, will happily draw features far smaller than any lithography tool can print, or place waveguides closer than the allowable spacing. The resulting shape works beautifully inside a computer and not at all in the real world.

The usual fix is to add a penalty term to the optimizer’s objective function — a sort of building inspector that, after each iteration, fines the architect for every beam that’s too thin or window too close. The trouble is that the inspector is heavy‑handed. The penalty can warp the search space, trapping the optimizer far from good solutions, and requires the engineer to tweak a schedule of hyperparameters — when to turn up the fine, when to relax it — with the patience of a gardener coaxing a reluctant vine onto a trellis. Even then, compliance is never guaranteed until the very last iteration.

The Koç team took a different path. They asked: what if the architect never sees an illegal shape to begin with? They built a generative model — a neural network that learns to map points from a compact latent space directly onto device layouts that already satisfy every design‑rule constraint. Train it once on a library of compliant geometries, and from then on every vector z you feed into the generator produces a layout that respects minimum feature size, minimum spacing, and the discretized binary nature of the material (silicon here, silica there). Optimization then proceeds within this learned manifold — a curved, constrained surface inside the much larger space of all possible permittivity distributions — and every step the optimizer takes stays firmly on the surface. You can no longer draw an illegal beam.

The generator’s architecture enforces compliance through a carefully orchestrated chain: cascaded upsampling stages that refine the design, a thresholding layer that snaps every pixel to pure silicon or pure silica, and a morphological closing operation that fills in tiny gaps and smooths out spiky protrusions. The whole pipeline is differentiable, so gradients can flow back from the electromagnetic simulation all the way to the latent code z, guiding the search toward higher performance. The building metaphor’s limits appear here, however: unlike a physical architect, the optimizer doesn’t “draw” in the intuitive sense — it moves through an abstract mathematical space where compliance is a property of the coordinate system, not a checklist.

The team validated the framework on a suite of representative silicon photonic devices. They designed broadband 50/50 power splitters, splitters with splitting ratios of 70/30 and 90/10, a photonic duplexer that separates two wavelength channels with sub‑decibel insertion losses, and a mode converter that transforms an incoming fundamental transverse electric mode into a higher‑order mode with near‑perfect fidelity. All of these were targeted for the 1,500–1,600 nm telecom band, a spectral window roughly twice the wavelength of visible red light. Across the board, the devices achieved state‑of‑the‑art performance metrics, matching or approaching the transmission efficiencies of conventionally optimized designs — but with a crucial difference.

The generator‑based optimization converged with a computational cost over five times lower than that of a pixel‑based baseline. In pixel‑based inverse design, the algorithm struggles with a gradual “binarization schedule” — a sequence of increasing pressure to push intermediate gray tones toward pure silicon or silica — which consumes iterations and often requires the designer to babysit the hyperparameters. The generator, by contrast, remains perfectly discrete from the very first iteration. No grays ever appear. Every intermediate layout, from the lucky initial guess to the final polished design, is already manufacturable. This means the engineer can stop the optimization at any point — if time or budget run out — and still have a device that could be sent to the foundry. That is a luxury the penalty‑based world never afforded.

Why does this matter beyond the lab? Photonics foundries operate on wafer‑scale processes where design‑rule violations are not suggestions; they are hard stops. A single sub‑minimum gap or a sharp corner that falls below the lithography threshold can render an entire wafer run useless, wasting months and tens of thousands of dollars. The Koç team’s framework, because it enforces compliance intrinsically, removes that risk. They demonstrated compatibility with both electron‑beam lithography platforms (with typical minimum features in the tens of nanometers) and photolithography platforms (with coarser, hundred‑nanometer‑scale rules), all by simply swapping the generator model trained for each foundry’s specific rulebook. The same optimization algorithm drives both; only the underlying manifold changes.

This hints at a deeper shift in how we might think about computational design. Instead of treating fabrication constraints as an external burden that we must later satisfy, the manifold‑based approach treats them as a fundamental property of the representation — a “language” of shapes that the designer can speak but never violate. The road from concept to chip becomes shorter not because we run the optimizer faster, but because we eliminate the inspection and re‑inspection loop that follows it.

The team has made the trained generator models and the optimization pipeline publicly available, so other groups can drop in their own electromagnetic solvers and foundry rule decks. The road ahead is not without challenges: the generator must be retrained for each new fabrication process, and the latent space may not yet span every exotic geometry that future applications demand. But the direction is clear. When compliance is no longer a negotiation with a finicky algorithm — when the very space of possibilities is shaped to exclude the impossible — nanophotonic design begins to feel less like coaxing and more like composing.

— Yanjiang

Yanjiang is an online editor of LoomSci.com.

References