The Photonic Design Flow

The parts bin is full: MZIs, phase shifters, rings, modulators, detectors, lasers. How does anyone turn a folder of ideas into a diced, packaged, working photonic chip? The answer is a design flow transplanted, organ by organ, from electronics — the same foundry model, the same GDSII masks, the same shuttle runs that the chip-design world perfected — but with several organs conspicuously missing. Electronic design rides on a tower of abstraction: RTL is synthesised into gates, gates into standard cells, cells placed and routed by algorithms, billions of devices that a human never looks at individually. Photonics has the bottom of that tower — foundries, PDKs, mask tooling — and almost none of the top: no HDL, no synthesis, no standard cell. A photonic chip has tens to thousands of components, every one placed by a person or a script that a person wrote, every geometric micron of it meaningful. This closing lesson walks the flow end to end and marks, honestly, where the pavement stops.

The flow at a glance

Four stations, one loop: when the tested chip disagrees with the simulation — and on a first spin it will, somewhere — the loop closes back to design, and the months start again. Everything in this lesson is about making each pass count.

PDKs and component libraries

As in electronics, you do not design for a process in general — you design for one foundry's process design kit. A photonic PDK contains the layer stack (silicon thickness, etch depths, doping implants, the Ge and heater modules), the design rules, and — its crown jewels — a component library: parameterised, silicon-validated building blocks. The foundry has fabricated and measured its grating couplers, splitters, phase shifters, ring filters and detectors across process corners, and publishes each as a black box with a layout cell and a measured S-parameter model (ports in, ports out, wavelength-dependent complex amplitudes — the same formalism as microwave engineering). The professional habit this teaches is humility: your design's novelty budget is finite, so spend it on the one custom device your idea actually needs and take everything else, gratefully, from the library. A circuit made of validated parts fails only in its new part — a circuit made of home-rolled couplers fails everywhere at once.

Simulation splits into two regimes that must not be confused. FDTD (finite-difference time-domain) solves Maxwell's equations on a grid through one component — exact, and exactly unaffordable at chip scale, since memory and time grow with volume over \Delta x^3: halving the grid step costs 8× the memory, and a whole chip at nanometre resolution would need more RAM than exists. So the flow uses FDTD once per novel component to extract its S-matrix, then simulates the full chip at circuit level: multiply and cascade the S-matrices, sweep wavelength, done in seconds. It is the photonic incarnation of "characterise the cell once, then compose" — with the phasor and matrix arithmetic of this module doing the composing.

// Circuit-level thinking in one page: cascade a link's components at 1550 nm. // Each part contributes loss (dB); the laser sets the input power (dBm). type Part = { name: string; lossDb: number }; const link: Part[] = [ { name: "laser -> coupler", lossDb: 1.5 }, { name: "modulator (IL)", lossDb: 3.0 }, { name: "waveguide, 2 cm", lossDb: 2 * 1.5 }, // 1.5 dB/cm { name: "ring filter", lossDb: 0.5 }, { name: "chip -> fibre", lossDb: 1.5 }, ]; let p = 10; // 10 dBm from the laser console.log("launch: " + p.toFixed(1) + " dBm"); for (const part of link) { p -= part.lossDb; console.log("after " + part.name + ": " + p.toFixed(1) + " dBm"); } const sensitivity = -18; // receiver needs -18 dBm console.log("margin over receiver: " + (p - sensitivity).toFixed(1) + " dB");

That ten-line "simulation" — losses add in dB, margin is what's left — is genuinely how first-pass link budgets are done, and running out of margin on paper is a great deal cheaper than running out on silicon.

Layout, DRC, and the shuttle to the fab

Layout produces the same artifact electronics ships: a GDSII file, polygons on numbered layers, one per mask step. But photonic geometry is a different dialect. Electronic layout is Manhattan — rectangles on a grid; photonic layout is curves — spirals, adiabatic tapers, S-bends, ring gaps — drawn as many-sided polygons on a nanometre grid, usually generated by Python frameworks rather than drawn by hand. Design-rule checking changes dialect too: alongside min-width and min-space, photonic DRC polices bend radii (too tight radiates light into the substrate), coupler gaps, and grid-snapping artifacts on curves. What it cannot police is function: photonics has only a weak analogue of electronics' LVS-and-extraction safety net, so "DRC clean" means manufacturable, not working.

Then the design goes to the fab — almost never alone. A full mask set costs from hundreds of thousands of dollars upward, so prototypes ride multi-project-wafer (MPW) shuttles: the foundry tiles dozens of customers' designs onto one reticle, everyone pays for their few square millimetres, and the shuttle departs on a published schedule a few times a year. A seat costs thousands rather than millions; the price is cadence — miss the tape-in date or botch the design, and the next boat is months away. Design reviews before a shuttle deadline have the atmosphere of a launch window.

Months later the wafers return, and the flow's most photonic-specific stage begins: test and packaging. Wafer-level testing works by pointing fibre probes at grating couplers — the one coupler type that fires vertically, so chips can be measured before dicing, thousands of structures overnight on an automated prober. Packaging is where photonics still hurts: a fibre must be glued to the chip edge holding sub-micron alignment through decades of thermal cycling, lasers must be attached or bonded, the electronics wire-bonded alongside — and this assembly, not the silicon, commonly dominates unit cost. Electronics solved its packaging with solder-ball self-alignment decades ago; photonics is still buying its way out one fibre at a time.

Rough current numbers, for calibration: a passive-silicon MPW seat of ~25 mm² runs on the order of €10–20k through brokers like Europractice or CMP on foundry platforms such as imec, AMF, AIM Photonics or GlobalFoundries; add active layers (junctions, Ge detectors, heaters) and the seat multiplies; a dedicated full-mask run starts in the high hundreds of thousands. The shuttle model is the same social technology that MOSIS pioneered for student CMOS chips in the 1980s — amortise the mask set across everyone on board — and it is quietly the reason university groups can afford to tape out photonic processors at all. The strategic consequence of "a few boats per year" is a design culture electronics has largely forgotten: you cannot iterate your way out of carelessness, so photonic groups over-instrument every tape-out — de-embedding structures, test rings, parameter sweeps of every gap and width — packing the margins of their seat with the experiments that will make the next seat's design right first time.

An electronics habit that must be surrendered at the border: treating layout as mere implementation of a schematic that carries the real meaning. In a photonic chip the geometry is the physics. A waveguide's width sets its effective index directly — a 1 nm width error shifts n_{\text{eff}} enough to move a ring resonance by a full linewidth (~100 pm), which is why no two nominally identical rings on one chip resonate at the same wavelength, and why every ring in Module 5's weight banks carries a trimming heater. A coupler gap drawn 10 nm wide changes its splitting ratio; a bend drawn slightly tight radiates a fraction of a dB that a schematic will never show. There is no synthesis tool re-deriving your geometry from intent, and only weak extraction checking that geometry against intent afterwards. The mitigations are cultural as much as technical: take library cells wherever possible, simulate with the PDK's measured corner data, design for post-fab trimming, and treat every hand-drawn nanometre as a design decision — because the fab will faithfully manufacture your mistakes at nanometre precision.

Module complete — where this goes next

This module built the parts and, now, the pipeline that turns parts into chips. Notice what the flow lacks by electronics' standards — synthesis, abstraction layers, mature packaging, self-correcting toolchains — because the shortfall is not a footnote: it is a standing tax on every photonic system in the rest of this course, paid in tape-out months and calibration heaters. From here the course stops building components and starts computing with them: Module 4 begins with the beamsplitter read as a matrix, and the MZIs of this module — laid out, fabricated, packaged and locked by exactly the flow you now know — become the tiles of programmable linear algebra.