Waveguide layout: bend radius, width, and paths on a photonic chip

You are trying to make a strip that light will follow around a chip, the way light follows a fiber. On a screen the strip looks like a metal route: a bar, a corner, another bar. Light is not a spare electron looking for the shortest right-angle path. It lives in a waveguide: a strip of higher-index material that holds a mode. The picture at the top is that chip in a lab, a fiber lined up to the edge so light can enter. Width sets the mode. A tight bend radiates. A rough sidewall scatters. A strip about 0.45 µm wide with a bend radius of several micrometers may keep the mode. The same strip with a bend of about 1 µm can dump the power into the slab. Bend radius is part of the device, not a routing convenience.

The corner is the circuit

If you jog the strip around an obstacle the way you would jog a clock wire, the corner becomes an antenna. Light leaves the strip and the detector goes dark, or a resonator sits at the wrong color because the path length changed. A 90-degree corner that is legal for metal can be illegal optically. There is no separate wire you can move at the end without changing the part.

Both drawings can look like successful routes. The second picture is the bend you want to see under a microscope: a smooth curve, not a stair of Manhattan jogs. Path length along that curve is phase. Phase is the circuit.

What else the strip feels

Neighbors matter. A metal heater over the strip tunes the index with temperature. A dummy strip too close steals power. Two arms of an interferometer need the same length, the same bends, and the same neighbors. A path in the editor — centerline plus width — is a convenient pencil. The factory still receives polygons on a grid. Too few vertices and the “arc” is a polygon that scatters. Too many and fracture later produces slivers.

Faceting that arc onto a coarse grid is how a smooth curve in the second picture becomes a glitter of edges on the mask. Width error of a few tens of nanometers shifts a filter. The geometry-rule deck may allow it. The mode will not.

Crossing two waveguides like metal traces is a scatter site unless the process has a designed crossing cell. Copying a path from a board tool, where width is current in copper, into a photonic cell, where width is a mode, keeps the path object and changes the physics.

The fiber in the first photograph is the reminder: light entered from outside, and it will only stay in the strip if the drawing respected the bend. The maze of other colors on a full chip is usually heaters, contacts, and keep-outs. Hide them when you are judging the curve.

The fiber also has a size the strip must meet. A mode in a 0.45 µm silicon waveguide does not fill a fiber core that is many micrometers across. Someone draws a taper, a grating, or an edge coupler so the light can cross that border. That coupler is a cell with a length and a width schedule, not a last-minute flare. If the fiber in the photograph is glued to the wrong edge, or the taper faces the other way, the chip is dark and the layout still “passes” a metal spacing deck.

In the editor, select the tightest bend and read its radius from the path. Write that radius, the strip width, and the wavelength into the cell name before you fracture anything. A radius near 1 µm on a 0.45 µm strip can dump the mode into the slab even when the curve looks smooth. A radius of 5 µm and a radius of 10 µm are different devices at 1.55 µm, even when both sit comfortably beside a metal pitch. If the radius you read is the tight one, near 1 µm, change it. Leave the new radius in the cell name so the next edit does not square the corner off to quiet a design-rule report. On the chip the fiber enters, that named curve is the path the light actually takes. A square jog that satisfies a metal spacing deck still dumps the mode, and a power meter at the far edge goes dark while the report stays clean. After you change the radius, read it back from the path and confirm the cell name still carries that new number before the file is fractured. A name that still says 1 µm over a path you have opened to several micrometers will send the tight bend you thought you had replaced.

Read the radius, then put it in the name

Select the tightest bend and read the radius from the path, not from memory. A 0.45 µm strip with a radius near 1 µm can dump the mode into the slab at 1.55 µm even when the curve looks smooth and a metal spacing deck is quiet. A radius of 5 µm and a radius of 10 µm are different devices at that wavelength. If the number you read is the tight one, change it, then read it back from the path.

Write the radius, the 0.45 µm width, and 1.55 µm into the cell name before fracture. A name that still says 1 µm over a path you opened to 5 µm will send the bend you thought you had replaced. On the chip the fiber enters, that named curve is the path the light takes. A square jog that satisfies a metal rule still leaves the power meter at the far edge dark. The fiber glued to the wrong edge fails the same way, with a layout that still “passes.”