What is OPC, and why doesn't the photomask look like the chip?

OPC stands for optical proximity correction. Before a photomask is made, its pattern is changed on purpose so it no longer matches the design drawing, and that is exactly what makes the pattern printed on the wafer come out close to the design. The reason is that the lines on a chip are narrower than the wavelength of the light used to print them, and the light "rounds off" the pattern. OPC builds that rounding in ahead of time.

How a photomask prints its pattern onto a wafer is in "What is photolithography?", and the mask itself is in "What is a photomask?". This article covers one thing: why the pattern on the mask has to be changed first.

Why not just draw the mask exactly like the design?

Because the light won't print it that way. The deep ultraviolet light most scanners use today has a wavelength of 193 nanometers, while the finest lines on a chip are only a few tens of nanometers wide, narrower than the wavelength.

When light passes through a very narrow gap, it spreads out. This is called diffraction ("What is extreme ultraviolet (EUV) light?" explains it). The narrower the gap, the more it spreads. The scanner's lens can only collect part of that spread-out light, and misses the light that spreads the widest. The finest details of the pattern are carried by exactly that light.

So the printed pattern looks like a fine drawing traced with a thick marker: the overall shape is there, but sharp points come out round and thin parts come out blurry.

What goes wrong in the print?

• Corners round off: a right angle in the design prints as a rounded corner.
• Line ends pull back: the tip of a thin line prints shorter than drawn. Once the two corners at the tip round off, the tip shrinks back.
• Dense and isolated lines print differently: two lines drawn the same width come out with a different line width if one sits among many neighbors (dense) and the other sits alone (isolated).

The third one is where the word "proximity" comes from. How a line prints depends not just on the line, but on what's near it, because light spreading from the neighbors reaches it and adds up. This is called the optical proximity effect, and it is what OPC corrects.

What happens without it? A line end that pulls back too far may no longer reach the connection it should land on, and then the circuit is broken. And when the same kind of line comes out wider in one place and narrower in another, switches that should be identical can end up faster or slower.

How does OPC "draw it wrong on purpose"?

It's like an archer who knows the wind will push the arrow to the right, so aims a little to the left. If you know how the light will smear the pattern, you change the mask in the opposite direction. The common changes:

• Small squares on corners: an outside corner gets a small extra square stuck on, so it rounds off less (it still rounds a little).
• Longer line ends: the tip of a line is drawn longer and a bit wider, like a small hammer, so after it pulls back it lands in the right place.
• Line width set by the neighbors: isolated lines and dense lines are drawn at different widths on the mask, so they print the same.
• Thin lines that don't print: next to an isolated pattern go a few lines so thin they don't print themselves. They only make the light behave as if the pattern were in a dense area. Draw them too wide and they print too, so they have to be sized carefully.

How does OPC work, step by step?

Step 1: Take the design and make it the target

Once a chip design is finished and goes through tape-out, the pattern for each layer goes to the people who prepare the mask. This drawing is "what the wafer should look like in the end", and every later step compares against it. Make a mask from it as it is, and the print comes out like the last section: round corners, short lines, dense and isolated lines that don't match.

Step 2: Have a computer predict what will print

The computer treats the drawing as a mask and calculates what pattern is left after the light passes through, goes through the lens, hits the photoresist and is developed. Making a real mask, printing a real wafer and measuring it is slow and expensive; the computer can run through the whole layer first. The rules it calculates with (the model) are tuned beforehand: test patterns are printed on wafers and measured with an electron microscope until the model's answers match the real ones. With a model that's off, however much you change the mask, the changes won't come out right.

Step 3: Compare with the target and change the mask where they differ

Every edge is cut into many short pieces, and each piece is compared: if the print falls short, that piece on the mask is pushed out a little; if it bulges out, it's pulled in a little. The squares on corners and the hammers on line ends are added in this step too; the thin lines alongside go in during this step or before it. Skip it, and corners round off, line ends pull back, and connections can be missed.

Step 4: Predict again, and repeat until it's close enough

The changed drawing is calculated again to see how far each piece still is from the target. Pushing one piece also changes how the pieces next to it print, so it takes several rounds until every piece is within the allowed range. Stop after one round, and some places are often overcorrected while others still fall short.

Step 5: Check: will two lines run into each other?

The whole layer is calculated once more, and not only at the best focus and the best amount of light, but also a little off: the check looks for lines so close they merge (a short), lines squeezed so thin they almost break, and thin helper lines that print by themselves. Scanners drift a little every day, so a pattern that is only right under ideal conditions can fail in production. Find out after the mask is made, and you have to make a new one.

Step 6: Send it to the mask shop

The checked pattern is turned into a file the mask-making machines can read and sent to the mask shop. With all its small squares and short pieces, the corrected pattern holds far more data than the original, so the mask takes longer and costs more to make. Once a mask is made, the pattern on it can't be redrawn; a change means a new mask.

Why does it take so much computing?

Because there is so much of it. A single mask layer has more shapes than anyone can count. Every edge is cut into several pieces, and each piece needs its own light calculation; pushing one piece affects its neighbors, so it takes several rounds; and the whole thing is checked again at several focus and light settings. A chip has dozens of layers, and the finest ones all need this, so it takes many computers working together.

Early on, it worked more like a lookup table: this wide a line, this much space next to it, add this much. Once patterns got complicated, the table could never cover them all, so the work moved to simulating piece by piece. Around the late 1990s, when lines started to be narrower than the wavelength of the light, OPC for whole chips came into use.

Is OPC the only trick?

No. There are fixes on the exposure side too: a layer of water between the lens and the wafer (immersion lithography), splitting a pattern that's too dense into two or three separate exposures (multiple patterning), or switching to extreme ultraviolet light with a far shorter wavelength (see "What is photolithography?" and "What is extreme ultraviolet (EUV) light?"). These are usually used together with OPC, and even extreme ultraviolet still needs OPC.

In fab24hr

Tap "Clock in", choose "A day in the fab", then on the "Your role today" card tap "R&D line", swipe to the R&D process engineer and tap "Litho" under the name: you're the litho R&D process engineer (marked Beta). The four shifts follow the time you open the game, and the afternoon shift, "Two Windows", is this article's story. The next mask set for the contact layer is finalized next week. The last search for exposure settings only measured the densely packed holes, but another part of the chip has isolated holes too. Measure both kinds together, and the isolated holes have a narrower range that works; the range where both kinds come out right is only a small patch. At the 16:00 mask meeting you give litho's view: add assist features next to the isolated holes, which means another processing run on the mask data and a mask two days late, or leave the mask on its original date and have the production fab measure one more monitor wafer every shift to watch focus.

The only correction the game asks you to decide on is the assist features; the "Know a bit more" part of the Assist feature glossary card says they belong to optical proximity correction. The piece-by-piece computing and redrawing isn't shown. In the prologue of the main story, The life of a chip, "Tape-out", the design of Glowmere goes to the fab and becomes a set of masks. The companies, people, tools and figures in the game are fictional; what OPC does and why follows public, general sources. For other jargon, see Chip fab jargon in plain English.