What is CMP, and why do wafers need to be polished flat?

CMP stands for chemical mechanical polishing, also called chemical mechanical planarization. It flattens the surface of a wafer: a liquid called slurry softens the surface while a spinning polishing pad grinds off the high spots. A chip is built up one layer at a time, and every layer leaves the surface uneven. Unless it's polished flat first, the pattern for the next layer can't be printed sharply. So a wafer goes through CMP many times before it becomes chips.

Why does the wafer surface get uneven?

A chip is a bit like a building: transistors at the bottom, then layer after layer of metal wiring on top. Each layer goes roughly like this: lay down a material, print a pattern on it, cut away what the pattern doesn't protect, then lay down the next one.

A new film follows the shape of whatever is under it. Put a tablecloth over a bumpy table and the cloth is bumpy too. For metal wiring, trenches are often dug into the insulating layer first and then the whole wafer is covered in metal, so there's an extra layer on top of everything outside the trenches.

Left alone, the height differences build up layer after layer and can reach a few tenths of a micrometer. That sounds tiny, but for the next step it's already too much. For how films are built up, read What is thin film deposition?.

Why does an uneven surface spoil the next layer?

Chip patterns are printed by lithography: a scanner uses a set of lenses to project the circuit pattern onto the wafer. Lenses have a limit: the pattern is sharp only within a narrow band of height.

Take a close-up of a flower on your phone and you'll see it: focus on one flower, and things in front of it and behind it go blurry. The range that stays sharp is called the depth of field. A scanner has a range like this too, called the depth of focus, and it's very shallow. One university lecture gives an example from an older generation of scanners: only about 0.35 micrometers up or down, hundreds of times thinner than a hair. The more advanced the lens and the finer the lines it prints, the shallower it usually gets.

So once the height differences on the surface are about as large as the depth of focus, the same pattern prints sharp in some places and blurry in others. Where it's blurry, the printed line width is off. After polishing, the whole wafer sits inside the range that can be brought into focus. For how the printing works, read What is photolithography?.

How does CMP polish a wafer?

A CMP tool has a few main parts:

• Polishing pad: a pad larger than the wafer, stuck to a big turntable (the polishing disc). It's usually a plastic (polyurethane) full of tiny pores, a little like felt to the touch.
• Polishing head: holds the wafer upside down, face to the pad, and presses it down.
• Slurry: water with very fine particles and chemicals in it, dripped onto the pad during polishing.
• Conditioner: a disc covered in tiny diamonds that scrapes the pad now and then to keep it rough.

The turntable and the head both spin, and the wafer is polished against the pad. Why does that make it flat? Picture sanding a bumpy wooden board using a flat block covered in sandpaper. The bumps touch the sandpaper first, so they're ground off first; the dips are reached later. CMP works the same way: high spots press harder and are polished faster, so the difference in height keeps shrinking. As a rule, the harder you press and the faster it spins, the faster it polishes.

Why chemical and mechanical?

Chemicals alone would be like soaking the whole wafer: the high spots and the low spots dissolve together, and the surface stays just as uneven. Grinding alone would be like using coarse sandpaper hard: you get it flat, but it scratches easily and can damage the material underneath. The lines on a chip are a thousand times thinner than a hair and can't take that kind of damage.

CMP combines the two. The slurry changes a thin top layer so it becomes softer and easier to remove; the pad and the fine particles only need to brush over it to carry that layer away. It's a bit like scrubbing a pan with cleaning paste: the paste loosens the burnt bits, and a sponge wipes them off without a steel brush.

How does it know when to stop?

More isn't better with CMP. Often the goal is to stop just as a certain layer underneath is exposed: polish the metal until it's left only in the trenches, for example, or until you reach a harder-to-polish layer placed underneath on purpose.

The wafer faces down, so the tool can't see it and has to go by a signal. This is called endpoint detection. Two of the common signals:

• How hard the motor works: when the surface changes to a different material, the friction changes, and so does the force the motor has to put out.
• Reflected light: light shines up through a small window in the pad, and metal and insulator reflect it differently.

When the signal turns, you've reached it. Often the tool polishes a little more as set, to clear away scattered residue; this is called overpolish. Polish too little and metal is left where it shouldn't be, so neighboring lines can end up connected. Polish too much and the lines get thinner, so their resistance goes up.

What usually goes wrong in CMP?

• Dishing: a wide metal line comes out sunk in the middle, like a shallow dish. The pad isn't perfectly hard and sags a little into wide areas, and the metal and the insulator next to it polish at different speeds, so a bit more comes off the middle. Where many thin lines are packed together, the whole patch can also end up lower; this is called erosion.
• Scratches: a larger hard bit caught between wafer and pad gets dragged into a scratch. Metal can get into a scratch and connect things that shouldn't touch, which hurts yield. So the slurry supply, the pad and the conditioner all need watching.
• Leftover particles: a freshly polished surface is covered in slurry particles and polishing debris. Particles left there become defects in the next layer.
• Slowing down: as a pad wears, its surface gets glazed and smooth, and the same polishing time removes less. So a monitor wafer is run regularly and its thickness measured before and after, to see whether the polishing speed has changed. Pads and conditioners both wear out and are replaced when their time comes, as part of maintenance (PM).

Why must the wafer be cleaned right after polishing?

A wafer coming off CMP is always dirty: slurry, particles and debris are all still on it. So it's cleaned right away, usually scrubbed with very soft brushes and water to get the particles off. Whatever isn't washed off becomes a defect in the next layer. Why fabs clean wafers so many times is covered in What is wafer cleaning, and why are wafers cleaned so many times?; for where CMP sits in the whole flow, see What do the seven process areas of a chip fab do?.

In fab24hr

Tap "Clock in", choose "A day in the fab", then on the "Your role today" card tap "Production fab", swipe to the process engineer and tap "CMP" under the name: you're the CMP process engineer (marked Beta for now), and the slurry, pad and conditioner settings are yours. Late at night you're on call at home when the night technician phones: CMP-02's endpoint detection never triggered, every wafer ran to the time limit before stopping, and you decide whether to put the lot on hold or let it be cleaned and go on. On the morning shift, the amount CMP-01's monitor wafers lose has dropped for three days straight, and the pad and conditioner are near the end of their life: stop for two hours and replace them now, or lengthen the polishing time first? On the afternoon shift, the step before you has a new sputtering target, the film is thicker, and lot C-4105's wide metal lines come out sunk in the middle. On the evening shift, the brush clean module at CMP-02's exit jams, and the wet wafers that have just been polished must not dry out.

In the same area, the CMP equipment engineer looks after the polishing discs, the polishing heads, the conditioners and the slurry supply, and on the afternoon shift has to work out from curved scratches on a monitor wafer where the slurry went wrong. The R&D line has a CMP R&D process engineer too, with a single tool for the whole line, CMP-R1. In chapter 2 of the main story, The life of a chip, "Station by Station", on day 4 you polish down to the silicon nitride on CMP-R1 and choose between stopping the moment it touches and polishing a little more as the recipe says. The companies, people, tools and numbers in the game are fictional; what CMP is, how it polishes and what goes wrong come from public, general sources. For other jargon, see Chip fab jargon in plain English.