How are silicon wafers made? From sand to a polished wafer

Silicon wafers are made like this: the silicon dioxide in sand (silica sand) is turned into silicon and purified until it is extremely pure. The silicon is melted, a small seed crystal is dipped in, and the seed is pulled slowly upward while it turns, growing a long, round crystal whose atoms are all lined up in neat rows. That crystal is called an ingot. The ingot is ground round, sliced into thin discs, ground flat, etched to remove the damage left by sawing, and polished to a mirror finish. After cleaning and inspection, the wafers are boxed and shipped to a chip fab. Growing the ingot is called crystal growth, or crystal pulling.

This article covers only how the wafer itself is made. For what happens after it reaches the fab, see "How are computer chips made?".

What are the steps in silicon wafer manufacturing?

There are roughly nine. Here they are in order.

Step 1: Turn sand into silicon

Sand and quartz are mostly silicon dioxide, silicon bound to oxygen, and the two have to be pulled apart first. Lumps of quartz go into an electric arc furnace with carbon such as coke or wood chips and are heated to around 2,000 °C; the carbon carries the oxygen off as carbon monoxide gas, and silicon is left behind. This silicon is about 98 to 99 percent pure and is called metallurgical-grade silicon. The last one or two percent is mostly iron and aluminum, which is still far too dirty for chips.

Step 2: Purify it until it is extremely pure

The metallurgical-grade silicon is crushed and reacted with hydrogen chloride to make trichlorosilane, a compound that is liquid at room temperature, so it can be distilled again and again, the way spirits are, to separate out the impurities. It is then broken down at high temperature, and silicon builds up layer by layer on heated thin silicon rods. The result is polysilicon: many small crystals stuck together, each pointing a different way, but already at least 99.9999999 percent pure ("nine nines"), with roughly one impurity atom per billion. It has to be this pure because the boron and phosphorus a chip gets on purpose are, in the lightly doped places, about one atom per hundred million silicon atoms (see "What is a semiconductor?"). Unwanted impurities must be far fewer than that, or the silicon won't conduct the way it should, and metal impurities make transistors leak.

Step 3: Melt it, dip in a seed, and pull it up while turning

The most common way is the Czochralski (CZ) method. Chunks of polysilicon go into a quartz crucible and are melted in an argon-filled furnace at a little over 1,400 °C (silicon melts at about 1,414 °C). To make the whole ingot p-type or n-type, a tiny amount of boron or phosphorus goes into the melt now. A small seed of single-crystal silicon is lowered until it just touches the surface. The seed and the crucible turn slowly in opposite directions while the seed is pulled up very slowly. As the molten silicon freezes under the seed, it carries on the seed's arrangement of atoms, layer after layer. The seed is like the person at the front of a line: everyone behind lines up after it. So the whole ingot is one single crystal.

At the start, a short, thin "neck" is pulled. When the seed first touches the hot melt, the sudden heat makes some atoms line up wrong. Because the start is so thin, those faults run out to the surface within the neck and stop there instead of carrying on into the ingot. Then the crystal is widened to the target diameter, grown at that width, and finally narrowed and lifted off the melt.

The pull is very slow: around a millimeter a minute, only a few centimeters an hour. An ingot can grow to about 2 meters long. Going by the density of silicon, an ingot 30 cm across and 1.5 meters long weighs over 200 kg, and the pulling alone takes about a day. Because it keeps turning as it grows, it comes out round (see "Why are wafers round?"). Pull too fast, or let the temperature waver, and the atoms don't have time to line up; faults appear, and that part may not be usable for chips.

Most wafers are grown this way. Another method, float-zone (FZ), uses no crucible and gives purer silicon, but its crystals can't be made as large, so it is used mostly for power devices.

Step 4: Cut off the ends, grind the outside, mark the direction

The two ends of the ingot are pointed, so they are cut off. The width of the middle also varies a little, so the outside is ground to an exact diameter: 300 mm for a 12-inch wafer (see "What do 12-inch and 8-inch wafers mean?"). Then the direction of the crystal is measured, and a small groove is ground along the side. After slicing, that groove is the small notch on each wafer's edge (smaller wafers get a flat edge instead), which tools use to turn every wafer to the same angle. Skip this, and wafers come out in different sizes, don't sit right in the tools, and nobody can tell which way they face.

Step 5: Slice into thin discs

Many thin steel wires run side by side, moving back and forth, carrying an abrasive or coated with diamond. The ingot is pressed into them and cut into many slices at once, like an egg slicer. Thin wire means a narrow cut, so less silicon turns into dust; each slice is also cut a little thicker than the finished wafer, to leave material for later grinding. However it's cut, the surface is left with saw marks, and under them a damaged layer knocked out of order, so the slices aren't usable yet. (Cutting a finished wafer into chips is a different job; see "How is a wafer cut into chips?".)

Step 6: Round the edges, grind flat, etch off the damage

First the edges are rounded: a sharp corner chips at the slightest knock, and the wafer can crack from there. Then both faces are ground flat so the thickness is even across the whole wafer. Finally the wafers go into an acid or alkaline solution that dissolves the saw marks and the damaged layer from grinding. You can't see that layer, but its atoms are already out of order; leave it, and on the surface, the neat crystal from step 3 is wasted.

Step 7: Polish to a mirror

The wafer is pressed onto a turning polishing pad while a slurry with very fine particles drips onto it: the chemicals soften the surface a little and the particles wipe it away, until it shines like a mirror. It works the same way as CMP inside a fab. It has to be this flat because the scanner that later prints the circuits is only in focus within a very shallow range of height; if the surface rises and falls too much, the same pattern prints sharp in some places and blurry in others (see "What is CMP?"). Some wafers also get a thin, neatly ordered layer of silicon grown on top after polishing; these are epi wafers (see "What is epitaxy?").

Step 8: Clean and inspect

After polishing, the surface still carries slurry particles and traces of metal, which are washed off with chemicals and ultrapure water (see "What is wafer cleaning?"). Then the thickness is measured, along with how flat the wafer is, overall and over each small area. A laser scans the whole wafer: the mirror-like surface reflects the light straight back, but where it hits a particle or a scratch, the light scatters to the side, and the machine records a spot. One particle in the wrong place can ruin a chip (see "How clean is a cleanroom?"), so wafers that fail don't ship.

Step 9: Seal them in a box and ship them to the fab

Good wafers go into sealed boxes made for shipping. For 12-inch wafers it's the kind that opens at the front, and it keeps out dust and knocks along the way. At the fab, the wafers are moved into the fab's own wafer pod, the FOUP (see "How do wafers move around a fab?"), and put on the line. That is wafer start (see "What is a wafer start?").

Why does it have to be a single crystal?

In a single crystal, all the silicon atoms across the wafer face the same way in neat rows, like a room tiled with every tile pointing the same direction. Polysilicon is like many small patches tiled separately, and the seams between them don't line up. Electrons move easily through the neat rows but get caught at the seams, and current leaks more easily near them. A chip can hold hundreds of millions of transistors or more, and each one needs to behave about the same, so chips are built on a single crystal.

In fab24hr

fab24hr has no wafer plant: the game begins once the wafers are in the fab, waiting for wafer start. The closest is chapter 1 of the main story, The life of a chip, "How to Split 12". It's the day before wafer start, and you're the R&D integration engineer on the R&D line, deciding at the split meeting how to split the first 12 wafers of Glowmere, a fictional hearing-aid chip. In chapter 2, "Station by Station", Day 1 is called "Wafer start": the 12 wafers are cleaned at the wet bench WET-R1, then go into the furnace FURN-R1. When you load the boat, you choose between a monitor wafer at the top, middle and bottom, or just one in the middle. The game describes a monitor wafer as a blank silicon wafer, one with no pattern on it yet.

In the codex, under "Glossary", "Wafer" also has a "Know a bit more" that tells how single-crystal silicon is pulled. The company, people, chip and tool IDs in the game are fictional; how wafers are made comes from public, general sources, simplified. For other jargon, see Chip fab jargon in plain English.