How does a chip fab know each step went right? Metrology vs inspection

A chip fab knows whether a step went right through two kinds of checks: metrology and inspection. After each important step, some of the wafers are checked. Metrology asks "how much": how thick a film is, how wide a line is (line width), how well this layer lines up with the one below (overlay). Inspection asks "is there anything wrong": it scans the whole wafer for extra particles and scratches, or for patterns with a missing corner or a break. The numbers go on a chart, and when they drift, someone goes looking for the cause.

Why check along the way?

A wafer goes through hundreds of steps over several weeks before it can be powered up and tested to see whether its chips work. If only then you found out that step 50 had gone wrong, weeks of work would be wasted, and other wafers that went through the same tool could be ruined too.

So a fab works like a building site: each floor is inspected before the next one goes up, instead of finding a crooked column on the third floor once the building is finished. The sooner a problem is caught, the fewer chips are lost and the higher the yield (see What does "yield" mean in chip news?).

Some mistakes can still be fixed. If litho prints a layer off target, the resist can be stripped and the layer printed again before etch; once it's etched, there's no going back. That's why the measurement right after litho matters so much.

What's the difference between metrology and inspection?

	Metrology	Inspection
The question	How much?	Is anything wrong?
What you get	A number: how thick, how wide, how far off	Where the suspicious spots are, and how many
How it works	Measures a few chosen spots very precisely	Scans the whole wafer for anything that shouldn't be there
It's like	Using a ruler	Proofreading a whole page for typos

The two usually work together: a layer can be exactly the right thickness and still carry three specks of dust. Both have to be right for the chip to be good.

How is film thickness measured? With light

Shine light on a transparent film, and part of it reflects off the top surface while part goes in and reflects off the bottom. The two reflections overlap: some colors get stronger, others cancel out. That's why soap bubbles and oil on water show rainbow colors, and the colors depend on how thick the film is.

The tool records how much of each color comes back, then works backwards with a model: how thick must the film be to reflect light like this? Another method shines light in at an angle and looks at how the direction the light vibrates in changes after reflecting. It's called ellipsometry, and it can measure even very thin films. Neither one touches the wafer.

Metal doesn't let light through, so for metal films it is common to measure how well they conduct instead (sheet resistance). Each wafer is measured at many points to see whether the center and the edge are equally thick. For how films are grown, see What is thin film deposition?.

How is line width measured? With an electron microscope

The finest lines are only a few tens of nanometers wide, or even narrower. Visible light has wavelengths of about 400 to 700 nanometers, and an ordinary optical microscope can't make out details much smaller than about 200 nanometers. Electrons have far shorter wavelengths, so line width is usually measured with an electron microscope.

The tool takes a picture of the line from directly above, finds where its two edges are in the image, and works out how wide it is. The industry often calls this number the CD (critical dimension). When imaging photoresist, the electron energy is turned down so the surface doesn't build up charge and distort the image.

How is overlay measured?

Why overlay matters is covered in What is photolithography?. Measuring it is simple: in the scribe lines between chips, every layer leaves a dedicated mark, for example a large square on the layer below and a small square printed inside it on this layer. An optical microscope takes a picture, and the tool checks whether the small square sits exactly in the middle, and if not, by how much and in which direction. The offset is passed back to the scanner so that the next lot is corrected.

How are defects found? By comparing with the chip next door

A bare wafer with no pattern has a flat surface, so it is the easiest. A laser scans across it; when the light hits a particle, it scatters sideways, the detector sees a flash, and that's one particle. It's like sunlight coming into a room and lighting up dust you normally can't see. For how cleanrooms keep particles down, see How clean is a cleanroom?.

A patterned wafer is harder: its surface is already covered in lines. Luckily, every chip on a wafer carries exactly the same pattern. So the tool lays the image of one chip over the image of the chip next to it and subtracts: whatever is the same cancels out, and whatever is left is a possible defect. It's like a spot-the-difference puzzle: put two pictures side by side and the difference jumps out.

This method has a blind spot. If every chip has the same problem in the same place, as can happen with a speck on the photomask, the subtraction cancels that out too. Those are found by comparing with the design data or by inspecting the photomask itself.

Optical inspection scans fast but can't see clearly what a spot actually is. So once the suspicious locations are known, an electron microscope goes to those coordinates for close-ups, to tell a particle from a scratch or from two lines stuck together, and sorts them one by one.

Why isn't every wafer measured?

Measuring doesn't make a wafer any better; it only confirms. Every measurement takes time, and metrology tools are expensive. Measure every wafer at every point, and lots pile up in front of the metrology tools, so each wafer takes longer to finish.

So most of it is sampling. A lot is usually one box of up to 25 wafers; a few are picked, and each is measured at several points. At some steps, when a tool has been stable all along, some lots skip the measurement entirely.

Sampling has a price too. If a tool slowly drifts between two measurements, every wafer it processed in that window may be bad and has to be tracked down lot by lot. So it's a trade-off between measuring more and feeling safe, and measuring less and moving faster; when a tool hasn't been stable lately, more lots get measured.

Apart from the product, there are monitor wafers: not customer wafers, but wafers used only to check the tool. One is run through the tool on a schedule to measure film thickness or count particles and see whether the tool itself is changing. The qual run after a tool is repaired works the same way.

What happens to the numbers?

Each measured lot becomes one point on a control chart. A point outside the limits, or several in a row drifting to the same side, counts as abnormal. Being outside the control limits and being outside the product spec are two different things (OOC and OOS; the glossary explains the difference, so it isn't repeated here).

When something is abnormal, the work stops until the cause is found: the lot is put on hold and handled step by step by the OCAP. Often the lot is measured again, because a changed number can mean the process changed, or that the metrology tool itself has drifted. That's why metrology tools also regularly measure a reference wafer whose values are known, to make sure they haven't drifted. Once it's clear the process is the cause, the engineers look at the tool and at the steps before it.

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 "Etch" under the name: you're the etch process engineer. On the morning shift, the handover sheet has a control chart nobody has acted on: on metrology tool MET-01, eight post-etch line width points in a row sit below the center line, each lower than the last, yet none is outside the limits. Your inbox has three OOC lots waiting for your call; later that morning Ya-Ting calls about the lot in Metrology: release it, or keep it? After the 16:30 review meeting, you can go back to the metrology tool for another look: with the part replaced, is the line width back on the center line? A random event can also bring another tool's film thickness chart drifting steadily upward.

The litho and thin film process engineers each have a shift that starts with a measurement too: for one, an overlay out of spec; for the other, wafers warped more than usual. The company, people, tool numbers and figures in the game are fictional; what metrology and inspection do, and how they measure, follows public, general sources. For other jargon, see Chip fab jargon in plain English.