What is wafer scrap? When can't a wafer be saved, and who decides?

What does wafer scrap mean?

Wafer scrap means deciding that a wafer can't be saved and pulling it out of its lot, so it goes no further. It isn't up to whoever happens to be standing there: first the facts are checked, then the person responsible for the lot decides, and someone signs to approve it.

The international quality management standards define scrap as an action on a nonconforming item that stops it from being used as originally intended, for example recycling or destruction. The point isn't "throwing it away" but "making sure it is never treated as product again".

Scrap is often decided one wafer at a time. For the whole sequence after a problem (hold first, then release, rework or scrap), see "What is a lot hold, and who decides whether to release, rework or scrap?". This article covers only the scrap decision itself.

When does a wafer get scrapped?

Broadly, two ways: a step changed the wafer itself and got it wrong, or the wafer itself is damaged.

• Over-etched, or implanted with the wrong atoms: what etch removes can't be put back, and what ion implantation drives in can't be picked out (see "What is rework, and why can litho be redone when etch can't?").
• Wrong step, wrong recipe: a step skipped, a step done twice, the wrong recipe chosen. A university course on yield lists this kind of mis-processing, together with wafers damaged in handling, as the two sources of whole-wafer losses.
• Broken wafers: a wafer that is broken or cracked can't go into a tool again.
• Over the Q-time, and no rework possible: it waited too long and the surface has already changed.
• Out of spec, and no way back: for example, the film left after etch is already thinner than the spec.

On the other hand, out of spec doesn't automatically mean scrap: after a second measurement and a look at the records, a wafer judged still usable runs on with a deviation.

What happens before a wafer is scrapped?

Scrap can't be undone, so it usually goes in this order.

Step 1: Hold first, don't throw anything away

What: put the lot on hold and label it, leave the wafers where they are, and save the tool's log. Why: a wafer that looks ruined may still be saved, and the wafer itself is evidence for finding the cause. Otherwise: once it's gone, it can't come back even if it turns out to be savable, and the root cause search loses its evidence.

Step 2: Check wafer by wafer

What: go through the tool log and the measurements to see what each wafer actually went through and what it measures, then decide each wafer's fate. This is called disposition. Why: wafers in the same lot often had different things happen to them. Otherwise: scrap the whole lot and good wafers go too; miss one and a bad wafer travels on with the rest.

Step 3: Look for another way first

What: can it be reworked? Can it run on with a deviation? Can it be put to another use, such as being cut open for analysis? Why: the quality standards list more than one way to handle a nonconforming item, and scrap is only one of them. Otherwise: useful wafers get thrown out, and the lot comes up short for no reason.

Step 4: Count forward: is it worth finishing?

What: look at how many steps are left, how much tool time they will take, and roughly how many good chips the wafer would end up with. Why: money already spent is gone whether the wafer is scrapped or not, so the decision has to look forward. One research paper also points out that a lot with only a few wafers left still takes up tool time, and sometimes scrapping it to make room for a new full lot works out better. Otherwise: you keep going only because "so much has gone into it already" and spend more and more, or you throw away wafers that were still worth finishing.

Step 5: Sign-off and records

What: wafers to be scrapped go to someone with the authority to sign; the record says which wafers, at which step, why and who approved, and the lot's wafer count in MES is updated. Why: the quality standards require a record that describes the nonconformity, the action taken and who decided it. Otherwise: the count on the books no longer matches the box, and deliveries, root cause work and the scrap rate are all calculated wrong.

Who decides, and why does scrap need a signature?

A common split: the process engineer for that step explains what happened; process integration looks at every wafer's history and decides what happens to it (see "What does a process integration engineer (PIE) do, and how is it different from a process engineer?"); QA checks that the records and sign-offs are complete (see "What does QA (quality assurance) do in a chip fab?"). When delivery or quality is affected, the customer is told, and some cases need the customer's agreement.

The signature is there because scrap can't be taken back: one more person looking means one less chance of a wrong call. What it's called and how high the signature goes differ from fab to fab.

Why is scrap rate watched so closely?

The closer a wafer is to finished, the bigger the loss when it's scrapped; the glossary card "Scrap" already explains why. Apart from the money, scrap also means:

• Fewer wafers to ship: making up the shortfall means a new wafer start at the very first step, and months of waiting.
• Tool time used for nothing: every step a scrapped wafer passed through took up a tool.
• A clue to a problem: when scrap piles up at one step, on one tool or for one reason, it's time to go back and find the cause (see "What is root cause analysis, and how do 8D and the 5 Whys work?").

So fabs track what share of the wafers started make it all the way to the end, usually called line yield. Research on fab performance often reads it alongside yield, cycle time and on-time delivery. Yield counts the share of good chips on a finished wafer (see What does "yield" mean in chip news?); line yield counts whether the whole wafer made it through.

What happens if you watch only the scrap rate?

An often-quoted line known as Goodhart's law says that when a measure becomes a target, it stops being a good measure. Watch only the scrap rate and you may see:

• Wafers that should be scrapped run on and are only thrown out at final test, after paying for every step in between.
• Scrap not recorded as it happened, so the lesson that most needed learning is lost.

So scrap rate is read together with yield and cycle time, and only scrap that is recorded honestly leads to the cause.

Where do scrapped wafers go?

1. Marked and set apart first: anything judged nonconforming is clearly labeled and kept away from good material, so it can't be mixed back in or shipped.
2. Some go to analysis first: a cross-section shows which layer went wrong (see "What is failure analysis (FA), and how do engineers find out why a chip failed?").
3. Patterns destroyed: a product wafer carries a circuit design, so even as scrap it can't leave the fab as it is.

Wafers that are never shipped, such as monitor wafers and dummy wafers, are mostly stripped, polished and used again (see "What are monitor wafers and dummy wafers, and how does a fab tool prove it's still OK?").

In fab24hr

Tap "Clock in", choose "A day in the fab", then on the "Your role today" card tap "Production fab" and swipe to the quality engineer; which of the four shifts you get depends on the time you open the site. On the afternoon shift, "How to Record It", a manual rework etch on ETCH-05 next door was meant for 3 wafers, but the tool had the whole lot of 25 selected and was stopped only on the third. Among the things to preserve first, one choice is "Throw away the first three wafers now", and under it: "Scrap is for Integration to decide. If you throw them away now, they cannot be brought back." Ya-Ting in Integration passes on the disposition: the first three are scrapped, the 3 to be reworked get decided after the rework and a measurement, and the 19 untouched wafers continue. You compare the tool and MES section by section, then decide how to record it: split into "confirmed" and "to be confirmed", write all 25 wafers as scrapped, or write the rest as "normal". Whatever was written wrong has to be changed back at the 16:30 review meeting.

As the process integration engineer, if your rework etch instruction on the afternoon shift leaves out the line 'For a manual run, press "Deselect all" first and select only wafer 7', the tool may run a few extra wafers, and you decide wafer by wafer whether each continues or is scrapped; the ones to be scrapped still need sign-off. As the implant R&D process engineer on the R&D line, the late-night shift brings a wafer implanted with the wrong recipe: "Wafer 7 is already implanted and can't come out. Keep it or scrap it?"

The company, people, lot numbers, tools and figures in the game are fictional; how scrap is decided, who signs and where scrap goes follow the general wording of quality management standards, university course notes and published papers, and real practice differs from fab to fab. For other jargon, see Chip fab jargon in plain English.