Play

← fab24hr Blog

How chips are made8 min read

What is endpoint detection? How etch and CMP tools know when to stop

Endpoint detection (often shortened to EPD) means a tool keeps watching a signal while it etches or polishes: as soon as the layer it is working on is gone and the layer below shows, the signal bends, and the tool stops at that bend. Etch tools mostly watch the light the plasma gives off; CMP tools often watch the motor current and the light reflected from the wafer.

Why does a tool need to know when to stop?

Etching removes a layer by cutting it away; CMP removes it by polishing it off. Both have to stop at just the right place. How each one works is in "What is etching?" and "What is CMP?"; this article is only about stopping.

  • Too little: what should be separated isn't. A thin film of metal left between two metal lines joins them; a hole that should go all the way through doesn't, and the layers above and below never connect.
  • Too much: the layer underneath gets eaten too. Lines get thinner and their resistance goes up, or something built earlier, further down, gets damaged.

Why isn't timing it accurate enough?

The simplest way is to time it: the thickness to remove, divided by how much comes off per minute. The trouble is that both numbers change:

  • Every wafer's film is a different thickness: the film from the previous step is a bit thicker on this wafer and a bit thinner on that one, and the center and the edge of one wafer differ too.
  • Every tool, every day, runs at a different speed: deposits slowly build up on an etch chamber's walls, a polishing pad gets slicker with use, and the speed keeps drifting.

How does an etch tool know it's through?

Watching the plasma's light

A plasma glows, and its colors depend on what is in it (see "What is plasma?"). The material being etched is knocked off, turns into gas and mixes into the plasma, and certain colors of light get brighter. When the layer is gone and the material below shows, those colors dim; the gas that was busy etching it is no longer being used up, so its own light gets brighter instead. A spectrometer (an instrument that splits light into its colors and measures how bright each one is) looks in through a small window in the side of the chamber and watches those colors the whole time.

This method struggles with small areas: when the parts being etched take up only a small share of the wafer (tiny holes, for example), the light changes only a little and is hard to see.

Watching reflected light

The other way is to shine a beam of light on the wafer and look at what comes back. When light hits a clear film, it reflects once off the top surface and once off the bottom, and the two reflections add up, sometimes reinforcing each other and sometimes canceling out, like the colors on a soap bubble. As the film gets thinner, the reflected light swings between bright and dim; count the swings and you know how much has been etched. When the film is gone, the swings stop.

Why keep etching after the endpoint?

At the moment the signal bends, often only part of the wafer has broken through: the film isn't the same thickness everywhere, and the etch isn't equally fast everywhere. So the tool etches a little longer to make sure every hole reaches the bottom. This is over-etch.

How does a CMP tool know it's there?

During CMP the wafer is pressed face down on the polishing pad, so the tool can't see it. Common signals include these two:

  • Motor current: the harder the motor has to work to turn, the more current it draws. When the metal is polished through and the insulating layer below shows, the friction changes and the current bends with it. It's like wiping a table with a cloth: the cloth feels different on a glass top than on bare wood.
  • Reflected light: there's a small window in the polishing disc, and light shines up through it onto the wafer. Metal and insulator reflect light differently, so when the surface changes material, the reflected light changes.

Some tools have other sensors as well and watch several at once. After the bend, the tool often polishes a little longer, as set, to clear stray leftovers.

Why does the bend sometimes go missing?

To decide that the signal has "bent", a tool often checks whether it has crossed a line set in advance, called the threshold. Set it too high and a small bend goes unseen; set it too low and a bit of noise makes it stop early. Common reasons the bend gets smaller or less clear:

  • The two layers polish about the same, so the friction barely changes.
  • The area being removed is small, so the signal is small too.
  • The pad has changed: as it wears and its surface gets slick, the signal can become less steady.

And sometimes there is no bend at all: the material is the same all the way down, for example when a layer of oxide is only being polished flat and thinner. The signal stays flat, so the only way is to polish for a time worked out from the measured removal rate. That's why the removal rate is checked every day with monitor wafers (see "What are monitor wafers and dummy wafers?").

What happens when endpoint never fires? Step by step

A tool doesn't keep going forever just because it hasn't seen a bend. Besides the endpoint, the recipe (the settings the tool runs) also sets a time limit, like a safety net.

Step 1: the tool hits its time limit, stops and calls someone

  • What: when time is up, the tool stops, raises an alarm and calls the person who looks after it.
  • Why: without an endpoint, the tool doesn't know whether this wafer is done.
  • Otherwise: it keeps going, and the layer underneath is eaten away.

Step 2: hold this lot, and keep the next one out

  • What: put this lot on hold; lots queued for the same tool go to another tool or wait.
  • Why: nobody knows yet whether the problem is the signal or the tool.
  • Otherwise: the next lot goes in and may miss its endpoint too. How a held lot is judged is in "What is a lot hold?".

Step 3: measure the lot

  • What: measure how thick the remaining film is and whether anything is left on the surface (see "How does a chip fab know each step went right?").
  • Why: running to the limit can mean too much (the speed was normal, the signal was just missed) or too little (the speed dropped, and time ran out before it was done).
  • Otherwise: the call is a guess.

Step 4: judge wafer by wafer

  • What: wafers within spec are released; those that fell short can sometimes be polished or etched again (rework); those that went too far can't be put back, and if they can't be used they're scrapped. This is disposition.
  • Why: each wafer in the same lot may be in a different state.
  • Otherwise: judge too loosely and bad wafers move on; too strictly and good ones are thrown away. Which steps can be redone is in "What is rework?".

Step 5: find out why endpoint never fired

  • What: pull up the signal curves for that stretch and check one thing at a time: did the signal itself shrink; are the sensor, the window and the wiring all right; did the pad, the film from the previous step or the threshold change.
  • Why: until the cause is found, the next lot will do the same.
  • Otherwise: switching to time only to get by doesn't hold up, because the speed is changing and the time can't be pinned down. Changing the stop condition means changing the recipe, so it needs sign-off from the people responsible (see "What is change control in a chip fab?"); and once it's fixed, the tool still needs a qual run before lots go back in.

In fab24hr

Tap "Clock in", choose "A day in the fab", then on the "Your role today" card tap "Production fab", swipe to the equipment engineer and tap "CMP" under the name: you're the CMP equipment engineer (marked Beta). Which of the four shifts you get depends on the time you open the site. On the late-night shift, "The Unbroken Sensor", it's just after 02:30 and you're on-call at home when De from the night shift calls: CMP-02's endpoint detection never fired, every wafer ran to the time limit before it stopped, and the 25 wafers of C-4102 are still soaking inside the tool. Then you go through the records: the motor current and the reflected light both show no bend at the same time, yet the sensor self-check passed, and the pad has polished well above the count that was set. Whether the next lot, C-4103, switches to time only is for Pin-Yu in CMP process to decide; you tell her whether the tool can hold up.

You can also open the chemical mechanical polishing tool directly: the small window in the polishing disc is where the reflected light is read.

The company, people, tool IDs, lot numbers and figures in the game are fictional; what endpoint detection is, how it's read and what happens when it doesn't fire are based on public, general sources, and the actual practice differs from fab to fab. For other jargon, see Chip fab jargon in plain English.

Where you'll meet it in the gameAs the CMP equipment engineer (production fab › equipment › CMP), your phone rings just after 02:30 on the late-night shift: CMP-02's endpoint detection never fired, the tool ran to its time limit, and the 25 wafers of C-4102 are still soaking inside.

In fab24hr, endpoint detection stays silent at 02:30. Is the sensor broken, or has the surface being polished changed? You read it from the curves yourself.

Play as the CMP equipment engineer and find out why endpoint never fired

Teachers: this article can go with "fab24hr in the classroom" in class.

Cite this article"What is endpoint detection? How etch and CMP tools know when to stop" fab24hr, updated . https://fab24hr.com/en/blog/endpoint-detection/
Free to use in class and for non-commercial teaching: on screen, in handouts, printed for students. Please credit fab24hr. Do not sell it or present it as your own work. Spot a mistake? Report it

More articles

More on this site