What is APC? How a chip fab corrects itself after measuring: feedback and feedforward

APC stands for advanced process control. In a chip fab, it means letting a system make small adjustments to a process by fixed rules: after each lot of wafers is processed and measured, the system compares the numbers with the target, works out how much to change the settings for the next lot, and sends them to the tool. The most common kind is called run-to-run control (often written R2R). When people in a fab talk about a correction for the next lot, this is usually what's doing the math. The letters APC mean other things in other fields; this article is only about process control in a chip fab.

Why keep adjusting? Tools slowly drift

A tool follows a fixed set of settings called a recipe (see "What is a recipe in semiconductor manufacturing?"). But the same settings don't always give the same result this month as next: deposits slowly build up on the inner walls of an etch chamber, so the etch speed changes; a CMP polishing pad gets smoother with use, so the same time removes less.

It's like an oven that's a few years old: same temperature, same time, and the results slowly change. Left alone, line width and film thickness creep away from target bit by bit. Regular maintenance (PM) brings a tool back; the slow drift between two rounds of maintenance is what APC looks after.

How does feedback work? Step by step

Feedback means looking at the tool's own latest results and correcting the next lot. One loop goes like this:

Step 1: Run this lot with the current settings

• What: The tool runs the recipe. Most of the numbers are locked; only a few agreed in advance are left for the system to adjust, such as the polish time or the scanner's alignment position.
• Why: If only the agreed numbers move, then when the result changes, you know which number did it.
• If not: If anything can be adjusted, nobody can trace what happened when something goes wrong.

Step 2: Measure

• What: A few wafers are measured for line width, film thickness, or how well this layer lines up with the one below (overlay; see "What is overlay?"). How it's measured: "How does a chip fab know each step went right?".
• Why: The system can only correct from numbers. Without a measurement, it doesn't know how far off things are.
• If not: With no new numbers, it can only keep running the last settings.

Step 3: Compare with the target

• What: Subtract the target from the measurement. Say the target is 40 and you measure 41: that's 1 over.
• Why: What you adjust is "how far off, and which way".
• If not: A measurement on its own doesn't tell you which way to adjust.

Step 4: Work out the next lot's adjustment by the rules

• What: Instead of correcting the whole difference at once, the system takes a weighted average of the last few lots' errors, with newer lots counting more: for example, the newest lot counts 30% and everything before it, rolled together, counts 70%. Then it uses a relationship measured in advance (for example, about how much more one extra second of polishing removes) to turn that into the next lot's settings.
• Why: Every measurement has noise mixed in: the measurement itself jumps a little, and no two wafers are exactly alike. Averaged out, one jump only nudges the settings a little; when a tool is really drifting, the errors keep landing on the same side lot after lot, and they add up.
• If not: Correct too hard and you swing back and forth with the noise: this lot happens to be high, so you push all the way down; the next lot's noise happens to be low, on top of the push, so it comes out even lower, and it ends up swinging more than if you'd left it alone. Correct too gently and you can't keep up with the drift.

It's like adjusting the water in a shower: crank it every time it feels cold, crank it back every time it feels hot, and it's scalding one moment and freezing the next. Turn it a little, wait, then check, and it settles.

Step 5: Run the next lot with the new settings, then measure again

• What: The new numbers go to the tool, the next lot runs with them, and after that it's back to Step 2. Every adjustment is recorded.
• Why: The tool keeps changing, so the correction has to keep up.
• If not: Correct once and walk away, and the tool keeps drifting until the error creeps back.

And what is feedforward? News from the step before

Feedforward works the other way round: instead of looking at the tool's own results, it looks at the lot coming in and adjusts before running it.

Say the step before deposits a film, and measures this lot a little thicker than usual. That number is passed on to the next step, CMP, which polishes a little longer, and the lot comes out on target. Etch can also adjust a little based on the line width measured at the step before. Think of cooking rice:

• Feedback: this pot came out too hard, so add a little more water next time.
• Feedforward: this bag is new-crop rice and still moist, so use a little less water before it goes in.

The two are often used together: feedforward handles "what's coming in is different", and feedback handles "my own tool is changing".

"Correct the next lot" and "leave it alone inside the limits": isn't that a contradiction?

If you've read "What is an SPC control chart?", you might ask: that article says not to adjust a tool for small wobbles inside the control limits, because it only makes things worse; this one says every lot gets a correction. The two don't clash. The difference is how you adjust:

• The "don't adjust" in SPC: a person sees one point a little off, adjusts the tool by hand, and corrects the full amount. One point is mostly noise; acting on it as if it were real just adds the wobble in a second time.
• The correction in APC: it follows rules designed in advance, adjusts only the agreed numbers, averages several lots, and moves only a little each time. It's aimed at slow drift in one direction; a single burst of noise moves it only slightly.

The two ideas were put together from the start: in the early 1990s, the university research that first proposed run-to-run control was about combining statistical process control with feedback control. Today they're still often used together: the results after APC are plotted on a control chart as usual.

What you see	Most likely	What to do
Small ups and downs from lot to lot	Noise	Leave it; APC moves only a little
Lot after lot creeping one way	The tool is drifting	APC corrects it by the rules
A sudden big jump	Something broke or changed	The control chart flags it (OOC), and people stop to find the cause

When can't you leave it to the system?

APC assumes two things: the measurements are right, and the tool is only changing slowly, not suddenly a different tool. When either isn't true, a person needs to look:

• The metrology tool itself is off: the system thinks the product is off, corrects for a wrong number, and pushes the process further and further off. So when a number suddenly jumps, it's often measured again first.
• After maintenance or a part change: the tool is in a new state, and the average of the last few lots no longer describes it. It needs a qual run first, and the system needs a fresh starting point.
• A new product, or one that hasn't run in a long time: there are no recent lots to average. Often one wafer is run ahead and measured before deciding how to run the rest.
• A correction that keeps growing in one direction: it often means something in the tool is wearing out or building up, and the correction is only hiding it. What it needs is maintenance, not more correction.

How is APC different from FDC?

In short, FDC watches the tool's sensor readings while it runs (pressure, temperature, power) and calls someone as soon as one crosses a threshold; it's asking "has something gone wrong with the tool?". APC looks at the measured results after the run and works out "how should the next lot be adjusted to hit target?". Many sources count both as part of APC; this article is about the adjusting half.

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 "Litho" under the name: you're the litho process engineer. On the morning shift, P-3A08's overlay is out of spec: you read the arrows on the measurement map, try shift, rotation and scale corrections, find the one that leaves only short arrows, and add it to every lot after it. The layer below was exposed on a different scanner, so you also choose: run the most critical layers on the same scanner from now on, or add the difference between the two scanners into the correction and let the correction keep chasing it. Whether P-3A08 itself gets reworked is your call too.

As the litho R&D process engineer on the R&D line, the morning shift is the first alignment to a new layer, with no earlier lots to go by. As the CMP process engineer, on the afternoon shift the step before has put in a new sputtering target and its film comes out thicker; that news needs to reach CMP first, which is the idea behind feedforward.

In the game, you make these corrections yourself; in a real fab, the small everyday corrections are often worked out by a system following rules, and people look after the rules and the exceptions. The companies, people, tool numbers and figures in the game are fictional; what APC is follows public, general sources, and practice differs from fab to fab. For other jargon, see Chip fab jargon in plain English.