What is FDC? How a chip fab keeps watch on its tools' sensors

FDC stands for fault detection and classification. In a chip fab, it's the system that keeps watch on the tools: as a tool runs each wafer, it keeps recording pressure, temperature, power and gas flow, and FDC compares those numbers with "what normal looks like". When one goes out of range, it sends an alert to call someone, and helps work out what kind of problem it is. The letters FDC mean other things in other fields; this article is only about tool monitoring in a chip fab.

What does a tool keep recording?

Take an etch tool. While it runs, the sensors inside keep measuring:

• Pressure: how much gas pressure there is in the sealed chamber.
• Power: how much electrical power goes in to light the plasma, and how much doesn't get in and bounces back (for plasma, see What is plasma?).
• Gas flow: how much of each gas flows in per minute, often written in sccm.
• Temperature: for example, how hot the stage under the wafer is.
• Helium on the back: the thin layer that carries heat away from the wafer (helium backside pressure). The higher its flow, the more is leaking, often because the wafer isn't sitting tightly.

These numbers are logged reading by reading over time, often about once a second, so every wafer leaves a whole set of curves, far too many for people to watch: only a computer can keep up.

Think of a fitness band that tracks your heart rate: a fast heartbeat while you're running is normal; the same heartbeat while you're sitting still is strange. FDC works the same way. What matters is "at this step, what is this number usually?"

Semiconductor manufacturing textbooks split measurements into two kinds: measuring the wafer (how thick a film is, how wide a line is) and measuring the tool (pressure, power, temperature). Metrology looks at the first kind; FDC looks at the second.

Why not just wait for measurement?

Because measurement takes time, and only a few wafers get measured. You find out whether a film thickness or line width is right only after the wafers are done and sent to be measured (see How does a chip fab know each step went right?). The monitor wafers a tool runs on a schedule only come every so often, too (see What are monitor wafers and dummy wafers?). Between two measurements, a tool may already have run several lots.

FDC looks at every wafer, and it looks while the work is happening. One paper puts the difference this way: when checks are done afterwards on samples, a problem can affect many wafers; when the tool is watched during processing, the problem can be caught on the very wafer being processed, so far fewer are affected.

How does FDC work? Step by step

Step 1: Record the curves for every wafer

• What: as each wafer runs, the sensor readings are saved as curves, labeled with the tool, the chamber, the wafer and which step of the recipe it was on.
• Why: the same reading is naturally different at different recipe steps; you can only compare once you know which step and which wafer.
• Otherwise: keep only one overall average, and you can't tell which step went wrong or find which wafers were affected.

Step 2: Compare with "what normal looks like"

• What: from a stretch of past wafers that came out fine, set the range each step normally falls in. That range is the threshold. A common way is to cut the curves at the recipe steps and compare a few numbers that sum up each piece, such as the average.
• Why: a tool wobbles a little on every wafer, so you need to tell "the usual wobble" apart from "something really is different".
• Otherwise: set the thresholds carelessly, and they either go off all the time or stay quiet when they should go off.

Step 3: Out of range, send an alert

• What: when a reading crosses its threshold, FDC sends an alert to the person looking after that tool, saying which tool, which reading and which wafer.
• Why: the sooner someone knows, the better the chance the next wafer won't be ruined too.
• Otherwise: the number quietly goes out of range, nobody sees it, and by the time measurement catches it, several lots have gone by.

Step 4: Work out what kind of problem it is (classification)

• What: decide what this alert is: something really broken (for example, a part that controls one gas's flow is leaking, or a part has worn), the sensor itself cutting out for a moment, or a normal jump that always happens (for example, the moment the recipe moves on to the next step). This is the C in FDC: classification.
• Why: different problems need different people and different fixes.
• Otherwise: treat every alert as real and people run themselves ragged; treat every one as a false alarm and the real one gets dismissed.

Step 5: Once it's confirmed, follow the OCAP

• What: follow the response plan written in advance, the OCAP: stop the tool, put the wafers that might be bad on hold, and notify the engineer. After the repair, do a qual run, and only when it passes is the tool released.
• Why: stop the damage first; the cause can be found a little later.
• Otherwise: keep running while you investigate, and more wafers go bad while you're looking.

What an OCAP says is in What does OCAP mean in a chip fab?; how a held lot gets decided is in What is a lot hold?.

How tight should the thresholds be?

This is the hardest part of FDC.

• Too tight: it goes off at any normal wobble. A paper that studied etch tool data put it this way: the normal variation is already large, while the real problems are often only slightly different. When alerts come too often, people get into the habit of dismissing them. Industrial alarm management standards call alarms that go off too often or for no need "nuisance alarms"; when there are many of them, the one that really matters sounds exactly like the rest.
• Too loose: things go quiet, but the slow kind of decline slips through. For example, an electrostatic chuck wears a little each day, a little more helium leaks from behind the wafer each day, and every wafer is still in range. A check that only asks "is this one point over the line?" is weak at catching small, slow shifts like that, so you also watch the trend: is each wafer a little higher than the last?

Thresholds aren't set once and done, either: after a PM or a part replacement, the curves may change shape. Setting and maintaining thresholds takes a lot of work, so there is research on letting the system learn from the data what normal looks like, aiming for fewer false alarms and fewer missed ones at the same time.

How is FDC different from SPC and APC?

The three acronyms often show up together, but they look at different things:

	What it looks at	When	The question it answers
FDC	The curves from the tool's sensors	While processing, every wafer	Is the tool acting strange?
SPC	Measured results, plotted on a control chart	After processing and measuring	Is the process still behaving as usual?
APC	How far measured results are from target	After measuring, before the next lot	How much should the next lot be adjusted?

How to read an SPC control chart is in What is an SPC control chart?; how APC works out corrections is in What is APC?. When FDC decides whether a reading is out of range, it often borrows the idea of a control chart, except that what goes on the chart is the tool's numbers, not measured values. Many sources also count FDC as part of APC.

The three are often used together: FDC catches "something is wrong with the tool" while it runs, SPC watches "is the process drifting?", and APC brings the next lot back to target. The data FDC keeps can also be used to schedule maintenance by the tool's actual condition (see What is PM (preventive maintenance)?).

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 "Etch" under the name: you're the etch equipment engineer. On the evening shift, "False Alarms All Night", you stay late to write a repair report and FDC pages you one alert after another: ETCH-05's reflected power jumps for the second the recipe changes steps; MET-01's stage vacuum drops to 0 and comes straight back, because the sensor signal cut out. For each one you choose: go and look, or dismiss it. Kai texts that the etch thresholds were set in a hurry half a year ago; whether to loosen them by 20% is up to you. If you don't, at 21:00 one more alert comes in, for the helium flow on Chamber A of ETCH-03: it didn't jump up, it crept up.

In chapter 5 of "The life of a chip", "The Last Box", you're the etch R&D process engineer on the R&D line: the POR to be handed to the production fab has one box left, the alarm limit for helium flow. Three months of the line show three spikes: a loose sensor connector, the first wafer after a PM, and only one from a worn electrostatic chuck. Write 12 and it goes off three times in three months, once for real; write 16 and it never goes off. Chapters 6 and 7 are the production fab two months later: at 02:47 ETCH-03 goes down, and looking back, the helium flow has been creeping up for three days, and the yellow-light warning was dismissed 6 times.

The companies, people, tool numbers and figures in the game are fictional; what FDC is and how it's used follows public, general sources, and practice differs from fab to fab. For other jargon, see Chip fab jargon in plain English.