Process Control and SPC

Lesson 4 ended on chamber matching, keeping multiple instances of nominally identical equipment behaving consistently. This lesson is about the broader discipline that question belongs to: statistical process control (SPC), the methodology a fab uses to tell, from measured data alone, whether a process is behaving consistently, and whether that consistent behavior is actually good enough for the product's requirements. Both questions matter, and they are not the same question: a process can be perfectly stable and consistently miss its target, or it can be centered right on target but swing unpredictably from lot to lot.

Reading time: about 9 minutes.

5.1 Control charts

A control chart plots a measured process parameter (a CD, a film thickness, a particle count, anything tracked regularly) against time or lot sequence, with a center line (commonly the process average) and upper and lower control limits, typically set at three standard deviations above and below the center line, calculated from the process's own historical, in-control data, not from the product's specification limits. This distinction matters: control limits describe what the process itself normally does when behaving consistently; specification limits describe what the product actually needs. A point outside the control limits signals the process has changed in some statistically identifiable way, worth investigating, regardless of whether that point is still comfortably inside spec.

5.2 Western Electric rules

A single point outside the three-sigma control limits is the most obvious out-of-control signal, but a process can also show subtler, non-random patterns within the limits that still indicate something has changed. The Western Electric rules are a widely used set of pattern-based tests applied to a control chart, each flagging a different kind of non-random behavior: for example, a run of several consecutive points all on the same side of the center line (suggesting a sustained shift in the process mean, even if no single point crosses a control limit), or a run of points steadily trending in one direction (suggesting a gradual drift, such as a slowly degrading component or an aging consumable), or two out of three consecutive points falling in the outer third of the control-limit range on the same side (a milder but still statistically unlikely pattern under normal random variation). None of these patterns individually proves the process has actually failed, but each is specifically chosen because it is unlikely to occur by chance alone if the process is truly behaving randomly and consistently, making it worth an engineer's attention when it appears.

5.3 Process capability: Cp and Cpk

A stable, in-control process (5.1-5.2) is not automatically a capable one, capability asks a different question: given the process's natural variation, how well does it actually fit inside the product's specification limits? Two standard capability indices answer this:

• Cp compares the spread of the specification limits to the spread of the process's natural variation, specifically the specification range divided by six standard deviations of the process. Cp says nothing about where the process is centered, only how wide the spec is relative to the process's own spread.
• Cpk accounts for centering as well, taking the smaller of two calculations, (upper spec limit minus process mean) and (process mean minus lower spec limit), each divided by three standard deviations. A process can have a good Cp (comfortably narrow relative to spec) but a poor Cpk if it is badly off-center, running close to one spec limit even though its spread alone would easily fit between both limits.

A process with Cpk well above 1 is comfortably capable, with its natural variation well clear of both spec limits even accounting for centering; a Cpk near or below 1 means the process's own variation, even if statistically stable, is genuinely at risk of producing out-of-spec material, a capability problem distinct from (and generally harder to fix than) an out-of-control signal, since a capability problem often needs the process itself improved or re-centered, not just a single assignable cause found and corrected.

5.4 Run-to-run control and FDC

Two further, more automated control layers build on the same statistical foundation:

• Run-to-run (R2R) control: rather than waiting for a control chart to flag an out-of-control signal after the fact, an R2R system automatically adjusts a tool's recipe settings for the next run, based on measured results from recent runs, specifically to keep the process centered on target continuously, compensating for slow, predictable drifts (for example, gradual pad wear in CMP, lesson 9 of the Lithography, Etch & CMP path) before they accumulate into an out-of-control or out-of-spec excursion.
• Fault detection and classification (FDC): rather than monitoring only a wafer's final measured result, FDC monitors a tool's own real-time process sensor data (chamber pressure, RF power, gas flow, temperature traces, and similar signals) during the run itself, flagging an abnormal sensor trace as a potential fault even before the wafer is measured afterward, and in many cases classifying what kind of fault the abnormal pattern likely indicates. FDC catches some problems earlier than a post-process measurement alone could, directly at the point where the anomaly actually occurred.

5.5 Interview questions

What is the difference between control limits and specification limits on a control chart? Control limits are calculated from the process's own historical in-control variation (commonly three standard deviations from the mean) and describe what the process itself normally does; specification limits describe what the product actually needs, and the two are generally different values.

Why would an engineer act on a Western Electric rule violation even if every point is still inside the control limits? Patterns like a sustained run of points on one side of the center line are statistically unlikely under truly random, stable behavior, and can flag a real process shift (for example a gradual drift) before any single point crosses a control limit, catching the problem earlier.

What is the difference between Cp and Cpk? Cp compares specification spread to process spread without regard to centering; Cpk also accounts for how centered the process is within the spec limits, and can be much lower than Cp if the process is running close to one limit.

What does run-to-run control do, and what kind of problem is it good at compensating for? It automatically adjusts a tool's recipe settings for the next run based on recent measured results, well suited to compensating for slow, predictable drifts like gradual consumable wear before they accumulate into an excursion.

How does FDC differ from monitoring a wafer's final measured result? FDC monitors a tool's own real-time process sensor data during the run itself, flagging an abnormal trace as a potential fault (and often classifying its likely cause) before the wafer is even measured afterward, catching some problems earlier than post-process measurement alone could.