The Day the CPK Report Came Out, the Whole Room Fell Silent for Three Seconds
I remember years ago, we were introducing a new product, and everyone was putting in their best effort. The first batch of wafers came out, and the yield wasn't bad, but as soon as the CPK report was released, a certain critical parameter directly dropped to 1.08, with a DPMO of 6210. The entire conference room instantly fell silent, so quiet you could hear the buzzing of the lights. The boss frowned and asked, "Is this data real, or is our measurement system playing tricks?" Frankly, that was a good question, because sometimes poor process capability isn't due to the process itself being bad, but rather because you haven't clearly distinguished which part is true variation and which part is noise introduced by the measurement system.
What's the Problem?
To put it simply, any measurement data we observe actually consists of two parts: one part is "the true variation of the process itself," meaning the differences caused by your equipment, materials, parameter settings, etc.; the other part is "measurement system variation," which could stem from unstable measurement equipment, different operator techniques, or inherent errors in the measurement method itself.
If you don't separate the measurement system's variation from the total variation, then when you see a low CPK, you're very likely to attribute all problems to the process, and then foolishly adjust parameters and modify equipment, only to find the CPK hasn't moved after much effort. In other words, you're spending effort solving a "non-existent process problem," because the root cause of the issue is your measurement tool itself.
How Is It Actually Done?
The most common method is to perform a Gage R&R (Repeatability & Reproducibility) study. Simply put, it involves selecting a few representative wafers and having different operators measure them multiple times using the same measurement equipment. Then, through statistical analysis, the total variation is decomposed into "variation caused by the equipment (Repeatability)" and "variation caused by the operators (Reproducibility)."
For example, if your Gage R&R results show that measurement system variation accounts for over 30% of the total variation, then congratulations, your process capability appears poor likely just because your measurement system isn't stable enough. In this scenario, instead of altering process parameters, you should first find ways to improve the stability of the measurement equipment, train operators, or revise the measurement SOP. Once you filter out the noise from the measurement system, you might be surprised to find that your process's true CPK is much better than you imagined.
The Most Common Pitfalls
The most common pitfall I've seen is that after completing a Gage R&R study, if the values aren't satisfactory, people directly "calibrate" the measurement equipment and then redo the study until the numbers are pleasing. This is self-deception! Honestly, calibration only makes your measurement equipment "accurate," but it doesn't necessarily solve the issues of "repeatability" and "reproducibility." I've even seen departments deliberately select wafers with "very stable measurement results" to conduct the Gage R&R study just to make the numbers look good. This completely defeats the purpose of Gage R&R, as it cannot reflect real-world variation. You must use samples and operating methods that are as close as possible to actual production conditions for evaluation to obtain meaningful results.
One Thing You Can Do Today
Go back and check: how long has it been since you performed a Gage R&R study for the measurement systems of your critical parameters?