That day, when the CPK report came out, the room fell silent for three seconds, and I thought, "Oh no, we have to retest again."
I still remember a few years ago, we had a new product ready for shipment, but the reliability test report showed that the CPK values for several key parameters were only around 1.08. The PM and my boss looked grim because this implied a potentially low yield, and clients would certainly complain. We all looked at each other, and finally, the boss's gaze fell on me: "Ah-De, what do we do about this? Am I supposed to tell the client that our product yield is 'just decent'?" To be honest, my heart was sinking then, because this situation clearly meant finding the root cause and then retesting, which would cost both money and time!
Where's the problem? It's not that you haven't tested enough; it's that you've tested incorrectly!
Often, when product reliability falls short of expectations, the first reaction is: "Test a few more batches!" or "Extend the test duration!" But in essence, you're merely spending more time and resources to confirm "it will fail," rather than discovering "why it fails" and "how to make it not fail." This is like having a cold and constantly taking pain relievers without seeing a doctor to find the root cause. In a semiconductor factory, we often face dozens or even hundreds of variables—just temperature, humidity, voltage, frequency, material batches... the permutations and combinations of these alone could keep you testing forever. DOE (Design of Experiments) is a method to efficiently identify key factors and their impact on results amidst a multitude of variables. In other words, DOE enables you to find the most effective solutions with the fewest resources.
How is it actually done? Stop relying on trial and error!
Frankly, before using DOE, everyone relied on "experience" or "trial and error." For instance, if a product failed easily at high temperatures, we'd try lowering the temperature slightly or changing the heat sink. But doing so, you only know "A has an effect," not whether the "interaction between A and B" might have a greater impact, or if there are other hidden variables.
How does DOE help you? For example: Suppose we have a new packaging process and want to know how much "oven temperature," "baking time," and "epoxy brand"—these three factors—affect the product's "warpage rate."
- Define the Response Variable: This is the outcome we want to measure, in this case, the "product warpage rate."
- Define Factors and Levels:
* Baking Time: e.g., Short (30min), Medium (45min), Long (60min).
* Epoxy Brand: e.g., Brand A, Brand B.
- Select Experimental Design: Typically, "Full Factorial" or "Fractional Factorial" designs are used here. If there aren't many factors and levels, a full factorial experiment can cover all combinations, allowing you to see all interactions. If there are many factors, a fractional factorial experiment can effectively reduce the number of experimental runs.
- Analyze Results: The software will tell you which factors have the greatest impact on warpage rate and which factors have interactions. For instance, you might discover that under "high temperature" and "long baking time," epoxy from Brand B exhibits particularly severe warpage. This is an "interaction" that would have been difficult to discover using traditional trial-and-error methods.
The key, therefore, is that DOE allows you to precisely pinpoint the root cause of a problem from data, rather than guessing blindly.
The Most Common Pitfall: Flawed Experimental Design Renders Even Perfect Data Useless
The most common pitfall I've seen is incorrectly setting factors and levels. Once, a colleague wanted to study the effect of solder quantity on welding strength. He set three levels: 1mg, 2mg, and 3mg. After the experiment, the data showed that solder quantity had little impact. However, it was later discovered that the difference between 1mg and 3mg was actually very small; the true problematic values should have been extreme ones like 0.5mg and 5mg. His levels were set within a range where the "effect was not obvious," so naturally, he couldn't detect anything significant. Therefore, the selection of factor levels must be based on your understanding of the process or preliminary small-scale trials (Pilot Run). If your experimental design itself is flawed, then no matter how sophisticated the subsequent statistical analysis, it will merely be analyzing meaningless numbers.
One Thing You Can Do Today
Think about the most troublesome reliability issue you're currently facing, and list 2-3 "factors" you believe might influence it.