The day the CPK report came out, the room fell silent for three seconds
I still remember several years ago, the coating thickness on our machine A was consistently unstable, with Cpk values struggling between 0.8X and 1.0X. The boss would stare at that curve during daily meetings, his face as dark as the coating. At that time, our engineering team was truly at our wits' end. We finally identified a few potential influencing factors, such as substrate temperature, gas flow, and power. We conducted several rounds of traditional DOE (Design of Experiments) to find the optimal combination, but after each adjustment, the Cpk still swung like a pendulum, sometimes improving slightly, sometimes dropping significantly. At that point, we genuinely thought, "What the heck? How am I supposed to adjust this?"
What was the problem?
Frankly, our traditional DOE at the time was like the "blind men touching an elephant." You only adjust one parameter at a time, or try only a few parameter combinations. This is like searching for an exit in a maze, only able to try one direction at a time, without knowing how these directions influence each other. A slightly higher substrate temperature might require a corresponding adjustment in gas flow to achieve optimal results. Individually, they might seem fine, but together they clash.
Central Composite Design (CCD) is precisely designed to solve this problem. Unlike traditional DOE, which only observes "presence or absence of influence," CCD crucially helps you locate the "optimal operating conditions." It not only tells you the influence of each parameter but also reveals the "interaction" relationships between parameters, and most importantly, where the "optimal point" lies. In other words, it helps you draw an "operating map," allowing you to immediately see where the "holy grail" is.
How is it actually done?
At that time, we applied CCD to the coating problem on machine A. We selected three key factors: substrate temperature (X1), gas flow (X2), and power (X3).
- Define Factor Ranges: First, set the operating range for each factor. For example, we set substrate temperature between 150°C and 200°C; gas flow between 200 sccm and 300 sccm; and power between 1000W and 1500W.
- Design Experimental Points: CCD automatically generates a series of experimental points. These points include:
* Axial Points: Points slightly extending outwards from the midpoint of each factor, aimed at capturing non-linear relationships.
* Center Points: The midpoint of all factors, run multiple times to evaluate experimental repeatability error.
- Run Experiments, Collect Data: Run the machine according to the experimental points arranged by the CCD software, and record the coating thickness, uniformity, and Cpk value for each run.
- Data Analysis: Input the data into statistical software (like Minitab or JMP), which will help you build a mathematical model. This model can tell you how much each factor, and even the interactions between factors, influences the results. Most importantly, it will plot a "response surface," showing you which parameter combination yields the highest Cpk value for your coating thickness.
After CCD analysis, we discovered a strong interaction between substrate temperature and power; high temperature with low power, or low temperature with high power, both caused the coating thickness to deviate from the target. Finally, we found an optimal operating point: substrate temperature 175°C, gas flow 250 sccm, power 1200W. The results showed the coating Cpk soaring to 1.35, and DPMO dropped from 6210 to below 200. All the engineers present genuinely breathed a sigh of relief.
Most common pitfalls
To be frank, when I first started using CCD, my most common mistake was "setting the factor ranges too broadly." Thinking I could test more at once, the resulting data either had extreme points that couldn't be run at all (due to machine limitations) or produced results that were too outlandish, falling outside the range we aimed to optimize.
Another pitfall is "too many factors." The number of experimental points in CCD increases geometrically with the number of factors. If you select five factors, the experimental points will explode in number, you'll run experiments until you cry, and the analysis will become much more complex. Therefore, it's crucial to first use traditional DOE or experience to filter down to 3 to 5 most critical factors. Frankly, you're not conducting an academic paper; you're solving factory problems, and efficiency is paramount.
One thing you can do today
Think about a machine you currently struggle with, and list 2-3 parameters you believe are most critical.