The day the CPK report came out, the room fell silent for three seconds, then exploded
I remember ten years ago, when I was still a PE. A measurement report came out for a batch of wafers from a new process, where a critical dimension had a CPK of only 1.08, calculating to 6210 DPMO. In the meeting room, everyone's faces turned green, and after three seconds of suspended animation, someone started shouting, "What the hell? The parameters were clearly optimized!" Being young then, I couldn't understand it; we had run multiple rounds of Design of Experiments (DOE), identified the optimal values for each parameter, so why did the combined result turn out like this?
Where was the problem? It wasn't that the parameters were wrong; it was that the design "wasn't robust enough"!
Simply put, our previous mindset often focused only on how to bring the average value to the target or minimize variance. But processes are never that ideal, are they? A one-degree fluctuation in ambient temperature, slightly more machine vibration, or even an operator's shaky hand due to a bad mood could instantly derail your perfect parameters. So the key question is, can your design "withstand" these external disturbances? This is the core spirit of "Robust Design."
In other words, Robust Design isn't about eliminating all noise (which is impossible), but about ensuring your product or process performs stably even in a noisy environment. Genichi Taguchi's parameter design is an excellent tool for this, teaching you how to find parameter combinations that are "noise-immune."
How is it actually done? Using "Signal-to-Noise Ratio" to find noise-immune factors
The most brilliant aspect of the Taguchi method is its introduction of the "Signal-to-Noise Ratio (SN Ratio)." This SN ratio is somewhat similar to the communication SN ratio we've heard of before, but the concept is slightly different. It actually measures how strong your process or product's resistance to "noise" is. A higher SN ratio means your design is more stable and less susceptible to external disturbances.
- Define the problem and objective: What problem do you need to solve? What is the objective? For example, stabilizing a certain film thickness at 1000 Å.
- Identify control factors and noise factors: Control factors are parameters you can adjust, such as reaction time, temperature, and pressure. Noise factors are factors you cannot control or are difficult to control, but which affect the outcome, such as ambient temperature fluctuations and raw material batch variations.
- Design the experimental matrix: Taguchi typically uses Orthogonal Arrays to design experiments, using the fewest number of trials to identify the impact of control factors on the SN ratio. At the same time, noise factors should also be incorporated into the experimental design, deliberately amplifying these disturbances to see if your design can withstand them.
- Calculate the Signal-to-Noise Ratio: Calculate the SN ratio for each experimental combination. There are several methods for calculating the SN ratio, such as "Nominal-is-Best" (target value is best), "Smaller-the-Better" (smaller is better), and "Larger-the-Better" (larger is better); choose one that aligns with your objective.
- Analyze SN Ratio and Mean: Identify which control factors have the greatest impact on the SN ratio, then examine their impact on the mean. Our strategy is to first select parameter combinations that maximize the SN ratio to stabilize the process, and then use other parameters that have little impact on the SN ratio but a significant impact on the mean to adjust the mean to the target value.
For example, if you find that reaction time has a significant impact on the SN ratio, and a longer time results in a higher SN ratio, then you should choose a longer reaction time. Then, if you find that pressure has little impact on the SN ratio but a significant impact on the average film thickness, you can use pressure to fine-tune the film thickness to 1000 Å.
The Most Common Pitfall: Focusing Only on the Mean, Not the SN Ratio
I made this mistake before. After running DOE and getting the data, the first thing I did was look for which parameter could bring the film thickness to 1000 Å, and which could minimize variation. As a result, the parameters were set, and small-scale production was fine, but once mass production started, wow, the daily film thickness fluctuated like a rollercoaster, making you doubt everything.
Frankly, this is a classic case of "focusing only on the mean, not stability." We are accustomed to adjusting parameters to the "optimal point," but this "optimal point" might be as narrow as a knife-edge, and even a slight deviation can lead to failure. Taguchi design aims to help you find a "plateau"—a range of parameters that is more lenient, where results remain stable even with minor fluctuations.
So, stop foolishly staring only at the mean and standard deviation; consider whether your process is "robust enough."
One Thing You Can Do Today
Next time you conduct a DOE, try to incorporate "noise factors" into your consideration, thinking about what disturbances might affect your process.