That Day the CPK Report Came Out, the Entire Room Was Silent for Three Seconds
I still remember several years ago, when we switched to a new material on our machine, wanting to see if it affected the product's electrical performance. Everyone excitedly designed a DOE, running several batches with new and old materials interchangeably, and then waited for the results. When the CPK report came out, the old material's CPK unexpectedly dropped from the original 1.35 to 1.08, and DPMO soared directly to 6210! The entire room went silent for three seconds, and I thought, "It's over, I'm going to be nailed to the wall." My boss's face turned green, questioning what went wrong. After checking for a long time, the equipment parameters hadn't changed, and the operating SOP was also correct. Finally, we discovered that the problem actually stemmed from the "experiment order."
Where Did the Problem Lie? Has Your Experiment Been "Hijacked"?
Simply put, you might think you're just measuring material differences, but your experiment results could actually be hijacked by other "hidden variables." The most common of these is the "time" variable. Think about it: if you run all old material batches in the morning and all new material batches in the afternoon, can you truly say the difference is due to the material? Could it actually be machine temperature rise after prolonged operation, variations in operating methods due to shift changes, or even environmental changes between batches? In such cases, your experiment results become a mix of material effects and time effects, making it impossible to differentiate them.
Therefore, the key is to always consider Randomization and Blocking when conducting DOE experiments. Randomization means scrambling your experiment execution order to prevent certain factors from being fixed in specific time slots or under specific conditions. Blocking is when you know certain variables are unavoidable but you don't want them to influence your main factor analysis, so you isolate them for consideration.
How to Do It in Practice? Order in Chaos is the Way to Go
The simplest method is to use random numbers to arrange the experiment order. Suppose you have two materials, A and B, running three batches each, totaling six batches. If you run them as A1, A2, A3, B1, B2, B3, you're very likely to fall into the time pitfall. But if you use a random number generator to arrange an order like B2, A1, B3, A3, B1, A2, then the effect of time on the material results will be "averaged out," making it less likely for a unidirectional bias to occur.
To give a more practical example, when we validate new process parameters, we often encounter situations where there's only one machine. If your experiment design is to run all low-temperature combinations first, then all high-temperature combinations, then effects like machine temperature rise and consumable wear during the process might be mistakenly attributed to temperature differences. In such cases, you can consider blocking the "temperature" factor. That is, under each temperature condition, randomize the order of other parameters. Alternatively, divide the day into two blocks, morning and afternoon, and randomize the experiment sequence within each, which helps separate daily variations from the intrinsic effects of the material itself.
The Most Common Pitfall: Taking Shortcuts Can Indeed Lead to Trouble
The most ridiculous situation I've encountered was, to save time or manpower, allowing the same engineer to run experiments scheduled for the entire day from start to finish, then having another engineer take over for the latter half of the next shift. As a result, when the data came out, the results from the first half and the second half were like two different stories. Upon investigation, it was discovered that the two engineers had subtle differences in their operating habits, and the machine's performance also started to drift after running continuously for over a dozen hours. At this point, if you haven't randomized and you mix the effects of these two "blocks" together, you'll find that no matter how you adjust the parameters, product yield will be like a rollercoaster ride, making it utterly impossible to understand which parameter is truly influential. Frankly, seeking temporary convenience often ends up costing ten times the time to clean up the mess.
One Thing You Can Do Today
Before your next DOE, take three minutes to randomize the experiment order.