What's the most efficient design to screen a large # of factors w/ as few runs as possible at industrial scale and why?

[#1]

For instance. If the purpose of a screening design is to (largely) test for main effects w/ as minimal runs as possible - and I want to screen 7 factors w/ a minimal # of runs, why isn't it a good idea to simply change ea of those factors 1 time on 7 runs?  If that's not the most efficient way, would running a saturated design w/ 8 runs for 7 factors be the most efficient way to test for main effects, or is there some other suggestion & why? Are saturated designs preferred for industrial DOE where you don't have the ability to do many runs & need to be as efficient as possible? 

[#2 · Accepted solution]

Hi @mjz5448,

There is no perfect design for every screening, it's a matter of factors type, number of factors, facility to change, noise and variability of the response, prior knowledge, etc...

If you consider OFAT for a screening, this topic might help you understand how this could be a bad idea : Solved: Re: Main effect screening design vs. OFAT: which is best? - JMP User Community

Supposing you have 7 factors, no mixtures factors and no constraints :

• If you have 7 factors and want to do the minimum number of runs, a screening factorial approach would require 8 runs minimum to estimate the intercept and the 7 main effects. You'll have aliases between your main effects and any higher order terms (factors interactions), so you won't have the possibility to detect interactions between your factors.
• You may also be interested in Group Orthogonal Supersaturated Design, where for 7 factors you may need only 6 runs. Depending on the experimental cost, two experiments less (than a screening factorial design) may be valuable in your context (but would require a careful planning and grouping of the factors). More infos about GO SSD :
  https://community.jmp.com/t5/Discovery-Summit-Tucson-2019/Construction-Properties-and-Analysis-of-Gr...
  https://community.jmp.com/t5/JMP-Blog/Group-Orthogonal-Supersaturated-Designs/ba-p/229877

The choice of the design will be strongly influenced by the experimental budget. If you can afford only <10 runs, you'll very likely end up with (D/A-) optimal screening design. If you can afford more runs, it could be interesting for you to augment the run size of your optimal design, or even try a Definitive Screening Design (for 7 continuous factors, a minimum of 17 runs is required, and 4 extra runs (for a total of 21runs) are recommended). You can try different run size and designs and compare them to choose the one with the best compromise for your topic.

The more information and certainty wanted, the higher experimental cost needed.

More generally, you may find this talk about modern screening designs from Bradley Jones highly interesting, and it may provide you some ideas and perspective for your use case :  

Hope this first answer will help you,

[#3]

Wow, thank you for the detailed response Victor_G!

I've never heard of "Group Orthogonal Supersaturated Design" before, but will look into it. 

I'm largely constrained by the number of runs I can perform, as I don't have the ability to run a lot of trials, and don't have pilot plant or R&D resources easily available. so it seems I'll look into the (D/A-) optimal screening designs as well. 

I'm assuming a lot of the minimal screening designs are dependent on how in control the process is as well in order to detect any effects right? 

Thanks again for the links!

[#4]

You're welcome, I'm glad it helped you !

GO-SSD may be perhaps more relevant with higher number of factors to screen, for example in the presence of several dozens of process factors on a manufacturing/chemical process line. This design type really excels to identify the vital fews among the trivial many.

I'm not sure about the relevance of this design for 7 factors, my safe bet would perhaps be more on (D/A-) Optimal designs, as they enable some flexibility on the run size.

It depends what you imply behind "assuming the process is in control". For responses measurement, your measurement process should be as controlled and stable as possible, so that the main effects detected by the DoE are really active, and not the consequences of unstable measurements (or alternatively, that you don't have too much noise/variability that would "hide" active main effects). Depending on the variability of the measurements and the number of experiments in the DoE, the facility to detect active main effects (power) will be different : lower measurement variability and higher run size improve the power to detect main effects.

Concerning the factors range, don't hesitate to be a little bolder than usual, in order to define ranges where they would be a possible variation due to your factors. If your factors ranges are narrow and your process fully in control, you may end up with no significant factors detected.

I hope this complementary response may help you,

[#5]

Some other thoughts:

1. You might first try directed sampling (components of variation studies) to understand which components have the most leverage and assess measurement systems at the same time.  This allows for assessing consistency/stability of the process while determining leveraged components.  I can't tell you how many experiments I have seen where the experimenter was experimenting on the wrong components.

2. A question...do you want to "pick a winner" in your study or are you trying to understand causal structure? If the latter, you will need to iterate.  The first experiment is intended to help you design a better experiment.  This will likely include screening a large number of factors using bold level settings (typically low resolution).  Once you have determined active/significant factors, then you can work on developing a useful model.

3. Don't neglect noise (e.g., repeats for short-term noise and replicates for long-term noise).  This is often the failure mode that dooms experimentation.  The experiment is run once in a narrow inference space (noise is held constant).  The conclusions from such experiments seldom are useful when the noise inevitably changes in the future.  So efficiency of the design isn't just the design structure/space.

4. In any case, there are a number of considerations for selecting the most effective and efficient designs given your current knowledge, number of factors is just one.  My suggestion is to design multiple experiments, consider what information can be obtained from each.  Weigh this against resource constraints and then choose one.

[#6]

Hi statman, 

Thanks for your response.

For your question 2. I'm trying to screen factors in order to narrow them down for a further full-factorial, and possible response-surface design. 

Thanks for your advice in point number 4. I'll take a look and try to weigh the different options as best I can.