Small design with many parameters

[#1]

Hi,

I am reading this paper: Full article: A high-throughput media design approach for high performance mammalian fed-batch cultu...

where authors mention that they constructed DoE with 16 runs using 43 parameters. If I understood correctly, they basically designed 80 random DoEs, choosing the one that best minimizes the correlations between factors. I was wondering if something similar can be done in JMP? Using the Custom design, it is not possible to construct anything smaller than the number of parameters in the model, which is obvious. But is it possible to populate model randomly with only certain main effects and/or 2nd powers, and force JMP to construct smaller designs? Or does anybody else have some other idea how to do something similar in JMP? 

Thanks.

[#2 · Accepted solution]

Yes, JMP can create a super-saturated screening design with Custom Design. See the documentation.

[#3]

Many thanks Mark, the trick was to choose "If Possible" in the Model section.

Still I find it quite interesting that the authors, if I understood correctly, used 80 of such designs (generated randomly), choosing the one with lowest correlations between factors. And their design (from Table 1) looks pretty well populated with low-middle-high values for parameters, while JMP when using super-saturated screening designs tends to only propose low and high values.

[#4 · Accepted solution]

Hi @aatw,

If you need to enforce more or less strongly the presence of middle values for factors, simply use the Discrete Numeric (3 levels) or Categorical (3 levels) factor type. If you use categorical factor type, you can switch the factor type back to numeric continuous after design generation (or use the Convert Labels to Codes utility to switch the nominal values quickly to continuous).

Here are the comparative correlation maps results with the original design, a D-Optimal supersaturated design (with 3-levels discrete numeric factors) and a D-Optimal supersaturated design (with 3-levels categorical factors) :

Attached you'll find the designs compared with the scripts for the correlation maps.

See similar discussions about enforcing a specific number of levels for DoE factors :

DOE with 3 levels for continuous factors 

Inquiry about Experimental Design with JMP 

Number of factor levels in I-Optimal design   

You can also look at Group Orthogonal Supersaturated Designs if you are in an early screening stage.

Hope this complementary answer will help you,

[#5]

Many thanks Victor, very helpful.