Inquiry about Experimental Design with JMP

[#1]

Hello,
I have a concern. I am a beginner in experimental design, and I am conducting formulation trials for medicines.

To start, I screened a number of factors and variables to narrow down the factors for my experimental design.

I would now like to optimize. The parameters are as follows:

• 3 responses: Y1, Y2, Y3
• 4 factors, of which 3 are continuous (each with 3 levels), and 1 is categorical with 2 levels.

I am a beginner with JMP and am wondering whether I should use an optimal DOE plan, a classical plan, or a definitive screening design.

Additionally, I am struggling to set the number of levels for the continuous factors, which are at 3 levels.

My question is, in my case, how many experiments would be the minimum required to be representative?
What is the ideal model for my experimental design? How many replicates would you recommend?

I sincerely thank you in advance for your help.

[#2 · Accepted solution]

Hi @PolygonGiraffe5,

Welcome in the Community !

As the name suggest, a Definitive Screening Design is a powerful screening design used at the beginning of a project, with a high number of potential factors, in order to screen active and important factors and optimize response(s), if the number of active factors is sufficiently low. Classical designs like Response Surface Design can be used for optimization, but they won't take advantage of the previous experiments you already have done, and may be less flexible in terms of necessary experiments number. Another last option could be the use of Space-Filling Designs, that are more flexible as they don't rely on a pre-specified model, but can be more costly to run, as the points are generated in order to cover the experimental space homogeneously.

If you have already done a screening phase with the factors identified for the optimization phase, you could try to directly use these preliminary runs from your screening design and Augment your Design on your active factors with various strategies, like model-based augmentation or model-agnostic (space-filling) augmentation.

In case of model-based augmentation, the number of levels of your continuous factors will depend on the complexity of the assumed model : if you specify quadratic (=2nd order) terms for factor A, then 3 levels will be needed for this factor. Selecting this augmentation and specifying a Response Surface model (with main effects, 2-factors interactions and quadratic effects) could be a good first start to create your design.

In case of space-filling augmentation, the number of levels for your continuous factors will depend on the type of space-filling and more importantly on the total number of runs allowed : the more runs in total, the more levels each factor will have.

About the question of "representativeness" and replicates (more generally about design creation options), this is a question you have to answer with domain expertise and statistical tools like Compare Designs platform : What is your objective (explainability, predictivity, both ?) and precision required (maximum and average prediction variance for example) ? What is your maximum experimental budget ?
These questions can help you find the best compromise when evaluating and comparing your designs with the Compare Designs platform.

Hope these answers will help you,

[#3]

Hello @Victor_G ,

Thank you for your detailed and clear response; it will be very helpful for decision-making.

I am still having difficulty modifying the number of levels for continuous factors, as it seems limited to 2 in JMP.

Thank you once again for all the clarifications you’ve provided.

Best regards,

[#4]

Hi @PolygonGiraffe5,

Concerning the number of levels for continuous factors in your design, you can check the solutions from the post force levels in DoE 

In Custom design platform, you can :

• Specify in the model higher order effects to "force" the design generation to introduce additional levels (number of levels depending on the highest orders of the terms introduced in the model),
• Use Discrete Numeric factor type (particularly if you want the different levels to not be equidistant) and specify the appropriate terms in the model (JMP should do it by default with these terms specified as "If Possible).

Hope this will clarify and solve your problem,

[#5]

It's perfect!
I was able to set three levels per factor for the discrete numerical factors. By default,  in JMP model incorporates the term "If Possible" setting in its configuration.

thanks again for your help