Multiple > 3 component Mixture DOE allowing only ternary mixtures

[#1]

I need to create a mixture DOE with 7 components where only ternary mixtures are allowed. There is no specific model required; goal is screening potential interesting ternary mixtures. With 7 components 35 ternary combinations are possible: is launching a 7-factor Space Filling Design and only keeping the ternary combinations an option? Or, before launching the DOE, using the disallowed 2, 4, 5, 6 and 7 factor combinations? Or starting from the 35 three-component combinations and nesting the mixture factors? Any other suggestion?

Many thanks for input!

[#2]

Great to read that design space profiler works for mixtures! As far as I remember this was not yet the case in JMP 17

[#3]

You're right, in JMP 17 the Design Space Profiler was not designed to handle Mixture factors : Design Space Profiler (JMP 17.1)

[#4]

Important imrovement!

[#5]

In our team we decided to start with a lattice DOE so mixtures as shown below will be obtained for all ternary combinations of the 7 components:

Question: Is the model based on these limited combinations already strong enough so that with the Design Space Profiler reliable control ranges can be specified? Is it not necessary first to augment the lattice DOE to a space filling design or full Scheffé model DOE for achieving a sufficiently powerful model? 

[#6 · Accepted solution]

Hi @frankderuyck,

It's almost impossible to answer your question apriori.

For a model to be "strong enough" it depends on:

• Your objective: The precision you would like to achieve, with a measurable objective like having a RMSE < threshold or MAPE < threshold (%)
• The type of responses you're modeling and its complexity : non-linearity, strong discontinuity, etc...
• The experimental and measurement error/"noisyness" : experimental uncertainty/variance, measurement error/variance, etc...
• The model type you have chosen and its adequacy to the type/complexity of response
• ...

Proceeding sequentially is a good idea in any case, and even if your model is not "good enough", there are a lot of things you can do on your results and model predictions to debug your model and augment your design. Doing some error analysis could help you figure out if some area of your experimental space are systematically incorrectly predicted, or if the errors are "homogeneously" distributed.
This error analysis can then guide your augmentation strategy : augmenting in a Space Filling strategy if the errors are homogeneously distributed and the model is not fitting well on the entire design space, or a more located augmentation if the errors are located in specific area.

Hope this answer may help you,