nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

Next steps after DSD?

mjz5448

I completed a 6 factor DSD recently, & found only 4 factors to be significant.

Below is what the prediction profiler looks like after reducing my model, showing a quadratic effect for 1 of the factors.

My model has 4 main effects, 2 two-factor interactions, & 1 quadratic effect.

My question is, what are the next steps to a DSD? Should I go straight to an RSM design? So fit 2^4 factorial w/ center points, and go from there? Or is there something else to be done w/ a DSD?

DSD prediction profiler.png

Accepted answers

Byron_JMP

I would do a validation run to make sure the model predicts reality.

If it looks good, then I might think about an RSM with the three active factors with ranges outside of the DSD range to see where the edge of failure is. (..or see where yield tapers off.)

It looks like a small increase in X1 would have a pretty dramatic effect on yield?

All comments

Byron_JMP

I would do a validation run to make sure the model predicts reality.

If it looks good, then I might think about an RSM with the three active factors with ranges outside of the DSD range to see where the edge of failure is. (..or see where yield tapers off.)

It looks like a small increase in X1 would have a pretty dramatic effect on yield?

Ressel

What about a resolution IV fractional factorial with two center points for replication? It seems to me you already have a decent understanding of the model effects?

mjz5448

I thought fractional DOEs were mostly for screening? Are you saying do a 2nd screening design, or can I measure lack of fit w/ those center points w/ a fractional design w/ fewer runs?

mjz5448

Thanks for the suggestion. I'll plan on validating the optimal conditions that the prediction profiler suggests & go from there. I was wondering, since I have curvature in 1 factor, does that mean I'm close to the optimum & should incorporate axial pts in my next design, or just start w/ the 2^k DOE w/ center points?

Ressel

Hm, not aware that they were mostly for screening but I had guessed from the number of effects you think you know that a fractional factorial design, perhaps with different factor settings would give sufficient information to build the next model iteration. I suspect this would mean fewer runs than a RSM, which means it saves time and it would still give you information on lack of fit. The center points appear necessary to capture the nonlinearity shown in your screenshot, and reproduction also seems beneficial. On the other hand, if you can perform experiments fast at low budget, than you are not restricted by the nr of runs. Difficult to tell without knowing more of the context.

Note, that the DoE platform in JMP allows to comparison between different designs, which may make the decision which step to take next easier.

Edit: Lack of fit in JMP help.

P_Bartell

I don't disagree with anything my former colleague @Byron_JMP or @Ressel have recommended. But I'll take a slightly different approach. I'm also going to assume you've got some practical problem you are trying to solve? With everything you've done to date, ask this question: "Can I solve the problem at hand?" If the answer is 'yes', then why execute follow up designed experiments? If the answer is 'no' then move on to additional investigative work.

Mark_Bailey

If you decide to pursue further experimentation, there is no need to start with a new design, Open your DSD data table and select DOE > Augment Design. This tool allows you to determine the optimal set of new runs to add to the existing design,

mjz5448

Thanks Mark. I wasn't aware of that option. I'll definitely look into & make use of it!