Latin Hypercube Design

[#1]

Hi, I'm planing a DOE for numerical simulations (CFD simulations).

I'm thinking on using Latin Hypercube Design, for 2 responses and 3 continuous factor.

How should I determine the minimum number of runs? I notice that in the software I can chose  the number of runs and it has an impact on the design diagnostics.

There is an recommendation to determine the number of runs? My simulations are quite computational expensive, so I need to use the minimum number of runs possible to keep the quality of the results.

[#2]

Welcome to the community.  You might want to try Jmp's Custom Design Platform (particularly if you have a model in mind):

https://www.jmp.com/support/help/en/16.2/?os=mac&source=application#page/jmp/custom-designs.shtml

I actually don't understand "computationally expensive".  Do you mean the simulation takes a significant amount of computer time to run?

[#3]

Yes, by computationally expensive I meant that the simulations take quite a long time to run, so it is important to keep the number of runs to a minimum.

[#4]

You will likely use a general interpolating function like that available in the Gaussian Process platform for JMP. You might start with a small space filling design and see how it fits. If the response changes (relatively) slowly, this data and resulting model might suffice for your prediction purposes. If not, you can use DOE > Augment Design to add more runs to improve the model performance.

[#5]

Using the augment design, can I use the results of the previous simulations? Or all the runs change?

[#6]

Augmentation adds runs to the existing design, but in a new data table. Nothing lost! Try it!