Latin Hyper Cube

All About UQ UQ Discussion Forum
1

Hi
I have just question about is any criteria or formulla to identify the number of design points using in Latin Hyper Cube please
Regards
Hazim

2

Dear Hazim,

There is no general rule. How many design points you need depends on your problem.

Best regards,
Styfen

3

Dear Styfen

Many thanks for your answer , so if I have 4 design variables how many DOE points I need please .

Regards

Hazim

4

Dear Hazim,

The number of design variables is only one out of many factors that determines how many point you will need. Further, it is not possible to make suggestion without knowing what you are intending to do with these points. Are you using them to evaluate your true model to then build a surrogate? Do you evaluate an existing surrogate model on them? Are you performing a MC study? How expensive is it to evaluate these point? How complex is the response of your model? Is it even non-smooth? You probably see that it is simply not possible to give you a single number that works for every problem. But as a general advise I would suggest to check convergence. If you are building a surrogate you might use e.g. cross-validation. If you are evaluating a surrogate you can check the mean and variance of your solution.

Best regards,
Styfen

5

Hi
Many thanks for your help; I want to generate a surrogate model based on CFD simulations results, also I want to use MC simulation to propagate the uncertainty so first if you can help which code I should use in UQ examples to deal with external results and I did not find MC in the examples.
Regards
Hazim

6

Dear Hazim,

If you already have your Results you can read them into Matlab and pass via a user-defined input to build your surrogate model as shown in this example:

If you don’t have your results yet, UQLink provides you with a wrapper to directly combine your CFD software with UQLab:

The MC study would be performed with the surrogate. Therefore, you can define and then generate your inputs. See the following example:

These inputs are then evaluated on your surrogate model.
This step is shown in the “pce-truss-data-set” example I referred to above. Of course this works similarly for the other surrogate model provided by UQLab.

Best regards,
Styfen