What type of design did you use for your experiment? If you used a 2-level factorial design, you can't get a quadratic model from it. You need a minimum of 3-levels for that. 

@ Andrew, what you mean is, say 1, 0 and -1 levels, right.

In that case how many iterations are required, i mean 3 variables and 2 levels, 2**3 = 8 iterations are required right.

Thanks,

Rajesh.

@ Shiv, thanks. I will go through these links.

@Raj

You are correct. If you ran a full factorial (8-runs), you can add axial points in a second block. This will require another 8 samples.

If you are in the design phase, you can use a Box-Behnken design. This requires about 15 samples. 

@ Andrew, thanks for the update.

Ofcourse design expert automatically fit your data and calculate fit for all model.

You have to choose best one

@ Asaimonat, thanks for the update.

First of all, you need to define the experimental designs that you will use. I suggest to use The CCD (Central Composite Design). In the case of your experiment (3 factors and 2 levels) you need: 8 factorial points, 6 axial points ((+/-alfa,0,0);(0,+/-alfa,0);(0,0,+/-alfa)) with alfa=square root of 3, and 2-5 central points (0,0,0). This design is the most commom, but you should verify the experimental region and see that these points (the axial points) are viables.

If the rage was strict (-1,+1), the best region could be the cuboidal, then you could use the Face Center Cube (CCD with alfa=1).

@ Martin, thanks for the info sir.

A Box Behnken design will also serve you well. The CCD Martin suggests will use about 15-20 samples. The Box-Behnken design will use 15-17 samples. It depends if you want to test 5 levels or 3. My preference is to use the BBD. The axial points in the CCD are only replicated once. IF you make a mistake, or those values are off in the proper direction, it can make a quadratic response look linear. 

@ Andrew, thanks for the info. I will check this up.

So, if I have run 2^2 full factorial design (A: pressure, B: temperature) with 3 cental point trials, can't I build a second order model? Using the central points, I can understand if a curvature is present..

Do I necessarily need the CCD, i.e. I have to run the axial point trials?

Thank you for the answer.