Very well explained by Dr Debopam Ghosh

I think any of the two designs you mention would be appropriate for your problem, in box behnken design mapping uncoded factors to coded factors is easier, while if you are interested in rotatable designs CCD is the better option.

Regardless which you use you will end up with an approximation, a quadratic response surface approximation to the actual response surface function over the experimental space under consideration.

This approximate response surface you have to analyse to determine its nature and its stationary points, which may be optima or saddle points.

For optimization purpose, you have to go with iterative evolution on the experimental space till you narrow down to region of sufficient curvature (significant lack of fit test).

To move into the region of curvature you can go with steepest ascent coupled with 2 level factorial designs for planar surface /distorted planar surface modeling of the results, and then going with these 2nd order designs to probe the region of experimental space showing significant curvature.

Sequential simplex optimization is another approach for the evolutionary search part of RSM, but it cannot give a pinpointed optimal point, so in region of sufficient yield of response variable ( say 80 % or more) it is better to sbandon it and go with steepest ascent approach mentioned above.

You can look into Design Expert Instruction manual, online materials available at stat ease website or contact and consult with their team of experts regarding any queries you may have about running the experiment using the software.

Hope this helps

D.Ghosh

I agree with Dr Debopam Ghosh

Very well explained by Dr Debopam Ghosh

Thanks for valueable advice