Is there a way to do Morris sensitivity analysis on existing sets of inputs and outputs without using PCE. If yes, can you please show me how I should implement this in uqlab?
Thanks in advance
Unfortunately, no! Morris method is based on a very particular sampling schemes with ad-hoc “trajectories”.
The opposite is possible (using Morris evaluations to build a PCE afterwards), although not optimal since the experimental design is not at all space filling.
In my opinion, Morris method should be left to the “history” section of textbooks on sensitivity, as the proper use of surrogate models in sensitivity analysis (and more specifically sparse polynomial chaos expansions) leads to more accurate results, is more efficient (by 1 to 2 orders of magnitude) and applicable at reasonable cost up to hundreds of input dimensions.
Thank you very much @bsudret for your answer.