I am currently working on the UQ of two complex numerical models A and B. I can obtain 500 y_A and y_B with reasonable time. My goal is to find the correlation coefficient between y_A and y_B, i.e., r_AB, and conduct sensitivity analysis on r_AB. I am thinking about using metamodels to generate y_A’ and y_B’ and then find r_AB and Sobol’ indices for SA. Is there any ways to justify the usage of metamodels? I think in order to justify it, I should some how demonstrate that the metamodels will provide more accurate results than just a few hundreds outputs y_A and y_B generated by the numerical model. I am not sure if my thinking is correct. Looking forward to your comments. Thank you very much.