@m-balesdent I played with your kriging script with compared Cobyla from NLopt with our Cobyla
and found that one you freeze the rng seed (using LHS from OT instead) there is not much difference (contrary to your first version of the script):
solver=TNC rmse=0.8885311127870625
solver=Cobyla rmse=0.0022978724463603986
solver=NLopt.LN_COBYLA rmse=0.00311640242441848
solver=NLopt.LN_NELDERMEAD rmse=0.0014765812010112487
So the conclusion here is that the scaling that NLopt may be using does not affect the kriging on this example.