Surrogate model does not fit well to computational model

[ayilmaz]

Dear UqLab Team,

I am trying to build a surrogate model with my experimental data. My computational model, which has 4 input parameters, has resonance-like results (which has an important issue in my computational model) (see figure 1).
I want to learn that how can I cover the resonance-like behavior with the surrogate model also. I mean, how my surrogate model should also exhibit resonance-like behavior?

Thanks in advance

Adem,

[bsudret (Bruno Sudret)]

@ayilmaz

Is your data coming from physical experiments (i.e. possibly having some noise), as opposed to computer simulations which would be free of noise? This could be a reason.

More generally, if your output is a (discretized) function of the frequency for each experiment, the resonances and anti-resonances are usually not aligned any more, which makes the direct computation of a surrogate for each frequency difficult. Techniques exist to solve this issue, see e.g.

• Yaghoubi, V., Marelli, S., Sudret, B. & Abrahamsson, T.,Sparse polynomial chaos expansions of frequency response functions using stochastic frequency transformation, Prob. Eng. Mech., 2017, 48, pp. 39-58 (preprint here).
• Mai, C. V. & Sudret, B., Surrogate models for oscillatory systems using sparse polynomial chaos expansions and stochastic time warping, SIAM/ASA J. Unc. Quant., 2017, 5, pp. 540-571 (preprint here).

Good luck!
Bruno

[ayilmaz]

Dear @bsudret

Thank you for your response.

My data is coming from computer simulations. In the beginning, the computation was done at 60 discrete points, later I increased this number to 1000 points (with the help of the continuous range option offered by the program).

Anyway, I will search for the techniques you mentioned.

Best Regards