1: \begin{abstract}

2:   In optical nano metrology numerical models are used widely for parameter

3:   reconstructions. Using the Bayesian target vector optimization method we fit a

4:   finite element numerical model to a Grazing Incidence X-Ray fluorescence data

5:   set in order to obtain the geometrical parameters of a nano structured line

6:   grating. Gaussian process, stochastic machine learning surrogate models, were

7:   trained during the reconstruction and afterwards sampled with a Markov chain

8:   Monte Carlo sampler to determine the distribution of the reconstructed model

9:   parameters. The numerical discretization parameters of the used finite element

10:   model impact the numerical discretization error of the forward model. We

11:   investigated the impact of the polynomial order of the finite element ansatz

12:   functions on the reconstructed parameters as well as on the model parameter

13:   distributions. We showed that such a convergence study allows to determine

14:   numerical parameters which allows for efficient and accurate reconstruction

15:   results.

16: \end{abstract}

17: