Structural identification and structural health monitoring often require the development of surrogate models. Either for their computational efficiency or the possibility of offline computation of the surrogate model. The latter is beneficial in monitoring where access to the licensed software may be limited. The basis for fitting a surrogate model are the evaluation points, which are sampled over the domain of the selected parameters and evaluated by the parametrised finite element model. When the quantities of interest are the modal properties, additional challenges linked to the mode classification arise. Common approaches to classifying the computed modes into fitting groups include frequency-based ordering, mode-assurance-criterion-based ordering, and mode tracking. However, none of these approaches offers a reliable method for splitting the modes into well-distinguished fitting groups in case of mode degeneration phenomena such as mode crossing and mode veering. The MOSAIC approach to surrogate modelling has been recently proposed to combat the issues connected to mode degeneration. This paper provides a case study evaluating the accuracy of the surrogate model trained according to the MOSAIC approach and comparing it to a benchmark model using artificial neural networks. The surrogate model is trained on the evaluation points obtained from a detailed finite element model of a tall timber building. The results of the comparison show a promising potential for improved surrogate modelling accuracy adopting the MOSAIC approach.