Soil-structure interaction is a key aspect when calculating the dynamic response of nuclear structures in front of ground motions induced by earthquakes. However, modelling the seismic response in the time-domain using FEM-based models to represent the soil volume is computationally expensive because of the large number of elements involved. For this reason, it is a common practice to employ 1-D models, which allows an explicit solution of the dynamic response of the system in a fraction of time. Yet, due to the importance of nuclear structures, sometimes it is mandatory to employ 3D FEM-based models. In addition, based on probabilistic estimations of the seismic hazard at a site, it is generally necessary that the seismic response of the model comply with the spectral ordinates of the input action at the surface level. In doing so, specific accelerograms acting at the bedrock should be developed. If these signals are obtained by deconvolution using the 1D model, and then applied to the 3D one, the response calculated at the surface level will always be somewhat different from the target free field motion (the one considered in the deconvolution process at the surface level). There are several sources, like the Poisson effect in the 3D model, explaining this difference. Therefore, it is necessary to modify the input signal so that the dynamic response obtained with the FEM model fit the target spectrum. This article presents an innovative approach to achieve this goal, developing an enhanced transfer function stemming from the 1D and 3D models. Two seismic scenarios and a specific soil profile have been used as benchmark. Results show that this approach provides an adequate solution for the studied problem.