Machine Learning techniques open important ways of development of physical models in almost every field. Performances reached rely on two major pillars: the (physical) model and the data. When a huge amount of data is available, the intrinsic characteristics of the chosen algorithm become less critical. On the other hand, with limited amount of data, all the human knowledge on the system to be modelled becomes critical to exploit.

One of the multiple applications of Machine Learning (ML) technics concerns the meta-models. Indeed, in this paper, we show how we can bypass a computation scheme by using clever regression models. The approach is performed on a system from which we want to know instantaneously the first natural frequencies without performing each time finite elements (FE) computation. We study the performance reached by studying the number of training cases required to teach the algorithm how to link inputs to outputs within a satisfying accuracy. Different algorithms are tested with very encouraging results as going into higher dimensional problem. The final aim of this study is to provide global guidelines for the most efficient Machine Teaching.