A machine learning based strategy is proposed for creating parametric surrogate models from parametrized finite element model simulation results. In the first major step, a unified nodal data structure is created from the topologically inhomogeneous set of finite element simulations. This is achieved by utilizing re-sampling and the coherent point drift method for node registration of the different designs. In the second major step, a parametric surrogate model is trained for predicting the initial coordinates using a fully-connected feed-forward neural network. Two different recurrent neural network modeling approaches are presented and compared for the prediction of various field quantities with different degrees of complexity and non-linearity. For the first proposed modeling approach, a node-by-node prediction is applied, where the time series of each structural node is predicted independently via a compact long-short term memory (LSTM) model. For each node, the initial coordinates of the node are used as additional input features. For the second modeling approach, an all-at-once prediction is applied, where the time series of all structural nodes are predicted at once by training an LSTM model on a reduced output space obtained by principal component analysis (PCA). Output fields exhibiting moderate non-linearity could be well predicted by both approaches, but only the node-by-node approach allowed an accurate generalized representation of a strongly non-linear and narrow field quantity representing the observed crack patterns within the finite element structure. The existence of a new yet unobserved crack pattern could be identified by the node-by-node approach and confirmed by subsequently running the corresponding FE simulation.