The development of automotive structures is strongly related to the design for crashworthiness; corresponding evaluations are done by explicit FEM leading to a high computational effort. Today, optimizations and uncertainty quantifications (reliability, robustness) complement this, increasing further the numerical effort. This motivates the search for machine learning approaches such that (i) the number of structural concepts and (ii) the effort for the simulations per concept are reduced. The latter objective, the reduction of computational effort, is addressed via non-linear model order reduction techniques (MOR) based on so-called snapshot matrices. The former objective, the reduction of the number of different concepts, can be fulfilled by learning from prior vehicle concepts. First results using (a) a transfer learning approach [2] and (b) a vector cloud representation of prior conceptual improvements [6] will be discussed.