The ongoing increase in computational capabilities has been successfully utilized for the engineering task of determining an optimal structural design, allowing for high-fidelity simulations of possible design parameter combinations. However, the more detailed these simulation models are, the more expensive a single performance prediction becomes, making the determination of the optimal design a very time-consuming task. In addition, with uncertainty modeling and propagation as an essential part of the robustness analysis of a design, the computational effort is yet again increased. By replacing the expensive simulation with a fast, capable metamodel, the structural design process can be significantly expedited. The prediction accuracy and, therefore, the usability of such a metamodel is strongly dependent on a high-quality set of data samples representing the relationship of input and output over the entire design space. Acquiring these samples is also an expensive process, therefore, it is important to select each sample in such a way as to maximize the information gain. Given the mostly limited or non-existing prior knowledge about what regions in the design space are of interest, i.e. show a highly nonlinear response, adaptive sampling strategies have been developed, which iteratively select new sample points automatically based on the existing data without further intervention by the user. In this contribution, we present a novel adaptive sampling strategy, which is constituted of a combination of an exploration and an exploitation criterion. The predictions’ variances of an ensemble of Artificial Neural Networks is used to detect regions of interest, whereas a Kernel Density Estimation of the existing samples allows for a detection and filling of those parts of the design space which exhibit a low sample density. The capability of the approach is shown for different test functions as well as an engineering example.