Materials whose unconventional properties are determined primarily by their internal organization —so-called metamaterials— underwent significant progress over the past decades. In particular, it has been shown that the Poisson's ratio in lattice-based metamaterials can be pushed to negative values for specific geometries. Simultaneously, by harnessing the mechanical instability of curved beams within the architecture, energy storing and energy harvesting become feasible. Since both of these manifestations of unusual behavior rely on the geometries of the components defining the metamaterials, they can be easily tuned through the rational selection of beam shapes. While predicting the mechanical behavior of beams and metamaterials is straightforward with numerical methods, finding designs with specific behaviors is more challenging. To address this inverse problem, we present a machine learning method based on Generative Adversarial Networks (GANs). To describe the beams, we employ Bézier curves that enable defining complex geometries while maintaining an efficient representation with a low number of parameters. In particular, we demonstrate that this representation, combined with GAN-based machine learning, allows for designing curved beams that match certain characteristics like snapping force and energy. Furthermore, through experimental measurements, we show that our model is capable of making reliable predictions in real-world settings, despite being trained only on simulation data.