Computer-aided engineering (CAE) allows to analyze and optimize both products and processes and furthermore provides insight into their reliability. CAE can, for example, facilitate the automatic optimization of engineering designs. Design optimization problems are constituted by design objectives, physical constraints, and geometrical representations. The given geometries together with the provided physics can then form the basis for predictive simulations — whose results can be further exploited to automatically optimize designs with respect to the defined objectives. In the past, many different approaches towards CAE in general and design optimization in particular have been explored. In terms of predictive simulations, apart from application-specific methods, an important field is the generation of reduced-order models in order to accommodate the many-query scenario of automatic design optimization. In the context of geometrical representations, these include vertex-based and spline-based methods. Despite the past successes, the value of these approaches can be further enhanced based on the advent and maturing of machine learning techniques such as deep learning [1]. Deep learning is attracting an increasing amount of attention in all fields of CAE and has, in fact, some interesting capabilities: it can seamlessly combine models and data and allows to identify lower-dimensional parameterizations. Consequently, all three above-mentioned constituents of automatic design tasks can benefit from deep learning. We will investigate the benefits of deep learning in automatic design problems that originate from the field of manufacturing. Possible examples are deep neural networks in the context of CAD-based modeling of geometries or physics-informed predictions of solidification in manufacturing processes [2].