Decisions in engineering design are closely tied to the 3D shape of the product. Limited availability of 3D shape data and expensive annotation present key challenges for using artificial intelligence in product design and development. In this work, we explore transfer learning strategies to improve the data-efficiency of geometric reasoning models based on deep neural networks as used for tasks such as shape retrieval and design synthesis. We address the utilization of problem-related and un-annotated 3D data to compensate for small data volumes. Our experiments show promising results for knowledge transfer on mechanical component benchmarks.