Geometric deep learning is a young and promising field of research that offers innovative solutions for the automated understanding and characterization of the geometry, shape and topology of individual parts and assemblies. The automated understanding of these aspects is crucial for downstream tasks such as classification, segmentation and the recognition of design intent. This master's thesis will explore the exciting challenge of encoding geometry information using machine learning techniques from current literature in an autoencoding task and identify the advantages and disadvantages of different methods for single part geometries. Subsequently, a modification shall be found to make these methods rotation-equivariant.