Deep learning (DL) has been highly successful on numerous tasks in computer vision, natural language processing and audio analysis. These are all examples of structured data defined on an underlying Euclidean grid-like domain. However, in applications concerning 3D data, structured representations are often inefficient and unnatural, so automation of 3D geometry processing and analysis lags severely behind. This poster introduces an early-stage PhD project proposal investigating how DL architectures and their training routines relate to suitable geometric representations. This work lies at the intersection of DL on 3D data, splines as a potentially interesting representation due to several favourable properties, and diverse methods within automated and evolutionary DL as a practical means of achieving efficient, performant and transparent architectures and training routines.