In this work, a generative adversarial network (GAN) was trained on images generated from mathematical equations which define surfaces that exist in high-dimensional ambient spaces. Thousands of these surfaces, known as Fermat hypersurfaces, were produced and then projected into two-dimensional space in order to construct a dataset upon which the GAN was trained. Shell is an audiovisual projective piece that is concerned with investigating the latent space generated by these images of projected Fermat hypersurfaces. It serves not only as an exploration into abstract mathematical spaces from which we can begin to gain an intuitive understanding of through visualization, but is also concerned with the form produced by these manifolds and how said form relates to our concepts of beauty and harmony. Audio is generated by sonifying the structural relations between constituent parts of the images. The work is self-sufficient; the latent space naturally unfolds without user input. Oculus Touch controllers are provided which allow for some limited control over the speed of movement through and location in the latent space with which users can interact.