AI and Machine Learning are becoming particularly useful for the exploration of the design alternatives and can offer a range of advantages when applied to the exploration of innovative forms of extra-terrestrial infrastructure under uncertain environmental conditions. This paper focuses on building an unsupervised machine learning model (convolutional autoencoder) capable of detecting patterns in-and differentiating between-different arch shapes and contours for extraterrestrial outposts. Foremost, detailed discussions of the model’s architecture and input data are presented. The variation of arch shapes and contours between cluster centroids in the learned latent feature space is determined, opening the door for design optimizations by moving towards clusters with more desirable features. Finally, a regression model is built to investigate the relationship between the input geometric variables and the latent space representation. It is proved that the autoencoder and regression models produce arch shapes with logical structural contours given a set of input geometric variables. The results presented in this paper provide essential tools for the later development of an automated design strategy capable of finding optimal arch shapes for extra-terrestrial habitats.