Abstract
The paper aims at addressing a common limitation in dataset availability when designers attempt to use deep learning or machine learning models for generative design in three-dimensions. It proposes an alternative non-parametric and 'small data' machine learning approach that is capable of utilising single (or few) inputs for model training and design generation. The paper extends the Markov random field (MRF) statistical model by first illustrating its naive architectural design appropriation before contrasting it with the proposed algorithmic improvements. Using increasingly complex and real architecture voxel models as input examples for the experiments, the strengths and weaknesses of the proposed method 'substitutional sampling' is discussed alongside the results. The paper thus seeks to provide a means to investigate and better understand the ways in which machine learning models such as MRF could be strategically adopted and effectively harnessed for the generation and exploration of three-dimensional forms in the architecture domain.