Color-Patterns to Architecture Representation
Often an apparent complex reality can be extrapolated into certain patterns that in turn are evidenced in natural behaviors (whether biological, chemical or physical). The Architecture Design field has manifested these patterns as a conscious (inspired designs) or unconscious manner (emerging organizations). If such patterns exist and can be recognized, can we therefore use them as genotypic DNA? Can we be capable of generating a phenotypic architecture that is manifestly more complex than the original pattern? Recent developments in the field of Evo-Devo around gene regulators patterns or the explosive development of Machine Learning tools could be combined to set the basis for developing new, disruptive workflows for both design and analysis.
This study will test the feasibility of using conditional Generative Adversarial Networks (cGANs) as a tool for coding architecture into color pattern-based images and translating them into 2D architectural representations.
This patterns are understood as a way to store and establish relations in data that can be scale aware while having fractal properties or adaptive while following a very limited set of rules. This images can be the result of extremely complex processes, outcomes that can be understood as phenotypes, where astonishing geometry emerges through the combination of simple rules. We will use a database with thousands of patterns and asociate isometric volumes to train a NN under the pix2pix toolset, based on Generative Adversarial Networks (GANs).
The conditional in GAN models derives from the need to force the training in one direction to avoid a certain lack of control on modes of the generated data. This can be achieved by feeding data that needs to be a condition on both the generator and discriminator. In the case of this experiment, it is by using paired images for evaluating the results.
The use of ML algorithms and NN has allowed adding more complexity and case studies to the topic of computer-generated architecture. Relevant recent works in this category worth mentioning are explorations from Stanislas Chaillou , using the proven creativity abilities of GaNs to generate different realities of the drafting process, all applied to floor plan designs. Within the same line of work, Huang & Zheng  propose codification for architectural elements (Figure 1), allowing a smoother learning experience for the NN, based on previous works done by Zheng et al. (2017) for generating urban and city-scale planning. Other examples from Mohammad et al.  on applying GaNs to design façades explore the relationship and symbiosis between architects and AI for making design decisions, which have proven to be successful.
After several tests defined by diferent kind of patterns, the main conclusion is that the initial hypothesis can be easily achieved by the cGAN, opening a world of possibilities around complex pattern encoding and architecture representation at many levels. The majority of tests support this claim and have shown the importance of data-base sizes and, especially the amount of echoes run. Despite the differences between pattern and isometric, every test improved as new resources where added.
The authors believe that the capacity of cGANs to generate content extremely quickly in a tremendously simple process encourages their integration into the workplace. Websites or apps can run the trained GANs in the face of the difficulty of installing and learning certain software. Calculating without calculating could mean a change of the paradigm in terms of simulation in fields such as structures, sunshine, energy efficiency and so forth.
Author Contributions: Conceptualization, D.N.; software, D.N. and P.C.; validation, O.C. and P.C.; investigation, O.C., P.C. and D.N.; resources, D.N. and P.C.; data curation, X.X.; writing—original draft preparation, D.N.; writing—review and editing, D.N. and O.C.; visualization, D.N. and P.C.;