Decrypting the Black Box: a designer’s journey into the infinite possibilities of ML in design

One of the most significant and revolutionary methods in modern computer science is machine learning (ML) [1]. It entails the unsupervised statistical inference of models from data. In design applications, these models hold great promise for creating and synthesizing new and sometimes innovative design alternatives. It enables creative practitioners—designers, artists, and architects—to produce novel results based on sample data. Machine learning is a powerful tool for solving complex problems, but it can also be difficult to understand and use effectively. 

This thesis is my personal journey into the infinite possibilities of ML in design, more concretely into the possibilities of generative systems of volumetric forms. During this thesis, I have been exploring and documenting the challenges and opportunities of ML as a design tool, trying to make it more comprehensible and accessible for other designers. Through a series of three design-research experiments, I have investigated the potential of ML, more concretely, the use of the latent space — lower-dimensional vector space that captures the underlying structure or patterns in the data that the model is trying to generate— learned by ML models for design exploration, reflecting on not only my personal but also empirical takeaways from each experiment and providing guidance for readers who wish to repeat the experiments or try new ones inspired by my findings. 

In experiment 1, I compare the parametric space with the feature space extracted from an ML model and show that the later can be used as tool for design exploration. Experiment 2 is a more exploratory project inspired by the mismatches between the way computers and humans understand everyday objects. With the training of an ML model in a chair dataset and the subsequent fabrication of one of the outputs, I explore the fundamental shift in design that the use of computational tools will have in the future, as well as the idea of authorship. Experiment 3 is an inquiry into alternative approaches to interacting with generative systems based on the interpretation and visualization of the latent space. 

Overall, this thesis offers a personal perspective on the challenges and opportunities of using generative systems as a design tool and presents the latent space as an alternative way to interpret and analyze volumetric shapes. This thesis serves as a starting point for another creative practitioner that wants to embark on the journey of using machine learning in design. My experiments can serve as guidance for those interested in exploring the potential of latent space as a design exploration space, or for those who want to explore new ways of using machine learning to create richer forms. Through this thesis, I demonstrate that machine learning can be a powerful tool for improving creativity in design and offer insights into how designers can harness this technology to generate innovative and exciting outcomes.