Harvard Graduate School of Design Thesis 2022
Throughout architectural history, different tools of design have affected the culture of architectural production. While drawings and visual imagery often act as a primary form of contemporary representation, architecture cannot be reduced to a single mode. The cyclical tension between the conceptual and material relies on a multimodal process originating from semantics. Whether built form or text, both can be seen as a form of architecture that rely on a necessary conceptual dimension. This thesis questions what opportunities arise within the cyclical translations and differences between these modes of representation, in particular language, through the use of machine-learning algorithms.
Emergent multimodal neural networks are capable of learning visual concepts from natural language supervision. They can be instructed in common language to generate images, which represents a historical moment of convergence between image and text processing. These models will bring forward a fundamental shift in the way language and its articulation can be used within the creative process. By exploring the potentials and limitations of these machine-learning models, from text to image to 3D geometries, the thesis seeks to uncover new relationships between language and architecture. Ultimately, these new collaborative human-machine interactions augment, rather than limit the agency and creativity of the architect within this process.
Thesis Advisor : Andrew Witt, Jose Luis Garcia del Castillo Lopez