T4T LAB 2022.

Patrick Danahy Distinguished Invited Professor.

Coordinated locally by Gabriel Esquivel as well as teaching the Theory Seminar for the lab.

Teaching Assistants: Shane Bugni


Theoretical Premise

“Relations of exteriority, in which assemblage components are self-subsistent and retain autonomy outside of the assemblage in which they exist” Manuel De Landa.

In contemporary surveying, the encounter with industrial or agricultural types comes in the form of unified shells or wholes, as the objects are scanned in their isolated setting in the field or lot. These scanned objects produce a new ontological definition for these industrial assemblages, where components do not exist, and the autonomy of parts is not read. Algorithmic disassembly of parts provides a defamiliarized version of the whole where parts can regain their autonomy within a new assembly logic. Readings of both the continuous and the discrete will be challenged through machine learning models trained on character synthesis and segmentation, assembly, or disassembly.

The Project

 This lab addressesed the latent qualitative character present within objects, and assemblages through machine learning and aesthetic explorations. Opposing historical methods for stylistic classification, generative adversarial networks (GAN’s) parse data networks through data clustering in hyper-space allowing for non-symbolic forms of formal and ornamental classification. In The Sympathy of Things, Lars Spuybroek discusses the connection between movement, growth, figure and structure as a method of developing sympathy in ornamental systems. Students used Agent-based computational methods to construct relational networks that can be used as operators navigating the balance between object and field readings within ornamental compositions. This studio studied relational aesthetics and qualitative effect in ornamental language by combining the two previously mentioned computational methods (GAN’s and Agent-based algorithms). The success of contemporary approaches to machine learning in design relies heavily on the construction of evaluator systems. Students will build unique methods for evaluation using machine learning models, correlating functional and programmatic outputs directly with aesthetic and ornamental effects. The figural and representational ornamental languages of the Art Nouveau, the structurally aligned language of the gothic and the movement and broken edge of the Baroque will all be studied using high-resolution 3D scans built into large datasets for machine learning, alongside more historical approaches to ornamental grammars from Owen Jones, John Ruskin, and others. Typological form was read into the work through deconstructed elements and sectional profiles, which will work within agent-based coding methods to build familiarity into relational formal systems. These agent-based relational networks operate across scales, to deploy ornamental instances, site-design approaches, tectonic assemblages, and structural framing. This type of flattening in the approach across scales seeks to unify the design approach while using the latent space characterization of GAN’s to build unique qualitative effect at multiple scales.

Using a series of algorithmic processes, we can decode the relations of parts and make substitutions to create new assemblies with non-continuous readings of whole, lying between a full autonomy of parts and continuous clarity of the whole. The flattened ontological synthesis provided by many modern machine learning models provides methods for recombination or refiguring wholes, but often struggles with the part-to-part relations. By discretizing the solution space, we can flash between the mesh-whole and its tiled parts, using substitution models of machine learning and perceptual similarity to draw relations between parts and typological, familiar sets. The works of Bernd and Hilla Becher established a new objectivity from collections of industrial and agricultural objects. These stable objective photographic works stand as a base for disassembly and reconfiguration in the early works of this studio, like their 3D counterparts of scanned abandoned agricultural objects in the field to be used in the later portions of this studio. The process of disassembly defamiliarizes the original object toward a flat reading of parts, without hierarchy. Conditional machine learning models trained on continuity, similarity and correlation reconfigure parts into new wholes and the resulting reading lies somewhere between parts and their respective wholes, or figures and their respective elements.

The results of the studio built upon and deepen an understanding of formal and stylistic languages and classifications within architecture, as well as build sympathetic relational effects using machine learning algorithms. Latent qualitative character will be studied and addressed through Jane Bennett’s writings on Vibrant Matter, which have previously discussed with her in connection to the deployment of GAN latent space organization in the pursuit of a non-symbolic qualitative classification system. Embedded within these discussions will be topics of discreteness and continuity and part to whole relations, where algorithmic notions of discrete and continuous influenced an approach to tectonics. This last topic was an underlying discussion throughout the studio as a theoretical understanding of assemblage, objects, parts, and their relationship to deconstructed tectonics and fabrication later in the design process.


The sympathetic output of the studio will focus on the development of unique ornamental languages and will carefully study topics including flora, figural beings, ecology, growth media, and existing stylistic references to develop familiar and sensitive sympathetic relationships between aesthetics and ecological and contextual thinking.


Students were asked to pursue an understanding of collective character and assembly through a perceptual similarity model for comparative object pairing. This will require image and geometric sets to be carefully curated to work with the tiled indexing methods and then will later provide the collections for conditional machine learning model training to produce new readings of wholes.