Speculative Hybrids: Investigating the generation of conceptual architectural forms through the use of Machine Learning.
Translated title
Speculative Hybrids
Author
Pouliou, Panagiota
Term
4. term
Education
Publication year
2022
Submitted on
2022-06-03
Pages
88
Abstract
Design thinking retter sig mod komplekse problemer uden en entydig definition, uden et klart slutpunkt for arbejdet, og hvor resultatet afhænger af, hvordan problemet formuleres. De fleste designprocesser er i dag præget af digitalisering. For at undersøge nutidigt arkitektonisk design analyserer vi, hvordan digitale værktøjer kan styrke processen, og vi foreslår en metode til at udforske problemrummet. I projektet Speculative Hybrids bruger vi maskinlæring til at generere konceptuelle arkitektoniske former, der tager højde for stedsspecifikke regler (fx lokale bestemmelser for højde og afstand). Mere specifikt anvender vi generative adversarial networks (GAN'er), en type AI-model hvor to netværk konkurrerer for at lære at skabe nye eksempler, til at producere annoterede punktskyer—3D-samlinger af punkter med tilhørende etiketter. Værktøjets iterative karakter understøtter designprocessen ved hurtigt at foreslå et spektrum af plausible muligheder og hjælpe med at strukturere rumlige problemstillinger, hvilket kan gøre problemløsningen mere effektiv.
Design thinking addresses complex problems with no single, clear definition, no obvious endpoint, and outcomes that vary with how the problem is framed. Most design processes today are shaped by digital tools. To explore contemporary architectural design, we analyze how such tools can enhance the process and propose a method for exploring the problem space. In the project Speculative Hybrids, we use machine learning to generate conceptual architectural forms that respect site-specific regulations (for example, local limits on height or setbacks). Specifically, we implement Generative Adversarial Networks (GANs)—a class of AI models in which two networks compete to learn to create new examples—to produce annotated point clouds, that is, 3D collections of points with labels describing their properties. The iterative nature of this tool supports the design process by rapidly proposing a range of plausible options and helping to structure spatial problems, leading to more efficient problem solving.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Documents