Interlacing Latent Features: Synthesis of Past and Present in Architectural Design through Artificial Intelligence in a Case Study of Japanese Houses

Abstract

Machine Learning (ML) algorithms have shown great promise for expanding the conventional limits of human perception, thereby augmenting the architect's imagination and design agency. This thesis extrapolates global implications of Artificial Intelligence (AI) in architecture that challenge the trends of globalization and standardization. Through case studies, an ML-enhanced approach is demonstrated, integrating contemporary Japanese houses with elements of historical context and cultural heritage. Initially, datasets of their plan and façade images are scraped from the internet, curated, and annotated with Japan's six latest historical periods. These datasets are employed to train an image classification model. This model quantitatively predicts the likelihood of the dataset houses belonging to each historical period. Subsequently, these datasets are utilized to fine-tune Stable Diffusion with Low-Rank Adaptation (LoRA), synthesizing past and present styles in response to specific period prompts. Images generated by the fine-tuned model, which offer design suggestions, are dissected into layers representing different architectural elements. These elements, interpreted by me, are restructured into a three-dimensional model to construct novel residential typologies learned from both historical and contemporary styles. This case-study intervention suggests the potential of AI application in architectural design to promote cultural diversity, sustainability, and the continuity and enrichment of design heritage.