Kayıp kentsel dokuyu üretmek: kentsel tasarımda bir araç olarak üretken çekişmeli ağları keşfetmek

This research investigates the potential of Generative Adversarial Networks as a design tool for generating the lost urban fabric. Focusing on the city center of Hatay, Türkiye, which experienced extensive structural loss due to the 2023 Türkiye-Syria earthquakes, the study employs a Pix2PixHD architecture to generate urban patterns. Using pre-disaster maps as the training dataset, the research aims to generate urban patterns that reflect Hatay's architectural identity and spatial continuity. The thesis adopts a hybrid methodology, combining quasi-experimental and case study approaches. It begins with an explanation of the fundamentals of Machine Learning and their relevance to urban studies, followed by detailed documentation of the model training and validation processes. The generated outputs are analyzed using quantitative metrics, including the Fréchet Inception Distance (FID) score to evaluate the model's ability to replicate urban patterns, and the Structural Similarity Index Measure (SSIM) to assess visual realism and structural integrity of the outputs. The study further investigates how different training dataset scales (1:2000, 1:3000, and 1:5000) impact the model's performance. This research demonstrates the potential of GANs as a tool for urban design in a post-disaster context and provides insights into how dataset scale influences the success of the model in generating realistic urban patterns.