Deep Neural Archives: Artificially Augmented Design in Historic Environments Deep Neural Archives: Artificially Augmented Design in Historic Environments
Physical architecture archives have been an integral part of architectural practice since the nineteenth century. In Northern Greece, the historic center of Thessaloniki was destroyed by a fire in 1917. This event resulted in a vast urban regeneration plan and process that called for quick and efficient reconstruction of the city.
The architecture archive left behind documenting the dynamic and rapid development of the city has been explored and analyzed by architecture historians for many decades. Recent breakthroughs in artificial intelligence, and specifically in deep neural networks, have paved the way for exploring this data under a new, artificially augmented light.
One can theorize that archives constitute the predecessors of big data, as they were used by humans to store and safeguard information of importance. The architecture archive of Thessaloniki’s historic center is used in this research as a dataset for generative AI algorithms to explore the latent space of the city’s architectural style and vocabulary.
Urban design policies and legislation introduced in the 1960s enabled the demolition of many historic buildings in favor of greater residential density and higher buildings. Ultimately, the image of the city appears fragmented and incoherent due to consecutive urban transformation plans.
An artificially augmented approach to redesigning parts of the city can bridge the city’s Ottoman, eclectic, and Art Deco heritage with the demands of contemporary urban life. The research explores the city’s mid-war architecture archive as the primary dataset to assess how new technologies and different deep neural networks can reinterpret architectural heritage.
The experiments conducted with the mid-war archive of Thessaloniki can possibly serve as the training dataset for a pre-trained model, which can then be used with different architectural archives as potential datasets.