Paper short abstract:

Cities are vulnerable to heat as the climate changes. AI systems can analyse environmental data and feed generative models; designing cities to reduce heat. This paper will present machine designed urban areas intended to reduce heat, comparing their thermal performance and form to human cities.

Paper long abstract:

Urban areas are particularly vulnerable to climate change. However, current urban planning tools and practices have limited capacity to predict the effects of projected larger-scale temperature changes on existing and planned urban areas with local precision. “Heatmapper” is a mapping application which models and forecasts land surface temperature for urban areas, using a deep learning computer vision model which integrates remote sensing, meteorological, and urban-morphological data to forecast local temperature.

Heatmapper allows users to tune parameters to assess the effects of proposed land use change (e.g parks, buildings, canals etc) on local temperatures during heatwaves. Accordingly, it is also possible to build generative models optimised for specific outputs (e.g. lowest average temperature) which suggest changes to the urban landscape.

Remote sensing data can present an attractive opportunity for training models, but is also fraught with issues of generalisation and reproducibility, as well as the surveillance capitalist logic which underlies the creation of these platforms. A generative model which learns from the best quality remote sensing data will reproduce urban design as it exists in the wealthy countries which provide such data – potentially excluding urban forms from elsewhere. As humans and computers work together - this poses questions over algorithmic and human agency, and the embedding of bias in this human-machine collaboration. This paper will present algorithmically designed cities which utilise Heatmapper, and compare these to real cityscapes to better understand the role machines can play as urban designers.