A machine learning-enhanced design optimizer for urban cooling

Abstract

Urban cooling becomes a priority in urban planning and design practices. Limited by the slow running speed and prescriptive nature, existing computational tools such as simulation and optimization are yet to be fully integrated in the design decision-making process. This paper describes the Machine Learning-Enhanced Design Optimizer (MLEDO), a novel workflow in search of optimal design option for urban cooling. A physics-based simulation model was developed to assess the cooling performances of a large database of urban design variations. The database was used to train an Artificial Neural Network model, which was then linked with a Genetic Algorithm to rapidly identify optimal design options. The MLEDO workflow was evaluated using a new development urban site against a traditional Simulation-based Genetic Algorithm Design Optimizer (SGADO) as well as human designers. MLEDO outperformed the latter two in terms of efficiency and the performance of the optimal design options. It can also quantify the importance of design parameters in their contribution to cooling performances, which can be used to enhance the understanding of human designers and inform design revisions. MLEDO has the potential to be further developed into a software tool in support of early-stage urban design.