Multi-objective optimization of urban environmental system design using machine learning

The efficacy of urban mitigation strategies for heat and carbon emissions relies heavily on local urban characteristics. The continuous development and improvement of urban land surface models enable rather accurate assessment of the environmental impact on urban development strategies, whereas physically-based simulations remain computationally costly and time consuming, as a consequence of the increasing complexity of urban system dynamics. Hence it is imperative to develop fast, efficient, and economic operational toolkits for urban planners to foster the design, implementation, and evaluation of urban mitigation strategies, while retaining the accuracy and robustness of physical models. In this study, we adopt a machine learning (ML) algorithm, viz. Gaussian Process Regression, to emulate the physics of heat and biogenic carbon exchange in the built environment. The ML surrogate is trained and validated on the simulation results generated by a state-of-the-art single-layer urban canopy model over a wide range of urban characteristics, showing high accuracy in capturing heat and carbon dynamics. Using the validated surrogate model, we then conduct multi-objective optimization using the genetic algorithm to optimize urban design scenarios for desirable urban mitigation effects. While the use of urban greenery is found effective in mitigating both urban heat and carbon emissions, there is manifest trade-offs among ameliorating diverse urban environmental indicators.