Cities will be exposed to a substantial increase in extreme weather events as global warming continues, thereby affecting the microclimate they create for their citizens. The microclimate in cities is shaped by the built environment, proposed and developed by architects and urban designers, and constrained by the zoning rules imposed by urban planners. These stakeholders go through a design process where the most impactful changes can only be made very early on; however, for those specific stakeholders, there are currently no computational tools to estimate the microclimatic impact of their designs at this crucial stage in early design. This thesis introduces a decoupled approach to simulating outdoor thermal comfort, motivated by global sensitivity analyses. For this, computational fluid dynamics and ray tracing processes are streamlined and validated, which are required to simulate the wind velocity and mean radiant temperature in urban areas. Further, a surrogate model driven by a generative adversarial network is introduced, demonstrating near-instantaneous performance feedback in very early design. In three case studies, the contributions are used in practice to (1) engage in building-scale architectural design, (2) run parametric studies and optimization for urban design, and (3) inform city-scale urban policy. In conclusion, the latest machine learning techniques and the opportunities they provide will revolutionize the way we engage with environmental performance simulation in the urban design process.