A Two-Stage Hierarchical Optimization framework for Efficient Generative Urban Design

Performance-driven design (PDD) is frequently constrained by the high computational cost associated with optimizing a large number of design variables against multiple objectives with varying computational costs. This paper presents a novel two-stage hybrid optimization framework to address this challenge. The framework first utilizes a Genetic Algorithm (GA) to efficiently conduct a global search, identifying a candidate pool of designs that adhere to fundamental constraints such as zoning and economic viability. Subsequently, Bayesian Optimization (BO) is deployed to perform a targeted local search, refining these candidates by optimizing computationally intensive metrics, such as outdoor thermal comfort. A case study involving a residential block design demonstrates the effectiveness of the method. Our framework improved Outdoor Thermal Comfort Autonomy (OTCA) to 72. 5\% from a baseline of 63. 8\%, while simultaneously reducing the required computational time by 83.8\% relative to a benchmark-accelerated GA. This work provides a scalable and efficient paradigm for complex architecture optimization, enabling a more effective balance between economic, regulatory, and environmental performance in the pursuit of sustainable urbanism.