Title: Enhancing Climate-Responsive Urban Planning Through U-Net-Driven Quality-Diversity Optimization

Abstract:

Designing urban environments that adapt to climate conditions necessitates a delicate equilibrium between construction density and the flow of cold air. However, the high computational cost of physics-based climate simulations typically restricts planners to evaluating fewer than ten manually created designs. While Quality-Diversity (QD) algorithms can systematically map the design space, their practical application relies heavily on the use of surrogate models. This study introduces a novel approach by substituting a sluggish, regulation-based physics simulator with a spatial deep-learning surrogate based on U-Net architecture, integrated within an offline MAP-Elites framework. We conduct a systematic comparison between this spatial methodology and a conventional Gaussian Process (GP) surrogate, analyzing their performance under varying training data strategies, specifically quasi-random Sobol sampling versus active QD bootstrapping.

The findings indicate that scalar GP surrogates suffer significant failures when trained on random data, necessitating costly, actively generated QD archives to achieve generalization. Conversely, the spatial inductive bias inherent in the U-Net enables it to robustly learn the underlying physics mapping with an $R^2$ value of 0.996, regardless of the training data origin. This capability permits offline QD optimization to secure highly accurate fitness rankings ($\rho = 0.994$) using merely a single batch of random training samples. Implemented within the open-source OpenSKIZZE tool, this pipeline is capable of generating thousands of diverse, climate-evaluated building layouts in less than ten minutes.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC