U-Net speeds climate-adaptive urban layout design 1000x

Optimizing urban layouts for climate adaptation is a computationally expensive task: physics-based simulations are so slow that planners typically evaluate fewer than ten manual designs. A team led by Alexander Hagg et al. from the University of Applied Sciences Bonn-Rhein-Sieg, Germany, proposes replacing the slow simulator with a spatial deep-learning surrogate (U-Net) inside an offline MAP-Elites algorithm. The key innovation is the U-Net's spatial inductive bias, which allows it to learn the underlying physics mapping directly from random training samples (quasi-random Sobol sampling), achieving an R² of 0.996. In contrast, traditional Gaussian Process (GP) surrogates fail catastrophically when trained on random data, requiring expensive, actively generated quality-diversity archives to generalize.

The resulting pipeline, integrated into the open-source OpenSKIZZE tool, can generate thousands of diverse, climate-evaluated building layouts in under ten minutes. The optimization achieves highly accurate fitness rankings (Spearman ρ = 0.994) using only a one-time batch of random training samples, bypassing the need for iterative, expensive data collection. This dramatic acceleration opens up new possibilities for climate-adaptive urban planning, where designers can systematically explore the trade-off between building density and cold-air ventilation. The work was accepted to the open-access venue arXiv and is published under arXiv:2606.04658.