[Paper] U-Net-Accelerated Quality-Diversity Optimization for Climate-Adaptive Urban Layouts

Source: arXiv - 2606.04658v1

Overview

The paper presents a new pipeline that couples a U‑Net deep‑learning surrogate with a MAP‑Elites quality‑diversity (QD) optimizer to generate climate‑adapted urban layouts at interactive speed. By replacing a costly physics‑based climate simulator with a spatial neural network, the authors make it feasible to explore thousands of design alternatives in minutes—something that was previously limited to a handful of manually evaluated concepts.

Key Contributions

• Spatial surrogate model: Introduces a U‑Net architecture that learns the full 2‑D climate response (cold‑air ventilation) from layout images, achieving an $R^2$ of 0.996.
• Surrogate‑enabled MAP‑Elites: Integrates the U‑Net into an offline MAP‑Elites loop, producing a high‑quality, diverse archive of building layouts in under ten minutes.
• Training‑data analysis: Systematically compares random Sobol sampling vs. active QD bootstrapping, showing that the spatial inductive bias of the U‑Net makes it robust to the source of training data, unlike scalar Gaussian‑process (GP) surrogates.
• Open‑source toolchain: Releases the full pipeline as part of the OpenSKIZZE platform, allowing practitioners to plug‑in their own simulators or datasets.
• Empirical validation: Demonstrates that fitness rankings from the surrogate‑driven QD run correlate strongly with the true simulator ($\rho = 0.994$), confirming practical fidelity.

Methodology

1. Data Generation: A one‑off batch of layout–climate pairs is created using a high‑fidelity regulatory climate simulator. The layouts are represented as raster images (building footprints, streets, green spaces).
2. Surrogate Training:
   • U‑Net: A convolutional encoder‑decoder network that predicts a spatial field of ventilation performance for any input layout.
   • GP Baseline: A scalar Gaussian‑process model that predicts a single fitness value per layout.
     Both models are trained on either Sobol‑sampled random layouts or on archives generated by an active QD process.
3. Quality‑Diversity Optimization (MAP‑Elites):
   • The surrogate replaces the expensive simulator inside the MAP‑Elites loop.
   • The algorithm explores a predefined behavior space (e.g., building density vs. ventilation) and stores the best‑performing layout for each cell, building a “illumination map” of the design space.
4. Evaluation: After the offline QD run, a subset of the archive is re‑evaluated with the true simulator to measure ranking correlation and overall surrogate accuracy.

Results & Findings

• Surrogate Accuracy: The U‑Net consistently reaches $R^2 = 0.996$ on held‑out test data, regardless of whether it was trained on random or actively generated samples.
• GP Fragility: Scalar GP models only perform well when trained on actively generated QD archives; with random data they collapse, requiring many more expensive simulations.
• Ranking Fidelity: The surrogate‑driven MAP‑Elites archive yields a Spearman rank correlation of $\rho = 0.994$ compared to the ground‑truth simulator, meaning the ordering of layouts is virtually unchanged.
• Speed: The entire pipeline produces thousands of diverse, climate‑evaluated layouts in <10 min on a commodity GPU, a >100× speed‑up over direct simulation.
• OpenSKIZZE Integration: The authors packaged the workflow into an easy‑to‑use Python library, enabling rapid prototyping for urban planners and developers.

Practical Implications

• Rapid Ideation: Architects and city planners can now generate and compare thousands of climate‑responsive layout concepts in real time, supporting data‑driven decision making.
• Design Exploration at Scale: Developers of generative design tools can embed the U‑Net surrogate to provide instant feedback on ventilation performance, reducing reliance on costly CFD or regulatory simulators.
• Plug‑and‑Play Surrogates: Because the surrogate operates on raster images, it can be swapped for other spatial performance metrics (e.g., solar exposure, flood risk) with minimal code changes.
• Cost Savings: By eliminating the need for iterative high‑fidelity simulations, firms can cut compute budgets dramatically while still delivering scientifically grounded designs.
• Open‑Source Ecosystem: The OpenSKIZZE release encourages community contributions, such as extending the behavior space, adding new climate variables, or integrating with GIS pipelines.

Limitations & Future Work

• Domain Specificity: The surrogate is trained on a single climate region and building code; transferring to a different climate or regulatory context will require retraining with new simulation data.
• Resolution Constraints: The U‑Net operates on fixed‑size raster inputs; very large city blocks may need tiling or hierarchical modeling.
• Physical Fidelity: While ranking is preserved, absolute performance values may still deviate from the true simulator, which could matter for compliance checks.
• Active Learning Extensions: The authors suggest exploring on‑the‑fly data acquisition (e.g., querying the physics simulator only for high‑uncertainty layouts) to further reduce training costs.
• Multi‑objective Extensions: Future work could integrate additional objectives (energy consumption, daylighting) into the QD archive, testing how the surrogate scales with more complex behavior spaces.

Authors

• Alexander Hagg
• Tania Guerrero
• Dirk Reith

Paper Information

• arXiv ID: 2606.04658v1
• Categories: cs.NE, cs.LG
• Published: June 3, 2026
• PDF: Download PDF