AI designs residential yards that better balance comfort, looks, access, and nature

A hybrid AI system generated landscape layouts for three Changsha residential sites and selected options that scored closer to professional designs across multiple goals.

Editorial illustration — not from the study.

Researchers say designing residential landscapes involves balancing competing goals—such as microclimate comfort, visual quality, functional accessibility, and ecological service provision—and that most early-stage workflows do not handle these trade-offs systematically. The team built a hybrid generative AI approach that creates a range of structurally coherent yard and site layouts based on site boundaries, building layouts, and density constraints.

Each generated option was evaluated using a four-part performance pipeline—microclimate modeling, an image-based aesthetic scoring system, graph-based accessibility measures, and ecological service estimators. Using a multi-objective optimization method, the study searched for Pareto-optimal configurations in the AI’s latent space and then used decision support to pick balanced recommendations.

Optimized landscapes score higher

In three residential parcels in Changsha (high-, mid-, and low-density), the hybrid AI produced lower Fréchet Inception Distance scores (24.6–31.4) and higher morphological diversity (LPIPS 0.412–0.461) than standalone GAN or VAE baselines. The optimized schemes scored 6.2% above professional human-designed benchmarks on a composite performance index on average, with the biggest margin on the most spatially constrained site. Ablation and sensitivity analyses indicated that each component of the framework contributed and that the composite rankings were robust when weights were perturbed.

Balancing design trade-offs

The work offers a way to generate and compare early landscape design options using multiple performance goals at once, rather than treating them separately. Within the tested Changsha settings, the method identified balanced recommendations that more closely matched professional benchmarks on a composite score.

Model-based tests and caveats

The evidence is based on model-based evaluation: generated designs were judged by a microclimate surrogate model, a CNN-based aesthetic scorer, graph-theoretic accessibility metrics, and ecological service estimators, with optimization using NSGA-III and selection support using TOPSIS. The authors note that all three parcels share one subtropical humid climate, so the results are framed as a model-based advantage rather than validated performance across different climates.

// Source

Scientific Reports · 2026 · DOI: 10.1038/s41598-026-62192-0

View paper (DOI)Open access versionPublished 2026-08-01

Authors: Jiyu Wang

Institutions: Sias University

A flying robot used touch to perch on unfamiliar structures

Researchers developed an aerial robot that uses touch to refine its position and perch on structures it has not encountered before. The system succeeded in more than 99% of simulated trials and completed 26 real-world perching trials despite corrupted position estimates.

4 hours ago·3 views

Abstract editorial illustration accompanying the article: Belgium’s low-carbon energy plans could cut climate pollution but raise other risks

Engineering & Technology

Belgium’s low-carbon energy plans could cut climate pollution but raise other risks

A modeling study of Belgium’s 2045 energy system found that all near-optimal low-carbon pathways met the climate target but exceeded limits for ecotoxicity and particulate matter under the most permissive assessment approach. The pathways had similar costs but differed in whether they relied mainly on electrification, hydrogen, renewable power capacity or a mix of strategies.

1 day ago·49 views

Engineering & Technology

A new traffic-planning method handles more controls across complex road networks

Researchers developed a way to calculate traffic sensitivities efficiently in road networks with different vehicle classes, merges, diverges and congestion spillback. Tests on a sample network found the approach was particularly efficient as the number of control parameters increased.

1 day ago·3 views