New Generative AI Framework MMCN Predicts Future Urban Layouts to Support Sustainable City Planning

JAIST researchers develop MMCN, a memory-aware AI framework that predicts future city layouts with high spatial coherence for sustainable urban development.

By: AXL Media

Published: Mar 24, 2026, 9:17 AM EDT

Source: Information for this report was sourced from Japan Advanced Institute of Science and Technology

Bridging the Gap Between AI and Urban Design

As global urbanization accelerates at an unprecedented pace, city planners face the daunting task of making long-term infrastructure decisions that will define a city's environmental sustainability for decades. Traditional design methods often struggle to synthesize the complex, interconnected dynamics of building height, road networks, and historical development patterns. To address this, a research team led by Associate Professor Haoran Xie has introduced the Memory-aware Multi-Conditional generation Network (MMCN). This framework is designed to mimic the professional workflow of urban planners, offering a predictive tool that captures how cities evolve as integrated systems rather than fragmented patches.

The Architecture of Spatial Coherence

The MMCN model utilizes a diffusion-based generative architecture enhanced by three specialized modules. A multi-conditional control mechanism allows diverse urban factors to guide the AI, while a semantic prompt fusion module encodes specific data types like building density and transportation layouts. Crucially, the system includes a spatial memory embedding component. This module preserves contextual information from neighboring urban regions, ensuring that the generated layouts maintain continuity across large areas. By using edge-stitching loss functions during training, the model prevents the "fragmented" look common in earlier AI-generated maps, producing smooth transitions between different city zones.

Testing the Model in Rapidly Evolving Environments

To train and validate the framework, researchers utilized multi-temporal spatial data from Shenzhen, recognized as the most rapidly developing city in China. The model was tasked with reconstructing and forecasting urban layouts in 512 × 512-pixel patches. In comparative testing against baseline AI models—including Pix2Pix and CycleGAN—MMCN achieved a Structural Similarity Index (SSIM) of 0.885. Qualitatively, the MMCN-generated maps displayed realistic, well-organized building clusters and continuous road networks, whereas other models frequently produced duplicated structures or disconnected, unusable transportation paths.