Large-scale synthetic environments are essential for advancing smart cities research. However, existing urban benchmarks typically prioritize either visual photorealism or geometric fidelity, often neglecting the material-level semantics and physical attributes required for simulations. To address this gap, this paper introduces UR-MAT , a modular design framework and a multimodal synthetic dataset specifically built to bridge computer vision and physics-based modelling. We present a unified construction workflow that integrates georeferenced geometry, physically based rendering (PBR) materials, and domain-specific metadata (e.g., electromagnetic properties based on ITU-R standards) into a traceable data representation. The resulting dataset encompasses eight heterogeneous urban scenarios, providing spatially aligned RGB imagery, depth maps, semantic segmentation masks, and 3D point clouds. The utility of UR-MAT is validated through cross-domain experiments: (i) a quantitative baseline analysis on semantic material segmentation across fourteen material classes, where the best-performing model achieved 0.787 mIoU and 0.968 pixel accuracy, and (ii) electromagnetic ray-tracing simulations, which verify that the dataset’s material annotations yield physically consistent and deterministic radio propagation patterns. By enabling perception and simulation tasks within a single coherent framework, UR-MAT serves as a resource for developing perception-driven communication strategies and next-generation Digital Twins.

The development of smart city applications increasingly relies on datasets that capture not only urban geometry but also the material properties that influence real-world interactions. However, most existing datasets lack detailed semantic labels and physical attributes, limiting their applicability to tasks such as signal propagation modeling and material-aware analysis. In this paper, we propose a structured process and a tool pipeline for generating material-aware synthetic datasets from custom 3D urban scenes. Our approach integrates 3D modeling, physically-based rendering, and automated data capture to produce multimodal outputs, including RGB images, depth maps, segmentation masks, structured metadata, and annotated 3D meshes. We demonstrate the process and pipeline instantiation using the Louvre Museum district as a running scenario, capturing the site's complex architecture and material diversity. A preliminary ray-tracing simulation in MATLAB further demonstrates how the generated data can support material-aware applications. By bridging geometry, appearance, and physical information, the proposed solution can contribute to advancing simulation-ready datasets for smart city research.

We present UR-MAT, a multimodal, material-aware synthetic dataset for urban scene understanding and physics-based simulation. UR-MAT comprises seven diverse outdoor environments, ranging from historic districts to modern office areas, reconstructed from OpenStreetMap data and procedurally enhanced in Unreal Engine. Each scene includes semantically structured 3D meshes and physically based rendering (PBR) materials annotated with electromagnetic properties such as permittivity, reflectance, and attenuation. The dataset provides spatially aligned RGB images, material segmentation masks, depth maps, point clouds, camera poses, and 3D mesh files in .glb format. All data is generated through a deterministic, reproducible pipeline integrating OSM2World, Unreal Engine, and UnrealCV. UR-MAT supports a wide range of research tasks, from semantic segmentation and 3D reconstruction to material-aware electromagnetic simulation (e.g., mmWave propagation). We also release two utility scripts to extract mesh-material relationships and assign physical metadata, enabling dataset extension and reproducibility. By bridging computer vision and physical modeling, UR-MAT serves as a testbed for multimodal AI and signal-aware urban simulation.