A physical-digital integration framework for environmental simulation through deep learning: Wind flow implementation

This research introduces a novel four-layer framework that bridges the gap between design with physical models and real-time environmental analysis in architecture. While physical models remain essential for spatial comprehension and tactile design exploration, their disconnect from environmental performance assessment limits their utility in sustainable architecture. Our framework addresses this challenge through four integrated layers: (1) a physical layer for tangible model manipulation, (2) a digital layer for real-time spatial recognition, (3) an AI processing layer for environmental simulation, and (4) an interaction layer for visualization and control. We demonstrate this framework through wind flow analysis implementation, developing a multimodal pix2pix model that achieves wind flow prediction with SSIM values of 0.754 and PSNR of 22.630, trained on 603 apartment complexes across five South Korean cities. The digital layer employs ArUco markers for robust object detection, while the interaction layer integrates the Mixtral-8x7b language model for natural parameter control through a web-based interface. Physical prototyping and user evaluation validate the framework's effectiveness, confirming its ability to preserve intuitive design workflows while providing immediate environmental feedback. By integrating physical modeling with real-time analysis, the system demonstrates significant potential for transforming architectural practice, education, and stakeholder engagement, while establishing a foundation for expanded environmental assessment capabilities.