Fluid-TrafficNet: A Physics-Informed Deep Learning Framework for Geometric Evaluation of Road Intersections

Evaluating the geometric efficiency of road intersections typically relies on computationally expensive microsimulations, which hinders rapid iteration during earlystage urban planning. This paper proposes Fluid-TrafficNet, a novel framework that evaluates road topology from a fluid dynamics perspective. Drawing an analogy between traffic flow and fluid mechanics, we utilize the SUMO simulation engine to generate a synthetic dataset comprising 200 diverse intersection geometries, labeled with fluidity scores based on average vehicle waiting times. A Convolutional Neural Network (CNN) is designed to learn the mapping from static top-down road imagery to these fluidity indices. Experimental results demonstrate that the model achieves a Root Mean Square Error (RMSE) of approximately 12.5 on the test set. Furthermore, we validate the model on real-world satellite imagery, where it successfully distinguishes between high-efficiency interchanges (Score > 84) and congestion-prone orthogonal intersections (Score < 76). This work demonstrates the feasibility of using physics-informed deep learning for zero-shot geometric evaluation of transport infrastructure.