Chapter 2: Physics-informed machine learning: Applications in smart transportation

Advancements in machine learning have significantly influenced various sectors, including transportation engineering. Machine learning techniques, particularly deep learning, have been applied to optimise routing decisions, predict traffic patterns and enhance safety measures within transportation systems. Building on these data-driven approaches, physics-informed deep learning (PIDL) integrates traditional deep learning with fundamental physical laws, offering transformative potentials for transportation engineering. This chapter explores how PIDL combines deep neural networks with principles from traffic flow theory to enhance traffic state estimation and prediction. It begins by elucidating the physical principles governing traffic dynamics and detailing the architecture of a PIDL neural network. Through case studies, the chapter demonstrates that PIDL surpasses traditional deep learning models in accuracy and computational efficiency, even with limited observed traffic data. This efficiency in utilising computational resources and training data makes PIDL particularly suitable for real-time intelligent vehicle applications. The chapter goes on to address practical challenges such as sensor reliability during traffic data collection, the impact of varying levels of connected autonomous vehicles (CAVs) penetration rate, and communication limitation with technologies such as dedicated short-range communications (DSRC). This chapter aims to guide engineers and students by providing a detailed model development process, encompassing training and testing phases, and offering relevant code scripts and datasets to facilitate further exploration and understanding.