Application of physics-informed deep learning in transportation: A review

A framework that integrates the advantages of Machine Learning (ML) with insights from governing physical equations is called Physics-Informed Deep Learning (PIDL). This paper comprehensively examines the existing scientific literature on PIDL in transportation. The objective is to identify emerging techniques of PIDL, their potential applications, and classify them using a taxonomy. The PRISMA flowchart is employed to facilitate the systematic literature search by establishing specific inclusion and exclusion criteria. Various keyword searches are conducted within the Scopus and Web of Science databases, accompanied by examining relevant references and citation analysis to identify eligible documents. A total of 157 and 11 documents were identified through the initial databases search and the forward and backward snowballing methods, respectively. Among them, 91 documents were selected for full-text reviews, resulting in 64 being ultimately included in the study. The identified literature was categorised into four distinct groups according to the nature of the records, which include traffic-related (55% of documents), vehicle-related (27%), personalrelated (8%), and miscellaneous (10%) categories, reflecting the transportation system’s components specifically individuals, vehicles, and traffic networks. This paper also identified potential research gaps in PIDL as applied to transportation and proposed different approaches to overcome these potential research gaps.