Physics-informed neural networks to advance pavement engineering and management

Abstract

Physics-informed neural networks (PINN) are machine learning (ML) algorithms that can bridge the gap between our understanding of physical phenomena and the corresponding empirical observations. This paper discusses applications of physics-integrated ML to advance pavement engineering. To demonstrate an example, a PINN model was pretrained to approximate the simulation of vehicles’ suspension responses to longitudinal road profiles. The parameters of the outer layers were finetuned to adapt model output to the standard International Roughness Index (IRI), while keeping the pretrained inner layers to preserve the embedded physical knowledge of the suspension behaviour. The PINN model showed low bias and standard error in predicting IRI values on training, test, and an independent dataset from an autonomous vehicle study by the Ford Motor Company. This approach to reconcile and supplement infrequent survey data with spatiotemporally continuous data (from connected vehicles) can enhance data-driven practices for pavement design, maintenance and asset management.