PINN-AFP: A novel C-S curve estimation method for asphalt mixtures fatigue prediction based on physics-informed neural network

Abstract

The accurate prediction of fatigue life of asphalt mixture is the key to the design of long-lasting durable pavement. Currently, a critical aspect influencing the accuracy of life prediction methods for asphalt mixtures, based on Viscoelastic Continuum Damage Mechanics (VECD), is the precision of the damage characteristic (C-S) curve. Within the VECD framework, the C-S curve of asphalt mixture is regarded as an intrinsic material property. However, it has been observed in engineering applications that the C-S curve of the material exhibits notable sensitivity to varying loading conditions. The nonlinear fitting method based on a small number of experiments is difficult to fully characterize the fatigue performance of the material, while a large number of complete material fatigue tests are expensive in time and money. Based on the above problems, a physics-informed neural network embedded in VECD, PINN-AFP, is proposed. It can accurately predict the complete material C-S curve based on a small amount of pre-fatigue data of the material, thus achieving accurate prediction of the fatigue life of asphalt mixture. The case study uses the fatigue test data of AC-25 as an example, and the results demonstrate that the proposed PINN-AFP has strong generalization ability and prediction accuracy, achieving the state-of-the-art in the mainstream machine learning and deep learning methods with an average 5.2% fatigue life prediction error.