Multiscale modeling and characterization of adhesion between bitumen and aggregate Phase II: Utilizing Physics-Informed Neural Network to Predict Fatigue Performance of Asphalt Mixtures

• Applicant: Professor Dr.-Ing. Pengfei Liu
• Subject Area: Construction Material Sciences, Chemistry, Building Physics
• Term: since 2021
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 459436571

In Phase I of this project, comprehensive multiscale experimental and advanced numerical simulation methods were successfully developed to better understand the adhesion between bitumen and aggregates. The investigations provided critical insights into the fundamental mechanisms governing adhesion and debonding in bituminous materials. However, despite significant progress in addressing adhesion-related challenges, a major issue remains: accurately predicting the fatigue behavior of bituminous materials across different scales, from nanoscale asphalt chemistry to mesoscale crack development. As fatigue damage progressively weakens pavement structures and poses substantial risks to public transportation safety, addressing this issue has become a pressing concern for German pavement scientists. Thus, Phase II aims to innovatively bridge this gap by leveraging the experimental data and simulation techniques from Phase I, in combination with Physics-Informed Neural Networks (PINNs), to establish a unified multiscale framework. This framework will utilize molecular composition data of bitumen, along with other key parameters, to predict the mesoscale fatigue performance of asphalt mixtures accurately and efficiently. The integration of physical laws across various scales—Newton's second law, the Viscoelastic Continuum Damage (VECD) model, and Paris law—into the PINN model's loss functions will ensure accurate cross-scale predictions. At each scale, data from prior research, relevant literature, and particularly from the experimental tests and validated numerical models developed in Phase I, will be utilized to train the neural network. Leveraging these extensive datasets will minimize the need for new experiments in Phase II, ensuring the project remains within budget and on schedule. Only a limited number of additional fatigue tests will be conducted to validate and refine the model. The Pareto optimization will be employed to balance loss functions across scales, further enhancing model efficiency and accuracy. The key innovations of Phase II include the development of this comprehensive PINN multiscale framework, the improved integration of asphalt chemistry and fatigue behavior through physical laws, and the application of advanced optimization techniques. These advancements are expected to significantly reduce reliance on extensive physical testing, accelerate material development, and contribute to the sustainable construction of durable, long-lasting pavements.

• DFG Programme: Research Grants