Exploring physics-informed recurrent neural networks for constitutive modeling of concrete pavements

Abstract

This study proposes a preliminary investigation into the use of physics-informed recurrent neural networks for constitutive modeling of concrete. Traditional empirical models for concrete often face limitations in capturing complex, path-dependent behaviors due to the constraints of experimental datasets. To address this, we explore the integration of recurrent neural network architectures, with foundational principles from continuum mechanics, including thermodynamic consistency and objectivity. This theoretical framework aims to establish a foundation for future developments of hybrid models that reduce reliance on extensive experimental data while maintaining physical interpretability. The work highlights key challenges, including incorporating physical constraints into machine learning architectures and addressing the material's inherent complexity, setting the stage for further empirical validation and refinement.