Physics-Informed neural network based inversion and prediction of natural chloride diffusion in uncracked and cracked concrete systems
Year: 2026
Type: article
Abstract: Studying chloride diffusion in concrete is essential for predicting structural durability and designing corrosion-resistant materials and structures. While analytical models and finite element methods can simulate diffusion, they typically require large and high-quality datasets and do not possess advantages in parameter identification. Physics-Informed Neural Network, which integrates Fick 2nd law with initial and boundary conditions, offer a promising alternative. It not only replicates diffusion behavior accurately but also enhances the fitting of experimental data via a data-driven loss term and enable inverse estimation of diffusion related parameters. This paper outlines three key advantages of using this new method for problem-solving about chloride diffusion in concrete: (1) robustness to noise and low data requirements for one-dimensional inverse estimation of diffusion coefficients; (2) strategy integrates data, physics, and engineering insights for parameter inversion.; and (3) extended physics-informed frameworks with weak constraints for cracked concrete. Overall, Physics-Informed Neural Network provides a robust numerical tool for efficient durability assessment and the design of corrosion-resistant and resilient concrete structures. For self-healing concrete, the proposed framework effectively estimates diffusion coefficient in healed crack and accurately predicts long-term diffusion behavior, contributing to the optimal design and evaluation of self-healing materials.
Source: Computers & Structures
Institutions: Politecnico di Milano, Shenzhen University
Language: English
Subfield: Civil and Structural Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $4,200
Award: G8194877290
OpenAlex ID: W7125810109