The durability of materials has long been a hot topic in civil engineering. This paper used a new approach to foster current research in two aspects: identifying important properties of concrete and simulating the transport of chloride in concrete. Machine learning methods have proven to have remarkable abilities in approximating complex functions, which provides a potential tool for realizing the above ideas. This paper used deep neural networks to identify important parameters of concrete via datasets and simulate the transport process by integrating physical constraints into models. Results showed that physical constraints could be well combined with deep neural networks to accurately identify materials’ properties (relative error < 1%) and simulate the aggression of chloride. This study also demonstrated that machine learning models could be interpretable. In addition, this paper also explored the influence of time intervals of detection datasets, the size of datasets, and the noise in datasets (-5%~5%) on the convergence of identification. Results showed that short time intervals and big datasets can accelerate convergence. This method can resist the influence of noise in datasets. In addition, with exact physical constraints, this method can predict the performance of materials without the assistance of any dataset.

Accurate prediction of structures’ durability has long been a hot topic in civil engineering. This paper proposed a new solution to this topic, i.e., integrating physical constraints into machine learning models. Machine learning methods have proven to have remarkable abilities in approximating complex functions, which provides a potential tool for finding the model that complies with all physical constraints. This paper used this method to predict the carbonation status of concrete. Partial differential equations governing the diffusion process of CO2 in concrete were used as physical constraints, which consider the effects of aggregate, load, and interface transition zone. Then, this study integrated physical constraints into deep neural network models and established CO2’s one- and two-dimensional diffusion models. Results showed that physical constraints could be well combined with deep neural networks and this new method can accurately simulate the diffusion of CO2. This method has great potential in solving multi-physics problems.