Physics-informed deep neural network for modeling the chloride diffusion in concrete

Abstract

Chloride diffusion in concrete is a complex chemo-physical process and it is of pivotal importance to forecast the initiation time of corrosion. But limited equations are accessible to simulate the chloride diffusion considering the nonlinear chloride binding capacity of concrete. This study proposes a physics-informed deep neural network to simulate the chloride diffusion mechanism and forecast the distribution of chloride concentrations in concrete. Physical laws are formulated as a loss term to guide the training process and to mitigate the data required for model training. The physical constraint loss (based on the governing equation and boundary conditions) and the training loss (based on the neural network) are then fused to produce the loss function. Different experimental cases are first employed to validate the proposed model and then the model is compared to numerical applications and purely data-driven approaches. Results revealed that the proposed model could effectively simulate the chloride transport behavior and forecast the diffusion coefficient of concrete with high precision over other states of art methods. The application of this method to the temporal and spatial domains of the chloride concentration within the concrete samples showed its potential as a powerful tool for investigating concrete properties.