Long-term prediction of surface chloride content of marine concrete structures using inverse physics-informed neural networks

Abstract

In a complicated marine environment, the surface chloride content on reinforced concrete structures exhibits stochastic temporal fluctuations. Traditional models of surface chloride content only consider the change trend over time without considering the stochastic temporal fluctuations, leading to a lack of comprehensiveness and accuracy in prediction results. To address these limitations, a physics-data-driven framework, coupled with a time-dependent deterministic and fluctuation function, is proposed to reconstrue the holistic trend and local fluctuations of the surface chloride content with limited data. The fluctuation characteristics of the surface chloride content curve, arising from environmental action and structural resistance, are extracted using the Fourier-cosine decomposition method. An inverse physics-informed neural networks is proposed to enhance the predictability of stochastic and discrete time series. In this physics-informed neural network, a surface chloride content probabilistic model is used as the surrogate model to reconstrued the stochastic temporal fluctuations. The boundary loss function is formulated as the physical constraints of the aforementioned inherent-external stochastic temporal fluctuations, while the true value loss is generated as the discrepancy among the measured data, PINN prediction data, and surrogate model data. Utilizing the measured data and PINN, the surrogate model is refined, in which the initial chloride content, the accumulation rate, and the fluctuation phases of surface chloride content are updated. The application of this model to over 16-year data of naturally exposed concrete specimens demonstrate its effectiveness and accuracy for predicting surface chloride content with the relative error controlled by 10 %.