Machine learning-based probabilistic prediction model for chloride concentration in the interfacial zone of precast and cast-in-place concrete structures

Reasonable prediction of chloride concentration in the interfacial zone of precast and cast-in-place concrete structures is crucial for assessing the durability of such structures. In this paper, the machine learning-based deterministic model is first proposed to predict the chloride concentration in the interfacial zone using seven machine learning methods. Then, an uncertainty prediction model of chloride concentration in the interfacial zone is developed considering the aleatory uncertainty of input parameters and epistemic uncertainty of model itself. Finally, the effectiveness of the proposed prediction model is validated by using the collected database containing 2505 sets of data. The research results indicate that the deep neural network (DNN) model performs the best in predicting chloride concentration in the interfacial zone among the seven machine learning (ML) models, achieving an R2 value as high as 0.9847. The proposed uncertainty prediction method not only demonstrates high prediction accuracy but also effectively accounts for uncertainty in the prediction process. The obtained prediction interval coverage probability (PICP) of the proposed method reaches to 0.9457 for the collected database. Moreover, rising dropout rate can improve the interval coverage at the expense of interval clarity, but the overall prediction performance still improves. Additionally, the prediction performance of uncertainty model for chloride concentration in interfacial zone is more sensitive to epistemic uncertainty than aleatory uncertainty in this case study.