Prediction of Chloride Ingress Profiles in Concrete by Machine Learning

Abstract

Machine learning offers a very promising tool for predicting unknown output of future instances, when trained on a database containing input and output values. However, the predictions of these methods are very limited to the scope of the data. They are often very unreliable in extrapolating results outside the used training experiments, making it unsuitable for direct extrapolation analysis. Prediction of chloride ingress at exposure times relevant for desired service life requires extrapolation in the temporal domain. This paper proposes a solution to this issue by a preprocessing step, which causes chloride profiles from multiple exposure times to largely overlap in a master curve. In a reassessment scenario these overlapping curves could be used to predict further chloride ingress. By applying machine learning on the preprocessed raw data, master curves can be predicted for new compositions and exposure environments relevant in a design scenario. In a postprocessing step the master curves are then converted to predicted chloride profiles. The method was trained on data from 6 north European exposure sites sampled after 2–21 years and tested against chloride profiles obtained from the submerged zone of old Danish bridges sampled after 30, 31 and 34 years. Chloride ingress predictions, which could in principle be obtained in the design stage, results for a basic ML model applied on OPC concrete as well as a refined method applied on FA concrete in root mean square errors at the level of uncertainties between chloride profiles measured at different locations for the same concrete and exposure time after 30–34 years. It is concluded that there is a huge potential for using machine learning in the design stage to predict chloride profiles in concrete when data is pre- and post-processed.