Quantitative description of chloride ingress in concrete using machine learning algorithms

Abstract

This study employs machine learning (ML) algorithms to quantitatively describe the spatiotemporal evolution of chloride ion content in concrete. This phenomenon is of paramount importance for the maintenance of concrete infrastructure components, e.g., bridges, as chloride ion ingress is the most common cause of corrosion in reinforcing and prestressing steel. The development and training of the ML algorithms are driven by a comprehensive dataset, consisting of 918 experimental measurements and incorporating a range of environmental and concrete mixture properties as input features. A variety of ML algorithms are considered, namely, linear regression, least absolute shrinkage and selection, multilayer perceptron artificial neural network, support vector machine, Gaussian process regression, random forest, extreme gradient boost, and a voting regressor combining the various algorithms. The latter was found to be the most promising approach for the adopted dataset in reproducing chloride content measurements from real-life structures, as it incorporates a variety of algorithms. The proposed ML framework provided insights into the optimal concrete mixture design to enhance the serviceability of critical arterial infrastructure components in harsh environments. The accuracy of a design-oriented mathematical model in the fib (International Federation for Structural Concrete) code for describing chloride ingress was also investigated and found to systematically underestimate the chloride content.