Machine learning for predicting concrete carbonation depth: A comparative analysis and a novel feature selection

Abstract

The accurate prediction of concrete carbonation depth is essential to prevent concrete from cracking and corrosion. However, identifying the critical parameters affecting carbonation depth is challenging due to the process’s complexity and the involvement of many variables. In this investigation, a range of computational techniques was employed to predict the carbonation depth of concrete, including artificial neural networks, random forests, decision trees, and support vector machines, on a dataset of 37 variables. Detecting the optimal variables to maximize the prediction accuracy is a complicated task. To address this, a novel technique called MOEA/D-ANN (Multi-objective Evolutionary Algorithm based on Decomposition and Artificial Neural Networks) was introduced in this study. This technique effectively determines the optimal significant variables, resulting in superior predictive performance. The effectiveness of the introduced technique was assessed by employing a conventional feature selection method called Regression Relief Feature Selection (RReliefF) as well. The results indicate artificial neural network was the most accurate model based on prediction accuracy when using the optimal feature set found by MOEA/D-ANN. Additionally, the MOEA/D-ANN approach demonstrated the ability to decrease the training time required for the modeling process. The findings demonstrate the usefulness of the MOEA/D-ANN method in developing accurate and efficient models for predicting concrete carbonation depth. Furthermore, this approach can be extended to other important properties of concrete, thereby improving the construction and maintenance of concrete structures.