Predicting the Mechanical Strengths of Steel Fiber–Reinforced Concrete Using Machine-Learning Methods and Feature Selection Techniques

Abstract

The purpose of this article is to predict the mechanical strength of concrete reinforced with three types of macrosteel fibers: straight; crimped; and hooked-end. This paper compares well-known machine learning algorithms to predict the compressive, tensile, and flexural strength of steel fiber–reinforced composites (SFRC) based on experimental training data. Also, different feature selection techniques are integrated and compared to find the optimal features. For this purpose, 2,283 experimental multinational databases are extracted from the state-of-the-art and preprocessed using normalization. Three feature selection methods, i.e., lasso, ridge, and principal component analysis, are used considering different fiber types. It was found that the XG-boost had the best performance, and the supported vector machines were not as good as others. Also, three strategies for the feature selection technique based on the mentioned database division are investigated. The effect of the feature selection techniques and methods is described in the text. The proposed winner strategy can improve prediction performance by up to 30% on average. This research brings an exclusive comparative prediction based on adequate training data, which enables the reader to know the machine learning algorithm to predict the compressive, tensile, and flexural strength of steel.