Machine learning algorithms for the prediction of the strength of steel rods: an example of data-driven manufacturing in steelmaking

Abstract

Analytical models based on physical metallurgy are of limited ability to predict the strength of steel due to the complexities of steelmaking. This paper presents the results obtained using Machine Learning procedures to predict the tensile strength of steel rods manufactured in an electric arc furnace. The available dataset includes 5540 observations (tensile tests) and 97 features (fabrication parameters) monitored during the different stages of the process (electric arc furnace, ladle furnace, continuous casting and hot rolling). The following regression algorithms have been implemented: Multiple Linear Regression, K-Nearest Neighbors, Classification and Regression Tree, three Ensemble Methods (Random Forest, Gradient Boosting and Adaboost) and Artificial Neural Networks. The fine-tuned Random Forest, provided an R2 of 0.775 and a mean absolute percentage error of 0.76% in the test dataset. After optimization, the Feature Importance and the Permutation Importance algorithms showed that chemical variables have the greater influence on the material strength. The quantitative influence of these variables was represented through Partial Dependence Plots. In short, this research has enabled validating a series of Machine Learning models that provide the necessary information for a correct decision-making to optimize the strength of the steel rods.