Machine learning-driven predictive modeling of mechanical properties in diverse steels

Abstract

This study explores the application of machine learning (ML) in steel design using a small dataset of various steel grades that include 13 key elements and three critical mechanical properties. Random forest (RF) models were systematically evaluated for their robustness and effectiveness in predicting the stress-strain of steel properties. Moreover, other alternative approaches, such as support vector machines, extreme gradient boosting machines, and artificial neural networks, were also evaluated to ensure that the predictions made by the RF model are as accurate as possible. To assess the bias-variance trade-off, 1-seed and random 100-seeds with 80/20 train/test split, and leave-one-out cross-validation for all datasets were conducted. The results demonstrated that the RF models are accurate and reliable, achieving low bias and variance while delivering predictions comparable to, and in some cases better than, those obtained in studies with larger datasets. The analysis revealed a trade-off between strength and ductility, with elongation negatively correlated with yield strength and ultimate tensile strength. This study highlights the feasibility of using small, realistic datasets to develop effective ML models for predicting mechanical properties in steel design. The methodology can also be readily extended to investigate processing-property relationships in other systems, offering a versatile approach for advancing materials science through data-driven techniques.