An effective multitask neural networks for predicting mechanical properties of steel

Abstract

An effective multitask neural network was carried out to predict the mechanical properties of steels based on their chemical compositions. The multitask neural network model outperforms other neural networks, and conventional algorithms. It achieved high prediction accuracy for both tensile strength and elongation, with R2 values of 0.9204 and 0.9409, respectively. Benefiting from the strong inter-task relationships, the multitask neural network enhances performance and parameter efficiency by sharing a potent representation across tasks. Additionally, we analyzed the influence of chemical composition on mechanical properties using the model's parameters, providing valuable insights into the relationship between different chemical compositions and the mechanical properties of steels.