Optimized Predictive Modeling for Accurate Hardness Profiles in Hot-Rolled Alloy Steel Manufacturing Processes

The alloy steel’s Mechanical Properties (MPs) are most influenced by its composition of chemicals and the hot rolling process factors. However, modeling the interactions between these components is extremely difficult due to the rolling process’s complexity and dynamic nature as a nonlinear system. To overcome these challenges, this work introduces an advanced approach for optimizing predictive modeling in Hot-rolled Alloy Steel (HRAS) manufacturing, focusing on enhancing accuracy in hardness profiles and refining process parameters. The proposed approach utilizes the Temporal Dynamic Graph Neural Network (TDGNN), with the main objective of improving the mechanical characteristics of HRAS and enhancing accuracy and reliability. The TDGNN is employed to forecast the mechanical characteristics. The proposed method is implemented and compared with existing techniques on the MATLAB platform, including Physics-informed Neural Networks (PINNs), Deep Neural Networks (DNNs), and Convolutional Neural Networks (CNNs), demonstrating superior performance. The proposed TDGNN approach achieved a prediction accuracy of 97%, significantly outperforming existing methods such as CNN (71%), PINN (79%), and DNN (88%). It also recorded the lowest Mean Square Error (MSE) values for key MPs: 0.0003 for Tensile Strength (TS), 0.0002 for Yield Strength (YS), and 0.0003 for Elongation (EL), along with minimal Mean Absolute Errors (MAEs), confirming its superior prediction capability. The analysis further shows that the most frequent prediction errors for YS and EL fall within the range of −2% to 0%, indicating highly reliable forecasts with minimal underestimation. These results emphasize the notable performance improvements of the proposed method, reinforcing its ability to outperform conventional approaches in accuracy and reliability.