Hybrid Data Driven Deep Learning Framework for Material Property Prediction

The research presents a hybrid approach of regression modeling with data-driven analysis for predicting steel's mechanical properties by analyzing the effects of composition on strength. The study fills the gap of models in accurately predicting steel's performance based on composition since traditional methods cannot fully capture complex relationships between alloying elements and material properties. Various regression models have been used for predicting material properties, such as Linear Regression, Random Forest Regression, Support Vector Regression (SVR), XGBoost Regression, and Neural Networks, and in this paper, Graph Attention Transformer Network (GAT-TransNet) is proposed. Incorporating novel graph attention into the transformer architecture model, GAT-TransNet handles complex data relationships and improves predictive accuracy. Data-driven analyses are also carried out alongside regression analysis to establish how alloying elements, such as carbon (C), manganese (Mn), and chromium (Cr), affect steel's mechanical properties strength, yield strength, hardness, and ductility. The study established that the GAT-TransNet model outperformed other regression models, with an R² score of 0.95, the lowest MAE of 1.40, and an MSE of 4.41, thus underscoring its superior predictive capability compared to existing models. Data-driven insights show that manganese hardens and increases wear resistance, while chromium enhances corrosion resistance and increases tensile strength. This has great importance for optimizing specific steel compositions for industrial applications. Combining machine learning methodologies with composition analysis, this study complements predictive modeling for steel properties with material design and promises better efficiency and targeting in steel production.