International Scientific Journals
of Scientific Technical Union of Mechanical Engineering "Industry 4.0"

This study presents a machine learning framework for predicting the mechanical properties of 2xxx series Al-Cu alloys (2024, 2219, 2524) across 11 temper conditions. Monte Carlo augmentation generated 8,800 synthetic samples from compositional specification ranges of a literature-mined dataset. Three regression models, Random Forest, Gradient Boosting, and SVR-RBF were evaluated via 5-fold cross-validation (CV) to predict ultimate tensile strength (UTS), yield strength (YS), and elongation. All models achieved coefficient of determination R² > 0.991, with Mean Absolute Error MAE ≤ 7.4 MPa for UTS, ≤ 5.4 MPa for YS, and ≤ 0.51% for elongation. Feature importance analysis revealed that temper condition encoding dominated predictions (>75% importance), while individual compositional features contributed <5% each. The high predictive accuracy reflects the effectiveness of the augmentation scheme in capturing withingroup property–composition–temper relationships, though generalization to unseen alloy–temper conditions remains to be validated. The results illustrate the potential of combining corpus-mined data with Monte Carlo augmentation for rapid alloy property screening.

This study develops a machine learning framework for predicting the mechanical properties of 6xxx series Al-Mg-Si alloys (6061, 6063, 6082) across seven temper conditions. A hybrid dataset of 860 real measurement-based and 4,200 Monte Carlo augmented samples was generated from a literature-mined dataset. Random Forest (RF), Gradient Boosting (GBR), and Multilayer Perceptron (MLP) models were evaluated via 5-fold cross-validation (CV). RF achieved the best or comparable accuracy: coefficient of determination R² = 0.80 (ultimate tensile strength), 0.92 (yield strength), and 0.83 (elongation). Feature importance analysis showed that alloy type and temper encodings dominated predictions (>90% combined), while individual compositional features contributed <3% a result partly attributable to the augmentation strategy, which decoupled measured compositions from their corresponding property values. Learning curve analysis confirmed model convergence above 2,000 samples.