Machine Learning Prediction of Density for Binary Mg-Containing Phases

Magnesium (Mg) alloys face a critical challenge in balancing performance optimization and unintended density increases caused by high-density secondary phases. To address this, machine learning was employed to predict the density and volume of Mg-containing binary phases, aiming to guide lightweight alloy design. Using 211 experimentally observed data points, five machine learning (ML) algorithms—Random Forest (RF), Support Vector Machine (SVM), Artificial Neural Network (ANN), K-Nearest Neighbors (KNN), and Bayesian Ridge (Bayes)—were trained and tested. Quantitative results showed that RF achieved exceptional performance in volume prediction, with a testing coefficient of determination (R²) exceeding 0.96 and a mean absolute error (MAE) of 41.0 ų, while SVM outperformed others in density prediction with a testing R² of 0.885 and MAE of 0.421 g/cm³. Feature importance analysis revealed that atomic count is the primary determinant of phase volume, whereas density prediction depends on the synergistic interaction of relative atomic mass and stoichiometric ratio, as further validated by SHapley Additive exPlanations (SHAP) analysis. This work establishes a physics-informed predictive model that accelerates the development of lightweight Mg alloys by mitigating high-density secondary phases, and can be extended to other alloy systems.