Constructing porosity database for Al−Si alloy castings through 3D cellular automata model and machine learning
A coupled three-dimensional cellular automata (CA) model was employed to predict hydrogen porosity in Al−Si alloy castings as a function of thermal boundary conditions. Simulations quantified porosity distribution across cooling rates from 0.25 to 50 °C/s at an initial hydrogen content of 3.0×10−3 mL/g, generating a comprehensive porosity defect database. Four machine learning algorithms—support vector machine (SVM), random forest (RF), K-nearest neighbors (KNN), and gradient boosting machine (GBM)—were trained and compared for each porosity characteristic to identify the optimal model. For porosity percentage prediction, the KNN model achieved a determination coefficient (R2) of 0.94, root mean square error (RMSE) of 0.035, and mean absolute error (MAE) of 0.022 on the test set. Experimental validation via optical microscopy confirmed that average equivalent porosity diameter and porosity percentage predictions fell within 20% error. The model demonstrates superior performance compared to single nonlinear function fits and other simulation approaches, offering a pathway to reduce simulation time while enhancing prediction accuracy for porosity size distribution in large casting components. The database and coupled CA-ML framework provide a robust tool for mapping porosity defects in industrial Al−Si castings, addressing a critical need for reliable quality control in automotive and aerospace applications.