Machine Learning Prediction of Porosity in Low-Pressure-Cast AlSi7Mg Alloy with Varying Recycled Aluminum Content

Industrial AlSi7Mg produced by low-pressure casting was evaluated with 50–95 wt.% recycled material under two conditions: no intentional Mn correction and Mn correction targeted at Mn/Fe ≈ 0.5. The porosity dataset contained 200 observations; for 60–95 wt.% recycled material, 160 records from 16 experimental conditions also included sample-specific chemical composition. Without Mn correction, MeanPor rose from 0.160% to 6.111%, whereas Mn correction reduced porosity under most conditions. Eight hypothesis-driven candidate feature architectures and 13 model families were assessed by grouped Leave-One-Scrap-Level-Out cross-validation, with hyperparameters tuned only on training data. A second-order regression with predefined interactions using recycled material content, Mn-correction variant, and sample-specific Fe gave the lowest out-of-fold (OOF) error for MeanPor (coefficient of determination (R²) = 0.9910, root mean square error (RMSE) = 0.1691, mean absolute error (MAE) = 0.1508). The reduced Sludge Factor descriptor, SF_Cr0 = (1 × %Fe) + (2 × %Mn), also improved prediction and reached R² ≈ 0.986. SHAP (SHapley Additive exPlanations) was used to describe model attribution without causal interpretation. At 90 and 95 wt.% recycled material, Mn correction coincided with higher V values and more regular pore cross-sections. Sample-specific chemistry improved porosity prediction within the investigated process window; independent validation is required before use outside that range.