Composition-Aware SDAS Prediction in Recycled Aluminum Alloys via Physics-Informed Machine Learning Guided by Analytical Solidification Physics

The mechanical performance of secondary aluminum alloys depends on Secondary Dendrite Arm Spacing (SDAS). Commercial casting simulations accurately predict local thermal history but typically neglect the influence of compositional variability on SDAS by using fixed material constants. This study introduces a physics-informed machine learning framework to bridge macroscopic process simulation and microscopic solidification physics. A computational Design of Experiments covering 500 AlSi7 alloy variants was generated, and a theoretical SDAS ground truth was calculated using an analytical model incorporating the growth restriction factor. A Gradient Boosting Regressor surrogate was trained to predict the physics-informed SDAS from thermal and chemical inputs. The analysis reveals a solute sensitivity gap, where standard simulations misestimate SDAS by up to 20% for high-impurity batches. The surrogate model captures this variance (R2=0.95, MAE=0.24 mu m), enabling rapid, composition-specific microstructural prediction without additional simulation cost. This approach supports the reliable simulation of casting with secondary alloys, where the composition can be hardly considered constant.