Integrating phase-field simulations with deep learning to predict properties from pressure-tailored microstructures in Al-Si alloys

High-pressure solidification offers a promising route for tailoring microstructures and enhancing alloy properties, yet a quantitative understanding of the microstructure evolution and its consequences for mechanical properties under such extreme conditions remains scarce. Thus, this study establishes a computational framework that integrates phase-field simulations with deep learning to directly bridge the processing, microstructure, and mechanical properties of high-pressure solidified Al-20Si (in wt.%) alloy, aiming at clarifying the pressure-tailored microstructural evolution and enabling microstructure-informed property prediction. The quantitative phase-field simulations, incorporating pressure-dependent thermodynamic descriptions and atomic mobilities, were performed to model the entire solidification process across a pressure range of 1 bar to 3.0 GPa. The results reveal a pressure-induced transition of the primary (Si) to (Al), a significant refinement of the eutectic spacing, and a shift in eutectic growth mode from cooperative to divorced. To directly link microstructure to properties, a hybrid deep neural network combining a convolutional neural network encoder and an artificial neural network regressor was developed, enabling an end-to-end prediction of ultimate tensile strength (UTS), yield strength (YS), and hardness from phase-field generated microstructures. The model exhibits high predictive accuracy, with coefficients of determination reaching 0.94, 0.85, and 0.98 for UTS, YS, and hardness, respectively, alongside low mean absolute errors. This integrated framework provides a generalizable approach for microstructure-informed property prediction, extendable to a broad range of materials systems.