Machine learning based phase-field simulation framework for multi-component alloy solidification

The computational efficiency of the phase-field method in simulating multi-component alloy solidification is strictly limited by the iterative solution of nonlinear quasi-equilibrium thermodynamic equations. To overcome this bottleneck, this study develops a machine learning-based phase-field (MLPF) framework. We systematically evaluated several machine learning models to determine the optimal surrogate for directly predicting quasi-equilibrium thermodynamic data (QETD). The neural network (NN) model was identified as the most effective solution, offering an optimal balance between predictive accuracy, model size, and inference speed. By integrating the surrogate model into the multi-phase-field solver, simulations of Al–Zn–Mg–Cu–Zr alloy solidification were conducted. The MLPF framework achieved a computational acceleration of approximately 300 times for millimeter-scale domains while maintaining high physical fidelity relative to conventional methods. Furthermore, the simulation elucidated the coupling mechanism between cooling rate and microsegregation. While high cooling rates induce severe solute pile-up in the liquid due to diffusion limitations, they simultaneously promote solute trapping within the solid matrix. This mechanism effectively alleviates terminal solidification segregation, providing new insights into the non-equilibrium solidification of complex aluminum alloys.