Accelerating phase-field simulation of multi-component alloy solidification by shallow artificial neural network

Low computing efficiency is a significant barrier in the phase-field modeling of multi-component alloys coupled with the CALPHAD (CALculation of PHAse Diagram) method. The fundamental issue is that the quasi-equilibrium thermodynamic data (QETD) required for calculating the chemical driving force must be acquired by repeatedly solving a large number of nonlinear equations. In this work, a novel method is developed to predict the QETD by a shallow neural network in a straightforward manner, so circumventing the repetitive numerical calculation of nonlinear quasi-equilibrium thermodynamic conditions. The numerical evaluation of a Ni-Cr-Al ternary alloy demonstrates that the proposed scheme can decrease the computing consuming to about 1/80 of that required by the conventional phase-field method when the computational domain size reaches the millimeter scale, while accurately reproducing the equiaxed dendritic growth and solute segregation kinetics during polycrystalline solidification. The method presented in this work will provide an effective tool for modeling the microstructure evolution of complex materials involved in practical engineering applications.