The phase field (PF) model has been widely used for microstructure simulations during the solidification of multicomponent alloys. For multicomponent alloys, phase compositions need to be solved iteratively at every time step during phase field simulations because of the constraint of the phase equilibrium condition. The procedure requires frequent access to the thermodynamic database, which is time intensive. To accelerate the simulations, the machine learning method is used to predict the phase compositions. The machine learning method was implemented with graphics processing unit (GPU). The phase compositions of binary, ternary, and technical superalloys during solidification were predicted. A speedup ratio of 20.39 was obtained compared with the Newton iteration method for solving the phase compositions of the Al‒Cu alloy. The prediction accuracy of the proposed method was evaluated. The machine learning method was subsequently applied to dendrite growth simulations via a phase field model. The dendrite tip radius versus growth velocity was examined via an analytical model, and good agreement was achieved. Finally, equiaxed dendrite growth was simulated, and the effect of the cooling rate was analyzed, which demonstrated the capacity of the proposed method.