Accelerating three-dimensional phase-field simulations via deep learning approaches

Phase-field modeling (PFM) is a powerful but computationally expensive technique for simulating three-dimensional (3D) microstructure evolutions. Very recently, integrating machine learning into phase-field simulations provides a promising way to reduce calculation time remarkably. In this study, we propose a deep learning model that combines a convolutional autoencoder with a deep operator network to predict 3D microstructure evolution by using 2D slices of the 3D system. It is found that the deep learning model can shorten the calculation time from 37 min to 3 s after the initial training, while skipping 5-time steps, and reduce the phase-field simulation time by 31% in entire calculation of the evolution process. Interestingly, this model achieves good accuracy in predicting 3D microstructures by utilizing only 2D information. This work demonstrates the efficiency of machine learning in accelerating phase-field simulations while maintaining high accuracy and promotes the application of PFM in fundamental studies.