EUROPEAN CONGRESS AND EXHIBITION ON ADVANCED MATERIALS AND PROCESSES - EUROMAT 2021

In the phase-field simulation of columnar dendrite growth, it is essential to treat multiple dendrites because columnar structure is formed through competitive growth between multiple dendrites. Meanwhile, the computational cost of phase-field simulation is very high due to the diffuse interface model. Thus, we developed a large-scale phase-field simulation scheme of columnar dendrite growth by introducing multiple GPUs parallel computing [S. Sakane, et al., IOP Conf. Ser. Mater. Sci. Eng., 84 (2015) 012063.]. Using the scheme, we simulated competitive growth between multiple columnar dendrites of single crystal [T. Takaki, et al., Acta Mater., 118 (2016) 230-243.], bi-crystal [T. Takaki, et al., ISIJ Int., 56 (2016) 1427-1435.], and polycrystal [T. Takaki, et al., Materialia, 1 (2018) 104-113.]. Also, we enabled permeability prediction of liquid flow in columnar dendritic structure [T. Takaki, et al., Acta Mater., 164 (2019) 237-249.]. As mentioned above, we succeeded the large-scale phase-field simulations using the multiple GPUs parallel computing. On the other hand, columnar dendrite growth simulation with large primary arm spacing is a challenging topic, and we need further efficient computational scheme.
In this study, we implement the multiple GPUs parallel computing for the adaptive mesh refinement (AMR) method with the aim of further accelerating large-scale phase-field simulations of columnar dendrite growth. Here, we also introduce the dynamic load balancing, which keeps the computational load between GPUs constant. Through simulations of directional solidification of a binary alloy, acceleration and accuracy of the implemented AMR method are evaluated. We also confirm the scalability in weak scaling test, and demonstrate the usefulness of the developed method in the simulation of columnar dendrite growth with large primary arm spacing.