Computationally efficient phase-field studies combining simulation sampling and statistical analysis

TL;DR

This paper demonstrates that decomposing large-scale phase-field simulations into smaller, parallelizable simulations can significantly reduce computational costs while maintaining accuracy, especially by focusing on precipitate density as a key parameter.

Contribution

It introduces a method combining simulation sampling and statistical analysis to efficiently perform large-scale phase-field studies without substantial accuracy loss.

Findings

Decomposition of simulation boxes accelerates computations.

Precipitate density is a key parameter controlling accuracy.

Large-scale simulations can be replaced by smaller, parallel simulations.

Abstract

The trade-off between accuracy and computational cost as a function of the size and number of simulation boxes was studied for large-scale phase-field simulations. For this purpose, a reference simulation box was incrementally partitioned. We have considered diffusion-controlled precipitation of delta prime in a model Al-Li system from the growth stage until early ripening. The results of the simulations show that decomposition of simulation box can be a valuable computational technique to accelerate simulations without substantial loss of accuracy. In the current case study, the precipitate density was found to be the key controlling parameter. For a pre-set accuracy, it turned out that large-scale simulations of the reference domain can be replaced by a combination of smaller simulations. This shortens the required simulation time and improves the memory usage of the simulation considerably, and thus substantially increases the efficiency of massive parallel computation for phase-field applications.