DNS of the Early Phase of Oblique Droplet Impact on Thin Films with FS3D

TL;DR

This study uses Direct Numerical Simulations to investigate the early crown formation during oblique droplet impacts on thin films, focusing on numerical solver efficiency and parallelization to enhance simulation performance.

Contribution

It introduces optimized multigrid solver cycles and hybrid parallelization techniques, significantly improving DNS computational efficiency for droplet impact simulations.

Findings

F- and V-cycle multigrid methods affect simulation speed

Hybrid MPI and OpenMP parallelization increases scaling and efficiency

Achieved a 12.4-fold increase in cycles per hour with new methods

Abstract

Spray impacts occur in several environmental and technical applications. The impact of droplets at different angles onto walls covered with a thin film of the same liquid can be regarded as an elementary process here. Direct Numerical Simulations (DNS) provide an important contribution to the understanding and modeling of the impact outcome, which might be associated with the formation of a crown and ejection of secondary droplets. Thus, we gain detailed information about, e.g. the flow field and shape of the interface, which are not accessible in experiments. This chapter presents a DNS study of the early crown formation mechanisms present at an oblique droplet impact on a thin film, as well as a grid study showing the resolution required to resolve the impact's details. Highly resolved simulations in large domains require continuous development of the numerical solver's efficiency. The performance of different cycles of the multigrid (MG) solver for the solution of the pressure Poisson equation was compared, and a F-cycle was added. Furthermore, we implemented a hybrid MPI and OpenMP parallelisation, which both increases the scaling limit further. Additionally, studies on the strong and weak scaling are conducted. The choice of the F- and V- cycle in the MG-solver and the additional hybrid parallelisation increased the achieved computed cycles per hour (CPH) by a factor of 12.4 compared to the formerly employed setup. The nodes efficiently usable were increased by a factor of 16. Both, the close to linear scaling regime for the strong scaling and the almost constant performance regime for the weak scaling were increased by this factor.