Cerfacs Enter the world of high performance ...

Abstract

Computational fluid dynamics is an important tool that helps to understand natural phenomena and to design, operate or optimize complex engineering processes in industry. Today the reliability of predicting fluid flows with numerical methods depends on the usage of massive supercomputers. With the continuous increase in computational power year by year, the question of how to efficiently use computational resources becomes more and more important. Improving existing practices involves a better understanding of the underlying physical mechanisms as well as optimizing the algorithms that are used to solve them with robust and rapid numerical methods. The Lattice-Boltzmann method (LBM) is a mesoscopic approach to approximate the macroscopic equations of mass and momentum balance equations for a fluid flow.

The objective of this study is to apply this concept to large scale problems and present its capabilities in terms of physical modelling and computing efficiency. As a validation step, computational models are tested against referenced theoretical, numerical and experimental evidences over a wide range of hydrodynamic conditions from creeping to turbulent flows and granular media. Turbulent flows are multi-scale flows that required fine meshes and long simulation times to converge statistics. Special care is taken to verify the fluid-solid interface for dispersed two-phase flows. Two main setups are examined – the non-reacting, swirling flow inside an injector and a particle-laden flow around a cylinder.

Swirling flows are typical of aeronautical combustion chambers. The selected configuration is used to benchmark three different large eddy simulation solvers regarding their accuracy and computational efficiency. The obtained numerical results are compared to experimental results in terms of mean and fluctuating velocity profiles and pressure drop. The scaling, that is the code performance on a large range of processors, is characterized. Differences between several algorithmic approaches and different solvers are evaluated and commented.

Next, we focused on particle laden flows around a cylinder as generic configuration for the interaction of a dispersed phase and flow hydrodynamic instabilities. In the framework of this study the particles are fully coupled to the fluid flow and fully resolved within the computational domain. It has been shown that the viscosity of a suspension increases relative to the particle volume fraction and for a certain range of particle material and concentration, this a fairly good model of interphase coupling. This phenomenon only occurs in numerical simulations that are able to describe finite size effects for rigid bodies. Comparing global flow parameters of suspensions at different particle volume fractions and sizes have shown that these features of the flow can be obtained for an equivalent single phase fluid with effective viscosity. Starting from neutrally buoyant particles the transition to granular flow is investigated. By increasing the relative density of particles, the influence of particle inertia on the equivalent fluid prediction is investigated and the contribution of particle collisions on the drag coefficient for varying relative densities is discussed. Conclusions are drawn on the code performance and perspectives of more complex configurations.