Error-driven adaptive mesh refinement for unsteady turbulent flows in spectral-element simulations

Abstract

The simulation of turbulent flows requires high spatial resolution in potentially a priori unknown, solution-dependent locations. To achieve adaptive refinement of the mesh, we rely on error indicators. We assess the differences between an error measure relying on the local convergence properties of the numerical solution and a goal-oriented error measure based on the computation of an adjoint problem. The latter method aims at optimizing the mesh for the calculation of a predefined integral quantity, or functional of interest. This work follows on from a previous study conducted on steady flows in Offermans et al. (2020) and we extend the use of the so-called adjoint error estimator to three-dimensional, turbulent flows. They both represent a way to achieve error control and automatic mesh refinement (AMR) for the numerical approximation of the Navier–Stokes equations, with a spectral element method discretization and non-conforming h-refinement.The current study consists of running the same physical flow case on gradually finer meshes, starting from a coarse initial grid, and to compare the results and mesh refinement patterns when using both error measures. As a flow case, we consider the turbulent flow in a constricted, periodic channel, also known as the periodic hill flow, at four different Reynolds numbers: Re = 700, Re = 1400, Re = 2800 and Re = 5600. Our results show that both error measures allow for effective control of the error, but they adjust the mesh differently. Well-resolved simulations are achieved by automatically focusing refinement on the most critical regions of the domain, while significant saving in the overall number of elements is attained, compared to statically generated meshes. At all Reynolds numbers, we show that relevant physical quantities, such as mean velocity profiles and reattachment/separation points, converge well to reference literature data. At the highest Reynolds number achieved (Re = 5600), relevant quantities, i.e. reattachment and separation locations, are estimated with the same level of accuracy as the reference data while only using one-third of the degrees of freedom of the reference. Moreover, we observe distinct mesh refinement patterns for both error measures. With the spectral error indicator, the mesh resolution is more uniform and turbulent structures are more resolved within the whole domain. On the other hand, the adjoint error estimator tends to focus the refinement within a localized zone in the domain, dependent on the functional of interest, leaving large parts of the domain marginally resolved.