Output-based error estimation and mesh adaptation for unsteady turbulent flow simulations

Abstract

This paper presents a method for estimating output errors and adapting computational meshes in simulations of unsteady turbulent flows. The chaotic nature of such problems prevents a stable unsteady adjoint solution, and existing regularization techniques are costly for large simulations. The method presented foregoes the unsteady adjoint and instead relies on a field-inversion machine-learning (FIML) framework, which only requires unsteady primal solutions without full-state storage or checkpointing. The FIML model yields an adjoint for the averaged solution, which is combined with an averaged unsteady residual to obtain an output error estimate and adaptive indicator. This error estimate is shown to be accurate when the FIML model augments the original unsteady equations with corrections that are not excessively large. The unsteady residual comes from sampling fine-space residual evaluations during the unsteady simulation. A novel objective function based on an adjoint-weighted residual is presented for the field inversion to improve the ability of the FIML model to predict output errors and the domain-interior state. The localized output error drives adaptation of the mesh size and approximation order. Results for three aerodynamic problems ranging in Reynolds number demonstrate accuracy of the error estimates and efficiency of the computational meshes when compared to other adaptive strategies, including uniform and residual-based refinement.