Robust Adaptive High-Order RANS Methods MASSACHUSETTS INGTWFUTE OF TECHNOLOGY by Jun Kudo OCT 0 9 201 Sc.B., Brown University (2010) LIBRARIES Submitted to the School of Engineering in partial fulfillment of the requirements for the degree of Master of Science in Computation for Design and Optimization at the MASSACHUSETTS INSTITUTE OF TECHNOLOGY September 2014 @ Massachusetts Institute of Technology 2014. All rights reserved. Signature redacted Author ............................... Scho$l of Engineering June 20, 2014 Signature redacted Certified by............................. David L. DAofal Professor of Aeronautics and Astronautics Thesis Supervisor Signature redacted Accepted by ............................... adjiconstantinou Ni Professor of Mechanical Engineering Co-Director, Computation for Design and Optimization 2 Robust Adaptive High-Order RANS Methods by Jun Kudo Submitted to the School of Engineering on June 20, 2014, in partial fulfillment of the requirements for the degree of Master of Science in Computation for Design and Optimization Abstract The ability to achieve accurate predictions of turbulent flow over arbitrarily complex geometries proves critical in the advancement of aerospace design. However, quantitatively accurate results from modern Computational Fluid Dynamics (CFD) tools are often accompanied by intractably high computational expenses and are significantly hindered by the lack of automation. In particular, the generation of a suitable mesh for a given flow problem often requires significant amounts of human input. This process however encounters difficulties for turbulent flows which exhibit a wide range of length scales that must be spatially resolved for an accurate solution. Higherorder adaptive methods are attractive candidates for addressing these deficiencies by promising accurate solutions at a reduced cost in a highly automated fashion. However, these methods in general are still not robust enough for industrial applications and significant advances must be made before the true realization of robust automated three-dimensional turbulent CFD. This thesis presents steps towards this realization of a robust high-order adaptive Reynolds-Averaged Navier-Stokes (RANS) method for the analysis of turbulent flows. Specifically, a discontinuous Galerkin (DG) discretization of the RANS equations and an output-based error estimation with an associated mesh adaptation algorithm is demonstrated. To improve the robustness associated with the RANS discretization, modifications to the negative continuation of the Spalart-Allmaras turbulence model are reviewed and numerically demonstrated on a test case. An existing metric-based adaptation framework is adopted and modified to improve the procedure's global convergence behavior. The resulting discretization and modified adaptation procedure is then applied to two-dimensional and three-dimensional turbulent flows to demonstrate the overall capability of the method. Thesis Supervisor: David L. Darmofal Title: Professor of Aeronautics and Astronautics 3 4 Acknowledgments I would like to thank all those who have made this thesis possible. I would first like to thank my adviser, Prof. David Darmofal, for giving me the opportunity to work with him and for all his encouragement throughout my graduate study. I would also like to thank Dr. Steven Allmaras for his help and insight along the way. In addition, I would like to recognize Marshall for teaching me various useful tools and for his unwavering dedication to our software maintenance. This thesis would not have been possible without the help of the past and present generations of the ProjectX team. I would like to thank them all for their many contributions that formed the foundation of ProjectX that enabled this work. I would specifically like to thank: Huafei, for helping me get started; Masa, for laying down the foundation of the adaptive framework; Josh, for helping me set up my threedimensional meshes and geometries; my office mates, Steven and Phil, for always being available for discussion on any topic; and Carlee, Jeff, Savi, and Yixuan, for bringing in some much needed fresh energy into the lab. On a more personal note, I would also like to thank all of my friends that have helped make my life during my graduate studies more fun than it might otherwise have been. Special thanks are directed towards Dan and Kenny who helped me stay sane when the going got tough and Allison who acted like my personal graduate adviser when I became frustrated. I would like to send my most sincere thanks to Melissa for all her love and support over the past two years. I do not know how to fully express my gratitude and am looking forward to spending more time together in Boston in the upcoming years. Finally, I would like to acknowledge the financial support of the DOD NDSEG Fellowship program and NASA (NASA Cooperative Agreement #NNX12AJ75A, nical monitor Dr. Harold Atkins). 5 tech- 6 Contents 1 Introduction 13 1.1 M otivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 1.2 O bjective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 1.3 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 1.3.1 High-Order Methods and Discontinuous Galerkin Methods . . 15 1.3.2 Error Estimate and Adaptation . . . . . . . . . . . . . . . . . 16 Thesis Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 1.4 2 Discretization of the RANS Equations 21 2.1 The RANS Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.2 The SA Turbulence Model . . . . . . . . . . . . . . . . . . . . . . . . 22 2.2.1 Baseline Model . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.2.2 Modifications to Baseline Model . . . . . . . . . . . . . . . . . 24 Spatial Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2.3.1 Inviscid Discretization . . . . . . . . . . . . . . . . . . . . . . 28 2.3.2 Viscous Discretization . . . . . . . . . . . . . . . . . . . . . . 28 2.3.3 Source Discretization . . . . . . . . . . . . . . . . . . . . . . . 29 Temporal Discretization and Solution Technique . . . . . . . . . . . . 29 2.4.1 30 2.3 2.4 3 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . Output Error Estimation and Continuous Mesh Framework 33 3.1 Error Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 3.1.1 34 Dual-Weighted Residual Method and Localization . . . . . . . 7 3.2 4 Continuous Mesh Framework . . . . . . . . . . . . . . . . . . . . . . 35 3.2.1 Metric-Conforming Meshes . . . . . . . . . . . . . . . . . . . . 36 3.2.2 Mesh-Conforming Metric Fields . . . . . . . . . . . . . . . . . 37 Metric Field Optimization using Local Error Sampling and Synthesis Framework 4.1 4.2 4.3 5 6 39 Model Definition . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . 39 4.1.1 Continuous Relaxation . . . . . . . . . . . . . . . . . . . . . . 39 4.1.2 Metric Manipulation Framework . . . . . . . . . . . . . . . . . 41 4.1.3 Cost M odel . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 4.1.4 Surrogate Error Model from Local Error Sampling and Synthesis 42 Optimization of Surrogate Model . . . . . . . . . . . . . . . . . . . . 44 4.2.1 Heuristic Optimization of Surrogate Model . . . . . . . . . . . 45 4.2.2 Modifications to Metric Optimization Procedure . . . . . . . . 47 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 4.3.1 r'-Type Corner Singularity 51 4.3.2 RAE 2822 Transonic RANS-SA 4.3.3 Supersonic flow over NACA 0012 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 . . . . . . . . . . . . . . . . 55 Turbulent Aerodynamic Problems 63 5.1 Three-element MDA 30P-30N, Subsonic . . . . . . . . . . . . . . . . 63 5.2 3D Zero Pressure Gradient Flat Plate, Subsonic . . . . . . . . . . . . 66 5.3 3D Duct, Subsonic 70 . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusion 77 6.1 Sum m ary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 6.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 A Adaptation From Under-resolved Meshes A.1 Test Cases and Initial Meshes A.2 Solution on Initial Meshes A.3 Adaptive Results 81 . . . . . . . . . . . . . . . . . . . . . . 81 . . . . . . . . . . . . . . . . . . . . . . . . 82 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 8 List of Figures 1-1 General Framework for PDE Solver with Mesh Adaptation . . . . . . 16 2-1 Comparison of Modified Diffusion Coefficients 27 2-2 Comparison of convergence for flow over a NACA 4412 using negative . . . . . . . . . . . . . v variants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-1 Element size h versus the distance of the element centroid from the corner for optimized meshes for r' singularity problem with a = 2/3 . 4-2 53 Comparison of error indicator and degrees of freedom history for transonic flow over a RAE 2822 Airfoil 4-3 31 . . . . . . . . . . . . . . . . . . . 54 Cumulative computation time taken to obtain primal solution and error indicator estimate on a given iterate's mesh for transonic flow over a RAE 2822 airfoil 4-4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mesh adaptation history for transonic flow over a RAE 2822 Airfoil using the heuristic metric optimization method 4-5 . . . . . . . . . . . . 58 Adaptation history for supersonic flow over a NACA 0012 Airfoil using the gradient-based optimization method 4-8 57 Adaptation history for supersonic flow over a NACA 0012 Airfoil using the heuristic optimization method . . . . . . . . . . . . . . . . . . . . 4-7 56 Mesh adaptation history for transonic flow over a RAE 2822 Airfoil using the gradient-based metric optimization method . . . . . . . . . 4-6 55 . . . . . . . . . . . . . . . . 58 Select adapted meshes (iterations 17 and 18) and respective Mach number distributions generated with the heuristic optimization method, NACA0012 supersonic flow. . . . . . . . . . . . . . . . . . . . . . . . 9 59 4-9 Select adapted meshes (iterations 19 and 20) and respective Mach number distributions generated with the heuristic optimization method, NACA0012 supersonic flow. . . . . . . . . . . . . . . . . . . . . . . . 60 4-10 Select adapted meshes and respective Mach number distributions generated with the gradient-based optimization method, NACA0012 supersonic flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 5-1 Initial mesh, Three-element MDA flow. . . . . . . . . . . . . . . . . . 63 5-2 Convergence of drag output for the three-element MDA flow; reference value obtained from p = 3, 5-3 doftarget = 250k adapted solution . . . . . Comparison of adapted p = 1 and p = 2 meshes with similar error levels, Three-element MDA flow . . . . . . . . . . . . . . . . . . . . . 5-4 . . . . . . . . . . 65 p = 2, doftarget = 120k adapted mesh and respective Mach number distribution, overview, Three-element MDA flow. 5-6 65 Pressure coefficient comparison of adapted p = 1 and p = 2 meshes with similar error levels, Three-element MDA flow. 5-5 64 p = 2, doftarget . . . . . . . . . . . 66 = 120k adapted mesh and respective Mach number distribution, far field, Three-element MDA flow. . . . . . . . . . . . . 67 5-7 Initial structured mesh, 3D flat plate flow. 67 5-8 Convergence of drag output for the 3D flat plate flow; reference value 5-9 . . . . . . . . . . . . . . . obtained from two-dimensional simulations . . . . . . . . . . . . . . . 68 P=2, 25k dof optimized mesh, 3D flat plate flow. 69 . . . . . . . . . . . 5-10 Mach number distribution of 2D center slice, initial mesh, 3D flat plate flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 5-11 Mach number distribution of 2D center slice, P=2, 25k dof optimized mesh, 3D flat plate flow . . . . . . . . . . . . . . . . . . . . . . . . . 70 5-12 Initial structured mesh, subsonic duct flow. . . . . . . . . . . . . . . . 71 5-13 Drag adaptation history for p = 2, subsonic duct flow . . . . . . . . . 72 5-14 P=2, 100k dof optimized mesh, subsonic duct flow. 73 10 . . . . . . . . . . 5-15 Mach number distribution, 2D view along duct at X = 0.25, subsonic duct flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 5-16 Mach number distribution, 2D cross sectional view at outflow, 2.0 units downstream from leading edge, subsonic duct flow . . . . . . . . . . . A-1 Initial isotropic mesh with no boundary layer resolution - 460 Elements 75 82 - A-2 Initial anisotropic mesh with significant boundary layer resolution 460 Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 A-3 Initial pu, T wall normal profile, 1.5 units downstream of the leading edge, isotropic mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 A-4 Initial pu, T wall normal profile, 1.5 units downstream of the leading edge, anisotropic mesh . . . . . . . . . . . . . . . . . . . . . . . . . . 84 A-5 Drag adaptation history for turbulent boundary layer flow starting with isotropic mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84 A-6 DOF adaptation history for turbulent boundary layer flow starting with isotropic mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 A-7 Drag adaptation history for turbulent boundary layer flow starting with anisotropic m esh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 A-8 DOF adaptation history for turbulent boundary layer flow starting with anisotropic mesh . . . . . . . . . . . . . . . . . . . . . . . . . . 86 A-9 Heat flux adaptation history for scalar boundary layer flow starting with isotropic mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 A-10 DOF adaptation history for scalar boundary layer flow starting with isotropic mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 A-11 Heat flux adaptation history for scalar boundary layer flow starting with anisotropic mesh . . . . . . . . . . . . . . . . . . . . . . . . . . 87 A-12 DOF adaptation history for scalar boundary layer flow starting with anisotropic mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 88 12 Chapter 1 Introduction 1.1 Motivation With the evolution of numerical algorithms and substantial advancements in computational power, the presence of Computational Fluid Dynamics (CFD) tools throughout industry and academia has steadily increased in the past few decades. These numerical tools promise the ability to rapidly and accurately simulate flows that may otherwise be prohibitively expensive to test experimentally. Despite this potential, many prevalent CFD software packages lack efficiency when high accuracy simulations are required and require a significant amount of user involvement for mesh generation. Aerodynamic flow, and in particular turbulent flow, can exhibit features with a wide range of lengths scales and singularities that can potentially interfere with the realization of accurate and robust CFD. The American Institute of Aeronautics and Astronautics (AIAA) organizes drag prediction workshops (DPW) with the purpose of evaluating the capability of current state of the art CFD for turbulence modeling on industrially relevant geometries and flow conditions. The starting meshes for the geometries of interest are generated based on industry's best practices and these meshes are uniformly refined in an attempt to achieve mesh independence. The results from the workshops display that the current CFD capabilities are insufficient for engineering applications; the uncertainties associated with the drag results are unacceptably high and to properly reduce the 13 spread of the results, the starting meshes must be uniformly refined to levels that make the simulations intractably large for routine analysis [32, 30, 52, 53]. Further troubling are the discoveries presented by Mavriplis [36] who demonstrated that even for very large grids, asymptotic results appear to be different for different families of self-similar grids. Adequate resolution of the relevant flow features by uniform refinement proves difficult when the flow exhibits a wide range of scales as seen in high-speed turbulent flows. Therefore, even for very large best practice meshes, the flow solutions may contain significant discretization errors if important regions of the flow are not well resolved. These results demonstrate that there is a significant need for improvements in the automation and efficiency of the current CFD methods. Spatial discretization error can significantly contribute to the inaccuracies in turbulent flow simulations and worse yet, can prove to be very elusive to even expert grid generators. Global refinement can potentially reduce this error but proves inefficient due to the very large grid sizes required to resolve the relevant flow features. High-order, adaptive RANS methods present an attractive path to alleviate these issues by automatically and efficiently resolving turbulent flows. 1.2 Objective The objective of this work is to develop a robust high-order, adaptive method for the simulation of high-speed turbulent flows and to demonstrate the performance of this method on two-dimensional and three-dimensional aerodynamic cases. 14 1.3 1.3.1 Background High-Order Methods and Discontinuous Galerkin Methods The goal of high-order methods is to achieve higher fidelity solutions at a lower cost. Generally, a discretization method has an error convergence behavior of E oc 0(h') where E is a measure of error, h is a measure of the mesh size, and r is the error convergence rate. In the aeronautical industry, the most prevalent methods for simulating flows are second-order finite volume methods (FVM) which exhibit a convergence rate of r = 2. However, as accuracy requirements become more stringent, second-order accurate methods may prove to be insufficient as they begin to require intractable amounts of degrees of freedom. High-order methods attempt to alleviate this issue by improving the simulation efficiency by increasing the convergence rate r. In this work, high-order methods are those that achieve r > 2 (for L 2 error). In finite-volume and finite-difference schemes, higher order accuracy is generally achieved by extending the approximation stencil. However, this extension introduces many complications including difficulties in parallelization and boundary condition treatment and the need for time-integration methods with stronger stability properties. Finite element schemes on the other hand provide a particularly attractive framework for achieving high-order convergence. For these schemes, the convergence rate r can be improved by increasing the degree of the basis polynomials. The methods are also naturally applicable to unstructured meshes which are useful for the tessellation of more complex geometries of interest. While finite element methods offer a conceptually simple path to high-order accuracy, the continuous Galerkin method is unstable for use on convection-dominated problems. As such, in this work, a discon. tinuous Galerkin (DG) finite element method is used The first DG method was introduced by Reed and Hill for scalar hyperbolic equations [44] in 1973. The error analysis for this work was later provided by LeSaint 15 and Raviart [31] and Richter [45]. This DG method was then extended to nonlinear hyperbolic problems by Chavent and Salzano [14] by incorporation of the Godunov's flux. Cockbur, Shu, and co-authors [17, 16, 15, 18] extended the method to non-linear systems of hyperbolic equations by combining the DG spatial discretization with a Runge-Kutta explicit time integration. The development of the DG methods for elliptic problems began with the interior penalty methods by Arnold [3] and Wheeler [57]. More recently, Bassi and Rebay developed the so-called BRI [6] and later the BR2 [7] DG discretization of the diffusive operator. 1.3.2 Error Estimate and Adaptation Mesh adaptation methods present an attractive alternative to current prevalent CFD practices in which a mesh is painstakingly created by empirical best practices. Adaptation promises to significantly reduce the amount of human intervention to produce a more reliable output prediction. These methods fundamentally rely on a definition of an error indicator which localizes and identifies the regions that need discretization modifications. With these error indicators, the process attempts to reduce the output error estimates via the adaptation machinery. The generalized adaptive framework can be seen in Figure 1-1. Problem Compute flow Estimate ,definition and outputs output error 0 Outputs o Error Estimate, Adapt mesh to control error Figure 1-1: General Framework for PDE Solver with Mesh Adaptation The process is initialized with a problem definition which involves an initial mesh, relevant boundary and initial conditions, and an output function of interest. With these inputs, the PDE is solved on the current grid and an output error is estimated. If the error is larger than a provided tolerance, an adaptation algorithm is used 16 to generate a new mesh according to the localized error estimates. The process is then repeated on this adapted mesh and continued until the target error tolerance is achieved. Error Estimation Error estimation forms the foundation for any adaptation framework by properly identifying the elements with large error contributions. Various a posteriori error estimation techniques have been developed in an attempt to best achieve these goals. These methods, which include estimates based on gradients [4], energy [1], and interpolation error [21], do not generally work for hyperbolic problems where upstream perturbations and errors can pollute the solution downstream [50]. For these convection dominated flows, output-based error estimation techniques which explicitly incorporate the adjoint associated with the output have been developed. The adjoint (or dual) solution represents the sensitivity of an output with respect to residual perturbations and as such, helps to identify the flow regimes that are critical for accurate prediction of the output. In this work, the dual-weighted residual (DWR) method proposed by Becker and Rannacher [9, 10] is used. Adaptation Given a measure of error, the goal of the adaptation procedure is to modify the current discretization to decrease the given output error estimate. For finite element methods, these modifications can largely be classified into three types: h-adaptation, p-adaptation, and hp-adaptation. In h-adaptation, the elements' shape and size are modified; with this process, elements can be moved, sub-divided or agglomerated to achieve different spatial distributions of shape and size. For h-adaptation methods, the size and orientation information of each element is often formulated as a metric tensor field [23, 55]. In p-adaptation, the polynomial approximation order of the elements is modified while keeping a constant triangulation. Although p-adaptation has been shown to be more efficient for sufficiently smooth flows, these methods are not particularly 17 well-suited for triangulations that currently contain insufficient resolution of the flow. The hp-adaptation methods modify both the triangulation and the polynomial approximation order and is attractive in that these methods can potentially combine the benefits of both h and p adaptation. However, the choice between resizing the element and locally changing the scheme's discretization order is far from trivial and has been the focus of much research [28, 48, 13]. For simplicity and with the goal of automatic mesh generation in mind, this work focuses solely on h-adaptation techniques. Anisotropic h-adaptation methods were originally largely driven by estimates of the directional interpolation error of a representative scalar. Peraire et al. [42] determined the desired mesh space based on estimating the Hessian of the density field. Similarly, Venditti and Darmofal [55] used the Hessian of the Mach number coupled with the DWR method to construct an anisotropic adaptation algorithm for the Navier-Stokes equations. Fidkowski [24] later generalized Venditti's approach to higher-order discretizations based on high-order derivatives of the Mach number. While shown to be successful, these adaptation algorithms heuristically assume that the mesh anisotropy is fundamentally governed by the directional interpolation error of one scalar quantity and do not take into account the anisotropy of the adjoint. More recent research has sought to eliminate this heuristic assumption by focusing on anisotropic adaptation algorithms that more directly target the output error. Formaggia [26, 25] used an output-based a posteriori error analysis combined with Hessian-based interpolation error estimates to create an indicator that includes the anisotropy of the element. Park [40] directly targeted the output error through local mesh operators. Yano and Darmofal [58] proposed the Mesh Optimization via Error Sampling and Synthesis (MOESS) algorithm, which performs adaptation by solving a continuous constrained optimization problem that attempts to obtain a metric field that minimizes the error estimate. In this thesis, we use Yano's MOESS framework with modifications to the optimization algorithm. 18 1.4 Thesis Overview This thesis describes the development of a high-order, h-adaptive, DG discretization for the RANS equations coupled with the negative SA turbulence model. The thesis makes the following contributions: " Demonstration of the negative SA turbulence model as described in [2] in a DG setting " Modifications to the metric optimization step of the MOESS algorithm for increased robustness in the adaptation procedure " Application of the adaptation algorithm to high-order discretizations of the RANS equations for 2D and 3D turbulent flows Chapter 2 begins with a review of the RANS equations and the SA turbulence model. This chapter also details several modifications to the SA model recommended by Allmaras [2] to increase the general robustness of its discretization. The chapter concludes with a detailed description of the discretization of the RANS-SA system and with a test case that demonstrates the increased robustness of the modified negative SA model. Chapter 3 provides a review of the output error estimation and continuous mesh framework as background for the adaptation algorithm used in this work. Chapter 4 reviews the MOESS framework and associated mechanisms as introduced by Yano [58] and details modifications made to the metric optimization portion of the original algorithm. Chapter 4 concludes with a demonstration of the modified optimization on representative model problems with comparisons to the original heuristic optimization procedure. Chapter 5 concludes with multiple two-dimensional and three-dimensional turbulent test cases demonstrating the performance of the adaptive algorithm. Chapter 6 ends with conclusions and suggestions for future work. 19 20 Chapter 2 Discretization of the RANS Equations This chapter describes the RANS equations, the baseline and negative SA turbulence models, and the discretization of the entire coupled system. 2.1 The RANS Equations The RANS equations are derived by applying Reynolds decomposition to the instantaneous compressible Navier-Stokes equations and then time-averaging the resulting equation set. Given a flow variable, f(x, t), we define the time averaging (Reynolds-averaging) procedure, f(x), as f(x) = lim T- oo Sto+ tO f (x, t) (2.1) The flow variables can then decomposed into mean and fluctuating parts, e.g. the density can we written as: + P P= (2.2) where p' is the fluctuating part. For compressible flows, we also define a convenient density-weighted time averag21 ing (Favre-averaging) procedure, f(x), as - f(x) rlim p(x, t)f (x, t)dt P -oto, (2.3) Applying the time averaging procedure to the compressible Navier-Stokes equations, we get the form of the RANS equations used in this work: ( =0 ai )= 2(p +pt)axj a- Xik , (2.5) ~-[ = axj [cp Pr + a Prt axj (2.4) 3axx (lk + a axj (2.6) 2i(p + pt) s-82 - - at (Pii)1i ax3 at a+ axi 3 aXk 'ij where p is the density, ui is the component of velocity in the i direction, p is the pressure, e is the internal energy, h is the enthalpy, T is the temperature, sij = 1 Ou+ is the strain rate tensor, p is the dynamic viscosity, pt is the dynamic eddy viscosity, Pr is the Prandtl number, and Prt is the turbulent Prandtl number. In these equations, the Reynolds stress tensor that emerges from time-averaging the flow equations is approximated using the Boussinesq eddy viscosity assumption. The Boussinesq approximation relates the Reynolds stress to the mean flow viscous stress tensor and introduces an additional unknown in the form of the eddy viscosity, pt. Since equations (2.4) through (2.6) contain an additional unknown in the form of this eddy viscosity, pt, the flow equations cannot be solved without an additional equation for closure. 2.2 The SA Turbulence Model The RANS system can be completed with an addition of a turbulence model that describes the transport of the eddy viscosity, pt. 22 In this work, the SA turbulence model is used. For reference, we present the baseline model as given in [49] and then proceed to illustrate the negative variations [38, 2] used and compared in this work. 2.2.1 Baseline Model The baseline model takes a form of a transport equation for a working variable, v, which is related to the eddy viscosity by x3 , x 3 1= where v = p/p is the kinematic viscosity. x (2.7) - 1t = pufv 1 The SA working variable i; obeys the transport equation -D+T +- DtP 1 [V.- ((v + 7)Vi) +cb2 (Vi' 2 ] (2.8) where the production P, wall destruction D and trip term T are P = Cb1(I - ft2) S, D= (c1f - ft2 T [] = ft1(aU)2 (2.9) S is the modified vorticity and is given as V fv2, K2d2 The function fw j v2 1 x + xfVi (2.10) is F f g + C6 s g = r + cw2(r' - r), S2d2 (2.11) The trip and laminar suppression terms are [d 2 + gd2] 23 ft2 = ct3 exp (-ct4X 2 ) fti = ctigt exp (-ct2 IL (2.12) where d is the distance to nearest wall, 0.41, Cw1 1.2,c 4 = Cbl/ 2 Cbl + (1 + Cb2)/O, c,2 = 0.3, = Cw3 = 0.1355, - = 2/3, Cb2 = 0.622, K = 2, cui = 7.1, cti = 1, Ct2 = 2, ct3 = = 0.5. An equivalent compressible conservation form can be constructed by combining the SA transport equation with the mass conservation equation, a (p ;)+ -(pUV) = p(P- D +T)+ 1 C ]b 2 (+)p-V21 V -[p(v + D)V~] v) + ; p T (2.13) In this work, all cases are run fully turbulent with no effort to model transition and as such, the trip terms are omitted in our discretization (T = 0). 2.2.2 Modifications to Baseline Model The discretization of the SA equation given in (2.13) can permit non-physical solutions that must be handled. Here we present the modifications proposed in [2] to handle these situations and the final resulting conservation form used in this work. Modified Vorticity Physically, the modified vorticity S S + 9 should always be a positive value such that the production term is always a positive quantity. However, in a discrete setting, the baseline modified vorticity can become zero or negative because the fv2 closure function is negative over a range of X and this numerical phenomena can cause robustness problems. To remedy this possibility, a modified form of S is used in this work, I ~S +9 3 ;>-c = S + _ )S-S' S 2S~c+C,,,S (c. 3-2c. 2 2 s >_ -+C2S < -C,2S (2.14) with c, 2 = 0.7 and cv3 = 0.9. This function results in a form of S that is identical to the original formulation for 5 > 0.3S, is C 1 continuous and positive for all non-zero S. 24 Negative iY Model Analytically, the exact solution of (2.13) is non-negative given non-negative boundary and initial conditions. However, this property is not always obtained on coarse grids and transient states where the discrete solution may exhibit negative turbulence solutions. To ensure that the eddy viscosity always obtains a non-negative value, the rela- tionship in (2.7) is modified to be >0 {Pvfv1i 0.0 (2.15) < 0 0, Since negative i values can emerge on these discrete solutions, a continuation of the SA equation for negative E; must be constructed to deal with these potential undershoots. The negative SA model used in this work is given by DiDDE =P. -Dn 1 + T+ - [V - ((V + Lfn)WD) + Cb2 (V ;)2] (.6 where the modified production P.,, is given by Cb(1 ft 2 ) V Pn = 1(Cb1 - ct 3)S > 0 (2.17) L; 5 0, the modified destruction Dn is given by Dn (c,1.-%f2) (Cwifw - ft2 ) 'j 2 L; > 0 (2.18) and the diffusion modification term fn is given as 1.0 fn = 0 0 25 (> (2.19) with ci = 16. This analytic continuation is C 1 continuous with respect to V at V = 0. Furthermore, the model ensures that the original positive SA model is unchanged for V > 0 and forces negative L transients back towards zero. The model is also constructed with the goal of mitigating nonlinearity. Again, an equivalent compressible conservation form can be constructed by combining the equation (2.16) with the mass conservation equation, C9 -(p)+V t 1 - (puV) = p(P. - D,)[+ - Cb. [V]+ a where the modified diffusion coefficient, n = pt(1 + (V)2 21q- -(v+ a f,)Vp-VV (2.20) xfn) Comparison of Negative SA Models We highlight the key differences between the model given in (2.20) and the negative extension given in [38]. This earlier variant of the present formulation presents the continuation in the same form as (2.20) and uses the same modified definitions of S and pt given in (2.14) and (2.15) respectively. The only differences lie in the negative extensions of production term P, and the modified diffusion coefficient q. The variant provided in Oliver's work defines the production term as P C1(1 - f12) cb1SLgn >0 (2.21) V < 0, where 1000X22 gn = 1 - (2.22) 1+ x and defines the modified diffusion coefficient as {q tt (I X) V(2.23) p(T + X +oiX2) u < 0, The modified diffusion coefficients are plotted in Figure 2-1 26 200 - Negative SA --- Oliver SA 150- - 100 - 0 -0 -15 -10 -5 0 5 4u Figure 2-1: Comparison of Modified Diffusion Coefficients at 2.3 Spatial Discretization We can re-write RANS equations coupled with the SA turbulence equation as a general system of conservation laws of the form OU+ V - Tc""(u) - V - (A(u)Vu) = S(u, Vu) (2.24) where u = [p, pui, pE, pL;]T is the state vector, Te"" is the inviscid flux, A(u)Vu is the viscous flux, and S is the source term. We seek a solution in a finite-dimensional approximation space Vhp defined on a triangulation Th consisting of non-overlapping elements r, of characteristic size h of the domain Q. Formally, the function space Vhp is defined as, V P = fv E [L (Q)]''I oVC fq E [PP(ref)]', VK E T} (2.25) where r is the state rank, PP denotes the space of polynomials of order p and f,, denotes the q-th degree polynomial mapping from the reference element to the physical element K. The weak form associated with the DG approximation of the conservation 27 laws now follows as, find uh,p(ti) Vhp E E Vh,p such that + lRh,p(Uhp, Vh,p) = 0, (2.26) VVhp E V where Rh,p consists of convective, diffusive, and source contributions JZh,p (Uh,p, Vh,p 2.3.1 conv(uh,p, Vh,p) p (Uh,1 (Uhp, Vh,p) ERh1 purce (Uh,p, Vh,p) (2 Inviscid Discretization The discretization of the inviscid terms is given by p"(w, zjvvT.Fcn"(w)+ v) =- (v+ _ T- H(w+, w-, i+) vT7b.-i KETh (2.28) where (.)+ and (.) are values on opposite sides of a face f, n+ is the normal vector pointing from + to -, H is the numerical inviscid flux for an interior face, and yb is the numerical inviscid boundary flux. 1i and aQ are the set of interior and boundary faces, respectively. In this work, the interior face numerical flux function uses Roe's approximate Riemann solver [46] of the form W- + 2yonvw+ oe (W H(w+, w-, n+)= 1~,,l++ -. ire(W+) + n- - co""(w-))+ |A"'+ - y +-I W (2.29) where ARoe is the flux Jacobian matrix computed about the Roe's mean state. The boundary flux state Ub 2.3.2 Fb is in general a function of the interior state w+ and the boundary which itself is a function of w+ and the prescribed boundary conditions. Viscous Discretization The viscous terms are discretized using the second method of Bassi and Rebay [7, 8]. To simplify the notation, we define the jump, [1, and average, {-}, operators for scalar 28 ~) s and vector F quantities as: s (+_ ) [S+ (g+ +s-~) [N = (V+-' 6 6= I(-+ 2 + V'-) - - 2 The Bassi and Rebay viscous flux can now written as Rdiff(w v) Z j Vv T - (A(w)Vw) KETh [Iw]W. {A T (w)Vv} faj [(w+ - - [v]T- ({A(w)Vw} + 77 ub)T(AT(ub)v+) . - + f ii,(w)})] (w) where Fv' is the numerical viscous boundary flux, Fr and rf are the auxiliary variables, and qf is the stabilization parameter. For all cases shown in this work, the stabilization parameter 7f is set to a conservative value of 6.0. 2.3.3 Source Discretization The source terms are discretized using the mixed form presented in [5] which is asymtotically dual-consistent [39]. The discretization is given by vV'- S(w, Vw + r"(w)) Rr"(wv) (2.30) KETh where the global lifting operator rgo is given as rhb(w) = ';if,,(w) + 2.4 '(w) (2.31) fEon fEri Temporal Discretization and Solution Technique To fully define the spatial discretization, a basis must be selected for the space V[. In this work, we use a nodal Lagrange basis that is element-wise discontinuous such 29 that the discrete solution has a solution of the form Us(t)#j(x) Uh(X, t) = (2.32) i=j where U E RN. The spatially discrete system is now given as: M- dU dt + R(U) = 0 (2.33) where M is the mass matrix and R is the spatial residual vector. Since this work focuses on steady problems, the temporal discretization is used only to improve the performance of the solver by marching the solution in time from the initial state U(0) to the steady state. In particular, a first-order backward Euler is used for time integration such that the complete discrete system is given by 1 IM(Um+l - Um ) + R(Um+1) = 0 At (2.34) Newton's method is applied at each time step and the resulting linear system is solved using the GMRES algorithm [47] with an in-place block-ILU(0) factorization [22] with minimum discarded fill reordering. For smaller linear systems, a sparse direct solver (UMFPACK [20]) is used. 2.4.1 Numerical Results We compare the presented negative SA model against the earlier negative variant presented in [38] on a model problem from the NASA Langley Turbulence Modeling Resource website. Test Case - 2D NACA 4412 Airfoil Trailing Edge Separation To demonstrate the effects of the modifications, an example case is solved using both versions of the negative SA model. The case is M.. = 0.09, Re, = 1.52 x 106, a = 13.870 flow over a NACA 4412 airfoil. For each model, the RANS-SA system 30 105 - P1 P2 P3 --- P1 -- -P2 ---P3 - i 10 ca -0 - Negative SA Negative SA Negative SA Oliver SA 0UUI.I - 4000 -- Oliver SA Oliver SA 3000 . .. -5 *s 10- Negative SA Negative SA Negative SA Oliver SA Oliver SA Oliver SA R E I- z P1 P2 P3 - - - P1 - - - P2 - - - P3 - 20001000 - - -Vill 1010 0 20 40 Iteration 60 80 100 0 2 20 40 60 Iteration 80 100 (b) Computation Time (a) Nonlinear Residual Convergence Figure 2-2: Comparison of convergence for flow over a NACA 4412 using negative T variants is discretized using the discretization presented in Section 2.3 and the solution is computed using p = 1, 2, 3 polynomials. To focus only on the dependence on the modifications to the turbulence model, UMFPACK is used to solve the linear systems. Figure 2-2a shows the nonlinear residual history of the NACA 4412 case computed on a coarse mesh (7168 elements). For the p = 2,3 solves, the presented i; modifications mitigate the overall nonlinearity which helps to decrease the number of nonlinear iterations required for convergence. Figure 2-2b demonstrates that this modified SA equation does not noticeably modify the work required at each iteration. The drag computed using the two models differs by only 0.02 percent. The presented negative iY modifications have little effect on the converged solution as evidenced by the final drag values. This is expected as both models preserve the positive i portion of the SA equation. However, the presented model better mitigates nonlinearity which results in a more robust scheme. 31 120 32 Chapter 3 Output Error Estimation and Continuous Mesh Framework In this chapter, we review the key ingredients required for the mesh adaptation algorithm used in this work. Specifically, we present the dual-weighted residual method proposed by Becker and Rannacher [9, 10] to estimate the output error and the continuous mesh framework [34, 35] as a means to control this error. 3.1 Error Estimation Let the output of interest be denoted by J = J(u), where u E V is the exact solution to the governing PDE, and J(-) : V -+ R is the output functional. Given a DG solution uhp E Vh,p, an approximation to the desired output is given by Jh,p = Jh,(Uh&,) (3.1) where Jh,p : Vh,p -+ R is the discrete functional. The objective of the error estimation is to approximate the true error in the output functional, Etrue J - Jh, = J(u) - Jh,p(uh,p). 33 (3.2) In this work, the output error estimation is achieved using the the dual-weighted residual (DWR) method proposed by Becker and Rannacher [9, 10] 3.1.1 Dual-Weighted Residual Method and Localization In the dual-weighted residual method, the output error can be expressed as Etrue = J - Jh,p = -lZhp(UhP, V), (3.3) Vw E W. (3.4) where 7P E W = V + Vhp is the adjoint satisfying j'[u, uh,],w, Here, 1' W x W -+ R and 7 = W - [u, [U) Uh,p](W), R are the mean-value linearizations defined by j [u, Uhp] (W, v) = 1Z'[( - 9)u + Ouhp(W, v)dO [u, Uhp](W) = 1 '(1 - )u + OUh,pI(w)dO, where 1'[z](-,-) and J'[z](.) denote the Fr'chet derivative of lh,p(,) and Jhp() with respect to the first argument evaluated about z. As equation (3.4) involves an infinite dimensional space W and the exact solution u, the adjoint solution '0 E W is not computable in general. In this work, for the purpose of error estimation, the true dual solution 0 is estimated with an approximate adjoint 7ph,f E Vh,p computed Vvh, E Vf (3.5) from a linearization about Uhp, 1R',Z[uh,p](Vh,, #L'h,,3) where Vp D Vhp where = J,f[u,],p][Vh,], P = p + 1 is the enriched space. The DWR error estimate using this surrogate adjoint is then given by Etrue ~ -Rh~p 34 (Uh,p, VNh,fi) (3.6) For the purpose of adaptation, a localized error estimate is also defined for each element r, Rhp (Uh~p,7 (3.7) h,pi 1 A conservative error estimate for the output of interest is then obtained by the summation of the locally positive error estimates: e q (3.8) rET 3.2 Continuous Mesh Framework While the DWR method gives a localized output error for each element, adaptation further requires connecting this elemental error contribution to the element's size and orientation. For this effort, we reformulate the anisotropic information for a simplex K as a metric tensor MK, which is a symmetric positive definite (SPD) matrix that encodes the element's size and orientation [27, 54]. elemental metric tensors, {MK}KETh, metric field {M()} E-E From the collection of these a continuous spatially varying Riemannian can be constructed [12, 34] which provides a continuous interpretation of the discrete mesh. Given a mesh Th, the field {MK}KET is uniquely defined and reversely, given a metric tensor field, a family of non-unique metric-conforming triangulations can be constructed. The output error for the DG discretization can be shown to be a function of the metric tensor field [58] and this fact completes the foundation for the adaptation process. A metric-based adaptation algorithm can now be constructed that strives to reduce the output error by manipulating the continuous metric tensor field and constructing the corresponding metric-conforming discrete meshes. In the following two subsections, we review the continuous mesh framework [34, 35] by more rigorously defining the duality between the Riemannian metric field and the corresponding discrete mesh. 35 3.2.1 Metric-Conforming Meshes A Riemannian metric field {M(x)}E 0 is a smoothly varying field of symmetric positive definite matrices on Q C Rd. The metric field introduces a distance function such that the length of a segment ab from point a E Q to point b E Q under the metric is given by: fM(ab) = / abT M(a + bs)b ds (3.9) With this definition of length under the metric, we can now formally define what it means for a discrete mesh to be metric-conforming. A mesh conforms to a metric if each edge, e, of the triangulation is close to unit length under the metric field {M()}xE- and if every element, ib, satisfies a measure of quality. Specifically, a metric-conforming mesh satisfies the edge-length condition, I fM(e) < vf2, Ve E Edges(Th) (3.10) /d Eet(M(X))dX) E [a, 1] with a > 0 (3.11) and the element-quality condition, 2 ()= ZeEEdges(rc) M where QM is the element quality measure. As noted before, for a given {M(x)}XE, a family of non-unique metric-conforming meshes with similar geometric characteristics can be generated. In this work, we use the Bidimensional Anisotropic Mesh Generator (BAMG) [11] developed by INRIA to generate all two-dimensional metric-conforming meshes and Edge Primitive Insertion and Collapse (EPIC) [37] developed by The Boeing Company for three dimensions. For problems with curved geometries, the linear mesh is globally curved using linear elasticity to properly represent high-order geometric information [38, 43]. 36 3.2.2 Mesh-Conforming Metric Fields Conversely, given a mesh Th, the discontinuous field {M},,ETh can be uniquely defined which then can be used to reconstruct a continuous metric field, {M(x)}XEQ, represented by metrics associated with the vertices of the triangulation, {MV}VEv where V is the set of vertices. That is, for a given tessellation, it is possible to find a continuous metric field that conforms to the mesh. We start with the construction of {M}KETh which is termed the element-implied metric. The element-implied metric MK of a simplex element K is a metric under which each edge of the element is unit length. Specifically, the element-implied metric, M., is a SPD matrix such that, 'eTMe = 1, Ve E Edges(r,) (3.12) For higher-order curved elements for which the implied metric spatially varies within the element, we formulate a singular element-implied metric by taking the value from the centroid of the element. The vertex-based metrics are then calculated by performing an affine-invariant average of the elemental metrics of the elements around the vertex in question, i.e. MV = mean afm({M re (3.13) (v)), where w(v) is the set of elements surrounding the vertex v and the affine invariant mean [41] is defined as, meanaffinv ({M}KEw(v)) = arg min M E log (M-1/ 2 MM-1/ 2 ) 2|$ (3.14) KEw(v) Finally, we define a continuous metric field over an element K as a weighted affine invariant mean of the vertex metrics, M(x) = w,(x)I|log (M-1/2MM-1/ 2 ) 1|2, argmin M vEV(s) 37 x E K (3.15) where w,(x) is the barycentric coordinate corresponding to the vertex v. With this recovery algorithm, we are now able to reconstruct a continuous metric field given a discrete mesh. The geometric duality described here serves as the foundation for the adaptation algorithm detailed in Chapter 4 in which we manipulate the continuous metric description of the current mesh to generate a metric-conforming mesh for the next iteration in an attempt to lower the output error. 38 Chapter 4 Metric Field Optimization using Local Error Sampling and Synthesis Framework In this chapter, we first review the metric optimization framework proposed by Yano and Darmofal [59] used in this work. We then present modifications to the opti- mization process to increase the robustness of the global adaptation process and demonstrate the modifications on some simple 2D test cases. 4.1 Model Definition The adaptation framework used in this work is the Mesh Optimization via Error Sampling and Synthesis (MOESS) algorithm developed by Yano and Darmofal [59]. In the following sections, we review the MOESS algorithm with a focus on the optimization. of the resulting statement. 4.1.1 Continuous Relaxation The overarching goal of mesh adaptation is to iteratively improve the mesh or triangulation Th to obtain a better output prediction. To eliminate trivial solutions of 39 arbitrarily refined meshes, we are only interested in "better" triangulations smaller than a specified degrees of freedom (DOF). This goal can be stated as an optimization problem to find the optimal triangulation 7* that minimizes the error subject to a degree of freedom constraint: T* = arg min e(Th) subject to C(*Th) < doftarget (4.1) 7h where e(-) is the error functional and C(-) is the cost functional that computes the number of degrees of freedom for a given Th. Since a mesh is defined by nodes and their connectivity, the optimization problem as formulated above is a discrete problem and as a result is largely intractable. Here we use a continuous relaxation technique of this problem proposed by Loseille and Alauzet [33] in which the discrete triangulation is described by a continuous metric field A4 {M(x)}E.-. The relaxed optimization problem's objective now becomes, find the optimal metric field A4* that minimizes the error subject subject to the same DOF constraint: M4* = argmine(.A4) subject to C(M) < doftarget (4.2) In view of the continuous mesh framework, the cost functional C(-) is now given as C(M) = jf det(M(x))dx, (4.3) where cp is the degree of freedom associated with a reference element normalized by the size of the reference element. The coefficients associated with triangular and tetrahedral elements are 2(p + 1)(p + 2) and c" = / (p+ 1)(p + 2)(p + 3) (4.4) To estimate the error functional, e(M), we use a locality assumption such that the total error functional results from the sum of the local elemental error contributions 40 in and that each of these contributions is a function of the elemental metric tensor: E(A4) ~:- E 77r(Mr.) 4.1.2 (4.5) Metric Manipulation Framework In our mesh adaptation procedure, we are fundamentally interested in manipulating the metric field in order to obtain a mesh that better approximates the true solution. More specifically, we are interested in controlling the change in the approximation in a given direction. In the metric framework, this approximation difference can be seen as the change in the directional lengths measured under the metric fields: h(e;M) = h(e; Mo) /eTM1/2e\ 1/2 (4.6) eTM1/2e Standard Euclidian manipulation of the entries of the metric tensors (i.e. M M = 6M) proves unsuitable as the updates 6M do not strongly correlate with changes in this approximability. We instead use a tangent vector S E Symd which arises from endowing the tensor space with an affine-invariant Riemannian metric [41] and allows us to more strongly control the modifications to the directional lengths. The change in the metric tensor from M 0 to a new configuration M under this affine invariant framework is given as, M(S) =M1/ 2 exp(S)M 1 2 (4.7) / where exp(-) is the matrix exponential. The fractional change in the directional length is bounded by the magnitude of S, exp (- \(41Amhe;n(()S)) IISIi2) <{exp1 (-Amax(S) 2 2 ~h(e; M0) < exp (2( IAmin(S) < exp /1iiIF (4.8) where 11-I|F, is the Frobenius norm, such that this choice of S (referred to as the step matrix in this work), allows us to control the change in the directional approximability 41 SF by controlling its magnitude. 4.1.3 Cost Model With the metric manipulation framework introduced in 4.1.2, an element-wise cost model p,, in terms of the step matrix is constructed by directly integrating the continuous local cost function over an element: p.(S.) = c, det (M2 exp(S-)M 2 = pK 0 exp tr (S.) (4.9) The global cost is now simply the sum of these elemental cost contributions: C(.M) = C(S.) = E 4.1.4 p. (Sn) (4.10) Surrogate Error Model from Local Error Sampling and Synthesis As the error functional is generally not known, a surrogate model is constructed from a sampling procedure. This construction is achieved by sampling how the elemental error changes with respect to changes in the element's configuration. For an element eE (2, we consider i E [1, ..., nconig] configurations formed by locally splitting the edges to obtain subdivided meshes ni. By convention, no corresponds to the original configuration. For each ith configuration, an elemental-wise local problem is then solved: find U" E Vhp(i) such that , , ='0, VVri E Vh,p(ri), where the local semilinear form (, (4.11) prescribes the boundary fluxes on ri by assuming the solution on the neighboring elements does not change. We then prescribe a localized error estimate corresponding to the subdivided mesh ri by recomputing 42 the localized DWR error estimate as (4.12) 94 lht,p(Uh,, 1= V#h,P1.e) I Each subdivided mesh is associated with an elemental metric M, by performing an affine-invariant average of the implied elemental metrics of the newly formed elements. As a result, the sampling procedure generates a set of metric-error pairs, {MKq, 7}. As introduced in 4.1.2, we can now characterize the change from the original metric tensor Mno to the new configuration MK, using the affine invariant framework: = log (M-1/ 2MniMA 1 /2 ) , i = 1,..., nconfig (4.13) Similarly, we measured the associated changes in the error as fa, = log (77,, /q ) i = 1, ... , nconfig (4.14) Thus with the affine invariant framework, we can construct step matrix and error change pairs {Sri, fa} which now be used to construct our surrogate error model. The MOESS algorithm constructs a linear error function of the form, fr(Sx) = tr(RnS.). where RK is synthesized from the pairs {S,., f,,,} (4.15) through least-squares regression. The local error model in terms of S, is then given as: 7 (S.) = qA exp(tr(RS)) (4.16) This local error model effectively represents how the local error changes with respect to local element shape and size changes as encoded by SK. 43 4.2 Optimization of Surrogate Model The last step of the adaptation process is to perform an optimization of the Riemannian metric field {M}sEQ with the constructed surrogate cost and error models to minimize the error. Consistent with the continuous mesh description provided in Section 3.2.2, {M}XEQ is described by the vertex metric tensors {M}vEV. We again describe the change from the original vertex metric tensor by using the affine invariant framework such that, M,(S,) = M0KJ exp(S )M , (4.17) where S, E Symd is the vertex step matrix. Thus the objective becomes to find the vertex step matrices that minimize the estimated error. As the surrogate error and cost models given in (4.16) and (4.19) respectively are in terms of the element step matrix SK, we require a relationship between S,, and S, to transform these models to be in terms of the the design variables. For this work, we construct the elemental step matrix SA, via an arithmetic average of vertex step matrices: (4.18) V Sr = {SV}vEv() = VEV(rK) Substitution of this relationship into the cost model yields the cost constraint in terms of the vertex step matrices: Pr. ({Sv}vEv(K)) C({= (4.19) r.E Th Similarly, the error model used as our objective function becomes 6(f{SV}VEV) = E 7 (f{SV}VEV~r) (4.20) Since this error model is constructed through local samples taken from the original mesh, the allowable change in the metric field in one adaptation iteration must be limited. As specified in (4.8), the choice of the step matrices as the design variables 44 allows us to control the change in the approximability in any direction by controlling the magnitude of the matrix. For this effort, we apply constraints to limit the magnitude of S,: ISvlF 5a v (4.21) Since the presented local splitting procedure effectively explores fractional directional length changes of 2, we require that the requested edge length h(e; M (S)) satisfies: h(e; M (S)) <2 2 h(e;kMo) - (4.22) - 1 Therefore, we limit the change in approximability to 2 in any direction by setting a = 2 log(2) The surrogate optimization problem for the optimal Riemannian metric field is now given as: {S*} = arg min E({SV} EV) s.t. C({Sv}ve) ISvIIF < a, (4.23) dOftarget (4.24) Vv E V (4.25) The gradient of the error and cost functions with respect to a given vertex step matrix S, is 6C je r. ({S V~ ovE() r.Ew(v)- Vr) R. (4.26) . 6SV {S}VEv(x)) 21V(r,)l (4.27) where w(v) is the set of elements that have v as one of their vertices. 4.2.1 Heuristic Optimization of Surrogate Model We now review the heuristic optimization method presented by Yano [58] and identify some limitations of the method that this work later addresses. 45 Summary of Algorithm For convenience, the step matrix S, is decomposed into the trace and trace-free parts, S, = s,I + S where sv = tr(S,)/d and 5, (4.28) is the trace-free part of S,. The heuristic method approaches the optimization problem by initially assuming that the current configuration is sufficiently close to the optimal configuration such that the constraints given in (4.25) are inactive. Via steepest descent type updates, the approach attempts to distribute the available mesh degrees of freedom such that the investment to any element results in the same marginal improvement .in the error and proceeds to change the trace-free part of the step matrix in an attempt to achieve stationarity with respect to shape change. The specifics of the heuristic algorithm to approximately solve the optimization problem are as follows: 0. Set Js = a/ntep /65 1. Compute vertex derivatives, &/6s, 1 6,s, local Lagrange multiplier A_ = (6,/6sv)/(6C/6sv) about {Sv}VEv 2. Update the isotropic part of Sv according to: " Refine the top 30% of the vertices v with the largest AV by setting Sn+1/3 = Sv +6sI " Coarsen the top 30% of the vertices v with the smallest AV by setting Sn+1/3 = S- -6sI 3. Update the anisotropic part of S, according to + 4. Rescale Sh+21 3 according to Svn where 0 is selected to obtain a metric field with 5. Set n = n +1. = S n+2/3 + doftarget Ifn < nstep go to step 1. 46 01 S+ 1 /3-6s(6/65)/(6/6s) Limitations The algorithm as presented above has been shown to automatically and efficiently generate metric requests to better resolve outputs of interest with no a priori assumptions on a wide range of problems [19, 56, 60]. However, this procedure exhibits a few limitations which can inhibit the realization of robust automated adaptation. The heuristic isotropic update shown in step 2 of the algorithm performs well when the Lagrange multipliers are well distributed such that the vertices with high sensitivity in the local error with respect to added degrees of freedom are refined and the vertices with low sensitivity are coarsened. However, since the selection process for coarsening and refining are independent of the actual distribution, this procedure can exhibit adverse behavior when the error is dominated by a small percentage of the total vertices. In this case, the vertices with high sensitivity will be properly selected for refinement but the 30% selection protocol will also assign refinement to vertices that should either be coarsened or untouched. The converse effect can occur if there are many vertices with large Lagrange multipliers such that the process coarsens vertices with high sensitivity. Along these lines, in the extreme case where R, = 0 such that A, = 0 Vv E V, the presented heuristic algorithm will still apply coarsening and refinement factors to the current metric even when there is no sensitivity. Another potential robustness issue is the fact that the updates and the final resulting step matrices do not necessarily remain within the error sampling space. Although the Js factor is sized such that the trace of every step matrix remains bounded, the resulting step matrix after the anisotropic update, Sj 2 3 1 , has no guarantee to satisfy the edge length constraint given in (4.22). The final scaling step further exacerbates the problem if the initial mesh is much smaller with respect to the doft.get. 4.2.2 Modifications to Metric Optimization Procedure To alleviate some of the identified issues with the heuristic optimization procedure, we choose to employ a formal gradient-based optimization algorithm in lieu of the 47 heuristic method to exactly solve the non-linear optimization problem. Here we are only interested in modifying the optimization process and not the underlying surrogate modeling. However, a straight forward application of a nonlinear optimization algorithm on the problem as posed in (4.23) proves to be largely intractable as the number of non-linear constraints scales with the number of vertices. In this section, we explore alternate optimization statements that satisfy the necessary requirements. We are specifically interested in the following properties for our optimization for practical and robust output-based mesh adaptation: " The problem must be able to be solved in a reasonable amount of time. Since the number of design variables S, will naturally scale with the number of vertices of the initial configuration, this property requires that the number of non-linear constraints remain constant with increasing mesh size. " For robustness, the design space of the optimization must be within the error sampling space. That is to say, the final resulting metric request must satisfy the edge length constraint given in (4.22). " To minimize the number of adaptation iterations necessary to obtain an optimal mesh, the design space of the optimization should be as large as possible while again satisfying the edge constraints. Metric Optimization with Global Penalty Term To avoid the issue of a growing number of non-linear Frobenius norm constraints, we replace the individual vertex norm constraints with a penalty term that is added to the objective function. This penalty term consists of a penalty parameter p, Vv E V that is multiplied by the measure of the constraint violation. To avoid possible numerical issues with the derivatives when IISVIIF = 0, the penalty is placed on the square of the Frobenius norm. The surrogate optimization problem for the optimal Riemannian 48 metric field using this penalty method is now given as: = argmin ({SV} {S}vk ) + {Sv}Vv s.t. pvO(| Sv|1 - (2 log(a)) 2 ) (4.29) VEV C({Sv}VEv) < doftarget (4.30) where p is a penalty vector of length IVI and 4(x) is a simple quadratic penalty term given as: O(W = 0, S2, if <0 (4.31) if X > 0 This penalty function penalizes the objective function whenever a magnitude of a vertex step matrix grows larger than the allowed bounds. However, a formal optimization of (4.29) can result in spurious shape change requests which do not noticeably improve the solution and can introduce unnecessary difficulties to the mesher. According to the MOESS surrogate error model, if R, = 0, the error can always be reduced by choosing a 'k such that tr(RKSK) < 0. In practice, this characteristic of the error model leads to maximum possible shape changes in almost every vertex metric even when the respective error reduction is insignificantly small. As such, we include an additional global constraint that forces the optimization process in some sense to focus on metric changes to those vertices with the highest error sensitivities: ||SvlI < #|VJ(2log(a))2 (4.32) vEV If 3 < 1, this constraint introduces a global limit to how much the original configuration can change measured by the sum of squares of the Frobenius norms and effectively forces the optimization to focus only on prescribing changes that most reduce the error. 49 The final optimization statement is now: {sv*}Isy= arg min E({S}vE) + E {SV}vEV s.t. pVq(HISv |12 - (2log(a)) 2) (4.33) vEV C({S}VEv) 5 doftarget S S<I I VI(2log(a)) 2 (4.34) (4.35) VEV With this method, we now solve a series of unconstrained (with respect to the edge length constraints) optimization problems and increase the penalty parameter if needed after every iteration. The algorithm to solve the optimization problem is as follows: 0. Set p = pinit and {Sv}vE = 0 1. Solve (4.33) to obtain {S*} Ev 2. Post-process {S*}VEv to test if (4.22) is satisfied with the resulting step matrices. Here we present two possible post-processing tests: (a) Eigenvalue Method: Calculate A (Sv) and test if IA (S,) < 2 log (a). (b) Edge-Based Method: Calculate h(ei; Mv) = h(ei; MOJ exp(S *MOK) where ej are the eigenvector directions of MO,v and test if the requested edge lengths satisfy (4.22). We note here that the edge-based method is less conservative in that the limit on the change in approximability might not be satisfied in every direction e. The eigenvalue method however does ensure this constraint is satisfied in every direction. 3. For each S* that breaks the constraint, increase the respective penalty term by a prescribed factor: pv = 0 pv where 0 > 1. If no S* breaks the constraint, exit with {MI}Mv = 4. Initialize {SV}vEV = {Sv*}VEV and go back to step 1. 50 {M /2 exp(S* )M1/2}v- For the adaptive cases presented in this work, at each iteration, the initial total error is normalized to unity and the initial penalty vector pi is set to 1 x 10-. Unless otherwise noted, the eigenvalue post-processing method is used. For solving the actual optimization problem for a given p, we use the globallyconvergent method-of-moving-asymptotes (MMA) algorithm [51] as implemented in NLopt [29]. For convenience, in this work, we refer to the optimization statement as given in (4.33) and the respective solution procedure as the gradient-based metric optimization method. 4.3 Numerical Results We present numerical examples of applying the MOESS adaptation algorithm with the gradient-based optimization method to select problems. We first verify the ability of the modified adaptation algorithm to produce optimal meshes in the L 2 error control setting. We then move to two-dimensional aerodynamic problems with which a comparison between the presented algorithm and the heuristic optimization method as presented in Section 4.2.1 is made. 4.3.1 r'-Type Corner Singularity We start by applying the gradient-based optimization method to a simple L 2 -projection problem with a canonical singularity for verification of the modifications to the adaptation algorithm. The L 2 approximation problem consists of finding a solution Uh,p E Vh,p that minimizes the square of the L 2 projection error: Uh,p= arg min Vh,p (U - vh, )2dx. (4.36) f For this case, we consider a general form of the singularities found at geometric corners of solutions to elliptic equations given by u(r, 0) = r' sin [a(O + Oo)] 51 (4.37) where r is the Euclidean distance from the corner, a > 0 is the singularity strength, and 0 is the offset angle. The optimal mesh for this r' function for a degree-p polynomial approximation as shown by Yano [58] consists of isotropic elements with a size distribution given by h(r) = Cr- (4.38) We apply the gradient-based adaptation algorithm to the L 2 projection problem of the r' corner singularity function with a = 2/3 and compare the resulting optimized h distributions to the analytically derived distributions. For each solution order p, the numbers of degrees of freedom considered are: p = 1, doftarget p = 3, dOftarget = = {600 900}, {2000 3000}, For the case presented here, the adaptation process is started on a square isotropic mesh with the solution's L 2 error as the output of interest. Figure 4-1 shows the resulting distribution of h against r for the optimized meshes. Here, the element size h is calculated based on the volume (h = det(M )- 1 / 4 ) and the distance r is measured from the corner to the centroid of the element. The optimal valies of h and r vary linearly in the log-log space with an optimal grading coefficient of kan" = 0.44 and 0.67 for p = 1 and 3, respectively. The meshes produced through the adaptation procedure exhibit grading factors of k = 0.45 and 0.66 for p = 1 and p = 3, respectively with 300 elements and demonstrate that the modified adaptive algorithm is able to obtain the optimal p-dependent grading automatically. 4.3.2 RAE 2822 Transonic RANS-SA To characterize the differences between the two presented metric optimization methods on aerodynamic problems, we first consider a transonic RANS flow that exhibits various flow features with a wide range of scales. Specifically, we consider turbu- lent transonic flow over an RAE 2822 airfoil with a freestream Mach number of 52 100 10 k =0.47 0.45 kanalt= 0.44 1010-2 10 -1 ,2. 10 *. -elem 10 34- 10 = 200 elem = 300 .-- 100 1 10- r -elem - 10-3 elem = 200 = 300 - 10 10- k= 0.66 k =0.66 kanayli = 0.67 10 100 r (a) p =1 (b) p = 3 Figure 4-1: Element size h versus the distance of the element centroid from the corner for optimized meshes for r' singularity problem with a = 2/3 M... = 0.737, a Reynolds number of Rec = 6.5 x 106, with an angle of attack of 2.79'. The RANS equations are solved using a p = 2 DG discretization and the meshes are adapted for drag on the surface with a doftarget = 80, 000. Figure 4-2 shows the error estimate and degrees of freedom for 20 iterations of adaptation using the two optimization methods. Both procedures are able to significantly lower the error estimate over the course of the adaptation history to effectively obtain the same estimated error level but do so along very different trajectories. The progression of meshes generated from the heuristic and the gradient-based optimization methods can be seen in Figures 4-4 and 4-5 respectively. The heuristic optimization method very aggressively refines the elements in the locations where the MOESS error model at that iteration detects high localized error contributions. It is clear from visual inspection of the initial mesh to the mesh of iteration 2, that the edge lengths of the triangulation change much more rapidly than the factor of 2 sampled during the local solve procedure. Due to this nature of the metrics extending beyond the sampling space, the transition meshes more frequently contain spurious shape features that significantly change over multiple adaptive iterations. This phenomena is exemplified by meshes 3 through 15 on which the spatial location of the clustered anisotropy for the shock significantly 53 10, x10 Gradient-Based Heuristic2 \- 10- 10' 10 10 - o4 o-e . 10-5F0 2-x - 10' 0 2 4 6 8 10 12 14 16 18 20 0 Adaptation Iteration 2 4 6 8 10 12 Adaptation Iteration (a) Error Indicator vs Adaptation Iteration - -0- 0 14 Gradient-Based Heuristic 16 18 (b) Degrees of Freedom vs Adaptation Iterati on Figure 4-2: Comparison of error indicator and degrees of freedom history for transonic flow over a RAE 2822 Airfoil shifts as flow becomes better resolved. This aggressive nature of the heuristic updates also pushes the global degrees of freedom to rapidly increase as seen in Figure 4-2b. Furthermore, since each iteration's primal initialization is performed with a projection of the previous iterate's converged primal state, the rapid changes in the triangulation can result in significant difficulties during the primal solves. Figure 43a displays the cumulative computation time taken to obtain a primal solution on a given mesh; this quantity includes the time taken for every primal and dual solve for the previous iterations. At adaptation iteration 3 for the heuristic process, this graph reveals that the primal solve requires a significant amount of computational effort for convergence due to a poor initialization that stems from rapid changes in the mesh. The mesh sequenced generated from the gradient-based optimization method is instead characterized by smooth transitions from mesh to mesh as a result of the strict enforcement of the edge length constraints. Although the process takes a significantly larger number of iterations, the lower degrees of freedom during most of the adaptation history coupled with better initializations due to slower changes in the mesh allow for all 20 iterations to finish before the completion of the 13th iterate of the heuristic process. As shown in Figure 4-3b, both methods result in similar error indicator levels with roughly the same amount of computation time. 54 20 10 14 - -0; 12 Gradient-Based Heuristic - -- Gradient-Based -0- Heuristic 10 0d E 100 10 - S20 4- 10- E 20 e 2 4 6 8 10 12 14 Adaptation iteration 16 18 20 0 2 4 6 8 10 12 Cumulative Computation Time (s) (a) Cumulative Time vs Adaptation Iteration (b) Error Indicator vs Cumulative Time Figure 4-3: Cumulative computation time taken to obtain primal solution and error indicator estimate on a given iterate's mesh for transonic flow over a RAE 2822 airfoil 4.3.3 Supersonic flow over NACA 0012 To quantify the importance of finely controlled mesh adaptation for aerodynamic flows, we consider a supersonic Euler flow over a NACA 0012 airfoil. The freestream Mach number is Mo, = 1.3, and the airfoil is at 0' angle of attack. equations are solved using a p = The Euler 1 DG discretization and the meshes are adapted for drag on the surface with a dOftarget = 20, 000. Figure 4-6 shows the drag and drag estimate behavior for 20 iterations of adaptation using the heuristic optimization method. The solutions generated through this method exhibit significant oscillations in both the actual Cd and the Cd error indicator values; the last five solutions of this sequence exhibit a Cd spread of over eight drag counts. The lack of convergence in the error estimate indicates that the adaptation process is unable to settle on a family of optimized meshes characterized by similar metric fields. Figure 4-7 shows the history that results from starting the adaptation procedure from the same initial mesh but using the gradient-based metric optimization method. Here, the generated sequence of solutions progressively reduces the error estimate until a family of optimal meshes with similar error levels is obtained. These results demonstrate a distinct lack of Cd oscillations relative to the heuristic results with a 55 14 X1 Cd (a) Initial Mesh (b) Iteration 1 (c) Iteration 2 (d) Iteration 3 (e) Iteration 4 (f) Iteration 5 flll 11 ME- (g) Iteration 10 777 (h) Iteration 15 (i) Iteration 20 Figure 4-4: Mesh adaptation history for transonic flow over a RAE 2822 Airfoil using the heuristic metric optimization method spread of less than 0.5 drag counts over the last 5 iterations. Since the formal metric optimization procedure stays within the error sampling space, the gradient-based method more robustly settles at this family of optimal meshes. Select drag-adapted meshes generated through the heuristic optimization method and their respective Mach number distributions can be seen in Figures 4-8 and 4-9. The meshes generated through the heuristic method visually demonstrate the inability 56 (a) Initial Mesh (b) Iteration 1 (c) Iteration 2 (d) Iteration 3 (e) Iteration 4 (f) Iteration 5 (g) Iteration 10 (h) Iteration 15 (i) Iteration 20 Figure 4-5: Mesh adaptation history for transonic flow over a RAE 2822 Airfoil using the gradient-based metric optimization method of the process to converge on an optimal mesh for this highly mesh-sensitive problem. For this particular flow, both the bow shock and the oblique shock contribute to the difficulties encountered by the mesh adaptation scheme as the meshes exhibit significant differences in the resolution of the features. Figure 4-10 displays two optimized meshes with similar error indicator values generated through the gradientbased method. As exemplified by these very similar meshes and their resulting Mach 57 0.1 0.1 *dolm dottargal 0.099 0.098 ~~~~-~O~04M"-" -"'* M-2' -0000 0,010 0.097 0O 15000 0.096 C 2 2 CP 0.095 U- 0.001. 0 0 0.094 - * 0.093 0.0001 10000 4000 0.092 0.091 n nQ, 5 10 15 0 20 5 (a) Cd 10 15 20 Adaptation Iteration Adaptation Iterations vs Adaptation Iteration (b) Cd Error Indicator and Degrees of Freedom vs Adaptation Iteration Figure 4-6: Adaptation history for supersonic flow over a NACA 0012 Airfoil using the heuristic optimization method 0.1 0.1 . - -dot 2964 0.099 0.098 0.014 0.097 CU 0.096 Q LO 0.095 15000 L- 0 0.001 2 0.094 10000 L, 0.093 0.0001 5000 0.092 0.091 " 0.09 W- 5 10 15 20 0 Adaptation Iterations (a) Cd 5 10 15 26 Adaptation Iteration vs Adaptation Iteration (b) Cd Error Indicator and Degrees of Freedom vs Adaptation Iteration Figure 4-7: Adaptation history for supersonic flow over a NACA 0012 Airfoil using the gradient-based optimization method number distributions, the gradient-based method is able to reach and settle on this set of optimized triangulations. 58 (a) Mach number distribution, iteration 17 0.1 0.2 0.3 0.4 0 0.8 0.7 0.8 0,9 1 1.1 (b) Mesh with Eind 4.3 x (d) Mesh with Eind 4.8 x 10-5, iteration 18 10-4, iteration 17 1.2 1.3 1.4 1.5 (c) Mach number distribution, iteration 18 Figure 4-8: Select adapted meshes (iterations 17 and 18) and respective Mach number distributions generated with the heuristic optimization method, NACA0012 supersonic flow. 59 (a) Mach number distribution, iteration 19 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 (b) Mesh with Eind = 6.5 x 10-4, iteration 19 1.4 1.5 (d) Mesh with Eind = 6.9 x 10- 5 , iteration 20 (c) Mach number distribution, iteration 20 Figure 4-9: Select adapted meshes (iterations 19 and 20) and respective Mach number distributions generated with the heuristic optimization method, NACA0012 supersonic flow. 60 0.1 0.2 0.3 04 0.5 0. 0.7 0.8 0.9 1 1.1 1.2 13 1.4 1.5 (a) Mach number distribution, iteration 19 01 02 0.3 0.4 0.5 0.6 07 0.8 09 1 1.1 2 1. (b) Mesh with Eind 4.5 x (d) Mesh with 3.9 x 10-5, iteration 20 10-5, iteration 19 1.4 1.5 (c) Mach number distribution, iteration 20 Eind Figure 4-10: Select adapted meshes and respective Mach number distributions generated with the gradient-based optimization method, NACA0012 supersonic flow. 61 62 Chapter 5 Turbulent Aerodynamic Problems In this chapter, we demonstrate the overall robustness of the discretization and modified adaptation procedure by applying the resulting adaptive solver to a range of two-dimensional and three-dimensional turbulent aerodynamic problems. 5.1 Three-element MDA 30P-30N, Subsonic We first consider a two-dimensional turbulent flow over a three-element McDonnell Douglas Aerospace (MDA) airfoil (30P-30N), with a freestream Mach number of moo= 0.2, an angle of attack of a = 16', and a Reynolds number based on chord of Re, = 9 x 106. The initial coarse two-dimensional mesh used for this case is shown in Figure 5-1. The curved airfoil geometry is represented using q = 3 simplex elements. Figure 5-1: Initial mesh, Three-element MDA flow. For this flow, the inflow and outflow boundaries are specified via characteristic 63 farfield boundary conditions. An adiabatic no-slip condition is imposed on the surface of the airfoil. For each solution order p, the numbers of degrees of freedom considered are: p = {1, 2}, dOftarget = {40000 60000 90000 120000}, For each p-doftarget combination, a family of optimized meshes are generated starting from the shown initial mesh. The performance of each p-doftarget is assessed by averaging the error obtained on five realizations of meshes in the family. Adaptation is performed on the drag on all three-elements of the airfoil. Figure 5-2 shows the resulting drag error against the number of degrees of freedom. The p = 2 discretization 0-2 102 -X-P1 -e-P2 10-3w W- 1.88 10- 10 12.7 10 3.87 ' 10 -2.5 10 h = 1/(dof)" 2 10 3 -23 Figure 5-2: Convergence of drag output for the three-element MDA flow; reference value obtained from p = 3, dOftarget = 250k adapted solution outperforms the p = 1 discretization for the entire range of doftarget considered. The coarsest p = 2 solution containing approximately 40, 000 degrees of freedom exhibits a lower error level than the finest p = 1 solution with roughly 120,000 degrees of freedom. The respective optimized meshes can seen in Figure 5-3; while both meshes focus on resolving the same physical phenomena, the p = 2 discretization is able to more accurately calculate the drag on a much coarser mesh. For this relatively smooth case, the optimal output error convergence rates of c xV are recovered for both the p = 1 and p = 2 discretizations. The adaptation algorithm applied to this complex two-dimensional case not only 64 (a) p (b) p 1, doftarget 2, doftarget 120k mesh with E = 3.11 x 10-4 40k mesh with E 3.28 x 10-4 -12 - Figure 5-3: Comparison of adapted p = 1 and p = 2 meshes with similar error levels, Three-element MDA flow. -10. -10- -8- -8- -6- -6- -4 -4 -2 -2 0- -0.2 0 0 0.2 0.4 06 0.8 1 1.2 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 x/c x/c (a) c, distribution of p = 1, doftaget = 120k mesh (b) c, distribution of p Figure 5-4: Pressure coefficient comparison of adapted p = 1 and p similar error levels, Three-element MDA flow. 2, doftarget = 40k mesh = 2 meshes with targets the boundary layers formed on every element but also the off-body features that prove crucial for resolving the interaction between the elements. As shown in Figure 5-6, the process results in significant anisotropic refinement in both the boundary layers and the wakes to effectively capture the interactions of slat with the 65 main element and the main element with the flap. Furthermore, a far field view of 0.75 0.65 0.55 0.45 0.05 (a) Mach number distribution (b) Optimized mesh Figure 5-5: p = 2, doftarget = 120k adapted mesh and respective Mach number distribution, overview, Three-element MDA flow. this optimized mesh (Figure 5-6) reveals significant refinement along the stagnation streamline as well as the wake. These meshes showcase the importance of mesh adaptation in complex flows; the spatial distribution of the element's size and shape required to effectively capture the boundary layer phenomena and the flow interaction of the airfoil elements proves difficult to predict a priori. The combination of mesh adaptation and the p = 2 high-order discretization shows significant advantages over the p = 1 discretization for the flow considered here. 5.2 3D Zero Pressure Gradient Flat Plate, Subsonic We now consider a M, = 0.2, ReL = 5 x 106 turbulent flow over a three-dimensional flat plate. The problem is produced by extruding a two-dimensional case into the third dimension. The baseline two-dimensional problem contains a flat plate of length 2L 66 0.75 0.65 0.55 0.45 0.35 0.25 0.15 0.05 (a) Mach number distribution (b) Optimized mesh Figure 5-6: p = 2, doftarget = 120k adapted mesh and respective Mach number distribution, far field, Three-element MDA flow. preceded by a short lead-in section of length IL such that the leading edge of the flat plate does not start directly at the inflow wall. The initial two-dimensional mesh and the resulting extruded three-dimensional mesh can be seen in Figure 5-7. Y LX (a) 2D side view (b) 3D view Figure 5-7: Initial structured mesh, 3D flat plate flow. 67 For the extruded problem in three dimensions, the inflow boundary is specified by a total temperature, total pressure, and a zero flow angle while the outflow is defined by a static pressure definition. An adiabatic no-slip condition is imposed on the flat plate itself while a slip boundary condition is imposed on every other wall. For each solution order p, the numbers of degrees of freedom considered are: p = 1, doftarget ={5000 10000 15000 20000}, p doftarget 10000 15000 25000}, = 2, = {5000 As before, for each p-doftarget combination, a family of optimized meshes are generated and is assessed by averaging the error obtained on five realization. The output adapted is the drag on the three-dimensional flat plate. Figure 5-8 shows the resulting drag error against the number of degrees of freedom 1 . For this case, the p = 2 102 -6- P1 -e- P2 10-3 10 S10-5 -3.98 r - 106 10-7 10~8 101 h = 1/(dof)" 3 10 1.2 Figure 5-8: Convergence of drag output for the 3D flat plate flow; reference value obtained from two-dimensional simulations discretization outperforms the p = 1 discretization for all but the coarsest doftarget considered. In fact, the finest p = 1 solution with roughly 30, 000 degrees of freedom exhibits a higher error level than the p = 2 solution containing almost half the degrees of freedom with approximately 16,000. For this smooth flow, the optimal 'These results were obtained using the edge-based method instead of the eigenvalue method during the post-processing step of gradient-based metric optimization as described in Section 4.2.2 68 output error convergence rate of c oc hIP is recovered such that for even more accurate solutions, the p = 2 discretization would exhibit far greater efficiency than the p = 1 discretization. A resulting optimized p = 2 mesh can be seen in Figure 5-9. As exemplified by this Y (a) 2D side view (b) 3D view Figure 5-9: P=2, 25k dof optimized mesh, 3D flat plate flow. optimized mesh, the adaptation process results in optimized meshes with significantly different characteristics relative to the initial mesh. The algorithm applied to this case properly detects that the flow exhibits no spanwise variation and automatically generates meshes with large amounts of anisotropy in this direction; the resulting optimized meshes contain only 1 to 2 elements along the z-direction while the initial mesh begins with 6 cells. Furthermore, in an attempt to mitigate the effect of the leading edge singularity, the adaptation algorithm very aggressively clusters elements at this transition point. The initial structured mesh fails to properly capture the flow at the leading edge due to inadequate resolution as shown in Figure 5-10. However, the adapted meshes with this strong grading of elements properly resolve the initial formation of the boundary layer as shown in Figure 5-11. 69 0.01 0.03 0.05 0.07 0.09 0.11 0.13 0.15 0.17 0.01 0.03 0.05 0.07 0.09 0.11 0.13 0.15 0.19 0.01 0.0006 0.008 0.0005 .17 0.19 0.0004 0.006 0.0003 0.004 0.0002 0.002 0.0001 0 II..II,I,II.II | I,, ,| 0 .1 0.5 x i 1 [ , , 1.5 , L | 2 i -0.06 , -0.04 -0.02 0 0.02 0.04 0.06 0.08 x (a) Boundary layer on initial structured mesh (b) Close up at leading edge Figure 5-10: Mach number distribution of 2D center slice, initial mesh, 3D flat plate flow. 0.01 0.03 0.05 0.07 0.09 0.11 0.13 0.15 0.17 0.19 0.01 0.008 0.006 0.004 0.002 0 0 1 0.5 1.5 2 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 x x (a) Boundary layer on p = 2 adapted mesh (b) Close up at leading edge 0.08 Figure 5-11: Mach number distribution of 2D center slice, P=2, 25k dof optimized mesh, 3D flat plate flow 5.3 3D Duct, Subsonic This case presents a M,, = 0.2, ReD = 1 x 106 turbulent flow through a square duct case of channel height and width D. The flow considered here develops over a length 2D. 70 As in the 3D flat plate case, the three-dimensional initial mesh is produced by extruding a two-dimensional structured mesh into the third dimension. Again, we incorporate a lead-in section of length 'D such that the walls of the duct do not immediately start at the inflow boundary. Here, we take advantage of the inherent symmetry of the problem and only consider one quarter of the duct. The initial twodimensional mesh and the resulting extruded three-dimensional mesh can be seen in Figure 5-12. KY X I 'l ." (a) 2D view of inlet and outlet (b) 3D view Figure 5-12: Initial structured mesh, subsonic duct flow. For this duct flow, the inflow boundary is specified via a characteristic farfield boundary condition while the outflow is defined by a static pressure definition. Adiabatic no-slip conditions are imposed on walls of the duct while slip boundary conditions are imposed on every other wall. We consider the solution order p = 2, with a target degrees of freedom of 100000 with the drag on the walls of the duct as the output of interest and generate a family of optimized meshes starting from the initial structured mesh. Figure 5-13 shows the resulting CD and CD error estimate adaptation histories. For this case, the meshes generated by EPIC overshoot the requested degrees of freedom by a factor of two to three. The adaptation process applied to this problem results in optimized meshes with 71 1-nX 10-3 0.01 1.64e+06 1.960.001 1 . 1.94 0 1.92 I1.28e+06 ... U- 0 1 ()0.00 9.20e+05 1.90 1.88 5.60e+05 1.86- 0 2 4 6 8 10 12 14 0 16 2 4 Adaptation Iterations (a) cD 6 8 10 12 14 16 Adaptation Iteration vs. adaptation iteration (b) CD error estimate convergence Figure 5-13: Drag adaptation history for p = 2, subsonic duct flow error estimates over 2 times smaller but with almost 6 times fewer degrees of freedom. A resulting optimized p = 2 mesh can be seen in Figure 5-14. Similar to the presented 3D flat plate meshes, the duct meshes generated through the adaption process contain significant clustering at the leading edges to mitigate the effect of the edge singularities. As before, the initial structured mesh fails to properly capture the flow at the leading edge as shown in Figure 5-15a while the adapted meshes are able to resolve the initial formation of the boundary layer as seen . in Figure 5-15b Furthermore, the optimized meshes gain higher accuracy per degree of freedom by effectively aligning the triangulation to better capture the boundary layer phenomena of the flow. As can be seen in the outflow comparison presented in Figure 5-16, the adaptation significantly coarsens the initially fine elements that lie outside the boundary layer to properly have the mesh reflect the current size of the boundary layer at that cross section. The optimized mesh also contains significantly coarser elements at the corner of the duct where the two boundary layers .intersect but still is able to properly resolve this flow feature. The mesh resolution required to properly resolve this corner would prove difficult to identify a priori but the adaptation algorithm autonomously changes the triangulation to efficiently capture this phenomena. 72 X y \ it x Y IV I _________ ~i~I~ ~ -- 4I 14 1r."' (a) 2D view of inlet (b) 2D view of outlet Y (c) 3D view Figure 5-14: P=2, 100k dof optimized mesh, subsonic duct flow. 73 W-:A 0.02 0.02 0.015 0.015 >0.01 >-0.01 0.005 0.005 I , , 0 0 , . 0.5 I II I I 1 Z 1.5 II . I I . II . , 0 2 0 (a) Initial mesh 0.5 , I I, , 1 z I, 1.5 , I 2 (b) p = 2 optimized mesh Figure 5-15: Mach number distribution, 2D view along duct at X = 0.25, subsonic duct flow 74 0.06 0.08 0.1 0.12 0.2 0.14 0.16 D.02 0.18 0.04 0.06 0.08 0.1 0.12 0.14 0.16 OAS8 02 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 .I 0 , , I 0.1 i I 0.2 0.3 x I . 0 , 0.4 0.5 ,i , ,, ,i ,, 0 0.1 i ,, ,, i ,, ,, i 0.2 x (a) Initial mesh 0.02 0.04 0.06 0.08 0.1 i , I,I II I 0.02 0.04 0.5 0.3 0.4 0.5 (b) p = 2 optimized mesh 0.02 0.12 0.14 0.16 0.18 0.2 0.03 0.03 0.02 0.02 0.01 0.01 0 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.16 0.2 0 0 0.01 0.02 x 0. 03 0 0.01 x (c) Initial mesh - Close up of duct corner 0.02 0.03 (d) p = 2 optimized mesh - Close up of duct corner Figure 5-16: Mach number distribution, 2D cross sectional view at outflow, 2.0 units downstream from leading edge, subsonic duct flow 75 76 Chapter 6 Conclusion 6.1 Summary This work investigated steps towards the realization of a robust high-order adaptive Reynolds-Averaged Navier Stokes (RANS) method for the analysis of turbulent flows. Specifically, modifications were made to the RANS discretization and the Mesh Optimization via Error Sampling and Synthesis (MOESS) framework [58] for increased robustness of the high-order adaptive solver. To improve the robustness associated with the RANS discretization, changes to the negative continuation of the Spalart-Alimaras turbulence model proposed by Allmaras [2] were reviewed and compared against an earlier variant presented by Oliver [38]. The introduced negative SA modifications were tested on a subsonic, turbulent flow over a NACA4412 airfoil with a high angle of attack and shown to improve the convergence of the nonlinear solver. In addition, the metric optimization framework proposed by Yano and Darmofal [59] was expanded to better control the mesh-to-mesh transition during the adaptation procedure. For this effort, the metric optimization step of the original algorithm was changed to a penalty-based statement and a gradient-based solution method was proposed. We presented numerical examples of applying the MOESS adaptation algorithm with this gradient-based optimization method to select problems. We first verified the ability of the modified adaptation algorithm to produce optimal meshes 77 in the L 2 error control setting applied to a canonical corner singularity function. We then moved to two-dimensional aerodynamic problems with the purpose of characterizing the improvements of the modified algorithm. Through a visual comparison of the intermediate meshes generated through the adaptation procedure for transonic turbulent flow over an RAE2822 airfoil, the modified algorithm was shown to much more smoothly transition the initial mesh to an optimal mesh with shock and boundary layer resolution. Next, an inviscid supersonic flow over an NACA0012 airfoil demonstrated the improved mesh-convergence behavior of the gradient-based optimization method relative to the baseline heuristic algorithm. Finally, to demonstrate the overall robustness of the solver, the resulting discretization and modified adaptation procedure was applied to a range of two-dimensional and three-dimensional turbulent aerodynamic problems. A subsonic turbulent flow over a three-element McDonnell Douglas Aerospace airfoil demonstrated the ability to robustly transition from a initial coarse mesh to a turbulent boundary layer mesh for both p = 1 and p = 2. The optimal error convergence rates were observed for both solution orders studied. Next, the flow over a three-dimensional flat plate at subsonic, turbulent conditions was considered. This case exhibited the capability of the adaptation process to properly detect the lack of spatial variation in the spanwise direction and to automatically generate meshes with large amounts of anisotropy in this direction. Again, optimal error convergence rates were observed for both p = 1 and p = 2. Finally, the ability of the algorithm to capture three-dimensional boundary layer effects was demonstrated in a subsonic turbulent flow through a square duct. 6.2 Future Work Improvements to the local error model Throughout this work, the MOESS framework as presented in Chapter 4 assumes a linear error model. Potential benefits could be realized with the optimization formulation if a higher-order error model were used instead. This model could be constructed 78 by either obtaining more error samples or by reinterpretation of the samples already taken during the local solve procedure. Adaptation for significantly under resolved meshes The error estimate can significantly underestimate the error on coarse meshes on which solution features are completed under resolved. On these meshes, the rates calculated through the local sampling procedure can push the triangulation towards nonoptimal metric fields. This problem can be alleviated by either improving the robustness of the error estimate within the current sampling-based adaptation framework or by incorporating an adaptation strategy that adequately detects under-resolution and adjusts the error model accordingly. A brief view of this problem is provided in Appendix A. Improvements to the metric optimization statement The presented optimization method and respective solution procedure was shown to be able to generate optimized metric requests which were then passed to a mesh generator to ultimately lower the error estimate. However, this optimization statement contains no constraints or controls pertaining to the realizability of the resulting metric. Although the Riemannian metric field is represented as a piecewise linear function on the triangulation, the presented adaptation procedure can still request metrics that can be arbitrarily difficult for mesh generators to produce corresponding metric-conforming meshes. This potential problem may be remedied by formulat- ing additional constraints or post-processing steps to introduce requirements on the smoothness of the metric fields. Three-dimensional RANS simulations This work demonstrated the capability of a high-order adaptive RANS method to both efficiently and automatically resolve turbulent flows on select two-dimensional and three-dimensional aerodynamic flows. However, the method as presented still is not robust enough to tackle three-dimensional problems of arbitrary complexity. 79 Discontinuous Galerkin methods still present significant robustness challenges that must be overcome before they are effectively utilized to solve the more complex threedimensional flows and geometries encountered in industry. Furthermore, robust threedimensional metric-based meshing is still an active problem that also serves as an existing bottleneck to the realization of high-order adaptive RANS. 80 Appendix A Adaptation From Under-resolved Meshes The turbulent flat plate case presented in Section 5.2 yielded optimized meshes with significant refinement in the boundary layer and at the leading edge when starting with an initial mesh with decent boundary layer resolution. However, the characteristic decreases in the error estimates from iteration to iteration observed in this case is not retained when starting the procedure from a much coarser mesh with significant under-resolution of the boundary layer. This appendix details the effects of initializing the process with a significantly under-resolved mesh on a similar set of flat plate problems. A.1 Test Cases and Initial Meshes Similar to the case in Section 5.2, we consider a Mo, = 0.2, ReL = 5 x 106 turbulent flow over a flat plate. The baseline two-dimensional problem contains a flat plate of 2L preceded by a short lead-in section of length 0.3L. As before, the respective three-dimensional meshes are produced by extruding the two-dimensional meshes into the third dimension. Here, we study both the two-dimensional and three-dimensional cases. To study the effects of under-resolution, we initialize the adaptation procedure 81 with an isotropic mesh containing no boundary resolution and an anisotropic mesh that contains similar normal spacing to the initial mesh shown in Section 5.2. The isotropic mesh and the anisotropic mesh can see in Figures A-1 and A-2 respectively. For further comparison, we also study a scalar convection-diffusion problem with the same ReL on the same meshes where the state on the plate is set to zero and the freestream state is set to unity. We consider the solution order p = 1, with a target degrees of freedom equal to the degrees of freedom of the initial mesh (1380 for 2D and 11040 for 3D) with the drag (heat flux for the scalar case) on the walls of the plate as the output of interest and generate meshes starting from these initial structured meshes. - 1.5 o -> .0.5-IZZV 0 0 0.5 1 15 2 x Figure A-1: Initial isotropic mesh with no boundary layer resolution - 460 Elements A.2 Solution on Initial Meshes The resulting pu and scalar T wall normal profiles on the two initial meshes can be seen in Figures A-3 and A-4. For the isotropic mesh, the weak imposition of the no-slip boundary condition results in essentially slip wall behavior for all four cases. On the initial anisotropic mesh, the no-slip condition is more clearly enforced. 82 k 1.5 0.0004 0.5 00002 0 0.5 1 1.5 x 2 2.1 (a) Initial Mesh . 2. j.1 0 2.15 x 2.2 (b) Close up of near-wall resolution Figure A-2: Initial anisotropic mesh with significant boundary layer resolution - 460 Elements 0.3 0 0.25 E 0.2- 8r 0.15- 0 0 0 3D 2D 3 2D RANS - pu RANS - p u Cony 01 - T Cony Dili -T 0, CO0.1 0 Z 0.05- 0 0 0.2 0.4 0.6 0.8 1 p u, T Figure A-3: Initial pu, T wall normal profile, 1.5 units downstream of the leading edge, isotropic mesh A.3 Adaptive Results The adaptation results for the RANS cases can be seen in Figures A-5 and A-7. The respective DOF histories can be seen in Figures A-6 and A-8. When starting from the under-resolved initial mesh, the adaptation severely struggles to reach the optimal triangulations. Whereas the two-dimensional RANS case eventually flatlines at a minimized error estimate value after 50 adaptation iterations, the three-dimensional case requires upwards of 70 iterations to reach this similar error level. When starting from 83 0.3 x 0.25E 0 3D RANS - p u 2DRANS-pu 0 3D Conv Diff - T S20 onv Diff - 3D RANS - p u 2D RANS - p u 3D Conv Difl - T 2D Conv Ditt - T 0 .045 - 0 0 0.04 T E 0.2 0 S0.15 C 0 0 .035 0.03 -6 40 0 .025 0.02 0 750.1 0 .015 0 z 0.05 0 0.2 0.4 0.8 0.6 0 001 00 02 04 .5. 0 0.2 0.4 0.6 pu, T 1 pu, T i (a) Wall profiles 0.8 I (b) Close up of profiles Figure A-4: Initial pu, T wall normal profile, 1.5 units downstream of the leading edge, anisotropic mesh the anisotropic boundary-layer resolved mesh, the adaptation procedure converges more rapidly for both the three-dimensional and two-dimensional problems. For both 2D and 3D flows and both initial meshes, the adaptive algorithm is eventually able to push the triangulations towards optimality. x 10, -A 4 10' 30 RANSolrpc -3D t - 3.8 RANS - -0-- 2D RANS - Isoropic Isotropici 10, 3.6 3.4 3.2 0 0 3 10 2.8 2.6 10-s 2.4 2.2 0 10 20 30 40 50 60 70 80 Adaptation Iteration 0 10 20 30 40 50 60 70 Adaptation Iteration (a) Cd vs. adaptation iteration (b) Cd error estimate convergence Figure A-5: Drag adaptation history for turbulent boundary layer flow starting with isotropic mesh The adaptation results for the scalar cases can be seen in Figures A-9 and A-11. The respective DOF histories can be seen in Figures A-10 and A-12. For the two- dimensional scalar flat plate cases, starting with the under-resolved mesh does not 84 80 3000 3 x 10, ,'x, 2500- 2.5- A - - 1500 -1.5 o 2 - A 2000- 0 11 1000 - 0 10 20 30 40 50 60 --- 500 d1 70 80 0 10 20 30 Adaptation Iteration 40 50 60 di 70 80 Adaptation Iteration (a) 2D RANS flow with doftarget 1380 (b) 3D RANS flow with doftarget= 11040 Figure A-6: DOF adaptation history for turbulent boundary layer flow starting with isotropic mesh i X 10-3 [ - 4 e i i1 10 RANS - Anisropic 2D RANS - Anisolropic -3D - 3.8 - 3D RANS - Anisoropic 2DRANS-Anisotropici 10 3.6 0 3.4 3.2 m 10 2.6- ; 10 2.42.2 2 0 10 20 30 40 50 60 70 80 Adaptation Iteration 10 0 10 20 30 40 50 60 70 Adaptation Iteration (a) Cd vs. adaptation iteration (b) Cd error estimate convergence Figure A-7: Drag adaptation history for turbulent boundary layer flow starting with anisotropic mesh seem to adversely affect the mesh convergence as much as seen in the two-dimensional and three-dimensional turbulent flows. While the error indicator convergence does still require an additional 5-6 iterations for the isotropic mesh, this increase is significantly smaller than those observed for the RANS cases. In three-dimensions, starting with the isotropic mesh does require a significant increase in the iterations required to obtain error indicator convergence relative to starting with the anisotropic mesh. 85 80 -x 10, II 2.5 J 2500- I 2000- I 2 U- U- 0 0 0 0 1500 - 1.5 1 0 10 0 10 20 30 40 so 60 7 8 Adaptation Iteration 0 01 0 10 20 30 40 50 60 70 80 Adaptation Iteration (a) 2D RANS flow with doftarget = 1380 (b) 3D RANS flow with doftarget 11040 Figure A-8: DOF adaptation history for turbulent boundary layer flow starting with anisotropic mesh 2 WI (. - 3D Conv Diff -4--~~~ 20 Conv Diff-Iorpc- . 1()-2 Isopic - -o 10, 10! 10, 1 . 10-2 -+- 3D Conv Di - soropic 2D Conv Difi - :sotropic- 10, - 10 I 0-100 I 1 5 L 10 15 20 A 25 I 30 35 40 Adaptation Iteration 10 I 0 5 10 15 20 25 30 35 Adaptation Iteration (a) Heat flux vs. adaptation iteration (b) Heat flux error estimate convergence Figure A-9: Heat flux adaptation history for scalar boundary layer flow starting with isotropic mesh Further investigation is still required to understand why the adaptation scheme on the RANS equations (especially in three-dimensions) struggles significantly more relative to the scalar equations when starting from an under-resolved mesh. The problem is likely to be a combination of the weaker control of DOF in three-dimensional mesh generation as well as more fundamental robustness issues of the error estimates for the RANS equations. 86 40 . x 10 . suaa. 2500 2.5 2000 2 U- U- 0 0 0 0 1.5 1000 01 0 5 10 15 20 25 30 35 05 '0 40 Adaptation Iteration (a) 2D scalar flow with doftarget = 1380 5 10 15 20 25 Adaptation Iteration 30 35 440 (0 (b) 3D scalar flow with doftarget =11040 Figure A-10: DOF adaptation history for scalar boundary layer flow starting with isotropic mesh 16 x 10 e - 12 3D Conv Dil - Anisoiropic -3D Conv _g.--2D Conv . 10-2 2D Conv Dili - Anisotropic Dill - Anisoropic, Dill - Anisotropic- - 14- 10~ 10 UCO 8 10, 6 4 20 5 10 15 20 25 Adaptation Iteration 30 35 40 10~-I 0 5 10 15 ' 20 ' 25 i 30 1 35 Adaptation Iteration (a) Heat flux vs. adaptation iteration (b) Heat flux error estimate convergence Figure A-11: Heat flux adaptation history for scalar boundary layer flow starting with anisotropic mesh 87 40 x 3000 2500- 10 2.5- 2 2000LL U- 0 0 0 0 1500 1.5 1000- 500 0 10 15 20 25 30 35 40 Adaptation Iteration 0 5 10 15 20 25 30 35 Adaptation Iteration (b) 3D scalar flow with doftarget (a) 2D scalar flow with doftarget = 1380 11040 Figure A-12: DOF adaptation history for scalar boundary layer flow starting with anisotropic mesh 88 40 Bibliography [1] Mark Ainsworth and Bill Senior. An adaptive refinement strategy for hp-finite element computations. Appl. Numer. Math., 26:165-178, 1998. [2] Steven R. Allmaras, Forrester T. Johnson, and Philippe R. Spalart. Modifications and clarifications for the implementation of the Spalart-Allmaras turbulence model. Seventh international conference on computational fluid dynamics, Big Island, Hawaii, 2012. [3] D. N. Arnold. An interior penalty finite element method with discontinuous elements. SIAM J. Numer. Anal., 19:742-760, 1982. [4] T. J. Baker. Mesh adaptation strategies for problems in fluid dynamics. Finite Elements Anal. Design, 25:243-273, 1997. [5] F. Bassi, A. Crivellini, S. Rebay, and M. Savini. Discontinuous Galerkin solution of the Reynolds averaged Navier-Stokes and k-w turbulence model equations. Comput. & Fluids, 34:507-540, May-June 2005. [6] F. Bassi and S. Rebay. A high-order discontinuous finite element method for the numerical solution of the compressible Navier-Stokes equations. J. Comput. Phys., 131:267-279, 1997. [7] F. Bassi and S. Rebay. GMRES discontinuous Galerkin solution of the compressible Navier-Stokes equations. In Karniadakis Cockburn and Shu, editors, Discontinuous Galerkin Methods: Theory, Computation and Applications, pages 197-208. Springer, Berlin, 2000. [8] F. Bassi and S. Rebay. Numerical evaluation of two discontinuous Galerkin methods for the compressible Navier-Stokes equations. Internat. J. Numer. Methods Fluids, 40:197-207, 2002. [9] R. Becker and R. Rannacher. A feed-back approach to error control in finite element methods: Basic analysis and examples. East-West J. Numer. Math., 4:237-264, 1996. [10] R. Becker and R. Rannacher. An optimal control approach to a posteriori error estimation in finite element methods. In A. Iserles, editor, Acta Numerica. Cambridge University Press, 2001. 89 [11] H. Borouchaki, P. George, F. Hecht, P. Laug, and E Saltel. Mailleur bidimensionnel de Delaunay gouverne par une carte de metriques. Partie I: Algorithmes. INRIA-Rocquencourt, France. Tech Report No. 2741, 1995. [12] Frank J. Bossen and Paul S. Heckbert. A pliant method for anisotropic mesh generation. In 5th Intl. Meshing Roundtable, pages 63-74, Oct. 1996. [13] Nicholas K. Burgess and Dimitri J. Mavriplis. An hp-adaptive discontinuous Galerkin solver for aerodynamic flows on mixed-element meshes. AIAA 2011490, 2011. [14] G. Chavent and G. Salzano. A finite element method for the iD water flooding problem with gravity. J. Comput. Phys., 42:307-344, 1982. [15] B. Cockburn, S. Hou, and C. W. Shu. TVB Runge-Kutta local projection discontinuous Galerkin finite element method for conservation laws IV: The multidimensional case. Math. Comp., 54:545-581, 1990. [16] B. Cockburn, S. Y. Lin, and C. W. Shu. TVB Runge-Kutta local projection discontinuous Galerkin finite element method for conservation laws III: One dimensional systems. J. Comput. Phys., 84:90-113, 1989. [17] B. Cockburn and C. W. Shu. TVB Runge-Kutta local projection discontinuous Galerkin finite element method for scalar conservation laws II: General framework. Math. Comp., 52:411-435, 1989. [18] B. Cockburn and C. W. Shu. The Runge-Kutta discontinuous Galerkin finite element method for conservation laws V: Multidimensional systems. J. Comput. Phys., 141:199-224, 1998. [19] David L. Darmofal, Steven R. Allmaras, Masayuki Yano, and Jun Kudo. An adaptive, higher-order discontinuous galerkin finite element method for aerodynamics. AIAA CFD Conference, 2013. [20] Timothy A. Davis. Algorithm 832: UMFPACK V4.3 - An unsymmetric-pattern multifrontal method. ACM Transactions on mathematical software, 30(2):196199, 2004. [21] L. Demkowicz, Ph. Devloo, and J. T. Oden. On an h-type mesh-refinement strategy based on minimization of interpolation errors. Comput. Methods Appl. Mech. Engrg., 53:67-89, 1985. [22] Laslo T. Diosady and David L. Darmofal. Preconditioning methods for discontinuous Galerkin solutions of the Navier-Stokes equations. J. Comput. Phys., 228:3917-3935, 2009. [23] Julien Dompierre, Marie-Gabrielle Vallet, Yves Bourgault, Michel Fortin, and Wagdi G. Habashi. Anisotropic mesh adaptation: towards user-independent, mesh-independent and solver-independent cfd. part iii. unstructured meshes. Internat. J. Numer. Methods Fluids, 39(8):675-702, 2002. 90 [24] Krzysztof J. Fidkowski and David L. Darmofal. An adaptive simplex cut-cell method for discontinuous Galerkin discretizations of the Navier-Stokes equations. AIAA 2007-3941, 2007. [25] L. Formaggia, S. Micheletti, and S. Perotto. Anisotropic mesh adaptation with applications to CFD problems. In H. A. Mang, F. G. Rammerstorfer, and J. Eberhardsteiner, editors, Fifth World Congress on ComputationalMechanics, Vienna, Austria, July 7-12 2002. [26] Luca Formaggia, Simona Perotto, and Paolo Zunino. An anisotropic a-posteriori error estimate for a convection-diffusion problem. Comput. Visual Sci., 4:99-104, 2001. [27] Wagdi G. Habashi, Julien Dompierre, Yves Bourgault, Djaffar Ait-Ali-Yahia, Michel Fortin, and Marie-Gabrielle Vallet. Anisotropic mesh adaptation: towards user-independent, mesh-independent and solver-independent CFD. part I: general principles. Internat. J. Numer. Methods Fluids, 32:725-744, 2000. [28] Paul Houston and Endre Sili. A note on the design of hp-adaptive finite element methods for elliptic partial differential equations. Comput. Methods Appl. Mech. Engrg., 194:229-243, 2005. [29] Steven G. Johnson. The nlopt nonlinear-optimization package. ab-initio .mit . edu/nlopt. http:// [30] Kelly R. Laflin, John C. Vassberg, Richard A. Wahls, Joseph H. Morrison, Olaf Brodersen, Mark Rakowitz, Edward N. Tinoco, and Jean-Luc Godard. Summary of data from the second AIAA CFD drag prediction workshop. AIAA 2004-0555, 2004. [31] P. LeSaint and P. A. Raviart. On a finite element method for solving the neutron transport equation. In C. de Boor, editor, MathematicalAspects offinite elements in partial differential equations, pages 89-145. Academic Press, 1974. [32] David W. Levy, Thomas Zickuhr, John Vassberg, Shreekant Agrawal, Richard A. Wahls, Shahyar Pirzadeh, and Michael J. Hemsch. Data summary from the First AIAA Computational Fluid Dynamics Drag Prediction Workshop. Journal of Aircraft, 40(5):875-882, 2003. [33] Adrien Loseille and Fred6ric Alauzet. Optimal 3D highly anisotropic mesh adaptation based on the continuous mesh framework. In Proceedings of the 18th InternationalMeshing Roundtable, pages 575-594. Springer Berlin Heidelberg, 2009. [34] Adrien Loseille and Frederic Alauzet. Continuous mesh framework part I: Wellposed continuous interpolation error. SIAM J. Numer. Anal., 49(1):38-60, 2011. [35] Adrien Loseille and Frederic Alauzet. Continuous mesh framework part II: Val- idations and applications. SIAM J. Numer. Anal., 49(1):61-86, 2011. 91 [36] D. J. Mavriplis. Results from the 3rd Drag Prediction Workshop using the NSU3D unstructured mesh solver. AIAA 2007-256, 2007. [37] Todd Michal and Joshua Krakos. Anisotropic mesh adaptation through edge primitive operations. AIAA 2012-159, 2012. [38] Todd A. Oliver. A Higher-Order, Adaptive, Discontinuous Galerkin Finite Element Method for the Reynolds-averaged Navier-Stokes Equations. PhD thesis, Massachusetts Institute of Technology, Department of Aeronautics and Astronautics, June 2008. [39] Todd A. Oliver and David L. Darmofal. An analysis of dual consistency for discontinuous Galerkin discretization of source terms. ACDL report, Massachusetts Institute of Technology, 2007. [40] Michael A. Park. Anisotropic Output-Based Adaptation with Tetrahedral Cut Cells for Compressible Flows. PhD thesis, Massachusetts Institute of Technology, Department of Aeronautics and Astronautics, 2008. [41] Xavier Pennec, Pierre Fillard, and Nicholas Ayache. A Riemannian framework for tensor computing. Int. J. Comput. Vision, 66(1):41-66, 2006. [42] J. Peraire, M. Vahdati, K. Morgan, and 0. C. Zienkiewicz. Adaptive remeshing for compressible flow computations. J. Comput. Phys., 72:449-466, 1987. [43] Per-Olof Persson and Jaime Peraire. Curved mesh generation and mesh refine- ment using Lagrangian solid mechanics. AIAA 2009-0949, 2009. [44] W. H. Reed and T. R. Hill. Triangular mesh methods for the neutron transport equation. Technical Report LA-UR-73-479, Los Alamos Scientific Laboratory, 1973. [45] G. R. Richter. An optimal-order error estimate for the discontinuous Galerkin method. Math. Comp., 50:75-88, 1988. [46] P. L. Roe. Approximate Riemann solvers, parameter vectors, and difference schemes. J. Comput. Phys., 43(2):357-372, 1981. [47] Youcef Saad and Martin H. Schultz. GMRES: A generalized minimal residual algorithm for solving nonsymmetric linear systems. SIAM Journal on Scientific and Statistical Computing, 7(3):856-869, 1986. [48] P. Solin and L. Demkowicz. Goal-oriented hp-adaptivity for elliptic problems. Computer Methods in Applied Mechanics and Engineering, 193:449-468, 2004. [49] Philippe R. Spalart and Steven R. Allmaras. A one-equation turbulence model for aerodynamics flows. AIAA 1992-0439, January 1992. 92 [50] Endre Siili and Paul Houston. Adaptive finite element approximation of hyperbolic problems. In T.J. Barth, M. Griebel, D. E. Keyes, R. M. Nieminen, D. Roose, and T. Schlick, editors, Lecture Notes in Computational Science and Engineering: Error Estimation and Adaptive DiscretizationMethods in Computational Fluid Dynamics, volume 25. Springer, Berlin, 2002. [51] Krister Svanberg. A class of globally convergent optimization methods based on conservative convex separable approximations. SIAM Journal on Optimization, pages 555-573. [52] John C. Vassberg, Edward N. Tinoco, Mori Mani, Olaf P. Brodersen, Bernhard Eisfeld, Richard A. Wahls, Joseph H. Morrison, Tom Zickuhr, kelly R. Laflin, and Dimitri J. Mavriplis. Summary of the Third AIAA CFD Drag Prediction Workshop. AIAA 2007-260, 2007. [53] John C. Vassberg, Edward N. Tinoco, Mori Mani, Ben Rider, Tom Zickuhr, David W. Levy, Olaf P. Brodersen, Bernhard Eisfield, Simone Crippa, Richard A. Wahls, Joseph H. Morrison, Dimitri J. Mavriplis, and Mithuhiro Murayama. Summary of the Fourth AIAA CFD Drag Prediction Workshop. AIAA Paper 2010-4547, 2010. [54] D. A. Venditti and D. L. Darmofal. Adjoint error estimation and grid adaptation for functional outputs: Application to quasi-one-dimensional flow. J. Comput. Phys., 164(1):204-227, 2000. [55] D. A. Venditti and D. L. Darmofal. Anisotropic grid adaptation for functional outputs: Application to two-dimensional viscous flows. J. Comput. Phys., 187(1):22-46, 2003. [56] Z.J. Wang, Krzysztof Fidkowski, Rmi Abgrall, Francesco Bassi, Doru Caraeni, Andrew Cary, Herman Deconinck, Ralf Hartmann, Koen Hillewaert, H.T. Huynh, Norbert Kroll, Georg May, Per-Olof Persson, Brain van Leer, and Miguel Visbal. High-order cfd methods: current status and perspective. International Journalfor Numerical Methods in Fluids, 72(8):811-845, 2013. [57] M. Wheeler. An elliptic collocation-finite element method with interior penalties. SIAM J. Numer. Anal., 15:152-161, 1978. [58] Masayuki Yano. An Optimization Framework for Adaptive Higher-Order Discretizations of Partial Differential Equations on Anisotropic Simplex Meshes. PhD thesis, Massachusetts Institute of Technology, Department of Aeronautics and Astronautics, June 2012. [59] Masayuki Yano and David Darmofal. An optimization framework for anisotropic simplex mesh adaptation: Application to aerodynamic flows. AIAA 2012-0079, January 2012. 93 [60] Masayuki Yano and David L. Darmofal. An optimization-based framework for anisotropic simplex mesh adaptation. J. Comput. Phys., 231(22):7626-7649, September 2012. 94