Computational aerodynamics, as a branch of the computational fluid dynamics (CFD), is the latest entry among the tools of the

aerodynamic sciences, barely 40 years old. In the 1950s, with matured numerical methods and rudimental computers (by today's standards !), the first solutions were attempted (ex. simple laminar boundary layers); late in the 1960s the first panel method solutions of practical problems (airfoils and wings) could be produced in relatively short time. At the present date, computations are routinely performed with multi-block Navier-Stokes solvers for problems as complex are fighter aircraft in complete configurations (Agarwal, 1999), military aircraft with rotor-fuselage interactions, multi-stage turbomachinery, etc.

importance of Aerodynamic Theory

Experimental aerodynamics is historically the oldest approach to aerodynamics. It was only when the Wright brothers invented the wind tunnel at the beginning of the century that rational data could be gathered and analyzed. Theoretical methods appeared at a later time (late 1910s) to provide information, usually in a closed mathematical form, that could be applied to general cases. Theoretical research that set outstanding guidelines include Kutta and Joukowsky (airfoil theory), Prandtl (boundary layer), Munk ( thin airfoil, airship theory), Theodorsen (unsteady airfoil theory), Wieselberger-Betz (lifting line theories), Glauert (propeller theory), von Karman (vortex street), Hayes (linearized supersonic theory), Whitcomb (transonic area rule), and many others. With the invention of the digital computer (also called digital wind tunnel) the range of possibilities has been steadily increasing. The computer itself created a new branch of aerodynamics and applied mathematics: numerical methods.

Importance of Experimental Methods

Experimental methods are still considered the ultimate answer to aerodynamic problems, although they require expensive equipment and can be cumbersome from the point of view of the model accuracy. Theoretical methods are not very useful unless the equations can be solved in general cases and provide quantitative answers. Back in the days when there was no computing facility calculations were performed by hand. Old papers showed plenty of tables of data (few a few decimal digits) resulting from hand computations. Numerics is the approach we follow more and more. Since numerical models are also tied the the computer resources available, it has become a field of fast progress in the last twenty years. One of the fundamental events in the development of numerical methods appeared in a study of the late 1920s. The paper due to Courant, Friedrichs and Lewy set the basis of stability of the solution to certain type of PDEs by numerical methods (CFL condition).

Fluid Dynamic Models

With respect to speed and viscosity, the computational methods are governed by different sets of equations. Incompressible low speed flows are governed by the Laplace equation for the velocity potential, that is an equation of elliptic type. The problem is closed with Dirichlet or Neumann boundary conditions (or both). Methods of solution are either the boundary element methods (panel methods) or finite element methods. Both can be written in discrete form using the Green identity for the Laplace equation. The equations for subsonic compressible inviscid flows are still elliptic as long as the entire flow field is subsonic.They switch to hyperbolic in the supersonic pockets defined by the shock waves. Methods of solution have been developed to get around this difficulty. The equations for inviscid supersonic flows are of the hyperbolic type, therefore they require initial data in the time-like direction. Viscous flows are of a wide variety of their own. The classical boundary layer equations (Prandtl) are of parabolic type and require initial conditions at one end of the integration domain.

Numerical Methods
Numerical methods for CFD are mostly concerned with the solution of system of partial differential equations, but the field is so broad as to include convergence acceleration methods (multi-grid, relaxation, artificial viscosity, etc.), stability control, pre-conditioning. The equations must be classified prior to attempting their solution. Besides countless articles in archival journals, there are entire books (references below) devoted only to numerical schemes for a wide range of equations, therefore it is not the purpose of this simple note. l the methods are broadly classified into finite differences, finite volumes, and finite element.

Methods fow Potential Flows

A fundamental development for transonic potential flows is the upwind differencing, to overcome the symmetry of the solution obtained by central schemes (this scheme in fact is unable to distinguish a flow reversal). Central differencing can be used in subsonic flows and the resulting system solved by a line relaxation process. The switching, however, may violate the conservation of mass in presence of shock waves. The problem is solved by

introducing appropriate artificial viscosity. Due to the slow convergence characteristics of the numerical method, several acceleration techniques have been proposed and applied: These include factorization of the difference operators, and multi-grid techniques.

Methods for the Euler Equations

Most numerical techniques for the Euler equations rely on time-dependent formulations to reach a steady-state solution (but there are also non-linear Newton-Raphson methods). This strategy allows the treatment of steady and unsteady flows with the same formulation and non-oscillatory schock-capturing schemes. Time-stepping techniques of the early days include the McCormack scheme. For steady state solutions the way the calculation procedes is inessential, therefore any fast time-stepping will do it. Current methods are classified as explicit (Lax-Wendroff, multi-stage methods, such as leap-frog, Adams-Bashfort and Runge-Kutta) and implicit (LU decomposition, alternating-direction ADI, relaxation). The space discretization needs dissipation terms, in order to damp numerical oscillations and to have more effective shock-capturing at transonic speeds. Dissipative schemes come in a wide variety, that include total variation diminishing (TVD), with or without flux limiting, flux-vector splitting, etc.

Methods for the Navier-Stokes Equations

Methods for the full Navier-Stokes equations are of the same variety as for the Euler equation. The additional viscous terms can be approximated by one adequate numerical differentiation technique. The NSE that are usually solved are in fact Reynolds-averaged equations, due to the large disparity of length scales between turbulence and integration domain. At the Reynolds numbers of interest in aerodynamics flows become easily turbulent and closure equations are needed. This is presently one of the great difficulties in CFD.

Parallel Computing
Given the sheer size of many CFD problems arising in industrial environments, aerodynamic components and processes, the step from sequential to parallel/vector programming is a necessary one. This requires fundamental changes in the hardware, in the language compiler, besides rational computer programming (the latter one to gain the maximum advantage from both hardware and compilers).

Hardware/Software
Parallel computers and clusters of sequential (single-processor) computers have been made available. One of the main ideas being pursued is the multiple-instruction multiple data (MIMD) processing. The parallel CFD consists in distributing grid blocks to N different processors (nodes); performing CFD computations on each node; and finally combining the results from N nodes. The goal is to achieve linear speed up of the computer codes (a code shared by N processors wuould be N times faster.)

Meshing
Meshing
Introduction Mesh classification Structured mesh generation Algebraic grid generation Elliptic grid generation

Unstructured mesh generation Special topics Mesh adaptation

First cell height calculation

Introduction
The partial differential equations that govern fluid flow and heat transfer are not usually amenable to analytical solutions, except for very simple cases. Therefore, in order to analyze fluid flows, flow domains are split into smaller subdomains (made up of geometric primitives like hexahedra and tetrahedra in 3D and quadrilaterals and triangles in 2D). The governing equations are then discretized and solved inside each of these subdomains. Typically, one of three methods is used to solve the approximate version of the system of equations: finite volumes, finite elements, or finite differences. Care must be taken to ensure proper continuity of solution across the common interfaces between two subdomains, so that the approximate solutions inside various portions can be put together to give a complete picture of fluid flow in the entire domain. The subdomains are often called elements or cells, and the collection of all elements or cells is called a mesh or grid. The origin of the term mesh (or grid) goes back to early days of CFD when most analyses were 2D in nature. For 2D analyses, a domain split into elements resembles a wire mesh, hence the name. An example of a 2D analysis domain (flow over a backward facing step) and its mesh are shown in pictures below.

The process of obtaining an appropriate mesh (or grid) is termed mesh generation (or grid generation), and has long been considered a bottleneck in the analysis process due to the lack of a fully automatic mesh

generation procedure. Specialized software progams have been developed for the purpose of mesh and grid generation, and access to a good software package and expertise in using this software are vital to the success of a modeling effort.

|| Introduction | Mesh Classification >>