This action might not be possible to undo. Are you sure you want to continue?

# Lecture notes on Computational Fluid Dynamics

Dan S. Henningson Martin Berggren January 13, 2005

Contents

1 Derivation of the Navier-Stokes equations 1.1 Notation . . . . . . . . . . . . . . . . . . . . 1.2 Kinematics . . . . . . . . . . . . . . . . . . 1.3 Reynolds transport theorem . . . . . . . . . 1.4 Momentum equation . . . . . . . . . . . . . 1.5 Energy equation . . . . . . . . . . . . . . . 1.6 Navier-Stokes equations . . . . . . . . . . . 1.7 Incompressible Navier-Stokes equations . . 1.8 Role of the pressure in incompressible ﬂow . 2 Flow physics 2.1 Exact solutions . . . . . . . . 2.2 Vorticity and streamfunction 2.3 Potential ﬂow . . . . . . . . . 2.4 Boundary layers . . . . . . . 2.5 Turbulent ﬂow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 7 8 14 15 19 20 22 24 29 29 32 40 48 54 59 59 62 65 66 69 71 71 72 73 74 76 76 80 82 83 85 85 86 87 88 88

3 Finite volume methods for incompressible ﬂow 3.1 Finite Volume method on arbitrary grids . . . . . 3.2 Finite-volume discretizations of 2D NS . . . . . . 3.3 Summary of the equations . . . . . . . . . . . . . 3.4 Time dependent ﬂows . . . . . . . . . . . . . . . 3.5 General iteration methods for steady ﬂows . . . .

4 Finite element methods for incompressible ﬂow 4.1 FEM for an advection–diﬀusion problem . . . . . . . . . . 4.1.1 Finite element approximation . . . . . . . . . . . . 4.1.2 The algebraic problem. Assembly. . . . . . . . . . 4.1.3 An example . . . . . . . . . . . . . . . . . . . . . . 4.1.4 Matrix properties and solvability . . . . . . . . . . 4.1.5 Stability and accuracy . . . . . . . . . . . . . . . . 4.1.6 Alternative Elements, 3D . . . . . . . . . . . . . . 4.2 FEM for Navier–Stokes . . . . . . . . . . . . . . . . . . . 4.2.1 A variational form of the Navier–Stokes equations 4.2.2 Finite-element approximations . . . . . . . . . . . 4.2.3 The algebraic problem in 2D . . . . . . . . . . . . 4.2.4 Stability . . . . . . . . . . . . . . . . . . . . . . . . 4.2.5 The LBB condition . . . . . . . . . . . . . . . . . . 4.2.6 Mass conservation . . . . . . . . . . . . . . . . . . 4.2.7 Choice of ﬁnite elements. Accuracy . . . . . . . . . A Background material A.1 Iterative solutions to linear systems . A.2 Cartesian tensor notation . . . . . . A.2.1 Orthogonal transformation . A.2.2 Cartesian Tensors . . . . . . A.2.3 Permutation tensor . . . . . . A.2.4 Inner products, crossproducts A.2.5 Second rank tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . and determinants . . . . . . . . . . . . 3 . . . . . . .

95 . 95 . 97 . 98 . 99 . 100 . 100 . 100

. . . . . . . . 109 . . . .2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 B. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .A.2. . . 104 B Supplementary material 107 B. . . .7 Gauss & Stokes integral theorems . . 102 A. . . . . . . . . . .1 Syllabus 2005 . . . . . . . . .8 Archimedes principle . . . . . . . .2 Study questions .6 Tensor ﬁelds . . . . . .2. . . . . . . . . . . 103 A. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

August 2004 Dan Henningson Martin Berggren . Erik St˚ alberg and Ori Levin has typed most of the L TEXformulas and has created the electronic versions of most ﬁgures.Preface These lecture notes has evolved from a CFD course (5C1212) and a Fluid Mechanics course (5C1214) at the department of Mechanics and the department of Numerical Analysis and Computer Science (NADA) A at KTH. Stockholm.

.

from 1 to 3.Chapter 1 Derivation of the Navier-Stokes equations 1. v. ∂x1 ∂x2 ∂x3 = ∂2 ∂2 ∂2 + 2+ 2 2 ∂x ∂y ∂z and the following deﬁnitions are used ν ρ p − kinematic viscosity − density − pressure see ﬁgure 1.1 for a deﬁnition of the coordinate system and the velocity components. This implies that a repeated index is summed over. The Cartesian tensor form of the equations can be written ∂ui + u ∂ui j ∂t ∂xj ∂ui ∂xi = − 0 1 ∂p ∂ 2 ui +ν ρ ∂xi ∂xj ∂xj = where the summation convention is used. u3 ) the nabla operator is deﬁned as = the Laplace operator is written as 2 ∂ ∂ ∂ .1 Notation ∂u + (u · ∂t ·u 1 p+ν ρ The Navier-Stokes equations in vector notation has the following form )u = = − 0 2 u where the velocity components are deﬁned u = (u. ∂x ∂y ∂z = ∂ ∂ ∂ . w) = (u1 . . as follows ui ui = u 1 u1 + u 2 u2 + u 3 u3 Thus the ﬁrst component of the vector equation can be written out as 7 . u2 . .

Lagrange coordinates: every particle is marked and followed in ﬂow. t x1 Figure 1.2.ˆ ˆ ui xi0 . u1 xi0 v.e. x2 ˆ P t=0 z. i=1 ∂u1 ∂u1 1 ∂p ∂u1 ∂u1 + u1 + u2 + u3 =− +ν ∂t ∂x1 ∂x2 ∂x3 ρ ∂x1 ∂u ∂u ∂u ∂u 1 ∂p +u +v +w =− +ν ∂t ∂x ∂y ∂z ρ ∂x ∂ 2 u1 ∂ 2 u1 ∂ 2 u1 + + ∂x1 2 ∂x2 2 ∂x3 2 or ∂2u ∂2u ∂2u + 2 + 2 ∂x2 ∂y ∂z 2u 1. t and the velocity of the particle is the rate of change of the particle position. x3 Figure 1.1: Deﬁnition of coordinate system and velocity components PSfrag replacements x. x1 w. u2 w. is . We begin by considering the motion of a ﬂuid particle in Lagrangian coordinates. u3 ri xi 0 t . where the particle path of P. x3 u. x1 2 x y.y. u1 PSfrag replacements x. i. The independent variables are xi 0 ˆ t − initial position of ﬂuid particle − time ˆ ri = r i xi 0 .2: Particle path. x2 v.2 Kinematics Lagrangian and Euler coordinates Kinematics is the description of motion without regard to forces. u3 z. the coordinates familiar from classical mechanics. u2 u. see ﬁgure 1.

t) = FE ri xi . As an example we consider the acceleration of a ﬂuid particle in a steady converging river.3.e. ﬂuid ﬂows past point. The raft which is following the ﬂuid is accelerating although the ﬂow ﬁeld is steady. Consider the quantity F following ﬂuid particle. This has been experienced by everyone in a raft in a converging river. Euler coordinates: consider ﬁxed point in space. i.4 during time dt. we still need to consider the rate of change of quantities following a ﬂuid particle. is easily found by noting that the particle position is expressed in ﬁxed space coordinates xi . t). The position of P2 can by Taylor expansion be expressed as P2 : ˆ ˆ ri dt + dri dt = ˆ = ri (0) + ∂rti (0) dt + dri (0) + ∂ˆ ∂ ˆ ( dri ) ∂t 0 ˆ dt ˆ ˆ = xi 0 + dxi 0 + ui (0) dt + dui (0) dt . t and (xi . ˆ The relation between Lagrangian and Euler coordinates. t) is now considered as a function of the coordinate xi and time t.e. see ﬁgure 1. t Material derivative t = ˆ t Although it is usually most convenient to use Euler coordinates. This leads to the following deﬁnition. t The rate of change of F following a ﬂuid particle can then be written ∂FE ∂ri ∂FE ∂t ∂FE ∂F ∂FE = + = · · + ui ˆ ˆ ˆ ∂xi ∂ t ∂t ∂ t ∂t ∂xi ∂t D as Based on this expression we deﬁne the material derivative Dt ∂ ∂ ∂ ∂ D = + ui = + (u · ) ≡ ˆ Dt ∂t ∂xi ∂t ∂t In the material or substantial derivative the ﬁrst term measures the local rate of change and the second measures the change due to the motion with velocity ui . Material derivative: rate of change in time following ﬂuid particle expressed in Euler coordinates. t = FE (xi . The independent variables are xi t − − space coordinates time Thus the ﬂuid velocity ui = ui (xi . ˆ at the time xi = r i xi 0 . The acceleration is deﬁned aj = Dui ∂ui ∂ui = + uj Dt ∂t ∂xj which can be simpliﬁed in 1D for stationary case to ∂u ∂u Du = +u Dt ∂t ∂x Note that the acceleration = 0 even if velocity at ﬁxed x does not change. i. xi 0 . t . a= Description of deformation Evolution of a line element ˆ Consider the two nearby particles in ﬁgure 1.ui = ∂ri ˆ ∂t Note here that when xi 0 changes we consider new particles. where ˆ ˆ F = FL xi 0 . Instead of marking every ﬂuid particle it is most of the time more convenient to use Euler coordinates.

u3 x1 x2 xi0 ˆ ri xi0 . x2 z. x1 y.3: Acceleration of ﬂuid particles in converging river. t ˆ P t=0 x u1 u2 > u 1 Figure 1.4: Relative motion of two nearly particles. t 0 ˆ ui xi . u1 v. u3 x1 x2 xi0 ˆ ri xi0 . dr i x1 dtˆ . u2 w. x2 z.PSfrag replacements x. t ˆ P t=0 x u1 u2 > u 1 x2 dxi0 d ui t dˆ P2 ˆ ui dt P1 xi0 ˆ t ri d x3 Figure 1. x3 u. PSfrag replacements x. t 0 ˆ ui xi . x1 y. u1 v. x3 u. u2 w.

e. no deformation. diﬀerence in velocity ∂ ( dri ) = dui in Lagrange coordinates to an equation for a material ˆ ∂t D ( dri ) Dt = = = dui {expand in Euler coordinates} ∂ui drj ∂xj where ∂ui ∂xj drj − − change in velocity with spatial position diﬀerence in spatial pos. describes the deformation and is considered in detail below.where we have used ∂ ( dri ) ˆ ∂t = = ∂ ∂ ∂ri dxj 0 = ˆ ∂xj 0 ∂ t ∂xj 0 ∂ui dxj 0 = dui ∂xj 0 ∂ri ˆ ∂t dxj 0 and where ∂ui ∂xj 0 dxj 0 dui We can transform the expression line element in Euler coordinates − − − change of ui with initial pos. eij . ∂xj ∂xj ∂ui = ξij + eij + eij ∂xj eij where eij is the deformation rate tensor and ξij eij eij = = = ∂uj 1 ∂ui anti-symmetric part − 2 ∂uj ∂xi 1 ∂ui ∂uj 2 ∂ur traceless part + − δij 2 ∂xj ∂xi 3 ∂ur 1 ∂ur δij isotropic part 3 ∂ur ∂ui . of particles Relative motion associated with invariant parts We consider the relative motion dui = ∂ui ∂ui drj by dividing in its invariant parts. diﬀerence in initial pos. i. whereas the ∂xj anti-symmetric part can be written in terms of the vorticity ωk and is associated with solid body rotation. The anti-symmetric part can be written The symmetric part of .e. i.

dxi0 ˆ dui dt ˆ r i dt ˆ dri dt P1 P2 r2 x2 R2 π 2 + dϕ12 R1 R x3 r3 Figure 1. where we deﬁne Rl k l rk = = Rδkl Rl + k component k of side l ∂uk l R dt ∂xj j dul k relative motion of side l = R δkl + ∂uk dt ∂xl deformed cube First. ωk = 1 2 ω2 1 − 2 ω1 kji ∂ ui } ∂xj − 1 ω3 2 0 1 2 ω1 0 Consider the deformation of the small cube in ﬁgure 1.5: Deformation of a cube. we consider the deformation on side 1. which can be expressed as dR1 = r1 − R = The inner product can be expanded as ∂u1 dt ∂x1 2 1 1 rk rk − R 1 1 rk rk = 1+ + ∂u2 dt ∂x1 2 + ∂u3 dt ∂x1 2 R2 = {drop quadratic terms} = R2 1 + 2 ∂u1 dt ∂x1 .5. 3 x1 r1 [ξij ] = 0 ∂u1 ∂x2 ∂u1 ∂x3 1 −2 −1 2 1 2 ∂u1 ∂x2 − − ∂u2 ∂x1 ∂u3 ∂x1 − 0 ∂u2 ∂x1 1 2 1 2 ∂u1 ∂x3 ∂u2 ∂x3 − − 0 ∂u3 ∂x1 ∂u3 ∂x2 1 −2 ∂u2 ∂x3 − ∂u3 ∂x2 = {use ω = 0 1 = 2 ω3 1 2 ω2 Deformation of a small cube × u .

We use the trigonometric identity cos π π π + dϕ12 = cos · cos dϕ12 − sin · sin dϕ12 ≈ − dϕ12 2 2 2 which allow us to obtain the ﬁnial expression dϕ12 = −2e12 dt Thus the deformation rate of angle between side 1 and side 2 depends only on traceless part of Third. Second. This can be expressed as r1 r2 r3 − R 3 1+ = R3 ∂u1 ∂x1 dt ∂u2 ∂x1 dt ∂u3 ∂x1 dt ∂ui ∂xj . dR 1 becomes ∂u1 ∂u1 dt − R = R 1 + dt + . This can be expressed as cos π + dφ12 2 = = = = r1 · r 2 1 2 1 1 2 2 = r k rk · rm rm · r n rn |r1 | |r2 | δk1 + ∂u1 ∂x2 ∂u1 ∂x2 −1/2 −1/2 ∂uk ∂uk ∂u1 dt δk2 + dt · 1 + 2 dt ∂x1 ∂x2 ∂x1 ∂u2 ∂u1 ∂u2 dt + dt 1− dt 1− dt ∂x1 ∂x1 ∂x2 ∂u2 + dt = 2e12 dt ∂x1 1+2 ∂u2 dt ∂x2 −1/2 where we have dropped quadratic terms. we consider the deformation of the angle between side 1 and side 2. − R ∂x1 ∂x1 dR1 = = R R 1+2 ∂u1 dt = Re11 dt ∂x1 which implies that dR1 = Re11 dt ∂ui Thus the deformation rate of side 1 depends on e11 .. In summary..We have dropped the quadratic terms since we are assuming that dt is small. both traceless and isotropic part of ∂xj . we consider the deformation of the volume of the cube. Thus we have dV = R3 err dt ∂ui and the deformation rate of volume of cube (or expansion rate) depends on isotropic part of ∂xj . dV = 1 ∂u1 ∂x2 dt 2 + ∂u2 dt ∂x ∂u3 ∂x2 dt 1 ∂u1 ∂x3 dt ∂u2 ∂x3 dt 3 + ∂u3 dt ∂x − R3 = = = R3 1 + ∂u2 ∂u1 dt 1+ dt ∂x1 ∂x2 ∂u1 ∂u2 ∂u3 dt R3 + + ∂x1 ∂x2 ∂x3 ∂uk R3 dt ∂xk 1+ ∂u3 dt − R3 ∂x3 where we have again omitted quadratic terms. the motion of a ﬂuid particle with velocity ui can be divided into the following invariant parts .

6: Volume moving with the ﬂuid. where a volume element describing the change in volume between V at time t and t + ∆t can be written as u · n∆t = uk nk ∆t ⇒ dV = uk nk ∆t dS ¡¡¡¡¢ ¢ ¢ ¢¡¢¡¢¡¢¡ ¡¢¡¢¡¢¡ ¡¢¡ ¡ ¢ ¢¡ ¡ ¡ ¡¢ ¡ ¡ ¡ ¡¢ ¡ ¡¢ ¡¡ ¡¢¡¢¡¢¡¡¢¡¢¡¢¡¡¢¡¡¢ ¡ ¡ ¡ ¡¢¡ ¡ ¡ ¡¢¡ ¡¢¡¡ ¢¡¡¡¡ ¡¡¡¡ ¡¡ ¡ ¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡ ¢ ¢¡¢ ¡¢ ¡¢ ¡ ¡¢ ¡¢ ¡¢ ¡ ¡¢ ¡ ¡ ¢¡ ¢¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡¡ ¡¡¡¡¢¡¡¡¡¢¡¡¢¡ ¡¢¡¢¡¢¡ ¡¢¡¢¡¢¡ ¡¢¡ ¡¢ ¢¡¡¡¡¢¡¡¢¡ ¢¡¢ ¡¢ ¡¢ ¡¡¢ ¡¢ ¡¢ ¡¡¢ ¡¡ ¢¡ ¢¡ ¡ ¡ ¡¢¡ ¡ ¡ ¡¢¡ ¡¢¡¡ ¡¡¡¡ ¢¡¢¡¢¡¢¡ ¡¢¡¢¡¢¡ ¡¢¡ ¡¢ ¡ ¡ ¡ ¡¢¡ ¡ ¡ ¡¢¡ ¡¢¡¡ ¡¢¡¢¡¢¡ ¡¢¡¢¡¢¡ ¡¢¡ ¡ ¢¢¡ ¢¡¡¡¡¢¡¡¢¡ ¢¡ ¢ ¡ ¢ ¡ ¢ ¡¡ ¢ ¡ ¢ ¡ ¢ ¡¡ ¢ ¡¡¡ ¡¡¡¡¢¡¡¡¡¢¡¡¢¡ ¢¢¡¡¡¡¢¡¡¡¡¢¡¡¢¡ ¡¢¡¢¡¢¡ ¡¢¡¢¡¢¡ ¡¢¡ ¡ ¢¢¡ ¡ ¢ ¡ ¢ ¡ ¢ ¡¡ ¢ ¡ ¢ ¡ ¢ ¡¡ ¢ ¡¡¡ ¢¡¢¡¢¡¢¡ ¢ ¢ ¢ ¢ ¢ ¢ ¢ ¢ ¢¡ ¢ ¡ ¢ ¡ ¢ ¡¢¡ ¢ ¡ ¢ ¡ ¢ ¡¢¡ ¢ ¡¢¡¡ ¢¡¡¡¡ ¢¡¡ ¢¡ ¢¡ ¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡ ¢ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡¡ ¢ ¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡¢¡ ¢ ¡¢ ¡¢ ¡¢ ¡ ¡¢ ¡¢ ¡¢ ¡ ¡¢ ¡ ¡ ¢¡ ¢¡¡¡¡ ¢¡¡¡¡ ¢¡¡ ¢¡ ¢¡ ¡ ¢ ¡ ¢ ¡ ¢ ¡¡ ¢ ¡ ¢ ¡ ¢ ¡¡ ¢ ¡¡¡ V (t+∆t) V (t + ∆t) − V (t) S (t + ∆t) (t) u n V (t + ∆t) solid body translation solid body rotation volume constant deformation volume expansion rate Tij (t + ∆t) dV − V (t) 1 V (t+∆t) Tij (t + ∆t) dV − V (t) Tij (t) dV Tij (t + ∆t) dV .3 Reynolds transport theorem Volume integral following with the ﬂuid Consider the time derivative of a material volume integral. a volume integral where the volume is moving with the ﬂuid. i.e. We obtain the following expressions 1 lim ∆t→0 ∆t ∆t→0 ∆t D Dt V (t) Tij dV = = lim = 1 + Tij (t + ∆t) dV − Tij (t) dV ∆t V (t) V (t) 1 ∂Tij lim dV Tij (t + ∆t) dV + ∆t→0 ∆t ∂t V (t+∆t)−V (t) V (t) The volume in the ﬁrst integral on the last line is represented in ﬁgure 1.π 2 R3 R2 R1 r3 r2 r1 + dϕ12 V (t) S Figure 1.6. i) ii) iii) iv) ξij = eij = 1 2 ui ∂ui ∂uj − ∂xj ∂xi 1 = −2 kij ωk 1 2 ∂uj 2 ∂ur ∂ui + − δij ∂xj ∂xi 3 ∂ur eij = 1 3 ∂ur δij ∂ur 1.

1. eq. that the time rate of change of momentum in a material region = sum of the forces on that region. i.) B0 ¡ ! ∂ ∂tij ∂ ∂tij Dtij ∂ρ ∂ ¨¨ (ρtij ) + (uk ρ tij ) = tij ¡ + ρ + tij ¨¨ k ρ) + ρuk (u =ρ ∂t ∂xk ∂t ∂t ∂x ∂xk Dt ¡ ¨k which implies that Reynolds transport theorem becomes D Dt V (t) ρ tij dV = V (t) ρ Dtij dV Dt where V (t) again is a material volume.e. The quantities involved are: Fi Ri ρui body forces per unit mass surface forces per unit area momentum per unit volume . We have D Dt V (t) ρ dV = V (t) ∂ ∂ρ + (uk ρ) = 0 ∂t ∂xk Since the volume is arbitrary. The terms in the expression can be given the following interpretations: 1 2 3 4 : : : : accumulation of mass in ﬁxed element net ﬂow rate of mass out of element rate of density change of material element volume expansion rate of material element By considering the transport of a quantity given per unit mass. This surface integral can in turn be changed back to a volume integral by the use of Gauss (or Greens) theorem. i.This implies that the volume integral can be converted to a surface integral. Conservation of mass By the substitution Tij → ρ and the use of the Reynolds transport theorem above we can derive the equation for the conservation of mass. We have D Dt V (t) Tij dV = ∆t→0 lim Tij (t + ∆t) uk nk dS + S(t) V (t) ∂Tij dV ∂t = S(t) Tij uk nk dS + V (t) ∂Tij dV = {Gauss/Green’s theorem} ∂t = V (t) ∂Tij ∂ + (uk Tij ) dV ∂t ∂xk which is the Reynolds transport theorem. the following must hold for the integrand 0= ∂ ∂ρ ∂uk ∂ρ ∂ρ ∂uk Dρ + +ρ (uk ρ) = + uk +ρ = ∂t ∂xk ∂t ∂xk ∂xk Dt ∂xk 1 2 3 4 where we have used the deﬁnition of the material derivative in order to simplify the expression.e. Tij = ρtij . The integrand in the theorem can then be written 0 (cont. we can simplify Reynolds transport theorem.4 Momentum equation Conservation of momentum The momentum equation is based on the principle of conservation of momentum.

V (t + ∆t) − V (t) ˆ r i dt nIII(k) S (t + ∆t) ˆ dri dt u dl P1 n P2 V (t + ∆t) x1 R(nI ) x2 x3 R3 nII nI R2 R1 dS r3 2 R(nII ) r r1 π + dϕ12 2 Figure 1. Here R(n) is the surface force per unit area on surface dS with normal n. . The stress tensor Remove a ﬂuid element and replace outside ﬂuid by surface forces as in ﬁgure 1.8. irrespective of volume forces or acceleration terms. Thus the surface forces acting on a small ﬂuid particle has to balance. In order to do that we have to deﬁne the stress tensor.7: (t) V Momentum balance for ﬂuid element S (t) V (t + ∆t) − V (t) n S (t + ∆t) u n R V (t + ∆t) R Figure 1. To proceed Ri must be investigated so that the surface integral can be transformed to a volume integral. We can put the momentum conservation in integral form as follows D Dt V (t) ∆s (k) ρui dV = V (t) ρFi dV + S(t) Ri dS Using Reynolds transport theorem this can be written ρ V (t) Dui dV = Dt V (t) ρFi dV + S(t) Ri dS which is Newtons second law written for a volume of ﬂuid: mass · acceleration = sum of forces. Momentum conservation for the small ﬂuid particle leads to ρ Dui dS dl = ρFi dS dl + Ri nI dS + Ri nII dS + j j Dt III(k) ∆s(k) dl Ri nj k Letting dl → 0 gives 0 = Ri nI dS + Ri nII dS j j Now nj = nI = −nII which leads to j j Ri (nj ) = −Ri (−nj ) implying that a surface force on one side of a surface is balanced by an equal an opposite surface force at the other side of that surface. see ﬁgure 1. Note that it is a general principle that the terms proportional to the volume of a small ﬂuid particle approaches zero faster than the terms proportional to the surface area of the particle.7.8: Surface force and unit normal.

9: Deﬁnition of surface force components on a surface with a normal in the 1-direction. x2R1 z. i ˆ dxi0 ˆ dui dt We now divide the surfaceˆforces into components along the coordinate directions. Momentum balance require the surface forces to balance R3 in the element. T32 ) + dϕ w. x3 r3 Figure 1. x1R y. Tij is the i-component t the surface force on a surface dS with a normal in the j-direction. u1 r2 x2 v. as in ﬁgures 1. t ∆t) (t + ˆ T12 ui xi0 . The areas of the surface elements are related by x1 x2 dSj = ej · n dS = nj dS x3 where dS is the area of the slanted surface. as in P2 ﬁgure 1.10: Deﬁnition u dt of surface force components on a surface with a normal in the 2-direction.9 and r i dt 1. u3 12 2 V x1 (t) T22 S x2 (t) V (t + ∆t) −xi0 (t) V ˆ ri Sxi0 . T22 . t u ˆ= 0 n P t V (t + ∆t) x T32 u1 u2 > u 1 x1 x2 x3 xi0 Figure 1. with corresponding deﬁnitions for the force components on a surface with a normal in the 3-direction.11.10.11: Surface force balance on a ﬂuid particle with a slanted surface. . u. T21 . T31 ) n V (t + ∆t) P2 x1 T21 x2 x1 T11 x3 T31 PSfrag replacementsR3 2 x. of P1 Consider a ﬂuid particle with surfaces along the coordinate directions cut by a slanted surface.u P1 R(e1 ) = ( T11 . we have R2 R1 3 r0 = R1 dS − T11 n1 dS − T12 n2 dS − T13 n3 dS r2 which implies that the total surface force Ri can be written in terms of the components of the stress r1 tensor Tij as π + dϕ12 2 V (t) R = T n + T n + T n = T n i i1 1 i2 2 i3 3 ij j S (t) V (t + ∆t) − V (t) e2 S (t + ∆t) R u n n V (t + ∆t) R1 dS1 dS2 dS3 e1 e3 Figure 1. u2 r1 π R(e2 ) = ( T12 . dri dˆ Thus.

viscous stress tensor. often=0 For a Newtonian ﬂuid.e. directed inward V (t + ∆t) − V (t) τij . shear Further consideration of the moment balance around a ﬂuid particle show that T ij is ax symmetric tensor. We thus divide the r stress tensor into an isotropic part.e. i. the hydrodynamic pressure p. For an isotropic ﬂuid. Recall that the invariant symmetric parts are eij = 1 2 1 ∂ur ∂ui ∂uj 2 ∂ur + − δij + δij ∂xj ∂xi 3 ∂xr 3 ∂xr eij eij where eij is the volume constant deformation rate and eij is the uniform rate of expansion. u2 > u 1 x1 Tij = Tji x2 x3 Momentum equation xi0 ˆ i dt Using the deﬁnition of the surface force in terms of the stress tensor. We have π + dϕ12 p 2 V (t) Tij = −pδij + τij S (t) p . and a part depending2 on the motion of r1 the ﬂuid.hydrodynamic pressure.P ˆ Here the diagonal components are normal stresses oﬀ-diagonal components are t = 0 stresses. u1 i. we thus have the following relationship between the viscous stress and the strain (deformation rate) τij = 2µeij = µ which leads to the momentum equation ρ ∂p ∂ Dui =− + µ Dt ∂xi ∂xj ∂ui ∂uj 2 ∂ur + − δij ∂xj ∂xi 3 ∂xr + ρFi ∂uj 2 ∂ur ∂ui + − δij ∂xj ∂xi 3 ∂xr . We have x1 x2 ∂Tij Dui = ρFi + ρ x3 Dt ∂xj R3 R2 Pressure and viscous stress tensor R1 r3 For ﬂuid at rest only normal stresses present otherwise ﬂuid element would deform. the viscous stress tensor is a linear function of the invariant parts of e ij . the momentumuequation can now be dxi0 written ˆ dui dt ˆ r i dt ∂Tij Dui dV ˆ dV = ρFi dV + Tij nj dS = ρFi + ρ Dt ∂xj dri dt V (t) S(t) V (t) V (t) P1 P2 The volume is again arbitrary implying that the integrand must itself equal zero. τij = λeij + 2µeij where we have deﬁned the two viscosities as µ (T ): λ (T ): dynamic viscosity (here T is the temperature) second viscosity. depends on ﬂuid motion S (t + ∆t) u n V (t + ∆t) p p Newtonian ﬂuid It is natural to assume that the viscous stresses are functions of deformation rate e ij or strain.

directed outward Using Reynolds transport theorem we can put the energy conservation in integral form as D Dt V (t) 1 ρ e + ui ui dV = 2 V (t) ρFi ui dV + S(t) [ni Tij uj − ni qi ] dS = V (t) ρFi ui + ∂ (Tij uj − qi ) dV ∂xi ˙ Compare the expression in classical mechanics. We have the following deﬁnitions 1 ρ e + 2 ui ui dV energy of particle.5 Energy equation The energy equation is a mathematical statement which is based on the physical law that the rate of change of energy in material particle = rate that energy is received by heat and work transfers by that particle.e. with e the internal energy work rate of Fi on particle force · velocity ρui Fi dV uj Rj dS ni qi dS work rate of Rj on particle heat loss from surface. We obtain ρ D 1 ∂τij ∂p ui ui = ρFi ui − ui + ui Dt 2 ∂xi ∂xj Thermal energy equation is then found by subtracting the mechanical energy equation from the total energy equation. We have ∂ ∂ ∂qi 1 D (pui ) + (τij uj ) − e + ui ui = ρFi ui − Dt 2 ∂xi ∂xi ∂xi The mechanical energy equation is found by taking the dot product between the momentum equation and u. with qi the heat ﬂux vector. ) From the integral energy equation we obtain the total energy equation by the observation that the volume is arbitrary and thus that the integrand itself has to be zero. ρ De ∂ui ∂ui ∂qi = −p + τij − Dt ∂xi ∂xj ∂xi The work of the surface forces divides into viscous and pressure work as follows ρ − ∂ (pui ) ∂xi = −p ∂ui ∂xi − ui ∂p ∂xi 2 1 ∂ (τij uj ) ∂xi = τij ∂ui ∂xj + uj ∂τij ∂xi where following interpretations can be given to the thermal and the mechanical terms 1: 2: thermal terms ( force · deformation ): heat generated by compression and viscous dissipation mechanical terms ( velocity · force gradients ): gradients accelerate ﬂuid and increase kinetic energy .1. i. where the momentum equation is m u = F and the associated kinetic energy equation is (u · u) work rate ( work m d 2 dt = = = F·u force · velocity force · dist.

the equation describing the conservation of momentum and the equation describing the conservation of energy.The heat ﬂux need to be related to the temperature gradients with Fouriers law qi = −κ ∂T ∂xi where κ = κ (T ) is the thermal conductivity. This allows us to write the thermal energy equation as ρ De ∂ui ∂ = −p +Φ+ Dt ∂xi ∂xi κ ∂T ∂xi where the positive deﬁnite dissipation function Φ is deﬁned as ∂ui 1 = 2µ eij eij − ∂xj 3 1 ∂uk δij 3 ∂xk 2 Φ = = τij ∂uk ∂xk 2 2µ eij − >0 Alternative form of the thermal energy equation can be derived using the deﬁnition of the enthalpy h=e+ We have Dh De 1 Dp p = + − 2 Dt Dt ρ Dt ρ Dρ Dt − ρ ∂xi i ∂u p ρ which gives the ﬁnal result ρ Dp ∂ Dh = +Φ+ Dt Dt ∂xi κ ∂T ∂xi To close the system of equations we need a i) ii) thermodynamic equation equation of state e = e (T. p) p = ρRT simplest case: e = cv T or h = cp T where cv and cp are the speciﬁc heats at constant volume and temperature. At this time we also deﬁne the ratio of the speciﬁc heats as γ= cp cv 1. We have Dρ ∂uk +ρ =0 Dt ∂xk ρ ρ where ∂p ∂τij Dui =− + + ρFi Dt ∂xi ∂xj κ ∂T ∂xi ∂ui ∂ De = −p +Φ+ Dt ∂xi ∂xi .6 Navier-Stokes equations The derivation is now completed and we are left with the Navier-Stokes equations. They are the equation describing the conservation of mass. respectively.

unknown ρ 1 ui 3 p 1 e 1 T 1 7 equations continuity momentum energy thermodyn. ρu3 . This form of the equation is usually termed the conservative form of the Navier-Stokes equations. ρu1 . ρF2 . ρui Fi ) is the vector of the right hand sides and G(i) = T T ∂ ∂ Dtij = (ρtij ) + (uk ρtij ) Dt ∂t ∂xk κ ∂T ∂xi These are all of the form or ∂U ∂G ∂G ∂G + + + =J ∂t ∂x ∂y ∂z ρui ρu1 ui + pδ1i − τ1i ρu2 ui + pδ2i − τ2i ρu3 ui + pδ3i − τ3i ∂T ρ (e + ui ui /2) ui + pui + κ ∂xi − uj τij is the ﬂux of mass. momentum and energy. ρF1 . Φ and the thermodynamic relation and the equation of state for a gas e = e (T. gas law 1 3 1 1 1 7 τij ∂ui ∂xj ∂uj 2 ∂ur ∂ui + − δij = µ ∂xj ∂xi 3 ∂xr = τij Equations in conservative form A slightly diﬀerent version of the equations can be found by using the identity ρ to obtain the system ∂ρ ∂ ∂t + ∂x (uk ρ) = 0 k ∂ ∂ ∂p ∂τij (ρui ) + (ρuk ui ) = − + + ρFi ∂xk ∂xi ∂xj ∂t ∂ ∂ 1 ∂ ∂ ∂ 1 = ρFi ui − (pui ) + (τij uj ) + ∂t e + 2 ui ui + ∂xk ρuk e + 2 ui ui ∂xi ∂xi ∂xi ∂U ∂G(i) + =J ∂t ∂xi where U = (ρ. ρu2 . p) p = ρRT We have 7 equations and 7 unknowns and therefore the necessary requirements to obtain a solution of the system of equations. J = (0. ρF3 . ρ (e + ui ui /2)) is the vector of unknowns. . respectively.

1.7

Incompressible Navier-Stokes equations

The conservation of mass and momentum can be written Dρ + ρ ∂ui = 0 Dt ∂xi Dui ∂p ∂τij ρ =− + + ρFi ∂xi ∂xj Dt τ = µ ∂ui + ∂uj − 2 ∂uk δ ij ij ∂xj ∂xi 3 ∂uk

for incompressible ﬂow ρ = constant, which from the conservation of mass equation implies

∂ui =0 ∂xi implying that a ﬂuid particle experiences no change in volume. Thus the conservation of mass and momentum reduce to ∂ui + uj ∂ui = − 1 ∂p + ν 2 ui + Fi ∂t ∂xj ρ ∂xi ∂ui ∂xi = 0 since the components of the viscous stress tensor can now be written ∂ ∂xj ∂ui ∂uj 2 ∂uk + − δij ∂xj ∂xi 3 ∂xk = ∂ 2 ui = ∂xj ∂xj

2

ui

We have also introduced the kinematic viscosity, ν, above, deﬁned as ν = µ/ρ The incompressible version of the conservation of mass and momentum equations are usually referred to as the incompressible Navier-Stokes equations, or just the Navier-Stokes equations in case it is evident that the incompressible limit is assumed. The reason that the energy equation is not included in the incompressible Navier-Stokes equations is that it decouples from the momentum and conservation of mass equations, as we will see below. In addition it can be worth noting that the conservation of mass equation is sometimes referred to as the continuity equation. The incompressible Navier-Stokes equations need boundary and initial conditions in order for a solution to be possible. Boundary conditions (BC) on solid surfaces are ui = 0. They are for obvious reasons usually referred to as the no slip conditions. As initial condition (IC) one needs to specify the velocity ﬁeld at the initial time ui (t = 0) = u0 . The pressure ﬁeld does not need to be speciﬁed as it can be obtained once the i velocity ﬁeld is speciﬁed, as will be discussed below.

**Integral form of the Navier-Stokes eq.
**

We have derived the diﬀerential form of the Navier-Stokes equations. Sometimes, for example when a ﬁnite-volume discretization is derived, it is convenient to use an integral form of the equations. An integral from of the continuity equation is found by integrating the divergence constraint over a ﬁxed volume VF and using the Gauss theorem. We have ∂ (ui ) dVF = ∂xi ui ni dSF = 0

S

V

An integral form of the momentum equation is found by taking the time derivative of a ﬁxed volume integral of the velocity, substituting the diﬀerential form of the Navier-Stokes equation, using the continuity equation and Gauss theorem. We have ∂ ∂t

VF

ui dV =

VF

∂ui dV = − ∂t

VF

**1 ∂p ∂ 2 ui ∂ (uj ui ) + −ν ∂xj ρ ∂xi ∂xj ∂xj dV = −
**

SF

dV =

−

VF

1 ∂ui ∂ uj ui + pδij − ν ∂xj ρ ∂xj

p ∂ui ui uj nj + ni − ν nj dS ρ ∂xj

Tw

T0 , U 0

L

Figure 1.12: Examples of length, velocity and temperature scales. Note that the integral from of the continuity equation is a compatibility condition for BC in incompressible ﬂow.

Dimensionless form

It is often convenient to work with a non-dimensional form of the Navier-Stokes equations. We use the following scales and non-dimensional variables x∗ = i xi L t∗ = tU0 L u∗ = i ui U0 p∗ = p p0 ∂ 1 ∂ = ∂xi L ∂x∗ i U0 ∂ ∂ = ∂t L ∂t∗ ⇒

where ∗ signiﬁes a non-dimensional variable. The length and velocity scales have to be chosen appropriately from the problem under investigation, so that they represent typical lengths and velocities present. See ﬁgure 1.12 for examples. If we introduce the non-dimensional variables in the Navier-Stokes equations we ﬁnd 2 ∗ U 2 ∂u∗ p0 ∂p∗ νU0 ∂ 2 u∗ U0 ∂ui i · ∗ + 0 u∗ i = − + 2 j L ∂t L ∂x∗ ρL ∂x∗ L ∂x∗ ∂x∗ j i j j U0 ∂u∗ i · =0 L ∂x∗ i ∂ui + uj ∂ui = − p0 ∂p + ν 2 ∂t ∂xj ρU0 ∂xi U0 L ∂ui =0 ∂xi

2 U0 L U0 /L = ν νU0 /L2 2

2 Now drop * as a sign of non-dimensional variables and divide through with U 0 /L, we ﬁnd

We have deﬁned the Reynolds number Re as Re =

“inertial forces” “viscous forces”

which can be interpreted as a measure of the inertial forces divided by the viscous forces. The Reynolds number is by far the single most important non-dimensional number in ﬂuid mechanics. We also deﬁne the pressure scale as

2 p0 = ρU0

which leaves us with the ﬁnal form of the non-dimensional incompressible Navier-Stokes equations as ∂ui + uj ∂ui = − ∂p + 1 ∂t ∂xj ∂xi Re ∂ui =0 ∂xi

2

¥ ¥ ¥ ¥ £¥ ¥ ¥ ¥ ¥ ¥ £¤£¤£¤£¤¤£¤£¤£¤£¤£¤ ¤¤¤¤¤¤¤¤¤¤£¥

T0 U0

S (t + ∆t) u n V (t + ∆t) p

Tw L

§ § § § ¦§ § § § § § ¦¤¦¤¦¤¦¤¤¦¤¦¤¦¤¦¤¦¤ ¤¤¤¤¤¤¤¤¤¤¦§

ui ui

**Non-dimensional energy equation
**

We stated above that the energy equation decouple from the rest of the Navier-Stokes equations for incompressible ﬂow. This can be seen from a non-dimensionalization of the energy equation. We use the deﬁnition of the enthalpy (h = cp T ) to write the thermal energy equation as DT Dp = + Φ + κ 2T Dt Dt and use the following non-dimensional forms of the temperature and dissipation function ρcp T∗ = This leads to DT ∗ Dt∗ T − T0 Tw − T 0 Φ∗ = L2 2 Φ U0 µ

= =

µ κ · · ρU0 L µcp 1 1 · · Re P r

∗2

∗2

T∗ +

T∗ +

2 U0 cp (Tw − T0 ) M a2 c

a2 0 p (Tw −T0 )

2 U0 µ Dp∗ + φ∗ cp (Tw − T0 ) Dt∗ ρU0 L

Dp∗ 1 + Φ∗ Dt∗ R

where a0 speed of sound in free-stream and M a = to obtain

U0 is Mach number. We now drop ∗ and let M a → 0 a0

DT 1 1 = · · 2T Dt Re P r µc where the Prandtl number P r = κp , which takes on a value of ≈ 0.7 for air. As the Mach number approaches zero, the ﬂuid behaves as an incompressible medium, which can be seen from the deﬁnition of the speed of sound. We have a0 = γ , ρα α= 1 ∂ρ ρ ∂p

T

where α is the isothermal compressibility coeﬃcient. If the density is constant, not depending on p, we have that α → 0 and a0 → ∞, which implies that M a → 0.

1.8

Role of the pressure in incompressible ﬂow

The role of the pressure in incompressible ﬂow is special due to the absence of a time derivative in the continuity equation. We will illustrate this in two diﬀerent ways.

Artiﬁcial compressibility

First we discuss the artiﬁcial compressibility version of the equations, used sometimes for solution of the steady equations. We deﬁne a simpliﬁed continuity equation and equation of state as ∂ρ ∂ui + =0 and p = βρ ∂t ∂xi This allows us to write the Navier-Stokes equations as ∂ui + ∂ (uj ui ) = − ∂p + 1 ∂t ∂xj ∂xi Re ∂ui ∂p +β = 0 ∂t ∂xi Let p βu βv u p + u2 uv u= e= f = v uv p + v2 w uw vw

2

ui

βw uw g= vw p + β2

Any wi in Ω can uniquely be decomposed into wi = u i + where ∂ui = 0. Let wi = u i The inner product between ui gives 0= Ω (1) (1) + ∂p(1) ∂p(2) (2) = ui + ∂xi ∂xi − ui (2) + ∂ p(1) − p(2) ∂xi and ui (1) − ui (2) ui (1) − ui (2) 2 + ui (1) − ui (2) ∂ p(1) − p(2) ∂xi dV = Ω ui (1) − ui (2) 2 dV Thus we have ui (1) = ui . w.be the vector of unknowns and their ﬂuxes. B = ∂f = u 0 ∂u 0 u 0 0 β 0 0 v u 0 . ∂xi ∂p ∂xi on ∂Ω. into a function ui that is divergence free and parallel to the boundary and the gradient of a function. Proof. We start by showing that ui and ui . u i · ni = 0 i. u. u ± v. v. B.e. The Navier-Stokes equations can then be written ∂u ∂e ∂f ∂g 1 + + + = ∂t ∂x ∂y ∂z Re or 2 Du 2 ∂u ∂u ∂u 1 ∂u +A +B +C = ∂t ∂x ∂y ∂z R 0 0 0 0 . v2 + β where the matrices above are deﬁned 0 β 1 2u ∂e = A= ∂u 0 v 0 w 0 ∂g 0 C= = ∂u 1 0 0 0 1 0 0 1 0 0 0 v 0 0 0 0 1 0 β u 2v w 0 0 0 1 0 0 0 v Projection on a divergence free space Second we discuss the projection of the velocity ﬁeld on a divergence free space. D = 0 0 2v 0 0 w v 0 Du the eigenvalues of A. (2) p(1) = p(2) + C . We begin by the following theorem. w ± w2 + β √ Note that the eﬀective acoustic or pressure wave speed ∼ β → ∞ for incompressible ﬂow. u. ∂p ∂xi = Ω ∂p is orthogonal in the L2 inner product. they are and w. here called p. v ± u2 + β . ∂xi ∂ ∂p ui dV = (ui p) dV = pui ni dS = 0 ∂xi ∂xi Ω ∂Ω The uniqueness of the decomposition can be seen by assuming that we have two diﬀerent decompositions and showing that they have to be equivalent. Theorem 1. C are the wave speeds of plane waves in the respective coordinate directions.

P ∂p ∂xi ∂p =0 ∂xi Applying the orthogonal projector to the Navier-Stokes equations. ui is divergence free and parallel to boundary. i. Apply to Navier-Stokes equations Let P be orthogonal projector which maps wi on divergence free part ui .13: Projection of a function on a divergence free space. Pwi = ui .13. U 0 Tw L T0 U0 π 2 Gradient ﬁelds wi ∂p ∂xi ui Divergence free vectorﬁelds // to boundary Figure 1. We can ﬁnd an equation for a p with above properties by noting that 2 p= ∂wi ∂xi in Ω with ∂p = wi ni on ∂Ω ∂n has a unique solution. We then have the following relations wi = Pwi + Pui = ui . we have P ∂p ∂ui + ∂t ∂xi = P −uj ∂ui 1 + ∂xj Re 2 ui Now. to project wi on divergence free space let wi = ui + 1) solve for p : ∂wi = ∂xi ∂p ∂xi 2 p. Note also that (wi − ui ) ⊥ui . if ui = w i − we have 0= ∂ui ∂wi ∂2p = − . ∂p =0 ∂n 2) let ui = wi − This is schematically shown in ﬁgure 1.e. Now.r + dϕ12 V (t) S (t) V (t + ∆t) − V (t) S (t + ∆t) u n V (t + ∆t) p T0 . thus . ∂xi ∂xi ∂xi ∂xi ∂p ⇒ ∂xi 0 = u i ni = w i ni − ∂p ∂xi ∂p ∂n and To sum up.

1 ∂ui + −uj ∂xj Re wi 2 ∂uj ∂ui ui = − ∂xi ∂xj Evolution equation for the divergence It is common in several numerical solution algorithms for the Naiver-Stokes equations to discard the divergence constraint and instead use the pressure Poisson equation derived above as the second equation in the incompressible Naiver-Stokes equations. but satisﬁes a heat equation. A priori we have none speciﬁed. we have ∂ui ∂ui 1 = P −uj + ∂t ∂xj Re wi 2 The pressure term in the Naiver-Stokes equations ensures that the right hand side w i is divergence free. Thus. if and only if it is zero everywhere on the boundary. but they have to be chosen such that the divergence of the velocity ﬁeld is zero on the boundary. something seen in the analysis of the artiﬁcial compressibility equations. Since it is the pressure term in the Navier-Stokes equations ensures that the velocity is divergence free. For the stationary case the solution obeys the maximum principle. which links the velocity ﬁeld in the whole domain instantaneously. has its maximum on the boundary. This may be a diﬃcult constraint to satisfy in a numerical solution algorithm. the divergence is zero inside the domain. If this is done we may encounter solutions that are not divergence free. Thus we ﬁnd an evolution equation without the pressure. this has implications for the boundary conditions for the pressure Poisson equation. i. we have an inﬁnite wave speed for pressure waves. We ﬁnd 2 ui p= ∂ ∂xi thus the pressure satisﬁes elliptic Poisson equation. according to the result above. This can be interpreted such that the information in incompressible ﬂow spreads inﬁnitely fast.e. a function satisfying the Laplace equation. We can ﬁnd a Poisson equation for the pressure by taking the divergence of w i . Consider the equation for the evolution of the divergence. P ∂ui ∂ui = ∂t ∂t 2 ui = 2 ui since 2 ui need not be parallel to the boundary. We have ∂ ∂t ∂ui ∂xi = 1 Re 2 ∂ui ∂xi Thus the divergence is not automatically zero.P However. found by taking the divergence of the momentum equations and substituting the Laplacian of the pressure with the right hand side of the Poisson equation. This states that the maximum of a harmonic function. .

.

. is driven by a pressure diﬀerence p 0 −p1 between ˆ r i dt the inlet and the outlet of the channel. x3 u.Chapter 2 z. where the constant is seem to be zero from the V (t + ∆t) − V (t) boundary conditions. meaning that eﬀects of inlet conditions have disappeared. u3 = 0 P2 ∂t x1 and that the ﬂow is fully developed. i. see ﬁgure 2. t ˆ P t=0 Flowuxphysics 1 u2 > u 1 x1 x2 x3 2. u2 w.e. ˆ dri dt P1 ∂ = 0. steady ﬂow. by applying the assumptions above. S (t + ∆t) The steady momentum equations read u n ρu ∂u1 + ρu ∂u1 = − ∂p + µ 2 u V (t + ∆t) 1 2 1 ∂x1 ∂x2 ∂x1 p ρu ∂u2 + ρu ∂u2 = − ∂p + µ 2 u − ρg 2 2 1 T0 . We have x2 x3 u1 = u1 (x2 ) . reduce to L T0 U0 x2 L Gradient ﬁelds wi ∂p h ∂xi ui ree vectorﬁelds // to boundary p0 > p 1 x1 p ©¨©¨©¨©¨©©¨©¨©¨©©¨©©¨©©¨©¨©©¨©©©¨ ©©©©©©©©©©©©©©©©©©©¨©©¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ©¨©¨©¨©¨©©¨©¨©¨©©¨©©¨©©¨©¨©©¨©©©¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ©©©©©©©©©©©©©©©©©©©¨©©¨ ©©©©©©©©©©©©©©©©©©©©© 29 1 Figure 2.1: Plane channel geometry. t 0 ˆ ui xi .exact solution for channel ﬂow dxi0 ˆ dui dt The ﬂow inside the two-dimensional channel.1. U 0 ∂x1 ∂x2 ∂x2 Tw which.1 Exact solutions xi0 ˆ ui dt Plane Pouseuille ﬂow . u3 x1 x2 xi0 ˆ ri xi0 . The continuity equation reads R1 r3 0 0 r2 U ∂u2 U ∂u3 ∂u1 r1 + =0 + π ∂x ∂x2 3 ∂x + dϕ12 1 2 V (t) where the ﬁrst and the last term disappears due to the assumption of two-dimensional ﬂow and that S (t) the ﬂow is fully developed. We will assume two-dimensional. This implies that u2 = C = 0. u1 v. u2 = u2 (x2 ) R3 2 R The boundary conditions are u1 = u2 = 0 at x2 = ±h.

The velocity becomes x2 u1 =1− umax h Flow rate From ﬁgure 2. giving the parabolic velocity proﬁle u1 = h2 dP x2 · 1− 2µ dx1 h 2 Evaluating the maximum velocity at the center of the channel we have umax = − h2 dP · >0 2µ dx1 2 if ﬂow is in the positive x1 -direction. The momentum equation in the streamwise direction implies 0=− Since P (x1 ) and u1 (x2 ) we have 1 dP d2 u1 = const.2 we can evaluate the ﬂow rate as dQ = u · n dS. x2 ) = dx2 µ dx1 2 We can integrate u1 in the normal direction to obtain u1 = 1 dP 2 x + C 1 x2 + C 2 2µ dx1 2 d2 u1 dP +µ dx1 dx2 2 The constants can be evaluated from the boundary conditions u1 (±h) = 0.ui Divergence free vectorﬁelds // to boundary L h x1 x2 p0 > p 1 p1 dS u n u·n ∆t Figure 2.2: Volume of ﬂuid ﬂowing through surface dS in time ∆t 0=− ∂p d2 u1 +µ ∂x1 dx2 2 ∂p 0=− − ρg ∂x2 The momentum equation in the x2 -direction (or normal direction) implies p = −ρgx2 + P (x1 ) ⇒ ∂p dP = ∂x1 dx1 showing that the pressure gradient in the x1 -direction (or streamwise direction) is only a function of x1 . (independent of x1 . Integrating over the channel we ﬁnd h Q= S ui ni dS = {channel} = −h u1 dx2 = − 4 2h3 dP · = humax 3µ dx1 3 .

t) which using continuity and the boundary conditions imply that normal velocity v = 0. t) ∂2u ∂y 2 = 0 0≤y≤∞ = u0 t > 0 = ν = 0 t>0 We look for a similarity solution. We have τ12 = τ21 = µ implying that τwall = τ Vorticity The vorticity is zero in the streamwise and normal directions. 0) u(0.e. This is a time dependent problem where plate velocity is set to u = u0 at time t = 0. we introduce a new dependent variable which is a combination of y and t such that the partial diﬀerential equation reduces to an ordinary diﬀerential equation.3: Stokes instantaneously plate. i. where p0 is the atmospheric pressure at the plate surface. ω1 = ω2 = 0 and has the following expression in the x3 -direction (or spanwise direction) ω3 = ∂u1 du1 2umax · x2 ∂u2 − =− = ∂x1 ∂x2 dx2 h2 x2 =−h Stokes 1:st problem: instantaneously started plate Consider the instantaneously starting plate of inﬁnite horizontal extent shown in ﬁgure 2. Let du1 2µ · umax · x2 =− dx2 h2 2µ umax h = = = Figure 2. This gives us the following diﬀusion equation.e.L h x1 x2 p0 > p 1 p1 u n dS u · n∆t y u0 p0 x Wall shear stress The wall shear stress can be evaluated from the velocity gradient at the wall. We assume that the velocity can be written u = u(y. i. . The momentum equations reduce to ∂u ∂t 0 1 ∂p ∂2u +ν 2 ρ ∂x ∂y 1 ∂p −g − ρ ∂y − The normal momentum equation implies that p = −ρgy + p0 . initial and boundary conditions for u ∂u ∂t u(y. t) u(∞.3.

5. It is related to the average angular momentum of a ﬂuid particle and the swirl present in the ﬂow. 0) = f (∞) = 0 = f (0) = 1 u0 u0 This equation can be integrated twice to obtain η u(∞. we have η 2 2 f =1− √ eξ dξ = 1 − erf(η) π 0 This solution is shown in ﬁgure 2. as well as the coeﬃcients in front of the t-terms have to be equal. ∂v ∂u u0 −η2 − =√ e ∂x ∂y πνt The vorticity is shown in ﬁgure 2. we have bt−1 η bt−1 = κνC 2 t2b where κ is a proportionality constant. Note that the time dependent solution is diﬀusing upward. Note again that the time dependent solution is diﬀusing upward.2 Vorticity and streamfunction Vorticity is an important concept in ﬂuid dynamics. We have u(0. ω3 = ω z = 2. both as a function of the similarity variable η and the unscaled normal coordinate y. . We transform the time and space derivatives according to f (η) = ∂ ∂t ∂ ∂y ∂2 ∂y 2 = = = ∂η d d = Cbytb−1 ∂t dη dη ∂η d d = Ctb ∂y dη dη d2 C 2 t2b 2 dη If we substitute this into the equation for u and use the deﬁnition of η.u η = C · y · tb u0 where C and b are constants to be chosen appropriately. This gives 1 2 We choose κ = −2 and obtain the following b=− C=√ 1 −2κν f + 2ηf = 0 y η= √ 2 νt d where = dη .4. Here the maximum of the vorticity is also changing in time. a ﬂow with circular streamlines may have zero vorticity and a ﬂow with straight streamlines may have a non-zero vorticity. The boundary and initial conditions also have to be compatible with the similarity assumption. both as a function of the similarity variable η and the unscaled normal coordinate y. the coeﬃcient of the two η-dependent terms have to be proportional to each other. we ﬁnd an equation for f d2 f df = νC 2 t2b 2 dη dη where no explicit dependence on y and t can remain if a similarity solution is to exist. from the inﬁnite value at t = 0 it is diﬀusing outward in time. t) u(y. t) = f (∞) = 0 u0 f = C1 0 eξ dξ + C2 2 Using the boundary conditions and the deﬁnition of the error-function. However. The exponents of these expressions. From the velocity we can calculate the spanwise vorticity as a function of the similarity variable η and the unscaled normal coordinate y. Thus.

4: Velocity above the instantaneously started plate.8 0. which is deﬁned as Γ= C ui dxi = C ui ti ds In ﬁgure 2. Using Stokes theorem we can transform that integral to one over the area S as ∂uk ωi ni dS dS = Γ= ijk ni ∂xj S S This allows us to ﬁnd the following relationship between the circulation and vorticity dΓ = ω i ni dS.4 0.5 0.5 0.9 1 ωz πνt √ πνt u0 = ω (η) ˜ 0 0 0.2 0. Example 1.6 0. b) U (y.9 1 ω (η)·u0 ˜ √ πνt Vorticity and circulation The vorticity is deﬁned mathematically as the curl of the velocity ﬁeld as ω= In tensor notation this expression becomes ωi = ijk ×u ∂uk ∂xj Another concept closely related to the vorticity is the circulation.6 0.8 0.3 0.rﬁelds // to boundaryn Divergence free vectorﬁelds // to boundary n dS dS L L u · n∆t u · n∆t hx hx x1 y x1 y √2 xu x y η = √2 0 yp=> p 4νt η u0 t 4νt p0 > p 1 0 1 p0 p0 y p1 p1 η = √4νt u u √n n y = η 4νt dS dS t u u u · n∆t u · n∆t u0 u0 x x Figure 2.4 0.2 0.1 0.2 0 3 2 1 0 0.5: Vorticity above the instantaneously started plate. t) for various times. which can also be written dΓ = ω i ni dS Thus.4 1. The circulation of an ideal vortex.4 0.1 0.6 1.2 0.5 0.7 0.4 0. u0 u0 p0 p0 y y η = √4νt η = √4νt 2 8 1.2 1 0.8 1.6 √ y = η 4νt t η y ωz = ω (η)·u0 ˜ √ πνt u u0 √ y = η 4νt ty η u u0 7 6 5 4 3 t 2 0.9 1 2 8 1.8 0. t) for various times. b) ωz (y. a) ωz πνt/u0 as a function of the similarity variable η.7 0.6 0.7 0. a) U as a function of the similarity variable y y η.9 1 0 0 0.3 0. one interpretation is that the vorticity is the circulation per unit area for a surface perpendicular to the vorticity vector.8 0.6 0.2 1 7 6 5 4 0.6 1. Let the azimuthal velocity be given as u θ = C/r.4 0.8 0.5 0.8 0.2 0 t 1 0 0.3 0.8 1.6 0.1 0.4 1. We can calculate the circulation of this ﬂow as 2π Γ= C uθ r dθ = C 0 dθ = 2πC Thus we have the following relationship for the ideal vortex uθ = Γ 2πr .6 we show a closed curve C. ωz √ 0.2 0.7 0.1 0.4 0.3 ωz = = ω (η) ˜ u0 √ Figure 2.

We have ∂ui ∂ui ∂p 1 2 + uj =− + ui ∂t ∂xj ∂xi Re We slightly modify this equation by introducing the following alternative form of the non-linear term uj ∂ ∂ui = ∂xj ∂xi 1 uj uj 2 + ijk ωj uk ∂ pqi ∂xq ) this can readily be derived using tensor manipulation.6: Integral along a closed curve C with enclosed area S. q and ∂xq ∂xi is symmetric in ∂2 =0 ∂xq ∂xi The ﬁrst term and the last term can be written as derivatives of the vorticity. The ideal vortex has its name from the fact that its vorticity is zero everywhere. the vorticity equation becomes ∂ωp ∂ωp ∂up 1 + uk = ωj + ∂t ∂xk ∂xj Re or written in vector-notation Dω = (ω · Dt )u + 1 Re 2 2 ωp ω . Derivation of the Vorticity Equation We will now derive an equation for the vorticity. and we now have the following from of the vorticity equation ∂ 1 2 ∂ωp + pqi ( ijk ωj uk ) = ωp ∂t ∂xq Re The second term in this equation can be written as ipq ijk ∂ ∂ ∂ ∂ (ωj uk ) = (δpj δqk − δpk δqj ) (ωj uk ) = (ωp uk ) − (ωj up ) = ∂xq ∂xq ∂xk ∂xj uk ∂ωp ∂uk ∂ωj ∂up + ωp − up − ωj ∂xk ∂xk ∂xj ∂xj Due to continuity ∂ωj ∂xj = ∂ 2 uq jpq ∂xj ∂xp = 0 and we are left with the relation ipq ijk ∂ ∂ωp ∂up (ωj uk ) = uk − ωj ∂xq ∂xk ∂xj Thus. since i. except for an inﬁnite value at the center of the vortex. We take the curl ( equations and ﬁnd ∂ ∂t pqi of the momentum ∂ui ∂xq ∂ui ∂xq + pqi ∂2 ∂xq ∂xi 1 uj uj 2 + pqi ∂ ( ∂xq ijk ωj uk ) =− pqi ∂2p 1 + ∂xq ∂xi Re pqi 2 The second and the fourth terms are zero.√ u0 y = η 4νt t η y ˜ (η)·u ωz = ω√πνt0 t dxi = ti ds S C Figure 2. q pqi pqi is anti-symmetric in i. We start with the dimensionless momentum equations.

Example 2. [t] = T ⇒ [ νt] = L [ν] = T . ωy = 0. We have ωx = 0.7. where we have used the following dimensional relations √ L2 . Diﬀusion of vorticity: the Stokes 1st problem revisited.Components in Cartesian coordinates In cartesian coordinates the components of the vorticity vector in three-dimensions become ex ω= ∂ ∂x ey ∂ ∂y ez ∂ ∂z = u v w ∂w ∂v − ∂y ∂z ex + ∂u ∂w − ∂z ∂x ey + ∂u ∂v − ∂x ∂y ez For two-dimensional ﬂow it is easy to see that this reduces to a non-zero vorticity in the spanwise direction. where η = √ e πνt 2 νt √ where the factor νt can be identiﬁed as the diﬀusion length scale. ωz = ∂v ∂u − ∂x ∂y This simpliﬁes the interpretation and use of the concept of vorticity greatly in two-dimensional ﬂow. It is possible to show that × ω = 0 ⇐⇒ viscous forces = 0 This means that without viscous forces there cannot be a vorticity ﬁeld in the ﬂow which varies with the coordinate directions. and is given by u0 −η2 y ω=√ . but zero in the other two directions. This is expressed by the following relationship ∂τij =µ ∂xj 2 ui = −µ ijk ∂ωk ∂xj This can be derived in the following manner. t) and that it thus is a solution of the following diﬀusion equation ∂2ω ∂ω =ν 2 ∂t ∂y The time evolution of the vorticiy can be seen in ﬁgure 2. First the diﬀusion of vorticity and second the advection or transport of vorticity. From the solution of Stokes 1st problem above. We have −µ ijk ∂ωk ∂xj = = = = −µ ijk kmn ∂ ∂ un ∂xj ∂xm ∂ 2 un ∂xj ∂xm −µ(δim δjn − δin δjm ) µ µ ∂ 2 ui ∂xj ∂xj 2 ui Terms in the two-dimensional vorticity equation In two dimensions the equation for the only non-zero vorticity component can be written ∂ω3 ∂ω3 ∂ω3 + u1 + u2 =ν ∂t ∂x1 ∂x2 2 ω3 We will take two examples elucidating the meaning of the terms in the vorticity equation. Vorticity and viscocity There exists a subtle relationship between ﬂows with vorticity and ﬂows on which viscous forces play a role. we ﬁnd that the vorticity has the following form ω3 = ωz = ω(y.

i

**ree vectorﬁelds // to boundary L h x1 x2 p0 > p 1 p1 u n dS u · n∆t x y u0 p0 y η = √4νt √ y = η 4νt t η y ω (η)·u0 ˜ √ ωz = πνt t
**

u u0

η

y = η 4νt t ηy y ˜ (η)·u ωz = ω√πνt0 t

y = √4νt u √ u0

u0 p0

8

7

6

5

4

3

t

2

1

0

0

0.1

0.2

0.3

0.4

ω

0.5

0.6

0.7

0.8

0.9

1

Figure 2.7: A sketch of the evolution of vorticity in Stokes 1st problem. v, y

ω=0 x, u Figure 2.8: Stagnation point ﬂow with a viscous layer close to the wall. In this ﬂow the streamlines are straight lines and therefore the non-linear term, or the transport/advection term, is zero. We now turn to an example where the transport of vorticity is important. Example 3. Transport/advection of vorticity: Stagnation point ﬂow. The stagnation point ﬂow is depicted in ﬁgure 2.8, and is a ﬂow where a uniform velocity approaches a plate where it is showed down and turned to a ﬂow parallel two the plates in both the positive and negative x-direction. At the origin we have a stagnation point. The inviscid stagnation point ﬂow is given as follows inviscid: u = cx v = −cy ⇒ ωz = ω = ∂v ∂u − =0 ∂x ∂y

Note that this ﬂow has zero vorticity and that it does not satisfy the no-slip condition (u = v = 0, y = 0). In order to satisfy this condition we introduce viscosity, or equally, vorticity. Thus close to the plate we have to fulﬁll the equation ∂ω ∂2ω ∂2ω ∂ω +v =ν + u ∂x ∂y ∂x2 ∂y 2 The two ﬁrst terms represent advection or transport of vorticity. We will try a simple modiﬁcation of the inviscid ﬂow close to the boundary, which is written as u = xf (y) v = −f (y) This satisﬁes the continuity equation, and will be zero on the boundary and approach the inviscid ﬂow far above the plate if we apply the following boundary conditions u = v = 0, u ∝ cx, v ∝ −cy, y=0 y→∞

The vorticity can be written ω = −xf (y). We introduce this and the above deﬁnitions of the velocity components into the vorticity equation, and ﬁnd −f f + f f + νf =0

0

η=

y = η 4νt t η y F,˜√ , F (η)·u ωz = ωFπνt0 t η

√y 4νt u √ u0

p0

1.5

F

1

F

0.5

F

0 0 1

η

2

Figure 2.9: Solution to the non-dimensional equation for stagnation point ﬂow. F is proportional to the normal velocity, F to the streamwise velocity and F to the vorticity. The boundary conditions become f =f =0 f ∝ cy, f ∝ c, An integral to the equation exists, which can be written f

2

y=0 y→∞

− f f − νf

= K = {evaluate y → ∞} = c2

This is as far as we can come analytically and to make further numerical treatment simpler we make the equation non-dimensional with ν and c. Note that they have the dimensions [ν] = and that we can use the length scale dependent variables as L2 , T [c] = 1 T √ νc to scale the independent and the

ν/c and the velocity scale y ν/c f (y) √ νc

η F (η)

= =

**This results in the following non-dimensional equation F with the boundary conditions F = F = 0, F ∝ η, η=0 η→∞
**

2

− FF − F

=1

In ﬁgure 2.9 a solution of the equation is shown. It can be seen that diﬀusion of vorticity from the wall is balanced by transport downward by v and outward by u. The thin layer of vortical ﬂow close to the wall has the thickness ∼ ν/c.

**Stretching and tilting of vortex lines
**

In three-dimensions the vorticity is not restricted to a single non-zero component and the three-dimensional vorticity equation governing the vorticity vector becomes ∂ωi 1 ∂ui ∂ωi + uj + = ωj ∂t ∂xj ∂xj Re

advection of vorticity 2

ωi

diﬀusion of vorticity

η=

y = η 4νt t η y ˜ (η)·u ωz = ω√πνt0 t

4νt u √ u0

√

dr

C(t)

Figure 2.10: A closed material curve. The complex motion of the vorticity vector for a full three-dimensional ﬂow can be understood by an analogy to the equation governing the evolution of a material line, discussed in the ﬁrst chapter. The equation for a material line is ∂ui D dri = drj Dt ∂xj Note that if we disregard the diﬀusion term, i.e. the last term of the vorticity equation, the two equations are identical. Thus we can draw the following conclusions about the development of the vortex vector i) Stretching of vortex lines (line ω) produce ωi like stretching of dri produces length

ii) Tilting vortex lines produce ωi in one direction at expense of ωi in other direction

**Helmholz and Kelvin’s theorems
**

The analogy with the equation for the material line directly give us a proof of Helmholz theorem. We have Theorem 1. Helmholz theorem for inviscid ﬂow: Vortex lines are material lines in inviscid ﬂow. A second theorem valid in inviscid ﬂow is Kelvin’s theorem. Theorem 2. Kelvin’s theorem for inviscid ﬂow: Circulation around a material curve is constant in inviscid ﬂow. Proof. We show this by considering the circulation for a closed material curve, see ﬁgure 2.10, Γm =

C(t)

ui dri

We can evaluate the material time derivative of this expression with the use of Lagrangian coordinates as follows DΓm Dt = D ui dri Dt C(t) Lagrange coordinates ˆ ˆ ui = ui (x0 , t) = ui (x0 (m), t) i i 0 ˆ 0 ˆ) ri = ri (xi , t) = ri (xi (m), t ∂ri d u dm ˆ m i ∂m dt ∂ ∂ri ∂ui ∂ri ui dm + ˆ ∂m ˆ ∂ t ∂m m ∂t m

= = =

dm

**The last term of this expression will vanish since ui
**

m

∂ ∂ri ˆ ∂ t ∂m

dm =

m

ui

∂ui dm = ∂m

m

∂ ∂m

1 ui ui 2

dm = 0

we have dx u dx dy = ⇐⇒ = ⇐⇒ u dy − v dx = 0 dy v u v and in three dimensions we can write dy dz dx = = u v w For two-dimensional ﬂow we can ﬁnd the streamlines using a streamfunction.12 and ﬁgure 2. v) t dx dy x. Thus we ﬁnd the following resulting expression DΓm Dt = C(t) Dui dri Dt − C(t) = C(t) 1 ∂p + ∂xi Re 2 2 ui dri Re → ∞ = 1 Re ui dri → 0. see ﬁgure 2. Which shows that the circulation for a material curve changes only if the viscous force = 0 Streamfunction We end this section with the deﬁnition and some properties of the streamfunction. We can introduce a streamfunction ψ = ψ(x.11. In two dimensions. on streamlines We can also quickly check that the deﬁnition given above satisﬁes continuity. We can make the following derivation dψ = = = = ∂ψ ∂ψ dx + dy ∂x ∂y ∂ψ ∂ψ assume u = . u Figure 2. v=− and check for consistency ∂y ∂x u dy − v dx 0.0 √ uy.13 in the following derivations of the volume ﬂux: .11: A streamline which is tangent to the velocity vector. y) with the property that the streamfunction is constant along streamlines. v y = η 4νt t η y ˜ (η)·u ωz = ω√πνt0 u = (u. We have ∂2ψ ∂2ψ ∂u ∂v + = − =0 ∂x ∂y ∂x∂y ∂y∂x Another property of the streamfunction is that the volume ﬂux between two streamlines can be found from the diﬀerence by their respective streamfunctions. We use ﬁgure 2. To deﬁne the streamfunction we need the streamlines which are tangent to the velocity vector.

12: Integration paths between two streamlines. B QAB = A B ui ni ds (nx u + ny v) ds A = = = A 1 ds = . and ﬁnally we deal with two-dimensional potential ﬂows.ωz = t n dS u · n∆t x y u0 p0 y η = √4νt u √ πνtu B Ψ = ΨB S n A Ψ = ΨA Figure 2. dy nx B − ds 1 = dx ny B A (u dy − v dx) = dψ = ψB − ψA 2. √ u0 y = η 4νt t η y ˜ (η)·u ωz = ω√πνt0 t dy ds nx ny |n| = 1 − dx Figure 2. Working with the velocity potential. but also restrict the validity of the solutions. a so called potential.3 Potential ﬂow Flows which can be found from the gradient of a scalar function. for example. will greatly simplify calculations. In three dimensions we have ui = ∂φ ∂xi or u= φ with the corresponding relations in two dimensions u = v ∂φ ∂x ∂φ ∂y = There are several consequences of this deﬁnition of the velocity ﬁeld. will be dealt with in this section. then discuss simpliﬁcations of the momentum equations when such potentials exist. We start with a deﬁnition of the velocity potential.13: The normal vector to a line element. . Velocity potential Assume that the velocity can be found from a potential. where analytic function theory can be used.

e. for two diﬀerent ﬂow situations.e.14. ∂uk ∂2φ ωi = ijk = ijk =0 ∂xj ∂xj ∂xk ii) Potential ﬂow is not inﬂuenced by viscous forces. Bernoulli’s equation can be used to calculate the pressure when the velocity is known. Integrate the momentum equations along a streamline. see ﬁgure 2. The velocity potential u = polar coordinates can be written ur = ∂φ ∂r u θ = 1 ∂φ r ∂θ φ in . derived earlier. We have ti ∂ ∂xi p 1 uk uk + + gx3 2 ρ ds = ijk ui uj ωk ds = 0 u ⇒ p 1 2 (u + u2 + u2 ) + + gx3 = constant along streamlines 3 2 2 1 ρ Thus. We have ∂τij =µ ∂xj 2 ui = −µ ijk ∂ωk ∂xj iii) Potential ﬂow satisﬁes Laplace equation. uj and obtain the following equation ∂ui ∂ + ∂t ∂xi 1 p uk uk + + gx3 2 ρ = ijk uj ωk 2 ui − gδi3 ∂ui ∂ = ∂xj ∂xi 1 uk uk 2 − ijk uj ωk +ν ijk ∂ωk ∂xj This equation can be integrated in for two diﬀerent cases: potential ﬂow (without vorticity) and stationary inviscid ﬂow with vorticity. which is seen by the relationship between viscous forces and vorticity. which can be seen by taking the curl of the gradient of the velocity potential. ∂φ i) Potential ﬂow. which is a consequence of the divergence constraint applied to the gradient of the velocity potential. We start by rewriting the momentum equation ∂ui ∂ui 1 ∂p =− +ν + uj ∂t ∂xj ρ ∂xi using the alternative form of the non-linear terms. we have ∂φ 1 ∂ p + uk uk + + gx3 = 0 ⇒ ∂xi ∂t 2 ρ p ∂φ 1 2 + (u1 + u2 + u2 ) + + gx3 = f (t) 3 2 ∂t 2 ρ ii) Stationary. Example 4. i. Ideal vortex: Swirling stationary ﬂow without vorticity. i. Introduce the deﬁnition of the velocity potential ui = ∂xi into the momentum equation above. inviscid ﬂow with vorticity. i.i) Potential ﬂow have no vorticity. ∂ui ∂2φ = = ∂xi ∂xi ∂xi 2 φ=0 Bernoulli’s equation For potential ﬂow the velocity ﬁeld can thus be found by solving Laplace equation for the velocity potential.e. The momentum equations can then be used to ﬁnd the pressure from the so called Bernoulli’s equation.

potential to ﬁnd the streamfunction. we have Γ ∂ψ =− ∂r 2πr We can easily ﬁnd the streamfunction for this ﬂow. i. which in polar coordinates becomes 2 φ= 1 ∂ r ∂r 1 ∂φ r ∂r + 1 ∂2φ =0 r ∂θ2 The ﬁrst term vanishes since ∂φ ∂r = ur = 0.e. Integrate the resulting equation twice and we ﬁnd φ = Cθ + D The constant D is arbitrary so we can set it to zero. In polar coordinates the streamfunction is deﬁned ur = 1 ∂ψ r ∂θ u = − ∂ψ θ ∂r ⇒ ψ=− Γ ln r 2π · u = 0. Using the deﬁnition of the circulation we ﬁnd uθ = C . The velocity potential satisﬁes Laplace equation. as we will show in the following. We use the solution from the velocity The streamfunction is constant along streamlines resulting in circular streamlines. see ﬁgure 2. x y u0 r p0 y η = √4νt u √ u0 y = η 4νt t η y ˜ (η)·u ωz = ω√πνt0 t Figure 2. Two-dimensional potential ﬂow using analytic functions For two-dimensional ﬂows we can use analytic function theory.15 We can now use Bernoulli’s equation to calculate the pressure distribution for the ideal vortex as 1 2 u + p = p∞ 2 θ ⇒ p = p∞ − Γ2 8π 2 r2 Note that the solution has a singularity at the origin.15: Streamlines for the ideal vortex.14: Integration of the momentum equation along a streamline.e. Any analytic function represents a twodimensional potential ﬂow. . i. In fact. complex valued functions. to calculate the velocity potential. r C= Γ 2π Note that the streamfunction satisﬁes continuity.h u t x1 xη 2 pω(η)·uy 0 > p0 1 ωz = ˜√πνt 1 p u t n t= u u dS u · n∆t Figure 2.

F (z) exists and is unique ⇐⇒ F (z) is an analytic function The fact that a derivative of a complex function should be unique rests on the validity of the Cauchy– Riemann equations.16. y + ∆y) − φ(x. y) + iψ(x. Velocity potential φ for two-dimensional ﬂow is deﬁned u = ∂φ ∂x Using the continuity equation gives ∂u ∂v + = ∂x ∂y u v = = v = ∂φ ∂y 2 φ=0 The streamfunction ψ is deﬁned ∂ψ ∂y − ∂ψ ∂x 2 By noting that potential ﬂow has no vorticity we ﬁnd −ω = − ∂v ∂y + = ∂x ∂y ψ=0 Thus. ∂φ ∂ψ ∂φ ∂ψ = and =− ∂x ∂y ∂y ∂x Using the Cauchy-Riemann equations we can show that the real φ and imaginary part ψ of the complex potential satisfy Laplace equations and thus that they are candidates for identiﬁcation with the velocity potential and streamfunction of a two-dimensional potential ﬂow. y) + iψ(x + ∆x. i. We can utilize analytic function theory by introducing the complex function F (z) = φ(x. y) where z = x + iy = reiθ = r(cos θ + i sin θ) which will be denoted the complex potential. y) + iψ(x. y + ∆y) − iψ(x. both the velocity potential φ and the streamfunction ψ satisfy Laplace equation. y)] i∆y→0 i∆y ∂φ = ∂ψ real part ∂x ∂y ⇒ ∂ψ 1 ∂φ ∂ψ ∂φ i = ⇒ =− imaginary part ∂x i ∂y ∂x ∂y Thus we have the Cauchy–Riemann equations as a necessary requirement for a complex function to be analytic. y)] = lim ∆x→0 ∆x 1 = lim [φ(x. We can derive those equations by requiring that a derivative of a complex function should be the same independent of the direction we take the limit in the complex plane. We have 2 ∂ ∂ψ ∂ ∂ψ ∂2φ ∂ φ = = =− 2 ∂x2 ∂x ∂y ∂y ∂x ∂y 2 ∂ ψ ∂ = 2 ∂y ∂y ∂φ ∂x = ∂ ∂x ∂φ ∂y =− ∂2ψ ∂x2 . We have 1 F (z) = lim [F (z + ∆z) − F (z)] ∆z→0 ∆z 1 [φ(x + ∆x. y) − φ(x. y) − iψ(x.We start by showing that both the velocity potential and the streamfunciton satisfy Laplace equation.e. We now recall some useful results from analytic function theory. see ﬁgure 2.

as u = U cos α v = U sin α Example 6.17. Calculate the velocity ﬁeld from the complex potential F = U e−iα z. θ) − iuθ (r.16: The complex plane and the deﬁnitions of analytic functions. Thus any analytic complex function can be identiﬁed with a potential ﬂow. we also have to show that the Cauchy-Riemann equations are consistent with the deﬁnition of the velocity potential and the streamfunciton. see ﬁgure 2. see ﬁgure 2. The Cauchy–Riemann equations are ∂ψ ∂φ u= = ∂x ∂y which recovers the deﬁnitions of the velocity potential and the streamfunction. In polar coordinates the complex potential becomes Arn einθ = Ar n cos(nθ) + iAr n sin(nθ) Thus the velocity potential and the streamfunction can be identiﬁed as φ = Ar n cos(nθ) ψ = Ar n sin(nθ) . The complex velocity becomes W (z) = F = U e−iα = U cos α − i sin α which gives us the solution for the velocity ﬁeld in physical variables. Calculate the velocity ﬁeld from the complex potential F = Az n . We can now deﬁne the complex velocity as the complex conjugate of the velocity vector since W (z) = ∂φ ∂ψ dF = +i = u − iv dz ∂x ∂x v = ∂φ = − ∂ψ ∂y ∂x It will be useful to have a similar relation for polar coordinates.18. where the real part is the velocity potential and the imaginary part is the streamfunction.η= y = η 4νt t η y ω (η)·u0 ˜ √ ωz = πνt t x √y 4νt u √ u0 u0 p0 iy i∆y ∆x Figure 2. θ)] e−iθ cos θ−i sin θ We will now take several examples of how potential velocity ﬁelds can be found from analytic functions. Example 5. In this case we have u = ur cos θ − uθ sin θ v = ur sin θ + uθ cos θ W (z) = u − iv = [ur (r. In addition to satisfying the Laplace equations.

17: The deﬁnition of polar coordinates.18: Constant velocity at an angle α. U0 Gradient ﬁelds wi ∂p ∂xi ui Divergence free vectorﬁelds // to boundary L h x1 x2 p0 > p 1 p1 u n dS u · n∆t x y u0 p0 y η = √4νt y = η 4νt t η y ˜ (η)·u ωz = ω√πνt0 t Figure 2. √ u u0 α . U 0 θ Tw x L T0 Figure 2.u · n∆t P2 x xy 1 x2 u0 x3 p0 √y 3 R η = 4νt u R2 u1 √ R0 y y = η 4νt r3 t 2 v rη 1 ry π dϕ12 2 + (η)·u ˜ ωz = ω√πνt0 u ˜ V (t) S (t) t uθ V (t + ∆t) − V (t) S (t + ∆t) ur u u n V (t + ∆t) r p T0 .

r1 π + dϕ12 2 and we can calculate the complex velocity as V (t) S (t) (z) = F = nAz n−1 W V (t + ∆t) − V (t) = nAr n−1 einθ e−iθ S (t + ∆t) = nAr n−1 [cos(nθ) + i sin(nθ)]e−iθ u n This gives us the solution for the velocity in polar coordinates as V (t + ∆t) ur = nAr n−1 cos(nθ) p uθ = −nAr n−1 sin(nθ) T0 .21: Stagnation point ﬂow when n = 2.xi y =u · n∆t η 4νt Ψ=0 ˆ ui dt x t 0 y dxiη Ψˆ= 0 dui uy d0 t ˜ (η)·u 0 ˆ p rω√ dt 0 ωz = i πνt y η = √4νt ˆ dri dt ut π u1 √ P0 n y = η 4νt P2 Ψ=0 t x1 η x2 Figure 2.19.22) is u · n∆t ur = nAr n−1 cos(nθ) ⇒ {θ = 0} ⇒ u = ur = nAr n−1 = Cxn−1 x y The wedge angle is n−1 2π u0 ϕ = 2π − = 2π n n p0 y η = √4νt u ∂p √ u0 y = η 4νt t η y ω (η)·u0 ˜ √ ωz = πνt Figure 2. The velocity along the wedge edge (x axis dS in ﬁgure 2.19: Flow towards a corner. ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! t . The streamfunction is zero on the walls and on L the stagnation point streamline. ui Figure 2. Here the analysis can be done in Cartesian Divergence coordinates free vectorﬁelds // to boundary L F = Az 2 h W (z) = F = 2Az x1 x2 = 2A(x + iy) p0 > p 1 u = 2Ax p1 ⇒ v = −2Ay u n Figure 2.20 shows the uniform ﬂow over a ﬂat plate which results when n = 1. We have U0 Gradient ﬁelds ψ = Ar n sin(nθ) = 0 ⇒ θ = πk wi n ∂xi Figure 2.22 shows the ﬂow towards a wedge when 1 ≤ n ≤ 2. U 0 Tw This is a ﬂow which approaches a corner. see ﬁgure 2. xy 3 ω (η)·u0 ˜ √ ωz = πνt 3 R R2 t Ψ=0 R1 r3 r2Figure 2. By solving for the zero streamlines we can calculate the angle between a T0 wall aligned with the x-axis and the stagnation point streamline.21 shows the stagnation point ﬂow when n = 2.20: Uniform ﬂow over a ﬂat plate.

Example 7. Line vortex.L y = η 4νt h t x1 η x2 py > p1 0 ϕ ˜ (η)·u p1 ωz = ω√πνt0 u t n dS u · 2. x y u0 p0 y η = √4νt √ u u u0 √ u0 y = η 4νt t η y ˜ (η)·u ωz = ω√πνt0 t Figure 2. F (z) = = = m ln z 2π m ln(reiθ ) 2π m [ln r + iθ] 2π The complex velocity can be written as The velocity potential and the streamfunction become φ = m ln r 2π ψ = mθ 2π W (z) = F = m −iθ m = e 2πz 2πr which gives the solution in polar coordinates.22: Figuren∆t The ﬂow over a wedge when 1 ≤ n ≤ 2. as ur = m 2πr uθ = 0 We can now calculate the volume ﬂux from the source as 2π Q= 0 m r dθ = m 2πr Example 8. We now list several other examples of potential ﬂow ﬁelds obtained from analytic functions. Line source.23.23: Line source. see ﬁgure 2. F (z) = = −i Γ ln z 2π Γ [θ − i ln r] 2π .

where the viscous eﬀects are small. However.24: Ideal vortex. The velocity potential and the streamfunction become φ= Γ θ 2π ψ = − Γ ln r 2π W (z) = F = −i Γ −iθ e 2πr √y 4νt u √ u0 x y u0 p0 The complex velocity can be written as which gives the solution in polar coordinates.24. there will be a layer close to the solid walls where the viscous terms are important even for very high Reynolds number ﬂow.η= y = η 4νt t η y ˜ (η)·u ωz = ω√πνt0 t Figure 2. The potential ﬂow around a cylinder with circulation has the complex potential F (z) = U z + U 2 Γ z r0 −i ln z 2π r0 Example 12. Airfoils. The Euler equations are obtained if Re → ∞ in the Navier-Stokes equations. In particular. see ﬁgure 2.4 Boundary layers For ﬂows with high Reynolds number Re. In order to satisfy the no slip condition we need to add the viscous term. Thus. 2. . thus a smaller number of boundary conditions can be satisﬁed when solving the Euler equations. the Euler equations. the no ﬂow condition can be satisﬁed at a solid wall but not the no slip condition. as ur = 0 uθ = Γ 2πr F (z) = m πz Example 9. most of the ﬂow can be considered inviscid and a simpler set of equations. can be solved. The potential ﬂow around a cylinder has the complex potential F (z) = U z + U 2 r0 z Example 11. This layer will be very thin and the ﬂow in that region is called boundary layer ﬂow. The dipole has the complex potential Example 10. The inviscid ﬂow around some airfoils can be calculated with conformal mappings of solutions from cylinder with circulation. the highest derivatives are present in the viscous terms.

we expect large velocity gradients near walls. where the no slip condition is fulﬁlled.26. δ ω≈0 U→ } U δ L x. By analogy to the Stokes solution for the instantaneously starting plate. as Inviscid : U velocity scale for u. We begin this section by estimating the thickness of the viscous region in channel with inviscid inﬂow (zero vorticity).26: Boundary layer proﬁle close to a solid wall. y BL : : : : U α L δ velocity scale for velocity scale for length scale for length scale for u v x y The pressure scale is assumed to be ρU 2 for both the inner and the outer region.25. i. Since we expect this region to be thin for high Reynolds numbers. ω (η)·u0 ˜ η √ πνt = 0 √y 4νt u √ u0 L v y.25: Estimatet of the viscous region for an inviscid vorticity free inﬂow in a plane channel. We choose diﬀerent length scales inside boundary layer and in inviscid region as indicated in ﬁgure 2. u Figure 2.e. The unknown velocity scale for the normal velocity α is determined using continuity ∂v ∂u + =0 ∂x ∂y O U L O α δ which implies that α U = L δ ⇒ α= δ U L In this manner both terms in the continuity equation have the same size. we can estimate the distance √ the vorticity has diﬀused from the wall toward the channel centerline as δ ∼ νts . v : L length scale for x. We obtain the following estimates for the size of the terms in the normal momentum equation . Where ts is time of a L ﬂuid particle with velocity U has travelled length L. ts = . see ﬁgure 2.ωz = }δ t ω = η 4νt ≈0 y U→ t η L y ω (η)·u0 ˜ ωz = √πνt Figure 2. Let us use these scalings in the steady momentum equations and disregard small terms. Thus we can estimate δ/L as U √ 1/2 δ 1 νL ν 1/2 1 νts = ⇒ = = =√ → 0 as Re → ∞ L L L U UL Re We will now divide the ﬂow into a central/outer inviscid part and a boundary layer part close to walls.

We have u ∂u ∂u 1 ∂p ∂2u +v =− +ν 2 ∂x ∂y ρ ∂x ∂y In summary. y → ∞) = = = 0 0 ue (x) outer inviscid ﬂow where the initial condition is at x = x0 and the equations are marched in the downstream direction. We have the no ﬂow and the no slip conditions at the wall and the u = u e in the free stream. the pressure is constant through the boundary layer. the only choice that gives a consistent approximation. Next we estimate the size of the terms in the streamwise momentum equation as →x: u ∂u ∂x + v ∂u ∂y = − 1 ∂p ρ ∂x U2 L 1 + ν ∂2u ∂x2 + ν ∂2u ∂y 2 U2 L 1 δU U · L δ 1 ν ·U L2 1 Re ν U δ2 1 · Re L δ 2 The ﬁrst three terms are order one and the fourth term negligible when the Reynolds number become large. By using continuity one can show that once the u0 is given the momentum equation can be written in the form − ∂v + f (y) v = g (y) ∂y which can be integrated in the y-direction to give the initial normal velocity. y = 0) u (x. There are several things to note about these equations. First. Second. In order to keep the last term. only pressure term O(1). y = 0) v (x. . vin cannot be given as its own initial condition. given by the inviscid outer solution. Thus no boundary condition for v in the free stream can be given. the system of equations is a third order system in y and thus only three boundary conditions can be enforced. Thus we are left with the parabolic equation in x since the second derivative term in that direction is neglected. the steady boundary layer equations for two-dimensional ﬂow is 2 u ∂u + v ∂u = − 1 ∂p + ν ∂ u ∂x ∂y ρ ∂x ∂y 2 ∂u ∂v + =0 ∂x ∂y with the initial (IC) and boundary conditions (BC) IC : u (x = x0 . y) = uin (y) BC : u (x.↑y: ∂v ∂x 1 δU U· L L u δ L 2 + ∂v ∂y δU 1 δU · · L δ L v δ L 2 = − 1 ∂p ρ ∂y U2 δ + ν ν L2 ∂2v ∂x2 · δU L + ∂2v ∂y 2 ν δU δ2 L ν 1 Re 1 1 Re δ L 2 where the last line was obtained by dividing all the terms on the second line by U 2 /δ. we have to assume that δ 1 ∼√ L Re which is the same estimate as we found for the extent of the boundary layer in the channel inﬂow case. which implies that ∂p =0 ∂y ⇒ p = p(x) Thus. When R → ∞ and δ/L is small.

27: Boundary layer ﬂow over a ﬂat plate with zero pressure gradient. We give two below v=− ∂η ∂Ψ = −u0 δ f + δf ∂x ∂x y = −u0 δ f − f δ δ = u0 (ηf − f ) δ $$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$ "#"#"#"#"#"#"#"#"#"#"#"#"#"#"#"#"#"#"#"#"#"#"#"#"#"#"#"#"#"#"#"#"# "$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$#"$# #################################"$"$ U→ y. y) u (x.Figure 2. The boundary layer equations become ∂Ψ ∂ 2 Ψ ∂3Ψ ∂Ψ ∂ 2 Ψ · − · =ν 3 ∂y ∂x∂y ∂x ∂y 2 ∂y with appropriate initial and boundary conditions. y = 0). The boundary layer equations become 2 u ∂u + v ∂u = ν ∂ u ∂x ∂y ∂y 2 ∂u ∂v + =0 ∂x ∂y IC : BC : u (x = x0 . y = 0) u (x. ∂y v=− ∂Ψ ∂x which implicitly satisﬁes continuity. y) = u0 δ (x) f (η) We can now evaluate the various terms in the momentum equation. Similarity solution We will now look for a similarity solution. i. the outer inviscid ﬂow is ue (x) = u0 . We have 1 p (x) + ρu2 (x) = const 2 e which implies that − due 1 dp = ue ρ dx dx Blasius ﬂow over ﬂat plate Consider the boundary layer equations for a ﬂow over a ﬂat plate with zero pressure gradient.e. since the thickness of the boundary layer is zero from the inviscid point of view. ue (x) = uinv (x. Fourth. i. y = 0) v (x. Let u= ∂Ψ = u0 f (η) .e. the pressure gradient term in the momentum equation can be written in terms of the edge velocity ue using the Bernoulli equation. y → ∞) = = = = u0 0 0 u0 We introduce the two-dimensional stream function Ψ u= ∂Ψ . u L U δ u0 u0 . ∂y η= y δ (x) which by integration in the y-direction implies Ψ (x. v x. Third. the boundary layer solution in the free stream approaches the inviscid solution evaluated at the wall.

v x. the coeﬃcient in front of the f f term has to be constant. We have f (η99 ) = 0.4 f = u u0 0.9 ⇒ √ x.28: Similarity solution giving the Blasius boundary layer proﬁle.2 0.6 0.L δ ω≈0 U→ y. With the help of ﬁgure 2. We are left with the Blasius equation and boundary conditions f BC : 1 + ff = 0 2 f (0) = f (0) = 0 f (∞) = 1 The numerical solution to the Blasius equation gives the self-similar velocity proﬁle seen in ﬁgure 2. This implies f + u0 δδ 1 = ν 2 which implies that δ (x) = ⇒ δδ dx = 1 2 ν dx u0 ⇒ 1 2 1 νx δ = 2 2 u0 f δ f η δ + u2 (ηf − f ) = νu0 2 0 δ δ δ νx . The boundary layer thickness can be evaluated as the height where the velocity has reached 99% of the free stream value.28. u η=√ yL νx/u0 U δ u0 } 5 Blasius proﬁle 4 3 2 1 0 0 0.8 1 Figure 2.99 ⇒ η99 = δ99 νx u0 = 4. η ∂2Ψ = −u0 f δ ∂x∂y δ If these and the other corresponding terms are introduced into the momentum equation we have the following equation for f −u2 f f 0 which can be simpliﬁed to u0 δδ ff = 0 ν For a similarity solution to be possible.9 νx u0 where it can be seen that the boundary layer thickness grows as Displacement thickness and skin friction coeﬃcient Another measure of the boundary layer thickness is the displacement δ ∗ of the wall needed in order to have the same volume ﬂux of the inviscid ﬂow as for the ﬂow in the boundary layer. δ99 = 4. where we have chosen the constant of integration such that δ = 0 u0 when x = 0.29 we can evaluate this as y0 0 y y u dy ≡ u0 dy = δ∗ 0 u0 dy − δ ∗ u0 which implies that .

664 =√ Rex where the last terms again is the Blasius value. .72 A quantity which is of large importance in practical applications is the skin friction coeﬃcient. since they become singular at that point. become 02102102102102102102102102102102110210210210210210210210210210210210211021021021021021111111111111111111111102 02 02 02 02 02 02 02 02 02 02 02 02 02 02 02 02 02 02 02 02 02 02 02 02 u f = u0 Blasius proﬁle δ∗ νx/u0 Separation point δ∗ = 0 ∞ 1− u u0 dy = y=0 u ∂u ∂u 1 dp ∂2u +v =− +ν 2 ∂x ∂y ρ dx ∂y +u ∂3u ∂u ∂ 2 u ∂3u ∂4u − +v 3 =ν 4 2 2 ∂x∂y ∂x ∂y ∂y ∂y ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) (&(&(&(&(&(&&(&(&(&(&(&(&&(&(&(&(&(& &&&&&&()&&&&&&&()&&&&&&() ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' %&%&%&%&&%&%&%&&%&%&%&&%&%&&%&%&%& &&&&%'&&&&%'&&&&%'&&&%'&&&&%' t L η=√ y νx/u0 }δ u f = u0 ω≈0 Blasius proﬁle U→ y. If we take the momentum equation take the normal derivative and use continuity we have u ∂2u ∂3u ∂2u +v 2 =ν 3 ∂x∂y ∂y ∂y If we again take the normal derivative and use continuity we can write this as 1 ∂ 2 ∂x ∂u ∂y 2 This expression evaluated at the wall. The skin friction at the wall is τw = µ ∂u ∂y = µu0 ρu2 0.332ρu2 0 0 f (0) = f (0) = √ u0 x δ Rex ν where the last term is the expression evaluated for Blasius ﬂow. If we evaluate this expression for the Blasius boundary layer we have δ ∗ = 1.30: Boundary layer separation caused by a pressure force directed upstream. The boundary layer equations cannot be integrated past a point of separation. see ﬁgure 2. with the use of the boundary conditions. u L U δ u0 η=√ y U δ u0 δ∗ x νx u0 ∞ 0 [1 − f (η)] dη νx u0 . This can be used to evaluate the skin friction coeﬃcient as Cf = τw 1 2 2 ρu0 0. This can be seen from the following manipulations of the boundary layer equations. a so called adverse pressure gradient.ωz = √ πνt Figure 2. v x. Boundary layer separation ∂p A large adverse pressure gradient ∂x > 0 may cause a back ﬂow region close to wall.29: The displacement thickness δ ∗ ⇐= Pressure force Figure 2.30.

invalidating the downstream parabolic nature of the equations assumed by the downstream integration of the equations. as above. For this purpose we deﬁne an ensemble average as 1 Ui = ui = lim N →∞ N (n) N ui n=1 (n) where each member of the ensemble ui is regarded as an independent realization of the ﬂow. 2. In addition. However. the point of separation is the limiting point after which there is back ﬂow close to the wall. let U i = ui . In practical situations it is very important to be able to predict the point of separation. We ﬁnd the Reynolds average equations . This means that there is information propagating upstream.5 Turbulent ﬂow Reynolds average equations Turbulent ﬂow is inherently time dependent and chaotic. in applications one is not usually interested in knowing full the details of this ﬂow. where ∂u ∂y =0 y=0 this expression is y=0 √ = κ xs − x We can see that the boundary layer equations are not valid beyond point of separation since we have a square root singularity at that point. for example when an airfoil stalls and experience a severe loss of lift.1 ∂ 2 ∂x ∂u ∂y 2 =ν y=0 ∂4u ∂y 4 y=0 ±κ2 which can be integrated to yield ∂u ∂y 2 y=0 = ±κ2 x + C ∂u ∂y At the point of separation x = xs . Divide the total ﬂow into an average and a ﬂuctuating component ui as ui = ui + u i p=p+p and introduce into the Navier-Stokes equations. but rather satisﬁed with the inﬂuence of the turbulence on the averaged ﬂow. i. We are now going to derive an equation governing the mean ﬂow Ui . Now. and drop the . We ﬁnd ∂ui ∂ui ∂u 1 ∂p 1 ∂p ∂u ∂ui ∂ui + + uj + uj i + uj + uj i = − − +ν ∂t ∂t ∂xj ∂xj ∂xj ∂xj ρ ∂xi ρ ∂xi We take the average of this equation. useing ui = 0 which gives ∂ui 1 ∂p ∂ui + uj =− +ν ∂t ∂xj ρ ∂xi ∂ui ∂xi = 0 ∂ ui uj ∂xj u j ui = ui · u j = 0 2 ui + ν 2 ui 2 ui − where the average of the continuity equation also has been added.e.

an approximation only true for very simple turbulent ﬂow. i. We have Rij = 2 kδij − 2νT E ij 3 The ﬁrst term on the right hand side is introduced to give the correct value of the trace of R ij . the eﬀect of the Reynolds stress components on in the Reynolds average equations have to be modelled. the main Reynolds stress component becomes uv = −νT where. uT is a characteristic turbulent velocity scale and l is a characteristic turbulent length scale. and the equations for those would in turn involve even more unknowns. The simplest consistent model is to introduce a turbulent viscosity as a proportionality constant between the Reynolds stress components and the strain or deformation rate tensor. νT νT = u T l Here. E ij = 1 2 ∂Ui ∂Uj 2 ∂Ur δij + − ∂xj ∂xi 3 r ∂x where we have assumed incompressible ﬂow. This closure problem implies that if one cannot calculate the complete turbulent ﬂow. Thus the eﬀect of the turbulence on the mean is through an additional stress Rij which explicitly written out is u2 [Rij ] = [ui uj ] = uv uw uv v2 vw uw vw w2 ∂Ui 1 ∂P ∂ ∂Ui + Uj =− + (τij − ui uj ) ∂t ∂xj ρ ∂xi ∂xj ∂Ui ∂xi The diagonal component of this stress is the kinetic energy of the turbulent ﬂuctuations k = Rii /2 = ui ui /2 = 1 2 u + v 2 + w2 2 Turbulence modelling One may derive equations for the components of the Reynolds stress tensor. It is introduced here only to point out the analogy between the molecular viscosity and the turbulent viscosity.e. in analogy with kinetic gas theory. However. in analogy with the stress-strain relationship in for Newtonian ﬂuids. Here the deformation rate tensor is calculated based on the mean ﬂow. i. The turbulent viscosity νT must be modelled and most numerical codes which solved the Navier-Stokes equations for practical applications uses some kind of model for νT . they would involve additional new unknowns. ∂U ∂y . velocity ﬁelds of the form U (y).where ui uj is the Reynolds stress. but is only interested in the mean.e. For simple shear ﬂows. Introducing this into the Reynolds average equations gives ∂Ui ∂Ui ∂ + Uj = − ∂t ∂xj ∂xj P 2 1 ∂P + k δij + 2 (ν + νT ) E ij = − + (ν + νT ) ρ 3 ρ ∂xi 2 Ui where the last equality assumes that νT is constant. In kinetic gas theory l would be the mean free path of the molecules and uT the characteristic velocity of the molecules.

We can thus evaluate uT and ﬁnd νT uT = l dU dy ⇒ νT = l 2 dU dy Deﬁne τw as the shear stress at the wall and uτ = uT = u τ . We have now introduced a number of new unknowns which have to be modelled. Zero-equation models The simplest model for uT and l is to model them algebraically. wakes and shear layers we can take ν T as constant in normal direction or y-direction and to vary in a self-similar manner in x. In wall bounded shear ﬂow we can take l as Prandtl’s mixing length (1920’s). called the log-region.u f = u0 Blasius proﬁle δ∗ x Separation point ⇐= Pressure force νx/u0 y uT l Figure 2. 5 5 5 5 5 5 5 5 5 3443434434344343443434 43544435444354443544435 l =κ·y = −ui uj − x τw /ρ as the skin friction velocity. In simple free shear ﬂows like jets. With U 1 = ln y + C.31.41. as Prandtl’s estimation of mixing length we get uT uτ dU = = dy l κy ⇒ κ is the von Karman constant. the Reynolds average equation. We can derive such an equation by taking the previously derived equation for Ui + ui and subtracting the equation for U . We ﬁnd uT = k= ∂ui ∂ui ∂Ui ∂ui ∂ + Uj = −uj − uj + ∂t ∂xj ∂xj ∂xj ∂xj p ∂ui − δij + ν + u i uj ρ ∂xj We now take the average of the inner product between the turbulent ﬂuctuations u i and the above equation. uτ κ −ν . the next level of modelling uses the assumption that 1 ui ui 2 and a diﬀerential equation is used to calculate the turbulent kinetic energy.31: Prantdl’s mixing length l is the length for which a ﬂuid particle displaced in the normal direction can be expected to retain its horizontal momentum. approximately equal to 0. and ﬁnd ∂ ∂t 1 ui ui 2 ∂ ∂xj 1 ui ui 2 ∂ ∂Ui + ∂xj ∂xj 1 ν ∂ puj − ui ui uj + ui ui ρ 2 2 ∂xj ∂ui ∂ui ∂xj ∂xj + Uj where is the mean dissipation rate of the turbulent kinetic energy. Thus the mixing length is proportional to the distant from the wall and the region close to the wall in a turbulent ﬂow has a logarithmic velocity proﬁle. One-equation models Instead of modelling the characteristic turbulent velocity algebraically. This is a normal distance over which the particle retains its momentum and is depicted in ﬁgure 2.

ui uj = 2 kδij − νT 3 k 3/2 ∂Ui ∂xj + ∂Uj ∂xi . This is based on the above model for k but instead of modelling a partial diﬀerential equation is derived governing the turbulent dissipation. such that the dimension is correct.Second.-model. We ﬁnd Dk Dt rate of increase = (ν + νT ) 2 k diﬀusion rate − i ui uj ∂Uj ∂x generation rate − dissipation rate Two-equation models The most popular model in use in engineering applications is a two equation model. We have = cD 2 First.e. we model the Reynolds stress as previously indicated. This equation has the form as the equation for k. i. 1 1 νT ∂k − ui ui uj − puj = 2 ρ P rk ∂xj where P rk is the turbulent Prandtl number. Many new model constants need to be introduced in order to close the equation. l since the dimension of is [ ] = L . with a diﬀusion term. a gradient diﬀusion model is used for the transport terms. Here cD is a modelling constant. the so called k. a generation term and a dissipation term. k 2 . T Third.e. For simplicity we write the resulting equation for the turbulent kinetic energy for constant ν T . i. Once k and is calculated the turbulent viscosity is evaluated as a quotient of the two. We have ⇒ νT = Cµ · where Cµ is a modelling constant. the turbulent dissipation is modelled purely based on dimension using k and l.

.

dy) dy − dx dy n dl = ( dy.j + ui.j ) /2 vu 2 Blasius proﬁle δ∗ y x Separation point ⇐= Pressure force x dx y l uT νx/u0 ∆yab = yb − ya x dl = ( dx.2.1 Finite Volume method on arbitrary grids y u0 Equations with 1st order derivatives: the continuity equation p0 y η = √4νt In ﬁnite volume methods one makes approximations of the diﬀerential equations in integral form.j ≈ (vi+1.1: Normal vector to the curve l. L We make use of the following deﬁnitions and approximations U 1 uab δ ui+ 2 . − dx). 59 . Recall that the integral from of ﬁrst y=η this equation can be written 4νt t η ∂ (uk ) dV = uk nk dS = 0 y ∂xk ω (η)·u0 ˜ V S √ ωz = πνt Chapter 3 Finite volume methods for incompressible ﬂow Evaluating the normal vector ni to the curve l for two-dimensional ﬂow. − dx) Figure 3.j ) /2 = ∆xab = xb − xa u0 y η=√ u f = ab 0= vi+ 1 . see ﬁgure 3.L T0 U0 Gradient ﬁelds wi ∂p ∂xi ui Divergence free vectorﬁelds // to boundary L h x1 x2 p0 > p 1 p1 u n dS u · n∆t x 3. We start u with an equation with only√ u0 order derivatives.j + vi. v We now apply this to an x.1.j ≈ (ui+1. u arbitrary two-dimensional grid deﬁned as in ﬁgure 3. we have the t normal with length dl as n dl = ( dy. the continuity equation. This implies that L }δ ω≈0 uk nk dl = (u dy − v dx) = 0 U→ S S y.

∆xbc = −∆xda which implies that the approximation of the continuity equation can be written (u dy − v dx) S ≈ 1 ui+ 1 .j ∆ycd − vi− 2 . Note that the grid lines need not be parallel to the coordinate directions.j+ 1 ∆xbc + 2 2 1 1 ui− 1 .e. j Aabcd d i − 1. We can now evaluate the integral associated with the two-dimensional version of the continuity equation for the surface deﬁned by the points abcd.j− 1 ∆xda = 2 2 2 1 1 (ui+1.j ∆yab − ∆ycd − ∆xab + i+ 1 .j− 1 ∆yda − vi. j i. ui+1.j vi.j − ui−1. j a i + 1.j ∆yab + vi.i − 1. j − 1 S i.j− 2 .j+ 2 + vi.2: Arbitrary grid.j ) ∆yab + (vi.j i− 1 . j − 1 Figure 3.j 2 ∆xcd + 1 i− 2 .j+1 − vi.j − ui− 1 . We have (u dy − v dx) S ≈ 1 1 ui+ 2 .j+ 1 2 1 i.j− 2 ∆xda 2 2 On a cartesian grid we have the relations ∆xab = ∆xcd = ∆ybc = ∆yda = 0 and ∆ycd = −∆yab .j−1 ) ∆xda 2 2 Note that dividing with ∆yab · ∆xda gives a standard central diﬀerence discretization of the continuity equation. i. j + 1 c i. with analogous expressions for the quantities on the bc.j+1 − vi. j − 1 ∂Φ ∂Φ dy − dx ∂x ∂y ∆ybc − ∆yda − ∆xbc + i.j − ui−1. j + 1 i − 1.j−1 + =0 2∆xda 2∆yab Equations with 2nd derivatives: the Laplace equation We use the Gauss or Greens theorem to derive the integral from of the two-dimensional Laplace equation. cd and da faces.j ∆xcd + ui. We ﬁnd 0= V ∂2Φ dV = ∂xj ∂xj S ∂φ nj dS = {2D} = ∂xj If we approximate this integral in the same manner as for the continuity equation we have ∂Φ ∂x ∂Φ ∂x ∂Φ ∂y ∂Φ ∂y ∂Φ ∂x ∂Φ ∂x ∂Φ ∂y ∂Φ ∂y 1 i+ 2 .j ∆xab + ui.j+ 2 ∆ybc − vi.j 2 767676767676767 8 8 8 8 8 8 8 687777777768 8767676767676767 8 8 8 8 8 8 8 67777777768 767676767676767 8 8 8 8 8 8 8 687777777768 687777777768 68 68 68 68 68 68 68 767676767676767 8 8 8 8 8 8 8 687777777768 767676767676767 8 8 8 8 8 8 8 687777777768 767676767676767 8 8 8 8 8 8 8 687777777768 767676767676767 8 8 8 8 8 8 8 687777777768 8 8 8 8 8 8 8 8 6767676767676767 7777777768 c b d a i.j− 2 b i + 1.j ∆yab − vi+ 1 .j+ 1 2 ∆xda 1 i. j + 1 i + 1.

−Φ dx) and the area evaluated as half magnitude of cross products of diagonals.j− 2 ∆y + 1 i− 2 .j Φi.j + Φi. e. Aab ≡ Aa b c d = and the value Φa is taken as the average Φa = 1 (Φi. j.j+1 Ii.j Gi. i+ 1 .j + Φi−1. ∂Φ along the other cell faces and substituting back we obtain a 9 point ∂x ∂y formula which couples all of the nine points around the point i.j+1 + + Ci.First derivatives are evaluated as mean value over adjacent control volumes (areas in the two-dimensional case).j ) ∆y · 1 ∆x · ∆y 1 i+ 2 .j Φi.j+1 Fi.j+ 2 1 i. Bij .j Φi+1.j ≈ 1 Aa b c d abcd (Φ dy. we obtain N equations which in matrix form can be written Ax = b Here xT = (Φ11 .j Φi+1. of the form Ai. Φmn ) and b contains the boundary conditions. i. .j + Φi.j Φi−1. see ﬁgure 3. The matrix A is a N × N with the terms Aij .j − Φi.j 2 ∂Φ ∂y 1 i+ 2 . For the threedimensional case we have 1 V V 1 ∂Φ dV = ∂xj V Φnj dS ⇒ S which in the two-dimensional case is approximated by ∂Φ ∂x where Φ dy ≈ Φi+1.j+1 − 2Φi.j and so on for the rest of the terms.j+ 2 We can evaluate the derivatives as ∂Φ ∂x = (Φi+1. Φ12 . .j ) ∆y ∆x + (Φi. which implies that ∂Φ ∂Φ dy − dx ∂x ∂y ≈ (Φi+1.j ∆ya b + Φb ∆yb c + Φi. .j−1 Φi+1.j−1 + Φi+1. we instead work out the expression for cartesian grids in two-dimensions. etc.j + Φi+1.2.e. as coeﬃcients. i.j−1 ) ∆x ∆y Dividing with the area ∆x · ∆y we obtain the usual 5-point Laplace formula.j + =0 2 ∆x ∆y 2 which of course also can be written in the matrix form shown for the general case above.j Φi−1.j + Φi−1.j−1 + = Ei. We use Gauss theorem over the area deﬁned by the points a b c d .j ∆yc d + Φ∆yd a abcd .j Φi−1.j Φi.j 0 + If we have a total of N = m × n interior nodes. Φ13 . We ﬁnd ∂2Φ ∂2Φ + ∂x2 ∂y 2 ∂Φ ∂x − ∂Φ ∂x dx dy = ∂Φ ∂Φ dy − dx ∂x ∂y ∂Φ ∂y − ∂Φ ∂y ≈ ∆x 1 i.j−1 ) 4 1 ∆xd b · ∆ya c − ∆yd b · ∆xa c 2 Evaluating the derivatives ∂Φ .j − 2Φi.j−1 Hi. .j + + Di.j+1 − 2φi. Φi. Rather than showing the general case in more detail.j − 2φi.j i+ 1 .j Φi+1.j + Φi. .j−1 + + Bi.

j+1 − (uv)i.j + ui.j + =0 2∆x 2∆y The discretization of the streamwise momentum equation needs both ﬁrst and second derivatives.j−1 ) ∆x = 0 2 2 which is equivalent to the ﬁnite diﬀerence discretization or vi.j+1 − vi.j = (−1) i+j · g (t) . j − 1 i. Recall that the ﬁnite volume discretization of the continuity equation becomes 1 1 (ui+1.j pi.j i i+1 Figure 3.j pi+1.j+1 − pi.3.j vi.2 Finite-volume discretizations of 2D NS Co-located Cartesian grid We will apply the ﬁnite volume discretizations to the Navier-Stokes equations in integral form.j + ui−1.j − 2vi. which in two-dimensions can be written (u dy − v dx) = 0 ∂u dS = − ∂t ∂v dS = − ∂t u2 dy − uv dx + p dy − uv dy − v 2 dx − p dx − 1 Re 1 Re ∂u ∂u dy − dx ∂x ∂y ∂v ∂v dy − dx ∂x ∂y where we have used the earlier derived result n = ( dy. The reason is that one solution to these equations are in the form of spurious checkerboard modes.a a i − 1.j − ui−1.j + 2 ∆x ∆y 2 =0 This simple discretization cannot be used in practice without some kind of modiﬁcation.j+1 − vi.j+1 − 2ui.j − ui−1. Consider the following solution to the discretized equations ui.j−1 + + + ∂t 2∆x 2∆y 2∆y − 1 Re vi.j−1 ui+1. 3.j−1 vi+1.j = vi.3: Cartesian ﬁnite-volume grid where the velocity components and the pressure are co-located.j − pi−1. j − 1 j ui.j − ui−1. Thus we obtain the following result for the u-component 2 (uv)i. In addition we choose a Cartesian grid in order to simplify the expressions.j + vi−1. We start with a co-located grid which means that the ﬁnite volume used for both velocity components and the pressure is the same.j−1 u2 ∂ui.j+1 − 2vi.j−1 ui+1.j ) ∆y + (vi.j pi. j − 1 i + 1.j−1 ∂vi.j vi.j + vi.j + ∆x2 ∆y 2 =0 In a similar manner the discretization for the v-component becomes 2 2 (uv)i+1.j+1 − vi. see ﬁgure 3. p = (−1) i+j .j − 2ui. ∆x = ∆y .j i+1.j + + + ∂t 2∆x 2∆y 2∆x − 1 Re ui. We have seen that the ﬁnite volume discretizations of both are equal to standard second order ﬁnite diﬀerence approximations. − dx).j − (uv)i−1.

j pi+1. If it is input into the equation for u and v we ﬁnd the expression 8 dg + g=0 dt Re · ∆x2 ⇒ g = e− Re·∆x2 8t Thus we have a spurious solution consisting of checkerboard modes which are damped slowly for velocity and constant in time for the pressure. j − 1 i + 1.j pi.j−1/2 i i− 1 2 i+ 1 ui−1/2.ui−1/2.j+1 − 2ui+1/2.j+ 1 − vi.j − =0 2 Re ∆x Re ∆y 2 @9@9@@9@9@9@ A A A A A 9A@@@9A@@@@9A@9 @9@9@@9@9@9@ A A A A A 9A@@@9A@@@@9A@9 @9@9@@9@9@9@ A A A A A 9A@@@9A@@@@9A@9 9A 9A 9A 9A 9A 9A @@@9A@@@@9A@9 A A A A A A F@ G G G FG G 9@9@9@@9@9@9@ @@F@F@F@@@ @@@9A@@@@9A@9 FG@@@@@FG@FG A@@@9A@@@9A@9A@9 FG@@@@@FG@FG A A A A F@ G G G FG G 9@9@9@@9@9@@ @@F@F@F@@@ A A A F@ G G G FG G 9@9@9@@9@9@@ @@F@F@F@@@ @A@@9A@A@@9A@9A@9 FG@@@@@FG@FG 9A 9A 9 9A 9 9 @ FG FG FG BCFG @@A@9A@@A@A@9A@9 @C@C@C@@@ FG @@B@B@B@@@ BC@BCFG@@@@@BCFG@BCFG BC@BC@@@@@BC@BC BC BC BC BC @@B@B@B@@@ C C C BC BC@BC@@@@@BC@BC @@B@B@B@@@ C C C DEBC BC@@E@E@E@@@ @@D@D@D@@@ DE@DEBC@@@@@DEBC@DEBC @@D@D@D@@@ E E E DE DE@DE@@@@@DE@DE @@D@D@D@@@ E E E DE DE@DE@@@@@DE@DE DE DE DE DE DE @DE@@@@@DE@DE vi.j pi.j + ui+1/2.j−1 1 ui+3/2.j− 1 ∆x = 0 2 2 1 1 ui+ 2 .j j i+1 vi.j i+1 v pi.j−1/2 i− 1 vi.j − ui. j−1 ence condition.j vi.j j+ 1 2 j j− 1 2 j−1 i i+ 1 2 i+1 i+ 3 2 ∆x + vi. j −a 1 i − 1. j i i+1 ui.j 2 i i+2 1 i− 1 i+ 1 3 2 i+ 2 i +1 1 ji + 1 i+ 2 3 2 i+ 2 j +1 1 ji + 1 3 j−2 i+ 2 2 j j −1 1 j+1 2 j−2 j j−1 j−1 2 ence condition.j−1/2 1 2 T S¡ T¡ST S¡ ¡ST QPQPRR QPRQPR QPQPRR RR PP I H¡ I¡I H¡ ¡HIH i. Staggered cartesian grid To eliminate the problem with spurious checkerboard modes. control volume for normal momentum equation (v) .j−1/2 pi.j−1/2 u2 ∂ui+1/2.j − ui− 2 . which satisﬁes the divergence constraint. where the control volumes for the streamwise. and can be avoided if they are coupled by one sided diﬀerences or if they are damped with some type of artiﬁcial viscosity. j − 1 i i.j i.j i− control volume for divergence condition. m equation (u) .j − ui− 2 . j − 1 j i+ i + 1.j+1/2 2 ui−1/2. These modes are a result of the very weak coupling between nearby ﬁnite volumes or grid points. The control volume for the continuity equation is centered abound the pressure point and the discretization becomes 1 1 ui+ 2 .4: Cartesian ﬁnite-volume grid where the velocity components and the pressure uses staggered grids. spanwise and pressure are diﬀerent.j+1/2 − (uv)i+1/2.j + + + − ∂t ∆x ∆y ∆x 1 ui+1/2.j pi.j+1/2 pi.j or The control volume for the streamwise velocity is centered around the streamwise velocity point and the discretization becomes 2 (uv)i+1/2. This mode will corrupt any true numerical solution and yields this discretization unusable as is.j ∆y + vi.j+1/2 ui−1/2. j − 1 j i − 1.4. Figure 3. control volume for streamwise momentum equation (u) .j pi.j vi.j+1/2 vi.j ui.j vi. j −a 1 i.j i+1.j i ui.j vi.j vi.j pi.j − pi. j − a 1 i + 1.j − 2ui+1/2. j − 1 i ui.j vi.j+ 1 − vi.j− 1 2 2 ∆y =0 . we can use a staggered grid as in ﬁgure 3.j + ui−1/2.j vi.

The control volume for the normal velocity is centered around the normal velocity point and the discretization becomes 2 2 (uv)i+1/2.j .j+1/2 1 vi. For this discretization no checkerboard modes possible. consist given values of u and v. 3 − v1.j+1/2 − (uv)i−1/2. BC in ﬁgure 3.1 . 3 = 2v 1 . so called Dirichlet conditions.0 .j+1/2 . 2 V V V V V V V V V V V V V V V V UVUV U&U&U&&U&U&U&U&U&U&&U&U&U&U&U&U&U&&& V&V&V&&V&V&V&V&V&V&&V&V&V&V&V&V&V&&& U&U&U&&U&U&U&U&U&U&&U&U&U&U&U&U&U&&& &&&UVUV&&&&&&&UVUV&&&&&&&&UVUV&&UVUV C 1 2 i 1 3 u 3 .j+1/2 .j−1 2j j−1 2 Inﬂow A 3 v1.1 Figure 3.5 for a deﬁnition of the points near a boundary.j+1 − vi. 3 ). the pressure or the convective terms. for example.3 = 2 1 v 3 + v0.1 = 2 The boundary conditions on the inﬂow. Boundary conditions To close the system we need to augment the discretized equation with discrete versions of the boundary conditions. We can ﬁnd this value 2 as follows 3 0 = u2. ui. 2 = v2. requires v0. (uv)i+1/2. The normal points are located on the boundary so that the no ﬂow condition simply become 1 v1. (uv)i+1/2.0 2 2 ⇒ ⇒ Solid wall 3 u 3 . .1 3 u 2 .j+1/2 + vi.0 = −u 2 .j ∂vi.j+1 .j + + + − ∂t ∆x ∆y ∆y 1 vi+1. We can ﬁnd this value 2 as follows 1 v2. since there is no decoupling in the divergence constraint.5.j+1/2 . consists of the no ﬂow and the no slip conditions.5: Staggered grid near the boundary.j+1/2 vi. 1).1 2 v0.j terms need to be interpolated from the points where the corresponding velocities are deﬁned. . .j . The boundary conditions on the solid wall. i.1 + u 2 . 2 where the u2 i+1. (uv)i−1/2.j+1/2 − 2vi. See ﬁgure 3.5. 3 2 2 1. u 1 .2 . u = v = 0.e.j−1/2 − =0 2 Re ∆x Re ∆y 2 2 where the (uv)i+1/2. 2 1 B 0 0 p1.j−1/2 terms need to be interpolated from the points where the corresponding velocities are deﬁned. AB in ﬁgure 3. 2 . for example. 2 2 3 The evaluation of the v-equation at the point (1.j+3/2 − 2vi. ui. The streamwise points are located on the boundary so that the boundary condition is simply consist of the known values u 1 . i. 1 = · · · = 0 2 3 The evaluation of the u-equation at the point ( 3 .j+1/2 + vi−1. u2 i.e. requires u 2 .j+1 − pi.j+1/2 pi. 3 2 2 2 2 .

j+ 2 + b(u.j+ 2 Re∆y 2 4∆y i+1.j 4∆y i+1.j − 2ui+ 2 .j+ 2 vi.j+ 1 can in a similar way be written 2 1 Bi.j−1 1 ui+ 2 .j 2 1 + Bi. For the streamwise and normal velocity we ﬁnd 1 3 u− 2 .j+1 − pi.j pi+1.j− 1 2 2 − 2 R ∆y + = {expand and interpolate using grid values} 1 1 = b(u.j + =0 ∂t ∆x ∂vi.j− 1 u2 i+1.j 1 + Ai+ 2 .j Re∆x2 4∆x i− 2 .j − ui. i.j + 4∆x i+ 2 . v)nb vnb nb .j+ 2 ) − i.j− 1 ) + 2 2 2 2∆x i+ 2 .j+ 1 pi.j ) + (v + ui+ 1 .j ∆y 1 1 + vi−1.j+ 1 2 − R ∆x2 = (uv)i+ 1 .j+ 2 − 2vi.j+ 2 + =0 ∂t ∆y D =0 i.j + ui− 2 .j− 2 − vi. 3 2 2 Note that no reference is made to the pressure points outside the domain.j+ 1 − (uv)i+ 1 .j− 2 Re∆y 2 = 1 1 = a(u.2 = u 2 .j can be written 2 Ai+ 1 .3 Summary of the equations The discretized equations derived for the staggered grid can be written in the following form 1 ∂ui+ 2 . so that in this formulation no boundary conditions for p is needed so far. v)i+ 2 .j Re∆x2 1 1 1 1 1 + v 1 1 + v 1 1 (v ui+ 2 .j + a(u.j + ui+ 1 . the boundary conditions may consist of ∂u = ∂x = 0.j i.5 is an outﬂow boundary. v0.j + ui− 1 . 3.j+ 2 + vi.j+ 1 − (uv)i− 1 .2 .j+ 3 − 2vi.j ui+ 2 .j+ 1 2 2 2 2 2 2 vi.j+ 2 Re ∆x2 ∆y 2 1 1 1 1 3 1 u 3 − u 1 − ui+ 2 . 3 = v1.∂v If AB in ﬁgure 3.j+ 1 − vi+1. These conditions should be evaluated at i = 2 .j = 2 2 (uv)i+ 1 .j 2 2 2 2 + ∆x ∆y 3 1 1 1 ui+ 1 .j+1 − vi. If this is the case particular care has to be taken so that those conditions do not upset the divergence free condition.j 2 2 2 − − Re ∆x2 Re ∆y 2 = = {expand and interpolate using grid values} 1 1 1 2 1 1 − v 1 1 (u 3 − ui− 2 .j+1 + (v ui+ 2 . However.j+ 2 = ∆x 1 1 vi+1.j where the expression for the Ai+ 1 . v)i. where the sum over nb indicates a sum over the nearby nodes. The expression for Bi.j+ 2 1 vi.j− 1 ) − 2 4∆y i+1.j − pi. v)nb unb nb The last line summarizes the expressions.e. depending on the particular discretization of the time derivative and other the details of the solution method one may need to augment the the above with boundary conditions for the pressure. ∂x 1 so called Neumann conditions.j−1 + i.j+1 − 2ui+ 2 .

i. v) 0 Gx fu d v + 0 B(u. linear algebraic operator from stremwise pressure gradient linear algebraic operator from normal pressure gradient linear algebraic operator from x-part of divergence constraint linear algebraic operator from y-part of divergence constraint vector of source terms from BC With obvious changes in notation. The velocity at the next time level thus becomes . Sometimes this method and its generalizations are also called splitting methods. we can write the system of equations in an even more compact form as d dt u 0 + N (u) D G 0 u p = f 0 If this is compared to the ﬁnite element discretization in the next chapter.j+ 2 and b(u. v) Gy v = fv dt 0 Dx Dy 0 p 0 vector of unknown streamwise velocities vector of unknown normal velocities vector of pressure unknowns non-linear algebraic operator from advective and viscous terms of u-eq. discrete version of pressure Poisson equation. non-linear algebraic operator from advective and viscous terms of v-eq. v)nb .j− 1 2 2 ∆y Here we we have used the deﬁnitions u v p A(u.j = ∆x + vi.4 Time dependent ﬂows The projection method For time-dependent ﬂow a common way to solve the discrete equations is the projection or pressure correction method. v)i+ 1 .j 2 Di. v) Gx Gy Dx Dy f - These equations. We start to discretize the time derivative with a ﬁrst order explicit method n+1 − un u + N (un ) un + G pn+1 = f (tn ) ∆t Dun+1 = 0 Then we make a prediction of the velocity ﬁeld at the next time step u∗ not satisfying continuity. v)i. including the boundary conditions.j+ 1 − vi. can be assembled into matrix form. v)nb are analogous to a(u. it is seen that the form of the discretized matrix equations are the same. We have ∗ n u − u + N (un ) un = f (tn ) ∆t n+1 u − u∗ + G pn+1 = 0 ∆t Dun+1 = 0 Next.j and a(u. We illustrate the method using the system written in the compact notation deﬁned above. apply the divergence operator D to the correction step DG pn+1 = 1 D u∗ ∆t Here DG represents divgrad = Laplacian. 3. and 2 therefore not explicitly written out.j − ui− 1 . v) B(u. and then a correction involving p such that Dun+1 = 0.1 where the expressions for b(u. The expression originating from the continuity equation is 1 ui+ 2 . We have u u A(u.e.

j+ 1 pi.j i. and the corresponding expressions for un+1.j 2 + =0 ∆t ∆x n+1 ∗ n+1 n+1 vi.5 we have 2 2 pn+1 − pn+1 1 2. In general it is preferred to make the time discretization ﬁrst for the full Navier-Stokes equations and afterward split the discrete equation into one predictor and one corrector step.j+ 2 n + Bi.j is Di. If we take the vector equation for the correction or projection step and project it normal to the solid boundary in ﬁgure 3.j i.j−1 + = ∆x2 ∆y 2 ∆t ∗ where Di.1 5 3 2 2 ∆x + ∗ ∗ v2. centered abound the point (2.j i+1.e. 3 − vΓ 2 ∆y if we substitute the value of pn+1 − pn+1 from the boundary condition obtained for the pressure into the 2.0 ∗ =− v n+1 − vΓ 1 ∆y ∆t 2. we can choose n+1 ∗ vΓ = v2. To see this we write the pressure Poisson equation. A number of other splitting methods have been devised for schemes with higher order accuracy for the time discretization.j i. This is a discrete version of the projection of a function on a divergence free space discussed in the section about the role of the pressure in incompressible ﬂow.j and vi.j+1 − pi.j i−1.j+ 1 = 0 2 ∆t n+1 ui+ 1 .j+ 2 − vi. In general 2 it is important to choose this boundary condition such that discrete velocity after the correction step is divergence at the boundary.j = 0 i+ 2 ∆t n ∗ 1 1 vi.j 2 2 + =0 ∆t ∆y 2 and the projection step as n+1 By solving for un+1.1 2. .j − u∗ 1 . that 2 this is a numerical artifact and does not imply that the real pressure gradient is zero at the boundary.j pn+1 − pn+1 i+ 2 i+1. see the discussion in the end of the section on the role of the pressure.j+ 1 − vi.j and i− 1 i+ 1 n+1 vi.j i. in the expression for the divergence constraint Di.1 1. ∗ for this particular discretization it turns out that we can choose vΓ arbitrarily since it completely decouples from the rest of the discrete problem. Thus we ﬁnd that we can choose the value of vΓ arbitrarily. In particular. Since the discretization of the time derivative is explicit there is no other place where the values of v ∗ or u∗ on the ∗ boundaries enter.j evaluated using the discrete velocities at the prediction step. close to the boundary.1 2.j+ 1 above and using this. 1 is an unknown value of the velocity in the prediction step at the boundary. See Peyret and Taylor for details.2 2.1 2.j+1 i. Note. in which case we obtain a Neumann boundary condition for the pressure.j 2 2 + An 1 . u∗ and v ∗ .1 − u∗ . We ﬁnd pn+1 − 2pn+1 + pn+1 pn+1 − 2pn+1 + pn+1 1 2. 2 ∗ ∗ where vΓ = v2. However.j − ui+ 1 . In order to be able to discuss the boundary condition on the pressure Poisson equation we write the prediction and correction step more explicitly. we need additional boundary conditions.un+1 = u∗ − ∆tG pn+1 Note here that un+1 is projection of u∗ on divergence free space.1 + = 2 2 ∆x ∆y ∆t u∗ .1 2. i. Using the staggered grid discretization the prediction step can be written ∗ n ui+ 1 .0 3.j− 1 . When the discrete pressure Poisson equation is solved we need values of the unknowns which are outside of the domain.1).0 ∗ above equation we can se that vΓ cancels in the right hand side of the pressure Poisson equation. 1 . however.j we ﬁnd the following explicit expression for the 2 discrete pressure Poisson equation ∗ pn+1 − 2pn+1 + pn+1 pn+1 − 2pn+1 + pn+1 Di.

substituting back the deﬁnitions of λ and α we ﬁnd the following stability conditions 1 U 2 ∆tRe ≤ 1 2 0 2∆t ≤ 1 Re∆x2 .Time step restriction For stability it is necessary that prediction step is stable. This implies that Φ (s) = − λ2 − α2 s + λ2 − α ≤ 0 ∀s : 0 ≤ s ≤ 1 Φ (s) is linear function of s and thus the the limiting condition on Φ (s) is given by its values at the two end points of the s-interval. V0 . It thus suﬃces to consider the equation ∂u ∂u ∂u 1 + U0 + V0 = ∂t ∂x ∂y Re For simplicity we will here consider one-dimensional version ∂u 1 ∂2u ∂u + U0 = ∂t ∂x Re ∂x2 which in discretized form can be written un+1 = un + ∆t −U0 j j un − u n 1 un − 2un + un j+1 j−1 j+1 j j−1 + 2∆x Re ∆x2 2 u where we have chosen to call the discrete points where the u-velocity is evaluate for x j . α= . numerical evidence indicates that this is suﬃcient. together with the deﬁnitions 2∆t U0 ∆t . In the predictor case the momentum equations decouple if we linearize them abound the solution U 0 . We ﬁnd Φ (0) = λ2 − α ≤ 0. We introduce un = Qn · ei·ωxj j into the discrete equation. which implies that ⇒ λ2 ≤ α ≤ 1 Φ (1) = α2 − α ≤ 0 For the two dimensional version of the predictor step Peyret and Taylor [8] ﬁnd the following stability limits 1 U 2 + V 2 ∆t · Re ≤ 1 0 4 0 4∆t ≤ 1 Re∆x2 Note that this is a stronger condition than in the one-dimensional case. ξ = ω · ∆x ∆x Re ∆x2 This implies that the ampliﬁcation factor Q can be written λ= Q = 1 − iλ sin (ξ) − 2α sin2 We ﬁnd that the absolute value of Q satisﬁes |Q| where s = sin2 ξ 2 ξ 2 = = 1 − 4 λ2 − α 2 s 2 + 4 λ2 − s 2 s 1 − 4s λ2 − α 2 s − λ 2 + α ≤ 1 .

. let y = B y which implies that ˆ AB y = b ˆ and use the iteration scheme on this modiﬁed equation. as in the Gauss-Seidell method for example.3. However. For this method we have Ax = b A=L−U (L − U ) x = b L xn+1 = U xn + b Application to the steady Navier-Stokes equations When the distributive iteration method is applied to the steady Naiver-Stokes we have A= N (u) D G 0 we can obtain AB in the following block triangular form AB = if the matrix B has the form B= I 0 −N (u)−1 G I ≈ I 0 ˜ −N (u)−1 G I N (u) D 0 −DN (u)−1 G ˜ where N (u)−1 is a simple approximation of the inverse of the discrete Navier-Stokes operator. in many cases it is more eﬀective to use other iteration methods. Note that R is a discrete Laplace like operator. We will here introduce a general distributive iteration method and apply it to the discretized steady Navier-Stokes equations. and iterate the solution in time to obtain a steady state.5 General iteration methods for steady ﬂows Distributive iteration For steady ﬂows we can still use the methods introduced for unsteady ﬂows. Now. We now split AB = M − T where M= Q D 0 R and where Q is an approximation of N (u) and R is an approximation of −DN −1 G. If B = I we obtain a regular iteration method where residual r k is used to update y k by inverting M . a simpliﬁed part of A. We divide the matrix AB in the following manner AB = M − T and use this to deﬁne the iteration scheme M B −1 y k+1 = T B −1 y k + b = M B −1 y k + b − A y k which can be simpliﬁed to y k+1 = y k + BM −1 b − A y k rk ⇒ M y = Ty + b ˆ ˆ where r k is the residual error in the k:th step of the iteration procedure. For the steady case the discretized equations can be written in the form N (u) D G 0 u p = f 0 which is of the form A y = b.

where the velocities and the pressure is updated uk+1 = uk + ωu δu pk+1 = pk + ωp δp The under relaxation is needed for the method to converge. Thus we have the following iteration steps 1. we have the distribution step BM −1 rk δu δp =B δu ˜ δp ˜ which becomes = ˜ δ u − N −1 G δ p ˜ ˜ δp ˜ 4. we calculate the residual b − A yk = f 0 − N (uk ) D G 0 uk pk = k r1 k r2 2. . Third. First.The SIMPLE (Semi-Implicit Method for Pressure Linked Equations) by Pantankar and Spalding (1972) can be thought of as a distributive iteration method. thus ωu and ωp is usually substantially lower that one. we calculate M −1 rk which implies k Q δ u = r1 ˜ ⇒ k δ u = Q−1 r1 ˜ k ˜ R δ p = r2 − D δ u ˜ 3. there is an under relaxation step. Fourth. Wesseling [10] indicates that by choosing ˜ N (u)−1 Q R = [diagN (uk )]−1 = N (uk ) = −D[diagN (uk )]−1 G we have a method which is essentially the original version of the SIMPLE method. Second.

j Γ of the domain can be decomposed into two portions Γ0 and Γ1 . 71 . j − 1 method. as in ﬁgure 4. We i. j + 1 i + 1.1b) i i+ 1 2 ∂u =0 on Γ1 . The Dirichlet j−1 2 isothermal condition. whereas the Neumann condition ∂u = 0 boundary condition u = 0 on Γ0 is called an ∂n j−1 on Γ1 is an adiabatic condition. 1 j i i+1 4. We assume that the boundary pi.1 ∂u v1. Incompressibility yields that ∂xj = 0.j velocity ﬁeld Uj is known (from a previous numerical solution. However.1c) i+1 ∂n 3 i+ 2 ∂Uj +1 2 The coeﬃcient ν > 0 is the thermal jdiﬀusivity.1 for instance. so that nj Uj = 0 on Γ1 .j We consider the steady ﬂow of an incompressible ﬂuid in a bounded domain Ω.j−1/2 ∂u ui−1/2. j − 1 introduce FEM ﬁrst for a scalar advection–diﬀusion problem before i + 1.1).1 0 n 1 Γ0 2 i j Ω Inﬂow Solid wall Γ1 dx dy − dx dl = ( dx. This simpliﬁes the analysis but is not essential. We are vi. 3 v = + v 2 u. dy) n dl = ( dy.1: A typical domain associated with the advection–diﬀusion problem (4.1a) Uj 1 − ν 2 u = f i −∂xj 2 Chapter 4 Finite element methods for incompressible ﬂow Figure 4. A Deriving the integral form that is used to form the ﬁnite-element discretization requires some vector B calculus. and we assume that its pi. (4. j i + 1. (4.j vi. j i. j − discussing the Navier–Stokes equations. j + 1 i.1 FEM for an advection–diﬀusion problem ui. j + 1 c c b b i − 1. a the integral form—called the variational form—is usually diﬀerent from the one used for the ﬁnite-volume i − 1. modeled by the advection–diﬀusion problem interested to compute the temperature ﬁeld vi. We assume that j Γ1 is a solid wall.u=0 on Γ0 . Moreover. (4. diﬀerent integral forms may be used for the same equation depending on circumstances such as boundary conditions. (4.2) 2 ∂xi ∂xi ∂xi ∂xi p1.j+1/2 u in the domain. j Aabcd d d a The ﬁnite-element method (FEM) approximates an integral form of the governing equations. The product rule of diﬀerentiation yields C u3 ∂v ∂u ∂ 2 . − dx) i − 1. for instance).j in Ω.

Note that the boundary conditions are incorporated into the variational problem. For this. and utilizing expression (4. a weak solution may not be a classical solution since functions in V may fail to be twice continuously diﬀerentiable. Multiplying equation (4.1).) ∂n ∂xi Let v be any smooth function such that v = 0 on Γ0 . Let M be the number of triangles and N the number of nodes in Ω ∪ Γ1 . The Neumann condition on Γ1 is deﬁned implicitly through the variational form and is therefore called a natural boundary condition.3) Integrating expression (4. that is. as in ﬁgure 4.1) (also called a classical solution) satisﬁes the varational form vUj Ω ∂u dΩ + ν ∂xj Ω ∂v ∂u dΩ = ∂xj ∂xj f v dΩ Ω (4. The requirement that the gradient-square of each function in V is bounded should be intepreted as a requirement that the energy is bounded. (4. v = 0 on Γ0 .5) shows that a classical solution is a weak solution.4) yields ∂u vUj dΩ − ν ∂xj ∂u vUj dΩ + ν ∂xj ∂v ∂u dΩ − ν ∂xj ∂xj 0 ∂u dΓ. However.2) and applying the divergence theorem on the left side yields v Γ ∂u dΓ = ∂n Ω ∂v ∂u dΩ + ∂xi ∂xi v Ω 2 u dΩ.4) (Note that ∂u ∂u = ni . integrating over Ω. v ∂n f v dΩ = Ω Ω v Ω 2 u dΩ = Ω Ω Γ where the last term vanishes on Γ0 due to the requirements on v. u.1.5). Let h be the largest side of any triangle.5) for each smooth function v vanishing on Γ0 We will now “forget” about the diﬀerential equation (4. V = v: ∂xi ∂xi Ω The space V is a linear space.1 Finite element approximation Let us triangulate the domain Ω by dividing it up into triangles. The derivation leading up to expression (4. and directly consider the variational form (4. that is. The variational problem will thus be Find u ∈ V such that ∂u ∂v ∂u dΩ + ν dΩ = vUj ∂xj ∂xj ∂xj Ω Ω f v dΩ Ω ∀v ∈ V.1a) with v.Recall the divergence theorem Ω ∂wi dΩ = ∂xi ni wi dΓ. Let Vh be the space of all functions that are .1c) We have thus shown that a smooth (twice continuously diﬀerentiable) solution to problem (4. we need to introduce the function space ∂v ∂v dΩ < +∞.2.6) is called a weak solution to advection–diﬀusion problem (4. v ∈ V ⇒ αu + βv ∈ V ∀α.1). (4. The Dirichlet boundary condition on Γ0 is explicitly included in the deﬁnition of V and is therefore referred to as an essential boundary condition. 4. . and on Γ1 due to boundary condition (4. β ∈ R. Γ (4.continuous on Ω.6) A solution to variational problem (4. the nodes on Γ0 are excluded in the count.

the black nodes.6).vanishing on Γ0 . .1 0 1 2 i j Inﬂow Solid wall n Γ0 Γ1 Ω Figure 4. 3 2 p1. m = 1. . N uh (xi ) = n=1 u(n) Φ(n) (xi ) . N in equation (4. . ∂xj ∂xj Cmn = Ω Φ(m) Uj ∂Φ(n) dΩ.2: A triangulation. Note that Vh ⊂ V . Here.1. Ω (4. Substituting expansion (4.11) .3: The “tent or “hat basis function associated with continuous. Once the value of a function uh in Vh is known at all node points. ∂xj (4. we may choose vh = Φ(m) ∈ Vh for m = 1. (4. which are functions in Vh that are zero at each node point except node n.8) where u(n) denotes the value of uh at node n.10) This is a linear system system Au = b. where it is unity.9).linear on each triangle.2 The algebraic problem.p1.7) yields N u(n) n=1 Ω vh U j ∂Φ(n) u(n) dΩ + ν ∂xj n=1 N Ω ∂vh ∂Φ(n) dΩ = ∂xj ∂xj f vh dΩ Ω ∀ vh ∈ Vh . We get a ﬁnite-element approximation of our advection-diﬀusion problem by simply replacing V by Vh in variational problem (4. . 4. Thus. it is easy to reconstruct it by linear interpolation. where A = C + νK and Kmn = Ω ∂Φ(m) ∂Φ(n) dΩ. . .1 2 v1. .1 0 1 2 i j Inﬂow Solid wall n Γ0 Γ1 Ω Γ0 Γ1 u 3 . (4. . piece-wise linear functions. .3). This interpolation can be written as a sum of the “tent” basis functions Φ (n) (ﬁgure 4. The ﬁnite-element method is a Galerkin approximation in which the subspace is given by piecewise polynomials. (4.7) Restricting the function space in a variational form to a subspace is called a Galerkin approximation. Assembly. Note that expansion (4. which then becomes N u(n) n=1 Ω Φ(m) Uj ∂Φ(n) dΩ + ν u(n) ∂xj n=1 N Ω ∂Φ(m) ∂Φ(n) dΩ = ∂xj ∂xj f Φ(m) dΩ. N . that is: Find uh ∈ Vh such that ∂uh ∂vh ∂uh vh U j dΩ + ν dΩ = ∂xj ∂xj ∂xj Ω Ω f vh dΩ Ω ∀vh ∈ Vh . . the number N are Figure 4.9) Since each basis function is in Vh .8) into FEM approximation (4.8) forces uh = 0 on Γ0 since it excludes the nodes on Γ0 in the sum.

The components of K are Kmn = Ω Note that Kmn vanishes for most m and n.4).12). n being the three nodes of triangle p and adding these 9 contribution to matrix K. but numerical quadrature is usually needed. for the particular case of the structured mesh depicted in ﬁgure 4. (p) 4.5). Let M be the number of internal nodes in the horizontal as well as the vertical direction (the solid nodes in ﬁgure 4. and the area of each triangle is |Tk | = h2 /2. By the expressions in X`XXX`X`XX XXWXXXWXWX a `a a a WY Y WY Y Y `aXXX`aXXX`aX`a XWYXXXWYXXXWY a`aX`a `XaX`a`a`a `a `a WYWYWY WY WYWYWY WY WY X`XX`aXaXaX`aX`a XWYXYXXWYXYXYXWY XaXXX`X`XX XXWXXXWXWX `aX`XX`aXaXaX`aX`a XWYXYXXWYXYXYXWY XaXXX`X`XX XXWXXXWXWX `aX`XX`aXXX`aX`a XWYXXXWYXXXWY X`XX`aX`X`X`aX`a XWYXWXXWYXWXWXWY aXXXaXaXX XXYXXXYXYX a`a a a WYWYWY Y WYWYWY Y Y `XX`X`XX XXWXXXWXWX a`aXaXXXaXaXX XXYXXXYXYX `X`XX`aX`X`X`aX`a XWYXWXXWYXWXWXWY aXaXXXaXaXX XXYXXXYXYX `X`XXX`X`XX XXWXXXWXWX XXX`a`aXXX`a`aX`a`a XWYWYXXXWYWYXXXWYWY `a`a WY WY m m+2 1 M h f Φ(1) dΩ Ω i j Inﬂow Solid wall n Γ0 Γ1 Ω M2 b= . u(n) The union of all triangles constitutes the domain Ω and that the triangles do not overlap. Kmn = 0 only when m and n are associated with nodes that are nearest neighbors. the “stiﬀness matrix” (a term borrowed from structural mechanics).m+2 = 0. Km.1. we will compute the elements of K. Since the functions are piecewise linear. (p) so Kmn is easy to compute exactly. Matrix K can thus be computed by looping (p) all M triangles. To compute matrix element (4. since basis functions Φ(m) and Φ(m+2) do not overlap (ﬁgure 4. Note that the derivatives of Φ(n) are constant on each triangle.4: An example mesh that gives a particular simple stiﬀness matrix K u(1) . any integral over Ω can be written as a sum of integrals over each triangle. In fact. ∂xj ∂xj p=1 (p) M ∂Φ(m) ∂Φ(n) dΩ ∂xj ∂xj (4. .) The width of each “panel” is then h = 1/(M + 1).h Figure 4. the gradient is constant on each triangles and may thus easily be computed by ﬁnite-diﬀerences in the coordinate direction (ﬁgure 4. Matrix C and vector b can also be computed by element assembly. For instance. . Thus. f Φ(N ) dΩ Ω Tp ∂Φ(m) ∂Φ(n) (p) dΩ = Kmn .3 An example As an example.12) . we need expressions for the gradients of basis functions Φ (m) and (m+1) Φ . calculating the element contribution Kmn for m. except if Uj and f happen to be particularly “easy” functions (such as constants). .4. u= . ∂Φ(m) ∂Φ(n) dΩ = ∂xj ∂xj p=1 M Kmn = Ω Note that Kmn vanishes as soon as m or n is not a corner node of triangle p. Thus Kmn contributes to the sum only when m and n both are corner nodes of triangle p.4.

m m+2 1 1 Φ 1 (m) −h. Thus. h..0 on T1 . 1 h 0 Figure 4.. ..0 T2 Φ(m+1) h 0 on all other triangles T1 1 . . 1 j on T2 . 0 1 .m+1 = Ω Φ(m) · Φ(m+1) dΩ = k=1T k Φ(m) · Φ(m+1) dΩ = − and in the same manner.m 2 m+2 T3 i on T1 .m−1 = Km..6: The support for basis functions Φm and Φm+1 are marked with vertical and horizontal stripes respectively.6 yields 2 Km. that is. Altogether. −h 1 h. it is only on these two triangles that both functions are nonzero at the same time.13) . − . 2 h h h 2 .−1 h h 0. n T1 T5 on T5 . h Φ(m+1) = 1 1 on T2 . and T the M -byM matrix 4 −1 −1 4 −1 . h T4 Inﬂow T2 on T3 . b b f@f cb f b b f b b b b b b b b b b b bc e@e@@f cccccccccccccccccbb @f@@e@f@f@@ef ef@e@ef@f@e@e@ef@ef@e d b e@e@@ @f@ef@f@ccccccccccccccccccccccccccccdbdbdbdbdbdbdbdbdbdbdbdbdbdbdbe@f@f@ef@@e b cb fef@e@@e@e@e@@ef e@f@ef@@f@f@ef@@e d f@e@@f cbdcc e@e@@ e@f@@e@f@f@@ef @e@ef@f@ccccccccccccccccccccccccccccdbdbdbdbdbdbdbdbdbdbdbdbdbdbdc e@e@ef@@e b cb @f@@e@f@f@@ef ef@e@ef@f@e@e@ef@@e @e@@f@e@e@@ef f@@e@f@f@@ ef@f@ef@f@ccccccccccccccccccccccccccccdbdbdbdbdbdbdbdbdbdbdbdbdbdbdbe@f@f@ef@ef@e b cb f@e@@e@e@e@@ef@ e@@efef@@ccccccccccccccccccccccccccccdbdbdbdbdbdbdbdbdbdbdbdbdbdbdc @@efef@@ee d c b cb fef@ef e@f@@ef cbd ef ef ef @e@ef@@f@f@ef@@e @f@@f cc e@e@@ef ef@e@ef@f@ccccccccccccccccccccccccccccdbdbdbdbdbdbdbdbdbdbdbdbdbdbdbe@f@f@ef@@e b cb @f@@e@e@e@@ef d ef@e@ef@f@f@f@ef@@e @f@@e@e@e@@ef ef@e@ef@@ccccccccccccccccccccccccccccdbdbdbdbdbdbdbdbdbdbdbdbdbdbdc @@ef@@e b cb @e@ef@e@f@e@ef@ef@e f@@f@e@f@@ef d df e@f c c cbe@f@e@@ df d d fef@f@@f@dcb e@dbc b e@dbdf@b ef@dbd@b ef@ e@e@ef@e@ccccccccccccccccccccccccccccdbdbdbdbdbdbdbc e d cd d ef@ef@@ef@d ef@d ef@@d ef@ d d d d d @@efef@@c@c@ced fef@c@ced e cccccd 2 + ∂Φ(m) ∂y 2 dΩ 1 2 h2 1 |T1 | − 2 |T2 | = − 2 = −1. 0 1 1 − . . we have Kmm = Ω ∂Φ(m) ∂Φ(m) dΩ = ∂xj ∂xj 6 k=1T k ∂Φ(m) ∂x 1 1 1 1 1 1 |T1 | + 2 2 |T2 | + 2 |T3 | + 4 |T4 | + 2 2 |T5 | + 2 |T6 | 2 h h h h h h 1 h2 =8 2 =4 h 2 = The common support for basis functions Φm and Φm+1 is only the two triangles marked in ﬁgure 4. Solid wall Φ(m) on T4 . h on T1 . .5. K becomes the block-tridiagonal M 2 -by-M 2 matrix T −I −I T −I . ﬁgure 4. h h 0.6. −I T −I −I T where I is the M -byM identity matrix. K= .5: The gradient of test function Φ for m being the (black) middle node.m−M = −1. (4. Γ0 Γ1 on T6 .. T6 Ω on all other triangles.. − 1 h 1 = −h. Utilizing the gradient expressions given in ﬁgure 4. −1 4 −1 −1 4 Km. T = .m+M = Km. . . 2 M (m) M Figure 4. . . (m) 1 h Φ = on T2 . . a typical row in the matrix–vector product Ku becomes 4um − um+1 − um−1 − um+M − um−M .

5 Stability and accuracy We have deﬁned the ﬁnite-element approximation and showed that the ﬁnite-element approximation yields a linear system Au = b that has a unique solution. However. we recognize the classical ﬁve-point stencil for the negative Laplacian − 2 . after dividing expression (4. vh ≡ 0. Then vh = N N N n=1 v (N ) Φ(n) . We will study the stability and accuracy of the numerical solution. if vh = Const. v T Cv = 0 for each v. A positive-deﬁnite matrix is a particular example of a nonsingular matrix.that is. A is positive deﬁnite.11) that Kmn = Knm . K is a symmetric matrix (K T = K). Let vh ∈ Vh . for a general unstructured mesh.13) with h2 . ∂xi ∂xi with equality if and only if = 0.1.1. Integration by parts also shows that matric C is skew symmetric (C T = −C). A common technique in ﬁnite-element codes is therefore to use sparse representation. v)V = Ω u· v dΩ = Ω ∂u ∂v dΩ. The number of nonzero elements on each row does not grow (on average) when the mesh is reﬁned (since the number of nearest neighbors in a mesh does not grow). Thus our ﬁnite-element discretization is for this particular mesh equivalent to a ﬁnite-volume or ﬁnitediﬀerence discretization. It is immediate from expression (4. that is. Theorem 1. matrix element A mn = 0 only in the columns where n represents a nearest neighbor to node m. that is. Basic deﬁnitions and inequalities The most basic integral norm is 1/2 v L2 (Ω) = Ω v 2 dΩ For the current problem. However. Thus From Theorem 1 also follows that A is positive deﬁnite. Thus. that is. .4 Matrix properties and solvability Matrix A is sparse: the elements of martix A are mostly zeros. FEM usually uses integral norms for analysis of stability and accuracy. . v (N ) )T .14) . . that is. Now the question is whether the approximation is any good. We start by reviewing some mathematical concepts that will be used in the analysis. that is. 4.7) has a unique solution. that is. . to store only the nonzero elements and pointers to matrix locations. To see this. Denote v = (v (1) . A is not symmetric whenever U j = 0. We thus conclude that the linear system resulting from the ﬁnite-element discretization (4. At row m. note that the skew-symmetry of C yields that T v T Cv = v T Cv = v T C T v = −v T Cv v T Av = v T (K + C) v = v T Kv + v T Cv = v T Kv > 0 for each v = 0. the matrices become sparser and sparser as the mesh is reﬁned. ∂xi ∂xi (4. v T Kv > 0 for each v = 0. K is positive deﬁnite. v T Kv = m=1 n=1 N v (n) Ω ∂Φ(n) ∂Φ(m) dΩ v (m) ∂xi ∂xi ∂ v (m) Φ(n) dΩ = ∂xi n=1 N = ∂ v (m) Φ(m) Ω m=1 ∂xi ∂vh ∂xi Ω ∂vh ∂vh dΩ ≥ 0. a more fundamental quantity is the energy norm 1/2 v and associated inner product V = Ω | v|2 dΩ = Ω ∂v ∂v dΩ ∂xi ∂xi 1/2 (u. Recall that a linear system Au = b has a unique solution for each b if the matrix is nonsingular. 4. Proof. But since vh |Γ0 = 0. the ﬁnite-element discretization will not give rise to any obvious ﬁnite-diﬀerence stencil.

it expresses the norm in terms of something that is optimized.7) yields uh Uj Ω ∂uh dΩ + ν ∂xj Ω ∂uh ∂uh dΩ = ∂xj ∂xj f uh dΩ. The Poincar´ inequality holds only if Ω is bounded in at least one e direction. g)L2 (Ω) ≤ f 1/2 L2 (Ω) g L2 (Ω) . ∂xi ∂xi 2 V that is. ∂xj . ≤C v Ω ∂v ∂v dΩ ∂xi ∂xi holds also when Stability Γ0 dΓ = 0. for each nonzero. various exotic norms can be deﬁned. Ω (4. . Thus 1/2 f L2 (Ω) = Ω f 2 dΩ = max g=0 Ω Ω f g dΩ 1/2 g 2 dΩ = max g=0 Ω f g dΩ L2 (Ω) g The rightmost expression is a “variational characterization” of the L2 (Ω) norm. the inequality v 2 dΩ ≤ C v 2 dΩ + Ω Ω 2 L2 (Ω) Ω Ω ∂v ∂v dΩ. Choosing g = f yields equality in above expression. By inserting derivatives in the nominator and/or denominator of the variational characterization. However. square-integrable g. The smallest possible value of C grows with the size of Ω. whereas for functions it reads 1/2 Ω 1/2 f g dΩ ≤ f 2 dΩ Ω Ω g 2 dΩ . .The Cauchy–Schwarz inequality for vectors a = (a1 .15) implies that. . (4.16) Ω | wi | dΩ This norm will be of crucial importance when we analyze the stability of FEM form the Navier–Stokes equations. that is. Theorem 2 says that the energy norm is stronger than then L 2 (Ω) norm. The Cauchy–Schwarz inequality (4. Choosing vh = uh in equation (4.17) Note that uh Uj Ω ∂uh dΩ = uT Cu = 0. f 2 dΩ Ω ≥ Ω Ω f g dΩ 1/2 g 2 dΩ . (4. the inequality does not hold if Γ0 dΓ = 0. an ) is |a · b| ≤ |a||b| = |a · a|1/2 |b · b|1/2 .15) which can be denoted (f. diﬀerent from the L 2 (Ω) norm and of importance for the Navier–Stokes equations is p p exotic = max wi Ω ∂wi ∂xi 2 dΩ 1/2 . Also. There is a C > 0 such that e v 2 dΩ ≤ C v for each v ∈ V . Remark 1. . A norm. The Poincar´ inequality relates the energy and L2 (Ω) norms: e Theorem 2 (Poincar´ inequality).

18) that is.6) and subtracting equation (4. ﬁgure 4. This is a “true” orthogonality condition when Uj ≡ 0: ν Ω ∂vh ∂ (u − uh ) dΩ = 0 ∂xi ∂xi ∀vh ∈ Vh .m m+2 1 h u V Vh uh Figure 4. (4. Orthogonality property (4.7) implies that the FE approximation is optimal in the energy norm when Ui ≡ 0: u − uh V ≤ u − vh V ∀vh ∈ Vh . When Uj = 0. Cf. ν Ω ∂vh ∂ (u − uh ) dΩ + ∂xi ∂xi vh U j Ω ∂ (u − uh ) dΩ = 0 ∂xj ∀vh ∈ Vh . two orthogonal vectors: a · b = 0 . the error u − uh is orthogonal to each vh ∈ Vh with respect to the inner product (4.6) yields the “Galerkin orthogonality”.7: An orthogonal projection on a line. that is. (4.4). there is a C > 0 such that Cea’s lemma holds. (4. since C = −C T (using the notation of § 4. shows that uh V C f L2 (Ω) ν where the constant C does not depend on h. Note that the vector u − uh is the shortest of all vectors u − vh for vh ∈ Vh . u − uh ≤ u − vh ∀vh ∈ Vh . the FEM approximation is no longer optimal in this sense. we have shown that u h cannot blow up as the discretization is reﬁned.20) Matrix A is nonsymmetric in this case.7) from(4.7. Orthogonality relation (4. ≤ Error bounds Choosing v = vh ∈ Vh ⊂ V in variational problem (4.18) means that uh is an orthogonal projection of u on Vh . .1.18) (cf. That is. As illustrated in ﬁgure 4.19) The matrix A is symmetric in this case (since C = 0).14). Expression (4.17) thus becomes ν Ω ∂uh ∂uh dΩ = ∂xi ∂xi uh 2 V f uh dΩ Ω 1/2 1/2 Ω (Cauchy–Schwarz) ≤ (Poincar´) ≤ C e f 2 dΩ Ω 1/2 u2 dΩ h ∂uh ∂uh dΩ ∂xi ∂xi 1/2 f 2 dΩ Ω L2 (Ω) Ω ≤C f Dividing with uh V uh V . u − uh V ≤ C u − vh ν V ∀vh ∈ Vh . However. as the next example illustrates. Note that the approximation may be far from optimal when ν is small. the orthogonal projection on a subspace yields the vector that is closest to the given element in the norm derived from the inner product in which orthogonality is deﬁned. which is the same as saying that the numerical solution is stable.

whereas the dashed line shows the FE approximation for h = 1/200. as illustrated for a 1D case in ﬁgure 4. Example 1. for suﬃciently smooth v. An explanation of this phenomenon is that the directions of ﬂow is not accounted for in the standard Galerkin approximation. Each of the conditions below is suﬃcient for approximation property (4.21) The problem is advection dominated. 2 h Figure 4.1 Γ1 0 0. The solid line in ﬁgure 4. the viscosity coeﬃcient in equation (4. The interpolant is the function Πh v ∈ Vh that interpolates the function v ∈ V at the node points of the triangulation.3 Vh n 0. Both these methods modify the basic Galerkin idea to account for ﬂow-direction eﬀects.6 i 0. (4. A strategy to preserve mesh quality when reﬁning is to subdivide each triangle into four new triangles by joining the edge midpoints (ﬁgure 4.7 m + 2 0.4 Inﬂow V Solid wall 0. Upwinding is used to prevent oscillations for ﬁnite-diﬀerence methods.2 u Γ0 uh 0. v − Πh v L2 (Ω) ≤ Ch2 v − Πh v V ≤ Ch “second order error” “ﬁrst order error” (4.94 0. for instance streamline diﬀusion and discontinuous Galerkin.98 1 Ω x 2 M M Figure 4. Various methods exist to obtain similar eﬀects in FEM.8 shows the FE solution h = 1/1000.96 0.92 0. Approximation and mesh quality Estimates (4.9.5 j h 0.M 0 M 0.19) and (4. Galerkin approximations correspond to ﬁnite-diﬀerence central schemes. The result of example 1 is typical: under-resolved standard Galerkin methods may generate oscillations for advection-dominated ﬂows in areas of sharp gradients. Interpolants are used to study approximation properties.22) to hold.9 0. A boundary layer with sharp gradients forms close to x = 1.8 1 m 0.9: An interpolant Πh v interpolates v at the nodal points. A mesh quality assumption is needed for (4. The essential point is: beware of “thin” triangles. that is. Next step in assessing the error is an approximation problem: How well can functions in V h approximate those in V ? This step is independent of the equation in question.10). 1).22) The constant C contains integrals of the second derivatives of v.20) are independent of the choice of Vh ! We only used that Vh ⊂ V .21) is quite small: ν = 1/500. do not let the shape deteriorate as the mesh is reﬁned. 3 1 2 u V Vh u uh ˆ v03 P t = 1. m+2 v1.8: Under-resolved standard Galerkin ﬁnite-element approximations may yield oscillations for m advection-dominated ﬂows. Let us consider the above FEM applied to the 1D problem − 1 u +u =1 500 u(0) = u(1) = 0 in (0.22) to hold: . Approximation theory yields that.

the functions are deﬁned uniquely by specifying the function values at each node. as in ﬁgure 4. The convergence rate is of optimal order : it agrees with approximation property (4. y) = a + b x + c y + d xy Quadrilaterals on a “logically rectangular domain” (a distorted rectangle. An example of a ﬁnite-element space Vh on quadrilaterals is globally continuous functions varying linearly along the edges. 4. but not the reverse. 3D A quadrilateral is a “skewed” rectangle: four points connected by straight lines to form a closed geometric object. Nonoverlapping quadrilaterals can be used to “triangulate” a domain. the solution may still be bad for a particular. An alternative is to keep the mesh ﬁxed and increase the order of the polynomials at each triangle (“p-method”).1 2 i j Inﬂow Solid wall n Γ0 Γ1 Ω Figure 4. As in the triangular case. in this case. However. Accuracy of the FE solution Error bounds (4.19) and (4.10: A subdivision that preserves mesh quality. where it is needed (say. Improving Accuracy Accuracy is improved by reﬁning the mesh (“h-method”).20) together with approximation property (4. in areas of large gradients).11) yields a regular structure to the matrix A. piecewise quadratics yields a third-order-accurate solution (in the L2 (Ω) norm). This may .22) and is the best order that in general can be obtained for. recall example 1. This second-order convergence rate requires a mesh quality assumption and a suﬃciently smooth solution u (since the constant C contains second derivatives of u). The V quotient between the diameter of the largest Vh circle that can be inscribed and the diameter u of the triangle should not vanish as h → 0. uh Condition (ii) implies (i). too-crude mesh. as in ﬁgure 4. Automatic methods for adaptation are common. This may be done adaptively.22) implies u − uh Further analysis (not provided here) yields u − uh L2 (Ω) V M m m+2 1 h V Vh u uh ≤ Ch ≤ Ch2 . This space reduces to piecewise bilinear functions for rectangles with edges in the coordinate directions: vh (x. continuous. The functions are deﬁned by interpolation into the interior of each quadrilateral. M2 M m (i) (“Maximum angle condition”). to prevent the matrices to become too large. piecewise-linear functions. where the mesh is reﬁned in certain regions and the order of approximations is increased in other. The largest anm+2 gle of any triangle should not approach 180◦ as 1 h → 0.11. These strategies may be combined in the “h–p method”. For instance.6 Alternative Elements. h (ii) (“Chunkiness parameter condition”).1.

2 1 2 p1.11: A logically rectangular domain triangulated with quadrilaterals. it is harder to generate quadrilateral meshes automatically for complicated geometries.12). and • allow eﬃcient solution of the linear system (particularly for uniform. M2 M M m m m+2 m+2 1 1 h h V V Vh Vh u u uh uh Figure 4.1 p1. • give high accuracy (particularly if the mesh is aligned with the ﬂow). . Triangular and quadrilateral meshes generalize in 3D to tetrahedral and hexahedral meshes (ﬁgure 4.12: In 3D. triangular and quadrilateral elements generalize to tetrahedral (left) and hexahedral (right) meshes.1 Ω 2 M 0 0 M 1 1 m 2 2 m+2 i i j j 1 h Inﬂow Inﬂow V Solid wall Solid wall Vh n n Γ0 Γ0 u uh Γ1 Γ1 Ω Ω M 2 Figure 4. rectangular meshes). Advantages and limitations with tetrahedral and hexahedral meshes are similar to corresponding meshes in 2D. Compared to triangular mesh.

23b)! We now turn to the Navier–Stokes equations for the steady ﬂow of an incompressible ﬂuid.2 FEM for Navier–Stokes 1 h V Γ Vh 0 u uh uj 0 Γ0 1 2 i j Ω Inﬂow Solid wall Γ0 n Γ0 Γ1 Ω Figure 4. The mathematical problem in diﬀerential form is Example 2 (“the counting argument”). as the following example indicates. Naive approximations of the Navier–Stokes equations are bound to produce disappoing results.23b). There are I number of panels in each direction and therefore 2(I − 1)2 degrees of freedom (two times the number of black dots) for the velocity vector ﬁeld. This condition introduces new complications not present in the advection–diﬀusion problem of section 4. Γ1 = ∅—and the mesh of ﬁgure 4.23b). which is more than the degrees of freedom for the velocity! Thus u h = vh = 0 is the only FE function satisfying the incompressibility condition (4.23c) .1. The boundary condition on Γ1 = ∂Ω \ Γ0 is of no obvious physical signiﬁcance.23a) (4. The left side of equation (4. Assume that we use continuous. Γ1 (4.14. but is useful as an artiﬁcial outﬂow condition. The Navier–Stokes equations is a system of nonlinear advection–diﬀusion equations together with the incompressibility condition (4. on Γ0 in Ω. M2 Γ0 M m m+2 1 h V Γ0 Γ Vh 0 u uh Γ0 Γ1 Ω Γ0 ∂p 1 ∂ui + − ∂xj ∂xi Re 1 ∂ui nj + pni = 0 − Re ∂xj ∂ui =0 ∂xi ui = g i 2 ui = f i in Ω.23b) to be satisﬁed on each triangle yields 2I 2 equations. demanding equation (4. vh for the x.13. A simple example domain is depicted in ﬁgure 4. is here constant on each triangle. that is ∂uh ∂vh + ∂x ∂y yy ww xx dd uu vv ee ff ss tt rr qqq p iXpiip p p piip p piip p p piip p piip p p piip p piip p p p p piip p p p p p p p p XXXiXiXXXiXXXiXiXXXiXXXiXiXXXiXXXiXiXXiXiXXXiXiXiXiXiXiXXiXiX pXXXiXiXXXiXXXiXiXXXiXXXiXiXXXiXXXiXiXXiXiXXXiXiXiXiXiXiXXiXiX iXXipXpXpXXipXpXXipXpXpXXipXpXXipXpXpXXipXpXXipXpXpXipXpXpXXipXpXpXpXpXpXpXipXpXpXip XXXpXpXXXpXXXpXpXXXpXXXpXpXXXpXXXpXpXXpXpXXXpXpXpXpXpXpXXpXpX ipXXpiXiXiXXpiXiXXpiXiXiXXpiXiXXpiXiXiXXpiXiXXpiXiXiXpiXiXiXXpiXiXiXiXiXiXiXpiXiXiXip ip ip ip ip ip ip ip ip XXipXXXXipXXXipXXXXipXXXipXXXXipXXXipXXXipXXXXipXXXXXXXipXXXip pipXipipip p p ipipip p ipipip p p ipipip p ipipip p p ipipip p ipipip p p p p ipipip p p p p p p p p iXXXiXiXXXiXXXiXiXXXiXXXiXiXXXiXXXiXiXXiXiXXXiXiXiXiXiXiXXiXiX XXipXpXpXXipXpXXipXpXpXXipXpXXipXpXpXXipXpXXipXpXpXipXpXpXXipXpXpXpXpXpXpXipXpXpXip XXXiXiXXXiXXXiXiXXXiXXXiXiXXXiXXXiXiXXiXiXXXiXiXiXiXiXiXXiXiX ipXXipXpXpXXipXpXXipXpXpXXipXpXXipXpXpXXipXpXXipXpXpXipXpXpXXipXpXpXpXpXpXpXipXpXpXip XXXiXiXXXiXXXiXiXXXiXXXiXiXXXiXXXiXiXXiXiXXXiXiXiXiXiXiXXiXiX ipXXipXpXpXXipXXXipXXXXipXXXipXXXXipXXXipXXXipXXXXipXXXXXXXipXXXip ip XXipXiXiXXipXiXXipXiXiXXipXiXXipXiXiXXipXiXXipXiXiXipXiXiXXipXiXiXiXiXiXiXipXiXiXip pXXiXiXXXiXXXiXiXXXiXXXiXiXXXiXXXiXiXXiXiXXXiXiXiXiXiXiXXiXiX iXXpXpXXXpXXXpXpXXXpXXXpXpXXXpXXXpXpXXpXpXXXpXpXpXpXpXpXXpXpX ipip p ipip p p ipipip p ipipip p p ipipip p ipipip p p p p ipipip p p p p p p p p pipXXXpXpXXXpXXXpXpXXXpXXXpXpXXXpXXXpXpXXpXpXXXpXpXpXpXpXpXXpXpX iXXipXiXiXXipXiXXipXiXiXXipXiXXipXiXiXXipXiXXipXiXiXipXiXiXXipXiXiXiXiXiXiXipXiXiXip pXXXpXpXXXpXXXpXpXXXpXXXpXpXXXpXXXpXpXXpXpXXXpXpXpXpXpXpXXpXpX iXXXiXiXXXiXXXiXiXXXiXXXiXiXXXiXXXiXiXXiXiXXXiXiXiXiXiXiXXiXiX XXipipXXXXipipXXXipipXXXXipipXXXipipXXXXipipXXXipipXXXipipXXXXipipXXXXXXXipipXXXipip ipip ipip ipip ip ip ip ip ip h gh h h gh h gh h h gh h gh h h gh h gh h h h h gh h h h h h h h h gXXXgXgXXXgXXXgXgXXXgXXXgXgXXXgXXXgXgXXgXgXXXgXgXgXgXgXgXXgXgX XXghXXXXghXXXghXXXXghXXXghXXXXghXXXghXXXghXXXXghXXXXXXXghXXXgh 4.14: The black nodes are the degrees of freedom for the velocity components. Thus. piecewise-linear approximations u h . which are fewer than the number of triangles. and consider a situation with ﬂow in a bounded domain Ω.13: A simple example of a domain.23d) (4. on Γ1 .Figure 4.and y-components of the velocity ﬁeld.23b) (4. Consider a case with only Dirichlet boundary conditions—that is.

but only ∼ 2I 2 vertices).3) on the left sides yields ∂zi ∂ui ∂ui dΓ = dΩ + zi 2 ui dΩ (4.26) n j zi ∂xj ∂xj ∂xj Γ Ω Ω and nj zj p dΓ = Γ Ω zi ∂p dΩ + ∂xi p Ω ∂zi dΩ. Now. whereas (iii) leads to “non-conforming” approximations related to ﬁnite-diﬀerence and ﬁnite-volume approximation with “staggered mesh”. To avoid this last complication and simplify the exposition. in the last equality. we will only consider homogeneous boundary conditions. (i) Use more degrees of freedom for the velocity—for instance continuous. . which.24) and (4. Also note that ui = gi on Γ0 whereas zi vanishes on Γ0 .23b) with a smooth function q and integrating yields q Ω ∂ui dΩ = 0 ∂xi (4. for instance. The product rule of diﬀerentiation yields ∂ui ∂zi ∂ui ∂ zi = + z i 2 ui (4. piecewise quadratics on each triangle—to balance the large number of equations associated with the incompressibility condition. multiplying equation (4.26) and (4.Following strategies can be used to tackle the problem of example 2.23d). we ﬁnd that zi fi dΩ = Ω Ω zi u j ∂ui dΩ + ∂xj zi Ω ∂p 1 dΩ − ∂xi Re zi Ω 2 ui dΩ n j zi ∂ui dΓ − ∂xj p Ω = Ω 1 ∂ui dΩ + zi u j ∂xj Re zi u j ∂ui 1 dΩ + ∂xj Re Ω ∂zi ∂ui 1 dΩ − ∂xj ∂xj Re ∂zi ∂ui dΩ − ∂xj ∂xj Γ ∂zi dΩ + ∂xi nj zj p dΓ (4. and utilizing (4. Multiply both sides of equation (4. we conclude that a classical solution (a twice continuously diﬀerentiable solution) to the Navier– Stokes equations satisﬁes the variational forms (4.28) and (4. ∂xi where. This also gives more degrees of freedom for the velocities. integrating over Ω.2. Similarly as for the advection–diﬀusion problem.27) which are the integration-by-parts formulas needed to derive our variational form.23a) with zi .28) Γ = Ω p Ω Ω ∂zi dΩ. The pressure space is H = L2 (Ω) = q| q 2 dΩ < +∞ Ω . (iii) Associate the velocity with the midpoints of edges instead of the vertices.24) ∂xj ∂xj ∂xj ∂xj and ∂ ∂p ∂zj (zj p) = zi +p .25) and applying the the divergence theorem (4. could be gravity or magnetic forcing for a electrically conducting ﬂuid. we need new integration-by-part formulas. the boundary integrals vanish on Γ0 since zi = 0 on Γ0 and on Γ1 due to boundary condition (4. without explicit reference to the diﬀerential equations. Strategies (i) and (ii) are the “standard” methods. A diﬀerence with advection–diﬀusion problem is that we here need diﬀerent function spaces for the velocity components and the pressure. gi = 0. 4. (There are ∼ 3I 2 edges.29) for all smooth functions z i and q such that zi vanishes on Γ0 . (ii) Accept that the discrete velocities only is approximately divergence-free. Thus.27). Moreover. ∂xj ∂xi ∂xj (4. ∂xi (4.29) Thus. let zi be a smooth vector-valued function on Ω that vanishes on Γ0 .25) Integrating expressions (4.1 A variational form of the Navier–Stokes equations To derive the variational form. the equations will be “driven” only by the right-hand side fi . the variational form will be used to deﬁne weak solutions.

and important issues are still unresolved. Note that the steady problem may have several solutions. Ω For the analysis in section 4. so that weak solutions are classical solutions if data is smooth enough. – smooth data yield smooth solutions. there is a $ 1 000 000 prize for the person that can develop a theory that can satisfyingly explain their behavior in three space dimensions! A short summary of what is known is the following.and the space of velocity components is V = u| ∂u ∂u dΩ < +∞ and u = 0 on Γ0 ∂xi ∂xi . As for the advection–diﬀusion problem. but the more complicated condition (4. even in 2D. or if singularities can evolve from smooth data. • in three space dimensions: – a weak solution exists for any square-integrable f and for any Re. – It is not known whether smooth data always yields a smooth solution. T ∗ ).2. The energy norm for elements of V is denoted u whereas the L2 (Ω) norm is 1/2 V = Ω ∂ui ∂ui dΩ ∂xj ∂xj 1/2 . Note that the natural boundary condition is diﬀerent than for the advection–diﬀusion problem. The mathematical properties of the Navier–Stokes equations are complicated.23d). T ). – It is not known whether the weak solution always is unique (besides during some initial interval (0. The essential boundary conditions are here that ui = 0 on Γ0 and is explicitly assigned through the deﬁnition of V . the pressure is only possible to specify uniquely up to an additive constant. Remark 2. Note that the natural boundary condition here not just is a Neumann condition. The variational problem deﬁning weak solutions to equation (4. Saying that u ∈ V is equivalent to saying that that each velocity component u i ∈ V . u L2 (Ω) = Ω ui ui dΩ . . where the size of T ∗ is unknown). We assume that both boundary conditions really are active. implicitly speciﬁed through the variational form.4. In the case of only Dirichlet boundary condition. u2 ). that is. For the unsteady problem on a given. in 2D as well as 3D.23) is Find ui ∈ V and p ∈ H such that ∂ui ∂zi ∂ui 1 zi u j dΩ + dΩ − ∂xj Re ∂xj ∂xj Ω Ω p Ω ∂zi dΩ = ∂xi zi fi dΩ Ω ∀zi ∈ V Ω ∂ui dΩ = 0 q ∂xi ∀q ∈ H This is a variational problem of mixed type: two diﬀerent spaces are involved. The steady problem has at least one weak solution. the boundary conditions are incorporated in the variational formulation. In fact. ﬁnite time interval (0. The natural boundary conditions are the boundary conditions on Γ1 . it will be convenient to work with the space V of the velocity vector u = (u1 . that Γ0 constitute the whole boundary. and • in two space dimensions: – a unique weak solution exists for any square-integrable f and for any Re.

. Then. . v (N u) T .30a) (4.31) and (4. . y) . ∀zh ∈ Vh .32) Now insert expansions (4.2. .33) − Nu p(n) n=1 Ω (n) ∂ϕ(m) (n) ψ dΩ = ∂y Ω Nu ϕ(m) f2 dΩ ∂ϕ(n) dΩ = 0 ∂y m = 1. (4. We postpone for a while the question on how to choose the subspaces V h and Hh . In 2D and in components. . Nu ∂ϕ(m) ∂ϕ(n) ∂ϕ(m) ∂ϕ(n) + ∂x ∂x ∂y ∂y Ω v (n) n=1 Ω ϕ(m) uh Np ∂ϕ ∂ϕ + vh ∂x ∂y (n) (n) dΩ + 1 u (n) v Re n=1 N dΩ (4. . Nu u(n) n=1 Ω ψ (m) ∂ϕ dΩ + v (n) ∂x n=1 ψ (m) Ω m = 1. and replace V with Vh and H with Hh in the variational form. . . y) . .30b). vh ∈ Vh and ph ∈ Hh such that zh u h Ω ∂uh ∂uh + vh ∂x ∂y dΩ + 1 Re Ω ∂zh ∂uh dΩ − ∂xj ∂xj ph Ω ∂zh dΩ = ∂x Ω zh f1 dΩ zh f2 dΩ Ω ∀zh ∈ Vh .34) . vh ) + Re K G(y) v = b(y) . choose z h = ϕ(m) for m = 1. .33) in the matrix–vector form (x) 1 0 G(x) C(uh . . . . equations (4. . (4. y) = n=1 p(n) ψ (n) (x. Np Placing the coeﬃcients into column matrices. y) . . we can write the system (4. we obtain Find uh .32) into equations (4. u = u(1) . Hh ⊂ H of piecewise polynomials. in (4. vh in basis for Vh yields Nu Nu uh (x. . choose qh = ψ (m) for m = 1. (4. y) = n=1 u(n) ϕ(n) (x. . . y) = n=1 v (n) ϕ(n) (x. Np . vh (x.30b) (4. . v = v (1) . . . In (4.30a) and (4. Nu .30).2 Finite-element approximations To obtain a FEM.31) Likewise. . p(Np ) T T . . expanding uh . 4. .30c). u(N u) p = p(1) . we choose subspaces Vh ⊂ V . . vh ) + Re K b u 1 0 C(uh .30c) zh Ω ∂vh ∂vh uh + vh ∂x ∂y ∂vh ∂uh + ∂x ∂y 1 dΩ + Re Ω ∂zh ∂vh dΩ − ∂xj ∂xj ∂zh ph dΩ = ∂y Ω qh Ω dΩ = 0. . we obtain Np ph (x. . . . . .30) become Nu u(n) n=1 Ω ϕ(m) uh Np ∂ϕ(n) ∂ϕ(n) + vh ∂x ∂y ∂ϕ ψ (n) dΩ = ∂x (m) dΩ + 1 u (n) u Re n=1 N ∂ϕ(m) ∂ϕ(n) ∂ϕ(m) ∂ϕ(n) + ∂x ∂x ∂y ∂y Ω dΩ − Nu p(n) n=1 Ω ϕ(m) f1 dΩ Ω m = 1.3 The algebraic problem in 2D Expanding ph in basis for Hh .4. . p 0 −G(x)T −G(y)T 0 (4.2. .

the Stokes equations are much easier to analyze than the full Navier–Stokes equations. Due to the lack of nonlinear advection. (4. it is guaranteed the the last block row of the matrix is the negative transpose of the last block column.37b) uh V = Ω ∂uh ∂uh i i dΩ ∂xj ∂xj ≤C f L2 (Ω) . Indeed. As for the Navier–Stokes equations. vh ) = Ω ϕ(m) uh ∂ϕ(n) ∂ϕ(n) + vh ∂x ∂y (y) bm = Ω dΩ. e We thus conclude from expression (4.36b) The Stokes equations is a limit case of the Navier–Stokes equations for very low Reynolds numbers and assume that inertia can be neglected compared to viscous forces. ∂ϕ(m) dΩ. We obtained the same structure of the nonlinear equation for the ﬁnite-volume discretization. ϕ(m) f2 dΩ. Cmn (uh .37a) (4.where the matrix and vector components are Kmn = Ω (x) Gmn = − Ω ∂ϕ(m) ∂ϕ(n) ∂ϕ(m) ∂ϕ(n) + ∂x ∂x ∂y ∂y ∂ϕ(m) dΩ. but also ﬂow of common ﬂuids (like air and water) at the micro scale. (4. lubrication. creeping ﬂow. h Choosing zi = uh in equation (4. This implies that the approximations will not blow up when the mesh is reﬁned. in makes sense to require that the velocity and pressure approximations satisfy the same bounds. Mathematically.36a) Ω ∂uh qh i dΩ = 0 ∂xi ∀ q h ∈ Hh . That is. The counting argument of example 2 indicates that not any choice is good. these are linear equations.35) Matrix K represents an approximation of the discrete negative Laplacian − 2 . C(uh . Thus. there exists a unique weak solutions for each square-integrable fi both in two and three space dimension. the structure of the Navier–Stokes equations are directly reﬂected in the nonlinear system (4. We will limit the discussion to the Stokes equations.36b). respectively. uh 2 V = Ω uh fi dΩ ≤ u i L2 (Ω) f L2 (Ω) ≤C u V f L2 (Ω) (4. (4. we obtain i ∂uh ∂uh i i dΩ − ∂xj ∂xj ph Ω Ω ∂uh i dΩ = ∂xi Ω ∂uh ∂uh i i dΩ = ∂xj ∂xj uh fi dΩ. Thus.4 Stability So far.36a). that is. i Ω where the ﬁrst equality follows from equation (4. except that here. 1/2 1/2 ph L2 (Ω) = Ω p2 dΩ h ≤C 1/2 fi fi dΩ Ω =C f L2 (Ω) .34). ψ (n) ∂x (y) Gmn = − Ω dΩ.38) where the Cauchy–Schwarz inequality yields the ﬁrst inequality. the approximating spaces Vh and Hh have not been speciﬁed. we require that Vh ⊂ V and Hh ⊂ H.39) . The Stokes equations may be used to model very viscous ﬂow. G(x) and G(y) the discrete gradient . (4.2. The reason to start the analysis on the Stokes equation is that a stable Stokes discretization is necessary to obtain a stable Navier–Stokes discretization. 4. Recall that each function in V and H is required to be bounded in the energy and the L2 (Ω) norm.38) that uh V ≤C f L2 (Ω) . and the Poincar´ inequality the second. (4. ψ (n) ∂y b(x) = m Ω ϕ(m) f1 dΩ. and G(x)T and G(y)T the discrete divergence. whose ﬁniteelement formulation reads Find uh ∈ Vh and ph ∈ Hh such that i Ω h ∂zi ∂uh i dΩ − ∂xj ∂xj h ∂zi dΩ = ∂xi ph Ω h zi fi dΩ Ω h ∀ zi ∈ Vh . vh ) the discrete ∂ ∂ advection operator u ∂x + v ∂y . where the constant C should not depend on h.

That is. for each qh ∈ Hh . Condition (4. then it is immediate that ph L2 (Ω) ≤ C f L2 (Ω) for some C > 0. we will of course obtain inaccurate velocity approximations. an eﬀect that usually manifests itself as wild pressure oscillations. ∂z h . (4. The derivation above proves thus that the LBB-condition is suﬃcient for the pressure to be stably determined by data: Theorem 3.39)) ≤ C z h V f V + z h V f L2 (Ω) L2 (Ω) + z h V f L2 (Ω) . the pressure (before discretization) is stably determined by data.40) V Dividing expression (4. the LBB condition. any Galerkin approximation of the velocity in the Stokes equations is stable! For bad choices of subspaces. The LBB condition thus means that a zero discrete pressure gradient implies a zero value of the pressure. Conversely.3). it means that the discrete pressure gradient may be small even when the pressure is not small: The discrete gradient cannot “feel” that the pressure is large.37a).43) is satisﬁed for Vh = V and Hh = H. The LBB-condition is a condition on the pairing of spaces Vh and Hh that requires the exotic norm to be stronger than the L2 (Ω)-norm for each function qh ∈ Hh .5 The LBB condition h ∂zi dΩ = ∂xi h ∂zi ∂uh i dΩ − ∂xj ∂xj h h ∂zi ∂zi dΩ ∂xj ∂xj V h Rearranging equation (4. Not any combination H h and V h yield stable pressure approximations in the sense of (4.43) then there is a C > 0 such that 1/2 Ω 1/2 p2 dΩ h ≤C Ω |f |2 dΩ holds. 1/2 α Ω 2 qh dΩ ≤ max 1 zh V Ω 0=zi ∈Vh qh h ∂zi dΩ. Vh . since their energy norm is bounded by the given data fi .2.16). We will derive a compatibility condition between the velocity and pressure spaces. but they will not blow up as the mesh is reﬁned. ∂xi (4. Thus. Regarding the pressure. the situation is more complicated. The LBB condition (4.37a). 4. such that. assume that there is an α > 0 independent of h. ph Ω Ω Ω 1/2 h zi fi dΩ (Cauchy–Schwarz) ≤ Ω Ω L2 (Ω) ∂uh ∂uh i i dΩ ∂xj ∂xj f L2 (Ω) 1/2 1/2 1/2 + Ω h h zi zi dΩ fi fi dΩ Ω = zh uh V + zh (Poincar´) ≤ z h V uh e (by (4. independently of h.2. if the LBB condition is not satisﬁed. from which it follows that there is a C > 0 such that 0=zi ∈V max 1 zh V Ω ph h ∂zi dΩ ≤ C f ∂xi L2 (Ω) .36a) implies that for each zi ∈ Vh .that is.41) We have thus found a bound on the discrete pressure in the “exotic” norm (4. However. That is.40) with z h yields that ph Ω h ∂zi dΩ ≤ (C + 1) f ∂xi L2 (Ω) 1 zh V h for each zi ∈ Vh .42) says that our exotic norm is stronger than the L2 (Ω)-norm for ph . (4.42) for some α > 0 independent of h. The expression Ω qh ∂xii dΩ represents the discrete gradient of qh (see § 4. we were aiming for the bound (4. if it happens that α ph L2 (Ω) ≤ max 1 zh V Ω 0=zi ∈Vh ph h ∂zi dΩ ∂xi (4. necessary to yield stable pressure approximations. Assume that the LBB-condition holds for the spaces Hh .

This pair does not satisfy the LBB condition and yields an unstable discretization. A Galerkin approximation is thus globally mass conservative if 1 ∈ Hh . whose h nodal values are equal to the elements of pc is called a checkerboard mode. there is at least one nonzero Np -vector pc such that Gpc = 0. to the ith row of vector Gpc (notice from the expressions in ﬁgure 4. that is.15: A checkerboard mode for equal-order interpolation on a regular triangular mesh. consider the element contribution. Accuracy Now we consider a regular triangulation of a 2D domain Ω. if ∂uh i dΩ = 0. so Γ ρ ni ui dΓ = 0. A velocity approximation uh is is globally mass conservative if i Γ ni uh dΓ = 0. we use continuous functions that are linear on each triangle both for the functions in H h and Vh .Ω M2 M m m+2 1 −1 0 1 0 1 −1 0 h V Vh −1 u uh Γ0 Γ1 1 Ω 0 1 −1 −1 0 1 Figure 4. Thus. the system matrix (4. There is no net mass-ﬂux through the boundary for a constant-density ﬂow.5.1). Recall from equation (4. Figure 4. that is.1.15 depicts the nodal values of a checkerboard mode for equal-order interpolation. it will not be locally mass conservative if the pressure approximations are continuous and piecewise linear.2.30c) that qh Ω ∂uh i dΩ = 0 ∂xi ∀qh ∈ Hh . if constant functions can be represented exactly in Hh . the FE approximation will be locally mass conservative. ∂xi K for each element K in the mesh.6 Mass conservation The integral Γ ρ ni ui dΓ yields the total ﬂux of mass(in kg/s) through the boundary.34) is singular : the discrete gradient matrix G has linearly dependent columns.1. Not all ﬁnite-element approximations are locally mass conservative. In particular.5 that ∂φi is constant on each triangle): ∂x − Tn pc h ∂φi ∂φi dΩ = − ∂x ∂x Tn Tn pc dΩ ≡ 0. To see this. Corresponding function pc ∈ Hh . All reasonable FE approximation satisfy this. Ω ∂xi by the divergence theorem (4.7 Choice of ﬁnite elements. equivalently. A ﬁnite-element approximation is locally mass conservative if ∂uh i dΩ = 0. i or. That is.3). However. 4.2. associated with a triangle Tn . h . “Equal-order interpolation” Let us start by trying the approximation space we used for the advection–diﬀusion problem (section 4. 4. a triangulation satisfying a mesh quality condition of the sort discussed in section 4. If the pressure approximations consist of piecewise constants on each element.

since v is not a function for discontinuous v. so that uh ∈ Vh can be expressed as uh (xi ) = m=1 ui φ(m) (xi ). whereas • uh .. piecewise linear on each element. Let the functions in H h be piecewise constant on each element. on each triangle. Indeed. • ph ∈ Hh is linear. This (and other) checkerboard mode will pollute the solution of the h Stokes as well as Navier–Stokes equations. (bi)linear velocities Now we consider a triangulation either of triangles or rectangles. which immediately implies that the LBB cannot h hold for a space Hh that includes pc . The pressure approximations need not to be continuous in order for H h ⊂ H.16. we have uh = a + b x + c y + d xy + e x2 + f y 2 The six coeﬃcients a–f are uniquely determined by uh ’s values at the triangle vertices and midpoints (ﬁgure 4. However. Checkerboard modes similar as in ﬁgure 4. There are two kinds of basis functions here. as can be seen from ﬁgure 4. This choice leads to a locally mass-conserving approximation. consising of piecewise polynomials of degree k for pressure and k + 1 for velocity (k ≥ 1) on each triangle. piecewise quadratic functions interpolates the nodal values of Nu ﬁgure 4. Then.h V Vh u uh Γ0 Γ1 Ω Figure 4. The Taylor–Hood pair Here. since the mean value of the function pc vanishes on each triangle. Thus. but not locally mass conservative. on each triangle.16: The evaluation points at each triangle for the continuous. whereas for quadrilateral elements. continuous functions that are piecewise bilinear on each rectangle comprise Vh .16). However. the constant pressure—bilinear velocity space on rectangles is in widespread engineering use! “Fixes” have been developed to take care of pressure oscillations. if Vh ⊂ V . The Taylor–Hood pair can be generalized to higher order. The LBB condition is satisﬁed for the Taylor–Hood pair. the functions in Hh and Vh are globally continuous and.18. Moreover. In case of triangular elements.15. for triangular elements. Remark 3.17 and 4. For instance. illustrated in ﬁgures 4. the space Vh will be continuous functions. the pressure and velocity approximations will be continuous functions. . piecewise quadratic velocity approximations of the Taylor–Hood pair. and the following error estimates holds for a smooth solution to the Navier–Stokes equation: u − uh p − ph L2 (Ω) L2 (Ω) ≤ Ch2 ≤ Ch3 The Taylor–Hood pair is globally. Nu A basis φ(m) (xi ) m=1 of continuous. the velocity approximations cannot be discontinuous. the LBB conditions is not satisﬁed for this unstable discretization.43) vanishes for qh = pc . Still. the counting argument of example 2 applies here and indicates problems with this choice of elements. vh ∈ Vh is quadratic. The right h side of the LBB condition (4. the checkerboard mode can be ﬁltered out from the discrete pressures.15 exist. Constant pressure. common iterative methods to solve the resulting discrete equations (such as the projection schemes) implies a stabilization of the discretization.

However. we here consider diﬀerent Ω meshes for Ω velocity and pressure.1 p1.j Ω vi.1 v1.21.to approximate the shape of the boundary.j+1/2 O(h2) m+2 vi.j+1/2 m+2 vi.j m corresponds to an edge-midpoint node pi. + 2 the following error estimates holds for m m and m+2 1 1 h h V V Vh Vh u u uh uh Γ0 Γ0 Γ1 Γ1 Ω Ω Figure 4. j − 1 Solid wall Γ0 j Γ1 n i Γ0 Ω Γ+ 1 i1 u Ω i. This is ﬁne 0 1 for piecewise-linear approximations since the error anyway is O(h2 ).j M 2 4. jVh 1 − Inﬂowu i + 1. For instance. a straight-forward triangulation of a domain with a curved 0 boundary will lead to a boundary-condition error at the mid nodes of O(h2 ) (ﬁgure 4.1 C 2 u 3 . but each pressure triangleM subdivided into four velocity triangles is M2 as in ﬁgure 4. The most common strategy for curved elements is known i j j as Isoparametric Elements: piecewise polynomials of the same degree as used in the FE approximation are Inﬂow used toInﬂow approximate the shape of the boundary (ﬁgure 4. as in the Taylor–Hood pair.j h 1 i− 2 V i Vh 1 i+ 2 u i+1 uh 3 i+ Γ0 2 j+1 Γ1 2 j jΩ 1 − 2 .j h 1 i− 2 V i Vh 1 i+ 2 u i+1 uh 3 i+ Γ0 2 j+1 Γ1 2 j jΩ 1 − m+2 11d 2h a ia V i − 1.j− Figure 4. B B ary condition.j−1/2 1 ui−1/2. ity and pressure.1 higher-order approximations.19: 1Approximating a curved boundary with of the sameA order as used for the FE approximations A a triangle sides leads to an O(h2 ) error of the bound. j j 1 u j−1 i. piecewise-linear functions for both veloc.21: Subdividing a pressure triangle into four four velocity rectangles yields a stable approximation velocity triangles yields a stable approximation with with continuous piecewise-bilinear functions for both continuous.17: The basis function φ(m) (xi ) when v M 2 i.j m vi.j M p m i.19). j − 1 Solid wallh u j n0 Γ ΓΓ i 0 1 i+ Γ1 1 Ω ui. Both the velocity and pressure approximations consist of continuous. 2 Using curved elements restores convergence rate. C u 3 . M m This is a stable approximation satisfying the LBB condition.j−1/2 1 ui−1/2.j Figure 4.20).22: Subdividing a pressure rectangle into Figure 4. 3 Inaccuracies in approximations of the domain boundary may reduce the convergence rate of p1.20: 1Isoparametric elements use polynomials j− Figure 4. j j 1 − i. 3 2 2 v1. this inaccuracy at the 1 2 boundary results in a half-order reduction of the convergence rateiof piecewise quadratic approximations. Inﬂow uh i + 1.j M m corresponds to a corner node pi.velocity and pressure.j vi. M2 piecewise-linear functions on triangles.18: The basis function φ(m) (xi ) when Figure pi. 2 Remark 4. 2 m+2 0 d 1 1d h a V a i Vh − i − 1. wall Solid Solid wall n n Γ0 A stable approximation with piecewise-linear velocities and pressures Γ0 Γ1 Instead of Γ1 using diﬀerent approximation orders.

Basis function φj is linear on each triangle and continuous along edge j (but discontinuous along four edges!). Thus. Tn zh u h n=1T n M z h vh ph ∂zh dΩ = ∂y Ω zh f2 dΩ Tn Tn Tn qh n=1T n ∂uh ∂vh + ∂x ∂y dΩ = 0 where M is the number of triangles. There is also a version of this discretization for rectangular meshes: Each pressure rectangle is then subdivided into four velocity rectangles. However. y). N is the total number of edges. consider basis functions. associated with the edges. y) = j=1 vj φj (x.22. Calculating the matrix elements (4. Figure 4. we deﬁne the nonconforming FE approximation: Find uh . A nonconforming method The discussion after example 2 suggested a strategy to address the problem with the overdetermined incompressibility condition. To accomplish this.1 h V Vh u uh Γ0 Γ1 Ω j 1 h V Vh u uh Γ0 Γ1 Ω Figure 4. y i are the coordinates of the mid point of edge i. constant) function. namely to use nodal points on the edge mid points for the velocity. vh (x. Note that uh will be continuous only at edge mid points in general. If x i . for i = j.24: Approximations spanned by the basis functions (4. in general. y). y) = j=1 uj φj (x. but are one order less accurate. Violating a property like Vh ⊂ V is called a variational crime. vh ∈ Vh and ph ∈ Hh such that M zh u h n=1T n M ∂uh dΩ + ∂x n=1 ∂vh dΩ + ∂x n=1 M M z h vh Tn 1 ∂uh dΩ + ∂y Re n=1 ∂vh 1 dΩ + ∂y Re n=1 ∀ q h ∈ Hh . excluding those on Γ1 . Here. Letting Hh be piecewise constant on each triangle and Vh deﬁned as above.23: The basis functions associated with the velocity approximation for a nonconforming but stable discretization. ∀ zh ∈ Vh . uh restricted on each element is a square-integrable function. of the type depicted in ﬁgure 4. Now deﬁne the space of the velocity components Vh from all possible expansions N N uh (x. y i ) = 1 0 for i = j. a smooth solution to the Navier–Stokes equation: u − uh p − ph L2 (Ω) L2 (Ω) ≤ Ch2 ≤ Ch These approximations use the same number of unknowns and the same nodal locations as for the Taylor– Hood pair. we have φj (xi . .35) is however somewhat easier than for the Taylor–Hood pair. Both the velocity and pressure approximations consist of continuous and piecewise-bilinear functions.23. Vh ⊂ V since uh is not a square-integrable (in fact. as in ﬁgure 4.23) yields piecewiselinear functions that are continuous only along at the edge midpoints. M M Tn ∂zh ∂uh dΩ − ∂xj ∂xj n=1 ∂zh ∂vh dΩ − ∂xj ∂xj n=1 M M ph ∂zh dΩ = ∂x Ω zh f1 dΩ ∀ zh ∈ Vh .

7–4. quite similar to equation (4.The fact that Vh ⊂ V complicates the analysis of this discretization. for instance continuous.30). This leads to a substantial simpliﬁcation when designing an algorithm to solve the discretized equations. . uh = (uh .44) yields that · uh = 0.46b) for some small > 0. (4. . and equations (4. Other nice properties with this method is that it is locally mass conservative. .46a) Ω ∂qh ∂ph dΩ + ∂xi ∂xi Ω ∂uh qh i dΩ = 0 ∂xi ∀ qh ∈ Hh . for instance. (4. does not increase the accuracy.46) is stable. non-linear advection– diﬀusion problem. (4. The main advantage with stabilization methods is the simpliﬁed implementation. and choosing (z h . Unfortunately. the approximation is stable. although not as accurate as the Taylor–Hood pair. vh ) = j=1 uj η j . We will present a very simple idea of this sort for the Stokes equations.2. but the method introduces an extra parameter that may need judicious tuning. By a clever combination of basis functions of the type illustrated (j) (j) in ﬁgure 4. More advanced stabilization methods avoids this inconsistency by using a term that vanishes as h → 0.36) with Find uh ∈ Vh and ph ∈ Hh such that i Ω h ∂zi ∂uh i dΩ − ∂xj ∂xj ph Ω h ∂zi dΩ = ∂xi h zi fi dΩ Ω h ∀ zi ∈ Vh . we have used the same basis for the uh and vh components. These manipulations complicate the implementation. choose zi = uh .44) into equations (4.46b) formally inconsistent by O( ) with corresponding variational form.30c). the construction of the basis functions η j is not straightforward.30) reduce to M M dNu uj j=1 Ω η i · (uh · ) η j dΩ + ν j=1 uj Ω ηi · η j dΩ + j=1 uj Ω η i · (uh · ) η j dΩ = Ω η i · f dΩ. Note also that the increased number of pressure unknowns.23. Stabilization of unstable discretizations Stability of the choices of elements presented in sections 4. An alternative is to use the same approximations for velocity and pressure (equal-order interpolation) combined with a stabilization method. wh ) for the velocity having the property · ηj = Expanding the velocity in this basis. including the present. wh ) = η i . Nevertheless. But maybe the most interesting with this discretization is that it allows for a construction of a divergence free velocity basis. piecewise-linear functions on a triangular mesh. (4.7 relied on manipulations of the local polynomial order. In all types of discretizations discussed so far. one of the great challenges for any solver of the Navier–Stokes equations is how to treat the incompressibility condition (4. or the conformity of the method. Equation (4.2. h To see that approximation (4.. First choose the same type of approximations for the pressure and the velocity components. in fact. and replace equations (4. M . Inserting expansion (4.2.7. The Navier–Stokes equations then reduces to a vector-valued. the continuity equation and the pressure variable vanishes. qh = ph and add equations (4. it is possible to construct a new vector-valued basis functions η j = (zh . as compared with the discretization in section 4.45) for i = 1. The added resolution introduced is ﬁltered away by the introduction of the stability term.10). This circumvents the LBB condition.46a) i . . the meshes. and quite easy to implement. M ∂wh ∂zh + ∂x ∂y (j) (j) = 0.

∂uh ∂uh i i dΩ + ∂xi ∂xi ∂ph ∂ph dΩ = ∂xi ∂xi uh fi dΩ i 1/2 1/2 Ω Ω Ω (Cauchy–Schwarz) ≤ (Poincar´) ≤ C e ≤C Ω uh uh dΩ i i ∂uh i Ω fi fi dΩ Ω 1/2 1/2 ∂uh i ∂xi ∂xi dΩ Ω fi fi dΩ ∂uh ∂uh i i dΩ ∂xi ∂xi 1/2 1/2 Ω ∂ph ∂ph dΩ + ∂xi ∂xi fi fi dΩ Ω . both the velocity and pressure approximations cannot blow up as h → 0.46b). even though the discretization is unstable for = 0.and (4. . Ω Dividing and squaring yields the stability estimate ∂uh ∂uh i i dΩ + ∂xi ∂xi ∂ph ∂ph dΩ ≤ C ∂xi ∂xi fi fi dΩ. Ω Ω Ω Since C > 0 does not depend on h.

.

j+1 + Φi.1.j+1 − 2Φi.u h Γ0 k+1 k+1 2 Φk Φk i+1. the error is reduced by O h2 each iteration.j = V 2 (1 + β 2 ) Vh Gauss-Seidel uses this expression to update Φi.j + Φi−1.j + Φi.j + β i.j (using the already updated values calculated in same u iteration).j + Φi−1. We can solve for Φi.j−1 Γ1 Φk+1 = i.j+1 + Φi. 1 Φi+1. Requiring an error reduction to O h2 the number of iterations must satisfy ρm = O h2 which implies that m=O ln(h) ln(ρ) =O ln(h) ln(1 − O(h2 )) = O −h2 ln(h) To proceed we note that we have N = n2 unknown interior nodes.j .j + β 2 (Φi. see Figure A. Thus 95 ite : old values ra tio : new values j n di re : to be updated ct io n j+1 .j j−1 2 j−1 A B C u 3 .e. 3 2 p1.j 2 (1 + β 2 ) Ω Appendix A Background material i−1 i i+1 j−1 This iteration method has a slow convergence. it is possible to show Φ − Φ k ≤ ρm Φ − Φ 0 where ρ = 1 − O h2 h = ∆x = ∆y e.1 Iterative solutions to linear systems n Γ0 Γ1 Gauss-Seidel method Ω The Laplace equation on a Cartesian grids can be discretized as M2 M Φi+1.1 0 1 2 i j Inﬂow Solid wall A. i.j + β 2 (Φi.1 2 v1.j−1 ) = 0 m m+2 where β = ∆x/∆y.j−1 ) h Φi.j + Φi−1.j − 2φi.i.

j i. Thus we are left with an equation implying that the Laplace operator on the correction is equal to the negative of the residual.g.restriction operator from grid 1 to grid 2 1 I2 .1: Domain for the solution of the Laplace equation. Γ0 Γ1 y Ω : to be updated : new values1 : old values n i−1 i i+1 j−1 j j+1 iteration direction 1 1 n 1 Figure A. Let LΦi.j − solution at iteration level k − correction to Φk to true discrete solution i.e. x Multigrid method A general way to accelerate the convergence of e. The idea behind the multigrid method is that a slowly converging low-frequency on a ﬁne grid is a fast converging high-frequency on a coarse grid. The prolongation operator I 2 may be deﬁned by linear interpolation for the intermediate values.j−1 Φi+1.j Φk i.j + Φi−1. we use every other value in each direction.j since LΦi.j − 2Φi.j+1 − 2Φi. This residual equation is solved on a coarser grid and the correction is interpolated onto ﬁner grid.prolongation operator from grid 2 to grid 1 1 deﬁne it as an injection. i.j = −LΦk ≡ −Ri.j .e.j = Φi.j ∆Φi.j + Φi.j + ∆x2 ∆y 2 and deﬁne the correction to the solution to the true discrete solution in the following manner Φi. i.j This implies that L∆Φi. the Gauss-Seidel method is provided by the Multigrid method. an average of right and left values = Φk + ∆Φi. We note that the Gauss-Seidel provides good smoothing of the local error.Ω 1 M2 h = ∆x = ∆y = ≈ n−1 = N −1/2 M n+1 m which implies that m+2 1 m = O (N ln(N )) h and that the total work is V Vh W = O N 2 ln(N ) u uh since the work per iteration is O (N ).j = 0. but converges slowly since it takes time for boundary information to propagate into the domain.j i. In order to deﬁne the algorithm we use the following deﬁnitions I transfer operator between grids: 2 An example of a restriction operator I1 is to 2 I1 .

. Vectors and tensors are objects which are independent of the coordinate system they are represented in.3 we have the unit vectors |ek | = 1 k = 1.j i. Example δkk = δll = δ11 + δ22 + δ33 = 3 δij aj = δi1 a1 + δi2 a2 + δi3 a3 = ai . A.j = −I1 Ri.j and goto 1.j i. 3 The scalar product of two unit vectors is the Kronecker’s delta. at the end of 1. a substantial improvement over the Gauss-Seidel method. The values marked with × in ﬁgure A. Φi.j iterated k times 2 2.j iterated k times. ∆Φ0 = 0 (starting value) (here ∆Φi.2: Two level grid. δkl ek · el = δkl = and the position vector r is 3 1 0 k=l k=l r = x 1 e1 + x 2 e2 + x 3 e3 = k=1 xk ek = x k ek . 2. The last expression make use of Einstein’s summation convention: when an index occurs twice in the same expression. the expression is implicitly summed over all values for that index.2 are then taken as an average of the averages of the intermediate values. Wesseling [10] estimated the work of the multigrid method to be W = O (N ln N ).j ⇒ LΦk = Ri.j = Φk + I2 ∆Φi.2 Cartesian tensor notation Tensor notation is a compact way to write vector and tensor formulas.j 1 3. In the Cartesian coordinate system in Figure A.j is deﬁnied on grid 2) i. Now we can deﬁne the following two level scheme: 1. LΦi. L∆Φi. convergence is checked.: old values i−1 i i+1 j−1 j j+1 iteration direction x y 1 n y ﬁne grid (1) coarse grid (2) j x i Figure A. and top and bottom values.

x3 )x3 e k · e1 x1 + e k · e2 x2 + e k · e3 x3 ⇒ xk = el · ek xl = cl k xl = clk xl we conclude that ﬁrst index refer to primed system! ⇒ xk = e k · el xl = c k l xl . Assume identical origin. l = 2) cl k = el · ek = cos(xl .4. The unit vectors are orthogonal ek · el = δkl and a position vector is independent of coordinate system. Since cos(x2 . a = 0. xl ).i y = x2 i i+1 j j−1 ﬁne grid (1) j j coarse grid (2) +1 r iteration direction x y e1 e1 1 x = x1 n e3 e2 x z = x3 y i Figure A. r = x l el = x l el . A. x1 ) = cos(x1 . xk ) = ck l From ek · r = x l ek · e l = x l ek · e l Example xk = = ckl xl = ck1 x1 + ck2 x2 + ck3 x3 = cos(xk .2.4: Two diﬀerent coordinate systems. x2 ) in general. j ﬁne grid (1) x2 coarse grid (2) x2 e2 e2 e1 e1 x1 a e1 e3 x3 x1 e3 x3 Figure A.3: Cartesian coordinate system. x1 )x1 + cos(xk .1 Orthogonal transformation Consider two arbitrary oriented coordinate systems as in Figure A. Component k of r can be written as ek · r = xl ek · el = xl el · ek = xl δkl = xk where we have the transformation tensor ck l = ek · el = cos(xk . x2 )x2 + cos(xk . we have (for k = 1.

. δkl is isotropic. .2 Cartesian Tensors Def.5: Projection of e1 x1 on e2 .n x y i j ﬁne grid (1) coarse grid (2) e1 x1 e1 e2 e2 · e1 x1 = cos (x2 .2. i. δlm for vectors R = 1 ak = ckl al and for a tensor of second order we have R = 2 δkl = ckm cln δmn = ckm clm = δkl . In the same manner we have xk δkm xm = clk xl clm xm = clk clm xm ⇒ (clk clm − δkm ) xm = 0. which transform as a vector in each of its R indices. e2 x2 .. Example For scalars the rank is R = 0 a · b = ak bk = ckl ckm al bm = al bl .. An outer product generates a tensor Tkl Srs = ckm cln crp csq Tmn Spq and an inner product generates a tensor Tkl Sls = ckm cln clp csq Tmn Spq = ckm csq Tmn Snq δnp . = ckr cls cmp ... x1 ) x1 = c21 x1 Figure A. Trsp.5. The relation between ckl and cml can be studied through xk δkm xm = ckl xl cml xm = ckl cml xm ⇒ (ckl cml − δkm ) xm = 0. A visualization of the projection of e1 x1 on e2 can be seen in Figure A.e. A. identical in all coordinate systems. which for component 2 reads x2 = e 2 · e1 x1 + e 2 · e2 x2 + e 2 · e3 x3 x2 is projection of e1 x1 . .. a Cartesian tensor of rank R is a set of quantities Tklm. Tklm. e3 x3 on e2 .

T kl = T(kl) + T[kl] .3.2. where the antisymmetric part is 1 T[kl] = −T[lk] = (Tkl − Tlk ) 2 and the symmetric part 1 T(kl) = T(lk) = (Tkl + Tlk ) .1) A.2. 3 . The inner product of the vectors a and b is The cross product of a and b can be written as a×b= e1 a1 b1 e2 a2 b2 e2 a3 b3 = (a2 b3 − a3 b2 )e1 + (a3 b1 − a1 b3 )e2 + (a1 b2 − a2 b1 )e3 = εijk ei aj bk and the determinant of the tensor aij is |{aij }| = = a11 a21 a31 a12 a22 a32 a13 a23 a33 = εijk a1i a2j a3k 1 1 εijk (ai1 aj2 ak3 + ai3 aj1 ak2 + ai2 aj3 ak1 − ai2 aj1 ak3 − ai3 aj2 ak1 − ai1 aj3 ak2 = εijk εpqr aip ajq akr . crossproducts and determinants a · b = δij ai bj = ai bi = a1 b1 + a2 b2 + a3 b3 . (3.3). If we set s = l in (A. (3. 6 6 We have the permutation relation εklm εksp = δls δmp − δlp δms .3 Permutation tensor even 1 odd −1 permutation of (1.1) we get εklm εklp = (3 − 1)δmp = 2δmp and if we also use p = m we end up with εklm εklm = 6 The permutation tensor can also be used to derive the following vector relation (a × b) · (c × d) = εijk aj bk εilm cl dm = (δjl δkm − δjm δlk )aj bk cl dm = aj cj bk dk − aj dj bk ck = (a · c)(b · d) − (a · d)(b · c) (A.1) (2. the isotropic part being 1 T kl = Tmm δkl 3 and the traceless part is 1 T kl = T(kl) − Tmm δkl .2.A. (2.3 we have (1. T (kl) = T kl + T kl .1). (1.2.2.5 Second rank tensors All second rank tensors can be decomposed into symmetric and antisymmetric parts. 3) if klm is = no 0 even permutation odd permutation no permutation ε123 = 1 ε213 = −1 ε221 = 0 εklm As examples of permutations of 1.1.4 Inner products. 2.2) at least 2 subscripts equal A.3. 2 The symmetric part can be decomposed into isotropic and traceless part.1.2.2).3).

⇒ T[lm] = In order to sum up. In the Navier-Stokes equations we have ∂ui the velocity gradient ∂xj . traceless and entisymmetric parts we get S(kl) T[kl] = S kl T kl = 1 1 S(kl) T[kl] + S(lk) T[lk] = S(kl) T[kl] − S(lk) T[kl] = 0 2 2 = 1 1 Smm Tkk − Tkk δll 3 3 =0 1 1 Smm δkl T(kl) − Tmm δkl 3 3 ⇒ Skl T[kl] Skl T(kl) Skl T kl Skl T kl = = = = S[kl] T[kl] S(kl) T(kl) S kl T kl S kl T kl . Example We have the second rank tensor which can be divided into a symmetric part 1 {Tkl } = 4 7 1 = 3 5 2 5 8 3 6 9 5 7 9 −2 −1 . an arbitrary tensor of rank 2 can be divided into the following parts Tkl = T(kl) + T[kl] = T kl + T kl + T[kl] = 1 1 Trr δkl + T(kl) − Trr δkl + 3 3 isotropic traceless 1 εikl di 2 antisymmetric Properties of the above parts are invariant under transformation. The antisymmetric part of this tensor describes rotation.If we project the isotropic. 3 Since we only need three components to completely describe an antisymmetric tensor of rank R = 2 it can be represented by its dual vector di = εijk Tjk = $$$$ + εijk T[jk] εijk T(jk) The ﬁrst term on the right hand side is zero since εijk is antisymmetric in any two indecies. the isotropic part the volume change and the traceless part describes the deformation of a ﬂuid element. and each part is invariant under transformation T[kl] = ckr cls T[rs] = −ckr cls T[sr] = −clr cks T[rs] = −T[lk] Trr = crp crq Tpq = δpq Tpq = Tpp 1 T kl = ckp clq T pq = ckp clq T(pq) − Trr δpq 3 1 = T(pq) − Trr δkl . Multiply both sides by εilm 1 εilm di 2 εilm di = εilm εijk Tjk = (δlj δmk − δlk δmj ) Tjk = Tlm − Tml = 2T[lm] Note: There are 3 independent components in T[lm] and di . 0 T(kl) and an antisymmetric part T[kl] 3 5 7 0 = 1 2 −1 0 1 .

√ x k xk . x3 ) ≡ Tij (xk ) with the following properties • partial derivative transforms like a tensor ∂ ∂xk ∂ ∂ = cpk = ∂xp ∂xp ∂xk ∂xk • gradient adds one to tensor rank ∂ Tij ∂xk where (xk = cpk xp ) • divergence decreases tensor rank by one ∂T1l ∂T2l ∂T3l ∂ Tkl = + + ∂xk ∂x1 ∂x2 ∂x3 • curl ∂uk ∂xj ( × u)i = εijk which can be compared with the expression for the determinant.2. 7 4 −2 {di } = 4 . since εklm is antisymmetric in kl and Example Show that ∂2 ∂xk ∂xl f (r) = 1 f (r) where {r}k = xk and r = |r| = r = { f (r)}k df ∂r ∂(xp xp )1/2 ∂ f (r) = =f = ∂xk dr ∂xk ∂xk 1 1 1 −1 ∂ f (xp xp ) 2 (xp xp ) = f (δpk xp + xp δpk ) = f xk = 2 ∂xk 2r r 1 fr r k Example Show that ×( ×( × F) × F) = i ( · F) − ∆F.6 Tensor ﬁelds We have the tensor ﬁeld Tij (x1 . Example Show that ·( × Ψ) = 0. −2 A. x2 . = εijk ∂ ∂ ∂ ∂ ∂ ∂ εklm Fm = εijk εklm Fm = εkij εklm Fm = ∂xj ∂xl ∂xj ∂xl ∂xj ∂xl .and a traceless part The symmetric part can then be divided into a isotropic 5 T kl = 0 0 T kl −4 = 3 5 part 0 5 0 0 0 5 The dual vector of the antisymmetric tensor is 3 5 0 7 . ·( × Ψ) = ∂ ∂Ψm ∂2 εklm = εklm Ψm = 0 ∂xk ∂xl ∂xk ∂xl is symmetric in kl.

7. (δil δjm − δim δjl ) Example Show that if u = Ω × r then Ω = ¯ ¯ ¯ εijk ∂ ∂ ∂ ∂ ∂ ∂ ∂ ∂ ∂ ∂ Fm = Fj − Fi = Fj − Fi = ∂xj ∂xl ∂xj ∂xi ∂xj ∂xj ∂xi ∂xj ∂xj ∂xj ×u ∂ . ∂xp Example f n dS = S V f dV . if Tklm nm = εklm al nm . 2 ∂xj ( · F) − ∆F i ∂ ∂ ∂ uk = {uk = εklm Ωl rm } = εijk (εklm Ωl rm ) = εijk εklm Ωl rm = ∂xj ∂xj ∂xj ∂ ∂ Ωl rm = (δil δjm − δim δjl ) Ωl rm = ∂xj ∂xj ∂ri = δil } = ∂xl εkij εklm ∂ ∂ ∂rm ∂ri ∂rm Ωi rm − Ωl ri = {Ω independent of x} = Ωi − Ωl ={ =1+1+1=3 ∂xm ∂xl ∂xm ∂xl ∂xm 3Ωi − Ωl δil = 2Ωi A.y 1 n x y i j ﬁne grid (1) coarse grid (2) n S V dS = n dS Figure A. see Figure A. This can be generalized to Tklm np dS = S V ∂ Tklm dV .2.6. of which dl is the line element as in Figure A. then a · dl = C S ( × a) · n dS . This means.7 Gauss & Stokes integral theorems Let n be the outward pointing normal to the closed surface S bounding the simply-connected volume V .6: Volume V bounded by the close surface S. show that εijk uk = 2Ωi . a × n dS = S V × a dV Stokes theorem states that for a diﬀerentiable vector ﬁeld a if an open simply-connected surface S is enclosed by a closed curve C. Gauss integral theorem then states a · n dS = S V · a dV ∂ak dV ∂xk or ak nk dS = S V for a diﬀerentiable vector ﬁeld a.

Figure A. Consider a water cylinder with volume dV and mass dm.8. The pressure at a distant x 3 from the surface in a ﬂuid is then p = ρgx3 + p0 where p0 is the atmospheric pressure.8: An object with mass M immersed in a ﬂuid illustrating Archimedes principle.8.7: The open surface S enclosed by the contour C. n p0 x y i n j ﬁne grid (1) dV dm coarse grid (2) x3 M dA Figure A.2. m the mass of equivalent volume of water and M the mass of the object in Figure A. Πk = − ρgx3 nk dS = −ρg ∂xk S V . Let Π be the buoyancy force resulting from the pressure on the object. The force from the water cylinder is dm g = dV ρg = x3 ρg dA pressure where g is the gravity constant and ρ the water density.8 Archimedes principle Gauss theorem can be used to calculate the buoyancy on an object immersed in a ﬂuid (Archimedes circa 281–212 BC). or ak dxk = C S εklm nk ∂am dS ∂xk this can be generalized to Tklm dxp = C S εijp ni ∂ Tklm dS ∂xj Example Let Tklm = f.j n i ﬁne grid (1) i+1 coarse grid (2) j−1 S j j+1 iteration direction dl C x y 1 Figure A. In the generalized Stokes theorem we then get S ∂f εijp ni ∂xj dS in right hand side f dl = C S n× f dS A. The force on a small element dS of the body resulting from the pressure is dΠk = −gρx3 nk dS where the minus sign stems from the outward pointing normal n. The buoyancy force can then be calculated through ∂ x3 dV = −ρgδk3 V.

This can then be used to calculate the total moment around G form pressure forces on the body M Gk = − S ρgεklm (xl − xGl ) x3 nm dS = −ρgεklm V −ρgεklm V If a part of the ﬂuid is frozen (without changing in density) this body is in equilibrium in the ﬂuid: F = M = 0. The buoyancy force is Π3 = −ρgV = −g m see Figure A.9: The forces acting on an object immersed in a ﬂuid. The moment around the center of gravity G of the homogenous body produced by the pressure force dΠ acting on an element dS of the body is dMG = (r − rG ) × dΠ as in Figure A. m is the mass of equivalent volume of n water and M the mass of the body. [δlm x3 + δm3 xl − δm3 xGl ] dV = −ρgεkl3 ∂ ∂x3 (xl x3 ) − xGl dV = ∂xm ∂xm V xl dV − xGl V = 0 . This means that the object loses some weight as the water is displaced.mg i x y i+1 j−1 i j j j grid ﬁne+ 1 (1) coarse grid (2) iteration direction x y Mg 1 Figure A.10. x y i j ﬁne grid (1) G coarse grid (2) r − rG n Figure A. The moment of a force F at some arbitrary point with position vector r from the point of application of the force is M = r × F.9.10: Moment around G from pressure forces acting on the object.

.

Appendix B Supplementary material B. Tue 18/1 8-10.se. henning@mech.kth. (TAP) Copies from Randall J LeVeque. Fluid dynamics I: LB Introduction and outline of the course.se. tel 790 7161 Erik St˚ alberg (ES). 2. E51. homeworks (required) 3×5 = 15p. Gustavsson. Lecturers: Dan Henningson (DH). cont. Well-posedness. parabolic and elliptic equations.se. Computational Fluid Mechanics and Heat Transfer.mech. tel 790 6696 Luca Brnadt (LB). tel 790 7161 Literature: Copies from J.kth. Mekanik. tel 790 9004 Katarina Gustavsson (KG).A. Web links http://www. Basic numerics I: KG Mathematical behavior of hyperbolic.kth. Lecture notes on Basic Numerics Grading: Test total of 50p. Pletcher. Thu 20/1 8-10. R. Tannehill.se. Fluid dynamics II: LB Derivation of the governing equations . Mekanik. luca@mech. eriks@mech. Mekanik. 107 . E51. > 50 (5).. Taylor & Francis. Lecture notes on Computational Fluid Dynamics K. D. Fininte Volume Method for Hyperbolic Problems. E35. 2002 (LeV) D. > 40 (4).1 Syllabus 2005 Computational Fluid Dynamics (5C1212). KG Lecture notes 3.kth. 5p. NADA. 1997. Total points > 30 (3). Anderson. Cambridge University Press.kth.H. Henningson.se/ Course plan 1. Derivation of the governing equations. Wed 19/1 15-17. katarina@nada.C.

Basic numerics II: KG Discretization by ﬁnite diﬀerences. TAP: 4. Compressible ﬂow in 2D. Lax-Wendroﬀ. Thu 27/1 8-10. Flow around an airfoil. 14 12. D34. Rankine-Hugoniot relations. weak solutions. Shock formation. Lax-Friedrichs. Fri 28/1 15-17. Wed 2/2 15-17. Fri 4/2 15-17. 16. Q35. Burgers equation. Compressible ﬂow V: KG Numerical methods for Euler equations. entropy conditions. equivalence with ﬁnite-diﬀerences. 8. Compressible ﬂow IV: KG High resolution schemes for conservations laws.dissipation and dispersion. Shock tube problem. Compressible ﬂow I: KG Introduction to compressible ﬂow. Some thermodynamics. Compressible ﬂow II: KG Numerical methods for scalar conservation laws. Numerical/physical boundary conditions. E52. Fri 11/2 15-17.1.6 7. E51 . convection-diﬀusion equation. Tue 8/2 08-10. order of accuracy. Fluid dynamics III: LB Derivation of the governing equations . cont.4. . 15.1-3. D34. TAP: 3.9. Tue 25/1 8-10. convergence and stability (von Neumann analysis). Properties of the equations. Basic numerics IV: KG Numerical methods for model equations.1-4. Tue 15/2 08-10. First and second order numerical methods such as upwind.6. L21. role of the pressure: artiﬁcial compressibility and projection on divergence free space. E35. Wed 9/2 15-17. LeV: 11 10.3. Properties of the numerical scheme such as CFLcondition. 6. First order methods. Riemann invariants. 3 11. Basic numerics III: KG Numerical methods for model equations related to diﬀerent levels of approximation of Navier Stokes equation: linear wave equation. Analysis of discretized equations.5. MacCormack. LeV: 8-8.3-6. ﬂuid phenomena. conservative/non-conservative form. etc. Mon 31/1 13-15.3. Euler equations. Wed 16/2 15-17. Finite volume and ﬁnite diﬀerence methods II: DH Introduction to incompressible ﬂow. Modiﬁed equation . LeV: 14 Scalar conservations laws: Conservation. linear systems: Gauss-Seidel as smothers for multi-grid.8. 9. E51. System of conservations laws LeV: 4. LeV: 6 13. non-uniqueness. simpliﬁed equations 8. algebraic mesh generation by transﬁnite interpolation.1-4. conservation and TVD.3 5. Compressible ﬂow VI: KG Numerical methods for Euler equations. Grids.2. Q36. LeV: 14 14. TAP: 6. Fri 21/1 15-17 L21. E3. Fluid dynamics IV: LB Dimensionless form. Compressible ﬂow III: KG Numerical methods for Euler equations: MacCormack and artiﬁcial viscosity for non-linear systems. cont. Navier-Stokes in integral form. Wed 26/1 15-17. cont. Finite volume and ﬁnite diﬀerence methods I: DH Laplace equation on arbitrary grids. V33. 3. LeV: 3. E52.7. LeV: 8. Boundary conditions.

Stability. • Homework 2.17. Finite elements I: DH An advection–diﬀusion problem. due 15/2: Numerical methods for non-linear conservation laws. Stability. Mixed variational form. Q33. 23. stability and accuracy. due 22/2: Running compressible Euler code. diﬀusion. 0) = = 0 1 0 − ∞ < x < ∞. due 1/3: FV homework + numerical project. Finite volume and ﬁnite diﬀerence methods III: DH Staggered grid/volume formulation + BC. E35. due 31/1: Homework on NS equations. Finite-element approximations. Fri 18/2 15-17. x≤0 otherwise t>0 2. Tue 1/3 13-15. mass conservation. Finite elements II: DH Navier–Stokes equations. Galerkin and FE approximations. Sat 2/4 9-13. E35. Deﬁne and use the Rankine-Hugoniot jump condition to compute the shock speed for the following problem ut + uux u(x. 20. Shock-tube. discrete Poisson pressure eq. • Homework 3. 19. Galerkin and FE approximations. the LBB condition. ﬂow over a bump. Galerkin and FE approximations. Thu 24/2 8-10. • Homework 4. weak solutions. The shock is moving with speed s and the state to the left is given by uL and the state to the right by uR . the LBB condition. EXAM. due 31/1: Numerical methods for model equations. 22. matrix assembly. Wed 2/3 15-17. Fri 25/2 15-17. • Homework 5. mass conservation. the algebraic problem. Q34. 18. Unsteady equations: projection and MAC method. the algebraic problem. Variational form of the equation. the algebraic problem. E31. Why do we need an entropy condition ? 3. B. Finite elements IV: DH Navier–Stokes equations. due 8/3: FEM homework. Homeworks • Homework 1A. Deﬁne the entropy condition for a scalar conservation law. ut + f (u)x = 0 − ∞ < x < ∞. Dispersion. Tue 22/2 8-10. Finite volume and ﬁnite diﬀerence methods IV: DH Time step restrictions. Diﬀerent schemes. E31. . the algebraic problem. the LBB condition. Steady equations: distributive iteration and SIMPLE methods. • Homework 1B. Write a diﬀerence approximation (in 1D) on conservative form and deﬁne the notation. Stability analysis. Finite elements III: DH Navier–Stokes equations. essential and natural boundary condition. 21. E31. Mixed variational form. Mixed variational form. Stability.2 Study questions 1. t>0 with a convex ﬂux function f (u). mass conservation.

Solve ut + ( u2 )x = 0. 9. What is the condition on the n × n real matrix A(u) for the system ut + Aux = 0 to be hyperbolic ? 11. u ) around a motionless gas. 5. ∆t n (u − un ) j−1 ∆x j Show that this approximation is not stable even though the CFL condition is fulﬁlled. Assume k/h constant. Why is this a desirable property ? 6. b) What type of equations is this? c) What kind of behavior can we expect from the solution? 8. a) Write down the modiﬁed equation. Deﬁne a total variation decreasing (TVD) method. Investigate the one-sided diﬀerence scheme un+1 = un − a j j for the advection equation ut + aux = 0 Consider the cases a > 0 and a < 0. 2 u(x. suggest another way of solving B. (B. can be linearized by considering small perturbations (ρ .1) by the upwind scheme on the following form. Apply Lax-Friedrichs scheme to the linear wave equation ut + aux = 0.1. un+1 = un − j j ∆t n n u (u − un ) j−1 ∆x j j a) Will the numerical solution converge to the analytical solution? b) If not. yields the approximation un+1 = un + j j a∆t (uj+1 − uj−1 ) 2∆x a > 0. What does Lax’s equivalence theorem state? 10. a) Prove that the scheme is consistent and ﬁnd the order of accuracy. b) Determine the stability requirement for a > 0 and show that it is unstable for a < 0. The barotropic gas dynamic equations ρt + ρux = 0 1 ut + uux + px = 0 ρ where p = p(ρ) = Cργ i and C a constant.2) . 7.4. A the three-point centered scheme applied to ut + aux = 0. 0) = 1 0 x<0 x≥0 (B.

1. see the ﬁgure below. a) Draw the domain of dependence of the solution to the system (B.. n = 0. c) Determine the characteristic variables in terms of ρ and u. 1.. b) The system is solved numerically on a grid given by xj = j∆x.3) is given by ρ u + t 0 a2 /ρ0 A ρ0 0 ρ u = 0.1t n+1 x y i t ﬁne grid (1) n coarse grid (2) j−1 j j+1 a) Let ρ = ρ0 + ρ and u = u0 + u where u0 = 0. j = 0. 2. The linearized form of (B. 13. a and ρ0 are constants.4) in a point P in the x-t plane. a and ρ0 are constants.4) g hgh opopop opop opop opop mnm op jij i . . 2.. To solve Euler equations in 1D ρt + ρux + uρx = 0 1 ut + uux + px = 0 ρ pt + ρc2 ux + upx = 0 How many boundary conditions must be added at (motivate your answer) inﬂow boundary when the ﬂow is a) Supersonic b) Subsonic outﬂow boundary when the ﬂow is c) Supersonic d) Subsonic kk ll (B. and tn = n∆t.3) for t > 0.. Draw the domain of dependence of the numerical solution at P (in the same ﬁgure as a)) of the three-point scheme in the case when i) the CFL condition is fulﬁlled ii) the CFL condition is NOT fulﬁlled.3) a hyperbolic system? Motivate your answer. Hint: Start from the characteristic formulation. e) Given initial conditions at t = 0 and let −∞ < x < ∞ (no boundaries) ρ(0. x) = sin(x) u(0. 12. Assume that P is a grid point.characteristic formulation. Linearize the system (B. x) = 0 determine the analytical solution to (B.2) and show that this yields the following linear system (the primes has been dropped) ρt + ρ0 ux = 0 a2 ut + ρx = 0 ρ0 where a is the speed of sound..3) (B. b) Is the system given by (B. using an explicit three-point scheme. x where a is the speed of sound. d) Determine the partial diﬀerential equations the characteristic variables fulﬁll ..

Ut + F (U)x = 0 15. b) Deﬁne the 2-level multigrid method for the Laplace equation. State the diﬃculties associated with the the ﬁnite-volume discretizations of the Navier-Stokes equations on a colocated grid? and show the form of the spurious solution which exist. From the diﬀerential form of the Navier-Stokes equations obtain a) the Navier-Stokes equations in integral form used in ﬁnite-volume discretizations. c) a stable choice with piecewise linear velocities. b) Describe and interpret the LBB condition. Discrete elements pairs a) constant pressure-bilinear velocities. show that the pressure term disappears and recover an equation for the pressure from the gradient part. b) a variational form of the Navier-Stokes equations used in ﬁnite-element discretizations. b) derive the FV discretization for Laplace equation on a Cartesian grid. 17. 23. b) Explain the concept of essential and natural boundary conditions. c) discuss how to treat noslip and inﬂow/outﬂow boundary conditions. 25. 21. show that any Galerkin velocity solution is stable. Iterative techniques for linear systems. 18. b) the Taylor-Hood pair. b) write down the FV discretization of the Navier-Stokes equations on a staggered cartesian grid. a) Deﬁne Gauss-Seidel iterations for the Laplace equations. 22. a) Show that a vector ﬁeld wi can be decomposed into wi = u i + ∂p ∂xi where u is is divergence free and parallel to the boundary. Describe the ideas behind a ﬂux splitting scheme for solving a non-linear hyperbolic system of equations. 20.g. Derive the ﬁnite element (FEM) discretization for Laplace equation on a Cartesian grid and show that it is equivalent to a central diﬀerence approximation. e.14. a) How is a ﬁnite-element approximation deﬁned? b) Explain how to convert a FEM discretization to an algebraic problem. a) Deﬁne an appropriate staggered grid that can be used for the discretization of the Navier-Stokes equations. c) show that both are equivalent to a central diﬀerence approximation for Cartesian grids. 24. . a) Derive the ﬁnite volume (FV) discretization on arbitrary grids of the continuity equation (∂u i /∂xi = 0). a) Write down the appropriate function spaces for the pressure and velocity used to deﬁne weak solutions to the Navier-Stokes equations. a) By choosing zh = uh in the ﬁnite element approximation of the Stokes problem. that the Navier-Stokes equations yield νK + C(u) GT G 0 u p = f 0 19. 16. b) Apply this to the Navier-Stokes equations. Choice of elements.

a) Motivate the use of an appropriate form of the advection-diﬀusion equation as a model equation for stability analysis. a) Deﬁne a simple projection method for the time dependent Navier-Stokes equations d dt u 0 + N(u) D G 0 u p = f g b) show in detail the equation for the pressure to be solved at each time step and discuss the boundary conditions for the pressure.26. b) derive the time step restrictions for the 1D version of that equation. Time dependent ﬂows. Time step restriction for Navier-Stokes solutions. 27. . c) state the 2D equivalent of that restriction.

.

2001.H. Cohen. [3] C. 1998. [6] P. [7] R. Second Edition. Computational Methods for Fluid Flow. Academic Press. T. Jr. [4] C. T. 1995. Springer-Verlag. Acheson. Computational Fluid Dynamics. Computational Techniques for Fluid Dynamics Volume 1. Second Edition. 1995. Oxford University Press. Dornseifer and T. SIAM Monographs on Mathematical Modeling and Computation. Springer-Verlag. [9] J.C. Pletcher. Taylor and Francis. Anderson. A practical Introduction. Taylor. The basics with applications. Springer-Verlag. [10] P. 1991. Elementary Fluid Dynamics. 1991. Peyret. [2] J. Springer-Verlag. Fletcher. Second Edition. Wesseling. 2002. Fluid Mechanics. Griebel. Computational Techniques for Fluid Dynamics.J. 1983.L.D. Principles of Computational Fluid Dynamics. 1997. Neunhoeﬀer. Wiley-Interscience.K. Computational Fluid Mechanics and Heat Transfer.M. D. Anderson and R.A. Second Edition.J.Bibliography [1] D..J. Second Edition. Numerical Simulation in Fluid Dynamics. [5] M. [8] R. Tannehill.A. Kundu and I. 115 .A. Incompressible Flow. Panton. Mc Graw-Hill.D. 1990. Fletcher. Volume 2.