MACE60711

Three Hours

UNIVERSITY OF MANCHESTER

COMPUTATIONAL FLUID DYNAMICS

January 2022

00:00 – 00:00

Answer ALL questions.

Electronic calculators may be used, provided that they cannot store text.

©The University of Manchester, 2022.

1 of 5
 MACE60711

Question 1
Consider the semi-discretisation of the linear advection equation

∂u ∂u
+c =0
∂t ∂x
on a regular mesh {xm : xm = m∆x}. Two finite difference approximations of different
orders of accuracy are considered for the discretisation of the left hand boundary influenced
grid point x0 .
The two schemes under consideration are
Scheme (I):
∂u 1
≈ (au0 + bu1 + cu2 )
∂x x0 ∆x
and Scheme (II):
∂u 1
≈ (mu0 + nu1 + pu2 + qu3 )
∂x x0 ∆x
where (a, b, c) and (m, n, p, q ) are coefficients to be determined.

a). For the stencil (I) above, find the values of (a, b, c) in order to produce an O((∆x)2 )
accurate approximation to the first derivative at x0 .
[5 marks]

b). For the stencil (II) above, find the values of (m, n, p, q ) in order to produce an O((∆x)3 )
accurate approximation to the first derivative at x0 .
[8 marks]

c). For the sets of coefficients calculated in parts (a) and (b), calculate and sketch the
spectral response for both differencing schemes. In each case, compare the behaviour
with the idealised response up to the Nyquist frequency. On your sketches, make clear
which curve is which, identify the Nyquist frequency, and make clear your axes.
[12 marks]

2 of 5
 MACE60711

Question 2
Consider the linear system
AF = B
Iterative solvers use the sequence F (1) , F (2) , ...F (r) → F sol and can be generically described
by the formula below for suitable matrices P and N :

P F (r+1) = N F (r) + B

a). Give the definitions of the error vector and the residual vector.
[3 marks]

b). Show that, for consistency, the matrices P and N should be chosen such that

A = (P − N )
[5 marks]

c). Explain what general properties one might want the matrix P to have. As examples,
show how the Jacobi and Gauss-Seidel methods could be used to obtain forms for P .
[5 marks]

d). Show that the error in one iteration, ε(r+1) , can be related to that in the previous itera-
tion, ε(r) , as follows:
ε(r+1) = P −1 N ε(r)

Hint: Note that, thanks to the consistency condition in part (b), one may replace B by
an expression using N , P and F sol in the iteration formula

P F (r+1) = N F (r) + B

[6 marks]
1/k
e). The spectral radius of a matrix M is defined by ρ(M ) = lim Mk . Show that for
k→∞
the above iterative solution method to converge it is necessary that ρ(P −1 N ) < 1.
(Note that the spectral radius can also be defined as the largest eigenvalue of the
matrix: ρ(M ) = maxi [|λi |])
[6 marks]

3 of 5
 MACE60711

Question 3
A 1-dimensional steady fluid flow is governed by the continuity and U -momentum equations:

∂
(ρU) = 0 (Q3.1)
∂x  
∂ 2

∂P ∂ ∂U
ρU = − + µ (Q3.2)
∂x ∂x ∂x ∂x

where ρ is the density, µ the viscosity, P pressure, and U may here be assumed to be
positive. This flow is to be modelled numerically, on a uniform grid as shown in Fig. Q3.1,
with a fully collocated, cell-centred, storage arrangement for the variables.

i−1 i i+1
∆y

∆x
Figure Q3.1

a). Show that using a standard finite volume discretization, with a first order upwind ap-
proximation for convection terms, and second order central differences for other gradi-
ents, the above momentum equation may be approximated by


(ṁ + 2µ∆y/∆x) Ui = (µ∆y/∆x) Ui+1 + (ṁ + µ∆y/∆x) Ui−1 + ∆y Pi−1/2 − Pi+1/2
(Q3.3)
where ṁ is the mass flow rate, ρU∆y , and subscripts i + 1/2 and i − 1/2 denote
a quantity evaluated at the cell faces between nodes i and i + 1, and i − 1 and i,
respectively.
[10 marks]

b). If the cell-face pressures Pi+1/2 and Pi−1/2 are approximated using standard linear
interpolation, show that the discretized pressure contribution to equation (Q3.3) can
be written as (∆y/2) (Pi−1 − Pi+1 ).
[2 marks]

c). If viscosity is neglected, and the pressure varies linearly with x, so P = −Bx for
some constant B , the exact solution of the original differential equation can be written
as U = (B∆y/ṁ)x + C for some constant C . Show that the discretized equation is
consistent with this exact solution (eg. take Ui−1 = (B∆y/ṁ)xi−1 + C and evaluate
what the discretized equation would then give for Ui ).
[5 marks]

d). Show that the same discretized solution as in part (c) would also be returned if we
took a pressure field of the form Pi = −Bxi + (−1)i ∆P for some constant ∆P , and
hence deduce that the numerical solution would allow unphysical pressure oscillations
to develop.
[5 marks]

e). Could the above pressure oscillations develop if the pressure was stored on the velocity
cell faces, instead of at the cell centres? Give brief reasons.
[3 marks]

4 of 5
 MACE60711

Question 4
A differential equation describing the pure diffusion of some scalar property φ may be written
as
∂φ ∂2φ
=Γ 2
∂t ∂x
Consider the discrete solution to this equation φnk = φ(n∆t, k∆x), obtained on a set of
regular grid points {xm : xm = m∆x}. A fourth order accurate approximation (i.e. O((∆x)4 ))
to the spatial derivative at grid point xk may be written as

∂2φ 1
≈ (aφk + b (φk+1 + φk−1) + c (φk+2 + φk−2)) (Q4.1)
∂x2 xk (∆x)2

where a, b and c are constants to be determined.

a). Calculate the coefficients a, b and c in equation (Q4.1) to provide an approximation

b). Using the forward Euler scheme in time

∂φk φn+1 − φn
≈ + O(∆t)
∂t tn ∆t

and the approximation you derived in part (a) of this question, apply Von-Neumann
stability analysis to derive a stability criterion (if one exists) to the fully discrete sys-
tem. How does this criterion (if it exists) compare with the ‘classical’ diffusion limit of
Γ∆t/(∆x2 ) ≤ 1/2?
[17 marks]

********** END OF EXAMINATION PAPER **********

5 of 5