Acta Polytechnica Hungarica

Investigation of Inverse Design Method for

Internal Flows

Árpád VERESS, Csilla TÓBIÁS

Budapest University of Technology and Economics, Department of Aeronautics,
Naval Architecture and Railway Vehicles, Stoczek u. 6., H-1111 Budapest,
Hungary, [email protected]
Abstract: Recently developed iterative CFD-based inviscid compressible flow solver with
inverse design method has been verified and tested by bump-in-channel flow cases in present
work. The software’s capabily for internal flows has been examined for further extension of
its usability. Possible application of the inverse design method using Stratford’s flow limiting
theory has been investigated on internal wall (e.g. diffuser) optimization. The outcome of the
study provides more information about the conditions and capabilities of the presented
inverse design method, which can significantly contribute to the successful application of the
coupled flow modelling and optimisation process.

Keywords: inverse design method; internal flows;optimization

1 Introduction
Nowadays, computer technology development and increasing computational
capacity make possible to solve and analyse high compute-demanding problems in
an acceptible time. This way computational flow modelling becomes more and
more widely spread. Applying CFD (Computational Fluid Dynamics) techniques
the number of experiments hence development cost and time can be reduced
considerably. With optimization tool integration the development process and
product efficiency can be further improved.
In industrial applications optimization of flow channel geometries is a frequently
occuring task. Flow losses can be reduced by eliminating stagnation points or
separation of the flow. Appropriate geometry design helps to establish required flow
characteristics that is essential for diffusers, inlet channels and water-cooling
schemes.

–1–
Veress et al. Investigation of Inverse Design Method for Internal Flows

1.1 Optimization Methods

1.1.1 Direct Optimization Techniques

These methods’ aim is to find an optimal solution by using gradient-based methods,
stochastic algorithms (e.g. evolutionary strategies, genetic programming) or even
other optimization methods. During the process a large number of flow simulations
may be calculated to determine the final state. This results in high computational
capacity demand which significantly increases with the number of design variables.

1.1.2 Inverse Design Based Optimization

In case of a specific set of the inverse design-type methods, the geometry
modification is based on the prescribed set of the pre-defined variables at the wall
by simple, fast and robust algorithms, which makes them especially attractive
amongst other optimization techniques. The wall modification can be completed
within much less flow solutions for inverse design techniques than for direct
optimization methods. Hence, the inverse design methods typically being much
more computationally efficient and they are very innovative to be used in practice.
The main drawback of inverse design methods is that the designer should create
target (optimum in a specific sense) pressure or velocity distributions that should
correspond to the design goals and meet the required aerodynamic characteristics.
However, it can be difficult to specify the expected pressure or velocity distribution
that satisfies all design goals. Also, one cannot guarantee that an arbitrarily
prescribed pressure/velocity distribution will provide mechanically correct surfaces
or bodies (airfoils without trailing edge open or cross over for example). [1]
The calculation process of the developed iterative type inverse design method is
shown in Fig. 1. The procedure, first of all, requires an initial geometry and a
required pressure distribution (preq) along the wall to be modified. The prescribed
distribution can be the goal function of an optimization method or it can come from
the industrial experiences and/or theory. The iterative cycle starts with the direct
solution of the inviscid Euler solver. Completing the convergence criteria, if the
target conditions are not reached, a new (opening) boundary condition is applied at
the solid boundary to be redesigned or optimized. The required pressure distribution
(preq) is imposed at the solid wall boundary, which is become locally opening as
inlet or outlet, depends upon the evolved pressure differences between the boundary
and computational domain. The outcome of this analysis is a velocity distribution
along the wall, which is not necessarily parallel with it. The final step of the cycle
is the wall modification. The wall becomes parallel with the local velocity vector
corresponds to a new streamline of the flow field. The mentioned steps are repeated
until the target distribution is reached by the direct analysis and so the new geometry
is available[2].
Recently, the inverse methods are based on the Euler and Navier-Stokes equations,
and the wall modification algorithm is completed by means of transpiration

–2–
Acta Polytechnica Hungarica

technique (Giles and Drela [3], Demeulenaere [4] and De Vito [5]). Although these
methods are primarily dedicated to the design of airfoils, wings and turbomachinery
cascades, it has also been applied for design of duct geometries (Cabuk and Modi
[6]). [7]

START

Initial Optimization
geometry criteria satisfied

Optimize pressure Impose optimal

distribution pressure distribution STOP
(for preq) and setup by Stratford limiting
model for morphing flow theory (preq)

Inverse analysis
Generate CFD mesh with opening wall
for prescribed set BC for Vn
distribution

Perform direct Wall modification

analysis algorithm

Figure 1

Procedure of the iterative inverse design

1.2 Stratford’s Separation Prediction

The inverse design methods require optimal pressure or velocity distributions in
some respects to determine the adherent geometry. The pressure gradient should
have been limited at each discretized point to be just below the condition of causing
separation. There are several existing methods for predicting separation such as
Goldschmied, Stratford, Head, and Cebeci-Smith. Due to the good accuracy, simple
expressions and conservative characteristics for predicting separation, Stratford’s
method has been used in followings. [7]
Stratford has derived an empirical formula for predicting the point of separation in
an arbitrary decelerating flow over a flat plate at the order of Re=10E6,


C p x dC p dx  12

S (1)
10 6
Re
1 10

where the canonical pressure distribution is

–3–
Veress et al. Investigation of Inverse Design Method for Internal Flows

p  p0
Cp  (2)
1
 0 u02
2

and if d 2 p dx2  0 then S = 0.39 or if d 2 p dx2  0 then S =0.35. Additionally,

C p  4 /7 . x is the distance measured from the leading edge of the plate and
Re  u0 x  . If the pressure rise in the flow begins at point x0 (it is the position of
minimum pressure and maximum velocity) then the left-hand side of eq. (1) grows.
When it reaches the limiting value of S, separation is said to occur. If S is held at its
limiting value of 0.39 for d 2 p dx2  0 eq. (1) reduces to an ordinary differential
equation for C p (x ) . It is evident from eq. (1) that the equation describes a flow that
is ready everywhere to separate. Stratford presents the following solutions [8],

 
C p  0.645 0.435Re0 1 5 x x0 1 5  1 
2n
for C p  n  2 n  1 (3)

and
a
Cp  1  for C p  n  2  n  1 (4)
x x   b
0
12

Figure 2

Stratford limiting flows at two values of unit Reynolds number [9]

In that two-part solution, x 0 is the start of pressure rise, Re0  u0 x0  , x is the

distance measured from the very start of the flow, which begins as flat-plate,
turbulent flow. The number n is a constant that Stratford finds to be about 6. The

–4–
Acta Polytechnica Hungarica

quantities a and b are arbitrary constants used in matching values and slopes in the
two equations at the joining point, C p  n  2  n  1 . Of course, eq. (3) describes
the beginning of the flow and eq. (4) the final part. The flow is an equilibrium flow
that always has the same margin, if any, against separation. Two families of such
flows have been computed; they are shown in Fig. 2.[9].[7]

2 Numerical Method

2.1 Governing Equations

Due to the aeronautical application with the assumption of no separation, the
conservative form of the unsteady 2D compressible Euler equations have been used
as governing equations for rotational flow modelling, which are given in Cartesian
coordinate system by [10]
U F U  G U 
  0, (5)
t x y

in  (x,y), where x, y Є R and t Є R+. The conservative variables and convective

fluxes are given by

    u   v 
   2   
 u   u  p   vu 
U    , F (U )   , G (U )   v 2  p  , (6)
v uv 
     
 E   uh   vh 
   0   0 

where  is the density, u and v are the Cartesian components of velocity and p
is the static pressure. The specific total energy and enthalpy are the following:

1 p u2  v2  p u2  v2
E  , h0   . (7)
 1  2  1  2

The  is the ratio of specific heat. [7]

2.2 Boundary Conditions

Theory of characteristics is used to determine the number of physical and numerical
boundary conditions. In case of subsonic inlet condition there are three ingoing and
one outgoing waves. Hence, 3 parameters; total pressure, total temperature and flow
angle are imposed as a physical boundary condition.

–5–
Veress et al. Investigation of Inverse Design Method for Internal Flows

 W n1   p  c 2   0, on the curve dn dt  V n 


 
 W n2    V s  0, the curve dn dt  V n
on 

 3    p  cV  0, on 
the curve dn dt  V n  c 
(8)
 n  
W n
 W 4    p  cV  0, on the curve dn dt  V n  c 
 n   n
Converting the characteristic problem into the normal and tangential direction the
dn dt  Vn  c curve has a negative slope, hence the fourth compatibility
equation belongs to Wn( 4) (8) is considered as a numerical boundary condition in
its total derivative form,

 p V   p V 
  c n     c n V n  c   0 . (9)
 t t   n n 
The combination of the energy and adiabatic Poisson state equation is used to derive
the second equations for missing variables. The Newton-Raphson method has been
implemented to solve the system equations for p and V n at the new time level. The
temperature and the components of the velocity vector are easily to be recovered by
using ideal gas law and inlet flow angle, while the tangential velocity component is
kept to be constant.
The characteristic method has also been applied at the subsonic outlet to determine
the unknown variables. In this case, there are two outgoing and one ingoing
characteristics, hence the equations belong to first and the third variable; Wn(1) and
Wn(3) (see (8)) are used. The static pressure at new time level is given as a physical
boundary condition, hence the system of equations can directly be solved for density
and normal velocity at the next time level. The tangential velocity is also supposed
to be constant due to the shear wave Wn( 2) .
Soft solid wall convergence accelerator technique has been used to make the
convergence faster based on [11]. The basic mechanism of the model is to consider
massless spring damper system at the solid wall, which allows in-, or outflow
locally. Hence, the disturbances are not reflected from the wall, but damped, which
can help to reach the same convergence criteria by less iterations compared with
mirror or characteristic type solid wall. The normal velocity is expected to be zero
at the wall, hence one outgoing and one ingoing characteristics are considered.
Equation belongs to dn dt  Vn  c characteristics (8) is considered as a numerical
boundary condition. The massless spring-damper system represents the soft solid
wall [11],
Vn Vn  p
  , (10)
t  c t
where k  1  is stiffness and   1  is the damping factor. The optimal
parameter values for the spring-damper system is   0.5 and  t  100 to
minimise the wave reflection. The (9) and (10) are solved for the boundary pressure

–6–
Acta Polytechnica Hungarica

and normal velocity at the new time level. The tangential velocity is kept constant
regarding the first cell adjacent to the wall. The temperature is calculated by
adiabatic Poisson sate equation because the wall is considered to be isentropic over
the time evolution. The number of arithmetic operation is decreased and the
computational time can further be decreased by the presented method. [12]

2.3 Finite Volume Discretization

Finite volume discretization has been applied in the present article, by which the
system of eq. (5) is integrated over a control volume  bounded by interface 
and applying the Gauss divergence theorem gives,
 
t 

 
Ud  Hnd  0 , (11)

 
   
where n  n x , n y is the local outward pointing unit normal, H  Fe x  Ge y and

Hn is given by (12),

 Vn    
   
   uVn  pn x   u 
H n  Hn    and U   
vVn  pn y v
   
 V H   E 
 n   
 u   v 
 2   
 u  p   vu 
with F (U )   and G(U )   2 . (12)
uv  v  p 
   
 uH   vH 
   
In order to pass from continuous to a discrete form, a choice about the type of
representation of the solution vector over the finite volume has to be made.
Replacing the second integral by a summation over the number of faces N f of the
chosen control volume i, j , eq. (11) can be written in the following semi-discrete
form for the point i, j ,
Nf

 H 
1
U i, j   n i , j ,k i , j ,k   in, j , (13)
t  i, j k 1
where H n i, j ,k is the total inviscid flux normal to the cell interface with the length
of i, j ,k cell boundary exchanged between points i, j and k .  i,n j represents the
residual, the scalar value of the surface integral in (11). In case of upwind
differencing schemes, the quantity H n i, j ,k are characterized by a flux function
~ ~

H n , which takes into account the sign of the Jacobian matrices. The H n U L , U R 

–7–
Veress et al. Investigation of Inverse Design Method for Internal Flows

can be evaluated by a linear wave decomposition, in where an unique average state

denoted by a hat, originally developed by Roe, of the left and right states exist [13],
~
 1
      
H n U L , U R  H n U L  H n U R  Dˆ n U L , U R U R  U L ,
2
 (14) 
The method of Yee [5] has been used to satisfy entropy condition. The method of
Roe is highly non-dissipative and closely linked to the concept of characteristic
transport. It is one of the most powerful linear Riemann solvers due to the excellent
discontinuity-capturing property including shear waves. The MUSCL (Monotone
Upstream Schemes for Conservation Laws) approach is used for higher order spatial
extension and MinMod limiter for preserving monotonicity [14]. For the minimum
computational storage and the large stability range with the optimal choice of its
parameters, the 4th order Runge-Kutta method is used to solve the time derivatives
of the conservative variables. The coefficients of the scheme is derived to maximize
the CFL number. In order to optimize the time step behind the stability criterion,
local time stepping has been used. [12]

2.4 Wall Modification Algorithm

While the incoming and outgoing velocity distribution is given at the solid wall,
based on the inverse mode of the analysis, the last step of the iterative design cycle
is the modification of the geometry. The new position of the solid boundary
coordinates are calculated by setting the wall parallel to the local velocity vector:
i
 vk 
y i ( xi )    u
k  Le k
x k  ,

(15)

where u and v are the Cartesian components of the velocity vector. The wall
modification starts from the inlet stagnation point till the outlet stagnation point and
completed in vertical directions (see Fig. 2.). [12]

Figure 2

Schematic view of the wall modification process

–8–
Acta Polytechnica Hungarica

3 Verification of the Direct Solver

As a verification the direct solver’s result of pressure coefficient distribution
evolved from flow over sinusoidal profile is compared to the computational results
of two independent RANS solvers (CFL3D and FUN3D) using Spalart-Allmaras
turbulence model for compressible flow. [15][16]
The numerical boundary conditions are the following: inlet total pressure:
ptot,in=104190.471 [Pa], inlet total temperature: Ttot,in = 302.4 [K], outlet static
pressure: pstat,out = 101325 [Pa] and inlet flow angle: α = 0°. The definition of the
4
  x  
sinusoidal bump is: y  0.05 sin      , for 0,3≤x≤1,2. [15]
  0.9 3  
The channel length is 51.5 m and height is 5 m. The numerical mesh size of the
model is 2060x180 in i and j directions respectively.
The geometry of the flow field and the boundary conditions are shown in Fig.3 and
with streamlines and Mach-number distribution in Fig.4.

Figure 3

Flow field geometry and boundary conditions of the verification case [15]

–9–
Veress et al. Investigation of Inverse Design Method for Internal Flows

Figure 4

Mach-number distrubution of the verification case

The verification is based on the comparison of the pressure coefficient distribution

[16] at the lower wall in which the pressure coefficient of point i is:
pi  p
C p ,i  (where pi is the resulting static pressure in point i,
0.5p M 2
p = 101325 [Pa] is the inlet static pressure ,  is the adiabatic exponent with value
of 1.4, and M  = 0,2 is the Mach-number of the inlet flow). The pressure coefficient
distribution deriving from the flow analysis of the direct solver and the distribution
given by verification case are illustrated in Fig.5

– 10 –
Acta Polytechnica Hungarica

0.4

0.2

0
Cp [-]

-0.2

-0.4

-0.6
verification data
direct solver result
-0.8
0 0.25 0.5 0.75 1 1.25 1.5
x [m]

Figure 5

Pressure distributions of the verification case [16]

The difference between the values from the applied verification case and the
examined solver’s analysis is not significant, therefore the evaluated method is
acceptable in engineering point of view.

4 Test of the Inverse Design Method

The purpose of this part is to gain more information about the limits of the usability
of the inverse design software.
The test cases use similar geometry in a certain sense: the definition of the
sinusoidal profile can be specified as

hmax  x  x start  h
y sin  2       max , (16)
2  x end  x start 2 2

where hmax is the maximum height of the sinusoidal profile, xstart is the start point
and xend is the end point of the profile.

4.1 Maximum Reconstructable Profile

The boundary conditions are the following: inlet total pressure ptot,in=107853.4 [Pa],
inlet total temperature Ttot,in = 293.15 [K], inlet flow angle α = 0° and outlet static
pressure pstat,out = 101325 [Pa] over 100x70 H-type mesh. Consequently, the Mach-
number M=0.3. The channel length is 10 m and height is 7 m. The definition of the
sinusoidal profile:

– 11 –
Veress et al. Investigation of Inverse Design Method for Internal Flows

0 .7  x  3 .5   0 .7
y sin  2      (17)
2  6 .5  3 . 5 2  2

The range of the inverse design is 3.5≤x≤6.5. Maximum height of the sinusoidal
profile is 0.7 m.
At first, direct flow analysis is performed to determine the pressure distribution over
the specified sinusoidal profile. Then this pressure distribution is imposed at a flat
wall.as the target distribution. The iterative cycle (direct analysis – opening – wall
modification) is repeated until the required pressure distribution is reached.
The streamlines and the Mach-number distributions of the target and the result
profiles are shown in Fig.6. Fig. 7 illustrates the pressure distributions of the initial,
the target and the result profiles.

Figure 6

Flow field geometry and Mach-number distribution of the maximum reconstructable profile

– 12 –
Acta Polytechnica Hungarica

110 000

105 000

100 000
pstat [Pa]

95 000

90 000

85 000
initial
result
target
80 000
0 1 2 3 4 5 6 7 8 9
x [m]

Figure 7

Initial, result and target pressure distribution of maximum reconstructable profile

At the end of the design cycles the target and the result pressure distributions equal
to each other. It can be stated that the inverse design software functions in
accordance with the expectations.
For higher sinusoidal profiles the inverse design could not provide the required
result, because the flow is unable to follow up the geometry precisely hence
recirculations appear. Therefore, the profile is not characterized perfectly by the
developing pressure distribution. The same situation occurs over profiles with
sudden change of direction (e.g trapezoidal profile) even in case of a much smaller
profile height. As the flow velocity increases the height of the maximum
reconstructable profile decreases. It can be assumed that inverse design can generate
a profile from the prescribed pressure distribution if that pressure distribution
represents a completely profile-following flow.

4.2 Maximum Smoothable Profile

In order to inspect the assumption, pressure distribution of a flat wall is defined to
reach from extreme high initial profiles.
The boundary conditions are the following: inlet total pressure ptot,in=107853.4 [Pa],
inlet total temperature Ttot,in = 293.15 [K], inlet flow angle α = 0° and outlet static
pressure pstat,out = 101325 [Pa] over 100x70 H-type mesh. Consequently, the Mach-
number M=0.3. The channel length is 10 m long and 7 m high.
5.65  x  3.5   5.65
y sin  2      (18)
2  6 .5  3 .5 2  2

The range of the inverse design is 3.5≤x≤6.5.

– 13 –
Veress et al. Investigation of Inverse Design Method for Internal Flows

The streamlines and the Mach-number distribution of the initial profile are shown
in Fig. 8. Fig.9 illustrates the pressure distributions of the initial, the target and the
result profiles.

Figure 8

Flow field geometry and Mach-number distribution of the maximum smoothable profile

Figure 9

Initial, result and target pressure distribution of maximum smoothable profile

– 14 –
Acta Polytechnica Hungarica

At the end of the design cycles the target and the result pressure distributions equal
to each other even in case of a formerly problematic profile. It can be stated that
the inverse design software functions in accordance with the expectations.
In all cases mesh refining may provide more accurate results.

5 Geometry Optimization
The limited pressure distribution is specified by Stratford’s separation prediction
method according to the following boundary conditions: inlet total pressure
ptot,in=107853.4 [Pa], inlet total temperature Ttot,in = 293.15 [K], inlet flow angle α =
0° and outlet static pressure pstat,out = 101325 [Pa] (equivalent to M=0.3) over
170x70 H-type mesh. To define the pressure distribution, the minimum pressure
coefficient is Cp=-3.278 connoting pstat=80398 [Pa] static pressure value.
The channel length is 17 m and height is 7 m. The definition of the initial geometry:
0 .8  x  3 . 5   0 .8
y sin  2      for 3.5≤x≤6.5 (19/a)
2  6 . 5  3 .5 2  2

y  0 otherwise. (19/b)

The range of the inverse design is 5≤x≤13.5.

The pressure distribution from Stratford’s method is prescribed as target pressure
distribution of the inverse design method. Thus, the flow is able to follow the
obtained geometry as the pressure increases in the decelerating stream.
The initial profile and the optimized profile are shown in Fig.10 and Fig.11
respectively. Their pressure distributions along with the target distribution are in
Fig.12.

– 15 –
Veress et al. Investigation of Inverse Design Method for Internal Flows

Figure 10

Flow field geometry and Mach-number distribution of the initial profile

Figure 11

Flow field geometry and Mach-number distribution of the result profile

– 16 –
Acta Polytechnica Hungarica

110 000

105 000

100 000
pstat [Pa]

95 000

90 000

85 000
initial
result
target
80 000
0 2 4 6 8 10 12 14 16
x [m]

Figure 12

Initial, result and target pressure distribution of diffuser

Conclusions
Recently developed inverse design-based optimization procedure has been verified
and tested for internal flows. A bump-in channel case from international literature
is performed to verify the direct solver. Testing part consists of determining the
maximum reconstructable and smoothable sinusoidal profile. Typically, the
software is able to reproduce the expected profile until recirculation does not occur
over the geometry.
Additional investigation on potetial applications of inverse design method is
implemented in form of diffuser optimization. The requested pressure distribution
is gained from Stratford’s separation prediction method. Flow over result geometry
is without recirculation and follows the profile along the entire designed wall.
Through these investigations more information about the conditions and capabilities
of the inverse design method can contribute to the successful application of the
coupled process of the flow modelling and optimisation.
References
[1] Á. Veress, J. Rohács: Application of Finite Volume Method in Fluid
Dynamics and Inverse Design Based Optimization, Finite Volume Method -
Powerful Means of Engineering Design, ISBN 978-953-51-0445-2, ch.1.,
2012
[2] O. Leonard; R. Van den Braembussche: Subsonic and Transonic Cascade
Design, AGARD-VKI Special Course on Inverse Methods in Airfoil Design
for Aeronautical and Turbomachinery Applications, May, 14-18, 1990,
Belgium

– 17 –
Veress et al. Investigation of Inverse Design Method for Internal Flows

[3] M. Giles; M. Drela: Two-Dimensional Transonic Aerodynamic Design

Method AIAA, Journal 25(9),1987, pp. 1199-1206.
[4] A. Demeulenaere: An Euler/Navier-Stokes Inverse Method for Compressor
and Turbine Blade Design, Von Kármán Institute for Fluid Dynamics,
Inverse Design and Optimisation Methods, Lecture Series, 1997-05, pp. 1-
45.
[5] L. De Vito, R. Van den Braembussche, H. Deconinck, A Novel Two-
dimensional Viscous Inverse Design Method for Turbomachinery Blading,
International Gas Turbine and Aeroengine Congress and Exhibition,
Amsterdam, PAYS-BAS, 03/06/2002, 2003, vol. 125, n2, pp. 310-316.
[6] H. Cabuk, V. Modi: Optimum Plane Diffusers in Laminar Flow, J. Fluid
Mechanics, 237, 1992, pp. 373-393.
[7] Á. Veress, T. Felföldi,T. Gausz,L. Palkovics: Coupled Problem of the
Inverse Design and Constraint Optimization, Applied Mathematics and
Computation, DOI: 10.1016/j.amc.2011.08.110, Paper In Press, Corrected
Proof. 2011
[8] B. S. Stratford: The Prediction of Separation of the Turbulent Boundary
Layer, Journal of Fluid Mechanics, Vol. 5., 1959, USA, pp. 1-16.
[9] A. M. O. Smith: High-Lift Aerodynamics, Journal of Aircraft, Vol. 12 No.
6, USA ,1975, pp. 501-530.
[10] M. Manna, A Three Dimensional High Resolution Compressible Flow
Solver, PhD thesis, Catholic University of Leuven, 1992
[12] Á. Veress, T. Gallina, J. Rohács: Modified Soft Solid Wall Convergence
Acceleration Technique for 2D Euler Equations
[13] P. L. Roe: Approximate Riemann Solvers, Parameter Vectors, and
Difference Schemes, Journal of Computational Physics, Vol. 43, 1981, pp.
357-372.
[14] H. C. Yee: A class of high-resolution explicit and implicit shock-capturing
methods, VKI lecture series 1989-04, March 6-10, 1989; NASA TM-
101088, Feb. 1989.
[15] http://turbmodels.larc.nasa.gov/bump.html (Nov.16, 2013)
[16] http://turbmodels.larc.nasa.gov/bump_sa.html (Nov.16, 2013)

– 18 –