Diffuser and Nozzle Design Optimization by Entropy

Generation Minimization
1. Introduction
Diffusers and nozzles may either be the last part of a flow system, releasing the fluid into the
ambient, or a conduit component within a flow system. These are two very different situations
and as a consequence optimization occurs with respect to different aspects. The main purpose of
using an end diffuser may be a high pressure recovery, while for a nozzle, however, it may be the
homogeneity of the velocity profile, as for example with a nozzle in an open wind tunnel.
When diffusers and nozzles are internal conduit components within a flow system, their purpose
is to increase or decrease the flow cross sections as shown in Figure 1. On a length L the
hydraulic diameter changes from Dhu to Dhd with the indices u and d for upstream and
downstream. Then the otherwise fully developed flow is influenced by the components within an
upstream and downstream length Lu and Ld, respectively.
Figure 1. Basic design of diffusers and nozzles within a flow system; 1 , 2 locations
beyond the influence of the conduit component.

With Dhu, Dhd and L prescribed, the open question is which shape of the wall (different from the
straight walls in Figure 1) will be the best. Here “best” means what can result in the least loss of
total head due to the presence of the diffuser or nozzle.
This is the situation analyzed here: How to find the optimum conduit design by varying the wall
shape of an internal diffuser or nozzle with prescribed values of D hu, Dhd and L for a certain
Reynolds number of the flow. The optimization target is to minimize the head loss coefficient
with a value as low as possible. The main purpose, however, is not only to provide the
information about the best diffuser or nozzle design but also to introduce a method for a
systematic optimization based on the least entropy generation in the flow field.

2. Head Loss, Dissipation and Entropy Generation

Diffusers and nozzles, used as internal flow system components, are characterized by their head
loss coefficient K. This coefficient describes the conversion of mechanical energy to thermal
energy by a dissipation process.
A reasonable definition of K therefore is

(1)
with ϕ as specific dissipation due to the presence of the conduit component in the flow system.
From a thermodynamic point of view dissipation of mechanical energy corresponds to a
generation
of entropy by this irreversible process. Thus K can also be expressed in terms of the specific
overall
entropy generation s including the mean temperature Tm, i.e.,

(2)
Details of how to determine ϕ and s will be given later. First, four general remarks should be
made with
respect to the above Equations (1) and (2).
1. A K-value, defined for all kinds of conduit components like bends, trijunctions but also straight
pipe segments, often is defined in terms of pressure differences that occur along that component.
For example a straight pipe of diameter D and length L is characterized by a friction factor f with

(3)
assuming that the flow is fully developed and therefore dp/dx = Δp/L holds. This definition, as
well as all definitions

(4)
with Δp as “pressure loss” are appropriate only in special situations. We recommend to use the
definition of K according to (1) for all conduit components and as the only version to define a
head loss coefficient, see [1] for more details, as well as the following point.
2. For conduit components with different cross section areas upstream and downstream (like with
diffusers and nozzles, c.f. Figure 1) there should be a careful interpretation with respect to
pressure
differences which are often measured in order to characterize the components. If Δp = p2 − p1 is
measured with p1 and p2, being the pressure in cross sections 1 and 2 in Figure 1, there is
• a change of pressure due to the change of cross section, Δp^12 CS, corresponding to the
difference in kinetic energy in both cross sections.
• a loss of total head due to losses in the flow field, Δp L^12 . This Δp L ^12 can formally be
subdivided into two parts, i.e.,

, Δp L^12,AT is the frictional loss in the adjacent tangents (index: AT) from the conduit component
to the cross section 1 and 2 ,respectively. It is determined under the assumption of a fully
developed and undisturbed flow in the tangents upstream and downstream. Then Δp12 L,CC
corresponds to the loss of total head due to (not in) the conduit component (index: CC). It can be
uniquely linked to the head loss coefficient K. It is the only part in

that is subject to changes when the shape of the diffuser or nozzle is changed (keeping Dhu
and Dhd unchanged), i.e., when an optimization is performed.
This is sketched in Figure 2 where X is a parameter of the geometry that can be altered
during the optimization process. With the definition of ΔpL^12,CC given above this part of
themeasured Δp (corresponding to the head loss coefficient) is independent of the exact location
of 1 and 2 in Figure 1, provided both cross sections are outside the zones of influence
upstream and downstream of the component.
In Figure 2b it has been anticipated that in a nozzle Δp L^12,CC can be negative, i.e., the nozzle
may reduce the head loss in the adjacent tangents compared to the assumed loss due to a fully
developed and undisturbed flow.
Figure 2. Typical overall pressure difference and its interpretation according to (5) for an
optimization with respect to an optimization parameter X (not specified here).

3. Losses from a thermodynamic point of view are best described as losses of exergy, sometimes
called losses of available work (exergy: maximum theoretical work obtainable from the energy
interacting with the environment to equilibrium). In a power cycle, for example, this is especially
important, since lost exergy is no longer available in the turbine of that cycle resulting in a
reduced cycle efficiency, see [2] for a comprehensive discussion of this issue.
Dissipation of mechanical energy, which is directly linked to the head loss coefficient in (2), is
(6)
with T m as mean temperature at which the dissipation process occurs. The specific entropy
generated, s, is related to the corresponding specific exergy loss eL by

(7)
so that the relation between ϕ and eL is

(8)
Here T0 is the environmental temperature which occurs in the so-called Gouy-Stodola theorem.
For these basic thermodynamic relations see standard textbooks like [3–5]. Note that all
specificvalues introduced so far (ϕ, s, eL) are their absolute values referred to the mass or more
convenient for an interpretation their absolute values per time, i.e., rates Φ ˙ , S, ˙ E ˙ L, referred
to the mass flux m ˙
in the flow system.
Since according to (8) it depends on the temperature level Tm how much exergy is lost with a
certain dissipation the head loss coefficient K defined in (1) is not a general measure for the
exergy lost in a conduit component.
Hence, in addition to K a second coefficient called exergy loss coefficient should be introduced.
With KE defined as

(9)
there now are two loss coefficients K and KE related to each other by

(10)
If, for example, the head loss occurs in a power cycle at an upper temperature level Tm, the
temperature ratio may be Tm/T0 = 3. Thus, the exergy loss is only one third of what it would be
if the same flow would have the ambient temperature T0 instead of Tm.
4. During the optimization process described hereafter, K-values of the diffusers and nozzles will
have to be determined. According to (2) this can either be done by the determination of ϕ
(dissipation rate per mass flux) or of s (entropy generation rate per mass flux). We definitely
prefer
s since it is the more fundamental quantity of the conversion process (mechanical → thermal
energy). A further argument in favor of the specific entropy is that then a common quantity exists
when also losses (of exergy) in a superimposed heat transfer process are accounted for, see [1] for
more details of this convective heat transfer situation.
3. Review of Literature
The scope of this paper is to provide a method for the design optimization of conduit components
like diffusers and nozzles, based on the head losses involved. As shown in the previous section
this can be accomplished by determining the entropy generation. Since s is closely related to the
second law of thermodynamics this method is called second law analysis (SLA). A literature
review with respect to this special aspect should, however, be accompanied by references about
entropy in general.
3.1. Literature about Entropy in General
There is a vast amount of textbooks and monographs about thermodynamics which all have major
parts with respect to entropy in it. Special books about entropy range from easy to read
introductions like [6–8] over more comprehensive books like [9] to very challenging ones like
[10,11].
3.2. Literature about Entropy Generation
The SLA approach in which irreversibilities are identified and determined is described and
applied in fundamental studies like [12–14], for example. Almost all studies that incorporate a
second law analysis refer to one of the several important contributions by Adrian Bejan, see
[12,15–17]. In these studies entropy generation is often determined as an integral value within a
finite solution domain, either by global balances or by integration of the local entropy generation
density. A general and systematic comparison of various approaches can be
found in [18,19].
In general, one can analyze a system as a whole or have a closer look to single components within
a complex system. Studies about whole systems like [20–22] aim at improving the system
performance, though there is no systematic optimization strategy involved, like in the special
study [23], for example.
Detailed studies about single components can be found in [24], where a vortex tube is analyzed,
in [25] with a study about a crossflow heat exchanger, or in [26] where rectangular ducts are
investigated, just to mention a few typical investigations out of the big number of studies based
on the SLA approach.
3.3. Literature about Optimization Processes
Optimal shapes for diffusers have been investigated in several publications. Almost every
optimization is based on the so-called pressure recovery factor cp which is basically the pressure
difference between the outlet 2 and the inlet 1 , non-dimensionalized with the product of the
kinetic energy based on
the cross section averaged velocity at the inlet and the density, i.e., the “dynamic pressure”. When
the flow is fully developed at 1 and 2 a diffuser optimized for a maximum cp will also be a
diffuser with a minimal value of K. Thus, designs from other publications can be used to compare
the optimal designs found here. Since, however, cp also includes losses in the tangents and the
difference in kinetic energy, see (5), the sensitivity of cp with respect to changes in the geometry
is rather small. If K is used instead, a higher precision can be expected, since no use of averaged
values, e.g., pressure or kinetic energy, at the inlet or outlet is made and integrating over a volume
will generally be more accurate than the determination of values at two distinct positions.
Detailed information about 2-dimensional diffuser optimization can be found in [27] and [28],
where c p is optimized using the standard k − turbulence model with wall functions and a similar
parameterization as discussed here for the wall shape. In [27] the so-called Response Surface
Approximation of cp based on a relatively small number of model evaluations is used as a
surrogate for repetitive model evaluations during a conventional search strategy. Different from
[27], a genetic algorithm is used in [28] in order to test a large variety of polynomial wall
parameterizations.
An extensive overview about optimization in research and industrial relevant applications using
CFD is given in [29], where various optimization methods are explained together with illustrative
examples. Especially the adjoint method is discussed in detail.

4. Entropy Generation as an Optimization Criterion

For diffusers and nozzles, dissipation and therefore the entropy generation due to this process is
the crucial quantity that should be optimized, i.e., minimized in the situation under consideration.
Before a strategy with respect to this optimization will be discussed in the next section, we first
show in detail how the overall entropy generation due to the components can be determined.

4.1. Determination of the Overall Entropy Generation Rate

As indicated in Figure 1 already, the conduit component has an influence on the upstream and
downstream flow, so that corresponding lengths Lu and Ld have been introduced. Within these
lengths the influence of the component can be felt by the otherwise fully developed and
undisturbed flow. In order to have a clearly defined length of influence upstream and
downstream, related lengths L ˆd and L ˆu are introduced. They are defined as those lengths up to
which 95% of the overall upstream or downstream influence occurs and are explained in more
details later. Altogether the situation sketched in Figure 3 appears: A part of the flow field, Vc,
being the interior of the conduit component and upstream as well as downstream parts of it, Vu
and Vd, have to be analyzed with respect to the (additional) entropy generation due to the
component.
Figure 3. Different parts of the flow field determined or influenced by the conduit
component. Vc: volume of the component itself; Vu: upstream volume affected by the
component; Vd: downstream volume affected by the component This entropy generation happens
locally due to velocity gradients under the action of a finite molecular viscosity μ. Its amount, see
[5] for example, in Cartesian coordinates is:

(11)

This local value (with units W/(3mK)) when integrated over a certain volume of the flow field
results in the overall entropy generation rate S ˙ or its specific value s = S/ ˙ m ˙ in this volume.
Outside of the diffuser or nozzle, however, only the additional entropy generation has to be
determined, i.e., only that parts of S ˙ that do not exist in the fully developed and undisturbed
flow in the adjacent tangents upstream and downstream. In Figure 4 the relevant entropy
generation rate S ˙ = ϕm/T ˙ m with its three parts is shown as the integral over S ˙ ,respectively.
Away from the conduit component asymptotically tends to zero so that a finite value has to be set
when the upstream and downstream lengths of influence should be determined. We therefore
define L ˆu and L ˆd as those lengths of influence within which 95% of the additional
entropy generation occurs. These are characteristic lengths. The integration to determine S ˙
according to Figure 4, however, still has to be performed over the entire length of influence, i.e.,
over Lu > L ˆu and Ld > L ˆd With S ˙ according to (11) S ˙ can only be determined for laminar
flows. When the flow is turbulent a RANS-approach (Reynolds averaged Navier-Stokes) is
appropriate and S ˙ is split into a mean and a fluctuating part:
With the numerical result from the RANS-equations (S ˙ ) according to (13) can be determined,
but not (S ˙ ), for which a turbulence model is required.
A simple model which basically relates (S ˙ ) to the turbulent dissipation rate ε and which can be
justified in the limit of infinite Reynolds numbers, see [30], reads

(15)
Since all RANS turbulence models provide the information about the turbulent dissipation rate ε
in the flow field, (S ˙ ) can be determined and together with (S ˙ ) used to find S ˙ for turbulent
flows.
In our investigations Menter’s k − ω-SST model is chosen due to its ability to integrate the
Navier-Stokes equations in the low Re regime near the wall without using damping corrections
and to adequately predict flow separation, see [31]. Thus, our model for (S ˙) is

(16)
with β = 0.09.
Figure 4. Determination of the overall entropy generation rate due to a conduit component.
Δϕu: additional specific dissipation upstream of the component; ϕc: specific dissipation
in the component; Δϕd: additional specific dissipation downstream of the component.
Vc conduit component upstream channel part Vu downstream channel part Vd

4.2. An Example
As an example and to demonstrate which information can be extracted from the numerical
solutions, the results for a certain nozzle will be discussed prior to incorporating it into an
optimization process. This nozzle with circular cross sections is shown in Figure 5. The Reynolds
number Re = umD/ν is defined with upstream quantities um1 and D1 and is Re1 = um1D1/ν. The
(polynomial) wall shape is
(17)
Four of the five constants ai are determined by

(18) (19)
with
(20)
which together with L can be varied in the subsequent optimization process. The final shape is
fixed, once L is chosen. Here the radius R m halfway in the nozzle is selected to be Rm = 0.375D1
and L is set to L = 2D1.
Figure 5. Nozzle with circular cross sections.
The main result of the second law analysis is the local entropy generation rate S ˙ according
to(12)–(14). With

(21)
the cross sectional entropy generation rate emerges, which is shown in Figure 6 for Re1 = 50000.
Farupstream and downstream these rates are constant due to the fully developed and undisturbed
(turbulent) velocity profiles. Non-dimensionalized with its upstream value S ˙ /S ˙1 continuously
grows from 1 toS ˙ /S ˙1 = 26.83 in this example.
Figure 6. Cross sectional entropy generation rate for the nozzle at Re1 = 50000 with
L = 2D1 and Rm = 0.375D1, dark: loss inside the nozzle, light: gain downstream of
the nozzle.

According to Figure 4 the overall entropy generation rate S ˙ due to the conduit component has
three parts:
• the additional entropy generation upstream, Δϕum/T ˙ m = Δsum ˙
• the entropy generation in the component (nozzle), ϕcm/T ˙ m = scm ˙
• the additional entropy generation downstream, Δϕdm/T ˙ m = Δsdm ˙ , which here is negative
All three parts are determined by integrating S ˙ /S ˙1 according to Figure 6 with respect to x/D1.
Here,the third part is negative since less entropy is generated downstream of the nozzle than it
would be in a fully developed and undisturbed flow right from the beginning of the downstream
tangent. For the head loss coefficient we thus get, c.f. (2),

(22)
for this example. The negative value of K may be unexpected. The reason again is that altogether
there is a lower entropy generation than it would be in the tangents with undisturbed flow. In
Figure 6 that means that the gain outnumbers the loss.
This result can be cross checked by considering the pressure distribution along the flow which
was discussed in detail in Section 2, c.f. (5), applying a Bernoulli equation, which takes into
account all flow losses. It reads, see for example [5],
(23)
In a one-dimensional version the specific kinetic energy would just be ekin = u2 m/2. If, however,
the actual time averaged turbulent velocity distribution u(r) is taken into account, we have with u

(24)

5. Nozzle Optimization
In Section 4.2, the nozzle according to Figure 5 was prepared to be optimized. Numerical details
ofthe CFD solutions will be given in Section 5.1. Since the nozzle length L and the radius Rm at
L/2 are variable, there are two optimization parameters Xi. They should be normalized with
respect to the range in which they can vary. Thus we get

If L is fixed by geometric constraints already, only one parameter X ˆ2 is left. For the more
general
case, however, an optimization procedure has to find the lowest value of the function K = K(X ˆ1,
X ˆ2).
A further extension would increase the number of optimization parameters by introducing more
intermediate radii that can vary like Rm used so far. In Section 6.2 for a diffuser we will fix the
length L and the wall shape by choosing six radii at equidistant positions on the centerline, so that
K = K(X ˆ1, X ˆ2, . . . , X ˆ6) has to be analyzed with respect to its optimum.
The optimization presented here should not only show the capabilities of an optimization
algorithm or
the existence of a global optimum. This has been done in earlier publications of different authors
already, see [27,28] for diffusers. Here, we especially want to highlight the advantages of the
SLA approach using
every available information of entropy generation as a field variable and the opportunities to
interpret the numerical results on a sound physical background.
5.1. Numerical Details
All calculations discussed in this paper are performed using the open source CFD toolkit
OpenFOAM version 1.6, see [32]. The structured grid exploits axial symmetry, which is applied
using OpenFOAM’s wedge boundary conditions at the opposing wedge faces. The boundary
condition implies a transformation leading to the appropriate gradients in Cartesian coordinates.
There is an angle of five
degree with only one cell layer in circumferential direction which reduces the grid size by a factor
of 1/72 compared to a corresponding full grid.
The domain is discretized with 660 steps in axial and 50 steps in radial direction with a
decreasing cell size towards the wall. An initial grid is set up with OpenFOAM’s blockMesh
utility with a constant radius in streamwise direction to be rescaled radially by a Matlab script
during the optimization. This can be achieved easily, since all point coordinates are stored as a
compressed ASCII file. A typical grid for a diffuser is shown in Figure 7.
Figure 7. Radially scaled structured numerical grid

At the inlet, profiles for u, k and ω for fully developed flow are provided. The profiles are gained
from a channel flow solution which is computed using periodic boundary conditions for u, p, k
and ω. Aprescribed pressure gradient is imposed in the direction of the main flow which is
adapted automaticallyto end up with a prescribed Reynolds number. As a pressure boundary
condition a fixed value is set at theexit. The coupling of pressure and velocity is done by the
OpenFOAM solver application simpleFoam,which uses the SIMPLE algorithm and computes,
after an in-house adaption, the integral of (12) in every outer iteration of the coupled solution
procedure. This value is needed to indicate convergence. Convergence is met when the relative
difference of S ˙ between two consecutive iterations is smaller than 1 × 10−7 and the relative
difference of the averaged inlet pressure is smaller than 1 × 10−5 at the same time. Using residuals
as a convergence criterion instead is no option since the correlation between the
residuals and changes in S ˙ can hardly be predicted. For the discretization of the equations,
gradient and interpolation schemes formally of second order are used, i.e., the OpenFOAM
schemes Gauss linear for the convective terms, Gauss linear corrected for the diffusive terms and
Gauss linear for the gradients.
5.2. Nozzle Optimization for the Polynomial Wall Shape, K = K(X ˆ1, X ˆ2)
To investigate how the K-value depends on X ˆ1 and X ˆ2, i.e., on L and Rm, discrete values for
both parameters are provided and all combinations are fed into the CFD-model. The nozzle length
L varies within the interval [D1/4, 4D1] so that X ˆ1 = (L/D1 − 0.25)/3.75.The midlength diameter
Rm varies within [0.3225D1, 0.4865D1] so that X ˆ2 = (Rm/D1 −0.3225)/0.164. For X ˆ1 we take 18
values, for X ˆ2 only 9, all of which equally distributed between 0 and 1 for X ˆ1 and X ˆ2,
respectively. In Figure 8 all 9 nozzle shapes are shown for L/D1 = 1 as an example.
Figure 8. Geometry of the polynomial shaped nozzle, here shown for L/D1 = 1, i.e., X ˆ1 =
0.75/3.75 = 0.2.
 Here optimization will be performed by calculating all 18 × 9 = 162 cases and analyzing the
resultswith respect to the optimal K-values. Calculations were performed on three cores of an
Intel Core 2Quad CPU at 2.83 GHz. Cases with three different lengths L were computed
simultaneously. Values for Rm were chosen in ascending order so that the result for a certain
length L and a radius Rm was a good initialization for the same L but the next Rm. CPU time
needed for computation was about four days per Reynolds number. K-values for the Reynolds
numbers Re = 10000, Re = 20000 and Re = 50000 were investigated.
The resulting K-values in the design space are shown for Re = 20000 in Figure 9. Quite generally
short nozzles have a lower K-value since the volume of the conduit component itself is smaller.
Yet,one could expect that a short nozzle has a stronger influence on the losses downstream. This
however is not the case as can be seen in Figure 10a compared to Figure 10b. In Figure 9b K-
values are shown for various lengths L. Optimal values, from now on called Kopt, for most of the
cases appear forthe Rm-values corresponding to X ˆ2 = 1, i.e., at the edge of the parameter range.
However, some lie within the [0, 1]-range. The overall optimum, marked by the square symbol,
appears for L/D1 = 1,Rm/D1 = 0.446, corresponding to X ˆ1 = 0.2, X ˆ2 = 0.69.
Figure 9. K-values for Re = 20000 in the nozzle design space. (a) Surface plot
K = K(L/D1, Rm/D1); (b) K = K(Rm/D1) for fixed values of L/D1.

The best solutions for Re = 10000, Re = 20000 and Re = 50000 have a very similar shape but they
differ in the K-value numbers. This Re-dependence of Kopt, which will be regarded in the subsequent
discussion prior to K, indicates that the inertia forces, i.e., the convective terms, which are the only
non-linear terms in the Navier-Stokes equations, do not fully dominate the flow behavior at moderate
Reynolds numbers. For each Reynolds number and a fixed value of L there is an optimal half length
radius Rm and thus a Kopt. These values are shown in Figure 11 for all three Reynolds numbers. Since
the Reynolds number is not a free parameter in the optimization process, optimal values are given for all
three cases, again marked by square symbols. They all appear almost for the same L/D1 (or X ˆ1) which
again shows that the overall optimization procedure is basically the same for different Reynolds numbers.
Figure 11. Comparison of nozzles with optimized K for prescribed L.

In order to see how different shapes affect the flow field downstream of the nozzle, the downstream
length Ld introduced in Section 4.1 can be plotted as a function of Rm and L. For Re = 20000 this
is done in Figure 12. As a general trend Ld decreases with increasing nozzle length L. However, this
effect is small and could be neglected in the design of a nozzle. According to Figure 10 most of the
disturbance in the downstream flow field is caused by the impact of the nozzle’s exit at the beginning
of the downstream channel. The K-values presented here are valid as long as the physical length of
the downstream channel is Ld/D1 > 11.5 for Re = 10000, Ld/D1 > 13.5 for Re = 20000 and
Ld/D1 > 14.5 for Re = 50000.
Figure 12. Downstream length of impact Ld for nozzles with Re = 20000.

The polynomial nozzle seems to be adequate due to its smooth shape and its rounded corners. In
order to evaluate the benefit of such a complicated geometry, the results are compared to those with the
most simply shaped nozzle of prescribed length, a straight nozzle. For Re = 20000 this comparison
reveals only a slight advantage of the polynomial nozzle in the case of high values for L, see Figure 13.
Otherwise the straight nozzle is equally good or even superior. It turns out, however, that almost identical
K-values for the straight and polynomial nozzles do not mean the same distribution of losses for both
cases. Figure 14 shows details of how the losses are distributed in both cases, and why they nevertheless
lead to the same K-value.
Figure 13. Optimal nozzles for straight and polynomial wall shapes; Re = 20000.

6. Diffuser Optimization
The examination of various diffuser designs will be started by asking whether a good nozzle
(determined just before) is also a good diffuser. It turns out that the answer is definitely “no” since
the physics in both cases are very different. The crucial aspect is that of flow separation, which in a
nozzle might happen as a local event but in diffusers may determine the flow structure completely.
This becomes obvious when the straight wall cases are compared. An optimal nozzle with straight
walls according to Figure 13 is very short, with L/D1 < 1. The corresponding diagram for a diffuser,
Figure 15, is very different. Good diffusers are very long, thus avoiding massive flow separation. In
Figure 15 the best diffuser is that for L/D1 = 11.6 when L ∈ [4D1, 11.6D1] is set. A further increase in
L/D1 would eventually lead to an increase in K due to the frictional losses within the diffuser. Note that
a diffuser with Re = 40000 corresponds to the nozzle with Re = 20000 when D1/D2 is set to 2 and
1/2, respectively, with the diameter in Re always being that of the conduit entrance. Then there is the
same mass flow rate for both cases. Also, Figure 15 shows that now K-values are positive since kinetic
energies for diffuser geometries cannot outbalance the pressure effects, c.f. Figure 11 for nozzles.
Figure 15. Optimal diffusers for straight wall shapes; Re = 40000.

6.1. Diffuser Optimization for the Polynomial Wall Shape; K(X ˆ1, X ˆ2)
As a first step the optimization is performed like for the nozzle in the previous section, i.e., by
assigning a value to the ratio L/D1 and systematically varying Rm/D1. Here the parameter ranges
are set to

In Figure 16a (c.f. Figure 9 for the nozzle) K in the design space is shown with the optimal diffuser
marked by the square symbol in Figure 16b. For all diffusers the losses can be illustrated by the
cross sectional entropy generation rates inside the diffuser and the additional generation rates outside.
Figure 17a shows these for diffusers with L/D1 = 4, Figure 17b for L/D1 = 11.6. The optimal diffusers
for each length are shown in Figure 18 for three different Reynolds numbers. The best diffusers occur at
the edge of the prescribed range for L/D1. Comparing these results (for Re = 40000) to those of the
straight wall diffusers in Figure 19 it turns out, however, that all straight diffusers are superior to those
with a polynomial shaped wall. It will be investigated next, whether this still holds for a non-straight
wall shape with more than two optimization parameters.
6.2. Diffuser Optimization for a Non-Straight Wall Shape; K(X ˆ1, . . . , X ˆ6)
Since it has been shown before that a single degree of freedom for the optimization might lead to an
unsatisfactory result, the shape for a fixed value of L/D1 is now described via six parameters X1 . . . X6
provided at six equally spaced positions on the diffuser centerline. These parameters are again linearly
scaled with a value X ˆi = 0 corresponding to Ri = D1/2 and X ˆi = 1 corresponding to Ri = D2/2.
However, these bounds were not reached during the optimization by the optimal and close to optimal
shapes. A monotonicity constraint for the wall is not applied as it is done in [27], since it turns out during
the nozzle optimization that a convergent part of the geometry could be beneficial for the reduction of
losses. The length of the diffuser is now fixed at L/D1 = 4 because a strong need for optimization is
given when losses are rather high due to the short length. The diameter ratio again is D2/D1 = 2 as in
the preceding examples.
In order to construct the wall shape from the discrete design variables X ˆi, a so called PCHIP
interpolation is chosen, which is a cubic piecewise interpolation with reduced overshoot for non-smooth
designs compared to splines and is available in Matlab R , see [33]. This is convenient since
Matlab R is used to generate the radially transformed structured grid for the repetitive computations
which are performed using OpenFOAM, see Section 5.1.
We decided to use a genetic algorithm for the systematic variation of the parameters X ˆi, since
this method is, besides its ability to deal with a complex design space, well documented and easy to
understand. For simplicity the implementation of the genetic algorithm from the Matlab Global
Optimization Toolbox is used, which offers a large variety of user settings. The population size
is set to be 12 individuals, i.e., twice the number of the individual geometry parameters, which represent
the “genome” of each individual. At every optimization step the population consists of one “elite child”
from a last iteration, six “crossover children” and five “mutant children”. The crossover children are
based on a linear extrapolation of two parents selected from the preceding generation by a roulette wheel
selection. Here, the child’s genome lies on a straight line, which crosses the genomes of both parents,
at a prescribed distance of 20% away from the better parent in outward direction. This choice seems
reasonable, when a comparably small population is used, since a single point crossover might lead to
many ill-designed diffusers with corrugated walls. The mutant children are gained from one parent
each by replacing each of its six parameters with a fixed probability of 5% by a random value which is
bounded by the upper and lower limit 0 ≤ X ˆi ≤ 1 for this parameter. This method is chosen to introduce
new design candidates at every step of the optimization to prevent premature convergence due to the
smooth crossover designs.
The best shape after 70 steps corresponding to a computational time of four weeks with at most four
individuals computed simultaneously is presented in Figure 20. This design has a K-value of 0.158 while
the K-number for the corresponding straight design is 0.191. Thus the introduction of the additional
parameters is worthwhile since the variation of the shape leads to a reduction of losses and not only to
their redistribution, as shown in Figure 21. In Figure 22 the six scaled coordinates and the corresponding
scaled K-value are shown for the best candidate of each of the 70 optimization steps. Here, K is scaled
according to K ˆ ≡ (K − Kmin)/(Kmax − Kmin) and is bounded between 0 and 1 corresponding to the
lowest and the highest value of K, respectively. Obviously there is a great influence of the wall shape on
the entropy generation in the diffuser which is mainly due to the different turbulent entropy generation
in the free shear layers. These layers can be identified in Figure 23 where the fluctuating part of the
entropy generation is shown for the optimal and the straight geometry. A similar optimal shape has been
found for a 2-dimensional planar case in [27]. It seems also to be valid for the rotational symmetric case
discussed here.
Figure 20. Optimum diffuser design with a PCHIP interpolation of the wall shape with six
radii provided at equidistant positions on the centerline for Re = 100000.
Figure 21. Distribution of the cross sectional entropy generation rates for the optimal and
the straight diffuser; Re = 100000.

Figure 22. Visualization of the investigated design space in parallel coordinates (only elite
individuals from every generation are shown); Re = 100000.