Energy 149 (2018) 875e889

Contents lists available at ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

Design optimization workflow and performance analysis for

contoured endwalls of axial turbines
Hakim T. Kadhim a, b, Aldo Rona a, *
a
Department of Engineering, University of Leicester, LE1 7RH, United Kingdom
b
Al-Dewaniyah Technical Institute, Al-Furat Al-Awsat Technical University, Iraq

a r t i c l e i n f o a b s t r a c t

Article history: Advances in computer-based optimization techniques can be used to enhance the efficiency of energy
Available online 12 February 2018 conversions processes, such as by reducing the aerodynamic loss in thermal power plant turbomachines.
One viable approach for reducing this flow energy loss is by endwall contouring. This paper implements a
Keywords: design optimization workflow for the casing geometry of a 1.5 stage axial turbine, towards mitigating
Flow control secondary flows. Two different parametric casing surface definitions are used in the optimization pro-
Design optimization
cess. The first method is a new nonaxisymmetric casing design using a novel surface definition. The
CFD
second method is an established diffusion design technique. The designs are tested on a three-
Contoured casing
Axial turbine
dimensional axial turbine RANS model. Computer-based optimization of the surface topology is
Kriging demonstrated towards automating the design process. This is implemented using Automated Process and
Optimization Workbench (APOW) software. Kriging is used to accelerate the optimization process. The
optimization and its sensitivity analysis give confidence that a good predictive ability is obtained by the
Kriging surrogate model used in the prototype design process tested in this work. A flow analysis con-
firms the positive impact of the optimized casing groove design on the stage isentropic efficiency
compared to the diffusion design and compared to the benchmark axisymmetric design.
© 2018 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).

1. Introduction series based curves in the streamwise and pitchwise directions to

define profiled endwall shapes. They demonstrated a clear reduc-
The demand for electrical energy is projected to continue rising tion in secondary flows and an increase in turbine efficiency. The
at substantial rates, due to the world's population growth and reduction of secondary flow losses is an active research area in
increased industrial activities. As most electrical energy is currently turbomachinery, as these losses represent approximately 40%e50%
produced in thermal power plants, advances in the design of of the estimated total aerodynamic loss in an axial turbine [4]. A
thermal cycles and of their individual components are required to detailed description of secondary flows and of their loss generation
ensure that this energy supply remains sustainable and affordable. is given in Refs. [5e9].
The axial turbine is a key component a in a thermal power plant. Germain et al. [10] used a combination of a streamwise decay
The overall cycle efficiency is significantly affected by the turbine function and of a pitchwise shape function that, when multiplied
performance [1] and it plays an important role towards meeting the together and scaled, defined contoured endwall surfaces. The
UNFCCC emissions goals. A variety of toolchains is used by axial shaping either accelerates the flow, which decreases the local static
turbine designers, in which the performance and the cost of the pressure, or decelerates the flow, which increases the local static
design are significantly affected by the parametrization and opti- pressure. This way, the endwall cross-passage flow can be reduced
mization stages in the workflow. by altering the pitchwise pressure gradient to reduce the associated
How the hub and the casing geometry is parametrized has a secondary flows [11].
significant effect on the optimization design process. Turgut and Praisner et al. [12] argued that there were disadvantages in
Camci [2] and Harvey et al. [3] adopted beta-spline and Fourier using shape functions to obtain a parametrized contoured endwall.
Simple shape functions, such as a sinusoidal curve, imply a pre-
conceived notion of the resulting endwall geometry. Praisner et al.
* Corresponding author. [12] therefore proceeded to parametrize their geometry based on
E-mail address: ar45@le.ac.uk (A. Rona).

https://doi.org/10.1016/j.energy.2018.02.001
0360-5442/© 2018 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
876 H.T. Kadhim, A. Rona / Energy 149 (2018) 875e889

Nomenclature h Blade span (m)

k Specific turbulent kinetic energy (m2 s
2 )
c Stator blade chord (m) yþ Dimensionless wall distance (
)
f Beta probability density function (
)
Ri Control points (
) Greek symbols
R Specific gas constant (J kg
1 K
1) a Shape factor (
)
w Casing groove width (radians) a1 Stator exit absolute yaw angle (degrees)
g Groove depth (
) b Scale factor (
)
r Radial distance (m) G Gamma probability density function (
)
rh Hub radius (m) h Minimum distance (m)
rt Casing radius (m) u Specific dissipation rate of k (m2 s
3 )
rtg Casing groove radius (m) m Maximum groove depth location (
)
s Parametrized surface (
) x Groove path curve (
)
x Axial coordinate (m)
q Pitchwise angular coordinate (radians) Acronyms
qg Groove pitchwise angle (radians) APOW Automated Process and Optimization Workbench
Pi Casing surface point (
) LE Leading Edge
xi x-coordinate of the ith point (m) NURBS Non-Uniform Rational B-Spline
yi y-coordinate of the ith point (m) PS Pressure Side
zi z-coordinate of the ith point (m) SS Suction Side
TE Trailing Edge
qi q coordinate of the ith point (radians)
UNFCCC United Nations Framework Convention on Climate
Usec Secondary flow velocity (m s
1)
Change
Ur Radial component of velocity (m s
1)

two-dimensional cubic splines in the streamwise and pitchwise 2. The Aachen Turbine test case
directions. This removed some of the restrictions from the pre-
scribed shape of the guide curves used in past work. Walraevens and Gallus [20] provide measurements in a 1.5-
The computer-based optimization of the shape-defining pa- stage axial turbine that are used for establishing a baseline CFD
rameters is a key enabler of performing axial and radial turbine model of the passage flow. These measurements were acquired
designs [13e16] and it is used in the design of non-axisymmetric downstream of the exit plane of the first stator, of the rotor, and of
endwalls. Sun et al. [17] used an aerodynamic optimization tech- the second stator, at the stations labelled 1, 2, and 3 in Fig. 1. These
nique based on combining endwall profiling parametrization and are the same axial planes at which CFD predictions are extracted for
global optimization, lowering secondary flow and profile losses. validation and comparison purposes in Section 3. Both stators and
Tang et al. [18] optimized the endwalls of a one-and-half stage the rotor are untwisted blades. The profile of the second stator, the
high-work axial turbine using a multi-island genetic algorithm, stagger angle, and the number of blades are the same as the first
decreasing the total pressure loss across the first stator by 10.7% stator. The rotor blades are unshrouded and have a tip clearance of
and increasing the stage efficiency by 0.4%. Kim et al. [19] optimized 0.4 mm. The low aspect ratio blading and the constant radius hub
individually a non-axisymmetric shroud and hub of a one-stage and casing design results in strong secondary flows. The second
high-pressure transonic turbine and achieved a 0.4% efficiency gain stator is clocked three degrees in the direction of rotation. The 1.5
by contouring the shroud alone. stage is tested at a mass flow rate of 7 kg/s. The cascade plane
In this paper, an OpenFOAM 3.2 Extend RANS k-u SST model of a schematic of the Aachen Turbine is given in Fig. 1 and key design
1.5-stage turbine is obtained and it is validated against experi- data are listed in Table 1.
mental measurements from RWTH Aachen. ANSYS ICEM CFD 18 is
used as the geometry and mesh generator.
A new casing surface parametrization is then introduced by a
novel surface definition method that draws from observations of
the typical pattern of secondary flows over the casing. This
parametrization uses the Beta function, which is the first applica-
tion of the Beta function for defining the casing surface of a tur-
bomachine. The new casing design technique is focused on
manipulating specific flow structures directly. The ensuing change
affects the surrounding pressure field. The casing design is imple-
mented using Automated Process and Optimization Workbench
(APOW) software, which provides an assessment of the design
sensitivity. It also makes the design process compatible with that of
GE Power.
An analysis of the predicted flow through the 1.5-stage turbine
is presented to verify the effectiveness of the design in mitigating
secondary flow structures and their associated loss. The prospective
Fig. 1. Schematic of cross-sectional view of the Aachen Turbine and detailed geometry
of using this surface definition method for a wider range of appli-
[20].
cations is presented in the conclusions.
 H.T. Kadhim, A. Rona / Energy 149 (2018) 875e889 877

Table 1
Design data of the Aachen Turbine test case.

Parameters Values

First and second stator Rotor

Tip diameter 600 mm 600 mm

Hub diameter 490 mm 490 mm
Passage height, h 55 mm 55 mm
Aspect ratio, h=s 0.887 0.917
Blade number 36 41
Tip clearance e 0.4 mm
Midspan blade pitch, t 47.6 mm 41.8 mm
Inlet flow angle measured from the axial plane 90.0 20.0
Design rotational speed, q_ e 3500 rpm

3. CFD method and model validation is added in Fig. 2 (b) for clarity. To investigate the effect of the
spatial discretization on the flow predictions, a mesh convergence
A three-dimensional (3-D) steady RANS k-u SST model is built of study is carried out using Richardson's extrapolation that is
the Aachen Turbine that is used later on as the baseline for evalu- generalized by Roache [21]. In this work, three meshes are used to
ating the effects of contouring the casing wall. OpenFOAM version discretize the first stator with a constant refinement ratio rm ¼ 2.
3.2 Extend with the steadyCompressibleMRFFoam solver generates These are a coarse mesh of 1.75 M cells (mesh 1), a mesh of inter-
the flow solutions using two mixing planes at the statorerotor mediate spatial refinement of 3.5 M cells (mesh 2), and a fine mesh
interfaces. The computational domain is pitchwise periodic and of 7 M cells (mesh 3). The grid convergence index (GCI) identifies to
only one blade pitch around the annulus is modelled. The stator what extent the predicted flow field approaches its asymptotic
blade is stacked radially at the trailing edge, while the rotor blade is value and therefore gives an assessment of the appropriateness of
stacked radially at its centroid that is located at x ¼ 25.265 mm and the spatial resolution level. In the present work, two GCI are
y ¼ 13.456 mm. The 3-D computational domain is generated using considered, based respectively on the averaged yaw angle and on
ANSYS ICEM CFD 18. The blading profiles are tabulated in Walra- the averaged total pressure loss coefficient. The GCI is calculated
evens and Gallus [20] as 116 (x, y) points for the stator and 133 (x, y) according to Wilcox [22]. The averaged yaw angle predicted with
points for the rotor, in two dimensions. These are mapped to a each of the three meshes is 72.3 , 71.6 , and 71.3 . The corre-
cylindrical reference system ðx; r; qÞ, where x is the axis of rotation, sponding values of the total pressure loss coefficient are 0.0561,
r the radial distance from x, and q ¼ 0 is through the stator blade 0.0570, and 0.0573. Table 2 shows that the GCI computed from the
trailing edge. averaged yaw angle predicted with the coarse mesh and that with
Each passage volume is initially divided into three hexahedral the intermediate mesh is 0.916. This is higher than the GCI
sub-domain blocks, as shown in Fig. 2 (a). ANSYS ICEM CFD 18 is computed from the averaged yaw angle predicted with the inter-
used to discretize the whole computational domain using hex- mediate mesh and that with the fine mesh, which is 0.394. A similar
ahedral unit volumes and to export the mesh as unstructured in the reduction in GCI is obtained for the CGI based on the average total
ANSYS Fluent .msh format. The mesh is then converted to Open- pressure loss coefficient. This indicates a reduction in the mesh
FOAM format for use by the steadyCompressibleMRFFoam solver. dependence of the numerical simulations in which the intermedi-
A simple H-mesh topology defines the first and the second ate and the fine meshes are used. The difference between the
stator blade passages and an O-mesh topology defines the rotor predicted value 4i using the ith mesh and the one obtained from
blade passage. The flow through the rotor tip clearance is modelled Richardson's extrapolation 4R is used to define the relative error
by adding an extra O-type block as shown in Fig. 2 (a). The mesh is as εi ¼ 4i 4
1
R
1.
clustered close to the solid walls to resolve the boundary layers flow
Table 2 shows the relative error εi using the ith mesh, based on
over the blade, the hub, and the casing. This clustering provides a
Richardson's extrapolation. For εi based on either a and Cpts , the
near-wall resolution of yþ z1.
difference in the relative errors between the intermediate mesh
Fig. 2 (b) shows the resulting three-dimensional computational
and the fine mesh is below 1%, therefore the mesh of intermediate
mesh. This covers one blade passage per stage and a second passage
spatial resolution is selected for the current study.
The flow is modelled as dry air, with constant specific heats,
under ideal gas assumptions. The specific heat ratio g ¼ 1:4, the
specific gas constant R ¼ 287 J kg
1 K
1, and the dynamic viscosity
depends on the flow temperature according to Sutherland's law.
The computational domain inlet is located 3.25 axial chords up-
stream of the stator blade leading edge, where the experimental
inlet average total pressure and total temperature are imposed. In
addition, a fully developed inflow compressible boundary layer
profile is imposed at the inlet over the casing. This profile is

Tip clearance
Block
Table 2
Grid Convergence Index with percentage error based on Richardson's extrapolation.

(a) (b) 4 rm GCI21 GCI23 ε1 (%) ε2 (%) ε3 (%)

a1 2 0.916 0.394 1:72 0:73 0:31

Fig. 2. (a) Schematic of the 1.5-stage turbine flow passage, (b) detail of the three-
Cpts 2 0.987 0.327 2.35 0.78 0.26
dimensional computational mesh.
878 H.T. Kadhim, A. Rona / Energy 149 (2018) 875e889

generated by the EDDYBL program of Wilcox [23] and includes outlined by Langston [6] that are often observed over turbine
profiles for k and u. The computational domain outlet is located endwalls, as it is reported in the introduction. Specifically, the
2.54 axial chords downstream of the second stator blade trailing inflow to stator cascade blades features a growing boundary layer
edge, where the radial profile of static pressure measured in on the casing, along the 143 mm long passage leading edge. The
experiment is imposed. The hub, the casing, and the blade surfaces upstream boundary layer bifurcates at the blade leading edge,
are modelled as no-slip adiabatic walls. Pitchwise periodic forming a saddle point. The location of this flow bifurcation is
boundary conditions are imposed over the remaining computa- highlighted by the circle in Fig. 4 (a). Due to the interaction of the
tional domain boundaries. The rotor blades are stationary in the casing boundary layer and the adverse pressure gradient from the
rotor frame of reference, which rotates at the constant rotor shaft blade potential pressure field, a horseshoe vortex is generated near
speed stated in Table 1. The mixing plane formulation by Jasak and the junction of the blade leading edge with the casing. The horse-
Beaudoin [24] is used 10 mm downstream of the first stator and of shoe vortex left (Vsh) and right (Vph) arms bend downstream into
the rotor exit planes. the passage on both pressure and suction sides as shown by ribbons
The flow state is extracted from the CFD predictions at the same in Fig. 4 (b). The Vph bundle of ribbons towards the right edge of
locations as in the experiment, these being 17 radial lines, with 38 Fig. 4 (b) shows that the pressure side arm of the horseshoe vortex
points on each line. Fig. 3 shows the comparison between moves across towards the suction side, under the influence of the
measured and calculated flow velocities 8.8 mm downstream of the passage pitchwise pressure gradient. It merges with the suction
first stator (plane 1) and of the rotor (plane 2). Values are pitch- side bundle of ribbons Vsh at around 0.55 axial chords and creating
averaged using a simple average. The CFD predictions appear to a larger vortex structure, the passage vortex, as described by
have captured the main trends in the experimental profiles. This Langston [6]. The next section introduces a non-axisymmetric
indicates that the CFD model is adequate for guiding the casing design of the stator casing to reduce the adverse effect of these
design optimization, which is the main purpose of this study. secondary flow features.
Fig. 4 (a) shows a flow visualisation over the axisymmetric A 4760-core High Performance Computer (HPC) cluster at the
casing through the first stator by near-surface limit streamlines. University of Leicester was used for the CFD simulations, by domain
This visualisation indicates the presence of the flow structures decomposition. The Message Passing Interface (MPI) version of

Fig. 3. Measured and calculated absolute (C1), tangential (Cw1), and axial (Ca1) velocity components, pitchwise averaged (a) 8.8 mm downstream of the first stator exit plane, and
(b) 8.8 mm downstream of the rotor exit plane.

Fig. 4. (a) Flow visualisation over the stator 1 axisymmetric casing showing the separation of the oncoming casing boundary layer (saddle point) by the streamlines; (b) The same
surface showing by the ribbons the pressure side horseshoe vortex interacting with its suction side branch.
 H.T. Kadhim, A. Rona / Energy 149 (2018) 875e889 879

OpenFOAM 3.2 Extend was used as the flow solver.

4. Contoured upstream stator casing design

It is of interest to explore a computer-based design process for

axial turbines in which the casing geometry is optimized using
modest user intervention within an acceptable timescale. To this
end, the workflow of Fig. 5 is adopted and implemented by the
Automated Process and Optimization Workbench (APOW) soft-
ware. The casing is contoured by applying a groove to it, while the
hub is kept axisymmetric.
Three surfaces are modelled mathematically to represent the
first stator casing delimiting one flow passage. These surfaces are
referred to as the blade to blade passage surface, the extended inlet
surface, and the extended outlet surface. 65 points from the RWTH
Aachen dataset define the blade pressure side and 47 points define
the blade suction side. These points Pi ¼ ðxi ; rt cos qi ; rt sin qi Þ are
projected on the casing cascade plane in MATLAB by the

 
function Pi ¼ xi ; p2
arctan yzii .
The projected points are interpolated using two smoothing cu- Fig. 6. Casing surface rendered as a three sets of sðu; vÞ.
bic splines by the MATLAB functions csaps and ppval. This gives the
cylindrical casing surface shown in Fig. 6 defined as the generalized
surface sðu; vÞ ¼ ½u; rt cos ðvÞ; rt sin ðvÞ . sðu; vÞ is re-stated in MAT-
LAB as a non-uniform rational B-spline (NURBS) surface. For this
Groove path
purpose, two open uniform knot vectors are used. The number of
knots m þ 1 and of the control points n þ 1 on the NURBS are
related as m ¼ n þ p, where p is order of the NURBS surface which is
restricted to 2 in ANSYS ICEM CFD. The internal knot vector values Flow direction
are equally spaced and identified using the parametrization weight
according to Piegl and Tiller [25]. Three NURBS surfaces are
generated and exported as IGES format files, which is a supported
input file format of ANSYS ICEM CFD. By using NURBS surfaces, the
cylindrical casing is represented as an exact geometry.
The next step is to define a grooved casing parametric surface.
The path of the groove shown in Fig. 7 is defined based on the
profiles of the turbine stator blades, inflated in the cascade plane. Mixing plane 1
This is achieved by offsetting the blade perimeter by distance a,
normal to the blade perimeter as Di ¼ Pi þ ni a, where ni is the
normal vector to the casing blade perimeter, defined based on the
secant between the lines defined by (Pi
1 ; Pi ) and (Pi ; Piþ1 ). The Fig. 7. Stator blade profiles ‘inflated’ in the annular cascade casing plane (gray open
circles) and interpolated groove path (black filled circles).
normal distance a is a chordwise linear interpolation between two
extrema (a1 , a2 ), where a1 ¼ 2.22 mm is located normal to the
blade leading edge and a2 ¼ 10 mm is located normal to the blade respectively the inflated blade pressure side and the inflated blade
trailing edge. Two subsets of Di are then interpolated using one a 
suction side. The groove path x x; qg is then defined by a linear
smoothing spline each, dp ðx; qÞ and ds ðx; qÞ, on r ¼ rt to represent

Create Execute Execute Execute

Start OPENFoam solver
MATLAB script MATLAB ANSYS ICEM

Create .rpl file Read output file

Copy baseline
ANSYS ICEM of and
geometry

End

Fig. 5. CFD simulation workflow, automated by APOW.

880 H.T. Kadhim, A. Rona / Energy 149 (2018) 875e889

a
interpolation between the inflated profiles as x x; qg ¼ cx ds ðx; qÞ þ extended inlet surface and the extended outlet surface as shown in
ð1
cx Þdp ðx; qÞ, where cx is the axial chord fraction of the inflated Fig. 10 and exported as an IGES file to ANSYS ICEM CFD.
blade profile. For each point Pðx; qÞ on the casing surface r ¼ rt , its
normal distance to the groove path is hðx; qÞ ¼ k xa ðxÞ
Pðx; qÞ k. 5. Contoured casing shape optimization
 2
The casing groove depth gðd; hÞ ¼ ho d
4 h2
d2 , where d is a The contoured casing definition of the previous section uses a
set angle in radians from the curve path. The groove width w ¼ 2d small number of free parameters to make the optimization process
   
varies along the groove path xa x; qg as w x; qg ¼ cx wte þ by APOW more treatable. Three main design variables are used for
ð1
cx Þwle , where wle is the user-defined groove leading edge the optimization process: the maximum groove depth location
 
width and wte is the user-defined groove trailing edge width. The along the groove path m x; qg , the groove width at the blade
groove path starts from upstream of the leading edge and ends at  
leading edge wle x; qg , and the groove width at the blade trailing
the first stator to rotor mixing plane. The mid-width groove depth  
  edge wte x; qg . The groove width at the blade leading edge wle was
ho ðxa Þ along the groove path xa x; qg is defined by the Beta dis- found to have a small effect on the total pressure loss [27], there-
tribution function (c.f. Devore [26]) f ðcx ; a; bÞ ¼ fore, this parameter is optimized segregated and last. The total
Gða þ bÞG
1 ðbÞG
1 ðaÞcax
1 ð1
cx Þb
1 ; 0  cx  1. The maximum pressure loss coefficient Cpt is used as two objective functions. The
groove depth is 3 mm located at x ¼ m, where m ¼ aða þ bÞ .
a
1 first objective function is the stator row total pressure loss
From this, the first stator casing radius is defined as coefficient

rt ; j hj > d
rtg ðx; qÞ ¼ and the stator 1 casing Pto
Pt1
rt þ g½d; hðx; qÞ ; jhj  d
Cpts ¼ (1)
surface is defined as sðu; vÞ ¼ u; rtg ðu; vÞcosðvÞ; rtg ðu; vÞsinðvÞ as Pt1
P1
shown in Fig. 8.
The second objective function is the stage total pressure loss
The parametric diffusion design method for turbine endwalls
coefficient
reported in Sun et al. [17] was applied to the Aachen Turbine so that
the performance of the turbine with the casing designed by this
Pto
Pt2
method could be compared to that from the casing with the opti- Cptr ¼ (2)
mized groove. The optimization of the parameters in the diffusion Pt2
P2
design method is presented in section five. This design uses a
where subscript 0 denotes the stator inlet plane, subscript 1 de-
cosinusoidal curve of half period in the circumferential direction as
notes the stator exit plane, subscript 2 denotes the rotor exit plane,
shown in Fig. 9 (a). The streamwise curve of Fig. 9 (b) defines the
(e) denotes a pitchwise averaged quantity, and (═) denotes a
maximum amplitude of the circumferential curve on different axial
pitchwise and spanwise averaged quantity. The contoured casing
planes. This streamwise curve is defined as CðR0 ; R1 ; …Rm ; Ytran Þ
P surfaces generated in MATLAB and imported as NURBS in ANSYS
¼ ki¼0 Bi;k ðuÞRi þ ð0; Ytran Þ, which is a non-uniform B-Spline with 7 ICEM CFD delimit the top of the computational domain of the first
control points Ri that are design variables. Bi;k is the B-Spline basis, stator. This is discretized in ANSYS ICEM CFD by maintaining similar
which is defined as meshing parameters to obtain the same mesh quality as for the
validation test case. The unstructured mesh is then converted to the
( OpenFOAM format. The same boundary conditions are applied as
1;ui uuiþ1
Bi;0 ðuÞ¼ for the validation test case and the numerical solution is iterated by
0;u;½ui ;uiþ1  OpenFOAM to convergence, as assessed by the reduction in the
u
ui u
u residuals by 10
5 with respect to their values at the start of the
Bi;k ðuÞ¼ B ðuÞþ iþkþ1 B ðuÞ;k1
uiþk
ui i;k
1 uiþkþ1
uiþ1 iþ1;k
1 computation.
The workflow of Fig. 5 is executed in batch mode for different
 
u is the node vector and Ytran is the transition distance in the values of wðx; qg Þ and m x; qg using APOW. This automation enables
circumferential direction. The B-Spline in the axial direction u using the workflow of Fig. 5 as part of a computer-aided optimi-
combines with the cosinusoidal curve in the circumferential di- zation workflow, which is shown in Fig. 11. The APOW Design of
rection v to generate the parametric surface sðu; vÞ. Experiments (DOE) is used to generate the design space, based on
Each sðu; vÞ is then re-stated as a NURBS surface, mated with the the Optimal Latin Hypercube (OLH) design technique. The

Fig. 8. Casing surfaces rendered as a set of three sðu; vÞ: (a) Surface with a guide groove, (b) surface shape driven by controlled diffusion as in Sun et al. [17].
 H.T. Kadhim, A. Rona / Energy 149 (2018) 875e889 881

(a) (b)
Fig. 9. The distribution used to generate the casing diffusion surface in the (a) pitchwise and (b) streamwise directions.

(a) (b)

Fig. 10. Non-axisymmetric casing NURBS imported in ANSYS ICEM CFD: (a) Groove design, and (b) diffusion design.

optimization analysis of the DOE is performed on a surrogate the next section. The optimum values of the design variables cor-
model. For this purpose, Kriging is selected as the surrogate model responding to the minimum value of Cpts shown in Figs. 12 and 13
type. The selection of the optimized design, for both the groove were noted. It will be shown by the end of this see section that
casing and the optimized diffusion casing, is based on the design these optimum values differ if the penalty function Cptr is used
parameter combinations that give the minimum stage total pres- instead.
sure loss coefficient in the Kriging model built from the CFD The optimization of the design variables is now considered us-
predictions. ing Cptr as the penalty function, since this relates more directly to
Fig. 12 shows the response function of the optimization method, the full stage performance and hence to the turbine efficiency. The
stated in terms of the stator row total pressure loss coefficient. The same parameter space sampling of 10 points is used as for Fig. 12.
changes in the response function are studied for different values of Fig. 14 shows there are only two sampling points in the region
 
the groove width at the blade trailing edge wte and of the axial around the optimal design point popt , using Cptr . This is not
location of the maximum groove depth m. Kriging provides a conducive to obtaining accurate predictions from the Kriging
continuous response function from a discrete sample of computa- model in this region. To examine and enhance the accuracy of the
tional fluid dynamics estimates of the total pressure loss Cpts . It Kriging model, a local adaptive sampling of 10 points was added to
enables to develop an insight of how the design parameters wte and the initial set of 10 points. The adaptive sample set is also generated
m influence the performance of the axial turbine across the full using the Optimal Latin Hypercube method. The extent of the re-
extent of the parameter space and to identify regions of high per- sampled region is constrained to include (to be tangent to) the
formance, with acceptable sensitivity. From the Kriging results, a 10% stage total pressure loss contour shown in Fig. 14. This smaller
region of low Cpts is identified over the range 0.025 radians  wte  sample space region includes the optimal design point ðpopt Þ from
0.05 radians and 0.4  m  0.8. It is of interest to explore the flow the initial set.
features characterising this parameter space, which are explored in The optimization process with the initial and adaptive sampling
882 H.T. Kadhim, A. Rona / Energy 149 (2018) 875e889

value from Fig. 15 is shown in Fig. 14 to be slightly different than

that from Fig. 13.
Fig. 16 (a) shows the predicted and the observed values of the
response function Cptr with a linear fit. Each prediction is obtained
by generating a Kriging model using 9 out of 10 points from the set
used in Fig. 13 and by computing the Cptr at the tenth point. The
predicted value is Cptr as determined from this Kriging model and
the observed value is Cptr as determined from CFD. This test is
designed to explore the sensitivity of the Kriging model on the
sampling of the ðwte ; mÞ parameter space. Fig. 16 (b) shows that the
linear fit lies very close to the bisector of the first quadrant, which
indicates that the Kriging model estimates consistently values close
to the ones from the CFD. The 50% confidence interval band,
determined from the norm of the regressed residuals, is shown by
the dashed lines. This band is shown to be reasonably narrow
 
around the Cptr at popt and is 0.22% of Cptr . With the adaptive
sampling procedure, Fig. 16 (b) shows that, by repeating the test
with 19 out of 20 points, the 50% confidence interval band width
reduces to 0.17% of Cptr . It is concluded that the dependence of Cptr
predicted by Kriging on the specific selection of the points that
inform the Kriging model is slightly reduced with 20 points and
there appears to be little scope for increasing the number of the
sampling points beyond 20 to seek further large improvements in
this agreement.
Table 3 presents the numerical values of Cptr corresponding to
Fig. 16 at ðpopt Þ. Table 3 reports a progressive reduction in the stage
pressure loss coefficient from the baseline configuration, using the
simple 10-point sample and then the 20-point adaptive sample
procedure. Adaptive sampling predicts a 3.5% further reduction in
Cptr . These predictions, from the Kriging models, are verified by
running CFD simulations at the ðwte ; mÞ location of ðpopt Þ. These
simulations predict Cptr values very close to the ones from the
Kriging models. The difference with the 10-point sample model is
0.2125%, which is arguably too small to be verified in experiment.
With a 20-point adaptive sample, the difference is 0.00276% and
therefore essentially immaterial. The percentage differences be-
tween the optimal stage total pressure loss prediction from Kriging
and from CFD for both sets confirm a good model data fit. More
details on DOE surrogate modelling techniques can be found in
Queipo et al. [28].
Fig. 11. Optimization workflow, implemented in APOW. From Fig. 16 and Table 3, the adaptive sampling has shown to be
able to improve the selection of ðpopt Þ over the initial design of 10
points, but the performance gain is rather small (3.5%). This
points (20 points) is shown in Fig. 15. The gradient in the response improvement is arguably worth the additional computational cost
function Cptr in the region that includes the set of points from the of the 20-point sample, which is about double of that from the 10-
adaptive stage is increased. The ðwte ; mÞ location of the optimal point sample. Therefore, for the remainder of this section, the

Fig. 12. Optimization of the casing groove using 10 sampling points based on the stator row total pressure loss coefficient.
 H.T. Kadhim, A. Rona / Energy 149 (2018) 875e889 883

Fig. 13. Optimization of the casing groove using 10 sampling points based on the stage total pressure loss coefficient.

This result seems to confirm that m and wte are the leading pa-
rameters in the design optimization process of the guide groove
casing, with wle being a comparatively less influential design
variable.
Fig. 18 (a) shows the response function Cptr obtained from
Kriging by varying R3 and R5 in Fig. 9. These points define axially
the contoured casing, by the B-spline curve of Fig. 9. The lowest Cptr
is shown in Fig. 18 (a) to be given with R3 and R5 set at approxi-
mately the same values of þ3 mm. This generates a hump close to
the blade suction side that mitigates the circumferential pressure
gradient over the turbine casing passage. A similar mitigation was
identified in Sun et al. [17] from the static pressure coefficient
distribution. Fig. 18 (b) reports the accuracy of the Kriging model for
this regression, using the same technique as that for Fig. 16. Close to
the lowest value of Cptr , the 50% confidence band appears to be
sufficiently narrow to give confidence in the predicted minimum
value of Cptr from the Kriging model used for optimizing R3 and R5 .
Each CFD solution used in the optimization process typically
took 10 h of computer wall time on 30 HPC cores. The wall time
used for obtaining each response function was about 200 h for the
groove design and 100 h for the diffusion design, as the CFD solu-
tions were run sequentially, due to the available computational
Fig. 14. Initial and adaptive sampling using the Optimal Latin Hypercube.
resources. Running the CFD solutions in parallel would have
reduced the wall time.
simpler 10-point sampling is used for optimizing the diffusion
surface casing design parameters.
6. CFD flow analysis
A feasibility study based on trial and error by Kadhim et al. [27]
indicated that larger beneficial changes to the stage average total
The validated simulation of the Aachen Turbine flow passage
pressure loss can be achieved by varying m and wte than by
from section three is used as the baseline for studying the effects of
changing wle , which motivated the optimization of the contoured
contouring the casing, using Tecplot 2017 as the CFD post-
endwall with respect to m and wte . It is now of interest to explore
processor. Fig. 19 (a) shows radial distributions of the pitch-
whether further reductions in the stage total pressure loss can be
averaged yaw angle a1 at the upstream stator exit plane, from the
obtained by tuning the groove width at the leading edge wle . The
mid-span to the casing. Corresponding distributions of stage total
casing with the guide groove optimized for m and wte using the 20-
pressure loss coefficient Cptr are shown in Fig. 19 (b). The a1 and Cptr
point samples is taken as the reference geometry. Five different
profiles are shown for the baseline, the optimized groove, and the
values of leading edge groove width wle are examined for this ge-
diffusion casing simulations. The yaw angle distribution in Fig. 19
ometry. Fig. 17 shows the stage total pressure loss coefficients
(a) above 0.9 of the blade span is mostly above the 70+ design
predicted by CFD with the five different leading edge groove
point for all three casing types. This results in an overturning of the
widths. The leading edge groove width used in the optimization of
flow near the casing. This is due to the cross-flow pressure gradient
Fig. 15 delivers the lowest stage pressure loss coefficient among the
on the casing boundary layer, which more easily turns the less
five CFD simulations. It can be seen that wle has a smaller effect on
energetic near-wall flow compared to the mid-passage flow. The
the stage performance compared to the other two variables used in
contribution to the stage total pressure loss coefficient from the
the optimization process. Fig. 17 shows that the maximum change
mid-span up to 0.9 blade span is predicted to be lowest by the
in the value of Cptr over the five CFD simulations is 0.0087. This is
diffusion casing design. However, the optimized groove casing
quite small compared to the change in Cptr obtained by varying m
design is shown to be the most effective in reducing the flow over-
and wte in the optimization process, as shown in Fig. 15. turning close to the casing, resulting in the most spanwise uniform
884 H.T. Kadhim, A. Rona / Energy 149 (2018) 875e889

(a) (b)
Fig. 15. Optimization of the casing groove using 20 sampling points, based on the stage total pressure loss coefficient. 10 points are from the adaption procedure.

Fig. 16. Observed and predicted values of the stage total pressure loss coefficient.

Table 3
Response function percentage improvement and error in the reduced order models (Kriging).

Response function Baseline Optimal_1 Optimal_2 % Improvement % Error_1 % Error_2

Cptr 11.6205 11.4139 11.4067 3.5 0.2125 0.00276

yaw angle among the three designs. Further away from the casing, Fig. 19 (b). At mid-span, a1 with the diffusion casing design is
between 50% and 70% of the blade height, Fig. 19 (a) shows that the predicted to be below the design turning angle of 70 and it is lower
a1 radial distributions from the axisymmetric and the optimized than a1 from the other two test cases. The under-turning of the flow
groove casing configurations are similar to one another, whereas at 50% blade span explains why the diffusion casing design reduces
the diffusion casing design gives a lower turning. This results in a Cptr at mid-span to a greater extent than the guide groove casing
change in the inlet angle to the rotor that in turns most probably design, as shown in Fig. 19 (b). The radial average of Cptr shown in
reduces the specific work output. The lower work rate extraction Fig. 19 is the objective function evaluated for the axisymmetric and
from the working fluid reduces the main flow total pressure drop for the contoured casing, as predicted by CFD. The reduction in the
across the rotor. This partially explains the lower Cptr predicted radial average of Cptr between the axisymmetric and the optimized
between 50% and 80% of the blade span with the diffusion casing guide groove casing design is 1.89%. The corresponding reduction
design compared to the other two configurations, as shown in between the axisymmetric and the optimized diffusion casing
 H.T. Kadhim, A. Rona / Energy 149 (2018) 875e889 885

shown in Fig. 20 to reduce the secondary kinetic energy coefficient

across the span, compared to both the baseline and the optimized
diffusion designs.
The stage isentropic efficiency is defined as

1
TTt2t0
hstage ¼ g
1 (4)
Pt2
1
Pt0
g

where Tt is the average total temperature and Pt is the average total

pressure across specific axial planes. The subscripts 0 and 2 denote
the axial planes identified in Fig. 1 that are used in Equation (4) to
evaluate hstage . These changes in the coefficient of secondary kinetic
energy were found to mirror the changes in the isentropic stage
efficiency of the turbine. Specifically, the optimized groove casing
was predicted to increase the stage isentropic efficiency by 1.13%
compared to the baseline axisymmetric casing. The isentropic stage
efficiency of the turbine using the diffusion design method was
predicted to marginally decrease, by 0.05%. This appears to point to
the importance of the flow over-turning highlighted in Fig. 19 (a)
Fig. 17. Effect of the groove pitchwise width at the leading edge on the stage average close to the casing, which the optimized groove casing design ap-
total pressure loss coefficient. pears to be able to mitigate whereas the diffusion design casing
appears not to mitigate.
A further insight into the effects of contouring the stator casing
design is 3.29%. wall on the passage flow is provided by Fig. 21. This figure compares
The coefficient of secondary kinetic energy Cske was also eval-
the contours of normalized axial vorticity ux ¼ ux Uups
1 c with
uated 8.8 mm downstream of the rotor exit plane and it is shown in
axisymmetric and with contoured casings, 8.8 mm downstream of
Fig. 20. Cske is defined as in Ingram et al. [11] as
the first stator exit plane. Fig. 21 shows that the optimized groove
casing design reduces the positive axial vorticity magnitude (þux )
2 þ U2
Usec
Cske ¼ r
(3) located at y ¼
0.046 m and z ¼ 0.295 m over the casing wall,
2
Uups which is indicative of a lower strain rate in the casing boundary
layer. This would be consistent with a reduction of the passage
where Uups is the reference stage inlet velocity, Usec ¼ Usinða
cross-flow, from the pressure side to the suction side, which sees a
amid Þ, and Ur ¼ U sin ðbÞ. amid is the rotor exit flow angle evaluated longer and more tortuous path over the optimized groove con-
at the blade mid-span in the stationary frame of reference and b is toured casing than over either the axisymmetric casing or the
the rotor exit flow angle in the moving (rotor) frame of reference. diffusion designed casing. The groove contoured casing design also
In Fig. 20, the region of high Cske between 65% and 85% span appears to mitigate and raise the minimum axial vorticity (
ux )
shows the presence of a rotor casing passage vortex whereas the associated to the casing passage vortex at y ¼
0.018 m and
region of high Cske between 90% and 98% span identifies the pres- z ¼ 0.286 m. The diffusion casing design shows a reduction in the
ence of tip leakage. An increase in the secondary kinetic energy size and peak magnitude of (þux Þ of an area of high positive axial
coefficient is predicted by using the optimized diffusion design at vorticity centered at y ¼ - 0.0125 m and z ¼ 0.294 m compared to
these spanwise locations compared to the baseline case. The Cske the baseline axisymmetric casing. This flow area appears to benefit
remains higher than the baseline across the remainder of the blade less from the groove contoured casing design, which appears to
span shown in Fig. 20. Conversely, the optimized groove design is mitigate the peak magnitude (þux Þ of but not the size of this region.

(a) (b)
Fig. 18. Optimization of the casing diffusion: (a) Kriging model response, (b) model quality indicator.
886 H.T. Kadhim, A. Rona / Energy 149 (2018) 875e889

Fig. 20. Predicted secondary kinetic energy coefficient pitch-averaged along the
normalized span at 8.8 mm behind the rotor.

Fig. 21. As a flow structure with secondary motion moves from the
stator to the rotor, it is typically cut through the rotating blades. In
the simulation, this process is modelled by the conversion of axial
vorticity into entropy through the mixing planes. Mitigating the
secondary flow structures would lead to a reduction in the turbu-
lent kinetic energy and this offers a plausible explanation for the
trends shown in Figs. 21 and 22.
Fig. 23 shows by contours the static pressure distribution on the
casing predicted with the three different endwall geometries of
Figs. 6 and 8. As the flow expands through the passage, Fig. 23
shows that the static pressure decreases to a minimum value
where indicated by the solid black arrows and then increases to-
wards the passage exit. This area of flow diffusion generates an
unwanted adverse pressure gradient that makes the flow over the
blade suction surface more prone to separate. Fig. 23 shows that the
suction side boundary layer sees almost the same exit pressure, as
indicated by the similar placement of the contour level indicated by
the dashed arrows in Fig. 23 (a), (b), and (c). Fig. 23 (c) shows that
the optimized groove casing generates a static pressure minimum
Fig. 19. Predicted pitch-averaged radial distributions of (a) yaw angle 8.8 mm down- of lower magnitude located further upstream compared to the
stream of the first stator exit plane, and of (b) stage total pressure loss coefficient axisymmetric design and the diffusion casing design. This reduces
8.8 mm downstream of the rotor exit plane. the adverse pressure gradient over the blade suction side, miti-
gating the secondary flow structure growth. The diffusion design is
shown in Fig. 23 (b) to broaden the area of low static pressure over
It is of interest to study how the normalized turbulent kinetic the blade suction side compared to the axisymmetric casing. The
energy changes through the stage as the casing is modified from the static pressure minimum occurs further downstream, which re-
baseline axisymmetric design to a contoured casing. To this end, the duces the blade surface over which the flow diffuses. This is a
normalized turbulent kinetic energy distribution behind the mix- positive feature, however, proximal to the suction side trailing
ing plane 1 and close to the rotor blade leading edge was deter- edge, the flow undergoes a more rapid diffusion to meet the
mined from the CFD results. The contours of normalized turbulent pressure equilibrium condition from the pressure side flow. While
kinetic energy of Fig. 22 show that the contouring affects the tur- this latter diffusion is undesirable, Fig. 21 (b) shows that the
bulent kinetic energy approaching the rotor close to the casing. resulting stator outflow streamwise vorticity is still lower and
Specifically, there is a slight difference in the turbulent kinetic therefore better than with an axisymmetric casing.
energy close to the casing between the axisymmetric casing and To gain a further understanding of the rotor tip leakage flow
the diffusion designed casing. A more pronounced change in the with three different first stator casing designs, the static pressure
normalized turbulent kinetic energy distribution is obtained with distribution over the rotor blade tip is considered [29]. The
the optimized groove casing design, by which the region of axisymmetric casing and the optimized diffusion casing designs
elevated turbulent kinetic energy near the casing is reduced in display similar normalized distributions of static pressure
radial size. This reduction is consistent with the reduction in the p ¼ 2pr
1
2
ups Uups in Fig. 24, which are characterized by a pressure
axial vorticity upstream of the mixing plane 1 that was shown in minimum well localized at about 70% axial chords from the blade
 H.T. Kadhim, A. Rona / Energy 149 (2018) 875e889 887

(a)
(a)

(b)

(b)

(c)
Fig. 21. Contours of normalized axial vorticity 8.8 mm downstream of the first stator
exit plane. (a) Axisymmetric casing, (b) diffusion casing, and (c) optimized groove
casing. (c)
Fig. 22. Contours of normalized turbulent kinetic energy downstream of mixing plane
leading edge. The optimized groove design has a less localized 1. (a) Axisymmetric casing, (b) diffusion casing, and (c) optimized groove casing.
pressure minimum where the pressure is higher than that of the
other two cases. This is indicative of a reduction in the pressure
gradient over the blade tip, from the pressure side to the suction 7. Conclusions
side, which in turn may reduce the tip leakage strength, as also
indicated by the change in the radial distribution of the secondary This paper presented the proof of concept of an improved pro-
kinetic energy coefficient shown in Fig. 20. cess for the design of turbines with a contoured casing. This process
includes attractive advances over more established turbine casing
888 H.T. Kadhim, A. Rona / Energy 149 (2018) 875e889

(a) (b) (c)

Fig. 23. Contours of static pressure in the cascade plane close to the casing. (a) Axisymmetric casing, (b) diffusion casing, and (c) optimized groove casing.

The new process, implemented using Automated Process and

Optimization Workbench (APOW) software, was shown to be able
to take advantage of this casing surface definition in its optimiza-
tion design loop. Specifically, the casing surface parametrization
(a) has both a lower set of parameters and produces topologically
smooth interfaces with the rest of the passage geometry. The re-
sults from the optimization and from its sensitivity analysis give
confidence that a good predictive ability is obtained by the Kriging
surrogate model used in the design process. The low cost overhead
of the model and its robustness are attractive for accelerating the
design iterations used in industry for turbomachines.
The new design process is shown by steady multi-row 3D RANS
k
u SST modelling to produce aerodynamic performance gains
with respect to the equivalent turbine stage with either an
axisymmetric casing or with a non-axisymmetric casing designed
(b) by a more established controlled diffusion method.
This computer-driven process appeals to the industrial design of
turbomachines due to its autonomy, as it required modest user
intervention once it was set up. The industry-wide adoption and
use of this technology would have significant economic and envi-
ronmental impacts. The automated optimization and component
performance gains support the growth of the turbomachinery in-
dustry, a global business forecast as growing to 4.16 billion US$ in
2020. The savings in fuel consumption and in CO2 emissions from
using a more efficient turbine with the contoured endwall
(c) will contribute to providing a more affordable and sustainable
energy supply while progressing towards the UNFCCC emissions
targets.

Acknowledgement

Fig. 24. Contours of static pressure on the rotor blade tip for different stator 1 casing
This work was undertaken under the auspices of the GE Powere
designs. (a) Axisymmetric casing, (b) diffusion casing, and (c) optimized groove casing.
University of Leicester framework agreement. Advice from Dr. N. Z.
Ince and Dr. M. Willetts, GE, is gratefully acknowledged. Funding by
contouring techniques. It uses a new casing surface parametriza- the Higher Committee for Education Development in Iraq (HCED) is
tion in which a guide groove, elicited from the natural path of the acknowledged. This research used the ALICE high performance
endwall secondary flows, manipulates specific flow structures computing facility at the University of Leicester. Graphical
directly. A set of parametric equations is used with the Beta dis- rendering software licenses were originally acquired with EPSRC
tribution function to design the casing groove path. This is a first support on Grant GR/N23745/01. The supply of experimental data
application of the Beta distribution function to the contouring of a for the 1.5-stage axial flow turbine “Aachen Turbine” under license
turbomachine casing. by RWTH Aachen is gratefully acknowledged.
 H.T. Kadhim, A. Rona / Energy 149 (2018) 875e889 889

References centrifugal turbine for Organic Rankine Cycle (ORC) applications. Energy
2017;140:1239e51.
[16] Persico G, Dossena V, Gaetani P. Optimal aerodynamic design of a transonic
[1] Al Jubori AM, Al-Dadah R, Mahmoud S. Performance enhancement of a small-
centrifugal turbine stage for organic Rankine cycle applications. Energy Pro-
scale organic Rankine cycle radial-inflow turbine through multi-objective
cedia 2017;129:1093e100.
optimization algorithm. Energy 2017;131:297e311.
€ [17] Sun H, Li J, Song L, Feng Z. Non-axisymmetric turbine endwall aerodynamic
[2] Turgut OH, Camci C. A nonaxisymmetric endwall design approach and its
optimization design: Part Idturbine cascade design and experimental vali-
computational assessment in the NGV of an HP turbine stage. Aero Sci Technol
dations. In: ASME turbo expo 2014: turbine technical conference and expo-
2015;47:456e66.
sition, American society of mechanical engineers; 2014. ASME Paper
[3] Harvey N, Brennan G, Newman D, Rose M. Improving turbine efficiency using
GT2014e25362.
non-axisymmetric end walls: Validation in the multi-row environment and
[18] Tang H, Liu S, Luo H. Design optimization of profiled endwall in a high work
with low aspect ratio blading. In: ASME turbo expo 2002: power for land, sea,
turbine. In: ASME turbo expo 2014: turbine technical conference and expo-
and air; 2002. ASME Paper GT2002e30337.
sition; 2014. ASME Paper GT2014e26190.
[4] Schobeiri M. Turbomachinery flow physics and dynamic performance. Ger-
[19] Kim I, Kim J, Cho J, Kang Y-S. Non-axisymmetric endwall profile optimization
many: Springer; 2005.
of a high-pressure transonic turbine using approximation model. In: ASME
[5] Coull JD. Endwall loss in turbine cascades. J Turbomach 2017;139(8):
turbo expo 2016: turbomachinery technical conference and exposition; 2016.
081004e12.
ASME Paper GT2016e57970.
[6] Langston L. Secondary flows in axial turbinesda review. Ann N Y Acad Sci
[20] Walraevens RE, Gallus HE. European reasarch community on flow turbulence
2001;934(1):11e26.
and combustion; ERCOFTAC SIG on 3D turbomachinary flow preditions. Test
[7] Ligrani P, Potts G, Fatemi A. Endwall aerodynamic losses from turbine com-
Case 1997;6. 1e1/2 stage axial flow turbine. In: Seminar and Workshop on 3D
ponents within gas turbine engines. Propulsion and Power Research
Turbomachinary flow prediction.
2017;6(1):1e14.
[21] Roache PJ. Perspective: a method for uniform reporting of grid refinement
[8] Wang HP, Olson SJ, Goldstein RJ, Eckert ERG. Flow visualization in a linear
studies. J Fluid Eng 1994;116(3):405e13.
turbine cascade of high performance turbine blades. J Turbomach
[22] Wilcox DC. Turbulence modeling for CFD. third ed. California: DCW Industries,
1997;119(1):1e8.
Inc.; 2006.
[9] Sieverding C. Recent progress in the understanding of basic aspects of sec-
[23] Wilcox DC. Companion software: turbulence modeling for CFD. third ed.
ondary flows in turbine blade passages. J Eng Gas Turbines Power
California: DCW industries, inc.; 2006.
1985;107(2):248e57.
[24] Jasak H, Beaudoin M. OpenFOAM turbo tools: from general purpose CFD to
[10] Germain T, Nagel M, Raab I, Schüpbach P, Abhari RS, Rose M. Improving ef-
turbomachinery simulations. In: ASME-JSME-KSME 2011 joint fluids engi-
ficiency of a high work turbine using nonaxisymmetric endwalls- part I:
neering conference; 2011. p. 1801e12. ASME Paper AJK2011e05015.
endwall design and performance. J Turbomach 2010;132(2):021007e9.
[25] Piegl L, Tiller W. The NURBS book. USA: Springer Science & Business Media;
[11] Ingram G, Gregory-Smith D, Rose M, Harvey N, Brennan G. The effect of end-
2012.
wall profiling on secondary flow and loss development in a turbine cascade.
[26] Devore JL. Probability and statistics for engineering and the sciences. USA:
In: ASME turbo expo 2002: power for land, sea, and air; 2002. ASME Paper
Cengage Learning; 2015.
GT2002e30339.
[27] Kadhim HT, Rona A, Leschke K, Gostelow JP. Mitigating secondary flows in a
[12] Praisner TJ, Allen-Bradley E, Grover EA, Knezevici DC, Sjolander SA. Applica-
1½ stage axial turbine by design a guide groove casing. In: Proceedings of the
tion of nonaxisymmetric endwall contouring to conventional and high-lift
23rd ISABE conference, international society of air-breathing engines; 2017.
turbine airfoils. J Turbomach 2013;135(6):061006e8.
Paper No. 22565.
[13] Da Lio L, Manente G, Lazzaretto A. Predicting the optimum design of single
[28] Queipo NV, Haftka RT, Shyy W, Goel T, Vaidyanathan R, Tucker PK. Surrogate-
stage axial expanders in ORC systems: is there a single efficiency map for
based analysis and optimization. Prog Aero Sci 2005;41(1):1e28.
different working fluids? Appl Energy 2016;167:44e58.
[29] Hilfer M, Ingram G, Hogg S. Endwall profiling with tip clearance flows. In:
[14] Meroni A, Andreasen JG, Persico G, Haglind F. Optimization of organic Rankine
ASME turbo expo 2012: turbine technical conference and exposition; 2012.
cycle power systems considering multistage axial turbine design. Appl Energy
ASME Paper GT2012e68488.
2017.
[15] Song Y, Sun X, Huang D. Preliminary design and performance analysis of a