0O UIF DPOKVHBUF

IFBU USBOTGFS
NFUIPE1XPHULFDO VWXG\ DQG DQ
DSSOLFDWLRQ FDVH RI RSWLPL]DWLRQ for D
JDV WXUELQH WUDQVLWLRQ piece FRROLQJ

4 "OHFMPW
WR REWDLQ WKH GHJUHH RI 0DVWHU RI 6FLHQFH
DW WKH 'HOIW 8QLYHUVLW\ RI 7HFKQRORJ\
WR EH GHIHQGHG SXEOLFO\ RQ )ULGD\ $SULO   DW  30

6WXGHQW QXPEHU 

3URMHFW GXUDWLRQ )HEUXDU\   ± $SULO  
7KHVLV FRPPLWWHH 3URI GU LU 6 .OHLQ 78 'HOIW VXSHUYLVRU
'U 5 3HFQLN 78 'HOIW FRVXSHUYLVRU
'U 6 +HUPHWK 6LHPHQV (QHUJ\ FRVXSHUYLVRU
'U % :HJQHU 6LHPHQV (QHUJ\ FRVXSHUYLVRU
'U + %D]\DU 78 'HOIW

7KLV WKHVLV LV FRQILGHQWLDO DQG FDQQRW EH PDGH SXEOLF XQWLO $SULO  

$Q HOHFWURQLF YHUVLRQ RI WKLV WKHVLV LV DYDLODEOH DW KWWSUHSRVLWRU\WXGHOIWQO

Acknowledgements

We’ve come a long long way together,

Through the hard times and the good,
I have to celebrate you baby,
I have to praise you like I should.

Fatboy Slim, “Praise You”

First, I want to express my deepest gratitude to my supervisor Prof. Sikke Klein, for letting me do
my Master’s thesis abroad and also for trusting me throughout the entire time. No one could have
expected what everyone was going to go through in the past year but his support and concise criticism
really helped me steer the project in the right direction, despite the many (physical) roadblocks that just
seemed to keep arising.
I would like to also thank my daily supervisor at Siemens, Dr. Sebastian Hermeth. First, for letting
me do my thesis in the combustion team, and second, for his generosity in sharing knowledge and
industry acumen, which were the key for me in building a wide network of contacts in Germany and
US, particularly Landon and Jeffrey, accelerating my progress in gaining crucial software skills and heat
transfer knowledge. My gratitude also goes to Dr. Bernhard Wegner. I sometimes wondered how he
managed (pun intended) to keep such a high level of calmness and professionalism when approaching big
uncertainties. My work benefits i n b ig p arts o f h im. I would a lso l ike t o t hank L ukasz f or b eing i n the
office w hen n o-one e lse was d uring t he p andemic a nd t he f un d aily t alks over l unch, g ently emphasizing
to us, students, that there is, in fact, a life outside of the simulation.
The friendships I built in Delft made the last two and a half years something I will never forget, and I
hope that I will meet again most of you soon. To my partner in crime in Berlin, Michael, the person to
probably suffer t he m ost f rom my o ccasional frustrations. H e was t here t o r eassure m e, w hile t he “few”
beers definitely helped (despite the neighbour’s bitterness). Fernando, Giaccomo, Herman, Marko, Lim,
Stefan, Diego, Akhil, and everyone else in Delft, also played a role in me reaching this point today, and
I hope I meet them soon somewhere across the globe.
To my sister and brother-in-law for their relentless support during the uncertain times in Berlin, and
also my family at home: it certainly was not easy for them, and I wish that someday I will be able to
“pay back” for their support, in one shape or another.
And last but surely not least, to Julia. I am sorry for being empty-headed at times. She made this
whole journey a lot more pleasant and I thank her for her endless patience.

Vielen Dank,
Dank je wel,
Thank you,

Svetoslav Angelov,
Burgas, April 2021.
Abstract

Gas turbines (GT) play a crucial role in the transition to fully sustainable energy. Conversely, the
expected growth of the GT market is met with more stringent regulatory requirements. Clearly, an
increasing turbine inlet temperature (TIT) is beneficial for the efficiency of GT cycles. With increasing
TIT, however, the magnitude of the thermal loads becomes unbearable, causing thermal fatigue of the
component in a transient setting such as GT start-up or shut-down. Therefore, it is crucial that the
thermal loads are well managed, which is typically achievable only by an adequate external and internal
cooling. Generally, such cooling systems are modelled in industry using an uncoupled heat transfer (HT)
approach, assuming that the heat coefficients (HTCs) are only influenced by the aerodynamics of the
flow field: HTCs are obtained in a separate computational fluid dynamics (CFD) simulation and are
extracted and mapped on the solid domain of the component.
Firstly, the report presents a comparison between the uncoupled approach and the increasingly popular
conjugate heat transfer (CHT) approach, in which the fluid and solid domains are solved in a single
simulation. To do so, the commercial code (STAR-CCM+) is first validated for both laminar flat plate
and turbulent offset jet in a steady-state setting. The latter is done by comparing three Reynolds-
Averaged Navier-Stokes (RANS) models. The popular wall treatment methods: blended and high-y +
wall functions (WFs) are also compared. The comparison between the two HT methodologies is done
by a parameter sensitivity study, involving the thermal conductivity of the solid (for both laminar and
turbulent cases), Turbulent Prandtl number and turbulence intensity. It was concluded that the realizable
k − ε model is the most accurate for this flow. Conjugation played a marginal role for all parameters.
with only a considerable difference in the stagnation regions, when the solid conductivity is low. The
results using a high-y + WF displayed high numerical dissipation due to the coarse mesh size and hence
the difference between the HT methodologies was artificially low. It is therefore advised that the study
of the high-y + WF is repeated on a different flow arrangement/setting (e.g. increased inlet velocity).
Secondly, the thesis details a study around the internal cooling of Siemens GT transition piece (TP),
using CHT. An introduction of perpendicular ribs into a fully developed turbulent channel flow can theo-
retically significantly increase the heat transfer coefficient through the introduction of large-scale vortices
downstream. An optimization routine was run with a hybrid algorithm to study the effect of different rib
configurations in a steady RANS simulation, aiming to minimize both mass flow and transition bond coat
temperature. The performance is compared to the baseline smooth channels used in the Siemens TP. No
better results were achieved for both objective functions in comparison to the baseline case mostly due
to the length of the channels. However, a considerable decrease in the coolant consumption was present
with a limited increase in the BC temperature. The study showed that the optimization of such ribs
can minimize the effect of skin friction (which influences the amount of coolant), and presents a possible
further improvement of the TP cooling, without relying solely on the experience of the designer.

Keywords: conjugate heat transfer, optimization, combustion equipment, cooling, channel turbulation.

i
Contents

Acknowledgements

Abstract i

List of Figures v

List of Tables x

1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Gas turbine introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Novel computational methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Project scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.1 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.2 Research questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Literature study 7
2.1 Gas turbine combustion system basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.1 Structural basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.2 Flow arrangement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.3 Heat transfer in combustion systems . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.4 Thermal loads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Conjugate heat transfer problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.1 Flat plate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.2 Turbulent offset jet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3 Combustion systems cooling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.1 Internal cooling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.2 External cooling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.4 Low cycle fatigue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4.1 Lifespan calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3 Optimization Terminology 29
3.1 Optimization algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.1.1 Simplex programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1.2 Deterministic and stochastic programming . . . . . . . . . . . . . . . . . . . . . . . 32
3.1.3 Genetic optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1.4 SHERPA ® . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Multi-disciplinary Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3 Screening and Design of Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

ii
 Delft University of Technology

4 Numerical Methodology 39
4.1 Numerics of conjugate heat transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2 Reynolds-Averaged Navier-Stokes Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.2.1 Standard k − ε model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.2.2 Realizable k − ε model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.2.3 SST k − ω model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.3 Calculation of heat transfer coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.4 Wall treatment functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

I Part 50
5 Numerical validation cases 51
5.1 Laminar flat plate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.1.1 Analytical solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.1.2 CFD setup and boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.1.3 Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.1.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.2 Turbulent offset jet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.2.1 CFD setup and boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.2.2 Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.2.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

6 Heat transfer methodology comparison 70

6.1 Flat plate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
6.2 Turbulent offset jet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.2.1 Turbulent Prandtl number . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.2.2 Turbulence intensity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.2.3 Thermal conductivity ratio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

II Part 89
7 Gas Turbine Transition Piece Analysis 90
7.1 Computational domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
7.2 Reynolds Analogy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
7.3 Parametrization of cooling topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
7.3.1 Mesh study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
7.4 Flow validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
7.5 Preliminary baseline comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

8 Ribs Optimization Study 106

8.1 Optimization set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
8.1.1 Optimization objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
8.1.2 Optimization parameters & range . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
8.2 Results & analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
8.3 Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

9 Conclusion 118
9.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
9.2 Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

Appendices 121

Energy & Process Technology Page iii

 Delft University of Technology

Appendix A Gas turbine basics 122

Appendix B Luikov solution 123

B.1 Direct heat transfer solution method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
B.2 Boundary layer approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
B.3 Numerical solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

Appendix C Turbulent offset jet mesh study 128

Appendix D Auxiliary contracted inlets study 132

D.1 Concept introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
D.2 Constant cross-section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
D.3 Adjusted cross-section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

Appendix E More on numerical methodology 135

E.1 Interpolation methods & coupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
E.1.1 Interpolation stencils . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
E.1.2 CHT coupling algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
E.2 More on turbulence models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
E.2.1 Derivation of the RANS equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
E.2.2 Realizable k − ε Damping function . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
E.2.3 Two-layer modification of the k − ε model . . . . . . . . . . . . . . . . . . . . . . . 142
E.2.4 Shear and stretching modification in the k − ω SST model . . . . . . . . . . . . . . 142
E.3 CFD roughness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

Appendix Bibliography 144

Energy & Process Technology Page iv

List of Figures

1.1 Illustration of the Siemens SGT-8000H gas turbine. The term ’PCS’ stands for Platform
Combustion System and was proposed by Siemens in 2005. [109] . . . . . . . . . . . . . . 1
1.2 Overall evolution over the years of the TIT. Obtained from Rao [128]. . . . . . . . . . . . 2
1.3 The decoupled approach (top) versus the conjugate approach (bottom). Dashed line is
used to indicate necessity for manual data extraction and mapping between simulation
tools. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Flow schematic in a reverse flow arrangement combustor, with a typical diffuser, plenum
chamber, and transition piece. Original unannotated image obtained from [103] . . . . . . 4
1.5 Thesis outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2.1 Main components in the combustion system of the Siemens SGT5-8000H [38] . . . . . . . 8
2.2 Typical transition piece geometries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Schematic of heat transfer loads in a combustor liner. Extraced from [6] . . . . . . . . . . 14
2.4 The one-dimensional CHT problem schematic . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.5 Temperature gradients at low and high Biot number settings in fluid and solid regions . . 16
2.6 The turbulent offset jet problem. Extracted from [95]. . . . . . . . . . . . . . . . . . . . . 18
2.7 Dimensionless wall temperature on the wall as a function of offset ratio (θw = (Twall −
′
T∞ )/(Tex − T∞ ), x = xIr ) for a conjugate heat transfer case [47]. . . . . . . . . . . . . . 19
2.8 Two examples of internal cooling design for GT turbine vanes; for reference, vane length
is 12.75mm [15]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.9 The three configurations used in the study of Siw et al. (top to bottom: triangular,
semicircular, circular) and the dimensional heat transfer coefficient distributions (domain
dimensions width, W =63.5mm; height, H=25.4mm; length, L=101.6mm) [111] . . . . . . 21
2.10 Rib shape configurations reported in [41] . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.11 Turbulence intensity contour plot from [17], indicating three key recirculation regions of
low HTC where hot spots occur, for flows over ribs (Re = 3 · 104 , e/Dh = 0.3, p/e = 10) . 22
2.12 Comparison of heat transfer performance (heat transfer augmentation N u/N uo on y-
coordinate versus friction factor f /fo ) for various types of internal cooling. Extracted
from [77] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.13 A typical array of a jet-impingement geometry. . . . . . . . . . . . . . . . . . . . . . . . . 24
2.14 Schematic of a cross-flow. Gc and Gj are the cross-flow and impingement jet velocities
respectively . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.15 Schematic of a film cooling process, illustrating the mixing process of a coolant with
mainstream flow. [91] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.16 Von Mises yield surface, along with the yield criterion function f , along with the increment
in plastic strain normal to the surface, and the resultant stress–strain curve obtained for
uniaxial straining (no hardening). Note at the yield surface, f = 0. . . . . . . . . . . . . . 25
2.17 Neuber shakedown method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.18 Wöhler SN curve (log-scale) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.19 Effect of mean stress presence on fatigue life [115]. . . . . . . . . . . . . . . . . . . . . . . 27

v
 Delft University of Technology

2.20 Constant mean strain ϵm effect on the mean stress levels σm . The relaxation is due to the
plastic deformation, and hence the relaxation rate depends on strain amplitude [115]. . . . 27

3.1 Families and classes of optimization algorithms (adapted from [80]) . . . . . . . . . . . . . 30

3.2 A polyhedron P and an extremum, highlighted in red. . . . . . . . . . . . . . . . . . . . . 31
3.3 Genetic evolutionary process [5]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.4 Simplified Pareto front illustration of a multi-objective trade-off study. . . . . . . . . . . . 33
3.5 Results from benchmark comparison of different optimization algorithms. SHERPA yielded
better results even at a very low number of evaluations [116]. . . . . . . . . . . . . . . . . 34
3.6 Pareto front of a MO (2 objectives) optimization and the equivalent optimization using a
weighted sum in a single-objective optimization. The dashed lines represent values of the
converted MO calculation. The resultant penalty decreases as the dashed line approaches
the Pareto front. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.7 Linear effect a), quadratic effect b), and interaction c). . . . . . . . . . . . . . . . . . . . . 36
3.8 A schematic of a full-factorial design with three factors and two levels each. . . . . . . . . 37
3.9 Illustrative representation of the effects of factor x1 . . . . . . . . . . . . . . . . . . . . . . 37

4.1 Domains and boundaries in a typical conjugate heat transfer problem . . . . . . . . . . . 40

4.2 Typical velocity profile of a turbulent boundary layer . . . . . . . . . . . . . . . . . . . . . 47
4.3 Velocity (a) and temperature (b) as a function of non dimensional wall distance displaying
wall treatment for viscous sub-layer, log layer, and the blending approach covering the
buffer layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

5.1 Schematic of boundary conditions and respective notation used for the analytical solution
of Luikov.[81] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.2 Problem domain schematic and boundary conditions. . . . . . . . . . . . . . . . . . . . . . 56
5.3 Heat transfer coefficient along the plate surface for Bi = 0.1, and Bi = 1 for each of the
two solution methods by Luikov [81]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.4 Mesh used in the validation and HT methodology comparison. Grid is of size 60x120 and
a first cell height of ∆ymin = 7.2e-05m with a logarithmic stretching factor of 1.061 . . . . 57
5.5 Interface temperature and heat transfer coefficient comparison with the differential heat
transfer approach [81], at Bi = 0.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.6 Temperature distribution across the fluid region (left) and solid (right) and comparison
to Luikov solution, with a Bi = 0.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.7 Comparison of velocity profiles to Blasius solution [106], at Bi = 0.1 . . . . . . . . . . . . 60
5.8 Interface temperature and heat transfer coefficient comparison with the differential heat
transfer approach [81], at Bi = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.9 Temperature distribution across the fluid region (left) and solid (right) and comparison
to Luikov solution, with a Bi = 1.0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.10 Comparison of velocity profiles to Blasius solution [106], at Bi = 1 . . . . . . . . . . . . . 62
5.11 Close-up of dark gray region in Figure C.2: schematic of boundary conditions used in the
turbulent offset jet problem. Domain width is 150mm (identical to experiment) . . . . . . 64
5.12 Interface temperature variation using countermeasures on the realizable k − ε model . . . 66
5.13 Comparison of eddy viscosity for all three models . . . . . . . . . . . . . . . . . . . . . . . 66
5.14 Comparison of interfacial temperature profile for all three models using finer mesh (M2)
and blended wall function approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.15 Nondimensional temperature in the wall-normal direction at (a) 3mm in front of the exit
nozzle and (b) X = 6.69 (midway in the recirculation region in the experiment). . . . . . 67
5.16 Decay of the maximum jet temperature downstream (SST model in red and k − ε model
in black). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.17 Comparison of interfacial temperature profile for all three models using coarse mesh (M6)
and high-y + wall function approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

6.1 Wall interface temperature (solid side) as a function of all conductivity ratios consider. . . 71

Energy & Process Technology Page vi

 Delft University of Technology

6.2 Velocity profile comparison for all thermal conductivity ratios at four different axial locations. 72
6.3 Nondimensional (a) velocity and (b) thermal boundary layers for all cases. . . . . . . . . . 73
6.4 Critical z criterion for all cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.5 Averaged Nusselt number, derived from Eq.5.23 (shape-preserving interpolant) . . . . . . 74
6.6 Effect of thermal conductivity ratio on boundary layer ratio (shape-preserving interpolant) 74
6.7 Comparison of interface temperatures using a decreased bottom plate surface temperature,
for all conductivities (a) - (e). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.8 Comparison of resultant wall y + profiles from the validation case and the methodology
comparison with increased jet velocity. The value of the wall y + is still ≤ 10 for the
blended WF approach, which is beneficial for the consistency of the study. . . . . . . . . . 77
6.9 Local wall interface temperature (solid side) as a function of P rt . . . . . . . . . . . . . . . 78
6.10 Local wall interface temperature (solid side) as a function of P rt . . . . . . . . . . . . . . . 78
6.11 Local wall interface temperature (solid side) as a function of P rt . . . . . . . . . . . . . . . 79
6.12 HTC as a function of P rt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.13 Local wall interface temperature (solid side) as a function of P rt . . . . . . . . . . . . . . . 80
6.14 HTC as a function of P rt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
6.15 Influence on surface averaged values of (a) temperature, (b) reattachment length, (c)
specified-y + heat transfer coefficient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.16 Local wall interface temperature (solid side) as a function of Tu(%). . . . . . . . . . . . . 82
6.17 HTC as a function of Tu(%). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.18 Turbulent kinetic energy for (a) Tu = 2%, (b) Tu = 15% using a fine mesh (M2) and the
blended wall function approach. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.19 Turbulent kinetic energy for (a) Tu = 2%, (b) Tu = 15% using a coarse mesh (M6) and
the high-y + wall function. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.20 Local wall interface temperature (solid side) as a function of Tu(%). . . . . . . . . . . . . 83
6.21 HTC as a function of Tu(%). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.22 Influence on surface averaged values of (a) temperature, (b) reattachment length, (c)
specified-y + heat transfer coefficient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.23 Local wall interface temperature (solid side) as a function of λ∗ . . . . . . . . . . . . . . . 85
6.24 HTC as a function of λ∗ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.25 Local wall interface temperature (solid side) as a function of λ∗ . . . . . . . . . . . . . . . 86
6.26 HTC as a function of λ∗ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.27 Influence on surface averaged values of (a) temperature, (b) reattachment length, (c)
specified-y + heat transfer coefficient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

7.1 Illustration of approach used for discretizing the transition piece geometry into a para-
metric slide of two channels. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
7.2 Top panel schematic and boundary conditions used, along with a close-up view of the chan-
nels, where relevant interfaces are illustrated using dashed lines. Red and blue are used
to indicate fluid region (hot and cold flows, respectively), whereas gray is the transition
metal region. Channel flow is shown in purple. . . . . . . . . . . . . . . . . . . . . . . . . 91
7.3 Bottom panel schematic and boundary conditions used. . . . . . . . . . . . . . . . . . . . 92
7.4 Types of HTC BCs and interfaces in the transition piece set-up. . . . . . . . . . . . . . . 92
7.5 Mapped HTC on hot and cold sides of the top and bottom panels, extracted from a full
midframe simulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.6 Roughness benchmarking for AM-channels of various build directions: friction factor f ,
Nusselt number (b) as a function of Re,and f and heat transfer augmentation for all tested
channels (c) from Snyder [112]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
7.7 Schematic representation of key parameters used in the rib geometry . . . . . . . . . . . . 95
7.8 Resultant flow structure in the rib geometry used (top panel used for illustrative pur-
poses). The structure closely resembles the flow structure commonly found in literature,
see Fig.2.11, with the exception of a recirculation region on top of the rib structure. . . . 96
7.9 Comparison of mesh resolution: M1 (a) and M5 (b), channel shown in blue. . . . . . . . . 97

Energy & Process Technology Page vii

 Delft University of Technology

7.10 Mesh convergence tracked in surface averaged temperatures: Bond coating (i.e. hot side
of transition metal) (a), thermal barrier coating (b), channel surface (c). . . . . . . . . . . 98
7.11 Outer-layer scaling at four locations downstream in one of the cooling channels inside the
bottom panel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
7.12 Schematic of a square duct with transverse ribs used by Wang [84]. The duct uses a
blockage ratio of 0.2 and pitch-to-height ratio of 4. . . . . . . . . . . . . . . . . . . . . . . 99
7.13 Comparison to DNS data of nondimensional streamwise velocity at three downstream
locations of the rib. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
7.14 Effect of HLP on cooling effectiveness, and the ratio of passive to active cooling require-
ment as a function of the cold side convective heat transfer coefficient. . . . . . . . . . . . 101
7.15 Local wall temperatures for bottom (a) and top (b) panels, with and without rib turbulators.102
7.16 Nondimensional bulk velocity and bulk nondimensional temperature along the two chan-
nels for both geometries. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.17 Bulk heat transfer coefficients on both panels for each channel: (a)-(b) top panel, (c)-(d)
bottom panel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

8.1 Flowchart of the automated design workflow in HEEDS. . . . . . . . . . . . . . . . . . . . 107

8.2 SHERPA Design performance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
8.3 Evolution of temperature OF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
8.4 Evolution of nondimensionalized cooling mass flow . . . . . . . . . . . . . . . . . . . . . . 109
8.5 Pareto front of top panel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
8.6 Pareto front of bottom panel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
8.7 Influence of rib parameters on bottom panel temperature and cooling . . . . . . . . . . . 113
8.8 Influence of rib parameters on top panel temperature and cooling . . . . . . . . . . . . . . 114
8.9 Nusselt number increase with Reynolds number . . . . . . . . . . . . . . . . . . . . . . . . 115
8.10 Skin firction increase with Reynolds number . . . . . . . . . . . . . . . . . . . . . . . . . . 115
8.11 Nusselt number against skin friction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
8.12 Z-score (scatter and distribution) of (a) Nusselt number and (b) Skin friction coefficient. . 116

B.1 All five meshes used for the independence study of the flat plate problem, in the vicinity
above the plate (no rear and front fluid extensions) . . . . . . . . . . . . . . . . . . . . . . 125
B.2 Temperature, energy, and thermal load residual values from M1 and M2, until iteration
number 120 (convergence for M1 reached at 260 iterations) . . . . . . . . . . . . . . . . . 126
B.3 Comparison of friction coefficient to Blasius analytical solution, along with heat transfer
coefficient, heat flux, and temperature at the interface with Luikov analytical solution
(Bi = 0.1). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
B.4 Mesh M1 with x-refinement (incl. extensions in fluid region). . . . . . . . . . . . . . . . . 127
B.5 Comparison of interface temperature obtained from refined mesh in the x-direction w.r.t
mesh without refinement and the analytical solution from Luikov (Bi = 0.1). . . . . . . . 127

C.1 Schematic of experimental arrangement, extracted from [95] . . . . . . . . . . . . . . . . . 128

C.2 Entire CFD domain with relevant dimensions (m) (light gray region is not included in
Figure 5.11, i.e. ambient room region in the experimental domain, having a width of 3.0m)128
C.3 All five meshes used for the independence study of the flat plate problem, in the vicinity
above the plate (no rear and front fluid extensions) . . . . . . . . . . . . . . . . . . . . . . 130
C.4 Convergence for M1, M2, M3, M4 and M5 . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
C.5 Axial velocity component convergence with experimental data at five downstream location 131

D.1 Schematic of the setup: auxiliary inlets are place downstream for both counter current
channels. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
D.2 Illustrative depiction of the idea of contracting the inlets downstream. . . . . . . . . . . . 133
D.3 Effect of number of inlets and inlet angle on average, maximum temperatures and coolant
consumption (nondimensionalized). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

Energy & Process Technology Page viii

 Delft University of Technology

D.4 Influence of the amount of contraction on the average, maximum temperatures and coolant
consumption (nondimensionalized). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
D.5 Kriging fits of numerical data: (a) nondimensionalized mass flow and (b) average bond
coating temperature. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

E.1 Nearest neighbour interpolation method illustration. . . . . . . . . . . . . . . . . . . . . . 136

E.2 Least square interpolation schematic. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
E.3 Flow chart of the hFTB method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
E.4 Flow chart of the hFFTB method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

Energy & Process Technology Page ix

List of Tables

2.1 Review of CHT studies on flat plates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3.1 Brief comparison of full-factorial and partial-factorial designs . . . . . . . . . . . . . . . . 38

4.1 Model coefficients used in the standard formulation of the k − ε model, according to
Launder and Sharma [70] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.2 Model coefficients used in the realizable formulation of the k − ε model . . . . . . . . . . . 44
4.3 Model coefficients used in the SST k − ω model . . . . . . . . . . . . . . . . . . . . . . . . 44
4.4 Summary of available heat transfer coefficients. . . . . . . . . . . . . . . . . . . . . . . . . 46
4.5 Wall functions for both approaches for turbulence parameters required by the RANS model 49

5.1 Calculated properties of air at 1000K. [65] . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

5.2 Resultant solid thermal conductivity from the two average Biot numbers . . . . . . . . . . 55
5.3 Summary of average values for temperature and heat transfer coefficient from both simu-
lations and the analytical solution [81]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.4 Sources of errors and unknowns in the numerical procedure followed for the CHT validation. 65

6.1 Resultant Biot numbers from heat conductivity ratios . . . . . . . . . . . . . . . . . . . . 71

6.2 Summary of average heat transfer coefficient, and boundary layers comparison for all
conductivities at Tb = 600K. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.3 Error comparison for average interface temperature for baseline and increased critical
temperature difference (T∞ − Tb ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

7.1 Non-dimensionalized transition channels effective area (Aef f /Ach ): validation of pressure
drop calculation versus test measurements. . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.2 Thermal conductivities used in the model. . . . . . . . . . . . . . . . . . . . . . . . . . . 94
7.3 Cross section of baseline (circular) and ribbed baseline (rectangular) channels used. . . . . 96
7.4 Mesh settings used for the top panel study. The channel with an inlet upstream is denoted
with 1, and vice versa. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
7.5 Computational cost comparison between all meshes. . . . . . . . . . . . . . . . . . . . . . 98
7.6 Maximum allowable temperatures and design margins . . . . . . . . . . . . . . . . . . . . 102
7.7 Comparison of two baseline cooling schemes . . . . . . . . . . . . . . . . . . . . . . . . . . 104

8.1 Comparison of maximum and minimum values achieved by SHERPA for both panels to
the values with baseline channel configuration (BR - baseline ribbed, BCIRC - baseline
circular). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

B.1 Interface boundary condition settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

B.2 Properties of meshes used for laminar flat plate study . . . . . . . . . . . . . . . . . . . . 124

C.1 Specification of mesh settings on target surfaces for M2 . . . . . . . . . . . . . . . . . . . 129

C.2 Mesh attributes of all meshes in the mesh independence study . . . . . . . . . . . . . . . . 129

x
 Delft University of Technology

E.1 Different interpolation methods in accordance to the source and target stencils. . . . . . . 135
E.2 Values and settings of the mapped interface tolerance . . . . . . . . . . . . . . . . . . . . 137

Energy & Process Technology Page xi

Chapter 1

Introduction

1.1 Motivation
1.1.1 Gas turbine introduction
While renewable technologies are becoming more and more popular, the gas turbine (GT) market was
responsible for roughly 30% of worlds’ energy production as of October 20201 , and is expected to grow
by at least 5.4% yearly by 2027 due the increased foreseen energy demand in the near future [79]. GTs
also offer flexibility compared to alternative power generation methods (easy start/stop operation). This
can be further employed to serve as a back up for intermittent renewable power, or potential to switch
to zero carbon fuels like hydrogen.

Transition piece
Can-annular
PCS burners

Multi-stage compressor

Multi-stage turbine

Figure 1.1: Illustration of the Siemens SGT-8000H gas turbine. The term ’PCS’ stands for Platform
Combustion System and was proposed by Siemens in 2005. [109]
.

On the other hand, governmental regulations on gas turbine components have increasingly become
more and more stringent in terms of emissions and pollutant levels. Thereby, in order to be able to
adhere to these regulations and remain competitive by adapting to rapidly changing demands, turboma-
chine manufacturers face a multi-faceted problem, the main two challenges of it being the shortening of
the design cycle and the necessary increase in the performance of their products.
1 Data provided by Organisation for Economic Co-operation and Development (OECD)

1
 Delft University of Technology

The latter could be achieved by addressing the improvement of various systems of the gas turbine.
An illustration of large a gas turbine is given in Figure 1.1, where the main stages in a gas-turbine
cycle are given: namely the compressor stage, combustion chamber, and turbine stage. From a societal
point of view, it can be mathematically shown that gas turbine cycles should be operated at a turbine
inlet temperature (TIT) as high as possible in order to maximize its thermodynamic efficiency. For the
case of the Joule-Brayton cycle (the idealized thermodynamic process in the gas turbine), at optimum
pressure ratio, the thermodynamic efficiency can be thus read as:
r
T2
ηth−dynamic = 1 − , (1.1)
T4
where T2 is the compressor inlet temperature and T4 is the TIT. the full derivation of the equation is
given in Appendix A. Clearly, the cooling requirement within gas turbines are hence also increased, due
to the resultant increase in heat flux on components at higher TIT, and the melting temperature of the
turbine material being lower. In addition to this, GT components exposed to such high temperatures
are more susceptible to a secondary failure mechanism, arising from temperature differentials developed
during the starting and stopping of the turbine. The resultant thermal expansion and contractions
produce thermal stresses, the cycling of which causes thermal fatigue. The overall trend of the further
increase of TIT in the industry is to remain, as shown in Fig.1.2, which amplifies the importance of novel
approaches to tackle the challenge.

2200
Engine turbine inlet temperature °K
Trent 900
Equiaxed crystal
2000
Turbine inlet temperature [K]

Polycrystal
Trent 500

gap compensated by
Single Crystal

tubrine cooling
1800 RB211-535E4B Trent 800
RB211-535E4
RB211-524B-02
1600 RB211-524-02
RB211-22B
Spey
1400 Conway
PWA1484 CMSX4 PWA1492
PWA1480
Conway-12 PWA1422
Tyne-12 R142 PWA1487
N4 N5
1200 MM247 R80H
Dart Avon-3

Avon-1
1000
1950 1960 1970 1980 1990 2000 2010
Certification Year

Figure
Figure1.2: Overall
2: Evolution evolution
of turbine over the years
inlet temperature of the
and metal TIT.temperature
operating Obtainedover
from Rao (Figure
the years [128].
reproduced by data from Rolls Royce jet engine).
A possible solution to the problem is simply increasing the cooling air rates until the cooling require-
ments are met. Moreover, increasingissue
The known OPR and TITincreasing
with increases NOxcooling
emissionsair
for flow,
a given however,
combustor technology
is in the[8].penalty
In Figureon
3, thermal
efficiency (η
the historical trend in Landing Takeoff cycle (LTO) NOx emissions based on ICAO engine certification databank is entropy
thermal = Shaft power/Heat addition), which is mainly a result of two factors: (1)
generation from coolant and hot burnt gas mixing, and (2) irrecoverability of the increased compression
illustrated
work on bleed air in[9]. the
For aexpansion
given periodprocess,
when the due
combustion
to itstechnology
usage inremained unchanged,
combustor andtheturbine
NOx emission
cooling. The
coolant mass flow (compressor bleed air) needs to be hence kept at a minimum.
increased with increase in OPR. The reduction in NOx emission due to the advancement in the combustion

technology can be seen in the figure. It is also evident from the figure that even with advanced combustion
1.1.2 Novel computational methods
technology, it is difficult to satisfy the ACARE objectives (depicted by the solid line at the bottom of the figure).
In today’s age of digitization, the challenge of higher TIT can be solved by both the rapidly changing
manufacturing andbreakthrough
Therefore, design processes.
technologiesIn orderarchitecture
in engine to exploit and the full potential
combustion of thebenewly
technology should explored.emerging manu-
facturing process such as additive manufacturing for example, the design and production methods must
also develop further in the field of gas turbines. Such novel methodologies allow for the employment of

Energy & Process Technology Page 2

 Delft University of Technology

cooling designs, and eventually deliver even better GTs.

With regards to the aforementioned improvement of such design cycles, computer-aided calculation
methods, mainly computational fluid dynamics (CFD) and computational structural mechanics (CSM)
have solidified their presence into the design process of modern manufacturers, together however with
the increased complexity of the simulations and computational demand. In addition, the fundamentals
in heat transfer in the design process are often neglected for the sake of reduction in computational
cost and also time efficiency of projects, and the iterative manner of the entire process has remained.
Thereby, it is important to introduce the topic of conjugate heat transfer (CHT), i.e. the simultaneous
solution of both solid and fluid regions, allowing for an enhanced design methodology that makes the
multiple-step process of calculating thermal stresses considerably shorter. Currently, the well established
practice consists of calculating heat transfer coefficients on all solid walls in adiabatic CFD simulations,
and later imposing the results along with the respective ambient temperature setting on the solid walls
in a separate CSM simulation as a convective boundary condition, with the overall procedure presented
schematically in Figure 1.3.

HTCs &
CFD simulation amb. temperature
CAD generation (adiabatic walls)
Solid FE simulation Solid FE simulation

(temperature) (stresses)

CHT simulation Temperature

CAD generation (interface coupling)
Solid FE simulation

(stresses)

Figure 1.3: The decoupled approach (top) versus the conjugate approach (bottom). Dashed line is used
to indicate necessity for manual data extraction and mapping between simulation tools.

This methodology is only applicable if the assumption that the heat transfer profile on the solid walls
is solely determined by the aerodynamics of the flow is valid. Thereby, by relying only on CFD-based
simulations, the heat loads are calculated with the assumption of a constant adiabatic temperature on
the solid region, which then leads to the arise of multiple concerns. The use of non-conjugate methods
completely omits the reality of fluid-structure interaction [83]. It is worthwhile mentioning that this
assumption may actually be a valid one in some cases, but the complete understanding of the effects
of conjugation require sound reasoning when and why the difference between the two approaches occur.
The use of one-way interaction between fluid and the solid neglects what happens in reality. Additionally,
it can be argued that the heat transfer coefficient does not depends only on the wall temperature level,
but also the upstream history of the boundary layer on the wall, which also could greatly influence the
uniformity on the wall temperature profile. The CHT method also eliminates the need for manual heat
transfer data from one finite volume (FV) mesh to solid finite element mesh (FEM), hence decreasing
component lead time. In this work, this is addressed by evaluating the difference between the two
approaches on two separate cases under various physical and numerical conditions.

Energy & Process Technology Page 3

 Delft University of Technology

Bur
ner Tr
ans
iti
onpi
ece HotTPgas Pl
enum

Fuelpump

Col
dcompr
ess
orf
low
Di
ffus
erwal
l Fl
owr ever
sal
tobuner

Figure 1.4: Flow schematic in a reverse flow arrangement combustor, with a typical diffuser, plenum
chamber, and transition piece. Original unannotated image obtained from [103]

With regards to the mutual connection between the two challenges (design cycle improvement and
gas turbine performance), the term multi-disciplinary optimization (MDO) has gained traction in recent
years. As previously mentioned, one way to meet such GT performance requirements, which is in the
scope of this project, is to enhance the cooling performance of the transition piece of one of Siemens
Heavy Duty GTs combustion systems (CS), visible in Figure 1.1, whereas the typical flow arrangement
is presented in Fig. 1.4. The optimization of the cooling topology would ideally lead to a reduction in
maximum and average temperatures on the hot side surface, and thus also increasing the lifetime of the
component. The multiobjective (MO) nature of such practices comes from the fact that the requirement
for minimizing the coolant mass flow rate is also present for reasons discussed earlier.

1.2 Project scope

1.2.1 Objectives
Given the widespread applicability of the uncoupled approach described above, a comprehensive com-
parison with the CHT method under a variety of conditions, to author’s knowledge, has not been per-
formed, which presents viable research field for this thesis. In addition, with regards to topologies used
for internal cooling of GT components, a number of surface devices such as ribs, dimples and fins are
commonly used in order to increase turbulent kinetic energy and mixing, which plays a crucial part of
effective convective heat transfer. The compromise between heat transfer augmentation and friction can
be achieved by a multi-objective optimization on the geometry parameters of the turbulation using a
novel hybrid algorithm. The main objectives of this thesis can be thereby formulated as:

1. Validate the CHT procedure in a commercial software package against canonical cases for both
laminar and turbulent flow and determine under what conditions the CHT methodology is recom-
mendatory over the commonly used uncoupled approach (e.g. adiabatic/isothermal heat transfer
coefficient method)

2. Develop an optimization methodology using CHT for the design of the cooling of a preselected GT
transition piece, and evaluate whether channel turbulation parameters can be optimized so that
the skin friction losses are minimized.

Energy & Process Technology Page 4

 Delft University of Technology

1.2.2 Research questions

To reach the two objectives mentioned, the following set of research questions was deduced:
1. Does the numerical CHT methodology in the CFD code used results in an accurate solution w.r.t
temperature profiles on the solid region on canonical test cases?
2. What is the best turbulence model for CHT in flows exhibiting similar characteristics as in a GT
transition piece (recirculation, reattachment)?
3. Do wall treatment methods and the physical properties of the flow affect the difference between
the CHT and decoupled approaches?
4. Can the internal cooling design of the transition piece be improved in terms of lifing and coolant
consumption compared to a baseline case using an optimized channel turbulation?

1.3 Methodology
The first three chapters (Chapter 2, Chapter 3, and Chapter 4) contain an elaborative literature
review, optimization terminology and numerical methodologies summaries necessary for the understand-
ing of the thesis contents. The literature review entails heat transfer mechanisms in combustion systems,
cooling techniques for gas turbines, with an emphasis on internal cooling, and calculation methods for
low-cycle fatigue.

The thesis paper can be split into two parts: the first part focuses on validation of the conjugate heat
transfer methodology and computational comparison between the two heat transfer approaches (uncou-
pled and conjugate). The findings of the validation studies (wall treatment and turbulence model) were
then applied to the GT transition piece, and subsequently the optimization of the cooling design of the
transition piece was performed. This is in the scope of the second part of this report. The structure is
presented schematically in Figure 1.5

Part I Part II
Transition
Background Canonical cases Conclusion
Piece

Chapter 1. Chapter 5.1 Chapter 7.1 – 7.4

Introduction Validation: Chapter 9
Domain,
Laminar flat plate Conclusions &
Chapter 2. parametrization
Recommendations
Literature review & flow validation
Chapter 5. 2
Chapter 3. Validation:
Optimization Turbulent offset Chapter 7.5
terminology jet Baseline cases
comparison
Chapter 4. Chapter 6.
Numerical Comparison of
methodology CHT and Chapter 8
uncoupled Ribbed channel
approach optimization

Figure 1.5: Thesis outline

In more detail, in Chapter 5, first, validation cases are introduced. The first validation study con-
sists of a simple laminar flow over a flat plate of finite thickness. The results at two physical settings
are compared to analytical solutions presented in literature. The second validation study is based on
experimental data of a turbulent offset jet over a flat plate. The study includes a comparison of popular

Energy & Process Technology Page 5

 Delft University of Technology

turbulence models (without parameter tuning) and the use of different wall treatment methods. The
resulting difference is discussed and recommendations are given and later used for the real-life case of the
TP. The next chapter (Chapter 6) includes a numerical comparison study on the two aforementioned
cases, where relevant physical or turbulence parameters are changed for the purpose of evaluating the
CHT solver’s sensitivity. The chapter ends by a discussion on the resultant differences between the
coupled and decoupled approach, by comparing relevant scalar fields in the vicinity of the wall and main
stream. To the author’s knowledge, no previous studies were conducted specifically for the comparison
between the uncoupled and conjugate method on a turbulent offset jet.

The second part of the study deals with the use of optimization methods for the design of an enhanced
cooling on a turbomachinery component. This is done by first comparing baseline cases of the currently
employed design by Siemens and new proposed turbulated channels, shown in Chapter 7. The turbu-
lation parameters are then fed into a proprietary hybrid optimization algorithm with the two competing
objectives of minimizing coolant consumption, and minimizing the surface averaged temperature of the
hot side of the component. The results are presented and discussed in Chapter 8. The second part
ends by proposing an adjustment in the cooling design that might lead to cooling enhancement. This is
done by the introduction of two design parameters, namely number of auxiliary cooling channels inlets,
and reduction of the area of the respective inlet downstream. Since the study is beyond the main two
objectives of the project, the results are resented in Appendix D

Finally, the conclusions for this thesis will be drawn in Chapter 9. Also, possible areas of improvement
are indicated in the form of recommendations for future research both in the field of cooling optimization
and CHT methods.

Energy & Process Technology Page 6

Chapter 2

Literature study

The aim of this literature study is to establish the current state-of-the-art research and practices,
relevant to the field of heat transfer in gas turbine combustion, conjugate heat transfer, and the sub-
sequent influence thermal loads resulting from combustion have on lifing in such components. Papers
addressing the life prolongation of such components, potentially by the enhancement or optimization of
some commonly used cooling techniques are discussed as well.

The study begins with basic overview of combustion systems and the resultant types of heat trans-
fers and heat loads which will effectively be applied to components in combustion systems, requiring
novel cooling methodologies. It further contains the history and techniques used for solving conjugate
heat transfer problems, namely a laminar flow over a flat plate and a turbulent offset jet. A proprietary
optimization algorithm is used in the thesis but it is important to keep track of what also has been
achieved in the past in the field of thermo(-mechanical) optimization and what frameworks have been
used in similar multidisciplinary investigations, and how do they compare to the approach and optimiza-
tion algorithm employed in this study. Fields and topics of potential research, where resources tend to
be deficient, are also pointed out.

The terminology of optimization and the most common types of optimization algorithms are outside
the scope of this chapter and are presented in detail in Chapter 3.

Herein after, purely resources depicting methods used in the heat transfer analysis performed by the
broad scientific community and research engineers in gas turbines are explained. Explanations of impor-
tant terms is included where deemed necessary.

2.1 Gas turbine combustion system basics

Gas turbines are internal combustion engines through which compressible fluids flow. They have a
high energy density and are used, for example, in the aviation and shipping sectors, or used as a mobile
drive system, and in stationary areas for generating electricity. A gas turbine basically consists at least
of a compressor, a combustion system and a turbine. Ambient air is sucked in in the compressor and
through several compressor stages it is compressed to a high temperature. The compressed air is mixed
with fuel in the combustion system and burned under almost isobaric conditions. The enthalpy of the
combustion gas is then extracted through the expansion of the gas in the turbine stage. The enthalpy
contained in the fluid is hence partially converted into rotational energy. Some of the energy is required
to drive the compressor, while any remaining usable rotational energy can be converted into electrical
energy with the help of a generator.

7
 Delft University of Technology

2.1.1 Structural basics

To analyse the heat loading and heat transfer mechanisms inside a combustion system, the system
structure and sequence of undergoing processes has to be understood first. Therefore, the combustion
system of interest in this project which is of the Siemens SGT5-8000H Heavy Duty Gas Turbine is shown
in the schematic below to illustrate key components and help visualize its operating cycle.

1 2 3 4 5

1. Main burner
2. Pilot burner
3. Flow sleeve
4. Flame tube
5. Transition

Figure 2.1: Main components in the combustion system of the Siemens SGT5-8000H [38]

The combustion system in Figure 2.1 is comprised of sixteen ring-shaped burners in total, arranged
circumferentially around the gas turbine. First, compressor air goes through the flow sleeve to the main
and pilot cone burners. Then, fuel is injected and mixed with the compressor air through the interior
of the pilot cone and swirled through the blades of a mixer, maximizing the homogeneity of the fuel-air
mixture and reducing the amount of pollutants released upon burning [131]. The purpose of the pilot
burner located in the center of the combustion system is generally speaking to suppress the inherent
combustion instabilities [76], which as a result leads to high frequency pressure fluctuations that exert
an additional strain on the structures of the combustion system. The mixture is then burned in the
flame tube. The transition piece (TP) is connected to the combustion basket (and hence flame tube),
from which the hot burnt gases are emitted. They are then guided and flow through the transition into
the first stage of the turbine blades. Hence, the purpose of the transition piece is to simply navigate
and accelerate the burnt hot gases into the turbine stage. The cross-section of a typical transition piece,
incl. the one used in this study changes from a circular to a quadrangle, the main purpose of which is
to enhance the uniformity of the temperature distribution at the outlet region of the transition piece.

A detailed look on the exact transition piece is not presented for confidentiality reasons. Similar TP
geometries to the one used in the study are found in open literature and are shown in Figure 2.2.

Energy & Process Technology Page 8

 Delft University of Technology

(a) (b)

Figure 2.2: Typical transition piece geometries

2.1.2 Flow arrangement

The combustor system of the studied TP employs the so-called reverse-flow arrangement, which has
been reported to lead to weight reductions and other technical benefits such as compactness and en-
hanced maintainability [25]. This is a plays a large role in the thermal management of the TP due to
the complexity of the cold compressor flow in the region outside the transition piece, commonly referred
to as the dump diffuser.

When it comes to literature regarding heat transfer inside the dump diffuser section in such an ar-
rangement, relatively limited studies are present. The oldest study found dates back to 1996 done by
Kapat [61] and Zhou at al. [132]. In the former a sub-scale diffuser model was used to study exper-
imentally the cold flow of the diffuser section reporting various causes of pressure losses and features
of the flow in key regions. The first computational study presented [132] yielded useful fundamental
information regarding the complicated flow structures present within a dump diffuser. The following
study of Kapat reported that air extraction does not accentuate flow nonuniformities, necessary in the
avoidance of hot spots on TPs [60]. Wang at al. later investigated the effect of extraction and concluded
that the propensity for flow reversal increases if the location and amount of air extraction is not opti-
mized for [124]. The most prominent recent studies available are performed by Wang et al. [122], in a
two-part paper with an experimental study and a numerical simulation part. In the former, Wang et al.
showed that the usage of a perforated sleeve (sheath), forming impingement cooling on the surface of the
transition results in a considerable reduction of transition wall temperature at the expense of an equal
increase in pressure drop, and in [123] it was shown that a similar sleeve can help harness temperatures
especially on the upper section near the turbine end.

The subsequent subsections will address the topic of types of heat transfer and the magnitude of heat
loads on combustion equipment for gas turbines of similar power output as the one of this study, ending
with a short discussion on the few key studies of transition piece heat transfer.

2.1.3 Heat transfer in combustion systems

Due to the comparatively low interest in research in gas turbine TPs, the section provides an overview
on heat transfer mechanisms found in the entirety of the gas turbine combustor, and not exclusively in
TPs. Hence, for the sake of this literature study, papers discussed entail heat transfer research aimed at
all components exposed to combustion gases (excluding. turbine vanes). It is also worthwhile mentioning
that due to the extensiveness of the field of heat transfer, generally only a small proportion is focused
exclusively on the subject of combustion. The majority of these studies are conducted on a completely
practical point of view. Therefore, a short discussion on the fundamental physics is of benefit and pre-

Energy & Process Technology Page 9

 Delft University of Technology

sented in the upcoming subsections. Moreover, the main focus when considering combustion processes
in industrial applications lies solely on gaseous radiation heat transfer and not heat transfer in general.
Chapter 3 of the book given in [57] by Buakal et al. presents a concise summary of heat transfer in
industrial combustion processes. A book by Schulz [44] lays a review in Chapter 4 mostly focused on
radiative heat transfer and convective cooling instead, with a brief discussion on numerical modelling
approaches.

Combustion flows are inherently of considerable complexity. They consist of a multitude of phenomena.
According to Viskanta [120], the processes can be summarized as: convective and radiation heat transfer,
molecular diffusion, turbulent/laminar multiphase fluid dynamics, chemical kinetics, nucleation, phase
change heat transfer (i.e. condensation and evaporation), and surface effects. Each of these processes can
be treated separately in order to produce an accurate mathematical combustion model. This has received
substantial research interest in the recent years and state-of-the-art combustion models are present in
literature [27], [52]. Each of the aforementioned mechanisms plays a partial contribution in accordance
to the specific case and conditions, and will therefore be examined separately.

Convection
In the case of heat transfer between solid and fluid, in cases where their respective temperatures
differ, one speaks of heat transfer by convection.It is a combination of energy transport by a macroscopic
movement of a fluid and individual random movements of molecules. Convection is driven by temperature
difference and the process can be simply expressed by Newton’s law of cooling:

q = h · (Tw − Tf ). (2.1)

The heat transfer coefficient (from here after referred to as HTC) h determines the intensity of the heat
transfer at an interface. The HTC is a related heat flux density, i.e. a quantity that is shared by all
variables of the process under consideration is influenced. Important influencing variables are geometric
conditions, flow speed, type of flow and surface properties. [33]. At the wall one can express the heat
flux balance as:  
∂T
h · (Tw − Tf ) = −λ · , (2.2)
∂n wall
and hence
−λ(∂T /∂y)|y=0
h= . (2.3)
Tw − Tf
The first principal dimensionless number which is essential for heat transfer analysis is the Nusselt
number. Heat transfer coefficient is proportional to the Nusselt number, which characterises the thermal
boundary layer, typically expressed as:
λ
h= · N u(Re, P r), (2.4)
l
with λ as the thermal conductivity of the fluid, l as the characteristic length, Re as the Reynolds number
and P r as the Prandtl number:
ρvl Inertial forces
Re = ∼ . (2.5)
µ Viscous forces
cp µ Momentum diffusivity
Pr = ∼ (2.6)
k Thermal diffusivity
Traditionally, the Nusselt number is the ratio of:

h·l Stagnant fluid conductive resistance

Nu = ∼ (2.7)
λ Convective resistance

Energy & Process Technology Page 10

 Delft University of Technology

and if the temperature gradient given in Eq.2.3 is nondimensionalized, and the resultant expression for
α is used, the following expression is obtained
 
d TTf−T
−Tw
w

Nu =
(2.8)
d nl
wall

Hence, the Nusselt number can also be regarded as the non-dimensional temperature gradient at the
wall [85].

The calculation of the heat transfer coefficient is in reality of combustion flow is fairly complicated
and most often a correct definition does not exist for the exact configuration of interest. Some heat
transfer correlations related to combustion are mentioned below, but such empirical formulations are
only useful in preliminary design if ever, as they provide only averaged quantities over the combustor
walls, and in most industrial cases, a more detailed temperature distribution along the wall is needed.
This is only achievable by numerical models for combustion applied with the appropriate level of sophis-
tication.

Relatively little research interest is present when it comes to convective heat transfer correlations for
(can-)annular combustion systems. The convective heat load in such an arrangement is non-uniform
and strongly depends on reaction and swirling flow dynamics. Lefebvre et al. ([72], [74]) performed the
first attempts to derive semi-empirical correlations aimed at heat transfer along combustor liner walls.
However, Lefebvre also stated that the the airflow in combustor liners exhibits a great amount of un-
certainties, including “the state of boundary layer development, and effective gas temperature”, which
effectively makes the choice “of a realistic model almost arbitrary”. Generally, same applies to TPs as
well. Hance, it was argued that when no film cooling is applied for the liner/TP, the correlation should
be comparable to fully developed turbulent channel flow. Lefebvre and Ballal hence proposed [74]:

k∞
hc = 0.02Re0.8
c,∞ . (2.9)
Dc
The vast majority of heat transfer studies in reactive flows for industrial applications are in fact belong
to the popular jet impingement problem, which has a wide variety of industrial applications and can be
considered to some extent analogous to processes in gas turbine combustors. For a turbulent impingement
jet, the correlation is given by Hustad [49]:
(  0.25 )
lj 0.6 0.35 0.15 Pre
Nu = 0.41 Reb,e Pre Tu (2.10)
λe Prw

particularly intended for flames by jets of CH4 and C3 H8 . A full comprehensive summary of various jet
impingement studies of flame flows prior to 2005 under a great variety of flow conditions and geometry
properties is presented by Chander and Ray, [20]. Griswold proposed a correlation for the heat transfer
coefficient for forced convection from the outside of a combustor wall, with a velocity of v against the
surface, in his book Fuels, Combustion and Furnaces [37]:
1 + 0.225v
hc = (2.11)
(tw − t∞ )

The author also proposes an empirical formula for natural convection from outside combustor wall as
well:
0.27
0.53C (Tw − T∞ )
hc = 0.18 (2.12)
[(Tw + T∞ ) /2]
where C is a shape constant (e.g. 1.79 for an arch, and 1.39 for a vertical wall).

Energy & Process Technology Page 11

 Delft University of Technology

A more commonly used correlation in design of gas turbine combustors makes use of a modified form of
the well-known Dittus-Boelter correlation [11], which is the empirical relation for heat transfer coefficient
in a fully developed turbulent flow in a circular tube:
nT
NuDittus-Boelter = 0.0235Re0.8
C,f Prf (2.13)
The exponent nt is set to 0.3 in cases where the wall temperature is lower than the free-stream temper-
ature. Otherwise it is 0.4. The correlation holds well for flows between 10, 000 < Re < 120, 000, and
temperature differences of wall to gas of only 56K, and flows properties also being taken as temperature
independent [75]. The gas properties are then evaluated at the film temperature. Lefebvre and Ballal
state that such modifications can vary from one use-case to another, and it should also be noted that
the correlation they derived (see Eq.2.9) is closely comparable to the Dittus-Boelter relation [74]. It
has been reported that the Dittus-Boelter correlation underpredicts the heat transfer coefficient in real
combustion processes, which is the case of the work of Patil et al. [94], where the author studies reactive
flow in a simulated scaled up annular combustor.

Since the formulation of the Dittus-Boetler correlation (published in 1930), a considerable understand-
ing has been gained when it comes to turbulent pipe flow heat transfer correlations. The addition of
flow swirl also helped understanding the characterization of heat transfer, particularly in gas turbine
combustors [127]. Heat transfer enhancement due to abrupt flow expansions has also been an insightful
endeavour in the understanding of heat transfer in combustors, mainly due to the presence of corner
recirculation regions upstream of the flow reattachment points ([7], [26]).

Conduction
When it comes to the heat transfer in the form of conduction, it is usually considered to be of smaller
influence on the calculation of heat transfer in combustion systems. Conduction is physically described
as the energy transfer occurring due to collisions of more-energetic particles with lower-energy particles
[23]. It can occur in gases, liquids and solids, where in each medium the underlying principles behind
the conduction process differ [23]. It is generally expressed with Fourier’s law of heat conduction, which
in three dimensions reads:
q = −λs ∇T (2.14)
where λs is the thermal conductivity of the solid region (e.g. combustor walls). The process is over-
whelmingly omitted in gas turbine combustion studies. On the other hand, it can become crucial when
calculating the heat losses through the outer wall of the combustor to the ambient. Conduction is in the
basis of conjugate heat transfer studies, as it will be shown later, but as explained earlier, such studies
are relatively few in the field of gas turbine combustor design.

Conduction is usually more carefully examined in microturbine combustion systems or when the surface-
area-to-volume ratio of the system examined is considerably larger than in large gas turbines. In such
systems the heat loss from the outer wall can become large and hence sensible calculation of thermal
efficiency becomes possible only when conduction is included in the analysis [16].

Again, similarly to convection heat transfer, conduction has been more widely studied in problems
involving flame impingement. Arrangement with internally cooled ([8], [9], and others) and uncooled
target surfaces ([32], [42], [108]) have been used for such studies, where a steady-state conduction through
the target surface has been produced.

Furthermore, Baukal [57] also noted that transient effects of conduction is usually also neglected, how-
ever it should be subjected to more research particularly in the firing up phase of the combustor. Other
studies that deal with conduction in gas turbine combustion systems focus on the field of contact ther-
mal resistance, which is closely related to the study of thermal barrier coatings (TBCs) and thermal
delamination, macro-cracking or spallation. Some studies reported in open literature entailing the use
of ceramic composites for gas turbine combustion applications are present in the works of Corman et al.
[24], Miriyala et al. [86], Nelson and Orenstein [89], Price [98], and others.

Energy & Process Technology Page 12

 Delft University of Technology

Radiation
Radiation is a fairly complicated subject and has received substantial amount of research interest over
the years. Viskanta and Menguc [120] noted that the main reasons behind the complexity of radiation in
combustors is the coupled nature of multiple phenomena: (i) as the system undergoes different operation
conditions, the surface effects change as well, (ii) turbulent mixing of gases and particles, and finally (iii)
irregularities in the temperature field. Soot formation, particle size distribution and agglomeration have
also been discussed as relevant factors too [66]. As pointed out above, radiation heat transfer is often
times discounted in thermal analyses of combustion systems, and are of little to no relevance in TPs.
Studies that validate such a choice are in [8], [20], and [110]. For that reason, herein, only a very basic
overview of the physics behind radiation is described, and the reader can refer to the review of Baukal
[57] for a more detailed study, where a variety of cases are discussed.

Radiation in general is the energy emission by a matter in solid, fluid or even gas state, primarily
due to the change of electron configuration in the molecules, which as a results reduces its internal
energy. This release of energy is done by a descrete energy quanta, i.e. photons, with an energy of
hf = hc/λ, where h is Plank’s constant and f is the frequency of the energy. Additional efforts were
put by Viskanta [120] in the understanding of radiative heat transfer (described briefly in the upcoming
paragraphs) along the contribution of Lefebvre [73].

q = σT 4 (2.15)

σ = 5.67 · 10−8 W/m2 K. (2.16)

The maximum heat flux from radiation is described by the law of Stefan-Boltzmann, and its magnitude
depend sheerly on its temperature, given in Eq.2.15

The upcoming subsection is solely focused on (semi-)empirical methods for estimating the resultant
heat loads of the previously described heat transfer mechanisms.

2.1.4 Thermal loads

While a considerable level of understanding on heat transfer phenomena has been gained over the
years, less light has been shed on combustor heat loads. With the growth of numerical methods over
the years, simulation models have become increasingly better and the application of Reynolds-Average
Navier Stokes (RANS) models to simple domains allow for quick results regarding heat loads in com-
bustors. When combined with supporting experimental measurements for validation, this has become a
strong toolchain for designers of gas turbine combustors.

To the author’s knowledge, the works of Lefebvre [74] and Gosselin et al. [36] present the only em-
pirical calculation methods of the heat loads on combustion components. A schematic is presented in
Figure 2.3. Using the assumption of a steady-state heat transfer on the wall, the heat transfer into the
wall can be balanced out by the heat transfer leaving, as given in Eq. 2.17. The emperical analysis is
summarized in the subsequent equations.

R1 + C1 = R2 + C2 = K1−2 (2.17)

λw
K1−2 = (Tw1 − Tw2 ) (2.18)
tw


R1 = 0.5σ (1 + εw ) εg Tg1.5 Tg2.5 − Tw1
2.5
(2.19)
Tg = T3 + ∆Tcomb (2.20)
−290P L(qlb )05 T8−1.5
Tg = T3 + ∆Tcomb εg = 1 − e (2.21)
L = 0.0691(C/H − 1.82)2.71 (2.22)

Energy & Process Technology Page 13

 Delft University of Technology

4
R2 = Zσ Tw2 − T34 (2.23)

 0.8
λg ṁg
C1 = 0.020 0.2 (Tg − Tw1 ) (2.24)
DL AL µ g
 0.8
λa ṁan
C2 = 0.020 0.2
(Tw2 − T3 ) (2.25)
Dan Aan µa
The term of the internal radiation is also comprised of soot and gas, as indicated in the schematic. The
two emissivities, of the wall and gas, respectively, are ϵw and ϵg . The temperature of the combustion gas
Tg is found by the sum of the annulus (compressor air) temperature and the resultant temperature rise
from the combustion process (∆Tcomb ), available in standard combustion tables for this purpose. The
value of the constant Z is dependent on the liner casing emissivity, which is interchangeable from one
material to another (0.4 for aluminium alloys and 0.6 for steel).

Figure 2.3: Schematic of heat transfer loads in a combustor liner. Extraced from [6]

The equation above hold only if the liner is uncooled. Thereby, in cases where film cooling is utilized,
only the convective heat load on the inside changes (C1 ), due to the resultant changes in the flow
characteristics in the vicinity of the wall [74]:

(ρU )a x
m= , Rex = Ua ρa (2.26)
(ρU )g µa

λa 0.7
if 0.5 < m < 1.3 : C1 = 0.069Rex (Tw,ad − Tw1 ) (2.27)
x
λa  x −0.36
m > 1.3 : C1 = 0.10 Re0.8 x (Tw,ad − Tw1 ) (2.28)
x s
The adiabatic wall temperature is calculated using the cooling effectiveness:

Tw,ad = Tg − η(Tg − Ta ) (2.29)

Lefebvre proposes two approaches for the calculation of the cooling effectiveness: turbulent-boundary-
layer-based or wall-jet-model-based.

Energy & Process Technology Page 14

 Delft University of Technology

The empirical relations concerning the final energy balance was also validated by Lefebvre and Her-
bert [71], in which both internal and external radiation were accounted for. The results were compared
to engine wall temperatures from an experiment on the primary zone. The resultant agreement was good
at various operating conditions. On the other hand, the validation study did not include an analysis on
the peak heat loads or distribution of the temperature field, caused by limited experimental data.

The approach from Gosselin et al. [36] does not alter the empirical relations used until now. Only
a small modification is that the combustor was divided into four zones in their study: recirculation zone,
primary zone, dilution zone. The loads were thereby calculated for each zone.

The complexity of the equations in both approaches accentuates the challenge of the selection of ap-
propriate reference mass flows, velocities and temperatures for relatively accurate calculations. For
example, in the early analysis of Lefebvre and Herbert [71], the mass flow in all crucial zones (combustor
inlet flow, primary zone flow, and swirler mass flow) was either estimated or calculated, which was in fact
crucial for the good agreement. Such data however most often is not available without prior experimental
or simulation work. The relations however, are still useful and are used to this day at early design stages
(e.g. work of Ramierz [102]).

2.2 Conjugate heat transfer problems

The term conjugate heat transfer (CHT) was first coined in 1961 by Perelman [96], addressing the
problem of heat transfer between a fluid and a solid, with an unknown initial interface condition, evaluated
from the solution of the heat transfer problem. The term conjugate heat transfer entails the simultaneous
solution of both conduction and convection, within the solid and in the thermal boundary layer close to
the wall, respectively. A one-dimensional CHT problem is depicted in Figure 2.4.

𝑣
𝑣∞ 𝑇
𝑇𝑓
𝑇𝑤𝑎𝑙𝑙 ℎ
𝐿 𝑞𝑤𝑎𝑙𝑙 𝜆𝑠

𝑇𝑠
𝑙
Figure 2.4: The one-dimensional CHT problem schematic

In addition to the Nusselt number described in Eq. 2.7, the second characteristic non-dimensional
number for CHT problems is the Biot number. It describes the ratio of the conductive over the convective
thermal resistances of the problem.
h·L Conductive resistance
Bi = ∼ (2.30)
λs Convective resistance
Consequently, a large Biot number could be used to indicate a high temperature gradient in the wall,
whereas in case of a low Biot number the temperature gradient in the boundary layer is the highest due
to to the high convective thermal resistance. This is also illustrated in the figure below.

Energy & Process Technology Page 15

 Delft University of Technology

𝑣 𝑣
𝑇𝑓 𝑇𝑓

𝑇𝑤𝑎𝑙𝑙 𝑇𝑤𝑎𝑙𝑙

𝑇𝑠 𝐵𝑖 ≫ 0 𝑇𝑠 𝐵𝑖 ≅ 0

Figure 2.5: Temperature gradients at low and high Biot number settings in fluid and solid regions

The fundamental difference with the Nusselt number is that in the latter case, the resistances are
related entirely to the f luid region, whereas for the former, it is the convective resistance of the f luid
and the conduction resistance of the solid region.

A comprehensive summary of all CHT work prior to 2018 is well presented by John [54], where analyt-
ical studies are discussed with relevant theory, and a set of some of the most important computational
studies is included, along with a strong direction to their application envelope. According to the author,
more emphasis is put on the computational side of CHT and presents various coupling techniques in
more detail. The review exhibits CHT as a great approach for achieving high-fidelity results for complex
problems, where analytical solutions do not work. The author also identifies the coupling algorithms as
critical part of the solution and distinguishes two types of coupling: (i) monolithic , and (ii) partitioned.
For a more mathematically-focused explanation of the conjugate heat transfer, the reader can skip to
Section4.1.

2.2.1 Flat plate

The analysis of conjugate heat transfer problem using a solid plate of finite thickness has been the
subject of research for a plethora of authors over the years, particularly in the early stages of CHT prob-
lems where the first analytical solutions were derived. The variety in proposed analytical solutions is
mainly due to the assumptions made for the velocity and temperature profiles in the boundary layer, the
most popular assumption of which is a linear velocity profile within the boundary layer. The complexity
of the solutions increased with more complicated geometries and hence did not present any practical
feasibility. In this section, both laminar and turbulent studies are included.

The first found publication with an experimental setup was published by Soliman [114]. The author
reported transient temperature responses with varying Reynolds numbers in the range of 500, 000 <
Re < 2, 000, 000. A well known problem in open literature is the Luikov problem [1], in which an
analytical solution was derived for a laminar flow over a plate with a bottom surface set at constant
temperature. The nondimensional Brun number (Brx = (λf /λs )(b/x)P rm Renx ) was introduced, and it
was noted that for Brun numbers > 0.1 (ratio between thermal resistance of the plate to that of the
convective boundary layer), the problem should be solved using the CHT method. The author also later
derived expressions for laminar flow in circular and planar tubes. Karvinen [62] derived new results for
conjugated heat transfer in a plate for both laminar and turbulent flow regimes. The author presented an
approximate method based on simplifications of the governing equations. The model was then compared
to experimental data by Sohal and Howel [113], in which the influence of radiation heat transfer was
taken into account. Later, Yu et al. [125]. derived a new solution method, which proved to be effective
for any Prandtl number. A proposed Nusselt number correlation was also compared to numerical data.
In the following study of Yu [129], additional conjugation parameters were introduced that seemed to
match well with numerical data from Keller’s box method differencing scheme [19]. Vynnycky, M. et al.
[121] proposed two analytical methods: an averaging method and a boundary-layer-based method, both
compared to numerical solutions at a wide range of three non-dimensional parameters (P r, Re, λ). The
methods allowed for a two-dimensional conduction in the plate, compared to Luikov’s solution where

Energy & Process Technology Page 16

 Delft University of Technology

the assumption for one-dimensional conduction is held. Mosaad, M. [87] proposed another analytical
solution based on the observation that the upper surface temperature of the plate is independent of the
Prandtl number. The condition for the Brun number above which conjugation is important was hence
adjusted from 0.1 (Luikov) to 0.15.

In this report, a laminar case is considered for the analysis. However, turbulence is of utmost prac-
tical importance in industry, and a summary of the work on turbulent CHT over flat plates is included
in the following paragraph.

Ames & Moffat [3] performed an experimental turbulence investigation over a flat plate problem includ-
ing conjugate heat transfer measurements. Thirty three years later Karvinen’s publication mentioned
above, the author presented a new semi-anaytical solution for a class of flat plate conjugate convective
heat transfer problems [63]. These included the popular Luikov problem, uniformly generated heat on
one surface, and a modified Luikov problem where both surfaces are under a convection boundary con-
dition. The author used the superposition principle and analytical expressions to couple the heat flux
and temperature distributions. A similar approach was followed by Shariatzadeh [56], where the author
also provided a semi-analytical solution for a conjugate turbulent forced convection boundary layer flow
over plates. The solution was based on the Differential Transform Method (DTM) used for solving
a non-linear integro-differential equation. Later, Hajmohammadi [39], using a temperature-dependent
thermal conductivity for the plate, performed numerical simulations for the sake of validating the model
of Shariatzadeh. In the paper, the author reported a single case in which a horizontal plate is heated with
uniform heat flux at the lower surface while being cooled at the upper surface under a laminar forced
convection flow. Lastly, Lindstedt & Karvinen [78] presented a new simpler semi-analytical solution for
problems with a surface at a constant temperature and forced/natural cooling applied on the other. It
is reported that the simplicity is achieved by assuming only one-dimensional conduction in the solid,
which leads only to small deviations from previous studies very near the leading edge, which in practical
engineering cases can be ignored.

Summary
The conjugate heat transfer problem over a flat plate has received considerable attention due to the
inherent geometric simplicity. An overview of the key papers discussed above is presented in Table 2.1
below.

Table 2.1: Review of CHT studies on flat plates

Author Analysis Flow type Comments

Soliman (1967) Experimental Turbulent Transient temperature response with 5e5 <
Re < 2e6
Luikov (1972) Fully analytical Laminar Analytical solution for a plate with a constant
temperature profile on the bottom surface
R. Karvinen (1978) Semi-analytical Turbulent/Laminar Comparison to experimental data by Sohal,
combining conduction, convection and radia-
tion heat exchange
K. Cole (1996) Fully analytical Laminar Laminar shear flow over a small heated strip
for electronic cooling; new scaling law using
Pe number
Karvinen (2011) Analytical Turbulent/Laminar Power sum series for different configurations
Shariatzadeh (2013) Semi-analytical Turbulent Analytical solution using DTM
Hajmohammadi (2013) Numerical Turbulent/Laminar Numerical FVM comparison to DTM analyt-
ical calculation
Lindstedt & Karvinen (2013) Analytical Turbulent/laminar Bottom isothermal surface, comparison with
previous studies

• (Semi-)analytical models developed over the years for turbulent and laminar CHT over flat plates
yield good agreement with numerical and experimental data.
• At the groundwork of the models however lie fundamental assumptions, which lead to discrepancies
more apparent at the leading edge of the plate

Energy & Process Technology Page 17

 Delft University of Technology

No specific criterion was found w.r.t a conjugation requirement: most of the studies entail a comparison
with the uncoupled approach using a narrow range of parameter variation.

2.2.2 Turbulent offset jet

The turbulent offset jet is a comparatively well researched problem with a multitude of experimental
research data, including turbulence and flow characteristics, heat transfer and conjugation set-ups. To
the author’s knowledge, the majority of available CHT data comes from practical experimental sources.
Plenty of computational studies, however, exist in open literature too. Some of the key studies including
a conjugate heat transfer approach are included below.

The first reported experimental study found in literature was by Nizou [90], where experimental in-
vestigation of a wall jet (i.e. offset ratio = 0) with a constant wall heat flux condition was conducted.
The heat transfer and momentum transfer for such arrangement were correlated and the difference be-
tween a wall jet boundary layer and turbulent boundary layer. In the same year, Hoch and Jiji [45]
studied the same problem with an additional imposition of freestream velocity, later providing an ana-
lytical solution for a laminar case. One of the most insightful studies in the matter was published later
by Pelfrey and Liburdy [95], with a detailed analysis of the mean flow and turbulent characteristics at
multitude of offset ratios (OR).

Figure 2.6: The turbulent offset jet problem. Extracted from [95]. The flow can be characterized by
three regions where the flow exhibits different behaviour. The recirculation zone (which is upstream
from the reattachment point), the impingement region, where the reattachment length is measured, and
the wall jet region where the flow is practically identical to a turbulent jet with an OR of 0. The wall
jet region is free of gradients of mean pressure, and is where the flow decelerates and spreads.

Briefly after the study of Pelfrey and Liburdy, Holland and Liburdy [47] used the same experimental
setup to study the thermal characteristics of offset jets. The key findings of the study concertn the reat-
tachment length (RL), and resultant interface temperature profiles, Nusselt numbers, and jet velocity
decay. RLs were found to coincide with the maximum Nusselt number locations, whereas the maximum
temperature location depends on the offset ratio, and approaches the RL with increasing ORs. The
interface temperature reduces/increases rapidly up to the reattachment point and thereafter it increas-
es/decreases. A typical profile is shown in Figure 2.7 below. The term Ir was introduced to scale the
axial location with respect to the RL for an OR=7 (i.e. for OR = 3, Ir = 1.77, as RL = 7.02, and
RL = 12.42 for OR = 7). Moreover, thermal energy within the recirculation zone is relatively uniform
for all ORs.

Energy & Process Technology Page 18

 Delft University of Technology

Figure 2.7: Dimensionless wall temperature on the wall as a function of offset ratio (θw = (Twall −
′
T∞ )/(Tex − T∞ ), x = xIr ) for a conjugate heat transfer case [47].

Vishnuvardhanarao and Das [118] performed an extensive numerical study by simulating the experi-
mental work of Holland and Liburdy using solely a k − ϵ turbulence model, with a detailed discussion
on velocity and temperature distribution in the domain are discussed. The same authors [119] then
performed another CHT on the same domain by imposing constant flux at bottom surface and sweeping
a number nondimensional numbers. The third paper from the sequence of studies by Vishnuvardhanarao
[29] was a CHT computation with a constant bottom wall temperature, investigating the influence of
P r number, offset ratio, thickness ratio, and conductivity ratio on the thermal and flow fields. A more
recent study using CHT was done by Fu [31] where results from various two-dimensional RANS models
were compared to LES data using STAR-CCM+. The author expressed a slight favour in the face of the
standard k − ϵ model.

Summary
Overall, the problem of a turbulent offset jet is wall established and a multitude of key findings w.r.t
heat transfer and flow behaviour have been solidified over the years of research. This includes:

• The static pressure distribution in the recirculation region is independent at sufficiently large
Reynolds numbers (Re > 11 · 103 ).
• The reattachment length does not depend on Re above Re = 15, 000.
• Not much focus on flow physics in the impingement region; studies agree on a rapid decay of e
local maximum streamwise mean velocity Um , and an increase in wall-normal jet half-width y0.5 ;
increasing OR seemed to lead to an increase in the streamwise velocity decay rate.
• Both lateral on wall-normal jet spreading are generally independent on OR.
• Overall, both k − ϵ and SST − ω models are able to predict the recirculation region extent rea-
sonably well; turbulence modelling predict the loci of the maximum streamwise velocity decay,
wall-normal location of Um , wall-normal half-width in the outer shear layer and the wall static
pressure distribution, partly due to the isotropic assumptions made in the models.
• The location of maximum Nusselt number coincides with the reattachment point for OR = 5, and
decreases downstream; a second increase is present upstream near the corner region. The maximum

Energy & Process Technology Page 19

 Delft University of Technology

and minimum locations of the Nusselt numbers, however, shift progressively downstream as the
offset ratio is increased.

In general, the amount of conjugate heat transfer studies is low, compared to the flat plate case. To
author’s knowledge, no comparison between the uncoupled and conjugate approach has been performed
in the past, particularly for the case of an turbulent offset jet.

2.3 Combustion systems cooling

The field of cooling studies in the field of gas turbines is one of the major research pillars in convective
heat transfer in general, with the age-old goal of augmenting heat transfer capabilities without resulting
in a penalty from a pressure drop point of view (the increase in pressure drop ∆p, however, directly
proportionate to the coolant mass flow rate, such that ṁc ∝ ∆p). According to Bunker [15], the field
comprises of five main subfields: (i) internal convective cooling, (ii) external surface film cooling, (iii)
materials selection, (iv) thermo-mechanical design, and (v) selection/conditioning of the coolant. The
subsection below would only focus on internal and external cooling geometries. Before more detailed
specifications of the geometries are given, it is worthwhile mentioning that in most gas turbine cooling
cases, both external and internal coolant flows are fully turbulent. It is a well-established fact that
turbulence enhances mixing [4] and hence accelerates heat transfer which is fully exploited in the design
of such components. A large proportion of studies on convective cooling in gas turbine components
focuses on turbine vanes, but the techniques used are also applicable to other moving or stationary
components in a combustion system (e.g. combustion liner or transition piece).

2.3.1 Internal cooling

Arguably, the most common mechanism reported in literature for internal channel cooling techniques is
the use of different turbulation topologies. With the advent of additive manufacturing, and the resultant
dimensionality reduction of the cooling, some turbulator arrangements have become increasingly more
complex [15]. Examples of internal cooling topologies achievable only by non-conventional methods like
AM are shown in Figure 2.8.

(a) Matrix cooling system (b) Longitudinal turbulation cooling system

Figure 2.8: Two examples of internal cooling design for GT turbine vanes; for reference, vane length is
12.75mm [15].

A fully detailed summary is given in the works of Ligrani [77]. The article mentions a multitude of
rib, fins, dimple and swirl generator geometries with a full comparison and discussion on relevant papers.
The review article eventually concludes that from the comparison between the cooling methodologies
obtained prior and after 2003, based on their respective thermal performance and friction factors, a only
a minimal improvement has been achieved in providing a better “spatially-averaged thermal protection.”

Energy & Process Technology Page 20

 Delft University of Technology

The main types can be divided into the following sections: (i) pin fin arrays, (ii) dimples, (iii) rib tur-
bulators, (iv) swirl chambers, (v) and combination devices, where swirl chambers have not undergone
major developments since 2003.

With regards to pin fins specifically, the geometry can be described as a pedestals generally arranged
into arrays and extending between two opposite walls of an internal cooling passage. Studies report
comparatively high pressure drop compared to other turbulators. Multiple types of cross-sections are
investigated in literature Some studies suggest that ‘‘the highest overall heat transfer coefficients, which
are spatially averaged over all pin fin and endwall surfaces, is produced by the triangular pin fin ar-
rangement’’ [111]. Siw et al. explains this by the increased shear and wake shedding caused by the
sharp edges on the triangular elements. The author examined the thermal performance of pin fins of
three cross sections, shown in Figure 2.9. This in turn enhances flow mixing, and thus an increase in the
turbulence transport is achieved.

Flow direction

Figure 2.9: The three configurations used in the study of Siw et al. (top to bottom: triangular, semi-
circular, circular) and the dimensional heat transfer coefficient distributions (domain dimensions width,
W =63.5mm; height, H=25.4mm; length, L=101.6mm) [111]

Tabulated ribs are similar to rectangular (not exclusively) bars mounted along a surface, at a certain
angle with respect to the bulk flow. Flow recirculation around a single rib are positioned just downstream
of each rib, and cover a considerable portion of the ribbed wall depending on the pitch/height ratio of
the ribs. Studies suggested that each of the recirculation zones is often considered to be a potential hot
spot because it is associated with locally lower surface Nusselt numbers. Parameters that influence the
performance of ribbed channels are namely the channel aspect ratio (wc /hc ), the rib blockage ratio (the
rib height to channel height e/Dh , the spacing expressed using the pitch-to-height ratio (p/e), the cross
sectional shape of the rib, the orientation of the ribs on the opposite wall, and the shape of the rib across
the channel (straight, V-shaped etc.). An example configuration of a study by Casarsa [17], is shown in
Figure 2.11.

The angle with respect to the bulk flow has been the subject of multiple studies. Some studies of
Lau et al. [68], [69] suggested that the best heat transfer augmentation is achieved when the angle is
equal to 45◦ , due to the stronger secondary flow induction, compared to other angles. With regards to
the cross section for the rib, it has been reported that the best cooling performance is achieved by the
so called backward-aligned delta-shaped ribs. The shape resulted in the lowest pressure drop and high-
est Nusselt number augmentation (N u/N uo ). Alternatively, wedge-shaped ribs also exhibit promising
results in studies. However they remain outperformed by backward-aligned delta-shaped ribs. The two

Energy & Process Technology Page 21

 Delft University of Technology

types are presented in Figure 2.10 below.

a) wedge-shaped rib b) backward-aligned delta rib

Figure 2.10: Rib shape configurations reported in [41]

Dimples on surfaces can also be protrusions, whereas multiple studies on flow structure resultant from
various shape of the dimple/protrusion are given in literature. An agreement on the flow structure on
dimples is lacking and is still an ongoing research, yet all studies reported an increase in heat transfer.
Generally, an increase in relative dimple depth produces increases in surface heat transfer rates, as
well as increases in pressure penalties downstream. Multiple papers however report the same overal key
features responsible for local heat transfer augmentation: (i) shedding of multiple vortex pairs, (ii) strong
secondary fluid motions within the vortex pairs, (iii) shear layer reattachment within the dimples.

Figure 2.11: Turbulence intensity contour plot from [17], indicating three key recirculation regions of low
HTC where hot spots occur, for flows over ribs (Re = 3 · 104 , e/Dh = 0.3, p/e = 10)

The main recent advancements in the internal cooling of gas turbine components and latest research
interest comes with the simultaneous use of different augmentation technologies in one channel. Dim-
ples together with rib-turbulators are for example investigated by Amano et al., whereas pin fins and
rib turbulators is given in Siw et al. Combinations of three geometries is also recorded. urata et al.
investigates passages with ribs, dimples and protrusions. Other studies also focus on the addition of
external cooling to such turbulation topologies. For example Yang et al reports effusion cooling (see

Energy & Process Technology Page 22

 Delft University of Technology

next subsection) with internal ribs. Results indicated that pin fin-dimple channels yielded a higher HTC
augmentation than purely pin fin channels. Each combination reported has a unique flow field and the
resultant thermal (and hence Nusselt number) fields is explained by the accompanying flow structures.
For a more detailed outlook, the reader can refer to the abovementioned studies.

Figure 2.12: Comparison of heat transfer performance (heat transfer augmentation N u/N uo on y-
coordinate versus friction factor f /fo ) for various types of internal cooling. Extracted from [77]

A direct comparison of each of the arrangement can be easily observed in the scatter plot given by
Ligrani et al. [77]. In the plot it becomes evident that the combination devices are not included. This
is due to the fact that the abovementioned studies did not include friction factor ratio data. A smooth
channel is also given to serve as a reference point for comparison. The scatter plot suggests that the best
tradeoff and hence best thermal performance is achieved for the rib geoemtry presented by Salameh and
Stunden [104]. Fractured and truncated ribs on the other hand showed very little increase in pressure
drop, but the resultant increase in heat transfer coefficients was not fully satisfactory either. On the far
right and middle of the plot are the pin fin arrangements which are more sensitive to pressure drops as
mentioned earlier.

2.3.2 External cooling

Since the goal of this study is mainly focused on an optimization of internal cooling passages, the
subject of external cooling is of little relevance. For the sake of logical coherence and completeness, a

Energy & Process Technology Page 23

 Delft University of Technology

brief overview is presented here.

The main types of external cooling for gas turbine components reported in literature are film cool-
ing, impingement cooling, and transpiration cooling. Jet impingement is a mechanism that utilises a
supply plenum air at an overpressure and impingement holes on an orifice plate, which serves for flow
confinement and is located at an offset from the heated surface. Areas of exposure to high thermal loads
are often times subjected to impingement cooling as it provides very high heat transfer coefficients. One
drawback, however, is that it reportedly could weaken the strength of the structure in the surrounding
region [40]. An impingement array is given in Figure 2.13, along with a side-view cut-through in Figure
2.14.

Figure 2.14: Schematic of a cross-flow. Gc and

Figure 2.13: A typical array of a jet- Gj are the cross-flow and impingement jet ve-
impingement geometry. locities respectively

An important phenomenon in jet-impingement cooling is cross-flow, which is the interference of the

regular cooling air with the impingement cooling air. Hence, the ratio of cross flow velocity to jet velocity
is an important parameter in the design of such cooling, along with the spacing in the jet array and hole
diameter.

Figure 2.15: Schematic of a film cooling process, illustrating the mixing process of a coolant with
mainstream flow. [91]

Film cooling generally refers to a secondary fluid which is injected in the area of interest. The resulting
flow structure is of high complexity and is influenced by multiple parameters. The flow mixing is
visualized in 2.15. Example parameters include the hole geometry configuration (exit shape, injection
angle, hole length and spacing, number of holes), mainstream conditions, such as mass flux ratio M ,
coolant density ratio DR, approach boundary layer and others. Some authors also investigated slots
instead of discrete holes for an injection geometry.

2.4 Low cycle fatigue

The aim of the design of a GT (combustion) component in the low-cyclic fatigue range is to determine
the service life under the local influence of cyclic plastic strain. This is done so that the components in
GT meet the mechanical integrity requirements, which are influenced by the thermal strain effects that
temperature gradients cause from the component’s cyclic thermal expansion[1]. The cyclic nature of
1 LCF is also a result of the mechnical cyclic stress. Such stresses arise from the mechanical loading of the component,

e.g. displacements, pressure loads etc, which most of the time account for the majority of the total stress.

Energy & Process Technology Page 24

 Delft University of Technology

the loading is based on the chosen operating cycles of the components (usually such cycle is the inter-
mittent shutting and starting of the entire GT). Low cycle fatigue arises due to thermal expansions of
a component The two key features of LCF analysis are the use of the total stress or strain amplitude
and the use of suitable material data in the form of stress or Wöhler strain curves (sample curve pre-
sented in Figure 2.18). The material data used in this work comes from the Siemens internal material
database. One duty cycle for the component considered in this work consists, in simplified terms, of
the state of rest, starting the Gas turbine, base load operation and shutdown of the gas turbine back to
its idle state. Under the assumption of the highest load in base load operation is used in this work to
determine a quasi-static analysis of the maximum stresses was carried out. For the hibernation of the
system it is generally assumed to be in a stress-free state at ambient temperature. As a general rule, the
time-dependent change in load must be determined by transient analyses for the starting and shutdown
processes.

The LCF analysis simulates the thermal fatigue of the combustion components; i.e. the thermal ex-
pansions and hence stresses caused by temperatures that change over time, which lead to the formation
and growth of cracks in the component. ([101], p.202) The real material behaviour shows a non-linear
relationship between stress and strain in this area. The multi-axial stress state can be compared with the
von Mises equivalent stress failure criterion of the material. The von Mises equivalent stress is calculated
according to the following equation, where σ1 to σ3 represent the principal stress components:
q
1 2 2 2
σV,M ises = √ · (σ1 − σ2 ) + (σ2 − σ3 ) + (σ3 − σ1 ) (2.31)
2
The Von Mises yield criterion is used to determine when yielding occurs. The criterion function and the
plane stress of the Von Mises yield surface is given in Figure 2.16 below.

Load point

Tangent to
yield surface
Yield surface
1
3 ′ 2
𝑓= 𝜎 : 𝜎′ − 𝜎𝑦 = 0
2
Elastic region

Figure 2.16: Von Mises yield surface, along with the yield criterion function f , along with the increment
in plastic strain normal to the surface, and the resultant stress–strain curve obtained for uniaxial straining
(no hardening). Note at the yield surface, f = 0.

In order to avoid the high computational effort of a nonlinear finite element analysis, a linear-elastic
calculation approach is used for simulations in this thesis. The result of the linear calculation method is
then approximated to the real elastic-plastic component behavior with a shakedown according to Neuber
correction method. The Neuber approximation is given by:
2
(Kt σn ) σ2  σ  1′
n
εplastic = = +σ , (2.32)
E E K′
where Kt is the stress concentration factor, σn is the nominal stress, σ is the local stress, and K ′ and
n′ are the Ramberg-Osgood parameters for cyclic loading. In case the stresses are calculated by a linear
elastic analysis, a proper determination of the cyclic loading requires a consideration of the reloading in
addition to the Neuber relationship for initial loading.

Energy & Process Technology Page 25

 Delft University of Technology

𝑙𝑖𝑛𝑒𝑎𝑟 − 𝑒𝑙𝑎𝑠𝑡𝑖𝑐

𝑆𝑡𝑟𝑒𝑠𝑠 𝜎
𝜎𝑛 𝐾𝑡

𝑁𝑒𝑢𝑏𝑒𝑟 𝑆ℎ𝑎𝑘𝑒𝑑𝑜𝑤𝑛 𝐸 Young’s Modulus

𝜎𝑛 Nominal stress
𝜎 Local stress
𝜎 𝐾𝑡 Stress concentration
𝐸 factor
𝑒𝑙𝑎𝑠𝑡𝑖𝑐 − 𝑝𝑙𝑎𝑠𝑡𝑖𝑐
𝜖𝑛 Nominal strain
𝜖 Local strain

𝜖𝑛 𝐾𝑡 𝜖
𝑆𝑡𝑟𝑎𝑖𝑛 𝜖

Figure 2.17: Neuber shakedown method

The shakedown for initial loading starts in the linear-elastic stress state and ends at the starting point
of the cyclic hysteresis curve, as shown in the figure above. Then, the cyclical process of loading and
unloading begins. The stress or strain range corresponds to twice the amplitude value of the stress or
strain of the initial load curve. The equation below describes the approximation by Neuber for cyclical
reloading;
2  1
(Kt ∆σn ) (∆σ)2 ∆σ n′
= + 2∆σ . (2.33)
E E 2K ′

2.4.1 Lifespan calculation

For the design of cyclically loaded components, the prediction of the service life is of great importance.
The stresses, strains and temperatures calculated in a finite element analysis are available for calculat-
ing the service life. Furthermore, empirically determined material coefficients from vibration tests are
known. Based on this data and the equations described in this section, the expected service life can be
thus calculated.

The basic equation describing the relationship between total strain amplitude and service life was de-
scribed by L. Coffin and S. Manson [115] and is:

σf′
(2Ni ) + ϵ′f (2Ni ) ,
b c
εa = (2.34)
E
| {z } | {z }
elastic plastic

where σf′ is the so-called vibration resistance coefficient, ϵ′f represents the cyclic ductility coefficient, b
σ′
is the fatigue strength component, and c is the cyclic ductility exponent. The factors ϵ′f and Ef are, per
convention, found from the inersection of the elastic or plastic strain Wöhler line with the vertical at
Ni = 0.5 [100]. Ni is thus also known as the number of tolerable oscillation cycles up to crack formation.
Furthermore, the exponents b and c, ϵ, and σ are temperature dependent material properties that are
empirically determined by vibration testing.

Energy & Process Technology Page 26

 Delft University of Technology

𝑆𝑡𝑟𝑎𝑖𝑛 𝑎𝑚𝑝𝑙𝑖𝑡𝑢𝑑𝑒
𝑝𝑙𝑎𝑠𝑡𝑖𝑐

𝑡𝑜𝑡𝑎𝑙

𝑒𝑙𝑎𝑠𝑡𝑖𝑐

𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐𝑦𝑐𝑙𝑒𝑠

Figure 2.18: Wöhler SN curve (log-scale)

The first term of the equation relates to the elastic strain amplitude, while the second part relates to
the plastic strain amplitude and represents a measure of the half-width of the stress-strain hysteresis.
Equation 2.34 has to be solved numerically. The Wöhler line is essentially derived from the superposition
of the elastic and plastic strain components shown in Figure 2.18 [100].

Figure 2.19: Effect of mean stress presence on fatigue life [115].

Figure 2.20: Constant mean strain ϵm effect on the mean stress levels σm . The relaxation is due to the
plastic deformation, and hence the relaxation rate depends on strain amplitude [115].

The presence of a mean stress σm can have a significant influence on the service life ([115], p.74).
Compressive mean stresses increase strength of the component, while on the other hand tensile mean
stresses reduce it [100] (refer to Figure 2.19). The mean stress σm is read off at the midpoint of the
stress-strain hysteresis. Several techniques have been developed in the attempt to incorporate the mean
stress influence into the Coffin-Manson equation (equation 2.34). In the following formulation, the mean

Energy & Process Technology Page 27

 Delft University of Technology

stress according to Morrow is taken into account, see equation 2.35. In this method, a mean stress
correction is applied to the elastic stress component of the Coffin-Manson equation. ([115], p.113)
!
σf′ σm
1 − ′ (2Ni ) + ε′f (2Ni )
b c
εa = (2.35)
E σf

However, since there is more mean stress relaxation at higher strain amplitudes due to larger plastic
strains, low-cycle fatigue region is less influenced by the mean stress effect [115].

Energy & Process Technology Page 28

Chapter 3

Optimization Terminology

The problem of mathematical optimization is a vast one, expanding over a multitude of scientific dis-
ciplines, incl. computational theory, statistic, mathematics etc., and has been studied for a considerably
long time: a monograph introducing the concept linear optimization technique for the first time was
written in 1939 by Kantonovich [59]. In this chapter, the author’s attempt to provide the reader with a
very brief overview of current state-of-the-art optimization algorithms, and where possible some relevant
mathematical background is herein presented, along with a relevant application envelope of each algo-
rithm. The list of optimization algorithms below is by no means complete and only serves to provide an
idea of the vast possibilities and the wide range of methods. The chapter also contains an overview on
the proprietary optimization algorithm by Siemens, which is at the groundwork of the optimization anal-
ysis in this project. The benefits of using such a hybrid optimization algorithm are also touched upon.
The chapter includes basic terminology regarding multidisciplinary optimization and, finally, statistical
design of experiments, which is presented as a method for efficient planning of simulation series.

3.1 Optimization algorithms

The hybrid nature of the proprietary optimization algorithm used in this project necessitates an ample
understanding of the basics of the most fundamental types of optimization algorithms. In its canonical
form, an optimization problem can be mathematically formulated as:

Minimize: f (−→x) (3.1)

→
−
Subject to: g ( x ) ≤ 0 j = 1..m (3.2)
j
hk ( →
−
x) = 0 k = 1..l (3.3)
x ( x ) ≤ x ≤ xu ( −
l →
i
−
i
→
x)
i i = 1..n (3.4)

 

 x1 

 x2 
Where →
−
x = (3.5)
..

 . 


 

xn
The following notation is used:
• The objective function (OF) (less formally put, the goal of the optimizer) is mathematically denoted
as f (−
→
x ). A common practice is to express the objective as a minimization problem, but the inverse
holds too (maximization). From an industrial perspective, the function can consider a wide variety
of objectives as long as a mathematical formulation exists.
• All design variables that can be modified are contained in the vector −
→
x . The last equation (3.4)
comprises of the lower and upper bounds of each of the respective design variables.

29
 Delft University of Technology

• Optimization can be unconstrained or constrained depending on the problem, expressed in the

inequality, i.e. the function gi (−
→
x ).
• The equality of the variable hk (− →
x ) can be used to express relations between design variables,
→
−
which tend to be explicit in x in most cases. Such relations can be eliminated when the choice
and number n of variable is properly selected.
A constrained optimization presents a challenge for direct solvers due to the inequality constraint and
can be often translated into an unconstrained pseudo-objective optimization:
X
m

2
f˜(−
→
x ) = f (−
→
x)+R· δj · g j ( −
→
x) (3.6)
j=1

Where δj = 0 if gj (−
→
x) ≤ 0
δj = 1 if gj (→
−
x) > 0
Such ’penalty-based’ methods use a multiplier R that puts weight to satisfying the constraints (genetic
algorithms discussed later for example use this technique). This however remains a weak formulation of
the problem and sometimes the constraints might remain unsatisfied.

The majority of optimization algorithms base the new design on the previously found one, such that:
→
− →
−
x i+1 = −
→
x i + αi · S i , (3.7)
→
−
where the iteration is indicated with the subscript i, and hence x i+1 is the new guess for optima on
the design space. The search direction can be expressed with Si and αi is a scalar used for representing
the amplitude of the change in the direction of the vector Si . Essentially, the difference in all possible
optimization algorithms lies in the manner in which previous iterations compute the amplitude and di-
rection vector in the search of new optimum, given in Equation 3.7.

Generally speaking, optimization algorithms can be classified depending on the order of derivatives
used of the OF. For example, zero order methods, sometimes called derivative free optimization (DFO),
include only algorithms which evaluate and use directly the value of the OF in search of an optimum.
First and second order methods respectively, which lie outside the scope of this report employ first and
second order derivatives of fi (−
→
x ). New optimization algorithms however are constantly being developed
and can be sometimes frouped into families. Figure 3.1 presents an overview of grouping of some of the
most popular optimization algorithms, also indicating the order of the respective algorithm.

Figure 3.1: Families and classes of optimization algorithms (adapted from [80])

Energy & Process Technology Page 30

 Delft University of Technology

For the sake of conciseness of the report, in the following subsections, only the types of optimization
algorithms present in the hybrid algorithm are presented in more detail: (i) simplex programming, (ii)
deterministic/stochastic programming, and (iii) genetic optimization. The term programming can be
used interchangeably with optimization in the context of algorithms and is not related to computer
programming but rather a set of logistic instructions.

3.1.1 Simplex programming

Linear programming refers to the solution approach used to solve optimization problems where all
the constraints are linear equalities or inequalities, along with a linear objective function. The most
common algorithm for such problems is the Simplex Method, which was first introduced in 1947 by
George Dantzig. Linear programs are most commonly expressed as:

Minimize: z = f (−
→
x) = −
→
c T−
→
x (3.8)
→
− →
−
Subject to: Ax ≤ b (3.9)
−
→
x ≥0 (3.10)

where A is an n × m coefficient matrix, comprised of the constraint coefficient vectors for each vari-
→
−
able (−
→ai ) and b is a vector of coefficients in Rn (n equations in m unknowns, m > n). The con-
straint set in a linear program defines a feasible region, which is in the form a polyhedron, coming
from the
 intersection of n-number half-spaces (−→ai ⊤ −
→
x ≤ bi ). The constraints set be expressed as
n −→ ⊤−
→
P := x ∈ R | a i x ≤ bi , i = 1 . . . , right} . An example polyhedron in two-dimensions is given in
Figure 3.2 below.

Figure 3.2: A polyhedron P and an extremum, highlighted in red.

An important mathematical concept in linear programming are the extreme points of the polyhedron.
They are defined such that
−
→
x k = λ−
→
x 1 + (1 − λ)−
→
x2 ⇔−
→
x1 =−
→
x2 =−
→
xk (3.11)

where the vectors − →

x 1 and −→
x 2 ∈ P, and the equation holds for any λ ∈ (0,1). The fundamental property
→
−
that follows for a polyhedron is then that it is a convex set. If it is assumed that P has p distinct
extrema −→x 1 , ..., −
→
x p . The convexity then is express as:
X
p X
p
−
→
x = λl x l , where λl = 1 and λl ≥ 0 l = 1, . . . , p (3.12)
l=1 l=1

Dantzig used this property of convexity, from which the fundamental theorem of linear programming
is derived: ”Suppose that the feasible region P to a linear program is a non-empty bounded set, i.e.,
an optimal solution exists. Then the optimal solution −
→x ∗ is an extreme point to the set defined by the

Energy & Process Technology Page 31

 Delft University of Technology

inequalities”.

The most important concept in Simplex Programming is another form of representation which states
that the Linear program has to be represented in the so called slack form, after ensuring that the poly-
hedron is in its standard form (the mathematical definition of the standard form lies outside the scope
of this subsection). The algorithmic flow then consists of converting one slack form into an equivalent
slack form whose objective value has not decreased, and has likely increased.

The slack form introduces an additional number of variable corresponding to the number of equations
that are present. Such variable are called basic variables, whereas the original variable in are non-basic.
The additional variables, hence, represent slack, in the sense that they correspond to how much slack is
given in the inequalities present in the original problem. Once a solution is found (called the basic solu-
tion) for the set of equations, the iterative search for an optimum begins. It is achieved by searching from
vertex to vertex. Each subsequent vertex is chosen such that the search direction is in the steepest fea-
sible direction (based on so called reduced cost calculation), an exercise commonly referred to as pivoting.

Simplex Programming is covered in an extensive amount of literature, and for all unknown terms in
the paragraphs above, the reader can refer to the following references where it is described in more
detail. This includes [12] and [30], and therefore no further quantitative explanation will be given about
it.

3.1.2 Deterministic and stochastic programming

The deterministic optimization contains the classical optimization algorithms of mathematics and is
largely based on linear algebra. One advantage of deterministic optimization is the faster convergence
compared to stochastic methods. This means that fewer simulations of the response variable are nec-
essary to achieve an optimized solution. Since the optimization is consistently based on mathematical
formulations and does not contain any random elements, the results are always clear and reproducible.
However, deterministic optimization has the disadvantage that by nature it can only optimize for one
objective. It is also required to have gradient information of the objective function, which includes a
considerable computational effort. The objective function also needs to be continuous and deterministic
methods display weak performance in a noisy environment. In addition, the deterministic algorithms
look for stationary points of the response variable, so that the solution may converge on a local optimum
and overlook a global optimum. ([18], p.77) This fact can be illustrated with a simple optimization
that starts at a certain point on a function curve and moves along the gradient to ever smaller function
values. If this optimizer reaches a local minimum, it stops because from this point on the function values
increase again. An underlying global optimum could hence remain undiscovered.

Stochastic optimization algorithms are methods that are based on chance or contain random elements.
Many stochastic algorithms are modeled on natural processes. Simplified mathematical models were
generated for various concepts from biology, physics or society, such as evolution, genetics or swarm
intelligence. The greatest advantage of stochastic optimization algorithms lies in the possibility of multi-
target optimization. So it is possible to optimize for more than one response variable. Furthermore,
global optima can also be easily found, since the entire design space is examined with “randomly” placed
samples ([18], p.103). The already mentioned slower convergence and the ambiguity of the solution are
disadvantageous.

3.1.3 Genetic optimization

One sub-area of stochastic optimization is optimization using genetic algorithms that imitate processes
from biological evolution. Genetic algorithms (GA) have the advantage that they can be applied to almost
any type of optimization problem. GA-base algorithms have been reportedly the most commonly used for
optimization problems involving Navier-Stokes equations in FV domains. Even complex relationships
do not pose a problem. The three main components of a genetic algorithm are crossing, mutation,

Energy & Process Technology Page 32

 Delft University of Technology

and selection. Each solution is encoded in what is called a chromosome. The crossing of two parents
(chromosomes) produces a new generation of children by interchanging parts of the chromosomes. In
the next step, a mutation takes place with a much lower probability.

𝐺𝑒𝑛𝑒𝑟𝑎𝑡𝑖𝑜𝑛 𝑖 𝑃𝑎𝑟𝑒𝑛𝑡𝑠 𝐶ℎ𝑖𝑙𝑑𝑟𝑒𝑛 𝐺𝑒𝑛𝑒𝑟𝑎𝑡𝑖𝑜𝑛 𝑖 + 1

1) Selection of fit data points 2) Crossbreeding

3) Mutation 4) Creation of new Generation i + 1

Figure 3.3: Genetic evolutionary process [5].

Individual parts of the chromosome are randomly changed. This ensures that random factor settings
are generated again and again and that the algorithm does not converge to a local optimum. The newly
generated solutions are sorted out based on their ”fitness”. Only the best solutions are passed on to the
next generation of parents. The iterative process is repeated until the maximum number of iterations or
the convergence criterion is reached [18].

𝑦2
Resultant designs
Pareto-optimal designs

Pareto Front

𝑦1
Figure 3.4: Simplified Pareto front illustration of a multi-objective trade-off study. Design points which
lie on top of the Pareto line are not dominated, and also do not dominate each other. On the other hand,
points that lie outside of the border, are dominated by the Pareto-optimal designs.

With the help of genetic algorithms, an optimization can be carried out for several goals. With multi-
objective (MO) optimization, there is no clear optimum, but a set of so-called Pareto optima. A Pareto
optimum is defined as the fact that no target variable can be improved without another target variable
deteriorating. The so-called Pareto Front is the sum of all Pareto Optima solutions, i.e. if the problem
is translated from a multi- to a single-objective optimization, the weighted sum of the objectives needs
to be minimized. For the sake of this projects optimization objective function this reads:

F (−
→
x ) = wtemperature · Ftemperature (→
−
x ) + wmass · Fmass (→
−
x ), (3.13)

which is analogous to Equation 3.15, representing the general form for a MO optimization. The relation-
ship is shown schematically in Figure 3.4 above. When the number of objectives in the study however
exceeds 2, the Pareto is not simply a single line anymore but a hyperplane. This leads to an increase in
the number geometries necessary for analysis, and the visualization becomes less trivial.

3.1.4 SHERPA ®
As mention earlier in the introduction chapter of this thesis, in this work a proprietary multi-objective
optimization algorithm SHERPA is used, which comes within Siemens’ Design Exploration Software

Energy & Process Technology Page 33

 Delft University of Technology

MDO HEEDS ®[116]. A very limited number of sources available in open literature makes use of this
algorithm. SHERPA’s search method employs a hybrid and adaptive algorithm, that makes use of mul-
tiple search strategies at once and adapts to the problem. The name stands for “Simultaneous Hybrid
Exploration that is Robust, Progressive, and Adaptive”. Siemens reports that a method’s participation
is diminished in the optimization if it is determined to be ineffective for the given search problem. It
is also worth mentioning that the developers of the algorithm do not claim that SHERPA is the best
approach for all problems. The method, however, has shown convincing efficiency in finding a global
optimum for a variety of practical engineering design problems.

A simplle benchmark problem (minimizing mass of a cantilever beam while also satisfying a constraint on
stress and deflection) is also used to compare the algorithm’s performance to numerous other ’baseline’
algorithms. The figures below shows that SHERPA manages to find an optimal solution in considerably
fewer iterations.

(a) Average best solution found over 50 runs versus (b) Standard deviation of the best solutions found
the number of allowable evaluations over 50 runs

Figure 3.5: Results from benchmark comparison of different optimization algorithms. SHERPA yielded
better results even at a very low number of evaluations [116].

Goldin et al. [35] also claimed that SHERPA is a hybrid combination of Nelder-Mead Simplex [88],
simulated annealing [50], quadratic programming [105] and a genetic algorithm [34]. Some of the reported
advantages include [97]:
• No iterative tuning of optimizer parameters is necessary

• Does not employ approximate response surface models and performs direct optimization
• Multiple strategies ensure that large and complex design space does hinder the efficiency of the
global and local optima search

3.2 Multi-disciplinary Optimization

The problem of multi-disciplinary optimization has gain considerable traction over the years, particu-
larly with the commercialization of high-performance computers. As the name suggests, MDO problems
are comprised of multiple disciplinary analyses. According to Zadeh [130], the core idea of MDO “pertains
to the decomposition of complex engineering systems into a set of smaller and less complex subsystems
together with formal approaches of accounting for system interactions and couplings.” While the term
multidisciplinary in literature is more relevant for system-level optimization types, the term is sometimes
interchangeable with multi-objective optimization. The MDO framework is characterized in [130]. In
such systems, it is desired to minimize/maximize several response variables at the same time, such as

Energy & Process Technology Page 34

 Delft University of Technology

the case in this thesis. Thereby, the correct mathematical formulation of such a system reads:

Minimize: fi ( −
→
x) i = 1..l
Subject to: →
−
gj ( x ) ≤ 0 j = 1..m
hk ( −
→ (3.14)
x)=0 k = 1..n
→
−
xlp ( x ) ≤ xp ≤ xup (−
→
x) p = 1..q

The vector −→x is the vector of design variables, consisting of the parameters of interest that can be modi-
fied. The function fi (−
→x ) is the respective objective function, specified by the subscript i. The constraints
of the optimization are given by the inequality equations, whereas the equality constraint contributes
the the relations between design variables, in an identical manner to a single-object optimization. The
only difference lies in the subscript of fi (−
→x ).

The multi-objective formulation can be easily converted to a single-objective by summing each indi-
vidual weighted objective:
X l
F (−
→
x) = w i · fi ( −
→
x) (3.15)
i=1

The function F (→
−x ) then corresponds to a single point on the Pareto front, depending on the choice of
weight values. This approach however diminishes the amount of insight obtained from the optimization
problem, despite its computational advantage. The set of equations in Eq. 3.14 are either converted into
a pseudo-objective functions (where the constraints are incorporated into a single equation, and hence an
unconstrained optimization is solved) or the strong formulation is used with an optimization algorithm.

Figure 3.6: Pareto front of a MO (2 objectives) optimization and the equivalent optimization using a
weighted sum in a single-objective optimization. The dashed lines represent values of the converted MO
calculation. The resultant penalty decreases as the dashed line approaches the Pareto front.

It is worthwhile mentioning that the choice among so-called ”Pareto optimal” solutions (solutions lying
on the Pareto front) to determine the ”favourite solution” is entirely delegated to the decision maker,
and is usually performed after the MO optimization is finished. The conversion from MO to a single-
objective less computations are needed to obtain the minimum, and the designer needs to observe only
the region in the vicinity of the optimum (presented as the region near the black dot in Figure 3.6. The
difficulty of using this approach comes in the determination of the weights. This becomes particularly
ambiguous if the objectives are in different fields (such as the case of the optimization problem in this
problem).

Energy & Process Technology Page 35

 Delft University of Technology

3.3 Screening and Design of Experiments

The field of Design of Experiments and screening could be used as an alternative powerful tool for op-
timization problems. The following contains a brief overview of the basics, necessary for the application
of so-called meta models used in optimization.

For a parametric model with n free parameters x, there is a large number of possible parameter combi-
nations. For each of these parameter combinations there is an answer y of the overall system.

(a) (b) (c)

Figure 3.7: Linear effect a), quadratic effect b), and interaction c).

The relationship between a parameter x and the answer y can in many cases be described with a math-
ematical function. This dependence of the system response on a parameter is called an effect. Linear
relationships between factors and response variables are called main effects, see Figure 3.7. A quadratic
effect can be seen in 3.7 b). If n factors influence one another, one speaks of N th order interaction.
Figure 3.7 c) shows the interaction between parameters x1 and x2 . For a high value of x1 , a different
function curve arises than for a low value of x2 : the parameters interact.

The determination of the main effects of many influencing factors is often more efficient, than the exact
specification of a higher-order term. Research experience has shown that for the majority of physical
problems, only a few interactions and nonlinear effects actually play a role.

In order to determine the answer y for a certain parameter combination, an experiment or a simulation
must be carried out. In practice, it is often not useful or either possible to look for every possible param-
eter combination due to the great computational effort involved. Statistical test planning is therefore
used for efficient planning and evaluation of test series, more commonly known as Design of Experiments
(or DoE).

A statistical test plan tries to determine the relationship between input parameters x and output pa-
rameters y as precisely as possible with the smallest possible number of experiments or simulations. The
input parameters of a DoE are also referred to as factors. Factors have at least two levels. The number
of levels describes how many different values each parameter x can assume. For a three-level design, each
factor x takes three values and one gets three response values for y. With such a test plan, quadratic
effects can be mapped, whereas mapping of high-order effects requires more levels, which becomes more
computationally expensive [5].

Energy & Process Technology Page 36

 Delft University of Technology

Figure 3.8: A schematic of a full-factorial design with three factors and two levels each.

Figure 3.8 displays a full-factorial design of three parameters. Each parameter in this design has two
levels, i.e. it takes on two different values, the minimum value (-) and the maximum value (+) of the
selected interval. This plan is sufficient to determine the main effects of the three parameters. The green
points represent the support points of the test plan, i.e. the points at which the output parameter y is
determined for a parameter combination x1 , x2 , x3 .

In a statistical test plan, several factors are strategically varied at the same time. The structure of
the test plans enables the effects to be clearly assigned to the input parameters despite simultaneous
variation. Standardized experimental plans are Orthogonal and Balanced. Orthogonal means that no
combination of two columns correlates with one another. Balanced means that the factor levels of any
factor and the settings of the other factors are evenly divided [5].

Figure 3.9: Illustrative representation of the effects of factor x1 .

Figure 3.9 shows that even a small effect of x1 can easily be determined by balancing the remaining
factors x2 and x3 . [5] Despite the simultaneous variation of all factors, the green block on the right is
the same on both sides of the scale. Hence, the difference between the red and yellow blocks corresponds
to the effect of x1 .

The concept of design screening is well put by Allen [2], according to whom screening is a process
that begins with an extensive list of potentially influential factors and ends with a smaller list of factors.
Once this process is finished, the remaining factors are likely to have a large effect on the mean response.
Screening is useful at the beginning of an examination in order to reduce the number of factors considered
and to make follow-up examinations more efficient. In most cases it is based on a partial factorial design.
A partial factorial test plan is a reduced, full factorial plan in which individual support points are not

Energy & Process Technology Page 37

 Delft University of Technology

Table 3.1: Brief comparison of full-factorial and partial-factorial designs

Full-factorial design Partial-factorial design

100% effort 50% effort

All main effects Main effects superimposed
All interactions with interactions

calculated in order to reduce the simulation effort.

In a full factorial test plan, measured values are recorded for all possible combinations of the factor
levels. With ns levels and k factors, n attempts are required:

n = nks (3.16)

Since the number of tests required increases sharply with the number of factors examined, factorial test
plans are often used in screening. The effort of the partial factorial test plan shown in Table 3.1 is half
of the full factorial test effort. In the fully factorial test plan shown, all main and interaction effects of
the three factors are mapped with eight tests. In this cas,e these are the main effects of the factors x1 ,
x2 and x3 , the interactions of the second degree between the factors x1 · x2 , x1 · x3 and x2 · x3 as well
as the triple interaction x1 · x2 · x3 . For the partial factorial design, only 50% of the tests are required
compared to the full factorial design, i.e. four tests. However, the main effects are overlaid with specific
interaction effects. With the reduced effort involved in the experiment, there is also a loss of information.
In a test plan with three factors and four tests, the following effects can no longer be separated:
• Main effect of x1 , and interaction of x2 · x3 ;
• Main effect of x2 , and interaction of x1 · x3 ;
• Main effect of x3 , and interaction of x1 · x2 .

In other words, it is not possible to differentiate between main effects and interaction effects of the second
degree. Assuming low interaction effects, the main effects of individual parameters can nevertheless be
estimated with a small number of tests [5]

The field of DoE expands further into so-called meta-models. The concept will only be briefly explained,
but for a detailed read, one can refer to [107] and [2]. Meta-models are used for an analytical calculation
of the response variable over the design space based on the results of the DoE. It thereby, allows for quick
and easy optimization without the necessity of running further simulations. They present suitability in
cases where the relationship betweem factors and response variable needs to be determined precisely and
when a subsequent optimization ought to be performed. Some meta-models include Central Composite,
Box Behnken, Latin hypercube, Optimal Space Filling and others.

Energy & Process Technology Page 38

Chapter 4

Numerical Methodology

The purpose of this section is to introduce the reader to governing equations used within the STAR-
CCM+ solver, starting from the mathematical formulation of the conjugate heat transfer methodology,
explaining into further details the available convective heat transfer coefficients, which will be later
used for the purpose of conducting a comprehensive comparison between the decoupled and conjugate
approach, and finishing off by describing common flow modelling models and wall treatment methods in
turbulent flows.

4.1 Numerics of conjugate heat transfer

Before introducing the numerical procedure followed in the commercial software package, first a co-
herent explanation of the mathematical formulation of conjugate heat transfer is presented.

The three fundamental conservation laws governing the physics of fluids are conservation of mass (con-
tinuity equation), conservation of momentum (aka the Navier-Stokes equations), and conservation of
energy. If Ωf is a fluid domain with a boundary ∂Ωf , then the three laws can be equated in the
differential form as:
∂ρf
+ ∇ · (ρf ⃗v ) = 0 (4.1a)
∂t
∂ρf ⃗v
+ ∇ · (ρf ⃗v ⊗ ⃗v ) = −∇p + ρf f⃗ + ∇ · τ (4.1b)
∂t
∂ρf E
+ ∇ · (ρf H⃗v ) = ρf f⃗ · ⃗v + ∇ · (τ · ⃗v ) − ∇ · (λf ∇Tf ) (4.1c)
∂t
in which, on the left hand side of the equations stands the density of the fluid, denoted ρf , the time
t, the velocity vector ⃗v , comprised of the scalars vx , vy , and vz , respectively, the pressure term p, the
total enthalpy and total energy, H and E, respectively. On the right hand side are the the body force
vector representing the combined body forces of rotation and gravity f⃗, the viscous stress tensor τ can
mathematically expressed as:  

2
τ = µ ∇⃗v + ∇⃗v T − ∇ · ⃗v I , (4.2)
3
also, the thermal conductivity of the fluid λf , and the temperature Tf . The energy term is defined as:

E = H − p/ρ, (4.3)

where Z T
2
H= Cp dT + |⃗v | . (4.4)
Tref

39
 Delft University of Technology

Based on the schematic shown in Figure 4.1, the contacts (boundaries) where the respective boundary
conditions are applied are defined such that Γ1 ∩ Γ2 = ∅, and the fluid boundary not in contact with the
solid boundary (i.e. ∂Ωf \ ∂Ωs ) is equal to Γ1 ∪ Γ2 , which leads to boundary conditions or Dirichlet and
Neumann type, applied to ∂Ωf \ ∂Ωs :

 Tf = f1 on Γ1
(4.5)
λf · ∂Tf = f2 on Γ2
∂n
with f1 being the specified temperature on the boundary, and f2 is the heat flux on Γ2

Figure 4.1: Domains and boundaries in a typical conjugate heat transfer problem

The solid region Ωs , with the domain boundary Ωs is described by the conservation energy equation
only, which in fact simplifies to the popular heat transfer equation. If the velocity is removed along with
the transient term, then:
∇ (λs ∇Ts ) = 0, (4.6)
with the thermal conductivity of the solid region λs . In a similar fashion, three possible boundary
conditions can be applied to the boundary non in contact with the fluid domain, namely Dirichlet,
Neumann or Robin, given respectively as:


 Ts = g 1 on ∆1


 ∂Ts
λs · = g2 on ∆2 (4.7)
 ∂n



 λs · ∂T s
= h · (Ts − T∞ ) on ∆3
∂n
where similarly ∂Ωs \ ∂Ωs = ∆1 ∪ ∆2 ∪ ∆3 . The heat transfer coefficient of the fluid is given as h, the
ambient temperature T∞ , specified temperature g1 and heat flux g2 on the boundaries ∆1 and ∆2 .

Finally, the boundary condition on the boundary where ∂Ωs intersects with Ωf , namely ∂Ωw = ∂Ωf ∩∂Ωs .
Thus, two boundary conditions are set: temperature and flux continuity:

 Tf = Ts
on ∂Ωw : (4.8)
λf · ∂Tf = −λs · ∂Ts
∂n ∂n
which finalizes the definition of the mathematical form of conjugate heat transfer, together with the
boundary conditions for the momentum conservation on the fluid region.

Energy & Process Technology Page 40

 Delft University of Technology

Now that it has been established that the heat flux is conserved at the interface, one can define two
additional boundaries adjacent to Ωs and write

q̇0 + q̇1 = −Su (4.9)

where q̇0 is the heat flux from the fluid through Boundary 0, and q̇1 is the heat flux leaving through
Boundary 1 into the solid (both shown in the schematic in Figure 4.1 as dashed lines, the thickness of
the boundary is exaggerated for the purposes of illustration). The term used for a specified heat source
is denoted with Su . Each of the aforementioned heat fluxes is then linearised:
4
q̇0 = A0 + B0 Tc0 + C0 Tw0 + D0 Tw0 (4.10a)
4
q̇1 = A1 + B1 Tc1 + C1 Tw1 + D1 Tw1 (4.10b)

with A, B, C, and D being the linearised heat flux coefficients. The subscript c or w indicates either cell
or wall value at the respective boundary number. For a single phase flow problem the coefficients are
composed of internally pre-calculated and manually specified coefficients (kept at 0 for all computations
in this study)

A = Ainternal + Aspecif ied (4.11a)

B = Binternal + Bspecif ied (4.11b)
C = Cinternal + Cspecif ied (4.11c)

The internal coefficients are evaluated in accordance to the physical processes undergoing in the system.
They are found using the equation for the diffusion flux at an interior cell face:

k∇T = k [∇Tc − (∇Tc · ds) α

⃗ ] − k⃗ αTw + 0Tw4 .
αTc + k⃗ (4.12)

The discretization of the diffusion flux and the derivation of Eq.4.12 is emitted here, but a full derivation
and more detailed explanation of the finite volume method could be found in the Appendix. where
a
α
⃗= , (4.13a)
a · ds
ds = xf − x0 . (4.13b)
The surface area vector of the respective cell is denoted with a and the vectors xf and x0 are used to
indicate the face vector and cell centroid vector from the coordinate frame origin, respectively.

Ainternal = k [∇Tc − (∇Tc · ds) α

⃗] (4.13c)
Binternal = −k⃗
α (4.13d)
Cinternal = k⃗
α (4.13e)

In cases where the temperature is different across the interface (e.g. a thermal barrier coating is included
without physically modelling the coating), an interface thermal resistance is set:

Su Tw1 − Tw0
q̇0 = − + , (4.14)
2 R
where the denominator in the second term on the right hand side of the equation is the resistance
(K/W ). The last term in Eq.4.12 representing radiation is set intentionally to zero for the purposes of
this explanation. Finally, the set of equations (4.9)-(4.10b) and (4.14) are solved for the four unknowns
either analytically or numerically. If the radiation term is included, STAR-CCM+ performs a numerical
calculation. If not, a linear system of equations is obtained and solved.

Energy & Process Technology Page 41

 Delft University of Technology

4.2 Reynolds-Averaged Navier-Stokes Modelling

In the present work, Reynolds-Averaged Navier Stokes (RANS) models are compared. The concept
of Reynolds-Avergaing comes from replacing the standard solution variables, ϕ, by a mean (averaged)
component and a fluctuation part
ϕ = ϕ + ϕ′ (4.15)
Upon substitution in Eq. 4.1b and in Eq. 4.1c, new terms are obtained, namely the Reynolds Stress
term:  ′ ′ 
u u u′ v ′ u′ w ′
τ = −  u′ v ′ v ′ v ′ v ′ w ′  (4.16)
u′ w ′ v ′ w ′ w ′ w ′
These new unknown quantities pose the challenge of the closure model, which is a fundamental part
of the theoretical description of turbulence, i.e. requiring another relationship to solve the problem.
RANS models are one approach to close the two terms and are based on the Boussinesq hypothesis. In
the Boussinesq hypothesis, the intrinsic assumption is made that the Reynolds stress anisotropy tensor,
A, is determined by the mean velocity gradients and the specific assumption is made that there is an
alignment between the Reynolds stress anisotropy tensor and the mean rate-of-strain tensor S, hence,
the following equation should hold:
2
A = τ + kI ≈ 2νt S (4.17)
3
1

S= ∇v + ∇vT (4.18)
2
where most importantly, νT is the eddy viscosity, and k = tr(τ ) is the turbulent kinetic energy. RANS
models rely on the Boussinesq hypothesis due to the fact that it gives acceptable results in simple shear
flows (Pope [2000]) as these flows match the specific assumption, but also, due to the added numerical
stability to the model from the Boussinesq approximation. However, it is known that in many engineering
applications and flow topologies, the Boussinesq assumption does not work for specific flow regions. The
RANS model used in this study are briefly explained in the upcoming sections below.

4.2.1 Standard k − ε model

The RANS models used in the project, employ the two-equaiton approach, i.e. two additional differen-
tial equations are solved for the transport of turbulence quantities, namely the turbulent kinetic energy
k and turbulent dissipation η. The formulation of the standard k − ε model, implemented in the CFD
code is the one by Jones and Launder [55], which for k and ε respectively reads:
  
Dk ∂k νt
= + ∇ · (kv) = ∇ · ν + ∇k + Pk − (ε − ε0 ) + Sk (4.19)
Dt ∂t σk
    
Dε ∂ε νt 1 ε ε0
= + ∇ · (εv) = ∇ · ν + ∇ε + Cε1 Pε − Cε2 − + Sε (4.20)
Dt ∂t σε Te Te T0
where S is a manually specified source term for both equations, Te = k/ε is the large-eddy time scale,
and Pk is the turbulence source term, comprised of turbulent and buoyancy production terms:
2 2 νt
Pk = νt S 2 − k (∇ · v) − νt (∇ · v)2 + β (∇T̄ · g) (4.21)
3 3 P rt

2 2 νt
Pε = νt S 2 − k (∇ · v) − νt (∇ · v)2 + Cε3 β (∇T̄ · g) (4.22)
3 3 P rt
√ √
with S being the modulus of the mean rate-of-strain tensor, S = |S| = 2S : ST = 2S : S. Following
the Boussinesq approximation, the turbulent viscosity term is approximated with:
k2
νt = ρCµ (4.23)
ε

Energy & Process Technology Page 42

 Delft University of Technology

Table 4.1: Model coefficients used in the standard formulation of the k − ε model, according to Launder
and Sharma [70]

Cµ σk σε Cε1 Cε2 Cε3

|vb |
0.09 1.00 1.30 1.44 1.92 tanh |ub|

If compressibility is included in the analysis, the code includes an additional production term, Ym =
− Cm kε
c2 , with c being the speed of sound and Cm is an additional model coefficient.

4.2.2 Realizable k − ε model

The Realizable formulation of the k − ε model contains a new transport equation for the turbulent
dissipation rate. Also, a variable damping function fµ expressed as a function of mean flow and tur-
bulence properties is applied to Cµ , and a second damping function f2 to Cε2 . This procedure lets the
model satisfy certain mathematical constraints on the normal stresses consistent with the physics of tur-
bulence (realizability). This concept of a damped Cµ is also consistent with experimental observations
in boundary layers.

The difference with the standard formulation is also expressed in the production term Pk and Pε by
the adoption of the curvature correction function fc , and also the damping functions f2 and fµ such that
 
2 2 νt
Pk = fc νt S 2 − k (∇ · v) − νt (∇ · v)2 + β (∇T̄ · g) (4.24)
3 3 P rt
νt

Pε = fc Sk + Cε3 β ∇T · g (4.25)
P rt
!
1
fc = min Cmax , p (4.26)
Cr1 (|η| − η) + 1 − min (Cr2 , 0.99)
 ε 2
η= (S : S − W : W) (4.27)
k


where W is the mean vorticity tensor, W = 12 ∇v − ∇vT .

The damping functions in the turbulent transport equations read

k
f2 = √ , (4.28)
k + νϵ

and
1
fµ = n √ h √ i √ o, (4.29)
S∗
Cµ 4 + 6 cos 1
3 cos−1 6√ 3
k
ε S:S+W :W
S∗ :S∗
∗
with S = S − 1
3 tr(S)I, applied to
k2
νt = ρfµ Cµ
. (4.30)
ε
The constants used are equal to the default values presented in the standard formulation of the model,
with the additional coefficients tabulated below. This model is substantially better than the Standard
k − ε model for many applications, and can generally be relied upon to give answers that are at least
as accurate. Both the standard and realizable models are available in STAR-CCM+ with the option of
using a two-layer approach, which enables them to be used with fine meshes that resolve the viscous
sublayer, which is explained in more detail in Section 4.4

Energy & Process Technology Page 43

 Delft University of Technology

Table 4.2: Model coefficients used in the realizable formulation of the k − ε model

Cε1 Cε2 σε Cr1 Cr2

η
max(0.43, η+5 ) 1.9 1.2 0.05 0.25

4.2.3 SST k − ω model

Another two-equaiton model is based on the specific rate of dissipation, ω, instead of the turbulent
dissipation rate, ε, which is sometimes referred to as the mean frequency of the turbulence (ω ∝ kε ). In
general, the kinetic energy and specific turbulent dissipation can be written in the form

Dk
= ∇ · [(ν + σk νt ) ∇k] + Pk − β ∗ fβ ∗ (ωk − ω0 k0 ) + Sk (4.31)
Dt
Dω 

= ∇ · ((ν + σω νt ) ∇ω] + Pω − βfβ ω 2 − ω02 + Sω (4.32)
Dt
The modification factors in the dissipation terms, fβ ∗ and fβ , are the free-shear modification factor
and vortex-stretching modification factors, further discussed in the Appendix. The so-called shear-stress
transport (SST) formulation adds an additional non-conservative cross-diffusion term ∇k · ∇ω, as shown
in the production term of ω. The addition of this term is triggered by a blending function F1 , which acts
as a trigger for the cross-diffusion term far from the walls but not near the wall. The blending functions
F1 and F2 (blending function in the formulation of the turbulent viscosity shown in Eq.(4.34)) are given
in the Appendix, along with the corresponding auxiliary relations necessary for their definition. The
turbulent kinetic energy production is identical to the k − ε model, whereas the production term of the
specific dissipation rate is given as:
  
2 2 1
Pω = γ S − (∇ · v) − ω∇ · v + 2 (1 − F1 ) σω2 ∇k · ∇ω
2 2
(4.33)
3 3 ω

The kinematic eddy viscosity is then calculated as:

 
α∗ a1
νT = k min , (4.34)
ω SF2

The model coefficients used are the default ones, shown in the table below.

Table 4.3: Model coefficients used in the SST k − ω model

a1 α∗ β∗ Ct β1 β2 σk1 σω1 σk2 σω2

0.31 1 0.09 0.6 0.075 0.0828 0.85 0.5 1 0.856

4.3 Calculation of heat transfer coefficients

Several heat transfer coefficients are available within the code, all shown in Eq.(4.40)-(4.39). These
include:

1. Heat transfer coefficient

2. Virtual local heat transfer coefficient
3. Local heat transfer coefficient
4. Specified y + heat transfer coefficient

Energy & Process Technology Page 44

 Delft University of Technology

For turbulent flows, the heat flux vector coming from the RANS equations is analogously calculated
using the Boussinesq approximation as shown below.
 
µt C p
q = − λf + ∇T (4.35)
Prt

The wall local value of the heat flux scalar is computed using the relevant scalars at the wall, such as:

Tw − T c
qw = |qw | = ρf cp,f uτ + (4.36)
Tc

The u∗ represents the velocity scale, which is calculated in accordance to the wall treatment method
+
chosen, as explained in the next section. This also holds for the RANS-averaged cell temperature T c .
The wall friction velocity is calculated using the non-dimensional tangential velocity component of the
velocity vector and the RANS averaged tangential velocity vector such that:
u∗
uτ = |vtangential | (4.37)
u+
with
vtangential = (vc − vw ) − [(vc − vw ) n] n (4.38)
where subscript c indicates cell centroid value, vw is the velocity at the wall (e.g. zero for no slip sta-
tionary wall) and n is the wall-normal vector. The value of u+ is also found using wall functions (see
next section on wall functions).

The first and most fundamental heat transfer calculation approach calculates the surface heat flux using
Eq.(4.41) with Tref = Tc , and then recasts for the heat transfer coefficient term such as:
qw
hbulk = (4.39)
Tw − Tref,h

The heat transfer coefficient is calculated using a preselected reference temperature Tref,h . Therefore, it is
worthwhile noting that this approach does not guarantee a positive value for the heat transfer coefficient
(i.e. negative for regions where the specified temperature is higher than the local wall temperature).
All other turbulent heat transfer coefficients mentioned earlier stem from the standard wall functions
(SWF). The second option is calculated without solving the energy equation in the fluid. Instead, only
flow information is used such that:
ρf (yc )cp,spec u∗
hvirtual = + (4.40)
T c (y + (yc ))
+
The value of T c is calculated using SWF. Since thermal material properties using this approach are
not available, a reference value for the specific heat cp,spec is specified, along with the molecular Prandtl
number P r and the specified turbulent Prandtl number P rt,spec . The necessity for such large number
of specified values makes this approach more ambiguous than the ones to follow. The wall surface heat
flux in Eq. 4.36 can be rewritten in the typical form of Newton’s cooling law:

qw = hlocal · (Tw − Tref ) (4.41)

The local heat transfer coefficient is hence always calculated and used internally in the code using SWF,
and Tref = T c from Eq. 4.41 above.

ρf (yc )cp,f (yc )uτ

hlocal = + (4.42)
T c (y + (yc ))

Energy & Process Technology Page 45

 Delft University of Technology

A similar alternative option is the specified y+ heat transfer coefficient, which is the HTC usually used
in the decoupled approach, since it allows to be exported to external codes. This is also the HTC used
in the decoupled approach for the turbulent offset jet (using a y + = 10). It is defined as:

ρf (yc )cp,f (yc )uτ

hy + = +
. (4.43)
+
T c yspec

The value is calculated at a explicitly specified value for y+ , which then allows for finding the reference
temperature Tref,y+ by using the value for the local heat flux and solving Newton’s law for cooling for
Tref :
qw
Tref,y+ = Tw − (4.44)
hy +
This definition of the reference temperature does not guarantee that it lies within the minimum and
maximum temperature range of the local fluid region. If Tref,y+ lies outside this temperature range,
then either the minimum or maximum temperature is used in its place, and the heat transfer coefficient
is recalculated accordingly.

A brief comprehensive comparison of the main differences of all four HTC approaches is given in Table
4.4.

Table 4.4: Summary of available heat transfer coefficients.

HTC Approach Definition Reference tem- Comment

perature
ρf (yc )cp,spec u∗
Virtual local HTC hvirtual = + Reference P r, P rt , Does not require energy
T c (y + (yc ))
and cp values are model to be activated; can
specified manually be used as an approxima-
tion of the Local HTC
ρf (yc )cp,f (yc )uτ qw
Local HTC hlocal = + Tlocal = Tw + hlocal Evaluated at the centroid
T c (y + (yc ))
of the cell next to the
boundary, that is h(y =
yc )
ρf (yc )cp,f (yc )uτ qw
Specified y + HTC hy + = + + Ty + = Tw + hy + Evaluated at a custom-
T c (yspec ) ref
specified value of y +
qw
HTC hbulk = Tw −Tref,h Manually specified Does not guarantee a pos-
on a rather empirical itive HTC value, choice
basis of reference temperature
is somewhat uncertain

4.4 Wall treatment functions

The importance of the boundary layer on the heat transfer characteristics of the flow has been well-
established throughout the years, particularly in the vicinity of the viscosity-affected region (viscous
sublayer and buffer layer, see Figure 4.2). Therefore, an accurate prediction of the flow across the
boundary layer is essential, particularly for conjugate heat transfer problems.

Energy & Process Technology Page 46

 Delft University of Technology

Figure 4.2: Typical velocity profile of a turbulent boundary layer

Wall functions provide an algebraically approximated value for main flow field quantities. The following
functions can be hence attributed:
v
1. Calculation of the wall shear stress (according to τw = ρu2τ |vtangential
tangential |
)

2. Calculation of the wall heat flux (as in eq.(4.41))

3. Imposition of values for non-dimensional turbulence quantities (ε+ or ω + , and Pk+ ) on the centroids
of the near-wall cells. Note that no wall treatment is available for the turbulent kinetic energy
term k + as it is strongly dependent on the Reynolds number of the flow. The definition is given
as k + = uk2 , and k is calculated from the initial conditions at the boundaries.
∗

The following three approaches are used in this thesis for the investigation of the conjugation effects in
turbulent flow: (i) High-y + treatment, (ii) All-y + treatment, and (iii) Two-layer all-y + . Within the code
used, the Two-layer all-y + treatment, which is available for the two-layer turbulence models, uses an ap-
proach that is identical to the all-wall treatment. However, specific values of turbulence dissipation rate
are imposed at the centroids of the near-wall cells to make it consistent with the two-layer formulation
of the underlying turbulence model. The high-y + approach relies on the assumption that the near-wall
cell lies within the log layer of the boundary layer, compared to the all-y + treatment which provides
a blending function, evaluating the field quantity depending on the region inside the inner layer (thus
blending the viscous sublayer and the log layers).

It should be noted that, however, not all wall treatments are available with every RANS model or
model variant. High Reynolds number models do not contain the ability to attenuate the turbulence in
the viscosity-affected regions and therefore only include a high-y + wall treatment. Low Reynolds number
models are offered only with a low-y + wall treatment and an all-y + wall treatment. The two-layer all-y +
wall treatment is only available with the two-layer k − ε model.

Two key non-dimensional variables are used, commonly referred to as wall units, namely the non-
dimensional wall-tangential velocity and the non-dimensional wall distance:
yu∗
y+ = (4.45)
ν
u
u+ = (4.46)
u∗
Each respective wall treatment approach proposes a formulation for five non-dimensional variables: ve-
locity, temperature, turbulent dissipation, turbulent kinetic energy, and production of turbulent kinetic
energy. The respective functions used for the High y + approach read:

u∗ = c1/4
µ k
1/2
(4.47)

Energy & Process Technology Page 47

 Delft University of Technology

1  ′ +
u+ = ln E y (4.48)
κ
where
′ E
E = (4.49)
f
with E = 9 being the log law offset, f is the roughness function, and κ = 0.42 is the von Kármán
constant. CFD roughness is an important parameter which is touched upon in the TP cooling analysis
due to the AM-related constraints of the cooling channels. The full formulation is given in the appendix
as well. The function for temperature similarly reads, with an additional term P :
 
1

T + = Pr ln E ′ y + + P (4.50)
κ

with the function P govenring the velocity at which the log layer and the viscous sub-layer of the thermal
profiles intersect [53]:
" 3/4 #  #
Pr Pr
P = 9.24 − 1 1 + 0.28 exp −0.007 . (4.51)
Prt Prt

(a) Non dimensional wall-tangential velocity u+ (b) Non-dimensional turbulent wall temperatureT +

Figure 4.3: Velocity (a) and temperature (b) as a function of non dimensional wall distance displaying
wall treatment for viscous sub-layer, log layer, and the blending approach covering the buffer layer

 
1
y+
ln 1 + κy + + C 1 − e−y /ym − + e−by
+ + +
u+ = (4.52)
κ ym
′
!  
1 E + κ 1
C = ln , b = 0.5 ym + + (4.53)
κ κ C ym
 
+
ym = max 3, 267(2.64 − 3.9κ)E ,0.0125 − 0.987 (4.54)
The temperature wall function is based on Kader’s law [58]:
   
1 1

T + = exp(−Γ) Pr y + + exp − Prt ln E y + + P (4.55)
Γ κ
4
0.01c (Pr y + )
Γ= (4.56)
1 + 5c Pr3 y +
c = exp(f − 1) (4.57)

Energy & Process Technology Page 48

 Delft University of Technology

The velocity scale needed for the shear stress calculation reads

µ |vtangential |
u∗ = γ + (1 − γ)Cµ1/4 k 1/2 (4.58)
ρy
with  
Red
γ = exp − (4.59)
11
√
ky
Red = . (4.60)
v

Table 4.5: Wall functions for both approaches for turbulence parameters required by the RANS model

Parameter High y + Blended approach

+
Turbulent dissipation, ε + 1
γ (y2k+2 ) + (1 − γ) κy1+
κy + q
Specific turbulent dissipation, ω + ω =√
+ 1 6 √ 1
β1 (y + )2 + β ∗ κy +
β ∗ κy +  
∂u+
Production of TKE, Pk+ Pk+ = γ µ+ t ∂y + + (1 − γ) κy +
1 1
κy +

Both high-y + and blended wall functions for velocity and temperature are illustrated in Fig.4.3. The
dashed lines are used to indicate the log-layer, i.e. modelling for y + ' 30 (high-y + function). The
respective functions for the turbulence variables are tabulated above in Table4.5.

Energy & Process Technology Page 49

 Part I

Canonical cases

50
Chapter 5

Numerical validation cases

The computational code used in this projects needs to be compared with analytical solutions or ex-
perimental data to test its validity and add to its credibility before embarking onto the optimization
case. Thereby, the physical modelling errors are quantified and carried over. Such errors may come
from simplifications in the mathematical model of the physical problem, assumptions in the analytical
solutions, or uncertainties in the experimental procedure. Hence, an ample evaluation of the program
accuracy is imperative prior to carrying out predictive simulations.

The chapter also includes the numerical comparison between the decoupled and conjugate heat transfer
approach, with the use of two types of boundary conditions on the wall for the decoupled approach,
specific to the flow case considered.

5.1 Laminar flat plate

As previously mentioned in literature, the problem of the conjugate heat transfer of flows over flat
plates, in particular, flows exhibiting laminar characteristics, presents a good suitability for solver vali-
dation purposes due to the existence of analytical solutions presented in literature.

Hence, the direct applicability and accuracy of numerical methods for conjugate heat transfer is tested
against solutions, the boundary conditions of of which most often include isothermal or adiabatic wall
conditions, and often a variety of simplifications around the velocity and temperature fields (as shown
later). In this study, the chosen analytical solution is given by Luikov [81] and is presented broadly in
the upcoming section, proceeded by the results and discussion.

5.1.1 Analytical solution

The analytical solution proposed by Luikov [81] comprises of two separate approaches:
• The direct solution of the partial differential heat transfer equation;

• The solution of the laminar boundary-layer equations, both applied in two dimensions (x-y).
For both cases the boundary conditions include a solid plate with a constant temperature of the bottom
surface set to Tb , and a free-stream of perfectly uniform temperature Ta = T∞ and velocity ua , both
imposed at the domain inlet (x = 0, y ∈ [0, ∞]) and far field (x ∈ [0, ∞], y = ∞]). It is also worthwhile
mentioning that the plate has a length l, which considerably higher than its thickness b. This allows for
the temperature profile within the plate region to be assumed linear (for small values of b. The governing
equations and their respective (shortened where deemed adequate) derivations are presented below. The
notations that are used in this chapter are also illustrated in Fig. 5.1 below.

51
 Delft University of Technology

Figure 5.1: Schematic of boundary conditions and respective notation used for the analytical solution of
Luikov.[81]

Differential heat transfer equation (DHT) approach

For an incompressible fluid, the steady state differential heat-conduction equation in a laminar boundary
layer reads:
∂Tf (x, y) ∂Tf ∂ 2 T (x, y)
vx + vy =α . (5.1)
∂x ∂y ∂y 2
The thermal diffusivity, α, is the ratio of the thermal conductivity and the product of the specific heat
capacity with the density of the fluid (k/ρCp ). For a full, more detailed derivation of the differential
heat transfer equation, the reader can refer to [85].

The boundary conditions illustrated in Fig. 5.1, are expressed as:

x = 0; Tf (0, y) = T∞ , (5.2a)
y = −b; Ts (−b, x) = Tb = const, (5.2b)
∂Tf (x, 0) ∂Ts (x, 0)
y = 0; −λf = −λs , (5.2c)
∂y ∂y
Tf (x, 0) = Ts (x, 0), (5.2d)
y = ∞; Tf (x, ∞) = T∞ = const. (5.2e)

The previously mentioned linear temperature distribution within the plate can be used from the heat
interface balance:
∂Tf (x, 0) λs
− + [T (x, 0) − Tb ] = 0. (5.3)
∂y bλf
Another key assumption used by Luikov is that the local longitudinal and transverse velocities are equal
to the their mean values:

vx = v x = const, (5.4a)
vy = v y = const. (5.4b)

Energy & Process Technology Page 52

 Delft University of Technology

The solution of (5.1) is published in [1] and [82], and reads:

   
T (x, y) − Tb K − 12 B λs y 1 1 λs y
= exp (K 2 − BK) · erfc K − B +
T∞ − Tb K −B λf b 2 2 λf Kb
 
1 1 1 λs y
+ erfc B− (5.5)
2 2 2 λf Kb
 
1 exp[(B/K)(λs /λf )(y/b)] 1 1 λs y
− · erfc B+ .
2 1 − (B/K) 2 2 λf Kb
The constants are presented in the Appendix, along with other constants related to the boundary layer
approach. and the local mean velocities in the streamwise and vertical directions in the boundary layer,
respectively, are given as:
v x = 0.66v∞ (5.5a)
 xv (− 21 )
∞
v y = 0.43v∞ (5.5b)
ν
The constants of 0.66 and 0.43 are derived using the similarity solution method from the corresponding
1/2
limit values of the auxiliary function f (ζ = (y/x)Rex ), which transforms the PDE-s system into an
ODE system. For more details on the analysis, one can refer to [1] (namely formula 3.1.19 and Fig. 3.3.).
This is a key assumption in the process of deriving an analytical solution for (5.1) using the conditions
in (5.2a)-(5.2e), and is one of the major differences between the numerical and analytical method, which
is discussed later.

The quantity of the local average Reynolds number, Rex , is expressed using (5.5a), which yields:
xv x xv∞
Rex = = 0.66 (5.6)
ν ν
and using the standard definition of the Prandtl number, P r:
µCp
Pr = (5.7)
λf

Boundary layer (BL) approach The analytical solution of heat transfer for laminar boundary
layers can be regarded more as an approximation, compared to the DHT method, and begins from the
integral heat transfer equation of a boundary layer:
Z δt  
∂Tf (x, 0)
(T∞ − Tf )vx (y)dy = α (5.8)
0 ∂y
Luikov uses the following assumption for the laminar boundary layer velocity distribution:
vx 3y 1  y 
= − . (5.9)
v∞ 2δ 2 δ
The temperature distribution in the fluid and solid regions is assumed to have linear and cubic expres-
sions, respectively as:
θ 1 = T f − T b = a 1 + b1 y + c 1 y 2 + d 1 y 3 , (5.10a)
θ2 = Ts − Tb = a2 + b2 y. (5.10b)
The boundary conditions are similar to (5.2a)-(5.2e), but in the form of θ1 and θ2 , such that:
∂θ
y = δt , θ1 = θ∞ = (T∞ − Tb ); = 0, (5.11a)
∂y
∂θ1 ∂θ2
y = 0, θ1 = θ2 = θw ; λf = λs , (5.11b)
∂y ∂y
y = −b, θ2 = 0. (5.11c)

Energy & Process Technology Page 53

 Delft University of Technology

Hence, from (5.11a)-(5.11c), (5.10a), and (5.10b):

3 T∞ − Tw 1 T ∞ − Tw 3
Tf (x, y) = Tw + ·y− · y , y ∈ [0, δt ] (5.12a)
2 δt 2 δt3
Tw − Tb
Ts (x, y) = Tw + · y, y ∈ [−b, 0]. (5.12b)
b
The wall surface temperature is denoted by Tw and reads
z
Tw (x) = Tb + (T∞ − Tb ) (5.13)
1+z
where the quantity z is:
3 λf b
z= , (5.14)
2 λs δt
with δt being the thermal boundary layer thickness. Equations (5.9) and (5.12a) can be subsequently
substituted into (5.8) which eventually yields:
r r
δt 3 13 1
·
3
= , (5.15)
δ 14 Pr
which is identical to the solution also presented in [28]. Similarly, the velocity boundary layer thickness is
expressed using the known relation, first derived by Blasius [106], using the similarity solution approach:
r
δ 280ν 4.64
= x= √ . (5.16)
x 13v∞ Rex
The Biot number is an another important nondimensional number. Hereby, the respective expression is
shown used for the analysis in this chapter. Using the Luikov’s boundary layer approach, which allows
for an analytical expression of the Biot number from the heat transfer coefficient h. To find the heat
transfer coefficient, one can fist express the wall heat flux (using the derivative w.r.t y of (5.12a))
∂Tf (x, 0)
q = −λf (5.17)
∂y
3 (T∞ − Tw )
= −λf (5.18)
2 δt
= h · (Tw − T∞ ), (5.19)

where (5.19) is from Newton’s Law of cooling. The convention used is a positive sign for expressing heat
entering the fluid domain. Further equating (5.18) to (5.19), and substituting δt with (5.15), yields:
3 1
h = λf · , (5.20a)
2 δt
√
√ Rex
= 0.332 · λf
3
Pr , (5.20b)
x
which is also identified in [106], where an ordinary heat transfer flat plate problem with a constant
wall temperature is considered (without taking into account for conjugation). Thereby, the effect of
conductivity of the solid is ignored. However, this assumption only holds for z ≪ 1, which was also
applied in the derivation of the thermal boundary layer from eq.(B.1g) (N.B. the limit value in the
denominator). Such a condition is satisfied by having a solid region conductivity considerably higher
than that of the fluid, which corresponds namely to cases with a low Biot number
h·b
Bi = . (5.21)
λs
One of the benefits of using the simplified boundary layer approach, is that the heat transfer coefficients
and hence the Biot number can be easily derived analytically. Due to the constant conductivity and

Energy & Process Technology Page 54

 Delft University of Technology

√
plate thickness, the Biot number is proportional to ( x)−1 .

A limitation of this approach that is worthwhile mentioning is the required thermal conductivity of
the solid for the method to be counted accurate. In the scope of this project, the averaged Biot number
is used, which takes the averaged heat transfer coefficient:
Z
1 L
h= hdx = (5.22)
L 0
0.664 √ 3
p
= λf P r ReL , (5.23)
L
where ReL is the Reynolds number at the plate trailing edge (x = L). The averaged Biot number thus
becomes:

h·b
Bi = (5.24)
λs
   p
b λf √
= 0.664 · · P r ReL .
3
(5.25)
L λs

Referring back to (5.15), the calculation of the mean Biot number using the boundary layer approximation
holds only for z ≪ 1 (i.e. values where the thermal conductivity of the solid region is considerable higher
than that of the fluid region). The results from the evaluated cases are presented in the following sections
and the fulfilment of this criterion is briefly discussed.

5.1.2 CFD setup and boundary conditions

The computational domain consists of a plate (0.2m × 0.01m), and two additional extensions with
a length of 5b at the entrance and outlet of the domain in order to ensure valid flow behaviour and
convergence. The fluid domain extends of a height of 10b.

For the purpose of this study, two values of Biot number were used, Bi = 0.1 and Bi = 1, in order
to investigate the applicability of the two analytical methods mentioned above and, respectively the
solver’s accuracy to be validated at two separate physical conditions.

Air is simulated as an ideal gas at 1.03bar, where the the necessary properties are calculated at 1000K,
with the respective values tabulated below.

The boundary conditions are applied and corresponding equations discretized and solved by using the
commercial CFD code in STAR-CCM+, for both fluid and solid domains.

Table 5.1: Calculated properties of air at 1000K. [65]

R (J/kgK) ρ (kg/m3 ) cp (J/kgK) γ µ (N s/m2 ) λf (W/mK) Pr

282 0.3525 1142.6 1.335 3.95e-5 0.06808 0.66

Table 5.2: Resultant solid thermal conductivity from the two average Biot numbers

Bi 0.1 1
λs (W/mK) 2.786 0.2786

A freestream boundary condition normal to the inlet plane is imposed using a velocity of 12m/s. The

Energy & Process Technology Page 55

 Delft University of Technology

resultant Reynolds number is hence:

ρair u∞ Lplate
ReL = = 21, 418 (5.26)
µair

According to [106], where for a plate with a sharp leading edge, the laminar–turbulent transition occurs
at a critical Reynolds number, Recrit = 320, 000, and hence one can safely conclude that a laminar flow
is present, which was a deliberate choice, and no turbulence modelling will be needed as a part of this
study.

Figure 5.2: Problem domain schematic and boundary conditions.

Using the equation for the average heat transfer coefficient (5.23), one can obtain h = 28.76W/m2 K.
and rearrange (5.25) for λs with the specified averaged Biot numbers and obtain the values in Table.
5.2. However, as previously mentioned, the criterion of z ≪ 1 is not met for Bi = 1 (z = 0.631). The
difference in the heat transfer coefficients calculated for both Biot number settings (and hence z) for
each solution approach, are plotted in Figure (5.3), which is created by using the formulae in [81] for the
local Nusselt numbers (given in the Appendix B).

Comparing the two solutions, the values for the heat transfer coefficient from the boundary layer ap-
proach, using Bi = 1 are greatly underestimated compared to the differential heat transfer approach, in
which the thermal conductivity of the solid is taken into account. For the latter, the two Biot number
settings lead to almost matching solutions. The mismatch for the boundary layer approach holds partic-
ularly well in the region after the leading edge approximately between x/L ∈ (0.05, 0.25). One can thus
find that the average Biot number in the BLE is in fact above one. For the sake of consistency in this
study however, the notation Bi = 1 is kept.

Energy & Process Technology Page 56

 Delft University of Technology

Figure 5.3: Heat transfer coefficient along the plate surface for Bi = 0.1, and Bi = 1 for each of the two
solution methods by Luikov [81].

Expanding on the boundary conditions applied to the domain (Figure 5.2), periodic wall boundary
conditions are applied on the surface in front of the plate and after the trailing edge. This is done in the
attempt to replicate as close as possible the boundary conditions applied in the studies of turbomachin-
ery components, such as the case of the transition, described in Section 7.1. All surfaces of the plate,
except for the fluid/structure interface are treated as adiabatic no-slip walls. The side walls of the fluid
domain, along with the top surface are also treated as slip-adiabatic walls.

The fluid/solid interface is the only surface where the no-slip boundary condition is present. The type
of interface used was a mapped contact interface, since it allows for data mapping when a conformal
match between the fluid and solid meshes is lacking. Such instances occur when different resolutions are
necessary on both sides of this interface. Such is the case for instance in this problem, where the heat
conduction in the solid region requires a coarser mesh compared to resolving the boundary layer in the
fluid domain. For further details on the choice of the CHT interface settings, one can refer to Appendix
B.

Figure 5.4: Mesh used in the validation and HT methodology comparison. Grid is of size 60x120 and a
first cell height of ∆ymin = 7.2e-05m with a logarithmic stretching factor of 1.061

The fluid region was modelled on a finite volume method domain using a laminar solver and a seg-
regated fluid enthalpy model (temperature is calculated from enthalpy using the equation of state) in
accordance with the SIMPLE algorithm. Variable gradients are calculated using the Hybrid Gauss-Least
Squares Method. The limiter of choice for the reconstruction gradients was Venkatakrishnan. Second-
order upwind differencing was employed for the convection scheme. The reader can refer to the theory of

Energy & Process Technology Page 57

 Delft University of Technology

FV discretization, including transport equations, differencing schemes and gradient calculation methods
in the Appendix, where a brief derivation and explanation of key terms is included.

A steady state finite element solver was applied to the solid region in order to solve for the solid temper-
ature. A nearest neighbour interpolation method was used to map data (temperature and heat transfer
coefficient) onto the solid domain. This method, compared to other methods such as the higher-order
stencil (alternative option in STAR-CCM+, which makes use of a distance-weighted, least-square ap-
proach) provides a quicker interpolation at the expense of accuracy. Due to the simplicity of the problem
and hence resultant meshes, it proves to be a reasonable choice. The shortened read on the numerics of
FEM is also included in the Appendix.

The convergence criteria for the simulation used were based on minimum residual values of the Conti-
nuity equation, Energy equation, and the Momentum equation (all directions). For all 5 equations the
value was set to 10−5 . A fully converged solution is thus achieved in 260 iterations for the finest mesh.

A structured mesh approach was used for both cases of Bi = 1, and Bi = 0.1. A mesh refinement
was strongly recommended in a similar study by Verstrate [117], and hence the resolution of the mesh
was chosen based on a mesh independence, which took both refinement in the -x and -y directions into
account. The chosen mesh is shown above in Fig. 5.4, along with the respective resolution at the wall.

5.1.3 Results and discussion

The results presented in this section consist of temperature in both solid and fluid domain and a com-
parison of the velocity in the laminar boundary layer to analytical solution. As previously mentioned,
two values of the averaged Biot number are used, 0.1 and 1. The comparison of the two simulation
methodologies (decoupled and CHT) are discussed later in Section 6.

An important remark regarding all post-processed figures presented below: the plots includes interface
quantities measured only from the FVM mesh, the reason being that it coincides with the interpolated
data on the FEM and thus avoiding confusion in the plots. Figure 5.8 presents the interface temperatures
and heat transfer coefficients at Bi = 0.1. The leading edge is primarily where discontinuities exist with
the analytical solution. This error tends to decrease downstream but remains firmly at an offset of ≈ 8K,
even at the trailing edge. This discontinuity is less apparent in the heat transfer coefficient calculation
where almost no difference is found after roughly x/L = 0.3. The large change in temperature that is
required over such a small length is the reason why such a difference is to be expected, particularly for
low Biot numbers. Additionally, a similar observation is reported by Verstraete [117], where the steep
temperature change is also predicted inaccurately.

An interesting observation coming from Figure 5.6 is present. The gradients of the thermal profiles
agree well with the analytical solution, despite the small difference. The difference however, tends to
become more pronounced in the downstream direction, which is contrary to what one might expect given
the results in Figure 5.8. The peak differences between the differential heat transfer approach for all
three locations in the downstream direction are 1.95%, 3.3%, and 3.6%, respectively.

Energy & Process Technology Page 58

 Delft University of Technology

900 400
Luikov B.L. solution Luikov B.L. solution
Luikov diff. h.t equation 350 Luikov diff. h.t equation
850 CHT CHT
300
800
250

750 200

150
700
100
650
50

600 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

(a) Interface temperature (b) Heat transfer coefficient

Figure 5.5: Interface temperature and heat transfer coefficient comparison with the differential heat
transfer approach [81], at Bi = 0.1

0.7 0
0.35 Luikov BLE, x/L = 0.25
Luikov diff.h.t., x/L = 0.25
0.6
0.3
-0.2 CHT, x/L = 0.25
Luikov BLE, x/L = 0.50
0.5 0.25
Luikov diff.h.t., x/L = 0.50
0.4 -0.4 CHT, x/L = 0.50
0.2
Luikov BLE, x/L = 0.75
Luikov diff.h.t., x/L = 0.75
0.3 0.15
840 860 880 900 920 940
-0.6 CHT, x/L = 0.75
0.2
-0.8
0.1

0 -1
600 700 800 900 1000 600 650

Figure 5.6: Temperature distribution across the fluid region (left) and solid (right) and comparison to
Luikov solution, with a Bi = 0.1

The velocity profiles of the CFD results are compared to the Blasius solution [106] at the previously used
locations downstream for both Biot number settings. The quantity η corresponds to the nondimensional
distance from the plate, given in correspondence with the coordinates in Fig. 5.1:
r
v∞
η=y , (5.27)
νx

in which y and x are the normal distance from the wall and the distance from the leading edge. Hence,
the velocity boundary layers are shown in Figure 5.7 and Figure 5.10, respectively.

Energy & Process Technology Page 59

 Delft University of Technology

1 1

Blasius BL solution
Blasius BL solution
0.8 x/L = 0.25 0.8 x/L = 0.1
x/L = 0.5
x/L = 0.75

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7

(a) x = 0.05, x = 0.10, x = 0.15 (b) Leading edge vicinity

Figure 5.7: Comparison of velocity profiles to Blasius solution [106], at Bi = 0.1

The velocity boundary layer thickness is found to be lower than the one proposed by Blasius, with the
tendency for the error to decrease downstream. This coincides with the interface temperature profiles,
but is contradicting the findings of the thermal boundary layer, which are displaying exactly the opposite
behaviour downstream.

When it comes to the computation of the case at Biot number equal to 1, the temperature and heat
transfer quantities from STAR-CCM+ are presented in 5.9 along with the direct solution of Luikov of
the differential heat transfer equation. In general, a good agreement is found with the CHT solution.
The largest difference is observed at the trailing edge, where the temperature is overpredicted by the
numerical method, and also in the vicinity of the first quarter of the plate upstream, where the value is
lower. The former is also where the maximum difference of 11K occurs.

Such phenomenon is also observed by Verstraete in [117] and Racca [99], where the former author
reported a similar increase of temperature in the second half of the plate in his attempts to study the
stability of various coupling algorithms. Coincidentally, this observation also happens for the CHT cou-
pling algorithm which is similar to what is used in this numerical study (Table B.1), namely the heat
transfer coefficient forward temperature back method (hFTB) method. In [117], the difference was ex-
plained by the remaining heat flux difference between solid and fluid domain at convergence. For a more
detailed read on the main differences between coupling schemes the reader can refer to the Appendix.

Energy & Process Technology Page 60

 Delft University of Technology

1000 400
Luikov B.L. solution Luikov diff. h.t equation
Luikov diff. h.t equation 350 CHT
950 CHT
300
900
250

850 200

150
800
100
750
50

700 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

(a) Interface temperature (b) Heat transfer coefficient

Figure 5.8: Interface temperature and heat transfer coefficient comparison with the differential heat
transfer approach [81], at Bi = 1

The temperature profiles in the y-direction for both fluid and solid region is again plotted in Figure
5.9 at the three previously mentioned downstream locations. The difference with the solution of the
differential approach is rather small. Overall the thermal boundary layer thickness is similar. An
interesting observation is that the temperature difference for x/L = 0.25 is the highest near the wall
(i.e. y/b < 1), and close to zero as the end of the boundary layer is approached. This is contrary to the
CHT prediction for x/L = 0.5 and x/L = 0.75, where the difference is close to zero near the wall, but
the thickness is underevaluated. This could partially explain the temperature profile at the interface,
in which case the gap with the Luikov solution is the smallest for the two aforementioned x-locations.
The difference, however, remains negligible. A similar observation is present in the solid region as well:
the gradient of the CHT solution is lower than the analytically calculated one as the leading edge is
approached, and virtually corresponds precisely to the Luikov solution further downstream.
0.7 0
0.3 Luikov BLE, x/L = 0.25
0.28
Luikov diff.h.t., x/L = 0.25
0.6 0.26

0.24 -0.2 CHT, x/L = 0.25

Luikov BLE, x/L = 0.50
0.5 0.22

0.2
Luikov diff.h.t., x/L = 0.50
0.18

0.4 0.16
-0.4 CHT, x/L = 0.50
0.14 Luikov BLE, x/L = 0.75
0.12
Luikov diff.h.t., x/L = 0.75
0.3 0.1
-0.6
880 900 920 940 960 980 CHT, x/L = 0.75
0.2
-0.8
0.1

0 -1
700 750 800 850 900 950 1000 600 700 800

Figure 5.9: Temperature distribution across the fluid region (left) and solid (right) and comparison to
Luikov solution, with a Bi = 1.0

The good fit of the thermal profiles with the analytical solutions can also be related to the velocity
boundary layer. The velocity profiles again at three distinct regions, similarly to the analysis at Bi = 0.1
above, are plotted in Figure 5.10. An overall good prediction is achieved by STAR-CCM+ when it comes
to the slope of the velocity within the boundary layer. The difference tends to increase upstream, where
the gradients are the highest. The close temperature fit at x/L = 0.1 in Fig. 5.8) can also be traced
down in Fig. 5.10(b): the difference in this case is smaller than in Fig.5.7, where a comparatively larger

Energy & Process Technology Page 61

 Delft University of Technology

divergence with the analytical solution was found in terms of interface temperature for the respective
location. The exact reason as to why this happens in conjugate heat transfer terms for both Bi = 0.1
and Bi = 1, however, is inconclusive and difficult to determine. A possible reasoning is the inherently
large gradients as previously discussed for Bi = 0.1.

1 1

Blasius BL solution Blasius BL solution

0.8 x/L = 0.25 0.8 x/L = 0.1
x/L = 0.5
x/L = 0.75

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7

(a) x = 0.05, x = 0.10, x = 0.15 (b) Leading edge vicinity

Figure 5.10: Comparison of velocity profiles to Blasius solution [106], at Bi = 1

The averaged values of the quantities on the centerline (not of the entire 2D surface) of the solid
interface are summarized in Table 5.3 below.

Table 5.3: Summary of average values for temperature and heat transfer coefficient from both simulations
and the analytical solution [81].

Bi = 0.1 Bi = 1
2
Twall (K) h(w/m K) Twall (K) h(w/m2 K)
CHT 641.2765 35.02 810.915 38.06
Luikov (diff.h.t.eq.) 634.7062 28.76 813.493 28.76

In both Biot number settings, the average values differ from the analytical solution, the larger error
being in the heat transfer coefficient.

The average heat transfer is overpredicted in both cases. The change in the average temperature for
Bi = 1, however, is smaller than the resultant increase for Bi = 0.1, given that ∆T (%) is ≈ 0.3% and
1%, respectively.

5.1.4 Summary
The numerical validation study investigated the accuracy of the STAR-CCM+ solver at two phys-
ical conditions on a simple flow problem where an analytical solution is present. Following the mesh
independence study (computed on a low Biot number setting), the following was observed:
• For the computation of the low Biot number setting, small differences were present in terms of the
interface temperature and heat transfer coefficient with the Luikov solution.

• The results of the Biot number of unity were comparatively better in terms of the interface temper-
ature prediction. It is assumed that this is primarily due to the lack of high temperature gradients
near the leading edge, owing to the high thermal conductivity of the solid region.

Energy & Process Technology Page 62

 Delft University of Technology

• In terms of velocity and thermal boundary layers, the case of high Biot number was also more
consistent with the analytical solution. On the other hand, the low Biot number computation
yielded thermal profiles that contradict the velocity profiles behaviour (boundary layer thickness).
• The exact reasons why the difference (never mind little) exist can be partially explained by the
assumptions used in the derivation of the analytical solution. The problem formulation (especially
in cases of low Biot numbers) introduces large gradients in the vicinity of the leading edge which
the solver finds hard to predict and opens room for further analysis as to why this happens even
with refinements in the x-directions (shown in the mesh study - see Figure B.5).

• However, for the purposes of this study the STAR-CCM+ CHT solver does a good job in predicting
the temperature profiles in both fluid and solid regions for both settings. The accuracy achieved in
terms of average values is also satisfactory to the prediction of thermal stresses and hence allows
for adequate lifing analysis, especially in cases of MDOs, where the focus lies on overall trends and
tradeoffs rather than exact solutions.

The introduction of turbulence in the upcoming section increases the complexity of the problem and
addresses different aspects more relevant to a turbine transition piece, presented in the sections to
follow.

5.2 Turbulent offset jet

5.2.1 CFD setup and boundary conditions
As for the validation of the turbulent offset jet a different procedure was followed: experimental data
for flow and temperature fields, as opposed to semi-empirical relations or solutions.

As a first step of the validation process was, similarly to the case with the flat plate, a mesh study
in which, the flow was validated against measurement velocity data presented by Pelfrey and Liburdy
[95], thus eliminating the effect of the mesh and boundary conditions as much as possible before conduct-
ing the numerical study. A code validation step followed in which the experimental interface temperature
on the plate was compared to the one predicted by the CHT solver in STAR-CCM+.

The computational domain that was initially utilised for this part of the validation process comprised
of a simple rectangular domain, using a structured directed mesh in both y- and x- directions. The
domain was very closely replicating the one used in the study by Vishnuvardhanarao [118], as mentioned
in Section 2.2.2, where no inlet duct was added for the jet itself.

The velocity profiles however were unsatisfactory in terms of accuracy with the experimental data.
Therefore, despite the inherent simplicity of the domain to discretize and solve using such a mesh, both
fluid and solid, an unstructured polyhedral grid was chosen for the fluid region in order to be consistent
with methodologies applied in turbomachinery applications, e.g. a combustor transition piece, which
most often cannot be discretied in a structured way due to the complexity of geometry and the complex
flow field. Additionally, the computational domain was altered so that it was attempted that it repli-
cate the experimental setup, described in the paper, using boundary conditions that would reproduce
the experimental boundary conditions with the highest possible fidelity, allowing room for assumptions,
where information was lacking.

The entirety of the computational domain used in this study is displayed in Figure C.2, followed by
a close-up of the region where the offset plate is contained within the plexiglass sidewalls. The dimen-
sions used are idenical to those in the experimental domain. The pressure outlet boundary conditions
were placed at a large distance from the jet discharge region so that they do not interfere with the
numerical results and to diminish the likelihood of reversed flow.

Energy & Process Technology Page 63

 Delft University of Technology

The schematic shown below in Figure C.1 provides information with regards to the experimental setup
by Pelfrey and Liburdy. The setup allows for a multitude of offset ratios using an adjustable plate.
The numerical study described here was computed using OR = 7. The choice of the setting was made
arbitrarily. The width of the exit nozzle is t = 12.5mm.

Similarly to the laminar flat plate case study, the FVM is used for the fluid region along with a tetrahe-
dral FEM mesh for the solid region. The zero-pressure gradient boundary condition (blue annotation in
Figure 5.11 was applied on all faces which do not correspond to walls or inlets. The interpolation stencil
and settings used for data transfer across the interface on the plate are identical to the ones used in
the laminar flat plate study (see Table B.1). An under-relaxation factor of 0.7, 0.3 and 0.9 are used for
the velocity, pressure and energy in the fluid region. A turbulence intensity I of 2% is specified at the
inlet for the kinetic energy equation, along a turbulent viscosity ratio (µT /µL ) of 10 for the dissipation
equation. The temperature specified is θ = 1, where θ is the nondimensionalized quantity:
T − T∞
θ= , (5.28)
T j − T∞

with T∞ and Tj being the ambient (300K) and jet outlet temperatures (385K), respectively. An additional
distinction between the continua set-up of the flat plate study is the calculation of the properties of air.
In this case, variable air properties are used and are calculated using Sutherland’s Law. This holds for
dynamic viscosity, thermal conductivity and molecular weight. The properties of the solid region were
all set to constant values. Due to the problem deifinition of the optimization task, which is intended for
a combustion component, an additional step was taken for the choice of the Turbulent Prandtl number.
The number changes from 0.9 at the wall to 0.4 in the free stream flow with a specified wall distance of
1.0mm and free stream distance of 5.0mm, using a linear interpolation for the in-between region. Such
manipulation of the Turbulent Prandtl Number is common for combustion flows in which the effects
on the resultant low turbulent scalar transport is to some extend mitigated. In a RANS combustion
simulation this is also done by decreasing the Turbulent Schmidt number, Sct , thus artificially increasing
the turbulent diffusion coefficient, Dt . Many studies confirm such considerations using a variety of
turbulence models. Examples of some are the works of He et al. [43], Ivanova et al. [51], and Baurle
[10].

Figure 5.11: Close-up of dark gray region in Figure C.2: schematic of boundary conditions used in the
turbulent offset jet problem. Domain width is 150mm (identical to experiment)

The value of 0.9 for the turbulent Prandtl number in the near-wall region is entirely based on turbu-
lence theory, which is consistently used in literature for air [67], [126]. However, it is widely known that

Energy & Process Technology Page 64

 Delft University of Technology

it varies significantly within a turbulent boundary layer [64]. The usage of a constant value nonetheless
is common for numerical simulations.

It is worthwhile mentioning that the thickness of the solid region was set to 0 for the mesh study,
where an adiabatic wall boundary condition was used, since the study of Pelfrey & Liburdy, used to
validate the mesh does not include conjugate heat transfer measurements, but only flow characteristics.
The thickness of the plate is 0.635cm and the material is balsa wood4 . The chosen turbulence model was
the realizable k − ε model with the blended WF approach. Parameters used for the mesh independence
study were reattachment length, Xcr , the maximum value of the coefficient of friction along the middle of
the interface, and a comprehensive comparison of the velocities with the experimental data from Pelfrey.
For the full analysis of the mesh independence study and results, refer to Appendix C.

In the following subsection, a comparison of the three turbulence models of choice will be discussed
(realizable k − ε, standard k − ε, SST k − ω), in which two wall-treatment methods are used: All-y +
(or Two-layer y + ) and High-y + wall functions for a fine and coarse mesh, respectively. This procedure
is also followed later in the comparison of the conjugate heat transfer approach to the decoupled ap-
proach, where of interest is the effect of flow properties and wall-treatment methods on conjugation
and under what conditions is the difference between the decoupled (adiabatic) and conjugate method
minimized/maximised.

5.2.2 Results and discussion

Given the experimental apparatus used in [47], the paper lacks a multitude of important details around
key parameters. The parameters are presented in Table 5.4 below. The results from all cases with the
respective measures taken are presented in 5.12, where the nondimensional wall temperature is given.

The results correspond to the expectations and there is no noticeable change in the behaviour of the
profile (no change of turbulence model). The temperature gradient in the recirculation, impingement,
and wall jet regions are almost identical. The difference consists only in a shift of the profile higher or
lower. The increase of the inlet velocity and also turbulence intensity lead to a seemingly identical level
of heat transfer augmentation. Further refinement of the solid mesh leads to an increase in heat flux
and respectively to wall temperature possibly due to the resultant interpolation at the interface. The
opposite effect is manifested by an increase in the thermal conductivity which lowers the temperature
gradients within the solid, explainable by the increase in heat flux across the boundary.

Table 5.4: Sources of errors and unknowns in the numerical procedure followed for the CHT validation.

Unknown parameter Experiment (Holland, Countermeasure

1990)
Turbulence intensity, I (%) Only Reynolds number and dis- Sweep of various values for
charge coefficient present I (%)
Thermal conductivity of plate Only thickness and material Two values of thermal
name conductivity of the given
material
Velocity inlet data No velocity data available for Velocity scaled down ac-
CHT experiment, only Re cordingly from Re number
Solid plate mesh - Mesh refinement in the
solid reigon

Therefore, for further analysis, the baseline settings described in the previous subsection will be used.
This includes a thermal conductivity of 0.0339W/mK, no mesh refinement (mesh presented in Figure

Energy & Process Technology Page 65

 Delft University of Technology

C.3e), turbulence intensity of 2%, and jet outlet velocity of 8m/s.

1
M2 (Baseline case: Realizable k-eps, All-y+)
Tu = 20%
0.9
Inlet velocity = 12m/s
Plate refinement
Plate refinement x2
0.8
Increased p
Holland & Liburdy (1990)
0.7

0.6

0.5

0.4

0.3
0 5 10 15 20 25 30 35 40

Figure 5.12: Interface temperature variation using countermeasures on the realizable k − ε model

Figure 5.14 indicates the results from all four computations on mesh M2, where a blended or low-
y + wall function modelling is applied to all three turbulence models investigated. The aerodynamic
characteristic of the results are discussed later, hence the averaging on the plot x-axis: ∆Xcr is intended
to take care of difference between the reattachment lengths for all models. Thereby, focus is put mainly
on the behaviour of the temperature before and after the reattachment point without emphasising on
the exact prediction, but rather on average temperatures. Thus, if one assumes that all computations
result in a reasonable match with Xcr , the results lie within a satisfactory region off the experimental
data.

10-5
10-6 9
1.4
Realizable k- , High-y+
Realizable k- , Two-layer 8 Standard k- , High-y+
1.2 Standard k- , Two-layer
SST k- , High-y+
SST k- , All-y+ 7
1 6

0.8 5

4
0.6
3
0.4
2
0.2
1

0 0
0 5 10 15 20 0 5 10 15 20 25 30 35 40

(a) Fine mesh (b) Coarse mesh

Figure 5.13: Comparison of eddy viscosity for all three models

In nominal terms, the difference at the peaks for the standard and realizable k − ε models is ≈ 2K and
≈ 4.2K, which considering the temperature difference of 85K between the bottom surface of the plate
and the jet outlet from the problem boundary conditions, is in good agreement. A more realistic fit is
achieved by both realizable and standard formulations of the k − ε models, compared to SST k − ω. The
difference with the SST k − ω model is larger by ≈ 7.34K. The reason for this can be traced back to the
formulation of the turbulent viscosity term in the model equations. The respective turbulent viscosities
along the interface are given in Figure 5.13. For all models, µt is at a local minimum at the reattachment

Energy & Process Technology Page 66

 Delft University of Technology

point, similar to other flow quantities (see friction coefficient in Figure C.4c), followed by a sharp increase
in the impingement region.

1
M2, Realizable k-eps, Two-Layer All-y+
M2, Standard k-eps, Two-Layer All-y+
0.9
M2, SST, All-y+
M2, SST, Low-y+
Holland & Liburdy (1990)
0.8

0.7

0.6

0.5

0.4

0.3
0 5 10 15 20 25 30 35 40

Figure 5.14: Comparison of interfacial temperature profile for all three models using finer mesh (M2)
and blended wall function approach

The values of the turbulent viscosity term from the k − ε, particularly in the near wall region, are
higher than the values computed by the SST k − ω. The difference is not overwhelming but is the reason
why the temperature increase on the surface is lower for the former model. When the turbulent viscosity
is low in the vicinity of the wall, the velocity of the flow in the region will become lower, which will
inherently lead to a drop the heat transfer rate. This results in in lower temperatures at the wall and
low heat fluxes through the solid walls. The alternative formulation of the turbulent viscosity term in
the SST turbulence model did not lead to a better prediction for this particular problem. However, this
greatly depends on the type of flow and its features, since the model has been reported to perform better
than the (realizable) k − ε model in multiple studies involving conjugate problems, [21] [48].

As it can be seen in Figure 5.15, where the temperature profiles in the wall-normal directions at two sep-
arate recirculation locations are compared to the experimental results of Holland and Liburdy [47], there
is no major difference in the predictions in the core of the flow for all computations. The similarity in the
results is due to the two-zonal behaviour of the SST model, which switches to the standard k − ε model
in the free-stream region. The difference in the temperature profiles however becomes more pronounced
near the wall, where the ω − SST model leads to large gradients, not present in the experimental data.

15 15 M2, Realizable k-eps, Two-layer All-y +

M2, SST, All-y+
M6, Realizable k-eps, High-y +
M6, SST, High-y +
10 10 Holland & Liburdy (1990)

5 5

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

(a) (b)

Figure 5.15: Nondimensional temperature in the wall-normal direction at (a) 3mm in front of the exit
nozzle and (b) X = 6.69 (midway in the recirculation region in the experiment).

Energy & Process Technology Page 67

 Delft University of Technology

Another important parameter is the thermal decay of the jet, which is presented as the decrease of the
maximum jet temperature downstream. The computed value is compared to the experimental data in
Fig.5.16. Evidently, neither of the models studied satisfy the real rate of decay. This is in accordance to
the expectations following also the slower rate of velocity decay compared to the experiment, as shown
previously in the mesh independence study. The realizable k − ε model tends to perform slightly better
in this regard. The difference however remains large.

1.1

0.9

0.8
M2, Realizable k-eps, Two-layer All-y+
0.7
M2, SST, All-y+
0.6 M6, Realizable k-eps, High-y+
M6, SST, High-y +
0.5 Holland & Liburdy (1990)

0.4
0 2 4 6 8 10 12 14 16 18

Figure 5.16: Decay of the maximum jet temperature downstream (SST model in red and k − ε model in
black).

1
M6, Realizable k-eps, High-y+ Wall Treatment
M6, Realizable k-eps, Two-layer Wall treatment
0.9
M6, Standard k-eps, High-y+ Wall Treatment
M6, SST, High-y+ Wall Treatment
M6, SST, All-y+ Wall Treatment
0.8
Holland & Liburdy (1990)

0.7

0.6

0.5

0.4

0.3
0 5 10 15 20 25 30 35 40

Figure 5.17: Comparison of interfacial temperature profile for all three models using coarse mesh (M6)
and high-y + wall function approach

Referring back to the temperature plot Figure 5.14, the temperature gradient in the x-direction is
similar for all models, where temperature tends to dissipate quicker than in the experiment. A signifi-
cant divergence from the physical result is the underprediction of the results for the SST k − ω model
in the recirculation region. The gradients, as a result, are much steeper for both types of wall function
treatment. As previously mentioned, the SST model does provide superior computations where adverse
pressure gradients and separations are present. However, it has been reported that the model can lead
to overpredicted turbulence levels in regions with large normal strain, i.e. stagnation zones and zones
with high acceleration [92], such as the recirculation zone in a offset-jet.

When it comes to the temperature profiles on the coarse grid, shown in Figure 5.17, the overall va-

Energy & Process Technology Page 68

 Delft University of Technology

lidity of the previous findings remain: the SST model leads to higher interface temperature and the
dissipation downstream does not match the one in the experiment. Varying the wall function treatment
leads to less visible effects on the SST model compared to the realizable k − ε. When the two-layer
formulation is used on the coarse grid in the latter, the model tends to predict the experimental data
fairly well in all regions. The difference, however, between the realizable and standard k − ε models
is diminished on the coarse grid. Additionally, the SST model does not display such large gradients
upstream in the recirculation region, and the curve is much more ’flatter’ than in the computations per-
formed on M2, particularly in the wall jet region. For future computations it is therefore recommended
to investigate the effect of different discretization schemes of the convection term, in order to address
this issue. Moreoever, no resources in open literature have been found discussing numerical dissipation
sensitivity of the k − ω and k − ε models.

5.2.3 Summary
In general, The exact boundary conditions are unknown and could have a noticeable effect on the
reattachment length. The goal of this study is to validate the CHT solver in STAR-CCM+ and by
doing so to choose the best option in terms of a RANS turbulence model and a wall treatment method.
Predicting an accurate reattachment length was not part of the objectives (RL could be predicted with
fine turbulence modelling tuning). The key findings from this validation case thereby include:

• Attempt was made to eliminate effect of BCs, by reproducing the experimental domain as closely
as possible. If the reattachment points are artificially matched (by shifting the results on the plot),
the CHT solver was a good agreement with the data in general.
• The resultant temperature profile for all computation cases considered is in very good agreement
with the experimental data from Holland & Pelfrey. The temperature profiles in the recirculation,
impingement, and wall jet regions are within the acceptable accuracy implying that the presented
turbulence models are suitable for predicting flows similar to the case of an offset jet flows.
• Coarser mesh study results in a more accurate solution in terms of both CHT and RL, which is
however due to numerical (not physical) dissipation

• A comparison between SST and realizable k − ε shows that it is more diffusive in the solid region
than the SST k − ω
• Both models displayed similar thermal dissipation downstream in the wall jet region
In general, the realizable k − ε model showed comparatively better agreement particularly in the im-
pingement region where the maximum temperature is found, and therefore will be used for the upcoming
study presented in Section 6 and also the optimization study of the combustion transition cooling.

Energy & Process Technology Page 69

Chapter 6

Heat transfer methodology

comparison

The section addresses the impact of conjugation on the computed results by performing a sensitiv-
ity study comparing the decoupled and CHT approaches. The sensitivity includes changing thermal
conductivity ratio. λ∗ = λs /λf . In addition, the problem of the turbulent offset jet also includes an
investigation of two key turbulence settings, namely the turbulent Prandtl number, P rt and the the
turbulence intensity, Tu(%). The total amount of simulations for the case of the turbulent offset jet
equals 13 cases for each HT methodology. The study was conducted for both coarse and fine wall mesh
setting. For each thermal conductivity ratio (λ∗ = 1 − 50), the turbulent Prandtl number, P rt , and
Turbulent intensity were kept at constant values of 0.9% and 2%, respectively, and vice versa, thermal
conductivity ratio was constant at 1 when varying P rt = 0.4 − 1, and Tu = 2% − 15%. In order to study
the heat transfer characteristics, results are presented in both interface profiles and averaged quantities,
and the effect of each parameter has been discussed.

The reasons for choosing the turbulent Prandtl number can be traced back to the definitions of the
wall treatment functions. There is also substantial relevance in reacting flows as explained in the ap-
pendix, and the applicability of Reynolds analogy in additively-manufactured channels, similarly to the
ones that will be used for the internal cooling of the TP. Hence, the effect that this parameter has on
the conjugation effects is of great interest. Moreover, inlet turbulence intensity provides a simplistic
approach into quantifying the effect of turbulence at the inlet. The transported variables at the flow
boundary are estimated using (for all k − ε models)
3

k= Iv2 (6.1)
2
ρCµ k 2
ε= (6.2)
(µt /µ) µ
The last parameter in the study is, as aforementioned, the conductivity ratio, which is an alternative
approach to using the Biot number. The use of the Biot number presents a somewhat level of vagueness
as explained in the next section.

6.1 Flat plate

In the present work, all computations are performed at the Reynolds number used for the validation
of the code, Re = 21, 418.

An alternative for the conventional Biot number is introduced: the mesh Biot number. The impor-
tance of such a modification is justified for the following reason: in CHT problems, it is known that

70
 Delft University of Technology

the ratio of the thermal conductivities plays a role (as previously established in the literature study,
also with the Brun number). The drawback of this approach however is, first, the use of a reference
temperature for calculating the heat transfer coefficient, and the characteristic lengths of the two media
is not taken into account. Hence, the mesh Biot number reads:

Kf (λf /∆yf )
Bi(∆) = = (6.3)
Ks (λs /b)

with ∆yf being the cell centroid wall-normal distance of the fluid region (FVM).

Table 6.1: Resultant Biot numbers from heat conductivity ratios

λ∗ Bi1 BiCHT Error(%) Bi(∆)

1 4.22 5.65 -34 275.86
5 0.84 1.11 -32 55.17
10 0.42 0.54 -28 27.59
20 0.32 0.26 -24 13.79
50 0.08 0.10 -20 5.52

The error tends to decrease as expected with increasing solid conductivity, which is traceable back
to the requirement of z ≪ 1. Figure 6.1 provides the solution for the interface temperature character-
ized by the variation of the conductivity ratio λ∗ . The decoupled flow solution, where an isothermal
wall condition was used is also plotted for reference with a broken line. The fluid film temperature
(Tf = T∞2+Tb ) was used as a reference temperature for the wall. Hence, one decoupled flow simulation
is used to subsequently map the heat transfer coefficient on the FEM solid solver.

The profile of the heat transfer coefficient achieved from a decoupled simulation (purely isothermal
CFD) closely replicates the one that is also visible from the boundary layer approach from Luikov (see
Figure 5.3), in which steep gradients are present in the leading edge and the profile tends to remain flat
downstream. It should be noted that this profile is indeed unique for the fluid studied, and is hence
dependent on the conductivity, Prandtl and Reynolds numbers, none of which are altered.
1000
*
= 1, CHT
950 *
= 1, Decoupled
= 5, CHT
900 *
= 5, Decoupled
*
850 = 10, CHT
*

800 = 10, Decoupled

*
= 20, CHT
*
750
= 20, Decoupled
*
700 = 50, CHT
*
= 50, Decoupled
*
650

600
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 6.1: Wall interface temperature (solid side) as a function of all conductivity ratios consider.

It is clear that the difference between the two methodologies increases as the conductivity of the solid
increases, which is in accordance to the theory of the lumped-capacitance analysis. The mean error along
the interface is then 0.4%, 1.1%, 1.8%, 2.4% for λ∗ = 50 − 5. For λ∗ = 1, however the mean error is
again reduced to 1.3%. The results comprise of surprisingly well predicted decoupled temperatures even
1 The value is calculated using the assumption that the average heat transfer coefficient stays constant (equation 5.23).

Energy & Process Technology Page 71

 Delft University of Technology

for large values of the Biot number well ≫ 0.1. On the other hand, the decrease in the difference for
low thermal conductivities (i.e. Bi ≫ 1) is due to the fact that the wall temperature gradient decreases
and local values approximate the value of the free stream temperature. This simplifies the boundary
condition for the problem so that the usage of a constant temperature on the wall in a decoupled simu-
lation (isothermal wall) is justifiable. The mesh Biot number underlines this observation: all values of
Bi(∆ ) ≫ 1.

The velocity profiles are presented in Figure 6.2. The difference tends to increase downstream for larger
conductivities, but remains negligible. This can be more clearly seen when the nondimensionalized ve-
locity boundary layer thickness ∆u is plotted, as presented in Figure 6.3. The two boundary layers,
temperature and velocity, δ and δt are nondimensionalized as:
δp
Re∞ , ∆u = (6.4)
L
δt p
∆t = Re∞ (6.5)
L
where δ and δt are numerically estimated from the computation using the 99% rule (location where 99%
of the respective field quantity value is reached). To reduce the potential for user error, the estimation
was performed using a Java macro within STAR-CCM+. The limitation however of this method comes
with the refinement of the mesh in y-direction, as the exact location where the criterion is met is rather
substituted with the last cell centroid location where ϕ ≤ 0.99ϕ. The nondimensional thermal boundary
layer is also shown in Figure 6.3, along with the value obtained from the Boundary Layer approach from
Luikov, see Eq.(5.16) and (5.15).

1.05 1.05

1 1

0.95 0.95
0.2 0.3 0.4 0.5 0.6 0.2 0.3 0.4 0.5 0.6

1.05 1.05

Decoupled
1 * 1
=1
*
=5
*
= 10
*
= 20
*
= 50
0.95 0.95
0.3 0.4 0.5 0.6 0.7 0.3 0.4 0.5 0.6 0.7

Figure 6.2: Velocity profile comparison for all thermal conductivity ratios at four different axial locations.

If one focuses solely on surface temperatures, then indeed, the two methodologies yield similar results.
However, when it comes to the convective boundary layer, the effect of conjugation clearly becomes more
prominent. Thermal conductivity of the solid begins to have a similar effect on the fluid behaviour as the
Prandtl number. The thermal boundary layer decreases with decreasing conductivity and diverges from
a CFD computation with an isothermal wall BC, partly tracable to the z criterion mentioned earlier.

Energy & Process Technology Page 72

 Delft University of Technology

The plot of the value of z (Eq.(5.14)) is shown below. The criterion of z ≪ 1 is matched for all settings
except λ∗ = 1 and thus explains the increasing difference with the boundary layer obtained from the
analytical solution with decreasing solid conductivity (z ∝ (λ∗ )−1 ). Hence, the conductivity ratio does
in fact provide some insight into the applicability of the decoupled method.

The effect is more clearly depicted as the ratio of the two boundary layers (ζ), shown in Figure 6.6.
Theoretically, the ratio is a function of the Prandtl number (see eq.(5.15)), as outlined in the above
paragraph. According to the Luikov BL method, it equals 1.12 for Pr=0.66. The value tends to be
assymptotically reached with increasing λ∗ . This however is purely due to the assumption taken for z
necessary for deriving the Luikov relationship, which does not necessarily hold a physical basis, but is
rather purely empirical.

5 5
* *
=1 =1
* *
=5 =5
4 * 4 *
= 10 = 10
* *
= 20 = 20
* *
= 50 = 50
3 3
CFD Luikov BLE

2 2

1 1

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

(a) (b)

Figure 6.3: Nondimensional (a) velocity and (b) thermal boundary layers for all cases.

5
*
=1
*
=5
*
4 = 10
*
= 20
*
= 50
*
3 Luikov BLE ( =1)

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 6.4: Critical z criterion for all cases

Energy & Process Technology Page 73

 Delft University of Technology

1.2

0.46
1.1

0.45
1

0.44
0.9

0.43
0.8

0.42
0.7

0.41
0.6
5 10 15 20 25 30 35 40 45 50

5 10 15 20 25 30 35 40 45 50

Figure 6.6: Effect of thermal conductivity ra-

Figure 6.5: Averaged Nusselt number, derived tio on boundary layer ratio (shape-preserving
from Eq.5.23 (shape-preserving interpolant) interpolant)

As demonstrated by the study, the resultant �usselt number is larger in cases where the heat flux to
the plate is accounted for (CHT). The increase is explainable by the decreasing surface temperature
of the wall T (x, 0) = Tw along x, from ∞ to T. The temperature driving force (i.e. temperature
differential between free-stream value and interface ∆T = T∞ − Tb ) is hence increasing with x. From
basic, theory of heat transfer it is known that this differential increases in the direction in the flow,
and thus the heat transfer coefficient (presented in the form of Nusselt numbers) also increases. For
the case of isothermal decoupled simulations where the heat flux into the plate is not accounted for,
∆T = const. One drawback of the approach in the study, however, is the critical temperature difference,
∆Tcr = T∞ − Tb = 400K = 0.4 · T∞ . This can also partially be used to explain the small difference
between the methods even for 0.1 << 1. Figure 6.7 presents the interface temperatures at a lowered
bottom surface temperature, Tb = 300K (∆Tcr = 700K), which clearly indicates that accuracy of the
decoupled approach for a laminar flow case is dependent not solely on the thermal conductivity ratio but
also on the critical temperature difference of the problem. A summary of the data is presented in Table
6.2.

Energy & Process Technology Page 74

 Delft University of Technology

1000 1000

950 900

800
900
CHT, Tb = 300K 700

850 CFD, Tb = 300K

CHT, Tb = 600K 600
CFD, Tb = 600K
800 500
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

(a) (b)
900 900

800 800

700
700
600
600
500

500
400

400 300
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

(c) (d)
800

700

600

500

400

300
0 0.2 0.4 0.6 0.8 1

(e)

Figure 6.7: Comparison of interface temperatures using a decreased bottom plate surface temperature,
for all conductivities (a) - (e).

Energy & Process Technology Page 75

 Delft University of Technology

Table 6.2: Summary of average heat transfer coefficient, and boundary layers comparison for all conduc-
tivities at Tb = 600K.

Average HTC, h ∆u (x = L) ∆t (x = L)
λ∗ CFD CHT Error (%) CFD CHT Error (%) CFD CHT Error (%)
1 32.58 38.44 -17.98 3.58 3.59 -0.28 4.05 2.93 27.65
5 32.58 37.85 -16.15 3.58 3.59 -0.28 4.05 3.73 7.90
10 32.58 36.88 -13.21 3.58 3.38 5.59 4.05 4.08 -0.74
20 32.58 35.76 -9.76 3.58 3.39 5.31 4.05 4.09 -0.99
50 32.58 34.52 -5.97 3.58 3.39 5.31 4.05 4.29 -5.93

Table 6.3: Error comparison for average interface temperature for baseline and increased critical tem-
perature difference (T∞ − Tb )

Average temperature, T w
∗
λ CHT CFD Error (%) ErrorTb =300K (%)
1 925.62 914.59 -1.21 -0.66
5 794.20 776.39 -2.30 -6.27
10 728.89 716.73 -1.70 -6.14
20 677.11 670.92 -0.92 -4.77
50 635.13 633.21 -0.30 -0.48

Investigating however only the interface temperature distribution is a rather simplistic approach to
comparing the two methodologies. The thickness of the thermal boundary layer showed that, the con-
ductivity ratio does in fact have influence, and it is imperative that CHT is used for the conductivity
envelop studied.

Energy & Process Technology Page 76

 Delft University of Technology

6.2 Turbulent offset jet

The common procedure followed in the case of turbulent flows is to employ an adiabatic wall boundary
condition for a decoupled simulation, extract the specified-y + heat transfer coefficient and map it onto
the FEM solid only solver (n.b., an isothermal wall in the laminar plate case).

40
Viscous sublayer
Buffer layer
Log layer
30 M2, u = 8m/s
jet
M6, u jet = 8m/s
M2, u jet = 20m/s
20
M6, u jet = 20m/s

10

0
0 5 10 15 20 25 30 35 40

Figure 6.8: Comparison of resultant wall y + profiles from the validation case and the methodology
comparison with increased jet velocity. The value of the wall y + is still ≤ 10 for the blended WF
approach, which is beneficial for the consistency of the study.

Another important adjustment is in the increase in the jet nozzle velocity to 20m/s. This was done
with the sole purpose of addressing the uncertainty of using the high-y + approach without the criterion
of y + > 30 being satisfied, and ensuring that the full range of the blended approach is tested (i.e. the
first cell is in the viscous sublayer at y + < 10 An example of the resultant first-cell y + values in the two
cases is shown Fig.6.8 above.

Prior to investigating the effects of the aforementioned paramters in the sensitivity study, it is use-
ful to compute the resultant difference in the validation study using the experimental data from Pelfrey,
should the decoupled approach was used instead.

The resulted temperature profiles on the wall are shown in Fig.6.9 and Fig.6.10. A comparatively low
difference is found between the two methodologies. For both, the temperature using the CHT approach
is higher than the decoupled approach in the recirculation and impingement regions. The opposite holds
for the blended WF approach in the wall jet region. The difference for both WF, however, is marginal. It
could be argued that it is due to the comparatively low thermal conductivity material of the solid plate
in the experiment (λ∗ = 1.04). From the previous laminar flow sensitivity study, it has been found that
conjugation plays no role in the HT process due to the low temperature gradients in the flow direction.
The same can also be assumed for the case of a turbulent flow. A more detailed discussion is made in
the upcoming sections where the previous λ∗ envelop is also applied for this problem.

Energy & Process Technology Page 77

 Delft University of Technology

1
CHT M2, Two-Layer All-y+
0.9 Decoupled M2, Two-Layer All-y+
CHT M6, High y+
0.8 Decoupled M6, High y+
Holland & Liburdy (1990)
0.7

0.6

0.5

0.4

0.3
0 5 10 15 20 25 30 35 40

Figure 6.9: Local wall interface temperature (solid side) as a function of P rt .

40

30

20

10

0
0 5 10 15 20 25 30 35 40

Figure 6.10: Local wall interface temperature (solid side) as a function of P rt .

Energy & Process Technology Page 78

 Delft University of Technology

6.2.1 Turbulent Prandtl number

The resultant temperature and specified y + heat transfer coefficient for the two-layer all-y + wall
function are plotted in Fig. 6.11 and Fig. 6.12 below. As expected and noticed in literature, interface
temperature tends to decrease with decreasing Turbulent Prandtl number (i.e increasing thermal eddy
diffusivity). Similarly to the laminar flat plate case, the temperature is higher along the entirety of the
plate length when using CHT approach. The divergene between the two methods is more pronounced in
the turbulent wall jet region (i.e. X > 20). With regards to heat transfer coefficients, the difference is less
visible. The values for both methods tend to remain identical for all regions, except for the impingement.
The decoupled approach slightly overpredicts the HTC and also results in a reattachment length which
is slightly lower than the one obtained from the conjugate method.

1
Prt = 0.4, CHT

0.9 Prt = 0.6, CHT

Prt = 0.6, CHT
0.8
Prt = 0.9, CHT

0.7 Pr = 1.0, CHT

t
Pr = 0.4, Decoupled
t
0.6
Prt = 0.6, Decoupled

0.5 Prt = 0.8, Decoupled

Prt = 0.9, Decoupled
0.4 Pr = 1.0, Decoupled
0 10 20 30 40 50 60 t

Figure 6.11: Local wall interface temperature (solid side) as a function of P rt .

120

100

80

60

40

20

0
0 10 20 30 40 50 60

Figure 6.12: HTC as a function of P rt .

Analogously, the observations mentioned above also hold for the high-y + approach, shown in Fig.6.13
and Fig 6.14: the temperature difference is higher in the wall jet region, whereas in the impingement
region, it tends to zero. The opposite is found for hy+ and interestingly, the differnce is augmented in
the impingement, in comparison to the blended wall treatment approach.

Energy & Process Technology Page 79

 Delft University of Technology

1
Prt = 0.4, CHT
0.9 Prt = 0.6, CHT
Prt = 0.8, CHT
0.8
Prt = 0.9, CHT
0.7 Prt = 1.0, CHT
Prt = 0.4, Decoupled
0.6
Prt = 0.6, Decoupled

0.5 Prt = 0.8, Decoupled

Prt = 0.9, Decoupled
0.4
0 10 20 30 40 50 60 Prt = 1.0, Decoupled

Figure 6.13: Local wall interface temperature (solid side) as a function of P rt .

120

100

80

60

40

20

0
0 10 20 30 40 50 60

Figure 6.14: HTC as a function of P rt .

In order to compare the conjugate and decoupled approach more broadly, the surface averaged tem-
perature, θw , reattachment point, Xcr , and the surface averaged specified-y + heat transfer coefficient
are shown in Fig. 6.15 across all P rt values.

Looking at the resultant difference between the HT methodologies, a concrete relationship with the
turbulent Prandtl number seems to be missing. The only evident variation is in the averaged surface
temperature. The absolute value however for all settings remains negligible and cannot therefore be
attributed solely to the effect of conjugation.

Energy & Process Technology Page 80

 Delft University of Technology

0.82 17

0.8
16
0.78
15
0.76

0.74 14

0.72 13

0.7
12
0.68
11
0.66

0.64 10
0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1

(a) (b)

80
Two-layer WF, CHT
Two-layer WF, Decoupled
75 High-y+ WF, CHT
High y+ WF, Decoupled

70

65

60

55

50

45
0.4 0.5 0.6 0.7 0.8 0.9 1

(c)

Figure 6.15: Influence on surface averaged values of (a) temperature, (b) reattachment length, (c)
specified-y + heat transfer coefficient

Energy & Process Technology Page 81

 Delft University of Technology

6.2.2 Turbulence intensity

Similar approach in the comparison is followed for the effect of turbulence intensity on conjugation
as well. As shown in Fig.6.16 and Fig.6.17, no clear influence on temperature and HTC, respectively�
is present. The maximum absolute difference percentagewise for θw and hy+ equals 18%(X = 60) and
66%(X = 11.5) at Tu=2% and Tu=15%, respectively for the Blended WF. Similarly, for the high-y +
WF treatment the values are 19.3%(X = 34) and 29.7%(X = 9.8) both at Tu=15%.

The small difference in the interface scalars is due to the inherent quick dissipation of the turbulence
intensity characteristic for such a flow problem. As it can be seen in Fig. 6.18 and Fig.6.19, where,
although not available for post-processing directly in a RANS model, one can trace back to turbulence
intensity via the turbulent kinetic energy (see eq. 6.1). The value of the TKE is little influenced par-
ticularly in the vicinity of the wall, and hence the heat transfer process is not affected. The value is
maximum in the impingement region, and when high turbulence intensity is considered (Tu= 15%), near
the jet exit region where the jet interacts with stationary fluid. Turbulene intensity tends to remain
constant in both inner and outer shearing layers for all cases.

On the other hand, if one considers the case of the high-y + treatment, the resultant distribution of
turbulent kinetic energy differs substantially. This however is not due to the wall treatment but rather
the numerical dissipation from the needed coarse mesh settings, pinpointing to one of the drawbacks
of this methodology,as shown earlier in the chapter. The resultant effect of Tu(%) is more pronounced
throughout the entirety of the jet near the nozzle and not exclusively at the shear layers. The turbulent
kinetic energy is also higher at the reattachment point compared to the case of low turbulent intensity.

1
Tu = 2%, CHT
Tu = 4%, CHT
0.9
Tu = 8%, CHT
Tu = 10%, CHT
0.8
Tu = 15%, CHT
Tu = 2%, Decoupled
0.7 Tu = 4%, Decoupled
Tu = 8%, Decoupled
0.6 Tu = 10%, Decoupled
Tu = 15%, Decoupled
0.5

0.4
0 10 20 30 40 50 60

Figure 6.16: Local wall interface temperature (solid side) as a function of Tu(%).

120

100

80

60

40

20

0
0 10 20 30 40 50 60

Figure 6.17: HTC as a function of Tu(%).

Energy & Process Technology Page 82

 Delft University of Technology

8 15 8 15

Turbulent Kinetic Energy (J/kg)

Turbulent Kinetic Energy (J/kg)

6 6
10 10

4 4

5 5
2 2

0 0 0 0
0 2 4 6 8 10 12 14 16 18 0 2 4 6 8 10 12 14 16 18

(a) (b)

Figure 6.18: Turbulent kinetic energy for (a) Tu = 2%, (b) Tu = 15% using a fine mesh (M2) and the
blended wall function approach.

8 8

6 6

4 4

2 2

0 0
0 2 4 6 8 10 12 14 16 18 0 2 4 6 8 10 12 14 16 18

(a) (b)

Figure 6.19: Turbulent kinetic energy for (a) Tu = 2%, (b) Tu = 15% using a coarse mesh (M6) and the
high-y + wall function.

1
Tu = 2%, CHT
Tu = 4%, CHT
0.9
Tu = 8%, CHT
Tu = 10%, CHT
0.8
Tu = 15%, CHT
Tu = 2%, Decoupled
0.7
Tu = 4% , Decoupled
Tu = 8%, Decoupled
0.6
Tu = 10%, Decoupled
Tu = 15%, Decoupled
0.5

0.4
0 10 20 30 40 50 60

0.8

0.75

0.7

16 18 20 22 24

Figure 6.20: Local wall interface temperature (solid side) as a function of Tu(%).

Thereby, the resultant effect of turbulence intensity is more pronounced when a coarse mesh is used,
from which the following observation is found: the difference between the heat transfer methodologies

Energy & Process Technology Page 83

 Delft University of Technology

tends to increase with increasing turbulence intensity. This more visible in the resultant average sur-
face temperatures, where a clear divergence is found between the solid and broken red lines in Figure
6.22a. The same hold also for the reattachment length, indicating that conjugation could play a role on
the aerodynamic characteristics of the flow should a coarse mesh be used. No clear trend of the heat
transfer coefficient can be established: the value tends to fluctuate between ≈53 and ≈50 with standard
deviations of 0.763 and 1.188 for the decoupled and conjugate approach, respectively.

120

100

80

60

40

20

0
0 10 20 30 40 50 60

Figure 6.21: HTC as a function of Tu(%).

0.82
17
0.8

16
0.78

0.76 15

0.74 14

0.72 13

0.7
12

0.68
11
0.66
10
0.64 2 4 6 8 10 12 14
2 4 6 8 10 12 14 Tu(%)
Tu(%)
(b)
(a)

53
Two-layer WF, CHT
Two-layer WF, Decoupled
High-y+ WF, CHT
52
High y+ WF, Decoupled

51

50

49

48

47
2 4 6 8 10 12 14
Tu(%)
(c)

Figure 6.22: Influence on surface averaged values of (a) temperature, (b) reattachment length, (c)
specified-y + heat transfer coefficient

Energy & Process Technology Page 84

 Delft University of Technology

Overall, the sensitivity study on turbulence intensity showed little to no influence on the effect of
conjugation particularly for a blended WF approach. The influence was more prominent when a coarse
mesh was used, where the the two HT methodologies yielded different results with increasing Tu(%).
According to author’s assumption however, this is also partially due to the effects of a coarse mesh setting
on computed flow parameters.

6.2.3 Thermal conductivity ratio

The results from the thermal conductivity ratio sensitivity study to an extend contradict the estab-
lished expectations from the laminar flat plate sensitivity study. In the latter it was established that
increasing the solid thermal conductivity to values of ≪ 0.1, lessens the effect of conjugation. This
was established not to be the case when considering turbulent conjugate heat transfer as shown in the
interface temperature at the wall in Fig.6.23 below. It is also worth emphasising on the difference in the
decoupled approach used in the two cases: an isothermal wall for the laminar plate and an adiabatic wall
in the case of the turbulent offset jet.

The difference between the two methodologies is mostly present in the impingement region and tends to
decrease in the wall jet region. The difference in the impingement seems to be dependent on the thermal
conductivity ratio with the tendency to increase at the peak values (X = Xcr ), whereas the wall jet
region values are irrespective of the solid conductivity.

1
*
= 1, CHT

*
= 5, CHT
0.8
*
= 10, CHT

*
= 20, CHT
0.6
*
= 50, CHT

*
= 1, Decoupled
0.4
*
= 5 , Decoupled

*
= 10, Decoupled
0.2
*
= 20, Decoupled

*
= 50, Decoupled
0
0 10 20 30 40 50 60

Figure 6.23: Local wall interface temperature (solid side) as a function of λ∗ .

140

120

100

80

60

40

20

0
0 10 20 30 40 50 60

Figure 6.24: HTC as a function of λ∗ .

The difference is diminished and close to virtually zero when a coarse mesh is used, as shown in
Fig.6.25. The influence of increasing thermal conductivity is also negligible when considering the heat
transfer coefficients, presented in Fig.6.24. The difference in the wall temperature is again the highest

Energy & Process Technology Page 85

 Delft University of Technology

near the impingement region and this time roughly remains constant. Again, these findings can be at-
tributed to the high numerical dissipation of the coarse mesh

For both mesh settings the difference between the two HT methodologies can be explained in an identical
fashion. In both cases, blended and high-y + , an identical heat transfer coefficient is exported and mapped
onto the FEM solver. Thereby, the temperature profiles for all decoupled computations are inherently
similar. The effect of the heat transfer coefficient augmentation for example, obtained in the blended
approach (see Fig.6.24) is not transferred onto the solid region, since the heat transfer coefficient that is
exported (no conjugation - broken orange line) is comparatively ’flat’ in the vicinity of the impingement,
which is in turn also visible in the wall temperature profile. The same applies to the coarse mesh, with
the exception of little influence on the heat transfer coefficient with increasing solid conductivity.

1
*
= 1, CHT

*
= 5, CHT
0.8
*
= 10, CHT

*
= 20, CHT
0.6 = 50, CHT
*

*
= 1, Decoupled
0.4 = 5 , Decoupled
*

*
= 10, Decoupled
0.2 = 20, Decoupled
*

*
= 50, Decoupled
0
0 10 20 30 40 50 60

Figure 6.25: Local wall interface temperature (solid side) as a function of λ∗ .

120

100

80

60

40

20

0
0 10 20 30 40 50 60

Figure 6.26: HTC as a function of λ∗ .

Another important finding is that, in fact, the CHT approach results in a higher temperature profile
when a fine mesh is used, whereas the opposite holds for a coarse mesh setting: the decoupled HT
approach results in higher temperatures. The difference is as mentioned more pronounced for the fine
mesh settings, but in both cases it remains negligibly small. This is also visible in the surface average
profiles in Fig.6.27a. The averaged values of the heat transfer coefficient scalar and the reattachment
length also reflects the abovementioned analysis.

Energy & Process Technology Page 86

 Delft University of Technology

0.9 17

0.8 16

0.7
15
0.6
14
0.5
13
0.4
12
0.3

0.2 11

0.1 10
5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50

(a) (b)

Two-layer WF, CHT

60 Two-layer WF, Decoupled
High-y+ WF, CHT
58 High y+ WF, Decoupled

56

54

52

50

48

5 10 15 20 25 30 35 40 45

(c)

Figure 6.27: Influence on surface averaged values of (a) temperature, (b) reattachment length, (c)
specified-y + heat transfer coefficient

Energy & Process Technology Page 87

 Delft University of Technology

6.3 Summary
The chapter included a comprehensive validation procedure and analysis of both laminar and turbulent
flow conjugate heat transfer methodologies, that ought to be followed in the upcoming part, where an
optimization is performed, followed by a comprehensive sensitivity study on both problems.

The laminar flow CHT solver in STAR-CCM+ was successfully validated and yielded results comparable
to an analytical solution within an acceptable accuracy for two Biot number settings. The difference
was discussed and concluded that it is due to the assumption used in the laminar solutions (incl. the
velocity from Blasius and temperature profiles from Luikov).

The validation study of the turbulent CHT problem showed that the realizable k − ε model resulted
in the most accurate solution when compared to experimental data, explainable by the different for-
mulation of the eddy viscosity term. The temperature profile for all turbulence models studied was
otherwise within an 8% difference from the experimental data (SST k − ω and standard formulation
of the k − ε models). This difference was shown to decrease to less than ≈ 5% when a coarser mesh
was used with a high-y + wall treatment, an effect attributable to the numerical dissipation from the mesh.

Moreover, the sensitivity study addressed two decoupled approaches: isothermal wall and adiabatic
walls for the laminar and turbulent flow problem, respectively. Contrary to the popular guideline for
conjugation use for Bi ≪ 0.1 and Bi ≫ 1, the study showed the error with the decoupled approach is in
fact negligible for the entire envelope, particularly for λ∗ (λs /λf ) > 1, i.e. a solid region with a thermal
conductivity considerably larger than that of the fluid region. The study pointed the importance of the
critical temperature difference (T∞ − Tb ) and proved that it is another important parameter that has
to be taken into account when choosing a HT methodology. With respect to the turbulent offset jet,
no firm dependency of the conjugation effect was found with respect to Turbulent Prandtl number and
turbulence intensity. The only visible apparent increase in the necessity for CHT was established at
increasing solid conductivities, contrary to the findings of the laminar flat plate, where increasing λ∗
yielded to essentially no difference. The difference was shown to be pronounced solely in the impinge-
ment regions of the jet. The difference in the wall jet and recirculation regions were also negligible.

The specific findings of the validation study are carried over to the optimization case of the transi-
tion piece cooling. In it, the realizable k − ε model is used along with a blended WF approach. Due to
the presence of internal cooling inside the GT component, the analysis employs the CHT approach, in or-
der to take the heat pick-up inside the cooling channels into account. The use of the decoupled approach
for internal cooling is hence strictly not recommended as it will result in grossly overestimated cooling
performance (channel wall temperature being higher than in reality). As discussed in the upcoming
chapter, the reasons for using the blended WF approach is pointed out too.

Energy & Process Technology Page 88

 Part II

GT Transition Piece

89
Chapter 7

Gas Turbine Transition Piece

Analysis

As initially mentioned in the introductory chapter of the this report, the study does not include the
full-scale transition. The reasons are explained into further detail in the upcoming sections, but they
can be overall summarized as a hedge against high computational costs of optimization studies of large
components. The drawback of this approach is mentioned as well. The chapter is comprised of a thorough
description of the computational domains used, description of the rib parameters, followed by a brief
comparison of the baseline flow fields of the baseline cases and a discussion on the findings.

7.1 Computational domain

Due to the size of the full transition piece, using a 1:1 scale model for the computational domain
would make the optimization task infeasible from a computational cost point of view. Therefore, the
component was simplified into two separate parts: top and bottom panels, as shown in Fig. 7.1. The
drawback of such approach its inability to predict the lifing enhancement of an actual transition piece.
However, it allows for easy parametrization and an insight into potential cooling improvement that can
be re-applied into a full-scale transition model, depending on the findings.

Figure 7.1: Illustration of approach used for discretizing the transition piece geometry into a parametric
slice (shown in red), consisting of two counter-current channels. Original unedited (author’s TP is
replaced with the Siemens 8000H TP) picture is obtained from [93]. The figure has an illustrative
purpose only, and components are not to scale.

90
 Delft University of Technology

The shape of the cooling channels and the shape of the panel are fully parametrized (with parameters
shown in the Appendix) allowing for both different channel designs and adaptable to accommodate the
entire range of transition shapes as well as TBC thickness. Each panel was made using a spline CAD
feature and consists of no curvature in the tangential direction (z-coordinate in the coordinate frame
shown in Figures 7.2-7.3 below). The key dimensions are also shown in the domain schematics.

As visible in Figures 7.2-7.3, the cooling consists of a pair of counter-current channels with a circular
cross section. The channels extract cool dump diffuser (i.e. compressor shell/plenum) air of tempera-
ture Tdif f from top and bottom parts of each transition panel and dump the coolant air in the hot gas
mixture from the combustion process, with a temperature of Tg , Fig. 1.4 presents a simplified overall
arrangement similar to the one used in the study. For confidentiality reasons, temperature and pressure
values are respectively nondimensionalised using:
T − Tdiff
θ= , (7.1)
Tg − Tdiff
p
Π= , (7.2)
pcomp
<u>
U∗ = , (7.3)
u∞
where pcomp is the dump diffuser pressure, and u∞ is the free-stream velocity from the combustor liner
exit, as displayed in Fig. 7.2. The computational domain has a constant thickness in all regions of 6mm,
with the baseline case (ciruclar channels) using a channel pitch distance of 3mm (equidistant from the
domain side walls). An additional extrusion is included at the end of the domain for convergence reasons.

Velocity inlet
• U* = 1.0 Stagnation inlet
• 𝜃 = 1.0 • 𝜃 = 0.0
• Tu = 0.01 • Π𝑡𝑜𝑡 = 0.07
• 𝜇𝑡 /𝜇 = 10.0
0.145m

Channel CHT interface

• Roughness H = 0.115mm
• Resistance = 3 ⋅ 10−4 𝑚2 𝐾/𝑊

BC-Transition CHT interface

{ (solid-solid)
TBC-BC CHT interface
(solid-solid)
• Resistance = 0.0 𝑚2 𝐾/𝑊

TBC Shell region

• Mapped HTC and 𝑇𝑎𝑚𝑏

Outlet
• Split ratio = 1.0
Pressure Outlet
• 𝜃 = 0.96
• Π = 0.0
• Tu = 0.01
• 𝜇𝑡 /𝜇 = 10.0
0.160m

Figure 7.2: Top panel schematic and boundary conditions used, along with a close-up view of the
channels, where relevant interfaces are illustrated using dashed lines. Red and blue are used to indicate
fluid region (hot and cold flows, respectively), whereas gray is the transition metal region. Channel flow
is shown in purple.

Energy & Process Technology Page 91

 Delft University of Technology

Velocity inlet
• U* = 1.0
• 𝜃 = 1.0
• Tu = 0.01
• 𝜇𝑡 /𝜇 = 10.0

Outlet Pressure Outlet

• Split ratio = 1.0 • 𝜃 = 0.96
• Π = 0.0
• Tu = 0.01
• 𝜇𝑡 /𝜇 = 10.0

Stagnation inlet
• 𝜃 = 0.0
• Π𝑡𝑜𝑡 = 0.07

0.356m
Figure 7.3: Bottom panel schematic and boundary conditions used.

In the numerical set-up used, only the channel wetted area is a conjugate (fluid-solid) interface (red
broken line in Fig. 7.2). The solid-solid interface between the TBC and BC is also a CHT interface, and
since no thermal resistance is added to the BC-TP interface, the BC essentially has no effect on the HT
performance. The figure below presents a simplified schematic of the types of BCs on the TP walls: it
can be said that the analysis hence uses a combined approach: decoupled HT (mapped hbulk and Tamb )
on the walls, and a CHT on the walls of the cooling channels.

Combustor side: known Temperature

HTChot: based upon previous full-scale simulation

TBC/bond coating/metal

Cooling channel CHT

Metal

Cooling side: known Temperature

HTCcold: based upon previous full-scale simulation

Figure 7.4: Types of HTC BCs and interfaces in the transition piece set-up.

The outlet BC at the start of the additional extension downstream in the hot region arises from the
fluid-fluid interface with the original one. Using an outlet BC allows the outflow conditions to not be
prescribed, but rather determined by the flow upstream of the outlet boundary. The split ratio of 1
indicates the fraction of the total outflow leaving at the given outlet boundary (i.e. entire flow) before
eventually reaching the pressure outlet.

The domain side walls for both cases are modelled using a translational periodicity boundary condi-
tion. This also applies to the area of the TBC shell region and transition piece at the end of the domain.
The hot and cold sides are mapped from a full GT mid-frame system model, allowing different regions of
the transition to be evaluated without extensive turning of the boundary conditions, i.e. only the values

Energy & Process Technology Page 92

 Delft University of Technology

of the heat transfer coefficients. The heat transfer coefficient is mapped at three locations along the hot
and cold sides and interpolated using a linear function. The distribution for both top and bottom panel
is shown in the Fig. 7.14 below.

5000 5000

4000 4000 Cold side

Hot side
3000 3000

2000 2000

1000 1000
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Figure 7.5: Mapped HTC on hot and cold sides of the top and bottom panels, extracted from a full
midframe simulation.

The origin and reasons for using such HTC values for extrapolating a convective BC is beyond the
scope of the project. Therefore, the values are not discussed and directly used in the current analysis.
The potential impact of the resultant distribution, however, on the cooling performance is touched upon
in the upcoming subsections. Due to lack of proprietary data on temperature profiles of the transition
piece, the temperature profiles are not validated. The flow inside the circular cooling channels is com-
pared to experimental data with a similar Reynolds number found in literature as a potential compromise
for a validation of the resultant temperature fields.

The table below examines the CFD-predicted effective area of the transition piece and compares it
to previously experimentally measured values using a test rig at Pennsylvania State University (PSU).
The bottom and top panel estimates are obtained by simply taking the submodel effective area and
multiplying by total channels for that panel.

Table 7.1: Non-dimensionalized transition channels effective area (Aef f /Ach ): validation of pressure drop
calculation versus test measurements.

Bottom Panel 69.5

Top Panel 84.2
Total 153.7
Measured (experimental) 152.5

The channel effective area was approximated using the Bernoulli principle, i.e.:
Nch ṁc
Aef f = p (7.4)
2∆p/ρc
with ∆p, the pressure drop across the transition, calculated using the explicitly calculated values at the
stagnation inlet and transition outlet (not the specified BCs values), and Nch is the number of channels
in the respective panel.

The assumptions made for building the computational model can be summarized as follows:
• Fluid flow is modelled using the 3D steady-state RANS equations;
• The model used is the realizable formulation of the k − ε model, in accordance with the CHT
validation studies presented earlier;

Energy & Process Technology Page 93

 Delft University of Technology

• The turbulence model employs the isotropic turbulence assumption for calculating the eddy vis-
cosity;
• A blended-wall y + approach is used for near-wall calculation of velocity, temperature and turbulent
flow field quantities;
• The model does not take radiation heat losses into account;
• The working fluid in the transition is considered as an ideal gas.
Additionally, an important distinction with the previously shown methodology has to be made:

A comparison between a FEM-FVM interpolation and a FVM-FVM conformal mesh interpolation

showed that there is no reduction in the accuracy of the results. There was, however, a small
reduction in the computational speed of a fully-converged simulation, which compounded into
hundreds of runs in an optimization task would potentially lead to a faster solution. Therefore,
for the sake of the computational cost of this study, a FVM mesh is used for the calculation
of the temperatures in the solid region.

Table 7.2: Thermal conductivities used in the model.

Region Material Function Value (W/mK)

TBC Proprietary YSZ compound Custom f (T ) Confidential
Transition metal Ni-based super alloy Constant 25
Coolant Air Sutherland’s law 0.026 (Tref = 20◦ C)

With regards to solver settings, a segregated fluid enthalpy with second order upwind convection
differencing scheme are used for the fluid region. As mentioned earlier, the Turbulent Prandtl number
is set according to the field function in the validation study: value is set at 0.9 near the wall, and 0.4
in the free-stream. With regards to the thermal conductivities used, for all three regions, the value is
tabulated above in Table 7.2. The proprietary field function for the TBC is omitted to avoid a conflict
of interest.

7.2 Reynolds Analogy

Further addressing the (rough) channel effective area, attention has to be paid to the addition of the
thermal resistance at the channel CHT interface, shown in red in Fig. 7.2. The results of previous
flow tests conducted Pennsylvania State University [112] indicate a relative roughness value of 0.31
for horizontally-printed cylindrical holes. Considering the measured hydraulic diameter of the cooling
channels of the TP, this results in a equivalent sand grain roughness of 175µm. The numerical relationship
between the equivalent (CFD) roughness height and the roughness function applied in the respective wall
treatment method is briefly explained in Appendix E.3. Normalizing both friction factor and Nusselt
number to smooth-wall values, it becomes apparent that friction factor augmentation is significantly
higher than the Nusselt number augmentation, i.e. the penalty for added roughness is 2x higher than
the heat transfer benefit in AM-channels. Therefore, the addition of a thermal resistance is used as a
compromise in the heat transfer enhancement from the metal to the coolant, while still accounting for
the penalty in the friction coefficient. The addition of thermal resistance at the interface hence leads to
a reduction in the heat flux to the coolant, which results in an artificially reduced cooling effectiveness
(i.e. higher temperatures on the channel walls). This can also be regarded as an adjustment for the
non-applicability of the Reynolds analogy, which states:
h Cf
St = = (7.5)
ρucp 2

Energy & Process Technology Page 94

 Delft University of Technology

with St, called the Stanton number. The non-proportionality between Nusselt number augmentation
and skin friction increase in rough channels has been widely reported and literature, whereas the above-
mentioned flow tests were validated against the data presented by Snyder [112]. The reason for the
disproportionality is well-put by Bons [14] where he assess the applicability of the Reynolds analogy in
GT flows and states that surface roughness introduces a large pressure drag component on the individual
roughness elements to the net skin friction measurement with only a modest corresponding increase in
heat transfer, which corresponds to a dramatic decrease of the Reynolds analogy factor (i.e. 2St/Cf ).
The respective increase in both skin friction and Nusselt numbers reported in the paper for various
AM-build directions are shown in Fig. 7.6(a)-(b), while Fig. 7.6 displays the trade-off between the two.

(a) (b) (c)

Figure 7.6: Roughness benchmarking for AM-channels of various build directions: friction factor f ,
Nusselt number (b) as a function of Re,and f and heat transfer augmentation for all tested channels (c)
from Snyder [112].

The validation of this approach, which is beyond the scope of this project, has been performed a priori
in-house at Siemens AG, where the resultant Nusselt numbers and friction factors are compared to the
experimental results from [112] shown in Fig 7.6, using a range of equivalent sand grain roughness and
the k − ω SST model. The exact results, however, are confidential and remain an intellectual property
of Siemens AG, only the final value of the added heat resistance of 3 · 10e-4m2 K/W is reported and used
in this study.

7.3 Parametrization of cooling topology

As mentioned in Section 2.3.1 and pointed out in Fig. 2.12, various rib arrangements have displayed
the best performance in terms of the trade off between friction factor and heat transfer enhancement.
Thus, for the sake of this study, a rectangular rib turbulation geometry will be used. The parametrization
is shown in the figure below.

Figure 7.7: Schematic representation of key parameters used in the rib geometry

Energy & Process Technology Page 95

 Delft University of Technology

Figure 7.8: Resultant flow structure in the rib geometry used (top panel used for illustrative purposes).
The structure closely resembles the flow structure commonly found in literature, see Fig.2.11, with the
exception of a recirculation region on top of the rib structure.

For the sake of this study the height an with of each rib is equal, as oppose to the schematic, i.e.
e1 = e2 = e, and b1 = b2 = b. The pitch of a rib on one side is denoted with p whereas the distance
between two ribs on the opposite sides of the channels is denoted with d, and later referred to as the
periodicity ratio. Hence, four key parameters are defined: rib height-to-channel height ratio, rib pitch-to-
rib height, distance-to-pitch ratio, and the aspect ratio of the rib (height-to-width). The values chosen
for each parameter, respectively, are shown below:

e/H = 0.35 (7.6)

p/e = 40 (7.7)
d/p = 0.5 (7.8)
e/b = 2 (7.9)
The reasoning behind the chosen values, is so that they lie in the middle of the design space exploration
later used for the optimization, as it will be shown. This, however, is not a requirement by the optimizer,
and does not guarantee a better performance. Hence, the procedure followed is to a certain degree arbi-
trary.

The same consideration applies to the width and height of the channel itself too. The full range of
heights and widths considered are presented later in Chapter 6. The values chosen are shown in Table
7.3, along with a comparison to the resultant hydraulic diameter (dh = 4P/A, with P being the channel
circumference) for both cases: baseline circular channel and ribbed rectangular channel.

Table 7.3: Cross section of baseline (circular) and ribbed baseline (rectangular) channels used. All values
are nondimensionalised, as a fraction w.r.t to the respective circular channel value.

Channel aspect ratio (H/w) A∗ch,rec A∗ch,circ d∗h,rec d∗h,circ

0.53 0.78 1.0 0.69 1.0

As it can be found from the height ratio and aspect ratio of each rib, a large proportion of the ribs
that will be generated in the design space for the optimization are beyond the manufacturing tolerance
capabilities of AM. This was done in order to increase the size of the design space and establish whether
the Pareto-optimal solutions perform better than the current design. The task is hence, more theoretical
rather than practical in the sense that concluding whether rib geometries can be adjusted to mitigate
the effect of friction was more important than the derivation of a set of a rib geometry that can be used
straight into manufacturing.

7.3.1 Mesh study

As mentioned above, no mesh independence study was performed for the domain with circular channels,
due to the simpler flow, which was expected to be captured well with the use of the wall function. Hence,

Energy & Process Technology Page 96

 Delft University of Technology

the resolution of the mesh was chosen based solely on the author’s experience, and the expectation that
a finer resolution would have little to no influence on the key parameters incl. average metal (BC)
temperature and coolant mass flow rate. When ribs are added however, the flow features are more
complicated as shown in the Literature study chapter earlier. Therefore, a mesh independence study was
deemed indispensable. Explain briefly settings and how they help with optimization task here

Table 7.4: Mesh settings used for the top panel study. The channel with an inlet upstream is denoted
with 1, and vice versa.

Average y +
Mesh ID Element count TSZ1 PLT2
Channel 1 Channel 2
M1 79.94M 8.09 8.29 0.03 0.075
M2 24.61M 11.06 11.36 0.05 0.1
M3 11.82M 11.29 11.38 0.07 0.1
M4 6.09M 10.92 11.21 0.1 0.1
M5 3.74M 15.7 16.0 0.125 0.15

(a) (b)

Figure 7.9: Comparison of mesh resolution: M1 (a) and M5 (b), channel shown in blue.

Three critical surface average temperatures are used to track the computational cost of each grid,
shown below in Fig. 7.10. Standard surface averaging is used as shown in the equation below:
Z P
1 θ f Af
Surface Average ≡ θf da = P (7.10)
a f Af

where θf is the cell face value of the temperature and Af is the face area. Increasing the domain
element count, particularly in the fluid region has no influence on the averaged temperatures on all three
key surfaces. The deviation for the bond coating, thermal barrier coating, and channel walls is merely
0.174%, 0.09%, and 0.276%, respectively. The velocity profiles inside the channel also did not display a
large difference, contrary to the expectation of numerical dissipation from the turbulent offset jet study.
The profiles are available to the reader in the Appendix.
1 Target surface size (mm)
2 Prism layer thickness (mm)

Energy & Process Technology Page 97

 Delft University of Technology

0.725 0.195

0.1945

0.194

0.72 0.1935

0.193

0.1925

0.715 0.192
0 2 4 6 8 0 2 4 6 8
10 7 10 7
(a) (b)

0.85
Channel 1
Channel 2

0.8

0.75
0 2 4 6 8
10 7
(c)

Figure 7.10: Mesh convergence tracked in surface averaged temperatures: Bond coating (i.e. hot side of
transition metal) (a), thermal barrier coating (b), channel surface (c).

With changing rib parameters, the mesh size would have to be adjusted in the automated mesh
generation process in the optimization. The result of the mesh independence study also provided a
guideline for choosing a wall mesh resolution inside the channels as a function of the size of the rib.
The provides a good compromise for conducting the optimization study, as recirculation regions are the
only flow structure that can be captured with an isotropic RANS approach, and with the low numerical
dissipation, it can be expected that they are well captured for a wide envelop of rib configurations. This
approach, however, lacks physical basis to some extent, since mesh size and a mesh-independent solution
does not guarantee an accurate solution.

Table 7.5: Computational cost comparison between all meshes.

Mesh ID Number of CPUs Wallclock (min)

M1 72 1465
M2 72 452
M3 24 435
M4 48 111
M5 24 129

The computational time as mentioned earlier has also been tracked to accelerate the optimization
process, with the computation wallclock time tabulated above. Thereby, based on cost and results it
can be concluded that mesh M4 is the best option for further analysis and a wall resolution of TSZ =
PLT = 0.7b is chosen for the optimization run. The average wall y + values also help validate the choice
for the blended WF approach.

Energy & Process Technology Page 98

 Delft University of Technology

7.4 Flow validation

As mentioned above, due to the lack of experimental apparatus and data, a rather crude approach was
followed for the validation of the flow field in the two cases. The velocity profile in the outer-layer (buffer
and logarithmic regions) were scaled in accordance to the defect law, shown on the y-axis. The data
was then compared to the Townsend’s (1976) outer flow similarity hypothesis for rough-wall flows first
presented in [46]. The term Ucl is used to denote the centerline velocity, i.e. U (y = Dh /2). The velocity
scaling collapses well, except at the inlet and outlet of the cooling channel, which is precisely after and
before the inlet and outlet rings, shown in Fig.7.3 which can be used to explain the non uniformity in the
velocity profiles. The slight curvature of the panel is an additional difference with a typical pipe flow.

20
x/L = 0.0
x/L = 0.312
x/L = 0.625
15 x/L = 1.0
-(1/0.421)ln(y/R)+1.20

10

-5
0 0.2 0.4 0.6 0.8 1

Figure 7.11: Outer-layer scaling at four locations downstream in one of the cooling channels inside the
bottom panel. Townsend’s similarity equation presented in the legend.

As shown by the mesh independence study, the exact solution of the flow inside the channels has little
impact on the overall cooling performance measurement, which in terms of the cooling performance num-
bers (η, P hi, HLP) is measured using the surface-averaged temperature. Thereby, the small deviations
from the defect law observed in the validation are expected to have a minimal impact the subsequent
results (baseline comparisons and rib optimization).

The procedure followed for the ribbed case is similar, where the velocity profiles are compared to DNS
data from Wang [84]. The author used a blockage ratio of 0.2 and a Reynolds number based on the
average bulk mean velocity over the streamwise direction of the ribbed duct (Ub ), Reb = 5600. The bulk
velocity (also used in the cooling comparison presented subsequently) is calculated as:
R P
ρ⃗v dV ρc⃗v Vc
Ub = R V
= Pc (7.11)
V
ρdV c ρc V c

where subscript c is used to indicate the value over a single cell.

Figure 7.12: Schematic of a square duct with transverse ribs used by Wang [84]. The duct uses a blockage
ratio of 0.2 and pitch-to-height ratio of 4.

Energy & Process Technology Page 99

 Delft University of Technology

4
RANS, x/ = 0.4
3 RANS, x/ = 1.0
RANS, x/ = 1.5
2 DNS (Wang, 2020), x/ = 0.4
DNS (Wang, 2020), x/ = 1.0
1 DNS (Wang, 2020), x/ = 1.5

-1

-2
-1 -0.5 0 0.5 1

Figure 7.13: Comparison to DNS data of nondimensional streamwise velocity at three downstream
locations of the rib.

The axial locations were nondimensionalized using δ = H/2, where x/δ = 0 is at the trailing edge
of the rib. A substantial difference is present with the DNS-solution. The reasons for the difference
discussed above however also hold in this case. An overview of the geometry used by Wang is shown in
Fig. 7.12. Additionally, the disagreement in the two cases in the rib geometry and Reynolds numbers,
particularly the latter (≈ 12, 000 in the ribbed channels of the bottom panel, based on the mean inlet
velocity), supplement the difference and explain why the streamsiwise velocity in considerably higher in
the region above the rib (y/δ > −0.3). The difference in the recirculation region in the lower half of the
channel is negligible.

7.5 Preliminary baseline comparison

Prior to conducting the comparative analysis of the two cooling schemes, multiple important parame-
ters have to be introduced first, which are not present to classical heat exchanger performance curves.

For transition piece heat transfer analyses the Heat Load Parameter (HLP) is an important parame-
ter that has to be adapted first to include technologies that are common in combustion hardware, but
not present in heat exchangers. Such includes TBC, and the presence of heat transfer to the midframe
flow. The heat addition from the hot gas path is equal to the sum of the heat picked up by the cooling
features in the component and the heat transferred form the part to the midframe flow:

qg = qc + qcs (7.12)

Subsequently, a thermal resistance network between the hot gas and the bond coat can be used to account
for the performance of the TBC:
Tg − Tm
qg =   (7.13)
1 tT BC
Ag hg + kT BC Ag

where the heat transfer coefficient on the hot side is hg and kT BC and tT BC are used for the thermal
conductivity and thickness of the thermal barrier coating, and Tm is the transition metal temperature.
Conversely, on the cold side of the transition, the heat transfer rate reads:

qcs = Acs hcs (Tcs − Tdiff ) (7.14)

using a simple heat transfer expression for the heat transfer pick-up inside the channels qc = ṁcp,c (Tc,o −
Tc,i ), substituting in (7.12) and rearranging, yields:
  
(Tcs − Tdiff ) 1 tT BC Tg − Tm
HLPaug = mc cp,c + Acs hcs + = , (7.15)
(Tc,o − Tci ) Ag h g kT BC Ag Tc,o − Tc,i

Energy & Process Technology Page 100

 Delft University of Technology

where the subscript aug is used to denote augmented. Historically, the HLP has been used to represent
a nondimensional coolant mass flow rate. The augmented HLP incorporates the heat transfer from
combustion parts to the midframe as well as the mass flow through the cooling features. As a result, the
augmented HLP represents a nondimensional cooling rate.

Another important term is the Cooling Effectiveness, defined as:

Tg − Tm
Φ= (7.16)
Tg − Tc,i

Coldside cooling contribution to HLP

Coolant contribution to HLP

Convective coldside
Convective coldside cooling
cooling only only
Combined coldside and internal cooling

Heat load parameter (HLP)

Figure 7.14: Effect of HLP on cooling effectiveness, and the ratio of passive to active cooling requirement
as a function of the cold side convective heat transfer coefficient. The internal cooling efficiency is
indicated using ηth = (Tc,o − Tc,i )/(Tm − Tc,i ), i.e. the actual channel heat pick-up to the theoretical
maximum

The figure above illustrates the cumulative nature of the augmented HLP. The thermal efficiency
curves for heat exchangers are plotted in the main figure on the left. The solid curves represent the opti-
mal designs where the part is cooled entirely by convective heat transfer to the midframe. The symbols
represent designs for the same range of cooling effectiveness (Φ) values which use channel cooling as well
as convective cooling. Since the augmented HLP captures the combined cooling effect of the various
technologies used in combustion hardware, the cooling designs collapse onto a single curve regardless of
the presence of channel cooling, the thickness of the TBC, or the amount of convective cooling to the
midframe. The inset in the figure shows the breakdown between the part of the augmented HLP that
corresponds to the midframe convective cooling (in green) and that of the channel cooling flow (in blue).
Each design meets a Φ value of roughly 0.8 and has a similar value of augmented HLP. However, each
design uses a vastly different amount of cooling air to do it. If the convective cooling to is sufficient, no
internal cooling features are required to meet the required Φ value. However, more cooling air is required
as the available midframe convective cooling decreases.

This adds up to the expectations of the cooling performance of the cooling channels inside both to
and bottom panels of the geometry considered in this study. Referring back to Fig. 7.14 it is clear that
the bottom panel might require less coolant air despite its larger length. On the other side, the heat
transfer coefficient of the top panel on the cold side is considerably lower, and hence, passive cooling will
be grossly insufficient, as visible in Fig.7.14.

The temperature design limits of the materials of the TP also need to specified. They are driven
by a combination of material properties and design requirements. They would normally be set at the
beginning of the design cycle.

Energy & Process Technology Page 101

 Delft University of Technology

Table 7.6: Maximum allowable temperatures and design margins

Cold side Bond coating TBC

θw,max 0.22 0.35 0.96
θmargin -0.39 -0.37 -0.4

The non-dimensional interface temperatures on the bond coating, TBC, and cold side surfaces of the
transition are shown below. The respective design limits are displayed with shaded areas.
1 1
Baseline, circ. - BC
Baseline, circ. - Coldside
Baseline, circ. - TBC
0.8 0.8
Baseline, ribs - BC
Baseline, ribs- Coldside
Baseline, ribs - TBC
0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

(a) (b)

Figure 7.15: Local wall temperatures for bottom (a) and top (b) panels, with and without rib turbulators.

In accordance to the expectations from the effect of the heat transfer coefficient on the cold sides, the
difference between the two cooling schemes is augmented for the bottom panel, where the HTC on the
cold side is considerably lower. In the top panel, the temperatures of the cold side and bond coating are
at or above the critical design limits for the majority of their length. For further investigation into the
reasons behind the lack of enhanced cooling, the bulk channel velocities is presented below along with
the nondimensional bulk temperatures of the cooling air.

As seen in the plot of bulk coolant temperature, a cross-over occurs at approximately x = 0.6 and
x = 0.7, for the ribbed and smooth channels of the bottom channels respectively. A region of extremely
low HTC on the cold side of the panel is present. As a result, not a lot of passive cooling and channels
pick up heat. Further downstream the channel, the cold-side HTC increase again, and the passive cool-
ing drops the metal temperature lower than the channel air temperature. This is also apparent on the
metal temperature plot, i.e. the location where the BC and cold side temperatures reach a maximum
values. The observation underlines the importance of effective cold side HTC. With regards to the top
panel, coolant stagnation where in reality, the channels only pick up heat but due to the low velocity, is
the reason behind the high panel temperatures. The considerable drop in the bulk velocity due to the
presence of ribs is also visible in the figure.

Energy & Process Technology Page 102

 Delft University of Technology

0.3
Channel 1 - Baseline, circ.
0.3
Channel 2 - Baseline, circ.
0.25 Channel 1 - Baseline, ribs
0.25
Channel 2 - Baseline, ribs
0.2
0.2

0.15
0.15

0.1
0.1

0.05
0.05

0
0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

(a) Bulk temperature top panel (b) Bulk temperature bottom panel
0.8 0.6

0.7
0.5
0.6

0.4
0.5

0.4
0.3

0.3
0.2
0.2

0.1 0.1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

(c) Bulk velocity top panel (d) Bulk velocity bottom panel

Figure 7.16: Nondimensional bulk velocity and bulk nondimensional temperature along the two channels
for both geometries.

The bulk heat transfer coefficient, calculated according to Eq. 4.39, are also presented below. Fol-
lowing the formulation of the turbulent wall heat flux in eq. 4.36, the resultant values of the bulk HTC
approach zero downstream and are negative on the cold side for both top and bottom panels, i.e. the
wall temperature of the channel is higher than the temperature of the RANS-averaged cell temperature,
+
T , computed from the respective wall treatment (despite choosing a value for θref = 0.0). The differ-
ence between the smooth and ribbed channels is the largest at the channel inlet, and tends to decrease
downstream the channel. This observation holds particularly well for the bottom panel. The augmenta-
tion due to the rib structures (in the presence of spikes) also tends to diminish downstream, putting an
emphasis on the effectiveness of turbulators for the given problem.

A summary of the obtained mass flow rates from the stagnation BC and key non-dimensional parameters
are presented in Table 7.7 below. The coolant mass flow is nondimensionalized due to confidentiality
reasons, such that
ṁ
ṁ∗ = (7.17)
ṁcirc.
A direct comparison between the top and bottom panels indicate that a smaller decrease in the Nusselt
numbers and friction coefficients for the top panel is present, when compared to the smooth circular
channel. A look at the wall temperature profiles, however, displays that the performance is worse than
in the bottom panel. This is due to the worse passive cooling in the form of mapped HTCs, which has a
more pronounced effect on cooling and hence despite the better internal cooling it is overwritten by the
HTCs.

Energy & Process Technology Page 103

 Delft University of Technology

3000
Cold side - Baseline, circ. Channel 2
3000
Hot side - Baseline, circ.
Cold side - Baseline, ribs
2000 Hot side - Baseline, ribs
2000

1000
1000

0 0

-1000 -1000
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

(a) (b)
Channel 1 Channel 2
3000 3000

2000 2000

1000 1000

0 0

-1000 -1000
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

(c) (d)

Figure 7.17: Bulk heat transfer coefficients on both panels for each channel: (a)-(b) top panel, (c)-(d)
bottom panel.

Table 7.7: Comparison of two baseline cooling schemes

Bottom Top
Smooth Ribbed Smooth Ribbed
ṁ∗ 1.0 0.29 1.0 0.34
Rech 43,932 12,501 19,187 6,493
N uch 519.69 113.72 244.56 86.88
C f,ch 351.29 55.33 488.4 92.5
Φ 0.76 0.73 0.71 0.64
HLP 4.42 3.56 3.41 2.18
θBC 0.25 0.28 0.29 0.36
max
θBC 0.39 0.43 0.41 0.47

It is thus safe to conclude that the chosen rib geometry as a turbulation feature inside the channel
is not effective for such channel lengths at such pressure drops. The performance is hence influenced
mainly by the following three parameters:
• Amount of passive cooling and heating, i.e. cold and hot sides HTCs respectively
• Channel length
• Channel curvature

Channel curvature has a less pronounced effect however in this case, since the channel length of the top
panel (despite its strong curvature) is almost half of the bottom one. It is worth noting that the results

Energy & Process Technology Page 104

 Delft University of Technology

are only preliminary. The full picture and potential HT improvements will be achieved upon completion
of the optimization loop, presented in the following Chapter.

Energy & Process Technology Page 105

Chapter 8

Ribs Optimization Study

The present chapter introduces the optimization problem of the transition piece and discusses the
findings, by considering the thermal performance in terms of the previsouly-introduced nondimensional
parameters used in combustion hardware, and the absolute value of the coolant air consumption. The
potential drawbacks of using this approach are also pointed out.

8.1 Optimization set-up

8.1.1 Optimization objectives
The optimization presented in the study is, as previously mentioned, a multi-objective unconstrained
optimization. Minimizing the average surface temperature, along with minimizing the requirement for
cooling mass flow comprise the two objectives, whereas the constraints are related to the feasibility
(i.e. parameters) of the rib geometry. The CHT method and required boundary conditions have been
described in detail in section 7.1.

Temperature (lifetime) objective

Ideally, in a full-scale transition, where appropriate BCs can be applied, the objective can be set to
maximizing the LCF lifetime. Due to the simplified geometry, such structural constraints cannot be
used, and as an alternative the average surface bond coating temperature is to be minimized. This is
also in accordance to the works of Chi et al. [22], in which was established that the most appropriate
objective function is the averaged outer wall temperature of a turbine rotor blade. The objective consists
of a target average surface temperature of the bond coating, θtar , set to 0.2 for both panels, which is
considerably low given the high inlet temperatures. Thereby, the objective function in Eq. 8.1 becomes
zero only if the target temperature value is met.


fθ ( −
→
x ) = max θ(−
→
x ) − θtar , 0 (8.1)

Minimizing fθ thereby corresponds to minimizing the surface temperature.

Cooling channel mass flow objective

A similar approach is also followed for the cooling mass flow objective. An increase in the consumption
of cooling air results in a lower cycle efficiency.
 
ṁ(⃗x) − ṁoffset
fṁ = max (8.2)
ṁoffset

The objective hence increases when the mass flow rate is higher than a specified offset, set at a non-
dimensional value of 0.015 and 0.017 for the top and bottom panels, respectively. The OF is therefore

106
 Delft University of Technology

minimized. The use of such target and offset values is not a firm requirement, but is rather used for
convenience and the avoidance of the SHERPA proposing a possible solution of ṁ for example equal to
zero.

8.1.2 Optimization parameters & range

In the optimization, the geometry parameters included in the study relate only to the channel cross
section and ribs. The choice of the parameter values defining the rib geometry in the previous section
was based on the mid values of the design space. The exact range is set as:

0.86 ≤ H/Hbaseline ≤ 1.44, (8.3a)

0.5 ≤ w/wbaseline ≤ 1.17, (8.3b)
0.2 ≤ e/H ≤ 0.4, (8.3c)
10 ≤ p/e ≤ 70, (8.3d)
0.3 ≤ d/p ≤ 0.7, (8.3e)
1 ≤ e/b ≤ 4. (8.3f)

The maximum allowable value for the width of the channel was set such that the minimal resultant metal
difference (nondimensionalised as a fraction of the baseline hydraulic diameter, dh,rec ) is 1, whereas the
minimum and maximum value of the height was to a certain degree arbitrary, while still keeping some
broad degree of manufacturability into consideration (e.g. H/Hbaseline = 0.125 height with a 0.2 block-
age ratio, will result in unrealistically low rib geometry even for the most sophisticated equipment). All
parameters were discretized as continuous in an interval of 50 values.

The comparatively low number of parameters (six) included presents a feasible opportunity to establish
whether the increase in friction factor can be mitigated for an optimized rib geometry. Possible addi-
tional parameters are the x- and y- coordinates of multiple points used in the definition of the spline
for each panel. Such an approach will potentially yield a panel shape that is less or more sensitive to
pressure drop. However, optimization methods suffer from what is called “the curse of dimensionality”,
which is a term introduced by R. Bellman in 1961 [13] to describe the problem arising by the exponential
increase in volume associated with additional dimensions to a mathematical space. Thereby, with a
bigger number of parameters, the optimization problem tends to converge to an optimum design more
slowly and the requirement for number of evaluations due to the increased design space increases. Given
that on average a CHT computation in STAR-CCM+ takes ≈ 2 hours to complete (see Table 7.5), it is
therefore desirable to obtain a better design in a number of evaluations as little as possible.

Objectives
Constraints SHERPA (HEEDS) Optimized designs
Parameters
Optimized parameters

CAD Generation

Parametric model
CHT computation
(STAR-CCM+)
Flow/temperature
solution
Yes
Converged

No

New Design
No

Figure 8.1: Flowchart of the automated design workflow in the MDO software HEEDS. The variables
are modified in input (.in) files, which are read by the solver running in batch mode. The results are
then obtained from an output (.out) file.

Energy & Process Technology Page 107

 Delft University of Technology

Due to the trade-off nature of the optimization algorithm, i.e. a Pareto optimization, SHERPA selects
the optimal designs based on whether they dominate other designs. A design is considered dominant
over another in cases where it has showed better performance in at least one of the objectives, and not
worse in all others. Thereby, SHERPA yields a so-called rank to such designs equal 1. Rank 2 is given to
the remaining designs and those that are dominated by any other design in that group. The reamining
designs are established under ranks of 3, 4, etc. following the same procedure. The full automated
workflow is presented in Fig. 8.1 above.

The settings used in the SHERPA optimization algorithm are listed as:
• The size of the archive set, which is maximum number of designs that are stored at any time is
set to 12 designs, based on the algorithm requirement if Nparameters < 10, otherwise the nearest
multiple of 4 (Nparameters = 6) is recommended; the archive set in the algorithm can be basically
interpreted as a certain cycle of designs, upon completion of which the algorithm learns and the
next cycle begins based on a new set of parameter values.
• Number of evaluations = 200, minimum requirement of SHERPA is that it equals at least 15
Nparameters
• The entire archive set of 12 designs is evaluated in parallel on a computational multiprocessor
cluster using 48 CPUs each, to speed up the process.

8.2 Results & analysis

The particular evolution history of both objective functions per iteration is of great interest for un-
derstanding the full scope of the results and deriving a coherent set of conclusions. This is presented
in Fig 8.3 and Fig 8.4 for the temperature and cooling flow objectives, respectively. Another important
parameter to track is the measured performance index of each design, shown in Fig. 8.2. During the
optimization, SHERPA provides each design with a performance rating, which is determined by the
extend to which the constraints are met, along with the value of the objective function. In the case
when no constraints are presented, the performance is only based on the objective value. Hence, the
weighted-sum performance can be then theoretically used to return the best designs (see eq. 8.4)

X  wi Si fi (⃗x) 
Nobj
Performance = (8.4)
i=1
normi

The sign of the objective is indicated with Si , i.e. value of +1 for maximization problem and -1 for
minimization, and wi is the linear weight for the ith objective (default value is 1).

-1

-1.5

-2

-2.5

-3 Bottom panel
Top panel
-3.5
0 20 40 60 80 100 120 140 160 180 200

Figure 8.2: SHERPA Design performance.

It is worth mentioning, however, that the performance rating is not used as a criterion by SHERPA
for choosing the values in subsequent designs during the search. The normalization coefficient therefore

Energy & Process Technology Page 108

 Delft University of Technology

plays a key role in calculating the performance. The value in this study was chosen to be the baseline
values (responses from the ribbed case in 7.5), and can explain the lack of convergence towards -1.

No clear convergence to a minimized value for both temperature and cooling flow is present either.
The temperature value for the top panel displays oscillations between 0.4 and 0.325, a range higher than
the average temperature of the baseline case with circular 1.5mm channels. The same applies to the
bottom panel. The fluctuations are less pronounced and the difference with the baseline case is lower.
Similar results are found with regards to the cooling flow. Interestingly, the value is successfully mini-
mized by a factor of almost 30 (the lowest nondimensionalized value from the optimization is ≈ 0.035g/s
compared to the baseline value). These observations lay down a comprehensive introduction to the full
picture of the results of the optimization.

Bottom panel
0.5
Top panel
, Baseline circ., bottom
0.45 BC

BC
, Baseline circ., top
0.4

0.35

0.3

0.25

0.2
0 20 40 60 80 100 120 140 160 180 200

Figure 8.3: Evolution of temperature OF

0.5

0
0 20 40 60 80 100 120 140 160 180 200

Figure 8.4: Evolution of nondimensionalized cooling mass flow

The average temperature objective versus the cooling mass flow rate objective is shown in Figure
8.5 and Fig. 8.6 for both bottom and top panels, respectively below. The envelop of all CHT results
successfully form a well-defined Pareto front in both cases. The colorbar on the right of the figures is used
to show the cooling effectiveness Φ of each design point. Results are in agreement with the expectations
that reducing the bond coating temperature (increasing the lifetime) requires a higher coolant mass flow.

Table 8.1: Comparison of maximum and minimum values achieved by SHERPA for both panels to the
values with baseline channel configuration (BR - baseline ribbed, BCIRC - baseline circular).

OF min max BR BCIRC1

Top panel fθ 0.12 0.19 0.16 0.09
fṁ 2.4 47.4 21.4 65.4
Bottom panel fθ 0.06 0.09 0.08 0.05
fṁ 1.4 35.4 17.4 56.2

Energy & Process Technology Page 109

 Delft University of Technology

As already established from the evolution plots above, none of the optimized geometries managed to
achieve an improved cooling system in terms of the averaged bond coating temperature. However, a thor-
ough design exploration is achieved and the reduction in mass flow rate is considerably large, particularly
when compared to the lower range of resultant values of the temperature objective. Hence, at it can be
seen in the Appendix, the histogram distribution of the values used for all design parameters indicate
that the majority of the design space is successfully explored.The highest and lowest values of the two
OFs are presented above in Table 8.1, where the values are compared to the ribbed and smooth baseline
geometries. In terms of the OFs on the other hand, the optimized geometries lie within 3% and 1% from
the currently used channels in the TP (baseline circular), for the top and bottom panels, respectively.
The cooling mass flow rate OF corresponding to the minimum of the computed temperatures is lower
by a factor of 1.4 and 1.6, respectively, which is a considerable improvement and can help reduce overall
GT emissions, only if however the entirety of the temperature profile of the BC is below the material
temperature limit.

The high negative values of the Pearson correlation coefficient (PCC) can be interpreted as Pareto
set, in which the majority of the designs are close to the Pareto front (assuming a linear Pareto front).
The PCC value is commonly calculated as:

cov(x, y)
ρ(y) = , (8.5)
σx σy

where cov is the covariance and σx and σy are the standard deviations of the predictor x and predictant y,
respectively. The Pearson correlation coefficient is then used to calculate the slope of the linear regression
line:
σy
β = ρ(y) (8.6)
σx
which is ultimately used for the full equation of the linear regression:

yi = β0 + βxi , (8.7)

with the y-intercept denoted β0 = y − βx.

1 Intentional abbreviation to avoid confusion with bond coating (BC).

Energy & Process Technology Page 110

 Delft University of Technology

Top panel
0.42
0.68
SHERPA Rank
0.4 1 0.67
2
BR 3
0.38 0.66
Linear fit: =-0.986
0.36 0.65

0.34 0.64

0.63
0.32 BCIRC

0.62
0.3
0.61
0.28
0 0.3 0.6 0.9

max
Figure 8.5: Pareto front of top panel. Note: θBC = 0.35.

Bottom panel
0.3
SHERPA Rank 0.735
1
0.29 BR 2
3
0.73
0.28 Linear fit: =-0.799

0.27 0.725

0.26 BCIRC
0.72

0.25
0.715

0 0.35 0.7 1.05

max
Figure 8.6: Pareto front of bottom panel. Note: θBC = 0.35.

To investigate the effects of the independent rib parameters, the value of the computed θBC and coolant

Energy & Process Technology Page 111

 Delft University of Technology

are plotted against each parameter individually. The plots are shown in Fig. 8.7 and Fig. 8.8. The two
panels do display slight differences in the PCC values, but the overall trends remain the same: blockage
has a negative effect on the temperature OF (i.e. temperature is increased with increasing blockage), and
possitive effect on the coolant mass flow rate OF. The increase in size of the recirculation regions inside
the channels, resulting from the higher ribs, contribute to higher friction, which explains the observation
mentioned above. Basically, with lower blockage ratios, the two OFs tend to converge to a single region:
see values for e/H < 0.225 below. The opposite holds for the pitch ratio, which is to expected. The
two OFs tend to converge after a p/e > 50. Increasing pitch ratio can also be interpreted as decreasing
the total number of ribs inside the channel. This reduces the skin friction losses and increases the mass
flow rate. An interesting observation is that the correlation in these two parameters is less pronounced
in the top panel. This can be attributed to the bigger effect passive cooling has on this panel and also
the strong curvature. Further investigation needs to be done to confirm this assumption however. On
the other hand, the periodicity and thickness ratio of the rib has a considerably less clear effect on the
two OFs, analogous to low PCC values, visible in the two abovementioned figure.

The results of both objectives are almost perfectly antisymmetric: the correlation coefficients ρ(OFi )
are approximately equal but with a negative sign, emphasizing on the strong coupling and trade-off
between the two objectives. This however should not be considered as a lack of improvement from
the optimization run. In a realistic problem setup, where the lifing is calculated using the appropriate
boundary conditions and full geometry, then the results will be more decoupled due to the introduction
of non-linearities from LCF (local stress regions, application of external loads, etc.). The results, how-
ever, are still indicative that heat transfer can be enhanced using varying turbulation features without a
considerable penalty in friction (in the form of low mass flow rate). A clearer picture is painted when the
HTC augmentation in the form of Nusselt number is plotted against the Skin friction, presented below.
Moreover, the plots are also beneficial for displaying the presence of outliers (i.e. high N u numbers
against a low value of Cf ) from the linear fit, which is desirable. Such data was recorded for the top
panel only and is shown below in Figures 8.9-8.11.

Energy & Process Technology Page 112

 Delft University of Technology

Cooling flow data

Cooling fit
0.7 Temperature 0.3data 0.7 0.3
Temperature fit

(mdot) = -0.67 (mdot) = 0.12

0.55 0.29 0.55 0.29
( ) = 0.75 ( ) = -0.11
0.4 0.28 0.4 0.28

0.25 0.27 0.25 0.27

0.1 0.26 0.1 0.26

0 0.25 0 0.25
0.2 0.25 0.3 0.35 0.4 0.2 0.3 0.4 0.5 0.6 0.7 0.8

0.7 0.3 0.7 0.3

(mdot) = 0.79 (mdot) = -0.11

0.55 0.29 0.55 0.29
( ) = -0.86 ( ) = 0.01
0.4 0.28 0.4 0.28

0.25 0.27 0.25 0.27

0.1 0.26 0.1 0.26

0 0.25 0 0.25
0 20 40 60 80 0 1 2 3 4 5

Figure 8.7: Influence of rib parameters on bottom panel temperature and cooling

Energy & Process Technology Page 113

 Delft University of Technology

Cooling flow data

Cooling fit
0.8 Temperature0.4 data 0.8 0.4
Temperature fit
0.39 0.39
0.7 0.7
(mdot) = -0.5 0.38
(mdot) = 0.09 0.38
0.6 ( ) = 0.55 0.6 ( ) = -0.09
0.37 0.37
0.5 0.36 0.5 0.36

0.4 0.35 0.4 0.35

0.34 0.34
0.3 0.3
0.33 0.33
0.2 0.2
0.32 0.32

0.1 0.31 0.1 0.31

0.2 0.25 0.3 0.35 0.4 0.2 0.3 0.4 0.5 0.6 0.7 0.8

0.8 0.4 0.8 0.4

( ) = -0.71 ( ) =0.41 (mdot) = -0.35
0.39 0.39
0.7 0.7
(mdot) = 0.64 0.38 0.38
0.6 0.6
0.37 0.37
0.5 0.36 0.5 0.36

0.4 0.35 0.4 0.35

0.34 0.34
0.3 0.3
0.33 0.33
0.2 0.2
0.32 0.32

0.1 0.31 0.1 0.31

0 20 40 60 80 0 1 2 3 4 5

Figure 8.8: Influence of rib parameters on top panel temperature and cooling

The range of values for Φ was split into four ranges, and then used to indicate the marker size of
all three plots to aid in the understanding of the results. The Nusselt number was calculated using
the average wall values for both specified-y + heat transfer coefficient and conductivity, the latter being
calculated using the average channel temperature and Sutherland’s law. A small difference between the
two channels in terms of variation with Reynolds number is present, but the data points eventually
converge as visible in Fig. 8.11. For all designs a strong linear trend is present, particularly for the
skin friction coefficient. For the case of cooling performance however, a lower correlation coefficient is
actually beneficial, as it can be used to display an increase in HTC, without a corresponding increase in
skin friction losses. The presence of such outliers can also be quantified with the so-called Z-score:
x − µx
Zx = (8.8)
σx
A distribution of the z-score of the difference between the predicted value (i.e. linear regression line)
and the computed value from the optimization run, which is more skewed or with a higher σ is therefore
desirable. The z-score of the skin friction difference however, is advantageous for the rib performance to
remain uniformly distributed with a very low σ as a countermeasure. The z-score for both N u number
and Cf are shown below in Fig.8.12.

Energy & Process Technology Page 114

 Delft University of Technology

110

100

90

80

70

60

50

40
Channel 1
30
Channel 2
Channel 1 fit
20
Channel 2 fit
10
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
104

Figure 8.9: Nusselt number increase with Reynolds number

220

200

180

160

140

120

100

80

60

40
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
104

Figure 8.10: Skin firction increase with Reynolds number

Energy & Process Technology Page 115

 Delft University of Technology

110

100

90

80

70

60

50

40

30

20

10
40 60 80 100 120 140 160 180 200

Figure 8.11: Nusselt number against skin friction

Nusselt number Skin friction coefficient

50 60
Channel 1
40 Channel 2

20

0 0
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3

200 200

150 150

100 100

50 50

0 0
-3 -2 -1 0 1 2 3 -5 0 5

(a) (b)

Figure 8.12: Z-score (scatter and distribution) of (a) Nusselt number and (b) Skin friction coefficient.

Figure 8.11 is hence a good summary of the objective of this optimization run. The plot displays a
low PCC ρ = 0.87, which is due to the wide combinations of rib parameters. A wide range of Nusselt
numbers for both channels are well above the linear regression line, proving that a rib geometry can be
customized in accordance of the given flow problem.

The overall cooling performance, however, of equipment used in gas turbines, does not depends solely
on the internal cooling HTC. As pinpointed previously, the issue of low cold side HTC on the TP is of
high impact on the cooling. Therefore, additional cooling adjustments are necessary to further decrease
the effect of skin friction as discussed in the next section.

Energy & Process Technology Page 116

 Delft University of Technology

8.3 Recommendations
It was observed that the surface average temperatures of the bond coating (i.e. transition metal)
on the top panel are above the maximum allowable design constraint of the material (see Table 7.6).
The maximum BC temperatures were not recorded during the optimization process, but considering the
results in Section 7.5 and particularly Fig. 7.15, one can assume that they are proportionally above the
average value and hence above the material limit for some of the designs on both panels. The optimiza-
tion, however, displayed a promising opportunity that such turbulation features can in fact be tailored
and help reduce the mass flow rate drastically, with a smaller penalty on temperature. Thereby, the
study points to the assumption that the temperature penalty can be further reduced, given the cooling
design is adjusted carefully. A possible adjustment is presented in Appendix D, where additional inlets
downstream are added, while their respective inlet cross-section area is progressively reduced to artifi-
cially throttle the flow downstream. Recommendations arising from that study led to the conclusion that
a large potential improvement in overall cooling (both in Φ and HLP values) is expected if in addition to
inlets, separate channel outlets are added as well, thus effectively reducing the channel length. A further
study in which the number of additional contracted inlets is matched with contracted outlets can help
validate this assumption.

The analysis presented the overwhelming effect cold-side cooling actually has on the transition piece.
Therefore, another recommendation is designers’ attempt to maximise the passive cold-side convective
cooling along with the use of internal cooling techniques. This is achievable, by maximising cold air cir-
culation in the plenum, and avoiding stagnation regions around the combustor inlet ring and the picture
frame on the turbine side.

Additionally, from the correlation plots it was found out that the periodicity and aspect ratio of the
ribs play approximately no role in the effectiveness of the cooling geometry. The range of values pro-
vided for the parameters with the highest influence, however, i.e. rib blockage and pitch were relatively
low. Therefore, a further study can increase the range for the abovementioned parameters along with a
higher maximum value of the channel height (the maximum value of the channel width chosen here does
not allow further increase for structural reasons). As previously explained the range of values provided
for some parameters was rather arbitrary. Therefore, in the current study, the maximum Dh = 1.38mm
is still lower than the circular channels currently used by Siemens (Dh = 1.5mm), and matching Dh in
both cases, circular baseline and the maximum value for the ribbed channel can eventually help conclude
whether the gap with the temperature value from the circular baseline case can be further reduced (an
approach that can be interpreted as an attempt to extend the Pareto front - see linear regression line in
Fig. 8.5 and Fig. 8.6).

Energy & Process Technology Page 117

Chapter 9

Conclusion

This final chapter of the report summarizes the approach followed in the study and the main key
findings. It also lays out further recommendations for potential fields of research that can shed light on
a broader spectrum of topics related to the objectives of the study

The study objectives were to validate the conjugate heat transfer methodology in a commerical software
for two canonical cases, and compare the conjugate heat transfer results with the uncoupled solution
under a wide envelope of conditions. The study also aimed at providing an insight whether an optimiza-
tion methodology, using a proprietary optimization algorithm, can lead to an optimized rib-turbulated
internal cooling channel geometry of the transition piece of an industrial gas turbine, such that lifing is
enhanced with the lowest mass flow rate possible.

9.1 Conclusions
Firstly, the conjugate heat transfer methodology in the commercial code package was successfully val-
idated. It included first a laminar flat plate case, the results of which were compared to an analytical
solution under two physical settings (Bi = 0.1 and Bi = 1). The model employed a finite element mesh
on the solid side in order to quantify the resultant difference when a mapped interface is used. The
results showed better convergence with the higher Biot number with the analytical solution, primarily
due to the lower temperature gradients at the leading edge. For both cases, the assumptions used in the
derivation of the analytical solution are in part the main cause of the discrepancy.

The validation of the turbulent flow field case (a turbulent offset jet) helped remove uncertainty be-
tween the available Reynolds-Averaged turbulence models, and also types of wall treatment. The study
showed that the resultant temperature profile from the realizable k − ε model yields the closest results
to an experimental measurement from literature. The reason for the better computation is the different
formulation of the turbulent viscosity, which was overpredicted for the k − ω model, particularly in the
stagnation regions, which led to higher temperature values. The runs using a high-y + treatment on a
coarser mesh suffered from high numerical dissipation, which artificially led to results being closer to the
experimental values. This artefact from the coarse grid points to the possible inappropriacy of the chosen
experimental setup: the low jet velocity in the experiment did not allow for robust testing of the high-y +
approach as the resultant solution is mesh dependent. Moreover, the resultant y + value on the wall was
in the buffer layer which is strongly advised against when suing this wall function. It is worth mentioning
however that the difference between the two approaches (blended and high-y + ) however remained small
in terms of maximum and minimum interface quantities and the overall interface temperature profile on
the plate.

Secondly, the subsequent part of the study: quantifying the relevance of the uncoupled approach showed
the following. When the thermal conductivity of the solid region was altered on the laminar flat plate

118
 Delft University of Technology

case, the error between the two heat transfer methodologies remained relatively low, with the tendency
to increase with lower solid conductivities, in accordance with the Lumped Capacitance theory. The
difference however remained negligible, particularly for averaged temperature values. Additionally it
was shown that increasing the difference between the cold (bottom wall of the plate) and hot tempera-
tures in the domain (T∞ − Tb ), can yield to an increase in the errors of the decoupled approach. The
maximum error however, remained small at ≈ 6%. More prominent differences were observed in the
thermal boundary layers.

When the heat transfer methodology was compared using a turbulent flow case on an offset jet, the
sensitivity study with the Turbulent Prandtl number (range between 0.4 - 1) showed no noticeable ef-
fect for both wall treatments, and the difference between the conjugate heat transfer and decoupled heat
transfer tended to remain constant. The second parameter studied was the turbulence intensity at the jet
inlet. The blended wall approach seemed to be independent of turbulence intensity. The only parameter
investigated that showed larger fluctuations was the reattachment length (the difference remained small
however at 3%). The lack of established tendency is even more pronounced for a high-y + approach. An
important observation is that the difference between conjugate heat transfer and the decoupled approach
seemingly increased with increasing Tu(%) for the high-y + approach. This observation, however, can
be solely an artefact from the coarse mesh size, rather than a result coming from the WF itself. The
third parameter was again, same as in the laminar flat plate case, the thermal conductivity of the solid
(presented as a conductivity ratio with the fluid region). In terms of average values, the difference in
temperature is negligible for both wall treatments. The largest divergence was at stagnation regions but
only apparent in the blended wall approach. For the high-y + treatment, the difference was artificially
softened, because of numerical dissipation, as discussed in the validation case, and remained rather small
even for intermediate solid conductivities. Hence, the effect of the solid thermal conductivity was negli-
gible for both local and averaged values on the wall when a coarse grid approach was used. In general,
a consistent observation for all sweeps tested on the blended wall function, showed that the interface
temperature and was higher when conjugation was included accounted for.

Finally, the last part of the Masters Thesis surrounded around the cooling study and optimization
of a gas turbine transition piece. A simplified domain was proposed, in order to reduce computational
costs, and analyse cooling performance only on the most thermally loaded parts of the transition. This
was done by, instead of calculating a full-scale mid-frame combustor model every time, the heat transfer
coefficients and ambient temperatures were mapped on the hot and cold sides of the transition piece.
The walls of the channels were the only conjugate heat transfer interface in the study. The addition of
turbulators (in the form of straight rectangular ribs perpendicular to the flow) resulted in a decrease
in the cooling performance for both top and bottom panels, with a more pronounced deterioration on
the top panel. This was due to the comparatively low cold side cooling on the panel, and the large role
that this component of the total cooling plays on the performance. Numerically, the increase in the
average bond coating temperature was 12% and 24% for the bottom and top panels, respectively, when
compared to the case of circular channels without turbulators. On the other hand, the improvement in
the consumption of coolant was reduced by a factor of 3.1 and 2.9. Since the temperature profiles for the
turbulated channels were very close to the temperature design limits, the performance of the baseline
ribbed geometry was deemed unsatisfactory and the possible improvement of the turbulated cooling is
only fully visible in after the optimization routine.

The optimization procedure helped further reduce the bond coating temperature for both panels com-
pared to the results from the initial rib configuration, specifically a reduction of 11% and 6.25% for
top and bottom panels, respectively. The resultant penalty for the mass flow rate was augmented by
approximately a factor of 2, but still remained considerably lower than the non-turbulated channel de-
sign used currently in the transition piece. In summary, the average temperature of the bond coating
were still unsatisfactory and higher than the non-turbulated channels, but the reduction in the coolant
mass flow rate was considerably higher. The results put an emphasis on the possibility to optimize
turbulation features and reduce the penalty on the pressure drop (expressed in mass flow in this study),

Energy & Process Technology Page 119

 Delft University of Technology

without solely relying on the experience of the designer. The ultimate cooling design choice is based
on the following: unless the average bond coating temperature has to be strictly lower than the previ-
ous configuration, one can use the lowest achieve value of the temperature, and considerably lower the
coolant consumption. A complete improvement in both mass flow rate and temperature profiles and,
thus, lifing, however, is only possible, however, with an additional measures in the cooling design, e.g.
additional inlets and outlets downstream, which will most likely lead to a temperature value close to the
nonturbulated circular channel at a reduced mass flow rate.

9.2 Recommendations
A list of further developments and improvements proposed for the methodology implemented in this
project potentially include:
• A different experimental set-up might yield different recommendations of a model and WF: the
results presented above are in no way binding to all flows, and in case of a different flow structure,
the appropriate model should be then re-chosen based on assumptions, and available literature
(unless validation data is present). A validation of the conjugate heat transfer methodology at
different settings is imperative to confirm or deny the finding that the realizable k − ε is the best
for conjugate heat transfer problems.
• The comparison between the two heat transfer methodologies can be expanded into sensitivity
studies including other nondimensional numbers or parameters. For example, it was briefly mention
that the the Turbulent Prandtl number study is relevant for reacting flows. The flows studied in
this project however were not reacting, and hence, a more interesting parameter to study is the
ratio between the turbulent Schmidt and turbulent Prandtl numbers.
• Additionally, with the advent of high performance computing, Large Eddy Simulations (LES) and
also Detached Eddy Simulations (DES) can be used to quantify the difference between the conjugate
heat transfer and decoupled approach. It can be, thus, found whether the behaviour of such large
flow structures, in an unsteady setup, omitted in the steady Reynolds-Averaged approach, plays a
big role on conjugation.
• With regards to the optimization study, a further step into addressing the trade-off between pressure
drop and temperature is using a more complex rib geometry. In Literature various rib configurations
were presented, such as wedge-shaped ribs, or backward-aligned ribs. The parametrization of such
ribs and a similar optimization analysis can shed more light on the turbulation technique offering
the lowest friction/heat transfer coefficient ratio.
• Finally, from the correlation plots of the rib parameters of this study, it was established that the
rib thickness and periodicity ratio play almost no role in both objectives. Therefore, one can run
additional computations with an extension of the specified range of blockage and pitch ratios to
establish whether the Pareto front also extends and achieves also better temperatures, and not
only mass flow, compared to the nonturbulated circular channel.

Energy & Process Technology Page 120

Appendices

121
Appendix A

Gas turbine basics

Derivation of the thermodynamic efficiency (at optimum pressure) of the Joule-Brayton cycle is herein
presented. The thermodynamic efficiency of the cycle is defined as the ratio of the extracted work to the
added heat (assuming no losses, i.e. ideal cycle):

Net work (T4 − T5 ) − (T3 − T2 )

ηth−dynamic = = , (A.1)
Heat added T 4 − T3
where T5 is the temperature at the end of the expansion process, and T3 is the compressor outlet
temperature. The expression can be further rearranged to
T2
ηth−dynamic = 1 − . (A.2)
T3

The net work (or the specific gas power, sgg, i.e. work per unit mass) is hence:

Ws,gg = cp (Tg − T5 ) = cp [(T4 − T5 ) − (T3 − T2 )] , (A.3)

and using the isentropic gas equation:

κ−1 T3 T4
PR κ = = , (A.4)
T2 T5
 
T4 T2
Ws,gg = cp T4 − − T3 + T2 . (A.5)
T3
Differentiating w.r.t T3 and equated to zero to find the peak value of specific power for a given temperature
ratio (T4 /T2 ), called the optimum pressure ratio, yields:
 
dWs,gg T4 T2
= cp −1 =0 (A.6)
dT3 T32

hence,
T32 = T2 T4 , (A.7)
which yields p
T3 = T2 T 4 , (A.8)
and finally substituting into eq.A.2:
r
T2
ηth−dynamic = 1 − .
T4

122
Appendix B

Luikov solution

B.1 Direct heat transfer solution method

The constants used in Eq. 5.5 (K and B) are given below:

λs x 1 1
K= (P r)(− 2 ) (Rex )(− 2 ) (B.1a)
λf b
q
vy
B= P rRex (B.1b)
vx
The exact formulation of the Nusselt number, and hence HTC is calculated using:
1 1 1
N uDHT
x = P r 2 (Rex ) 2 N (K, B), (B.1c)
π
where √
φ(K, B) − 12 πB erfc B
N (K, B) =
, (B.1d)
1− K B
− K√
1
π
φ(K, B) + 12 K
B
erfc 12 B
 
1B

φ(K, B) = 1 − (π)K exp K 2 − BK
2K
  (B.1e)
1
× erfc K − B .
2

B.2 Boundary layer approximation

The procedure omitted in the derivation of the thermal boundary layer thickness in Chapter 5 includes
the evaluation of the following integral:
Z δt "  3 #
∂ 3y 1 y
v∞ 1 − + (θ∞ − θw )
∂x 0 4 δt 2 δt
    (B.1f)
3 y 1  y 3 ∂θ1 (x, 0)
× − dy = α ,
2δ 2 δ ∂y

By evaluating the integral and applying some mathematical operations ultimately yields:
   3
∂ 1 1 10αz
2
= , (B.1g)
∂x (1 + z)z δ zδt v∞ (1 + z)

123
 Delft University of Technology

Given the limit value of the equation, limz→0 , (B.1g) becomes:

   3
∂ 1 1 10αz
= (B.1h)
∂x z2δ δt v∞

The Biot number can be calculated analytically, using the approximation of the HTC, such that:
  √
λf √ Rex
Bi = 0.332 · b
3
Pr . (B.1i)
λs x

The Nuselt number is calculated as:

 
x ∂T 3x
N uBLE = = (B.1j)
x
(T∞ − Tw ) ∂y 2 δt

B.3 Numerical solution

The energy coupling setting was chosen to explicit, since it allows for coupling of FVM-FEM meshes.
This includes a mapped static temperature on the fluid mesh interface, and mapped reference temperature
and mapped local heat transfer coefficient on the solid interface. The types of solvers for each domain
are briefly presented in the main part of the report. A summary of the boundary conditions settings
is given in Table B.1 A compact interpolation stencil was selected for connection method for the mesh
faces on the two sides of the interface. The interpolation stencil method on interfaces is briefly discussed
in Appendix E.1, including a short comparison with the alternative methods available in the code.

Table B.1: Interface boundary condition settings

Type Mapped Contact

Energy coupling Explicit
Energy source None
Contact resistance None
Mapped local HTC
Solid side
Mapped ref. temp.
Fluid side Mapped static temp.
Interpolation stencil Compact
Interpolation method Nearest Neighbour

The mesh study included five sizes tabulated below. The naming convention used for each mesh is
based on the resultant minimum y + value on the fluid/solid interface. To the author’s knowledge, the
y + is exclusively a turbulence based parameter, but will be solely used in this study for the sake of
convenience for addressing the level of mesh refinement in the vicinity of the wall. The results of each
mesh are compared in the following section. The number of elements in the z-direction was kept constant
at 10 for all five meshes. The mesh study focuses only on Bi = 0.1. A value for the Biot number of 1 will
only be used for the studies in Section 5.1.3, in which the findings of the validation study are discussed.

Table B.2: Properties of meshes used for laminar flat plate study

Grid size ∆ymin @wall (m) Stretching factor Approx. min. y +

60x120 7.2e-05 1.061 0.5
60x60 1.4e-04 1.115 1
60x40 7.2e-04 1.129 5
60x20 1.4e-03 1.235 10
60x5 1.4e-02 1.467 100

Energy & Process Technology Page 124

 Delft University of Technology

The results, indicating the accuracy of each mesh are shown below. From the resultant interface
quantities, it is clear that accuracy is gained, as expected, with increasing (logarithmic) refinement near
the interface. Since the difference between the first and second finest meshes (y + = 1 and y + = 0.1)
is essentially intangible and the computational speed gain is negligible (Figure B.2, for the sake of the
results of this study, presented in Section 5.1.3, and all upcoming studies, including the comparison of
the conjugate heat transfer method with the decoupled approach on this study case, are based on the
finest mesh (60×120 cell elements).

(a) 60×5 (b) 60×20

(c) 60×40 (d) 60×60

(e) 60×120

Figure B.1: All five meshes used for the independence study of the flat plate problem, in the vicinity
above the plate (no rear and front fluid extensions)

Energy & Process Technology Page 125

 Delft University of Technology

10 4
M1(y+ = 0.5) - Energy
M1(y+ = 0.5) - Thermal Load
M1(y+ = 0.5) - Temperature
M2(y+ = 1.0) - Energy
10 2 M2(y+ = 1.0) - Thermal Load
M2(y+ = 1.0) - Temperature

10 0

10 -2

10 -4
0 20 40 60 80 100

Figure B.2: Temperature, energy, and thermal load residual values from M1 and M2, until iteration
number 120 (convergence for M1 reached at 260 iterations)

Figure B.3: Comparison of friction coefficient to Blasius analytical solution, along with heat transfer
coefficient, heat flux, and temperature at the interface with Luikov analytical solution (Bi = 0.1).

Expanding on the mesh study results, it is evident that for M1 and M2, in particular, the difference
between the numerical solution and the analytical one is the highest in the vicinity of the leading edge,
where the gradients are the highest. The error hence grows from ≈ 1% to ≈ 2.5% upstream from x/Lref
= 0.25 to x/Lref = 0.0, respectively, for both M1 and M2. Therefore, an additional step taken towards
minimizing this error was adding mesh stretching also downstream in the x-direction (in a similar fashion
such that x+ = y + ) in both solid and fluid regions. The resulting mesh is presented below.

Energy & Process Technology Page 126

 Delft University of Technology

Figure B.4: Mesh M1 with x-refinement (incl. extensions in fluid region).

900
Luikov BLE
850 Luikov diff. h.t
M1(y+ = 0.5), with x-refinement
M1(y+ = 0.5)
800

750

700

650

600
0 0.2 0.4 0.6 0.8 1

Figure B.5: Comparison of interface temperature obtained from refined mesh in the x-direction w.r.t
mesh without refinement and the analytical solution from Luikov (Bi = 0.1).

However, the refinement near the leading edge did not lead to any benefit in terms of convergence
with the steep gradients present in the analytical solution, as shown in Figure B.5. The only noticeable
difference is the slight increase in the temperature downstream, however, it still remains negligible.
Hence, due to the lack of improvement in the vicinity where the difference is the highest, the subsequent
studies are done on mesh M1, as previously mentioned, having ensured that the solution is independent
on the mesh fineness.

Energy & Process Technology Page 127

Appendix C

Turbulent offset jet mesh study

The experimental domain that was attempted to be replicated is shown in Fig. C.1.

Figure C.1: Schematic of experimental arrangement, extracted from [95]

The entire CFD domain is shown below.

Figure C.2: Entire CFD domain with relevant dimensions (m) (light gray region is not included in Figure
5.11, i.e. ambient room region in the experimental domain, having a width of 3.0m)

128
 Delft University of Technology

Table C.1: Specification of mesh settings on target surfaces for M2

Control value
Surface (region)
TSS 1 (mm) GR2 PLT3
Ambient 250 1.05 N/A
Plexiglass walls 10 1.05 3
Interace 0.5 1.05 5
Plenum chamber walls 0.5 1.1 N/A

A short description of some key attributes of all meshes used for the study is shown in the table
below, followed by a set of comprehensive plots, to illustrate the mesh convergence. The meshes and a
table containing the respective mesh settings along all surfaces (table contains data only for mesh M2)
are illustrated in Figure C.3 and Table C.1, respectively. Clearly, the solution tends to be independent
around an mesh size of ≈ 8 × 106 . Figure C.4 displays the convergence of the peak value of the friction
coefficient along the centreline of the interface and the measured reattachment length.

Table C.2: Mesh attributes of all meshes in the mesh independence study

+ +
Reference ID Cell count ymax yave Cf Xcr
M1 15.9M 2.25 1.63 1.16 16.64
M2 10.3M 3.59 2.62 1.22 16.32
M3 4.9M 4.62 3.26 1.24 16.00
M4 2.4M 6.14 3.83 1.11 15.28
M5 1.5M 8.56 5.31 1.13 14.48
M6 226K 15.91 11.65 1.01 11.91

The red horizontal line represents the experimental reattachment length for OR=7 in the study by
Pelfrey, equal to 12.42. In the current study, it was measured by taking the axial location of the mini-
mum friction coefficient. All meshes overpredicted the reattachment length, which was found to be one
of the key differences between the studies in the results in Section 5.2.2. Later, it will be shown that this
finding is independent on the choice of RANS model, but mostly on the mesh size.

Meshes M1-M3 display very similar velocity and interface friction profiles. The resultant reattachment
lengths for meshes M4-M6 are visible lower and friction profiles are inherently more dissipative, which
in particularly becomes evident in the region downstream (∆X ≈ 7), using the friction profile as an
example. This observation can also be traced back to the normalized axial velocity profiles plotted in
Figure C.5 for four regions - two for recirculation region (X = 3, 6), impingement (X = 9), and wall jet
region (X = 15).
1 Target surface size (mm)
2 Growth rate factor
3 Prism layer thickness (mm)
4 No thermal conductivity is given in the paper.

Energy & Process Technology Page 129

 Delft University of Technology

(a) M5 (b) M4

(c) M3 (d) M1

(e) M2, including a zoomed view on the mesh at the interface

Figure C.3: All five meshes used for the independence study of the flat plate problem, in the vicinity
above the plate (no rear and front fluid extensions)

The dissipative effect of M5 is less apparent in velocity: the peak axial value is lower by between 5% to
10% in the impingement and wall jet regions, respectively. A more striking difference is that despite the
relatively good fit with experimental data for the friction coefficient along the plate, the axial velocity
for all finer meshes is higher than in the experiments following the recirculation region. Only mesh M6
seems to match the physical dissipation of the problem based on velocity data and is the closest to the
experimental attachment point.

2.6 M5
1 M3
18 Present study M2
2.5 Pelfrey and Liburdy (1986), OR = 7 M1
Pelfrey (1986)
0.8
2.4 16

2.3 0.6

14
2.2
0.4

2.1 12
0.2
2
10
0 5 10 15 0 5 10 15 0
0 5 10 15 20 25 30
10 6 10 6

(a) Maximum coefficient of friction (b) Reattachment length (c) Coefficient of friction along the
centreline on the interface

Figure C.4: Convergence for M1, M2, M3, M4 and M5

Energy & Process Technology Page 130

 Delft University of Technology

12 M1 12
M2
M3
10 M5 10
M6
Pelfrey and Liburdy (1986)
8 8

6 6

4 4

2 2

0 0
-0.5 0 0.5 1 -0.5 0 0.5 1

(a) x/t = 3 (b) x/t = 6

12 12

10 10

8 8

6 6

4 4

2 2

0 0
-0.5 0 0.5 1 -0.5 0 0.5 1

(c) x/t = 9 (d) x/t = 15

Figure C.5: Axial velocity component convergence with experimental data at five downstream location

Thereby, based on the brief analysis done above, mesh M2 was chosen for the purpose of the upcoming
studies for all computations involving All- or low-y + wall functions. The grid size provides satisfactory
results and very close to to M1, and ensures a grid-independent solution. The fine discretization around
the near wall region is also benefitial for the analysis of scalar fields in the subsequent studies.

Multiple additional measures were taken in terms of solver type/settings and parameters to address the
mismatch of experimental and numerical reattachment length, and hence axial velocities downstream.
Such include:
• Turbulence intensity sweep (no CHT enabled) - Constant air properties were applied to the
fluid; the jet temperature at the outlet was set equal to that of the ambient air; the free-stream
and near-wall turbulent Prandtl number were both set to a constant value of 0.9.

• Prandtl Number - Various values for the Prandtl number were also tested, under the assumption
that the varying thermal and momentum diffusivity would tangibly change the velocity profiles used
for validating the mesh and settings.
• Multi-component mixture, along with variable properties - Thermal conductivity and
dynamic viscosity were both algebraically modelled using Sutherland’s formulation.

All three abovementioned methods yielded no improvement in the accuracy of the simulations. This
is partially documented in the Appendix, where the different RL values are tabulated along with axial
velocities at the four aforementioned locations downstream.

Energy & Process Technology Page 131

Appendix D

Auxiliary contracted inlets study

D.1 Concept introduction

Since no improvement in the cooling was achieved, a proposed method was keeping the length of the
cooling channel the same as in the baseline, but with additional inlets positioned equidistantly down-
stream. The cross section of the generated aux. inlets was kept identical to the main original inlets
of the ribbed component inlets with d∗h,aux . Initally, the number of aux. inlets is swept from 1 to 3,
at 5 different angles, α = 30◦ , 45◦ , 60◦ , 90◦ , for each number, leading to 15 evaluations in total. An
illustration is shown below.

Figure D.1: Schematic of the setup: auxiliary inlets are place downstream for both counter current
channels.

The results of this approach are presented in Section D.2. Subsequently, a new parametrization on
the inlets was proposed with the goal of further reducing the BC temperature. By decreasing the area
of the channels downstream closer to the outlet, the inlet velocity will be higher and hence the coolant
can be artifically throttled, mitigating for the velocity decrease (arising from the pressure drop at the
end). The inlet area of the aux. inlets was parametrized and related to the main inlets using:
i
Ai = Cadjusted Amain . (B.1a)

The power of the area factor is equal to the number of the aux. inlet, e.g. the 2nd inlet downstream
uses an exponent equal to 2. Two aux. inlets per channel were used (i.e. only a quadratic relationship).
Three hydraulic diameters (of the main inlet), d∗h = 0.7, 0.9, 1 for each correction factor are used. A
schematic is shown below, to illustrate the concept of such artificial throttling.

132
 ௜
‫ܣ‬௜ ൌ ‫ܥ‬௔ௗ௝௨௦௧௘ௗ ‫ܣ‬௠௔௜௡ Æ ‫ܥ‬௔ௗ௝௨௦௧௘ௗ ՛ǡ ‫ܣ‬௜ ՛

• The power of the area factor is equal to the number of the aux. inlet:
e.g.: 2nd inlet downstream is pow 2

• 2 aux. inlets per channel were used (i.e. only a quadratic

Delft relationship)
University of Technology
• Three hydraulic diameters (of the main inlet) [1.0 1.3 1.6] mm for each correction factor

Hydraulic diameters of the main in

Main inlet

Aux. inlet (1)

Channel
Aux. inlet (2)

outlet
Cold inflow Heated outflow Width (mm) Height (mm)
1.5 0.8
1.875 1.0
2.25 1.2

X= 0.0m X= 0.107m X= 0.213m X= 0.32m

Figure D.2: Illustrative depiction of the idea of contracting the inlets downstream.

D.2 Constant cross-section

The influence on the average and maximum temperatures, along with the nondimensional coolant mass
flow (as a fraction of the mass flow in the circular smooth bottom panel). Clearly, temperature fields
are virtually independent on the number of inlets downstream. The effect of inlet angle is even smaller.
There seems to be a minimum for the average temperature but the difference is however negligible.
Interface temperatures showed that the decrease in temperature from the inlet angle is strongly localized
exclusively at the respective inlet location. The profile everywhere else was virtually identical.

1
0.275 = 30°
= 45°
0.27
= 60° 0.5
0.265 = 90°

0.26
0
1 2 3

0.5

0.48

0.46

1 2 3

Figure D.3: Effect of number of inlets and inlet angle on average, maximum temperatures and coolant
consumption (nondimensionalized).

Pumping more air downstream does not influence cooling performance (when the outlet number is
kept constant). Longitudinal gradients are actually introduced on the BC and cold side which will likely
have a negative/neutral effect on lifing. Improvement on average temperatures is negligible.

D.3 Adjusted cross-section

The results for each conrtaction factor and hydraulic diameter are shown below. The effect of area
reduction becomes more pronounced for larger hydraulic diameters when it comes to ṁ∗ . Same phe-
nomenon is observed for the average BC temperature, θBC . The maximum temperature near the exit
frame is reduced when flow is ‘more’ throttled’ (i.e smaller areas), thus the correlation b/w ṁ∗ and
max
θBC : the area-adjustment factor effect becomes negligible for small channel diameters, similarly to the
maximum temperature where the increase is also higher for larger diameter. However, the maximum
temperature is in fact lower, with a lower mass flow, which is the main benefit from this study. The

Energy & Process Technology Page 133

 Delft University of Technology

influence of area reduction however is still negligible for θBC when it comes to small hydraulic diam-
eter, dh,aux = 0.7. Additionally, the maximum difference between the obtained values for the average
temperatures are within the range of only 0.01 in nondimensional terms.

2 0.5
d *h,aux = 0.7

0.48 0.28 d *h,aux = 0.9

1.5
d *h,aux = 1

0.46
1 0.275
0.44

0.5
0.42 0.27

0 0.4
0.4 0.5 0.6 0.7 0.8 0.9 0.4 0.5 0.6 0.7 0.8 0.9

Figure D.4: Influence of the amount of contraction on the average, maximum temperatures and coolant
consumption (nondimensionalized).

(a) Fine mesh (b) Coarse mesh

Figure D.5: Kriging fits of numerical data: (a) nondimensionalized mass flow and (b) average bond
coating temperature.

Referring back to the constant channel length, the results prove that unless the channel length is
reduced, i.e. include at least more channel outlet, the achievable reduction in temperature and mass
flow from such artificial throttling is still limited. The benefit of reducing the maximum temperature
near the turbine inlet however, presents a promising path into a potential full (both mass-flow- and
temperature-wise) HT enhancement of the cooling system in the future.

Energy & Process Technology Page 134

Appendix E

More on numerical methodology

E.1 Interpolation methods & coupling

E.1.1 Interpolation stencils
Interpolation stencil properties and values from STAR-CCM+ Interpolation, as previously mentioned,
is required at the fluid-solid interface, where data is mapped from the FVM fluid mesh to the FEM
solid mesh and vice versa. Data interpolation is not solely limited to interface calculations however:
remeshing operations require interpolation of existing solution data onto a new mesh, or additionally,
mapping external (tabular) data onto a mesh, such as the case in the decoupled HT approach. The
available interpolation algorithms in STAR-CCM+ include
• Nearest Neighbor Interpolation
• Least Squares Interpolation
• Exact Imprinting Interpolation
• Approximate Imprinting Interpolation
• Shape function Interpolation
The choice of interpolation method is determined by the type of mesh from which data is interpolation
(source stencil) to the mesh points on which the data is interpolated (target stencil), as shown in the
table below.

Source stencil
Vertex Face Cell
Shape Function
Least Squares Least Squares
Vertex Least Squares
Nearest Neighbor Nearest Neighbor
Target stencil Nearest Neighbor
Least Squares
Shape Function
Nearest Neighbor Least Squares
Face Least Squares
Exact Imprinting Nearest Neighbor
Nearest Neighbor
Approximate Imprinting
Shape Function
Least Squares Least Squares
Cell Least Squares
Nearest Neighbor Nearest Neighbor
Nearest Neighbor

Table E.1: Different interpolation methods in accordance to the source and target stencils.

The available methods are briefly explained in the subsections to follow.

135
 Delft University of Technology

It is also worthwhile mentioning the mechanism used for connecting the mesh faces on the two sides
of the interface. Two options are given within the code: (i) compact and (ii) imprinted. The compact
scheme builds a connectivity map by finding the nearest neighbour of a face on the opposite side of
the interface, and then sets the interpolation weights based on the face area ratios from the map. The
imprinted scheme uses a connectivity map by finding all the faces on the opposite side of the interfaces
that overlap with a face. The method however, is less tolerant to mesh irregularities and poor mesh
quality, compared to the compact method.

In addition, the Explicit energy coupling was used for both laminar and turbulent offset jet studies,
in which the fluid interface boundary is automatically set to a temperature thermal boundary condition
(i.e. static temperature), while the solid side of the CHT interface uses a convection thermal boundary
condition. When the Laminar Flow model is selected, either the Mapped Local Heat Transfer Coefficient
method (for no time averaging) is used or the Mapped Averaged Local Heat Transfer Coefficient method
(for time averaging). When the Turbulent Flow model is selected, the default is set to the Mapped
Specified Y+ Heat Transfer Coefficient method.

Nearest Neighbour Interpolation

The nearest neighbor method sets the data on a target face to the data on the nearest source face. In
the following diagram, the centroid of the source face k is the closest point to the centroid of the target
face n. The data from face k is assigned to face n as the mapped value.

Figure E.1: Nearest neighbour interpolation method illustration.

Since the interface tolerance is manually set, the search algorithm is activated and tolerances that the
algorithm uses to find the neighbors to the cells have to be specified. The following properties are set:

• Proximity Check - when activated, the minimum proximity check occurs.

• Minimum Proximity - the relative distance (based on cell size) to the target cell. 1 is equivalent to
100%.
• Normal Directions check - when activated, the normal direction check occurs.

• Minimum Angle - the angle, in degrees, between the source and target face normals. When the
minimum angle is 0, the two normals point in the same direction, when 180, the normals point
toward each other.
The values used are tabulated below.

Energy & Process Technology Page 136

 Delft University of Technology

Proximity check Enabled

Minimum proximity 0.05
Normal directions check Enabled
Minimum angle 110.0

Table E.2: Values and settings of the mapped interface tolerance

Least Squares Interpolation

A cost function is implemented in the least squares method to approximate the solution field distribu-
tion near a target location. In the following diagram, the solution field has known values fi at the face
centroids i on the source stencil.

Figure E.2: Least square interpolation schematic.

The solution of the solution field is approximated to find the mapped value at the target point a by
first-order Taylor’s expansion:

f (x) ≈ f (xa ) + (x − xa ) · ∇f (xa ) . (B.1a)

The sum of the squared of the residuals to calculate f (a):

X
S= ri2 , (B.1b)
i∈F (c)

where the residual is the difference between the predicted value at the source locations i and their known
values fi
ri = f (xi ) − fi . (B.1c)

Exact Imprinting Interpolation

The exact imprinting name comes from the fact that the target cell face is imprinted on the source
mesh. The value at the target a is hence found from the resultant (from the imprint operation) normalized
facelet areas as interpolation weights: X
f a Aa = Ai f i . (B.1d)
i∈F (c)

The difference with the least square method comes from the fact that all faces on the imprint surface
contribute, as opposed to the least square interpolation where only the cells neighbouring the cell closest
to the target cell centroid contribute (i.e. 2, 3, 4, 5, 6, and 8 in Fig. E.2)

Energy & Process Technology Page 137

 Delft University of Technology

Approximate Imprinting Interpolation

The method is a variation of the exact imprint method, which makes it less computationally expensive.
The difference is that the method uses a proximity method instead of an exact intersection of the cells.
A drawback of the method is that it is not flux-preserving, and is hence only recommended if the target
mesh is considerably coarser than the source mesh.

Shape Function Interpolation

This interpolation method is only available for triangle and quadrilateral elements in 2D and tetra-
hedral and hexahedral elements in 3D for a source mesh, mapping to a FVM mesh. To evaluate the
mapped value at the target point n, the solution is interpolated at the vertices of element k using nodal
shape functions N (k): X
fn = fl Nl (ξn , ηn , χn ) , (B.1e)
l∈N (k)

where Nl is the shape function at vertex l, fl is the solution value at vertex l, and denotes the local
coordinates (ξn , ηn , χn of point n in the element k. The sum is over the set of vertices of element k,
N (k).

E.1.2 CHT coupling algorithms

Several explicit coupling algorithms are available according to literature. Explicit coupling accentuates
on the need for interpolation between the two sides of the interface. However, it allows for conjugate
heat transfer between two different mesh types (e.g. FVM for fluid and FEM for solid region), whereas
the implicit energy coupling option is only available when both source an target meshes are FVM, and
preferably the time-scales of both solid and fluid regions are similar. The drawback of the implicit energy
coupling is that in order to calculate the thermal stresses in a solid, the component has to be meshed
twice. In the subsections to follow, the most commonly used coupling algorithms are briefly discussed.

FFTB
The naming convention of all methods follow the quantities transferred relative to the fluid domain.
In the flux forward temperature back method (FFTB), the wall temperature distribution is imposed to
the fluid solver and the resulting heat flux distribution is imposed as a boundary condition to the solid
conduction solver. The heat flux imposition on the solid then predicts a new temperature for the fluid
region. The procedure is then iterated until convergence (i.e. the heat flux across the interface meets
the specified convergence criteria). Hence, at the i−th iteration:
i i
f em f vm
qwall = qwall (B.1f)
The subsequent interation then interpolates the previosuly obtained FEM temperature on the FVM side:
i+1 i
f vm f em
Twall = Twall (B.1g)
For stability, an alternative formulation of the method can also employ under-relaxation factors. The
equations of this techniques are omitted in this report.

TFFB
The temperature forward flux back method (TFFB) is the inverse method in which the heat flux distri-
bution is imposed as a boundary condition for the fluid computation and the resulting wall temperature
to the solid conduction solver. The updated heat flux is then returned as a boundary condition to the
fluid solver:
f vmi f emi
qwall = qwall (B.1h)
i+1 i
f em f vm
Twall = Twall (B.1i)
Convergence is obtained when the exchanged temperatures and heat fluxes are not changing anymore.

Energy & Process Technology Page 138

 Delft University of Technology

hFTB
The third explicit coupling uses Newton’s law of cooling to calculate the HTC to update the bound-
ary conditions on the solid FEM side and then reimpose the newly found temperature on the fluid
FVM side.The method is thus named the the heat transfer coefficient forward temperature back method
(hFTB).

The method is initialized with an imposed temperature on the FVM boundary. The adiabatic flow
solver then estimates the local HTC and the ambient fluid temperature according to
qwall = h · (Twall − Tf luid )
Thereby, an implicit relation is provided between Twall and qwall that can be used as a boundary condition
for the solid conduction computation. By using this method, the heat flux at the wall is automatically
adjusted by the wall temperature which is the advantage of this algorithm. The boundary condition
after the i−th iteration are hence:
 
f vmi f vmi
qwall = h Twall − Tfi luid (B.1j)
 
f emi f emi
qwall = h Twall − Tfi luid (B.1k)
i+1 i
f vm f em
Twall = Twall (B.1l)

Figure E.3: Flow chart of the hFTB method.

The first equation is used to compute, for a fixed value of h, the value of Tf luid in function of the value
f vm f vm
Twall and qwall wall defined by the fluid computation. The second equation is the boundary condition

Energy & Process Technology Page 139

 Delft University of Technology

f em f em
for the conduction calculation in the solid. This results in a new Twall and qwall on the solid wall.
The third equation defines the boundary condition for the next iteration of the fluid computation. The
flowchart above presents schematically the hFTB algorithm.

hFFTB
The heat transfer coefficient forward flux back method (hFFTB) is an alternative to HFTB where,
instead of temperature, the quantity returned to the FVM domain can be a heat flux. This results in a
new method with different stability properties and is called the hFFB method. The boundary conditions
can be re-written as:  
f vmi f vmi
qwall = h Twall − Tfi luid (B.1m)
 
f emi f emi
qwall = h Twall − Tfi luid (B.1n)
i+1 i
f vm f em
qwall = qwall (B.1o)

Figure E.4: Flow chart of the hFFTB method.

The flowchart above presents schematically the hFFTB algorithm.

Energy & Process Technology Page 140

 Delft University of Technology

E.2 More on turbulence models

E.2.1 Derivation of the RANS equations
In Chapter 4, the derivation of the Reynolds stress term, which arises due to the non-linearity of
the convective term was not included. This subsection of the appendix addresses the derivation of
the new term in the RANS equations, by showing explicitly the time-averaging procedure followed. In
this subsection Cartesian tensor notation is used as it provides a short-hand that is widely used in
fluid dynamics, especially in the context of classical turbulence analyses. Hence, the velocity vector for
example is expressed as
→
−v = (u, v, w)T = ui , where i = 1, 2, 3 (B.1p)
The NS equations (for an incompressible Newtonian fluid) are hence expressed as (the expression of the
divergence and gradient operators using the Cartesian tensor notation, one can refer to in any standard
turbulence textbook):
∂ui
= 0, (B.1q)
∂xi
∂ui ∂ui 1 ∂p ∂ 2 ui
+ uj = fi − +ν . (B.1r)
∂t ∂xj ρ ∂xi ∂xj ∂xj
Splitting the instantaneous field quantities (pressure, velocity, and body forces) into a mean and fluctu-
ating component u(x, t) = ū(x) + u′ (x, t), plugging into the original set of NS equations yields:

∂ (ūi + u′i )
= 0, (B.1s)
∂xi

∂ (ūi + u′i )
∂ (ui + u′i )
1 ∂ (p̄ + p′ ) ∂ 2 (ūi + u′i )

+ ūj + u′j = f¯i + fi′ − +ν . (B.1t)
∂t ∂xj ρ ∂xi ∂xj ∂xj
and knowing that the mean of the fluctuating component u′i = 0, and the double average of ui = ui :

∂ ūi ∂ ūi 1 ∂ p̄ ∂ 2 ūi ∂u′i u′j

+ ūj = f¯i − +ν − . (B.1u)
∂t ∂xj ρ ∂xi ∂xj ∂xj ∂xj

Then, using the definition of the mean-strain tensor, S, expressed in Cartesian tensor notation
 
1 ∂ ūi ∂ ūj
S = Sij = + , (B.1v)
2 ∂xj ∂xi

the equation becomes

∂ ūi ∂ ūi ∂ h i
ρ + ρūj = ρf¯i + −p̄δij + 2µS̄ij − ρu′i u′j , (B.1w)
∂t ∂xj ∂xj

where δij is the Kronicker delta function. The Reynolds averaged conservation, momentum and energy
equations, hence, in vector form (the default notation of this report) are:

∂ρ
+ ∇ · (ρv) = 0, (B.1x)
∂t
∂
(ρv) + ∇ · (ρv ⊗ v) = −∇ · p̄I + ∇ · τ + fb , (B.1y)
∂t
∂
(ρĒ) + ∇ · (ρĒv) = −∇ · p̄v + ∇ · τ v − ∇ · q + fb v. (B.1z)
∂t

Energy & Process Technology Page 141

 Delft University of Technology

E.2.2 Realizable k − ε Damping function

The damping function used in the realizable k − ε model of the coefficient Cµ follows the standard
formulation found in literature. It is expressed as a function of mean flow and turbulence properties, so
that certain mathematical constraints on the normal stresses are consistent with the physics of turbulence
(hence term realizable).

E.2.3 Two-layer modification of the k − ε model

When the blended wall function approach is used, a two-layer model modification is used and solves
for k but prescribes ε algebraically with distance from the wall in the viscosity dominated near-wall flow
regions. The dissipation rate near the wall is simply prescribed as
k 3/2
ε= , (B.1)
lε
where lε is the length scale function calculated according to the Wolfstein variant:
  
Red
cl d 1 − exp − , (B.1)
2cl
−3/4
with d being the wall normal distance and cl = 0.42Cµ . As per the suggestion of Jongen, a wall-
proximity indicator is used to combine the two-layer formulation with the full two-equation model, such
that:   
1 Red − Re∗y
λ= 1 + tanh , (B.1)
2 A
with Re∗y = 60, and A = 4.35. More importantly, a blending function for the turbulent viscosity on the
wall is used to implement the formulation according to Wolfstein such that:
 
µt
µt = λµt |k−ϵ + (1 − λ)µ , (B.1)
µ 2−layer
    
µt Red
= 0.42 Red Cµ1/4 1 − exp − (B.1)
µ 2−layer 70

E.2.4 Shear and stretching modification in the k − ω SST model

The blending functions F1 and F2 in Eq.4.33-4.34, needed in the formulation of the turbulent viscosity
and production of ω are defined as:
" √ ! !#4 
k 500v 2k
F1 = tanh  min max , , 2 , (B.1)
0.09ωd d2 ω d CDkω
 √ !!2 
2 k 500v
F2 = tanh  max , . (B.1)
β ∗ ωd d2 ω


with d being the wall-normal distance and CDkω = max ω1 ∇k · ∇ω, 10−20 is the cross-diffusion coeffi-
cient. By definition the SST model includes free shear-modifications including:
(
1 for χk ≤ 0
fβ ∗ = 1+680χ2k , (B.1)
1+400χ2
for χk > 0
k

which is often referred to as the dissipation limiter, and cross-diffusion limiter:

ρ ∂k ∂ω
CDω = σd , (B.1)
ω ∂y ∂y

Energy & Process Technology Page 142

 Delft University of Technology

where (
∂y ∂y ≤ 0
0, ∂k ∂ω
σd = 1 ∂k ∂ω (B.1)
8 , ∂y ∂y > 0

E.3 CFD roughness

The effect of added wall roughness is implemented in the wall functions by using the wall roughness
function f . The roughness parameter R+ is first defined as:
rρu∗
R+ = (B.1)
µ
with r being the sandgrain roughness height used in the computation. The roughness function f used is
then calculated given three conditions:

a ; R ≤ Rsmooth
+ +

   1 
 +
R −Rsmooth
+
+ +
f= B R+ −R + CR+ ; Rsmooth < R+ < Rrough (B.1)

+


rough smooth
+
B + CR+ ; R+ > Rrough

The exponent a is simply 

+ +


π log R /R
a = sin   smooth
 (B.1)
2 log R+ /R +
rough smooth

+ +
and the model coefficients C = 0.253, Rsmooth = 2.25, and Rrough = 90.

Energy & Process Technology Page 143

Bibliography

[1] A. Predvoditelev A. V. Luikov, A. Pomerantsev, and V. Bubnov. “Heat and Mass Transfer (Hand-
book)”. In: International Journal of Heat and Mass Transfer 16 (1973), pp. 1062–1063.
[2] Theodore Allen. Introduction to Engineering Statistics and Lean Six Sigma: Statistical Quality
Control and Design of Experiments and Systems. Jan. 2019. isbn: 978-1-4471-7419-6. doi: 10.
1007/978-1-4471-7420-2.
[3] F. E. Ames and R. J. Moffat. Heat Transfer With High Intensity, Large Scale Turbulence: The
Flat Plate Turbulent Boundary Layer and the Cylindrical Stagnation Point. 1990.
[4] Forrest E. Ames. “Turbulence Effects on Convective Heat Transfer”. In: Handbook of Thermal
Science and Engineering (2018), pp. 391–423. doi: 10.1007/978-3-319-26695-4_17.
[5] Mark Anderson and Patrick Whitcomb. “Design of Experiments”. In: Sept. 2010. isbn: 9780471238966.
doi: 10.1002/0471238961.0405190908010814.a01.pub3.
[6] Mehdi Bahador, T Nilsson, and Bengt Sundén. “On heat load calculations in gas turbine com-
bustors”. eng. In: Computational Studies. Vol. 5. WIT Press, 2004, pp. 345–357. doi: 10.2495/
HT040321. url: http://dx.doi.org/10.2495/HT040321.
[7] J. W. Baughn et al. “Local Heat Transfer Downstream of an Abrupt Expansion in a Circular
Channel With Constant Wall Heat Flux”. In: Journal of Heat Transfer 106.4 (1984), pp. 789–796.
doi: 10.1115/1.3246753.
[8] C. E. Baukal and B. Gebhart. “A Review of Flame Impingement Heat Transfer Studies Part
2: Measurements”. In: Combustion Science and Technology 104.4-6 (1995), pp. 359–385. doi:
10.1080/00102209508907728.
[9] C.E. Baukal and B. Gebhart. “Surface condition effects on flame impingement heat transfer”.
In: Experimental Thermal and Fluid Science 15.4 (1997), pp. 323–335. issn: 0894-1777. doi:
https://doi.org/10.1016/S0894-1777(97)00036-8. url: http://www.sciencedirect.com/
science/article/pii/S0894177797000368.
[10] R. A. Baurle. “Hybrid Reynolds-Averaged/Large-Eddy Simulation of a Cavity Flameholder: Mod-
eling Sensitivities”. In: AIAA Journal 55.2 (2017), pp. 524–543. doi: 10.2514/1.J055257. eprint:
https://doi.org/10.2514/1.J055257.
[11] T.L. Bergman et al. Fundamentals of Heat and Mass Transfer. Wiley, 2011. isbn: 9780470501979.
[12] Dimitris Bertsimas and John N. Tsitsiklis. Introduction to linear optimization. Athena Scientific,
1997.
[13] L. Bittner. “R. Bellman, Adaptive Control Processes. A Guided Tour. XVI + 255 S. Princeton, N.
J., 1961. Princeton University Press. Preis geb. $ 6.50”. In: ZAMM - Journal of Applied Mathemat-
ics and Mechanics / Zeitschrift für Angewandte Mathematik und Mechanik 42.7‐8 (1962), pp. 364–
365. doi: https://doi.org/10.1002/zamm.19620420718. eprint: https://onlinelibrary.
wiley.com/doi/pdf/10.1002/zamm.19620420718. url: https://onlinelibrary.wiley.com/
doi/abs/10.1002/zamm.19620420718.

144
 Delft University of Technology

[14] J. Bons. “A Critical Assessment of Reynolds Analogy for Turbine Flows ”. In: Journal of Heat
Transfer 127.5 (2005), pp. 472–485. issn: 0022-1481. doi: 10.1115/1.1861919. eprint: https:
//asmedigitalcollection.asme.org/heattransfer/article-pdf/127/5/472/5734392/472\
_1.pdf. url: https://doi.org/10.1115/1.1861919.
[15] R.s. Bunker. “Innovative gas turbine cooling techniques”. In: Thermal Engineering in Power Sys-
tems WIT Transactions on State of the Art in Science and Engineering (2008), pp. 199–229. doi:
10.2495/978-1-84564-062-0/07.
[16] H.l. Cao and J.l. Xu. “Thermal performance of a micro-combustor for micro-gas turbine system”.
In: Energy Conversion and Management 48.5 (2007), pp. 1569–1578. doi: 10.1016/j.enconman.
2006.11.022.
[17] Luca Casarsa, Murat Çakan, and Tony Arts. “Characterization of the Velocity and Heat Transfer
Fields in an Internal Cooling Channel With High Blockage Ratio”. In: Jan. 2002. doi: 10.1115/
GT2002-30207.
[18] Marco Cavazzuti. Optimization methods from theory to design ; scientific and technological aspects
in mechanics. Springer, 2013.
[19] Tuncer Cebeci and Peter Bradshaw. In: Physical and Computational Aspects of Convective Heat
Transfer (1988). doi: 10.1007/978-1-4612-3918-5.
[20] Subhash Chander and Anjan Ray. “Flame impingement heat transfer: A review”. In: Energy
Conversion and Management 46.18-19 (2005), pp. 2803–2837. doi: 10.1016/j.enconman.2005.
01.011.
[21] Yuming Chen, Frederik Arbeiter, and Georg Schlindwein. “A Comparative Study of Turbulence
Models for Conjugate Heat Transfer to Gas Flow in a Heated Mini-Channel”. In: Numerical Heat
Transfer, Part A: Applications 61.1 (2012), pp. 38–60. doi: 10.1080/10407782.2012.638524.
eprint: https://doi.org/10.1080/10407782.2012.638524. url: https://doi.org/10.1080/
10407782.2012.638524.
[22] Cooling Structure Optimization for a Rib-Roughed Channel in a Turbine Rotor Blade. Vol. Volume
3B: Heat Transfer. Turbo Expo: Power for Land, Sea, and Air. V03BT11A009. June 2013. doi: 10.
1115/GT2013- 94527. eprint: https://asmedigitalcollection.asme.org/GT/proceedings-
pdf/ GT2013 / 55157 / V03BT11A009 / 4225298 / v03bt11a009 - gt2013 - 94527 . pdf. url: https :
//doi.org/10.1115/GT2013-94527.
[23] Young Cho. “Handbook of Heat Transfer”. In: Jan. 1997.
[24] G. S. Corman et al. “Rig and Engine Testing of Melt Infiltrated Ceramic Composites for Com-
bustor and Shroud Applications”. In: Journal of Engineering for Gas Turbines and Power 124.3
(2002), pp. 459–464. doi: 10.1115/1.1455637.
[25] Riccardo Da Soghe et al. “Thermo Fluid Dynamic Analysis of a Gas Turbine Transition-Piece”.
In: June 2014. doi: 10.1115/GT2014-25386.
[26] P. A. Dellenback, D. E. Metzger, and G. P. Neitzel. “Heat Transfer to Turbulent Swirling Flow
Through a Sudden Axisymmetric Expansion”. In: Journal of Heat Transfer 109.3 (1987), pp. 613–
620. doi: 10.1115/1.3248132.
[27] Tarek Echekki and Epaminondas Mastorakos. Turbulent combustion modeling: Advances, new
trends and perspectives. Vol. 95. Jan. 2011. isbn: 978-94-007-0411-4. doi: 10.1007/978-94-007-
0412-1.
[28] E.R.G. Eckert, R.M.J. Drake, and R.M. Drake. Heat and Mass Transfer. International student
edition. McGraw-Hill, 1959.
[29] Vishnuvardhanarao Elaprolu and Manab Das. “Conjugate heat transfer study of incompressible
turbulent offset jet flows”. In: Heat and Mass Transfer 45 (July 2009). doi: 10.1007/s00231-
009-0486-9.
[30] F. A. Ficken. Simplex method of linear programming. Dover Publications Inc., 2015.

Energy & Process Technology Page 145

 Delft University of Technology

[31] Chen Fu, Mesbah Uddin, and Alex Curley. “Insights derived from CFD studies on the evolution
of planar wall jets”. In: Engineering Applications of Computational Fluid Mechanics 10.1 (2016),
pp. 44–56. doi: 10.1080/19942060.2015.1082505.
[32] Hideo Futami, Ryoichi Hashimoto, and Hiroshi Uchida. “The development of new catalyst and
heat transfer design method for steam reformer. (I).” In: Journal of the Fuel Society of Japan 68.3
(1989), pp. 236–243. doi: 10.3775/jie.68.236.
[33] VDI Gesellschaft. VDI-Wärmeatlas -. Wiesbaden: Springer Berlin Heidelberg, 2005. isbn: 978-3-
540-25503-1.
[34] D.E. Goldberg. “Genetic algorithms in search, optimization, and machine learning”. In: Choice
Reviews Online 27.02 (1989). doi: 10.5860/choice.27-0936.
[35] Graham Goldin et al. “HEEDS Optimized HyChem Mechanisms”. In: Volume 4B: Combustion,
Fuels and Emissions (2017). doi: 10.1115/gt2017-64407.
[36] P Gosselin, A De Champlain, and D Kretschmer. “Prediction of wall heat transfer for a gas
turbine combustor”. In: Proceedings of the Institution of Mechanical Engineers, Part A: Journal
of Power and Energy 213.3 (1999), pp. 169–180. doi: 10.1243/0957650991537527.
[37] J. Griswold. Fuels, Combustion, and Furnaces. Chemical engineering series. McGraw-Hill Book
Company, Incorporated, 1946.
[38] U. Gruschka et al. “ULN System for the New SGT5-8000H Gas Turbine: Design and High Pressure
Rig Test Results”. In: Jan. 2008. doi: 10.1115/GT2008-51208.
[39] Mohammad R. Hajmohammadi and S. S. Nourazar. “Conjugate Forced Convection Heat Transfer
From a Heated Flat Plate of Finite Thickness and Temperature- Dependent Thermal Conduc-
tivity”. In: Heat Transfer Engineering 35.9 (2014), pp. 863–874. doi: 10.1080/01457632.2014.
852896.
[40] Je-Chin Han, Sandip Dutta, and Srinath Ekkad. Gas turbine heat transfer and cooling technology.
Taylor Francis, 2013.
[41] Je-Chin Han, J. Joy Huang, and Ching-Pang Lee. “Augmented Heat Transfer in Square Channels
with Wedge-Shaped and Delta-Shaped Turbulence Promoters”. In: Journal of Enhanced Heat
Transfer 1.1 (1993), pp. 37–52. doi: 10.1615/jenhheattransf.v1.i1.40.
[42] M. N. A. Hawlader, S. K. Chou, and K. J. Chua. “Development of design charts for tunnel dryers”.
In: International Journal of Energy Research 21.11 (1997), pp. 1023–1037. doi: 10.1002/(sici)
1099-114x(199709)21:11<1023::aid-er309>3.0.co;2-a.
[43] G. He, Y. Guo, and A. Hsu. “The effect of Schmidt number on turbulent scalar mixing in a
jet-in-crossflow”. In: International Journal of Heat and Mass Transfer 42 (1999), pp. 3727–3738.
[44] High Intensity Combustors - Steady Isobaric Combustion. Wiley-VCH Verlag GmbH Co. KGaA,
2002. isbn: 9783527277315,9783527601998. url: http://gen.lib.rus.ec/book/index.php?
md5=6ba76199f62ca5e2413baecff2e26103.
[45] J. Hoch and L. M. Jiji. “Two-Dimensional Turbulent Offset Jet-Boundary Interaction”. In: Journal
of Fluids Engineering 103.1 (1981), pp. 154–161. doi: 10.1115/1.3240766.
[46] M. Hoffmeister. “Townsend, A. A., The Structure of Turbulent Shear Flow, Second Edition,
Cambridge. Cambridge University Press. 1976. 429 S., £ 15.50 A. Cambridge Monographs on
Mechanics and Appl. Mathematics”. In: ZAMM - Journal of Applied Mathematics and Mechanics
/ Zeitschrift für Angewandte Mathematik und Mechanik 56.9 (1976), pp. 448–448. doi: https:
//doi.org/10.1002/zamm.19760560921. eprint: https://onlinelibrary.wiley.com/doi/
pdf/10.1002/zamm.19760560921. url: https://onlinelibrary.wiley.com/doi/abs/10.
1002/zamm.19760560921.
[47] J.T. Holland and J.A. Liburdy. “Measurements of the thermal characteristics of heated offset jets”.
In: International Journal of Heat and Mass Transfer 33.1 (1990), pp. 69–78. issn: 0017-9310. doi:
https://doi.org/10.1016/0017-9310(90)90142-H.

Energy & Process Technology Page 146

 Delft University of Technology

[48] Matjaz Hriberšek et al. “Numerical computation of turbulent conjugate heat transfer in air heater”.
In: vol. 51. Jan. 2005, pp. 470–475. doi: 10.1615/ICHMT.2004.IntThermSciSemin.980.
[49] J. E. Hustad and O. K. Sonju. “Heat Transfer to Pipes Submerged in Turbulent Jet Diffusion
Flames”. In: Heat Transfer in Radiating and Combusting Systems (1991), pp. 474–490. doi: 10.
1007/978-3-642-84637-3_30.
[50] L. Ingber. “Simulated annealing: Practice versus theory”. In: Mathematical and Computer Mod-
elling 18.11 (1993), pp. 29–57. doi: 10.1016/0895-7177(93)90204-c.
[51] Elizaveta M Ivanova, Berthold E Noll, and Manfred Aigner. “A numerical study on the turbulent
Schmidt numbers in a jet in crossflow”. In: Journal of Engineering for Gas Turbines and Power
135.1 (2013).
[52] Muhammad Javed, Naseem Irfan, and Muhammad Ibrahim. “Combustion Kinetic Modeling”. In:
July 2010. isbn: 9783527628148. doi: 10.1002/9783527628148.hoc007.
[53] Chandra Laksham Vaidyaratna Jayatilleke. “The influence of Prandtl number and surface rough-
ness on the resistance of the laminar sub-layer to momentum and heat transfer”. In: 1966.
[54] Bibin John, P. Senthilkumar, and Sreeja Sadasivan. “Applied and Theoretical Aspects of Conju-
gate Heat Transfer Analysis: A Review”. In: Archives of Computational Methods in Engineering
26.2 (2018), pp. 475–489. doi: 10.1007/s11831-018-9252-9.
[55] W. P. Jones and B. Launder. “The prediction of laminarization with a two-equation model of
turbulence”. In: International Journal of Heat and Mass Transfer 15 (1972), pp. 301–314.
[56] Shariatzadeh Joneydi. “Analytical solution of conjugate turbulent forced convection boundary
layer flow over plates”. In: Thermal Science 20.5 (2016), pp. 1499–1507. doi: 10.2298/tsci140115062j.
[57] Charles E. Baukal Jr. “Heat Transfer in Industrial Combustion”. In: (2000). doi: 10 . 1201 /
9781420039757.
[58] B.A. Kader. “Temperature and concentration profiles in fully turbulent boundary layers”. In:
International Journal of Heat and Mass Transfer 24.9 (1981), pp. 1541–1544. issn: 0017-9310.
doi: https://doi.org/10.1016/0017-9310(81)90220-9. url: https://www.sciencedirect.
com/science/article/pii/0017931081902209.
[59] L. V. Kantorovich. “Mathematical Methods of Organizing and Planning Production - English
Transcript”. In: Management Science 6.4 (1960), pp. 366–422. doi: 10 . 1287 / mnsc . 6 . 4 . 366.
eprint: https://doi.org/10.1287/mnsc.6.4.366. url: https://doi.org/10.1287/mnsc.6.
4.366.
[60] J. S. Kapat, A. K. Agrawal, and T. Yang. “Air Extraction in a Gas Turbine for Integrated
Gasification Combined Cycle (IGCC): Experiments and Analysis”. In: Journal of Engineering for
Gas Turbines and Power 119.1 (1997), pp. 20–26. doi: 10.1115/1.2815551.
[61] J. S. Kapat et al. “Cold Flow Experiments in a Sub-Scale Model of the Diffuser-Combustor Section
of an Industrial Gas Turbine”. In: Volume 3: Coal, Biomass and Alternative Fuels; Combustion
and Fuels; Oil and Gas Applications; Cycle Innovations (1996). doi: 10.1115/96-gt-518.
[62] R. Karvinen. “Note on conjugated heat transfer in a flat plate”. In: Letters in Heat and Mass Trans-
fer 5.3 (1978), pp. 197–202. issn: 0094-4548. doi: https://doi.org/10.1016/0094-4548(78)
90005-X. url: http://www.sciencedirect.com/science/article/pii/009445487890005X.
[63] Reijo Karvinen and Antti Lehtinen. “Analytical Solution For A Class Of Flat Plate Conjugate
Convective Heat Transfer Problems”. In: Frontiers in Heat and Mass Transfer 2.4 (2011). doi:
10.5098/hmt.v2.4.3004.
[64] William Morrow Kays. Convective heat and mass transfer. Tata McGraw-Hill Education, 1993.
[65] J.H. Keenan and J. Kaye. Thermodynamic Properties of Air, Including Polytropic Functions.
Wiley, 1945.

Energy & Process Technology Page 147

 Delft University of Technology

[66] J.C Ku and K.H. Shim. “The effects of refractive indices, size distribution, and agglomeration on
the diagnostics and radiative properties of flame soot particles”. In: Heat and Mass Transfer in
Fires and Combustion Systems 148.3 (1997), pp. 105–115.
[67] L.D. Landau and E.M. Lifshits. Fluid Mechanics, by L.D. Landau and E.M. Lifshitz. Teoretich-
eskai�a�fizika. Pergamon Press, 1959.
[68] S. C. Lau, R. D. Mcmillin, and J. C. Han. “Heat Transfer Characteristics of Turbulent Flow
in a Square Channel With Angled Discrete Ribs”. In: Volume 4: Heat Transfer; Electric Power;
Industrial and Cogeneration (1990). doi: 10.1115/90-gt-254.
[69] S. C. Lau, R. D. Mcmillin, and J. C. Han. “Turbulent Heat Transfer and Friction in a Square
Channel With Discrete Rib Turbulators”. In: Journal of Turbomachinery 113.3 (1991), pp. 360–
366. doi: 10.1115/1.2927884.
[70] B.E. Launder and B.I. Sharma. “Application of the energy-dissipation model of turbulence to
the calculation of flow near a spinning disc”. In: Letters in Heat and Mass Transfer 1.2 (1974),
pp. 131–137. issn: 0094-4548. doi: https://doi.org/10.1016/0094-4548(74)90150-7. url:
https://www.sciencedirect.com/science/article/pii/0094454874901507.
[71] A. H. Lefebvre and M. V. Herbert. “Heat-Transfer Processes in Gas-Turbine Combustion Cham-
bers”. In: Proceedings of the Institution of Mechanical Engineers 174.1 (1960), pp. 463–478. doi:
10.1243/pime_proc_1960_174_039_02.
[72] Arthur H. Lefebvre. “Design Considerations In Advanced Gas Turbine Combustion Chambers”.
In: Combustion in Advanced Gas Turbine Systems (1968), pp. 3–19. doi: 10.1016/b978-0-08-
013275-4.50007-6.
[73] Arthur H. Lefebvre. “Flame radiation in gas turbine combustion chambers”. In: International
Journal of Heat and Mass Transfer 27.9 (1984), pp. 1493–1510. doi: 10.1016/0017-9310(84)
90262-x.
[74] Arthur H. Lefebvre and Dilip R. Ballal. “Gas Turbine Combustion”. In: (2010). doi: 10.1201/
9781420086058.
[75] J. H. Lienhard IV and J. H. Lienhard V. A Heat Transfer Textbook. 5th. Mineola, NY: Dover
Publications, Dec. 2019. 784 pp. isbn: 9780486837352. url: http://ahtt.mit.edu.
[76] Timothy C. Lieuwen and Vigor Yang. Combustion instabilities in gas turbine engines operational
experience, fundamental mechanisms and modeling. eng. Progress in astronautics and aeronautics
; v. 210. Reston, Va: American Institute of Aeronautics and Astronautics. isbn: 1-60086-680-8.
[77] Phil Ligrani. “Heat Transfer Augmentation Technologies for Internal Cooling of Turbine Com-
ponents of Gas Turbine Engines”. In: International Journal of Rotating Machinery 2013 (2013),
pp. 1–32. doi: 10.1155/2013/275653.
[78] Matti Lindstedt and Reijo Karvinen. “Conjugate heat transfer in a plate – One surface at constant
temperature and the other cooled by forced or natural convection”. In: International Journal of
Heat and Mass Transfer 66 (Nov. 2013), pp. 489–495. doi: 10.1016/j.ijheatmasstransfer.
2013.07.052.
[79] Mordor Intelligence LLP. “Industrial Gas Turbine Market - Growth, Trends, and Forecast (2020-
2025)”. In: May 2020.
[80] Morten Løvbjerg. “Improving particle swarm optimization by hybridization of stochastic search
heuristics and self-organized criticality”. PhD thesis. 2002.
[81] A.V. Luikov. “Conjugate convective heat transfer problems”. In: International Journal of Heat and
Mass Transfer 17.2 (1974), pp. 257–265. issn: 0017-9310. doi: https://doi.org/10.1016/0017-
9310(74)90087-8.
[82] A.V. Luikov. “Heat and Mass Transfer in Capillary-Porous Bodies”. In: Oxford: Pergamon, 1966,
pp. 305–340. isbn: 978-1-4832-0065-1. doi: https://doi.org/10.1016/B978- 1- 4832- 0065-
1.50011-8.

Energy & Process Technology Page 148

 Delft University of Technology

[83] Wall Temperature Effects on Heat Transfer Coefficient. Vol. Volume 3C: Heat Transfer. Turbo
Expo: Power for Land, Sea, and Air. V03CT14A003. June 2013. doi: 10.1115/GT2013-94291.
eprint: https://asmedigitalcollection.asme.org/GT/proceedings- pdf/GT2013/55164/
V03CT14A003/2420664/v03ct14a003- gt2013- 94291.pdf. url: https://doi.org/10.1115/
GT2013-94291.
[84] S. V. Mahmoodi-Jezeh and Bing-Chen Wang. “Direct numerical simulation of turbulent flow
through a ribbed square duct”. In: Journal of Fluid Mechanics 900 (2020), A18. doi: 10.1017/
jfm.2020.452.
[85] A.F. Mills. Heat Transfer. Irwin, 1992, pp. 402–404. isbn: 9780256076424.
[86] Narendernath Miriyala, Anthony Fahme, and Mark Van Roode. “Ceramic Stationary Gas Tur-
bine Program: Combustor Liner Development Summary”. In: Volume 4: Manufacturing Materials
and Metallurgy; Ceramics; Structures and Dynamics; Controls, Diagnostics and Instrumentation;
Education; IGTI Scholar Award (2001). doi: 10.1115/2001-gt-0512.
[87] M. Mosaad. “Laminar forced convection conjugate heat transfer over a flat plate”. In: Heat and
Mass Transfer 35.5 (1999), pp. 371–375. doi: 10.1007/s002310050338.
[88] J. A. Nelder and R. Mead. “A Simplex Method for Function Minimization”. In: The Computer
Journal 7.4 (1965), pp. 308–313. doi: 10.1093/comjnl/7.4.308.
[89] W. A. Nelson and R. M. Orenstein. “TBC experience in land- based gas turbines”. In: Journal of
Thermal Spray Technology 6.2 (1997), pp. 176–180. doi: 10.1007/s11666-997-0009-5.
[90] P. Y. Nizou. “Heat and Momentum Transfer in a Plane Turbulent Wall Jet”. In: Journal of Heat
Transfer 103.1 (1981), pp. 138–140. doi: 10.1115/1.3244407.
[91] Gordon C. Oates. The Aerothermodynamics of aircraft gas turbine engines. National Technical
Information Service, 1978.
[92] Mohamed Hatim Ouahabi, Mohamed Ichenial, and F. Benabdelouahab. “Evaluation of the tur-
bulence models at low Reynolds number for wind turbine blade design”. In: Aug. 2017.
[93] Jun Su Park et al. “Thermal Analysis of Cooling System in a Gas Turbine Transition Piece”. In:
vol. 5. Jan. 2011. doi: 10.1115/GT2011-45961.
[94] Sunil Patil et al. “Study of Flow and Convective Heat Transfer in a Simulated Scaled Up Low
Emission Annular Combustor”. In: Journal of Thermal Science and Engineering Applications 3.3
(2011). doi: 10.1115/1.4004531.
[95] J. R. R. Pelfrey and J. A. Liburdy. “Mean Flow Characteristics of a Turbulent Offset Jet”. In:
Journal of Fluids Engineering 108.1 (Mar. 1986), pp. 82–88. issn: 0098-2202. doi: 10.1115/1.
3242548.
[96] T.L. Perelman. “On conjugated problems of heat transfer”. In: International Journal of Heat and
Mass Transfer 3.4 (1961), pp. 293–303. issn: 0017-9310. doi: https://doi.org/10.1016/0017-
9310(61)90044-8.
[97] Siemens PLM.
[98] Jeffrey Price. “Advanced Materials for Mercury 50 Gas Turbine Combustion System”. In: (2008).
doi: 10.2172/991117.
[99] A. Racca, T. Verstraete, and L. Casalino. “Radial Turbine Thermo-Mechanical Stress Optimiza-
tion by Multidisciplinary Discrete Adjoint Method”. In: International Journal of Turbomachinery,
Propulsion and Power 5 (2020). doi: https://doi.org/10.1016/j.ijheatmasstransfer.2016.
05.041.
[100] D. Radaj and M. Vormwald. Ermüdungsfestigkeit: Grundlagen für Ingenieure. Springer Berlin Hei-
delberg, 2007. isbn: 9783540714590. url: https://books.google.bg/books?id=MAgeBAAAQBAJ.
[101] Bürgel Ralf, Maier Hans Jürgen, and Thomas Niendorf. Handbuch Hochtemperatur-Werkstofftechnik
Grundlagen, Werkstoffbeanspruchungen, Hochtemperaturlegierungen und -beschichtungen ; mit 66
Tabellen. Vieweg Teubner, 2011.

Energy & Process Technology Page 149

 Delft University of Technology

[102] D. Ramirez. “Heat Transfer and Flow Measurements in an Atmospheric Lean Pre-Mixed Com-
bustor”. In: 2016.
[103] Takeo Saitou, Yasuyuki Watanabe, and Shouhei Yoshida. Gas turbine combustor including a
transition piece flow sleeve wrapped on an outside surface of a transition piece. May 2012.
[104] Tareq Salameh and Bengt Sunden. “Comparison of Continuous and Truncated Ribs on Internal
Blade Tip Cooling”. In: Volume 4: Heat Transfer, Parts A and B (2012). doi: 10.1115/gt2012-
68028.
[105] K. Schittkowski. “NLPQL: A fortran subroutine solving constrained nonlinear programming prob-
lems”. In: Annals of Operations Research 5.1-4 (1986), pp. 485–500. doi: 10.1007/bf02739235.
[106] H. Schlichting (Deceased) and K. Gersten. “Boundary–Layer Equations in Plane Flow; Plate
Boundary Layer”. In: Boundary-Layer Theory. Berlin, Heidelberg: Springer Berlin Heidelberg,
2017, pp. 145–164. isbn: 978-3-662-52919-5. doi: 10.1007/978-3-662-52919-5_6.
[107] A. Schumacher. Optimierung mechanischer Strukturen: Grundlagen und industrielle Anwendun-
gen. SpringerLink: Springer e-Books. Springer, 2005. isbn: 9783540218876. url: https://books.
google.bg/books?id=PaE%5C_21AYJj4C.
[108] N. Selcuk, R. G. Siddall, and J. M. Beer. “Prediction of the effect of flame length on temperature
and radiative heat flux distributions in a process fluid heater”. In: J. Inst. Fuel; (United Kingdom)
48 (June 1975).
[109] SGT6-8000H: H-class Gas Turbine: Gas Turbines: Manufacturer: Siemens Energy Global. url:
https : / / www . siemens - energy . com / global / en / offerings / power - generation / gas -
turbines/sgt6-8000h.html.
[110] Robert Siegel et al. Thermal Radiation Heat Transfer. CRC Press, Dec. 2020. doi: 10.1201/
9780429327308.
[111] Sin Chien Siw, Mary Anne Alvin, and Minking Chyu. “Heat Transfer Enhancement of Internal
Cooling Passage With Triangular and Semi-Circular Shaped Pin-Fin Arrays”. In: ASME Turbo
Expo 2012: Turbine Technical Conference and Exposition (2012), pp. 493–503. doi: GT2012-69266.
[112] Jacob Snyder et al. “Build Direction Effects on Additively Manufactured Channels”. In: Journal
of Turbomachinery 138 (Dec. 2015). doi: 10.1115/1.4032168.
[113] Manohar S. Sohal and John R. Howel. “Determination of plate temperature in case of combined
conduction, convection and radiation heat exchange”. In: International Journal of Heat and Mass
Transfer 16.11 (1973), pp. 2055–2066. issn: 0017-9310. doi: https : / / doi . org / 10 . 1016 /
0017 - 9310(73 ) 90108 - 7. url: http : / / www . sciencedirect . com / science / article / pii /
0017931073901087.
[114] M. Soliman and H. A. Johnson. “Transient Heat Transfer for Turbulent Flow Over a Flat Plate
of Appreciable Thermal Capacity and Containing Time-Dependent Heat Source”. In: Journal of
Heat Transfer 89.4 (1967), pp. 362–370. doi: 10.1115/1.3614398.
[115] R.I. Stephens et al. Metal Fatigue in Engineering. A Wiley-Interscience publication. Wiley, 2000.
isbn: 9780471510598. url: https://books.google.bg/books?id=B2aAPVa1TloC.
[116] Red Cedar Technology. 2015.
[117] Tom Verstraete and Sebastian Scholl. “Stability analysis of partitioned methods for predicting
conjugate heat transfer”. In: International Journal of Heat and Mass Transfer 101 (2016), pp. 852–
869. issn: 0017-9310. doi: https://doi.org/10.1016/j.ijheatmasstransfer.2016.05.041.
[118] E. Vishnuvardhanarao and Manab Kumar Das. “Computation of Mean Flow and Thermal Char-
acteristics of Incompressible Turbulent Offset Jet Flows”. In: Numerical Heat Transfer, Part
A: Applications 53.8 (2007), pp. 843–869. doi: 10 . 1080 / 10407780701715760. eprint: https :
//doi.org/10.1080/10407780701715760.
[119] E. Vishnuvardhanarao and Manab Kumar Das. “Study of Conjugate Heat Transfer from a Flat
Plate by Turbulent Offset Jet Flow”. In: Numerical Heat Transfer, Part A: Applications 53.5
(2007), pp. 524–542. doi: 10.1080/10407780701678331.

Energy & Process Technology Page 150

 Delft University of Technology

[120] R Viskanta and M.P. Menguc. “Radiation heat transfer in combustion systems”. In: Progress in
Energy and Combustion Science 13.2 (1987), pp. 97–160. doi: 10.1016/0360-1285(87)90008-6.
[121] M. Vynnycky et al. “Forced convection heat transfer from a flat plate: the conjugate problem”. In:
International Journal of Heat and Mass Transfer 41.1 (1998), pp. 45–59. issn: 0017-9310. doi:
https://doi.org/10.1016/S0017-9310(97)00113-0. url: http://www.sciencedirect.com/
science/article/pii/S0017931097001130.
[122] Liang Wang and Ting Wang. “Investigation of the Effect of Perforated Sheath on Thermal-Flow
Characteristics Over a Gas Turbine Reverse-Flow Combustor: Part 1 — Experiment”. In: Volume
3C: Heat Transfer (2013). doi: 10.1115/gt2013-94474.
[123] Liang Wang and Ting Wang. “Investigation of the Effect of Perforated Sheath on Thermal-Flow
Characteristics Over a Gas Turbine Reverse-Flow Combustor: Part 2 — Computational Analysis”.
In: Volume 3C: Heat Transfer (2013). doi: 10.1115/gt2013-94475.
[124] T. Wang et al. “Effect of Air Extraction for Cooling and/or Gasification on Combustor Flow
Uniformity”. In: Volume 3: Coal, Biomass and Alternative Fuels; Combustion and Fuels; Oil and
Gas Applications; Cycle Innovations (1998). doi: 10.1115/98-gt-102.
[125] Yu Wen-Shing, Lin Hsiao-Tsung, and Hwang Tsung-Yuan. “Conjugate heat transfer of conduction
and forced convection along wedges and a rotating cone”. In: International Journal of Heat and
Mass Transfer 34.10 (1991), pp. 2497–2507. issn: 0017-9310. doi: https://doi.org/10.1016/
0017 - 9310(91 ) 90091 - R. url: http : / / www . sciencedirect . com / science / article / pii /
001793109190091R.
[126] Victor Yakhot, Steven A. Orszag, and Alexander Yakhot. “Heat transfer in turbulent fluids - I.
Pipe flow”. In: International Journal of Heat and Mass Transfer 30.1 (1987), pp. 15–22. issn:
0017-9310. doi: https://doi.org/10.1016/0017-9310(87)90057-3.
[127] Mehmet Yilmaz et al. “Heat transfer and friction characteristics in decaying swirl flow generated
by different radial guide vane swirl generators”. In: Energy Conversion and Management 44.2
(2003), pp. 283–300. doi: 10.1016/s0196-8904(02)00053-5.
[128] F. Yin and A. Gangoli Rao. “Performance analysis of an aero engine with inter-stage turbine
burner”. In: The Aeronautical Journal 121.1245 (2017), pp. 1605–1626. doi: 10.1017/aer.2017.
93.
[129] Wen-Shing Yu and Hsiao-Tsung Lin. “Conjugate problems of conduction and free convection on
vertical and horizontal flat plates”. In: International Journal of Heat and Mass Transfer 36.5
(1993), pp. 1303–1313. issn: 0017-9310. doi: https : / / doi . org / 10 . 1016 / S0017 - 9310(05 )
80099-7. url: http://www.sciencedirect.com/science/article/pii/S0017931005800997.
[130] Parviz Mohammad Zadeh and Mohadeseh Alsadat Sadat Shirazi. “5 - Multidisciplinary Design
and Optimization Methods”. In: Metaheuristic Applications in Structures and Infrastructures.
Ed. by Amir Hossein Gandomi et al. Oxford: Elsevier, 2013, pp. 103–127. isbn: 978-0-12-398364-
0. doi: https : / / doi . org / 10 . 1016 / B978 - 0 - 12 - 398364 - 0 . 00005 - X. url: http : / / www .
sciencedirect.com/science/article/pii/B978012398364000005X.
[131] Yan Zhao et al. “Numerical and Experimental Study of Geometry Effects on Fuel/Air Mixing and
Combustion Characteristics of a DLN Burner”. In: Oct. 2020. doi: 10.1115/POWER2020-16371.
[132] Dadong Zhou, Ting Wang, and William R. Ryan. “Cold Flow Computations for the Diffuser-
Combustor Section of an Industrial Gas Turbine”. In: Volume 3: Coal, Biomass and Alternative
Fuels; Combustion and Fuels; Oil and Gas Applications; Cycle Innovations (1996). doi: 10.1115/
96-gt-513.

Energy & Process Technology Page 151