Alma Mater Studiorum - Università degli Studi di Bologna

DOTTORATO DI RICERCA IN

MECCANICA E SCIENZE AVANZATE

DELL'INGEGNERIA
Ciclo XXVI
Settore Concorsuale di aerenza:

09/C1 INGEGNERIA DELLE MACCHINE E DEI SISTEMI ENERGETICI

Settore Scientico disciplinare:

ING-IND/08 MACCHINE A FLUIDO

Towards Centrifugal Compressor

Stages Virtual Testing
Presentata da:

Dott. Ing. Emanuele Guidotti

Relatore

Chiar.mo Prof. Ing. Giovanni Naldi

Coordinatore Dottorato

Chiar.mo Prof. Ing. Vincenzo Parenti Castelli

Esame nale anno 2014

 A thesis submitted to the
Graduate School
of the University of Bologna
in partial fulllment of the
requirements for the degree of

Ph.D.

in the Department of Industrial Engineering

of the College of Engineering

By

Emanuele Guidotti
B.S., Mechanical Engineering
University of Bologna, Italy, July 2006
M.Sc., Aerospace Engineering
University of Cincinnati, U.S.A., July 2008

Committee Chair: Professor Giovanni Naldi

`Essentially, all models are wrong, but some are useful.'

George E. P. Box
 iv

ABSTRACT

Requirements for centrifugal compressors performance continue to increase. To this

intent, prediction tools need to be at the state of the art to assess performance variations

between dierent designs and/or to predict, within tight tolerance, single stage or multi-

stage machines for design verication. Furthermore, the complex ow phenomena inside the

dierent components of centrifugal compressor stages need to be fully understood to allow

aerodynamicists to design new stages with increasing performance but at the same time to

produce reliable and robust machines.

However, ow features inside centrifugal compressor stages are very complicated

to simulate with numerical tools due to the highly complex geometry and varying gas condi-

tions all across the machine. For this reason, a big eort is currently being made to increase

the delity of the numerical models during the design and validation phases. Computa-

tional Fluid Dynamics (CFD) plays an increasing role in the assessment of the performance

prediction of centrifugal compressor stages. Historically, CFD was considered reliable for

performance prediction on a qualitatively level, whereas tests were necessary to predict

compressors performance on a quantitatively basis. In fact "standard" CFD with only the

ow-path and blades included into the computational domain is known to be weak in captur-

ing eciency level and operating range accurately due to the under-estimation of losses and

the lack of secondary ows modeling. This research project aims to ll the gap in accuracy

between "standard" CFD and tests data by including a high delity reproduction of the gas

domain and the use of advanced numerical models and tools introduced in the author's OEM

in-house CFD code. In other words, this thesis describes a methodology by which virtual

tests can be conducted on single stages and multistage centrifugal compressors in a similar

fashion to a typical rig test that guarantee end users to operate machines with a condence

level not achievable before. Furthermore, the new "high delity" approach allowed under-

standing ow phenomena not fully captured before, increasing aerodynamicists capability

and condence in designing high eciency and high reliable centrifugal compressor stages.
 vi

ACKNOWLEDGMENTS

When a journey is completed, one of the pleasure is to look over the past and

remember all the people who have been present and supported me along this long but

exciting road.

It is very dicult to overstate my gratitude to my Ph.D. advisor Prof. Giovanni

Naldi who helped to make turbomachinery and CFD fun for me. He always enabled me to

face the hard work with a smile on the lips. He has been an exceptionally good leader and

motivator. He has provided me encouragement, good teaching and lots of good ideas. I am

really glad that I have come to get know you in my life.

I love all the professors I have deal with at both University of Bologna and Uni-

versity of Cincinnati but a special mention goes to Prof. Orkwis and Prof. Turner. I owe a

debt of gratitude to both of you.

I'll always be grateful to GE Oil&Gas Nuovo Pignone for letting me publish all

the material presented in this thesis and for the nancial support. I'll be forever in debt

to my former and current managers at GE OIL&GAS Nuovo Pignone Libero Tapinassi and

Tommaso Dante Rubino for their sound advices and their constructive and critical spirit.

Thanks to my friends, which are too many to mention (a thesis has a page size

maximum, guys). So because the list might be too long and by fear of leaving someone out,

I will simply say thank you very much to you all. Some of you, anyway, are quite lucky:

thank you Pietro, Sandro and Andrea.

My deepest gratitude goes to my parents Rolando and Aurora and to my brother

Christian and to Zietta and all my relatives. They bore me, raised me and loved me. They

have provided everything that one could hope for in a family. And in the end my parents

has taught me everything that I truly need to know in life.

I dedicate this thesis to Paola and Martina. This thesis is yours too, and so the

spirit behind it, which in the end is everything. Whenever I try to describe my feelings for

you...no matter what I come up with, it only seems to capture half of the story.


 vii

Contents
List of Figures 1

List of Tables 4

Nomenclature 5

1 Introduction 8
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2 Test Cases Description 12

2.1 Centrifugal Compressors Classication . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Test Cases Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3 Experimental Setup and Data 16

3.1 Test Rigs Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.1.1 Single Stage Test Rig Description . . . . . . . . . . . . . . . . . . . . . 16
3.1.2 Type 2 and String Test Rigs Description . . . . . . . . . . . . . . . . . 17
3.2 Measurement Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.3 FRAPP Probe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

4 Numerical Models 24
4.1 Cavity Modeling in Centrifugal Compressors . . . . . . . . . . . . . . . . . . . 24
4.2 Computational Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.3 Computational Grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.4 Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.5 Solver Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

5 Numerical Results and Comparison with Test Data 34

5.1 Inuence of Numerical Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.2 Inuence of Boundary Conditions and Gas Model . . . . . . . . . . . . . . . . 42
5.3 Inuence of Geometry Fidelity Reproduction . . . . . . . . . . . . . . . . . . 43
5.4 Inuence of Leakage Flows Modeling . . . . . . . . . . . . . . . . . . . . . . . 44
5.5 Inuence of Surface Roughness . . . . . . . . . . . . . . . . . . . . . . . . . . 47
 viii

5.6 Multistage Eects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

5.7 Inuence of Unsteady Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

6 Flow Analysis 55
6.1 Impeller Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6.2 Diuser Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6.3 Return Channel Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.4 Cavity Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

7 Conclusions 83
7.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
7.2 Recommendations and Future Work . . . . . . . . . . . . . . . . . . . . . . . 84

Bibliography 85

A List of Publications 90
 1

List of Figures
2.1 Test cases on plane Mu-Φ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Test cases on plane Tin − Pin . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

3.1 Test Rig setup in intermediate stage conguration. . . . . . . . . . . . . . . . 18

3.2 Test Rig setup in axial inlet conguration. . . . . . . . . . . . . . . . . . . . . 18
3.3 Type 2 centrifugal compressors test facilities at GE OIL&GAS in Florence. . 19
3.4 String test centrifugal compressors test facilities at GE OIL&GAS in Massa. . 20
3.5 FRAPP probe. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.6 FRAPP installed at the exit of a centrifugal impeller. . . . . . . . . . . . . . . 23

4.1 CFD domain including cavities for single stage test rig. . . . . . . . . . . . . . 28
4.2 CFD Forward and back sides cavities grids before merging. . . . . . . . . . . 28
4.3 Cavity domain details shroud side. . . . . . . . . . . . . . . . . . . . . . . . . 29
4.4 Cavity domain details hub side. . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.5 Two stages computational domain. . . . . . . . . . . . . . . . . . . . . . . . . 30

5.1 Overall stage performance comparison between test and steady CFD for stage
ow coecient 0.0444. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.2 Computational domain for stage ow coecient 0.0444. . . . . . . . . . . . . 35
5.3 Overall stage performance comparison between test and high delity CFD for
stage ow coecient 0.0444 at design and o-design speeds. . . . . . . . . . . 36
5.4 Computational domain for stage ow coecient 0.0095. . . . . . . . . . . . . 37
5.5 Overall stage performance comparison between test and steady CFD for stage
ow coecient 0.0095. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.6 Computational domain for stage ow coecient 0.1600. . . . . . . . . . . . . 38
5.7 Overall stage performance comparison between test and steady CFD for stage
ow coecient 0.1600. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.8 Impeller trailing edge grid details and EGR contours. . . . . . . . . . . . . . . 40
5.9 Leakage ows eciency decrement across the full range of ow coecient stages. 46
5.10 Leakage ows work coecient eect with respect to no cavity ows across the
full range of ow coecient stages. . . . . . . . . . . . . . . . . . . . . . . . . 46
5.11 Comparison of leakage ow amount between CFD and correlations at dierent
operating conditions for the stage ow coecient 0.0444. . . . . . . . . . . . . 47
5.12 Computational domain for a reycling wheel in an ammonia process. . . . . . . 49
 2

5.13 Details of the back-to-back conguration. . . . . . . . . . . . . . . . . . . . . 49

5.14 Comparison between predicted performance from CFD and test results for a
reycling wheel in an ammonia process. . . . . . . . . . . . . . . . . . . . . . . 50
5.15 Streamlines for the back-to-back system coloured by velocity. . . . . . . . . . 50
5.16 Multistage centrifugal compressor for LNG application. . . . . . . . . . . . . . 51
5.17 Comparison between predicted performance from CFD and test results for
section 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.18 Computational domain for an over-hung impeller plus vaned diuser. . . . . . 53
5.19 Steady and unsteady CFD predictions and test data for an over-hung impeller
plus vaned diuser. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

6.1 FRAPP and CFD total pressure 2D map at section 20 comparison at design
ow rate. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
6.2 FRAPP and CFD total pressure 2D map at section 20 comparison close to
surge. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
6.3 FRAPP and CFD total pressure 2D map at section 20 comparison close to
choke. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
6.4 FRAPP and CFD YAW angle 2D map at section 20 comparison at design
ow rate. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
6.5 FRAPP and CFD YAW angle 2D map at section 20 comparison close to surge. 58
6.6 FRAPP and CFD YAW angle 2D map at section 20 comparison close to choke. 59
6.7 Yaw angle comparison between FRAPP and dierent CFD approaches at
impeller exit. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
6.8 Total pressure comparison between FRAPP and dierent CFD approaches at
impeller exit. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
6.9 Average tangential total pressure prole comparison between FRAPP and
high delity CFD. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
6.10 Average tangential yaw angle prole comparison between FRAPP and high
delity CFD. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6.11 Tangential total pressure prole comparison between FRAPP and high delity
CFD at 10% of the span. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6.12 Tangential total pressure prole comparison between FRAPP and high delity
CFD at 50% of the span. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
6.13 Tangential total pressure prole comparison between FRAPP and high delity
CFD at 90% of the span. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
6.14 FRAPP and CFD tangential averaged yaw angle at section 20 comparison
close to surge. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.15 FRAPP and CFD tangential averaged yaw angle at section 20 comparison
close to choke. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.16 Non dimensional total pressure proles at section 40 at design ow rate. . . . 66
6.17 Yaw angle proles at section 40 at design ow rate. . . . . . . . . . . . . . . . 66
6.18 Non dimensional static pressure proles at section 40 at design ow rate. . . . 67
6.19 Meridional velocity contours inside the diuser close to surge operating limit. 67
6.20 Meridional velocity contours inside the diuser at design operating point. . . 68
6.21 Meridional velocity contours inside the diuser close to choke operating limit. 68
 3

6.22 Yaw angle proles at diuser exit from CFD close to surge, design and close
to choke operating conditions. . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.23 comparison of average yaw angle at diuser exit between CFD and test close
to surge, design and close to choke operating conditions. . . . . . . . . . . . . 69
6.24 Flow angle prole at diuser inlet for high delity and standard CFD at
dierent operating points. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
6.25 Contours of radial velocity inside the diuser after the mixing plane position
for the high delity and standard CFD respectively. . . . . . . . . . . . . . . . 71
6.26 Static pressure recovery coecient inside the diuser for the CFD with and
without leakage ows included across the stage operating range. . . . . . . . . 72
6.27 Circumferential averaged meridional velocity contours from standard CFD
(left) and high delity CFD (right) at design point. . . . . . . . . . . . . . . . 73
6.28 Circumferentially averaged meridional velocity contours from standard CFD
(left) and high delity CFD (right) close to stall point. . . . . . . . . . . . . . 74
6.29 Circumferentially averaged meridional velocity contours from standard CFD
(left) and high delity CFD (right) close to choke point. . . . . . . . . . . . . 75
6.30 Component losses from CFD with and without cavities modeling at design
point. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.31 Total pressure loss coecient in return channel for test and high delity CFD. 76
6.32 Contours of entropy with streamlines super-imposed for the front cavity. . . . 78
6.33 Contours of entropy with streamlines super-imposed for the front cavity labyrinth
seals. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.34 Axial velocity distribution inside the forward cavity. . . . . . . . . . . . . . . 79
6.35 Radial velocity distribution inside the forward cavity. . . . . . . . . . . . . . . 80
6.36 Circunferential velocity distribution inside the forward cavity. . . . . . . . . . 80
6.37 Static pressure variation across the shroud cavity for CFD and experimental
data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.38 Static pressure variation across the hub cavity for CFD and experimental data. 81
6.39 Position of the static pressure probes inside both the cavities at hub and
shroud for a typical case. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
 4

List of Tables
3.1 Measurement uncertainties. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Percentage dierences between FRAPP and 5-Hole Probe . . . . . . . . . . . 22
 5

Nomenclature
Symbol Units Description
Cp − Static pressure recovery coecient

D m Diameter

ṁ kg/s Mass ow rate

Mu − Peripheral Mach number

P Pa Pressure

Q m3 /s Volumetric ow rate

R J/(kg ∗ K) Gas constant

T K Temperature

U m/s Peripheral speed

V m/s Absolute velocity

Y AW degree Absolute ow angle

y+ − Dimensionless wall distance

Z − Compressibility factor

Greek Symbols
Symbol Units Description
 m2 /s3 Turbulent dissipation

Φ − Flow coecient

γ − Ratio of specic heats

η − Eciency

κ m2 /s2 Turbulent kinetic energy

µ Pa ∗ s Viscosity
 6

θ − Circumferential direction

ρ kg/m3 Density

τ − Work coecient

ω 1/s Specic dissipation rate

ξ − Total pressure loss coecient

ψ − Head coecient

Subscripts and superscripts

in Inlet component

M Meridional component

out Outlet component

p Polytropic

r Radial component

s Static quantities

t Total quantities

1 Impeller inlet

2 Impeller exit

Acronyms
BPF Blade Passing Frequencies

CAD Computer Aided Design

CAES Compressed Air Energy Storage

CCS Carbon Capture and Sequestrations

CFD Computationa Fluid Dynamics

CTL Coal To Liquids

DNS Direct Numerical Simulation

EGR Entropy Generation Rate

EOR Enhanced Oil Recovery

FCC Fluid Catalytic Cracking

FFT Fast Fourier Transform

FRAPP Fast Responce Aerodynamics Pressure Probe

 7

GGI General Grid Interface

GMBS Global Multi-Block Surfaces

GTL Gas To Liquids

IGCC Integrated Gasication Combined Cycle

IGV Inlet Guide Vanes

JST Jameson-Schmidt-Turkel

KL Kato-Launder

LES Large Eddy Simulation

OEM Original Equipment Manufacturer

RANS Reynolds Averaged Navier Stokes

URANS Unsteady Reynolds Averaged Navier Stokes

SST Shear Stress Transport

 8

Chapter 1

Introduction
In the recent years, Computational Fluid Dynamics (CFD), has evolved enormously

and it is now used extensively by turbomachinery OEMs in designing the varoius compo-

nents.

Flow phenomena inside centrifugal compressor stages are very complicated to sim-

ulate with numerical tools due to the highly complex geometry and varying gas conditions

all across the machine. For this reason, the accuracy of CFD, for designing and predict-

ing centrifugal compressors performance is currently under a controversial and stimulating

debate.

1.1 Motivation
The idea of this work was born with the intent to push CFD capabilities towards

virtual testing of centrifugal compressor single and multistage machines.

Market trend requires centrifugal compressors OEMs to produce machines with

continuously increasing performance and end users to safely operate the compressors. In

particular, this translates in a constant eort for aerodynamicists to design new stages with

increasing ow capacity and higher peripheral Mach number but at the same time to produce

reliable and robust machines. The accuracy of prediction and design tools is critical for the

purpose.

On one hand, thermodynamic tests on single and multistage centrifugal compres-

sors are commonly conducted in centrifugal compressor OEMs as per API-617 to assess

machines performance. On the other hand, prediction tools need to be at the state of the
 9

art to predict, within extremely tight tolerances, compressor performance and avoid expen-

sive reworking during testing phase.

Historically, CFD was considered reliable for performance prediction on a qualita-

tively level, whereas tests were necessary to predict compressors performance on a quantita-

tively basis. In fact, "standard" CFD with only the ow-path and blades included into the

computational domain is known to be weak in capturing eciency level and operating range

accurately due to the under-estimation of losses and the lack of secondary ows modeling.

In the vision of the author, CFD is nowadays mature enough to represent a reli-

able tool for accurate centrifugal compressor performance prediction when an high delity

reproduction of the geometry and advanced numerical models are used into the simulations.

This research project aims to ll the gap in accuracy between "standard" CFD

and tests data by including a high delity reproduction of the gas domain and the use of

advanced numerical models introduced in the author's OEM in-house CFD code. In other

words, this thesis describes a methodology by which virtual tests can be conducted on single

stages and multistage centrifugal compressors in a similar fashion to a typical test rig that

guarantee aerodynamicists to design machines and end users to operate machines, with a

condence level not achievable before.

1.2 Overview
In the past few decades, signicant developments have been achieved in the nu-

merical models and the associated CFD algorithms. Furthermore, in the author's OEM,

a big eort has been done in the recent years, to develop automatic and robust meshing

tools that allow the aero designers to model additional features not included before into the

computational domain. Combined with the rapid developments in computer hardware in

both speed and memory which are becoming increasingly available at aordable prices, the

simulation of complete test rigs or even full compressors are increasingly becoming a reality.

Several centrifugal compressor single stages, ranging across dierent inlet ow coef-

cients and peripheral Mach numbers, applied to disparate processes (e.g. LNG, pipeliners,

barrel compressors, etc), as well as applications of multistage compressors, were selected as

test cases for this study. Both steady and unsteady simulations were conducted in order

to fully capture the time averaged and time accurate operating conditions. Moreover, de-

tailed geometrical features like llets and leading edge shape as well as the entire secondary
 10

cavity ows system were faithfully reproduced in the computational domain and advanced

numerical models were used in the setup with accuracy not possible before.

CFD predictions were compared with test results regarding both overall perfor-

mance and detailed ow features showing a very good agreement, not found in open lit-

erature for closed impeller stages. Furthermore, the new "high delity" approach allowed

understanding ow phenomena not fully captured before, increasing aerodynamicists capa-

bility and condence in designing high eciency and high reliable centrifugal compressor

stages. The increment in performance prediction accuracy with respect to "standard" CFD

is noticeable. Advantages of this approach is the reduced cost with respect to tests and the

possibility to virtually instrument the machine all along the computational domain where

it is not possible in the real compressor.

This thesis delves into the details of the CFD accuracy by introducing innovative

methodologies and capabilities for tri-dimensional numerical simulations. Several test cases

have been simulated and results will be presented throughtout this work to assess compu-

tational capabilities comparing predictions with test data. Finally a clearer understanding

of ow features inside the centrifugal compessors is claimed.

• Chapter 2 presents the details about the selected test cases and the strategy for nu-

merical validation.

• In chapter 3 the experimental setup and test data are provided as well as a description

of the GE OIL&GAS testing capabilities

• In chapter 4 is presented the calculations setup as well as a detailed explanation of the

theory behind the numerical models and tools used.

• In chapter 5 the results of the present work and the comparison between the exper-

imental data and the numerical simulations are shown. Furthermore, a review of

the implications of geometrical and numerical models with computational accuracy is

presented.

• Chapter 6 shows the details and the explanation of ow phenomena and secondary

ows inside centrifugal compressors stages.

• Finally chapter 7 surveys the results from all the chapters and summarizes them.

Future directions for research and improvements to the current work are also suggested.
 11

All throughout the thesis, a complete literature review of the current state of the art

in CFD simulations for centrifugal compressor stages is presented for each specic argument.
 12

Chapter 2

Test Cases Description

In the OIL&GAS market, centrifugal compressors are used extensively for very

dierent applications and in the most variable operating conditions. Centrifugal compressors

can be classied based on applications, casing types, number of stages, etc. Furthermore,

each single stage is classied based on some adimensional parameters.

Validation of the numerical models in this reasearch work has been conducted

on several dierent geometries, single and multistage compressors, dierent applications,

dierent working gas, dierent operating conditions like inlet pressure and so on.

2.1 Centrifugal Compressors Classication

Currently a separation in upstream, midstream and downstream sectors is often

used to categorize the dierent applications. A rough classication of the dierent sectors

is given here:

• Upstream: oil & gas i.e. gas lift, gas export, gas injection, gas gathering, gas treatment,

gas processing, CO2 injection, LNG, boil o, gas transport, liqueed petroleum gas.

• Midstream: gas transport, gas storage, fuel gas, CAES.

• Downstream: reneries, fertilizers, chemical & petrochemical, i.e. hydrogen produc-

tion, hydrogen recovery, hydro cracking. Desulfurization, FCC, propane dehydration,

methanol, olens, IGCC, Coal-to-Liquids, GTL-Syngas, ammonia, urea, nitric acid.

Industrial gases: air separation, nitrogen, oxygen, paper, coke oven, GTL, CTL.
 13

• Power generation: fuel gas, CCS.

A classication of centrifugal compressors based on the casings is:

• Vertically split casings

• Horizontally split casing

• Barrel compressors

• Overhung/integrally geared compressors

Moreover, the dierent compressors can be single stage, multistage with single or

multiple sections, back to back and so on.

As already highlighted before, gas and operating conditions can vary greatly: inlet

pressure can be from atmosferic to several hundred bar, gas can range from single component,

light gases like hydrogen to mixtures of heavy hydrocarbons.

A common classication for single stage description is given by dening the adi-

mensional parameters Φ (inlet ow coecient) and Mu (peripheral Mach number).

The ow coecient and peripheral Mach number are dened as equations ( 2.1)

and ( 2.2) respectively.

4 × Qt,1
Φ= (2.1)
π × D22 × U2

U2
Mu = p (2.2)
γ × R × Tt,1

Where Qt,1 is Volumetric ow rate (calculated using total density) at impeller inlet.

2.2 Test Cases Selection

As described in the previous paragraph, the applications, working gases, inlet con-

ditions and geometries can vary greatly in centrifugal compressors. For this reason to val-

idate the accuracy of current CFD it was necessary to select dierent cases. Furthermore,

dierent ow phenomena are predominant depending on the stage type; e.g. bi-dimensional

impellers are dominated by leakage and friction losses, instead tri-dimensional high ow co-

ecient stages by curvature and secondary ows losses. To this intent, several single stage
 14

and multistage centrifugal compressors were selected for numerical models validation and

comparison with test data, ranging across a full spectrum of dierent ow coecients and

peripheral Mach numbers. In particular below a summary of the dierent cases is provided:

• Working gas: air, CO2, propane, methane, hydrogen, nitrogen, ammonia, refrigerant

gases, hydrocarbon mixtures.

• Inlet pressure: from atmosheric up to several tens of bar.

• Inlet temperature: from several degrees below zero up to several tens of degrees Cen-

trigade.

• Mach number: from highly subsonic ows up to transonic ones.

• Inlet ow coecient: from very small bi-dimensional impellers up to mixed ow tri-

dimensional ones.

Furthermore, both single stage and multistage machines have been studied, both

over-hung, vaned diusers impellers with inlet guide vanes and vaneless intermediate stages

have been simulated.

Figures 2.1 and 2.2 show the dierent cases used to validate the CFD method-

ology presented here divided respectively on the plane peripheral Mach number-inlet ow

coecient (Mu-Φ) and compressor inlet temperature-inlet pressure (Tin − Pin ) respectively.

It can be seen that the full range of peripheral Mach numbers and inlet ow coecients of

most common industrial applications of centrifugal compressors has been covered. More-

over, if on one hand inlet atmospheric conditions of pressure and temperature are the ones

used the most in single stage validations, Reynolds eect has been also studied with several

validation cases, both on pressurized test rigs and full scale type 2 and string tests.
 15

Figure 2.1: Test cases on plane Mu-Φ.

Figure 2.2: Test cases on plane Tin − Pin .

 16

Chapter 3

Experimental Setup and Data

Depending on the test case, the experimental verication of compressor perfor-

mance was done at the GE Oil&Gas testing facilities either in Florence or Massa. In partic-

ular, in Florence at the Oil&Gas Technology Laboratory (OGTL) all the single stage tests

were performed whereas the type 2 tests were done in full scale test rigs in the same plant.

At Massa testing facilities the string tests were operated.

3.1 Test Rigs Description

All the tests have been performed by following API-617 regulations. In the follow-

ing paragraphs the description of the dierent test rigs will be shown with the overlapping

areas between the dierent test benches setup clearly highlighted.

3.1.1 Single Stage Test Rig Description

A typical single stage test bench, whose test cell is depicted in Fig. 3.1, consists

of a closed loop rotating rig, which is employed for performance measurements of many

centrifugal compressor designs; further details can be found in [1]. The test rig can be used

for both intermediate stage conguration or axial inlet conguration. For the intermediate

stage conguration, a multistage compressor is simulated through a "pseudo-stage" in the

ow-path upstream of the impeller. This is included in order to provide ow proles at

the impeller inlet typical of that expected in a multistage environment. The pseudo-stage

consists of a set of pre-swirl vanes followed by a return channel. Instead for the axial

inlet conguration, the test cell is modied according to Fig. 3.2 where inlet guide vanes
 17

and an outlet scroll replace the upstream pseudo-stage and the downstream return bend

respectively. The test rigs are instrumented so as to allow ange-to-ange measurements,

single component performance evaluation (e.g., impeller eciency and head, diuser recovery

and losses, etc.), as well as the detection of the occurrence of stall and surge (through

dynamic probes installed for this purpose). Measurements are taken at various locations

throughout the test rig. The instrumentation used is well established and can be tailored to

suit the needs of a specic study. For the purpose of this discussion, only the measurement

apparatus of direct relevance to performance evaluation is discussed. Pressure, temperature,

ow angle and velocity measurements were performed at each of the measurement locations

throughout the compressor. According to Fig. 3.1 and Fig. 3.2, the main measurements

sections are indicated as:

• Section 00: Stage inlet/IGV inlet

• Section 10: Impeller inlet

• Section 20: Impeller outlet/diuser inlet

• Section 40: Diuser outlet/return channel or scroll inlet

• Section 60: Stage outlet/return channel or scroll outlet

Single stage test rigs are usually operated in similitude conditions with real ma-

chines due to the impossibility to opearte at the same inlet pressure and with the same

working gas of on site machine.

3.1.2 Type 2 and String Test Rigs Description

Type 2 performance tests is conducted on real machine geometry. It can be consid-

ered as a laboratory-type uid test on real machine to conrm uid dynamic characteristics

of the compressor. It permits the use of a substitute test gas and accepts extensive devia-

tions between test and specied operating conditions. There are only a few limits on some

essential gas dynamic parameters of test conditions (compare to specied operating condi-

tions). Specic volume ratio and ow coecient should be within around 5% deviations.

There are some limits on machine Mach number and machine Reynolds number. The test

speed, capacity, mass ow, pressures, temperatures, power, etc are often totally dierent
 18

Figure 3.1: Test Rig setup in intermediate stage conguration.

Figure 3.2: Test Rig setup in axial inlet conguration.

from the specied operating condition speed. In Type 2 test, a suitable gas is identied

which does not lead to excessive power or discharge temperature and is readily and cheaply

available. Substitute gas such as air, nitrogen, CO2, CO2/He mixes, fuel gas, etc are used.

Safe operating speed, critical speeds, maximum allowable pressures, allowable temperatures

and other machine limits are considered in test condition selection. In type 2 test, test

Reynolds number is dierent compared to specied operating condition but it is still within

certain limits to keep governing friction formulations the same (same model and ow regime).

Based on theory, a correction to the test results is applied based on available gas dynamic

knowledge to estimate the friction eects of compressor performance in specied operating

 19

condition. All correction formulations are available in API-617 for estimation.

Instead a string test type 1 performance test is actually a shop performance test

in anticipated site condition. It is conducted with same gas as site (same gas with molec-

ular weight deviation below 2%). Generally pressure, temperature, compressor speed and

capacity permissible deviations are below around 4-8%.

Type 2 and string tests are, as expected, less instrumented than single stage model

tests. Usually, inlet and outlet sections are fully instrumented and vibration detection is

performed all throughout the test duration. On some special cases, like prototype cases,

compressor is heavily instrumented to have also inter-stage or inter-component measure-

ments whenever possible.

Figures 3.3 and 3.4 show the GE OIL&GAS testing facilities for both type 2 and

string tests on centrifugal compressors, respectively in Florence and Massa.

Figure 3.3: Type 2 centrifugal compressors test facilities at GE OIL&GAS in Florence.

3.2 Measurement Techniques

Measurement techniques and instrumentations are common to all the dierent test

rigs with particular indications for each setup given in the previous sections.

Total pressure measurements are made with Kiel probes, static pressure measure-

ments are made using wall taps (both at the hub and at the shroud) and shielded J-type

thermocouples are used for total temperature measurements. Flow angles are measured ei-

ther with three-hole or ve-hole probes, depending on the expected three-dimensionality of

 20

Figure 3.4: String test centrifugal compressors test facilities at GE OIL&GAS in Massa.

the ow. To minimize the intrusiveness and increase the resolution, axial-radial actuators

and a rotating conveyor (where instrumentation is installed) can be employed. The use of

multiple probes rake or single probe with traversing system depends on the fow channel size;

in fact for narrow channel is preferred to employ one single probe to minimize the aerody-

namics blockage given by the probe itself. Moreover, probe actuators allow both maximum

recovery of Kiel probes and "nulling mode" operation for multi-hole probes. The ow rate

is measured with an orice following the EN ISO 5167-1 standard, while the rig rotating

speed is measured using a magnetic pick-up based key-phasor. The uncertainties at the 95%

condence level associated with the steady measurements of each section are collected in

Table 3.1. The measurements accuracy, evaluated usign a calibrated nozzle, is ±0.5% of the
◦
kinetic head for both the total and static pressure and ±0.2 for the yaw angle.

3.3 FRAPP Probe

The FRAPP (Fast Responce Aerodynamics Pressure Probe) has been recently

introduced in the Oil&Gas Technology Laboratory (OGTL) and is employed in single stage

measurements at the exit of impeller, i.e. section 20, in order to obtain a highly detailed

reconstruction of ow eld. The FRAPP measurement concept arises from the matching of

classical pneumatic directional probes and piezoresistive fast-response pressure transducers.

While the former, like a cobra probe and 5 hole probe, can provide information about an

unknown steady ow eld, the latter oers fast-response, miniaturization, reliability and
 21

Measurement Section Absolute Uncertainty

Temperature 10 ± 0.2
Pressure 10 ± 35 Pa
Yaw Angle 10 ± 0.5 deg
Temperature 20 ± 0.2
Pressure 20 ± 155 Pa
Yaw Angle 20 ± 0.5 deg
Pitch Angle 20 ± 0.5 deg
Temperature 40 ± 0.2
Pressure 40 ± 155 Pa
Yaw Angle 40 ± 0.5 deg
Temperature 60 ± 0.2
Pressure 60 ± 155 Pa
Yaw Angle 60 ± 0.5 deg

Table 3.1: Measurement uncertainties.

low cost. This combination of features allows the unsteady evolution of the ow eld to

be measured. Despite the number of dierent congurations available, the highest degree

of miniaturization is achieved with single-sensor probes, which allow the measurement of

the 2D ow eld in a plane normal to the probe stem. Since the probe has a single hole,

only one measurement at a time can be performed. The ow eld is then reconstructed by

re-arranging three pressure measurements at dierent angles of rotation around the probe

axis, thus virtually reproducing the three-hole probe technique.

The probe considered here was developed around a commercial miniaturized pres-

sure sensor (Kulite XCQ-062, full scale 25 psi) to ensure high reliability, low cost and simpli-

ed manufacturing of the probe heads. The probe concept was developed at the Politecnico

di Milano over the last decade. The transducer is installed co-axially with the probe head

to obtain a minimum probe head diameter of 2.0 mm. The nal probe spatial resolution,

dened as the physical distance between the extreme positions of the tap, is 1.5 mm. The

single pressure tap on the probe head has a diameter of 0.3 mm. Detailed information

regarding FRAPP technology can be found in [2] and [3].

The instantaneous pressure signals is acquired at 1 MHz for a period of 1 second.

Raw pressure data are phase-locked to the rotor wheel and then phase-averaged to obtain

40 intervals on a single rotor-blade passing period (BPP). As a nal step, the ow properties

are derived by combining the dierent phase-averaged pressures.

 22

The unsteady ow quantities, originally measured in the absolute frame, are fur-

ther converted into relative quantities making use of the time-averaged total temperature

measured at station 20. Furthermore, other quantities can be derived, even if under some

assumptions, like the turbulent kinetic energy.

A validation of the FRAPP based technique was successfully carried out in a pre-

vious activity where a comparison between a FRAPP and a pneumatic 5-hole probe was

made [4]. Good agreement was observed in terms of both average values over the entire

ow eld and hub-to-shroud proles as also shown in Table 3.2 where the percentage errors

for the average results are listed. However, no detailed information or detailed ow-eld

reconstruction can be obtained with a 5-hole probe, whereas FRAPP can actually provide

information about the two-dimensional ow eld in terms of total pressure, static pressure

and yaw angle. Thanks to the unsteady character of this techniques, FRAPP data can be

used to derive the "steady" ow eld in the relative frame at the rotor exit.

Measurement Percentage error (FRAPP-5H)/5H

Total Pressure 0.10%

Static Pressure 1.80%
Yaw Angle 0.70%

Table 3.2: Percentage dierences between FRAPP and 5-Hole Probe .

Figure 3.5 shows the dynamic pressure probe, while Fig. 3.6 depicts the FRAPP

installed at the impeller exit in a typical test rig conguration; the red circle shows the

FRAPP location.
 23

Figure 3.5: FRAPP probe.

Figure 3.6: FRAPP installed at the exit of a centrifugal impeller.

 24

Chapter 4

Numerical Models
The three-dimensional and viscous nature of the ow eld inside a modern high

speed centrifugal compressor stage is dicult to simulate with numerical models. Never-

theless, during design phase, is fundamental to correctly predict the ow behavior and the

corresponding performance of the stage to validate the design accurately. Software tools,

ranging from simple one-dimensional to highly complex fully three-dimensional are used to

assess this issue. The advantage of one-dimensional tools is that they are fast and reliable for

an overall screening of the design space and overall performance prediction. However, accu-

rate three-dimensional computational models are necessary to simulate the ow structures

that are not captured by simpler models. The drawback of such complex tools is the exe-

cution time, that especially with the increasing complexity in modeled features, sometimes

can be not aordable during design space exploration. In addition, for radial machines, the

high curvature of streamlines makes the CFD computations even more dicult to simulate

with a good accuracy with respect to axial machines. Many validation cases are available

in open literature on centrifugal compressor stages like [5], [6], [7], [8], [9] to name a few.

However they are mainly related to open impellers or, for closed ones, they lack in capturing

losses generation correctly and so the overall performance of the stages due to a not delity

representation of the real machine and environment.

4.1 Cavity Modeling in Centrifugal Compressors

In closed centrifugal compressor impellers like the ones used as test cases in this

study, a gap remains between diaphragms and impeller trailing edge on both hub and shroud
 25

sides. Due to the high pressure at the trailing edge location those cavities need to be sealed

in order to avoid the ow passing through. The cavity at shroud side is usually connected to

the impeller leading edge and labyrinths are used to reduce as much as possible the leakage.

On the hub side the trailing edge gap is connected via a cavity up to the return channel

trailing edge. Labyrinths are used there as well. In over-hung impellers, the rear cavity

can be sealed or in a back-to-back conguration, is connected with the opposite impeller to

balance the axial thrust. In single stages, the back cavity can also used to balance inlet and

outlet pressure connecting the two parts. In mutistage machine the last impeller is often

connected to the inlet through a balancing line.

Leakage ows play a big role in the correct prediction of ow behavior in centrifu-

gal compressors. Up to few years ago, the modeling of cavities in centrifugal compressor

stages was not adopted and only the main ow-path was simulated. A new approach using

appendages and source terms has been proposed by this author in [10]: this enhanced model

allows to correctly capture overall performance and ow features with an almost negligi-

ble increase in computational and user time with respect to the model without any cavity

included. However, an additional source model needs to be added and a ow network re-

solved by external dedicated tools for each operating point. Furthermore, correlations used

in the tools need to be tuned for dedicated cases; 1D models are know to be weak in correct

windage losses prediction. Moreover the ow features inside the cavities are not simulated

and reproduced. In fact the knowledge of the ow inside the impeller cavities is fundamen-

tal, not only for the aerodynamic designers but also for the mechanical and rotordynamic

assessments of the stage. Axial thrust, aeromechanical behavior, swirl brakes design, seals

design and geometry optimization can be evaluated and predicted by the numerical tools if

the entire cavities are included in the model.

Few studies are available on centrifugal compressors with full cavities modeled in

open literature. Validation studies are mainly related to open impellers. For open im-

pellers, some attempts are reported with the aim to simulate the eects of ow control

devices on stage operating range. In the studies from Hunziker [11] and Tamaki [12], the

ow analysis inside such bleed slots for internal recirculation are shown and the numerical

results with respect to baseline conguration without control devices are discussed. Sun

et al. [13] simulated a backside cavity of an open centrifugal compressor impeller using a

source terms approach. The eects of the hub cavity on the impeller ow eld have been

reported. For closed impeller centrifugal compressor stages the study from Mischo et al.
 26

[14] shows a comparison of dierent shroud cavities geometry and the inuence on impeller

performance as well as comparison of overall performance of the stage with test data. Wang

et al. [15] studied the eects of dierent cavities geometries on overall performance in a low

ow coecient centrifugal compressor stage. In the recent past, the modeling of cavities in

centrifugal compressor stages using appendages and source terms was presented by Guidotti

et al. [10]. The main limitations in this approach is the reliability of tuned 1D tools for

correct prediction of the source terms at each operating point of the stage. The inuence of

with and without full cavity modeling at design point for one ow coecient stage on overall

performance and ow eld was shown in an earlier activity by Guidotti et al. [16] and [17]

and reported a substantial improvement in the accuracy of the prediction with full cavity

numerical model on a centrifugal compressor stage. Satish K. et al. [18], following the same

approach, showed an accurate agreement between advanced test data and numerical predic-

tion on dierent ow coecient centrifugal compressor stages using steady computations.

Lettieri et al [19] also showed a good agreement between test data and CFD with cavity

modeling.

4.2 Computational Setup

Results presented in this thesis were obtained by using the GE in-house CFD code

TACOMA. TACOMA (Turbine And COMpressor Analysis) is a GE proprietary code used

for computational uid dynamics simulations on axial and radial turbomachinery. TACOMA

is a 3D multi-block, multi-grid, structured, non-linear and linear Euler/Navier-Stokes solver

for turbomachinery blade rows. TACOMA is a cell-centered explicit ow solver based on

the so-called JST scheme [20]. Details of the scheme as well as validation cases can be

found in [21] and [22]. The solution is obtained via a multi-step Runge-Kutta explicit time

marching scheme with convergence acceleration via local time steps, residual averaging, and

V-cycle or W-cycle multigrid. In the present analysis, steady and unsteady 3D Reynolds-

Averaged Navier-Stokes (RANS and URANS) equations are solved. Both the two equations

κ−ω turbulence model developed by Wilcox [23] and the variation SST were used in the

computations. TACOMA uses by default the production modication of Kato and Launder

instead of the original one [24]. Also the curvature correction model that was found to

be useful on centrifugal compressor stage analysis as shown by Smirnov et al [7] has been

implemented into the code. Previous validations and modeling experiences with TACOMA
 27

on centrifugal compressor stages can be found in [25], [26], [27], [10], [28], [16], [17] and [29]

4.3 Computational Grid

Two dierent grid domains were meshed for each airfoil. In particular, the geometry

of the compressor main ow-path was modeled with the Numeca grid generator Autogrid by

importing the ow-path and sections at dierent spanwise locations for all the airfoils. The

geometry of the stage was exactly replicated from the test rig including llets. Structured

grids for impeller, diusers and/or return channel were generated. Instead the cavities

domains were meshed by a recently developed in-house GE tools. Finally all the blocks were

merged. The new automated in-house tools for cavity meshing requires a very short amount

of additional user time. The base airfoil grid is linked to a CAD software where the axis-

symmetric shape of the cavity is reproduced; the meshing tools uses the two domains to mesh

the cavity once some topology parameters are provided in a text le. The possibility inside

the new tools to split and merge the dierent blocks allows the user to create very accurate

grids in dierent domains and then manipulate them. Also the possibility to create templates

for dierent centrifugal stages is benecial for repeating designs. The computational domain

includes, depending on the test case, impeller, vaneless or vaned diuser, return channel,

hub cavity and shroud cavity, double inlet, balancing line, etc.

The entire CFD domain including cavities for a typical one stage simulation is

shown in Fig. 4.1. From right to left, it can be seen in sequence the preswirl, deswirl,

impeller and return channel vanes as well as the two cavities connecting impeller trailing

edge to impeller leading edge shroud side and return channel to impeller hub side.

The meshed forward and back sides cavities before merging are shown in Fig. 4.2.

Details of the grid inside the cavities including labyrinth seals at both shroud and

hub sides are shown in Fig. 4.3 and Fig. 4.4 showing the very good grid details including a

correct boundary layer grid clustering.

For multistage applications, the dierent computational stage domains are stacked

in sequence. Proles are exchanged through the dierent rotating/not rotating domains

with tangential averaging of main quantities at interfaces. To have a fairly similar numer-

ical resolution along the span direction and avoid numerical inaccuracies, a similar grid is

targeted on both sides of the interface. An example of a multistage application is showed in

 28

Figure 4.1: CFD domain including cavities for single stage test rig.

Figure 4.2: CFD Forward and back sides cavities grids before merging.

gure 4.5.

The grid size for each airfoil domain is the result of many grid sensitivity studies

and cases validation performed during past years in the author's OEM like the ones in [10],

[16], [17], [18] [25], [26] and [27]. To give a rough estimation of the sensitivity study resuts,

it has been found that below about 200k cells for each airfoil domain the main ow features

are not captured at all. Between 200k and 500k cells only blade loading distribution is well

reproduced. Up to 1000k cells are necessary for a good matching between numerical results
 29

Figure 4.3: Cavity domain details shroud side.

Figure 4.4: Cavity domain details hub side.

and test measurements on exit angles proles. More than 1000k cells are mandatory for

loss generation and mechanisms assessment. A at response of results was found at about

2.5 million of computational cells for each airfoil main ow-path domain. Furthermore the

increment in computational time and memory requirements with included meshed cavities

with respect to only main ow-path computations is about 30%. As already stated, negligible

addition in user time is required thanks to the new in-house automated tools.

Wall integration is used to capture the boundary layer and computational grids
 30

Figure 4.5: Two stages computational domain.

modeled with an average y+ less than 1 on all the wall surfaces to ensure good resolution of

viscous sub layer. In particular, to accurately model boundary layer development inside the

cavities, the in-house GE meshing tool allows inserting an O type grid block at each wall

inside the cavities. TACOMA Global Multi-Block Surfaces (GMBS) interfaces were imposed

at few domain interfaces, where one-to-one interface connections are not modeled. At GMBS

interface locations, the mesh on both sides was created in order to have fairly similar grid

clustering. GMBS (non conformal interfaces) allow not to have a one-to-one connection at

interface and also to realize the grid in the airfoil domain without any distortion especially

in the area close to the impeller trailing edge where the correct resolution of ow features is

mandatory for the correct loss modeling assessment. Airfoil/cavity interface can have blocks

extents and grid counts dierent on the two sides. Dierent patch count, size and shape can

be realized and a user friendly surfaces tagging system has been included. A modication

of the CFD code has been necessary to cope with non matching block interface. A new

advanced system of donor/acceptor cells across the interface with a cell area weighting ux

system has been included. The new feature has been tested inserting non matching blocks

in computational domains simulated previously and comparing the results with older block

matching computations with no appreciable dierences. Then the new system has been

extended to cope with cavity domains.

 31

Particular attention was also paid in the main ow domain to correctly dening the

mesh close to the leading and trailing edges and to ensuring that enough cells were placed

in the passage to capture the main structures of the ow.

4.4 Boundary Conditions

For accurate representation of test setup in CFD computations, the CFD model

inlet and outlet locations are maintained same as the experimental setup and the ow

proles at the inlet of the test campaign or previous plenum CFD simulations are applied

at the CFD domain inlet and outlet domains. In particular, tangential averaged proles

of total quantities and angles were used at compressor inlet section. Mass ow rate or

static pressure conditions were applied at the outlet depending on numerical stability at the

dierent operating conditions. In particular, static pressure at exit can lead to numerical

divergence (static instability) close to stall point, so mass ow rate outlet is used close to

stall operating condition and static pressure outlet is applied close to choke condition. All

the walls are modeled as no-slip and adiabatic. As the heat inside the cavities is dissipated

by leakage ows, there is no necessity to use isothermal boundary conditions at cavity walls

that is instead necessary when the uid is stagnant inside the cavities like in the case of a

backward cavity in over-hung applications where an o-ring is mounted at the exit.

Numerical computations are performed on single airfoil for the steady computa-

tions with periodic boundary conditions. A phase lag boundary condition is used in the

tangential direction for the unsteady runs. The algorithm implemented in the GE in-house

code TACOMA makes use of the phase-lag boundary conditions introduced by Erdos [30].

The key idea is based on the phase shift methodology, i.e. the assumption that the solution

at any particular passage in the blade row can be related to the solution at another passage

at an earlier time in the same blade row. This allows modeling of rotor-stator interaction

using only a single blade passage in each blade row without the necessity to change blade

count between adjacent blade rows and is useful for time-accurate analyses dominated by

unsteadiness at the adjacent blade passing frequency. This allows the unsteady simulation

to be performed by using only one blade from each of the two rows.

Ideal, linearly variable, or even real gas models have been used in the computations

based on the operating pressure and gas type used. Air, carbon dioxide, refrigerant gas,

propne, etc. have been setup in the models. For real gas applications, thermodynamics
 32

tables of specic heat, ratio of specic heats and viscosity are linked to the in-house CFD

code Tacoma and accessed during the computation. The Benedict-Webb-Rubin equation in

the version modied by Starling (BWRS) was used to create the gas table and to calculate

the performance.

4.5 Solver Convergence

Complex geometries, multistage coupling, blade-row interactions and many other

factors tends to stive CFD convergence capabilities for single and in particular for multistage

centrifugal compressor simulations. Furthermore, the introduction of cavity models inside

the computational domain represents a challenge for numerical stability due in particular

to the very dierent Mach number between main ow-path and cavities domains that can

be even two/three times. GMBS interfaces are used at few domain interfaces, where the

nodes on the two sides are not aligned. GMBS interface in TACOMA is modeled to conserve

the uxes across the interfaces. Fluxes on both sides of the interfaces were checked (in the

converged solution) and no variations were found up to several decimals. Furthermore for

the best results of the interpolation algorithm, the mesh on both sides was made in order

to have fairly similar length scales and both sides of the interface fully overlap.

Coupling of main-ow and cavities often produces numerical oscillations that can

persist regardless the number of iterations used, expecially at operating conditions where

unsteady eects are revelant. In these cases average scalar variables need to be monitored

until they reach a stable level. Local adjustments of the grid and/or numerical models

and time scale need to be used in order to reach a good convergence and overcome the

instabilities due to the dierent velocities between main-ow path and cavities.

Flow in the cavities travels mostly in the circumferential direction which means that

the signal responding to main ow eld change takes long time to propagate from one opening

of the cavity to the other and as a whole this induces a lower convergence rate with respect

to computational model without. Regarding steady computations both mathematical and

physical convergence was monitored during the runs and the simulations were stopped when

main ow quantities reached a stable pattern. In the unsteady runs Courant number was

constantly monitored during the transient computations and time step has been decided to

have the average value in an acceptable range, i.e. below 10. However, due to the dierence

in grid details and velocities inside the dierent computational domains, a sensible variation
 33

between cavity domains and main ow is expected. In fact the introduction of cavity models

inside the computational domain represents a challenge for numerical stability, particularly

due to very high dierence in Mach number between main ow path and cavities. The

average Mach number between main ow path and cavity domains can vary up to 2 to 3

times. Local adjustments of the grid and/or numerical models need to be used in order

to reach a good convergence and overcome the instabilities due to the dierent velocities

between main ow-path and cavities. For each time step, convergence has been achieved with

main ow quantities residuals down up to 4/5 orders of magnitude. Moreover, the steady

converged solution was used to initiate the ow for the transient computations. Suresh et al

[31] investigated also the stability of the Phase Lag method on turbomachinery simulations

for the TACOMA solver.

Transient methodology is also recommended in cavity ows modeling with high

unsteadiness like in the present study to overcome numerical instabilities due to time av-

eraged calculations on inherently unsteady ows. Finally, it is well known that Phase-Lag

boundary conditions require approximately an order of magnitude more iterations to con-

verge than full unsteady runs but the possibility to model only one periodic sector for blade

row decreases dramatically the time necessary to complete the simulation.

 34

Chapter 5

Numerical Results and Comparison

with Test Data
All throughout the following paragraphs, "standard" CFD is used to indicate the

hystorical process used to simulate centrifugal compressor stages, i.e. with only the ow-

path and blades included into the computational domain, whereas "high delity" is used to

indicate a complete delity reproduction of the geometry and the use of advanced numerical

models. All the dierent aspects between the two approaches will be discussed in details in

the following paragraphs.

Figure 5.1 shows the comparison of performance curves between standard CFD,

high delity CFD and test data for a single stage centrifugal compressor with external

impeller diameter of 390 mm., ow coecient 0.0444 and peripheral Mach number of 0.73.

The geometry of the stage, including llets and leakage ows, is reported in gure 5.2. In

particular polytropic eciency, work coecient and polytropic head are reported in the

performance curves. The black vertical line indicates the stage design ow coecient. The

curves for polytropic eciency, work coecient and polytropic head have been derived using

respectively equations ( 5.1), ( 5.2) and ( 5.3).

Hpol
ηp = (5.1)
∆H0

∆H0
τ= (5.2)
U2
 35

ψ = τ · ηp (5.3)

Figure 5.1: Overall stage performance comparison between test and steady CFD for stage
ow coecient 0.0444.

Figure 5.2: Computational domain for stage ow coecient 0.0444.

In the CFD simulations, total pressure and total temperature were mass averaged.

The curves are plotted versus the ow coecient. All the CFD simulations are steady in

time and circumferential averaged proles are exchanged between blade rows with dierent
 36

reference frame. As a criterion to assess compressor operating point close to stall ow rate

during CFD simulations, runs have been discarded when either the polytropic head curve

becomes at or when onset of numerical instabilities appear. This is not the same as the

onset of compressor instabilities like surge or stall, that can be detected during test with

dynamic pressure probes. However, numerical instabilities are considered to have a strong

connection with non desirable ow phenomena, such as separation or blockage extension.

The improvement in CFD performance prediction accuracy with high delity CFD

versus standard one is clearly visible. Standard CFD is the one with the highest eciency as

expected and also is the one with the largest stall and choke margins. In fact it is well known

that standard CFD tends to under-estimate losses with respect to test data if no additional

dissipation terms are used and secondary eects are not modeled properly. Instead, when an

high delity reproduction of the geometry and advanced numerical schemes are used, CFD

predicitons match very well test data.

The stage was tested also in o-design speed; in particular in gure 5.3 are reported

also performance curves for peripheal Mach number 0.5 (lower curve) and 0.85 (higher curve).

Figure 5.3: Overall stage performance comparison between test and high delity CFD for
stage ow coecient 0.0444 at design and o-design speeds.

The comparison clearly suggests that only by using advanced high delity CFD

accurate numerical predictions are achievable for eciency, head, and operating margin,

otherwise not possible with simplied models.

Figures 5.5 and 5.7 show the same comparison for a very low and very high ow
 37

coecient stages. In particular the test cases selected are a bi-dimensional stage with ow

coecient 0.0095 and peripheral Mach number 1.0 in over-hung conguration with vaned

diuser and a tri-dimensional stage with ow coecient 0.1600 and peripheral Mach number

0.85. Computational domains for the two stages are reported respectively in gures 5.4 and

5.6.

Also for these two extreme stages the agreement between test and CFD is very good.

The introduction of the new computational models allow to assess centrifugal compressor

performance stages with an accuracy not possible before. Moreover, the new methodology

has been applied to a large envelope of dierent ow coecient and peripheral Mach number

stages giving a very good agreement with test data with small deviations across the full

spectrum.

Figure 5.4: Computational domain for stage ow coecient 0.0095.

In the remained part of this chapter, the inuence of all the main actors of the

level of CFD accuracy will be described in details.

5.1 Inuence of Numerical Schemes

Numerical errors arise from dierent sources. In particular they can be divided in

three main categories:

• Grid resolution
 38

Figure 5.5: Overall stage performance comparison between test and steady CFD for stage
ow coecient 0.0095.

Figure 5.6: Computational domain for stage ow coecient 0.1600.

• Code stability

• Turbulence modeling

For the rst point, with the increasing in computer performance, numerical errors

due to inaccurate discretization are becoming nowadays very small. In fact, the accuracy

of numerical errors were strongly linked to the limited number of grid points that could

be used in the past. Several million of elements for one blade sector are often used for
 39

Figure 5.7: Overall stage performance comparison between test and steady CFD for stage
ow coecient 0.1600.

routinely calculations. Many grid independency studies have been performed during this

research work and the necessary computational domain discretization to achieve a good

numerical resolution has been described in the previous chapter. In fact, as pointed out

by Denton in [32], the most common numerical approximation is that the ow properties

vary linearly between two grid points and so the error is proportional to the square of the

grid spacing times the second derivative of the ow property concerned. A critical area for

numerical error in centrifugal compressors is the area downstream the blunt trailing edge of

the impeller. Regarding the trailing edge discretization, a C block, instead of the usual O

type one, is very useful in the grid topology to avoid dissipating areas and high numerical

errors. Details of the grid at the mid-span trailing edge as well as the contours of EGR

showing the structure of the two counter-rotating vortices is presented in gure 5.8 for a

medium ow coecient impeller. In fact, the grid at the trailing edge has to be ne enough

to resolve the counter rotating vortices and not to show false uid structures as also shown

in [33].

CFD codes are becoming more and more robust. Articial viscosity has been

used extensively in the past to stabilize the numerical algorithms. However, in the common

applications of industrial centrifugal compressors, codes robustness limits the use of articial

viscosity to very few tough cases.

 40

Figure 5.8: Impeller trailing edge grid details and EGR contours.

Turbulent ows occur at high Reynolds numbers, when the inertia of the uid

overwhelms the viscosity of the uid, causing the laminar ow motions to become unstable.

Under these conditions, the ow is characterized by rapid uctuations in pressure and ve-

locity which are inherently three dimensional and unsteady. Turbulent ow is composed of

large eddies that migrate across the ow generating smaller eddies as they go. These smaller

eddies in turn generates smaller eddies until they become small enough that their energy

is dissipated due to the presence of molecular viscosity. The full inuence of the turbulent

uctuations on the mean ow must be modelled. In order to use a larger mesh size and still

capture the turbulent ow, the Favre averaged Navier-Stokes equations are used.

For attached fully turbulent boundary layers all the turbulence models are enough

accurate to predict ow beahvior. All the considerations are related to RANS models,

whereas more rened methodologies like LES and DNS are too expensive to be used in

common applications. Bourgeois [8] et al. and Mangani et al. [34] revised accurately the

dierent turbulence models for centrifugal compressor applications. Just few considerations,

uncovered in open literature for centrifugal compressors, will be added by the present author.

In particular, in the present work both the κ−ω and the SST models have been used. It

has been found that the κ−ω is more sensitive to the incoming turbulence length scale

than the SST that is so recommended when inlet boundary conditions are not well known.

Furthermore, when dealing with high strain ows, e.g. close to choke conditions, the Kato-

Launder production limiter variation of both the κ−ω and the SST models has given

more accurate agreement with test data. In fact, Kato-Launder modication is an ad-hoc

modication of the turbulent production term in the turbulent kinetic energy equation.
 41

The main purpose of the modication is to reduce the tendency that many two equation

models have to over-predict the turbulent production in regions with large normal strain,

i.e. regions with strong acceleration or deceleration. For separated ows, the eddy viscosity

is not allowed to get as large with SST as it does with other κ−ω model (and much less

than κ −  models) so that in a ow tending to separate, SST will predict a larger separation
region than other models. It has been shown to work well for ows with small separations

whereas for ows with larger separations, SST tends to predict separations that may be

too large relative to experimental observations. LES calculations might be expected to give

better results than RANS calculations for separated regions but several years are still needed

before computer power will allow to perform these calculations in standard applications.

Furthermore a curvature correction modier, like the one proposed by Smirnov and Menter

[7] has been implemented in the GE in-house code. It has been found that the correction

has a clear inuence on high ow coecients stages whereas has negligible eects on low

ow coecient bi-dimensional stages. Also, the eects, are predominanly located at the

U-bend of multistage compressors where the accuracy in predicting the separation region

is fundamental for the assessment of downstream return channel performance. In fact,

analysis of the turbulent ow eld showed a clear increased in turbulence production near

concave surfaces and decreased production near convex surfaces with connected dierent

loss generation mechanisms.

Finally, transitional ows are often present in many practical cases. Transition

refers to the process when a laminar boundary layer becomes unstable and turbulent bound-

ary layer develops. There are two types of transition: natural transition, where inherent

instabilities in the boundary layer cause the transition and by-pass transition, where con-

vection and diusion of turbulence from the free-stream into the boundary layer cause the

transition. Most transitions in turbomachinery are by-pass transitions caused by free-stream

turbulence and other external disturbances like wakes, vortices and surface defects. Simu-

lating transition in a CFD code accurately is very dicult, sometimes almost impossible for

very complex cases. Recent transitional models claim the ability to correctly simulate the

eects of transition even if much work still needs to be done on this subject.
 42

5.2 Inuence of Boundary Conditions and Gas Model

Inlet and outlet boundary conditions are fundamental to correctly model centrifugal

compressor stages. Usually, tangential averaged proles of total quantities and angles are

used at stage inlet section. Total pressure, total temperature and ow angle proles were

applied as inlet conditions. Static pressure prole is applied at the outlet. A great amount

of test data is available in the GE Oil&GAS database and accurate boundary conditions

are always applied in the CFD runs. Furthermore, recently, a great eort has been put in

measuring inlet turbuence conditions. In fact, a wrong estimation of the turbulence level at

inlet can lead to an over-production or under-production of turbulent kinetic energy inside

the stage with an inaccurate prediction of entropy and at the end losses. Two equations

turbulence models require the specication of both turbulence intensity and length scale

or some derived quantities. Hot-wire anemometry systems are currently employed at the

author's OEM test benches to estimate turbulence level. The incoming turbulence length-

scale is often even more dicult to guess than the incoming turbulence level. The best

way of guessing a realistic incoming length-scale is to use the geometrical properties of the

upstream components. Fortunately the incoming turbulence length-scale is usually not that

important for the end results.

Single stages centrifugal compressors are routenely tested with air, carbon dioxide

or refrigerant gases. Multistage centrifugal compressors are instead operated with very

dierent gases and operating conditions that can range between light gases like hydrogen

and heavy hydrocarbon mixtures and between atmospheric pressure and several hundred

bar. In CFD gas models can vary from simple constant properties to real gas algorithms.

In many cases (e.g. air at atmospheric inlet pressure), the inaccuracy in using a constant

properties gas model is negligible. Challenging applications like carbon dioxide close to

critical conditions require instead a real gas properties model to accurately predict the

performance of the compressor. In between these two options, a linear γ gas model is often

a good compromise between accuracy and execution time. The two main parameters that

aects the choice of the gas model are the compressibility factor Z and the isoentropic

exponent γ. It has been found that up to few percent of variation inside the compressor for

both the parameters, the perfect gas model oers a good prediction of stage performance.

For higher variations, e.g. in the order of 5% or more, the inaccuracy tends to be not

negligible.
 43

5.3 Inuence of Geometry Fidelity Reproduction

A delity representation of the geometrical features is a critical factor in CFD

accuracy. The main sources of uncertainties in the geometry denition are usually related

to:

• Leakage ows modeling

• Fillets

• Leading edge, trailing edge and blade prole shapes

• Hot geometry and tip clearance for open impellers

Regarding the rst two points, up to few years ago, the grid generation software

packages did not allow to mesh the geometry properly. For the leakge fows modeling, in

the next paragraph, a detailed explanation of the implications on CFD accuracy will be

presented. For the impeller llets, it is clear that operating range, especially choke margin,

and the amount of work transferred from the impeller to the gas, are related to the llet

radius and shape. Furthermore, a connection between the capability of the impeller to

resist at high incidence levels close to the blade tip and the llet radius in that zone is well

established.

Blades prole and, in particular, leading edge shape, is another sources of possible

simulation inaccuracies. If the blade shape is often well modeled in the computational

domain with a reasonable number of grid points, the leading edge shape is more sensible to

possible mismatching between real geometries and simulated ones. In fact, inaccuracies in

the stagnation point caption in the virtual simulation can lead to a misleading prediction of

boundary layer development on the blade surface. To correctly model the rapid accelerating

and decelerating ow in that area a very ne grid is mandatory.

Manufacturing variabilities and uncertanty quantication are being studied in de-

tails by Panizza et al [35], [36] and [37] in the same author's OEM and the implications on

CFD predictions are explored. It has been shown that the scatter for polytropic eciency

is largest at surge and smallest near peak eciency, while it is largest near choke for work

coecient. Impeller exit width and angle are critical parameters for all responses at all ow

coecients. However, away from the design point, other parameters gain a considerable in-

uence, such as inlet angles for eciency towards surge, and impeller eye clearance for work
 44

coecient, again in the surge region. Again, inlet and outlet delity geometrical impeller

modeling has been shown to be fundamental for compressors performance quantication.

Finally, tip clearance shape is often not well know in open impellers especially

for transonic stages where the running blade overall shape should be estimated by stress

analysis in hot conditions. The present work is mainly focused on closed impellers and

already detailed implications of these uncertanties on endwall losses and stall margin are

present in open literature, mainly for aero-engines applications like in the work from Hah

[5].

5.4 Inuence of Leakage Flows Modeling

Regarding eciency, it is well known that CFD without cavities modeling under

predicts losses generation due to the lack in capturing main secondary ows eects. Re-

garding the operating range, it is clear that stall limit is inuenced by instability eects on

ow angle at impeller exit due to the leakage cavities. In addition to this, the perturbated

adverse pressure boundary layer at the impeller inlet shroud cavity re-injection implies a

dierent ow incidence that aects both eciency and stall margin. At overow condi-

tions an additional mass ow is recirculating inside the impeller due to the presence of the

shroud cavity that implies a shift of the operating conditions of the impeller towards choke

conditions. Work coecient is the least aected by the introduction of cavities modeling.

Cavity modeling plays a fundamental role in the performance prediction. In par-

ticular losses associate with leakage ows are relevant and this is expect to increase with

lower ow coecient stages. In fact leakage mass ow scales with impeller pressure ratio

and cavity geometry, in particular number of labyrinths, labyrinths clearance and cavity

roughness that do not scale with ow coecient. It is well know that centrifugal compressor

stages have the maximum eciency in the range of medium ow coecients (0.0444-0.0956).

Moving up in ow coecient stages, prole losses become higher; moving down skin friction

losses become predominant. A linear trend can be observed for cavity ow losses moving

from the highest ow coecient stages to the lowest ones. In Figure 5.9 and Figure 5.10 it

can be observed the eciency and work coecient dierence from CFD simulations for the

cases with cavities across the full spectrum of ow coecients and peripheral Mach numbers

usually used in practical applications. Performance data are reported to the same conditions

at design speed with respect to no cavity cases. Moving to lower ow coecient impellers,
 45

stage eciency decreases not only due to the increased dissipation inside the main ow path

but also the windage and leakage losses. If, for medium ow coecient impeller leakage

losses are in the order of 2/3%, for a very low one it can reach up to 6/7%.

As stated above also an eect on work coecient can be associated with the leakage

ow, even if quantitatively smaller with respect to the losses increase. In fact compressed hot

ow passing through the shroud cavity re-enters before impeller leading edge. Two eects

counter-act: the higher mass ow passing through the impeller due to the recirculating

forward cavity leakage, forces the impeller to work more on the right side of the characteristic

line. Instead the compressed and hotter re-entering leakage ow upstream leading edge goes

in the opposite direction. Depending on the stage design and compressor operating point one

or the other eect prevails and the work coecient is aected consequently. Furthermore,

the ow re-entering the main ow from the shroud cavity aects the swirl close to the

tip impeller and so the amount of work transfered by the impeller to the ow. In the

case reported in gure 5.1 very small variation with respect to the model without cavities

included can be seen all across the performance line.

Eects of leakage ows are also very important in o-design speeds when standard

CFD is not even able to match performance on a delta basis. In fact, as it can be noted in

gure 5.3, CFD is able to match with a very good accuracy not only design speed performance

curves but also o-design speed curves. This was possible only by including leakge ows into

the model since the mass ow through the front and back sides cavity changes accordingly

with dierent head coecient. Without including leakage ows into the model, only the

eect of o-design blading losses are captured leading to under or over prediction of eciency

decrement with respect to design speed.

In gure 5.11 is shown the comparison of leakage ow amount between CFD and

correlations at dierent operating conditions for the stage ow coecient 0.0444 described

before. Correlations used for assessing the leakage ow amount are derived from the original

work of Daily e Nece and available in open literature. The agreement between CFD and

1D correlations is good expecially at design and close to choke condition. Close to stall the

CFD is able to better segregate the 3D eects and the leakage amount increases sensibly.

For the stages at lowest ow coecients the width of the cavities become of the same order

of the main ow-path and combined with the higher pressure ratio achieved translates in

larger leakage ow that can reach several percentage points with respect to the main ow.

Finally, the correct modeling of the shape and position of the area close to the
 46

Figure 5.9: Leakage ows eciency decrement across the full range of ow coecient stages.

Figure 5.10: Leakage ows work coecient eect with respect to no cavity ows across the
full range of ow coecient stages.

interface between the leakage ow and the main ow-path is fundamental in order to release

the ow inside the compressor with the appropriate turbulent kinetic energy that will aect

the boundary layer thickness in an area of adverse pressure gradient of the ow that is prone

to separation.
 47

Figure 5.11: Comparison of leakage ow amount between CFD and correlations at dierent
operating conditions for the stage ow coecient 0.0444.

5.5 Inuence of Surface Roughness

Surface roughness inuence on predicatability accuracy can vary largely based on

compressor main source of losses and Reynolds number. For an high ow coecient tri-

dimensional impeller the eect of skin friction losses is much lower than a bi-dimensional low

coecient stage. Moreover, depending on stage operating Reynolds number, viscous terms

can become a large source of losses also for high inlet volumetric ow stages. The eect of

surface roughness in CFD modeling is clearly related to the boundary layer thickness. A

detailed sensitivity analysis, both experimental and numerical, at dierent surface roughness

and for dierent stage type has been conducted during this research projects. A transition

Reynolds number and stage ow coecient have been found at which the surfaces can not

be considered anymore hydrodinamicaly smooth. In fact, for a medium ow coecient

stage at atmospheric inlet pressure, the eect on performance of the roughness model can

be neglected; instead for a low coecient stage at several hundreds bar inlet pressure the

losses are under-predicted if the roughness model is not included in the numerical setup.

The eect of roughness model on the work transfered from the impeller to the uid it has

been found to be lower than the one on eciency. Another source of uncertainty in using the

roughness model is the implementation of the model itself in all the CFD codes; in fact the

value to specify for surface roughness is not the real roughness but the sand grain roughness.
 48

In open literature dierent relations can be found between sand grain roughness and real

roughness ranging from a value of 2 up to 10. Many validation cases have been run to assess

this information in the present study.

Surface roughness is also very important to correctly assess the stall margin into

the diuser. It is well know that very low roughness value could avoid the transition from

laminar to turbulent ow in some cases and so to limit left operating range.

5.6 Multistage Eects

All the numerical and discretization uncertainties described before become even

more critical when dealing with multi-stage turbomachinery simulations. In fact, it is su-

cient that one of the stages of a multistage centrifugal compressor is not correctly predicted

in terms of performance curves in the CFD simulation, that all the following stages will

work at a dierent operating point from which they have been designed. In other words, it

is sucient than one stage is not correctly modeled that a mismatching will derive. Not only

overall single stage performance needs to be correcty evaluated but also released ow proles

at the outlet of each stage needs to be faithfuly reproduced numerically to not aect the

behavior of following stages. Furthermore, in multistage smulations also additional features

like balance drum leakages, inter-phase leakages, shunt holes and so on, need to be modeled

in order to assess overall compressor performance properly.

Figure 5.12 shows the CFD computational domain of a recycling wheel for an

ammonia process. The compressor stacking is composed by two sections with inlet and

outlet scrolls for each section. The last stage of the rst section is connected to the recycling

wheel in a back-to-back conguration. The computational model is made by the recycling

wheel itself, the inlet and outlet scrolls and the last stage of the rst section. Furthermore,

all the leakage ows for both the impellers as well as the end drum have been modeled.

Details of the back-to-back conguration are shown in gure 5.13.

Aim of the study was to predict the performance of the recycling wheel. This

impeller is critical for the process since it sets the pressure at the exit and so the overall

process conditions. Figure 5.14 shows the comparison between predicted performance from

CFD and test results: the agreement is excellent. As expected the agreement was possible

only by delity reproducing the computational model. Furthermore, it is mandatory to

model the leakage ow passing through the end drum, as shown in gure 5.15 to predict
 49

Figure 5.12: Computational domain for a reycling wheel in an ammonia process.

Figure 5.13: Details of the back-to-back conguration.

not only impeller performance but also diuser and outlet scroll ow proles and in general

overall losses as it will be shown later.

One of the most critical applications in the OIL&GAS market are the compressors

for LNG process. An example is reported in gure 5.16. The compressor is made by 5 stages

and it is divided into four sections with one inlet plenum and three side-streams. Operating

conditions go from several degrees Cesius below zero and atmospheric pressure at inlet up to

several tens of bar and degrees Celsius at the outlet. In the computational model, to reduce
 50

Figure 5.14: Comparison between predicted performance from CFD and test results for a
reycling wheel in an ammonia process.

Figure 5.15: Streamlines for the back-to-back system coloured by velocity.

the size, the stationary components were not included like in the previous one but all the

inlet boundary conditions for the stages were extracted from previous CFD computations

of all the stationary components. A real gas model for propane was used. The balancing

line, connecting the outlet with the inlet, was taken into account in the model with a source

term approach. As per API-617, the dierent sections can be tested separately with some

acceptance on ow deviations. In gure 5.17 is reported the performance curves predicted
 51

by CFD and measured in the string test for the rst section that is the most critical. The

other sections are not reported here for semplicity but the overall considerations are still

valid. The top graph shows the polythropic head in meters, isntead the bottom one the

polythropic eciency. Both are plotted versus inlet volumetric ow. The stages are high

Mach, high capacity tri-dimensional ones. The agreemnt is very good between test and

CFD. Furthermore the operating range is correctly matched ensuring a safe opeartion of the

compressor on eld.

Figure 5.16: Multistage centrifugal compressor for LNG application.

Figure 5.17: Comparison between predicted performance from CFD and test results for
section 1.
 52

5.7 Inuence of Unsteady Flows

On one hand, when dealing with unsteady ows, uid dynamics computations

require long solution times and therefore require huge computing and human resources. On

the other hand, steady computations are not sucient to fully understand phenomena like

blade row interaction, stall, distortion and hot streak migration that are inherently unsteady.

Unsteady turbomachinery simulations require the modeling of a common pitch

between adjacent blade rows to apply instantaneous periodic conditions. However, to avoid

resonance eects commonly rotors and stators have dierent blade counts; for all the cases

where no common pitch can be found between rotor and stator, a full annulus simulation

is necessary. This requires computational and memory eorts not applicable in most of the

industrial cases. Using the phase-lag method, temporal periodicity due to adjacent blade

interaction can be modeled with single passage. Phase-lag methodology can be applied

for periodic blade rows, whereas a ful annulus model is required when no axis-symmetric

geometry are present.

Compressor operating left limit is in general the one with the lowest prediction

accuracy. This is connected to the increment in unsteadiness as long as the compressor

operating point moves towards surge. Steady computations are known to be weak in this

eort and transient analysis is required. This is particularly true when sudden and strong

phenomena happens like in impeller or diuser stall. Steady computations are accurate

enough to predict overall performance curves but they lack in capturing rotating stall phe-

nomena. In gure 5.18 is shown the computational model for an over-hung impeller plus

vaned diuser. It is well know that mixing plane approach under-estimates separated shroud

region at diuser inlet that is a relevant feature in surge margin assessment on this low ow

coecient, high Mach stage. Anyhow, there is also experimental evidence that stability is

aected by leakage ows, especially at shroud side cavity where the out coming ow aects

the angle prole: to this intent, unsteady simulations with cavity ows modeling are able to

closer match the experimental data with respect to steady simulations without.

Figure 5.19 shows the performance comparison for the stage between steady CFD,

unsteady CFD and test data. Looking at the test data it is clear that strong unsteadiness

happens suddenly moving towards left operating limit. Since the work transfered from the

impeller to the ow continues to increase, the stall seems happening inside the diuser.

Furthermore since dynamic pulsations of static pressure are stronger at section 40, i.e.
 53

Figure 5.18: Computational domain for an over-hung impeller plus vaned diuser.

diuser outlet than other locations and the frequency of sub-syncronous vibrations are the

ones common for stationary components, the conclusion is that some blades into the diuser

are too loaded and so stalled. Steady CFD is not able to model rotating stall; an indication

of aerodynaic instabilities comes looking at the convergence but the eect seems to be mild

compared to the real phenomena. instead unsteady CFD is able to predict with greater

accuracy the sudden drop in stage eciency due to the strong unsteadiness inside the diuser.

Furthermore, an FFT of the signal from usnteady CFD coming from virtual probes inside

the computational domain showed a comparable frequency for the phenomena with test

data.
 54

Figure 5.19: Steady and unsteady CFD predictions and test data for an over-hung impeller
plus vaned diuser.
 55

Chapter 6

Flow Analysis
In this chapter the ow eld inside the dierent components of a centrifugal com-

pressor will be analysed in details with the aim to understand and explain aerodynamics

phenomena not fully captured before.

6.1 Impeller Flow

Figure 6.1, 6.2 and 6.3 show the 2D maps at section 20 for total pressure from

FRAPP, high delity CFD (with leakage ows model included) and standard CFD (without

cavity ows model included) at design mass ow rate, close to surge and close to choke. The

test case reported here is a tri-dimensional impeller, single stage (impeller, vaneless diuser

and return channel) at medium ow coecient for general purposes applications already

described in gure 5.2.

The measurement plane is at section 20, i.e. downstream of the impeller trailing

edge, where the FRAPP measurements were taken. The maps refer to two blade pitches

An highly non uniform ow eld is presented from experimental data with huge defects

at some positions close to the hub and shroud surfaces. The region of highest pressure is

concentrated around mid-span. The pitch value equal to 1 corresponds to the blade suction

side. The impeller is rotating from right to left. Proles from FRAPP are time averaged.

Total pressure is normalized by the dierence between maximum and minimum value at the

measurement plane.

The high delity CFD is in very good agreement both qualitatively and quantita-

tively with FRAPP data; instead at some locations, standard CFD gives misleading results.
 56

Figure 6.1: FRAPP and CFD total pressure 2D map at section 20 comparison at design
ow rate.

Figure 6.2: FRAPP and CFD total pressure 2D map at section 20 comparison close to surge.

The capability to correctly reproduce the ow features at the impeller trailing edge with

CFD is fundamental for the design of diuser and downstream stationary components opti-

mization. In fact the correct design and optimization of downstream stationary components

requires an accurate modeling of ow phenomena at impeller trailing edge.

The absolute total pressure maps present a highly nonuniform pattern both along

the spanwise and tangential directions for both FRAPP and high delity CFD. The high
 57

Figure 6.3: FRAPP and CFD total pressure 2D map at section 20 comparison close to choke.

total pressure zone is contoured by a low pressure zone and a considerable defect close to

both hub and shroud. In particular, both the FRAPP and CFD total pressure maps present

a reduced total pressure zone in the region identied as the wake zone (dened by high

tangential ow). The low total pressure region at the hub has historically been one of the

most dicult areas for CFD to reproduce. This is mainly due to the high tangential ow

re-entering from the cavity at the hub. It should be noted here that all the previous analysis

in open literature did not show same agreement. The high pressure zone is located close to

shroud side at high and design ow coecients and moves towards mid-span at lower ones.

Moving away from design point the high pressure zone spreads and the non uniform pattern

is more pronounced in the circumferential direction.

Slightly higher dierences between experimental data and high delity CFD with

cavities modeling can be found quantitatively at close to stall and choke operating conditions

where unsteady eects are predominant.

Figures 6.4, 6.5 and 6.6 show the 2D maps at section 20 for YAW angle. The yaw

angle is calculated as per equation ( 6.1).

Vθ
yaw = tan−1 (6.1)
VM
Flow angles are plotted with respect to the radial direction and are positive if the

tangential velocity has same sign than the impeller peripheral velocity. More tangential ow
 58

angle is associated with greater values.

Figure 6.4: FRAPP and CFD YAW angle 2D map at section 20 comparison at design ow
rate.

Figure 6.5: FRAPP and CFD YAW angle 2D map at section 20 comparison close to surge.

In the yaw angle maps a very similar ow distribution between FRAPP and high

delity CFD can also be observed. In fact, both maps have a more tangential ow region in

the suction side-shroud corner than in the main core region of the ow. This distribution

corresponds to the classical two-zones of jet and wake model. The wake formation inside the

impeller is interesting due to its strong interaction with loss generation. In closed impellers
 59

Figure 6.6: FRAPP and CFD YAW angle 2D map at section 20 comparison close to choke.

the wake is usually positioned close to the center of the channel. The ow angle dierence

between the jet and wake zones has implications on velocity triangles at the impeller exit.

In particular the wake region is characterized, in the absolute reference of frame, by a

high tangential, high momentum ow that concentrates at the suction surface-shroud side

in agreement with [5], [38] and [39]. In fact in centrifugal compressor impellers a strong

secondary ow region develops through the passage and this low momentum ow region is

concentrated in the shroud surface suction corner whereas the ow is well energized near

the hub. This secondary ow detaches from the suction surface at the beginning of the

back-swept zone and moves away from the wall. The low momentum ow is responsible for

the highest losses in the impeller. This aects not only the impeller ow eld but also the

angle distribution at the exit of the impeller. The absolute ow angle at the diuser inlet

will be more tangential near the shroud and more radial near the hub region, leading to

a complex ow evolution in the diuser channel. In fact, this non-uniformity in the yaw

angle eld at the impeller outlet originates two counter-rotating vortical structures in the

diuser that reduce the component eciency. Moreover, the secondary ow region is due

not only to possible separated regions in the impeller, especially close to stall conditions,

but also to skewed boundary layers and secondary ows. Both CFD and FRAPP give a

more developed and dissipative pattern close to the tip as a result of the strongest ow

phenomena occurring in the impeller region. Maximum dierence between jet and wake

zones is about 30
◦ close to choke operating point. High tangential ow zones are associated
 60

with more dissipative ow behavior and higher turbulent kinetic energy. Higher tangential

ow region can be found, as expected, close to stall point. Turbulent kinetic energy (TKE)

can be derived from both FRAPP and CFD and can be linked to the ow structures shown

before: the pattern reveales a strip along the blade span and two regions of higher turbulent

kinetic energy are reported close to the endwalls with a peak value near to the hub as shown

also by [4]. Both regions are the result of the interaction between the impeller wake and the

strong dissipative secondary ows regions developed at the impeller trailing edge.

Figure 6.7 and Figure 6.8 shows the comparison at the measurements location of

the tangential averaged span-wise proles respectively for yaw angle and total pressure for

FRAPP data and both standard and high delity CFD simulations. The test case the

gures are referring about is still a tri-dimensional impeller, single stage (impeller, vaneless

diuser and return channel) at high ow coecient and high peripheral Mach number for

LNG applications shown in gure 4.1.

Figure 6.7: Yaw angle comparison between FRAPP and dierent CFD approaches at im-
peller exit.

An highly non uniform ow eld is presented from experimental data with huge

defects at some positions close to the hub and shroud surfaces. The region of highest pressure

is concentrated around mid-span. The impact of the high delity approach with full cavities

modeled on impeller exit prole is strong. In fact, the standard CFD approach has been

found historically to be weak in capturing correctly these proles.

At design point the ow direction in the hub cavity is from return channel trailing
 61

Figure 6.8: Total pressure comparison between FRAPP and dierent CFD approaches at
impeller exit.

edge towards impeller trailing edge due to the static pressure recovery in the stationary

parts being higher than the pressure losses associated with skin friction. High delity model

with full cavities meshed is able to capture the total pressure prole close both ot hub and

shroud sides. In the area close to the non-matching interfaces between impeller trailing edge

and both shroud and hub cavities, a strong area of recirculating ow develops. Reverse ow

forms across the interface that is well captured by the numerical model. In fact, for the

shroud cavity, the eect of the rotating blades is counter-acted by the recirculating ow in

the cavity. The impeller pushes the ow inside the cavity but the presence of labyrinths seals

mitigate the amount of ow passing through that recirculates close to the impeller trailing

edge determining the shroud defect shown in the total pressure prole. Also the eect of the

hub cavity on total pressure defect prole at mixing region is clearly visible. The leakge fow

coming from the return channel at very high swirl mixes wth the radial ow exiting from

the impeller creating a recirculation area responsible of the low pressure close to the wall.

Aerodynamic parameters in the cavities are almost constant due to the axis-

symmetry of the geometry except near the interfaces close to the impeller trailing edge

where a distorted pressure distribution inuences the uid ow uniformity. With full cav-

ities meshed also tangential proles of ow at interface can be correctly captured. Figures

6.9, 6.10, 6.11, 6.12 and 6.13 show the average tangential proles of yaw angle and total

pressure and the tangential proles of total pressure at 10%, 50% and 90%. Close to the wall
 62

locations, the eect of leakages is expected to have a stronger infuence on the ow proles.

The tangential direction refers to two blade pitches and the angle is plotted with respect to

the radial direction. The pitch value equal to 1 corresponds to the blade suction side. The

impeller is rotating from right to left. The simulated operating point is the design one of

the compressor stage descrived previously and the proles are plotted at section 20. The

agreement is very good with all the hills and valleys captured along the tangential direction.

It was also found that llets modeling plays a non negligible role in capturing the tangential

proles. Combined llets and cavities accurate meshing was possible only with non matching

interfaces especially at the trailing edge of the impeller due to the combined high curvature

of the ow-path and interaction of llets with blunt trailing edge. Also, due to the loss

generation near the wall regions, the total pressure and its amplitude of uctuation due to

the blade passing are lower in the shroud and hub regions than at the mid-span. Generally

CFD is more sensitive to pressure uctuations than experimental data.

Figure 6.9: Average tangential total pressure prole comparison between FRAPP and high
delity CFD.

For the same test case, also tangential averaged proles of YAW angle at o-design

points, i.e. close to stall and close to choke are presented in gures 6.14 and 6.15.

CFD show a tendency to slightly under-predict the spread of the quantity across

the span at o-design conditions even if the average value is in good agreement with test

data. This is in agreement with what can be expected from a steady simulation; in fact, the

time averaged of the main quantities tends to smooth out strong and sudden phenomena.

In fact, close to stall and close to choke unsteadiness become predominant and in particular

close to endwalls where the tri-dimensional eects are higher. However, the agreement on
 63

Figure 6.10: Average tangential yaw angle prole comparison between FRAPP and high
delity CFD.

Figure 6.11: Tangential total pressure prole comparison between FRAPP and high delity
CFD at 10% of the span.

average vaues is very good and overall performance can be then correcty predicted.

6.2 Diuser Flow

Centrifugal compressors are very dicult to simulate using CFD due to extremely

high curvature. This is particularly true for stationary components; standard CFD is his-

torically weak in capturing the correct ow features in downstream stationary components.

In gures 6.16, 6.17 and 6.18 respectively the non dimensional values of total

pressure, yaw angle and non dimensional static pressure at section 40 are reported , i.e.

diuser outlet, for both high delity CFD and standard CFD. The test case is the one
 64

Figure 6.12: Tangential total pressure prole comparison between FRAPP and high delity
CFD at 50% of the span.

Figure 6.13: Tangential total pressure prole comparison between FRAPP and high delity
CFD at 90% of the span.

described in gure 4.1.

The aerodynamic behavior of the diuser is strongly dependent by the leakage ows

and can be understood properly only if the modeling of shroud and hub cavities are included

in the computational domain. The amount of ow passing through the cavities depends, for

a certain geometry, on the pressure ratio delivered by the impeller for the shroud leakage

and by the combined eects of pressure recovery and total pressure losses of downstream

stationary components for the hub one. Moving from design operating point towards left

and right limit the amount of ow through both the cavities tends to increase and decrease

respectively. Furthermore in deep choke conditions the direction of the ow in the hub cavity

can reverse, i.e. going from impeller to the return channel.

 65

Figure 6.14: FRAPP and CFD tangential averaged yaw angle at section 20 comparison close
to surge.

Figure 6.15: FRAPP and CFD tangential averaged yaw angle at section 20 comparison close
to choke.

Figures 6.19, 6.20 and 6.21 show the contours plots of meridional velocity inside

the diuser with super-imposed the ow streamlines respectively close to surge operating

limit, at design point and close to choke operating points.

First thing to note is the absence, for the high delity CFD, of the low momentum

zone close to shroud side at the end of the diuser that has been shown in all previous

computations with standard CFD. From the streamlines pattern, it is easy to understand
 66

Figure 6.16: Non dimensional total pressure proles at section 40 at design ow rate.

Figure 6.17: Yaw angle proles at section 40 at design ow rate.

that the high swirling ow coming back into the domain from the hub cavity, pushes the

main ow towards the shroud side. This aects also the ow behavior downstream where a

region of low momentum ow develops close to the diuser hub due to the mixing of cavity

ow with main-stream.

Close to surge condition, the low momentum zone due to the mixing of the ow

re-entering from the hub cavity increases signicantly along the hub side and continues

downstream up to the diuser end. Without modeling cavity leakage domains, this phe-
 67

Figure 6.18: Non dimensional static pressure proles at section 40 at design ow rate.

Figure 6.19: Meridional velocity contours inside the diuser close to surge operating limit.

nomenon was explained in the past, with an apparent switch of diuser low momentum area

from shroud to hub at lower ow operating points as shown by Sorokes et al. [9]. Including

cavity domains into the model the picture that comes out looks dierent; in particular the

ow on the shroud side close to the diuser end tends to separate due to the diusion inside

the diuser and the higher tangential ow at the end of the impeller. The last phenomenon

is obviously correlated to diuser/impeller ratio and pinch amount.

Close to choke operating limit the inuence of both leakage ows is much smaller
 68

Figure 6.20: Meridional velocity contours inside the diuser at design operating point.

Figure 6.21: Meridional velocity contours inside the diuser close to choke operating limit.

and translates in lower separated regions both close to shroud and hub endwalls.

Figure 6.22 and 6.23 show the CFD proles of yaw angle at diuser exit and the

comparison of average value between high delity CFD and test respectively close to surge,

design and close to choke operating conditions. Yaw angle values are negative and with

respect to the radial direction that means more tangential ow towards the left.

The agreement between test and CFD is very good as shown. The average yaw

angle becomes more tangential moving from choke to surge as expected. The eects of cavity
 69

Figure 6.22: Yaw angle proles at diuser exit from CFD close to surge, design and close to
choke operating conditions.

Figure 6.23: comparison of average yaw angle at diuser exit between CFD and test close
to surge, design and close to choke operating conditions.

ow is clearly visible in the yaw angle proles when moving from choke to design point, the

ow becomes more tangential at both hub and shroud. Instead the low momentum region

close to hub side is visible for both design point and close to surge conditions with very close

values. At the shroud side the angle is more tangential for the close to surge condition due

to the strong adverse pressure gradient at the end of the diuser.

Boundary layer development in those areas are especially dicult to simulate be-
 70

cause they are usually formed by a combination of boundary layer growth, secondary ows

and leakage ows. As a result the boundary layer skewness introduces streamwise vorticity,

which has a large eect on the secondary ow in the downstream blade rows.

To also take into account the eect of diuser modeling when vanes are present,

another test case for a bi-dimensional impeller at ow ow coecient stage with vaned diuser

is show in gure 5.18. Figure 6.24 shows the spanwise, tangentially averaged, ow angle

prole at diuser inlet for high delity CFD and standard one at design point, close to choke

(right limit) and close to stall (left limit) operating points.

Figure 6.24: Flow angle prole at diuser inlet for high delity and standard CFD at dierent
operating points.

It is evident from the comparison that low momentum, high tangential ow is

sucked inside the shroud cavity for all the operating points. The same eect happens for

the back-side cavity even if to a lesser extent than the front-side one. Figure 6.25 shows the

contours of VM · |Vr | inside the diuser for the high delity and standard one cases close to

stall operating point. The horizontal black line represents the position of the vanes leading

edge into the diuser.

Even if it is well known that the mixing plane approach under-estimates separated

shroud region at diuser inlet that is a relevant feature in surge margin assessment, it is

clear from the gure that the an area of low radial velocity ow is present at diuser inlet

close to the endwalls in case of leakge ows due to the strong diuser/cavities interaction.

Instead for the no leakages simulations the area is not present and the ow develops inside
 71

Figure 6.25: Contours of radial velocity inside the diuser after the mixing plane position
for the high delity and standard CFD respectively.

the diuser without any interaction.

Also, when diuser is choked, the correct computation of the total pressure prole

at impeller exit is fundamental for the right limit assessment since diuser throat area,

aerodynamic blockage and gas properties do not vary signicantly. Cavity ows modeling

was found to be important for these purposes.

Finally in gure 6.26 is shown the trend of static pressure recovery coecient

inside the diuser for the CFD with and without leakage ows included across the stage

operating range. The static pressure recovery coecient is calculated as shown in equation

( 6.2). The black vertical line indicates the stage design ow coecient, the horizontal one

the 0 in the Cp scale.

P s40 − P s20
Cp = (6.2)
P t20 − P s20
It worth nothing here that in case of no cavities modeled inside the domain the

diuser shows a recovery also close to deep choke whereas the CFD with leakge ows included

shows an acceleration inside the diuser. Also towards the left limit the CFD with cavities

ow included tends to be more conservative and closer to test data.

Moving from impeller exit to diuser outlet, previous analysis on vaneless diusers

show that there is a strong interaction bewteen leakage ows and diuser and that only
 72

Figure 6.26: Static pressure recovery coecient inside the diuser for the CFD with and
without leakage ows included across the stage operating range.

modeling cavities inside the domain is possible to understand correctly ow behavior and

accurately predict the performance. In particular the low momentum, high tangential ow is

sucked inside the cavities and an area of recirculating/mixing ow develops after the impeller

trailing edge. This translates in a lower static pressure diuser recovery due to the increase

of separated regions close to the endwalls with fallouts in the downstream components

performance. Diuser inet metal angle distribution designed for a non perturbated ow

would lead to a wrong assessment of stationary components behavior.

Finally, when vanes are installed into the diuser, and velocity proles at diuser

exit are less aected by the modeling of cavities because are more inuenced by the turn-

ing given by the blades. Furthermore, the necessity to interpose a mixing plane between

impeller trailing edge and diuser leading edge due to dierent periodicity of the blades

smooth out the circumferential eects of the cavity ows presence on downstream station-

ary components. Finally, it is well know that outlet scroll imposes a strong circumferential

non-uniformities throughout the compressor and provides a downstream volume that can

allow surge and so it is necessary to model it for compressor dynamic behavior assessment

at surge. Nevertheless it was found in this case that static pressure distortions at section

40 were below 3% with respect to average value, i.e. the distribution of static pressure was

quite uniform at diuser exit.

 73

6.3 Return Channel Flow

The correct modeling of ow phenomena at diuser exit is very important for the

optimization of return channel. In particular Smirnov et al. [7] showed improvements in

ow modeling using the SST turbulence model with a curvature correction method and in

particular in the U-bend region. Kowalski et al. [40] using standard CFD showed a big

separated region in the U-bend of a centrifugal compressor stage that has never been found

in experimental activities as per author's knowledge.

Figure 6.27 shows the non dimensional circumferentially averaged contours of

meridional velocity from diuser inlet (after leakage cavities interfaces) up to U-bend exit

for the standard CFD high delity CFD respectively at design mass ow rate. Test case

reported here is the one described in gure 5.2

Figure 6.27: Circumferential averaged meridional velocity contours from standard CFD (left)
and high delity CFD (right) at design point.

The ow eld for the standard CFD shows a big area of low momentum ow close

to the shroud side in the U-bend region reported in the red circle in the picture. Instead the

case with cavities modeling does not show this phenomenon. The eect of the re-entering

ow at hub side is clearly visible (in red circle). The eects on the ow proles at impeller

trailing edge were discussed in the previous paragraphs; furthermore the high swirling ow

pushes the main ow towards the shroud side. This aects the ow behavior downstream

where a region of low momentum ow develops close to the diuser hub (reported in black
 74

circle at hub side) due to the mixing of cavity ow with main-stream. Further downstream

the low momentum region at shroud side inside the U-bend region reduces signicantly (in

black circle at shroud side of the U-bend).

Figure 6.28 and 6.29 shows the non dimensional circumferential averaged contours

of meridional velocity from diuser inlet (after leakage cavities interfaces) up to U-bend exit

for the standard CFD and high delity CFD respectively close to stall and close to choke.

Figure 6.28: Circumferentially averaged meridional velocity contours from standard CFD
(left) and high delity CFD (right) close to stall point.

The rst thing that can be observed, as expected, is the major inuence of leakage

ows modeling close to stall with respect to close to choke operating condition. In fact close

to choke the amount of leakage ow through the cavities is much smaller with respect to

close to stall as showed before for both hub and shroud cavities. This results in a lower

inuence on ow features dierences between the case with and without cavities. However

similar conclusions to the design point condition can be drawn even if at a lower extent.

Close to stall condition, the dierence in the circumferentially averaged meridional

velocity contour plot is remarkable between standard CFD and high delity one. In the

standard CFD, the ow is attached to the diuser hub and a large low velocity zone is

visible close to the diuser shroud. The ow starts to separate in the diuser pinch area

and develops all across the diuser and U-bend zones. On contrary, in the high delity

CFD, low velocity zone can be seen close to the diuser hub wall and main ow is pushed

towards the shroud side. Anyhow the ow starts to detach from shroud wall close to the
 75

Figure 6.29: Circumferentially averaged meridional velocity contours from standard CFD
(left) and high delity CFD (right) close to choke point.

diuser end due to the strong adverse pressure gradient; the separation seems to be a real

aerodynamic situation and dierent from what found at design point condition for the case

without cavities.

Figure 6.30 shows the component by component losses for the standard CFD and

high delity CFD. The eects of leakage ows are clearly shown all across the compressor

stage in reducing the eciency with respect to the case without. As expected, the main

dierence is found in the impeller, where the compressed gas leaks through the shroud cavity

re-entering in the main ow at impeller inlet. The dierence in ow proles at diuser inlet

as shown before aects the diuser losses, which in the case of cavities modeling are due

to mixing zone and low momentum areas. Losses in the return channel are also aected by

the ow leaving the main-stream at hub trailing edge even if at a lesser extent. U-bend is

the only component that shows higher eciency in case of CFD with cavities modeling, the

reduced low momentum zone close to the shroud side can be responsible of this eect as also

conrmed by ow proles at diuser exit showed in Figures 6.16, 6.17 and 6.18.

Finally the comparison between test and high delity CFD for the total pressure

loss coecient in the return channel is shown in Figure 6.31. The total pressure loss coe-

cient is calculated as shown in equation ( 6.3).

P t60 − P t40
ξ= (6.3)
P t40 − P s40
 76

Figure 6.30: Component losses from CFD with and without cavities modeling at design
point.

Figure 6.31: Total pressure loss coecient in return channel for test and high delity CFD.

It is well know that 1D model of return channel are not able to capture accurately

the losses across the entire compressor operating range. This is due to the weak correlations

of ow properties, e.g. ow angle, between impeller and return channel. Furthermore it has

been shown that standard CFD is not able to capture accurately the ow proles at the

diuser exit, leading into a wrong assessment of return channel losses. Instead, highdelity

CFD enabled capturing the ow behavior and overall losses all across the stage operating

conditions with a good accuracy, especially at overow.

 77

6.4 Cavity Flow

The modeling of the entire cavities allows to have a deeper insight in the leakages

ow eld. Direction of ow stream, in a single stag compressor, is always from trailing edge

to leading edge in the shroud cavity; instead for the hub one the direction depends on the

balance between pressure drop from impeller trailing edge to return channel trailing edge

and static pressure recovery of stationary components. Static pressure recovery coecient

is calculated as per equation ( 6.4):

P s60 − P s20
cp = (6.4)
P t20 − P s20
At compressor operating conditions where the pressure recovery coecient is pos-

itive the ow direction is from return channel to impeller, vice versa if negative. At design

operating point and towards left limit the pressure recovery coecient is positive. Towards

right operating limit at a certain ow coecient the increase of total pressure losses due

to skin friction becomes too high with respect to the stationary components static pressure

recovery and the ow inverts the direction.

Moreover, the ow in balance drum or inter-stage leakages is always from the higher

pressure section to the lower one. The ow there is similar to what will be described for

single stage.

In both shroud and hub cavities, inner wall rotates at the same peripheral speed

of the impeller, instead the outer one is stationary.

Figure 6.32 shows contours of entropy with streamlines super-imposed for the front

cavity of a bi-dimensional low ow coecient stage at high peripheral Mach number in

over-hung conguration, described in gure 5.18. It is clear that there is a strong area of

recirculating ow in proximity of the shroud interface. Reverse ow that forms across the

interface is well captured by the numerical model. In fact the eect of the rotating blades

is counteracted by the recirculating ow in the cavity. The impeller pushes the ow inside

the cavity but the presence of labyrinth seals mitigate the amount of ow passing through

that recirculates close to the impeller trailing edge. Figure 6.33 shows contours of entropy

with streamlines super-imposed for the front cavity labyrinth seals. Jets form between teeth

and lower faces, instead cavity vortices establish between dierent teeth. Steps in the cavity

geometry are used to destroy jets continuity.

 78

Figure 6.32: Contours of entropy with streamlines super-imposed for the front cavity.

Figure 6.33: Contours of entropy with streamlines super-imposed for the front cavity
labyrinth seals.

Figure 6.34, 6.35 and 6.36 show respectively axial, radial and tangential velocity

inside the forward cavity in the area at mid distance between cavity ow inlet and labyrinth

seals marked with a read line in gure 6.32. In gures 6.34, 6.35 and 6.36 the non

dimensional span is referred to the no dimensional distance between stationary and rotating

walls inside the cavity. Fluid particles close to the inner wall due to the no-slip condition

have a tangential velocity that is the same of the rotating wall and so centrifugal force much
 79

higher than the ones close to the outer wall that are at rest. As expected axial velocity

is the one with lowest magnitude. Radial velocity increases rapidly from 0 at the wall to

the maximum close to the wall and then becomes zero again at the center region; opposite

walls have opposite radial velocity direction. The ow is anti-symmetric with a very small

negative total average radial velocity due to the net leakage ow leaving the cavity. Static

pressure inside the cavity is almost equal along the axial and circumferential direction and

decreases gradually in the radial one moving downwards. An highly complex uid structure

develops and three main regions can be distinguished based on previous statements: a core

ow and two opposite areas close to the walls. The core ow has essentially zero velocity

gradient as shown in the radial velocity proles. The core ow can have dierent vortex

structures depending mainly on the cavity geometry. The uid close to the rotating wall

is centrifuged towards the cavity inlet whereas the uid close to the outer wall is moved

downwards. A perfectly corresponding behavior can be described for the rear cavity.

Figure 6.34: Axial velocity distribution inside the forward cavity.

Figures 6.37 and 6.38 show the comparison between test data and CFD for static

pressure variation inside the cavities at the inner and outer walls versus non dimensional

radius for a medium ow coecient stage in intermediate stage conguration. Radius 0 is

the minimum cavity radius and 1 is the maximum. The experimental data are area averaged

between two redundant probes for each measurement point. Figure 6.39 shows the position

of the static pressure probes inside both the cavities at hub and shroud for a typical case.

From CFD is clearly visible the pressure drop across the teeth seals for both cavities
 80

Figure 6.35: Radial velocity distribution inside the forward cavity.

Figure 6.36: Circunferential velocity distribution inside the forward cavity.

that is well captured. Generally the agreement between test and simulation is good. From

the static pressure drop across the cavity in the test data is possible to estimate the leakage

mass ows. At design point leakage mass ows are 0.6% and 0.15% of the total mass ow

respectively for shroud and hub cavities. This was found to be in very good agreement

with CFD as can be inferred from the static pressure distribution. For simulations with

coupled hub cavity and source terms like the one in the study of Sun [13], the static pressure

distribution is highly dependent on the prescribed mass ow that is not known a priori;
 81

Figure 6.37: Static pressure variation across the shroud cavity for CFD and experimental
data.

Figure 6.38: Static pressure variation across the hub cavity for CFD and experimental data.

usually a trial and error method is used until the expected pressure drop across the cavity

is found. Instead in this case, modeling all the cavity domain, no source terms needs to

be prescribed and the accuracy of the simulation greatly improves. Static pressure at hub

cavity on the return channel side is higher than the one close to impeller trailing edge

and this is connected to ow direction as explained before. It is also clear from the static

pressure distribution in the hub cavity the pumping eect of uid after the seal labyrinths.
 82

Figure 6.39: Position of the static pressure probes inside both the cavities at hub and shroud
for a typical case.

Furthermore the agreement between test and computational data allows to expect that

the amount of recirculating mass ow inside the shroud cavity is correctly assessed by

CFD. Axial thrust is another important information for bearings selection and rotordynamic

verication. The agreement in static pressure distribution between test and CFD allows to

derive important correlations for a priori verication. Prediction tools for leakage eects

can be tuned based on CFD results.

 83

Chapter 7

Conclusions
Aim of this thesis was to push CFD capabilities toward virtual testing. In other

words, this thesis describes a methodology by which virtual tests can be conducted on single

stages and multistage centrifugal compressors in a similar fashion to a typical test rig. In

fact, in the vision of the author, CFD represents nowadays an exceptional tool for centrifugal

compressors performance prediction and ow phenomena understanding.

7.1 Summary
Several centrifugal compressor single stages, ranging across dierent inlet ow coef-

cients and peripheral Mach numbers, applied to disparate processes (e.g. LNG, pipeliners,

barrel compressors, etc), as well as applications of multistage compressors, were selected as

test cases for this study. Both steady and unsteady simulations were conducted in order to

fully capture the time averaged and time accurate operating conditions. Moreover, detailed

geometrical features were faithfully reproduced in the computational domain and advanced

numerical models were used in the setup with accuracy not possible before.

CFD predictions were compared with test results regarding both overall perfor-

mance and detailed ow features. Advantages of this approach is the reduced cost with

respect to tests and the possibility to virtually instrument the machine all along the com-

putational domain where it is not possible in the real compressor.

Results of the present study can be summarized as follow:

• The so called high delity CFD is able not only to predict with an outstanding accuracy

overall compressor performance but also to reproduce local detailed ow features.
 84

• Particular care needs to be put in modeling accurately all the single components of

the centrifugal compressor. In prticular, geometrical features like llets, leakages ows,

leading edge and trailing edge discretization need to be faithfully reproduced into the

computational domain.

• Numerical algorithms need to be at the state of the art. Turbulence models, curvature

correction, roughness treatment, boundary conditions and real gas model are the main

actors for the nal outcome.

7.2 Recommendations and Future Work

Even if nowadays CFD is mature enough to represent a realible tool for centrifugal

compressors modeling, many areas need to be further explored and developed. Some working

in progress are reported here below:

• Unsteady simulations of full multistage centrifugal compressors are still too expensive

for common industrial applications. However, with the exponential increase in infras-

tructures happened in the recent years, in the vision of the author, this will become

reality in few years.

• Boundary layer transition is almost impossible to predict accurately with RANS mod-

els. LES applications would solve many uncertainties but many years need to be passed

before industrial applications will become available.

 85

Bibliography
[1] Ferrara, G., Ferrari, L., and Baldassarre, L., 2006. Experimental characterization of

vaneless diuser rotating stall. part v: inuence of diuser geometry on stall inception

and performance. In ASME Turbo Expo 2006. ASME Paper GT-2006-90693.

[2] Persico, G., Gaetani, P., and Guardone, A., 2005. Design and analysis of new concept

fast-response pressure probes. In Measurements Science and Technology, 16(9), pp

1741-1750.

[3] Persico, G., Gaetani, P., and Guardone, A., 2005. Dynamics calibration of fast-response

probes in low-pressure shock tubes. In Measurements Science and Technology, 16(9),

pp 1751-1759.

[4] Toni, L., Ballarini, V., Cioncolini, S., Persico, G., and Gaetani, P., 2010. Unsteady ow

eld measurements in an industrial centrifugal compressor. In 39th Turbomachinery

Symposium, pp 49-58.

[5] Hah, C., and Krain, H., 1990. Secondary ows and vortex motion in a high-eciency

backswept impeller at design and o-design conditions. Journal of Turbomachinery,

pp. 713.

[6] Smirnov, P., Hansen, T., and Menter, F., 2007. Numerical simulation of turbulent ows

in centrifugal compressor stages with dierent radial gaps. In ASME Turbo Expo 2007.

ASME Paper 2007-27376.

[7] Smirnov, P., and Menter, F., 2008. Sensitization of the sst turbulence model to rotation

and curvature by applying the spalart-shur correction term. In ASME Turbo Expo

2008. ASME Paper 2008-50480.

 86

[8] Bourgeois, J. A., Martinuzzi, R. J., Savory, E., and Zhang, C., 2011. Assessment of

turbulence model predictions for an aero-engine centrifugal compressor. Journal of

Turbomachinery, p. 133.

[9] Sorokes, J., Vezier, C., Pacheco, J., and Fakhri, S., 2012. An analytical and exper-

imental assessment of a diuser ow phenomenon as a precursor to stall. In ASME

Turbo Expo 2012. ASME Paper 2012-69122.

[10] Guidotti, E., Tapinassi, L., Toni, L., Bianchi, L., Gaetani, P., and Persico, G., 2011.

Experimental and numerical analysis of the ow eld in the impeller of a centrifugal

compressor stage at design point. In ASME Turbo Expo 2011. ASME Paper 2011-

45036.

[11] Hunziker, R., Dickmann, H., and Emmerich, R., 2001. Numerical and experimental

investigation of a centrifugal compressor with an inducer casing bleed system. Journal

of Power and Energy, 215(6), September, pp. 783791.

[12] Tamaki, H., 2010. Eect of recirculation device on performance of high pressure ratio

centrifugal compressor. In ASME Turbo Expo 2010. ASME Paper 2010-22570.

[13] Sun, Z., Tan, C., and Zhang, D., 2009. Flow eld structures of the impeller backside

cavity and its inuences on the centrifugal compressor. In ASME Turbo Expo 2009.

ASME Paper 2009-59879.

[14] Mischo, B., Seebass-Linggi, C., Ribi, B., and Mauri, S., 2009. Inuence of labyrinth seal

leakage on centrifugal compressor performance. In ASME Turbo Expo 2009. ASME

Paper 2009-59524.

[15] Wang, Z., Xu, L., and Xi, G., 2010. Numerical investigation of the labyrinth seal

design for a low ow coecient centrifugal compressor. In ASME Turbo Expo 2010.

ASME Paper 2010-23096.

[16] Guidotti, E., Tapinassi, L., Naldi, G., and Chockalingam, V., 2012. Cavity ow mod-

eling in an industrial centrifugal compressor stage at design and o-design conditions.

In ASME Turbo Expo 2012. ASME Paper 2012-68288.

[17] Guidotti, E., Tapinassi, L., and Naldi, G., 2012. Inuence of seal leakages modeling on

low ow coecient centrifugal compressor stages performance. In ATI Congress 2012.
 87

[18] Satish, K., Guidotti, E., Rubino, D., Tapinassi, L., and Prasad, S., 2013. Accuracy of

centrifugal compressor stages performance prediction by means of high delity cfd and

validation using advanced aerodynamic probe. In ASME Turbo Expo 2013. ASME

Paper 2013-95618.

[19] Lettieri, C., Baltadjiev, N., Casey, M., and Spakovszky, Z., 2013. Low-ow-coecient

centrifugal compressor design for supercritical co2. In ASME Turbo Expo 2013. ASME

Paper 2013-95012.

[20] Jameson, A., Schmidt, W., and Turkel, E., 1981. Numerical solution of the euler

equations by nite volume methods using runge-kutta time-stepping schemes. In AIAA

14th Fluid and Plasma Dynamics Conference.

[21] Holmes, D. G., Mitchell, B. E., and Lorence, C. B., 1997. Three dimensional linearized

navier-stokes calculations for utter and forced response. In ISUATT Symposium.

[22] Kapetanovic, S. M., N., F. F., G., H. D., and M., O., 2007. Impact of combustor exit

circumferential ow gradients for gas turbine prediction. In In Proceedings of the 5th

International Conference on Heat and Mass transfer.

[23] Wilcox, D. C., 1998. Turbulence Modeling for CFD. DCW Industries, Inc., La Cañada,

CA.

[24] Launder, B. E., M., and Kato, 1993. Modeling ow induced oscillations in turbulent

ow around a square cylinder. In ASME FED, 57, pp 189-199.

[25] Grimaldi, A., Tapinassi, L., Biagi, F. R., Michelassi, V., Bernocchi, A., and Guenard,

D., 2007. Impact of inlet swirl on high-speed high-ow centrifugal stage performance.

In ASME Turbo Expo 2007. ASME Paper GT-2007-27202.

[26] Scotti, A. D. G., Biagi, F. R., Sassanelli, G., and Michelassi, V., 2007. A new slip factor

correlation for centrifugal impellers in a wide range of ow coecients and peripheral

mach numbers. In ASME Turbo Expo 2007. ASME Paper GT-2007-27199.

[27] Aalburg, C., Simpson, A., Schmitz, M. B., Michelassi, V., Evangelisti, S., Belardini, E.,

and Ballarini, V., 2008. Design and testing of multistage centrifugal compressors with

small diusion ratios. In ASME Turbo Expo 2008. ASME Paper GT-2008-51263.
 88

[28] Ramakrishnan, K., Richards, S., Moyroud, F., and Michelassi, V., 2011. Multi-blade

row interactions in a low pressure ratio centrifugal compressor stage with vaned diuser.

In ASME Turbo Expo 2011. ASME Paper 2011-45932.

[29] Guidotti, E., Satish, K., Rubino, D., Tapinassi, L., Prasad, S., Toni, L., and Naldi, G.,

2014. Inuence of cavity ows modeling on centrifugal compressor stages performance

prediction across dierent ow coecient impellers. In ASME Turbo Expo 2014. ASME

Paper 2014-25830.

[30] Erdos, J. I., Alzner, E., and McNally, W., 1997. Numerical solution of periodic tran-

sonic ow through a fan stage. AIAA Journal, 15(11), pp. 15591568.

[31] Suresh, A., and Rangwalla, A., 2012. Stability of the phase-lag boundary condition

in turbomachinery simulations. In 48th AIAA/ASME/SAE/ASEE Joint Propulsion

Conference Exhibit 2012. 2417-2425.

[32] Denton, J., 2010. Some limitations of turbomachinery cfd. In ASME Turbo Expo

2010. ASME Paper 2010-22540.

[33] Cotroneo, J. A., Cole, T. A., and Hofer, D. C., 2007. Aerodynamic design and prototype

testing of a new line of high eciency, high pressure, 50 In ASME Turbo Expo 2007.

ASME Paper GT-2007-27315.

[34] Mangani, L., Casartelli, E., and Mauri, S., 2012. Assessment of various turbulence

models in a high pressure ratio centrifugal compressor with an object oriented cfd

code. Journal of Turbomachinery, p. 134(6).

[35] Panizza, A., Sassanelli, G., iurisci, G., and Sankaran, S., 2012. Performance uncertainty

quantication for centrifugal compressors: 1. stage performance variation. In ASME

Turbo Expo 2012. ASME Paper 2012-68036.

[36] Sankaran, S., Sassanelli, G., iurisci, G., and Panizza, A., 2012. Performance uncertainty

quantication for centrifugal compressors: 2. ange to ange variability. In ASME

Turbo Expo 2012. ASME Paper 2012-68220.

[37] Panizza, A., Rubino, D., and Tapinassi, L., 2014. Ecient uncertainty quantication

of centrifugal compressor performance using polynomial chaos. In ASME Turbo Expo

2014. ASME Paper 2014-25081.

 89

[38] Casey, M. V., Dalbert, P., and Roth, G., 1990. The use of 3d viscous ow calculations

in the design and analysis of industrial centrifugal compressors. ASME Paper 90 Gt-2.

[39] Clayton, R. P., Leong, W. U. A., Sanatian, R., Issa, R. I., and Xi, G., 1998. A numerical

study of the three-dimensional turbulent ow in a high-speed centrifugal compressor.

ASME Paper 98 Gt-49.

[40] Kowalski, S., Pacheco, J. E., Fakhri, S., and Sorokes, J., 2012. Centrifugal stage

performance prediction and validation for high mach number applications. In 41th

Turbomachinery Symposium.
 90

Appendix A

List of Publications
In appendix A it is reported the list of publications that have been presented or

submitted during this reserch work and object of the present thesis.

• Guidotti, E., Tapinassi, L., Toni, L., Bianchi, L., Gaetani, P., and Persico, G., 2011.

Experimental and numerical analysis of the ow eld in the impeller of a centrifugal

compressor stage at design point. In ASME Turbo Expo 2011. ASME Paper 2011-

45036.

• Guidotti, E., Tapinassi, L., Naldi, G., and Chockalingam, V., 2012. Cavity ow mod-

eling in an industrial centrifugal compressor stage at design and o-design conditions.

In ASME Turbo Expo 2012. ASME Paper 2012-68288.

• Guidotti, E., Tapinassi, L., and Naldi, G., 2012. Inuence of seal leakages modeling

on low ow coecient centrifugal compressor stages performance. In ATI Congress

2012.

• Satish, K., Guidotti, E., Rubino, D., Tapinassi, L., and Prasad, S., 2013. Accuracy of

centrifugal compressor stages performance prediction by means of high delity CFD and

validation using advanced aerodynamic probe. In ASME Turbo Expo 2013. ASME

Paper 2013-95618.

• Guidotti, E., Satish, K., Rubino, D., Tapinassi, L., Prasad, S., Toni, L., and Naldi,

G., 2014. Inuence of cavity ows modeling on centrifugal compressor stages per-

formance prediction across dierentow coecient impellers. In ASME Turbo Expo

2014. ASME Paper 2014-25830.

 91

• Guidotti, E., Baldassarre, L., Rubino, D., Tapinassi, L., 2014. On the accuracy of

centrifugal compressor systems performance prediction by means of high delity CFD.

In 43rd Turbomachinery Symposium.