Modelling of Gas Foil Bearings Towards Controllable Operation Multi-Domain Analysis

Downloaded from orbit.dtu.
dk on: 5 02, 2023
Modelling of Gas Foil Bearings Towards Controllable Operation Multi-domain Analysis
von Osmanski, Alexander Sebastian
Publication date:
2020
Document Version
Publisher's PDF, also known as Version of record
Link back to DTU Orbit
Citation (APA):
von Osmanski, A. S. (2020). Modelling of Gas Foil Bearings Towards Controllable Operation Multi-domain
Analysis. Technical University of Denmark. DCAMM Special Report No. S273
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright
owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.
 Users may download and print one copy of any publication from the public portal for the purpose of private study or research.
 You may not further distribute the material or use it for any profit-making activity or commercial gain
 You may freely distribute the URL identifying the publication in the public portal
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately
and investigate your claim.
DTU Mechanical Engineering
Department of Mechanical Engineering
Modelling of Gas Foil Bearings

Towards Controllable Operation
Multi-Domain Analysis
Alexander Sebastian von Osmanski
PhD Thesis
PhD Thesis
Modelling of Gas Foil Bearings

Towards Controllable Operation
Multi-domain Analysis
Sebastian von Osmanski
Kongens Lyngby, Denmark, April 2020

PhD thesis:
Modelling of Gas Foil Bearings Towards Controllable Operation: Multi-domain Analysis
April 2020
ISBN: 978-87-7475-591-3
DCAMM Special Report No. S273
Danish title of the thesis:
Modellering af Gasfolielejer med Henblik på Regulerbar Drift: Multidomæneanalyse
PhD student:
asvosm@mek.dtu.dk or sebastian@osmanski.dk
ORCID: 0000-0002-7008-0670
Principal supervisor:
Professor Ilmar Ferreira Santos, Dr.-Ing. dr.techn. livre-docente
ifs@mek.dtu.dk
ORCID: 0000-0002-8441-5523
Industrial supervisor:
Jon Steffen Larsen, PhD
jonsteffen.larsen@gea.com
Assessment committee:
Associate professor Jon Juel Thomsen, ph.d., dr.techn. (chairman)
Technical University of Denmark (DTU)
Assistant professor Jürg Alexander Schiffmann, PhD.
École Polytechnique Fédérale de Lausanne (EPFL)
Professor Mihaï Arghir, PhD, Dr. Eng.
Université de Poitiers
Front page illustration:
Full spectra of simulated unbalance responses for the group’s GFB test rig with 60 g mm
unbalance at one end. The diagram covers ±333 Hz (horizontal axis) and comprise 1221
steady state time series of 10 s duration in the range 8–20 kRPM (vertical axis).
DTU Mechanical Engineering

Section of Solid Mechanics
Technical University of Denmark
Nils Koppels Allé, Bld. 404
2800 Kongens Lyngby, Denmark
Phone (+45) 45 25 42 50
info@mek.dtu.dk
www.mek.dtu.dk
Abstract
The feasibility of combining gas lubrication with compliant surfaces has been known
since at least the 1960s. These principles are combined in the Gas Foil Bearing
(GFB) offering a versatile, mechanically simple, oil–free and environmentally
friendly mechanism for the support of lightweight high-speed rotating machinery.
The applicability of GFBs is, however, restricted by several factors; a limited
number of start–stop cycles, an inherently low level of damping and rich hard-
to-predict dynamics. This project is motivated by the latter and is aiming at
improving the fundamental understanding of GFB dynamics and the available
modelling tools. Furthermore, it is an objective to leverage the improved modelling
capacity to push towards augmentation of GFBs with injection, eventually enabling
a mechatronic GFB with adaptable properties.
The first two papers are devoted to the modelling of friction. A truss-based
bump foil model, a beam-based top foil model and a smooth friction model are com-
bined and it is concluded that the foil mass must be included to retain simultaneity.
It is demonstrated that a dynamic friction model is not sufficient to explain the
observed discrepancies between simulations and experimental observations, and it
is hypothesized that a model allowing sticking would be required. The third paper
presents an extended perturbation method treating the foil degrees of freedom
explicitly. This is demonstrated to provide onset speeds of instability in better
agreement with the non-linear time integration than the classical perturbation
technique in which only the rotor states are perturbed. The fourth paper presents
an all-new GFB simulation tool. This includes a finite volume-based discretization
of the Reynolds Equation coupled to generic foil and rotor models along with
clearly defined domain interfaces. Importantly, the paper also details the assembly
of the system Jacobians. The fifth paper deals with eigenvalue analysis of the
system Jacobians. The predicted stability limits are shown to exactly match those
of the extended perturbation, while diverging from the classical perturbation as
the compliance level increases. Furthermore, Campbell diagrams are extracted,
and the multi-domain modes are visualized. In the final paper, a scheme for mod-
elling the injection mass flow based on interpolation from Computational Fluid
Dynamics results is coupled to the bearing code. The feasibility of adding injection
to an existing GFB supported test rig is investigated, leading to a predicted 20 %
increase of the stability limit and reduced sub-synchronous vibrations.
Resumé (Danish)
Muligheden for at kombinere gassmøring med fleksible overflader har været kendt
i hvert fald siden 1960’erne. Disse principper forenes i gasfolielejet, som udgør et
alsidigt, mekanisk simpelt, oliefrit og miljøvenligt ophæng til let roterende maskineri
ved høje hastigheder. Anvendeligheden af gasfolielejer indsnævres dog af flere
forhold; et begrænset antal start/stop cyklusser, lav dæmpning og en omfangsrig
svært forudsigelig dynamik. Dette projekt er motiveret af sidstnævnte og har som
mål at forbedre den grundlæggende forståelse af gasfolielejers dynamik samt de
tilgængelige modelleringsværktøjer. Ydermere er det en målsætning at udnytte
de forbedrede modeller til arbejde i retning af gasfolielejer med indsprøjtning,
ultimativt for at muliggøre et mekatronisk gasfolieleje med adaptive egenskaber.
De første to artikler omhandler friktionsmodellering. En gitterbaseret bølgefo-
liemodel, en bjælkebaseret topfoliemodel og en differentiabel friktionsmodel kobles
og det konkluderes, at foliens masse skal medtages for at bibeholde en simultan
løsning. Det demonstreres, at en dynamisk friktionsmodel er utilstrækkelig til at
forklare uoverensstemmelserne med eksperimentelle observationer, og det antages
at en model med statisk friktion ville være nødvendig. Den tredje artikel præ-
senterer en udvidet perturbationsmetode, hvori foliens frihedsgrader behandles
eksplicit. Dette leder til stabilitetsgrænser i bedre overensstemmelse med ikke-
lineær tidsintegration end den klassiske pertubationsmetode, hvori kun rotorens
frihedsgrader perturberes. Den fjerde artikel dokumenterer et ny beregningsværk-
tøj til gasfolielejer. Det omfatter en ”Finite Volume” baseret diskretisering af
Reynolds’ ligning koblet til generelle folie- og rotormodeller med præcist definerede
domænegrænseflader. Desuden er det væsentligt, at artiklen beskriver opbygningen
af systemets Jacobi-matricer. Den femte artikel omhandler egenværdianalyse af
Jacobi-matricerne. Det demonstreres, at de beregnede stabilitetsgrænser stemmer
nøjagtigt overens med forudsigelserne fra den udvidede perturbationsmetode, imens
afvigelsen fra den klassiske perturbation stiger med folieudbøjningen. Slutteligt
genereres Campbelldiagrammer og multidomæne-modalformerne visualiseres. I den
sjette og sidste artikel præsenteres en metode til modellering af indsprøjtningsflow.
Denne er baseret på interpolation af ”Computational Fluid Dynamics” resultater og
kobles til lejekoden. Muligheden for at tilføje indsprøjtning til en eksisterende for-
søgsopstilling med gasfolielejer undersøges og det forudsiges, at stabilitetsgrænsen
kan forøges med 20 % samtidig med, at de sub-synkrone vibrationer formindskes.
Preface
This thesis is submitted in partial fulfilment of the requirements for obtaining
the Danish PhD degree at the Technical University of Denmark (DTU). The
thesis consists of an introduction including a brief summary, a conclusion and six
research papers. The PhD project was conducted at the Department of Mechanical
Engineering (MEK), Section of Solid Mechanics (FAM) in the period from January
2017 to April 2020 under the supervision of Professor Ilmar Ferreira Santos and
Jon Steffen Larsen from GEA Process Engineering A/S. I would like to thank
the department for funding the work and GEA for the time Jon has allocated
to supervision. Furthermore, the project allowed me to visit Professor Minel J.
Braun at the University of Akron, Ohio, USA in the fall of 2018. I am thankful
that this opportunity was granted to me and for the hospitality and guidance I
received during the stay.
When I was first taken under Ilmar’s wing and invited into his electro-
mechanical research group during my graduate studies in 2014, I hardly imagined
this to be the prelude to a six-year journey. I am, however, deeply thankful for
the ride! Ilmar, your enthusiasm, creativity and inexhaustible optimism has been
a great source of inspiration. I sincerely hope that you will continue teaching for
many years to come so that other students may share this experience. Thank you
for our countless talks and for showing me that academia can also be about drink-
ing (wheat) beer or caipirinhas. Admittedly, I have occasionally been challenged
by your Brazilian mindset, but your warmth and sympathetic character more than
makes up for this. To my co-supervisor Jon; you have been an exemplary mentor
along the entire journey. Throughout my ”academic upbringing”, your orderliness,
focus and pragmatism have formed an ideal counterweight to Ilmar’s seemingly
perpetual flow of new ideas. Thank you for the countless drafts covered in red ink
that you have returned, our myriad email exchanges and for our many talks about
technical as well as non-technical matters.
I would like to thank all my colleagues in the research group for sharing so
many hours at DTU, for our discussions and rewarding exchange of ideas. In
particular, thanks to my friends and colleagues Nikolaj A. F. Dagnæs, Jonas
Lauridsen, Alejandro de Miguel and Alejandro C. Varela with whom I have shared
many good moments both in and outside business hours. Thanks to the present
as well as former members of the ”lunch club”; Andreas J. Voigt, Bo B. Nielsen,
x Preface
Cian Conlan-Smith, Hannibal Overgaard, Mikael M. Eronen, Thomas T. Paulsen

and Vergilio T. S. Del Claro for our daily breaks and (occasional) talks about
non-bearing-related topics. Thanks to Cesar A. L. L. da Fonseca for the time here
in Denmark as well as for showing us Rio. Likewise, a special thanks should go to
Troy Snyder for taking care of Mona and me while in Akron.
Of the students that I have had the chance to supervise, I would like to thank
Signe Tophøj Heineman and Henrik Øllgaard Larsen for the countless hours spent
struggling with my computational tools to provide me with invaluable feedback.
Also, thanks to Lars Molzen for providing experimental results and for helping me
understand and operate the test rig.
Lastly, a profound thanks should go to my family and friends outside DTU for
their patience (which has been put to the test!) and everlasting support. And to
the most important of them all; my beloved wife Mona: While giving me the time
and space needed to pursue the PhD degree, you have managed to complete your
own studies, take time off to go with me to the US, form our home and give birth
to our daughter. I have no words for my heartfelt gratitude.
Kongens Lyngby, Denmark, 10 April 2020

Contents
Abstract v
Resumé (Danish) vii
Preface ix
Contents xi
Thesis Nomenclature xv
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Literature Overview . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Included Publications . . . . . . . . . . . . . . . . . . . . . . . . 17
1.4 Original Contributions . . . . . . . . . . . . . . . . . . . . . . . 19
2 Conclusive Remarks & Future Challenges 25

2.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
Publications 31
P1 A Fully Coupled Air Foil Bearing Model Considering Friction

– Theory & Experiment 33
Nomenclature for publication P1 . . . . . . . . . . . . . . . . . . . . . 33
P1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
P1.2 Modelling of the Rigid Rotor . . . . . . . . . . . . . . . . . . . . 38
P1.3 Modelling of the Fluid Film . . . . . . . . . . . . . . . . . . . . 40
P1.4 Modelling of the Foil Structure . . . . . . . . . . . . . . . . . . 42
P1.5 Coupled System of ODEs . . . . . . . . . . . . . . . . . . . . . . 51
P1.6 Results & Discussion . . . . . . . . . . . . . . . . . . . . . . . . 53
P1.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
P1.A Rotor Model Matrices . . . . . . . . . . . . . . . . . . . . . . . 68
P1.B Bump Foil Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . 68
xii Contents
P1.C Validation of Bump Model . . . . . . . . . . . . . . . . . . . . . 68

P1.D Top Foil Matrices and Work Equivalent Load . . . . . . . . . . 69
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
P2 On the Incorporation of Friction into a Simultaneously

Coupled Time Domain Model of a Rigid Rotor Supported by
Air Foil Bearings 71
P2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
P2.2 The Rotor–Bearing System . . . . . . . . . . . . . . . . . . . . . 75
P2.3 Mathematical Model of the Rotor and the Fluid Film . . . . . . 75
P2.4 Modelling of the Foil Structure . . . . . . . . . . . . . . . . . . 78
P2.5 Structure of the Assembled Equation System . . . . . . . . . . . 82
P2.6 Results & Discussion . . . . . . . . . . . . . . . . . . . . . . . . 83
P2.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
P3 The Classical Linearization Technique’s Validity for Compli-

ant Bearings 93
P3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
P3.2 The extended perturbation method . . . . . . . . . . . . . . . . 95
P3.3 Review of the classical two-DOF perturbation . . . . . . . . . . 101
P3.4 Solution strategy . . . . . . . . . . . . . . . . . . . . . . . . . . 102
P3.5 Results & discussion . . . . . . . . . . . . . . . . . . . . . . . . 102
P3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
P4 Modelling of Compliant-type Gas Bearings: A Numerical

Recipe 109
P4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
P4.2 Rotor Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
P4.3 Foil Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
P4.4 Fluid Film Domain . . . . . . . . . . . . . . . . . . . . . . . . . 116
P4.5 Domain Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . 122
P4.6 System Assembly . . . . . . . . . . . . . . . . . . . . . . . . . . 124
P4.7 Steady-state and Transient Solution . . . . . . . . . . . . . . . . 127
P4.8 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
P4.9 Conclusions & Future Aspects . . . . . . . . . . . . . . . . . . . 129
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Contents xiii
P5 Multi-domain Stability and Modal Analysis Applied to

Gas Foil Bearings: Three Approaches 133
P5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
P5.2 Bearing Configuration and Basic Modelling . . . . . . . . . . . . 138
P5.3 Perturbation Methods . . . . . . . . . . . . . . . . . . . . . . . 145
P5.4 Direct Coupling of the Domains . . . . . . . . . . . . . . . . . . 153
P5.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
P5.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
P5.A Jacobian Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . 175
P5.B Foil Structure Mass . . . . . . . . . . . . . . . . . . . . . . . . . 177
P5.C Extended Perturbation Without Foil Mass . . . . . . . . . . . . 178
P5.D Mode Shape with Degenerated Orbit . . . . . . . . . . . . . . . 178
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
P6 Gas Foil Bearings with Radial Injection: Multi-domain Stabil-

ity Analysis and Unbalance Response 181
P6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
P6.2 Hybrid Bearing Configuration . . . . . . . . . . . . . . . . . . . 189
P6.3 Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
P6.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
P6.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
P6.A Isentropic Injection Flow . . . . . . . . . . . . . . . . . . . . . . 248
P6.B Rotor Model Matrices and Forces . . . . . . . . . . . . . . . . . 249
P6.C Foil Structure Matrices and Pressure Load Mapping . . . . . . . 250
P6.D Modified Reynolds Equation Finite Volume Residual Derivatives 252
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
Bibliography 259
Thesis Nomenclature
Abbreviations
AMB Active Magnetic Bearing FV Finite Volume

BC Boundary Condition GFB Gas Foil Bearing
CFD Computational Fluid HGFB Hybrid Gas Foil Bearing
Dynamics IC Initial Condition
DAE Differential/Algebraic IVP Initial Value Problem
Equation ODE Ordinary Differential Equation
OSI Onset Speed of Instability
DOF Degree of Freedom
RE Reynolds Equation
EOM Equation of Motion RPM Revolutions per Minute
FD Finite Difference SEFM Simple Elastic Foundation
FE Finite Element Model
Symbols
(Û) Time derivative, d/dt ε Eccentricity

C Radial clearance ψ Film state variable, ψ = ph
do Injection orifice diameter
f (z, t) Coupled system function
h Film height
p Film pressure z Coupled system state vector
t Time M (z) Generalized mass matrix
CHAPTER 1
Introduction
1.1 Motivation
A significant portion of the technology that has formed our modern society since the
industrial revolution relies on rotating machinery. Mechanical energy conversion
most often involves a transformation from linear to angular motion requiring
the support of rotating machine components, be it crankshafts, impellers, gears,
drive shafts etc. The fundamental task of keeping these components in place has
fostered a plenitude of bearing variants relying on diverse physical mechanisms
such as rolling, viscous shearing or magnetic levitation. The oil lubricated variants
based on rolling elements or oil-based fluid films have been developed to near-
perfection and have demonstrated their reliability and high load carrying capacities
through more than a century of widespread use. This has led to their well-earned
dominance in most mechanical systems, such as internal combustion engines, large
scale turbines and electric motors.
Since the mid twentieth century, technological developments have encouraged
oil-free alternatives based on (electro-)magnetism or gas lubrication. They have
in common the ability to support rotors at very high speeds with low frictional
drag. In itself, the absence of oil represents a significant advantage as it precludes
auxiliary oil cooling and filtration systems, mitigates the risk of process fluid
contamination and allows for operation in extreme environments. In recent years,
the Active Magnetic Bearings (AMBs) have undergone a rapid development and
demonstrated great potential, but they are inherently limited by their complexity,
spatial requirements, power supply and need for auxiliary systems including sensors,
controllers, amplifiers, backup bearings etc. For applications where these issues
are problematic or prohibitive, the Gas Foil Bearing (GFB) presents an interesting
alternative. The GFB belongs to the class of hydrodynamic bearings but is
characterised by its compliant inner surface with a micrometre-scale process gas
film separating the bearing from its rotating counterpart. The low viscosity of gas
implies a vanishing frictional loss even at high speeds and gaseous lubricants do
not tend to melt, vaporize, burn, decompose or freeze as is the case with other
lubricants. The bearing type is self-acting, mechanically simple and requires no
auxiliary systems. Furthermore, GFBs have no practical upper speed limit like the
”DN” number observed for rolling element bearings (though speed will be limited
2 1 Introduction
by stability or risk of thermal runaway) and mean-time-between-failures in excess

of 100 kh have been reported during field usage. The oldest and most widespread
commercial application of GFBs is in aircraft air cycle machines, but GFBs have
likewise been applied in turbo-compressors, cryocoolers and micro-turbines of
ever-increasing sizes. The present work focuses on bump-type GFBs, but other
designs do exist [21].
The components of a GFB, namely the bearing housing along with top, bump
and shimming foils, are illustrated in fig. 1.1. The shown bearing is an industrial
three-pad GFB manufactured by Siemens, two of which are presently installed in
a test rig supporting a 21 kg shaft at up to 30 kRPM. The rig was constructed
in 2012 by Jon S. Larsen (co-supervisor of the present PhD project) and mimics
a Siemens-ACC directly driven oil-free compressor rated at 120 kW. Since then,
the trend of using GFBs in increasingly powerful compressors has continued, and
Korean-based TurboMax currently offers a directly driven GFB supported turbo
blower rated at 600 kW. Nevertheless, despite seven decades of research efforts
and multiple successful applications, the development of GFBs is still inextricably
linked with trial-and-error methodologies and the understanding of their dynamics
is limited.
Any model striving to predict the dynamics of a GFB supported rotor must
cope with three distinct domains as illustrated in fig. 1.2. The rotor can be
described using well-established methods based on point masses, rigid bodies or
finite elements and does not contribute any particular challenges in the present
context. For many widely used process fluids, dry ambient air being one, the
behaviour of the lubricating gas film can be described using the well-established
isothermal Reynolds Equation (RE) assuming ideal gas behaviour. The foil model
presents a greater challenge and can be modelled in numerous ways using one, two
or three dimensions with or without viscous and/or frictional dissipation. As a
whole, the present challenges related to GFB modelling can roughly be traced to
three sources:
Frictional effects in the foils: The foil contributes damping to the overall
rotor–bearing system through sliding friction, but the quantification of
this effect is still an active area of research. The majority of GFB models
are hence limited to equivalent viscous damping with empirically determined
coefficients. At the same time, friction influences the effective foil stiffness
(especially for bump-type GFBs). This stiffening effect of friction is poten-
tially equally significant to frictional dissipation, as it affects the damping
properties of the gas film.
Pressure–compliance interaction: Compared to other bearing types showing

compliance effects, the attainable deformation in GFBs has a much stronger
1.1 Motivation 3
(c)
(b) (d)
(a)
Figure 1.1: Rendering of Siemens three-pad segmented GFB with a nominal

radius of 33.5 mm and length of 53.0 mm: (a) Stainless steel foil for clearance
adjustment (shimming); (b) bump foil (Inconel X750); (c) PTFE coated top foil
(Inconel X750); and (d) bearing housing (AISI 316L).
(a) (b) (c)
Figure 1.2: Three distinct domains to be handled in a GFB model: (a) the rotor;
(b) the lubricating gas film; and (c) the compliant foil structure.
4 1 Introduction
influence on the bearing dynamics. This means that a reasonable model for
the compliant foil structure is needed and its coupling to the fluid domain
becomes essential. The latter is noteworthy, as many existing bearing
prediction codes handle the domains in a segregated fashion where the
equations have been decoupled. Such a solver structure is sketched in fig. 1.3
and requires very small load/time steps to adequately capture the domain
interaction. The latter furthermore implies the need for temporal convergence
studies.
Non-linear nature of the response: Rotordynamics research has traditionally

focused on linear damping and stiffness coefficients, i.e. condensation of the
bearing characteristics into eight (possibly speed and/or frequency dependent)
scalar values. This methodology has likewise been attempted with GFBs
but is fundamentally deficient. This is illustrated in fig. 1.4 showing two
experimentally obtained waterfall diagrams from the aforementioned test rig;
one with the shaft assembly’s residual unbalance only and one with a rather
heavy unbalance mass added to one extremity of the shaft. The residual
unbalance response in fig. 1.4(a) is essentially that of a linear bearing, but
the response in fig. 1.4(b) shows significant sub-synchronous components
stemming from non-linear phenomena. In fact, the relative prominence of the
Initial conditions: ε0 , εÛ0 , etc.
Primary rotor system solver: Algebraic (steady) or IVP (transient)
Film height If applicable

p(t), ε(t), etc.
Solve Reynolds equation Solve other domains:

(non-linear for gas bearings) Temperature, deformation, etc
Fluid film forces
Postprocessing:
Output
Bearing coefficients etc
Figure 1.3: General structure of many, if not most, existing rotordynamic codes
in which the governing equations for the involved domains are decoupled and
solved in a segregated fashion.
1.2 Literature Overview 5
sub-synchronous amplitudes in fig. 1.4(b) is likely to be underestimated since

mechanical and electrical/magnetic run-out at the bearing locations have not
been compensated. For both unbalance levels, this implies that the measured
synchronous vibration amplitudes are too high and is the reason for the
visible (commensurate) higher order harmonics. As the effects causing these
sub-synchronous vibrations cannot be captured by a linear model, non-linear
analysis must be applied to adequately predict GFB behaviour.
In this light, the overall objective of the present thesis is to contribute to the
fundamental understanding of GFB dynamics and to improve the available model-
ling tools. Second, it is an aim of this thesis to extend the modelling capabilities
to encompass GFBs with injection, ultimately to facilitate the development of an
electromechanical GFB with adaptable properties. Arguably, the addition of an
injection system would compromise one of the bearing type’s main advantages,
namely its simplicity, but it is plausible that the achievable improvements would
justify this. In hybrid operation, i.e. without feedback control, injection would
allow heavier rotors to be supported, reduce wear during start–stop and possibly
increase permissible speed ranges. By adding active control, the effective damping
could likely be increased significantly thus constituting an alternative to other
currently investigated damping-enhancing configurations such as GFBs with sealed
squeeze film dampers. The potential of GFBs with injection, be it hybrid or
actively controlled, is hence substantial and such bearings could represent an
important next step towards sustainable support of heavy rotating shafts.
1.2 Literature Overview

The theoretical foundation for gas bearing modelling was essentially established
when RE from 1869 [150] was extended to encompass compressible fluids by
Harrison [58] in 1913. Widespread practical interest in gas lubrication did, however,
not emerge until the mid-1950s where it proved favourable to especially two
applications: gas circulators for nuclear power reactors and gyroscopes for inertial
navigation systems required during prolonged submerged operation by nuclear
submarines [145]. By the end of the decade, gas lubrication had gained popularity
and found usage in applications such as memory storage drums, precision grinders,
hand tools and turbo expanders [47].
The earliest investigation of a ”foil bearing” is probably that by Blok and
Van Rossum [14] from 1953 considering cellophane wrapped halfway around
a 6 cm journal at up to 1500 RPM. At the same time, compliance effects in
combination with gas lubrication was identified at IBM as a nuisance in magnetic
tape drives, where the tape would lift off the recording heads at high speeds
6 1 Introduction
)
µm
8
p(
6
Am
0.5X 1X 4
2
0
25
20
Ω (kRPM)
15
10
0 100 200 300 400 500

Frequency (Hz)
(a)
)
µm
10.0
p(
7.5
Am
0.5X 1X 5.0
2.5
0.0
25
20
Ω (kRPM)
15
10
0 100 200 300 400 500

Frequency (Hz)
(b)
Figure 1.4: Waterfall diagrams from experimental coast downs of a rotor suppor-
ted by two GFBs: (a) with residual unbalance (<ISO G2.5) and (b) with 45 g mm
added to one end, corresponding to approximately 5 times the G2.5 limit. The
experimental data was generated by Lars Molzen as part of his MSc project [115].
Notice that no run-out compensation has been applied.
[51, 106]. During the following years, rapid developments within manufacturing
methods, experimental equipment and computerized solution techniques strongly
facilitated the development of compliant type gas bearings. This led to Garrret
AiResearch developing the first commercially applied GFB for the environmental
control system onboard the McDonnell Douglas DC-10 in the late 1960s [1]. For
this application, the GFB proved superior to existing rolling element bearings and,
today, environmental control systems in virtually all jet airliners utilise GFBs.
Despite success in some areas, the bearing type remains somewhat a niche
technology, but its range of applications is currently widening. The GFB is thus
finding increased usage in various applications involving (directly driven) high-
speed rotating machinery such as turbo blowers or auxiliary power units. For
this reason, the development in recent years has pushed for GFBs to support
increasingly heavy machinery. This is evident when comparing a study from
2007 by Walton et al. [177] on the development of a 100 hp blower to the present
product catalogue of Korean-based Turbomax offering an 800 hp GFB supported
blower. In a slightly different field, NASA strives to create a completely oil-free
turbine engine enabled by GFBs expecting this to allow up to 15 % engine weight
savings [38, 119]. Another interesting perspective is the use of GFBs in automotive
turbochargers. This could be imminent, and if so, would represent a first case of
high-volume serial production of GFBs [39].
1.2.1 Structural Models

In early work from 1977, Koepsel [82] and Zorzi [186] present an elasto-hydrodyn-
amic solution for both a leaf-type foil bearing and a thrust foil bearing. This is
based on a Finite Difference (FD) solution of the isothermal compressible steady
state RE which is iteratively coupled to a Finite Element (FE) solution of the
elasticity problem using plate or beam elements. Considering the available compu-
tational resources at the time, the presented model is impressive. Convergence
of the fluid–structure interaction is said to cause problems, but journal bearing
results using a 35 × 5 hydrodynamic grid and a 160 Degree of Freedom (DOF) foil
structure are presented showing reasonable agreement to experimental data.
In 1983, two papers on compliant journal and thrust bearings, respectively,
were presented by Heshmat et al. [61, 62]. In this work, the foil compliance was
represented using a Winkler type foundation model, i.e. with a stiffness per unit
area, coupled to the isothermal compressible steady state RE through the film
height function. In this approach, the compliant surface is effectively replaced by
a set of radially acting and mutually independent springs and no additional level
of iteration is required for the fluid–structure coupling. Additionally, this work
established the representation of compliant bearing clearance as a sum of three
components: a hardly measurable nominal clearance, the ”rigid” contribution
8 1 Introduction
due to journal movement and the contribution from foil compliance. Using FD,
this allowed RE and hence the static equilibrium of the rotor to be solved while
taking into account the dynamic pressure–compliance interaction. By numerically
perturbing these solutions, it was furthermore demonstrated how linear bearing
stiffness coefficients could be obtained. This approach to compliance modelling in
GFBs is commonly referred to as the Simple Elastic Foundation Model (SEFM)
and remains a popular choice due to its simplicity and computational efficiency.
The inherent assumption of zero dissipation made in [61, 62] was remedied by Peng
and Carpino [132] in 1993 by introducing a mechanical loss factor, thus allowing
the SEFM to capture dissipative contributions from the foil. In the same paper
[132], the SEFM was linked to the analytical perturbation method by Lund [110,
111], thus bypassing the error-prone numerical perturbations. Using these and
various other modifications, the SEFM has since been applied in e.g. [68, 71, 74,
79–81, 92, 93, 101, 135, 136, 162, 175].
As originally used by Heshmat et al. [61, 62], the foundation modulus, or stiff-
ness, of the SEFM is often calculated using the analytical single-bump expression
by Walowit and Anno [176]. This neglects the top foil stiffness, the stiffening
effect of the foil fixation, bump–bump interactions and, importantly, frictional
effects. Accordingly, later theoretical as well as experimental studies have found
this expression to severely underestimate the foil stiffness for an actual multi-bump
foil strip [83, 84, 89, 94]. This was partly addressed by Iordanoff [71] who presented
an enhanced model to predict SEFM stiffness coefficients distinguishing between
”welded” and ”free” bumps while attempting also to capture the stiffening effect of
Coulomb friction. Peng and Carpino [134] presented a model where the damping
caused by dynamic Coulomb friction at the bump–housing contact points was
included through a circumferentially varying equivalent SEFM damping. Others
have implemented the SEFM with coefficients based on experimental data [81] or
separate higher-fidelity models [87, 90].
The stiffness distribution in most GFBs varies along the axial and/or circum-
ferential directions. This can either be unintentional as caused by the foil welding,
sagging or similar, or intentionally engineered to promote certain properties. Based
on the spatial stiffness distribution, DellaCorte and Valco [37] introduced a now
widely accepted categorisation of GFBs into first, second and third generations.
Heshmat et al. [65] and other affiliates from Mohawk Innovative Technology, Inc.
have subsequently presented fourth, fifth and sixth generation GFBs distinguished
by their coatings and permissible temperature ranges, but these extended defini-
tions have not gained currency among other researchers. An experimental study
documenting the stiffness distribution for a commercial second-generation bearing
is given by Rubio and San Andrés [153]. In order to model second and higher gen-
eration bearings, several authors have thus implemented the SEFM with spatially
varying coefficients [68, 74, 93]. Such distributions are essential to the properties
of modern GFBs as evident from the study by Schiffmann and Spakovszky [162].
This predicted that an enhancement of two orders of magnitude to a rotordynamic
stability parameter could be obtained by numerically optimizing the shim-locations
and hence the distribution of stiffness.
1.2.2 Beyond the Simple Elastic Foundation Model

The SEFM remains a convenient choice, but its limitations have given rise to a
variety of alternative and more comprehensive foil structural models. A large
portion of these are concerned with improvements to the top foil modelling using
one-dimensional beam elements [22, 77, 93, 104], two-dimensional plate elements
[68, 98] or three-dimensional shell elements [23, 25, 123]. Though serving to couple
neighbouring foil points and possibly adding some stiffening effects of the top
foil, these models remain linear and dependent on coefficients for the underlying
SEFM-like bump foil support. Interestingly, the opposite strategy is pursued in
[100] where the top foil is neglected while the bump foil is modelled using beam
elements. A model where both the bump foil and the top foil is discretized using
shell elements is given in [102] or in [9] additionally including large deformation
theory. The latter is intended to capture buckling, especially in the bump foil,
and such phenomena are indeed predicted by the analysis in [9]. This is, however,
unlikely to be encountered for commonly applied foil geometries as such large
deformations would lead to rapid deterioration.
Ku and Heshmat [83] presented a non-linear structural model based on analyt-
ical expressions for the stiffness of each bump but taking into account bump–bump
interactions and Coulomb friction. The model was, however, not immediately
suitable for direct coupling into a bearing code and rather strived to produce
equivalent linear coefficients to be used with an SEFM-like model.
San Andrés and Kim [156] presented a detailed comparison of two top foil models
based on one-dimensional beam and two-dimensional plate elements, respectively.
In both cases, the top foil FE mesh was aligned to the FD discretization used for
solving RE (78 by 10 divisions). The bump foil was modelled as a series of axially
distributed springs (with stiffness given from Iordanoff [71]) circumferentially
spaced according to the bump pitch. The model hence represents an improvement
over the pure SEFM as it captures sagging and foil fixation effects, but bump–
bump interactions and frictional dissipation are still neglected. It is found that
both models lead to slightly lower predictions for bearing stiffness and damping
than the SEFM, presumably due to the sagging-induced ”ripples” between bump
summits allowing gas to escape. Furthermore, the two-dimensional plate model
is found to overestimate the top foil deflection, while the one-dimensional beam
element provides both the lowest computational cost and the best agreement to
experimental data. This is in agreement with previous work by Carpino et al.
10 1 Introduction
[23] concluding that membrane stiffness (omitted in the two-dimensional plate

formulation in [156]) serves to prevent axial film height variations.
In the model presented by Carpino and Talmage [26] and Peng and Carpino
[133], the principle of virtual work is applied to develop a non-linear FE-based foil
model incorporating circumferential displacements and Coulomb friction. This was
coupled through an iterative scheme to the steady state and perturbed RE for the
hydrodynamic gas film and used to predict dynamic bearing coefficients. It was
concluded that Coulomb friction contributes only a minor damping contribution
unless the orbit is large or the frictional coefficient small.
A very comprehensive model is due to Fatu and Arghir [42, 43] who implemented
an entire GFB in the commercial FE software suite Abaqus. This included rigid
body representations of the rotor and bearing sleeves, the top and bump foils
as three-dimensional shell elements utilizing large-displacement theory and the
fluid film through a Petrov–Galerkin FE discretization of the unsteady RE with
upwinding. A similar structural model, likewise implemented in Abaqus, but
without the fluid–structure interaction is presented by Żywica [187, 188] and
Żywica and Bagiński [190]. The primary purpose of the extensive model by Fatu
and Arghir [42, 43] was to study the influence of geometrical manufacturing errors
on the dynamic bearing characteristics. Even minor manufacturing errors were
found to have a significant impact due to the induced lack of contact between
foil layers. This matches the findings by Shalash and Schiffmann [163, 164] who
likewise investigated the sensitivity to geometrical manufacturing errors. The
latter thus concluded that a 1 % error in the manufactured bump radius would
entail a 5 % error in compliance. Currently, the majority of models in the literature
assume perfect bearing geometries, but if GFBs are to gain currency and, hence,
be manufactured at lower cost, further research in this direction seems necessary
to improve the design’s robustness.
An interesting family of truss like models exists which are intermediate in com-
plexity to the simplistic SEFM-descendants and comprehensive three-dimensional
FE representations. The model by Le Lez et al. [94] thus represents the bump foil
using a number of interconnected springs requiring approximately three DOFs per
bump. Later, a similar model utilizing a bar/spring element and two rigid links
per bump was presented by Feng and Kaneko [45]. The computational expense
of these models (if not considering friction) is comparable to the SEFM, but the
physical behaviour of the corrugated foil is more closely resembled, including
bump–bump interactions and tangential displacements. The truss model by Le Lez
et al. [94] was originally coupled to a simple frictional model relying on a stick/slip
bookkeeping algorithm. This was later expanded with a regularized friction model
by Petrov and Ewins [137] in [95] and more recently it has been coupled to a
comprehensive frictional modelling approach inspired by contact mechanics [4, 5].
The latter, furthermore, includes the modelling of closed/open contacts allowing
e.g. manufacturing errors to be modelled.

Yet another approach to foil modelling is offered by Hassan and Bonello [59].
They implemented a dynamic model of the bump foil using the commercial FE
software Ansys Mechanical and extracted a modal representation using five modes.
The analysis was furthermore experimentally validated. Frictional effects were
not included, but it is argued that the circumferential displacements are available
and hence friction could be added. It is furthermore concluded that the effect of
bearing curvature on the bump foil stiffness is negligible.
1.2.3 Model Structure

As suggested by the focus of this literature review until now, the main challenges in
GFB modelling relate rather to the foil structure than to the fluid film. The latter
has been successfully modelled using the compressible RE to which analytical, FD,
FE and Finite Volume (FV) solutions are amply documented in the literature.
For typical Mach numbers, the isothermal assumption is justifiable [172] and the
cavitation related Boundary Condition (BC) issues encountered in oil lubrication
[111] are not relevant for gasses. The inertia terms (advective as well as temporal)
neglected when deriving RE have given rise to some work comparing RE to more
comprehensive Computational Fluid Dynamics (CFD) solutions, especially for
oil lubricated bearings [31, 169]. For GFBs, Braun et al. [22] presented a two-
dimensional (radial and circumferential directions) transient Navier–Stokes solution
for an infinitely wide bearing coupled to a beam element top foil supported by
SEFM-like springs. No direct comparison to RE based solutions was provided,
but the results did not suggest inertia to be significant. Imagining a very high
excitation frequency, the inertia terms might become important, but within the
range of Reynolds numbers typically encountered in hydrodynamic gas bearings,
the RE presents a reasonable approximation [172]. It is thus natural that GFB
models are typically distinguished primarily by the approach applied for the foil
structure [21].
Other important distinctions should, however, be made to navigate the variety
of GFB models. Unlike the linear incompressible RE used for oil lubricated
bearings, the compressible RE is non-linear (though relatively weakly for the
commonly assumed combination of isothermal conditions and ideal gas behaviour).
To simulate a full GFB, couplings between the rotor, fluid and foil domains must
furthermore be established adding complexity and possibly supplying additional
non-linearity. A gas bearing simulation must hence encompass a non-linear solver
at some stage. This is, however, not to be confused with the character of the
intended rotordynamic analysis, where a distinction should be made between linear
and non-linear approaches. In a linear analysis, as performed by e.g. [68, 79–81, 87,
93, 101, 132, 175], the objective is a linear representation of the bearing in terms
12 1 Introduction
of eight stiffness and damping coefficients as functions of rotor speed and, possibly,
the frequency of oscillation. Subsequently, the coefficients can then be applied to
perform classical rotordynamic calculations to assess the Onset Speed of Instability
(OSI) or simulate an unbalance response. The family of non-linear analyses is
manifold. Studies of time series obtained from non-linear time integrations are
presented in [6, 10, 18, 19, 69, 90, 139], a numerical bifurcation analysis is given in
[105] and a study of limit cycles based on harmonic balancing is presented in [19,
139].
Second, a distinction should be made between models formulated in a segreg-
ated/uncoupled fashion and fully coupled models providing simultaneous solutions.
The majority of the presented literature belongs to the segregated category and
thus employ solver structures similar to the illustration in fig. 1.3. In a linear
analysis using such a structure, the first step is to obtain a solution either to
the steady RE if applying numerical perturbations or to the zeroth order RE in
case an analytical perturbation is employed. An outer loop (Newton–Raphson or
similar) searches for the rotor equilibrium, while an inner loop solves RE at a given
rotor position. If compliance is absent (rigid bearing) or introduced implicitly,
no further loops/couplings are required, but more elaborate foil structure models
usually introduce an additional level of iteration (or several) for the fluid–structure
coupling. Avoiding this extra level of iteration is a significant advantage of the
SEFM, but other models sharing this property do exist. One example is presented
by Carpino and Talmage [25] who developed a hybrid quadrilateral four-node
44-DOF FE containing both the hydrodynamic model (isothermal steady RE)
and a foil model (SEFM-supported cylindrical shell elements). Having obtained
the static solution, regardless of the employed structural model, this is either
perturbed to obtain resulting changes in bearing forces or the corresponding linear
first order equations are solved for the dynamic pressure field(s).
Non-linear rotordynamic analysis is likewise possible using a segregated struc-
ture. In that case, the outermost loop represents a time marching algorithm, i.e.
an Initial Value Problem (IVP) solver for the rotor system of Ordinary Differential
Equations (ODEs). Additionally, the unsteady terms of RE (if included) are
substituted by temporal FD approximations as needed. Such a non-linear transient
analysis was performed as early as 1967 by Castelli and McCabe [27] for a gas
lubricated tilting pad bearing. It was highlighted that the analysis was ”similar
to making an experimental run” and that ”computer time is cheap compared to
test rig failures”. Nevertheless, such analysis has not gained popularity until more
recent years due to the rapid development of available computational resources.
Using a segregated solver structure, Wang and Chen [178] presented a non-linear
analysis of a rigid gas bearing where the unsteady RE was discretized using FD
schemes in both the temporal (backward difference) and the spatial (central differ-
ence) dimensions. They integrated the system explicitly in time till reaching steady
state conditions allowing limit cycle and bifurcation analysis. Arghir et al. [6]
implemented an FV based RE discretization, notably using an unstructured grid,
and calculated the fluid film reaction forces due to sinusoidal rotor displacement
perturbations. The bearing coefficients could then be obtained from the time
series using a least squares method allowing the amplitude dependency to be
assessed. For GFBs, non-linear time integrations using segregated solvers have
been presented by e.g. Hoffmann et al. [69], Kim [74], and Song and Kim [170].
From an implementation standpoint, the segregated structure is easy to handle
since the involved models for each domain can be treated independently and with
few limitations. Historically, it has been an advantage that the equation systems
belonging to each domain could be solved separately putting less demand on
computer memory. From a mathematical point of view, the structure implies that
none of the involved solver routines possess full knowledge of the problem and that
a time lag between the solution variables is inevitable in time-integrations. The
latter entails a requirement for small time steps and temporal convergence studies,
while the former can lead to convergence issues since the direction and magnitude of
steps taken by one solver might not be feasible to the other domains. Furthermore,
a larger number of iterations in the inner loops are usually spent sub-optimally
trying to achieve converged sub-domain solutions at every incremental state passed
through by the outer solvers.
In 2013, Pham and Bonello [139] presented a ”dynamical system representation”
of a GFB model. Instead of decoupling the domain equations and transforming
RE into an algebraic equation system using FD for the unsteady terms, all three
domains were cast into systems of explicit first order ODEs and coupled directly,
i.e. as a system of the form zÛ = f (z, t). The ODE representation of RE was
obtained using an alternative state variable ψ = ph (as also used in e.g. [27]),
while a first order ODE was included for each discrete circumferential foil point
using the SEFM’s stiffness and damping. The fully coupled formulation has
multiple advantages. It discards with the time lag between variables and allows
standardized, readily available and highly efficient numerical tools to be employed
for time integration as well as static solutions. An explicit Runge–Kutta method
is used in [139], but the format is well suited for implicit time stepping algorithms
which usually allow for significantly larger time steps. This is particularly beneficial
if computationally efficient means for obtaining the coupled system Jacobian matrix
are available. A method for obtaining the Jacobian based on symbolic computations
is detailed by Bonello and Pham [19] and used in conjunction with an implicit time
integrator. In this paper, it is furthermore demonstrated how the dynamic system
representation can be used to perform linear analysis of a static equilibrium based
on the full Jacobian eigenvalues. This ability is unique to the fully coupled models.
More recently, Bonello [15, 16] further demonstrated how a Campbell diagram can
be extracted from such eigenvalues.
14 1 Introduction
Following the original work of Pham and Bonello [139], simultaneous formu-
lations have been applied by multiple authors. Baum et al. [10] presented a
simultaneously formulated GFB model incorporating a static foil model and per-
formed bifurcation analysis based on the (FD approximated) Jacobian eigenvalues.
Larsen and Santos [90] presented a model of an industrial rotor supported by two
GFBs with RE discresized using FE. This was successfully integrated in time using
a variable step-size implicit solver, but likewise relying on FD approximations to
the Jacobian. Leister et al. [104] presented a simultaneously formulated model
encompassing an FD discretization of RE, a point mass rotor and three variations
of the SEFM. An explicit Euler time integration scheme was applied, i.e. without
expressions for the Jacobian. A very interesting model was presented by Gu et al.
[52] to describe a rigid rotor supported by two flexibly supported GFBs. The
alternative state variable ψ = ph and explicit inversion of the mass matrices were
circumvented by casting the rotor, bearing housing, SEFM foil structure and fluid
film ODEs into implicit form, i.e. as M (z) zÛ = f (z, t). Importantly, this allowed
the Jacobian matrix to be written directly without an intermediate step requiring
symbolic calculations. Obviously, the implicit format is impractical in combination
with an explicit time integrator as a non-linear system of equations needs be solved
to obtain zÛ . A main virtue of the simultaneous formulations is, however, the ability
to utilize implicit time integrators where such a solution is required in any case.
From a research perspective, a simultaneous framework provides slightly less
flexibility than a segregated structure since all components must be represented in
terms of ODEs. For some algorithms, e.g. the flow charts for determining stick/slip
states presented by some authors [4, 100], this is inconvenient or puts certain
requirements on the remaining model. For example, it was found in the present
PhD project (publications P1 and P2) that the inclusion of foil inertia is necessary
in order to incorporate a frictional model into a simultaneous framework, a finding
that has later been utilized by Leister et al. [105]. For modern computers, the
advantages of the simultaneous formulation are, however, significant in terms of
solution time and available standard analysis methods.
1.2.4 Hybrid Gas Foil Bearings

GFBs are widely recognized as a promising technology for future oil-free tur-
bomachinery. However, application of GFBs to support increasingly heavy rotor
assemblies implies many technical challenges due to their limited load capacity,
low damping capabilities and the potential for significant wear during start–stops.
Enhanced coatings and carefully engineered foil properties have provided significant
improvements to the passive GFBs, but as the technology matures, hybridization
is becoming advantageous or even necessary for further advances.
A hybrid foil–magnetic bearing was proposed by Heshmat et al. [63] in which

an AMB was mounted axially adjacent to a GFB. A now-expired patent was
filed in 1998 by the same authors describing a similar configuration, but with the
GFB and AMB axially co-located [64]. Theoretically, the combination appears
attractive. Magnetically levitating the rotor at low speeds would allow the wear on
foil and journal surfaces to be eliminated. By designing the AMB controller with
the operating point of the GFB in mind, the hydrodynamic load carrying capacity
could be exploited at higher speeds to allow a less powerful design than required for
a stand-alone AMB. At the same time, the AMB could serve to supply additional
damping in speed-ranges problematic for the GFB and the GFB would double as
a backup bearing handling the event of a rotor drop. Potentially, this could even
allow for continuous operation following an AMB failure. Swanson et al. [171]
presented a promising experimental investigation of such a bearing configuration,
but no practical applications seem to have appeared and subsequent research is
limited [11, 73, 140, 181]. In practice, it is questionable if the advantages of a
hybrid GFB–AMB would justify its added complexity compared to a pure AMB.
Likewise, the permissible temperature range of such a hybrid bearing would be
limited by the AMB thus forfeiting a significant GFB advantage.
Another approach to hybridization of GFBs is through the addition of injection
as commonly used for rigid gas bearings [12, 29, 48, 143]. Theoretically, this
could provide advantages similar to those of a GFB–AMB hybrid, but with less
complexity. The first analysis of injection in combination with compliance is from
Eshel [41]. While of limited value in relation to modern GFBs, he did present an
analysis of a cylindrical hydrostatic foil bearing arrangement (resembling magnetic
tape running over a recording head) while considering two axial line sources.
Carpino and Peng [24] modelled another early example of such a configuration in
1994 comprising a purely hydrostatic GFB with an incompressible lubricant being
supplied from the journal centre through capillary tubes to recessed pockets on the
journal surface. Another configuration, intended for a compliant thrust bearing, is
presented by Bosley and Miller [20] where air is injected to the backside of the foil.
Radil and Batcho [147] and Shrestha et al. [167] experimentally investigated
the feasibility of air injection as a means for thermal management as an alternative
to axial flow cooling. In both setups, air was thus fed into the groove between
leading and trailing pad edges. The method is considered viable as it is comparable
in performance to axial flow cooling or slightly better but provides smaller thermal
gradients.
A Hybrid Gas Foil Bearing (HGFB) with injection through orifices attached
to the top foil was first introduced by Kim and Park [77]. Their prototype was
based on a compression spring type GFB from [170]. This was augmented with
four steel supply tubes attached to 0.5 mm orifices in the top foil using rubber
connection hoses and epoxy glue. The HGFB is shown experimentally to provide
16 1 Introduction
a slight increase in load carrying capacity even for a limited supply of injected air
flow (compared to what is usually supplied for cooling) and theoretical predictions
indicate the OSI to be significantly improved. The bearing coefficients of the same
HGFB configuration are calculated as a function of injection pressures in [85]. A
three-pad preloaded HGFB with three 1 mm orifices is presented and analysed
by Kim and Lee [75] designed to carry a 25 kg 101.6 mm rotor at 25 kRPM. As
opposed to the previous design, the steel orifice tubes were laser welded to the top
foil. The same bearing design was further investigated in [78] with loads up to
445 N at 10 kRPM. It was shown that 1000 start–stop cycles could be endured
without visible wear for a 356 N load. In [179], the dynamic coefficient of the
same bearing was measured experimentally using a floating bearing setup showing
reasonable agreement with theoretical predictions.
Using, again, the three-pad HGFB mentioned above, Yazdi and Kim [182]
introduced a ”controlled hybrid mode”, in which the lowermost injector (at 180◦ ) is
shut off above 6 kRPM providing inflow only from the topmost injectors at 60◦ and
300◦ . The operational mode does not apply closed-loop control and should hence
more accurately be described as being ”regulated”. Nevertheless, the hybrid mode
is experimentally shown to significantly delay the appearance of sub-synchronous
vibrations, but limited agreement to theoretical predictions is achieved. A similar
study aiming to locate the optimal locations of the injection orifices is presented
by Yazdi and Kim [183] for a single-pad HGFB utilizing the same on–off injection
scheme.
A novel 110 mm HGFB design incorporating eight hermetically sealed squeeze
film dampers is presented by Delgado and Ertas [36] and Ertas and Delgado
[40] aiming at turbomachinery in the MW range. It is highlighted that previous
efforts related to actively controlled injection in rigid gas bearings [117, 142] have
demonstrated significant improvements in terms of damping to be attainable,
but that the rigid bearings are unsuitable for supporting heavy high-performance
turbomachines. Experimental results as well as numerical predictions are promising,
but the design is relatively complex and involves many components.
1.2.5 Injection Modelling

For injection modelling in oil bearings, it is common to describe the mass flow
using an assumption of fully developed incompressible Hagen–Poiseuille flow [28,
158]. This assumption has likewise been applied for gas injection by Morosi and
Santos [117] and Pierart and Santos [142] in combination with variable discharge
coefficients, while Rowe [152] provides tabulated expressions for the volumetric flow
through various restrictor geometries including corrections for the compressibility.
A more commonly applied analytical expression for gas injection can be derived
by assuming an isentropic process from a supply reservoir with stagnant conditions.
1.3 Included Publications 17
The flow is hence assumed to be adiabatic, steady and inviscid with the flow
rate being limited by compressibility effects. This is in stark contrast to the
assumptions implied by the Hagen–Poiseuille flow where viscosity, and hence wall
shear, is the limiting factor. The isentropic inflow assumption for injection in gas
bearings can be traced to as early as 1964 where Lund [109] calculated the orifice
mass flow in a hydrostatic gas journal bearing using these formulas. Later, the
same assumption was applied by multiple authors, e.g. [6, 12, 30, 35, 57, 60, 75, 77,
85, 97, 120, 184], often crediting a ”Mechanical Technology Inc. (MTI) gas bearing
design manual” from the late 1960s. When deriving the isentropic expression, the
flow is assumed to be limited at a certain choking area and the flow through this
cross section is assumed uniform. This requires the choking area and a correction
factor to be defined. The latter is usually denoted as a ”discharge coefficient”,
”inherent compensation factor”, ”adiabatic efficiency factor” or ”vena contracta
coefficient” attempting to describe the physical phenomenon being compensated
for. Often, this coefficient is furthermore assumed to vary with the pressure ratio.
Depending on the injection system geometry, various different expressions for the
choking area are applied. Arghir et al. [6] and Kim and Park [77] use the curtain
area in the fluid film defined as πdo h, while Kim and Lee [75], Kumar and Kim
[85], and Lee and Kim [97] use the nominal curtain area πdoC and Neves et al.
[120] use the orifice area πdo2 /4. More involved expressions for the choking area
considering both orifice and inherent restriction (=curtain) areas are presented
by Delgado and Ertas [35], Hassini et al. [60], and Zhang et al. [184]. These are
furthermore characterised by a separate model for the injection recess. In [60],
a study on the discrepancy between CFD simulated mass flows as a function of
area relations and pressure ratios is given. It is shown that an almost constant
discharge coefficient is sufficient if the restricting area is correctly defined. Neves
et al. [120] likewise present a comparison of isentropic injection mass flows to
CFD predictions. They find a constant correction factor of 0.88 to be adequate
for sonic/choked conditions, while a linear function of pressure ratio (ranging
from 0.83 to 0.88) is used for subsonic conditions. Other similar studies can be
found in the literature with varying conclusions, accentuating the requirement
for separate numerical or experimental investigations to be performed for each
injector geometry. Alternatively, the use of the isentropic injector model should be
abandoned entirely in favour of more comprehensive fluid models for the injector
area or the entire fluid film.
1.3 Included Publications

At the Technical University of Denmark, a PhD thesis can be submitted in one of
two formats: as a self-contained monograph or as a collection of scientific articles
18 1 Introduction
accompanied by a synopsis. As recommended, the present thesis has been prepared

in the latter form and should hence be regarded as a synopsis linking the following
papers:
P1 S. von Osmanski, J. S. Larsen, and I. F. Santos. “A fully coupled air foil

bearing model considering friction – Theory & experiment”. In: J. Sound
Vib. 400 (2017), pp. 660–679. doi: 10.1016/j.jsv.2017.04.008
P2 S. von Osmanski, J. S. Larsen, and I. F. Santos. “On the incorporation of

friction into a simultaneously coupled time domain model of a rigid rotor
supported by air foil bearings”. In: Tech. Mech. 37.2-5 (2017), pp. 291–302.
doi: 10.24352/UB.OVGU-2017-105
P3 S. von Osmanski, J. S. Larsen, and I. F. Santos. “The Classical Linearization

Technique’s Validity for Compliant Bearings”. In: Proc. 10th Int. Conf.
Rotor Dyn. - IFToMM vol.1. Ed. by K. L. Cavalca and H. I. Weber. Vol. 60.
Mechanisms and Machine Science. Springer International Publishing, 2019,
pp. 177–191. isbn: 978-3-319-99261-7. doi: 10.1007/978-3-319-99262-4
_13
P4 S. von Osmanski, J. S. Larsen, and I. F. Santos. “Modelling of Compliant-type

Gas Bearings: A Numerical Recipe”. In: 13th Int. Conf. Dyn. Rotating
Mach. (SIRM 2019). Ed. by I. Santos. February. Kgs. Lyngby: Technical
University of Denmark (DTU), 2019, pp. 14–27. isbn: 978-87-7475-568-5
P5 S. von Osmanski, J. S. Larsen, and I. F. Santos. “Multi-domain stability

and modal analysis applied to Gas Foil Bearings: Three approaches”. In: J.
Sound Vib. (2020), p. 115174. doi: 10.1016/j.jsv.2020.115174
P6 S. von Osmanski and I. F. Santos. “Gas Foil Bearings with Radial Injection:
Multi-domain Stability Analysis and Unbalance Response”. (submitted for
publication in the Journal of Sound and Vibration, 7 April 2020)
The thesis includes postprints of the publications P1 to P5 which have already

been published and a preprint of publication P6 which has been submitted for
publication at the time of writing. The original contributions of each paper will
be highlighted in the subsequent section.
Notice that the original nomenclatures (if any) are included for each paper as
these are not entirely consistent, whereas the original paper bibliographies are left
out to allow consistent numbering of references. The reference numbers in the
included article post/preprints are thus referring to the thesis bibliography.
1.4 Original Contributions 19
1.4 Original Contributions

The contributions of the thesis can be grouped as follows:
1.4.1 Publications P1 and P2: Foil Structure & Friction

The objective of this first part of the project is related to foil modelling and the
influence of friction. Using the SEFM with carefully chosen stiffness and damping
coefficients, the supervisors of the present project had already achieved notable
agreement with experimental results in terms of unbalance response [90], but the
underlying ”engineering assumptions” were difficult to generalize.
In publication P1, the radially acting spring–dashpot pairs of the SEFM are
thus replaced by an alternative foil structure model aiming to capture directly
the stiffening and dissipative effects of friction. The bump foil is represented
by a warren-like truss model by Le Lez et al. [94] as it encompasses bump–
bump interactions, provides coefficients based on measurable quantities, includes
tangential sliding and is feasible for time integration. The truss is coupled to a
top foil model based on Euler–Bernoulli beam elements and a dynamic smooth
friction model is added. This forms a new compliant foil structure model which
is combined with an FE based discretization of RE and a four-DOF rigid rotor
previously used in [88, 90, 91]. The resulting GFB model lacks analytical Jacobian
matrices, but is formulated as a system of explicit first order ODEs, allowing the
state equations to be simultaneously solved. By adding a static foil model, a system
of Differential/Algebraic Equations (DAEs) would have emerged being solvable
only in the absence of friction. Since a frictional model is included, instantaneous
expressions for the sliding velocities are required to retain simultaneity of the
solution thus requiring differential equations. First order foil ODEs could have
been formed using viscous dampers alone, but this would imply time-discontinuous
velocities and undermine the objective of abandoning viscous dissipation. Instead,
the foil mass is included to form Equations of Motion (EOMs) for the foil with
readily available continuous velocities. The entailed practical issues related to
ill-conditioning and undesired high-frequency dynamics are addressed using non-
dimensionalization and a stiffness proportional damping. A main contribution
of publication P1 is thus the reconciliation of a sliding frictional model with a
simultaneous solution framework for GFBs.
It is demonstrated that a dynamic, i.e. continuously sliding, frictional model
is not sufficient to capture the experimentally observed unbalance response in
terms of sub-synchronous vibration components. On the contrary, it is shown that
promising results can be obtained if all housing–foil contact points are constrained
to mimic permanent sticking. In this configuration, the foil behaves excessively
20 1 Introduction
stiff and contributes zero damping, suggesting that the stiffening effect of friction
is at least as important as the dissipative one.
Essentially, publication P2 addresses the same fundamental challenges related to
frictional modelling and is based on the same model. A more general discussion on
the issues related to a massless model with friction is provided and the implications
of a DAE system are elaborated. Furthermore, the influence of the dynamic
frictional model on the rotor’s equilibrium position is investigated in greater detail.
Jointly, the results contribute the finding that sticking must be a prevalent state.
This, on the other hand, would imply that sliding friction – which is providing the
widely accepted dissipative properties of the foil – must be localized to a subset of
bumps and hence that a larger portion of the damping is supplied by the gas film
than previously anticipated.
1.4.2 Publication P4: General Modelling and Jacobians

The implementation applied in the former work suffers from several limitations
including: (a) it is tailored to a specific test rig comprising two identical GFBs
supporting a rigid shaft; (b) it exists in several parallelly maintained versions using
different foil models; and (c) the explicit equation format complicates a general
formulation of the Jacobian matrices.
In publication P4, a generic approach to GFB modelling based on a fully
implicit equation form and its implementation is presented. The three domains
– fluid film, compliant structure and rotor – are treated independently and their
interfaces are clearly defined in terms of mapping matrices. This allows the
Jacobians to be directly formulated and a number of rather programming-intensive
optimizations to be reused across different rotor–bearing systems. Support for
non-uniform fluid film meshes being non-coincident with the foil mesh is also added
allowing disjoint convergence studies. In order to facilitate a later inclusion of
radial gas injection, an FV based discretization (inspired by [6]) is substituted for
the FE based one, thus allowing upwinding to be applied more easily and global
conservation to be guaranteed.
The implicit format of the equations furthermore implies that the unsteady
and steady terms become simple additions to the system residual. This allows the
exact same function and Jacobian routines to be used both for time-integration
(IVP solution) and for obtaining static equilibria (algebraic equation solution).
Finally, selected results are validated against an existing rigid gas bearing code
as well as the former SEFM-based GFB code.
1.4.3 Publications P3 and P5: Perturbation Methods

In publication P3, an extension of the classical two-DOF perturbation method for
obtaining linear bearing coefficients is presented. The purpose of this extension is
to explain the discrepancies between non-linear time-integration and perturbation
results in terms of the OSI observed in e.g. [88].
Classically, in the method developed by Lund, only the rotor DOFs are per-
turbed. When the method has later been applied to compliant bearings, the foil
deformation has thus been implicitly included. This preferential treatment of the
rotor DOFs is hard to justify in a compliant bearing, where the fluid film is just
as susceptible to foil deformation as to rotor movement. The starting point is
hence to include the foil deformation explicitly in the Taylor series expansion,
analogously the treatment of pad DOFs in tilting pad journal bearings [157]. This
is achieved in a straightforward manner where the discrete foil DOFs are perturbed
directly, meaning that a complex-valued first order pressure field is added for every
foil DOF (in the order of a 100). This causes the computational complexity of the
extended version to increase manifold, but it is demonstrated that the predicted
OSIs are in better, but still imperfect, agreement with results from the SEFM
based time-integration code by Larsen and Santos [90]. This improved level of
agreement suggests that the discrepancy could be explained, at least partly, by
the omitted pressure–compliance interaction in the Taylor series.
In publication P5, the general GFB model from publication P4 is augmented
with an FE based RE discretization (including analytically formulated Jacobians)
and several issues with regard to the consistency between the models compared in
publication P3 are identified. The issues are related to the axial averaging of the
pressure in the first order equations, the enforcement of the Gümbel condition and
the choice of frequency in the time-domain equivalent viscous damping coefficient.
Having addressed these, it is demonstrated in publication P5 that the OSIs found
from eigenvalue analysis based on the coupled system Jacobian are in perfect
agreement with those from the extended perturbation method over a wide range
of compliance levels. This is compared to the classical two-DOF perturbation
method where the OSI discrepancy increases with the compliance level. As already
indicated by the results in publication P3, it is therefore concluded that the
omission of the pressure–compliance interaction term(s) in the Taylor series is
fundamentally erroneous for compliant type bearings with significant flexibility.
This is a primary contribution of the paper as well as the thesis, since the two-DOF
perturbation has been widely applied for GFB analysis since the early 1990s.
It should be emphasized that while managing to capture the pressure–compli-
ance interaction similarly to the fully coupled model, the presented extension is
much too computationally expensive to be workable in practice. Depending on
the foil model, this could possibly be remedied by perturbing the coefficients of a
22 1 Introduction
Chebyshev polynomial or similar instead of discrete DOFs directly. The purpose

of the extension was, however, never to present a feasible perturbation method for
compliant bearings, but rather to expose the root-cause of the classical method’s
deficiency for compliant bearings. In a GFB with high compliance levels, the linear
bearing coefficients are of limited value in any case and one should resort to fully
coupled models.
1.4.4 Publication P5: Multi-Domain Modal Analysis

Apart from the perturbation results highlighted above, it is a novel contribution of
publication P5 to perform and visualize modal analysis across all three included
domains. Each eigenvector is thus decomposed into its rotor, foil and pressure
components and these are visualized over an oscillation cycle. The resulting modes
are compared to those from the classical perturbation providing rotor movements
only and the extended perturbation providing rotor and foil movements. This
makes the domain interaction clearly visible and serves to expose the coupled
nature of GFB dynamics.
1.4.5 Publication P6: Injection

In existing work on gas bearing injection, the injected mass flow is calculated from
analytical expressions with constant or varying correction factors. In publication P6,
a method for determining the injection mass flow based on direct interpolation from
CFD results is presented. The mass flow is thus interpolated from a pre-generated
map of steady state solutions to the three-dimensional Navier–Stokes equations
for a submodel of the injector. Computationally, the map is relatively expensive
to calculate, but the subsequent use in rotor–bearing simulations is very efficient.
An injection flow map comprising 2940 simulations covering a five-dimensional
parameter space spanned by rotational velocity, injector opening, equivalent ec-
centricity, fluid film pressure and injection pressure is presented and discussed.
For a given configuration encountered in a rotor–bearing steady state solution or
time-integration, the injected mass flow and its derivatives with respect to the
map parameters are calculated using an N-dimensional piecewise multi-linear inter-
polation scheme from the literature. This allows the injection flow contributions
to the FV residuals as well as to the Jacobians to be efficiently calculated.
This approach is applied to model the previously mentioned GFB test rig
augmented with an existing gas injection system presently attached to a rigid
gas bearing. A hybrid operation scheme is suggested with injection only from
the topmost injector (akin to the configuration in [182]) and this is shown to
provide a 20 % increase to the predicted stability limit along with a reduction of
sub-synchronous amplitudes. This result indicates that GFB dynamics can be
affected significantly and hence provides an indication of the potential of more

elaborate hybrid or actively controlled injection schemes. Additionally, a simplistic
aerostatic configuration with equal pressurisation of all three injectors is presented
and it is predicted that the 21 kg rotor can be levitated at 0 RPM if supplying
approximately 3.25 bar. This is a noteworthy result, as an aerostatically aided lift-
off could serve to reduce or avoid devastating foil–rotor contact during start–stop
cycles for heavy HGFB supported rotors.
To provide a benchmark for the hybrid case, the stability limit and multi-
domain modal analysis results are likewise presented for the case without injection.
Such results have previously been reported for the exact same rotor–bearing
configuration by Bonello [16] allowing for a direct comparison. Despite multiple
modelling differences, the results are found to be in near-perfect agreement serving
to validate both models.
Finally, it is observed from the combined Campbell diagrams and unbalance
response plots that the abrupt appearances, bifurcations and disappearances of
sub-synchronous vibration components coincide with the intersections of the half
speed and the natural frequencies. Though observed both for the hybrid and
passive cases, no attempt has been made to prove this in general. If, however, these
crossings could be shown to generally govern the appearance of sub-synchronous
vibrations in GFB systems, this would provide valuable knowledge for the design of
HGFB controllers. Using feedback control, it is possible to manipulate the natural
frequencies online, meaning that this could provide a means for avoiding/reducing
sub-synchronous vibrations in HGFBs.
CHAPTER 2
Conclusive Remarks
& Future Challenges
2.1 Conclusions
The PhD project documented in the present thesis including the enclosed public-
ations was fundamentally focused on the theoretical aspects of GFB modelling,
with a particular focus on the improvements necessary to pursue hybridization
and subsequent augmentation with feedback control. The work has comprised
the development and implementation of multiple mathematical models in the
frequency as well as the time-domain. The work has been documented in six
articles (of which five are published at the time of writing) forming the basis of
the present thesis. The conclusions can be grouped as follows:
2.1.1 The Role of Friction

An initial objective of the project was an enhanced foil structural model discarding
with the use of empirically determined coefficients as needed by the SEFM and its
descendants. A model capturing the structural properties well was implemented
and coupled into a simultaneous solution framework. The work identified the
necessity of including the foil mass in order to combine friction with a simultaneous
formulation and managed to address the consequent challenges. The results
indicated the stiffening effect of friction to be (at least) as significant as the direct
dissipative effect. Unfortunately, this also implied that the chosen dynamic friction
model was incapable of correctly capturing the foil behaviour, suggesting a model
encompassing frictional sticking to be necessary. This conclusion is interesting
from a modelling perspective, but also questions the fundamental role of friction in
a GFB. If sticking is in fact a prevalent state in the foil structure, sliding friction
can hardly be a root-cause of significant damping as commonly assumed.
26 2 Conclusive Remarks & Future Challenges
2.1.2 General Modelling of Gas Foil Bearings

The barrier-of-entry for theoretical work on GFBs is relatively high compared
to other bearing types due to their non-linear characteristics and coupling of
multiple domains with different challenges. At the onset of the present PhD
project, simultaneous formulations were already available [19, 52, 90], but the
present work has contributed a very general and highly efficient framework for
such models. Importantly, this includes formulations of the essential analytical
Jacobian matrices without the need for configuration-specific symbolic calculations.
Mathematically, the main contribution is a clear-cut formulation of the domain
interfaces in terms of mapping matrices, but the efficiency of the model should
likewise be attributed to its implementation in C code. Hopefully, this will serve
to reduce the necessary footwork in future GFB related projects, as the inclusion
of extended sub-models e.g. to include manufacturing defects [42, 43, 163, 164],
an enhanced frictional model [4, 5, 105] or a modal foil representation [59] would
be relatively simple. The possibility of publishing the code is currently being
investigated.
2.1.3 Perturbation Techniques for Compliant Bearings

Linear coefficients for hydrodynamic bearings are commonly found semi-analytically
using the perturbation technique initially proposed by Lund [110] in the 1960s (and
later presented in more familiar formats [108, 111]). Using an implicit compliance
model, the perturbation has also been widely applied to compliant type bearings
since the early 1990s. In the present project, this method is demonstrated to
be fundamentally erroneous in the presence of considerable flexibility, as the
pressure–compliance terms of the underlying Taylor series are unintentionally
omitted. An extension of the perturbation is thus presented including these terms
in a very straightforward manner. For a very stiff bearing, the additional terms
vanish implying both perturbation methods to predict OSIs in agreement with
a fully coupled model. When increasing the compliance level, and hence the
significance of the pressure–compliance interaction terms, the OSIs predicted from
the classical method diverge while those from the extended perturbation remain
in perfect agreement with the coupled model. The study is based on a benchmark
single-pad first generation GFB with the foil stiffness varied from 1/10 to 1000
times the nominal value from Walowit and Anno [176]. For the softest case, the
OSI discrepancy reaches 13 % which may not be alarming, but no attempt has
been made to quantify the sensitivity of the deviation to other bearing design
parameters.
Even though the extended perturbation method is shown to work, it is way
too computationally expensive for practical application. This could possibly be
2.1 Conclusions 27
improved upon, but the essential conclusion and recommendation is to base GFB
analysis on coupled models rather than linear bearing coefficients.
2.1.4 Multi-Domain Stability and Modal Analysis

Based on the presented coupled modelling framework, it has been demonstrated
how the static equilibrium of the rotor–bearing system can be obtained and the
Jacobian matrices evaluated at this point. Based on these Jacobians, the stability
limits and multi-domain modal results have been calculated. The latter is a novel
contribution, as it extends to all three domains, thus allowing the coupled nature
of the equations to be examined in a very tangible way.
Evaluation of stability based on Jacobian matrix eigenvalues is more mathem-
atically rigorous and less sensitive to numerical uncertainties than the exponential
growth/decay resulting from time-integrations. However, it has been shown that a
rather fine discretization of the fluid film is needed for convergence of the stability
limit (especially for non-uniform meshes) leading to large Jacobian matrices. In
terms of stability alone, the structure of the leading mode is indifferent, and it is
hence sufficient to locate the eigenvalue with the largest real part. In order to filter
the eigenvalues to extract a Campbell Diagram, it is, however, necessary to solve
for the full set of eigenvalues and vectors. At the time of writing, the eigenvalue
problems thus take several times longer than solving for the static equilibria and
generating the Jacobians.
2.1.5 Hybrid Gas Foil Bearings

In the existing literature on modelling of gas bearing injection, analytical ex-
pressions are used for the injected mass flow as a function of nozzle geometry
and upstream/downstream pressures. As these expressions are not sufficiently
accurate, constant or varying correction factors are typically added based on CFD
or experimental results. In the present work, the identification of such correction
factors is bypassed by interpolating the injection flow directly from a pre-calculated
five-dimensional map of CFD results. In terms of computational effort, the map is
relatively expensive to generate, but this only has to be done once for each injector
geometry. The subsequent use in GFB simulations is very efficient.
Apart from the injection mass flow itself, the applied piecewise multi-linear in-
terpolation scheme is capable of supplying the gradients with respect to the mapped
parameters. The Jacobian contributions to the rotor–bearing equations can thus
be calculated allowing eigenvalue analysis and ensuring the added computational
expense for HGFB systems to remain limited.
Using the parameter-map approach to injection modelling, a yet-to-be-built
hybrid configuration based on the aforementioned GFB test rig and an existing
28 2 Conclusive Remarks & Future Challenges
injection system currently mounted on a rigid gas bearing has been simulated. It is
predicted that the 21 kg rotor can be levitated aerostatically, i.e. at 0 RPM, which
could allow the bearing design to support an even heavier rotor assembly without
suffering from devastating wear during start–stop cycles. For a hybrid operational
mode without feedback control, a significant OSI improvement of the order of
20 % as well as a reduction in sub-synchronous vibration amplitudes are predicted.
The potential of injection to manipulate GFB dynamics is thus indicated to be
substantial.
2.2 Future Work

Frictional model The structural foil model with friction from publications P1
and P2 should be included into the general GFB simulation tool presented
in publication P4 and used in publications P5 and P6. To remedy the
limitations of the purely dynamic friction model, it should furthermore be
enhanced to encompass sticking and ideally foil separation. This could
possibly be achieved by adapting the frictional model proposed by Arghir
and Benchekroun [4, 5] to fit into the simultaneous solution framework.
Shell-based top foil model A top foil model based on three-dimensional shell
elements as proposed by [123] has already been implemented, but currently
leads to highly ill-conditioned Jacobian matrices. This jeopardizes the
accuracy of time integrations and prohibits stability analysis. Using an
alternative non-dimensionalization or other form of regularization, it should
be possible to correct this.
Top foil curvature The shell-based top foil is currently implemented both in
an unfolded (essentially two-dimensional) version and a more realistic three-
dimensional curved configuration. This allows the influence of curvature
(and hence membrane stresses) on the top foil stiffness to be studied and
preliminary calculations have revealed a significant difference between the
foil deformation patterns in these two configurations. Further studies could
contribute new insights to the influence of the top foil thickness and the
relative merits of different top foil modelling techniques.
Hybrid operation The results shown in publication P6 indicate the significant

potential of a fixed hybrid configuration. An adjustable/regulated injection
system is simpler and could be realised at a lower cost than a truly controllable
one. It would hence be interesting to investigate the improvements in terms
of static load-carrying capacity or stability attainable with a fixed or slowly
2.2 Future Work 29
varying injection configuration. Furthermore, experimental validation of the

predicted injection flow would be of great value.
Feedback The presented model is capable of generating the state and input
Jacobian matrices. A linear time-invariant representation of the system
around its static equilibrium is hence readily available, allowing for standard
controller synthesis methods to be applied. To allow for a practical imple-
mentation, it would most likely be necessary to add the piezo valve dynamics
(two equations per injector) as described by Morosi and Santos [117], but
this would represent a minor addition. Notice that some issues related to
hysteresis in the piezo valves have been observed which can be remedied by
applying closed–loop control [174]. The potential of an actively controlled
HGFB to support higher loads or run at higher speeds could be significant.
For a rigid gas bearing supporting a flexible rotor, Theisen et al. [174] thus
achieved a ninefold damping improvement by the addition of feedback control
and Pierart and Santos [143] demonstrated how this allowed the safe passing
of the first two critical speeds.
Publications
PUBLICATION P1
A Fully Coupled Air Foil
Bearing Model Considering
Friction – Theory & Experiment
This chapter is a postprint of the identically titled article [125] published in the
Journal of Sound and Vibration.
Sebastian von Osmanski, Jon S. Larsen, Ilmar F. Santos*
Department of Mechanical Engineering, Technical University of Denmark, 2800
Kgs. Lyngby, Denmark
* Corresponding author, ifs@mek.dtu.dk
Abstract
The dynamics of air foil bearings (AFBs) are not yet fully captured by any model.
The recent years have, however, seen promising results from nonlinear time domain
models, and simultaneously coupled formulations are now available, avoiding the
previous requirements for undesirably small time steps and temporal convergence
studies.
In the present work, an alternative foil structure model is substituted for the
simple elastic foundation model to avoid its inherent limitations. The new foil
model is based on a truss representation from the literature, but incorporates
the foil mass and a dynamic friction model. As a consequence of the friction
model’s velocity dependency, the foil mass is included to obtain a set of differential
equations that can be coupled to the rotor and fluid domains while allowing a
simultaneous solution.
Considerations leading to a practically applicable implementation are discussed
and numerical results are compared with experimental data. The model predicts
natural frequencies and mode shapes well, but it is not yet capturing the unbalance
response when friction is considered. Possible causes for this discrepancy are
discussed and it is suggested that sticking is a more prevalent state than previously
assumed.
34 P1 A Fully Coupled Air Foil Bearing Model Considering Friction...
Nomenclature for publication P1
AFB Air Foil Bearing hc, h̃c Film height (compliant),

CG Center of Gravity h̃c = hc /C
DAE Differential/Algebraic hr , h̃r Film height (rigid), h̃r = hr /C
Equation hs, h̃s Slope height, h̃s = hs /C
EOM Equation of Motion I Mass moment of inertia
FD Finite Difference It Top foil area moment of inertia
FE Finite Element kb Structural bump stiffness per
FRF Frequency Response Function unit area
LOM Light Optical Microscopy k j, dj Truss stiffness and damping,
ODE Ordinary Differential Equation j ∈ {1, 1b, 2, 3, 3b, 4}
PDE Partial Differential Equation kt Asperity stiffness
RE Reynolds Equation L, L̃ Bearing length, L̃ = L/R
SEFM Simple Elastic Foundation l0 Bump half length
Model l1, l2 Distance from CG to bearings
DOF Degree of Freedom l3, l4 Distance from CG to discs
(Û) Time derivative, d/dτ lt Top foil length
(Ü) Time derivative, d 2 /dτ 2 m Mass
∇ Gradient, ∇ = {∂/∂θ, ∂/∂z̃} mb Bump foil mass per bump
∇· Divergence Nb Number of bumps
( )e Element specific Np Number of bearing pads
(˜) Nondimensional quantity p, p̃ Film pressure, p̃ = p/pa
p1e, p2e Pressure values at nodes
A, B Bearings
pa Ambient pressure
a Scalar field quantity
p̃m Nondimensional mean axial
C Radial clearance
pressure
Eb Young’s modulus of bump foil
R Journal radius
material
Rb Bump radius of curvature
Et Young’s modulus of top foil S Compressibility number,
material
S = 6µω/pa (R/C)2
e, ε Journal eccentricity
Sb Bump foil pitch
components, ε = e/C
t Physical time
F, F̃ Bearing force components, tb Thickness of bump foil
F̃ = 1/(pa R2 )F tt Thickness of top foil
fµ Friction force U Unbalance kg·m
fN Normal force vr Relative sliding velocity
h, h̃ Film height, h̃ = h/C W, W̃ Static load components,
hb Bump foil height W̃ = 1/(pa R2 )W
Nomenclature for publication P1 35
x, y, z, z̃ Cartesian coordinates, z̃ = z/R f, f̃ Bearing force vector,

xr Relative displacement f = {f AT , fBT }T , f̃ = 1/(pa R2 )f
α Bearing position fµ Vector of friction forces
γ Friction function smoothing fp Vector of pressure forces
parameter fub, f̃ub Unbalance force,
µ Dynamic viscosity f̃ub = 1/(pa R2 )fub
µf Coefficient of friction g() Nonlinear vector function
νb Poisson’s ratio of bump foil h̃c Foil deformation vector
material p̃ Pressure vector
νt Poisson’s ratio of top foil ψ Film state vector
material r Residual vector
ω Angular speed of journal s Advection vector, s = {S, 0}T
Φ Fluid domain ũ Foil structure state space
ψ Film state variable vector
(nondimensional), ψ = p̃ h̃ w, w̃ Load vector, w̃ = 1/(pa R2 )w
x Foil displacement vector
ρt Density of top foil material
y Global state vector
ρb Density of bump foil material
z1 Rotor displacement vector,
τ Dimensionless time, τ = ωt z1 = ε
θ Circumferential angle z2 Rotor velocity vector, z2 = εÛ
θ0 Curvelinear coordinate, 0 Zero matrix
θ 0 = θR
A f , Ã f Foil structure system matrix
θ̃ Dimensionless circumferential B Shape function derivatives
coordinate, θ̃ = θ 0/R = θ matrix
θ0 Bump half angle D f , D̃ f Foil structure damping matrix
θi Inlet slope end angle for i-th Γ Fluidity matrix
pad Gr , G̃r Rotor gyroscopic matrix,
θj Truss transmission angle, θ d or G̃r = ω2C/(pa R2 )Gr
θ db I Identity matrix
θl First pad leading edge angle Kb Bump foil structure stiffness
θ li Leading edge angle for i-th pad matrix
θs Inlet slope extend K f , K̃ f Foil structure stiffness matrix
θt First pad trailing edge angle Kt Top foil structure stiffness
θ ti Trailing edge angle for i-th pad matrix
ξi, η j Gauss points M f , M̃ f Foil structure mass matrix
ζ Damping ratio Mr , M̃r Rotor mass matrix,
ae Vector of field quantities at M̃r = ω2C/(pa R2 )Mr
element nodes Mt Top foil mass matrix
ε Eccentricity vector N Shape function matrix
P1.1 Introduction
Vibrations of rotor–bearing systems have been subject to investigation at least
since Rankine [148] published his findings on lateral vibrations in 1869. Later,
the mathematical description of fluid film bearings was enabled by the findings of
Reynolds, [150] and the course of modern rotordynamics was charted by Jeffcott
[72] in 1919.
While the Reynolds equation (RE) was already extended to encompass com-
pressible lubricants by Harrison [58] in 1913, the practical interest in gas lubrication
can be traced back to the mid-1950s. At this point, especially two applications
appeared to which gas lubrication became attractive. One was inertial navigation
systems to be used during prolonged submerged operation by nuclear submarines.
The second was gas pumps located inside the radioactive gas circuits of nuclear
power reactors [145]. Along with advances within experimental equipment and the
emergence of computerised numerical solutions, these new applications strongly
facilitated the development of practical gas lubrication. During the 1960s, the first
gas bearings with compliant surfaces appeared and the air foil bearing (AFB) type
was first put into production by Garrret AiResearch in the late 1960s [1]. Since
then, AFB technology has evolved significantly, and more recent efforts even strive
towards creating a completely oil-free turbine engine using AFBs [119].
The compliant inner surface of AFBs counteracts some of the inherent issues of
rigid gas bearings, such as the small tolerance to shaft growth and misalignment.
Unfortunately, AFBs can also introduce undesirable nonlinear phenomena into the
dynamics of a rotor–bearing system, and lateral vibrations stemming from such
nonlinearities are often the limiting factor of an AFB design. Reliable means for
predicting the nonlinear unbalance response are hence necessary for further spread
of the technology. f Although nonlinear transient analysis of oil lubricated bearings
was performed as early as 1967 [27], the majority of the gas bearing literature
concerns perturbation solution methods. These apply to the frequency-domain
and rely on linearisation of the reaction forces around the static equilibrium to
effectively replace the bearing and fluid film with a spring–dashpot system. Some
of the first to apply such methods to compliant type gas bearings were Heshmat
et al. [62], who included the compliance by introducing a linear elastic function of
the fluid film pressure directly into the steady compressible RE. Even though the
top foil was disregarded entirely and any interaction between neighbouring points
in the foil structure was neglected, it allowed the equilibrium state to be obtained
while taking into account the dynamic interaction between foil compliance and
fluid film pressure. This approach to introduce compliance has later been referred
to as the simple elastic foundation model (SEFM), and in 1993, it was linked by
Peng and Carpino [132] to the perturbation method given by Lund [110, 111].
This combination has subsequently been applied directly and along with various
P1.1 Introduction 37
SEFM extensions by many authors, e.g. [68, 71, 74, 79–81, 93, 101, 175].
The perturbation methods are, however, inherently restricted to an assumed
small-amplitude periodic motion [17], and recent work [88] additionally suggests an
inadequacy in the usually applied Taylor series expansion of the pressure field. One
possible way of overcoming these limitations is nonlinear time-domain integration,
which is becoming increasingly practicable due to the growing computational
resources available. Some of the first to present transient simulations of compliant
gas bearings were Grau et al. [50], and more advanced foil structure models have
been used for time-simulations by Lee et al. [99] and Le Lez et al. [96]. A primary
challenge for these models is the pressure and film thickness temporal derivatives,
which are approximated using finite differences (FDs). This inevitably introduces a
time-lack between the variables, rendering the models non-simultaneous as stated
in [18, 19]. The time-lack issue can, to some extent, be made up for by using
very small time steps, but the solution will remain dependent on the step size and
convergence studies are, strictly speaking, necessary [90].
Bonello and Pham [18, 19] solved this issue by substituting an alternative
fluid state variable for the film height–pressure product. The substitution allows
the unsteady compressible RE to be treated as a system of ordinary differential
equations (ODEs) and hence to retain the true simultaneously coupled nature of the
state variables. This approach has later been used to obtain the unbalance response
of a rigid rotor supported by industrial three-pad AFBs with good agreement to
experimental data [90, 91].
While FD and finite element (FE) spatial discretisations are used in [18, 19]
and [90, 91], respectively, both rely on the SEFM incorporating a loss factor. This
implies that: (a) the foil structure’s energy dissipation is modelled as being viscous;
(b) the stiffness is linear and independent of both deformation and frequency; and
(c) neighbouring points in the foil are assumed to deform independently. These
are rather crude approximations which calls for the adoption of more realistic foil
structure models.
A possible bump foil model for this purpose, which is still sufficiently simple to
be suitable for time-domain simulation, is suggested by Le Lez et al. [94] based on a
truss with member stiffness coefficients derived from the foil geometry. In contrast
to the radially acting spring–dashpot pairs of the SEFM and its relatives, the truss
model takes into account bump–bump coupling and it includes circumferential
displacements facilitating the inclusion of friction.
In the present paper, the truss based bump foil model from [94] is coupled
to a simple one-dimensional top foil model similar to that used in [155], the FE
discretised unsteady compressible RE previously used by the authors [88, 90, 91]
and a smoothed dynamic friction model as discussed in e.g. [113, 118, 137]. Even
though the foil dynamics are widely accepted to be negligible as argued in e.g.
[95], the foil mass is furthermore included in order to retain the simultaneity of
the equation system. This is important as it distinguishes the presented model

from existing coupled models with friction in the literature.
The purpose of the proposed model extensions is twofold: Firstly, to increase
the generality of the AFB model, and secondly, to mimic the physical mechanism
for energy dissipation more closely. While the agreement to experimental data
presented in [90] was already notable, this was achieved through a careful choice
of foil stiffness and viscous dissipation coefficient based on a separate FE bump
foil model and a number of ”engineering assumptions”. The truss based foil model,
on the contrary, is based on directly measurable geometrical quantities and relies
on much fewer assumptions.
It should be noted that while the parameters and results in the following
are reported using SI-units, any consistent system could have been applied as
the mathematical model and its numerical implementation rely on dimensionless
quantities, marked with ”∼”, only.
P1.2 Modelling of the Rigid Rotor

The modelled test rig has previously been presented by the authors [87] and
comprises a hollow shaft supported by two identical three-pad AFBs as illustrated
in fig. P1.1. The shaft is mounted with plane discs at its extremities to which
unbalance mass can be added. The operational range of the rig is 15 to 30 kRPM
and the lowest free-free natural frequency of the assembled shaft is found to
be approximately 1050 Hz, hence a rigid shaft model is deemed adequate. The
resulting four degrees of freedom (DOFs) are expressed by the nodal position
vector holding the instantaneous position of the shaft at the bearing locations A, B
as
ε = {ε Ax, ε Ay, εBx, εBy }T . (P1.1)
Using eq. (P1.1) and under the assumption of small amplitude vibrations in the x-
and y-directions, the equation of motion (EOM) can be written in dimensionless
form as the system of second order ODEs
M̃r εÜ − G̃r εÛ = w̃ − f̃ − f̃ub, (P1.2)
where M̃r and G̃r denote the dimensionless mass and gyroscopic matrices of the
rotor, respectively, and f̃ub is the vector of unbalance forces given in appendix P1.A.
The static load contribution is contained in w̃, and f̃ = {f̃ AT , f̃BT }T is the reaction
force vector from the bearings. The latter is given by integration of the fluid film
pressure p̃ in each bearing α = A, B as
( ) ∫ ∫ ( )
F̃x L̃ 2π cos θ
f̃α = = ( p̃ − 1) d θ̃dz̃, (P1.3)
F̃y 0 0 sin θ
P1.2 Modelling of the Rigid Rotor 39
Unbalance B Unbalance A
Permanent
Shaft magnets
Bearing B Bearing A
CG z
l2 l1
l4 l3
x
(a)
θ
θl
θs
hs
Wx
h
Wy
ey y
ex
ω
θt
(b)
Figure P1.1: Schematics and nomenclature of a rigid rotor supported by foil

journal bearings: (a) Shaft, bearings and rotor disks for unbalance masses; and
(b) detailed view of bearing geometry.
where L̃ = L/R is the dimensionless bearing length and θ̃ = θ 0/R = θ, z̃ = z/R are
the dimensionless circumferential and axial coordinates, respectively.
Furthermore, introducing the rotor state space vectors z1 = ε and z2 = εÛ ,
eq. (P1.3) can be recast into first order form as

zÛ 1 0 I z1 0
= + . (P1.4)
zÛ 2 0 M̃r−1 G̃r z2 M̃r−1 (w̃ − f̃ + f̃ub )
P1.3 Modelling of the Fluid Film

The pressure in each bearing can be obtained by solving the unsteady isothermal
Reynolds equation for compressible fluids. This is a partial differential equation
(PDE) nonlinear in the pressure which can be written in dimensionless vector form
[55] as
d
∇ · p̃ h̃3 ∇ p̃ = ∇ · p̃ h̃ s + 2S (P1.5)

p̃ h̃
dτ
where S = 6µω/pa (R/C)2 is the compressibility number, s = {S, 0}T is the
advection vector and the film height h̃ is divided into a rigid and a compliant
contribution as first suggested by Heshmat et al. [62]:
h̃ = h̃r (ε x, ε y, θ̃) + h̃c . (P1.6)
The film height in the undeformed bearing h̃r depends on the rotor eccentricity
components ε x, ε y and the circumferential coordinate θ̃. For a segmented bearing
with inlet slopes and the nomenclature as illustrated in fig. P1.1(b), the film height
contribution h̃r can be written as
3
 1 + ε x cos(θ) + ε y sin(θ) − h̃s θ−θi , θ li ≤ θ ≤ θi


h̃r (ε x, ε y, θ̃) = θs , (P1.7)

 1 + ε x cos(θ) + ε y sin(θ),
 θi < θ ≤ θ ti

where the leading, inlet region and trailing edge angles of the i-th pad in a bearing
with Np pads are
2π 2π 2π
θ li = θ l + (i − 1), θi = θ s + θ l + (i − 1) and θ ti = θ t + (i − 1). (P1.8)
Np Np Np
Notice from eq. (P1.7) that a cubic inlet slope function is used in order to mimic
beam bending and hence to make ∂ 2 h̃r /∂ θ̃ 2 continuous at θ = θi . This property is
desirable as eq. (P1.5) contains second order spatial derivatives of h̃.
The compliant height term of eq. (P1.6) could be supplied by a variety of foil
structure models with varying dependencies. As it will be shown in section P1.4,
P1.3 Modelling of the Fluid Film 41
the foil model introduced in the present work results in a h̃c function depending
on the applied pressures and the state variables from the entire pad. This is a
fundamental difference from the strictly pointwise dependency assumed in the
SEFM and aims to reflect the continuous nature of the physical foil structure.
P1.3.1 Discretisation of Reynolds Equation

The spatial discretisation of the fluid film PDE eq. (P1.5), and hence the conversion
into a system of ODEs, is achieved using FE, but FD or finite volume methods
could likewise have been applied. Furthermore, the alternative state variable
ψ = p̃ h̃ is introduced to maintain the simultaneity of the equation system as it
was introduced in [18, 19] and subsequently used by the authors in [88, 90, 91].
Performing the partial substitution with ψ in eq. (P1.5), the fluid film PDE
can be expressed as
∇ · p̃ h̃3 ∇ p̃ − ∇ · p̃ h̃ s − 2S ψÛ = 0 (P1.9)

which is then spatially discretised following a standard Bubnov-Galerkin FE

procedure with implementation of an isoparametric element formulation [32].
Upon discretisation [91], this yields a system of nonlinear equations for the film
state derivative vector ψ
Û on the element level
Û e = re
Γe ψ (P1.10)
with the fluidity matrix Γ and residual vector r given on the element level as
∫ ∫ ∫
Γ = 2S
e
N N dΦ , r = −
T e
B p̃ h̃ B dΦ · p̃ +
T 3 e
BT s h̃N dΦ · p̃e (P1.11)
Φe Φe Φe
and where N and B denote the element shape function matrix and its derivatives.
The element vectors and matrices are expanded to structure size by the usual
element mapping symbolised by summations [32]:
(P1.12)
Õ Õ Õ Õ
r= re ; p̃ = Û =
p̃e ; ψ Û e; Γ =
ψ Γe
e e e e
where the integrals are numerically integrated using a quadrature rule [32]. The
scalar field quantities p̃, h̃ are calculated in the respective Gauss points (ξi, η j )
using the interpolation functions as:
a(ξi, η j ) = N(ξi, η j )a e (P1.13)
where a and a e are the scalar field quantities and nodal vectors, respectively. Note
that the right hand side of eq. (P1.10) is denoted re , which is in fact the residual
that needs to be minimised in order to find the static equilibrium of the journal.
An efficient method for this minimisation is given in [93].
P1.3.2 Fluid Film Boundary Conditions

Ambient pressure is enforced at the outer edges of the bearing and at the leading
and trailing pad edges. This can be written in dimensionless form as
p̃(θ̃, L̃/2) = p̃(θ̃, − L̃/2) = 1 (P1.14)

p̃(θ̃ li, z̃) = p̃(θ̃ ti, z̃) = 1 for i = 1, 2, . . . , Np . (P1.15)
The pressure p̃ is, however, not included explicitly in the fluid film ODEs,
meaning that eqs. (P1.14) and (P1.15) cannot be enforced directly. Differentiating
Û̃ from which it can be seen
ψ = p̃ h̃ with respect to time, one obtains ψÛ = pÛ̃ h̃ + p̃ h,
Û̃ which is continuously enforced in the simulation.
that p̃ = 1 corresponds to ψÛ = h,
As the two AFBs of the rotor–bearing test rig are assumed perfectly aligned
and as the model is limited to small rotations of the shaft, the pressure profile
is consequently symmetric about the bearing midplane. This is exploited to
reduce the computational burden by including only one half of each bearing in
the simulation. In practice, this implies that one of the edges in eq. (P1.14)
is dropped to effectively enforce ∂ p̃/∂z̃ = 0 [91] and that the integrated fluid
film reaction forces should be multiplied by two. In the process of integrating
the fluid film pressures across the mesh to obtain f̃α , the Gümbel condition is
furthermore enforced. This means that sub-ambient pressures, i.e p̃ < 1, are
discarded effectively rendering these regions inactive.
P1.4 Modelling of the Foil Structure

The presented foil structure model is based on the fundamental assumption that
the film height can be treated as being constant in the axial direction. This is in
line with experimental results [154] and allows the foil structure to be modelled in
two dimensions affected by the mean axial pressure over the bearing length L [93,
135, 156]. The three key-components of the model, namely friction, bump foil and
top foil, are covered in the following sections.
P1.4.1 Friction
Energy dissipation due to sliding friction in the foil structure is widely assumed to
be of major importance to the properties of AFBs, e.g. [1, 70, 96, 155], and the
effective radial bump foil stiffness is strongly dependent on the stick/slip conditions
at the housing contacts. Therefore, it is obvious to seek a model relying on friction
to supply stiffness and damping, instead of the widely applied constant stiffness
coefficients and viscous dissipation approximations.
P1.4 Modelling of the Foil Structure 43
Numerical simulation of friction is difficult due to the nonlinear behaviour of

the friction force near zero velocity, and various approaches to this have been
suggested. The friction models found in the literature can be roughly divided
into three categories: stick-slip bookkeeping with alternating boundary conditions
[94, 99, 173]; nonlinear springs with moving reference points [89]; and continuous
dynamic friction force approximations [95, 113, 124, 137].
As in Coulomb’s law of friction, the stick-slip bookkeeping models differentiate
between static and dynamic friction regimes to apply either boundary conditions
or dynamic friction forces. One such model is suggested for AFB simulation by
Lee et al. [99] and relies on an algorithm for continuously evaluating the stick/slip
state at each friction point. This concept is a challenge to numerical stability
as non-smooth, or even discontinuous, (in time) reaction and friction forces are
hardly avoidable. Also, the determination of the state transition times becomes a
difficult challenge for a system with many friction interfaces, as a change of state
for one friction point influences the remaining ones.
To cope with the classical issue of determining the friction force at zero velocity,
Larsen et al. [89] presented an alternative model based on nonlinear springs with
continuously updated reference points. The model was employed in a quasi-static
framework where the times of slide direction changes were known a priori. However,
it has proven difficult to apply in a dynamic framework due to the requirement
of instantaneous detection of direction shifts and discontinuities in the resulting
friction forces.
The third category of friction models are characterised by approximating the
sign function in the Coulomb friction law using a smooth function of the relative
sliding velocity dxr /dt = vr . Several approximating functions for this purpose
are found in the literature, including the hyperbolic tangent, the inverse tangent
and fractions similar to vr /|vr |, but with some smoothing parameter added to the
denominator.
Petrov and Ewins [137] present a very interesting model based on an inverse
tangent approximation taking into account also the asperity stiffness kt . This is
given as the ODE with time t as the independent variable
π fµ

dfµ 1
= kt vr − tan , (P1.16)
dt γ 2µ f fN
where fµ is the friction force, fN is the normal force, γ is a smoothing parameter

to the sign approximation and µ f is the coefficient of friction. The treatment
of friction forces as state variables is well-suited for the present purpose, but
numerical difficulties have been encountered as the normal force is often close
to zero in certain regions of the bearings. These issues could possibly be solved
by introducing a preload as suggested in [137], but for the present, a simpler
approach is taken where the friction force is given explicitly as the hyperbolic
tangent approximation with the smoothing parameter γ
fµ = fN µ f tanh (γvr ) . (P1.17)
This gives a friction force with no distinction between dynamic and static regimes,
but it could be extended using a more sophisticated expression for mimicking other
friction phenomena as suggested by Makkar et al. [113]. The approximation of
eq. (P1.17) to the sign function for increasing values of γ is illustrated in fig. P1.2.
A common characteristic shared by all of the assessed friction models is velocity
dependency. The stick-slip bookkeeping models require either a direct evaluation
of the sliding velocity or an indirect one through comparison of positions between
time steps. The mentioned nonlinear spring based model requires the velocity
to be monitored in order to correctly move a reference position, and the velocity
dependency of the continuous dynamic models is self-evident. To the best know-
ledge of the authors, no friction model exists that does not share this dependency,
meaning that a simultaneously formulated AFB model including friction is not
realisable without a dynamic foil structure model.
P1.4.2 Bump Foil Structural Model

The bump foil is modelled as a warren-like truss with member properties derived
from the foil geometry sketched in fig. P1.3(a). The model is illustrated for an
Nb = 3 bump foil strip in fig. P1.4, and requires 4(Nb + 1) DOFs, of which Nb + 3 are
µf
fµ /fN [-]
γ = 102 γ = 105
−µf
γ = 104 sign
−2000 −1500 −1000 −500 0 500 1000 1500 2000

−1
Sliding velocity vr [µm s ]
Figure P1.2: The fµ / fN ratios resulting from eq. (P1.17) using different values
of the smoothing parameter γ compared with the sign function. For reference,
±2000 µm s−1 corresponds to the maximum velocity reached by a 0.5X frequency
sliding motion with an amplitude of 2 µm at 20 kRPM.
used for reaction force calculations only. The model uses six stiffness coefficients
k 1 , k 2 , k 3 , k 4 , k 1b , k 3b and two force transmission angles θ d , θ db , where the ”b”
postfix coefficients are particular to the last bump, i.e. that at the trailing edge.
These eight coefficients are calculated using 33 analytical expressions resulting
from a tedious derivation based on Castigliano’s second theorem as presented by
Le Lez et al. [94].
The test rig bump foil strips are unchanged from the previous work by the
authors, but the foil geometry has been further investigated to minimise uncer-
tainties. The dimensions included in the commonly used schematics in fig. P1.3(a)
are updated using a series of light optical microscopy (LOM) photos of the actual
bump foil as the one shown in fig. P1.3(b). Knowing the LOM photo resolution
from calibrations, the bump foil height hb , the bump foil pitch Sb and the foil
thickness tb are found to be 7.0 mm, 1.15 mm and 0.13 mm, respectively. The
superimposed dashed arc results from a circle fit and indicates a radius of curvature
Rb of 5.7 mm. No well-defined flat sections are present meaning that the half bump
length l0 given in the schematics is not directly measurable. This ambiguity is
interesting considering e.g. the analytical bump foil stiffness by Walowit and Anno
[176]
3
Eb t b
(P1.18)
−1
kb = 1 − νb2 ,
2Sb l0
which indicates an inverse proportionality to l0 cubed. Furthermore, the lack of flat
intermediate sections reduces the foil–housing contact regions to lines, contrary to
areas or line-pairs often assumed in foil friction analyses [83, 89]. For consistency,
l0 is therefore calculated as the half length of a chord with sagitta hb in a circle
with radius Rb resulting in 3.43 mm and implying a bump half angle θ 0 of 37◦ .
Using these updated dimensions, the stiffness coefficients and transmission angles
for representing the present bump foil geometry are given in table P1.1.
A thorough validation of the truss model was presented in [94], but to substan-
tiate its applicability to the present bump foil geometry and to validate its current
implementation, a comparison is made to results from a plane FE model (see
appendix P1.C). The FE model geometry is extracted directly from fig. P1.3(b)
and two load cases are included in which the bump foil strip is uniformly loaded in
Table P1.1: Stiffness coefficients and transmission angles of the truss structure
obtained from the analytical expressions in [94] and foil geometry in table P1.3.
Stiffness coefficients MN m−1 Angles [ ◦ ]

k1 k2 k3 k4 k 1b k 3b θd θ db
31.55 0.4285 2.289 4.992 12.53 1.755 19.97 23.77
tt
tb hb
Rb
θ0 l0
Sb
(a)
hb = 1.15 mm
tb = 0.13 mm
Rb = 5.7 mm
Sb = 7.0 mm
(b)
Figure P1.3: The applied bump foil schematics compared with a LOM photo
of the actually used bump foil: (a) Schematics of the foil structure; and (b)
light optical microscopy (LOM) photo of one bump using 20 times magnification
superimposed with measurements.
its initial straight and fitted curved state, respectively. This results in the normal,
i.e. radial, bump stiffness listed in table P1.2. The effect of fitting the foil strip
into the bearing housing is found negligible (less than 2.5 %), and the agreement
to the truss model is generally good. The observed discrepancy (42 %) for the
first bump is ascribed to a 4 mm flat section of the foil strip at the leading edge,
which is not taken into account by the truss coefficients. To capture this, a set of
alternative coefficients similar to those for the last bump could be developed for
the first one, but this has not been attempted. For comparison, the widely used
analytical expression eq. (P1.18) predicts a uniform stiffness of 0.88 GN m−3 .
P1.4.2.1 Inclusion of Bump Foil Mass

Inclusion of the foil mass is generally avoided for at least two good reasons: Firstly,
the foil dynamics are insignificant to the desired rotordynamic response, and
secondly, the low foil mass results in very high natural frequencies, posing a
challenge to the numerical integration. Neglecting the foil mass, a set of purely
algebraic equations is provided by the presented truss model, meaning that it
Table P1.2: Comparison of effective stiffness resulting from the truss model and
from the plane FE model presented appendix P1.C.
Effective normal stiffness k b for each bump GN m−3

Model 1 2 3 4 5 6 7 8 9
Truss, coefficients 3.4 3.2 3.3 3.2 3.2 3.3 3.1 3.7 1.7
from table P1.1
Plane FE model, 2.4 3.6 3.2 3.3 3.3 3.3 3.1 4.5 1.6
straight state
Plane FE model, 2.4 3.7 3.2 3.3 3.3 3.4 3.1 4.4 1.6
fitted state
has no inherent notion of time and hence that only (quasi-)static results can be
obtained from its solution. If this foil model was coupled directly to the shaft and
fluid film ODEs, the coupled equation system would obtain a singular Jacobian
matrix, meaning that instead of being a system of ODEs, it would belong to the
wider class of differential/algebraic equation (DAE) systems [7]. Even though
the properties of DAEs are not as well understood as those of ODEs and no
general guarantees regarding solution existence and uniqueness can be given [144],
numerical solvers do exist for initial value problems for DAE systems [54, 67, 165,
166].
For the present purpose, however, such a DAE formulation would lack the
sliding velocities necessary for the coupling to a friction model. Theses could be
reconstructed from the displacements through a number of previous time steps
using FD, but this would break the simultaneity of the solution and reintroduce
the need for temporal convergence studies. Instead, the foil mass is included to
transform the foil structure algebraic equations into ODEs providing the velocities
directly. Considering the schematics in fig. P1.3(a), this is achieved by estimating
the foil mass per bump as the arc and flat section lengths multiplied by the foil’s
axial extension, thickness and density as
mb = (2θ 0 Rb + Sb − 2l0 ) Ltb ρb, (P1.19)
and lumping 12 mb to each DOF of the truss. This is a rough approximation

compared with the distribution of mass in the physical bump foil, but since an
accurate prediction of the bump foil dynamics is not of primary interest, this is
deemed adequate. For comparison, a nine-bump truss model with lumped masses
and the present bump foil dimensions, as given in table P1.3, has its first three
natural frequencies at 2.12, 4.97 and 5.52 kHz, while the aforementioned plane FE
model equipped with consistent mass matrices gives 2.09, 5.62 and 6.16 kHz.
It should be noted that despite the present foil being relatively soft compared
with other foil geometries usually found in the literature, its natural frequencies
are still well separated from the frequency band of interest, in this case ranging
up to approximately 500 Hz. To alleviate the numerical burden of resolving the
high-frequency modes of the bump foil, a stiffness proportional viscous damping is
therefore added to the structure. As a main objective has been to discard viscous
damping in favour of frictional dissipation, the added structural damping should be
limited not to become significant within the frequency range of interest, but even a
slight damping has proven very beneficial to the integration. At present, a stiffness
proportional damping is added to provide a damping ratio of ζ = 0.001 at 500 Hz,
resulting in modal damping ratios for the first three natural frequencies of 0.004,
0.010 and 0.011. Simulations have been performed using both ζ(500 Hz) = 0.0001
and ζ(500 Hz) = 0.005 without notable changes in the results.
P1.4.3 Top Foil Structural Model

A top foil model is included to provide the distribution of fluid film pressure
onto the underlying bump foil. The top foil is modelled using plane two-node
Euler–Bernoulli beam elements without longitudinal stiffness, very similar to the
one-dimensional model given in [155]. The stiffness matrix of such an element is
given by eq. (P1.D1).
As the radial top foil deflection corresponds to the compliant height hc in the
fluid film equations, the top foil velocities are necessary for enforcing boundary
conditions as discussed in section P1.3.2. For this reason, the top foil mass is
also included. This could be achieved using a consistent mass matrix, but since
the number of elements in this matrix is decisive for the number of arithmetic
operations required in each time step, it has proven advantageous to use a diagonal
mass matrix instead. An approximate lumped matrix is given in eq. (P1.D1).
Furthermore, a stiffness proportional damping is added to the top foil as described
for the bump foil structure in the previous section.
The beam representation of the top foil implies that curvature effects and
membrane forces are neglected and that the bump foil–top foil contacts reduce
to singular points. This implies an overestimation of the sagging effect and,
according to San Andrés and Kim [155], the top foil Young’s modulus should
be multiplied by a factor of four to obtain a level of sagging comparable to the
experimental results from [154]. Sagging has been shown to neither influence
the linear stiffness coefficients [155] nor the obtained steady state equilibrium
position [93] significantly, but it implies very large pressure gradients challenging
the numerical integration. In the present work, artificial stiffening of the top foil,
i.e. the mentioned multiplication by four, is avoided as far as possible, but for
the cases to be presented with pinned bump foil, very high levels of sagging have
prohibited numerical convergence without it.
P1.4.4 Assembled Foil Structural Model

The bump and top foil elements are assembled into a single system representing
the foil structure of an entire bearing as illustrated for a single pad with three
bumps in fig. P1.4. The bump and top foils are rigidly connected, meaning that
foil separation is not permitted, and the top foil element length is chosen to match
that of the fluid film. The upper leading edge node of the foil, which is pinned in
the original model by Le Lez et al. [94], is flexibly supported by the linear and
torsional springs of stiffness k 5 and k6 . Currently, the stiffness of these supports
are chosen equivalent to those of a top foil beam element, i.e. k5 = 12Et It /lte 3 and
k 6 = 4Et It /lte .
A proper nondimensionalisation of the foil structure has proven critical to the
time integration, especially with regard to the chosen length scale. It would have
been convenient to use the same length scale as for the fluid film quantities, i.e. the
clearance C, but this has proven unsuitable. Instead, the top foil element length lte ,
which is in the order of 18 to 12 of the bump pitch, is used while the time and force
scales are shared with the fluid film. Using these scales, the nondimensional foil
structure displacements are collected in the vector x̃α for each bearing α = A, B
and the foil structure is represented by the nondimensional mass, damping and
stiffness matrices M̃ f , D̃ f and K̃ f . The vector of compliant heights h̃c needed in
the fluid model, and the main purpose of the foil model, is hence a subset of the
foil structure displacement vector, h̃c ⊂ x̃. The piecewise linear fluid film pressure
is acting on the top foil beam elements through the vector of nondimensional work
equivalent nodal loads f̃ p (p̃α ), calculated using eq. (P1.D2).
To include the frictional forces, a vectorised version of the scalar friction force
function eq. (P1.17) is obtained following four steps: (a) calculating the relevant
normal forces from the bump foil’s dynamic equilibrium, taking into account the
reversed direction of the normal forces at the top foil interfaces; (b) equating
negative normal forces to zero; (c) associating the obtained normal forces to
their relevant sliding DOFs; and (d) calculating the friction force for each sliding
DOF using eq. (P1.17). This procedure is represented by the friction force vector
f̃ µ x̃α, xα, xα, p̃α .
Û̃ Ü̃

With the external forces represented by f̃ p and f̃ µ , the foil structure’s EOMs
can be written as a system of nonlinear ODEs as
M̃ f xÜ̃ α + D̃ f xÛ̃ α + K̃ f x̃α = f̃ µ x̃α, xÛ̃ α, xÜ̃ α, p̃α + f̃ p (p̃α ) . (P1.20)

P1 A Fully Coupled Air Foil Bearing Model Considering Friction...
Piecewise linear
pressure profile
k6,d6 Consistent
nodal loads
k4,d4 fµ k3,d3 fµ k3b,d3b fµ
k5,d5 k1,d1 k1,d1 k1b,d1b
k1,d1 k1,d1 k1b,d1b
θd θd θdb
k2,d2 fµ k2,d2 fµ k2,d2 fµ
θ'
Beam element Generalised mass Bar element Rigid link
Figure P1.4: Illustration of foil structure model for three bumps
50
P1.5 Coupled System of ODEs 51
By introducing the dimensionless foil state space vector
(P1.21)
T
ũα = x̃Tα xÛ̃ Tα ,

the EOM eq. (P1.20) can be reformulated into a system of first order ODEs as
Ã f
z }| { ( )

0 I
0
uÛ̃ α = ũα + Û̃ α, p̃α + f̃ p (p̃α ) , (P1.22)

−M̃−1 −M̃−1 M̃−1 ,

f K̃ f f D̃ f f f µ ũ α u
which constitutes the governing equation of the foil structure including friction. The
linear contribution is represented by the nondimensional and constant foil structure
matrix Ã f while the nonlinear contribution is given by the vector functions f̃ p and
f µ.
P1.5 Coupled System of ODEs

Three sets of first order ODEs representing the rotor, fluid and foil structure
domains have been established as given by eqs. (P1.4), (P1.10) and (P1.22).
Introducing the global state vector
(P1.23)
T
y = ψTA ψTB ũTA ũTB zT1 zT2 ,

these domains can be coupled to obtain a single system of nonlinear first order
ODEs as
0 ··· ··· ··· ··· 0
 ψÛ A
  ..

..   ψA  gψÛ A(ψ A, z1, z2, ũ A) 
. . 
    


ψÛ
 ··· ··· ··· ··· 
ψ

 

g (ψ , z , z , ũ )


ψB
  ... .. ...  

 B

   
 B

 
 Û B 1 2 B


. Ã f 0
    
 uÛ̃ A 
 
···  ũ A 
  g Û̃ (ψ , z1, ũ A, uÛ̃ A) 

= . + Û̃ B )  . (P1.24)
    

uA A
uÛ̃ B   .. .. .. 
ũB  Û̃ (ψ B, z1, ũ B, u
guB

  . 0 Ã f . 0    
 
zÛ 1  . .. z 0
      
  ..
1
     
. ··· ···
    

 zÛ 2 
  0 I     
 z2   M̃ (w̃ − f̃ + f̃ub ) 
  −1


−1 r
0 · · · · · · 0 M̃r G̃r 0
      
The nonlinear functions on the right hand side are defined from eq. (P1.10) as
Û (ψ α, z1, z2, ũα ) = Γ α rα
gψα −1
(P1.25)
and from eq. (P1.22) as
( )
0
Û̃ (ψ α, z1, ũα, u
guα Û̃ α ) = , (P1.26)

M̃−1 , Û̃ α, ψα, z1 + f̃ p ψα, ũα, z1

f f µ ũ α u
respectively.
P1.5.1 Numerical Implementation – an Overview
The numerical integration of eq. (P1.24) is more time-consuming than for its
SEFM based counterpart previously presented by the authors in [88, 90, 91],
and an efficient implementation is necessary to obtain reasonable solution times.
The implementation used for the present work is written mainly in C and the
system of ODEs are solved using the ”CVODE” linear multistep solver from
the Sundials package [67]. A substantial reduction of integration time has been
obtained by applying an iterative solution scheme to solve the linear system of
equations required for each Newton–Raphson iteration within each time step. The
Jacobian matrix is approximated numerically, but it is expected that an analytical
expression for this, as used by Bonello and Pham [19], would provide a further
significant speed-up.
Each evaluation of eq. (P1.25) requires the solution of a linear system of

equations. This is achieved using LAPACK’s band form routines to initially
factorise the fluidity matrix Γα and subsequently to obtain the solution from back-
substitution. For the assembly of rα , it has proven beneficial to use a regular mesh
as this allows the constant parts of the matrix–vector products from eq. (P1.11)
to be shared between elements.
For evaluating eq. (P1.26), the reaction forces need to be obtained from the
bump foil structure’s dynamic equilibrium, requiring three matrix–vector products,
and the inverse mass matrix should furthermore be multiplied by the sum of friction
and pressure forces. To efficiently obtain these products, a compressed-sparse-row
matrix format is used.
A limitation in the current implementation is the transfer of forces between

the fluid film and top foil elements, as this requires the nodes of the two models
to line up circumferentially. This means that separate convergence studies cannot
be performed and hence that one of the domains must be meshed denser than
necessary.
The numerical results presented in the following are calculated on a workstation

equipped with a 3.5 GHz Intel® XeonTM processor. For each bearing, the fluid film
has been discretised using 819 nodes, while 576 degrees of freedom have been used
for the foil structure. Including the rigid shaft, this gives 2798 equations. The
time required to solve this system is strongly dependent on the model parameters,
especially those related to damping and friction, but one second of simulation
starting from a rotor-drop currently requires a few hours of computation.
P1.6 Results & Discussion 53
P1.6 Results & Discussion

The following three sections present a number of results regarding the friction
model, the frequency response of the rotor–bearing system and the unbalance
response. All of these are based on the test rig data as given in table P1.3.
P1.6.1 The Foil Model

The first results treat the friction function smoothing parameter γ. Too low
a value effectively renders the function linear and provides viscous dissipation,
while too high a value results in a very stiff problem difficult to solve. A series
of rotor drops from the bearing centres is performed for varying values of γ at
20 kRPM, resulting in a maximum foil deformation at the first drop as depicted
in fig. P1.5. The simulations are continued for 0.5 s while the transient part of
the rotor trajectory decays and a steady state orbit forced by the unbalance is
formed. In fig. P1.6, the mean axial pressure p̃m is plotted as a function of the foil
deformation hc at θ = 180◦ for three different values of γ. This point corresponds
to the summit of bump three in the second pad, which is in the heaviest loaded
region as seen in fig. P1.5. Setting γ to zero, effectively deactivating the friction
model, the foil behaves linearly and no friction-induced hysteresis is present as
shown in fig. P1.6(a). Fitting a line to the last 0.125 s reveals a local stiffness of
3.2 GN m−3 , which is very close to the statically obtained values from table P1.2.
Setting γ = 102 , the friction model is activated and a hysteresis loop opens up,
although with rather smooth corners. A fit to the last 0.125 s gives a line that
passes diagonally through the hysteresis loop indicating an increase in effective
stiffness to 4.9 GN m−3 as illustrated in fig. P1.6(b). Further increasing γ, the
corners of the hysteresis loop sharpen and the effective diagonal stiffness increases
to 6.8 GN m−3 , as shown for γ = 105 in fig. P1.6(c). The same stiffness value
and loop shape is obtained using γ = 104 and γ = 106 , hence γ = 104 is deemed
adequate. Notice, however, that this value should be scaled consistently to the
nondimensionalisation.
The right and left edges of the hysteresis loop are often associated to sticking,
but it should be emphasised that the current model is not capable of producing a
true stick-phase as a friction force is only present for non-zero sliding velocities.
The stiffness experienced on the right edge of the shown hysteresis loops is around
10 GN m−3 , which is dependent on the coefficient of friction µ f , but true sticking
would provide an even higher stiffness. If the foil–housing contact nodes are pinned,
a situation corresponding to permanent sticking is created providing no frictional
dissipation at these nodes. Repeating the same rotor drop simulation with this
configuration, the deformation–pressure plot shown in fig. P1.6(d) is obtained. As
the bump foil structure is stiffened significantly, the tendency towards sagging is
Table P1.3: Geometry, material properties and operating conditions of the

Siemens foil bearing test rig.
Shaft assembly
Bearing A to CG, l1 201.1 mm Mass, m = mx = my 21.1166 kg
Bearing B to CG, l2 197.9 mm Polar moment of 30.079 × 10−3 kg m2
inertia, Izz
Unbalance A to CG, l3 287.2 mm Transverse moment 525.166 × 10−3 kg m2
Unbalance B to CG, l4 304.0 mm of inertia, Ixx = Iyy
Bearing configuration
Bearing radius, R 33.50 mm First pad leading edge, θ l 30◦
Bearing length, L 53.00 mm First pad trailing edge, θ t 145◦
Radial clearance, C 40 µm Slope extend, θ s 30◦
Number of pads, Np 3 Slope height, hs 50 µm
Fluid properties
Viscosity, µ 1.95 × 10−5 Pa s Ambient pressure, pa 1 × 105 Pa
Bump foil properties
Bump foil thickness, tb 0.13 mm Bump foil pitch, Sb 7.00 mm
Bump foil half length, l0 3.43 mm Bump foil height, hb 1.15 mm
Young’s modulus Eb 207 GPa Poisson’s ratio, νb 0.3
Radius of curvature, Rb 5.7 mm Coefficient of friction, µ f 0.05
Density, ρb 8280 kg m−3 Bump half angle, θ 0 37◦
Top foil properties
Top foil thickness, tt 0.254 mm Poisson’s ratio, νt 0.3
Young’s modulus Et 2.07 × 1011 Pa Density, ρt 8280 kg m−3
increased. The very high pressures reported in fig. P1.6(d) are hence caused by a
local pressure peak near the bump summit following a sagging-induced converging
zone. The fitted line indicates an effective stiffness in the order of 26 GN m−3 .
P1.6.2 Natural Frequencies

The frequency response of the physical test rig has been thoroughly investigated.
This provides a reference for validation of the numerical model. The experimental
frequency response functions (FRFs) have been obtained using shaker excitation at
the shaft extremities while maintaining a constant angular velocity. The excitation
force amplitude was chosen to give a maximum displacement amplitude of 8 µm in
both the vertical and horizontal directions, as this provided an operable signal-
to-noise ratio, while the behaviour remained reasonably linear. The experimental
procedure is further described in [87], while the current results are given from
1200
1000
800
600
400
200
Foil height [µm]

0
0 50 100 150 200 250 300 350 400
θ [◦ ]
(a)
1170
1160
1150
1140
1130
1120
1110
Foil height [µm]

1100
160 180 200 220 240 260
θ [◦ ]
(b)
Figure P1.5: Excessive foil structure deformation at ε x, ε y = 1.98, 0.121 occurring 5.6 ms after the rotor is dropped
from the centre: (a) The full foil structure; and (b) zoom to the midmost pad.
55
(b)
(a)
2.5 3.2 GN m−3 2.5 4.9 GN m−3
Mean pressure p̃m [-]

2.0 2.0
1.5 1.5
1.0 1.0
0 10 20 30 0 10 20 30
Deformation hc [µm] Deformation hc [µm]
(c) (d) 10
2.5 6.8 GN m−3 26 GN m−3

8
2.0 6
1.5 4
2
1.0
0 10 20 30 0 10 20
Deformation hc [µm] Deformation hc [µm]
Figure P1.6: Hysteresis curves at the summit of bump three in the second pad segment (θ = 180◦ ) for 0.5 s of
simulation from a rotor drop from the centre using different smoothing parameter values: (a) γ = 0; (b) γ = 102 ; (c)
γ = 105 ; and (d) pinned foil–housing contact nodes. The dashed lines are fits to the last 0.125 s and indicate the local
effective foil stiffness. Notice the different vertical scale in (d). This is made with pinned foil–housing contact nodes
56 to simulate sticking which implies high sagging-induced pressure peaks.
[161].
The numerical receptance FRF matrix is obtained by applying 0–300 Hz linear
chirps at the extremities with force amplitudes chosen to provide maximum
displacement amplitudes of 8 µm. The magnitude and phase of the diagonal
elements of this FRF matrix are shown in fig. P1.7. The first two identifiable modes
are located at ≈ 79 Hz and ≈ 105 Hz, respectively, with the first being a cylindrical
and the second a conical mode. The third mode at ≈ 146 Hz is much more heavily
damped and completely absent when exciting in the horizontal y-direction. It
has not been possible to identify the expected fourth mode theoretically, but this
has not been observed experimentally either. A comparison to the experimentally
observed natural frequencies for the three identifiable modes is given in table P1.5.
P1.6.3 Unbalance Response

The unbalance response has been obtained experimentally by attaching various
unbalances to the A-disc, accelerating the rotor to 27 kRPM, disabling the electrical
motor and measuring the response during the following coast down. The coast
down approach is chosen to eliminate any contribution from the motor. The
waterfall diagram obtained using the maximum applied unbalance of 45.5 g mm is
shown in fig. P1.8 for the vertical vibration in bearing A. The synchronous and
a number of super-synchronous components are present at all velocities, while
various sub-synchronous components, mostly arranged symmetrically about one
half the synchronous speed (0.5X), appears and disappears sporadically. A single
0.5X vibration emerges at 9.5 kRPM before it briefly disappears and bifurcates into
two branches around 11.5 kRPM. At 14.6 kRPM, the two branches reunite only
to bifurcate yet again around 16.2 kRPM and eventually reunite at 22.5 kRPM.
These rotational speeds are marked with thick black lines.
The waterfall diagram is obtained numerically by simulation of 1.3 s from
a rotor drop from the centre for every 250 RPM. Discarding the first 0.3 s of
each simulation, steady state is assumed to have been reached and FFTs are
calculated for the remaining 1 s. The residual unbalance is estimated to be less
Table P1.4: Comparison of experimentally observed and numerically predicted

natural frequencies at 20 kRPM
Mode 1 2 3
Shape Cylindrical Conical Cylindrical
Experimental observations [Hz] 64–77 104–119 153–158
Numerical [Hz] 78–80 103–108 146–147
Mid-interval deviation [%] +12 −5 −6
10−4 eAx eBx

eAy eBy
|H( f )|
10−5
0 50 100 150 200 250 300

Frequency [Hz]
(a)
0
eAx eBx
Phase [ ◦ ]
eAy eBy
−90
−180
0 50 100 150 200 250 300
Frequency [Hz]
(b)
Figure P1.7: The diagonal elements of the numerically obtained receptance FRF
matrix, i.e. the response of each DOF when it is excited directly.
than ±2.5 g mm at each end, and even if the unbalance mass is assumed to be
added in the angle giving the highest possible effective unbalance of U A = 48 g mm
and UB = −2.5 g mm, the model is unable to predict the sub-synchronous unbalance
response correctly. The predicted waterfall diagram is shown in fig. P1.9, and even
though some sub-synchronous vibration is present in the 7–10 kRPM range, this is
dissimilar to the experimentally observed pattern of bifurcations and reunifications.
Increasing the coefficient of friction µ f , it is possible to shift this RPM range with
sub-synchronous vibrations slightly upwards, but no qualitative changes to the
waterfall appearance are obtainable.
One possible explanation for the absence of sub-synchronous activity could
be an overestimation of the frictional damping. To test this, the waterfall is
reproduced with γ = 0 resulting in the waterfall diagram shown in fig. P1.10.
The damping is clearly lowered, as a 75 Hz natural frequency becomes visible
and the solver has difficulties completing the integration for velocities less than
10 kRPM. The sub-synchronous vibrations visible from 13.5 to 14.8 kRPM resemble
the experimental results reasonably well, but much is still left to be desired.
]
10 [µm
t ude
8 li
0.5X 1X 2X 6
p
4
Am
2
0
25
20
15
10
Rotational speed [kRPM]
0 100 200 300 400 500

Frequency [Hz]
Figure P1.8: Experimentally obtained waterfall diagram for the vertical vibration in bearing A showing the
unbalance response during a coast down with 45.5 g mm added to the A disc. The velocities marked with thick black
lines are 9.5, 11.5, 14.6, 16.2 and 22.4 kRPM, respectively.
59
]
µm
20.0
[
17.5
ude
15.0
0.5X 1X 2X 12.5
plit
10.0
7.5
Am
5.0
2.5
0.0
25

20
15
10
5
0 100 200 300 400 500
Frequency [Hz]
Figure P1.9: Waterfall diagram obtained numerically with properties as given in table P1.3 and unbalance masses
of U A = 48 g mm and UB = −2.5 g mm. The 0.5X vibration appears at 7.3 kRPM and bifurcates at 8.3 kRPM, as
marked with thick black lines. No sub-synchronous vibration is visible after 10 kRPM.
60
]
[µm
20
it ude
15
l
0.5X 1X 2X
p
10
Am
5
25
20
15
10
0 100 200 300 400 500

Frequency [Hz]
Figure P1.10: Waterfall diagram from the same configuration as in fig. P1.9, but friction disabled, i.e. γ = 0.
61
A second possible explanation could be related to the aforementioned lack of

true sticking. If sticking is in fact a prevalent state, the current model will both
underestimate the stiffness and overestimate the frictional dissipation. This is
tested by performing the simulation yet again with the foil–housing contact nodes
pinned, corresponding to permanent sticking. This is obviously unrealistic and it
eliminates the frictional dissipation at the foil–housing interface entirely, but it
serves to indicate what prevalent sticking would imply. The waterfall diagram from
this simulation is shown in fig. P1.11, and the similarity to the experimental results
is notable even though the solver has difficulties integrating past 22 kRPM due to
the low damping. The 0.5X vibration emerges at 9 kRPM, bifurcates at 12 kRPM,
reunites at 14.25 kRPM and re-bifurcates at 15.5 kRPM. Zoomed views on the
region with sub-synchronous vibrations is also shown in fig. P1.12. The qualitative
behaviour of the simulated results are the same as observed experimentally, and a
comparison of the angular velocities at the bifurcation events are given in table P1.4.
These angular velocities are predicted within approximately 5 %, but the final
reunion of the two branches around 22.5 kRPM is not captured and the branches
are too widely spaced. It should also be noted that the increased bump foil stiffness
implied by the pinning causes excessive sacking which produces very large pressure
peaks. Hence, the numerical solution has necessitated an artificial stiffening of the
top foil to smooth the pressure profile, but this has been found not to influence
the remaining results.
Lastly, it is worth showing the difference in obtained steady state orbits between
the three assessed cases, i.e friction, no friction and pinning. These are shown for
11 and 20 kRPM, respectively, in fig. P1.13. The two non-pinned cases give almost
identical one-periodic orbits with mean eccentricity ratios (ε x, ε y ) = (1.25, 0.11)
and (ε x, ε y ) = (1.22, 0.14) at 11 kRPM, and (ε x, ε y ) = (1.16, 0.21) and (ε x, ε y ) =
(1.13, 0.22) at 20 kRPM. The pinned case gives much larger orbits with multiple
clearly defined periods and with mean eccentricity ratios of (ε x, ε y ) = (0.60, 0.15)
Table P1.5: Comparison of the angular velocities of four clearly defined events
between the experimental data and the numerical simulation with pinned bump
foil.
Experimental Numerical (pinned)
Event
fig. P1.8 fig. P1.11
First appearance of 0.5X [kRPM] 9.5 9
First bifurcation [kRPM] 11.5 12
Reunification [kRPM] 14.6 14.25
Second bifurcation [kRPM] 16.2 15.5
]
[µm
40 lit ude
0.5X 1X 2X 30
p
20
Am
10
0
25
20
15
10
0 100 200 300 400 500

Frequency [Hz]
Figure P1.11: Waterfall diagram from the same configuration as in fig. P1.9, but with pinned foil–housing contact
nodes, as it is also the case in fig. P1.6(d). The velocities marked with thick black lines are 9, 12, 14.25 and 15.5 kRPM,
respectively.
63
[ µm]
itude
10
8
6
Ampl
0.5X 1X 4
2
0
17

16
15
14
13
12
11
10
9
60 80 100 120 140 160 180
Frequency [Hz]
(a)
[µ m]
itude
25
20
15
Ampl
0.5X 1X 10
5
0
17
16
15
14
13
12
11
10
9
60 80 100 120 140 160 180
Frequency [Hz]
(b)
Figure P1.12: Zoomed views on the region with sub-synchronous activity from
the experimental and the numerical (with pinned bump foil) waterfalls: (a) From
fig. P1.8, i.e. the experimental data; and (b) from fig. P1.11, i.e. the pinned
numerical case.
P1.7 Conclusions 65
at 11 kRPM, and (ε x, ε y ) = (0.29, 0.12) at 20 kRPM.

Moreover, for 20 kRPM, the equilibrium points reached, if removing the un-
balance, are furthermore indicated for each case. It is noticeable that the same
static equilibrium (ε x, ε y ) = (1.17, 0.22) is reached for both γ = 0 and γ = 104 ,
even though a much higher effective stiffness is present in the latter, as previously
shown in fig. P1.6. This is ascribed to the dynamic nature of the applied friction
model, which cannot provide a friction force without a non-zero sliding velocity.
P1.7 Conclusions
The paper has presented a nonlinear time domain model of a rigid shaft supported
by AFBs. The compliant foil structure is included using an extended model based
on the original idea of Le Lez et al. [94], coupled to a simple top foil beam model.
The frictional energy dissipation at the top foil–bump foil and bump-foil–housing
interfaces is included using a dynamic friction model and, in order to achieve a fully
simultaneous solution of the state variables, the foil structure mass has also been
included. The latter is argued to be a necessary addition since all friction models
known to the authors encompass velocity dependencies, and hence a timescale is
necessary. Separately, simultaneous models and models including friction already
exist in the literature, but the presented model succeeds in reconciling these
properties. The model has been demonstrated to be numerically realisable, though
dependent on an efficient implementation.
Based on directly measurable quantities and without the ”engineering assump-
tions” related to the SEFM, the dynamics of the rotor–bearing system is captured
well in terms of natural frequencies and mode shapes. The unbalance response,
however, has not yet been captured satisfactorily using the current friction model.
It is postulated that this is due to the characteristics of the applied dynamic friction
model, as it tends to provide low-lying orbits corresponding to the frictionless
situation. It is inconvenient that the absolute placement of the journal is hard to
obtain experimentally within the required accuracy, and hence, that it is difficult
to assess whether the orbits obtained with friction are realistic. If the bump
foil is pinned to the housing, effectively disabling friction at these contacts while
providing a constant stiffness similar to that at sticking, promising results has
been achieved. This is not a workable approach for general AFB simulation, but
it suggests that sticking, which cannot be captured accurately using the dynamic
friction model, is in fact a prevalent state. This is interesting as widespread
sticking would be incompatible with the generally accepted perception of frictional
dissipation as a main source of damping in AFBs.
Regarding friction, further work should be focused on incorporating a friction
model similar to e.g. eq. (P1.16), in order to further assess the sticking behaviour.
Pinned γ =0 γ = 104
0.0
0.5
εx
1.0
1.5
−1.0 −0.5 0.0 0.5 1.0 1.5

εy
(a)
Pinned γ =0 γ = 104
Pinned, Uα = 0 γ = 0, Uα = 0 γ = 104 , Uα = 0
−1.0
−0.5
0.0
εx
0.5
1.0
1.5
−1 0 1
εy
(b)
Figure P1.13: Comparison of steady state orbits with friction (fig. P1.9), without
friction (fig. P1.10), and with pinned foil–housing contact nodes (fig. P1.11): (a)
11 kRPM; and (b) 20 kRPM, where the equilibrium points reached if removing the
unbalance are furthermore shown.
P1.7 Conclusions 67
Also, experimental recordings of the bump foil sliding in operation would be of great
value. Regarding the top foil, its stiffness has been found negligible in comparison
to that of the bump foil, but numerical difficulties have been experienced due to
excessive sagging causing high local pressure gradients. The overestimated sagging
is partly caused by the omission of curvature effects and membrane forces, but
comparing the model’s pointwise top foil supports to the non-zero width contact
zones of the actual foil as sketched in fig. P1.14, it is possible that a contribution is
also stemming from an overestimation of the free span. This would be important,
as the deflection of a uniformly loaded clamped–clamped beam is proportional
to its free span to the power of four. The excessive sagging has previously been
investigated in [155], where a correction factor on the top foil’s Young’s modulus
was introduced by fitting to experimental data, but a more generic approach is
desirable. A point of attention for further work should hence be the implementation
of an improved top foil formulation, possibly based on shell elements as presented
in [121], if not only to provide numerical stability and speed up the simulation.
p
Free span
hc
Sb
Figure P1.14: Illustration of top foil sagging. The top foil–bump foil contact
zones have non-zero width, meaning that the effective free span between bump
summits becomes smaller than the bump pitch Sb . The dashed (red) triangles
illustrate the bump foil truss model which supports the top foil beam model only
at singular points spaced Sb apart.
P1.A Rotor Model Matrices

The mass and gyroscopic matrices for a rigid rotor can be written as
 l 2 m x + I yy 0 l1 l2 m x − I yy 0 
 2
l2 m y + I x x
2

1  0 0 l1 l2 m y − I x x 
Mr = 2  ,
l12 l1 l2 m x − I yy 0 l12 m x + I yy 0
2m + I

0 l1 l2 m y − I 0 l y
(P1.A1)
x x x x
 
 1 
 0 −Izz 0 I zz 


1  Izz 0 −Izz 0 
Gr = 2  ,
l12  0 Izz 0 −Izz 
−Izz 0 Izz 0 

where l12 = l1 + l2 . The nondimensional form of the mass and gyroscopic matrices
and the mass unbalance vector are:
U Aω2 cos τ 


 
ω C ω C pa R sin τ 
2 2
 
2
 
G̃r = , = , = (P1.A2)

 
G M̃ M f̃
pa R2
r r
pa R2
r ub

 UB ω cos τ 
2

 p R2 sin τ 
 
 a 
P1.B Bump Foil Matrix

The truss structure members are modelled as two-node bar elements with four
DOFs. The stiffness matrix of a bar element with axial stiffness k j (representing
k 1 , k 2 , k 3 , k 4 , k 1b or k 3b ) and angular orientation θ j (representing θ d or θ db ) is
given from [32] as
cos2 θ j cos θ j sin θ j − cos2 θ j − cos θ j sin θ j 
sin θ j − cos θ j sin θ j − sin2 θ j 
 2
Keb = k j  , (P1.B1)

cos θ j
2 cos θ j sin θ j 
sin2 θ j

 
 
where the elements below the diagonal are given from symmetry.
P1.C Validation of Bump Model

The FE model used for validation is created using a commercial software package
and utilises plane quadratic quadrilateral elements, includes large displacement
theory and assumes a state of plane strain. Its geometry is extracted directly from
the LOM photo in fig. P1.3(b), but a 4 mm flat section is added to the leading
P1.D Top Foil Matrices and Work Equivalent Load 69
end as present on the actual foil strips. Discarding friction, a nine-bump foil
strip is clamped at its leading edge and subjected to uniform normal loads at
the bump summits up to the equivalent of 200 kPa in two different cases. In the
first case, the undeformed and stress-free foil strip is supported by a straight rigid
surface as shown in fig. P1.C.15(a). In the second case, the foil strip is preloaded
to fit a curved rigid surface, representing the bearing hosing as illustrated in
fig. P1.C.15(b). For both cases, a converged mesh is used.
In the applied load range, giving normal deformations up to 125 µm, no signi-
ficant geometrical nonlinearity is observed and the stress level stays well below
the yield strength.
P1.D Top Foil Matrices and Work Equivalent Load

The top foil is modelled as two-node Euler–Bernoulli beam elements with rotational
and lateral translation DOFs only. For an element of length lte , Young’s modulus Et ,
depth L, thickness tt , density ρt and cross-sectional moment of inertia It = 12
1 3
tt L,
the stiffness and lumped mass matrices are given from [32] as
12 6lte −12 6lte  12 0 0 0 
4lte 2 −6lte 2lte 2  ρt Ltt lte e2
   
Et It  l 0 0  , (P1.D1)
Kt = e 3  , Mte =
e
 
 t
lt  12 −6lte  24 
 12 0 

 4lte 2  
 lte 2 
where the elements below the diagonals are given from symmetry.
Based on the linear element shape functions, a laterally working tapered
pressure can be shown to give the work equivalent nodal loads

 21p1e + 9p2e 

lte Llt 2p1 + 3p2 
 
 e e e 
fp = , (P1.D2)
e

 
60  21p1 + 9p2 
e e
 −l e 2pe + 3pe 

 

 t 1 2 
where p1e and p2e are the pressure values at the element nodes, lte is the beam
element length and L is the beam element width (normal to the load direction).
References
All references have been collected in the thesis bibliography on page 259.
(a)
(b)
Figure P1.C.15: The plane element FE model in its two load cases: (a) Foil
strip loaded in its undeformed, straight state; and (b) foil strip loaded after being
fitted into the bearing housing.
PUBLICATION P2
On the Incorporation of Friction
into a Simultaneously Coupled
Time Domain Model of a Rigid
Rotor Supported by Air Foil
Bearings
journal Technische Mechanik. An initial version of the manuscript was presented
at the SIRM 2017 conference on 15–17 February 2017 in Graz, Austria.

Abstract
Despite decades of research, the dynamics of air foil bearings (AFBs) are not
yet fully captured by any model, suggesting that the fundamental mechanisms
of the AFB and their relative merits are not yet fully understood. The recent
years have seen promising results from nonlinear time domain models, allowing
the dynamic pressure–compliance interaction and the unsteady terms of the
compressible Reynolds equation to be considered.
By including the simple elastic foundation model (SEFM) in a fully coupled
simultaneous time integration, the dynamics of a rotor supported by industrial
AFBs have previously been modelled by the authors, leading to good agreement
with experimental results. In this paper, the authors investigate the substitution
of the SEFM for a new foil structure model which is based on directly measurable
72 P2 On the Incorporation of Friction into a Simultaneously Coupled...
quantities and includes frictional energy dissipation in the foil structure. An

important finding is that the incorporation of a friction model into the global
model cannot be reconciled with a simultaneous time solution without the inclusion
of the foil inertia. The resulting AFB model allows the effects of friction on AFB
performance to be directly examined and leads to the questioning of friction’s role
and its significance to the operation of AFBs.
AFB Air Foil Bearing hr , h̃r Film height (rigid), h̃r = hr /C

CG Center of Gravity hs, h̃s Slope height, h̃s = hs /C
DAE Differential/Algebraic I Mass moment of inertia
Equation k Stiffness
FE Finite Element k j, dj Truss stiffness and damping,
ODE Ordinary Differential Equation j ∈ {1, 1b, 2, 3, 3b, 4}
SEFM Simple Elastic Foundation L, L̃ Bearing length, L̃ = L/R
Model l0 Bump half length
DOF Degree of Freedom l1, l2 Distance from CG to bearings
(Û) Time derivative, d/dτ l3, l4 Distance from CG to discs
(Ü) Time derivative, d 2 /dτ 2 m Mass
∇ Gradient, ∇ = {∂/∂θ, ∂/∂z̃} Np Number of bearing pads
∇· Divergence p, p̃ Film pressure, p̃ = p/pa
A, B Bearings pa Ambient pressure
C Radial clearance p̃m Nondimensional mean axial
e, ε Journal eccentricity pressure
components, ε = e/C R Journal radius
Eb Young’s modulus of bump foil Rb Bump radius of curvature
material S Compressibility number,
Et Young’s modulus of top foil S = 6µω/pa (R/C)2
material Sb Bump foil pitch
fµ Friction force t Physical time
fµ f Friction coefficient smoothing tb Thickness of bump foil
function tt Thickness of top foil
fN Normal force function vr Relative sliding velocity
h, h̃ Film height, h̃ = h/C W, W̃ Static load components,
hb Bump foil height W̃ = 1/(pa R2 )W
hc, h̃c Film height (compliant), x, y, z, z̃ Cartesian coordinates, z̃ = z/R
h̃c = hc /C xj Generalised degree of freedom
xj Relative displacement f, f̃ Bearing force vector,

α Bearing position, α = A, B f = {f AT , fBT }T , f̃ = 1/(pa R2 )f
γ Friction function smoothing fµ Vector of friction forces
parameter fp Vector of pressure forces
µ Dynamic viscosity fub, f̃ub Unbalance force,
µf Coefficient of friction f̃ub = 1/(pa R2 )fub
νb Poisson’s ratio of bump foil g() Nonlinear vector function
material p̃ Pressure vector
νt Poisson’s ratio of top foil ψ Film state vector
material r Residual vector
ω Angular speed of journal s Advection vector, s = {S, 0}T
ψ Film state variable ũ Foil structure state space
(nondimensional), ψ = p̃ h̃ vector, ũ = {x̃T xÛ̃ T }T
ρt Density of top foil material w, w̃ Load vector, w̃ = 1/(pa R2 )w
ρb Density of bump foil material x Foil displacement vector
τ Dimensionless time, τ = ωt y Global state vector
z1, z2 Rotor state vectors, z1 = ε,
θ Circumferential angle
z2 = zÛ 1
θ0 Curvelinear coordinate,
θ 0 = θR 0 Zero matrix
A f , Ã f Foil structure system matrix
θ̃ Dimensionless circumferential
coordinate, θ̃ = θ 0/R = θ D f , D̃ f Foil structure damping matrix
Γ Fluidity matrix
θ0 Bump half angle
Gr , G̃r Rotor gyroscopic matrix,
θj Truss transmission angle, θ d or
θ db G̃r = ω2C/(pa R2 )Gr
θl First pad leading edge angle I Identity matrix
θs Inlet slope extend K f , K̃ f Foil structure stiffness matrix
θt First pad trailing edge angle M f , M̃ f Foil structure mass matrix
ζ Damping ratio Mr , M̃r Rotor mass matrix,
ε Eccentricity vector, M̃r = ω2C/(pa R2 )Mr
{ε Ax, ε Ay, εBx, εBy }T (˜) Nondimensional quantity
P2.1 Introduction
Practical application of gas lubrication appeared in the mid-1950s driven by
its attractiveness to several emerging technologies and facilitated by improved
experimental equipment together with the development of computerised numerical
methods [145]. The first gas bearings with compliant inner surfaces appeared
in the mid-1960s and the air foil bearing (AFB) was introduced industrially by
Garrret AiResearch in 1969 [1]. The AFB offers several advantages compared to
conventional rigid gas bearings, and it is a key component in NASA’s efforts towards
creating a completely oil-free turbine engine [119]. NASA is interested in the
AFB technology’s weight-saving potentials in rotorcrafts and its high-temperature
capabilities, but AFBs also present an environmentally friendly alternative in
many applications of oil-lubricated high-speed rotating machinery.
The compliant nature of AFBs does, however, complicate the modelling of
its dynamic characteristics and is capable of introducing undesirable nonlinear
features. As the performance of AFB supported rotor–bearing systems is often
limited by nonlinear phenomena, such as sub-synchronous vibrations driven by
unbalance, reliable means for predicting the response are an essential prerequisite
for further spread of the technology.
The majority of the literature on AFB modelling rests on the original contri-
butions by Heshmat et al. [61, 62], who introduced the simple elastic foundation
model (SEFM). The original SEFM, as well as the refined version by Peng and
Carpino [132], was applied in a perturbation method framework introduced by
Lund [110] and hence relied on a linearisation of the reaction forces to effectively
replace the bearing and fluid film with a spring–dashpot system. Such analyses
are inherently restricted to an assumed small-amplitude periodic motion [17], and
recent work [88] additionally suggests an inadequacy in the applied Taylor series
expansion of the pressure field. Another commonality shared throughout much
of the literature is equivalent viscous models for the energy dissipation in the
compliant structure. This approximation is pivotal since sliding friction in the foil
structure is widely assumed to constitute a major source of damping [1, 70, 96,
155] and hence to be essential in the workings of the AFB.
Nonlinear time domain integration circumvents the limitations of the perturba-
tion techniques and provides a basis for the incorporation of foil structure models
without the assumption of viscous dissipation. Applying different foil structure
models, but based on a decoupling of the fluid, rotor and foil structure equations,
time domain models have been presented by for example Hoffmann et al. [69], Le
Lez et al. [96], and Lee et al. [99]. This approach introduces a demand for very
small time steps and temporal convergence studies, which has been overcome using
simultaneous formulations [18, 19, 90, 91]. Several promising results have been
presented from these simultaneous models, but they are, however, still relying on
the SEFM and hence on the assumption of viscous dissipation.
In recent work by the authors [125], a fully coupled simultaneously formulated
AFB model including friction has been presented. The current work provides
additional discussions on the necessity of foil mass inclusion and an assessment
of three effects related to friction: (a) displacement of the static equilibrium; (b)
introduction of a dynamic foil stiffness; and (c) dissipation of energy through
sliding friction.
P2.2 The Rotor–Bearing System 75
P2.2 The Rotor–Bearing System

The presented model and the derived considerations are based on a test rig pre-
viously presented and described by the authors [87, 88, 90, 91, 125]. The rig
comprises a near-symmetrical hollow rotor supported by two identical second
generation Siemens AFBs as sketched in fig. P2.1a. The illustrated permanent
magnets are part of the electrical drive capable of rotating the shaft to approxim-
ately 30 kRPM and the discs at each of the shaft’s extremities allow unbalance
mass to be added. As observed from the AFB geometry (fig. P2.1b), the foil
structure is segmented into three pads fixed to the bearing housing at their leading
edges. The dimensions and mechanical properties as used throughout this paper
are listed in table P2.1.
P2.3 Mathematical Model of the Rotor and the Fluid

Film
As the dynamics of the support structure is not considered, modelling of the
rotor–bearing system requires three domains to be assessed: the rotor, the fluid
film and the compliant structure of the AFBs. This paper concerns mainly the
latter of these, hence only a brief exposition of the applied rotor and fluid film
models will be made.
The operational range of the rig is limited by the electrical drive to 30 kRPM,
while the lowest free–free natural frequency of the assembled shaft is found to be
approximately 1050 Hz, hence a rigid shaft model is deemed adequate. This gives
a four degrees of freedom (DOFs) model which is considered a system of first order
ordinary differential equations (ODEs) as

zÛ 1 0 I z1 0
= + , (P2.1)
zÛ 2 0 M̃r−1 G̃r z2 M̃r−1 (w̃ − f̃ + f̃ub )
where the state vectors hold the rotor displacements and velocities at the bearing
positions as
z1 = ε = {ε Ax, ε Ay, εBx, εBy }T and z2 = zÛ 1 = εÛ . (P2.2)
The system matrix contains the dimensionless mass and gyroscopic matrices of
the rotor, M̃r and G̃r , while w̃, f̃ and f̃ub represent the static load, integrated fluid
film reaction forces and unbalance forces, respectively.
The fluid film formed between the shaft and compliant inner surface of the AFBs
is assumed to be governed by the isothermal, compressible, transient Reynolds
Unbalance B Unbalance A
Permanent
Shaft magnets
Bearing B Bearing A
CG z
l2 l1
l4 l3
x
(a)
θ
θl
θs
hs
Wx
h
Wy
y
ey
ex
ω
θt
x (b)
Figure P2.1: Schematics and nomenclature of a rigid rotor supported by AFBs:

(a) shaft, bearings and rotor discs for unbalance masses; and (b) detailed view of
the bearing geometry.
P2.3 Mathematical Model of the Rotor and the Fluid Film 77
Table P2.1: Geometry, material properties and operating conditions of the

Siemens AFB test rig.
Shaft assembly
Bearing A to CG, l1 201.1 mm Mass, m = mx = my 21.1166 kg
Bearing B to CG, l2 197.9 mm Polar moment of 30.079 × 10−3 kg m2
inertia, Izz
Unbalance A to CG, l3 287.2 mm Transverse moment 525.166×10−3 kg m2
Unbalance B to CG, l4 304.0 mm of inertia, Ixx = Iyy
Radial clearance, C 40 µm Slope extend, θ s 30◦
Number of pads, Np 3 Slope height, hs 50 µm
Fluid properties
Viscosity, µ 1.95×10−5 Pa s Ambient pressure, pa 105 Pa
Bump foil properties
Bump foil thickness, tb 0.13 mm Bump foil pitch, Sb 7.00 mm
Bump foil half length, l0 3.43 mm Bump foil height, hb 1.15 mm
Young’s modulus Eb 207 GPa Poisson’s ratio, νb 0.3
Radius of curvature, Rb 5.7 mm Coefficient of friction, µ f 0.05
Density, ρb 8280 kg/m3 Bump half angle, θ 0 37◦
Top foil properties
Top foil thickness, tt 0.254 mm Poisson’s ratio, νt 0.3
Young’s modulus Et 2.07 × 1011 Pa Density, ρt 8280 kg/m3
equation:
d
∇ · p̃ h̃3 ∇ p̃ = ∇ · p̃ h̃ s + 2S p̃ h̃ , (P2.3)

dτ
where S = 6µω/pa (R/C)2 is the compressibility number, s = {S, 0}T is the
advection vector and the film height h̃ is divided into a rigid and a compliant
contribution as first suggested by Heshmat et al. [62]:
h̃ = h̃r (ε x, ε y, θ̃) + h̃c . (P2.4)
The rigid contribution h̃r depends on the initial undeformed bearing geometry
as illustrated in fig. P2.1b and is given by e.g. Osmanski et al. [125], while the
compliant contribution h̃c is treated in the following sections.
Following a partial substitution of ψ for p̃ h̃ as introduced in Bonello and Pham
[18, 19], the fluid film partial differential equation eq. (P2.3) is spatially discretised
using a finite element (FE) scheme. This gives a system of nonlinear ODEs in the
film state variable time derivative vector ψÛ α for each bearing α = A, B
Û α = rα ψÛ α, ψα , (P2.5)

Γα ψ
where the fluidity matrix Γα is constant for a given angular velocity, while the
residual vector rα depends on both the pressures, the film heights and the film
heights’ temporal derivatives.
P2.4 Modelling of the Foil Structure

In the first time domain model presented by Larsen and Santos [90] and Larsen
et al. [91] as well as in the presented models by Bonello and Pham [18, 19], the
compliant height contribution h̃c in eq. (P2.4) is supplied using the SEFM. This is
numerically efficient, but implies that (a) the foil structure’s energy dissipation is
modelled as being viscous using an equivalent rotor-speed based loss factor; (b)
the stiffness is linear and independent of both deformation and frequency; and
(c) neighbouring points in the foil are assumed to deform independently. Other
authors, such as Hoffmann et al. [69] and San Andrés and Kim [156], have applied
the SEFM merely to the underlying bump foil structure while incorporating more
comprehensive models for the top foil. In this case, the top foil’s bending stiffness
couples the deflections of neighbouring points, but any bump–bump interaction
effects are still neglected. An objective of the present work has hence been to
discard the SEFM entirely in favour of a more general foil structure model. This
model should facilitate the inclusion of friction, allow for a simultaneous solution
of the equation system and be sufficiently efficient to permit time integration.
P2.4.1 Friction Models

Sliding friction in the foil structure is widely assumed to be an important mechanism
in AFBs [1, 70, 96, 155]; consequently, a friction model is included in the present
work. Time integration of friction phenomena is difficult due to the nonlinear
behaviour of the friction force near zero velocity and/or zero normal force. The
potentially applicable friction models available in the literature can be roughly
divided into three categories: stick-slip bookkeeping with alternating boundary
conditions [94, 99, 173], nonlinear springs with moving reference points [89] and
continuous dynamic friction approximations [95, 113, 124, 137].
The stick-slip bookkeeping models introduce a differentiation between static
and dynamic friction regimes in which either a boundary condition or a dynamic
friction force is applied. These models hence rely on a continuous evaluation of the
stick/slip states, and the abrupt changes inevitably caused by the change of state
pose a challenge to numerical stability due to non-smooth, or even discontinuous,

reaction and friction forces.
A friction model relying on nonlinear springs with moving reference points
has also been suggested [89] and shown to perform well in a quasi-static setting.
The model handles the classical issue of determining the friction force at zero
velocity, but has proven difficult to apply in a time domain framework due to the
requirement of instantaneous detection of direction shifts.
The friction model applied in the current work belongs to the group of con-
tinuous dynamic friction approximations. These are based on expressions for the
friction force fµ of the form
fµ = fN fµ f (vr ) , (P2.6)
where fN is the normal force function and fµ f (vr ) is a smooth function of the
sliding velocity vr approximating the sign function. In the literature, various
different functions can be found serving as sign approximations, including the
inverse tangent, fractions similar to vr /(γ+|vr |) and the hyperbolic tangent. The
latter is used in the present as
fµ f (vr ) = µ f tanh (γvr ) , (P2.7)
where µ f is a dynamic coefficient of friction and γ is a smoothing parameter

controlling the slope near vr = 0 and hence the level of approximation. As it can
be seen, eq. (P2.7) provides no distinction between static and dynamic friction,
but this could be achieved using the extended version given by Makkar et al. [113].
Note that while the particular choice of friction model and smoothing function
is debatable, an important point is that all of the assessed approaches share the
common characteristic of velocity dependency. This is, to the best knowledge of
the authors, the case for all existing and suitable friction models.
P2.4.2 Structural Models

The compliant structure of the Siemens AFB consists of a bump foil and a top foil.
For the present purpose, a simple one dimensional Bernoulli–Euler beam model is
utilised for the top foil, as the main point of attention is the supporting bump foil
structure. This approach leaves out any axial film height variations, but this has
been shown to be a reasonable assumption [156].
For modelling of the bump foil, a straightforward plane FE approach requires
several thousand DOFs per bump [89], and is hence precluded from time integration
purposes. A model reduction technique could possibly be applied, but here an
efficient equivalent model by Le Lez et al. [94] is used instead. In this model, the
bump foil is represented using bar elements forming a simple truss with member
stiffness coefficients k 1 , k 2 , k 3 , k 4 , k 1b , k3b and force transmission angles θ d , θ db .
These coefficients are calculated from 33 analytical expressions given by Le Lez

et al. [94] based on Castigliano’s second theorem, and their values for the present
geometry are given by Osmanski et al. [125].
In table P2.2, the effective radial stiffness of the truss model for a nine-bump
foil strip with dimensions from table P2.1 is compared to results from a plane
FE model based on a very accurate replication of the actual foil geometry. To
emphasise the significance of the boundary condition at the foil–housing contacts,
results are included for both rolling and pinned supports at this interface. The
truss coefficients from Le Lez et al. [94] are based on the case of rolling supports,
meaning that the bumps are allowed to slide circumferentially with no frictional
resistance. Considering that the widely used expression by Walowit and Anno
[176] predicts a uniform stiffness of 0.88 GN/m3 , the observed agreement between
the truss and plane FE models is very good. If the foil–housing contacts are
pinned, i.e. restrained from circumferential sliding, the effective truss stiffness is
increased more than fivefold, while the plane FE model stiffness increases at least
tenfold. A similar stiffening was observed by Feng and Kaneko [45] and should be
kept in mind as the two cases correspond to the extreme cases of zero friction and
permanent sticking, respectively.
The truss model is formulated as a static model, meaning that it is governed
by a system of algebraic equations. Coupling any (quasi-) static model directly
to the differential equations governing the fluid film and the velocity dependent
friction model gives rise to certain issues. These become evident if considering the
simple mechanical system sketched in fig. P2.2. It comprises a single point mass
supported by four massless springs affected by a friction force in a configuration
similar to the bump foil truss model. Writing out the governing equations in first
Table P2.2: Effective radial stiffness (for a uniform pressure) resulting from the
truss model compared to a plane FE model for the two cases of (frictionless) rolling
and pinned housing contact nodes.
Effective normal stiffness for each bump [GN/m3 ]

Condition Model 1 2 3 4 5 6 7 8 9
Truss model 3.4 3.2 3.3 3.2 3.2 3.3 3.1 3.7 1.7
Sliding
Plane FE 2.4 3.6 3.2 3.3 3.3 3.3 3.1 4.5 1.6
model
Truss model 19.8 19.8 19.8 19.8 19.8 19.8 19.8 19.8 11.0
Pinned
Plane FE 42.5 43.6 43.7 43.8 43.8 43.8 43.7 43.3 40.5
model
order form, the following system can be obtained:
0   2kc2 0 −kcθ2 0 0   x1
θ

   

   2
 
0 0 k 3 − 2c kc s −k 0 x2
   
θ θ θ

 
   
 

=  −kcθ kcθ + k

  
 2 2
 
 

0 kcθ sθ 0 0  x3
xÛ4  0 0 0 0 1   x4

    


 
  
 

 xÜ4 
   0 k/m 0 −k/m 0

 xÛ4


(P2.8)

    


 0 


0
 
cθ = cos θ d,

 

− fN (x1, x2, x3 ) fµ f ( xÛ3 )

 

where
sθ = sin θ d .
0

 


 

0

 

 
If friction is discarded, the upper three rows of eq. (P2.8) governing the massless
truss are purely algebraic and hence provide neither velocities nor accelerations.
Together with the two differential equations governing the point mass, this con-
stitutes a system of semi-explicit differential/algebraic equations (DAEs). The
DAEs represent a superset of the ODEs and are, in general, more troublesome
since no guarantees on solution uniqueness or existence can be given as is the case
for ODEs (see Petzold [138] and Poulsen et al. [144]). Without friction, eq. (P2.8)
is nevertheless very easy to solve. It could be condensed and solved as two first
order ODEs, or it could be solved as is using a DAE solver.
Introducing friction, the solution of eq. (P2.8) becomes considerably more
troublesome since the velocity required to determine the orientation and size of
the friction force is not available. Obviously, this could be reconstructed using
information from previous time steps using finite difference, but this would violate
the requirement for a simultaneous formulation and reintroduce the demand for
temporal convergence studies. In the case of a strictly positive normal force, the
system could be considered as an implicit ODE (or DAE), but for any reasonable
approximation to the sign function, this system is too stiff for practical purposes.
In the actual AFB model, the case of zero normal force, implying zero friction
force for any sliding velocity, would furthermore have to be spanned leaving the
very structure of the equation system state dependent.
From a physical point of view, the fundamental issue is the lack of inertia to
smooth out the displacements caused by the rapidly changing friction forces in the
vicinity of zero sliding velocity. As a remedy, it is therefore natural to introduce
the foil mass, even though this is per se insignificant to the overall rotordynamic
response. Lumping the bump foil mass onto the truss structure (giving a diagonal
mass matrix) the equations are remoulded from algebraic to differential with sliding
velocities directly available. Coupling the obtained bump foil differential equations
to the (also dynamic) Bernoulli–Euler beam top foil model and the friction model,
the overall foil structure is assembled as visualised in fig. P2.3. Notice that viscous
dampers have also been introduced in the truss. These are principally undesired
as a main objective is to model frictional instead of viscous dissipation, but a
slight structural damping has proven numerically necessary due to the very high
natural frequencies in the foil ranging up to around 500 kHz. The frequency range
of interest for the rotordynamic response goes to around 500 Hz, and is hence well
separated from the first natural frequency of the foil structure at around 2 kHz.
This allows a proportional damping, providing a damping ratio of ζ = 0.001 at
500 Hz, to be introduced to effectively dampen out the foil structure dynamics
while leaving the dynamics of interest virtually unaffected.
Collecting the DOFs of the foil structure for each bearing α = A, B into the foil
state vector ũα , the system of first order nonlinear ODEs governing the structure
in fig. P2.3 can be written as
Ã f
z }| {

0 I 0
uÛ̃ α = ũα + , (P2.9)
−M̃−1
f K̃ f −M̃−1
f D̃ f M̃−1
f fµ ũα, uÛ̃ α, p̃α + f p (p̃α )
where M̃ f , D̃ f and K̃ f are the mass, (proportional-) damping and stiffness matrices
of the foil structure, respectively. The vector function f µ represents the friction
forces at the contact nodes given from eq. (P2.6) and the vector function f p
represents the work equivalent nodal loads on the top foil stemming from the fluid
film pressure p̃α .
P2.5 Structure of the Assembled Equation System

The three sets of first order ODEs representing the rotor, fluid and foil structure
domains given from eq. (P2.1), eq. (P2.5) and eq. (P2.9) are now coupled. For
this purpose, the global state vector
(P2.10)
T
y = ψTA ψTB ũTA ũTB zT1 zT2 ,

is introduced, using which a single system of nonlinear first order ODEs can be
written as
0 ··· ··· ··· ··· 0
 ψÛ A   ..

..  ψA  gψÛ A(ψ A, z1, z2, ũ A) 
. . 
   


ψÛ

 ··· ··· ··· ··· 
ψ

 

g (ψ , z , z , ũ )


ψB
  ... .. ..  

 B

   
 B

 
 Û B 1 2 B 

. Ã f 0 .  ũ A
    
Û̃ (ψ A, z1, ũ A, u A )
 uÛ̃ A 
 
···
 
guA Û̃ 
= . + Û̃ B )  . (P2.11)
 
 
 
 

uÛ̃ B   .. .. .. 
ũB  Û̃ (ψ B, z1, ũ B, u
guB

  . 0 Ã f . 0    
 
Û
z . .. z1  0
      
 zÛ 2   ..
1
    
. ··· ···
     

  0 I   z   M̃−1 (w̃ − f̃ + f̃ ) 
 
 
 
 
2 ub 

r
0 · · · · · · 0 M̃r−1 G̃r 0
    
x4
m
k
x2
x1
fN (x1, x2, x3 ) fµ f ( xÛ3 )
k k
θd
k x3
Figure P2.2: Mechanical system illustrating the challenges of using a massless

foil model. It comprises a point mass (governed by a second order differential
equation) supported by a massless truss (governed by four algebraic equations)
subject to a friction force.
The nonlinear functions gψα Û on the right hand side of the upper equations repres-
enting the fluid film are defined from eq. (P2.5), while those in the midmost rows
representing the foil structure, guαÛ̃ , are given as the nonlinear part of eq. (P2.9).
It should be noted that the numerical time integration of the coupled equation
system given from eq. (P2.11) is a nontrivial task. To make the presented foil
model extension practically feasible, considerable prior optimisation of the SEFM
based time integration code has been necessary. Through these efforts, the SEFM
based simulation times has been reduced from days to minutes; but with the new
foil structure, especially the friction model, the relevant simulations nevertheless
take in the order of 24 hours to complete.

To provide insight into the behaviour of the foil model and the influence of friction,
a rotor drop from the bearing centres with a high level of unbalance at 20 kRPM is
simulated for 0.5 s with and without friction. In fig. P2.4, the mean axial pressure
p̃m is plotted as a function of the top foil deformation hc at θ = 180◦ . This point
is in the heaviest loaded region and coincident with the summit of bump three
in the second pad. Setting γ to zero, effectively deactivating the friction model,
the foil behaves linearly and no friction-induced hysteresis is present. Fitting a
line to the last 125 ms reveals a local stiffness of 3.2 GN/m3 , which is very close
to the statically obtained values from table P2.2. Activating the friction model
P2 On the Incorporation of Friction into a Simultaneously Coupled...
ω
Journal surface
Fluid film (grey area)
Consistent nodal loads Fluid pressure
hr (θ 0)
k 6, d6
θ0
k 4, d4 fµ k 3, d3 fµ k3b, d3b fµ
k 5, d5 k 1, d1 k 1, d1 k 1b, d1b
k1, d1 k 1, d1 k1b, d1b
θd θd θ db
k 2, d2 k2, d2 k 2, d2
fµ fµ fµ
Beam element Generalised mass Bar element Rigid link
Figure P2.3: Illustration of the foil structure model for three bumps interacting with the journal through the
generated fluid film (grey area). k j , d j denote the truss element stiffness and damping, while θ j are the transmission
angles and fµ represents the friction forces. Notice that the last bump uses different coefficients than the remaining
ones.
84
using γ = 104 , a hysteresis loop opens up and the fit now gives a line passing
diagonally through the hysteresis loop indicating an increase in effective stiffness
to 6.8 GN/m3 .
In fig. P2.5, the vertical eccentricity ratio in bearing A is plotted during the
first and last 40 ms of the rotor drop simulation both with and without friction.
In the transient part, the inclusion of friction lowers the displacement amplitudes,
but almost identical steady states are eventually reached. This indicates that the
equilibrium position is determined by the structural stiffness alone. This is in the
order of 3–4 GN/m3 and hence much lower than the equivalent SEFM stiffness of
9 GN/m3 used by Larsen and Santos [90]. The value of 9 GN/m3 was based on a
number of ”engineering assumptions” and was intended by Larsen and Santos [90]
to represent the dynamic foil stiffness, but as the applied model made no distinction
between static and dynamic behaviour, this resulted in equilibrium points lying
much higher than those obtained from the present model. If the effective static
stiffness of the foil structure is in fact closer to 9 GN/m3 , as suggested by the
agreement to experimental results, the foil structure must, at least partly, be
sticking.
fig. P2.6 shows the circumferential sliding displacement for the first four bump
foil–housing contact nodes in the second pad during the first and last 40 ms. The
presence of friction is evident in both plots from the characteristic flattened peaks
and valleys related to the sign change of the friction forces. As both the mean
displacements and the dynamic displacement amplitudes increase along the pad
(this is also true for the remaining five bumps), the frictional energy dissipation
will be largest for the bumps closest to the trailing edge.
The original motivation for introducing a friction model was twofold: (a) to
circumvent the requirement for an empirically determined and constant equivalent
stiffness; and (b) to avoid the inclusion of an empirical mechanical loss factor. It
is evident from fig. P2.4 that the friction model provides damping and that it
increases the effective dynamic stiffness. It has, however, also been shown that
the friction induced dynamic stiffness does not affect the equilibrium position.
This means that the present AFB model results in steady state rotor eccentricities
determined solely by the structural foil stiffness, while its dynamic characteristics
at this equilibrium are influenced by friction. The effect is sketched in fig. P2.7,
where the left plot shows the radial deflection of a bump subjected to the load
profile shown to the right. When loading up, point ”A” is reached tracking the
upper edge of the global hysteresis loop, where the effective stiffness is that of
the sliding structure plus a frictional contribution. Ramping down the load, a
line parallel to the right edge of the global loop is tracked to point ”B” while the
direction of sliding and hence the frictional force is reversed. Ideally, the effective
stiffness is here close to that of a pinned bump, as no sliding should occur while
the friction force changes direction. Note that the friction force crosses zero and
2.5 3.2 GN/m3

2.0
1.5
1.0
0 5 10 15 20 25 30 35
Deformation hc [µm]
2.5 6.8 GN/m3

2.0
1.5
1.0
0 5 10 15 20 25 30
Deformation hc [µm]
Figure P2.4: Hysteresis curves at the summit of bump three in the second pad
segment (θ = 180◦ ) for 0.5 s of simulation from a rotor drop from the centre with
and without friction. The dashed lines are fits to the last 0.125 s and indicate the
local effective foil stiffness.
0.0
γ =0
0.5 γ = 104
εAx [-]
1.0
1.5
2.0
0 5 10 15 20 25 30 35 40
Time [ms]
0.9
γ =0
1.0
γ = 104
1.1
εAx [-]
1.2
1.3
1.4
460 465 470 475 480 485 490 495 500
Time [ms]
Figure P2.5: Vertical eccentricity ratio ε Ax with friction (γ = 104 ) and without
(γ = 0) friction. The first 40 ms after the rotor drop are shown to the left (at
the top in the postprint), while the final 40 ms, where steady state has set in, are
shown to the right (at the bottom in the postprint).
Circumferential displacement [µm]
60
40
#4 at 198°
20 #3 at 186°
#2 at 174°
0 #1 at 162°
0 20 40
Time [ms]
Circumferential displacement [µm]
#4 at 198°
30
20 #3 at 186°
10 #2 at 174°
0 #1 at 162°
460 480 500

Time [ms]
Figure P2.6: Circumferential displacement, i.e. sliding, of the bump foil–housing
contact nodes of the four leading bumps in pad segment two. The left plot (upper
plot in the postprint) shows the first 40 ms after the rotor drop, while the right
plot (lower plot in the postprint) shows the final 40 ms of the simulation where
steady state has been reached.
P2.7 Conclusion 89
changes direction as the dashed line indicating the sliding bump stiffness is crossed.
Further decreasing the load from ”B”, the global hysteresis loop is tracked until
”C” while the contact slides. Increasing the load to ”D”, the same stiffness as
between ”A” and ”B” is experienced, but as the load is lowered from ”D” towards
”E”, a frictional force sign change is required without any sliding. This is not
possible for the dynamic friction model, meaning that the contact point will drift
towards the frictionless equilibrium for each sign change made within the global
loop. These drifts are marked in red in the plot (the short horizontal arrows) and
even though their magnitude is dependent on the smoothing parameter, point ”F”
will eventually be reached. At this point, the oscillations take place around the
same equilibrium as would have been reached without friction.
P2.7 Conclusion
The paper has presented an alternative foil structure model for AFB simulation
based on a truss representation of the bump foil originally proposed by Le Lez
et al. [94]. The top foil is added using a simple one dimensional Bernoulli–Euler
beam and a dynamic friction approximation is included to model frictional energy
dissipation at the top foil–bump foil and bump foil–bearing housing interfaces.
The usually applied foil structure models are static, i.e. represented by algebraic
equations, but it is argued that the combination of a simultaneous solution in time
and a friction model requires the inclusion of the foil inertia. This is achieved by
augmenting the bump foil truss and top foil beam elements with lumped masses
and subsequently to eliminate the entailed very high natural frequencies using
stiffness proportional damping.
The new foil structure model is coupled to a nonlinear time domain model of a
rigid shaft supported by two AFBs as a replacement for the SEFM. This allows
the SEFM’s inherent assumptions of viscous dissipation, constant stiffness and
decoupled neighbouring points to be abandoned, and the empirically determined
equivalent stiffness and loss coefficients of the SEFM can be replaced by directly
measurable quantities.
The numerical results from the coupled model demonstrates that energy is
dissipated by the friction model and that an increased effective dynamic stiffness
is introduced. It is, however, evident that the dynamic stiffness caused by friction
does not affect the obtained steady state position in the current model. This
is reasonable considering the applied dynamic friction approximation, as this is
not capable of representing true sticking and hence will allow net sliding until
the equilibrium position dictated by the structural stiffness is reached. This is
interesting, as the equivalent SEFM stiffness used by Larsen and Santos [90] to
obtain unbalance responses in very good agreement to experimental results, was
P2 On the Incorporation of Friction into a Simultaneously Coupled...
Load
Load
Structural stiffness (sticking) Structural
stiffness
(sliding) A
A F D F
D
E E
B
C B C
Displacement Time
Figure P2.7: Sketch of the frictional drift present for the dynamic friction model. The right plot depicts a
hypothetical load profile applied to a single bump and the left plot is the resulting displacement. (For interpretation
of the references to colour in this figure, the reader is referred to the web version of this paper.)
90
P2.7 Conclusion 91
based on the dynamic foil structure stiffness. In the SEFM, no distinction is made
between the static structural stiffness and the friction induced dynamic stiffness,
meaning that much higher foil compliances and rotor eccentricities are obtained
using the new model than was the case using the SEFM [90].
The agreement to experimental results achieved using the SEFM, with a
constant equivalent stiffness much higher than that of a sliding bump foil structure,
suggests that sticking is in fact a prevalent state. This would imply that the
orbits predicted using the new foil structure model are too low-lying and, most
importantly, that frictional dissipation is not as significant as generally assumed.
References
PUBLICATION P3
The Classical Linearization
Technique’s Validity for
Compliant Bearings
This chapter is a postprint of the identically titled article [128] presented at the
10th IFToMM conference on Rotor Dynamics in Rio de Janeiro, Brazil on 23–27
September 2018 and published in Springer’s Mechanisms and Machine Science vol.
60.
Sebastian von Osmanski1 , Jon S. Larsen2 , Ilmar F. Santos1,*

1 Technical University of Denmark, Kongens Lyngby, Denmark
2 GEA Process Engineering A/S, Søborg, Denmark
Abstract
The Gas Foil Bearing (GFB) is a promising and environmentally friendly technology
allowing support of high-speed rotating machinery with low power loss and without
oil or electronics. Unfortunately, GFBs provide limited damping, making an
accurate prediction of the Onset Speed of Instability (OSI) critical. This has
traditionally been assessed using linearised coefficients derived from the perturbed
Reynolds Equation with compliance included implicitly. Recent work has, however,
revealed significant discrepancies between OSIs predicted using these techniques
and those observed from nonlinear analysis. In the present work, the perturbation
method’s underlying assumption on the pressure field is investigated by including
the hitherto neglected pressure–compliance dependency directly. This leads to an
extended perturbation akin to that commonly applied to tilting pad bearings and
is shown to predict OSIs with much better agreement to time integration results.
The extended perturbation method is cumbersome, but serves to highlight the
error introduced when applying the classical perturbation method — as developed
for rigid bearings by J. W. Lund — to GFBs.
94 P3 The Classical Linearization Technique’s Validity for Compliant Bearings
P3.1 Introduction
Due to the limited damping provided by GFBs, an accurate prediction of the
Onset Speed of Instability (OSI) remains critical to their application. In order to
predict the lateral vibration response of GFB-supported rotors, and hence their
stability, it is possible to apply linear [68, 79–81, 87, 93, 101, 132, 175] as well
as nonlinear [10, 18, 19, 69, 90, 139] approaches. In the linear approach, the gas
film forces are fundamentally represented by equivalent springs and dampers with
coefficients derived from a linearisation of the Reynolds Equation (RE) around
one or several states of equilibrium. The calculation of such gas film coefficients
can be achieved by a numerical perturbation or analytically as proposed by Lund
[110]. Peng and Carpino [132] were among the first to apply Lund’s perturbation
technique to GFBs and such analyses have since been performed by numerous
authors [68, 80, 175]. Some of these have furthermore compared their results to
nonlinear analyses showing varying levels of agreement. In recent work by the
authors [88], this has been investigated by comparing the OSI of an industrial
GFB-supported rotor predicted from the classical frequency domain technique to
the stability limits observed from nonlinear time integration. Here, using equivalent
numerical implementations for the two approaches, a significant discrepancy was
demonstrated and shown to be increasing with the level of compliance.
A possible root cause of the observed disagreement can be found in a primary
assumption of the classical perturbation approach, namely that the pressure field
depends exclusively on the rotor position and velocity. Assuming subsequently
the rotor to perform small harmonic oscillations, a Taylor series expansion can be
inserted into RE to solve for the eight bearing stiffness and damping coefficients.
In the present work, the importance of the — thus far neglected — pres-
sure–compliance dependency is investigated. The pressure field is thus assumed
to depend not only on the rotor position and velocity, but also on the degrees
of freedom (DOFs) representing the foil deflection. This additionally requires
the solution of a dynamic pressure field for each of the foil DOFs and hence
provides a coefficient matrix with contributions from each of these, analogous to
the coefficient matrices often used for tilting pad journal bearings. Using a simple
and widely studied GFB configuration [154] supporting a point mass as starting
point, OSI predictions from the extended perturbation method are compared to
results from both a classical perturbation method and a simultaneous nonlinear
time integration. These two reference models are identical to those previously
presented by the authors [88, 90, 93] and have been experimentally validated. It
should be emphasized that the vibrations occurring at the investigated OSI stems
from a self-excited instability and thus are related exclusively to the homogeneous
part of the equation system. This should not be confused with the onset of forced
subsynchronous vibrations caused by the unbalance excitation in conjunction with
P3.2 The extended perturbation method 95
the nonlinear GFB characteristics. The appearance of the latter is influenced by

the level of unbalance, while the former is not.
The additional terms included in the extended perturbation multiplies the
effort needed to attain the bearing coefficients and complicates their subsequent
interpretation, but the extension is demonstrated to provide results in much better
agreement to those obtainable from nonlinear time integration. The challenges of
the extension are discussion and the significance of the foil–compliance terms is
treated to illuminate the error introduced when neglecting them. In this sense,
the novel contribution of the present work is to identify the limits of validity for
the classical perturbation method with respect to compliance.
P3.2 The extended perturbation method

In addition to the eccentricities, the pressure field p is assumed to depend on N
DOFs. These dependencies can be collected in the vector q as
(P3.1)
T
q = e x e y w1 · · · wN ∈ R2+N ,

where e x and e y are the rotor eccentricity components and w j denotes the j-th
foil compliance DOF. Using eq. (P3.1), a first order Taylor series expansion of the
pressure field around a state (q0, qÛ 0 ) can be written as
∂p ∂p
p = p (q, q)
Û ≈ p (q0, qÛ 0 ) + (q − q0 ) + (qÛ − qÛ 0 ) . (P3.2)
∂q q0,qÛ 0 ∂ qÛ q0,qÛ 0

Defining q0 as the stationary point where qÛ = 0, and assuming all DOFs to

exhibit small harmonic oscillations around this point with frequency ωs as
eγ = eγ0 + ∆eγ eiωs t, γ = x, y and w j = w j0 + ∆w j eiωs t, j = 1, . . . , N, (P3.3)
the pressure field expansion from eq. (P3.2) becomes

N
(P3.4)
Õ
p = p0 + p x Λ x + p y Λ y + pw j Λ j ,
j=1
where
p0 =p (q0 ) , (P3.5)
∂p ∂p

pγ = + iωs , γ = x, y, (P3.6)
∂eγ q0 ∂ eÛγ q0
∂p ∂p

pw j = + iωs , j = 1, . . . , N, (P3.7)
∂w j q0 ∂ wÛ j q0
Λγ =∆eγ eiωs t, γ = x, y, (P3.8)

Λ j =∆w j eiωs t
, j = 1, . . . , N, (P3.9)
so that p0 ∈ R and pγ, pw j , Λγ, Λ j ∈ C.
P3.2.1 The film height function

For the present purpose of investigating the pressure–compliance dependency, the
rotor is assumed to be perfectly aligned and the foil is assumed to deform uniformly
in the axial direction. Axial film height variations are hence neglected and the
film height function becomes one-dimensional. This has previously been shown
to be reasonable [156], but the proposed perturbation method is not restricted to
such assumptions.
Referring to fig. P3.1(b), the film height is composed of a rigid contribution,
hr = C + e x cos θ + e y sin θ, (P3.10)
and a compliant contribution hc stemming from the deformation of the foil structure.
It is hence necessary to establish a relation between the foil DOFs w j and the
deformed shape of the top foil. The foil DOFs could be representing Fourier series
amplitudes or polynomial coefficients, but for now they will represent discrete
point deformations directly such that w j = hc θ j . To simplify the numerical
implementation, the points θ j are furthermore chosen to be coincident with the
circumferential discretization of the fluid film.
In order to evaluate the deformed shape between the discrete points, a simple
linear interpolation is applied that is consistent with the linear shape functions
applied for the finite element (FE) discretization of RE. A weight function is hence
defined for each foil DOF as illustrated in fig. P3.1(a) and given as
 θθ−θ
 j−1
if θ j−1 < θ < θ j
j −θ j−1




W j (θ) = θθ j+1−θ
−θ
if θ j ≤ θ < θ j+1 (P3.11)


 j+1 j


otherwise,

0


such that the continuous compliant film height contribution can be written as
N
(P3.12)
Õ
hc = w j W j (θ) .
j=1
This finally allows, when introducing the perturbations from eq. (P3.3), the
perturbed film height to be written as
N
(P3.13)
Õ
h = hr + hc = h0 + Λ x cos θ + Λ y sin θ + Λ j W j (θ)
j=1
θ j=1
j=N
Wj ∆θj
Ω
1
h Rwj
ey
ex y
Ry
θj−1 θj θj+1 θ
Rx
wj
x
(a) (b)
Figure P3.1: (a) Weight function associated with the j-th foil compliance DOF;
(b) Schematics of the GFB and illustration of the perturbed DOFs.
where the static film height h0 is given from
N
(P3.14)
Õ
h0 = C + e x0 cos θ + e y0 sin θ + w j0W j (θ) .
j=1
P3.2.2 Perturbation of Reynolds Equation

Defining the circumferential and axial coordinates Θ = θR and z in a bearing with
radius R, the isothermal, compressible and transient RE can be written as
∂ ph3 ∂p ∂ ph3 ∂p ∂ ∂

+ = (phU) + (ph) , (P3.15)
∂Θ 12µ ∂Θ ∂z 12µ ∂z ∂Θ ∂t
where µ is the gas viscosity and U = RΩ/2 is the circumferential gas film velocity.
Substituting eqs. (P3.4) and (P3.13) into eq. (P3.15) while neglecting higher order
terms, the zeroth and first order equations can be separated. The zeroth order
equation obtains the familiar form
! !
∂ p0 h03 ∂p0 ∂ p0 h03 ∂p0 ∂
+ = (p0 h0U) , (P3.16)
∂Θ 12µ ∂Θ ∂z 12µ ∂z ∂Θ
while the two first order equations for p x and p y can be written as
! ! !
∂ p0 h03 ∂pγ ∂ p0 h03 ∂pγ ∂ pγ h03 ∂p0
+ + +
∂Θ 12µ ∂Θ ∂z 12µ ∂z ∂Θ 12µ ∂Θ
! !
∂ pγ h03 ∂p0 ∂ ∂ p0 h02 fγ ∂p0
pγ h0U − iωs pγ h0 = − (P3.17)

−
∂z 12µ ∂z ∂Θ ∂Θ 4µ ∂Θ
!
∂ p0 h02 fγ ∂p0 ∂
+ p0U fγ + iωs p0 fγ, γ = x, y,

−
∂z 4µ ∂z ∂Θ
where fx = cos θ and fy = sin θ, and, finally, the N first order equations for pw j
related to the foil DOFs become
∂ p0 h03 ∂pw j ∂ p0 h03 ∂pw j
! ! !
3
∂ pw j h0 ∂p0
+ + +
∂Θ 12µ ∂Θ ∂z 12µ ∂z ∂Θ 12µ ∂Θ
! !
3
∂ pw j h0 ∂p0 ∂ ∂ p0 h02W j ∂p0
− pw j h0U − iωs pw j h0 = − (P3.18)
∂z 12µ ∂z ∂Θ ∂Θ 4µ ∂Θ
!
∂ p0 h02W j ∂p0 ∂
+ p0UW j + iωs p0W j , j = 1, . . . , N.

−
∂z 4µ ∂z ∂Θ
P3.2.3 The coefficient matrix

Having obtained p0 , pγ and pw j by a suitable numerical solution scheme, these
can be integrated to obtain the fluid film forces f . Using eq. (P3.4), these can be
written as
θ

 ∫ L ∫ 2π cos 
 (p − p ) Rdθdz 
a
θ
 


 0 0 sin 


θ + 1
∆θ
 
 ∫ L ∫ 1 2 1

(p − p ) Rdθdz Zee Zew
 
f= a
= f0 + q̃eiωs t, (P3.19)

 

0 θ −
1 2
1
∆θ 1
.
. Z we Z ww
.

 


 


 ∫ L ∫ θ N + 1 ∆θ N 

2
(p − pa ) Rdθdz 

 
θ 1
∆θ
 0 
 −
N 2 N 
where the static force is given as the vector
θ

 ∫ L ∫ 2π cos 
 (p − p ) Rdθdz 
0 a
θ
 


 0 0 sin 


θ + 1
∆θ
 
 ∫ L ∫ 1 2 1

(p − p ) Rdθdz
 
f0 = 0 a
∈ R2+N , (P3.20)

 

0 θ 1 − 2 ∆θ 1
1
.
.
.

 


 


 ∫ L ∫ θ N + 1 ∆θ N 

2
(p − p ) Rdθdz
 
0
 a

θ N − 2 ∆θ N
1
 0 
 
and the dynamic forces are expressed in terms of the perturbation vector
(P3.21)
T
q̃ = ∆e x ∆e y ∆w1 · · · ∆wN ∈ R2+N ,

along with the matrix Z with blocks given as

∫ L ∫ 2π
p x cos θ p y cos θ

Zee = Rdθdz ∈ C2×2, (P3.22)
0 0 p x sin θ p y sin θ
∫ L ∫ 2π
pw1 cos θ · · · pw N cos θ

Zew = Rdθdz ∈ C2×N , (P3.23)
0 0 p w1 sin θ · · · p wN sin θ
θ
 L 1 2 1
∫ ∫ + 1
∆θ
 0 θ − 1 ∆θ p x p y Rdθdz 

1 2 1
.
 
Zwe =  .
.  ∈ CN×2, (P3.24)
 
 L θ N + 12 ∆θ N
∫ ∫ 
p x p y Rdθdz

 0 θ − 1 ∆θ
 N 2 N 
∫ L ∫ θ 1 + 1
∆θ 1

 0 θ − 12∆θ pw1 · · · pw N Rdθdz 
 
1 2 1
.
 
Zww =  .
.  ∈ CN×N . (P3.25)
 
 L θ N + 12 ∆θ N
∫ ∫ 
pw1 · · · pw N Rdθdz

 0 θ − 1 ∆θ
 N 2 N 
Finally, Z is related to the dynamic stiffness and damping coefficients as

 kxx
 k xy k xw1 ··· k xw N 
 k yx
 k yy k yw1 ··· k yw N 
= Re (Z) =  1 ··· ∈ R(2+N)×(2+N),
 kw x kw y kw w k w1 w N 
Kfluid 1 1 1
 .. .. .. ... .. 
 . . . . 

 k w N x k w N y k w N w1 · · · k w N w N 
(P3.26)

 dx x
 d xy d xw1 ··· d xw N 
 dyx dyy dyw1 ··· dyw N 
1 
= Im (Z) =  1 ··· ∈ R(2+N)×(2+N) .
 dw x dw y dw w dw1 w N 
Dfluid 1 1 1
ωs  .. .. .. .. .. 
 . . . . . 

dw N x dw N y dw N w1 ··· dw N w N 
(P3.27)

P3.2.4 Including the foil stiffness and damping

The bearing coefficients in eqs. (P3.26) and (P3.27) are derived from the dynamic
pressure fields and thus contain merely the contributions from the fluid film. When
evaluating the displacement of the j-th foil DOF from its equilibrium w j0 , the
stiffness and damping of the foil itself should furthermore be considered.
Various elaborate foil structure models are available in the literature [45, 96, 125,
155], but the added complexity of these would not benefit the present comparison.
Instead, the Simple Elastic Foundation Model (SEFM), as introduced by Heshmat
et al. [62], is employed with a uniform baseline stiffness derived from the widely
applied expression by Walowit and Anno [176]
3
Eb t b
(P3.28)
−1
k= 1 − νb2 ,
2Sb l0
where Eb , Sb , tb , l0 and νb are the bump foil’s Young’s modulus, pitch, thickness,
half bump length and Poisson’s ratio, respectively. The damping is assumed
viscous and expressed simply as d = kη/ωs where η is a loss factor. A fundamental
assumption of the SEFM is that all points in the foil behave independently,
meaning that the foil structure is assumed not to contribute any cross couplings.
Furthermore, the foil is not directly affecting the e x, e y DOFs, meaning that the
stiffness and damping matrices stemming from the bump foil can be written as
Kfoil = LRk diag 0 0 ∆θ 1 · · · ∆θ N ∈ R(2+N)×(2+N),

(P3.29)
Dfoil = ηωs−1 Kfoil ∈ R(2+N)×(2+N),
where L is the axial length of the bearing and ∆θ j is the angular segment ascribed
to the j-th foil DOF as illustrated in fig. P3.1(b).
P3.2.5 The mass matrix

The rotor is modelled as a point mass, implying that neither polar nor transverse
inertia is considered. Combining the rotor mass with a rough estimate of the foil
mass ascribed to each foil DOF, the system mass matrix is constructed as
M = diag m x m y mfoil0 R∆θ 0
∈ R(2+N)×(2+N), (P3.30)

1 · · · mfoil R∆θ N
where m x = m y is the rotor mass and mfoil

0 is the average mass of the foil structure
per unit circumferential length.
P3.2.6 System assembly

Combining eqs. (P3.26), (P3.27), (P3.29) and (P3.30), the linearised 2+N equations
of motion for the rotor–bearing system can be written as
MqÜ̃ + (Dfoil + Dfluid ) qÛ̃ + (Kfoil + Kfluid ) q̃ = 0, (P3.31)
| {z } | {z }
≡D(ωs,Ω) ≡K(ωs,Ω)
P3.3 Review of the classical two-DOF perturbation 101
which can be recast into first order form using the state vector z = q̃ qÛ̃ as
T

0 I
zÛ = z, (P3.32)
M−1 K (ωs, Ω) M−1 D (ωs, Ω)
| {z }
≡A(ωs,Ω)
where 0 and I denote (2 + N) × (2 + N) zero and identity matrices, respectively.
P3.3 Review of the classical two-DOF perturbation

To emphasize the novel aspects of the proposed extended perturbation, a brief
review of the traditional two-DOF approach employed with variations by e.g. [68,
79–81, 87, 93, 101, 132, 175] is included. As originally introduced in the appendix
of [110], a fundamental assumption is for the pressure field to be given as a function
of rotor position and velocity only. This means that a Taylor expansion leads to a
perturbed pressure field given as
p = p0 + p x Λ x + p y Λ y , (P3.33)
which should be compared to eq. (P3.4) of the extended perturbation. The foil
compliance is then introduced implicitly through the film height function as
h = C + e x cos θ + e y sin θ + fhc (p) , (P3.34)
where fhc (p) is some function supplying the compliant film height as a function of
pressure. For the SEFM incorporating a loss factor, this would be
fhc (p) = k c−1 (p − pa ) where k c = k (1 + iη) . (P3.35)
Inserting the film height eq. (P3.34) and perturbed pressure eq. (P3.33) into
RE from eq. (P3.15), equations for p0 and pγ can be separated. The zeroth
order equation is identical to the one obtained for the extended perturbation in
eq. (P3.16), while the first order equations for pγ become
! ! !
∂ p0 h03 ∂pγ ∂ p0 h03 ∂pγ ∂ pγ h02 h0 + 3p0 k c−1 ∂p0

+ +
∂Θ 12µ ∂Θ ∂z 12µ ∂z ∂Θ 12µ ∂Θ
!
∂ pγ h02 h0 + 3p0 k c−1 ∂p0

∂

+ pγ h0 + p0 k c−1

− U
∂z 12µ ∂z ∂Θ
! (P3.36)
2f
∂ p 0 h 0 γ ∂p 0
−iωs pγ h0 + p0 k c−1 = −

∂Θ 4µ ∂Θ
!
∂ p0 h02 fγ ∂p0 ∂
+ p0U fγ + iωs p0 fγ, γ = x, y

−
∂z 4µ ∂z ∂Θ
where
h0 = C + e x0 cos θ + e y0 sin θ + k c−1 (p0 − pa ) . (P3.37)
which should be compared to eq. (P3.17) of the extended perturbation (the
differences are framed in square brackets). The classical technique hence requires
the solution of a nonlinear equation for p0 and two linear complex equations for
pγ from which the bearing coefficients can be extracted as
∫ L ∫ 2π
p x cos θ p y cos θ

k x x k xy d x x d xy
+ iωs = Rdθdz ∈ C2×2,
k yx k yy dyx dyy 0 0 p x sin θ p y sin θ
(P3.38)
involving only the two rotor DOFs (per bearing).
P3.4 Solution strategy

The zeroth order eq. (P3.16) is a nonlinear real equation in p0 where h0 is given
from eq. (P3.14). The assumption of an axially uniform deformation combined
with the SEFM means that when solving for p0 , the discrete foil compliance w j is
calculated in each iteration from the local mean axial static pressure as
1 1 L
∫
wj = p0 (θ j )dz − pa . (P3.39)
k L 0
The zeroth order equation is identical to that presented by the authors in [93] and
the same solution strategy based on an FE discretization and Newton-Raphson
iteration is employed. Notice that sub-ambient pressures are discarded when
integrating the pressure fields (the Gümbel condition). As described in [93], this
is meant to represent the effect of the top foil separation from the bump foil [62].
The 2 + N first order eqs. (P3.17) and (P3.18) for pγ and pw j are linear complex
equations which are solved using the FE discretization documented in [93]. Notice
that while [93] treats the solution to the pγ equations produced by an implicit
treatment of the foil compliance as in eq. (P3.36), this collapses into eq. (P3.17)
for k c−1 = 0 and the same solution method can be applied.
A converged mesh of 17 axial (over half the bearing length, as symmetry is
exploited) and 114 circumferential elements is used giving 2070 nodes. The number
of circumferential locations, and hence the number of pw j equation systems to be
solved, is thus N = 115.
P3.5 Results & discussion

The proposed method is evaluated by predicting the OSI for the simplest possible
case of a point mass supported by a single-pad GFB in the widely studied config-
P3.5 Results & discussion 103
uration [154] listed in table P3.1. The OSI is calculated using: (a) a simultaneous
nonlinear time integration; (b) a classical two-DOF perturbation; and (c) the
proposed extended perturbation. For the present bearing, eq. (P3.28) predicts a
foil stiffness of 4.6417 GN/m3 (case 2) which is evaluated along with a practically
rigid (case 1: 4641.7 GN/m3 ) and a more flexible (case 3: 2.3209 GN/m3 ) variant.
To put these stiffness levels into perspective, the maximum static foil deflection in
the three cases at 20 kRPM are 0.00, 0.21 and 0.40 times C respectively.
P3.5.1 Stability analysis: Frequency domain

To assess the linear stability limit of the rotor system described by eq. (P3.32),
a solution on the standard form z̃ = z̃0 eλt is substituted to produce the standard
eigenvalue problem A (ωs, Ω) z̃ = λz̃. Having solved this, an eigenvalue λi with
positive real part then indicates the corresponding eigenmode z̃i to be unstable.
Applying the classical two-DOF perturbation to a single bearing with coeffi-
cients given from eq. (P3.38), four eigenvalues in complex conjugated pairs are
obtained. By contrast, the extended perturbation in its current implementation
and with the applied mesh results in 2 (2 + 115) = 234 eigenvalues. As the eigenval-
ues of both methods depend on the oscillation frequency and the angular velocity,
the stability limit, i.e. the OSI, is characterised by the condition
Re (λi (ωs, Ω = ΩOSI )) = 0 ∧ Im (λi (ωs, Ω = ΩOSI )) = ωs, (P3.40)
Table P3.1: Properties of the GFB configuration, mainly from [154].
Journal
Load, W x = m x g 30 N Mass, m = m x = m y 3.059 kg
Bearing radius, R 19.05 mm Pad leading edge, θ l 0◦
Bearing length, L 38.10 mm Pad trailing edge, θ t 360◦
Radial clearance, C 31.80 µm
Fluid properties
Ambient pressure, pa 101.3 kPa Viscosity, µ 1.950×105 Pa s
Foil structure properties
Foil thickness, tb 0.1016 mm Bump pitch, Sb 4.572 mm
Bump half length, l0 1.778 mm Bump height, hb 0.5080 mm
Young’s modulus Eb 207.0 GPa Poisson’s ratio, νb 0.3
Density, ρb 8280 kg/m3 Loss factor, η 0.2
requiring the eigenvalue problem to be solved over a range of excitation frequencies

and angular velocities. Here it should be highlighted that one should keep track
of the eigenvalue order in between the (ωs, Ω) evaluation points. This is a trivial
task for the two complex conjugated pairs arising from the two-DOF approach,
but becomes a significant challenge for the 117 pairs obtained from the extended
perturbation. At present, this has been handled by comparing the structure of the
corresponding eigenvectors.
The result is shown in fig. P3.2, where the contours of the logarithmic decrement
(LD) δ = −2πRe (λ) /Im (λ) is plotted along with a single contour showing the
concurrence of the excitation frequency ωs and the damped natural frequency
ωd = Im (λ). The crossing of the δ = 0 contour is exactly where eq. (P3.40) is
fulfilled and thus marks the OSI. The shown plot is for one of the two eigenmodes
dominated by the rotor and is the one first becoming unstable. The mass of the
foil from eq. (P3.30) used in the eigenvalue solution is a rough estimate, but it has
been verified that this has a vanishing influence on the OSI.
A similar plot is shown in fig. P3.3, but obtained using the traditional two-
DOF perturbation method. The maps are similar, but the OSI is shifted more
than 4 kRPM and it is worth emphasizing the difference in computational cost.
While both stability maps are based on solutions to 2805 eigenvalue problems, the
extended perturbation requires solving 117 2070 × 2070 complex linear equation
35000 1.516
-0.0
5 0.05
30000
0.00 0.050
Ω (RPM)
25000
δ (–)
s 0.000
20000 ou
h ron
y nc
ωs = ωd
15000
S
−0.050
10000
−0.338
100 200 300 400
ωs (Hz)
Figure P3.2: Case 3 stability map based on the extended perturbation

and eigenvalue solutions over a 51 × 55 (ωs, Ω) grid. The OSI is marked at
(92.8 Hz, 27 424 RPM).
P3.5 Results & discussion 105
systems to obtain the coefficients forming each of the 234 × 234 eigenvalue problems
to be solved. In the traditional perturbation, each eigenvalue problem is merely
4 × 4 and the coefficients can be extracted from just two complex linear equation
system solutions. The effort related to solving the zeroth order equation for each
value of Ω is, however, the same for both methods. For the current implementation,
the two-DOF perturbation calculations for fig. P3.3 take around 45 s while the
extended perturbation calculations for fig. P3.2 take 10 times longer.
P3.5.2 Stability analysis: Time domain

In the nonlinear sense, the stability limit is characterised by a breakdown of
structural stability due to a small change in the angular velocity. At this point, the
system Jacobian evaluated at the singular point representing the static equilibrium
obtains an eigenvalue with zero real part. After this point, the behaviour will be
governed by the system’s nonlinear characteristics and should not be compared to
the linear solutions.
To assess the stability limit, eq. (P3.15) is FE discretized and simultaneously
integrated in time with zero unbalance as described in [88]. The implementation has
been carefully aligned to that used when solving the perturbed equations. A time
integration can hence be started from the static equilibrium found from eq. (P3.16)
35000 1.617
0.0
-0.05 0.05
0
30000
0.050
Ω (RPM)
25000
δ (–)
s 0.000
20000 ou
hron
y nc
ωs = ωd
15000
S
−0.050
10000
−0.175
100 200 300 400
ωs (Hz)
Figure P3.3: Case 3 stability map based on the two-DOF perturbation

and eigenvalue solutions over a 51 × 55 (ωs, Ω) grid. The OSI is marked at
(95.7 Hz, 31 769 RPM).
and the development of the oscillation amplitude can be observed. Such analysis
is, however, subject to several challenges. First, even for an asymptotically stable
state, the vibration amplitude of a numerical integration will never reach zero.
Instead, it converges towards a noise floor specific to the numerical implementation.
Initiating a time integration from a static equilibrium will hence always imply some
oscillation. Second, even before the true structural stability limit is reached at
Ω = ΩOSI , the static equilibrium state’s basin of attraction will gradually narrow,
increasing the possibility that an incidental numerical perturbation will push the
system away from the stable equilibrium.
To demonstrate the uncertainty related to the OSI using time integration, 1 s
simulations have been initialized from their static equilibria for every 10 RPM
in the range 23–25 kRPM. For each response it is easy to judge whether the
oscillation remains within the numerical noise level or grows exponentially. In
fig. P3.4(a), it is shown that some responses grow exponentially from around
23 kRPM, but that this is not persistent until 25.5 kRPM. This is, however, highly
dependent on the tolerances applied in both the steady state solver and the
subsequent time integration. More consistent results were found to be achievable
when perturbing the static eccentricity just enough to leave the numerical noise
floor. An initial perturbation of 10−10 on the eccentricity ratio (= 31.8×10−16 m)
thus results in tiny, but very clean oscillations from which the decay or growth
can be identified. Plotting the LD obtained from curve fitting to each time series
shows a monotonically decreasing curve with a distinct zero-crossing as shown in
fig. P3.4(b). The OSI identified using this method is denoted Ω̃OSI to distinguish
it from the more firmly defined OSI resulting from the frequency domain analysis.
P3.5.3 Comparison of OSIs

The OSIs obtained using the perturbation methods and time integration are listed
in table P3.2 for all three cases of foil stiffness. The excitation frequencies at the
OSI are likewise listed for comparison. For the time integration, the latter has been
extracted from frequency spectra based on 5 s simulations at angular frequencies
slightly above Ω̃OSI . The relative OSI discrepancies are plotted in fig. P3.5 as a
function of foil flexibility, and the extended perturbation is seen to provide results
better in agreement with the nonlinear time integration results.
P3.6 Conclusion
Discrepancies in OSI between the traditional perturbation approach and time
integration results were pointed out in [88], and this was postulated to be caused
by a deficiency in the pressure expansion. This has been investigated in the present
P3.6 Conclusion 107
Decay
0.025 24 015 RPM
0.000
δ [-]
−0.025
−0.050
Growth
23.0 23.5 24.0 24.5 25.0 24 25

Ω (kRPM) Ω (kRPM)
(a) (b)
Figure P3.4: (a) Stability of static equilibria judged from 1 s simulations for
every 10 RPM (case 2); (b) LDs obtained by curve fitting to 0.25 s simulation after
a rotor perturbation of C × 10−10 = 31.8×10−16 m for every 10 RPM (case 2).
Table P3.2: Comparison of obtained OSIs
Foil Time Two-DOF Extended

stiffness integration perturb. perturb.
Case
k Ω̃OSI ω̃s ΩOSI ωs ΩOSI ωs
GN/m3 (RPM) (Hz) (RPM) (Hz) (RPM) (Hz)
1 4641.7 20 868 112.8 21 731 107.5 21 731 107.5
2 4.6417 24 015 101.8 27 868 103.4 24 987 101.0
3 2.3209 25 118 91.7 31 769 95.6 27 424 92.8
ΩOSI /Ω̃OSI (-)
Two-DOF 26 %
1.2 16 %
Extended
9.2 %
1.1
4.1 % 4%
0.0 0.1 0.2 0.3 0.4

−1 3
Flexibility k (m /GN)
Figure P3.5: Foil structure flexibility versus the discrepancy in the predicted
OSI relative to the time integration result for both perturbation methods.
work by carrying through the perturbation method while including the additional
pressure–compliance terms in the Taylor expansion. In the presented form, these
terms are treated straightforwardly using a high number of discrete foil deflections
leading to a rather comprehensive model. This approach is debatable, as are
the concrete choices of GFB properties and numerical schemes. The extended
perturbation method has, however, been demonstrated to predict OSIs in much
better agreement to time integration results, indicating the pressure–compliance
terms to be significant.
Irrespective of the physical interpretation of the foil DOFs and their subsequent
implementation, the fluid film coefficients affecting each foil DOF are eventually
added to the contribution from the foil structure itself. The significance of the
pressure–compliance terms will hence decrease for an increasingly stiff foil structure,
matching the expectation for the two–DOF perturbation to be adequate for rigid
bearings. This has also been demonstrated to be the case using the present
implementation, as the methods converge for a very stiff foil.
When lowering the foil stiffness, the significance increases as the discrepancy
of the two-DOF method grows and the two perturbation methods diverge. The
baseline discrepancy of around 4 % is ascribed to the determination of the time
integration OSI, while the increased inconsistency of 9.2 % for case 3 should be
further investigated. The agreement is, however, still significantly better than the
26 % discrepancy of the classical two-DOF method.
In conclusion, caution should be exercised if the classical perturbation method
is applied to compliant bearings and the OSI of these should rather be assessed
using either nonlinear methods or an extended perturbation as presented here.
References
PUBLICATION P4
Modelling of Compliant-type
Gas Bearings:
A Numerical Recipe
This chapter is a postprint of the identically titled article [127] presented at the
13th International Conference on Dynamics of Rotating Machinery (SIRM 2019)
in Copenhagen, Denmark on 13–15 February 2019 and published in the conference
proceedings.
Sebastian von Osmanski1 , Jon S. Larsen2 , Ilmar F. Santos1,*

1 Dept. of Mechanical Engineering, Technical University of Denmark, 2800 Kgs.
Lyngby, Denmark
2 GEA Process Engineering A/S, 2860 Søborg, Denmark
Abstract
Despite its merits, the Gas Foil Bearing (GFB) suffers from several inherent
limitations which could likely be overcome using active radial injection. This
hence represents a natural next step, even though several issues relating to the
foil structure modelling remain unresolved. A prerequisite for the development
of any model-based feedback control scheme is a model capturing the effects of
gas injection on the system dynamics. The currently presented work on a generic
numerical recipe for GFB simulation is intended as a step towards such a model.
The aim of the present work is to consolidate the existing state of the art
knowledge on GFB modelling into a generic framework that can act as an efficient
platform for further research. By creating a structure where the same pieces of code
can be applied for simulation of a wide range of rotors supported by any number of
rigid or compliant type gas bearings — with or without injection — the available
experimental results can be utilized optimally for validation. Furthermore, the
programming intensive optimizations necessary to obtain tolerable solution times
can be more easily reused.
110 P4 Modelling of Compliant-type Gas Bearings: A Numerical Recipe
The framework consists of three domains treating the rotor, the foil structure
and the fluid film respectively, along with clear-cut domain interfaces based on
linear mappings. In the fluid domain, the film is modelled using the Modified
Reynolds Equation dizcretised using finite volume, leaving the injection flow to an
auxiliary model. Both the rotor and the compliant structure are represented in
generic state-space formats facilitating various different models for both domains.
The global system is solved both statically and in time using efficient general
purpose routines exploiting e.g. analytical Jacobian matrices and sparse direct
linear solvers.
BC Boundary Condition ∇· Divergence

CDS Central Differencing Scheme (˜) Non-dimensional quantity
CFD Computational Fluid Ab Bearing surface area
Dynamics Ab = [0, L] × [0, 2π]
CSR Compressed Sparse Row Ai In-film area of the i-th CV
CV Control Volume aj FV stencil coefficients
DAE Differential/Algebraic constituting AFV
Equation ftol Global force tolerance
DOF Degree of Freedom h, h̃ Film height, h̃ = h/C
EOM Equation of Motion L, L̃ Bearing length, L̃ = L/R
FD Finite Difference lf Foil domain length scale
FE Finite Element lr Rotor domain length scale
FV Finite Volume ltol Global length tolerance
GFB Gas Foil Bearing Mg Molar mass of gas
IC Initial Condition Minj,i Integrated injection flux for
IVP Initial Value Problem the i-th CV
LUDS Linear Upwind Differencing minj Injection flux field
Scheme mr Rotor domain mass scale
MRE Modified Reynolds Equation ncv Number of CVs (per bearing)
ODE Ordinary Differential Equation nb Number of GFBs
SEFM Simple Elastic Foundation nrdof Number of rotor domain DOFs
Model nub Number of unbalance masses
UDS Upwind Differencing Scheme p, p̃ Film pressure, p̃ = p/pa
(1st order) pa Ambient pressure, fluid
(Û) Time derivative, d/dτ pressure scaling
(Ü) Time derivative, d 2 /dτ 2 pf Foil domain pressure scale
∇ Gradient pi Pressure at i-th CV centre
ptol Global (derived) pressure h0,cve Initial film heights at CV edges

tolerance hc,cvc Compliant film heights at CV
R Journal radius centres
rFV,i FV equation residual for i-th hc,cve Compliant film heights at CV
CV edges
Ru Universal gas constant hr,cvc Rigid film heights at CV
Si In-film circumference of the centres
i-th CV hr,cve Rigid film heights at CV edges
Sτ Non-dimensional group, n Outward pointing unit normal
Sτ = 6R2 µωτ vector
C 2 pa
Tiso Isothermal gas temperature pα , p̃α Bearing α CV centre pressures,
t Physical time pα = pa p̃α
U Unbalance, mass × eccentricity r f ,α Foil structure residual in
xg, yg, zg Fluid film coordinates, yg is bearing α
curvelinear rFV,α FV residual vector for bearing
xr , yr , zr Rotor coordinates α
α Bearing index, α = 1, . . . , nb rr Rotor residual vector
µ Dynamic gas viscosity rscale Nondimensionalization scalings
Ω, Ω̃ Angular velocity Ω̃ = Ω/ωτ for rr
ωτ Characteristic frequency rtol Dimensional system residual
φ Phase of unbalance tolerances
ρg Density of gas ug , ũg Gas in-film velocity vector,
Dimensionless time, τ = ωt t
T
τ ug = u θ u z

θ Circumferential angle x f , x̃ f Foil structure DOF vector

θ̃ Circumferential coordinate, xr , x̃r Rotor DOF vector
θ̃ = yg /R z f ,α Foil structure state vector for
Ξ FV stencil, i.e. a set of CV bearing α
indices zr Rotor state vector
fb Bearing force vector Af Foil structure state-space
f f w,α Foil structure static loads in inertia
bearing α AFV FV coefficient matrix, Couette
f µ,α Foil structure friction forces in & Poiseuille
bearing α Ar Rotor state-space inertia
f p,α Foil structure pressure loads in Bf Foil structure state-space
bearing α stiffness/damping
frw Rotor static force vector BFV FV coefficient matrix, squeeze
fub Rotor unbalance force vector term
h0,cvc Initial film heights at CV Br Rotor state-space
centres stiffness/damping
Cr Rotor state-space gyroscopic H fb,p Mapping from bearing

matrix pressures to Tr f̃b
D f , D̃ f Foil structure damping matrix H fp,p Mapping from pα to T f f̃ p,α
Dr , D̃r Rotor damping matrix I Identity matrix
CFV FV coefficient matrix, local K f , K̃ f Foil structure stiffness matrix
expansion term Kr , K̃r Rotor stiffness matrix
Gr , G̃r Rotor gyroscopic matrix M f , M̃ f Foil structure mass matrix
Mr , M̃r Rotor mass matrix
Hcvc,r Mapping from x̃r to hr,cvc
Tr Rotor DOF to state-space
Hcvc, f Mapping from x̃ f ,α to hc,cvc mapping
Hcve,r Mapping from x̃r to hr,cve Tf Foil structure DOF to
Hcve, f Mapping from x̃ f ,α to hc,cve state-space mapping
P4.1 Introduction
Rotor–bearing systems supported by Gas Foil Bearings (GFBs) often display
complex dynamics originating from the non-linear behaviour of the compressible
gas film and its interaction with the compliant structure. To understand and
predict such a behaviour, several researchers have contributed improvements
to the available GFB modelling tools over the recent years. These include the
introduction of the simultaneous solution approach [19] and work on various foil
structure models [96, 121, 125]. Even though several issues remain uresolved for
conventional GFBs, especially regarding the foil structure damping, a natural next
step concerns active radial injection. This could potentially overcome some of the
passive GFB’s inherent issues and work on hybrid GFBs have been presented [182,
183].
The aim of the present work is to generate a fast and versatile simulation
tool for radial GFBs. This should be capable of modelling an arbitrary number
of rigid or compliant type gas bearings supporting a variety of rigid or flexible
rotors. The tool should rely on fast and readily available general purpose solvers
to provide simultaneous steady-state solutions as well as time integrations while
reusing as much code as possible. This allows the available base of experimental
and theoretical gas bearing results from the literature to be utilised optimally for
validation purposes.
A prerequisite for the development of any model-based feedback control scheme
would be a model capturing the effects of gas injection on the system dynamics.
The currently presented work on a generic numerical recipe for GFB simulation is
intended as a step towards such a model.
Journal
Bearing
housing
Bump foil
Top foil
Injection
pressure

Piezoelectric
actuator
(a)
xg
zg
yg = θR
yr
θ
zr
xr
(b)
Figure P4.1: (a): A three-pad GFB augmented with a single injector. (b):
Illustration of the discretized fluid film along with its coordinate system and the
rotor. Notice that the fluid film is unwrapped meaning that yg is curvilinear in
the illustration.
P4.2 Rotor Domain

The model is build to support any type of rotor representable in a general state-
space form using the rotor Degree of Freedom (DOF) vector x̃r as
0 M̃r xÛ̃ r

K̃r D̃r 0 −G̃r x̃r I
+ +Ω̃ − f̃rw + f̃ b + f̃ ub = rr ,
I 0 xÜ̃ r 0 −I 0 0 xÛ̃ r 0
| {z } |{z} | {z } | {z } |{z} |{z}
Ar zÛ r Br Cr zr Tr
(P4.1)
where M̃r , D̃r , G̃r and K̃r are the non-dimensional rotor mass, damping, gyroscopic
and stiffness matrices, respectively. The static forces are kept in f̃rw , while f̃ub
represents the unbalance forcing and f̃b holds the bearing forces. No restrictions are
made on the origin of the rotor equations which could represent a point mass with
just two linear displacement DOFs in x̃r , a rigid rotor with a mix of rotational and
linear displacement DOFs or a Finite Element (FE) based rotor with potentially
hundreds of DOFs. For some models, involving e.g. a reduced order rotor, a
more complicated Tr would be required. Notice that the momentum balance in
eq. (P4.1) is equated to a rotor residual rr , which should be minimized when
solving the non-linear system either in time as an IVP or as an algebraic system
in steady-state.
All quantities in the rotor domain are non-dimensionalized using a global
characteristic frequency ωτ , a rotor length scale lr and a rotor mass scale mr .
The frequency ωτ gives a non-dimensional angular velocity Ω̃ = Ω/ωτ which is
shared between the domains. It will often be sensible to define ωτ simply as the
angular velocity implying Ω̃ = 1, but other choices could be advantageous for
some simulations [19, 103]. The choice of rotor length and mass scales depends
on the specific rotor and the underlying mathematical model, but these can be
tuned to enhance the numerical properties of the rotor matrices. A point mass
rotor supported by a single GFB will have D̃r , G̃r and K̃r equal to zero, hence
the scales could be chosen to provide M̃r = I while maintaining sensible numerical
force values. For a more complex rotor, based on e.g. FE discretization with beam
elements, a good choice is less obvious.
The force vector fb contains the vertical xr and horizontal yr components of
the fluid film forces on the rotor stemming from all GFBs in the system and hence
has 2nb non-zero elements. For each bearing, the forces are found from the usual
integration of the projected fluid film pressures as
cos θ
∬
fb,xr
= (p − pa ) RdA where Ab = [0, L] × [0, 2π] . (P4.2)
fb,yr Ab sin θ
The force vector f̃ub contains the components of the unbalance forces due to
each unbalance mass U attached to the rotor and hence has 2nub non-zero elements.
P4.3 Foil Domain 115
For one particular unbalance mass attached to a node of the rotor, the unbalance
forces are found using the usual projections as
f˜ub,xr cos Ω̃τ + φ

= Ũ Ω̃2, (P4.3)
f˜ub,yr sin Ω̃τ + φ
where the non-dimensional unbalance mass is given as Ũ = U/(mr lr ).
P4.3 Foil Domain

Various different foil structure models are available in the literature of which the
majority are based on variations over the Simple Elastic Foundation Model (SEFM).
The available models include purely algebraic equation systems (stiffness only),
first order systems (including stiffness and damping) as well as full second order
systems including also mass and possibly friction. In the current implementation,
a format capable of containing all these variations is desirable. Therefore the foil
structural model in a bearing α is represented using the non-dimensional foil DOF
vector x̃ f ,α in the general state-space residual format as
0 M̃ f xÛ̃ f ,α

K̃ f C̃ f x̃ f ,α I
+ + f̃ f w,α + f̃ p,α + f̃ µ,α = r f ,α, (P4.4)
I 0 xÜ̃ f ,α 0 −I xÛ̃ f ,α 0
| {z } | {z } | {z } | {z } |{z}
Af zÛ f ,α Bf z f ,α Tf
which is suitable for the most general class of foil models. The non-dimensional
foil structure mass, damping and stiffness matrices are given as M̃ f , D̃ f , K̃ f
respectively, while the static loads are represented in f̃ f w,α , the fluid film pressure
loads are contained in f̃ p,α and the friction forces (if any) are held by f̃ µ,α . For a
non-zero mass matrix, this represents a system of Ordinary Differential Equations
(ODEs), but for many models, such as the widely applied case of an inertialess
SEFM incorporating a structural loss factor, eq. (P4.4) turns into a system of
Differential/Algebraic Equations (DAEs). If friction forces are not included,
the DAE structure could still be used directly as discussed in [126], but this is
unnecessarily complicated. Instead, the implementation also supports a decayed
version of eq. (P4.4) given simply as

C̃ f xÛ̃ f ,α + K̃ f x̃ f ,α + I f̃ f w,α + f̃ p,α = r f ,α, (P4.5)
|{z} |{z} |{z} |{z} |{z}
Af zÛ f ,α Bf z f ,α Tf
representing a system of first order ODEs for non-zero damping matrices or an

algebraic equation system in the case of zero damping. As for the rotor domain
equations, all quantities in the foil domain are non-dimensionalized. This is

achieved using the characteristic frequency ωτ , a foil domain length scale l f and a
foil domain pressure scale p f . For an SEFM-like model, the obvious choices for
these would be the clearance and ambient pressure, but for more complex models
other scales might be advantageous.
P4.4 Fluid Film Domain

To model the gas film separating the journal from the rigid or compliant surfaces
of the bearing, we will apply the Modified Reynolds Equation (MRE) under the
assumption of isothermal, isoviscous conditions. The compressibility is introduced
by assuming ideal gas behaviour providing the relation for the gas density ρg
Mg
ρg = p, (P4.6)
RuTiso
where Mg , Ru and Tiso are the (average) molar mass of the gas, the universal gas
constant and the isothermal gas temperature, respectively. A natural extension
would be to drop the assumption of isothermal conditions and simultaneously solve
the energy equation. This has been done for rigid gas bearings, a least using non-
simultaneous approaches [131], but for foil bearings the thermal BCs are less clearly
defined. The maximum temperature differences in the circumferential and axial
directions were found experimentally by Żywica et al. [189] for a specific GFB to
be approximately 35 ◦C and 2 ◦C, respectively. This would imply density variations
on the order of ±5 %, which is not considered sufficient to justify the inclusion
of the energy equation. If this was nevertheless to be attempted, experimental
validation of the resulting temperature field would be necessary to adjust the
applied thermal BCs.
P4.4.1 The Modified Reynolds Equation in Terms of Flows

To formulate a discretization scheme for the MRE, it is convenient to consider the
underlying velocity profiles. Expressing the journal surface velocity as uθ xg = 0 =
ΩR and assuming the pressure and viscosity to be constant across the film thickness
h, the velocity profiles can be derived from the Navier–Stokes equations using an
order of magnitude analysis followed by an integration across the film thickness
[56] as
T
h − xg
(P4.7)
T
−xg 2µg ∂∂p
h−x ∂p
ug = u θ u z = h−x
,

yg + ΩR −xg 2µg ∂z
h g
P4.4 Fluid Film Domain 117
giving the in–film fluid velocities in m/s as functions of the pressure gradients and
the journal surface velocity. Note that the axial velocity of the journal is neglected
leaving only the Poiseuille component in this direction.
The MRE can be obtained by inserting the velocity profiles into the continuity
equation augmented with a source term to represent injection and integrating it
across the film thickness as
∂ ρg
∫ h
+ ∇ · ρg ug dxg = minj, (P4.8)

0 ∂t
where ρg is the fluid density and minj = minj yg, zg represents an injection source

term field in kg/(m2 s). The inflow due to injection could be obtained using
analytical expressions, possibly including empirical correction factors, as suggested
by various authors [6, 116, 141, 182] or by an external Computational Fluid
Dynamics (CFD) routine, but this is outside the present scope.
Replacing the fluid density by its average across the film thickness and applying
Leibniz’ rule, eq. (P4.8) produces the MRE on the form
∂ ρg h3 ∂p ∂ ρg h3 ∂p ωR ∂ ρg h ∂ ρg h

− − + + = minj, (P4.9)
∂ yg 12µ ∂ yg ∂zg 12µ ∂zg 2 ∂ yg ∂t
which can alternatively be expressed in terms of the flow vector qg with units
kg/(m s) [33] as
∂ ρg h ρg h 3

ΩR
+ minj where qg = qθ qz = −
T T
∇ · qg = − ∇p + ρg h 1 0 .

∂t 12µ 2
(P4.10)
Equation (P4.10) can be non-dimensionalised using the clearance C and radius
R as through-film and in-film length scales respectively, the ambient pressure pa
as the pressure scale and a frequency ωτ as time scale. Assuming isothermal
conditions in the fluid film, the viscosity can be assumed constant and the gas
density can be expressed using the ideal gas law. This gives the version of the
MRE implemented in the model
∂ p̃ h̃

∇ · q̃g = −2Sτ + m̃inj, (P4.11)
∂τ
where
6R2 µωτ T T
Sτ = , q̃g = q̃θ q̃z = − p̃ h̃3 ∇ p̃ + Sτ Ω̃ p̃ h̃ 1 0

C 2 pa
(P4.12)
12R2 µRuTiso
and m̃inj = minj .
p2a C 3 Mg
Notice that before eq. (P4.11) is non-dimensionalised, insertion of the ideal gas
law and multiplication by RuTiso /Mg changes its units into J/(m2 s), such that
the conserved quantity is rather energy than mass. Subsequently, the equation is
divided by the viscosity µ resulting in the rather awkward unit J Pa s/(m2 s) or
N2 /m3 .
P4.4.2 Finite Volume Discretization

Having the MRE represented in terms of flows as given in eq. (P4.11), the Finite
Volume (FV) method can be applied and implemented following an intuitive
approach. Equation (P4.11) represents the differential form of the conservation
equation which can be integrated over a Control Volume (CV) i with in-film area
Ai and circumference Si . Applying the divergence theorem, this can be written as
∂
∫ ∬ ∬
q̃g · ndS = −2Sτ p̃ h̃ dA + (P4.13)

m̃inj dA,
Si Ai ∂τ Ai
where the three terms represent the flux across the CV boundary to the neighbour-
ing cells, the accumulation within the CV and the injection influx, respectively.
Limiting the implementation to rectilinear non-uniform grids, i.e. grids of quad-
rilateral CVs, the integral over the CV circumference in eq. (P4.13) can be split
into four components as illustrated in fig. P4.2(b). Representing the flow q̃g at
each face by its midface value and the unsteady term ∂τ p̃ h̃ by its cell centre∂

value (both being second order approximations), the FV residual equation for the
i-th CV can be written as
∂ h̃i ∂ p̃i

∆z̃i (q̃θ |e − q̃θ |w ) + ∆θ̃i (q̃z |n − q̃z |s ) + 2Sτ ∆z̃i ∆θ̃i p̃i + h̃i − M̃inj,i = rFV,i,
∂τ ∂τ
(P4.14)
where ∆θ̃i , ∆z̃i are the CV dimensions, p̃i , h̃i are the CV centre pressure and
film height and q̃θ , q̃z are the flow components evaluated at the midpoint of the
respective CV faces as indicated in fig. P4.2(a). For now, the integrated value of
the injection flux over the cell area is denoted simply as M̃inj,i since its treatment
depends on the applied injection model.
P4.4.3 Cell Face Fluxes

To evaluate eq. (P4.14), the components of the flow must be known at the CV
faces. These are given from eq. (P4.12) and hence depend on the face pressure,
face pressure gradient and face film height. The gradient at the midpoint of
each face is reconstructed using linear interpolation between the centres of the
two CVs sharing the face following a Central Differencing Scheme (CDS). This
p = pa
NN
N
n
w e
zr WW W CV i E EE
s
S
SS
θ̃ l q·n =0 θ̃ t θ̃
(a)
−2Sτ A ∂∂τ
∫ p̃h̃
dA
∫
q̃ n dS
Sn z z i
+ A m̃inj dA
i
∫
i
∆θ̃i
∫
Siw
q̃θ nθ dS
CV i
∫
∆z̃i Sie
q̃θ nθ dS
z̃
∫
q̃z nz dS
θ̃ Sis
(b)
Figure P4.2: The fluid film discretization. (a): Mesh of a single pad showing
the BCs and the naming convention for the neighbours and faces of the i-th CV.
(b): Contributions to the integrated conservation equation for the i-th CV.
approximation is of second order on a uniform mesh, but formally decays to first

order for a non-uniform mesh. For reasonable grid expansion rates it will, however,
still produce convergence characteristics close to that of a second order scheme [46].
Referring to fig. P4.2(a), the gradient at the eastern face is thus reconstructed
using the CV centre values i and E. If formal second order accuracy was requested,
one could apply the ”Quadratic Upstream Interpolation for Convective Kinematics”
scheme or a fourth-order CDS using three or four cell centres, respectively.
The face pressure values could likewise be reconstructed using CDS to provide
a second order accurate approximation. It is, however, argued by Arghir et al. [6]
that these ”convected pressures” should be treated using an upwind procedure
to stabilize the numerical solution. The approach presented in [6] is intended
for unstructured grids where the implementation of higher order upwind schemes
is complicated, thus the Upwind Differencing Scheme (UDS) relying on a single
upstream cell is used. The current work is limited to structured grids where it is
straightforward to include an additional upwind cell in the stencil to produce the
Linear Upwind Differencing Scheme (LUDS). This provides second order accuracy
for the face pressure, but has the potential to produce unbounded solutions when
used without limiters.
Considering the structure of the resulting equation system, the combination
of CDS for face pressure gradients and UDS for face pressure values gives a
computational molecule for each cell with five contributions: the cell itself and its
four immediate neighbours E, N, W and S. Using LUDS for the face pressures,
four additional contributions from the next set of neighbours E E, N N, WW and
SS should furthermore be included.
Prior to applying any of the upwind schemes, the flow direction at the given
face must be evaluated. Noting that the pressure is always positive, the sign of
the flow vector components can be judged from eq. (P4.12) as
3 ∂ p̃ 3 ∂ p̃

sign (q̃θ ) = sign − h̃ + Sτ Ω̃ h̃ and sign (q̃z ) = sign − h̃ , (P4.15)
∂ θ̃ ∂z̃
where the CDS approximations for the pressure gradients are used while the film
height at the face is supplied by the rotor and foil models. Knowing the sign of the
flow, the face pressure itself can be reconstructed using cell centres in the upwind
direction only.
At the outer boundaries where ambient pressure is prescribed (see fig. P4.2(a)),
corresponding to a Dirichlet type Boundary Condition (BC), only the face gradient
needs reconstruction. This is achieved using a one-sided FD scheme relying on
two CV centres into the mesh. If a symmetry condition is imposed at the bearing
mid plane, corresponding to a Neumann type BC, the flux at these faces should
simply be set to zero. Both types of BCs are used if the LUDS is requested to
reconstruct face pressures too close to a boundary for two upstream CV centres to
be available.
Lastly, the model implements a cyclic condition which is physically meaningful
for rigid and single-pad foil bearings. In this case, the cyclic boundary face fluxes
are reconstructed exactly as for the internal faces using the corresponding CVs
from the opposite end of the mesh.
P4.4.4 Finite Volume Residual Equation

Using the FD schemes described in the preceding section to evaluate the fluxes
across each CV face and plugging these into eq. (P4.14), one obtains a residual
equation on the form

a j p̃ j + 2Sτ ∆z̃i ∆θ̃i p̃i hÛ̃ i + h̃i pÛ̃i − M̃inj,i = rFV,i, (P4.16)
Õ
ξ∈Ξ
where Ξ is the set of stencil members making up the computational molecule, e.g.
Ξ = {i, E, N, W, S} for the case of a CDS+UDS combination. The stencil coefficients
a j depend on the present pressure values across the stencil, the current film heights
at the cell faces and the dimensions of the cell. In a steady-state simulation, the
cell centre temporal pressure derivative pÛ̃i is zero, while it is directly available in
a time integration. The film heights at the cell centre and faces as well as the
temporal derivative at the cell centre hÛ̃ i are given from the rotor and foil domains.
For a bearing α discretized using ncv CVs, the pressure state vector holding all
CV centre pressures can be written
(P4.17)
T
pα = p̃1, . . . , p̃ncv ,

using which the FV residual vector can be assembled for the entire bearing as
AFV pα, xr , x f ,α pα + BFV xÛ r , xÛ f ,α pα + CFV xr , x f ,α pÛ α + fFV = rFV,α,

| {z } | {z } | {z } |{z}
2Sτ ∆z̃i ∆θ̃ i hÛ̃ i p̃i 2Sτ ∆z̃i ∆θ̃ i h̃i pÛ̃i
Í
ξ ∈Ξ a j p̃ j M̃inj,i
(P4.18)
where AFV holds the FV coefficients originating from the Couette and Poiseuille
terms of the MRE, while the coefficients in BFV and CFV represent the squeeze
and local expansion terms respectively. The injection contribution is represented
by fFV which will generally depend on the bearing pressure and possibly variables
from the rotor and fluid domains.
P4.5 Domain Interfaces

To evaluate the FV residuals from eq. (P4.16), the film height h̃ is required at all
CV edges and centres. This contains three contributions: (a) a constant describing
the initial geometry such as clearance and pad inlet slopes; (b) a rigid contribution
stemming from the movement of the rotor; and (c) the compliant contribution due
to deformation of the foil structure (if simulating a compliant bearing).
Recalling the rigid contributions to be given merely as trigonometric projections
of the eccentricity components, written usually like e x cos θ + e y sin θ, this can
be represented as a constant linear mapping applied to the rotor DOF vector x̃r .
Collecting the rigid film heights from all CV centres and edges across all bearings
in the vectors hr,cvc and hr,cve respectively, these can be calculated as
hr,cvc = Hcvc,r x̃r and hr,cve = Hcve,r x̃r , (P4.19)
where the mapping matrices can be assembled offline knowing: (a) the angular
positions of the CV centres/edges; (b) the indices of the rotor DOFs representing
the eccentricity components in each bearing; and (c) the length scaling between
the two domains. The row dimensions correspond to the combined number of CV
centres/edges across all bearings while the column dimensions equal the number
of rotor DOFs nrdof , but only 2nb columns will contain non-zero elements.
Analogously to the rigid contributions in eq. (P4.19), the compliant film heights
at all CV centres and edges can be represented as constant linear mappings from
the foil DOF vectors x̃ f ,α . This can be written as the linear combinations
n oT n oT
hc,cvc = Hcvc, f x̃Tf ,1 . . . x̃Tf ,nb and hc,cve = Hcve, f x̃Tf ,1 . . . x̃Tf ,nb
(P4.20)
but the content of the mapping matrices will be dependent on the applied foil model.
The currently implemented SEFM places one foil node for each circumferential
CV centre position. In this case, Hcvc, f maps each foil DOF directly to all CV
centres sharing the same circumferential position and applies a length scaling.
The mapping to CV edges Hcve, f , on the contrary, will furthermore contain FD
coefficients interpolating from the foil DOF locations to the CV edges. Having
prebuild the mapping matrices and the constant contribution vectors h0,cvc and
h0,cve , the film heights can hence be found at runtime simply as
n oT
hcvc =h0,cvc + Hcvc,r x̃r + Hcvc, f x̃Tf ,1 . . . x̃Tf ,nb
n oT (P4.21)
hcve =h0,cve + Hcve,r x̃r + Hcve, f x̃ f ,1 . . . x̃ f ,nb .
T T
To evaluate the rotor residual from eq. (P4.1), the bearing forces must be
obtained by integration of the pressure as given in eq. (P4.2). For a single bearing,
P4.5 Domain Interfaces 123
this is evaluated numerically by summing the projected force contributions from

all CVs as
ncv
f˜b,x pa R2 Õ cos θ cvc,i

= ∆θ̃i ∆z̃i ( p̃i − 1) , (P4.22)
f˜b,y lr mr ωτ2 i sin θ cvc,i
where p̃i and θ cvc,i are the the i-th CV centre pressure and circumferential position
2
respectively, while l pma Rω2 is the inter-domain force scaling. Notice that eq. (P4.22)
r r τ
effectively applies the midpoint rule within each CV, but a higher order numerical
integration scheme could likewise have been applied. This would require the
trigonometric functions to be sampled and the pressure value to be reconstruc-
ted using FD at a number of locations within each CV, thus making the force
contribution from each CV dependent also on the neighbouring cells. Regardless
of the integration scheme applied within each CV, the scaling, integration and
summation can be represented using a constant linear mapping
T
Tr f̃b = H fb p pT
1 − 1, . . . , pnb − 1
T
, (P4.23)

calculating the state-space sized bearing force vector as a single matrix–vector

product. H fb p has row dimension equal to the rotor state-space, i.e. 2nrdof , and
column dimension matching the combined number of CVs across all bearings, i.e.
nb ncv . Generally, each bearing will give rise to two bearing force components,
meaning that H fb p will contain 2nb ncv non-zero elements. For GFBs it is common
to apply the Gümbel condition such that sub-ambient pressures are discarded
during the integration. In this case, H fb p becomes non-constant as the columns
corresponding to CVs with sub-ambient pressures should be zeroed. In practice,
this is achieved by prebuilding the full non-Gümbel H fb p and continuously restoring
a copy of this before zeroing out the sub-ambient columns.
Lastly, the pressure forces on the foil structure should be integrated to calculate
f̃ p,α in eq. (P4.4) or eq. (P4.5). For each bearing α, this can be achieved analogously
to the bearing force integration in eq. (P4.23) using a constant linear mapping as
T f f̃ p,α = H fp p pα, (P4.24)
where the exact content of H fp p depends on the foil model. For the currently
implemented SEFM, the mapping will simply calculate the average pressure at
each circumferential position multiplied by the surface area ascribed to each foil
2
DOF and the inter-domain force scaling ppa Rl 2 . The row dimension should match
f f
the number of foil states while the column dimension is ncv , but the number of
non-zero elements depends on the foil model.
The mapping matrices introduced in this section usually have sparse structures
and are used mostly for calculating matrix–vector products. They are hence
build and stored in Compressed Sparse Row (CSR) format to lower the storage
requirements and provide efficient matrix–vector products.
P4.6 System Assembly

For a rotor–bearing system comprising nb GFBs, the coupled state-space and
residual vectors are defined as
n oT
z = pT · · · pT zT · · · zT x̃ Û̃ rT
T x
1 n b f ,1 f ,nb r
n oT (P4.25)
r = rFV,1 · · · rFV,nb r f ,1 · · · r f ,nb rr
T T T T T
using which the global non-linear system of equations can be written on fully
implicit form like
fG (τ, z, zÛ ) = fG,t (τ, z, zÛ ) + fG,s (z) + fG,0 = r, (P4.26)
with a global residual function fG (τ, z, zÛ ) which can be partitioned into its transient,
steady and constant components fG,t (τ, z, zÛ ), fG,s (z) and fG,s (z) respectively. The
main advantage of this partitioning is to clearly separate out all time dependent
terms such that fG,t (τ, z, zÛ ) = 0 at steady-state, thus allowing the exact same
implementation of fG,s (z) and fG,0 to be used for steady-state solutions as well as
in time integrations.
It is common to treat IVPs for ODE systems on explicit first order form where
a system function evaluates the temporal derivative of the state vector directly.
In the current model, the fluid film FV equations could have been formulated
explicitly in terms of the alternative variable ψ = p̃ h̃, if this had been substituted
into eq. (P4.11) as demonstrated by Bonello and Pham [19]. For many cases,
explicit forms of the rotor and foil equations could likewise have been obtained
by inverting Ar and A f in eqs. (P4.1), (P4.4) and (P4.5). In the GFB models by
Larsen and Santos [90] and Gu et al. [52], the global systems are formulated on
linearly implicit form where the product of a ”mass” matrix and the state vector
derivative is provided by the system function. In [90], this matrix is constant
such that the system can be solved on explicit form by evaluating and inverting
it, while a dedicated solver for the linearly implicit form is used in [52] where
the ”mass” matrix is state dependent. In the present work, the fully implicit
form is maintained for several reasons. It allows the implementation to span the
largest possible variety of foil structure models, including those resulting in a
DAE system. Any explicit matrix inversions can be avoided and the use of only
physically meaningful variables simplifies the implementation of BCs as well as the
interpretation of errors and intermediate results. Furthermore, it is an advantage
when it comes to friction models that the variable zÛ is available when evaluating
the residual. The disadvantages of using the fully implicit form are mainly the
added complexity of defining consistent Initial Conditions (ICs) and a smaller
selection of available time integrators.
P4.6 System Assembly 125
Assembling the contributions from the three domains, the global equation
system can be written on the fully implicit format of eq. (P4.26) as
Û̃ r , xÛ̃ f ,1 p1 + CFV x̃r , x̃ f ,1 pÛ 1 

 BFV x
.
 
.
 
.

 


 

, + ,
 
Û̃ Û̃

B x x p C x̃ x̃ Û
p
 
FV r f ,nb nb FV r f ,nb nb 

 

 

 
A f zÛ f ,1 + T f f̃ µ p1, z f ,1 +

 

 .. 
.

 


 

+ ,
 



 A f Û
z f ,n b T f f̃ µ p n b z f ,n b





 

+
 

 A Û
z
r r T f̃
r ub (τ) 

| {z }
fG,t (τ, z, zÛ )
(P4.27)
AFV p1, x̃r , x̃ f ,1 p1

  fFV,1 

.. .
   
.
   
. .

 
 
 


 
 
 

AFV pnb , x̃r , x̃ f ,nb pnb
   
f
   
FV,n
   

 
 
 b  

 
 
 

B f z f ,1 + T f f̃ p (p1 ) + T f f̃ f w,1 = r,

 
 
 

 ..   .. 
. .

 
 
 


 
 
 

+
   



 B f z f ,n b T f f̃ p p n b







 T f f̃ f w,n


b


 
 
 

 B + Ω̃C z + T f̃ p , . . . , p   T f̃
   

 r r r r b 1 nb   r rw 
| {z } | {z }
fG,s (z) fG,0
where the upper rows represent the non-linear FV equations governing the fluid
film while the midmost rows can either be left out (to model a rigid bearing) or
represent a linear/non-linear foil structure. The lowermost rows hold the rotor
model. Notice that the injection contributions, if any, are currently located in
the constant part fG,0 , while most injection models would be state dependent and
hence should be relocated to the non-constant parts.
Depending on the system size, ranging potentially from a few hundred to tens
of thousands of states, and the system stiffness, it is advantageous or downright
necessary to formulate analytical expressions for the Jacobian matrices. This is
the case for steady-state solutions, but particularly when integrating in time. For
the steady-state case where fG,t (τ, z, zÛ ) = 0, the Jacobian with respect to the state
vector z is given as
 ∂AFV p1 · · · 0 ∂AFV p1
··· 0  ∂AFV p1
0
 ∂p1 ∂z f ,1  ∂x̃r
 .
.. . .. .. .. ... ....  ..
 . . . .  .
∂AFV pnb ∂AFV pnb ∂AFV pnb
 
 
∂fG,s  0 ··· ∂pnb 0 ··· ∂z f ,nb ∂x̃r 0 
=
Jz,s = H .

∂z fp p · · · 0 Bf ··· 0
. . .. .. .. ..
 
.. .. . . . .
 

 0 


 0 ··· H fp p 0 ··· Bf 

 H fb p 0 Br + Ω̃Cr 
(P4.28)

The upper left and midmost blocks reflect the changes in FV residuals within each
bearing as the pressure and foil deflection (governing the compliant film height)
vary within that same bearing. The upper right block gives the FV residual
changes due to the rotor position (governing the rigid film height). These blocks
are all state-dependent, i.e. non-constant. The blocks in the midmost rows are
constant and represent the changes in foil residuals within each bearing due to
changes in pressure and foil states within that same bearing. If a friction model
was included, this would likely cause additional non-constant contribution to the
left and midmost of these blocks. Notice that the foil residuals are not directly
affected by the rotor states, which is physically meaningful since the rotor and foil
domains are only coupled through the fluid film. The lower left block governs the
change in rotor residuals due to pressure changes across all bearings. As discussed
in relation to eq. (P4.23), this block is pressure dependent if the Gümbel condition
is applied, but constant if the full bearing surface is included in the pressure
integration. The lower right block is constant and contains the rotor stiffness,
damping and gyroscopic contributions governing the change in rotor residuals due
to changes in rotor states. Again, the nature of the foil–rotor coupling is evident,
as the rotor residuals are not directly affected by the foil states.
For the transient case, the Jacobian with respect to both the state vector and
to its temporal derivative are necessary. The Jacobian with respect to z is now the
sum of Jz,s = ∂fG,s /∂z from eq. (P4.28) and an additional contribution ∂fG,t /∂z
given as
 BFV · · · ∂CFV pÛ 1 ∂CFV pÛ 1

0 ∂z f ,1 ··· 0 ∂x̃r 0 
 . .. .. .. .. ..
 
 . .. ..
. .

∂fG,t  . . . . . .
= 0 ∂CFV pÛ nb ∂CFV pÛ nb
0 , (P4.29)

· · · BFV 0 ···

∂z  ∂z f ,nb ∂x̃r 

 0 0 0 


 0 0 0 

P4.7 Steady-state and Transient Solution 127
which represents the additional pressure and film height dependencies originating
from the local expansion and squeeze terms of the MRE. If a friction model
was included, this would possibly give rise to additional contributions in this
matrix. Since ∂fG,s /∂ zÛ = 0 by definition, the Jacobian with respect to the temporal
derivative of the state vector zÛ is given as
 CFV · · · ∂BFV p1 ∂BFV p1

 0 ∂ zÛ f ,1 ··· 0 ∂ xÛ̃ r
0 

 . .. .. .. .. .. .. ..
 .. . .

 . . . . .

 ∂BFV pnb ∂BFV pnb 
∂fG,t  0 · · · CFV 0 ··· ∂ zÛ f ,nb ∂ xÛ̃ r
0 
JzÛ = = , (P4.30)

∂ zÛ Af ··· 0
.. ... ..
 
. .
 

 0 0 


 0 ··· Af 


 0 0 Ar 

where the FV residual derivatives in the uppermost blocks are, once again, stem-
ming from the MRE’s local expansion and squeeze terms. A f represents either foil
inertia or damping depending on whether eq. (P4.4) or eq. (P4.5) is employed and
Ar holds the rotor inertia. Notice that the content of eq. (P4.30) is decisive to the
overall system structure. When JzÛ has full rank, the model is a system of ODEs,
while a singular JzÛ implies a DAE system.
P4.7 Steady-state and Transient Solution

The entire model is implemented in C where eq. (P4.27) is solved in steady-state
using the general purpose non-linear algebraic solver ”KINSOL” and integrated
in time using the ODE/DAE IVP solver ”IDA”. Both are readily available and
part of the ”SUNDIALS: SUite of Nonlinear and DIfferential/ALgebraic Equation
Solvers” developed by the Lawrence Livermore National Laboratory [67]. In both
cases, a Modified Newton iteration is employed to solve the non-linear equation
systems using the Jacobian matrices given in eqs. (P4.28) to (P4.30). Like the
mapping matrices, the Jacobians are constructed in the CSR format and the linear
equation systems emerging from the Newton iterations are solved using sparse LU
factorization provided by the KLU sparse solver library [34].
P4.7.1 Ill-conditioning and Tolerances

Equation (P4.27) presents a multi-domain problem, hence the numerical quantities
involved will have varying orders of magnitude. To some extent, this can be
alleviated using the scaling parameters within each domains, but the specification
of meaningful solver tolerances still requires each equation to be assessed separately.

Rotordynamics codes are often organized in a segregated fashion using an inner
routine to solve for the pressure distribution (and related fields, such as temperat-
ure) while an independent outer routine searches for the rotor static equilibrium
or steps the dynamic rotor equations forward in time. In these constructions,
separate equation systems are solved for each domain and the solver tolerances
should hence also be set separately. This can usually be achieved using scalar
values as the variables/residuals within each domain are homogeneous. In the
present model, all equations are solved simultaneously by a single solver, meaning
that a vector of absolute tolerances must be supplied.
This is achieved by supplying global absolute force and length tolerances ftol [N]
and ltol [m] which are then used to construct equivalent tolerances for each state
and residual quantity involved. As an example, the fluid film state variables
physically represent pressure with unit Pa. Assuming the absolute tolerances to
be specified in the vicinity of 1 N and 1 m respectively, the corresponding relative
tolerances can be added and multiplied by the derived base value for the relevant
quantity as

1N ftol 2ltol
ptol = + , (P4.31)
(1 m)2 1 N 1 m
which for realistic values ftol = 10−3 N and ltol = 10−8 m is completely dominated
by the force tolerance and gives ptol = 1.000 02×10−3 Pa. One could argue that
other base values than 1 N and 1 m should be used, but for now this is utilized as a
systematic way of producing tolerances across the equation system. This approach
is beneficial since the tolerances on some quantities, such as the FV residual given
in J Pa s/(m2 s) = N2 /m3 , are challenging to define based on physical intuition.
Collecting the system residual scaling factors implied by the non-dimensionali-
zations in the vector rscale and the corresponding absolute tolerances derived as
shown above in the vector rtol , the convergence criterion used for steady-state
solutions can be written as
||diag (rscale ) diag (rtol )−1 r||∞ < 1, (P4.32)
where || ||∞ denotes the infinity norm. When integrating in time using the IDA
solver, the tolerances are set on the local truncation error as estimated by the
integration routine within each time step [66]. This is done in a way similar to the
well-known Matlab integrators using a vector of absolute tolerances and a scalar
relative tolerance. The first is here calculated as diag (rscale )−1 rtol consistently to
eq. (P4.32), while the latter is set to around 10−6 .
P4.8 Results 129
P4.8 Results
The current implementation has been validated against two previously presented
codes treating rigid and compliant gas bearings respectively.
In the case of a rigid bearing, the midmost rows representing the foil structure
in eqs. (P4.25) and (P4.27), along with the corresponding Jacobian rows and
columns in eqs. (P4.28) to (P4.30), are simply left out. The same is the case
for the compliant film height contribution in eq. (P4.21). In this configuration,
the code has been used to simulate the rotor–bearing system presented in [142].
This comprises a flexible rotor weighing approximately 4 kg modelled using FE
which is supported by a ball bearing and a single rigid gas bearing. Placing an
unbalance of 10 g mm at the disc, this system has been integrated in time to obtain
the forced steady-state orbit at 5 kRPM. The same has been simulated using the
code presented in [117] utilising an FD discretization of the fluid film, a reduced
order rotor and an explicit time integration scheme. The resulting orbits are very
similar, as it can be seen from the comparison in fig. P4.3(a).
To validate the full model including a foil structure, the test rig presented in
[90] has been simulated. This is composed of a rigid shaft with a mass of 21.1 kg
supported by two identical three-pad GFBs. The SEFM is employed to model the
foil structure under the assumption of an axially uniform displacement field. For
the present code, this is implemented by locating foil nodes at each circumferential
CV position and to incorporate the calculation of the average pressure at each of
these locations in the pressure–foil mapping matrix from eq. (P4.24).
Simulating in time at 15 kRPM with unbalances of 40 g mm and −2.5 g mm at
the bearing locations, the forced steady-state orbits measured in the two bearings
are obtained as shown in fig. P4.3(b). These are compared to results from the
code presented in [90], where a FE discretized fluid film is used, and the orbits are
seen to agree well for both bearings.
P4.9 Conclusions & Future Aspects

In the paper, a recipe for a simultaneously formulated model for simulation of
rigid as well as compliant type gas bearings supporting a wide variety of rotor
configurations has been treated. This has included the mathematical formulation
as well as selected aspects of the numerical implementation.
Two main features of the presented model are the full coupling between the
domains and the residual form of the system. For steady-state solutions, the
presented formulation allows a single general purpose non-linear algebraic solver
to be used while given full system knowledge through the Jacobian. This is not
possible in the more common segregated solution approaches involving (at least)
P4 Modelling of Compliant-type Gas Bearings: A Numerical Recipe
Current code A
0.48 Current code B
0.7
Benchmark A
0.50 Benchmark B
0.52 0.8
εx [−]
εx [−]
0.54
0.9
0.56
0.58 1.0
0.60 Current code
Benchmark 1.1
0.62
0.500 0.525 0.550 0.575 0.600 0.625 0.0 0.1 0.2 0.3 0.4
εy [−] εy [−]
(a) (b)
Figure P4.3: Time integrated unbalance driven steady-state orbits compared to benchmark codes: (a): Flexible
rotor supported by a rigid gas bearing as described in [142] simulated as given in [117] at 5 kRPM with a 10 g mm
unbalance at the disc. (b): Rigid rotor supported by two identical GFBs (”A” and ”B”) using the geometry and
code described in [90] at 15 kRPM with 40 g mm and −2.5 g mm unbalances at the bearings.
130
P4.9 Conclusions & Future Aspects 131
two independent solvers. For time integrations, the same formulation allows all
state variables to be solved simultaneously in time using a general purpose IVP
solver. The residual, or fully implicit, form of the equation system has numerical
advantages as no explicit matrix inversions are necessary, but it is mainly chosen
to allow a wider variety of foil models. Unfortunately, a limited number of IVP
solvers are available for this form, and calculation of consistent ICs is required,
something which is not generally necessary for explicitly formulated systems.
In the future, several extensions to the model should be made. One is the
inclusion of reduced order rotors, e.g. through modal truncation, which will
primarily require the left eigenvectors or similar to be included in the Tr matrix
of eq. (P4.1). The resulting reduction in overall system size would usually be
limited, as even the full rotor would most often require much fewer states than the
fluid film, but removing the high frequency modes of a full FE rotor model would
allow the IVP solver to take much larger time steps. Another interesting extension
would be to include a bump foil model with frictional dissipation as described in
[125], possibly coupled to a top foil as described in [121].
The original and main motivation for developing the presented model has
been to prepare an efficient platform for research into GFBs with active radial
injection. For rigid bearings, several analytical expressions are available in the
literature, but the additional compliance–injection interaction in GFBs complicates
the identification of the empirical correction factors, or ”discharge coefficients”,
most often involved. This could possibly be circumvented by simultaneously
solving a miniature CFD model for the injection zones, but this has not yet been
proven feasible.
References
PUBLICATION P5
Multi-domain Stability and
Modal Analysis Applied to Gas
Foil Bearings: Three
Approaches
Journal of Sound and Vibration.

Abstract
The dynamic characteristics of rotors supported by Gas Foil Bearings (GFBs) are
commonly assessed using linear force coefficients, which are usually derived from
an analytical perturbation in combination with an implicit compliance model. This
remains common practice even though discrepancies have been reported between
the Onset Speed of Instability (OSI) predicted using such linear coefficients and
observations from non-linear time integration. For the first time, the present
paper pinpoints the root cause of this discrepancy by demonstrating an extended
perturbation, akin to that used for tilting pad journal bearings, to predict exactly
the same OSIs as obtained from an eigenvalue analysis based on the Jacobian of the
full non-linear system of equations. To demonstrate this, the OSIs predicted using
(i) the classical perturbation, (ii) the extended perturbation, and (iii) the Jacobian
eigenvalues are compared over a range of compliance levels covering essentially
rigid bearings to modern heavily loaded industrial GFBs. Using carefully aligned
implementations, the deficiency of the classical method is demonstrated to increase
with the compliance level (up to 13 % for the benchmark case), while the extended
134 P5 Multi-domain Stability and Modal Analysis Applied to Gas Foil...
perturbation provides OSIs in agreement with the Jacobian eigenvalues for all
compliance levels. Furthermore, the mode shapes attainable from the three
approaches, encompassing only the rotor, the rotor and foil, and the rotor, foil
and pressure respectively, are compared.
DOF Degree of Freedom k0 Nominal bump stiffness per

FE Finite Element unit area
GFB Gas Foil Bearing k Bump stiffness per unit area
NR Newton–Raphson kc Dynamic foil stiffness
OSI Onset Speed of Instability L, L̃ Bearing length, L̃ = LR−1
RE Reynolds Equation l0 Bump half length
RPM Revolutions per Minute m0f Average foil mass per
SEFM Simple Elastic Foundation circumferential length
Model mr Rotor mass
(Û) Time derivative, d/dτ nθ Number of circumferential
(Ü) Time derivative, d 2 /dτ 2 nodes
(˜) Non-dimensional quantity nz Number of axial nodes
() e Quantity at element level Pc (λ) Characteristic polynomial
A In-film element area p, p̃ Film pressure, p̃ = p−1
a p
C Radial clearance p0 , p̃0 Static pressure, p̃0 = p−1
a p0
e x,y , ε x,y Journal eccentricity
pγ , p̃γ Dynamic pressure, rotor DOF
components, ε = eC −1
γ, p̃γ = p−1a pγ
Eb Young’s modulus of bump foil
pw j , p̃w j Dynamic pressure, foil DOF j,
material
p̃w j = p−1
a pw j
fα Functions fx,y and fw j = W j (θ)
pa Ambient pressure, fluid
fx,y Functions fx = cos θ and pressure scaling
fy = sin θ
p̃m Mean axial pressure
fhc (θ, p) Axially uniform compliance
R Journal radius
function
Rb Bump radius of curvature
g Gravitational acceleration
Sτ Non-dimensional group,
h, h̃ Film height, h̃ = hC −1 Sτ = 6R2 µωτ C −2 p−1
a
h0 , h̃0 Static film height, h̃0 = h0C −1 Sb Bump foil pitch
hb Bump foil height t Physical time
hc, h̃c Film height (compliant), tb Thickness of bump foil
h̃c = hc C −1 w j , w̃ j j-th discrete foil deformation,
hr , h̃r Film height (rigid), h̃r = hr C −1 w̃ j = w j C −1
Wj j-th foil DOF film height fG Coupled system residual

weight function function
Wx Static vertical load fG,t Unsteady part of fG
X Oscillation amplitude fG,s Steady non-constant part of fG
z, z̃ Axial gas film coordinate, fG,0 Constant part of fG
z̃ = zR−1 f p , f̃ p Foil structure pressure loads,
zj j-th axial nodal coordinate f̃ p = p−1 −2
a R fp
∆θ j , ∆z̃ j Span associated with j-th grid frw , f̃rw Rotor static force vector,
location f̃rw = p−1 −2
a R frw
δ Logarithmic decrement
hr , h̃r Rigid film heights at FE nodes
ε x , ε y Eccentricity ratios
hc , h̃c Compliant film heights at FE
η Hysteretic foil damping
nodes
constant
p, p̃ FE nodal pressures, p̃ = p−1
a p
Λγ Rotor perturbation,
∆εγ eiω̃τ τ, γ = x, y r̃ Coupled system residual
r̃ f Foil structure residual
Λj Foil perturbation,
r̃ pγ FE source term, 1st order RE
∆w j eiω̃τ τ, j = 1, . . . , nθ
r̃RE FE residual
λ Eigenvalue
r̃r Rotor residual vector
λ̂ Pseudo eigenvalue
vi i-th eigenvector
µ Dynamic gas viscosity
w, w̃ Foil structure DOF vector,
νb Poisson’s ratio of bump foil
w̃ = C −1 w
material
w0 , w̃0 Foil structure static
Ω, Ω̃ Angular velocity Ω̃ = ωτ−1 Ω
equilibrium vector
ΩOSI Angular velocity at the OSI
x Generic DOF vector
ω f , ω̃ f Foil oscillation frequency,
x f , x̃ f Foil structure DOF vector,
ω̃ f = ωτ−1 ω f
x̃ f = C −1 x f
ωs , ω̃s Perturbation/whirl frequency,
xr , x̃r Rotor DOF vector, x̃r = C −1 xr
ω̃s = ωτ−1 ωs
xr0 , x̃r0 Rotor static equilibrium
ωτ Characteristic frequency
z Coupled system state vector
ρb Density of bump foil material
z0 Static system equilibrium
τ Dimensionless time, τ = ωt t
z f , z0f Foil structure state vector
θ Circumferential angle
θb Bump half angle zr Rotor state vector
θi i-th circumferential nodal Af Foil structure matrix
coordinate A0f Massless foil structure matrix
θl Pad leading edge angle AG Global linearised system
θt Pad trailing edge angle matrix
ζ Damping ratio A pγ FE coefficient matrix, 1st order
fb , f̃b Bearing force vector, RE
f̃b = p−1
a R fb
−2 Aλ Lambda matrix
ARE RE FE matrix, Couette & H fb p Bearing pressure to force

Poiseuille terms mapping
Ar Rotor state-space inertia I Identity matrix
B Shape function derivatives Jz,s Jacobian of fG,s with respect to
matrix z
Bf Foil structure
Jz,t Jacobian of fG,t with respect to
stiffness/damping
z
Bf0 Massless foil structure
JzÛ Jacobian of fG,t with respect to
stiffness/damping
zÛ
BRE RE FE matrix, squeeze term
Br Rotor state-space K (ωs,Ω) Generic frequency dependent
stiffness/damping stiffness
D (ωs,Ω) Generic frequency dependent K̃g,2DOF Gas stiffness, two-DOF
damping perturbation
D f , D̃ f Foil structure damping, K̃g,xDOF Gas stiffness, extended
D̃ f = ωτ Cp−1 −2
a R Df perturbation
D̃g,2DOF Gas damping, two-DOF K f , K̃ f Foil structure stiffness,
perturbation K̃ f = Cp−1 −2
a R Kf
D̃g,xDOF Gas damping, extended M Generic mass matrix
perturbation
M f , M̃ f Foil structure mass,
Dr , D̃r Rotor damping,
M̃ f = ωτ2Cp−1 −2
a R Mf
D̃r = ωτ Cp−1 −2
a R Dr
CRE RE FE matrix, local expansion Mr , M̃r Rotor mass,
term M̃r = ωτ2Cp−1 −2
a R Mr
Hr Mapping from x̃r to h̃r N Shape function matrix
H hc p Pressure to compliant film Z Extended perturbation
height mapping complex stiffness
P5.1 Introduction
The Gas Foil Bearing (GFB) offers a mechanically simple, self-acting and oil-free
means for support of lightweight high-speed rotating machinery, making it a perfect
and environmentally friendly fit for a wide range of green technologies. As GFBs
provide only limited damping, reliable tools for prediction of stability limits and
other rotor–bearing system characteristics remain essential to their application.
Fundamentally, the stability limit can be assessed using non-linear or linear analysis.
In the former case, the full non-linear system of equations is formulated and solved
either in a decoupled or simultaneous fashion. For a decoupled model where the
physical domains are treated separately, the stability limit can be assessed using
brute-force by integrating in time with steps sufficiently small to compensate for

the decoupling. For a simultaneous formulation, stability assessment using time
integration is likewise possible [88, 128], but it is more efficient in this case to base
the analysis on the Jacobian eigenvalues [19]. Having obtained these eigenvalues, it
has recently been established that Campbell diagrams can furthermore be extracted
[15, 16]. In a linear analysis, the gas film forces are represented by equivalent
force coefficients dependent on the operational conditions and excitation frequency.
Such force coefficients are found by perturbing the rotor from its static equilibrium
either numerically or analytically as proposed for rigid gas bearings by Lund [110]
in 1968. The analytical approach was extended to compliant bearings by Peng
and Carpino [132], who incorporated the compliance implicitly through the film
height function to effectively eliminate the foil equations of motion—an approach
that has since been applied extensively [68, 74, 79–81, 87, 93, 101, 132, 175].
Previous studies [69, 74, 88, 128, 146] have investigated the accuracy of
the Onset Speed of Stability (OSI) predicted from the perturbation method for
compliant type bearings and found varying levels of discrepancy. The main
contribution of the present work is, for the first time, to pinpoint the root cause of
this discrepancy by demonstrating an extended version of the perturbation method
to predict OSIs in agreement with results from an eigenvalue analysis based on
the full non-linear system Jacobian.
Similar comparative investigations have previously been published by the
authors [88, 128] indicating that the discrepancy of the perturbation OSI increases
with the compliance level. In previous works, the coupled system OSI was assessed
by time integration, whereas in the present work it is evaluated through system
Jacobian eigenvalues. Having expressions for the Jacobian matrices available, this
is far more efficient, but it does not change the predicted OSIs. More importantly,
three issues related to the consistency between the perturbation and time domain
models have been identified and addressed. The first and most critical change is
for the assumption of rotor-synchronous viscous foil damping in the time domain
model to be replaced by an iterative scheme ensuring that the foil oscillation
frequency matches the frequency of the unstable system mode. This means
that the equivalent time domain foil damping at the OSI now matches the foil
damping applied in the first order Reynolds Equation (RE). This has previously
been found to be negligible for a very stiff bearing [19], but has recently been
further investigated for higher compliance levels by Pronobis and Liebich [146]
proving this inconsistency to account for a significant portion of the reported
OSI discrepancies. In [146] the rotor-synchronous foil damping is maintained,
while the frequency domain damping expression is consistently adjusted leading
to an unprecedented level of agreement between the predicted OSIs (though still
deteriorating with compliance level). The strategy for obtaining consistency in
[146] is hence contrary to the one adopted in the present work where the damping
in both models is based on the instantaneous oscillation frequency. . The second

model change is related to the enforcement of the Gümbel condition in the first
order perturbation equations. This is now performed during the dynamic pressure
integration rather than through constraints on the dynamic pressure solution
(which would be adequate for a Reynolds-type boundary condition). Thirdly, the
compliant film height due to the dynamic pressure in the first order perturbation
equations is now calculated from the axially averaged dynamic pressure, exactly
as for the static film height in the zeroth order perturbation equation and for the
film height in the coupled model.
With the above model changes in place, the present paper presents OSIs
calculated for a benchmark rotor–GFB system for various compliance levels in
three different ways:
1. Using force coefficients obtained through the classical perturbation with an

implicit film height description to provide force coefficients related to the
rotor Degrees of Freedom (DOFs) only.
2. Using force coefficients obtained through an extended perturbation incor-

porating the foil compliance explicitly through a number of dedicated foil
DOFs akin to the pad DOFs (pitch, radial displacement and tangential
displacement) used in analysis of tilting pad journal bearings.
3. Using the eigenvalues of the Jacobian matrix of the full non-linear system of
equations.
Lastly, a secondary contribution of the presented work is the calculation and

visualization of the multi-domain mode shapes obtainable from the coupled system
formulation and, partly, from the extended perturbation. The article thus presents
the multi-domain mode shapes encompassing intertwined rotor, foil and pressure
states and compares these to the mode shapes attainable from the force coefficient
methods.
P5.2 Bearing Configuration and Basic Modelling

To facilitate an unambiguous study of the pressure–compliance dependency, the
present work is based on the simplest feasible rotor–GFB system. This comprises
a widely studied bearing configuration, the Simple Elastic Foundation Model
(SEFM) for the compliant structure and a point mass rotor. The mathematical
models applied in each domain are treated in the following sections and the
properties of the studied system are given in table P5.1. All quantities in all three
sub domains are non-dimensionalized using common scalings: the frequency ωτ ,
P5.2 Bearing Configuration and Basic Modelling 139
ambient pressure pa , the clearance C for radial lengths and the radius R for axial
and circumferential lengths.
P5.2.1 Fluid Film

Assuming ideal gas behaviour and defining the circumferential and axial coordinates
θ and z̃, the isothermal, compressible and transient Reynolds equation (RE) can
be written in terms of pressure p and film height h as
∂ 3 ∂ p̃ ∂ 3 ∂ p̃ ∂ ∂

+ = Sτ Ω̃ p̃ h̃ + 2Sτ

p̃ h̃ p̃ h̃ p̃ h̃
∂θ ∂θ ∂z̃ ∂z̃ ∂θ ∂τ
(P5.1)
6R2 µωτ
where Sτ =
C 2 pa
and µ is the gas viscosity. Following the Bubnov–Galerkin finite element (FE)
approach, eq. (P5.1) is discretized over half the bearing surface (to exploit symmetry
conditions) using a regular grid of nθ × nz nodes connected by bilinear quadrilateral
elements as previously described by the authors [91, 93]. The FE approach is
straightforward to combine with both implicit and explicit compliance models and
performs well despite the fact that it does not guarantee local conservation, as
could have been achieved using finite volume. Introducing the non-dimensional
Table P5.1: Properties of the GFB configuration, mainly from [154].
Journal
Load, W x = mr g 30 N Mass, mr 3.059 kg
Bearing radius, R 19.05 mm Pad leading edge, θ l 0◦
Bearing length, L 38.10 mm Pad trailing edge, θ t 360◦
Radial clearance, C 31.80 µm
Fluid properties
Ambient pressure, pa 101.3 kPa Viscosity, µ 1.950×10−5 Pa s
Foil thickness, tb 0.1016 mm Bump pitch, Sb 4.572 mm
Bump half length, l0 1.778 mm Bump height, hb 0.5080 mm
Young’s modulus Eb 207.0 GPa Poisson’s ratio, νb 0.3
Density, ρb 8280 kg/m3 Loss factor, η 0.2
θ j=1
j = nθ
∆θ j
Ω
Wj h
ey
1 ex y
Wx
wj
θ j−1 θj θ j+1 θ x
(a) (b)
Figure P5.1: (a) Weight function associated with the j-th foil compliance DOF;
(b) Schematics of the GFB and illustration of the perturbed DOFs.
nodal pressure vector p̃, the resulting element level Galerkin residual equations are
∫ ∫
1 T Û̃
−B p̃ h̃ B + Sτ Ω̃B h̃
T 3 T
N − 2Sτ N hNdAp̃ − 2Sτ e
NT h̃NdApÛ̃ e = r̃RE
e
,
Ae 0 Ae
(P5.2)
where N is the element shape function matrix, r̃RE is the element level residual
e
and B = ∂NT /∂θ ∂NT /∂z̃ . The system level residual equations are obtained
T
by the usual element summation. The film height h̃ is composed of a rigid and
a compliant contribution with the former given from the rotor DOF vector x̃r ,
while the latter depends either on the pressure p̃ for an implicit formulation or on
the foil DOF vector x̃ f for an explicit formulation. It is convenient to partition
eq. (P5.2) as
ARE p̃, x̃r , x̃ f p̃ + BRE p,

Û̃ xÛ̃ r , xÛ̃ f p̃ + CRE p̃, x̃r , x̃ f pÛ̃ = r̃RE, (P5.3)

such that ARE represents the Poiseuille and Couette terms, BRE holds the squeeze
term contribution and CRE the local expansion term. The Jacobians needed in
order to solve eq. (P5.3) are given in appendix P5.A, while the boundary conditions
are given as
p̃ (θ, 0) = p̃ (0, z̃) = p̃ (2π, z̃) = 1 and ∂ p̃/∂z̃|z̃= L̃/2 = 0. (P5.4)

P5.2.2 Rotor
The rotor is modelled as a point mass, i.e. with two DOFs and no gyroscopic effect.
The rotor model can be cast in state space residual form using the rotor DOF
vector x̃r = ε x ε y as
T
0 M̃r xÛ̃ r f̃rw + f̃b

0 0 x̃r
+ − = r̃r ∈ R4, (P5.5)
I 0 xÜ̃ r 0 −I xÛ̃ r 0
| {z } |{z} | {z } |{z}
Ar zÛ r Br zr
where M̃r = ωτ2Cp−1 a R diag({mr mr }) is the rotor mass matrix, f̃rw = pa R

−2 −1 −2
{W x 0} holds the static load and f̃b contains the bearing forces from the integrated
gas film pressure. The rotor is assumed to be perfectly aligned to the bearing
at all times (it includes no rotational DOFs) giving a one-dimensional rigid film
height contribution. Referring to fig. P5.1(b), this is given in the usual form as
h̃r (θ) = 1 + ε x cos θ + ε y sin θ, (P5.6)
where the trigonometric functions can be evaluated offline at the circumferential

positions needed by the gas film model allowing the rigid film heights to be
evaluated using the constant mapping
h̃r = 1 + Hr x̃r . (P5.7)
The bearing forces are found from the usual integration of the projected gas
film pressure approximated using a weighted sum over the nodal pressures as
2π nθ nz
L̃
cos θ cos θi Õ
∫ ∫
∆z̃ j p̃(θi, z̃ j ) − 1 , (P5.8)
Õ
f̃b =

( p̃ − 1) dA ≈ ∆θi
0 0 sin θ sin θi
i j
where ∆θi and ∆z̃ j are the weights assigned to a given nodal position by the
trapezoidal rule (chosen for consistency to the linear shape functions). The
integration of forces can hence be written as a mapping
(P5.9)
T
= H fb p (p̃ − 1) ,

f̃b 0
which is convenient when formulating the coupled system Jacobian. As is common

for GFBs, the Gümbel condition is applied meaning that sub ambient pressures
are discarded during the rotor force integration. This requires the columns of H fb p
corresponding to sub ambient nodal pressures to be zeroed, effectively making the
mapping pressure dependent. Notice that the Gümbel condition is one possible
(and simple) way of treating the sub ambient regions. The current research is
motivated by previous work by the authors [87, 88, 91, 93] on a three-pad bearing
for which the Gümbel condition is well-suited, hence the authors have a special
interest in this particular case. In general, one should carefully consider the foil
configuration in question (fixed leading/trailing edge, position of pad fixation points
compared to load direction etc.) to decide on the most appropriate treatment of
sub ambient regions as discussed in e.g. [16, 122].
P5.2.3 Compliant Structure

The compliant structure is modelled using the commonly applied SEFM [62]. This
is essentially a Winkler foundation model assuming neighbouring points to be
independent and the top foil to be inseparable from the bump foil. Furthermore,
the foil is assumed to deform uniformly in the axial direction as a function of the
axially averaged pressure, which has previously been shown to be a reasonable
assumption [156]. For more realistic foil modelling, a variety of elaborate foil
structure models are available in the literature [45, 96, 125, 155], but the added
complexity of these would complicate a strictly consistent implementation across
the different modelling approaches and blur the present comparison.
The SEFM is implemented using a uniform nominal stiffness given from the
widely applied expression [176]
3
Eb t b
(P5.10)
−1
k0 = 1 − νb2 ,
2Sb l0
where Eb , Sb , tb , l0 and νb are the bump foil’s Young’s modulus, pitch, thickness,
half bump length and Poisson’s ratio, respectively. The stiffness predicted by
eq. (P5.10) does not account for bump–bump interaction or friction and has been
found to underestimate the bump foil stiffness [89, 94]. For the present work,
however, the accuracy of the predicted nominal value is of limited importance as
the foil stiffness is varied between 1/10k 0 and 1000k0 .
The structural damping of the foil is assumed viscous as quantified by the
hysteretic damping constant η = k −1 dω f with ω f being the foil oscillation frequency
and d the equivalent viscous damping coefficient. When used with harmonic
perturbation methods, it is common [74, 81, 93] to combine the foil stiffness and
damping into a complex dynamic stiffness as

k c = k iωs ω−1
f η + 1 = k (iη + 1) ∈ C, (P5.11)
where the foil oscillation frequency ω f is dictated by the perturbation frequency ωs

yielding the second equals sign. For time domain models, however, the simulated
response will usually contain a variety of frequency components, but one still needs
to fix ω f at a certain value. For want of anything better, the most common choice
is to assume 1X foil oscillation [18, 19, 69, 74, 88, 90], i.e. ω f = Ω. This might
be exact for a non-resonant unbalance response, but only a rough approximation
in the presence of significant non-synchronous vibrations. Bonello and Pham [19]
investigated the influence of this issue for self-excited limit cycles by comparing
non-linear time integrations to results from a Harmonic Balance analysis and found
it to be negligible. This conclusion was, however, drawn for a practically rigid
bearing (236.5 GN/m3 ) with a small clearance (15.9 µm) and low foil damping
(η = 0.1) in which case the compliant structure has a vanishing influence on the
system dynamics compared to the gas film. More recently, Pronobis and Liebich
[146] investigated the same issue for higher compliance levels (up to around C/2)
where they proved it to have a significant impact. In [146], consistency is achieved
by setting ω f = Ω in eq. (P5.11), effectively giving k c = k c (ωs ), for use in the
perturbation method while defining the time domain viscous damping coefficient as
ηk Ω −1 (contrary to ηkω−1 f as used in the present work, see eq. (P5.13) below). The
approach taken in [146] allows the damping to be defined a priori and consistently
for both models providing a signifiant improvement of the agreement in predicted
OSIs. Though being consistent, the damping model in [146] causes the steady
state energy dissipation per oscillation cycle for a given amplitude X to become
frequency dependent as πkη ωΩs X 2 which, depending on the interpretation of η,
may or may not be desirable. In the present work, k c is kept constant as defined
from eq. (P5.11) providing the constant energy dissipation per cycle πkηX 2 , but
requiring the time domain OSI to be calculated iteratively since the set value
for ω f must match the resulting self-excited oscillation frequency. As most GFB
vibrations of interest appear around Ω/2, the difference in provided damping
between the two approaches is small compared to the remaining uncertainties
related to the loss factor, but results should not be compared directly between the
two approaches. Notice that the damping model applied in [146] allows consistent,
but frequency dependent, damping to be prescribed simultaneously across all
modes, while the present approach provides consistency at the particular frequency
ω f only. This is not a major concern when investigating the OSI as this is governed
solely by the leading mode, but the iterative approach required for each frequency
would become increasingly impractical if multiple modes were to be investigated.
In the present work, the SEFM is implemented both in an implicit and an

explicit version. In the former case (used for the perturbation methods), the
compliant film height is a direct function of the axially averaged pressure p̃m
(approximated using trapezoidal integration) as

∫ L̃
1 1
h̃c (θ) = ( p̃m (θ) − 1) = ( p̃ (θ, z̃) − 1) dz̃
k̃ c L̃ k̃ c 0
(P5.12)
| {z }
≡ fhc (θ, p̃−1)
nz
1 Õ
∆z̃ j p̃ θ, z̃ j − 1 .

≈
L̃ k̃ c j=1
In the explicit case, the compliance is given from x f containing nθ foil DOFs
representing pointwise radial deformations at the pressure nodes’ circumferential
positions. In this case, the stiffness and damping matrices can be written as
K f = LRk diag ∆θ 1 . . . ∆θ nθ ∈ Rnθ ×nθ ,

(P5.13)
D f = ηω−1
f Kf ∈ Rnθ ×nθ ,
where L is the axial length of the bearing and ∆θ j is the angular segment ascribed
to the j-th foil DOF as illustrated in fig. P5.1(b). The foil equation of motion
providing exactly the same dynamics as the implicit model can hence be written
in residual form as
D̃ f xÛ̃ f + K̃ f x̃ f −f̃ p = r̃0f ∈ Rnθ , (P5.14)
|{z} |{z} |{z} |{z}
A 0f zÛ 0f B 0f z0f
representing a first order system with time constant ηω−1 f . The vector f̃ p holds the
loading from the axially averaged gas film pressure minus the ambient pressure
acting on the back of the foil. This is obtained using trapezoidal integration as
∫ L̃ ∫ θ1 + 1 ∆θ1
2
p̃ (θ, z̃) − 1dθdz̃ 

 
0 θ − 1
 ∆θ

 1 2 1 
.
 
f̃ p = ..

 


 ∫ L̃ ∫ θ n + 1 ∆θ n 

θ 2 θ
 0 θ − 1 ∆θ p̃ (θ, z̃) − 1dθdz̃ 
 
(P5.15)
 
 nθ 2 nθ 
n
∆θ 1 j=1 p̃ θ 1, z̃ j − 1 ∆z̃ j 
Í z


 
.
 
.. = H fp p (p̃ − 1) .

 

≈ L̃
Í nz
j=1 p̃ θ nθ , z̃ j − 1 ∆z̃ j 
 
 ∆θ nθ
 


Even though this is expected to be insignificant to the overall system dynamics,
the proposed extended perturbation method furthermore requires the inclusion of
foil mass. A simple foil mass matrix can be obtained similarly to K f and D f in
eq. (P5.13) as
M f = m0f R diag ∆θ 1 . . . ∆θ nθ ∈ Rnθ ×nθ , (P5.16)

P5.3 Perturbation Methods 145
where m0f is the average mass of the foil structure per unit circumferential length
(see appendix P5.B). The foil equation of motion including mass can hence be
written in state space residual form as
0 M̃ f xÛ̃ f

K̃ f D̃ f x̃ f f̃
Ü̃ + Û̃ − p = r̃ f ∈ R2nθ . (P5.17)
I 0 xf 0 −I x f 0
| {z } |{z} | {z } |{z}
Af zÛ f Bf zf
P5.3 Perturbation Methods

P5.3.1 The Classical two-DOF Perturbation
This section gives a brief review of the classical two-DOF perturbation employed
with various variations by e.g. [68, 69, 74, 79–81, 87, 93, 101, 132, 175]. As
originally introduced in the appendix of [110], the fundamental assumption is for
the pressure field p̃ to be given as a function of rotor position x̃r and velocity xÛ̃ r
only. Assuming the rotor exhibits small harmonic oscillations with frequency ω̃s
around a stationary point x̃r0 as
x̃r = x̃r0 + ∆x̃r eiω̃s τ, (P5.18)
a Taylor series expansion leads to a perturbed pressure field given as
p̃ = p̃0 + p̃ x Λ x + p̃ y Λ y, (P5.19)
where
p̃0 = p̃ (x̃r0 ) ∈ R,
∂ p̃ ∂ p̃

p̃γ = + iω̃s ∈ C for γ = x, y (P5.20)
∂εγ x̃r0 ∂ εÛγ x̃r0

and Λγ = ∆εγ eiω̃s τ ∈ C for γ = x, y.

This perturbation is then substituted into RE from eq. (P5.1), while the foil
compliance is introduced implicitly through the film height function as
h̃ (θ) = h̃r (θ) + fhc (θ, p̃ − 1) , (P5.21)
where fhc provides the axial average of p̃ − 1 at the circumferential position θ

as defined from eq. (P5.12). Inserting the film height description and pressure
perturbation from eqs. (P5.19) and (P5.21) into eq. (P5.1), the zeroth order
equation for p̃0 can be extracted by collecting terms of order Λ0γ as
∂ 3 ∂ p̃0 ∂ 3 ∂ p̃0 ∂

+ = Sτ Ω̃ p̃0 h̃0 , (P5.22)

p̃0 h̃0 p̃0 h̃0
∂θ ∂θ ∂z̃ ∂z̃ ∂θ
while the first order equations for p̃γ correspond to the terms of order Λ1γ as
∂ 3 ∂ p̃γ ∂ 3 ∂ p̃γ ∂ ∂ p̃0

p̃0 h̃0 + p̃0 h̃0 + h̃ p̃γ h̃0 + 3 p̃0 fhc θ, p̃γ
2
∂θ ∂θ ∂z̃ ∂z̃ ∂θ 0 ∂θ
∂ ∂ p̃0 ∂

+ h̃02 p̃γ h̃0 + 3 p̃0 fhc θ, p̃γ − Sτ Ω̃ p̃γ h̃0 + p̃0 fhc θ, p̃γ

∂z̃ ∂z̃ ∂θ
(P5.23)
−2iSτ ω̃s p̃γ h̃0 + p̃0 fhc θ, p̃γ
∂ ∂ 2 ∂ p̃0 ∂ 2 ∂ p̃0

= 2iSτ ω̃s p̃0 fγ + Sτ Ω̃ ,

p̃0 fγ − 3 p̃0 h̃0 fγ − 3 p̃0 h̃0 fγ
∂θ ∂θ ∂θ ∂z̃ ∂z̃
γ = x, y,
where fx = cos(θ), fy = sin(θ) and
h̃0 (θ) = 1 + ε x0 cos θ + ε y0 sin θ + fhc (θ, p̃0 − 1) . (P5.24)
The zeroth order equation is a real-valued (η is set to zero) non-linear equation

corresponding to the steady part of eq. (P5.2) with p̃ = p̃0 . It is solved in a segreg-
ated fashion using two nested Newton–Raphson (NR) solvers. The outer solver
searches for the rotor equilibrium x̃r0 satisfying eq. (P5.5) using finite difference
approximations to the bearing force Jacobian ∂ f̃b /∂x̃r ∈ R2×2 (incorporating the
Gümbel condition as given from eq. (P5.9)). The inner solver solves eq. (P5.22)
for a given eccentricity using the analytical expressions for the Jacobian ∂r̃RE /∂p̃
provided in appendix P5.A.
The first order equation eq. (P5.23) is linear and complex valued and can be
discretized using FE following the same procedure as for the zeroth order equation
to yield
A pγ p̃γ = r̃ pγ , (P5.25)
where A pγ ∈ Cnθ nz ×nθ nz and p̃γ, r̃ pγ ∈ Cnθ nz . On the element level, these are given as
∫
1
Aepγ p̃γ = p̃0 h̃03 BT B+ h̃03 BT Bp̃0e N − Sτ Ω̃ h̃0 B
T
N + 2iω̃s Sτ h̃0 NT NdAp̃γe
Ae 0
∫
T 1
+ 3 p̃0 h̃0 B Bp̃0 N − Sτ Ω̃ p̃0 B
2 T e
N + 2iω̃s Sτ p̃0 NT NdAHehc p p̃γ
Ae 0
(P5.26)
∫
1
r̃epγ = Sτ Ω̃ p̃0 fγ BT − 3 p̃0 h̃02 fγ BT Bp̃0e − 2iω̃s Sτ p̃0 fγ NT dA. (P5.27)
Ae 0
which is solved using the LAPACK routines zgbtrf and zgbtrs [3]. The coupling
to p̃γ , i.e. to pressure DOFs not included in p̃γe , in eq. (P5.26) is due to the axial
pressure averaging given by Hehc p which is detailed in appendix P5.A, while the
boundary conditions are given as
p̃γ (θ, 0) = p̃γ (0, z̃) = p̃γ (2π, z̃) = 0 and ∂ p̃γ /∂z̃|z̃= L̃/2 = 0. (P5.28)
Having obtained the dynamic pressure fields, the gas film reaction forces can
be integrated to obtain the bearing coefficients as
∫ L̃ ∫ 2π
d˜x x d˜x y p̃ x cos θ p̃ y cos θ

k̃ x x k̃ xy
+iω̃s ˜ = dθdz̃ ∈ C2×2,
k̃ yx k̃ yy dyx d˜yy 0 0 p̃ x sin θ p̃ y sin θ
| {z } | {z }
K̃g,2DOF D̃g,2DOF
(P5.29)
using which, the linearised equations of motion for the two rotor DOFs around the
static equilibrium can be written as
M̃r xÜ̃ r + D̃g,2DOF ω̃s, Ω̃ xÛ̃ r + K̃g,2DOF ω̃s, Ω̃ x̃r = 0. (P5.30)

Compared to previous work by the authors [88, 93, 128], two important
changes should be mentioned. First, the point wise expression for the dynamic
compliant film height p̃γ k̃ c−1 has been replaced by fhc θ, p̃γ in eq. (P5.23) to
provide axial averaging in agreement with the calculation of the static film height
h̃0 as also described in [81]. Second, the Gümbel condition implies that sub
ambient regions should not contribute to the dynamic forces, i.e. p̃γ = 0 where
p̃0 < 1. Previously, this was introduced as constraints in eq. (P5.25) before its
solution, while the dynamic pressure in sub ambient regions is now discarded only
during the integration in eq. (P5.29) analogously to the static pressure integration
in eqs. (P5.8) and (P5.9). It should be emphasised that the truncation of the
dynamic pressure field p̃γ based on the static pressure p̃0 is only correct if the
coefficients are calculated at the static equilibrium, e.g. to assess its stability from
the eigenvalues. If the coefficients are calculated along a trajectory in a transient
simulation, the truncation should instead be based on the instantaneous pressure
field at each point of evaluation.
P5.3.2 An Extended Perturbation

A natural extension of the classical two-DOF perturbation is to introduce an
explicit dependency of the pressure field on a number of foil states, akin to the pad
DOFs used when analysing tilting pad journal bearings [2, 159]. Effectively, this
means that the pressure is assumed to vary as a function of foil deformation even in
the absence of rotor movement. To allow a direct comparison, this is implemented
using nθ discrete radial deformations w̃ j located at the same circumferential
positions as the pressure nodes. The deformed shape between these points is
evaluated using linear interpolation consistent with the linear shape functions of
the FE discretization of RE making the explicit formulation equivalent to the
implicit one in terms of the static deflection. A weight function describing the
film height contribution from each foil DOF is hence defined and illustrated in
fig. P5.1(a) as
θ−θ j−1



 θ j −θ j−1 if θ j−1 < θ < θ j


W j (θ) = θθ j+1−θ
−θ
if θ j ≤ θ < θ j+1 (P5.31)


 j+1 j


otherwise,

0


such that the continuous film height can be written as
nθ
(P5.32)
Õ
h̃ (θ) = 1 + ε x cos θ + ε y sin θ + w̃ j W j (θ) ,
j=1
where the discrete deformation at each foil DOF is calculated using the axially
averaged pressure defined in eq. (P5.12) as
w̃ j = fhc θ j , p̃ − 1 . (P5.33)

The assumption of small harmonic oscillations, which was applied to the rotor
DOFs in eq. (P5.18), is now applied also directly to the foil DOFs as
w̃ = w̃0 + ∆w̃eiω̃s τ where w̃ = w1 · · · wnθ , (P5.34)

T
such that a first order Taylor series expansion of the pressure field p̃ = p̃ (x̃r , w̃)
now provides 3 + nθ terms as
nθ
(P5.35)
Õ
p̃ = p̃0 + p̃ x Λ x + p̃ y Λ y + p̃w j Λ j ,
j=1
where p̃0 , p̃ x , p̃ y , Λ x and Λ y are defined from eq. (P5.20) as previously, except
that the former three should now be evaluated at x̃r0, w̃0 . The dynamic pressure
fields related to the foil DOFs are defined as
∂ p̃ ∂ p̃

p̃w j = + iω̃s ∈ C j = 1, . . . , nθ and
∂ w̃ j x̃r0,w̃0 ∂ wÛ̃ j x̃r0,w̃0 (P5.36)

Λ j = ∆w̃ j eiω̃s τ ∈ C j = 1, . . . , nθ .
Notice that p̃w j provides the pressure derivative at any point with respect to the
j-th foil deformation DOF at the frequency ω̃s . Fundamentally, this is no different
from the p̃ x and p̃ y fields supplying the pressure derivatives with respect to the
two rotor DOFs.
Inserting the film height description and extended perturbation from eqs. (P5.32)
and (P5.35) into RE from eq. (P5.1), the zeroth and first order equations can be
separated as was also the case for the two-DOF perturbation. The zeroth order
equation is unchanged, except for an alternative, but equivalent, static film height
expression
nθ
(P5.37)
Õ
h̃0 (θ) = 1 + ε x0 cos θ + ε y0 sin θ + w̃ j0W j (θ) ,
j=1
where w̃ j0 is evaluated with η = 0 as for the implicit formulation. Compared to

the two-DOF perturbation, the first order equations for the rotor DOF dynamic
pressure fields p̃γ are slightly simplified corresponding to setting fhc θ, p̃γ = 0 in

eq. (P5.23), reflecting that the dynamic compliance is no longer dictated by the
rotor movement. Additionally, nθ first order equations for the dynamic pressure
fields related to the foil DOFs are obtained, giving in total 2 + nθ first order
equations as
∂ 3 ∂ p̃α ∂ 3 ∂ p̃α ∂ ∂ p̃0 ∂ ∂ p̃0

p̃0 h̃0 + p̃0 h̃0 + 3
h̃ p̃α + 3
h̃ p̃α
∂θ ∂θ ∂z̃ ∂z̃ ∂θ 0 ∂θ ∂z̃ 0 ∂z̃
∂
−Sτ Ω̃ p̃α h̃0 − 2iSτ ω̃s p̃α h̃0

∂θ (P5.38)
∂ ∂ ∂ ∂ ∂

p̃0 p̃0
= 2iSτ ω̃s p̃0 fα + Sτ Ω̃ ( p̃0 fα ) − 3 p̃0 h̃02 fα − 3 p̃0 h̃02 fα ,
∂θ ∂θ ∂θ ∂z̃ ∂z̃
α = x, y, w1, . . . , wnθ ,
where fx = cos(θ), fy = sin(θ) and fw j = W j (θ). These can be FE discretized and

solved analogously to the two-DOF first order equations from eq. (P5.23). Having
obtained p̃0 , p̃γ and p̃w j , these can be integrated to obtain the gas film reaction
forces for each DOF as

θ

 ∫ 2π cos 
 ( p̃ − 1) dθ 
θ
 


 0 sin 


∫ L̃  θ1 + 2 ∆θ1

 ∫ 1 

f̃b ( p̃ − 1) dθ

=

 

θ −
1 2
1
∆θ 1 dz̃
f̃ p 0  ..
.
 


 


 ∫ θ n + 1 ∆θ n 

θ 2 θ
( p̃ − 1) dθ 

 
θ 1
 
− ∆θ
(P5.39)
 nθ 2 nθ 
cos θ

 ∫ 2π 
 ( p̃ − 1) dθ 
0
θ
 


 0 sin 


θ + 1
 
∫ L̃  ∫ 1 2 ∆θ 1

( p̃ − 1) dθ Z Z x̃r iω̃s τ
 
0 x x x w
= dz̃ + e ,

 

θ 1 − 2 ∆θ 1
1 r r r
0  .. Zwxr Zww w̃
.
 


 
 | {z }

 ∫ θ n + 1 ∆θ n 
 ≡Z
θ 2 θ
( p̃ − 1) dθ 
 
 θ nθ − 12 ∆θ nθ 0

 

where the complex stiffness matrix Z ∈ C(2+nθ )×(2+nθ ) is given as
 ∫ 2π p̃ cos θ p̃ cos θ ∫ 2π p̃w1 cos θ · · · p̃wn cos θ

x y θ
 dθ 0 dθ 
 0
 p̃ x sin θ p̃ y sin θ p̃w1 sin θ · · · p̃wnθ sin θ 

∫ L̃  ∫ θ1 + 1 ∆θ1 ∫ θ1 + 1 ∆θ1


Z=
 2
p̃ x p̃ y dθ 2
p̃ w · · · p̃ w dθ  dz̃

 θ 1 − 2 ∆θ 1
1
θ 1 − 2 ∆θ 1
1 1 n θ
0  .. .. 
 . . 

 ∫ θ n + 1 ∆θ n ∫ θ n + 1 ∆θ n 
θ θ θ θ
2
2

 p̃ x p̃ y dθ p̃w1 · · · p̃wnθ dθ 

 θ nθ − 2 ∆θ nθ
1
θ nθ − 2 ∆θ nθ
1 

 k̃ x x k̃ xy k̃ xw1 ··· k̃ xwnθ   d˜x x d˜xy d˜xw1 ··· d˜xwnθ 

 d˜yx d˜yy d˜yw1 d˜ywnθ 
  
 k̃ yx
 k̃ yy k̃ yw1 ··· k̃ ywnθ 
  ···
=  w1 x k̃ w1 y k̃ w1 w1 ···  +iω̃s  d˜w1 x d˜w1 y d˜w1 w1 ··· d˜w1 wnθ  .
 k̃ k̃ w1 wnθ  
 . .. .. ... ..  . .. .. ... .. 
 .. . . .  .. . . . 


  
 k̃ w x k̃ w y k̃ w w · · · k̃ wnθ wnθ  d˜w x d˜w y d˜w w · · · d˜wnθ wnθ 
 nθ nθ nθ 1  nθ nθ nθ 1
| {z } | {z }
K̃g,xDOF D̃g,xDOF
(P5.40)
Finally, the system equations of motion, which now comprise the rotor DOFs
as well as the nθ newly introduced foil DOFs, can be assembled as
xÜ̃ r xÛ̃ r

M̃r 0 0 0
0 M̃ f wÜ̃ + 0 D̃ f (ω̃s ) + D̃g,xDOF ω̃s, Ω̃ wÛ̃
(P5.41)
0 0 x̃r
+ + K̃g,xDOF ω̃s, Ω̃ = 0,
0 K̃ f w̃
where the foil damping is given from eq. (P5.13) with ω f = ωs . At present,
the foil mass is included in eq. (P5.41) to provide a straightforward format of
the extension not treating the foil DOFs any differently from the rotor DOFs.
If recasting the rotor equations into state space form, it would be possible to
rearrange eq. (P5.41) leaving out the foil mass as shown in appendix P5.C. This
would lead to a slightly altered eigenvalue problem requiring adjustments to the
solution procedure described in the following section, but would probably yield
similar results. The foil mass is, however, retained since this intuitively matches
the assumption of harmonic oscillations applied in eq. (P5.34) and ultimately is
shown not to affect the OSI.
P5.3.3 Calculation of Eigenvalues and Eigenvectors

Calculation of eigenvalues from the equations of motion resulting from the per-
turbation methods is complicated by the frequency dependency of the stiffness
and damping matrices. For a generic frequency dependent system
MÜx + D (ωs, Ω) xÛ + K (ωs, Ω) x = 0, (P5.42)
the eigenvalues λi at a given speed Ω are the roots of the characteristic polynomial
Pc (λi ) = 0
Pc (λ) = det λ M + λD (Im (λ) , Ω) + K (Im (λ) , Ω) , (P5.43)
2
where
| {z }
Aλ (λ,Ω)
while the corresponding eigenvector vi (for eigenvalues with multiplicity 1) is found

as the null space (or kernel) as
vi = null (Aλ (λi, Ω)) . (P5.44)
To find the eigenvalues at a given speed Ω from eq. (P5.43), the matrices
K (ωs, Ω) and D (ωs, Ω) must be evaluated over a range of ωs values from which
the determinant of Aλ can be calculated over a suitable grid of guessed real and
imaginary λ values. The local minima of the resulting response surface (currently
found using a simple filter) can then serve as initial guesses from which the
eigenvalues λi can be found using a standard local minimization routine (here the
downhill simplex algorithm as implemented in SciPy’s ”fmin”), relying either on
interpolation between the previously evaluated K (ωs, Ω) and D (ωs, Ω) matrices
or re-evaluation of these. Notice that a very similar approach to the calculation of
eigenvalues has recently been presented by Pronobis and Liebich [146]. Having
located an eigenvalue λi , finding the corresponding eigenvector then requires a
singular value decomposition of the resulting Aλ (λi, Ω) matrix to identify its null
space. Repeating this procedure to obtain the eigenvalues as functions of Ω, one

can construct a Campbell diagram and assess the stability limit.
If only the stability limit is of interest, a graphical approach to the solution is
possible as previously applied by the authors [88, 128]. Temporarily relaxing the
requirement for ωs to coincide with Im (λi ), the eigenvalue problem

λ̂i2 M + λ̂i D (ωs, Ω) + K (ωs, Ω) v̂i = 0 (P5.45)
can be solved for the pseudo eigenvalues λ̂i over a suitable ωs × Ω grid using a
standard eigenvalue solver. Sorting the obtained pseudo eigenvalues consistently,
the frequency and speed dependent pseudo eigenvalues λ̂i (ωs, Ω) can be obtained
and the stability limit ΩOSI can be found by evaluating the condition

Im λ̂i (ωs, Ω = ΩOSI ) = ωs ∧ Re λ̂i (ωs, Ω = ΩOSI ) = 0, (P5.46)
for each λ̂i . The left half of the condition requires the pseudo eigenvalue to satisfy
eq. (P5.43) and hence to represent an actual eigenvalue, while the right half
requires the same eigenvalue to be marginally stable. The fulfilment
of eq. (P5.46)
can be found graphically by superimposing the contour of Im λ̂i /ωs = 1 on the

contours of Re λ̂i as shown in fig. P5.2. It should be noted that the OSI defined
by eq. (P5.46) is not restricted to modes with any particular structure. For systems
resulting from the extended perturbation, the condition could hypothetically be
fulfilled by a mode with vanishing rotor participation. In this case, the static
equilibrium would have become unstable, but further analysis would be needed to
assess if the OSI (in a rotordynamics sense) had been found. This situation has,
however, not been encountered by the authors and all perturbation OSIs presented
in the present paper are defined by rotor-dominated modes, i.e. the modes that
would appear in a Campbell diagram.
For the 2×2 systems resulting from the two-DOF perturbation, both approaches
are relatively easy to apply. Only two eigenvalues are to be found and these are
usually well separated by frequency. The sorting of pseudo eigenvalues between the
(ωs, Ω) grid points required to generate the contour plots is hence trivial and the
evaluation of the characteristic polynomial in eq. (P5.43) is very cheap. However,
for the (2 + nθ ) × (2 + nθ ) systems resulting from the extended perturbation, both
methods become significantly more difficult. The number of eigenvalues increases
drastically and they are usually much more densely spaced. For the contour
plot approach, each individual eigenvalue problem becomes much more expensive
and a more elaborate scheme for sorting the eigenvalues (presently based on the
eigenvector structures) is required. For the root finding approach, the initial
Re (λ) × Im (λ) grid must be denser to reliably find and separate the requested
P5.4 Direct Coupling of the Domains 153
eigenvalues, the evaluation of eq. (P5.43) becomes rather expensive and the values
of Pc (λ) (away from the roots) explode due to the matrix size causing numerical
difficulties. The latter issue is currently tackled by calculating the logarithm of
the determinant, i.e. log (Pc (λi )) = log det (Aλ ), instead of the determinant directly.
Note that if only the OSI is of interest, only a few Re (λ) × Im (λ) grid points in
the Re (λ) direction around zero are sufficient. In the present work, the OSIs are
based on K and D matrices evaluated with an ωs resolution of 1 Hz for every 10
RPM followed by linear interpolation to locate the roots to an absolute accuracy
of 10−4 Hz.
P5.4 Direct Coupling of the Domains

As an alternative to the perturbation approaches, one can formulate the governing
equations for the rotor–GFB system as a set of coupled non-linear equations
allowing the full system dynamics to be captured. For the present rotor–bearing
system, the global state-space and residual vectors can be defined as
n oT n oT
z= p̃T zTf zrT and r̃ = r̃T T T
RE r̃ f r̃r , (P5.47)
while the governing equations for the three domains can be combined into the
global residual function fG (z, zÛ ) with transient, steady and constant components
fG,t (z, zÛ ), fG,s (z) and fG,0 . This allows the coupled non-linear system of equations
to be written in implicit form as
fG (z, zÛ ) =
,

A p̃, x̃ x̃ p̃
 BRE xr , x f p̃ + CRE x̃r , x̃ f pÛ̃ 
Û̃ Û̃


 


 RE
h r
iT f 

 
  0 

+ B f z f + H f p 0 (p̃ − 1) +
 
 
  
 

A f zÛ f T 0
p  T
Ar zÛ r
 
  

  

H f p (p̃ − 1)

 
 f̃rw 0  
| {z } |  b 
{z } | {z }
fG,t (z, zÛ ) fG,s (z) fG,0
= r̃,
(P5.48)
where the topmost row holds the FE gas film equations from eq. (P5.3), but with
the p̃ and pÛ̃ dependencies of BRE and CRE omitted since no implicit compliance
model is included. Instead, the compliance is governed by the foil equations of
motion from eq. (P5.17) (or eq. (P5.14)) placed in the midmost row. Lastly, the
lowermost row holds the rotor equations of motion from eq. (P5.5). Notice that
if unbalance forces were to be included, fG,t (and hence fG ) would furthermore
contain an explicit time dependency.
If attainable, system formulations on explicit first order form, i.e. where fG (z) =
zÛ , are usually preferred over the implicit form used in eq. (P5.48). Transforming the
rotor and foil equations to explicit form would merely require a one-time inversion
of A f and Ar , while an explicit formulation of the gas film equations could have
been obtained by substitution of the alternative state variable ψ = p̃ h̃ into the
unsteady terms of eq. (P5.1) [139]. The present implementation [127] does, however,
encompass a range of foil structure models where the foil mass, the foil damping
or both of these are disregarded. The latter results in differential/algebraic (rather
than ordinary differential) equation systems where an explicit formulation cannot
be obtained hence the implicit form is used.
For eq. (P5.48) to be efficiently solved and to form a firm basis for the stability
analysis, it is necessary to formulate analytical expressions for the Jacobian
matrices. For the steady-state case where fG,t (z, zÛ ) = 0, the Jacobian with respect
to the state vector z is given as
 ∂ARE p̃,x̃r ,x̃ f p̃ ∂ARE p̃,x̃r ,x̃ f p̃ ∂ARE p̃,x̃r ,x̃ f p̃


∂p̃ ∂z f ∂zr
∂fG,s  h
 
(P5.49)
iT
= =  HT 0 ,

Jz,s Bf 0
∂z  fp p


 H f p (p̃) 0 Br 
 b 
where the upper three elements are state-dependent and given in appendix P5.A.
Notice that the foil residuals are not directly affected by the rotor states and vice
versa, which is physically meaningful since the rotor is only coupled to the foil
through the gas film. For the transient case, the Jacobian with respect to both the
state vector and the state vector’s temporal derivative is necessary. The Jacobian
with respect to z receives the additional contribution
∂CRE x̃r ,x̃ f pÛ̃ ∂CRE x̃r ,x̃ f pÛ̃

xÛ̃ r , xÛ̃ f

∂fG,t  BRE
 
∂z f ∂zr 
Jz,t = = (P5.50)

∂z  0 0 0 


 0 0 0 

representing the pressure and film height dependencies originating from the local
expansion and squeeze terms. Finally, the Jacobian with respect to the temporal
derivative of the state vector zÛ is given as
∂BRE xÛ̃ r ,xÛ̃ f p̃ ∂BRE xÛ̃ r ,xÛ̃ f p̃

∂fG,t  CRE x̃r , x̃ f

 
∂ zÛ f ∂ zÛ r 
JzÛ = = . (P5.51)

∂ zÛ  0 Af 0 

 0 0 Ar 

The static equilibrium of the rotor system can be found by zeroing out the tran-
sient component of eq. (P5.48), i.e. setting fG,t (z, zÛ ) = 0, and solving the resulting
non-linear algebraic equation system. This is achieved using a C implementation
P5.4 Direct Coupling of the Domains 155
utilising the general purpose algebraic solver KINSOL from the SUNDIALS suite
[67]. Having aligned the implementations carefully, this leads to exactly the same
equilibrium positions as found by the segregated solver applied for the zeroth
order equations (to machine precision). As the segregated approach seems to be
common practice in bearing codes, it is worth mentioning that the coupled solver
converges in about the same number of iterations (on the order of 10) as used
by the segregated solver’s outer NR loop alone, thus saving potentially hundreds
of iterations taken by the inner solver. Notice, however, that while the linear
equation systems arising from the segregated solver’s inner NR steps are banded
(bandwidth 4nz , see eq. (P5.A2)), those arising from the coupled solver steps are
not (see eq. (P5.49)). To benefit computationally from the lower iteration count of
the coupled approach, it is hence necessary to employ a sparse linear solver such
as KLU [34], which is currently being used.
Having obtained the static equilibrium z0 (where zÛ 0 = 0) at a given speed Ω,
the behaviour of the dynamical system in the vicinity of this equilibrium can be
assessed from a linearisation, i.e. a first order Taylor series expansion, of eq. (P5.48)
as
∂fG,t (z, zÛ ) ∂fG,t (z, zÛ ) ∂fG,s (z)

zÛ + + (z − z0 ) = 0, (P5.52)
∂ zÛ z0,Ûz0 =0 ∂z z0,Ûz0 =0 ∂z z0 | {z }

| {z } | {z } | {z } ∆z
JzÛ (z0,Ûz0 =0) Jz,t (z0,Ûz0 =0) Jz,s (z0 )
which can alternatively be written in the more familiar format
zÛ = AG ∆z,
(P5.53)
where AG = −JzÛ (z0, zÛ 0 = 0)−1 Jz,t (z0, zÛ 0 = 0) + Jz,s (z0 )

allowing the (linear) stability of the full system to be judged from the eigenvalues
of AG as introduced for foil bearing systems in explicit form by Bonello and Pham
[19] and used for systems on linearly implicit form in [52]. It should be emphasised,
as it was also discussed in relation to eq. (P5.46), that for an unstable leading
eigenvalue of AG to define the OSI consistently to the perturbation methods, it
must be related to a mode with significant rotor participation. An unstable leading
mode dominated by the foil would predict a stability limit undetectable using the
two-DOF perturbation, while a pressure dominated unstable leading mode would
be invisible to both perturbation techniques. A case in which the stability limit
is dictated by a mode with vanishing rotor movement has not been encountered
by the authors and it is debatable whether it would ever occur in practice for
common GFB configurations. For tilting pad journal bearings, on the contrary,
pad flutter would represent exactly such a case, hence potentially devastating
instabilities arising from the compliant structure (or other included domains)
should not be precluded. In the present form, it is evident that AG depends on JzÛ
to be invertible. This is the case when eq. (P5.48) represents a system of ordinary
differential equations, as a singular JzÛ is exactly the defining characteristic of a
differential/algebraic equation system. The presented stability analysis can hence
not be applied directly to a system containing algebraic equations as originating
e.g. from a purely static foil model.
Two notes should be made regarding the foil equations when included into
the coupled model. As is, eq. (P5.48) includes the mass-augmented SEFM from
eq. (P5.17) in order to match the foil description used in the extended perturbation
method. It is, however, straightforward to replace this by the massless SEFM from
eq. (P5.14) to produce a foil model matching that used in the classical perturbation
method. Secondly, the foil damping entering eq. (P5.48) through B f depends on
the foil oscillation frequency ω̃s as discussed in section P5.2.3. For the coupled
model to be completely consistent with the perturbation methods at the OSI, ω̃s
must match the frequency of the mode becoming unstable. As this is unknown a
priori, one needs to locate the coupled system OSI in an iterative fashion while
updating the foil oscillation frequency. In the present work, this has been done
manually, leading to ω̃s matching the frequency of the mode becoming unstable
(±0.001 Hz) within a few iterations.
P5.5 Results
The stability limit of the benchmark rotor–bearing system with parameters as given
in table P5.1 is calculated in three different ways using 1) the classical perturbation
method, i.e. the eigenvalues of eq. (P5.30); 2) the extended perturbation method,
i.e. the eigenvalues of eq. (P5.41); and 3) the coupled system representation, i.e.
the eigenvalues of eq. (P5.53). For all three approaches, the bearing varies from
very soft to practically rigid using SEFM stiffness values between 1/10 and 1000
times the nominal foil stiffness of 4.6417 GN/m3 as calculated from eq. (P5.10).
P5.5.1 Eigenvalues and Mode Shapes from the Classical Two-DOF

Perturbation
For the case of nominal foil stiffness, the stability maps for the two eigenvalues
resulting from the classical perturbation method are shown in fig. P5.2. The
black line marked ωs = ωi is the contour where eq. (P5.43) is fulfilled by the
pseudo-eigenvalues and hence gives the actual system eigenvalues as a function of
rotational speed Ω. The stability is indicated by the contours of the logarithmic
decrement showing the OSI condition from eq. (P5.46) to be fulfilled for the lowest
P5.5 Results 157
mode at 23.62 kRPM with a frequency of 102.7 Hz and the second mode to be
well-damped throughout the plotted range. The same limit has been found by
solving for the characteristic equation roots based on interpolation between stiffness
and damping matrices evaluated over a grid with 10 RPM and 1 Hz resolutions.
Omitting the decay factor eRe(λi )t , the rotor mode shape resulting from a given
eigenvector vi with associated natural frequency ωi = Im (λi ) can be evaluated as
x̃r = x̃r0 + vi eiωi t + v̄i e−iωi t (P5.54)
providing an oscillation amplitude |vi | + |v̄i | = 2|vi | which has been normalized to
an infinity norm of C/5. Having solved for the corresponding eigenvectors from
eq. (P5.44), the mode shapes associated with the two modes in fig. P5.2 can thus
be plotted over an oscillation period, i.e. with ωi t going from 0 to 2π, as shown
in fig. P5.3 along with the static foil deflection. The mode becoming unstable is
clearly identified as a forward whirling mode, while the stable mode at 130.7 Hz
represents backward whirling. Notice that only the rotor oscillation is plotted,
as the classical perturbation provides eigenvectors vi corresponding to the rotor
DOFs only. Using the relations given by eqs. (P5.19) and (P5.21), it would be
possible to reconstruct and plot also the foil and pressure oscillation based on the
rotor DOFs, but these would be phase and frequency locked to the rotor along the
entire circumference and hence would not contribute additional information.
P5.5.2 Eigenvalues and Mode Shapes from the Extended

Perturbation
Performing the same analysis based on the extended perturbation (still with
nominal foil stiffness), very similar stability maps can be obtained, but the OSI
(still governed by the lowest mode) is not reached until 24.34 kRPM at 103.1 Hz.
The corresponding mode shapes are shown in fig. P5.4, now encompassing the
static foil deflection, the rotor movement and the foil oscillation. This is possible as
the obtained eigenvectors (now with length 2 + nθ ) encompass both rotor and foil
DOFs directly, such that a mode shape is now plotted over an oscillation period as

x̃r x̃
= r0 + vi eiωi t + v̄i e−iωi t (P5.55)
w̃ w̃0
where the decay factor eRe(λi )t has been omitted and the infinity norm of 2|vi | has
been normalized to C/5 as for the rotor-only modes in eq. (P5.54). Again, the
mode shapes are visualized over a full oscillation period with ωi t running from 0 to
2π allowing the phase between different DOFs to be read from the colours. Notice
e.g. from the mode shown in fig. P5.4(a) (which is the one dictating the OSI) that
the foil oscillation at 180◦ is approximately in counter phase to the foil at 270◦ .
30 1.626
5
-0.0
0.0
5
25 0.050
0.00
Ω (kRPM)
δ (–)
20
0.000
i
ωs = ω
15
−0.050
10
−0.360
50 100 150 200 250
ωs (Hz)
(a)
30 4.8
4.0
25
2.40
3.2
Ω (kRPM)
δ (–)
20
2.4
ωs = ωi
15 1.6
0.80
1.6
0.8
0
3.20
10
0.0
50 100 150 200 250
ωs (Hz)
(b)
Figure P5.2: Stability maps for a nominal foil stiffness (k = 4.6417 GN/m3 )
resulting from a two-DOF perturbation based on a 46 × 45 (ωs, Ω) grid. The
contours of the logarithmic decrement δ = −2πRe (λ) /Im (λ) are plotted along
with a single contour showing the concurrence of the excitation frequency ωs and
the natural frequency ωi = Im (λ). Equation (P5.46) is fulfilled at the crossing of
δ = 0 thus marking the OSI (for mode 1 at 102.7 Hz, 23.62 kRPM, while mode 2 is
stable).
P5.5 Results 159
2π
−1.0
ε=1
xf 0
−0.5
ωi t (rad)
εx (–)
0.0
π
0.5
1.0
0
−1.0 −0.5 0.0 0.5 1.0
εy (–)
(a)
2π
−1.0
ε=1
xf 0
−0.5
ωi t (rad)
εx (–)
0.0
π
0.5
1.0
0
−1.0 −0.5 0.0 0.5 1.0
εy (–)
(b)
Figure P5.3: Static deflection and mode shapes at 23.62 kRPM resulting from
a two-DOF perturbation corresponding to the stability maps in fig. P5.2: (a)
Forward mode with ωi = 102.7 Hz and ζ = −2.4×10−5 ; (b) backward mode with
ωi = 130.7 Hz and ζ = 0.19. Notice that the decay factor eRe(λi )t has been omitted
from the mode visualizations.
Considering the rotor alone, the orbits resulting from the extended perturbation
are very similar to those from the classical perturbation shown in fig. P5.3.
P5.5.3 Eigenvalues and Mode Shapes from the Coupled Model

Using the coupled model to solve for equilibrium positions in the range
8–30 kRPM provides the locus shown in fig. P5.5. At each of these equilibria, the
system Jacobian AG as given in eq. (P5.53) is assembled and its eigenvalues are
calculated. For the present case, the system comprises nz × nθ = 18 × 118 = 2124
pressure nodes providing 1972 pressure states (having eliminated prescribed nodes),
a state space foil model with 2nθ = 236 states and a state space rotor model with 4
states. The natural frequencies and damping ratios of the resulting 2212 eigenvalues
are plotted as functions of rotational velocity in fig. P5.6. Figure P5.6(a) reveals a
high number of mostly highly damped modes to be scattered sporadically over the
range 0–9.3 kHz with a single mode becoming unstable (marked by red stars) after
24.5 kRPM. From the corresponding damping ratios plotted in fig. P5.6(b), two
modes with damping ratios between -0.05 and 0.22 are seen to be clearly separated
from the rest. Isolating these two modes, the Campbell diagram shown in fig. P5.7
can be extracted. This filtering technique is simple and works well for relatively
uncomplicated systems such as the one presently studied. For more complex
systems or systems containing other lowly damped modes, a filtering technique
based on eigenvector structures might be necessary as discussed by Bonello [15,
16] when he introduced the Jacobian based Campbell diagrams. As expected, the
two modes remaining in the Campbell diagram represent backward and forward
whirl and the OSI is seen to be dictated by the forward mode becoming unstable
at 24.34 kRPM with a frequency of 103.1 Hz (found by interpolation between the
eigenvalues evaluated at steps of 10 RPM). This matches the result from the exten-
ded perturbation exactly. Notice from the Campbell diagram, that the backward
mode is apparently changing into a forward mode at 8 kRPM as also found in [15].
This is slightly misleading, as the backward orbit is actually collapsing into a line
(oriented at approximately θ = 15◦ ) with forward and backward amplitudes of
equal magnitude. Similar degenerated orbits was observed, though for a different
GFB configuration, as an intermediate stage for modes transitioning from forward
to backward whirl in [16]. For such a degenerated orbit, the notion of forward and
backward loses its meaning and the difference is only due to numerical differences
between the two components’ amplitudes as reported in appendix P5.D. Notice
that the nominal case resembles closely the case treated in [15] and that the results
agree well despite the slightly higher foil stiffness (4.739 GN/m3 ), exclusion of foil
mass, finite difference based RE discretization and rotor-synchronous foil damping
used in [15]. Good agreement is observed from the locus and Campbell diagrams
and the OSI predicted at 22.5 kRPM in [15] is comparable to the present result.
P5.5 Results 161
2π
−1.0
ε=1
xf 0
−0.5
ωi t (rad)
εx (–)
0.0
π
0.5
1.0
0
−1.0 −0.5 0.0 0.5 1.0
εy (–)
(a)
2π
−1.0
ε=1
xf 0
−0.5
ωi t (rad)
εx (–)
0.0
π
0.5
1.0
0
−1.0 −0.5 0.0 0.5 1.0
εy (–)
(b)
Figure P5.4: Static deflection and mode shapes at 24.35 kRPM resulting from the
extended perturbation: (a) Forward mode with ωi = 103.2 Hz and ζ = −3.3×10−4 ;
(b) backward mode with ωi = 130.8 Hz and ζ = 0.18. Notice that the decay factor
eRe(λi )t has been omitted from the mode visualizations.
The unfiltered eigenvalue map shown in fig. P5.6, on the contrary, looks different
from the result in [15] due to the high number of heavily damped high frequency
modes stemming from the foil inertia.
For the coupled model, each eigenvector vi contains information about the
oscillation of all three domains, meaning that the mode shapes can be evaluated as
n oT
z = p̃T zTf zrT = z0 + vi eiωi t + v̄i e−iωi t . (P5.56)
To plot these mode shapes, the full eigenvector is initially normalized to provide a
maximum displacement amplitude of C/5 (based on the rows representing foil and
rotor displacements). The resulting normalized pressure mode shape represents
a two-dimensional oscillating field which is awkward to visualize, hence the rows
representing the pressure along the bearing mid-plane are extracted. The mode
shape associated with the forward mode becoming unstable is thus plotted just
after the OSI in fig. P5.8. This shows (to the left) a rotor orbit and foil oscillation
very similar to those from the extended perturbation in fig. P5.4(a) and, to the
right, the corresponding oscillation of the mid-plane pressure field. The mode
shape associated with the backward mode also present in the Campbell diagram
is plotted, likewise at 24.35 kRPM, in fig. P5.9. In this mode shape, the foil is
closer to being in-phase along the entire circumference, but a phase difference is
still present. This is very similar to the corresponding mode extracted from the
extended perturbation in fig. P5.4(b), though it should be emphasized that the
30
le
Unstab
0.6
25
Stable
Ω (kRPM)
0.7
εx (–)
20
0.8 15
ε=1
0.9 x̃r0 10
0.2 0.3 0.4 0.5 0.6

εy (–)
Figure P5.5: Locus of rotor equilibrium points for a nominal foil stiffness
(k = 4.6417 GN/m3 ) obtained from the coupled model.
P5.5 Results 163
1.0
10000
Unstable mode
Stable mode 0.8
8000
0.6
ωi (Hz)
6000
ζ (–)
4000 0.4
2000 0.2
0 0.0
10 15 20 25 30
Ω (kRPM)
(a)
1.0
8000
0.8
6000
0.6
ωi (Hz)
ζ (–)
0.4 4000
0.2
2000
0.0
0
10 15 20 25 30
Ω (kRPM)
(b)
Figure P5.6: (a) Natural frequencies and (b) damping ratios of the system
Jacobian’s 2212 eigenvalues evaluated at the stationary points found at 8–30 kRPM
for a nominal foil stiffness (k = 4.6417 GN/m3 ).
P5 Multi-domain Stability and Modal Analysis Applied to Gas Foil...
175
150
125
ωi (Hz)
100
75
50 Forward 1X
Backward 0.5X
25
Unstable
0
10 15 20 25 30
Ω (kRPM)
Figure P5.7: Campbell diagram extracted from the full set of eigenvalues shown in fig. P5.6(a) based on damping
ratios. The stability limit is ΩOSI = 24.34 kRPM.
164
P5.5 Results 165
coupled model modes are calculated with the foil damping tuned to the leading
mode OSI frequency of 103.1 Hz. For the 131.3 Hz mode in fig. P5.9, the foil
damping applied in the coupled model will thus be approximately 30 % higher
than for the corresponding mode from the extended perturbation in fig. P5.4(b),
possibly explaining the slightly increased damping ratio of ζ = 0.20 predicted by
the coupled model compared to ζ = 0.18 from the extended perturbation. To allow
direct comparison of other modes than the leading one, the iterative tuning process
would either have to be repeated for each mode or a damping model like the one
presented in [146] should be used instead. In all cases, it should be highlighted
that the rotor orbits and foil deflection shapes related to both the forward and
backward modes plotted to the left in figs. P5.8 and P5.9 are very similar to those
presented in [15].
It can be seen from fig. P5.6(b) that the lowest damped mode not included
in the Campbell diagram (for the nominal foil stiffness) has a damping ratio of
around 0.7, meaning that this mode would contribute marginally to the system
response. Furthermore, as seen from the mode shape plotted in fig. P5.10, the
mode is dominated by foil–pressure interaction with virtually no rotor movement.
Measuring such a mode experimentally would hence be very difficult and it is
reasonable not to include such modes in the Campbell diagram. It should be
noted, however, that the damping ratios of the foil dominated modes depend
significantly on the foil parameters. If, e.g., the foil stiffness is lowered to 1/10 of
the nominal value, the damping ratio of the shown foil mode drops to about 0.4,
so it is plausible that a configuration could be found where the foil modes become
significant.
P5.5.4 Comparison of OSIs obtained from the three approaches

A main point of interest for the present work is to investigate the discrepancies
in predicted OSIs between the three methods at different compliance levels. Varying
the SEFM stiffness between 1/10 and 1000 times the nominal stiffness without
changing the bearing load, the OSI predicted by the coupled model along with
the corresponding static eccentricity and maximum foil compliance is given in
table P5.2. At the nominal foil stiffness, the static eccentricity is 0.7140 and the
maximum static foil deformation is about 0.2C, indicating a bearing with limited
influence from the compliance. Increasing the stiffness to 1000k 0 , the bearing
behaves essentially rigid, as it can also be seen from the mode shape of the unstable
mode shown in fig. P5.11. Decreasing the foil stiffness, the significance of the
compliance increases until the foil deformation contributes more than 90 % of the
clearance at 1/10k0 . The investigated range of foil stiffness values hence covers the
full spectrum from rigid gas bearings to modern heavily loaded industrial GFBs.
The OSIs predicted using the coupled system Jacobian eigenvalues and the
160 2π
−1.0
xf p
xf 0 p0
−0.5 140
ε=1 pa
Pressure (kPa)
ωi t (rad)
εx (–)
0.0 120
π
0.5
100
1.0
80
0
−1.0 −0.5 0.0 0.5 1.0 0 1 2 3 4 5 6
εy (–) θ (rad)
Figure P5.8: Mode with ωi = 103.2 Hz and ζ = −2.9×10−4 at 24.35 kRPM extracted from the full system Jacobian
using a nominal foil stiffness (k = 4.6417 GN/m3 ) where ΩOSI = 24.34 kRPM. The foil oscillation and rotor movement
(forward whirl) is shown to the left while the pressure field oscillation at the bearing centre plane is shown to the
right. The mode is scaled to provide a maximum displacement amplitude (ε x ) of C/5 and plotted over a full period
with the phase indicated by the colour (dark towards light) Notice that the decay factor eRe(λi )t has been omitted
166
2π
P5.5 Results
xf 160
−1.0
p
xf 0 p0
−0.5 ε=1 140 pa
0.0
120 π
εx (–)
ωi t (rad)
0.5
Pressure (kPa)
100
1.0
80
0
−1.0 −0.5 0.0 0.5 1.0 0 1 2 3 4 5 6
εy (–) θ (rad)
Figure P5.9: Mode with ωi = 131.3 Hz and ζ = 0.20 at 24.35 kRPM extracted from the full system Jacobian using
a nominal foil stiffness (k = 4.6417 GN/m3 ) where ΩOSI = 24.34 kRPM. The foil oscillation and rotor movement
(backward whirl) is shown to the left while the pressure field oscillation at the bearing centre plane is shown to the
with the phase indicated by the colour (dark towards light). Notice that the decay factor eRe(λi )t has been omitted
167
2π
−1.0
xf 140 p
xf 0 p0
−0.5 ε=1 pa
120
Pressure (kPa)
ωi t (rad)
εx (–)
0.0
100 π
0.5
80
1.0
60
0
−1.0 −0.5 0.0 0.5 1.0 0 1 2 3 4 5 6
εy (–) θ (rad)
Figure P5.10: Mode with ωi = 410.9 Hz and ζ = 0.73 at 24.35 kRPM extracted from the full system Jacobian using
a nominal foil stiffness (k = 4.6417 GN/m3 ) where ΩOSI = 24.34 kRPM. The foil oscillation and rotor movement (a
forward whirl of vanishing amplitude) is shown to the left while the pressure field oscillation at the bearing centre
plane is shown to the right. The mode is scaled to provide a maximum displacement amplitude (the foil deflection at
261.5◦ ) of C/5 and plotted over a full period with the phase indicated by the colour (dark towards light)). Notice
that the decay factor eRe(λi )t has been omitted from the mode visualizations.
168
2π
P5.5 Results
−1.0
xf p
160
xf 0 p0
−0.5 ε=1 pa
140
0.0 120 π
εx (–)
ωi t (rad)
Pressure (kPa)
0.5 100
80
1.0
0
−1.0 −0.5 0.0 0.5 1.0 0 1 2 3 4 5 6
εy (–) θ (rad)
Figure P5.11: Mode with ωi = 109.9 Hz and ζ = −2.3×10−4 at 18.82 kRPM extracted from the full system Jacobian
with the nominal foil stiffness multiplied by 1000 (k = 4641.7 GN/m3 ) giving ΩOSI = 18.81 kRPM. The foil oscillation
and rotor movement (forward whirl) is shown to the left while the pressure field oscillation at the bearing centre
plane is shown to the right. The mode is scaled to provide a maximum displacement amplitude (ε x ) of C/5 and
plotted over a full period with the phase indicated by the colour (dark towards light)). Notice that the decay factor
eRe(λi )t has been omitted from the mode visualizations.
169
Table P5.2: Eccentricity and maximum foil deformation at the OSI from the
fully coupled model.
Foil
Factor stiffness k ΩOSI εx εy ε max (hc ) θ max(hc )
GN/m3 (kRPM) (−) (−) (−) (−) (deg)
1000 4641.7 18.81 0.4747 0.3844 0.6108 0.0002321 200.0
100 464.17 18.88 0.4758 0.3839 0.6113 0.002315 200.0
1 4.6417 24.34 0.6088 0.3730 0.7140 0.2001 196.9
1/2 2.3209 27.36 0.7606 0.3914 0.8554 0.3835 196.9
1/4 1.1604 34.18 1.0496 0.4394 1.1379 0.7394 193.8
1/8 0.580 22 57.32 1.5573 0.5430 1.6492 1.4332 190.8
1/10 0.464 17 74.59 1.7875 0.5949 1.8839 1.7768 190.8
whirling frequencies of the modes becoming unstable are compared to the per-
turbation results in table P5.3. As evident from the calculated deviations shown
in table P5.4 and the comparative plots in fig. P5.12, the extended perturbation
OSIs match the coupled system OSIs for all compliance levels, while the classical
two-DOF perturbation matches only for the rigid bearing and underestimates the
OSI by ≈ 13 % for the softest case.
P5.5.5 Influence of the Foil Oscillation Frequency on the OSI

The coupled domain OSIs are calculated iteratively in order for the foil oscil-
Table P5.3: Comparison of OSIs obtained from the coupled system Jacobian,
the classical perturbation and the extended perturbation.
Foil Jacobian Two-DOF Extended

stiffness with ω f ≈ ωs perturbation perturbation
Factor
k ΩOSI ωs ΩOSI ωs ΩOSI ωs
GN/m3 (kRPM) (Hz) (kRPM) (Hz) (kRPM) (Hz)
1000 4641.7 18.81 109.8 18.81 109.8 18.81 109.8
100 464.17 18.88 109.9 18.87 109.9 18.88 109.9
1 4.6417 24.34 103.1 23.62 102.7 24.34 103.1
1/2 2.3209 27.36 94.1 26.08 93.6 27.35 94.2
1/4 1.1604 34.18 80.9 31.59 80.6 34.18 80.9
1/8 0.580 22 57.32 66.0 50.57 65.6 57.33 66.1
1/10 0.464 17 74.59 61.2 65.06 60.9 74.60 61.2
P5.5 Results 171
Two-DOF
70
Extended
k0
60 Jacobian
ΩOSI (kRPM)
50
40
30
20
100 101 102 103

Stiffness k (GN3 /m)
(a)
Deviation from Jacobian ΩOSI (%)
−2
−4
−6
−8
−10 k0
Two-DOF
−12 Extended
100 101 102 103

Stiffness k (GN/m3 )
(b)
Figure P5.12: Comparison of obtained stability limits as a function of foil

stiffness: (a) the OSI predicted from the three methods; (b) the deviation between
the coupled system OSIs and the perturbation method OSIs.
Table P5.4: Deviation of perturbation OSIs from the coupled system Jacobian
OSIs.
Foil stiffness Deviation from Jacobian ΩOSI (%)
Factor
k GN/m3 Two-DOF perturbation Extended perturbation

1000 4641.7 −0.0011 0

100 464.17 −0.033 −0.0011
1 4.6417 −3.0 −0.000 82
1/2 2.3209 −4.7 −0.033
1/4 1.1604 −7.6 0.012
1/8 0.580 22 −12 0.018
1/10 0.464 17 −13 0.021
lation frequency ω f to match the resulting whirling frequency (±0.001 Hz). This
differs from the commonly applied assumption of rotor-synchronous foil damping,
but is essential for obtaining OSIs in agreement with the perturbation methods.
This is evident from the results in table P5.5, where the coupled system OSIs
previously shown in table P5.3 are compared to corresponding OSIs calculated
without iteration using rotor-synchronous foil damping, i.e. by setting the foil
oscillation frequency equal to the rotational speed. For the rigid bearing, where the
foil has a vanishing influence on the system dynamics, the difference is negligible,
but the importance increases with the foil compliance level leading to a 51 %
Table P5.5: Change in predicted OSIs from the full system Jacobian when
assuming rotor-synchronous viscous foil damping, i.e. by setting ω f = Ω. The
results with ω f = ωs ±0.001 Hz (found iteratively) correspond to those in table P5.3.
Foil ωf =
ωf = Ω Deviation
stiffness ωs ± 0.001 Hz
Factor
GN/m3 (kRPM) (Hz) (kRPM) (Hz) (%) (%)
1000 4641.7 18.81 109.8 18.81 109.8 −0.0043 −0.0091
100 464.17 18.88 109.9 18.84 109.9 −0.19 −0.045
1 4.6417 24.34 103.1 21.81 99.7 −10 −3.3
1/2 2.3209 27.36 94.1 22.75 90.2 −17 −4.2
1/4 1.1604 34.18 80.9 24.80 77.0 −27 −4.8
1/8 0.580 22 57.32 66.0 31.63 61.9 −45 −6.2
1/10 0.464 17 74.59 61.2 36.46 57.2 −51 −6.5
P5.6 Conclusion 173
underestimation of the OSI for 1/10k 0 .
P5.5.6 Influence of the Foil Inertia on the OSI

Lastly, it should be noted that the coupled system OSIs are based on the system
formulation from eq. (P5.48) including the foil mass. The foil mass is likewise
taken into account by the extended perturbation, but not by the classical two-DOF
perturbation, meaning that any influence on the OSI by the foil mass would assign
the extended perturbation method an unintended competitive advantage. To make
sure this is not the case, the coupled system OSIs have additionally been calculated
using the massless, but otherwise equivalent, foil model from eq. (P5.14). The
results are compared to the previously shown OSIs from table P5.3 (which are
calculated with foil mass) in table P5.6, revealing the influence of the foil mass on
the OSI to be negligible.
P5.6 Conclusion
Compared to previous work by the authors [88, 128], the discrepancies between
the OSIs calculated from the classical perturbation method and those from a full
non-linear model are smaller and with opposite sign. This is caused by three model
changes:
• The coupled model OSI is now found with a foil frequency matching the
frequency of the mode becoming unstable (see table P5.5). This requires
an iterative calculation of the OSI, but provides consistency with the foil
Table P5.6: Change in predicted OSIs from the full system Jacobian when
removing the foil mass i.e. by degrading eq. (P5.17) into eq. (P5.14). The results
with foil mass correspond to those in table P5.3.
Foil SEFM SEFM

Deviation
stiffness with foil mass without foil mass
Factor
GN/m3 (kRPM) (Hz) (kRPM) (Hz) (%) (%)
1000 4641.7 18.81 109.8 18.81 109.8 0 0
100 464.17 18.88 109.9 18.88 109.9 0 0
1 4.6417 24.34 103.1 24.34 103.1 −0.0012 −0.000 97
1/2 2.3209 27.36 94.1 27.35 94.1 −0.049 −0.0014
1/4 1.1604 34.18 80.9 34.17 80.9 −0.022 0.0098
1/8 0.580 22 57.32 66.0 57.29 66.1 −0.048 0.016
1/10 0.464 17 74.59 61.2 74.55 61.3 −0.058 0.017
oscillation frequency employed in the first order equations. The same issue
was treated in a recent publication by Pronobis and Liebich [146], but
here consistency was achieved while retaining the rotor-synchronous viscous
damping by setting ω f = Ω in eqs. (P5.11) and (P5.13), effectively providing
a frequency dependent hysteretic damping.
• The Gümbel condition is enforced on the dynamic pressure field at the

integration stage, as for the static and time domain pressures, rather than
through constraints on the solution (as described in [93]).
• The dynamic compliance, i.e. the foil deflection due to the dynamic pres-
sure field, is now based on the axially averaged dynamic pressure. This is
consistent with the calculation of the static foil compliance from the axially
averaged static pressure and, importantly, with the foil formulation used
in the coupled time domain model where no distinction is made between
dynamic and static pressures.
For a very stiff bearing, the classical perturbation method has thus been shown
to provide exactly the same OSI as the coupled model. This was to be expected
since the method was introduced for use with rigid bearings. For the worst
case of a very soft bearing, the perturbation leads to an underestimation of the
stability limit by 13 % for the present benchmark rotor–bearing system. These
findings are comparable to other recently published results [146], where the classical
perturbation is found to predict OSIs in good, but declining, agreement to a time
domain model up to a compliance level of around C/3.
In the present work, it has been demonstrated that this discrepancy can be
eliminated entirely for all compliance levels by extending the perturbation to
include also the foil states and their governing equations of motion, similarly to
the treatment of pads in stability analysis for tilting pad journal bearings. In its
present format, the extended perturbation requires the inclusion of foil mass, but
this has been verified as not affecting the OSI by comparing with OSIs from a
coupled model incorporating a massless SEFM (see table P5.6). The deficiency
of the classical perturbation is hence not rooted in the omitted foil mass, but
fundamentally in the elimination of the independent foil states through the implicit
compliance description introduced by Peng and Carpino [132] in 1993. Though
evidently problematic to the OSI, the motivation for this simplification is easy to
understand considering the available computational resources at the time and it
should be highlighted that the efforts by Peng and Carpino [132] have enabled
GFB analysis for almost three decades.
From the presented benchmark results as well as the results from [146], it could
seem that the two-DOF perturbation approach could still be a useful simplification
for low compliance levels, but it is not possible to quantify the influence of other
P5.A Jacobian Matrices 175
bearing design parameters or the bearing geometry on the OSI discrepancy, and
thus whether smaller or larger OSI discrepancies are to be expected for more
realistic rotor–bearing systems. Furthermore, it has not been investigated how the
issue might influence other predicted characteristics, such as the unbalance response.
It can, however, be concluded that the elimination of foil states is fundamentally
erroneous for stability analysis as the pressure–compliance interaction is neglected.
A possible remedy is to apply the presented extended perturbation method, but
this is rather elaborate, computationally expensive and difficult to reconcile with
more realistic foil models. Furthermore, a main virtue of the perturbation approach
is lost, as the computational effort increases drastically due to the higher number
of first order equations and subsequent much larger eigenvalue problems. Instead,
stability analysis for GFB supported rotor–bearing systems should be based on
coupled equation systems, allowing stability assessment either through brute-force
time-integration or, ideally, based on Jacobian eigenvalues.
P5.A Jacobian Matrices

P5.A.1 Reynolds Equation Finite Element Residual Derivatives
As the element boundaries are fixed, Leibniz’s rule allows the partial derivative of
the RE FE residual from eq. (P5.3) with respect to the pressure to be written (on
the element level) as
∂r̃RE
e ∂ARE
e p̃e, x̃r , x̃ f p̃e ∂BRE
e Û̃ xÛ̃ r , xÛ̃ f p˜e e p̃, x̃r , x̃ f pÛ̃ e

p, CRE
= + + , (P5.A1)
∂p̃ ∂p̃ ∂p̃ ∂p̃
where
∂ARE
e p̃e ∫
= T
B − p̃ h̃3 B − h̃3 Bp̃e N
∂p̃ Ae |{z} | {z }
I II
(P5.A2)
e ∂ h̃

1 1
+ Sτ Ω̃ 2
N − 3 p̃ h̃ B p̃ + Sτ Ω̃ h̃ N dA,
0 | {z } ∂p̃ 0
| {z } IV | {z }
III V
∂BRE
e p̃e ∫
= −2Sτ Û̃ T N dA,
hN (P5.A3)
∂p̃ Ae | {z }
VI
∂CRE
e pÛ̃ e ∂ h̃
∫
= −2Sτ NT NpÛ̃ e dA, (P5.A4)
∂p̃ Ae ∂p̃
| {z }
VII
in which terms I, II, V and VI are non-zero at the element DOFs only, i.e. they
are 4 × 4 matrices corresponding to ∂/∂p̃e , while terms III, IV and VII (if present)
provide element Jacobian contributions to pressure DOFs not included in p̃e due
to the axial averaging of pressures.
For an implicit film height description where no x̃ f is present, compliance is
a function of pressure directly, thus giving rise to terms III, IV and VII. For the
present implementation where compliance is based on the axially averaged pressure
(referring to eq. (P5.12)), the compliant film heights at the element nodal positions
resulting from a pressure field p̃ can be written as
Í nz
p̃ θ 1e, z̃ j ∆z̃ j 

fhc θ 1e, p̃ 
 
   j=0 
. .
   
h̃c ( p̃) = . . = Hehc p p̃, (P5.A5)
e

 
 −1 −1

 

. ≈ L̃ k̃ c .
 fhc θ e, p̃   nz p̃ θ e, z̃ j ∆z̃ j 

  
 Í 
 
 4   j=0 4 
where Hehc p has dimension 4 × nθ nz , but contains non-zero contributions only from
the 2nz pressure DOFs affecting the film height in a given element. The film height
derivative with respect to pressure required in eqs. (P5.A2) and (P5.A4) can be
found using the shape functions as
∂ h̃ ∂ h̃c ∂ h̃ e
= = N c = NHehc p ∈ R1×nθ nz , (P5.A6)
∂p̃ ∂p̃ ∂p̃
meaning that terms III, IV and VII result in 4 × nθ nz matrices with non-zero
contributions to the four element residuals from the 2nz pressure DOFs affecting
the compliant height within the element.
For the coupled system approach, the compliant film height is given explicitly
from the foil states x̃ f implying ∂ h̃/∂p̃ = 0 and thus eliminating terms III, IV and
VII. In return, the derivatives with respect to the rotor and foil DOF vectors x̃r
and x̃ f are required. These can be written as
∂r̃RE
e ∂ARE
e p̃e e p
CRE Û̃ e
= +
∂x̃r ∂x̃r ∂x̃r
e ∂ h̃ ∂ h̃
∫ ∫
1
= B −3 p̃ h̃ B + Sτ Ω̃
T 2
N p̃ dA − 2Sτ NT NpÛ̃ e dA
Ae 0 ∂x̃r Ae ∂x̃r
(P5.A7)
∂r̃RE
e ∂Ae p̃e Ce pÛ̃ e
= RE + RE
∂x̃ f ∂x̃ f ∂x̃ f
e ∂ h̃ ∂ h̃
∫ ∫
1
= B −3 p̃ h̃ B + Sτ Ω̃
T 2
N p̃ dA − 2Sτ NT NpÛ̃ e dA
Ae 0 ∂x̃ f Ae ∂x̃ f
(P5.A8)
P5.B Foil Structure Mass 177
where ∂ h̃/∂x̃r is simply [cos θ sin θ], while ∂ h̃/∂x̃ f is a 1 × nθ row vector with
two non-zero elements corresponding to the two foil states coinciding with the
circumferential positions of the element’s pressure nodes. Finally, the residual
formulation requires also the partial derivatives with respect to the state vector
temporal derivative which are given as
∂r̃RE ∂CRE Û̃ e ∂Be p̃e ∂ hÛ̃

e e p ∫
= + RE
= −2Sτ NT h̃N + NT Np̃e dA, (P5.A9)
∂ pÛ̃ ∂ pÛ̃ ∂ pÛ̃ Ae ∂ pÛ̃
∂r̃RE ∂BRE p̃ ∂ hÛ̃
e ∫
= = −2Sτ NT Np̃e dA, (P5.A10)
∂ xÛ̃ r ∂ xÛ̃ r Ae ∂ xÛ̃ r
∂r̃RE ∂ hÛ̃
e
∂BRE p̃
∫
= = −2Sτ NT Np̃e dA, (P5.A11)
∂ xÛ̃ f ∂ xÛ̃ f Ae ∂ xÛ̃ f
Û̃ xÛ̃ , ∂ h/∂
where ∂ h/∂ Û̃ xÛ̃ and ∂ h/∂
Û̃ pÛ̃ are found analogously to ∂ h̃/∂x̃ , ∂ h̃/∂x̃
r f r f
and ∂ h̃/∂p̃. The leftmost term of eq. (P5.A9) is non-zero only at the columns
corresponding to the element DOFs.
Notice that in previous work by the authors [93], only terms I and V of
eq. (P5.A2) were included in the tangent matrix (yielding an inexact NR method)
and that the squeeze and local expansion derivative terms VI and VII are present
only for transient conditions. Furthermore, it should be mentioned that term VII
from eq. (P5.A4) and the rightmost term in eq. (P5.A9) are only present if an
implicit compliance formulation is used in a transient simulation. This combination
is not applied in the present work, but the terms are included for completeness.
P5.B Foil Structure Mass

Assuming the top foil shares the bump foil density and thickness, a rough estimate
of the collective foil structure’s average mass per unit circumferential length can
be obtained as
m0f = (2Rb θ b + Sb − 2l0 ) Ltb ρb Sb−1 + tb L ρb , (P5.B1)
| {z } |{z}
Bump foil Top foil
hb2 +l02
where the bump radius of curvature is given as Rb = 2l0 and the bump half
angle as θ b = arccos Rb .
Rb −h0
For the properties given in table P5.1 this gives
m0f = 65.44×10−3 kg/m.
P5.C Extended Perturbation Without Foil Mass

Even though this is believed by the authors to present a less intuitive approach,
it would be possible to formulate the extended perturbation equations of motion
from eq. (P5.41) without including the foil mass. Recasting the rotor equations
into state space form, these could be coupled to the first order (massless) rotor
equations as
 0 I 0  0 0 0 Ü̃ 
 xr 
  
ª 
+
©     

 M̃r 0 0  0  ® xÛ̃ r
 0 D̃g,xDOF ω̃s, Ω̃  

0 D̃ f (ω̃s ) 
  
 Û̃ 
« 0 ¬  w  (P5.C1)
 

 −I 0 0   0 0 0 Û̃ 
 xr 
 
ª 
+  0 0 0  +  0  ® x̃r = 0,
©   

K̃g,xDOF ω̃s, Ω̃  
« 0
 0 K̃ f   0  ¬  w̃ 
  
leading to an eigenvalue problem requiring a slightly different treatment than

currently described in section P5.3.3. In brief, the characteristic polynomial
function from eq. (P5.43) would change into
 −I λI 0 
Pc (λ) = det  λM̃r 0
©  
0 
0 λD̃ f (Im(λ)) + K̃ f 

« 0
(P5.C2)

 0 0 0 
 ª® ,
 
0
λD̃g,xDOF Im(λ), Ω̃ + K̃g,xDOF Im(λ), Ω̃ 



 0 ¬
and the structure of the obtained eigenvectors would change. This approach has
not been further pursued by the authors, but it would possibly have allowed similar
results to be obtained from the extended perturbation without including the foil
mass.
P5.D Mode Shape with Degenerated Orbit

Reading from the Campbell diagram in fig. P5.7, the second mode is starting out
as a forward whirl at 8 kRPM, but reversing into a backward whirl for ≥ 8.5 kRPM.
This is slightly misleading, as the orbit is simply collapsing into a straight line at
an angle of approximately θ = 15◦ as shown in fig. P5.D.13. For such a mode, the
backward and forward components of the rotor movement are of equal magnitude,
meaning that the categorization as a forward mode is merely due to numerical
differences between the calculated forward and backward magnitudes. Deciding
on a numerical tolerance, one could hence have introduced a third category for
modes showing this kind of degenerated orbit in fig. P5.7.
2π
−1.0
xf 180
p
xf 0 p0
160
−0.5
ε=1 pa
140
0.0
π
εx (–)
120
ωi t (rad)
0.5
Pressure (kPa)
100
1.0
P5.D Mode Shape with Degenerated Orbit
80
0
−1 0 1 0 1 2 3 4 5 6
εy (–) θ (rad)
Figure P5.D.13: Mode with ωi = 158.6 Hz and ζ = 0.2041 at 8 kRPM extracted from the full system Jacobian
with the nominal foil stiffness k = 4.6417 GN/m3 . The foil oscillation and rotor movement (an orbit degenerated into
a straight line) is shown to the left while the pressure field oscillation at the bearing centre plane is shown to the
with the phase indicated by the colour (dark towards light)). Notice that the decay factor eRe(λi )t has been omitted
179
References
PUBLICATION P6
Gas Foil Bearings with Radial
Injection: Multi-domain
Stability Analysis and
Unbalance Response
This chapter is a preprint of the identically titled article [130] submitted to the
Journal of Sound and Vibration on 7 April 2020.
Sebastian von Osmanski, Ilmar F. Santos*

Abstract
Gas Foil Bearings (GFBs) possess significant potential for oil-free and environ-
mentally friendly support of high-speed rotating machinery in a wide range of
applications. Advances in GFB design and solid lubricants have significantly
improved the capabilities of GFBs, but application in heavier machinery still
leaves many technical limitations to be overcome. A possible path is through
hybridization with radial gas injection as commonly applied in rigid gas journal
bearings. From a modelling perspective, modifications of the well-documented
Reynolds Equation have previously been proposed to include injection, but the
mass flow itself is difficult to predict. In the literature, this is often described
using idealized analytical expressions with carefully defined/calculated/identified
correction factors. In the present work, a parameter-map approach is presented
where the injection flow is interpolated from a five-dimensional pre-generated map
of mass flows calculated using a CFD sub-model for the injector valve. This allows
injection to be added into a fully coupled simultaneous GFB simulation framework.
Non-controlled hybrid mode simulations are presented for a yet-to-be-built version
182 P6 Gas Foil Bearings with Radial Injection: Multi-domain Stability...
of an existing industrial scale GFB test rig augmented with an existing injection
system. The presented results comprise multi-domain mode shape visualizations,
Campbell diagrams, and unbalance responses indicating significant improvements
attainable already from the hybrid mode. The addition of actively controlled
injection is thus considered feasible and may allow hybrid GFBs to support heavier
rotors than permissible by using the currently available passive variants.
BC Boundary Condition A, B Bearing designations

CG Center of Gravity ai,e, j j-th FD pressure coefficient for
CDS Central Differencing Scheme eastern face of i-th CV
CFD Computational Fluid bi,e, j j-th FD pressure gradient
Dynamics coefficient for eastern face of
CV Control Volume i-th CV
CVC Control Volume Centre C Radial clearance
CVE Control Volume Edge/face Cp Constant pressure specific heat
DOF Degree of Freedom capacity
FD Finite Difference c Speed of sound
FV Finite Volume di Beam element nodal
GFB Gas Foil Bearing displacement
HGFB Hybrid Gas Foil Bearing do Generic orifice diameter
LUDS Linear Upwind Differencing Et Young’s modulus of top foil
Scheme material
MRE Modified Reynolds Equation eαx, eαy GFB eccentricity components
OSI Onset Speed of Instability g Gravitational acceleration
RPM Revolutions per Minute h, h̃ Film height, h̃ = hC −1
SEFM Simple Elastic Foundation
hc, h̃c Film height (compliant),
Model
h̃c = hc C −1
(Û) Time derivative, d/dτ or d/dt
hent Specific enthalpy
(Ü) Time derivative, d 2 /dτ 2 or
hfilm Injector film height
d 2 /dt 2
∇· Divergence hi , h̃i Film height at i-th CV centre,
(˜) Non-dimensional quantity h̃i = hi C −1
( )e Quantity at element level hinj Injector valve clearance
( )0 Stagnant/upstream quantity hr , h̃r Film height (rigid), h̃r = hr C −1
∩ Intersection hs, h̃s Slope height, h̃s = hs C −1
A Area It Top foil area moment of inertia
Ai In-film area of the i-th CV Izz Rotor polar moment of inertia
I x x, I yy Rotor transverse moment of pinj Injection pressure

inertia qinj Injection source field
k Bump stiffness per unit area qθ Circumferential fluid film flow
L, L̃ Bearing length, L̃ = LR−1 qp Top foil pressure load per unit
l1, l2 Distance from CG to bearings length
l12 Bearing distance l12 = l1 + l2 qz Axial fluid film flow
l3, l4 Distance from CG to discs R Journal radius
lb,i Inner radius injector bore Rs Specific gas constant
length rFV,i , FV equation residual for i-th
lb,o Outer radius injector bore r̃FV,i CV
length Si In-film circumference of the
le Top foil element length i-th CV
lp,i Inner radius injector pin length sdp,i,e Constant FD pressure gradient
lp,o Outer radius injector pin contribution at eastern face of
length i-th CV
lrs Rotor domain length scale s p,i,e Constant FD pressure
M Mach number contribution at eastern face of
mÛ i Integrated injection flux for i-th CV
the i-th CV rb,i Inner injector bore radius
mÛ inj,i Mass flow through i-th injector rb,o Outer injector bore radius
mr Rotor assembly mass rfilm Film zone radius in injector
mrs Rotor mass scale sub-model
nCV Number of CVs (per bearing) rp,i Inner injector pin radius
nfe Number of top foil elements rp,o Outer injector pin radius
(per bearing) Sτ Non-dimensional group,
nfdof Number of foil DOFs (per Sτ = 6R2 µωτ C −2 p−1
a
bearing) Tiso Isothermal gas temperature
ninj Number of injectors (per t Physical time
bearing) tt Thickness of top foil
npad Number of GFB pads u Generic flow velocity
nrdof Number of rotor DOFs uα Bearing unbalance, mass ×
nθ Number of circumferential CVs eccentricity
nz Number of axial CVs wi i-th Gauss quadrature weight
p, p̃ Film pressure, p̃ = p−1
a p x, y, z Cartesian coordinates
pa Ambient pressure, fluid xf0 Static foil deformation
pressure scaling xf Foil deformation
pfilm CFD sub-model fluid film xr0 Static rotor displacement
pressure BC xr Rotor displacement
pi , p̃i Pressure at i-th CV centre, z, z̃ Axial gas film coordinate,
p̃i = p−1
a pi z̃ = zR−1
∆z̃ j Span associated with j-th CV θ ξi θ-coordinate coincident with

position Gauss point ξi
∆z̃v, j Span associated with j-th ξ Isoparametric coordinate
vertex position ξi i-th Gauss quadrature point
α Bearing index, α = A, B ζ Damping ratio
αs Injector shoulder angle ε x0 Constant contribution to ε x,α
∆θ j Span associated with j-th CV mapping
position ε x,α Equivalent injector ε x values
∆θ v, j Span associated with j-th in bearing α
vertex position fb , f̃b Rotor bearing force vector,
εαx , εαy GFB eccentricity ratios, f̃b = mrs
−1 l −1 ω−2 f
rs τ b
ε = eC −1 fb0 Constant contribution to fb
εx CFD sub-model eccentricity mapping
ratio fb,α Bearing α contribution to fb
η Hysteretic foil damping fG Coupled system residual
constant function
γ Specific heat capacity ratio fG,0 Constant part of fG
λi i-th eigenvalue fG,s Steady non-constant part of fG
µ Dynamic gas viscosity fG,t Unsteady part of fG
Ω, Ω̃ Angular velocity Ω̃ = ωτ−1 Ω finj (. . . ) Injection flow interpolation
ΩOSI Angular velocity at the OSI function
ω f , ω̃ f Foil oscillation frequency, f p,α Foil structure pressure loads
ω̃ f = ωτ−1 ω f f p0 Constant contribution to f̃ p,α
ωi i-th (damped) natural mapping
frequency frw , f̃rw Rotor static force vector,
ωτ Characteristic frequency f̃rw = mrs
−1 l −1 ω−2 f
rs τ rw
φi Beam element nodal rotation fub , f̃ub Rotor unbalance force vector,
ρ Density of gas f̃ub = mrs
−1 l −1 ω−2 f
rs τ ub
τ Dimensionless time, τ = ωτ t hinj,α Nozzle clearances in bearing α
θ Circumferential angle hc,CVC Compliant film heights at
θi i-th circumferential nodal CVCs
coordinate hc,CVE Compliant film heights at
θl Pad leading edge angle CVEs
θ min h Location of minimum film hr,CVC Rigid film heights at CVCs
height hr0,CVC Constant rigid film heights
θs Inlet slope extent (= C) at CVCs
θt Pad trailing edge angle hr,CVE Rigid film heights at CVEs
θ v, j j-th circumferential vertex hr0,CVE Constant rigid film heights
θ-coordinate (= C) at CVEs
Ûα
m CV injection mass flows BFV MRE FV matrix, squeeze term
Û inj,α
m Injector mass flows CFV MRE FV matrix, local
n Outward pointing unit normal expansion term
vector Df Foil structure damping
p AB ,p̃ AB CVC pressures for both DSEFM Foil structure SEFM damping
bearings, p̃ AB = p−1
a p AB Gr , G̃r Rotor gyroscopic matrix,
pα , p̃α CVC pressures, p̃α = p−1 a pα G̃r = mrs
−1 G
pinj,α Injection pressures in bearing r
α HCVC,x f Mapping from x̃ f ,α to hc,CVC

pfilm0 Constant contribution to HCVC,xr Mapping from x̃r to hr,CVC
pfilm,α mapping HCVE,x f Mapping from x̃ f ,α to hc,CVE
pfilm,α Injector film BC pressures in HCVE,xr Mapping from x̃r to hr,CVE
bearing α Hε,xr Mapping from x̃r to ε x,α
qg MRE flow vector Hε,x f Mapping from x̃ f ,α to ε x,α
r Coupled system residual H fb,p Mapping from p̃ AB to f̃b
rFV,α FV residual H fp,p Mapping from p̃α to f̃ p,α
r f ,α Foil structure residual
rq Element loads due to load q Hm̃inj,mÛ̃ Mapping from m Û̃ inj,α to m
Û̃ α
rr Rotor residual H pfilm,p Mapping from pα to pfilm,α
rSEFM Element loads due to the I Identity matrix
SEFM support Jz,s Jacobian of fG,s with respect to
vi i-th eigenvector z
xf Foil DOF vector Jz,t Jacobian of fG,t with respect to
x f ,α Foil structure DOF vector z
xr , x̃r Rotor DOF vector, x̃r = lrs −1 x
r
JzÛ Jacobian of fG,t with respect to
z Coupled system state vector
zÛ
z0 Static system equilibrium
AFV MRE FV matrix, Couette & Kf Foil structure stiffness
Poiseuille terms Kt Top foil element stiffness
B Shape function derivative Mr , M̃r Rotor mass, M̃r = mrs −1 M
r
matrix N Shape function matrix
P6.1 Introduction
Gas Foil Bearings (GFBs) are used in many lightly loaded high-speed compressors
and small turbines, taking advantage of an oil-free sustainable technology. The
application of GFBs in heavier rotating machines still requires many challenges to
be addressed and technical limitations to be overcome. These relate to the limited
load carrying capacity, wear during start–stops, and low damping at high rotor
speeds leading to instabilities and sub-synchronous vibrations due to non-linear
behaviour. Advances in solid lubrication and coatings along with optimizations

of the foil structure design have already significantly improved the operational
capabilities of passive GFBs. To overcome their remaining technical limitations and
facilitate the application in heavier rotating machines, passive [36, 40], regulated
[44, 182], and feedback controlled solutions [53] can be adopted. All technical
approaches to ”hybridization” have advantages and drawbacks, but a common
trend is an increase in complexity due to the added subsystems, regardless of them
being passive, regulated or feedback controlled. Hybridization is, nevertheless,
becoming advantageous or even necessary for further advances as the technology
evolves. One approach to GFB hybridization is through the addition of pressurized
injection. This approach has been used for several decades in externally pressurized
gas bearings with undeformable surfaces, or simply rigid gas bearings, as reported
in [12, 29, 48, 149, 152, 185].
Several control strategies exist to enhance fluid film bearing properties via
hybridization as elucidated in [160]. However, for gas bearings, Maamari et al. [112]
nicely condensed all strategies into two fundamental categories: i) gap geometry
control and ii) flow restriction control. In gap geometry control, the actuator force
is directly applied to move or deform the bearing surface leading to control of
the gap and hence of the aerodynamic pressure distribution. Examples of gap
geometry control in gas bearings have been presented by Al-Bender [13], Guan
et al. [53], Ro et al. [151], and Snoeys and Al-Bender [168] among others. Ro
et al. [151] use four iron core motors with permanent magnets to actively preload
a set of eight porous media air bearings. Al-Bender [13] investigated a novel
active gas bearing based on piezoelectrically controlled gap deformation or bearing
surface tilting. More recently, Guan et al. [53] mounted three pads of a GFB on
piezoelectric stack actuators to build actively controlled mechanical preloads. In
flow restriction control, the mass flow rate of the pressurized gas is controlled by
varying the restricting area of the injection supply without influencing the film
geometry. Morosi and Santos [117] and Pierart and Santos [142] designed and
tested a piezoelectrically controlled gas injector for a plane journal bearing. Using
four actuators coupled to adjustable cone restrictors, the gas flow rate into the
bearing gap could be actively controlled allowing a flexible rotor to safely cross its
first critical speed. Using a diaphragm valve to pneumatically control the supply
pressure, Ghodsiyeh et al. [49] also investigated a flow control system resulting
in a significant increase in static stiffness of the order of 40 %. Mizumoto et al.
[114] designed and tested a piezoelectric ”active inlet resistance” system for both
thrust and journal bearings. The piezoelectric actuation was implemented in a
high-speed spindle avoiding pneumatic hammer effects.
One of the first hybridization design solutions combining gas injection and
compliant surface bearings was from Eshel [41]. The authors presented an analysis
of a planar cylindrical hydrostatic foil bearing arrangement (resembling magnetic
tape running over a recording head) while considering two axial line sources. The
work is, however, of limited relevance for modern GFBs. Another example of such
a layout comprising a purely hydrostatic GFB is given by Carpino and Peng [24].
In this configuration, an incompressible lubricant is supplied from the journal
centre through capillary tubes to recessed pockets on the journal surface. With
the focus on a compliant thrust bearing, Bosley and Miller [20] designed another
setup where air is injected into the backside of the foil. Aiming at managing steady
state thermal effects, Radil and Batcho [147] and Shrestha et al. [167] carried out
experimental feasibility studies with air fed into the groove between leading and
trailing pad edges as an alternative to axial flow cooling. Aiming at adjusting static
and dynamic properties of a rotor–bearing system, a Hybrid Gas Foil Bearing
(HGFB) with injection through orifices attached to the top foil was designed by
Kim and Park [77]. Their prototype uses a compression spring type GFB from [170]
and four steel supply tubes attached to orifices in the top foil. A slight increase in
load carrying capacity is experimentally achieved using a relatively low injection
flow rate. Theoretically, the predictions indicated a significant improvement of
the Onset Speed of Instability (OSI). In [85] the authors investigated the bearing
force coefficients as a function of pressurized injection, and in [75], a three-pad
preloaded HGFB with three orifices is designed which is theoretically as well as
experimentally analysed in a number of papers. Experimental wear testing during
1000 start–stop cycles was carried out by Kim and Zimbru [78] with loads up to
445 N at 10 kRPM without visible wear for a 356 N load. In [179], the bearing force
dynamic coefficients were measured experimentally showing reasonable agreement
with theoretical predictions. In [182] the authors introduced a ”controlled hybrid
mode”, meaning that the lowermost injector (at 180◦ ) is shut off above 6 kRPM,
leaving pressurization, and thus static load generation on the rotor, only from
the topmost injectors at 60◦ and 300◦ . It is important to point out that a more
correct designation for such a mode of operation would hence be ”regulated”
rather than ”controlled”. Using this ”regulated hybrid mode”, the appearance of
sub-synchronous vibrations is delayed due to the static load contributed by the
pressurized topmost injector.
The most frequently applied analytical expression for gas injection is based on
the isentropic inflow assumption derived by assuming the flow to be steady and
inviscid with the flow rate being limited by compressibility effects. The isentropic
inflow assumption for injection in gas bearings can be traced to as early as 1964
where Lund [109] calculated the orifice mass flow in a hydrostatic gas journal
bearing using these formulas. More recently the same assumption has been used
by several authors, among these [6, 30, 35, 57, 60, 75, 77, 85, 97, 107, 120, 184].
It is important to recall that the flow is assumed to be restricted at a certain
choking area and that the flow through this cross section is considered uniform.
This requires the choking area and a correction factor — also known as ”discharge
coefficient”, ”inherent compensation factor”, ”adiabatic efficiency factor” or ”vena

contracta coefficient” — to be defined/calculated/identified, often depending on
the pressure ratio. Different expressions for the choking area are used in the
literature as this depends strongly on the injection system geometry. For an orifice
diameter do , Kim and Park [77] and Arghir et al. [6] use the immediate curtain area
in the fluid film defined as πdo h. Kim and Lee [75], Kumar and Kim [85], and Lee
and Kim [97] use the nominal curtain area πdoC, and Neves et al. [120] consider
the orifice area πdo2 /4. Delgado and Ertas [35], Hassini et al. [60], and Zhang et al.
[184] report more elaborate expressions for the choking area, taking into account
orifice as well as inherent restriction (=curtain) areas. The latter three models are
furthermore characterised by a separate model for the fluid film in the recess zone.
For example, Computational Fluid Dynamics (CFD) simulated mass flows are
calculated in [60] as functions of area relations and pressure ratios and compared
to mass flows predicted using the isentropic approach. It is concluded that an
almost constant discharge coefficient can be used if the restricting area is correctly
defined. Similarly, Neves et al. [120] compared isentropic injection mass flows
to CFD predictions for varying pressure ratios, concluding a constant correction
factor of 0.88 to be adequate for sonic/choked conditions, while a linear function
of the pressure ratio (ranging from 0.83 to 0.88) is used for subsonic conditions.
Multiple similar investigations can be found in the literature leading to varying
conclusions on the adequacy of the correction factors, substantiating the need
for separate numerical or experimental investigations of the correction factor to
be performed for each injector geometry. Alternatively, the use of the isentropic
injector model should be abandoned entirely in favour of more comprehensive fluid
models for the injector area alone or for the entire fluid film.
In this framework, the main original contribution of this article is to investigate

the non-linear and linearized lateral dynamic behaviour of a rigid rotor [90]
supported by HGFBs using a fully coupled model for rotor and GFBs based on
[127, 129] and including a more comprehensive model for the radial gas injection
based on CFD results and a mapping strategy. The procedure is general, relies on
knowledge about the injector geometry and is built aided by a systematic variation
of sensitive parameters affecting the pressure–flow relationship. The fully coupled
HGFB model — including regulated radial gas injection — is used to evaluate the
feasibility of HGFBs to increase the bearing load capacity and enhance the OSI.
The procedure furthermore serves to provide insights into the coupled behaviour
of the three domains, i.e. fluid, rotor, and foil structure, via multi-domain modal
analysis. This is the first attempt to build a more comprehensive and holistic
HGFB model suitable for coupling with feedback control laws for the piezoelectric
injectors, aiming at allowing heavier rotating machines to be supported on GFBs.
P6.2 Hybrid Bearing Configuration 189
P6.2 Hybrid Bearing Configuration

The current work is based on a test rig previously presented by the authors [87]
and subsequently analysed in several papers [88, 90, 91, 125]. It was designed
to mimic a 120 kW directly driven turbo blower with a nominal speed range of
15–30 kRPM and comprises a 21.17 kg hollow shaft with and approximate length
of 0.59 m. The shaft is near-symmetrically supported by two identically designed
GFBs with nominal diameter of 67 mm and length 53 mm. Sketches of the shaft
assembly and the GFBs including their nomenclatures are shown in fig. P6.1 and
additional details on the setup are provided in table P6.1.
The modelled injection system was originally designed for active control of a
rigid gas bearing with diameter and length of 40 mm and clearance of 25 µm carrying
approximately 40 N of static load from an overhung flexible rotor. The injection
system was first described in [117] and has later been subject to several theoretical
as well as experimental studies [141–143, 174]. In its original configuration, the
system comprises four adjustable cone restrictors located at 0◦ , 90◦ , 180◦ , and
270◦ each being controlled using a piezoelectric stack actuator as illustrated in
fig. P6.2(b). The cone restrictors are designed to allow nozzle clearances hinj in
the range of 0–45 µm while the air supply system is capable of delivering air at up
to 7 bar. Notice that the attainable nozzle clearances are comparable to the fluid
film clearance in both the originally targeted rigid bearing with C = 25 µm and
the investigated GFB with C = 40 µm. Depending on the valve setting and the
journal eccentricity relative to the nozzle position, the flow will thus be restricted
either by the cone section, the curtain area in the fluid film or by a combination
of these. The design is thus intermediate to an adjustable cone restrictor and an
inherent flow restrictor. A sketch of the injector geometry relevant to the model is
shown in fig. P6.3(a) and the dimensions are given in table P6.2.
The present work concerns the yet-to-be-constructed layout sketched in fig. P6.2(a)
with three such injectors added to the existing passive GFBs. This is similar to
the three-pad HGFB configuration presented in [75], but with the addition of
electromechanical actuators. The sketch is simplified, since an arrangement for
connecting the injector valve to the top foil would have to be added, possibly
in the shape of a short flexible pipe or hose. The use of three injectors instead
of four implies the vertical and horizontal directions to be inherently coupled,
thus significantly complicating a subsequent controller design. The configuration
is, however, interesting, since it would allow the proven and well-known passive
properties of the GFB to be maintained to the greatest extent possible. In terms
of practical application, reasonable passive properties would be an appreciable
advantage of a HGFB compared to other active bearing technologies such as active
magnetic bearings (for which separate backup bearings are further needed).
Disc B Disc A
Shaft Permanent magnets
Bearing B Bearing A
CG z
l2 l1
l4 l3
x
(a)
θ θl
θs
hs
h
y
ex
ey
Ω
θt
x
(b)
Figure P6.1: Schematics and nomenclature of the modelled shaft assembly and
the GFBs supporting it: (a) Shaft with discs, permanent magnets and bearings;
and (b) sketch of bearing geometry.
Table P6.1: Geometry, material properties and operating conditions of the GFB
test rig.
Shaft assembly
Bearing A to CG, l1 201.1 mm Mass, mr 21.17 kg
Bearing B to CG, l2 197.9 mm Polar moment of 0.030 08 kg m2
inertia, Izz
Disc A to CG, l3 287.2 mm Transverse moment 0.5252 kg m2
Disc B to CG, l4 304.0 mm of inertia, I x x = I yy
Radial clearance, C 40 µm Slope extent, θ s 30◦
Number of pads, npad 3 Slope height, hs 50 µm
Fluid properties
Viscosity, µ 1.95×10−5 Pa s Specific gas 28.71 J/(kg K)
Ambient pressure, pa 105 Pa constant, Rs
SEFM stiffness, k 8.8 GN/m3 Hysteretic foil 0.15
damping constant, η
Top foil thickness, tt 0.254 mm Foil oscillation Ω
frequency, ω f
Young’s modulus, Et 207.0 GPa Aerostatic case ω f 20 kRPM
Table P6.2: Geometry of the physical injector and the injector model as illustrated
in fig. P6.3.
Nozzle clearance, hinj 0–45 µm Inner bore radius, rb,i 1.0 mm

Outer bore radius, rb,o 3.0 mm Inner pin radius, rp,i 0.85 mm
Outer pin radius, rp,o 2.7 mm Pin length with inner 1.2 mm
radius, lp,i
Bore length with inner 1.37 mm Shoulder angle, αs 120◦
radius, lb,i
Bore length with outer 2.0 mm Radius of film zone, rfilm 2.5 mm
radius, lb,o
Rounding radius, bore–film 2.0 µm Rounding radius, 10 µm
edge other edges
Journal
Bearing housing
Bump foil
Top foil
(a)
Gas film
pressure Plastic pin
O-ring
Supply Belleville
pressure washers
Piezo
stack
actuator
(b)
Figure P6.2: Hybrid gas foil bearing configuration: (a) The existing three-pad
Siemens GFB augmented with injectors at the pad centres at 87.5◦ , 207.5◦ and
327.5◦ ; and (b) sketch showing the components and working principle of each
piezoelectric valve.
rb,i
rp,i
lb,i
lp,i
αs
hinj
rp,o
rb,o
(a)
2rfilm
RΩ
pfilm pfilm
lb,o
pinj pinj
(b)
Figure P6.3: Injector model geometry and boundary conditions: (a) Dimensions
of the rotationally symmetrical modelled part of the injector as given in table P6.2;
and (b) sketch of x y-plane cross section of the injector CFD model (hatched area)
where a part of the fluid film is included, thus breaking the rotational symmetry.
P6.3 Modelling
P6.3.1 The Fluid Film
The fluid film is modelled and discretized as described in [127], but a brief review
is included for completeness. Defining the circumferential and axial coordinates
θ and z̃, the isothermal, compressible, and transient Modified Reynolds equation
(MRE) for an ideal gas can be written as
∂ 3 ∂ p̃ ∂ 3 ∂ p̃ ∂ ∂

+ = Sτ Ω̃ p̃ h̃ + 2Sτ p̃ h̃ + q̃inj, (P6.1)

p̃ h̃ p̃ h̃
∂θ ∂θ ∂z̃ ∂z̃ ∂θ ∂τ
12R2 µRsTiso 6R2 µωτ
where q̃inj = qinj and Sτ = . (P6.2)
p2a C 3 C 2 pa
The spatially varying pressure, film height and injection source term fields are
denoted by p̃, h̃ and q̃inj , respectively. The gas viscosity is denoted by µ, Rs is
the specific gas constant and Tiso is the isothermal gas temperature. The equation
has been non-dimensionalized using the nominal bearing clearance C and radius
R as through-film and in-film length scales respectively, the ambient pressure pa
as pressure scale, and a frequency ωτ as time scale. In the present work, the
characteristic frequency is defined simply as the rotor angular velocity ωτ = Ω,
but separate definitions are maintained to allow aerostatic simulations (where
Ω = 0) or time-integration with varying speed. In order to formulate a Finite
Volume (FV) discretization, it is convenient to rewrite eq. (P6.2) in terms of the
flow vector q̃g as
∂ p̃ h̃

∇ · q̃g = −2Sτ − q̃inj,
∂τ (P6.3)
T T
where q̃g = q̃θ q̃z = − p̃ h̃3 ∇ p̃ + Sτ Ω̃ p̃ h̃ 1 0 ,

which can be integrated over a Control Volume (CV) with in-film area Ai and
circumference Si yielding
∂
∫ ∬ ∬
q̃g · ndS = −2Sτ (P6.4)

p̃ h̃ dA − q̃inj dA.
Si Ai ∂τ Ai
The fluid film is discretized using a non-uniform Cartesian grid to allow refinement
in the injection areas as illustrated for the first pad in fig. P6.4. On a CV level,
this allows the boundary flux terms (left hand side of eq. (P6.4)) to be split into
four line integrals approximated using midface values, while the unsteady term can
be represented by the Control Volume Centre (CVC) values (both being second
P6.3 Modelling 195
order approximations). The FV residual equation for the i-th CV can thus be
written as
∂ h̃i ∂ p̃i

∆z̃i q̃θ |i,e − q̃θ |i,w + ∆θ̃i q̃z |i,n − q̃z |i,s + 2Sτ ∆z̃i ∆θ̃i p̃i + h̃i + mÛ̃ i = r̃FV,i,

∂τ ∂τ
(P6.5)
where ∆θ̃i , ∆z̃i are the CV dimensions, p̃i , h̃i are the CVC pressure and film heights
and the integrated injection flux is denoted simply as mÛ̃ i . The fluid film flow
components q̃θ , q̃z are evaluated at the midpoint of the northern (n), southern
(s), western (w), and eastern (e) Control Volume Edges (CVEs) as indicated. As
defined from eq. (P6.3), the flow depends on the pressure, pressure gradients, film
heights and the angular velocity. The pressure gradients are reconstructed from the
CV centre pressures using a Central Differencing Scheme (CDS), while the pressure
values are found using a Linear Upwind Differencing Scheme (LUDS). The film
heights are composed of constant contributions from the nominal clearance and
linear pad inlet slopes as well as varying compliant and rigid contributions. The
latter are added through mappings from the foil and rotor domains. Depending on
the local flow direction, the collective face flux contribution to the FV residual for
the i-th CV given in Equation (P6.5) hence depends on six out of eight possible
neighbouring cells.
Each of the two bearings α = A, B are discretized using nθ nz = nCV cells.
0.8
0.6
z/R (−)
0.4
0.2
0.0
40 60 80 100 120 140
◦
θ( )
Figure P6.4: FV mesh for the first pad. The filled ∬ inner semicircle (blue) marks
the injection zone, i.e. where q̃inj , 0 and mÛ̃ i = A q̃inj dA is thus added to the
i
CV residuals. The larger semicircle (red) shows the extent of the CFD injector
sub-model and hence where the fluid film pressure is sampled and averaged to
provide the pressure BC p̃film .
Collecting the CV centre pressures to form the bearing pressure state vector
(P6.6)
T
p̃α = p̃1, . . . , p̃nCV ,

the assembled FV residual equations can be expressed for bearing α as

AFV p̃α, x̃r , x̃ f ,α p̃α + BFV xÛ̃ r , xÛ̃ f ,α p̃α + CFV x̃r , x̃ f ,α pÛ̃ α + m
Û̃ α = r̃FV,α,

| {z } | {z } | {z } |{z}
∆z̃i (q̃θ |e −q̃θ |w )+∆θ̃ i (q̃z |n −q̃z |s ) 2Sτ ∆z̃i ∆θ̃ i hÛ̃ i p̃i 2Sτ ∆z̃i ∆θ̃ i h̃i pÛ̃i mÛ̃ i
(P6.7)
where AFV holds the coefficients stemming from the MRE’s Couette and Poiseuille
terms, while the coefficients in BFV and CFV represent the squeeze and local
expansion terms, and m
Û̃ α contains the CV injection flow contributions collected as
Û̃ α = mÛ̃ i . . . mÛ̃ nCV T . (P6.8)

m
The partial derivatives of eq. (P6.7) required to assemble the system Jacobian
are detailed in appendix P6.D where some additional details on the implementation
are also included as required. For a detailed description of the interpolation schemes
and the treatment of Boundary Conditions (BCs), the reader is referred to [127].
P6.3.2 Rotor
The rotor assembly is modelled as a rigid shaft with nrdof = 4 Degrees of Freedom
(DOFs) collected in the vector
(P6.9)
T
x̃r = ε Ax ε Ay εBx εBy ,

using which, the rotor equations of motion can be cast in state space residual form
as
xÛ̃ r

I 0 0 I x̃r 0
− − = r̃r ∈ R8, (P6.10)
0 M̃r xÜ̃ r 0 Ω̃ G̃r xÛ̃ r f̃rw + f̃b + f̃ub
with r̃r being the residual to be minimized when solving the equations either
statically or dynamically. The rotor mass and gyroscopic matrices are denoted
by M̃r and G̃r , while f̃rw and f̃ub hold the static loads and the unbalance forces.
All quantities have been non-dimensionalized and are detailed in appendix P6.B.
The bearing forces f̃b are obtained through the usual integration of the projected
(relative) gas film pressure. This can be written for the bearing α components as
2π θ +1
nÕ nz +1
L̃
cos θ cos θ v,i Õ
∫ ∫
f̃b,α = ∆z̃v, j p̃(θ v,i, z̃v, j ) − 1 ,

( p̃ − 1) dA ≈ ∆θ v,i
0 0 sin θ sin θ v,i
i j
(P6.11)
P6.3 Modelling 197
where the nested summations represent a two-dimensional trapezoidal rule approx-

imation using the vertex pressures p̃(θ v,i, z̃v, j ) with weights ∆zv, j and ∆θ v,i (note
Ínz +1 Í θ +1
that j=1 ∆z̃v, j = L̃ and nj=1 ∆θ v, j = npad (θ t − θ l )). As the vertex pressures are
not directly available, these additionally need to be reconstructed from the CV
centre positions and BCs. This is achieved using CDS or LUDS depending on the
vertex location θ v,i, z̃v, j in the mesh. By assembling the FD coefficients needed
for the reconstruction of vertex pressures and the trapezoidal rule weights from
eq. (P6.11), the integration of bearing forces can be written as a non-constant
mapping
f̃b = f̃b0 + H fb,p (p̃ AB ) p̃ AB, (P6.12)
where the combined system state vector has been defined as
T
p̃ AB = p̃TA p̃T . (P6.13)

B
Notice that the pressure dependency of the mapping H fb,p is due to the enforced
Gümbel condition. This is implemented by discarding the vertices with sub-
ambient pressures in eq. (P6.11) thus requiring a reconstruction of H fb,p whenever
the pressure at a vertex crosses pa . Rebuilding the mapping matrix is computation-
ally cheap, but, being mathematically rigorous, the jump discontinuous pressure
dependency implies the bearing force Jacobian contribution to be given as
∂ f̃b ∂H fb,p
= p̃ AB + H fb,p, (P6.14)
∂p̃ AB ∂p̃ AB
where ∂H fb,p /∂p̃ AB is undefined where the pressure at any vertex is crossing pa and
otherwise zero. When assembling the system Jacobian, only the second term of
eq. (P6.14) is thus included implying that no crossings are assumed to be imminent.
This has, however, not created any noticeable challenges. The fundamental problem
is present in all numerical implementations relying on the Gümbel condition, but
usually it is blurred by the use of FD approximated Jacobian matrices. A possible
remedy would be to smooth the transition from sub-ambient to super-ambient
integration coefficients in H fb,p , but as the Gümbel condition is problematic in
itself, efforts to circumvent its use entirely would be more relevant.
In the present work, the rigid film height contribution caused by the rotor
movement is assumed axially uniform as the angle of the shaft is neglected.
Referring to the FV discretized MRE in eq. (P6.5), the film height must be
evaluated at the CVCs for use in the squeeze and local expansion terms, while the
Couette and Poiseuille terms depend on the film heights at the CVE midpoints.
The rigid film height contribution at the circumferential coordinate θ is given
as a trigonometric projection of the bearing eccentricity components, usually
written as hr (θ) = C + e x cos θ + e y sin θ. Having laid out the fluid film mesh, the
CVC/CVE θ-coordinates are known and these projections can be evaluated offline
and assembled to form the constant mappings
h̃r,CVC = h̃r0,CVC + HCVC,xr x̃r and h̃r,CVE = h̃r0,CVE HCVE,xr x̃r (P6.15)
| {z } | {z }
2nCV ×nrdof 2 npad nz +2nθ nz +nθ ×nrdof

providing the rigid film heights at all CVCs/CVEs in both bearings. Notice that
HCVC,xr and HCVE,xr also contain the conversion factor between the rotor and fluid
film length scales.
P6.3.3 Foil Structure

The top foil structure is modelled using Euler–Bernoulli beam elements with
rotational and lateral DOFs only, while the bump foil support is modelled using
the Simple Elastic Foundation Model (SEFM) with stiffness k and loss factor η.
The foil is divided into nfe foil elements across the entire circumference. Each node
contributes two DOFs, but the first node of every pad is constrained (leading edge
is fixed) leaving a foil DOF vector x f with length nfdof = 2nfe . The stiffness matrix
is composed of a beam contribution from the top foil and a support contribution
from the SEFM, while the damping matrix is due to the SEFM only. Using the
element matrices detailed in appendix P6.C, the global foil structure matrices can
be constructed using the standard element assembly procedure as
Õ
K̃ f = k̃t + k̃SEFM ,
e e
e
Õ Õ (P6.16)
D̃ f = e
d̃SEFM = ηω̃ f e
k̃SEFM,
e e
where the element matrices have been non-dimensionalized using the element
length le , ambient pressure pa , and time scale ωτ−1 . Notice that the foil oscillation
frequency ω f needed to calculate the equivalent viscous damping coefficient from
the loss factor η is chosen here simply as ω f = Ω, except for the aerostatic case
with Ω = 0 where this is set to ω f = 20 kRPM to avoid zero damping. The value
chosen for ω f affects the effective damping provided by the SEFM directly and
other choices would be more adequate for some cases, especially if comparing with
frequency domain methods as recently discussed in [129, 146]. Having condensed
out the constrained DOFs at the leading pad edges, the foil equations of motion
can be written in residual form as
D̃ f xÛ̃ f ,α + K̃ f x̃ f ,α − f̃ p,α = r̃ f ,α, (P6.17)
representing a first order system with time constant ηω−1
f . The work equival-
ent nodal forces can be obtained from the fluid film pressure as detailed in
P6.3 Modelling 199
appendix P6.C for each element. Using the element assembly procedure, the
constant mapping providing the full foil load vector in bearing α can be obtained
as
Hefp,p . (P6.18)
Õ Õ
f̃ p,α = f̃ p0 + H fp,p p̃α where f̃ p0 = e
f̃ p0 and H fp,p =
|{z} e e |{z}
nfdof ×nCV 4×nCV
Analogous to the calculation of rigid film heights from the rotor states in
eq. (P6.15), the compliant film height contributions should likewise be evaluated at
both CVCs and CVEs. Identifying the beam element spanning each circumferential
CVC/CVE θ-position, the element shape functions (given in appendix P6.C) can
be evaluated to provide the compliant film height contribution as a function of the
four element DOFs. The resulting foil DOF coefficients can be assembled to form
constant linear mappings from the bearing α foil state vector to the compliant
film height contributions as
h̃c,CVC = HCVC,x f x̃ f ,α and h̃c,CVE = HCVE,x f x̃ f ,α (P6.19)

| {z } | {z }
2nCV ×nfdof 2 npad nz +2nθ nz +nθ ×nfdof

for the CVCs and CVEs, respectively. Notice that HCVC,x f and HCVE,x f also
contain the conversion factor between the foil and fluid film length scales.
P6.3.4 Injection
When modelling the injection flow in gas bearings, the flow is commonly described
as an ideal gas flowing from a stagnant reservoir under isentropic process conditions.
This leads to the widely used closed-form expression [6, 30, 35, 57, 60, 75, 77,
85, 97, 109, 120, 184] for which a derivation is provided in appendix P6.A. The
flow is thus assumed adiabatic, steady, and inviscid with the flow being limited
at a characteristic restricting area by compressibility. Alternative closed-form
expressions based on fully developed Hagen–Poiseuille flow limited by viscous
shear have likewise been applied [117, 142, 152]. In both cases, the deficiencies
of the analytical expressions are compensated with (possibly varying) correction
factors aiming at reconciling the analytical predictions with results from more
comprehensive models or experience.
P6.3.4.1 Injector Sub-model and Calculation of Parameter Map Mass Flows

In the present work, an approach fundamentally different from the correction
factor aided analytical expressions is taken. Instead of fitting idealized isentropic
or Hagen–Poiseuille based predictions to CFD results using a correction factor,
the CFD results are directly used. This allows the complete injector geometry
to be modelled capturing the hinj and ε x dependent interaction between multiple
restricting cross sections and accounts for both compressibility and viscous shear.
Theoretically, the approach could also capture unsteady and turbulence effects, but
the assumption of laminar and steady injection flow from the analytical expressions
is currently retained.
To predict the injected mass flow, a three-dimensional sub-model of the injector
and a small portion of the fluid film is created as illustrated in fig. P6.3(b) using
the open-source CFD code OpenFOAM v1906. The flow is restricted when passing
through the narrowest cross sections of the injector sub-model which are found
in the fluid film (the curtain area) and the cone section. To adequately capture
the flow profile, these parts have been meshed using eight CVs through the cross
sections. This implies rather small cell dimensions which are largely dictating
the divisions and gradings of the remaining mesh necessary to obtain acceptable
cell geometry metrics (aspect ratios, determinants, face interpolation weights
etc.). The resulting mesh consists of 212400 CVs (209800 hexahedra and 2600
prisms in the centre region) and is shown in fig. P6.5. Constant temperature
(T = Tiso ) BCs are prescribed on all walls and steady state solutions are obtained
using the ”rhoSimpleFoam” compressible solver applying the SIMPLEC algorithm.
Convergence is defined either by the standard SIMPLE criteria (set to terminate
if the initial field residuals fall below 10−4 ) or a custom criteria requiring the
integrated flows in and out of the sub-model to match within 1 %.
A set of Python scripts are written to auto-generate and execute the Open-
FOAM model for a given set of input parameters. Initially, this was intended to
be coupled directly to the rotor–bearing code, but this proved unfeasible due to
high simulation times and frequent crashes of the CFD solver. Instead, the model
is evaluated over a structured grid of input parameters. Referring to fig. P6.3(b),
the parameters to vary are chosen as: (a) the axial nozzle clearance hinj which
can be physically varied using the piezo actuator as shown in fig. P6.2(b); (b) the
injection pressure pinj ; (c) the fluid film pressure pfilm representing the pressure at
a distance of rfilm from the injector centre; (d) the vertical ε x eccentricity for the
sub-model injector located at θ = 180◦ ; and (e) the angular velocity of the rotor Ω.
The mass flow is thus solved for all combinations of the parameter values listed in
table P6.3.
As to be expected for a compressible solver, the convergence has been found to
be very sensitive to the initial conditions, often leading to solver failures. This
is overcome by initially solving an ”easy” case from the central region of the
parameter space resulting in low flow velocities. The converged velocity, pressure,
and temperature fields can then serve as initial conditions for the neighbouring
parameter space grid points, i.e. the 10 cases with parameter sets differing only by
one step in a single direction. These 10 cases can then be solved with much less
hinj
P6.3 Modelling
αs
rb,i
rp,i
(a) (b)
Figure P6.5: Selected views of the CFD model mesh with injection parameters hinj = 35 µm, pinj = 400 kPa,
pfilm = 168 kPa, ε x = 0.0 and Ω = 40 kRPM: (a) x y-plane cross section at the cone to inner annulus transition
with selected dimensions from fig. P6.3(a) superimposed. The CVs are coloured by the steady state CV centre
velocity magnitudes without smoothing/averaging (flow from top towards bottom of the picture). The maximum
velocity corresponds to Mach 1 implying the flow is choked at the cone–inner annulus transition. (b) View of the
three-dimensional injector sub-model FV mesh comprising 212400 CVs (209800 hexahedra and 2600 prisms around
the centre axis).
201
tendency to diverge/crash and using fewer iterations than required if initialized

from ambient conditions. Repeating this process for every new solution obtained,
it is possible to automatically propagate through the parameter space with few
manual interventions. The process is, however, slightly complicated by the flow
velocities approaching sonic conditions in some regions of the parameter space.
To obtain converged solutions for the cases with high Mach numbers, it has
been necessary to augment the propagation scheme with a logic sensing the flow
velocities and switching on the ”transonic” option of rhoSimpleFoam. The flow
shown in fig. P6.5(a) is such an example as the maximum flow velocity corresponds
to Mach 1, indicating a choked flow at the cone to inner annulus transition. This
is as expected from the downstream/upstream pressure ratio of 0.42.
From an interpolation perspective, the resolution along each parameter dimen-
sion in table P6.3, i.e. the step sizes, is finer than necessary, but the small steps are
chosen to reduce convergence issues. The simulation time for the 2940 simulations
varies significantly, but averages at about 15 min. These have been executed in
parallel on a desktop pc supporting 12 threads giving a collective wall clock time
of the order of 60 h. This is considerable compared to the targeted subsequent
rotor–bearing simulations, but the parameter map only has to be calculated once
for a particular injector geometry. Furthermore, with the benefit of hindsight, the
number of simulations could have been reduced appreciably since the mass flow
has proven virtually indifferent to the angular velocity.
Originally, the map was intended to include injection pressures pinj up to
the limit of the currently installed air supply system at 7 bar and axial injector
clearances hinj up to the fully open setting at 45 µm. This did, however, turn
out to be impossible without violating the laminar flow assumption. The cross
Table P6.3: Details of the swept five-dimensional structured parameter space

with dimensions 7 × 4 × 5 × 7 × 3 = 2940 runs.
Dimension Parameter Levels Values
1 Nozzle clearance, hinj 7 5.0 µm, 10 µm, 15 µm, 20 µm,
25 µm, 30 µm and 35 µm
2 Injection pressure, pinj 4 150 kPa, 233 kPa, 317 kPa
and 400 kPa
3 Fluid film pressure, pfilm 5 90.0 kPa, 168 kPa, 245 kPa,
323 kPa and 400 kPa
4 Eccentricity, ε x 7 −0.9, −0.5, 0.0, 0.5, 0.7, 0.8
and 0.9
5 Angular velocity, Ω 3 0 kRPM, 20 kRPM and
40 kRPM
P6.3 Modelling 203
section of the innermost part of the injector cone, i.e. at the transition to the
inner annular section, forms a cylindrical surface with height hinj and radius rb,i .
Recalling the hydraulic diameter to be defined as four times the cross sectional area
2πrb,i hinj divided by the wetted perimeter 4πrb,i and noting that a given mass flow
mÛ inj implies the average flow velocity mÛ inj / 2πρrb,i hinj , the cross section Reynolds

number for a given mass flow can be written as
mÛ inj,i
Re = (P6.20)
πrb,i µ
if assuming the mass flow is uniformly distributed around the circumference.
Equation (P6.20) is independent of the height meaning that the same Reynolds
number is present in the fluid film curtain area sharing the radius rb,i (though the
circumferentially uniform flow assumption is more questionable due to the rotor
surface velocity). Both cylinders can be approximated as annular sections, where
the turbulent transition is commonly assumed to initiate at Reynolds numbers
from around 2000. An overview of the mass flows from all simulations are shown
in fig. P6.6(a) and the approximate equivalent Reynolds numbers calculated using
eq. (P6.20) are shown in fig. P6.6(b). It is evident that the maximum flow rates at
around 145 mg/s obtained for pinj = 400 kPa and hinj = 35 µm are already edging
the laminar assumption, meaning that a reliable mass flow prediction for larger
nozzle clearances and/or injection pressures would require a turbulence model.
The calculated mass flows are plotted as functions of each parameter in figs. P6.7
to P6.11. The left plots (upper plots in the preprint) contain the individual mass
flows for all 2940 simulations to give a sense of the distributions, while the right
plots (lower plots in the preprint) provide a measure of the parameter mean effects
in terms of the average mass flow for all cases sharing each parameter value.
Notice that the effects of the height parameters hinj and ε x are approximately
unidirectional meaning that the average signed mass flow for each parameter value
is close to zero. To bring out the mean effect, the average mass flows for these
two parameters have thus been calculated from the absolute mass flows (as also
indicated by the axis labels).
The interaction between the restricting areas can be observed from the hinj
and ε x plots in figs. P6.7 and P6.10. Noting the fluid film height at the sub-model
injector located at θ = 180◦ to be given approximately as hfilm = C (1 − ε x ), the
included eccentricity values from -0.9 to 0.9 imply curtain area heights in the range
4–76 µm. From fig. P6.7, it is seen that very small flows are obtained for all cases
with hinj = 5 µm regardless of the remaining parameter values, while the flow rates
vary increasingly when the nozzle clearance is enlarged towards hinj = 35 µm. This
matches the expectation of the flow to be restricted mainly in the cone for small
hinj values, while the effective restricting area varies also with ε x for larger values
of hinj . Consistently, it can be seen from fig. P6.10 that changes to the eccentricity
100
Mass flow (mg/s)
−100
0 500 1000 1500 2000 2500 3000

Simulation number
(a)
2500
Approximate Re (−)
2000
1500
1000
500
0
0 500 1000 1500 2000 2500 3000
Simulation number
(b)
Figure P6.6: Results from all 2940 CFD simulations: (a) Steady state mass
flows and (b) approximate Reynolds number in the annular shaped innermost part
of the cone and the curtain area as calculated from eq. (P6.20).
P6.3 Modelling 205
ε x have virtually no effect for ε x ≤ 0 (implying curtain heights ≥ 40 µm) as the

flow is here restricted by the nozzle clearance. Reducing the curtain height further,
the mass flow becomes increasingly sensitive to the eccentricity as the curtain area
restriction gains relative significance. It is also interesting to note that the mean
effects of hinj and ε x (for values larger than zero) appear closer to being linear as
would be expected from the compressibility limited isentropic flow than cubic, as
would be expected from a viscosity limited Hagen–Poiseuille flow. Though the
modelled flow contains both of these physical effects, the compressibility hence
seems to be dominating.
The interaction of hinj and ε x is likewise visualized in the contour plot in
fig. P6.12(a) showing the average absolute mass flow for each of the 7 × 7 (hinj, ε x )
combinations. For cases below the superimposed contour hinj /hfilm = 1/2, the
mass flow is restricted mainly in the cone and consequently varies only with hinj
(horizontally). Above hinj /hfilm = 2, the flow is restricted mainly in the fluid film
and the mass flow is seen to vary only with ε x (vertically).
Considering the pinj and pfilm plots in figs. P6.8 and P6.9, the two parameters
are seen to be heavily interdependent resulting in the seemingly arbitrary zero-
crossings in the mean effect plots. The influence on the mass flow of one pressure
parameter is completely dependent on the value of the other as this determines the
upstream/downstream roles and subsonic/choked conditions. It should be kept in
mind that while sharing the maximum value of 400 kPa, the minimum values and
distributions of the two pressures are not the same. The overall tendency of the
mass flow increasing with pinj and decreasing with pfilm is as expected, but the
mean effect plots are, to a large degree, artefacts of the included parameter levels.
To jointly analyse the influence of the two pressure parameters, the average
mass flow is calculated for each of the 4 × 5 (pinj, pfilm ) combinations as shown
in the contour plot in fig. P6.12(b). From the derivation of the isentropic flow
model in appendix P6.A, choked conditions are expected to arise when reaching
a downstream/upstream pressure ratio of around 0.53 (for dry air with γ = 1.4).
This implies that Mach 1 is reached in the restricting area and hence that further
decreasing of the downstream pressure will have no effect on the flow rate. In
fig. P6.12(b), the pressure ratio contours corresponding to choked inflow and
outflow are marked with Ma = 1. Between the two Ma = 1 contours, subsonic
conditions prevail and the flow varies with both pressures. The Ma = 0 contour
marks pinj = pfilm at which the flow is zero and reverses. Below the lowermost
Ma = 1 contour, the injection pressure is upstream (we have flow into the bearing)
and the flow is seen to be virtually independent of the downstream pressure pfilm .
Above the topmost Ma = 1 contour, the upstream/downstream roles have been
reversed (we have flow out of the bearing) and no dependency on the downstream
pressure pinj is evident.
Last, it is interesting to consider the influence of rotor velocity on the predicted
100
Mass flow (mg/s)
−100
5 10 15 20 25 30 35
hinj (µm)
(a)
Mean abs mass flow (mg/s)
40
30
20
10
0
5 10 15 20 25 30 35
hinj (µm)
(b)
Figure P6.7: Injection mass flow for all 2940 simulations as a function of nozzle
clearance hinj : (a) Individual mass flows for the 420 runs at each of the seven hinj
values; and (b) average absolute mass flow at each value of hinj .
P6.3 Modelling 207
100
Mass flow (mg/s)
−100
150 200 250 300 350 400

pinj (kPa)
(a)
Mean mass flow (mg/s)
20
10
−10
150 200 250 300 350 400

pinj (kPa)
(b)
Figure P6.8: Injection mass flow for all 2940 simulations as a function of injection
pressure pinj : (a) Individual mass flows for the 735 runs at each of the four pinj
values; and (b) average mass flow at each value of pinj .
100
Mass flow (mg/s)
−100
100 150 200 250 300 350 400

pfilm (kPa)
(a)
20
10
−10
−20
100 150 200 250 300 350 400

pfilm (kPa)
(b)
Figure P6.9: Injection mass flow for all 2940 simulations as a function of film
pressure pfilm : (a) Individual mass flows for the 588 runs at each of the five pfilm
values; and (b) average mass flow at each value of pfilm .
P6.3 Modelling 209
100
Mass flow (mg/s)
−100
−0.5 0.0 0.5

εx (−)
(a)
30
20
10
0
−0.75 −0.50 −0.25 0.00 0.25 0.50 0.75
εx (−)
(b)
Figure P6.10: Injection mass flow for all 2940 simulations as a function of
eccentricity, ε x : (a) Individual mass flows for the 420 runs at each of the seven ε x
values; and (b) average absolute mass flow at each value of ε x .
injector mass flows as shown in fig. P6.11. Even though the rotor velocity has been
observed to affect the flow patterns and cause recirculation in the injector recess,
the effective mass flow rate is virtually independent of the rotor velocity. For a
balanced study design, the average mass flow (at least for Ω = 0) would have been
zero, but the included pressure levels slightly favour inflow. The slight increase
from 3.1 mg/s at Ω = 0 to almost 3.5 mg/s at Ω = 40 kRPM could be indicating
a weak dependency, but the change is too small to be significant considering
the current solver tolerances. In conclusion, two thirds of the 2940 simulations
could have been omitted without any significant loss of accuracy. This might seem
counter-intuitive, but it should be emphasized that the results in fig. P6.11 consider
the isolated effect of the rotor surface velocity. As evident from fig. P6.10, the
injection mass flow is strongly dependent on the hydrodynamic pressure which, in
a full bearing, depends on the speed and gap geometry. This provides an indirect
velocity–flow dependency, which is inherently captured by the present modelling
approach.
P6.3.4.2 Obtaining the Injector Mass Flow in a Rotor–Bearing Simulation
In a rotor–bearing simulation, the mass flow from each injector is calculated from
the pre-generated map of CFD results described in the preceding section using
the piecewise multi-linear n-dimensional interpolation scheme given by Weiser and
Zarantonello [180]. Representing this interpolation by the function finj , the mass
flows through the ninj injectors in bearing α can thus be written as
Û̃ inj,α = mÛ̃ inj,1 . . . mÛ̃ inj,ninj T = finj hinj,α, pinj,α, pfilm,α, ε x,α, Ω (P6.21)

m
for which the derivatives ∂finj /∂hinj,α , ∂finj /∂pinj,α , ∂finj /∂pfilm,α , ∂finj /∂ε x,α and
∂finj /∂Ω are likewise provided by the interpolation scheme. Except for the scalar
rotational velocity shared by all injectors, the independent variables are vectors of
length ninj . At present, hinj,α and pinj,α are fixed, but these could be made variable
e.g. to apply feedback control. This is also the reason for having varied these in
the parameter map. In the CFD model, a constant fluid film pressure is prescribed
along the fluid film boundary as illustrated in fig. P6.3(b). It is thus an important
assumption in the present approach that the actual MRE fluid film pressure along
the radius rfilm from the injector centre, i.e. along the red circle shown in fig. P6.4,
can be treated as being uniform. For sufficiently small values of rfilm (here 2.5 mm),
this has been found reasonable. In practice, this is achieved by interpolating from
the CVC pressures to a number of locations on the rfilm periphery and averaging
these. The coefficients representing this procedure can be calculated offline and
P6.3 Modelling 211
100
Mass flow (mg/s)
−100
0 10 20 30 40
Ω (kRPM)
(a)
3.4
3.3
3.2
3.1
0 10 20 30 40
Ω (kRPM)
(b)
Figure P6.11: Injection mass flow for all 2940 simulations as a function of
angular velocity, Ω: (a) Individual mass flows for the 980 runs at each of the three
Ω values; and (b) average mass flow at each value of Ω.
P6 Gas Foil Bearings with Radial Injection: Multi-domain Stability...
0.9 400.0
0.8 hinj /h 79.6
film =2 33.6
1”
0.7
70.8
a=

hinj /h 25.2
0.5
”M
film
=1 62.0 322.5
16.8
pfilm (kPa)
53.2 8.4
hi
εx (−)
0.0
nj /
h 44.4 245.0 0” 0.0
fil
m a=
= 1 35.6 ”M −8.4
2
26.8 −16.8
167.5 1”
−0.5 18.0 a=
”M −25.2
9.2 −33.6
−0.9 0.4 90.0
5 10 15 20 25 30 35 150.0 233.3 316.7 400.0
hinj (µm) pinj (kPa)
(a) (b)
Figure P6.12: Contours of the injection mass flow for all 2940 simulations collapsed (by averaging) into two planes:
(a) 7 × 7 hinj × ε x grid providing 60 simulations per point to average. Notice that the average is taken from the
absolute mass flows. Contours showing three different ratios between the nozzle clearance and the effective film
clearance hfilm = C (1 − ε x ) have been superimposed. (b) 4 × 5 pinj × pfilm grid giving 147 simulations to average per
point. The Ma = 0 line marks pinj = pfilm where the flow is zero while Ma = 1 indicates the pressure ratios (≈ 0.53)
providing choked flow conditions according to the isentropic model.
212
P6.3 Modelling 213
assembled to form a mapping from the CVC pressures as
pfilm,α = pfilm0 + H pfilm,p p̃α, (P6.22)

| {z }
ninj ×nCV
where the vector pfilm0 and matrix H pfilm,p are both constant. The bearing eccentri-
city in the CFD model is represented by the vertical eccentricity ε x for an injector
located at θ = 180◦ in a rigid bearing. To incorporate compliance, an average
fluid film height across the injector area (marked in blue in fig. P6.4) is calculated
and converted to an equivalent eccentricity. A second important assumption of
the present approach is hence that film heigh variations within the injection zone
can be neglected. The procedure for obtaining these equivalent eccentricities can
likewise be represented using a mapping from the rotor and foil states as
ε x,α = ε x0 + Hε,xr x̃r + Hε,x f x̃ f ,α, (P6.23)

|{z} |{z}
ninj ×nrdof ninj ×nfdof
with ε x0 , Hε,xr and Hε,x f being constant.

Having calculated pfilm,α and ε x,α , the injector mass flows m Û̃ inj,α can be cal-
culated (=interpolated) from eq. (P6.21) and should then be distributed to the
CVs. Within the region assigned to each injector
i, the flux field q̃inj in eq. (P6.2)
is calculated simply as the average mÛ̃ inj,i / πr 2 . Subsequently, this is integrated
bi
over each CV using a 5 × 5 Gaussian quadrature scheme allowing the mass flow to
be approximated also for CVs lying only partly within the injection circle. This
integration can be represented by a constant linear mapping from the injector
mass flows to the CV injection flows in bearing α as
Û̃ α = mÛ̃ i . . . mÛ̃ nCV T = Hm̃ ,mÛ̃ m

Û̃ inj,α . (P6.24)

m inj
| {z }
nCV ×ninj
Notice that the even distribution of the injection flow across all CVs within the
orifice area is chosen as it is believed to represent a reasonable approximation.
Alternatively, one could have treated the injection as a point source adding the full
injection inflow to one cell similar to the procedure used (for an FD discretization
of MRE) by Liu et al. [107]. The difference in integrated pressure, and hence the
resulting force on the rotor, is small, but the point source approach implies a taller
and narrower pressure peak. This should be kept in mind when visually comparing
pressure profiles between different studies.
P6.3.5 Assembly of Equation System

The full system of equations can be obtained by assembling the models defined for
each domain: the MRE for the fluid film, the interpolation based injection model,
the rigid shaft model for the rotor, and the SEFM supported Euler–Bernoulli
beams for the foil structure. The state and residual vectors for the full system
with two bearings can thus be assembled as
n oT
z= p̃TA p̃T
B x̃Tf ,A x̃Tf ,B x̃rT Û̃xrT and
n oT (P6.25)
r̃ = r̃T
FV,A r̃T
FV,B r̃Tf ,A r̃Tf ,B T
r̃r ,
while the contributions from each model can be written in residual form as
Û̃ r , xÛ̃ f ,A p̃ A + CFV x̃r , x̃ f ,A pÛ̃ A 


 B FV x 
 BFV xÛ̃ r , xÛ̃ f ,B p̃B + CFV x̃r , x̃ f ,B pÛ̃ B 

 
 

 
 Û̃
D̃ f x f ,A

fG (τ, z, zÛ ) =

 


 D̃ f xÛ̃ f ,B 



 Û̃ r
x



 


 M̃ r Ü̃
x r − f̃ ub (τ)



| {z }
fG,t (τ,z,Ûz)
(P6.26)
, , + Û̃ A p̃ A, x̃r , x̃ f ,A 


 A FV p̃ A x̃r x̃ f ,A p̃ A m 

 0  
, , + Û̃ , ,
   
A p̃ x̃ x̃ p̃ m p̃ x̃ x̃ 0
   


 FV B r f ,B B B B r f ,B  

 
 


K̃ f x̃ f ,A − H fp,p p̃ A −f̃ p0 
   
+ + = r̃.

 
 
 

 K̃ f x̃ f ,B − H fp,p p̃B 
 
 −f̃ p0 

−xÛ̃ r 0 

 
 
 

 
 
 

Ω̃ Û̃ −
   

 − G̃ r x r − H fb ,p (p̃ AB ) (p̃ AB − 1) 
 
 f̃ rw


| {z } | {z }
fG,s (z) fG,0
The upper two rows represent the non-linear FV equations governing the fluid
film including the injection contributions. The two midmost rows represent the
linear foil structural model with pressure loads supplied through a mapping from
the pressure states. The lowermost rows hold the likewise linear state space rotor
model with the bearing forces mapped from the fluid domain, the constant static
loads and forces due to unbalance. Across all three domains, the system residual
function fG (τ, z, zÛ ) has been partitioned into its constant fG,0 , steady fG,s (z) and
transient/unsteady fG,t (τ, z, zÛ ) components.
In order to solve eq. (P6.26) and apply eigenvalue analysis, the system Jacobians
must be obtained. For the steady state case in which fG,t (τ, z, zÛ ) = 0, the Jacobian
P6.3 Modelling 215
with respect to the state vector z is given as
 ∂AFV p̃ A ∂AFV p̃ A ∂AFV p̃ A 

 ∂p̃ A ∂z f , A ∂x̃r 
0 0 ∂m Û̃ A 0
 + ∂ mÛ̃ A ∂m
+ ∂z
Û̃ A
+ ∂z
 
∂p̃ A

 f, A f, A
 ∂AFV p̃ B ∂AFV p̃ B ∂AFV p̃ B 
 ∂p̃ B ∂z f , B ∂x̃r

∂fG,s  0 ∂m 0 ∂m 0 
+ ∂∂x̃
Û̃ B Û̃ B Û̃ B

m
Jz,s= = + ∂p̃ B + ∂z f , B ,
∂z  r

 −H fp,p 0 K̃ f 0 
 0 

 0 −H fp,p 0 K̃ f 

 0 0 −I 
0
− Ω̃ G̃r 

 −H fb,p 0
(P6.27)

where the dependencies of AFV p̃α, x̃r , x̃ f ,α , mÛ̃ α p̃α, x̃r , x̃ f ,α and H fb,p (p̃ AB ) have

been omitted. As discussed in relation to eq. (P6.12), the latter dependency results
from the Gümbel condition.
For the transient case, the Jacobian with respect to both the state vector and
to its temporal derivative is necessary. The Jacobian with respect to z obtains an
additional contribution ∂fG,t /∂z due to the squeeze and local expansion terms of
the MRE and is given as
∂CFV pÛ̃ A ∂CFV pÛ̃ A

 BFV 0 0  0
 
∂z f , A ∂x̃r
∂fG,s ∂fG,t ∂fG,t  0 BFV ∂CFV pÛ̃ B ∂CFV pÛ̃ B
 
Jz = + where = 0 ∂z f , B 0 ,

∂x̃r
∂z ∂z ∂z  0 0 0


 
 0 0 0 
(P6.28)
 
while the Jacobian with respect to the temporal derivative of the state vector zÛ is
given as
 CFV ∂BFV p̃ A ∂BFV p̃ A

 0 ∂ zÛ f , A 0 ∂ xÛ̃ r
0 
∂BFV p̃B ∂BFV p̃ B
 0 CFV 0 0 
 
∂ zÛ f , B ∂ xÛ̃ r
∂fG,t  
JzÛ = = D̃ f 0 (P6.29)

∂ zÛ 0 0 

 0 D̃ f 

 I 0 
 0 0

 0 M̃r 
where the dependencies of BFV xÛ̃ r , xÛ̃ f ,α and CFV x̃r , x̃ f ,α have been omitted.

Some of the partial derivatives needed in eqs. (P6.27) to (P6.29) are self-explanatory,
such as the foil and rotor matrices, while the remainder are detailed in ap-
pendix P6.D.
If zeroing out the transient component of eq. (P6.26), the static equilibrium
position can be found using the Jacobian from eq. (P6.27). For time integra-
tions, the transient terms should be retained and the Jacobian from eqs. (P6.28)
and (P6.29) are additionally required. This is achieved using a C implementation
utilising the general purpose algebraic solver KINSOL and the implicit form initial
value problem solver IDA, both from the SUNDIALS suite [67]. While the system
matrices arising from one domain are very often banded, this is not the case for
the present coupled system Jacobian matrices. In order to efficiently solve the
linear equation systems arising during the non-linear iterations, it is thus necessary
to employ a sparse linear solver such as the currently applied ”KLU” [34]. For
further details on the implementation, the reader is referred to [127].
For a given static equilibrium z0 (where zÛ 0 = 0), the local dynamical system
behaviour is governed by the linearisation (a first order Taylor series expansion) of
eq. (P6.26) which can be written in terms of the Jacobian matrices as
∂fG,t (τ, z, zÛ ) ∂fG,t (z, zÛ ) ∂fG,s (z)

zÛ + + (z − z0 ) = 0, (P6.30)
∂ zÛ z0,Ûz0 =0 ∂z z0,Ûz0 =0 ∂z z0 | {z }

| {z } | {z } | {z } ∆z
JzÛ (z0,Ûz0 =0) Jz,t (z0,Ûz0 =0) Jz,s (z0 )
from which a generalized eigenvalue problem can be formed and solved directly
as −λi JzÛ vi = Jz,t + Jz,s vi or a standard eigenvalue problem can be obtained by
calculating −J−1 zÛ Jz,t + Jz,s vi = λi vi . In either case, the (linear) stability of the
full system around z0 can be judged from the resulting eigenvalues λi as originally
introduced for GFB systems by Bonello and Pham [19] and the multi-domain
mode shapes can be visualized from the eigenvectors vi as previously described by
the authors [129].
P6.4 Results
The results presented in the following sections are given for a passive case, i.e
without injection, and for a hybrid operational mode with the topmost injector
valve in both bearings open with hinj = 35 µm and pinj = 400 kPa. Additionally, a
few results are presented for an aerostatic case at 0 RPM with all injector valves
open. The passive case is chosen as it matches the results presented by Bonello
[16] allowing for a direct validation. The hybrid mode is similar to the ”controlled
hybrid mode” presented by Yazdi and Kim [182] and serves to push the rotor
downwards to enhance the rotordynamic stability. In a hybrid mode with injection
also from the lowermost injectors, the fluid film pressure would increase over
the entire circumference generally lowering the stability threshold unless actively
controlled. The present hybrid mode is thus chosen to illustrate the potential of
P6.4 Results 217
injection to manipulate the fluid film properties. The aerostatic case is included to
demonstrate the potential for enhancing the load carrying capacity and reducing
wear during start–stop cycles.
In order to ensure direct comparability, the passive, hybrid, and aerostatic cases
have been simulated using identical meshes. Studying the convergence of the OSI
with respect to the fluid film and top foil meshes have revealed the OSI to converge
slower than other system properties such as the equilibrium position. Furthermore,
the OSI has been observed to converge from above with respect to the number of
axial CVs nz , while from below with respect to the number of circumferential CVs
nθ . It is hence important to assess the two directions independently. Requiring five
CVs across the injection zone and a reasonable grading, the mesh shown for one
pad in fig. P6.4 with nz = 18 and nθ = 135 (across all three pads) has been found
to provide converged OSIs. Slightly fewer CVs are necessary if using a uniform
mesh. Regarding the top foil discretization, convergence is achieved already with
around five beam elements per pad, but since the cost of additional foil elements
is low, ten elements are used per pad giving nfe = 30. This results in 2430 CVs
(=fluid DOFs) and 60 foil DOFs per bearing. As the rotor contributes eight states,
the full system size is thus 4988.
P6.4.1 Steady State Foil Deformation, Film Height and Pressure

Profiles
In fig. P6.13, the steady state pressure and foil deformation profiles are shown
for both the passive and hybrid case at 20 kRPM in bearing A. For the passive
case in figs. P6.13(a) and P6.13(b), the equilibrium position for the rotor is at
(ε Ax, ε Ay ) = (0.8627, 0.2528) with the third pad (270◦ –25◦ ) being almost entirely
passivated. The minimum film height in the passive case is 10.8 µm. For the hybrid
case in figs. P6.13(c) and P6.13(d), the topmost injectors in both bearings have
been opened, both with clearance hinj = 35 µm and supply pressure pinj = 400 kPa.
This results in 139 mg/s of air being injected in each bearing pushing the rotor
equilibrium downwards to (ε Ax, ε Ay ) = (0.9302, 0.1914) and reducing the minimum
film height to 9.8 µm. The injection induced pressure build-up in pad three is
evident, but the maximum pressure in the second pad (150◦ –265◦ ) also increases
slightly from 177 kPa in the passive case to 186 kPa in the hybrid case.
The hybrid injection scheme forces the rotor downwards to promote dynamic
properties, but injection can likewise be applied to enhance the load carrying
capacity at low speeds. This is illustrated using a purely aerostatic case, i.e. with
Ω = 0 RPM, where all three injectors in both bearings are open with hinj = 35 µm.
The bearing A static eccentricities, minimum film heights, and injection mass
flows are given for four different injection pressures along with the rotor natural
180
120
160 0
Foil deformation (µm)

p (kPa)
Film height h (µm)

100
140
−2 80
120
100 −4 60
40
−6 xf
0.83
Nxfe
20
)
z /R (−
0.42 h
−8
0.00 100 200 300 400
50 100 150 200 250 300 350 400
θ (◦ )
θ (◦ )
(a) (b)
Figure P6.13: Steady state pressure, foil deformation and film height profiles in bearing A at 20 kRPM: (a,b)
passive configuration with static eccentricity (ε Ax, ε Ay ) = (0.8627, 0.2528); and (c,d) hybrid configuration (uppermost
injector open with hinj = 35 µm, pinj = 400 kPa) with static eccentricity (ε Ax, ε Ay ) = (0.9302, 0.1914) implying a mass
flow of 139 mg/s. Notice in the right plots that x f and Nxef denote nodal and element foil deformations respectively,
both referring to the left vertical scale, while h is the combined film height from compliant, rigid and inlet slope
contributions referring to the right vertical scale. Figure continued on the following page.
218
P6.4 Results
180
120
160 0
100
140
p (kPa)
−2 80
120
100 −4 60
40
Film height h (µm)
xf
0.83

−6
Nxef
)
20
0.42 h
−8
100 200 300 400
z /R (−
0.00
50 100 150 200 250 300 350 400
θ (◦ )
θ (◦ )
(c) (d)
Figure P6.13: Continued from previous page.

219
frequencies and damping ratios in table P6.4. The corresponding pressure profile,
film height, and foil deformation for pinj = 400 kPa are shown in fig. P6.14. At
this injection pressure, the rotor equilibrium at (ε Ax, ε Ay ) = (0.7026, −0.1017) is
closer to the bearing centre than the passive case at 20 kRPM and the minimum
film height of 14.9 µm is larger. This is achieved at the expense of a more heavily
deformed foil structure as evident from fig. P6.14(b), which could prove problematic
in a practical application. From the pressure profile in fig. P6.14(a), it is clear
that the bearing was designed with hydrodynamic operation in mind, as the inlet
slopes distort the aerostatic pressure peaks. Lowering the injection pressure, the
film height rapidly decreases indicating that a minimum of 325 kPa is necessary to
levitate the rotor when pressurizing all injectors, but less could possibly suffice if
reducing the injection from the topmost injectors.
From a modelling perspective, it should be noticed that the characteristic

frequency ωτ cannot be defined from Ω in the aerostatic case and thus has been set
simply to one. Furthermore, the eigenvalue analysis depends on JzÛ (see eq. (P6.29)),
and hence the foil damping matrix D̃ f (see eq. (P6.16)), to be non-singular and
thus requires ω f , 0. For the remaining results, rotor-synchronous foil damping
has been assumed (ω f = Ω), but for the aerostatic case this has been defined as
20 kRPM. Arguably, a value in the vicinity of the eigenvalues around 100 Hz would
have been more appropriate, but this has no significant effect on the stability of
the aerostatic equilibria.
Notice from the foil plots in figs. P6.13(b), P6.13(d) and P6.14(b) that the
shown foil deformation between the nodal locations follows the cubic element shape
functions as is also the case for the CVE/CVC compliant film height contributions
calculated using HCVC,x f and HCVE,x f . It should also be noted that while the
commonly applied assumption of an axially uniform foil deformation has previously
been shown to be reasonable in passive GFBs, this might be challenged in HGFBs
by the steep axial pressure gradient close to the bearing centre.
To judge the feasibility of the injection scheme, an indication of the minimum

power required to supply 139 mg/s of air at 400 kPa can be obtained by assuming
the air flow to be compressed from ambient conditions through an idealized
isentropic process yielding approximately 20 W (per bearing). As previously
mentioned, the modelled test rig is imitating a 120 kW blower and current efforts
strives to apply GFBs in MW-class machinery [36]. An air compression power
consumption of the order of 100 W is hence believed to be justifiable for many
applications. Furthermore, some applications would allow the compressed air
supply to be extracted from a later compression stage of the main process.
P6.4 Results
Table P6.4: Aerostatic results, i.e. at 0 RPM, with all three injectors in both bearings open at hinj = 35 µm. The
resulting eccentricities, minimum film heights, and injector mass flows are provided in bearing A. The listed natural
frequencies all represent stable modes and are obtained with ω f = 333 Hz corresponding to 20 kRPM.
pinj min h θ min h ε Ax ε Ay mÛ inj,1 –mÛ inj,3 ω1 –ω4 ζ1 –ζ4

(kPa) (µm) (deg) (−) (−) (mg/s) (Hz) (−)
400 14.9 145.0 0.7026 −0.1017 128, 78.6, 139 91.6, 114, 117, 145 0.11, 0.066, 0.14, 0.082
375 12.8 145 0.7656 −0.1035 116, 59.2, 128 89.1, 108, 114, 138 0.14, 0.093, 0.18, 0.12
350 10.4 180 0.8385 −0.1029 106, 41.3, 118 87.7, 101, 114, 129 0.21, 0.14, 0.26, 0.17
325 6.0 180 0.9521 −0.08485 96.6, 20.7, 107 92.0, 105, 120, 157 0.19, 0.55, 0.23, 0.44
221
0
300 100

p (kPa)
−2
Film height h (µm)

250
80
200 −4
150 60
−6
100
40
−8 xf
0.83
Nxfe
)
20
z /R (−
0.42 h
−10
0.00 100 200 300 400
50 100 150 200 250 300 350 400
θ (◦ )
θ (◦ )
(a) (b)
Figure P6.14: Steady state pressure, foil deformation profile, and film height in bearing A at 0 kRPM with
all three injectors in both bearings opened at hinj = 35 µm with pinj = 400 kPa). The resulting eccentricity is
(ε Ax, ε Ay ) = (0.7026, −0.1017). Notice in the right plot that x f and Nxef denote nodal and element foil deformations
respectively, both referring to the left vertical scale, while h is the combined film height from compliant, rigid and
inlet slope contributions referring to the right vertical scale.
222
P6.4 Results 223
P6.4.2 Campbell Diagrams and Stability Limits

Solving eq. (P6.26) for the rotor–bearing static equilibrium over a range of angular
velocities, the corresponding eigenvalue problems from eq. (P6.30) can be estab-
lished and solved for both eigenvalues and eigenvectors. For each angular velocity,
the full set of 4988 eigenvalues can be filtered to extract those with significant
rotor participation as discussed in [16, 129]. The resulting Campbell diagrams
for the passive and hybrid cases are shown in figs. P6.15(a) and P6.15(b) with
∆Ω = 250 RPM. For the passive case, the OSI is dictated by the first mode (cyl-
indrical forward whirl) becoming unstable at ΩOSI = 29.65 kRPM with a frequency
of 96.55 Hz, while a secondary instability occurs for the fourth mode (conical
forward whirl) at Ω = 39.07 kRPM with a frequency of 174.9 Hz. It should be
mentioned that the stability limits have been pinpointed by locally increasing the
Campbell diagram resolution to ∆Ω = 10 RPM, but this is not shown.
The passive case is nearly identical to the ”case 1” presented by Bonello [16]
where an explicit form of the coupled equation system is used, the Reynolds
Equation is discretized using FD, and the SEFM is applied for the foil structure.
In [16], a Campbell diagram with ∆Ω = 500 RPM and a range of 5–35 kRPM is
presented from which the OSI is identified in the interval 30.0–30.5 kRPM with
a frequency of 97.22 Hz (at 30.5 kRPM). The OSI has previously been shown by
the authors [129] to be increasing with the foil flexibility; therefore, the slightly
lower ΩOSI in the present work compared to [16] is attributed to the additional
foil stiffness contributed by the top foil beam elements (the SEFM parameters
are identical). This has been verified by temporarily removing the top foil beam
elements from the present model to leave only the plain SEFM. This shifts the
OSI to 30.25 kRPM, thus agreeing with [16], and predicts a first mode frequency
at 30.5 kRPM of 97.56 Hz deviating by 0.35 % from the result in [16].
All four rotor natural frequencies and damping ratios obtained at 30 kRPM
(including top foil) are compared to results from [16] in table P6.5. The predictions
agree well, except for the damping ratio of the lowly damped first mode deviating
by approximately −220 %. Both models predict this mode to be very close to its
stability limit with approximately the same damping ratio magnitude, but since
the OSI is slightly shifted as discussed above, the 30 kRPM evaluation point is
effectively placed on opposite sides of the OSI in the two models.
Figure P6.15(b) shows the Campbell diagram obtained with the hybrid injection
scheme activated. The general appearance is similar to the passive case, but all
four natural frequencies are slightly raised throughout the speed range due to the
increased bearing stiffness and the speed of several characteristic events has been
shifted. Notably, the second mode’s reversal from forward to backwards precession
is shifted from 18.25 kRPM in the passive case to 12.5 kRPM in the hybrid case;
but more importantly, both visible modal stability limits are shifted towards higher
200
150
ωd (Hz)
100
Forward 1X
50
Backward 0.5X
Unstable
0
5 10 15 20 25 30 35 40
Ω (kRPM)
(a)
Figure P6.15: Campbell diagrams extracted from the full set of 4988 eigenvalues by filtering at ζ = 0.4: (a) passive
configuration with the stability limit dictated by mode one (cylindrical forward whirl) at ΩOSI = 29.65 kRPM with a
frequency of 96.55 Hz and a secondary instability occurring for mode four (conical forward whirl) at Ω = 38.19 kRPM
with a frequency of 162.6 Hz; and (b) hybrid configuration with the stability limit dictated by mode one (cylindrical
forward whirl) at ΩOSI = 36.24 kRPM with a frequency of 107.9 Hz and a secondary instability occurring for mode
four (conical forward whirl) at Ω = 39.07 kRPM with a frequency of 174.9 Hz. Figure continued on the following page.
224
P6.4 Results
200
150
100
ωd (Hz)
Forward 1X
50
Backward 0.5X
Unstable
0
5 10 15 20 25 30 35 40
Ω (kRPM)
(b)

225
Table P6.5: Comparison of natural frequencies and damping ratios at 30 kRPM

between the present passive case (with the beam element top foil) and the ”case
1” results from Bonello [16].
From [16] Present results Deviation

Mode
ωi (Hz) ζ (−) ωi (Hz) ζ (−) ωi (%) ζ (%)
1 Cylindrical, 96.99 0.00092 96.83 -0.0011 -0.16 −2.2×102
forward
2 Conical, 121.57 0.081 121.5 0.080 -0.06 -0.22
backward
3 Cylindrical, 129.04 0.081 128.7 0.083 -0.26 3.4
backward
4 Conical, forward 165.75 0.044 165.2 0.047 -0.33 4.8
speeds. The OSI is dictated by the stability limit of the first mode being shifted to
ΩOSI = 36.24 kRPM with a frequency of 107.9 Hz, while the secondary instability
occurring for the fourth mode is shifted to Ω = 39.07 kRPM with a frequency of
174.9 Hz. Effectively, the OSI of the hybrid operational mode is thus increased
by 22 %. A detailed comparison of all four rotor mode natural frequencies and
damping ratios at 30 kRPM is given in table P6.6. Notice that the first mode
damping ratio is increased considerably despite the negative sign of the relative
change caused by the shifted OSI. The damping ratio of the second mode (conical
backward whirl) is increased by 9.3 %, while the third and fourth mode damping
ratios are actually lowered by 13 % and 9.7 % respectively.
Table P6.6: Comparison of natural frequencies and damping ratios at 30 kRPM

between the passive case also listed in table P6.5 and the hybrid configuration.
Passive Relative
Hybrid
Mode (from table P6.5) change
ωi (Hz) ζ (−) ωi (Hz) ζ (−) ωi (%) ζ (%)
1 Cylindrical, 96.83 -0.0011 103.7 0.016 7.1 −1.6×103
forward
2 Conical, 121.5 0.080 129.7 0.088 6.7 9.3
backward
3 Cylindrical, 128.7 0.083 137.9 0.072 7.1 -13
backward
4 Conical, forward 165.2 0.047 177.4 0.042 7.4 -9.7
P6.4 Results 227
P6.4.3 Mode Shapes at 30 kRPM

As detailed in [129], the multi-domain mode shapes can be visualized from the
eigenvectors vi solved from eq. (P6.30). Omitting the decay factors eRe(λi )t , this is
achieved by recalling the contribution from each mode to be given as
n oT
z= p̃TA p̃T
B x̃Tf ,A x̃Tf ,B x̃rT xÛ̃ rT = z0 + vi eiωi t + v̄i e−iωi t (P6.31)
allowing the components of the eigenvectors to be separated and visualized ap-

propriately for each domain. Notice that each eigenvector is scaled to produce a
maximum rotor displacement magnitude of C/5 in the visualizations.
The rotor components of the four modes are shown at 30 kRPM for the passive
and hybrid cases in figs. P6.16 and P6.17 respectively. As evident already from
the Campbell diagrams in the previous section, the GFB damping and stiffness
characteristics imply a somewhat peculiar ordering of the modes compared to most
conventional rotor–bearing systems. Nevertheless, the passive modes agree well
with those presented by Bonello [16]. The hybrid rotor modes are very similar,
but the major axis inclination of the two lowest modes is shifted slightly towards
the horizontal yz-plane, while the major axes of the two highest modes are rotated
slightly towards the vertical xz-plane.
The bearing A components of the four mode shapes are visualized across all
three domains in figs. P6.18 and P6.19, likewise at 30 kRPM. The left plots contain
the rotor and foil modal oscillations around the static displacements, while the right
plots contain the bearing mid plane pressure oscillation around the static pressure
profile. The foil and rotor deformations are likewise visualized in [16] showing
good agreement with the present results. Notice that the rather prominent ”wings”
of the foil deformation represent the pad inlet slopes with height 50 µm = 1.25C.
The effect of these is evident from the pressure profiles, especially in the second
pad, as the pressure increases steeply until making a small kink at the end of
the inlet region 30◦ after the leading edge. The passive and hybrid case mode
shapes are similar, with the injection driven pressure peak in pad three being
the most noticeable difference. The previously mentioned changes in rotor orbit
inclinations are likewise visible and the rotor static equilibrium can be seen to be
pushed slightly downwards.
P6.4.4 Unbalance Response

In practice, GFB supported rotor–bearing systems are often limited by sub-
synchronous vibrations rather than linear stability limits. It is hence appropriate
to also study the effect of the hybrid injection scheme on the unbalance response.
For this purpose, transient solutions to eq. (P6.26) have been obtained for 2.34 s (12
2π 2π
GFBs GFBs
−0.2 −0.2
−0.1 −0.1
εx (–)
εx (–)
ωd t (rad)
ωd t (rad)
0.0 0.0
π π
0.1 0.1
0.2 0.2
−200 −200
−100 −100
−0.2 0 ) −0.2 0 )
−0.1 100 m −0.1 100 m
0.0 0.0
0.1 200 (m 0 0.1 200 (m 0
εy (– 0.2 z εy (– 0.2 z
) )
(a) (b)
Figure P6.16: Rotor modes in passive configuration at 30 kRPM: (a) cylindrical forward mode at 96.83 Hz with
ζ = −0.0011; (b) conical backward mode at 121.5 Hz with ζ = 0.080; (c) cylindrical backward mode at 128.7 Hz with
ζ = 0.083; and (d) conical forward mode at 165.2 Hz with ζ = 0.047. The modes are plotted at 30 kRPM to allow
direct comparison to results in [16]. Notice that the decay factor eRe(λi )t has been omitted from the visualizations.
Figure continued on the following page.
228
P6.4 Results
2π 2π
GFBs GFBs
−0.2 −0.2
−0.1 −0.1
0.0 0.0
εx (–)
εx (–)
π π
0.1 0.1
ωd t (rad)
ωd t (rad)
0.2 0.2
−200 −200
−100 −100
−0.2 0 ) −0.2 0 )
−0.1
0.0 100 m 0
−0.1
0.0 100 m 0
0.1 200 (m 0.1 200 (m
εy (– 0.2 z εy (– 0.2 z
) )
(c) (d)

229
2π 2π
GFBs GFBs
−0.2 −0.2
−0.1 −0.1
εx (–)
εx (–)
ωd t (rad)
ωd t (rad)
0.0 0.0
π π
0.1 0.1
0.2 0.2
−200 −200
−100 −100
−0.2 0 ) −0.2 0 )
−0.1 100 m −0.1 100 m
0.0 0.0
0.1 200 (m 0 0.1 200 (m 0
εy (– 0.2 z εy (– 0.2 z
) )
(a) (b)
Figure P6.17: Rotor modes in hybrid operation at 30 kRPM: (a) cylindrical forward mode at 103.7 Hz with
ζ = 0.016; (b) conical backward mode at 129.7 Hz with ζ = 0.088; (c) cylindrical backward mode at 137.9 Hz
with ζ = 0.072; and (d) conical forward mode at 177.4 Hz with ζ = 0.042. The modes are plotted at 30 kRPM to
allow direct comparison with results in fig. P6.16. Notice that the decay factor eRe(λi )t has been omitted from the
visualizations. Figure continued on the following page.
230
P6.4 Results
2π 2π
GFBs GFBs
−0.2 −0.2
−0.1 −0.1
0.0 0.0
εx (–)
εx (–)
π π
0.1 0.1
ωd t (rad)
ωd t (rad)
0.2 0.2
−200 −200
−100 −100
−0.2 0 ) −0.2 0 )
−0.1
0.0 100 m 0
−0.1
0.0 100 m 0
0.1 200 (m 0.1 200 (m
εy (– 0.2 z εy (– 0.2 z
) )
(c) (d)

231
200 2π
−2
xf xr0 ε=1 p
xf 0 xr 180 p0
−1 pa
Pressure (kPa)
160
ωd t (rad)
εx (–)
0 140 π
120
1
100
2
80 0
−1 0 1 2 1 2 3 4 5 6 7
εy (–) θ (rad)
(a)
200 2π
−2
xf xr0 ε=1 p
xf 0 xr 180 p0
−1 pa
Pressure (kPa)
160
ωd t (rad)
εx (–)
0 140 π
120
1
100
2 80 0
−1 0 1 2 1 2 3 4 5 6 7
εy (–) θ (rad)
(b)
Figure P6.18: Multi-domain mode visualization in passive operation at 30 kRPM

shown in bearing A. The foil oscillation and rotor movements are shown to the
left while the pressure field oscillations at the bearing centre plane are shown to
the right, both with the decay factor eRe(λi )t omitted from the visualizations. The
modes are scaled to provide a maximum displacement amplitude of C/5 and plotted
over a full period with the phase indicated by the colour (dark towards light):
(a) Cylindrical forward mode at 96.83 Hz with ζ = −0.0011 (normalized by ε x in
bearing A); (b) Conical backward mode at 121.5 Hz with ζ = 0.080 (normalized
by ε x in bearing A); (c) cylindrical backward mode at 128.7 Hz with ζ = 0.083
(normalized by ε x in bearing A); and (d) conical forward mode at 165.2 Hz with
ζ = 0.047 (normalized by ε x in bearing B). Figure continued on the following page.
P6.4 Results 233
2π
−2
xf xr0 ε=1 p
200
xf 0 xr p0
−1 180 pa
Pressure (kPa)
ωd t (rad)
160
εx (–)
0 π
140
1 120
100
2
0
−1 0 1 2 1 2 3 4 5 6 7
εy (–) θ (rad)
(c)
2π
−2
xf xr0 ε=1 200 p
xf 0 xr p0
−1 180 pa
Pressure (kPa)
ωd t (rad)
160
εx (–)
0 π
140
1 120
100
2
0
−1 0 1 2 1 2 3 4 5 6 7
εy (–) θ (rad)
(d)
blocks of 400 samples at 2048 Hz) for every 250 RPM using three different unbalance
configurations: (a) u A = 2.5 g mm, u B = −2.5 g mm; (b) u A = 20 g mm, u B =
−2.5 g mm; and (c) u A = 60 g mm, u B = −2.5 g mm. For all three configurations,
the unbalance is added at the bearing positions providing the unbalance force vector
given in eq. (P6.B1). The time integrations are initiated from the corresponding
static equilibria and the initial 0.5 s is discarded to obtain steady state time series
before calculating the frequency spectra. Notice that while 0.5 s is sufficient to
eliminate transient vibrations for most of the speed range, the damping ratio of the
first mode goes towards zero when approaching the OSI meaning that the duration
of the transient approaches infinity. Strictly speaking, the 1.84 s time series close
to the OSI will contain a transient component from the first mode and thus be
non-steady state. Nevertheless, the spectra are calculated as complex Fourier
2π
−2
xf xr0 ε=1 200 p
xf 0 xr p0
180
−1 pa
Pressure (kPa)
160
ωd t (rad)
εx (–)
0 140 π
120
1
100
2 80
0
−1 0 1 2 1 2 3 4 5 6 7
εy (–) θ (rad)
(a)
2π
−2
xf xr0 ε=1 200 p
xf 0 xr p0
180
−1 pa
Pressure (kPa)
160
ωd t (rad)
εx (–)
0 140 π
120
1
100
2 80
0
−1 0 1 2 1 2 3 4 5 6 7
εy (–) θ (rad)
(b)
Figure P6.19: Multi-domain mode visualization in hybrid operation at 30 kRPM

shown for bearing A. The foil oscillation and rotor movements are shown to the
left while the pressure field oscillations at the bearing centre plane are shown
to the right, both with the decay factor eRe(λi )t omitted from the visualizations.
The modes are scaled to provide a maximum displacement amplitude of C/5 and
plotted over a full period with the phase indicated by the colour (dark towards
light): (a) Cylindrical forward mode with ωi = 103.7 Hz and ζ = 0.016 (normalized
by ε y in bearing A); (b) conical backward mode with ωi = 129.7 Hz and ζ = 0.088
(normalized by ε y in bearing B); (c) cylindrical backward mode with ωi = 137.9 Hz
and ζ = 0.072 (normalized by ε x in bearing A); and (d) conical forward mode with
ωi = 177.4 Hz and ζ = 0.042 (normalized by ε x in bearing B). Figure continued
on the following page.
P6.4 Results 235
2π
−2 220
xf xr0 ε=1 p
xf 0 xr p0
200
−1 pa
Pressure (kPa)
180
ωd t (rad)
εx (–)
160
0 π
140
1 120
100
2
0
−1 0 1 2 1 2 3 4 5 6 7
εy (–) θ (rad)
(c)
2π
−2 220
xf xr0 ε=1 p
xf 0 xr 200 p0
−1 pa
Pressure (kPa)
180
ωd t (rad)
εx (–)
160
0 π
140
1 120
100
2
0
−1 0 1 2 1 2 3 4 5 6 7
εy (–) θ (rad)
(d)
transformations of the complex displacement ε x + iε y to provide separate backward

and forward component amplitudes. These can be plotted separately as a double
sided spectrum but for the present analysis, the forward and backward amplitudes
are added to obtain the major axis amplitude. Contrary to the separate single
sided spectra of ε Ax and ε Ay , the major axis amplitude spectrum provides the
vibrational amplitude experienced in the bearing regardless of the orbit inclination.
Furthermore, the speed-dependent rotor natural frequencies from the Campbell
diagrams have been superimposed on the waterfall diagrams. This serves to aid the
interpretation of some vibrational components, but it should be emphasized that
the stiffness and damping properties of GFBs are amplitude dependent meaning
that significant oscillations will cause a shift in the effective natural frequencies
compared to those obtained from eigenvalue analysis (for infinitesimal oscillations).

Figures P6.20(a) and P6.20(b) show the waterfall diagrams with and without
injection for u A = 2.5 g mm, u B = −2.5 g mm being of the order of the residual
unbalance. For both cases, it can be observed that the synchronous vibration
at the critical speeds corresponding to the conical second and fourth modes are
higher than for the cylindrical first and third modes. This is to be expected from
the counter-phase arrangement of the unbalance masses. It should also be noticed
that the integrator has been able to reach the requested end time of 2.34 s up
to 31 kRPM for the passive case in fig. P6.20(a) and up to 37.5 kRPM for the
hybrid case in fig. P6.20(b) despite these having surpassed the predicted stability
limits of ΩOSI = 29.65 kRPM and ΩOSI = 36.24 kRPM. As the time integrator has
been confirmed to eventually crash (in both cases) if continuing the integration
for a longer period above the predicted OSI, this indicates the instability enters
relatively slowly.
By increasing the bearing A unbalance to u A = 20 g mm, a more realistic
level of unbalance is obtained for which the predicted unbalance responses are
shown in figs. P6.21(a) and P6.21(b). For both the passive and hybrid case,
the time integrator is now crashing approximately 500 RPM before the predicted
stability limits. From the passive case waterfall, the synchronous vibration is
clearly amplified across all four critical speeds with a maximum amplitude of
approximately 7.5 µm and settling at around 3 µm for higher speeds. Notice,
however, that the peak amplitudes are subject to some uncertainty since the height
and width of each peak depends on the exact location relative to the frequency bins.
Contrary to the essentially linear response observed for the lower unbalance level,
a number of sub-synchronous components now appear in the range 11–12.5 kRPM
with a maximum amplitude of around 10.5 µm at 12 kRPM and 88.5 Hz. From the
hybrid case response in fig. P6.21(b), the synchronous amplitude across the critical
speeds is comparable to the passive case, but slightly increased when passing the
fourth critical speed reaching a maximum amplitude of approximately 9 µm. The
sub-synchronous vibrations appear slightly later in the range 12–13 kRPM and with
significantly reduced amplitudes reaching a maximum of 3.5 µm at 12.5 kRPM and
92.8 Hz. The hybrid operational mode thus manages to reduce the sub-synchronous
vibration amplitude by around 66 % while maintaining a comparable synchronous
amplitude.
If further increasing the bearing A unbalance to u A = 60 g mm, the response
becomes even richer with sub-synchronous vibrations appearing over a wide speed
range. This is shown for the passive case in fig. P6.22(a) and hybrid case in
fig. P6.22(b). At this unbalance level, which is considered high for an actual
machine, the integrator was unable to reach the requested end time already
3–5 kRPM before the predicted OSIs. This, however, still means that the in-
tegration was successful for the hybrid case for 4.5 kRPM after the passive case
)
2.0
µm
(a) 1.6
p(
1.2
P6.4 Results
ωi
m
0.5X 1X
0.8
A
0.4
0.0
30
25
20
15
Ω (kRPM)
10
0 100 200 300 400 500

Frequency (Hz)
Figure P6.20: Bearing A unbalance responses (major axis amplitude) for the passive and hybrid cases based on 1.84 s
of simulation (after having discarded 0.5 s) for every 250 RPM with unbalances u A = 2.5 g mm and u B = −2.5 g mm:
(a) Passive case spanning 3–31 kRPM as the integrator was unable to reach the requested end time (2.34 s) from
31.25 kRPM and upwards. The natural frequencies from fig. P6.15(a) have been superimposed (black lines). (b)
Hybrid case spanning 3–37.25 kRPM as the integrator was unable to reach the requested end time (2.34 s) from
37.5 kRPM and upwards. The natural frequencies from fig. P6.15(b) have been superimposed (black lines). Figure
continued on the following page.
237
)
µm
(b) 2.0
1.6
p(
0.5X 1X ωi 1.2
Am
0.8
0.4
0.0
35
30
Ω (kRPM)
25
20
15
10
5
0 100 200 300 400 500
Frequency (Hz)
238
)
µm
10
(a) 8
p(
6
P6.4 Results
ωi
m
0.5X 1X
4
A
2
0
25
20
15
Ω (kRPM)
10
0 100 200 300 400 500

Frequency (Hz)
of simulation (after having discarded 0.5 s) for every 250 RPM with unbalances u A = 20 g mm and u B = −2.5 g mm:
(a) Passive case spanning 3–29 kRPM as the integrator was unable to reach the requested end time (2.34 s) from
29.25 kRPM and upwards. The natural frequencies from fig. P6.15(a) have been superimposed (black lines). (b)
Hybrid case spanning 3–35.5 kRPM as the integrator was unable to reach the requested end time (2.34 s) from
35.75 kRPM and upwards. The natural frequencies from fig. P6.15(b) have been superimposed (black lines). Figure
continued on the following page.
239
)
µm
(b) 10
p(
8
0.5X 1X ωi 6
Am
4
2
0
35
30
25
Ω (kRPM)
20
15
10
5
0 100 200 300 400 500
Frequency (Hz)
240
P6.4 Results 241
crashed. For the passive case, the synchronous vibration reaches its maximum
amplitude of approximately 21.5 µm at 7 kRPM and 117 Hz, while being close
to 20 µm across the critical speeds. Sub-synchronous components emerge in the
range 8.25–19.75 kRPM with the main branches bifurcating and reuniting several
times in-between. The maximum sub-synchronous amplitude of 28.5 µm appears
at 7 kRPM and 117 Hz. For the hybrid case in fig. P6.22(b), the synchronous vibra-
tion remains similar to the passive case with amplitudes close to 20 µm throughout
the critical speed range, but with a slightly increased maximum amplitude of
approximately 23 µm. Sub-synchronous vibrations appear at 9.5–20.75 kRPM, but
with a maximum amplitude of around 22 µm at 14.25 kRPM and 104 Hz. Though
these are still very high vibrational amplitudes unlikely to be tolerated in actual
operation, the sub-synchronous amplitude is hence reduced by around 20 % in the
hybrid operational mode.
P6.4.5 Sub-synchronous Vibration Bifurcation Points

As evident from the presented unbalance responses as well as previous studies, the
appearance of sub-synchronous vibrations in GFB supported rotor–bearing systems
depends on the excitation level. When increasing the unbalance, sub-synchronous
components initially appear at low speeds and gradually propagate to higher speeds
with various frequency branches alternately emerging, bifurcating, reuniting, and
disappearing at seemingly sporadic speeds. Figure P6.23(b) contains an alternative
version of the hybrid case unbalance response from fig. P6.22(b), likewise with
the natural frequencies from the Campbell diagram in fig. P6.15(b) superimposed.
From this view, it is noticeable how the bifurcation and reunifications of the sub-
synchronous branches coincide with the 0.5X crossings of the natural frequencies.
The first bifurcation at 11.25 kRPM happens as the 0.5X speed intersects the
first natural frequency, while the two emerged branches fade out and reunite at
14.75 kRPM as the 0.5X speed crosses the second natural frequency. A second
bifurcation occurs as the 0.5X speed intersects the third natural frequency and the
sub-synchronous vibrations disappear just before the crossing of the fourth natural
frequency. Following the trend from lower excitation levels, it is quite possible
that a higher, or differently distributed, unbalance excitation would further extend
the two sub-synchronous branches to actually reach the crossing. Even though the
transitions are less well-defined, possibly due to the higher vibrational amplitudes
affecting the natural frequencies, a similar pattern can be observed from the
corresponding plot for the passive case in fig. P6.23(a). For the lower unbalance
level with u A = 20 g mm shown in figs. P6.21(a) and P6.21(b), only a subset of the
sub-synchronous vibrations is visible, but the two visible branches do, however,
follow the same tendency and appear between the 0.5X speed’s intersections of ω1
and ω2 .
)
µm
(a) 25
p(
20
0.5X 1X ωi 15
Am
10
5
0
25
20
Ω (kRPM)
15
10
5
0 100 200 300 400 500
Frequency (Hz)
of simulation (after having discarded 0.5 s) for every 250 RPM with unbalances u A = 60 g mm and u B = −2.5 g mm: (a)
Passive case spanning 3–26.25 kRPM. At 26.5 kRPM, a large orbit dominated by a 139 Hz cylindrical forward whirl
with an amplitude around 70 µm appears which has been omitted from the plot. From 26.75 kRPM and upwards,
the integrator was unable to reach the requested end time (2.34 s). The natural frequencies from fig. P6.15(a) have
been superimposed (black lines). (b) Hybrid case spanning 3–31 kRPM as the integrator was unable to reach the
requested end time (2.34 s) from 31.25 kRPM and upwards. The natural frequencies from fig. P6.15(b) have been
superimposed (black lines). Figure continued on the following page.
242
P6.4 Results
)
(b) 25 µm
p(
20
ωi
m
0.5X 1X 15
A
10
5
0
30
25
20
15
Ω (kRPM)
10
0 100 200 300 400 500

Frequency (Hz)
243
For the presently studied system, it has thus been observed that the speed range
can be divided into five characteristic regions based on the natural frequencies
with distinct sub-synchronous behaviours as listed in table P6.7. The addition of
injection, especially if actively controlled, would allow the natural frequencies and
hence the 0.5X–ωi intersections to be manipulated. If the observed significance of
these intersections could be proven to hold in general, this would provide valuable
insights which could be exploited to suppress sub-synchronous vibrations through
controller design.
P6.5 Conclusion
The paper has presented simulated Campbell diagrams, multi-domain mode shapes
and unbalance responses for a previously studied rotor–bearing system supported
by two industrial three-pad GFBs with and without injection. The passive case, i.e.
without injection, is directly comparable to results from the literature [16] and thus
serves to validate the model. Despite modelling differences (fluid film discretization
scheme, equation format, and top foil model), the comparison demonstrated the
models agree well with all four rotor mode natural frequencies matching within
0.33 %. As a by-product of the comparison, the additional stiffness contributed by
the presently used beam element top foil was shown to adversely affect the OSI
(by approximately 2 %) matching previous findings by the authors [128, 129].
Using the validated passive results as a point of departure, the main contribution
of the present work is the addition of injection through a parameter-map approach.
As an alternative to the usually applied analytical mass flow expressions based on
Table P6.7: Characteristic speed ranges for the present system (passive as well
as hybrid) with respect to sub-synchronous vibrations observed from the combined
Campbell diagrams and unbalance response plots in figs. P6.23(a) and P6.23(b).
Speed range
Appearance of sub-synchronous vibrations
Ωstart Ωend
0 0.5X ∩ ω1 Absent or limited to half-speed whirl
0.5X ∩ ω1 0.5X ∩ ω2 Multiple components with two dominant branches
gradually diverging from 0.5X
0.5X ∩ ω2 0.5X ∩ ω3 Half-speed whirl, possibly with minor neighbouring
components
0.5X ∩ ω3 0.5X ∩ ω4 Multiple components with two dominant branches at
near-constant speed ratios.
0.5X ∩ ω4 ΩOSI None
25
(a)
101
P6.5 Conclusion
20
100
15
Ω (kRPM)
Amplitude (µm)
10 10−1
5 0.5X 1X ωi
10−2
0 100 200 300 400 500
Frequency (Hz)
Figure P6.23: Alternative views of the unbalance responses shown in fig. P6.22, i.e. major axis amplitudes in bearing
A based on 1.84 s of simulation (after having discarded 0.5 s) for every 250 RPM with unbalances u A = 60 g mm and
u B = −2.5 g mm: (a) Passive case response covering 3–26.5 kRPM as the integrator was unable to reach the requested
end time (2.34 s) from 26.75 kRPM and upwards; and (b) hybrid case response covering 3–31 kRPM as the integrator
was unable to reach the requested end time (2.34 s) from 31.25 kRPM and upwards. For both cases, the natural
frequencies from fig. P6.15 are superimposed (black lines) and logarithmic colour maps are used to visually amplify
low-amplitude details. Figure continued on the following page.
245
30
(b)
101
25
Amplitude (µm)
Ω (kRPM)
20 100
15
10−1
10
5 0.5X 1X ωi
10−2
0 100 200 300 400 500
Frequency (Hz)
246
P6.5 Conclusion 247
compressibility limited isentropic flow or viscosity limited Hagen–Poiseuille flow,

the mass flow through the injector is obtained using a three-dimensional steady
state CFD sub-model for the injector and a small part of the fluid film. Using
this, the injection mass flow was evaluated for 2940 cases across a five-dimensional
structured parameter space spanned by rotational velocity, injector clearance,
equivalent eccentricity, fluid film pressure and injection pressure. Compared to
the targeted rotor–bearing simulations, the map is computationally expensive, but
it only needs to be evaluated once for a given injector geometry. Subsequently,
the injected mass flow at a given configuration can be obtained efficiently using
an n-dimensional piecewise multi-linear interpolation scheme from the literature
[180]. Importantly, this also provides the mass flow state derivatives required to
assemble the rotor–bearing system Jacobian matrices and needed, in the future,
for controller design.
In the presented hybrid case, the rotor–bearing system is imagined to be
augmented with three piezoelectrically controlled injection valves from an existing
injection system designed for active control of a rigid gas bearing supporting a
flexible shaft. Results are presented for a regulated hybrid operational mode with
only the topmost injector valve open, similar to the configuration studied by Yazdi
and Kim [182], and for an aerostatic case with all injectors open at 0 RPM. The
mode shapes of the hybrid case were shown to be similar to the passive case,
except for the higher loaded third pad (where injection is added). From Campbell
diagrams, the OSI was found to be increased by 22 % from 29.65 kRPM in the
passive case to 36.24 kRPM in the hybrid case. Additionally, the hybrid injection
scheme was shown to have a positive effect on the appearance of sub-synchronous
vibrations. The sub-synchronous amplitude was thus observed to be reduced by
approximately 66 % for a medium unbalance level while approximately 20 % for
a high unbalance level, both at the expense of a slightly increased synchronous
amplitude. At 20 kRPM, the steady state mass flow supplied in the hybrid mode to
obtain these results was estimated to require of the order of 100 W which should be
compared to the 120 kW power rating of the imitated turbo blower. Furthermore,
it is shown that the 21.17 kg rotor can be levitated at 0 RPM with less than 4 bar
supplied to each injector. It can thus be concluded that injection has significant
potential to manipulate the GFB fluid film properties with reasonable power
requirements even for a relatively heavy rotor. Adding feedback to actively control
the injection flows, this potential could be further exploited to achieve additional
enhancements. In this setting, it should be highlighted that the current model is
capable of producing the injection pressure and nozzle clearance state derivatives,
thus paving the way for standard model-based controller synthesis methods to be
applied.
As a final remark, it should be noted that the included foil model is essentially
one-dimensional assuming axially uniform foil deformation. Considering the high
axial pressure gradients resulting from the injected air flow, this assumption
might be inadequate. Further work should hence be conducted to include a
three-dimensional shell element top foil allowing the axial film height variations
to be captured. Importantly, this should include membrane stiffness so as not to
overestimate the axial variation [23].
Acknowledgements
The work is financially supported by the Department of Mechanical Engineering at
the Technical University of Denmark for the project ”From Passive to Controllable
Gas Foil Bearings – Modelling & Control Design”.
P6.A Isentropic Injection Flow

A commonly applied gas injection model is based on the assumptions of ideal gas
behaviour and isentropic conditions. This allows the density ρ, temperature T,
and pressure p at one state to be related to another state 0 as [86]
γ γ−1 γ−1
p ρ T ρ p γ
= , and = = , (P6.A1)
p0 ρ0 T0 ρ0 p0
while the specific heat capacity at constant pressure Cp and speed of sound c can
be expressed as
γRs
Cp = and c = γRsT, (P6.A2)
p
γ−1
where Rs and γ are the specific gas constant and the ratio of specific heat capacities.
Combining these with the definition of stagnation enthalpy
1
hent,0 = hent + u2, (P6.A3)
2
the Mach number M in a gas flowing from a reservoir with stagnation properties
(subscript 0) can be expressed as
v
γ−1
u
t !
2 p0 γ
M= −1 , (P6.A4)
γ−1 p
from which it follows that the limiting condition for a subsonic flow is given as
γ
γ−1
p 2
M ≤1⇔ ≥ ≈ 0.5283 for dry air with γ = 1.4. (P6.A5)
p0 γ+1
P6.B Rotor Model Matrices and Forces 249
Equations (P6.A1) to (P6.A5) allow the flow mass leaving the reservoir through
a duct with restricting cross sectional area A to be written for the two cases of
subsonic and choked, i.e. maximum, flow conditions as
s
 γ2 1+γ γ
√
γ
γ−1
2γ p p
for 1 ≥ p
 2
ApM γ − ≥

Ap0 γ−1 γ+1

p0 p0 p0
mÛ γ = √ =√ .


γ
RsT RsT0  q 1
γ−1 γ−1
 2γ 2
for p 2


 γ+1 γ+1 p0 ≤ γ+1
(P6.A6)
Though presenting a convenient closed-form expression for the injection mass
flow, the assumptions made to produce eq. (P6.A6) are rather comprehensive.
Conditions in the supply system are assumed stagnant and the flow is assumed
adiabatic, steady, and inviscid with a clearly defined choking area A. When
applying eq. (P6.A6) in practice [6, 30, 57, 60, 76, 85, 107, 109, 120, 184], it is
common to apply a correction factor based on CFD simulations or other fluid
mechanical considerations. These corrections are typically denoted ”discharge
coefficients”, ”inherent compensation factors”, ”adiabatic efficiency factors” or
”vena contracta coefficients” depending on the physical phenomena compensated
for.
P6.B Rotor Model Matrices and Forces

The rotor equations are non-dimensionalized using a rotor mass scale mrs , rotor
length scale lrs , and the system time scale ωτ−1 . The mass matrix, gyroscopic
matrix, unbalance force vector and static load vector can thus be written as
 l 2 mr + I yy 0 l1 l2 mr − I yy 0 
 2
l2 mr + I x x
2

1 0 0 l1 l2 mr − I x x 
M̃r = ,

l12 mr + I yy

2
mrs l12 l1 l2 mr − I yy 0 0
2m + I
 

 0 l1 l2 mr − I x x 0 l1 r x x


cos τ 

 0 −Izz 0 Izz  
uA
 
Ω̃ τ (P6.B1)
 2
 
1  Izz 0 −Izz 0  sin
 
G̃r = , f̃ub (τ) = ,

 

cos τ 

2
mrs l12  0 Izz 0 −Izz  mrs lrs 

−Izz u
 B sin τ 
 
0 Izz 0 
 
  
 l2 

 

mr g 0

= ,
 

and f̃rw
mrs lrs ωτ2 l12 
 l1 

0
 
 
where l12 = l1 + l2 is the distance between the bearing locations. In the present
work, mrs = mr = 21.17 kg and lrs = 1 mm has been found to provide reasonable
conditioning numbers. The time scale is, as for the foil and fluid domains, defined
from the rotor angular velocity as ωτ = Ω, but this is kept adjustable to allow
aerostatic simulations or time integrations with varying speed.
P6.C Foil Structure Matrices and Pressure Load

Mapping
The foil structure is modelled using two-node Euler–Bernoulli beam elements with
rotational and lateral translation DOFs only. The element DOF vector and the
associated shape functions in the isoparametric coordinate ξ can hence be written
as
(P6.C1)
T
xef = d1e φ1e d2e φ2e ,

1 le 1 le
N (ξ) = (2 + ξ)(ξ − 1) 2
(1 + ξ)(ξ − 1)2
(2 − ξ)(1 + ξ)2
(ξ − 1)(1 + ξ) ,
2
4 8 4 8
(P6.C2)
where le is the element length. The relation between the physical coordinate x
and isoparametric coordinate ξ is given as ξ = 2x/le − 1 implying ∂ξ/∂ x = 2/le .
Using this, the curvature–displacement matrix can be calculated as
∂ 2 N (ξ) 4 ∂ 2 N (ξ) 6ξ 3ξ + 1

6ξ 3ξ − 1
B= = 2 = 2 − 2 , (P6.C3)
∂ x2 le ∂ξ 2 le le le le
from which the stiffness matrix of a beam element with Young’s modulus Et , width
L (equalling the bearing length), thickness tt , and cross-sectional moment of inertia
It = 12 tt L can be integrated as
1 3
12 6le −12 6le 

1
4le 2 −6le 2le 2 
∫  
le Et It Et It 
Kt =
e
B Bdξ = 3 
T
, (P6.C4)
2 −1 le  12 −6le 

 4le 2 
where the sub-diagonals are given from symmetry. The top foil beam elements are
furthermore supported by the bump foil modelled using the SEFM with stiffness k
and loss factor η. A viscous damping coefficient is calculated from the loss factor
at the foil oscillation frequency ω f . To represent the SEFM support, recall the
work equivalent consistent nodal forces produced by a distributed load q to be
defined as
le 1 T
∫
rq =
e
N q(ξ)dξ, (P6.C5)
2 −1
P6.C Foil Structure Matrices and Pressure Load Mapping 251
and note that for a given beam element deformation and velocity, the load (per
unit length) exerted by the SEFM support on the beam will be given as q =
LkNxef + Lkηω−1f NxÛ ef . Inserting this into eq. (P6.C5), the nodal loads resulting
from the SEFM support can be written as
le Lk
∫ 1 le Lkηω−1 ∫ 1
(P6.C6)
f
e
rSEFM = N NdξT
xef + NT Ndξ xÛ ef ,
2 −1 2 −1
| {z } | {z }
e
KSEFM e
DSEFM
providing the element SEFM stiffness and damping matrices. These can be readily
integrated to provide
156 22le 54 −13le 
1
 
4le2 13le −3le2 
∫
le Lk le Lk 
e
KSEFM = N Ndξ =
T
, (P6.C7)
2 −1 420  156 −22le 

 4le2 
e
DSEFM =ηω−1 f KSEFM,
e
(P6.C8)
where the sub-diagonals are given from symmetry.

Denoting the load (per unit length) on a beam element from the fluid film
pressure as qp (ξ), the work equivalent nodal loads due to the fluid film pressure
on the top foil can be obtained using eq. (P6.C5) as
∫ 1 n
−le L −le L Õ
f pe = T
N qp (ξ)dξ ≈ − wi NT qp (ξi ) , (P6.C9)
2 −1 2 i=1
where the approximate equality represents an n-point Gaussian quadrature with

weights wi and sample points ξi . At each sample point, the fluid film pressure load
is calculated as the axial average of p − pa as
−1 L
∫
qp (ξi ) = p θ ξi , z − pa dz

L 0
nz +1 (P6.C10)
1 L
∫
1 Õ
=pa − p θ ξi , z dz ≈ pa − ∆zv, j p θi, zv, j ,

L 0 L j=1
where θ ξi is the circumferential position coincident with ξi and the approximate

equality represents a trapezoidal rule summing the vertex pressures at the axial
Ínz +1
positions zv, j using the weights ∆zv, j (note that j=1 ∆zv, j = L). Last, the
p θi, zv, j pressures themselves need to be reconstructed from the CV centre

positions and BCs. This is achieved using CDS or LUDS depending on the location
θi, zv, j in the mesh. Combining the Gauss scheme summation from eq. (P6.C9),

the trapezoidal summation from eq. (P6.C10) and the Finite Difference (FD)
scheme summations needed to obtain each p θi, zv, j value, the loads can be
written as a constant mapping from the CV centre pressures in the relevant
bearing α as
f̃ pe = f̃ p0
e
+ Hefp,p p̃α, (P6.C11)
where the necessary non-dimensionalizations have also been incorporated. No-

tice that the shape functions contain cubic terms while the fluid film pressure
distribution will generally be linear between the adjacent cell centres for the
current FV scheme. If, for a given top foil beam element, all circumferential
pressure samples qp (ξi ) in eq. (P6.C9) are taken from a single CVC–CVC span,
the resulting integrand hence contains terms of order four requiring a three point
Gauss integration rule. It is, however, feasible to use much fewer circumferential
top foil divisions than fluid film CVs, hence the samples qp (ξi ) will generally be
extracted from multiple circumferential CVC–CVC spans. Strictly speaking, the
resulting qp (ξ) function will be non-polynomial, but the error can be lowered by
applying additional Gauss points. In the present implementation, a five-point
scheme has been found to perform satisfactorily, but since the mapping matrix is
only constructed once, the cost of adding additional points is low.
P6.D Modified Reynolds Equation Finite Volume

Residual Derivatives
The steady state system Jacobian structure is given in eq. (P6.27), while the
additional components of the transient case Jacobians are given from eqs. (P6.28)
and (P6.29). The upper two rows contain the partial derivatives of the MRE FV
residual from eq. (P6.7) for each bearing. The state vector derivatives in bearing
α can be decomposed as
∂r̃FV,α ∂AFV p̃α, x̃r , x̃ f ,α p̃α ∂BFV xÛ̃ r , xÛ̃ f ,α p̃α ∂ m

Û̃ inj,A p̃ A, x̃r , x̃ f ,A

= + + ,
∂p̃α ∂p̃α ∂p̃α ∂p̃α
| {z } | {z } | {z }
I II III
(P6.D1)
∂r̃FV,α ∂AFV p̃α, x̃r , x̃ f ,α p̃α ∂CFV x̃r , x̃ f ,α pÛ̃ α ∂ m

= + + ,
∂x̃ f ,α ∂x̃ f ,α ∂x̃ f ,α ∂x̃ f ,α
| {z } | {z } | {z }
IV V VI
(P6.D2)
∂r̃FV,α ∂AFV p̃α, x̃r , x̃ f ,α p̃α ∂CFV x̃r , x̃ f ,α pÛ̃ α ∂ m


= + + ,
∂x̃r ∂x̃r ∂x̃r ∂x̃r
| {z } | {z } | {z }
VII VIII IX
(P6.D3)
while those with respect to the temporal state vector derivative can be written as
∂r̃FV,α ∂CFV x̃r , x̃ f ,α pÛ̃ α

= , (P6.D4)
∂ pÛ̃ α ∂ pÛ̃ α
| {z }
X
∂r̃FV,α ∂BFV xÛ̃ r , xÛ̃ f ,α p̃α

= , (P6.D5)
∂ xÛ̃ f ,α ∂ xÛ̃ f ,α
| {z }
XI
∂r̃FV,α ∂BFV xÛ̃ r , xÛ̃ f ,α p̃α

= . (P6.D6)
∂ xÛ̃ r ∂ xÛ̃ r
| {z }
XII
Term I contains the pressure derivatives stemming from the Poiseuille and
Couette boundary fluxes. This can be assembled from the CV contributions as
∂ ∂ ∂ ∂

 q̃θ |1,e q̃θ |1,w q̃ z |1,n q̃ z |1,s

 ∆z̃1 − + ∆θ̃ 1 − 
 ∂p̃α ∂p̃α ∂p̃α ∂p̃α 
∂AFV p̃α  
.

= .
. ,

∂p̃α
∂ q̃θ |nCV,e ∂ q̃θ |nCV,w ∂ q̃z |nCV,n ∂ q̃z |nCV,s 
 
+ ∆θ̃ nCV

∆z̃nCV − −
| {z } 
∂p̃ ∂p̃ ∂p̃ ∂p̃

nCV ×nCV
α α α α

(P6.D7)
 
where each of the four ∂/∂p̃α terms for each CV holds three non-zero contributions
at any given time. Recalling the flow vector definition from eq. (P6.3), this can be
illustrated by expanding the eastern face flux term contribution to the i-th CV
residual as
∂ q̃θ |i,e ∂ ∂ p̃

= 3
− p̃|i,e h̃ |i,e + Sτ Ω̃ p̃|i,e h̃|i,e , (P6.D8)
∂p̃α ∂p̃α ∂θ i,e
| {z } | {z }| {z }
1×nCV Poiseuille flow Couette flow
where the face pressure and gradient are calculated from the surrounding CV
centre pressures (or BCs) using FD schemes. The gradient is reconstructed using
CDS providing dependencies on the i-th CV itself and its eastern neighbour which
are hence assigned non-zero coefficients ai,e, j calculated from the cell dimensions.
The face pressure is reconstructed using LUDS to provide upwinding. Depending
on the flow direction, the eastern face pressure gradient is thus depending either
on the i-th CV itself and its western neighbour or on the two eastern neighbours
which are hence assigned non-zero coefficients bi,e, j . If source terms are furthermore
included to represent boundary conditions, the FD reconstructions can be written
as
∂ p̃
(P6.D9)
Õ Õ
p̃|i,e = s p,i,e + ai,e, j p̃ j and = sdp,i,e + bi,e, j p̃ j ,
j=1...nCV
∂θ i,e j=1...nCV
where it should be noted that all coefficients and source terms must be calculated
separately for each face since the mesh is non-uniform. As the stencil is potentially
varying due to the upwinding scheme, the bi,e, j coefficients should be continuously
updated. Using eq. (P6.D9), the Jacobian contribution from the j-th CV centre
pressure to the i-th FV residual due to the eastern face flux term can thus be
written as
∂ q̃θ |i,e ∂ p̃

= − h̃ |i,e ai,e, j
3
+ bi,e, j p̃|i,e + Sτ Ω̃ai,e, j h̃|i,e, (P6.D10)
∂ p̃ j ∂θ i,e
while the contributions stemming from the remaining face fluxes q̃θ |i,w , q̃z |i,n , and
q̃z |i,s in eq. (P6.D7) can be formulated analogously. Adding these four contributions,
the i-th FV residual will thus be depending on the i-th CV itself and its four
immediate neighbours due to the CDS reconstructed face pressure gradient and
additionally on two of the four secondary neighbours due to the LUDS scheme for
the face gradients.
For transient conditions, the squeeze contribution is non-zero, meaning that
term II should be included. Assembling the contributions from all CVs, this
provides the diagonal matrix
∂BFV p̃α
= 2Sτ diag ∆z̃1 ∆θ̃ 1 hÛ̃ 1, . . . , ∆z̃nCV ∆θ̃ nCV hÛ̃ nCV . (P6.D11)
∂p̃α
| {z }
nCV ×nCV
The foil deformation derivatives of the Poiseuille and Couette boundary fluxes
in term IV and the local expansion in term V can be assembled from the CV
contributions as
∂AFV p̃α ∂CFV pÛ̃ α

+ =
∂x̃ f ,α ∂x̃ f ,α
| {z }
nCV ×nfdof
∂ q̃θ |1,e ∂ q̃θ |1,w ∂ q̃z |1,n ∂ q̃z |1,s

+ θ̃
 
∆z̃ 1 − ∆ 1 −
∂x̃ f ,α ∂x̃ f ,α ∂x̃ f ,α ∂x̃ f ,α
 
 
∂ (P6.D12)
 
h̃ 1
+2S θ̃
 
 τ ∆z̃ 1 ∆ 1 pÛ̃1 
 ∂x̃ f ,α

..
 
. ,
 

 
+ ∆θ̃ nCV

∆z̃nCV − −
∂x̃ f ,α ∂x̃ f ,α ∂x̃ f ,α ∂x̃ f ,α 


∂ h̃nCV

 
 +2S τ ∆z̃ n ∆ θ̃ n pÛ̃ n



CV CV
∂x̃ f ,α CV 

where ∂ h̃i /∂x̃ f ,α represents the row in HCVC,x f corresponding to the i-th CV. The
construction of the remaining terms can be exemplified by expanding the eastern
face flux term as
∂ q̃θ |i,e ∂ ∂ p̃

= 3
− p̃|i,e h̃ |i,e + Sτ Ω̃ p̃|i,e h̃|i,e
∂x̃ f ,α ∂x̃ f ,α ∂θ i,e
| {z } | {z }| {z }
1×nfdof Poiseuille flow Couette flow
(P6.D13)
∂ p̃ ∂ h̃|i,e

= p̃|i,e −3 h̃ |i,e i,e + Sτ Ω̃
2
,
∂θ ∂x̃ f ,α
|{z}
1×nfdof
where face pressure and gradients are calculated using eq. (P6.D9) and ∂ h̃|i,e /∂x̃ f ,α
is given as the row in HCVE,x f corresponding to the eastern face of the i-th CV.
The rotor displacement derivatives from term VII and VIII can be constructed
analogously as
∂AFV p̃α ∂CFV pÛ̃ α

+ =
∂x̃r ∂x̃r
| {z }
nCV ×nrdof
∂ q̃θ |1,e ∂ q̃θ |1,w ∂ q̃z |1,n ∂ q̃z |1,s

 
 ∆z̃ 1 − + ∆ θ̃ 1 − 

 ∂x̃ r ∂x̃ r ∂x̃ f ,α ∂x̃r


 ∂ h̃1 Û̃ 
(P6.D14)
 +2Sτ ∆z̃1 ∆θ̃ 1 p1 
∂x̃r
 
..
 
,
.
 

 
+ θ̃

∆z̃nCV − ∆ n −
∂x̃r ∂x̃r ∂x̃r ∂x̃r
CV

 
∂
 
 h̃ n 
 +2S τ ∆z̃ n ∆ θ̃ n
CV Û̃
p n



CV CV
∂x̃ r
CV 

where ∂ h̃i /∂x̃r represents the row in HCVC,xr corresponding to the i-th CV, while
the eastern face flux term is given as
∂ q̃θ |i,e ∂ ∂ p̃

= 3
− p̃|i,e h̃ |i,e + Sτ Ω̃ p̃|i,e h̃|i,e
∂x̃r ∂x̃r ∂θ i,e
| {z } | {z }| {z }
1×nrdof Poiseuille flow Couette flow
(P6.D15)
∂ p̃ ∂ h̃|i,e

= p̃|i,e 2
−3 h̃ |i,e + Sτ Ω̃ ,
∂θ i,e ∂x̃r
|{z}
1×nrdof
and ∂ h̃|i,e /∂x̃r represents the row of HCVE,xr corresponding to the eastern face of
the i-th CV. To calculate the injection derivatives from term III, VI, and IX, the
injector to CV mass flux mapping from eq. (P6.24), the interpolation function from
eq. (P6.21) and the state to injection map variable mappings from eqs. (P6.22)
and (P6.23) can be combined to express the state dependencies explicitly as
Û̃ α = Hm̃ ,mÛ̃ finj (hinj,α, pinj,α, pfilm0 + H pfilm,p p̃α, ε x0 + Hε,xr x̃r + Hε,x f x̃ f ,α, Ω).
m inj
| {z } | {z }
pfilm,α ε x,α
(P6.D16)
The partial derivatives can thus be written for bearing α as the matrix products
∂mÛ̃ α ∂mÛ̃ inj,α ∂finj

= Hm̃inj,mÛ̃ = Hm̃inj,mÛ̃ H pfilm,p, (P6.D17)
∂p̃α | {z } ∂p̃α | {z } ∂pfilm,α | {z }
|{z} | {z } | {z } n ×n
nCV ×ninj nCV ×ninj inj CV
nCV ×nCV ninj ×nCV ninj ×ninj
∂m Û̃ α ∂mÛ̃ inj,α ∂finj

= Hm̃inj,mÛ̃ = Hm̃inj,mÛ̃ Hε,x , (P6.D18)
∂x̃ f ,α | {z } ∂x̃ f ,α | {z } |{z}∂ε x,α |{z}f
n ×n
|{z} | {z }
nCV ×ninj nCV ×ninj inj fdof
ninj ×nfdof ninj ×nfdof ninj ×ninj
∂mÛ̃ α ∂mÛ̃ inj,α ∂finj

= Hm̃inj,mÛ̃ = Hm̃inj,mÛ̃ Hε,x , (P6.D19)
∂x̃r | {z } ∂x̃r | {z } ∂ε x,α |{z}r
|{z} | {z } |{z} ninj ×nrdof
nCV ×nrdof nCV ×ninj n nCV ×ninj
inj ×nrdof ninj ×ninj
where the mapping matrices Hm̃inj,mÛ̃ , H pfilm,p , Hε,x f , and Hε,xr are constant and
thus can be calculated offline, while the (diagonal) ninj × ninj injection Jacobians
∂finj /∂pfilm,α and ∂finj /∂ε x,α must be updated continuously from the interpolation
function. Notice that a naive dense treatment of the triple matrix products in
eqs. (P6.D17) to (P6.D19) would require nCV 2 n3 , n2 n3 , and n2 n3
inj CV fdof CV rdof multiplic-
ations, respectively. As these must be performed for every evaluation of the system
Jacobian, care should be taken to do this efficiently while exploiting the sparsity
patterns. In the present implementation, this is achieved using the compressed
sparse row format and a multiplication algorithm given by Bank and Douglas [8].
The terms X, XI, and XII are non-zero for transient conditions only. Term X
is the derivative of the MRE’s local expansion term with respect to the temporal
pressure derivative. For bearing α this provides the diagonal matrix
∂CFV pÛ̃ α
= CFV = 2Sτ diag ∆z̃1 ∆θ̃ 1 h̃1, . . . , ∆z̃nCV ∆θ̃ nCV h̃nCV . (P6.D20)

∂ pÛ̃ α
| {z }
nCV ×nCV
while the local expansion derivatives with respect to the foil and rotor displacements
from term XI and XII are given as
 ∂ hÛ̃ 1 
 ∆z̃1 ∆ θ̃ p̃
1 1

∂ x f ,α
Û̃
 
∂BFV p̃α
 
=2Sτ 
 .
..

 and
∂ xÛ̃ f ,α Û̃

∂h 
 
∆z̃nCV ∆θ̃ nCV p̃nCV nCV 
| {z } 
nCV ×nfdof
∂ xÛ̃ f ,α 
(P6.D21)


∂ hÛ̃ 1
 
∆z̃1 ∆θ̃ 1 p̃1
 
∂ Û̃
 
x
∂BFV p̃α r
 
=2Sτ 
 .
.. .

∂ xÛ̃ r 
∂ hÛ̃ nCV 
 
∆z̃nCV ∆θ̃ nCV p̃nCV
| {z } 
nCV ×nrdof


 ∂ xÛ̃ r 
where hÛ̃ i /∂ xÛ̃ f ,α and hÛ̃ i /∂ xÛ̃ r are given as the i-th rows of HCVC,x f and HCVC,xr
respectively.
References
Bibliography
[1] G. L. Agrawal. “Foil air/gas bearing technology - an overview”. In: Am. Soc.
Mech. Eng. ASME, 1997, p. 347. doi: 10.1115/97-GT-347.
[2] P. E. Allaire, J. K. Parsell, L. E. Barret, L. E. Barrett, and L. E. Barret. “A
pad perturbation method for the dynamic coefficients of tilting-pad journal
bearings”. In: Wear 72.1 (1981), pp. 29–44. doi: 10.1016/0043-1648(81
)90281-7.
[3] E. Anderson, Z. Bai, C. Bischof, L. S. Blackford, J. Demmel, J. Dongarra,
J. Du Croz, A. Greenbaum, S. Hammarling, A. McKenney, and D. Sorensen.
LAPACK Users’ Guide. 3rd ed. Philadelphia, PA: Society for Industrial
and Applied Mathematics, 1999. isbn: 978-0-89871-447-0. doi: 10.1137/1
.9780898719604.
[4] M. Arghir and O. Benchekroun. “A New Structural Bump Foil Model With
Application From Start-Up to Full Operating Conditions”. In: J. Eng. Gas
Turbines Power 141.10 (2019). doi: 10.1115/1.4044685.
[5] M. Arghir and O. Benchekroun. “A simplified structural model of bump-
type foil bearings based on contact mechanics including gaps and friction”.
In: Tribol. Int. 134.January (2019), pp. 129–144. doi: 10.1016/j.triboin
t.2019.01.038.
[6] M. Arghir, S. Le Lez, and J. Frêne. “Finite-volume solution of the com-
pressible Reynolds equation: linear and non-linear analysis of gas bearings”.
In: Proc. Inst. Mech. Eng. Part J J. Eng. Tribol. 220.7 (2006), pp. 617–627.
doi: 10.1243/13506501JET161.
[7] U. M. Ascher and L. R. Petzold. Computer Methods for Ordinary Differential
Equations and Differential-Algebraic Equations. January. Philadelphia, PA,
USA: Society for Industrial and Applied Mathematics, 1998, 314 s. isbn:
9780898714128. doi: 10.1137/1.9781611971392.
[8] R. E. Bank and C. C. Douglas. “Sparse matrix multiplication package
(SMMP)”. In: Adv. Comput. Math. 1.1 (1993), pp. 127–137. doi: 10.1007
/BF02070824.
260 Bibliography
[9] L. Barzem, B. Bou-Said, J. Rocchi, and G. Grau. “Aero-elastic bearing

effects on rotor dynamics: a numerical analysis”. In: Lubr. Sci. 25.7 (2013),
pp. 463–478. doi: 10.1002/ls.1218.
[10] C. Baum, H. Hetzler, and W. Seemann. “On the Stability of Balanced
Rigid Rotors in Air Foil Bearings”. In: 11th Int. Conf. Vib. Rotating Mach.
(SIRM 2015). Magdeburg, Germany, 2015, pp. 23–25.
[11] S. I. Bekinal, T. R. rao Anil, S. S. Kulkarni, and S. Jana. “Hybrid Permanent
Magnet and Foil Bearing System for Complete Passive Levitation of Rotor”.
In: Vib. Eng. Technol. Mach. Ed. by J. K. Sinha. Springer International
Publishing, 2015, pp. 939–949. doi: 10.1007/978-3-319-09918-7_83.
[12] G. Belforte, T. Raparelli, V. Viktorov, and A. Trivella. “Discharge coeffi-
cients of orifice-type restrictor for aerostatic bearings”. In: Tribol. Int. 40.3
(2007), pp. 512–521. doi: 10.1016/j.triboint.2006.05.003.
[13] F. Al-Bender. “On the modelling of the dynamic characteristics of aerostatic
bearing films: From stability analysis to active compensation”. In: Precis.
Eng. 33.2 (2009), pp. 117–126. doi: 10.1016/j.precisioneng.2008.06
.003.
[14] H. Blok and J. J. Van Rossum. “Foil bearing – New departure in hydro-
dynamic lubrication”. In: Lubr. Eng. 9.6 (1953), pp. 316–320.
[15] P. Bonello. “A new method for the calculation of the Campbell diagram of a
foil-air bearing rotor model”. In: 13th Int. Conf. Dyn. Rotating Mach. (SIRM
2019). Ed. by I. Santos. Kgs. Lyngby: Technical University of Denmark
(DTU), 2019, pp. 4–13. isbn: 978-87-7475-568-5.
[16] P. Bonello. “The extraction of Campbell diagrams from the dynamical
system representation of a foil-air bearing rotor model”. In: Mech. Syst.
Signal Process. 129 (2019), pp. 502–530. doi: 10.1016/j.ymssp.2019.04
.018.
[17] P. Bonello and H. M. Pham. “A receptance harmonic balance technique
for the computation of the vibration of a whole aero-engine model with
nonlinear bearings”. In: J. Sound Vib. 324.1-2 (2009), pp. 221–242. doi:
10.1016/j.jsv.2009.01.039.
[18] P. Bonello and H. M. Pham. “Nonlinear Dynamic Analysis of High Speed Oil-
Free Turbomachinery With Focus on Stability and Self-Excited Vibration”.
In: J. Tribol. 136.4 (2014), p. 41705. doi: 10.1115/1.4027859.
[19] P. Bonello and H. M. Pham. “The efficient computation of the nonlinear
dynamic response of a foil-air bearing rotor system”. In: J. Sound Vib.
333.15 (2014), pp. 3459–3478. doi: 10.1016/j.jsv.2014.03.001.
Bibliography 261
[20] R. W. Bosley and R. F. Miller. Hydrostatic augmentation of a compliant

foil hydrodynamic fluid film thrust bearing. US Patent 5,827,040 (filed Jun
14, 1996): United States Patent and Trademark Office, 1998.
[21] M. Branagan, D. Griffin, C. Goyne, and A. Untaroiu. “Compliant gas foil
bearings and analysis tools”. In: J. Eng. Gas Turbines Power 138.5 (2016),
pp. 1–8. doi: 10.1115/1.4031628.
[22] M. J. Braun, F. K. Choy, M. Dzodzo, and J. Hsu. “Two-dimensional dynamic
simulation of a continuous foil bearing”. In: Tribol. Int. 29.1 (1996), pp. 61–
68. doi: 10.1016/0301-679X(95)00035-3.
[23] M. Carpino, L. A. Medvetz, and J.-P. J.-P. Peng. “Effects of Membrane
Stresses in the Prediction of Foil Bearing Performance”. In: Tribol. Trans.
37.1 (1994), pp. 43–50. doi: 10.1080/10402009408983264.
[24] M. Carpino and J. P. Peng. “Theoretical performance of a hydrostatic foil
bearing”. In: J. Tribol. 116.1 (1994), pp. 83–89. doi: 10.1115/1.2927051.
[25] M. Carpino and G. Talmage. “A Fully Coupled Finite Element Formulation
for Elastically Supported Foil Journal Bearings”. In: Tribol. Trans. 46.4
(2003), pp. 560–565. doi: 10.1080/10402000308982664.
[26] M. Carpino and G. Talmage. “Prediction of Rotor Dynamic Coefficients in
Gas Lubricated Foil Journal Bearings with Corrugated Sub-Foils”. In: Tribol.
Trans. 49.3 (2006), pp. 400–409. doi: 10.1080/10402000600781416.
[27] V. Castelli and J. T. McCabe. “Transient Dynamics of a Tilting Pad Gas
Bearing System”. In: J. Lubr. Technol. 89.4 (1967), pp. 499–508. doi:
10.1115/1.3617043.
[28] A. Cerda Varela and I. F. Santos. “Validity of the modified Reynolds
equation for incompressible active lubrication”. In: Proc. Inst. Mech. Eng.
Part J J. Eng. Tribol. 230.12 (2016), pp. 1490–1502. doi: 10.1177/135065
0116638880.
[29] S. H. Chang, C. W. Chan, and Y. R. Jeng. “Numerical analysis of discharge
coefficients in aerostatic bearings with orifice-type restrictors”. In: Tribol.
Int. 90 (2015), pp. 157–163. doi: 10.1016/j.triboint.2015.04.030.
[30] C. H. Chen, T. H. Tsai, D. W. Yang, Y. Kang, and J. H. Chen. “The
comparison in stability of rotor-aerostatic bearing system compensated by
orifices and inherences”. In: Tribol. Int. 43.8 (2010), pp. 1360–1373. doi:
10.1016/j.triboint.2010.01.006.
[31] P. Y. P. Chen and E. J. Hahn. “Use of computational fluid dynamics in
hydrodynamic lubrication”. In: Proc. Inst. Mech. Eng. Part J J. Eng. Tribol.
212.6 (1998), pp. 427–436. doi: 10.1243/1350650981542236.
262 Bibliography
[32] R. D. Cook, D. S. Malkus, M. E. Plesha, and R. J. Witt. Concepts and

Applications of Finite Element Analysis. 4th ed. John Wiley & Sons, Inc.,
2001. isbn: 978-0-471-35605-9.
[33] V. Crastan and J. Devos. “Finite element solution of the three-dimensional
flow problem and of reynold’s equation for incompressible and compressible
fluids”. In: Nucl. Eng. Des. 22.2 (1972), pp. 225–232. doi: 10.1016/0029-
5493(72)90164-1.
[34] T. A. Davis and E. P. Natarajan. “Algorithm 907: KLU, A Direct Sparse
Solver for Circuit Simulation Problems”. In: ACM Trans. Math. Softw. 37.3
(2010), p. 17. doi: 10.1145/1824801.1824814.
[35] A. Delgado and B. Ertas. “Dynamic Force Coefficients of Hydrostatic Gas
Films for Recessed Flat Plates: Experimental Identification and Numerical
Predictions”. In: J. Tribol. 140.6 (2018), pp. 1–9. doi: 10.1115/1.4040114.
[36] A. Delgado and B. Ertas. “Dynamic Characterization of a Novel Externally
Pressurized Compliantly Damped Gas-Lubricated Bearing with Hermet-
ically Sealed Squeeze Film Damper Modules”. In: J. Eng. Gas Turbines
Power 141.2 (2019), pp. 1–10. doi: 10.1115/1.4041311.
[37] C. DellaCorte and M. J. Valco. “Load Capacity Estimation of Foil Air
Journal Bearings for Oil–Free Turbomachinery Applications”. In: Tribol.
Trans. 43 (2000), pp. 795–801.
[38] C. DellaCorte. “Oil-Free shaft support system rotordynamics: Past, present
and future challenges and opportunities”. In: Mech. Syst. Signal Process.
29 (2012), pp. 67–76. doi: 10.1016/j.ymssp.2011.07.024.
[39] C. DellaCorte and R. J. Bruckner. “Remaining Technical Challenges and
Future Plans for Oil-Free Turbomachinery”. In: J. Eng. Gas Turbines Power
133.4 (2011), p. 42502. doi: 10.1115/1.4002271.
[40] B. Ertas and A. Delgado. “Compliant Hybrid Gas Bearing Using Modular
Hermetically Sealed Squeeze Film Dampers”. In: J. Eng. Gas Turbines
Power 141.2 (2019), pp. 1–10. doi: 10.1115/1.4041310.
[41] A. Eshel. “Numerical solution of the planar hydrostatic foil bearing”. In: J.
Lubr. Technol. 101.1 (1979), pp. 86–91. doi: 10.1115/1.3453282.
[42] A. Fatu and M. Arghir. “Numerical Analysis of the Impact of Manufacturing
Errors on the Structural Stiffness of Foil Bearings”. In: J. Eng. Gas Turbines
Power 140.4 (2018), pp. 1–9. doi: 10.1115/1.4038042.
[43] A. Fatu and M. Arghir. Influence of manufacturing errors on the unbalance
response of aerodynamic foil bearings. Vol. 60. Springer International Pub-
lishing, 2019, pp. 281–291. isbn: 9783319992624. doi: 10.1007/978-3-319
-99262-4_20.
Bibliography 263
[44] K. Feng, Y. Cao, K. Yu, H. Guan, Y. Wu, and Z. Guo. “Characterization

of a controllable stiffness foil bearing with shape memory alloy springs”. In:
Tribol. Int. 136.December 2018 (2019), pp. 360–371. doi: 10.1016/j.trib
oint.2019.03.068.
[45] K. Feng and S. Kaneko. “Analytical Model of Bump-Type Foil Bearings
Using a Link-Spring Structure and a Finite-Element Shell Model”. In: J.
Tribol. 132.2 (2010), p. 21706. doi: 10.1115/1.4001169.
[46] J. H. Ferziger and M. Perić. Computational Methods for Fluid Dynamics.
3rd ed. Berlin, Heidelberg: Springer Berlin Heidelberg, 2002. isbn: 978-3-
540-42074-3. doi: 10.1007/978-3-642-56026-2.
[47] D. D. Fuller. “General review of gas-bearing technology”. In: Proc. 1st Int.
Symp. Gas Lubr. Bear. 1959, pp. 1–27.
[48] S. Gao, K. Cheng, S. Chen, H. Ding, and H. Fu. “CFD based investigation
on influence of orifice chamber shapes for the design of aerostatic thrust
bearings at ultra-high speed spindles”. In: Tribol. Int. 92 (2015), pp. 211–221.
doi: 10.1016/j.triboint.2015.06.020.
[49] D. Ghodsiyeh, F. Colombo, T. Raparelli, A. Trivella, and V. Viktorov. “Dia-
phragm valve-controlled air thrust bearing”. In: Tribol. Int. 109.December
2016 (2017), pp. 328–335. doi: 10.1016/j.triboint.2016.12.036.
[50] G. Grau, I. Iordanoff, B. B. Said, and Y. Berthier. “An Original Definition
of the Profile of Compliant Foil Journal Gas Bearings: Static and Dynamic
Analysis”. In: Tribol. Trans. 47.2 (2004), pp. 248–256. doi: 10.1080/05698
190439157.
[51] W. Gross. Gas Film Lubrication. 1st ed. New York: John Wiley & sons,
Inc., 1962.
[52] Y. Gu, Y. Ma, and G. Ren. “Stability and vibration characteristics of a rotor-
gas foil bearings system with high-static-low-dynamic-stiffness supports”.
In: J. Sound Vib. (2017). doi: 10.1016/j.jsv.2017.02.047.
[53] H. Q. Guan, K. Feng, Y. L. Cao, M. Huang, Y. H. Wu, and Z. Y. Guo.
“Experimental and theoretical investigation of rotordynamic characteristics
of a rigid rotor supported by an active bump-type foil bearing”. In: J. Sound
Vib. 466 (2020), p. 115049. doi: 10.1016/j.jsv.2019.115049.
[54] E. Hairer and G. Wanner. Solving Ordinary Differential Equations II:
Stiff and Differential-Algebraic Problems. 2nd ed. Vol. 14. Springer Series
in Computational Mathematics. Berlin Heidelberg: Springer-Verlag, 1996.
isbn: 978-3-642-05221-7. doi: 10.1007/978-3-642-05221-7.
264 Bibliography
[55] B. J. Hamrock. Fundamentals of Fluid Film Lubrication. McGraw-Hill Series

in Mechanical Engineering. New York: McGraw-Hill, Inc., 1994, p. 690.
isbn: 9780070259560.
[56] B. J. Hamrock, S. R. Schmid, and B. O. Jacobson. Fundamentals of Fluid
Film Lubrication. CRC Press, 2004. isbn: 9780429215315. doi: 10.1201/9
780203021187.
[57] D.-c. Han, S.-s. Park, W.-j. Kim, and J.-w. Kim. “A study on the charac-
teristics of externally pressurized air bearings”. In: Precis. Eng. 16.3 (1994),
pp. 164–173. doi: 10.1016/0141-6359(94)90121-X.
[58] W. J. Harrison. “The Hydrodynamical Theory of Lubrication With Special
Reference to Air as a Lubricant”. In: Trans. Cambridge Philos. Soc. 22.3
(1913), pp. 39–54.
[59] M. F. B. Hassan and P. Bonello. “A new modal-based approach for modelling
the bump foil structure in the simultaneous solution of foil-air bearing rotor
dynamic problems”. In: J. Sound Vib. (2017). doi: 10.1016/j.jsv.2017
.02.028.
[60] M. A. Hassini, M. Arghir, and M. Frocot. “Comparison Between Numerical
and Experimental Dynamic Coefficients of a Hybrid Aerostatic Bearing”.
In: J. Eng. Gas Turbines Power 134.12 (2012), pp. 921–931. doi: 10.1115
/1.4007375.
[61] H. Heshmat, J. A. Walowit, and O. Pinkus. “Analysis of Gas Lubricated
Compliant Thrust Bearings”. In: J. Lubr. Technol. 105.4 (1983), p. 638.
doi: 10.1115/1.3254696.
[62] H. Heshmat, J. A. Walowit, and O. Pinkus. “Analysis of Gas-Lubricated
Foil Journal Bearings”. In: J. Lubr. Technol. 105.4 (1983), pp. 647–655. doi:
10.1115/1.3254697.
[63] H. Heshmat, H. M. Chen, and J. F. Walton. “On the Performance of Hybrid
Foil-Magnetic Bearings”. In: J. Eng. Gas Turbines Power 122.1 (2000),
pp. 73–81. doi: 10.1115/1.483178.
[64] H. Heshmat, H. M. Chen, and J. F. Walton. Hybrid Foil-Magnetic Bearing.
US Patent 6,353,273 B1 (filed Sep 15, 1998): United States Patent and
Trademark Office, 2002.
[65] H. Heshmat, J. F. Walton, and J. L. Córdova. “Technology readiness
of 5th and 6th generation compliant foil bearing for 10 MWE s-CO2
turbomachinery systems”. In: 6th Int. Supercrit. CO2 Power Cycles Symp.
Pittsburgh, Pennsylvania, USA, 2018, pp. 1–29.
Bibliography 265
[66] A. C. Hindmarsh, R. Serban, and A. Collier. User Documentation for ida

v3.1.2. UCRL-SM-20. Center for Applied Scientific Computing Lawrence
Livermore National Laboratory, 2018.
[67] A. C. Hindmarsh, P. N. Brown, K. E. Grant, S. L. Lee, R. Serban, D. E.
Shumaker, and C. S. Woodward. “SUNDIALS: Suite of Nonlinear and
differential/algebraic equation solvers”. In: ACM Trans. Math. Softw. 31.3
(2005), pp. 363–396. doi: 10.1145/1089014.1089020.
[68] R. Hoffmann, T. Pronobis, and R. Liebich. “The Impact of Modified Cor-
rugated Bump Structures on the Rotor Dynamic Performance of Gas Foil
Bearings”. In: Vol. 7B Struct. Dyn. Vol. 7B. International Gas Turbine
Institute. ASME, 2014, V07BT32A012. isbn: 978-0-7918-4577-6. doi: 10.1
115/GT2014-25636.
[69] R. Hoffmann, T. Pronobis, and R. Liebich. “Non-linear Stability Analysis
of a Modified Gas Foil Bearing Structure”. In: Proc. 9th IFToMM Int. Conf.
Rotor Dyn. 2015, pp. 1259–1276. doi: 10.1007/978-3-319-06590-8_103.
[70] S. A. Howard and L. San Andrés. “A new analysis tool assessment for
rotordynamic modeling of gas foil bearings”. In: J. Eng. Gas Turbines
Power 133.2 (2011), p. 22505. doi: 10.1115/1.4001997.
[71] I. Iordanoff. “Analysis of an aerodynamic compliant foil thrust bearing:
Method for a rapid design”. In: J. Tribol. 121.4 (1999), pp. 816–822. doi:
10.1115/1.2834140.
[72] H. H. Jeffcott. “The Lateral Vibration of Loaded Shafts in the Neighbour-
hood of a Whirling Speed.—The Effect of Want of Balance”. In: London,
Edinburgh, Dublin Philos. Mag. J. Sci. 37.219 (1919), pp. 304–314. doi:
10.1080/14786440308635889.
[73] S. Jeong, D. Jeon, and Y. B. Lee. “Rigid Mode Vibration Control and
Dynamic Behavior of Hybrid Foil–Magnetic Bearing Turbo Blower”. In: J.
Eng. Gas Turbines Power 139.5 (2017), p. 052501. doi: 10.1115/1.40349
20.
[74] D. Kim. “Parametric Studies on Static and Dynamic Performance of Air
Foil Bearings with Different Top Foil Geometries and Bump Stiffness
Distributions”. In: J. Tribol. 129.2 (2007), p. 354. doi: 10.1115/1.254006
5.
[75] D. Kim and D. Lee. “Design of Three-Pad Hybrid Air Foil Bearing and
Experimental Investigation on Static Performance at Zero Running Speed”.
In: J. Eng. Gas Turbines Power 132.12 (2010), p. 122504. doi: 10.1115/1
.4001066.
266 Bibliography
[76] D. Kim and S. Park. “Hydrostatic air foil bearings: Analytical and ex-
perimental investigation”. In: Tribol. Int. 42.3 (2009), pp. 413–425. doi:
10.1016/j.triboint.2008.08.001.
[77] D. Kim and S. Park. “Hybrid Air Foil Bearings With External Pressuriza-
tion”. In: Tribology. Vol. 2006. ASME, 2006, pp. 63–69. isbn: 0-7918-4782-9.
doi: 10.1115/IMECE2006-16151.
[78] D. Kim and G. Zimbru. “Start-stop characteristics and thermal behavior of
a large hybrid airfoil bearing for aero-propulsion applications”. In: J. Eng.
Gas Turbines Power 134.3 (2012), pp. 1–9. doi: 10.1115/1.4004487.
[79] T. H. Kim and L. San Andrés. “Limits for High-Speed Operation of Gas
Foil Bearings”. In: J. Tribol. 128.3 (2006), pp. 670–673. doi: 10.1115/1.21
97851.
[80] T. H. Kim and L. San Andrés. “Analysis of advanced gas foil bearings with
piecewise linear elastic supports”. In: Tribol. Int. 40.8 (2007), pp. 1239–1245.
doi: 10.1016/j.triboint.2007.01.022.
[81] T. H. Kim and L. San Andrés. “Heavily Loaded Gas Foil Bearings: A Model
Anchored to Test Data”. In: J. Eng. Gas Turbines Power 130.1 (2008),
p. 012504. doi: 10.1115/1.2770494.
[82] W. F. Koepsel. Gas lubricated foil bearing development for advanced turboma-
chines, AFAPL-TR-76-114 vol 1. Tech. rep. Phoenix, Arizona: AiResearch
Manufacturing Company of Arizona, 1977.
[83] C.-P. R. Ku and H. Heshmat. “Compliant Foil Bearing Structural Stiffness
Analysis: Part I—Theoretical Model Including Strip and Variable Bump
Foil Geometry”. In: J. Tribol. 114.2 (1992), pp. 394–400. doi: 10.1115/1
.2920898.
[84] C.-P. R. Ku and H. Heshmat. “Compliant Foil Bearing Structural Stiffness
Analysis – Part II: Experimental Investigation”. In: Trans. Asme J. Tribol.
115.3 (1993), p. 364. doi: 10.1115/1.2921644.
[85] M. Kumar and D. Kim. “Parametric Studies on Dynamic Performance
of Hybrid Airfoil Bearing”. In: J. Eng. Gas Turbines Power 130.6 (2008),
p. 062501. doi: 10.1115/1.2940354.
[86] P. K. Kundu, I. M. Cohen, and D. R. Dowling. Fluid Mechanics. 5th ed.
Academic Press, 2012, p. 920. isbn: 9780123821003. doi: 10.1016/C2009-
0-63410-3.
[87] J. S. Larsen, A. J. T. Hansen, and I. F. Santos. “Experimental and theoret-
ical analysis of a rigid rotor supported by air foil bearings”. In: Mech. Ind.
16.1 (2015), p. 106. doi: 10.1051/meca/2014066.
Bibliography 267
[88] J. S. Larsen, I. F. Santos, and S. von Osmanski. “Stability of rigid rotors

supported by air foil bearings: Comparison of two fundamental approaches”.
In: J. Sound Vib. 381 (2016), pp. 179–191. doi: 10.1016/j.jsv.2016.06
.022.
[89] J. S. Larsen, A. C. Varela, and I. F. Santos. “Numerical and experimental
investigation of bump foil mechanical behaviour”. In: Tribol. Int. 74 (2014),
pp. 46–56. doi: 10.1016/j.triboint.2014.02.004.
[90] J. S. Larsen and I. F. Santos. “On the nonlinear steady-state response of
rigid rotors supported by air foil bearings – Theory and experiments”. In: J.
Sound Vib. 346 (2015), pp. 284–297. doi: 10.1016/j.jsv.2015.02.017.
[91] J. S. Larsen, B. B. Nielsen, and I. Santos. “On the Numerical Simulation of
Nonlinear Transient Behavior of Compliant Air Foil Bearings”. In: Proc. 11.
Int. Conf. Schwingungen Rotierenden Maschinen. Magdeburg, Germany,
2015, pp. 1–13.
[92] J. S. Larsen and I. Santos. “Compliant foil journal bearings-investigation of
dynamic properties”. In: Proc. 10th Int. Conf. Schwingungen Rotierenden
Maschinen. Berlin, Germany, 2013.
[93] J. S. Larsen and I. F. Santos. “Efficient solution of the non-linear Reynolds
equation for compressible fluid using the finite element method”. In: J.
Brazilian Soc. Mech. Sci. Eng. 37.3 (2015), pp. 945–957. doi: 10.1007/s4
0430-014-0220-5.
[94] S. Le Lez, M. Arghir, and J. Frêne. “A New Bump-Type Foil Bearing
Structure Analytical Model”. In: J. Eng. Gas Turbines Power 129.4 (2007),
pp. 1047–1057. doi: 10.1115/1.2747638.
[95] S. Le Lez, M. Arghir, and J. Frêne. “A dynamic model for dissipative
structures used in bump-type foil bearings”. In: Tribol. Trans. 52.1 (2008),
pp. 36–46. doi: 10.1080/10402000802065345.
[96] S. Le Lez, M. Arghir, and J. Frêne. “Nonlinear Numerical Prediction of Gas
Foil Bearing Stability and Unbalanced Response”. In: J. Eng. Gas Turbines
Power 131.1 (2009), p. 012503. doi: 10.1115/1.2967481.
[97] D.-H. Lee and D. Kim. “Design and Performance Prediction of Hybrid
Air Foil Thrust Bearings”. In: J. Eng. gas turbines power 133.4 (2011),
p. 042501. doi: 10.1115/1.4002249.
[98] D.-H. Lee, Y.-C. Kim, and K.-W. Kim. “The static performance analysis
of foil journal bearings considering three-dimensional shape of the foil
structure”. In: J. Tribol. 130.3 (2008), p. 31102. doi: 10.1115/1.2913538.
268 Bibliography
[99] D.-H. Lee, Y.-C. Kim, and K.-W. Kim. “The Dynamic Performance Analysis
of Foil Journal Bearings Considering Coulomb Friction: Rotating Unbalance
Response”. In: Tribol. Trans. 52.2 (2009), pp. 146–156. doi: 10.1080/1040
2000802192685.
[100] D.-H. Lee, Y.-C. Kim, and K.-W. Kim. “The effect of Coulomb friction
on the static performance of foil journal bearings”. In: Tribol. Int. 43.5-6
(2010), pp. 1065–1072. doi: 10.1016/j.triboint.2009.12.048.
[101] Y. B. Lee, T. H. Kim, C. H. Kim, N. S. Lee, and D. H. Choi. “Unbalance
Response of a Super-Critical Rotor Supported by Foil Bearings—Compar-
ison with Test Results”. In: Tribol. Trans. 47.1 (2004), pp. 54–60. doi:
10.1080/05698190490279038.
[102] A. Lehn, M. Mahner, and B. Schweizer. “Elasto-gasdynamic modeling of air
foil thrust bearings with a two-dimensional shell model for top and bump
foil”. In: Tribol. Int. 100 (2016), pp. 48–59. doi: 10.1016/j.triboint.201
5.11.011.
[103] T. Leister, C. Baum, and W. Seemann. “On the Importance of Frictional
Energy Dissipation in the Prevention of Undesirable Self-Excited Vibrations
in Gas Foil Bearing Rotor Systems”. In: Tech. Mech. 37.2-5 (2017), pp. 280–
290. doi: 10.24352/UB.OVGU-2017-104.
[104] T. Leister, C. Baum, and W. Seemann. “Computational Analysis of Foil
Air Journal Bearings Using a Runtime-Efficient Segmented Foil Model”. In:
Proc. ISROMAC (2016).
[105] T. Leister, W. Seemann, and B. Bou-Saïd. “Bifurcation Analysis of Rotors
on Refrigerant-Lubricated Gas Foil Bearings”. In: 13th Int. Conf. Dyn.
Rotating Mach. (SIRM 2019). Ed. by I. F. Santos. Kgs. Lyngby, Denmark:
Technical University of Denmark (DTU), 2019, pp. 46–56.
[106] L. Licht and A. Eshel. Study, fabrication and testing of a foil-bearing
rotor support system, NASA Contractor Report No. CR-1157. Tech. rep.
Washington, D. C.: NASA, 1968, p. 235.
[107] Z. S. Liu, G. H. Zhang, and H. J. Xu. “Performance analysis of rotating
externally pressurized air bearings”. In: Proc. Inst. Mech. Eng. Part J J.
Eng. Tribol. 223.4 (2009), pp. 653–663. doi: 10.1243/13506501JET510.
[108] J. W. Lund and K. K. Thomsen. “A calculation method and data for the
dynamic coefficients of oil-lubricated journal bearings”. In: Top. Fluid Film
Bear. Rotor Bear. Syst. Des. Optim. ASME, New York (1978), pp. 1–28.
[109] J. W. Lund. “The Hydrostatic Gas Journal Bearing With Journal Rotation
and Vibration”. In: J. Basic Eng. 86.2 (1964), p. 328. doi: 10.1115/1.365
3073.
Bibliography 269
[110] J. W. Lund. “Calculation of Stiffness and Damping Properties of Gas

Bearings”. In: J. Lubr. Technol. 90.4 (1968), pp. 793–803. doi: 10.1115/1
.3601723.
[111] J. W. Lund. “Review of the concept of dynamic coefficients for fluid film
journal bearings”. In: J. Tribol. 109.1 (1987), pp. 37–41. doi: 10.1115/1.3
261324.
[112] N. Maamari, A. Krebs, S. Weikert, and K. Wegener. “Centrally fed orifice
based active aerostatic bearing with quasi-infinite static stiffness and high
servo compliance”. In: Tribol. Int. 129.August 2018 (2019), pp. 297–313.
doi: 10.1016/j.triboint.2018.08.024.
[113] C. Makkar, W. E. Dixon, W. G. Sawyer, and G. Hu. “A New Continuously
Differentiable Friction Model for Control Systems Design”. In: Proc. 2005
IEEE/ASME Int. Conf. Adv. Intell. Mechatronics. Monterey, California,
USA: IEEE, 2005, pp. 600–605. doi: 10.1109/AIM.2005.1511048.
[114] H. Mizumoto, S. Arii, Y. Kami, K. Goto, T. Yamamoto, and M. Kawamoto.
“Active inherent restrictor for air-bearing spindles”. In: Precis. Eng. 19.2-3
(1996), pp. 141–147. doi: 10.1016/s0141-6359(96)00041-4.
[115] L. Molzen. “Experimental Modal Analysis of an Industrial Compressor Sup-
ported By Air Foil Bearings”. Master’s thesis. Kongens Lyngby, Denmark:
Technical University of Denmark, Dept. of Mechanical Engineering, 2016.
[116] S. Morosi and I. F. Santos. “On the modelling of hybrid aerostatic-gas
journal bearings”. In: J. Eng. Tribol. 225.7 (2011), pp. 641–653. doi: 10.11
77/1350650111399845.
[117] S. Morosi and I. F. Santos. “Active lubrication applied to radial gas journal
bearings. Part 1: Modeling”. In: Tribol. Int. 44.12 (2011), pp. 1949–1958.
doi: 10.1016/j.triboint.2011.08.007.
[118] N. Mostaghel and T. Davis. “Representations of Coulomb friction for
dynamic analysis”. In: Earthq. Eng. Struct. Dyn. 26 (1997), pp. 541–548.
doi: 10.1002/(SICI)1096-9845(199705)26:5<541::AID-EQE660>3.0.C
O;2-W.
[119] NASA. Creating a Turbomachinery Revolution: Research at Glenn Enables
an Oil-Free Turbine Engine, FS-2001-07-014-GRC. Tech. rep. Cleveland,
Ohio: Information and Publications Office, NASA Glenn Research Center,
2001.
[120] M. T. Neves, V. A. Schwarz, and G. J. Menon. “Discharge coefficient
influence on the performance of aerostatic journal bearings”. In: Tribol. Int.
43.4 (2010), pp. 746–751. doi: 10.1016/j.triboint.2009.11.001.
270 Bibliography
[121] B. B. Nielsen and I. F. Santos. “Transient and steady state behaviour

of elasto-aerodynamic air foil bearings, considering bump foil compliance
and top foil inertia and flexibility: A numerical investigation”. In: Proc.
Inst. Mech. Eng. Part J J. Eng. Tribol. 231.10 (2017), pp. 1235–1253. doi:
10.1177/1350650117689985.
[122] B. B. Nielsen. “Combining Gas Bearing and Smart Material Technologies
for Improving Machine Performance – Theory and Experiment”. PhD thesis.
Kongens Lyngby, Denmark: Technical University of Denmark, 2017.
[123] B. B. Nielsen, M. S. Nielsen, and I. F. Santos. “A layered shell containing
patches of piezoelectric fibers and interdigitated electrodes: Finite element
modeling and experimental validation”. In: J. Intell. Mater. Syst. Struct.
28.1 (2017), pp. 78–96. doi: 10.1177/1045389X16642537.
[124] J. T. Oden and J. A. C. Martins. “Models and Computational Methods
for Dynamic Friction Phenomena”. In: Comput. Methods Appl. Mech. Eng.
52.1 (1985), pp. 527–634. doi: 10.1016/0045-7825(85)90009-X.
[125] S. von Osmanski, J. S. Larsen, and I. F. Santos. “A fully coupled air foil
bearing model considering friction – Theory & experiment”. In: J. Sound
Vib. 400 (2017), pp. 660–679. doi: 10.1016/j.jsv.2017.04.008.
[126] S. von Osmanski, J. S. Larsen, and I. F. Santos. “On the incorporation of
friction into a simultaneously coupled time domain model of a rigid rotor
supported by air foil bearings”. In: Tech. Mech. 37.2-5 (2017), pp. 291–302.
doi: 10.24352/UB.OVGU-2017-105.
[127] S. von Osmanski, J. S. Larsen, and I. F. Santos. “Modelling of Compliant-
type Gas Bearings: A Numerical Recipe”. In: 13th Int. Conf. Dyn. Rotating
Mach. (SIRM 2019). Ed. by I. Santos. February. Kgs. Lyngby: Technical
University of Denmark (DTU), 2019, pp. 14–27. isbn: 978-87-7475-568-5.
[128] S. von Osmanski, J. S. Larsen, and I. F. Santos. “The Classical Linearization
Technique’s Validity for Compliant Bearings”. In: Proc. 10th Int. Conf. Rotor
Dyn. - IFToMM vol.1. Ed. by K. L. Cavalca and H. I. Weber. Vol. 60.
Mechanisms and Machine Science. Springer International Publishing, 2019,
pp. 177–191. isbn: 978-3-319-99261-7. doi: 10.1007/978-3-319-99262-4
_13.
[129] S. von Osmanski, J. S. Larsen, and I. F. Santos. “Multi-domain stability
and modal analysis applied to Gas Foil Bearings: Three approaches”. In: J.
Sound Vib. (2020), p. 115174. doi: 10.1016/j.jsv.2020.115174.
[130] S. von Osmanski and I. F. Santos. “Gas Foil Bearings with Radial Injection:
Multi-domain Stability Analysis and Unbalance Response”. (submitted for
publication in the Journal of Sound and Vibration, 7 April 2020).
Bibliography 271
[131] B. T. Paulsen, S. Morosi, and I. F. Santos. “Static, dynamic, and thermal

properties of compressible fluid film journal bearings”. In: Tribol. Trans.
54.2 (2011), pp. 282–299. doi: 10.1080/10402004.2010.538490.
[132] J. P. Peng and M. Carpino. “Calculation of Stiffness and Damping Coeffi-
cients for Elastically Supported Gas Foil Bearings”. In: J. Tribol. 115.1
(1993), pp. 20–27. doi: 10.1115/1.2920982.
[133] J. P. Peng and M. Carpino. “Finite Element Approach to the Prediction
of Foil Bearing Rotor Dynamic Coefficients”. In: J. Tribol. 119.1 (1997),
pp. 85–90. doi: 10.1115/1.2832484.
[134] J. P. Peng and M. Carpino. “Coulomb Friction Damping Effects in Elastic-
ally Supported Gas Foil Bearings”. In: Tribol. Trans. 37.1 (1994), pp. 91–98.
doi: 10.1080/10402009408983270.
[135] Z.-C. Peng and M. M. Khonsari. “Hydrodynamic analysis of compliant foil
bearings with compressible air flow”. In: J. Tribol. – Trans. Asme, 126.3
(2004), pp. 542–546. doi: 10.1115/1.1739242.
[136] Z.-C. Peng and M. M. Khonsari. “On the limiting load-carrying capacity of
foil bearings”. In: J. Tribol. 126.4 (2004), pp. 817–818. doi: 10.1115/1.17
92697.
[137] E. P. Petrov and D. J. Ewins. “Generic Friction Models for Time-Domain Vi-
bration Analysis of Bladed Disks”. In: J. Turbomach. 126.1 (2004), pp. 184–
192. doi: 10.1115/1.1644557.
[138] L. Petzold. “Differential/Algebraic Equations are not ODE’s”. In: SIAM J.
Sci. Stat. Comput. 3.3 (1982), pp. 367–384. doi: 10.1137/0903023.
[139] H. M. Pham and P. Bonello. “Efficient Techniques for the Computation of
the Nonlinear Dynamics of a Foil-Air Bearing Rotor System”. In: ASME
Turbo Expo 2013 Turbine Tech. Conf. Expo. Vol. Volume 7B. International
Gas Turbine Institute. San Antonio, Texas, USA, 2013, GT2013–94389.
isbn: 978-0-7918-5527-0. doi: 10.1115/GT2013-94389.
[140] M. N. Pham and H. J. Ahn. “Experimental optimization of a hybrid foil-
magnetic bearing to support a flexible rotor”. In: Mech. Syst. Signal Process.
46.2 (2014), pp. 361–372. doi: 10.1016/j.ymssp.2014.01.012.
[141] F. G. Pierart and I. F. Santos. “Steady state characteristics of an adjustable
hybrid gas bearing - Computational fluid dynamics, modified Reynolds
equation and experimental validation”. In: Proc. Inst. Mech. Eng. Part J J.
Eng. Tribol. 229.7 (2015), pp. 807–822. doi: 10.1177/1350650115570404.
272 Bibliography
[142] F. G. Pierart and I. F. Santos. “Active lubrication applied to radial gas

journal bearings. Part 2: Modelling improvement and experimental valida-
tion”. In: Tribol. Int. 96.12 (2016), pp. 237–246. doi: 10.1016/j.triboint
.2015.12.004.
[143] F. G. Pierart and I. F. Santos. “Lateral vibration control of a flexible
overcritical rotor via an active gas bearing – Theoretical and experimental
comparisons”. In: J. Sound Vib. 383 (2016), pp. 20–34. doi: 10.1016/j.js
v.2016.07.024.
[144] M. Z. Poulsen, P. G. Thomsen, and N. Houbak. “Structural analysis of
DAEs”. PhD thesis. Lyngby, Denmark: Technical University of Denmark,
2002.
[145] J. W. Powell. “A review of progress in gas lubrication”. In: Rev. Phys.
Technol. 1.2 (1970), pp. 96–129. doi: 10.1088/0034-6683/1/2/303.
[146] T. Pronobis and R. Liebich. “Comparison of stability limits obtained by
time integration and perturbation approach for Gas Foil Bearings”. In: J.
Sound Vib. 458 (2019), pp. 497–509. doi: 10.1016/j.jsv.2019.06.034.
[147] K. Radil and Z. Batcho. “Air injection as a thermal management technique
for radial foil air bearings”. In: Tribol. Trans. 54.4 (2011), pp. 666–673. doi:
10.1080/10402004.2011.589964.
[148] W. J. M. Rankine. “On The Centrifugal Force of Rotating Shafts”. In: Eng.
27 (1869), p. 249.
[149] J. C. Renn and C. H. Hsiao. “Experimental and CFD study on the mass
flow-rate characteristic of gas through orifice-type restrictor in aerostatic
bearings”. In: Tribol. Int. 37.4 (2004), pp. 309–315. doi: 10.1016/j.tribo
int.2003.10.003.
[150] O. Reynolds. “On the Theory of Lubrication and Its Application to Mr.
Beauchamp Tower’s Experiments, Including an Experimental Determination
of the Viscosity of Olive Oil”. In: Philos. Trans. R. Soc. London 177.0 (1886),
pp. 157–234. doi: 10.1098/rstl.1886.0005.
[151] S. K. Ro, S. Kim, Y. Kwak, and C. H. Park. “A linear air bearing stage with
active magnetic preloads for ultraprecise straight motion”. In: Precis. Eng.
34.1 (2010), pp. 186–194. doi: 10.1016/j.precisioneng.2009.06.010.
[152] W. B. Rowe. “Basic Flow Theory”. In: Hydrostatic, Aerostatic and Hybrid
Bearing Design. 1st ed. Elsevier, 2012, pp. 25–48. isbn: 9780123969941.
doi: 10.1016/B978-0-12-396994-1.00002-4.
[153] D. Rubio and L. San Andrés. “Bump-Type Foil Bearing Structural Stiffness:
Experiments and Predictions”. In: J. Eng. gas turbines power 128.3 (2006),
pp. 653–660. doi: 10.1115/1.2056047.
Bibliography 273
[154] D. Ruscitto, J. Mc Cormick, and S. Gray. Hydrodynamic air lubricated

compliant surface bearing for an automotive gas turbine engine {I} – Journal
bearing performance. Tech. rep. NASA CR-135368. Cleveland, Ohio: NASA
Lewis Research Center, 1978, p. 136. doi: 10.2172/6793795.
[155] L. San Andrés and T. H. Kim. “Improvements to the Analysis of Gas Foil
Bearings: Integration of Top Foil 1D and 2D Structural Models”. In: Vol.
5 Turbo Expo 2007. Vol. 5. International Gas Turbine Institute. Montreal,
Canada: ASME, 2007, pp. 779–789. isbn: 0-7918-4794-2. doi: 10.1115/GT
2007-27249.
[156] L. San Andrés and T. H. Kim. “Analysis of gas foil bearings integrating
FE top foil models”. In: Tribol. Int. 42.1 (2009), pp. 111–120. doi: 10.1016
/j.triboint.2008.05.003.
[157] I. F. Santos. “Aktive Schmierung zur Regelung von Rotorsystemen”. In:
Proc. SIRM, Schwingungen rotierenden Maschinen IV. Ed. by H. Irretier,
R. Nordmann, and H. Springer. Kassel: Vieweg, Braunschweig, Wiesbaden,
1997, pp. 37–47.
[158] I. F. Santos and F. H. Russo. “Tilting-Pad Journal Bearings With Electronic
Radial Oil Injection”. In: J. Tribol. 120.3 (1998), pp. 583–594. doi: 10.111
5/1.2834591.
[159] I. F. Santos, A. Scalabrin, and R. Nicoletti. “Ein Beitrag zur aktiven
Schmierungstheorie”. In: Proc. SIRM, Schwingungen rotierenden Maschinen
V. Ed. by H. Irretier, R. Nordmann, and H. Springer. Wien: Vieweg,
Braunschweig, Wiesbaden, 2001, pp. 21–30.
[160] I. F. Santos. “Controllable sliding bearings and controllable lubrication
principles-an overview”. In: Lubricants 6.1 (2018). doi: 10.3390/lubrican
ts6010016.
[161] I. Santos and L. Molzen. “On the Experimental Nonlinear Modal Analysis
of a Rigid Rotor Supported by Air Foil Bearings”. In: ASME Turbo Expo
2017 (submitted Oct. 2016). ASME International Gas Turbine Institute.
2017.
[162] J. Schiffmann and Z. S. Spakovszky. “Foil Bearing Design Guidelines for
Improved Stability”. In: J. Tribol. 135.1 (2013), pp. 1–11. doi: 10.1115/1
.4007759.
[163] K. Shalash and J. Schiffmann. “On the manufacturing of compliant foil
bearings”. In: J. Manuf. Process. 25 (2017), pp. 357–368. doi: 10.1016/j
.jmapro.2016.12.021.
274 Bibliography
[164] K. Shalash and J. Schiffmann. “Pressure Profile Measurements Within

the Gas Film of Journal Foil Bearings Using an Instrumented Rotor With
Telemetry”. In: J. Eng. Gas Turbines Power 142.3 (2020), pp. 1–9. doi:
10.1115/1.4044798.
[165] L. F. Shampine. “Solving 0 = F(t, y(t), y(t)) in Matlab”. In: J. Numer.
Math. 10.4 (2002), pp. 291–310. doi: 10.1515/JNMA.2002.291.
[166] L. F. Shampine, M. W. Reichelt, and J. A. Kierzenka. “Solving Index-1
DAEs in MATLAB and Simulink”. In: SIAM Rev. 41.3 (1999), pp. 538–552.
doi: 10.1137/S003614459933425X.
[167] S. K. Shrestha, D. Kim, and Y. C. Kim. “Experimental feasibility study of
radial injection cooling of three-pad air foil bearings”. In: J. Tribol. 135.4
(2013), pp. 1–9. doi: 10.1115/1.4024547.
[168] R. Snoeys and F. Al-Bender. “Development of improved externally pressur-
ized gas bearings”. In: KSME J. 1.1 (1987), pp. 81–88. doi: 10.1007/BF02
953383.
[169] T. Snyder and M. Braun. “Comparison of Perturbed Reynolds Equation and
CFD Models for the Prediction of Dynamic Coefficients of Sliding Bearings”.
In: Lubricants 6.1 (2018), p. 5. doi: 10.3390/lubricants6010005.
[170] J.-h. Song and D. Kim. “Foil Gas Bearing With Compression Springs:
Analyses and Experiments”. In: ASME J. Tribol. 129.3 (2007), pp. 628–639.
doi: 10.1115/1.2736455.
[171] E. E. Swanson, H. Heshmat, and I. Walton. “Performance of a foil-magnetic
hybrid bearing”. In: J. Eng. Gas Turbines Power 124.2 (2002), pp. 375–382.
doi: 10.1115/1.1417485.
[172] A. Z. Szeri. Fluid Film Lubrication. 2nd ed. Cambridge: Cambridge Univer-
sity Press, 2010. isbn: 9780511782022. doi: 10.1017/CBO9780511782022.
[173] F. A. Tariku and R. J. Rogers. “Improved Dynamic Friction Models for
Simulation of One-Dimensional and Two-Dimensional Stick-Slip Motion”.
In: J. Tribol. 123.4 (2001), pp. 661–669. doi: 10.1115/1.1331057.
[174] L. R. S. Theisen, F. G. Pierart, H. Niemann, I. F. Santos, and M. Blanke.
“Experimental Grey Box Model Identification and Control of an Active Gas
Bearing”. In: Vib. Eng. Technol. Mach. Ed. by J. K. Sinha. Cham: Springer
International Publishing, 2015, pp. 963–976. isbn: 978-3-319-09918-7. doi:
10.1007/978-3-319-09918-7_85.
[175] P. Vleugels, T. Waumans, J. Peirs, F. Al-Bender, and D. Reynaerts. “High-
speed bearings for micro gas turbines: stability analysis of foil bearings”.
In: J. Micromechanics Microengineering 16.9 (2006), S282–S289. doi: 10.1
088/0960-1317/16/9/S16.
Bibliography 275
[176] J. A. Walowit and J. N. Anno. Modern developments in lubrication mechan-

ics. London: Applied Science Publishers LTD, 1975. isbn: 9780853345923.
[177] J. F. Walton, H. Heshmat, and M. J. Tomaszewski. “Design and test
program in the developmen of a 100 HP oil-free high-speed blower”. In:
Proc. ASME Turbo Expo. 2007.
[178] C.-C. Wang and C.-K. Chen. “Bifurcation Analysis of Self-Acting Gas
Journal Bearings”. In: J. Tribol. 123.4 (2001), p. 755. doi: 10.1115/1.138
8302.
[179] Y. P. Wang and D. Kim. “Experimental identification of force coefficients
of large hybrid air foil bearings”. In: J. Eng. Gas Turbines Power 136.3
(2014). doi: 10.1115/1.4025891.
[180] A. Weiser and S. E. Zarantonello. “A Note on Piecewise Linear and Multi-
linear Table Interpolation in Many Dimensions”. In: Math. Comput. 50.181
(1988), p. 189. doi: 10.2307/2007922.
[181] B. Yang, H. Geng, Y. Sun, and L. Yu. “Dynamic characteristics of hybrid foil-
magnetic bearings (HFMBs) concerning eccentricity effect”. In: Int. J. Appl.
Electromagn. Mech. 52.1-2 (2016), pp. 271–279. doi: 10.3233/JAE-162083.
[182] B. Z. Yazdi and D. Kim. “Rotordynamic Performance of Hybrid Air Foil
Bearings With Regulated Hydrostatic Injection”. In: J. Eng. Gas Turbines
Power 140.1 (2017), p. 012506. doi: 10.1115/1.4037667.
[183] B. Z. Yazdi and D. Kim. “Effect of Circumferential Location of Radial
Injection on Rotordynamic Performance of Hybrid Air Foil Bearings”. In: J.
Eng. Gas Turbines Power 140.12 (2018), p. 122504. doi: 10.1115/1.4041
646.
[184] J. Zhang, D. Zou, N. Ta, Z. Rao, and B. Ding. “A numerical method for
solution of the discharge coefficients in externally pressurized gas bearings
with inherent orifice restrictors”. In: Tribol. Int. 125.December 2017 (2018),
pp. 156–168. doi: 10.1016/j.triboint.2018.03.032.
[185] Y. Zhou, X. Chen, and H. Chen. “A hybrid approach to the numerical
solution of air flow field in aerostatic thrust bearings”. In: Tribol. Int. 102
(2016), pp. 444–453. doi: 10.1016/j.triboint.2016.04.002.
[186] E. S. Zorzi. Gas lubricated foil bearing development for advanced turboma-
chines, AFAPL-TR-76-114 vol 2. Tech. rep. Phoenix, Arizona: AiResearch
Manufacturing Company of Arizona, 1977.
[187] G. Żywica. “The Static Performace Analysis of the Foil Bearing Structure”.
In: acta Mech. Autom. 5.4 (2011).
276 Bibliography
[188] G. Żywica. “The Dynamic Performance Analysis of the Foil Bearing Stru-
ture”. In: Acta Mech. Autom. 7.1 (2013), pp. 58–62. doi: 10.2478/ama-20
13-0011.
[189] G. Żywica, P. Bagiński, J. Kiciński, G. Ż, P. Bagi, and J. Kici. “Selected
operational Problems of high-speed Rotors supported by Gas Foil Bearings”.
In: Tech. Mech. 37.2-5 (2017), pp. 339–346. doi: 10.24352/UB.OVGU-2017
-109.
[190] G. Żywica and P. Bagiński. “Investigation of Gas Foil Bearings with an
Adaptive and Non-Linear Structure”. In: Acta Mech. Autom. 13.1 (2019),
pp. 5–10. doi: 10.2478/ama-2019-0001.
The feasibility of combining gas lubrication with compliant surfaces has been known since at
least the 1960s. These principles are combined in Gas Foil Bearings (GFBs) offering a versatile,
mechanically simple, oil-free and environmentally friendly mechanism for the support of
lightweight high-speed rotating machinery. The applicability of GFBs is, however, restricted by
several factors: a limited number of start-stop cycles, an inherently low level of damping and rich
hard-to-predict dynamics. This project is motivated by the latter and is aiming at improving the
fundamental understanding of GFB dynamics and the available modelling tools. Furthermore, it
is an objective to leverage the modelling capacity to push towards augmentation of GFBs with
injection, eventually enabling a mechatronic GFB with adaptable properties.
Specifically, the thesis provides a contribution to the modelling of the foil structure, proves a
commonly applied perturbation analysis method to be faulty, presents a simultaneously coupled
multi-domain mathematical model for GFB systems and showcases the potential of GFBs with
injection.
DTU Mechanical Engineering DCAMM

Section of Solid Mechanics Danish Center for Applied Mathematics
Technical University of Denmark and Mechanics
Nils Koppels Allé, Bld. 404 Nils Koppels Allé, Bld. 404
DK-2800 Kgs. Lyngby DK-2800 Kgs. Lyngby
Denmark Denmark
Phone: (+45) 4525 4250 Phone (+45) 4525 4250
Fax: (+45) 4525 1961 Fax (+45) 4525 1961
www.mek.dtu.dk www.dcamm.dk
April 2020 DCAMM Special Report No. S273
ISBN: 978-87-7475-591-3 ISSN: 0903-1685

Modelling of Gas Foil Bearings Towards Controllable Operation Multi-Domain Analysis

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Modelling of Gas Foil Bearings Towards Controllable Operation Multi-Domain Analysis

Uploaded by

Copyright:

Available Formats

Downloaded from orbit.dtu.

dk on: 5 02, 2023

Modelling of Gas Foil Bearings Towards Controllable Operation Multi-domain Analysis

von Osmanski, Alexander Sebastian

Link back to DTU Orbit

Modelling of Gas Foil Bearings

Modelling of Gas Foil Bearings

Sebastian von Osmanski

Kongens Lyngby, Denmark, April 2020

DTU Mechanical Engineering

Cian Conlan-Smith, Hannibal Overgaard, Mikael M. Eronen, Thomas T. Paulsen

Kongens Lyngby, Denmark, 10 April 2020

Resumé (Danish) vii

2 Conclusive Remarks & Future Challenges 25

P1 A Fully Coupled Air Foil Bearing Model Considering Friction

P1.C Validation of Bump Model . . . . . . . . . . . . . . . . . . . . . 68

P2 On the Incorporation of Friction into a Simultaneously

P3 The Classical Linearization Technique’s Validity for Compli-

P4 Modelling of Compliant-type Gas Bearings: A Numerical

P5 Multi-domain Stability and Modal Analysis Applied to

P6 Gas Foil Bearings with Radial Injection: Multi-domain Stabil-

AMB Active Magnetic Bearing FV Finite Volume

(Û) Time derivative, d/dt ε Eccentricity

by stability or risk of thermal runaway) and mean-time-between-failures in excess

Pressure–compliance interaction: Compared to other bearing types showing

Figure 1.1: Rendering of Siemens three-pad segmented GFB with a nominal

(a) (b) (c)

Non-linear nature of the response: Rotordynamics research has traditionally

Initial conditions: ε0 , εÛ0 , etc.

Primary rotor system solver: Algebraic (steady) or IVP (transient)

Film height If applicable

Solve Reynolds equation Solve other domains:

Fluid film forces

sub-synchronous amplitudes in fig. 1.4(b) is likely to be underestimated since

1.2 Literature Overview

0 100 200 300 400 500

0 100 200 300 400 500

1.2.1 Structural Models

1.2.2 Beyond the Simple Elastic Foundation Model

[23] concluding that membrane stiffness (omitted in the two-dimensional plate

e.g. manufacturing errors to be modelled.

1.2.3 Model Structure

1.2.4 Hybrid Gas Foil Bearings

A hybrid foil–magnetic bearing was proposed by Heshmat et al. [63] in which

1.2.5 Injection Modelling

1.3 Included Publications

accompanied by a synopsis. As recommended, the present thesis has been prepared

P1 S. von Osmanski, J. S. Larsen, and I. F. Santos. “A fully coupled air foil

P2 S. von Osmanski, J. S. Larsen, and I. F. Santos. “On the incorporation of

P3 S. von Osmanski, J. S. Larsen, and I. F. Santos. “The Classical Linearization

P4 S. von Osmanski, J. S. Larsen, and I. F. Santos. “Modelling of Compliant-type

P5 S. von Osmanski, J. S. Larsen, and I. F. Santos. “Multi-domain stability

The thesis includes postprints of the publications P1 to P5 which have already

1.4 Original Contributions

1.4.1 Publications P1 and P2: Foil Structure & Friction

1.4.2 Publication P4: General Modelling and Jacobians

1.4.3 Publications P3 and P5: Perturbation Methods

Chebyshev polynomial or similar instead of discrete DOFs directly. The purpose

1.4.4 Publication P5: Multi-Domain Modal Analysis

1.4.5 Publication P6: Injection

affected significantly and hence provides an indication of the potential of more

2.1.1 The Role of Friction

2.1.2 General Modelling of Gas Foil Bearings