Reduced Order Modeling For Large-Scale Linear Systems

by

Linyan Xiang

A dissertation submitted in partial fulfillment

of the requirements for the degree of
Doctor of Philosophy
(Mechanical Sciences and Engineering)
in the University of Michigan-Dearborn
2022

Doctoral Committee:
Associate Professor Cheol W. Lee, Co-Chair
Professor Oleg Zikanov, Co-Chair
Assistant Professor Hugo Casquero
Assistant Professor Zhen Hu
Professor Dewey Dohoy Jung
 Linyan Xiang
[email protected]
ORCID iD: 0000-0002-6243-909X

© Linyan Xiang 2022

 ACKNOWLEDGMENTS

I would like to express my greatest appreciation to my supervisors, Dr. Oleg Zikanov and Dr.
Cheol W. Lee. All the works are built upon the endless help, strong financial support, and careful
guidance of the two mentors along with doctoral career.
I acknowledge, gratefully, Professor Oleg Zikanov for his constant encouragement, precious
advice, and patience which made this work possible. Excellent guidance and helpful comments
are greatly appreciated. I have been lucky to have a professional and knowledgeable supervisor
who is passionate about research and life. I thankfully take the inspiring memo, and may it be with
me in the future, as the Hitchhiker’s Guide to the Galaxy says, don’t panic.
I also acknowledge many thanks to Professor Cheol W. Lee for inspiring me academically and
professionally. I hope that I can go to work and study in the future with a great profession, diligent
work attitude, and the pursuit of perfection.
I would like to thank my peers and friends for their inspiration. Looking forward to the opportu-
nity to continue working with you in the future. Special thanks to Dr. Xuan Zhang for encourage-
ment and support; thanks to Ruslan and Ali for their mutual motivation during the epidemic; and
Dr. Zhuo Wang for their excellence the inspiration from working together. Many special thanks to
my friends and colleagues for their support during the Ph.D. journey.
Thanks to the excellent managers who have provided me with opportunities in the industry,
their support has enabled me to get in touch with the impressive work experience in the industry.
Thanks to Dr. Shu for allowing me to work at Eberspächer in 2018; thanks to Dr. Hsu of GM for
the opportunity to dedicate contributions to the R&D process of the industry; thanks to Kerry and
Qi for providing the opportunity to have wonderful summer at Mathworks.
At the same time, thanks to the computing resources provided by Great Lakes HPC cluster and
technical consulting from Bennet Fauber. This research was supported in part through compu-

ii
tational resources and services provided by Advanced Research Computing at the University of
Michigan, Ann Arbor.
Finally, I would like to thank my parents for their support and trust in me, allowing me to follow
my path on my own. Without their support and encouragement, I would not be able to get to where
I am today.

iii
 TABLE OF CONTENTS

ACKNOWLEDGMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix

LIST OF ALGORITHMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii

ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiv

CHAPTER

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Statement of the Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.3 Proposed Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.4 Literature Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.5 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2 Theoretical and Numerical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.1 High-Fidelity Large-Scale Linear Systems . . . . . . . . . . . . . . . . . . . . . 12

2.1.1 Conjugate Heat Transfer in a Battery Pack of an Electric Vehicle . . . . . 12

2.1.2 Airborne Transmission of Respiratory Infections . . . . . . . . . . . . . 15

2.1.3 Consistent Steady-State Representation of Large-Scale Linear Systems . 18

2.1.4 Boundary Condition as an Input . . . . . . . . . . . . . . . . . . . . . . 19

2.2 Reduced-Order Modeling Methodologies . . . . . . . . . . . . . . . . . . . . . 21

iv
 2.2.1 Krylov-Subspace Method . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.2.2 Parametric Reduced-Order Modeling . . . . . . . . . . . . . . . . . . . 23

3 Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3.1 Indirect Arnoldi Algorithm for Single-Input Systems . . . . . . . . . . . . . . . 26

3.2 Indirect Arnoldi Algorithm for Multiple-Input Systems . . . . . . . . . . . . . . 28

3.3 Frequency-Shift Arnoldi Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 30

3.4 Parametric Reduced Order Modeling Algorithm . . . . . . . . . . . . . . . . . . 32

3.5 Change of Variable Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

4 Reduced-Order Modeling for Conjugate Heat Transfer in a Battery Pack . . . . . . . 35

4.1 Battery Pack Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

4.2 Error of ROM Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

4.3 Single-Input Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

4.4 Multiple-Input Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

4.5 Boundary Condition as Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

4.6 Parametric Reduced-Order Modeling . . . . . . . . . . . . . . . . . . . . . . . . 56

4.7 Systems With Time-Varying Parameters . . . . . . . . . . . . . . . . . . . . . . 58

4.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5 Reduced-Order Modeling for Airborne Transmission of Respiratory Infections . . . 64

5.1 Comparison Between Eulerian Approach and Lagrangian Approach Applied to

Aerosol Transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

5.2 ROM Generation and Optimization for Ventilated Indoor Environments . . . . . 75

5.2.1 Convergence Issues: Regarding the Residual of Transport Equation So-

lution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

5.2.2 ROM Generation Using Single-Input Algorithm . . . . . . . . . . . . . 76

5.2.3 ROM Generation Using the Frequency-Shift Algorithm . . . . . . . . . . 84

v
 5.3 ROM Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

5.3.1 Exhalation Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

5.3.2 ROM Validation Under the Designed Coughing Cycle . . . . . . . . . . 95

5.3.3 Time-Varying Process Predicted by ROM . . . . . . . . . . . . . . . . . 102

5.3.4 Risk Assessment Using ROM Data . . . . . . . . . . . . . . . . . . . . 103

5.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

6 Conclusion and Future Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

vi
 LIST OF TABLES

TABLE

4.1 Model specifications. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

4.2 The norms (in ◦ C) of the error at various time instants for the test case 1 (see Figures
4.2, 4.4 and 4.5). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

4.3 The norms (in ◦ C) of the error at r = 15 for the test case 2. . . . . . . . . . . . . . . 44

4.4 The computation cost of the ROM creation and transient simulations in the single
input case compared with the cost of the FOM simulations. . . . . . . . . . . . . . . . 45

4.5 The norms of the error vector (in ◦ C) for the test case 3 (see Figures 4.8–4.10). . . . . 48

4.6 The computation cost of the ROM creation and transient simulations compared with
the cost of the FOM simulation of transient process in the test case 3 lasting 1000
seconds. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

4.7 The norms (in ◦ C) of the error vector for the test case 4 (see Figures 4.11, 4.12 and
4.13). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

4.8 The norms of the error vector (in ◦ C) for the boundary condition as the input (see
Figures 4.16). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

4.9 The norms of the error vector (in ◦ C) for parametric ROM at mass flow rate =
0.02kg/s (see Figures 4.17 and 4.18). . . . . . . . . . . . . . . . . . . . . . . . . . . 58

4.10 The norms of the error vector (in ◦ C) for parametric ROM at time-varying mass flow
rates (see Figures 4.19). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

5.1 The quantitative study between Eulerian/Lagrangian methods at t=50s . . . . . . . . . 74

5.2 The quantitative study between Eulerian/Lagrangian methods at t=100s . . . . . . . . 75

5.3 The quantitative evaluation for ventilation ROM using Algorithm 1. . . . . . . . . . . 82

5.4 The 11 Sigma points for frequency-shift Krylov method . . . . . . . . . . . . . . . . 85

vii
5.5 The quantitative evaluation for ventilation ROM using Algorithm 3. . . . . . . . . . . 91

5.6 The norms at t=10, 50 and 220s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

5.7 Volume integration of passive scalar concentration over boxes in front of students at
t=2000s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

viii
 LIST OF FIGURES

FIGURE

1.1 Dissertation outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.1 Schematic representation of various modes of respiratory infection transmission. . . . 17

2.2 Breathing cone. Spatial distribution of the source intensity B(x) is shown. . . . . . . 18

2.3 Boundary cells illustration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

4.1 Schematic of the battery pack with dual coolant channels. . . . . . . . . . . . . . . . 36

4.2 Heat input signal for the test case 1 – single input in the form of a stepwise heating rate. 40

4.3 Location of the point within the battery pack, at which temperature is probed. (a)
Front view; (b) Side view. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

4.4 Comparison between ROM and FOM solutions for the test case 1 (see Figure 4.2).
Transient temperature signals at the probed point obtained in the FOM and the ROMs
with different values of r are shown. . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

4.5 Temperature observed in the middle cross-section of battery pack for the test case 1.
The heating input signal is shown in Figure 4.2. The results of ROM with r = 15
and FOM at different instants are presented. (a) FOM t=3s; (b) ROM t=3s. (c) FOM
t=402s; (d) ROM t=402s. (e) FOM t=800s; (f) ROM t=800s. . . . . . . . . . . . . . . 43

4.6 Results for the test case 2 with a realistic transient single input. The heat input sig-
nal (yellow line, right-hand side axis) and the temperature values at the probed point
obtained in the FOM and the ROM with r = 15 are shown. . . . . . . . . . . . . . . . 44

4.7 The Bode diagrams of the FOM and the ROM at different values of r for a single-input
system. (a) Magnitude plot; (b) Phase plot. . . . . . . . . . . . . . . . . . . . . . . . 45

4.8 Heating rates for the test case 3 – multiple inputs with two stepwise heating rates. . . 46

ix
4.9 Comparison between the ROM results with r = 16 and the FOM results for the test
case 3 (see Figure 4.8 for input signals). Temperature distributions in the middle cross-
section of battery pack are shown. (a) FOM t=3s; (b) ROM t=3s. (c) FOM t=510s;
(d) ROM t=510s. (e) FOM t=1000s; (f) ROM t=1000s. . . . . . . . . . . . . . . . . . 47

4.10 Comparison of ROM and FOM solutions for the test case 3 (see Figures 4.8 and 4.9.)
Transient temperature signals at the probed point obtained in the FOM and the ROMs
with different values of r are shown. . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

4.11 Heating rate input of the battery cells for the test case 4 – multiple inputs. The numbers
indicate the heating rates applied to individual cells. . . . . . . . . . . . . . . . . . . 49

4.12 Comparison between the ROM results at r = 50 and the FOM results for the test case
4 (see Figure 4.11 for input signals). Temperature distributions in the middle cross-
section of the battery pack are shown. (a) FOM t=400s; (b) ROM t=400s. (c) FOM
t=505s; (d) ROM t=505s. (e) FOM t=600s; (f) ROM t=600s. . . . . . . . . . . . . . . 51

4.13 Comparison between the ROM and FOM results for the test case 4 (see Figures 4.11
and 4.12). Transient temperature signals at the probed point obtained in the FOM and
the ROMs with different values of r are shown. . . . . . . . . . . . . . . . . . . . . . 52

4.14 Extruded layer method for BC as input. The extruded layer is marked in pink. . . . . . 54

4.15 Inlet temperature signal for boundary condition as input validation. . . . . . . . . . . 54

4.16 Comparison of ROM and FOM solutions for the boundary condition as input. Tran-
sient temperature signals at the probed point obtained in the FOM and the ROM of
order 20 are shown. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

4.17 Comparison between the parametric ROM results at MFR = 0.02kg/s and the FOM
results. Temperature distributions in the middle cross-section of battery pack are
shown. (a) FOM t=5s; (b) parametric ROM t=5s. (c) FOM t=400s; (d) parametric
ROM t=400s. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

4.18 Comparison of parametric ROM and FOM solutions at mass flow rate = 0.02kg/s.
Transient temperature signals at the probed point obtained in the FOM and the ROM
of order 15 are shown. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

4.19 Comparison of parametric ROM and FOM solutions under transient mass flow rates.
Temperature signals at the probed point obtained in the FOM and the parametric ROM
are shown. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

4.20 Reverse flow developing in the cooling channel with time-dependent mass flow rate . . 61

5.1 Model geometry and iso-surface of infectivity field C=10 at t=100s. . . . . . . . . . . 65

x
5.2 Illustration of (a) the classroom model and (b) the computational mesh used in the
CFD simulations. The imagine from [1] is represented with permission of AIP Pub-
lishing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

5.3 (a) Turbulent kinetic energy, (b) velocity magnitude distribution, and (c) velocity vec-
tors across a slice going through students 2, 5, and 8. The imagine from [1] is repre-
sented with permission of AIP Publishing. . . . . . . . . . . . . . . . . . . . . . . . . 67

5.4 Distribution of 1 µm aerosol particles in the classroom at different points in time for
the (a) student 5 source. The imagine from [1] is represented with permission of AIP
Publishing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

5.5 Transient Scalar Intensity signal of the aerosol cloud. . . . . . . . . . . . . . . . . . . 69

5.6 Initial scalar intensity source cloud (Source is marked in red box with 0.38x0.4 m2 in
area and 0.3m in height). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

5.7 The spatial distribution of infectivity field C at t=20s. . . . . . . . . . . . . . . . . . . 71

5.8 Infectivity field C at t=50s (on the left) and 100s (on the right) . . . . . . . . . . . . . 72

5.9 Line plot of total percentage of Infectivity Field C/Particle along central direction of
classroom. The plot at t=50s is shown in (a); the plot at t=100s is shown in (b). . . . . 72

5.10 Schematic diagram of the partition of the classroom . . . . . . . . . . . . . . . . . . . 73

5.11 Transient signal at the probed point in front of student 5 . . . . . . . . . . . . . . . . 78

5.12 Distributions of C along the line at y=2.25m, z=4.5m obtained by the FOM and by
SIMO ROMs with various r. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

5.13 Contour plot at t=5s in the cross-section z=4.5m under the signal of 5.5. ROM of (a)
r=5; (b) r=15; (c) r=30; (d) r=60; (e) r=85; (f) FOM . . . . . . . . . . . . . . . . . . . 80

5.14 Contour plot at t=50s in the cross-section z=4.5m under the signal of 5.5. ROM of (a)
r=5; (b) r=15; (c) r=30; (d) r=60; (e) r=85; (f) FOM . . . . . . . . . . . . . . . . . . . 81

5.15 Infectivity field C at t=100s. (a) Optimal ROM, with r=65, using SIMO algorithm,
and (b) FOM. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

5.16 Results of frequency-shift ROMs for the source signal in Fig 5.5 . . . . . . . . . . . . 86

5.17 Distributions of C along the line at y=2.25m, z=4.5m obtained by the FOM and by
frequency-shift ROMs with various r. . . . . . . . . . . . . . . . . . . . . . . . . . . 87

5.18 Contour plot at t=5s in the cross-section z=4.5m under scalar intensity signal 5.5.
Frequency-shift ROMs of (a) r=11; (b) r=22; (c) r=33; (d) r=66; (e) r=99; (f) FOM . . 88

xi
5.19 Contour plot at t=10s in the cross-section z=4.5m under scalar intensity signal 5.5.
Frequency-shift ROMs of (a) r=11; (b) r=22; (c) r=33; (d) r=66; (e) r=99; (f) FOM . . 89

5.20 Contour plot of error vector at t=10s. ROM of (a) r=11; (b) r=99. . . . . . . . . . . . 89

5.21 Contour plot at t=50s in the cross-section z=4.5m under scalar intensity signal 5.5.
Frequency-shift ROMs of (a) r=11; (b) r=22; (c) r=33; (d) r=66; (e) r=99; (f) FOM . . 90

5.22 Breathing and coughing cycles used in simulations. Source modulations u(t) corre-
sponding to breathing cycles (a) and coughing cycles (b) are shown. . . . . . . . . . . 95

5.23 Source modulation signal u(t) used in the numerical experiment. . . . . . . . . . . . . 96

5.24 Transient scalar signals, under the input of 5.23, at the probed points in front of stu-
dents whose frontal space are identified as ’high-risk’ area. . . . . . . . . . . . . . . . 97

5.25 Cross section at t=10, 50, 220s of ROM at (a), (c), (e) and FOM at (b), (d), (f). . . . . 98

5.26 Probe line signal of (a) t=10s,(b) 50s and (c) 220s at y=2.5 m (elevation) . . . . . . . . 99

5.27 Probe line signal of t=10s (a), 50s (b) and 220s(c) at y=1.3 m (elevation) . . . . . . . . 101

5.28 Time-varying source modulation signal u(t) in simulating a lecture. . . . . . . . . . . 103

5.29 Concentrations of infectivity field C at probe points in front of students under the input
signal of 5.28. Left axis data for student 5; Right axis-data for the other students. . . . 104

5.30 At t=2000s, (a) Iso-surface of passive scalar with iso-value of 0.8; (b) Cross-section
of C. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

5.31 Integration area in front of students. . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

5.32 The data for student 5 (the source) are excluded. . . . . . . . . . . . . . . . . . . . . 106

xii
 LIST OF ALGORITHMS

1 Indirect Arnoldi algorithm for single-input systems . . . . . . . . . . . . . . . . . 27

2 Indirect Arnoldi algorithm for multiple-input systems . . . . . . . . . . . . . . . . . 30
3 Frequency-shift Arnoldi algorithm for single-input systems . . . . . . . . . . . . . 32
4 Parametric Model Order Reduction algorithm . . . . . . . . . . . . . . . . . . . . . 33

xiii
 ABSTRACT

A large variety of physical phenomena can be described by large-scale systems of linear ordi-
nary differential equations (ODEs) obtained by one of the discretization methods, in particular one
of the methods of Computational Fluid Dynamics (CFD). The solution of such ODE systems is
relatively straightforward with well-developed methods, which makes the large-scale linear sys-
tems one of the powerful ways of analyzing physical phenomena. Their practical applicability
is, however, severely limited by the computational expense. Days or even weeks may be needed
to simulate an unsteady behavior of a system with typical 106 or more degrees of freedom. This
limits applications in many important areas, from the demand for extensive solution results for fast-
paced optimization design to the need for industrial online predictive control. Therefore, efficient
yet accurate models that approximate large-scale linear systems are critically needed. We focus on
two major application scenarios: thermal management system in battery packs of electrical/hybrid
electric vehicles and the prediction of airborne transmission of respiratory infections, e.g., SARS-
COV-2, in indoor environments. The reduced-order modeling (ROM) Krylov-subspace method is
developed to reduce the computational effort of CFD. It is based on the projection of the original
model onto a Krylov subspace by the Arnoldi-type algorithms. Versions of the method for both
single-input and multiple-input systems are presented. The algorithms do not require access the
original system matrix, which is usually inaccessible from commercial CFD software. The com-
parison between the results using the ROM and the original CFD models shows a reduction by a
factor of 103 in computational time without significant loss in the accuracy of the results.

xiv
 CHAPTER 1

Introduction

In this chapter, the motivation is introduced firstly, followed by the definition of two large-scale
linear systems. The reduced-order modeling is proposed as a tool of generation of a small yet
accurate models of original large-scale linear systems. Finally, the outline of the dissertation is
presented.

1.1 Motivation

The project is motivated by the demand for computing low-order approximated models of large-
scale linear systems. Firstly, a large variety of physical phenomena are modeled with linear, time-
invariant (LTI) systems. Such models are generally produced as a result of the discretization of the
governing equations, e.g., by the finite volume and finite difference methods, into a large number
of ordinary differential equations (ODEs). The solution of an ODE system can be achieved in
various ways, e.g., by the iterative methods. Thus, large-scale linear systems are often used in
analyzing physical phenomena from practical perspectives.

1.2 Statement of the Problems

The examples of such applications include the modeling of conjugate heat transfer in a battery
pack of an electric vehicle. Despite the advantage of high specific energy and energy density with
relatively low cost compared to other types, the Li-ion battery technology is not free from prob-

1
lems. One of the major concerns is the thermal behavior of a battery pack in its on-board operation
[2]. It is known that operating Li-ion batteries outside the normal temperature range negatively af-
fects their efficiency, safety, reliability, and lifespan [2, 3]. A battery thermal management system
(BTMS) capable of effective control of the heat transfer processes, so as the battery temperature
remains within the desired range is, thus, critically important for practical applications [2].
In order to play its role, a BTMS must be able to accurately describe the temperature field inside
a battery pack and predict the evolution of this field in response to variations of load, changes in
operation of cooling condition, and other factors. The conjugate heat transfer between batteries and
cooling liquid must be reproduced. High-fidelity computational fluid dynamic (CFD) simulations
using such numerical techniques as the finite volume method or finite element method are often
utilized. The numerical techniques are applied to discretize the system’s spatial and time domain
into numerous control volumes to generate a detailed reproduction of the heat transfer process
[4]. While important and often indispensable for high-fidelity analysis, the CFD approach is often
computationally unaffordable for rapid analysis. The reason is the large scale of the models (∼ 106
to ∼ 109 degrees of freedom), which requires tens to thousands of core-hours to complete a single
simulation. This severely limits the use of CFD for onboard control, extensive simulation studies,
and simulations coupled with electrochemical battery cell models [5].
A second example of large-scale linear system is the modeling of airborne transmission of res-
piratory infections, e.g., SARS-COV-2, in indoor environments. During the current epidemics,
society has shown a poor understanding of the mechanisms of transmission of respiratory infec-
tions. As the variants of COVID-19 virus continue to break out with stronger infectious risk even
to people who have already been vaccinated [6, 7]. The understanding of these mechanisms and
the ability to predict the transmission is essential for fighting the pandemics in many perspectives,
e.g., for making consistent health policy decisions, designing prevention facilities [8, 9]. A com-
prehensive review of various aspects of the airborne transmission methods of numerical analysis,
and open questions can be found in [10].
The study of transmission mechanism usually requires extensive laboratory or field experiments

2
on infection transmission using human subjects and actual virus [11]. However, this is either
impossible or prohibitively expensive due to safety concerns. In order to successfully predict the
airborne transmission of respiratory infections, a feasible alternative is to conduct the experiments
in the form of numerical simulations. The airborne transmission process after virus-laden droplets
are generated by a respiratory activity of an infected person can be modeled in that way. The
respiratory activity forms a turbulent cloud of the virus-laden droplets that have various sizes and
velocities. The droplets diameters range from 0.1 to 1000 µm. Droplets of large size descend to
the ground or other solid surface quickly due to their high settling speed. The Lagrangian approach
is usually applied to determine particle dispersion pattern and track the pathway of each individual
particle [12].
In our work, we only consider the behavior of small droplets (< 10µm) which is referred to
as the airborne transmission [10, 13]. Their own inertia is negligible (since the Stokes number
is below 10−4 ) and settling speed is small (less than 1 mm/s), and they travel suspended in air
currents. The propagation of small droplets in a turbulent air flow is studied in the form of particle
concentration distributions in indoor environment. Since we focus on the particle concentration
distributions, the Eulerian approach, which considers the particle cloud as a continuum where the
concentration of some particle-related characteristics, e.g., alive virions, is applied. The concentra-
tion is expressed by a scalar field evolution of which is determined by a transport equation [10, 14].
Details will be discussed later in Section 2.1.2.
It is important to conduct numerical experiments in order to [15, 16]:

• Predict the distribution of the airborne virus-laden particles from various configurations,
e.g., population distribution or density in the room, different droplet sizes, different mass
flow rates of ventilation;

• Understand the decay of virus viability in turbulence;

• Assess the risk of airborne transmission in an indoor space;

• Draft the public regulation to mitigate the indoor airborne transmission;

3
 • Understand the mechanism of infection.

Therefore, extensive numerical experiments are needed to study airborne transmission. How-
ever, the computation cost of each such experiment makes it difficult to accumulate a sufficiently
large number of data by performing parametric studies under extensive parameters.

1.3 Proposed Approaches

Solution of large-scale linear system modeling is common and typically required for a detailed
and accurate description of the physical phenomena. Such solutions are fairly expensive, which
limits the use of numerical modeling to be inefficient in practical applications. The systems in all
examples mentioned above consist of ODEs with more than 106 degrees of freedom. A simulation
of a transient process takes days of calculations even with parallel computations. Although such
high-fidelity models tend to accurately describe the behavior of the physical system, the computa-
tional expense makes it highly impractical.
One of the promising ways to save the computation cost while maintaining high fidelity in simu-
lations of large-scale linear system is the reduced-order modeling (ROM). The ROM approximates
a large-scale system by a small yet accurate system. Compared with large-scale linear systems, the
ROM has the critical advantages of efficiently predicting the system behavior while dramatically
reducing the computation cost. The work presented in this dissertation focus on development,
verification, and use of such models.

1.4 Literature Survey

Several attempts to develop an ROM for battery thermal management have been made before.
A control-oriented model based on the singular perturbation method was proposed in [17]. The
method predicts the battery cell’s internal temperature profile by applying the Laplace transform to
the one-dimensional boundary-value problem and further reducing it into a low-order linear model

4
in the frequency domain. However, the effect of heat removal from the cooling components is not
considered in the model, making its applications very limited.
In [18] and [19], ROMs are developed by parameter identification using experimental data
under different geometries, inlet flow velocities, and temperature. The models relatively accurately
predict temperature on the cell’s surface, in its core, and average temperature of the cell, which are
applied for control purposes. However, the extensive experiments under various system parameters
required to derive the models make their use ineffective. Besides, the models become impractical
when the complete temperature field is needed for the design and optimization of the BTMS.
An efficient reduced-order modeling for large-scale BTMS based on singular value decomposi-
tion (SVD) was presented in [20, 21]. The model was obtained by applying the SVD to snapshots
of high-fidelity CFD solutions. The method was shown to be effective. Its major drawback is
that the approximation accuracy strongly depends on the selection of snapshots, which makes it
impractical.
Comprehensive reviews of the ROM methods in general can be found in [22, 23, 24]. The
proper orthogonal decomposition (POD) method uses selected instantaneous states of the system
during its evolution as basis vectors. The method is applicable to linear and nonlinear systems,
but it is computationally expensive for large-scale systems [25]. The balanced truncation method
determines a balancing transformation by computing the controllability and observability Grami-
ans from the Lyapunov equations. The method can preserve the asymptotic stability of a stable
full-order modeling (FOM) while allowing the optimization of ROM by the error bound. The
computations needed to solve the high-dimension Lyapunov equations are, however, extensive,
which makes the method less suitable for large-scale systems [24, 25].
The projection-based Krylov-subspace method based on moment matching is considered to
be a good candidate for ROM development. The method computes the basis vectors that span
the Krylov subspace. The basis vectors are used to reduce the FOM into a low-order state-space
system such that the Taylor series coefficients of the transfer function (also called moments) are
matched between the ROM and FOM. The method is only applicable to linear systems, but known

5
to deliver good performance (see [26] for a review). Examples of its use are the developments
of reduced-order models for micro electro-mechanical system (MEMS) actuators and microflu-
idic chip thermal systems [27, 28], the near-range atmospheric dispersion [29], structural-acoustic
phenomena [30], and the aeroelastic analysis of turbomachines [31].
There are many algorithms available for construction of the Krylov subspace. One is the Padè
approximation via Lanczos (also known as ‘PVL’) algorithm [32], which generates two sequences
of basis vectors spanning the input and output Krylov subspaces. The method is impractical be-
cause the generated ROMs occasionally become unstable even though the original FOM system
is stable [33]. A better alternative is the Arnoldi algorithm, in which the modified Gram-Schmidt
process is applied to construct the orthonormal basis of the Krylov subspace [34].
A common feature of the Krylov-subspace method algorithms is that they require access to the
system matrix that contains the coefficients of the discretized governing equations of the FOM.
This hinders application of the methods for practical CFD analysis, in particular for the thermal
analysis in the BTMS, since most commercial CFD solvers, such as ANSYS Fluent [35] or STAR-
CCM+ [36], do not provide a user the access to the system matrix. While an accurate ROM is often
needed for the efficient thermal analysis in the BTMS and multiphysics modeling in the industrial
product lifecycle management environment, few studies have attempted the Krylov-method ROM
development using approaches that do not require access to system matrix. One such indirect
method based on the Arnoldi algorithm was reported in [29] to predict the hazardous pollutant’s
dispersion into the atmosphere. The method, however, was demonstrated for the single-input sys-
tems only. To our best knowledge, several Arnoldi-type algorithms developed for the multiple-
input systems using Krylov subspace are not applicable to the situations when system matrices are
not accessible [27, 37, 38].
We attempt to fill the gaps mentioned above by reviewing the previous works. With this purpose
in mind, the preliminary result includes the developed algorithms for large-scale linear dynamical
system describing the conjugate heat transfer. The major advantage of proposed approach is that it
does not need access to the system matrix. The method is based on the projection of the FOM onto a

6
Krylov subspace by the Arnoldi algorithm. Versions of the algorithm for single-input and multiple-
input systems are developed. The algorithms are applied to derive ROM models of conjugate heat
transfer in a battery pack. Single-input (when the system is controlled by a single heat source term)
and multiple-input (when components of a pack are controlled individually) cases are considered.
The ROM is used to predict the unsteady transient conjugate heat transfer in an entire battery
pack. The comparison between the predictions and the results of high-fidelity CFD modeling
shows a dramatic reduction in computational time without significant loss in the accuracy. The
developed techniques are not restricted to BTMS and can be employed, in general, for reduced-
order modeling of time-invariant linear dynamical systems including those describing conduction,
convection or conjugate heat transfer or transport of admixtures.
Concerning the second system considered in the dissertation, there have been extensive dis-
cussions of the airborne transmission of respiratory infections, since it is identified as one of the
major risks of infection [39]. Studies indicate that the airborne transmission at outdoor locations
are posing a light risk of infection [40]. Compared with that, the airborne transmission is identified
as a substantially risky scenario in indoor environments [41, 42].
In [43, 44, 45], it is found that the parameters of the droplets, such as their number, size,
initial velocity, and viral load, vary greatly depending on the type of activity (breathing, sneezing,
coughing, singing, etc.), and physical characteristics and the stage of infection of the source. As an
example, the droplet size can be anywhere between 0.1 and 1000 µm [10, 12, 44]. Droplets larger
than a few tens of µm have a significant settling speed, so they soon descend to the ground or
another solid surface 1 . There are also some droplets with a diameter between 10 µm and 100 µm
that will attach to the top of a possibly moving surface in a short time ∼ 10 min. There, virions
may remain infectious for several hours to several days depending on the type of the surface,
humidity, UV irradiation, and other factors. The last part of the droplets with a diameter of less
1
The typical distance traveled by such particles before attaching to a surface was determined in the 1930s and
1940s, most notably in [46]. It appears that those findings form the basis of the modern recommendation of the safe
person-to-person distance of 6 ft. The problem with the findings and, thus, the recommendation is that the threshold
was identified in [47] as a distance traveled by droplets of size 100 µm issued by an average coughing source. It is
now understood that smaller particles ejected with higher velocity (e.g., by sneezing) may travel 20 ft or even further
[48].

7
than 10 µm will travel in the indoor air with the airflow in the form of aerosol for a long period
of time and broad range of space. The proposed work will focus exclusively on this mechanism
of transmission, namely by droplets of 10 µm and smaller. Their own inertia is negligible (since
the Stokes number 2 is below 10−4 and settling speed is small (less than 1 mm/s), and they travel
suspended in air currents. Water quickly evaporates from their surface, so they become particles
consisting of virions and semi-solid residue. The viability and, plausibly, infectious potential of
the virions have been measured to decrease by half in about 1.2 h [49].
Considering that each particle contains a large number of virions, and that a small number of
(theoretically, one) virions are sufficient for infection, each particle retains infectious potential for
many hours. This scenario suggests the possibility of airborne transmission [50, 51]. A person
is infected after inhaling or contacting in some other way a small airborne particle generated,
possibly, long time ago, by a source located far away. The possibility appears especially realistic
in indoor environments, where strong recirculating flows of air are generated by ventilation systems
and heat sources (humans, computers, etc).
In a typical approach, (see, e.g., [48]), one estimates the number of potentially infectious doses
produced by a source over a time period and then evaluates the probability of transmission based
on the assumption that the doses are uniformly distributed in the room. The assumption and the ap-
proach itself are prone to criticism from the physical viewpoint. Turbulent transport from a source
of time-dependent intensity inevitably results in a distribution of particles, which is both strongly
non-uniform (high concentration in some areas and zero in others) and unsteady. These models
currently used in epidemiology ignore the uncertainties or treat them in an oversimplified way.
Computational Fluid Dynamics (CFD) approach offers a more accurate alternatives. Instead, one
of the viable ways to study the airborne transmission is achieved by a RANS (Reynolds-Averaged
Navier-Stokes) model that computes steady-state averaged velocity and turbulent diffusivity of air
flow fields [52] as a ‘base’. Then, either Lagrangian or Eulerian can be used to reproduce the prop-
agation of aerosol particles. Depending on the focus of research, these two methods are applied to
2
The Stokes number is defined as the ratio of the characteristic time of a particle to a characteristic time of the flow
or of an obstacle. It is a dimensionless number characterising the behavior of particles suspended in a fluid flow.

8
substantially different scenarios according to characteristics of the methods.
Lagrangian method is applied in [53] to analyze transmission in a bus with windows and doors
closed, at half-seated occupancy under COVID-19 restrictions, with the main HVAC at maximum
flow rate, the driver HVAC on low, and the defroster on medium. The work in [15] presents
CFD study of the influence of an alternate ventilation configuration on the possible flow path of
infectious aerosols using Lagrange method. The investigation in [44] computed the transport of the
droplets exhaled by the index patient at designed position. The bulk airflow pattern was calculated
using CFD RANS while the airborne droplet was modeled using a Lagrangian method. It can
simulate the trace of particles under different diameters, and even the process of thermodynamic
changes of the droplet during the propagation process, and therefore can predict the dispersion
pattern of the particles. However, a very dense mesh must be used otherwise the calculated path
will vary greatly due to poor quality of the grid.
A numerical model using RANS and the Eulerian approach is presented in [54] to study the
transient behavior of cough particles transport in a chamber. The work evaluated the risk of infec-
tion under different ventilation scenarios for which the conclusion was confirmed by experimental
data. The study [10] investigated the aerosol transmission under different droplet sizes consider-
ing the particle lingering over time in a generic public place. It assessed the risk of infection as
‘exposure time’ over indoor environment. Through solving the transport equation with Eulerian
method, the research in [55] estimated the transport of indoor contaminant. Better agreement with
respect to measured chamber concentrations could be found both qualitatively and quantitatively
over those using the uniform inlet velocity.
The Eulerian approach can predict the spatial distribution of a concentration field. The dis-
tribution can then be used in a statistical analysis. The result can be validated by comparing
both qualitatively and quantitatively with physical experiments. Through the spatial distribution
of aerosol, we can effectively predict the high-risk infection area. The Eulerian method is more
efficient than the Lagrangian method in prediction of particle distribution. The limitation of the
Eulerian method is its inability to reproduce the inertial effects of the particle evolution. It is,

9
therefore, recommended to apply the Lagrangian method for large particles, for which the iner-
tial effects are significant. The Eulerian method is best suited for studying propagation of small
particles characterized by small Stokes number.
Overall, due to a substantial computation effort, the Lagrangian method cannot efficiently han-
dle the amount of calculations in simulating series of activities. Because the number of particles
produced by speaking is significant as it is normally done continuously over a long period of time.
The simulation constantly exhales a huge number of particles to calculate its trajectory. This task
will consume a huge amount of computing resources for Lagrangian approach.
Since the focus of our research is on predicting the particle concentration distributions and
since the attention is limited to aerosol particles with St ≪ 1, we apply the Eulerian approach.
The method model the particles as a continuum where the concentration of particle-related char-
acteristics, e.g., alive virions, is represented by a scalar field in a transport equation [10]. Then, a
transport equation for the concentration of the droplets can be solved with precomputed velocity
and diffusivity fields. The intensity of the source can be derived from the experimental data on
expiratory droplet formation, and virus loading [44, 47, 48].
CFD modeling the airborne transmission is computationally expensive, e.g., [10]. It cannot,
therefore, be used in large-scale statistical studies taking into account variability of the source of
infection. It cannot also be used for real-time control of the strategies of prevention of infection.
An effective alternative explored in this work is to develop an ROM. It in general requires off-
line computation to approximate the original full-order model into a model that has much smaller
size. Once ROM is generated, however, it can be used in a large number of fast and accurate
predictions. To the best of our knowledge, the method of reduced-order modeling has never been
applied to analysis and prediction of indoor airborne transmission of respiratory infections. On this
basis, algorithms that use the Krylov subspace method are developed to build an efficient, reliable
ROM for the study of the field.

10
1.5 Objectives

As a summary, we will develop and apply reduced-order models for two configurations: the
heat transfer in the battery pack of an electric vehicle and the airborne transmission of respiratory
infections. The work is outlined in Figure 1.1. Each study will start with the modeling of a high-
fidelity large-scale system that predicts the physical behavior accurately. This part is categorized
as the ‘Modeling based on physics’. The second part identified in Figure 1.1 as ‘Reduced-order
modeling’ will include development, analysis and validation of ROMs. Efficient algorithms for
single-input and multiple-input scenarios will be developed and implemented to construct the ROM
that can accurately approximate the high-fidelity large-scale system. The accuracy of the ROMs
is confirmed in verification studies, in which predictions of the ROM and of the full-order models
are compared for realistic scenarios.

Figure 1.1: Dissertation outline

11
 CHAPTER 2

Theoretical and Numerical Models

As introduced in Chapter I, two examples of large-scale linear systems are considered. Al-
though different governing equations are used, the numerical solution leads to similar linear sys-
tems which can be written in the form of state-space representation as presented in 2.1. The
reduced-order modeling methodologies for large-scale linear systems are presented in 2.2.

2.1 High-Fidelity Large-Scale Linear Systems

2.1.1 Conjugate Heat Transfer in a Battery Pack of an Electric Vehicle

Firstly, we consider the conjugate heat transfer in the battery pack of an electric vehicle. The
proposed ROM method, while generally applicable to a wide array of linear problems, will be
applied and tested for the specific conjugate heat transfer system described in this section. The
system is a simplified model of a typical automotive battery pack. Details of battery chemistry
are ignored, and the effects of charging and discharging on heat transfer are simulated as time-
dependent internal heat sources distributed over the battery interiors. At the same time, all the
key features of a typical battery pack directly relevant to heat transfer and BTMS are accounted
for by the model. This includes a realistic set of physical properties and a realistic geometry with
multiple coupled subdomains: solid battery cells, casing, cooling plate, and cooling channels with
water flowing through them (see section 4.1).
The remaining discussion in this section has two parts: (i) the list of further simplifying assump-

12
tions and corresponding governing equations of heat transfer and (ii) the state-space representation
of the problem from a CFD dicretization.
The battery system includes coupled solid battery cells and liquid cooling substance domains.
The battery is cooled by heat transfer from solid battery cells to the liquid cooling substance. While
the heat generation rate in cells is determined by Joule heating, phase change, mixing effects and
electrochemical reactions [56], the electrochemical behavior of the battery cell and the effect of
temperature on cell behavior are not considered. Instead, since we focus on examining the cooling
performance of the pack design, the cumulative heat generation rate by battery cells is imposed as
input.
The battery system has multiple coupled subdomains, including solid battery cells and liquid
coolant zone. The heat transfer in the solid region is by conduction and, thus, described by the
linear partial differential equation for temperature:

∂T 1 1
= ∇ (κ∇T ) + Q, (2.1)
∂t ρCp ρCp

where T is the temperature, Q is the volumetric density of internal heat source, and ρ, Cp , κ are
the density and specific heat capacity of the solid, which do not depend on time and temperature,
but can vary in space since their values are different for different subdomains. The thermal con-
ductivity κ is also a function of space. It can also be anisotropic, so κ is a tensor, rather than a
scalar.
The heat transfer in the liquid subdomain is a non-linear process governed by coupled mo-
mentum, continuity, and energy equations. This problem can be simplified to a linear system of
equations with constant coefficients for temperature by adopting the following commonly valid
assumptions:

Negligible effect of natural convection. In many thermal management systems, the convection
heat transfer in liquid subdomains is dominated by forced convection. The temperature-
dependent buoyancy force in the momentum balance can be ignored.

13
Constant flow properties. We assume that the fluid is incompressible and Newtonian and that all
the physical properties, such as density, thermal conductivity, specific heat, and viscosity
are constant. This approximation is widely adopted in heat transfer analysis and should be
considered as adequate so far as the temperature variations within the system are not too
large (see [57] for a detailed discussion).

Steady-state flow velocity field The fluid flow is assumed to be fully developed under a certain
constant mass flow rate at the inlet. Thus, the flow velocity field U is steady-state. This is
justified by our examples considered in section 4.1, where we show that the values of the
Reynolds number are below or about the value Re ≈ 2000 of transition to turbulence. We
note that the assumption can be extended to the case of turbulent flows with time-independent
mean velocity, in which case a RANS turbulence model has to be applied (see, e.g. [52]).

The first two assumptions assure that the values of physical properties and velocity are
temperature-independent. The third condition implies that the coefficients are also independent
of time. The steady-state flow velocity field and, in the case of turbulence, eddy diffusivity can
be pre-computed using a CFD solver. The solver computes solutions of the equations of mass and
momentum conservation for incompressible fluid flows in the coolant channel. Then, the fields are
used to calculate the coefficients in the temperature equation for both FOM and ROM generation.
It must be noted that the computational time required for pre-computing the fluid flow is small in
comparison with the time needed to compute the steady-state temperature field. The flow calcu-
lations are limited to a small portion of the system consisting of the coolant channel with simple
geometry, which substantially reduces the computation effort. Overall, the assumptions decouple
the Navier-Stokes equations for velocity and pressure from the energy equation for temperature in
the fluid region. In each analysis of the system, a steady-state solution of the Navier-Stokes equa-
tions at given flow parameters, such as the flow rate, is calculated once to determine the velocity
field. The energy equation describing various heat transfer scenarios at the same flow can then be
considered as a linear equation with constant physical properties and constant three-dimensional
flow velocity field U (the eddy diffusivity term is omitted for simplicity):

14
 ∂T κ 1
+ U∇T = ∇2 T + Q. (2.2)
∂t ρCp ρCp

Our next step is to represent (2.2) as a linear time-invariant (LTI) system. The partial differential
equation (2.2) is discretized, in our case by the finite volume method, to produce a system of
ordinary differential equations for the values of temperature at n grid points (centroids of finite
volume cells). The system can be formally written as a state-space system with m input and n
state variables as shown later in this chapter.

2.1.2 Airborne Transmission of Respiratory Infections

The second physical system to be analyzed is the propagation of virus-laden droplets in turbu-
lent air flow. Airborne transmission of a respiratory viral infection, such as COVID-19, is con-
sidered. Macro droplets containing active virions are exhaled by an infected person [39]. The
parameters of the droplets, such as their number, size, initial velocity, and viral load, vary greatly
depending on the type of activity (breathing, sneezing, coughing, singing, etc). In line with the
intended focus on airborne transmission, the model only considers droplets of small size, i.e., di-
ameter below 10µm. We analyze the transport of such droplets in a turbulent air generated by an
indoor ventilation system.
The RANS (Reynolds-Averaged Navier-Stokes) model is used to study the flow dynamics of
ventilation system. The three-dimensional steady-state flow equations are solved to find the spa-
tial distributions of the mean velocity and turbulent diffusion coefficients. Then, the diffusion-
convection equation based on the Eulerian description is solved to predict the spatial-temporal
evolution of a scalar field representing the volume concentration of infectious virions in air. We
assume that:

Negligible effect of aerosol particles size on the air flow. Their own inertia is negligible (since
the Stokes number is below 10−4 ) and settling speed is small (less than 1 mm/s), and they
travel suspended in air currents. In this way, the air flow equations are decoupled from

15
 the transport equation. Once the flow velocity field and turbulent diffusivity of air flow are
found, it is frozen and used in calculating the transport of the aerosol particles

Droplet size does not change during the airborne transmission process. In the process of ex-
halation, a mucous droplet undergoes evaporation. The process reduces the diameter of the
droplet and makes it nearly vitrified. The process depends on the environment conditions
such as air humidity and temperature [58]. The evaporation process is very rapid, usually
lasting less than 0.3s. Therefore, we ignore the drying process and assume that the aerosol
cloud forming in the course of exhalation consists of fully evaporated droplets of fixed di-
ameters.

The exhaled aerosol forms a cloud in the shape of a cone. All the released aerosol is mixed to
a control volume of a breathing cone after the exhalation.

Steady-state flow velocity field The fluid flow is assumed to be fully developed under a certain
constant mass flow rate at the inlets of an indoor ventilation system. Thus, the mean flow
velocity field U and the turbulent diffusivity is steady-state. χt produced by a RANS model
are assumed to be constant [52].

The fate of virions in aerosol cloud. The viability and, plausibly, infectious potential of the viri-
ons contained within an aerosol particle have been measured to decrease by half in about 1.2
h [39, 49]

Incompressible flow with constant fluid properties The constant flow properties are assumed in
analyzing the transport of airborne.

In order to characterize the infection potential of the airborne aerosol, we introduce the infectivity
field C(x, t). This is scalar field representing the local value of the number of active virions per unit
volume of air. Its evolution in a turbulent flow is described by the convection-diffusion equation:

∂C
+ U · ∇C = ∇ · (χt ∇C) − σC + S (2.3)
∂t

16
where x is space location, t is time, σ is the rate of decay of virus viability. S(x, t) is the intensity
of the source, which can be derived from the experimental data on expiratory droplet formation and
virus load [44]. Time-independent velocity fields U(x) and turbulent diffusivity χt (x) of air flow
are precomputed using a CFD solver, where steady-state equations describing the conservation of
mass, momentum and energy as well as the equations for characteristics of turbulence in the form
of a RANS model are solved.

Figure 2.1: Schematic representation of various modes of respiratory infection transmission.

The exhaling forms a breathing cone-shaped cloud in front of the infected person illustrated
in Figure 2.2. Previous studies have shown that the geometry of the aerosol breathing cone varies
greatly depending on the person, climatic conditions, and exhalation behaviors. Our study defines a
generic exhaling cone with a high density along the cone’s axis while the aerosol gradual reduction
towards its edges. Along with cone’s axis, the peak value decreases exponentially with the distance
from the mouth. We hypothesize that the defined breathing cone intensity change proportionally
with the different exhaled intensities under different exhaling activities.

17
 Figure 2.2: Breathing cone. Spatial distribution of the source intensity B(x) is shown.

Based on the experimental data, the particle concentration exhaled by an adult is: from 2.4 to
5.2 cm−3 per cough or 0.004 to 0.223 cm−3 per second. The total number of the droplet expelled
ranged from 947 to 2085 per cough and 112 to 672 per second [58].

2.1.3 Consistent Steady-State Representation of Large-Scale Linear Sys-

tems

The equations have to be solved by CFD solvers, which convert (2.2) or (2.3) into the form
of state-space representation that consists of millions of ordinary differential equations. The set
of ordinary differential equations is built for the values of interests, i.e., values of temperature for
(2.2) and values of infectivity field for (2.3) at the points of a discretization grid, in the consistent
form:

ẋ(t) = Ax(t) + Bu(t) (2.4)

where x(t) ∈ Rn is the solution vector of grid-point values, the constant sparse matrix A ∈ Rn×n
is the discretization coefficients matrix of the ODE system from (2.2) or (2.3), which incorporates
the terms of the governing equations. We assume the asymptotically stable systems, so A is a

18
negative-definite matrix. The input matrix B ∈ Rn×m can be a vector for a single-input system
(m = 1) or a multi-column matrix for a multiple-input system (m ≥ 2). Each column vector in
B represents the spatial distribution of the system’s input, which is defined as a volumetric heat
source in (2.2) or volumetric source of infectivity in (2.3). The system (2.2) or (2.3) is solved
by a general CFD software, e.g., STAR-CCM+, Ansys Fluent, or OpenFOAM, as a ‘black box’
such that the system matrix A is not explicitly available. The systems can be approximated via
reduced-order modeling, which is described shortly.

2.1.4 Boundary Condition as an Input

For an ODE system where a Dirichlet boundary condition is imposed, it specifies the values
that a solution needs to take along the boundary of the domain. Such a boundary condition with a
fixed or time-varying value often plays an important role in the dynamics of the system and can be
considered as an input.
As an example, we consider the conjugate heat transfer of a battery pack, where the inlet tem-
perature of the coolant is applied. The temperature of the inlet coolant is one of the essential
parameters governing the conjugate heat transfer process of the system. Therefore, it would be
useful to treat it as an input control parameter in (2.4). The Dirichlet boundary condition is repre-
sented by discretization coefficients γ in the system matrix:

ẋ(t) = Ax(t) + Bu(t) + γ (2.5)

where B can be an input matrix that B = b1 . As we apply the inlet coolant temperature as an
input control parameter, γ can be represented as an additional input vector such that
 γ = b2 u2 .
 u1 
 
In this case, the system (3.4) can be represented as ẋ(t) = Ax(t) + b1 b2   where
u2
the boundary condition is represented by b2 = γ multiplied by a transient signal u2 , i.e., transient
boundary condition temperature TBC .
The spatial coefficients of boundary conditions γ need to be explicitly computed in order to

19
construct the boundary condition as an input control parameter. An additional extruded layer of
Finite Volume cells at the inlet is created. The computation of γ mimic the convective heat flux
at the inlet by volumetric heating within the extruded cells. Considering that the Finite Volume
Method (FVM) is used for the CFD solver, the discretized form of the energy equation is:

d X X k  
q̇

(ρT U) + [ρT (U · a)]f = ∇T · a + · V olume (2.6)
dt f f Cp
f Cp

where ρ, k and Cp are the density, heat conductivity and specific heat in the finite volume, T and q̇
are the temperature and volumetric heat source stored at the center of a finite volume cell. U is the
velocity field vector. The vector a represents the surface area vector. V olume is the cell volume.
In (2.6), the convective flux term at a boundary face is expressed by:

[ρTi (U · a)]f = ṁf · TBC (2.7)

where ṁf is the mass flow rate at the boundary face. TBC is the temperature value as boundary
condition at the finite volume.
The diffusive flux term in (2.6) at a boundary face is:

k k
[ ∇Ti · a]f = (TBC − Ti ) · α
⃗ ·a (2.8)
Cp Cp

a
where α
⃗= a·ds
and dS = xf − x0 is the distance between face center and cell center and Ti is the
temperature at the cell center.

20
 Figure 2.3: Boundary cells illustration

We see that the inlet temperature contributes to the free term of the discretization equation
as illustrated in Figure 2.3, if all the Finite Volume cells adjacent the boundary have the same
thickness, the contributions are the same for all boundary cells and can be rearranged as:

TBC (Cconv +Ccond )·ABC

VBC
−1
= TBC (Cconv + Ccond ) · δextrudedM esh
(2.9)
k −1
= TBC (ṁf + Cp
) · δextrudedM esh = F = constant

where Cconv , Ccond are the coefficients of convective (2.7) and diffusive (2.8) flux term from dis-
cretization equation at a boundary cell. ABC and VBC are the boundary surface area and the volume
of the boundary cell. δextrudedM esh is the thickness of the extruded layer.
Evidently, the same contribution to the discretization equation is achieved by the internal heat-
ing with volumetric density F within each extruded cell. We choose this interpretation and add the
heating as an input parameter while setting TBC in the CFD problem to zero.

2.2 Reduced-Order Modeling Methodologies

2.2.1 Krylov-Subspace Method

The method utilizes projection onto the Krylov subspace to reduce the FOM system (2.4) into
a system of lower order r. As discussed, e.g., in [26, 28, 29, 30, 34, 38, 59], the resulting ROM

21
matches the first r coefficients of the Taylor series expansion of the transfer function of the FOM.
The coefficients are called ‘moments’ in the reduced-order model, whereas the model reduction
based on the Krylov subspace is also called moment-matching model order reduction [38].
The Krylov subspace of order r is defined as [59]:

  n o
κr Ã, b̃ = span b̃, Ãb̃, ..., Ã(r−1) b̃ , (2.10)

where à is a constant matrix and b̃ is a starting vector generating the subspace. For the purpose of
deriving an ROM for conjugate heat transfer problems, which approximates the original system’s
behavior in the low frequency domain, the appropriate selection is à = A−1 and b̃ = A−1 B [27].
On the other hand, the ROM is generated in a broader range of frequency for airborne transmission
analysis, the details will be discussed shortly.
The model reduction is achieved by projection of the FOM solution vector x(t) on the Krylov
subspace [26]. With an orthonormal matrix V , whose column vectors span the subspace, this is
expressed as
x ≈ V xr , (2.11)

where xr is the vector of the approximate solution in the Krylov subspace.

The system (2.4) is approximated as

V ẋr = AV xr + Bu. (2.12)

The ODEs for xr (t) (the actual ROM) are obtained by the Petrov-Galerkin projection using another
orthonormal matrix W ∈ Rn×r . Following [29], one-sided Arnoldi algorithm is utilized, which
implies W = V , so W T V = I, where I is the identity matrix. Left-multiplying (2.12) by W T we
obtain

x˙r (t) = Ar xr (t) + Br u(t), (2.13)

where Ar = W T AV ∈ Rr×r and Br = W T B ∈ Rr×m .

22
 The first r/m moments of the transfer function of the ROM obtained in this manner match the
first moments of the full system [30, 29]. Since r ≪ n, the reduced order system (2.13) can be
solved with a much smaller computational effort. Once the system is solved, an approximation of
the solution state vector in the full-order space can be reconstructed as x(t) = V xr (t).
The two-sided Arnoldi algorithm, where W ̸= V , can potentially improve the moment-
matching property [38]. However, this is true only when the order of the outputs is considerably
smaller than the order of the original system [38]. Since the temperature distribution in the entire
battery pack is of primary interest in our problem, the output consists of n values. Therefore, only
the one-sided Arnoldi algorithm is employed in this study.

2.2.2 Parametric Reduced-Order Modeling

The parametric reduced-order modeling (pROM) is presented in this section. The model allows
us to generate ROM for an arbitrary value of the system parameter entering the system matrix A,
e.g. the mass flow rate of coolant, and input matrix B using interpolation from the reduced order
models already developed for several values of the parameter [60].
We firstly introduce the FOMs with k different parameters, which are obtained from several
mass flow rates:

ẋi = Ai xi + Bi u. (2.14)

where parameter-dependent matrices Ai ∈ Rn×n and Bi ∈ Rn×1 are spatial discretization coeffi-
cient in system matrices and heat input, respectively.
To start with the construction of pROM, k FOMs are reduced independently with its indivadual
projection matrices Wi , Vi . Thus, we have k developed ROMs, hereafter called local ROMs:

WiT Vi ẋr,i = WiT Ai Vi xr,i + WiT Bi u (2.15)

where the reduced state vectors xr,i in each ROMs belong to different vector spaces since they are

23
projected by different local projection matrices, i.e., xi ≈ Vi xr,i . To interpolate between them,
we need to transform the local ROMs into a common subspace. The transformation is achieved
by applying the singular value decomposition (SVD) to the local projection matrix dataset Vall =
 
V1 V2 ... Vk . The first r columns of the orthogonal matrix U from the SVD form the
transformation matrix R ∈ Rn×r . After R is formed, each local projection matrix Vi , Wi can
be transformed into a common reduced-order space, where every local ROMs have a consistent
interpretation. Each local projection matrix is transformed by the matrix R, so that the transformed
local ROM gives:

Mi Ti−1 ẋ∗r,i = Mi WiT Ai Vi Ti−1 x∗r,i + Mi WiT Bi u (2.16)

−1
where Mi = WiT R and Ti = RT Vi . x∗r,i is the transformed state vector, i.e., xr,i = Ti−1 x∗r,i .
Now, the transformation matrices Mi and Ti are computed to allow the transition from each
coordinates of local ROMs vector xr,i to a modified coordinate system x∗r,i that share the same
basis R. The pROM can be constructed by interpolation with the weighted coefficient ϖi . Given a
pROM with parameter p, the weighted coefficients are obtained such that ki=1 ϖi (p) = 1. Vari-
P

ous interpolation methods can be applied for the weighted coefficients such as linear or quadratic
interpolation. For example, when two local ROMs are selected to construct the pROM, linear
interpolation can be applied in calculating the weighted coefficients.
The parametric reduced-order model can be constructed as:

Xk Xk Xk
ϖi Mi Ti−1 ẋ∗r = ϖi Mi WiT Ai Vi Ti−1 x∗r + ϖi Mi WiT Bi u (2.17)
i=1 i=1 i=1

The pROM ODE is:

ẋ∗pmor (t) = Apmor x∗pmor (t) + Bpmor u(t) (2.18)

where the ROM system matrix Apmor = ( ki=1 ϖi Mi Ti−1 )−1 ( ki=1 ϖi Mi WiT Ai Vi Ti−1 ) and
P P

Bpmor = ( ki=1 ϖi Mi Ti−1 )−1 ( ki=1 ϖi Mi WiT Bi ). The state vector of the full-order system is
P P

24
 k
ϖi Vi Ti−1 x∗pmor (t).
P
reconstructed from parametric reduced-order state vector as
i=1
The results of pROM will be presented later which shows excellent approximation with original
high-fidelity FOMs.

2.2.2.1 Time-Varying Parameters Using the pROM Framework

We developed the applications with time-varying parameters when the system matrix A(t) and
input matrix B(t) change with time under the current framework of the pROM. The application
aims to predict the batteries’ thermal behavior under a time-varying mass flow rate signal. Since
the mass flow rate is changing over time, the state-space representation of the system becomes:

ẋ(t) = A(t)x(t) + B(t)u (2.19)

where A(t) is the system matrix that changes with time. We can apply the parametric model to
approximate the system matrix by interpolation (2.17).
The ROM for the time-varying MFRs is:

ẋr (t) = Ar (t)xr (t) + Br (t)u (2.20)

where Ar (t) and Br (t) are interpolated based on MFRs. The procedure is a generalization of the
interpolation procedure described in section 2.2.2. For every value of the time step tj = t0 + j∆t,
we use the instantaneous mass flow rate p(tj ) to compute the interpolation coefficients ϖi (tj ) and
then the matrices Ar (tj ) and Br (tj ). In this case, the parametric model can be applied to approxi-
mate the system matrix by interpolation. The discussion and limitation of the implementation will
be shown shortly.

25
 CHAPTER 3

Algorithms

We introduce the indirect Arnoldi algorithms for single-input systems in 3.1 and multiple-input
systems in 3.2. The frequency-shift Arnoldi-type algorithm is introduced in 3.3. The algorithm
developed for parametric reduced-order modeling is presented in 3.4. Finally, we present the
change of variable method in 3.5 for the systems, in which inhomogeneous Dirichlet boundary
conditions are applied.
The proposed reduced-order modeling (ROM) approach and the developed algorithms for
single-input and multiple-input systems in the current chapter have been successfully implemented
in our GM-funded project on conjugate heat transfer in the battery pack of an electric vehicle [61].
Further development of the algorithms for airborne respiratory transmission systems is carried out
by the frequency-shift algorithm.

3.1 Indirect Arnoldi Algorithm for Single-Input Systems

The indirect version of the Arnoldi algorithm proposed in [29] and presented in Algorithm
1 is used for problems with a single input, i.e., with m = 1. The version is different from the
original Arnoldi algorithm because it computes the reduced system matrix as Ar = V T X rather
than V T AV . The orthogonalization matrix X = AV is constructed simultaneously with V . It
contains the information from the system matrix A and makes direct access to A unnecessary.
As one can see in Algorithm 1, the column vector xi of X is generated as the basis vector vi−1

26
orthonormalized through the modified Gram-Schmidt process with respect to the previous basis
vectors in X.
The key (and the only computationally demanding) element of the algorithm is the calculation
of the matrix products ṽ1 = A−1 b1 and ṽi = A−1 vi−1 , i = 1, . . . , r. They can be obtained
without direct access to A by applying the CFD solver to steady-state heat transfer problems,
discretization of which corresponds to Aṽ1 = b1 and Aṽi = vi−1 . For example, the initialization
step of Algorithm 1 includes a CFD solution of the steady-state version of (2.4) with the source
term −b1 . The computed values of temperature at the grid points form the vector ṽ1 . The following
steps require similar CFD solutions of heat transfer problems with source terms determined by the
previously generated vector vi−1 .
The basis vectors vi are computed by taking the recursive orthogonalization of A−1 vi−1 to the
previously generated orthonormalized basis vectors through the modified Gram-Schmidt process.
The basis vectors xi are computed and orthonormalized to satisfy xi = Avi , accordingly. Once V
and X are computed, the ROM matrix is formed by Ar = V T X.

Algorithm 1: Indirect Arnoldi algorithm for single-input systems

1) Initialization:
ṽ1 = A−1 b1 ; x̃1 = b1 ;
v1 = ∥ṽṽ11 ∥ ; x1 = ∥ṽx̃11∥ ;
2) Arnoldi iterative process:
for i = 2 : r do
a)ṽi = A−1 vi−1 ;
b) x̃i = vi−1 ;
c) Orthogonalization:
for j = 1 : i − 1 do
h = ṽi vj ; ṽi = ṽi − hvj ; x̃i = x̃i − hxj ;
end
d) Normalization:
vi = ∥ṽṽii ∥ ; xi = ∥ṽx̃ii∥ ;
end
3) Set V and X:
V = [ v1 ... vr ]; X = [ x1 ... xr ];
4) Construct Ar :
Ar = V T X;

27
3.2 Indirect Arnoldi Algorithm for Multiple-Input Systems

Several Krylov-subspace methods have been developed to create ROMs for systems with mul-
tiple inputs (m ≥ 2). Successful attempts include, e.g., the block Arnoldi method [37] and the
two-sided Arnoldi method for multiple-input and multiple-output systems [38]. The block Arnoldi
method is analogous to the Arnoldi algorithm, but operates with the block matrix A−1 B. It effi-
ciently utilizes the block Arnoldi recursion through the orthogonalization process and basis defla-
tion [37]. Compared with that, the two-sided Arnoldi method constructs the matrices V and W
from the input and output Krylov subspaces of any order r independently of the number of inputs
and outputs.
The common drawback of these two methods is that they require access to the system matrix
A. To overcome this problem, a new method presented by Algorithm 2 is developed in this work.
It uses the one-sided Arnoldi approach and is suitable for large-scale multiple-input systems, for
which the system matrix is unavailable.
 
The column vectors in B = b1 · · · bm are utilized as starting vectors. At the initializa-
tion stage, the algorithm computes the first m orthogonal basis vectors vi by recursive orthogonal-
ization of the steady-state solution ṽi = A−1 bi with respect to the previously generated orthonormal
basis vectors. Note that the starting vectors in B must be linearly independent to ensure the orthog-
onality. In the following Arnoldi iterative process (see Algorithm 2), every newly generated basis
vector vi is computed through orthogonalization of the steady-state solution ṽi = A−1 vi−m with re-
spect to the previously orthonormalized basis vectors followed by normalization. It is required that
the newly generated basis vector is linearly independent of the existing basis vectors in V . If this
condition is not satisfied, the deflation procedure [38], by which the newly computed vector and its
derivative recursive vectors are deleted, is utilized. Practically, the near linear dependency of the
generated basis is identified as the situation, in which the norm of the newly computed orthogonal
vector becomes smaller than ε = 10−3 .
Since the system matrix A is not explicitly available, the orthogonalization matrix X = AV ,
such that Ar ≡ V T AV = V T X, is computed by the algorithm. Every new column vector xi of X

28
is the results of orthonormalization of x̃i = Aṽi . Throughout the process of orthogonalization and
normalization, the relationship of the basis vectors, xi = Avi , is kept valid as:

 T
−1 i−1
A vi−m − V h1 h2 · · · hi−1
Avi = A
∥ṽi ∥
 T (3.1)
xi−m − X i−1 h1 h2 · · · hi−1
= = xi ,
∥ṽi ∥

where hi is the coefficient of the Gram-Schmidt orthogonalization and V i−1 , X i−1 denote the
already constructed matrices containing the first i − 1 vectors of V , X. The Algorithm 2 thus
computes Ar , equivalent to the matrix Ar found by the conventional one-sided Arnoldi algorithm.

29
 Algorithm 2: Indirect Arnoldi algorithm for multiple-input systems
1) Choose starting vectors  bi (linearly independent):
Bn×m = b1 ... bm
2) Initialization of starting vectors:
for i = 1 : m do
ṽi = A−1 bi ; x̃i = bi ;
if i ̸= 1 then
for j = 1 : i do
h = ṽi vj ; ṽi = ṽi − hvj ; x̃i = x̃i − hxj ;
end
end
vi = ∥ṽṽii ∥ ; xi = ∥ṽx̃ii∥ ;
end
3) Arnoldi iterative process with deflation:
for i = m + 1 : r do
a)ṽi = A−1 vi−m ;
b) x̃i = vi−m ;
c) Orthogonalization:
for j = 1 : i − 1 do
h = ṽi vj ; ṽi = ṽi − hvj ; x̃i = x̃i − hxj ;
end
d) Deflation:
if ∥ṽi ∥ < ε then
m = m − 1;
if m ̸= 0 then
Continue;
else
Break;
end
vi = ∥ṽṽii ∥ ; xi = ∥ṽx̃ii∥ ;
end
4) Set V and X:
V = [ v1 ... vr ]; X = [ x1 ... xr ];
5) Construct Ar :
Ar = V T X;

3.3 Frequency-Shift Arnoldi Algorithm

We introduce a new algorithm base on Krylov subspace method: Frequency shift Arnoldi Al-
gorithm. The algorithm establishes a more accurate approximation for Krylov subspaces by the
selection of non-zero frequencies, i.e., σ ∈ C, in describing the moment matching [26].

30
 By defining the Krylov subspace of shifted frequency, the ROM approximates characteristic,
under the selections of frequency points, from system’s response within a broad range of frequency,
i.e.,
 
σ1 σ2 · · · σn . (3.2)

We define the Krylov subspaces as:

  n (r−1)
o
κr (Ã − σi I), b̃ = span b̃, (Ã − σi I)b̃, ..., (Ã − σi I) b̃ , (3.3)

where σi represents the selected frequency shift points that are distributed within a range of fre-
quency. The Taylor series expansion of moment at shifted frequencies (3.2) involves shifted mo-
ments given by (3.3). In the expansion of the Taylor series, one can choose frequency shift points
σ according to system’s characteristics to improve the approximation accuracy. For example, we
define σ at high-frequency range that helps ROM captures the system’s high-frequency dynamics.
Compared with that, ROM misses part of the high-frequency characteristic if ROM is built based
on SIMO algorithm in the previous sections. Note that systems with high-frequency response can
be approximated more accurately by frequency-shift algorithm.
In algorithm 3, we define a total of n frequency points, σs. Per each σ, we construct the
orthogonalized vector set following the Arnoldi process that spans (3.3). Each of the vector set
has m+1 ortho-normalized vectors, which indicate up to mth moments are matched between ROM
and FOM. The vectors from n vectors sets are organized from i=1 to n per each moment to form
vector set V̂ . As a final step, we get the projection matrix V by applying modified Gram-Schmidt
orthogonalization to vector candidates set V̂ . Each vector in V̂ are explicitly orthogonalized against
all the previous basis vectors. Note that V is orthonormal, and its columns form a basis that spans
every Krylov subspace in a defined frequency shift.
The basis vectors xi are computed as xi = Avi because one-sided Arnoldi algorithm is used.
Once V and X are computed, the ROM system matrix is formed by Ar = V T X.

31
 Algorithm 3: Frequency-shift Arnoldi algorithm for single-input systems
n
 sampling points are  selected:
σ1 σ2 · · · σn
1) Initialization:
for i = 1 : n do
ṽi = (A − σi I)−1 b1 ;
v̂i = ∥ṽṽii ∥ ;
end
2) Generation of Krylov subspace at each frequency shift by Arnoldi iterative process:
for k = 1 : m do
for i = 1 : n do
a) Getting CFD solution:
ṽkn+i = (A − σi I)−1 v̂(k−1)×n+i ;
b) Orthogonalization:
for j = 1 : k do
h = ṽkn+i v̂(j−1)×n+i ;
ṽkn+i = ṽkn+i − hv̂(j−1)×n+i ;
end
c) Normalization:
v̂kn+i = ∥ṽṽkn+i
kn+i ∥
;
end
end
3) Set V and X:
a) Apply
 Gram-Schmidt Orthogonalization
 to vectors set:
V̂ = v̂1 v̂2 · · · v̂(m+1)n ;
b) Get matrices:
V = [ v1 ... v(m+1)n ];
4) Construct Ar :
Ar = V T AV ;

3.4 Parametric Reduced Order Modeling Algorithm

The construction of the parametric reduced-order model is executed following the algorithm 4.
Note that ϖi are the weighting coefficients for interpolation as described in Section 2.2.2.

32
 Algorithm 4: Parametric Model Order Reduction algorithm
1) Choice of R:
Apply Singular Value Decomposition (SVD) to the local projection matrix collection Vall .
 
Vall = V1 V2 ... Vk ;
Choosing R as the first r columns of U from the SVD of Vall .
Vall = U ΣV T
(where U is a k × k real orthonormal matrix, Σ is a k × n rectangular diagonal matrix
with non-negative real numbers on the diagonal, and V is an real unitary matrix.)
2) Calculation of matrices M and T for local ROMs:
for i = 1 : n do
a)Ti = RT Vi ;

−1
b)Mi = WiT R ;
end
3) Construction of pMOR based on weighing coefficients ϖi :
Pk −1 ∗
Pk T −1 ∗
Pk T
i=1 ϖi Mi Ti ẋr = i=1 ϖi Mi Wi Ai Vi Ti xr + i=1 ϖi Mi Wi Bi u;

4) Reconstruction of pMOR :
∗
= ki=1 ϖi Vi Ti−1 ;
P
Cpmor
yr∗ = Cpmor
∗
x∗pmor

3.5 Change of Variable Method

The change of variable method can be applied when the input vector u(t) in system (2.4) has
time-invariant components. For example, if the system with m = 2
 
 u1 
 
ẋ(t) = Ax(t) + b1 b2  
u2

33
has a time-varying signal u1 and a constant signal u2 , it can be rewritten as

ẋ(t) = Ax(t) + Bu(t) + γ, (3.4)

where γ = b2 u2 = const and Bu(t) = b1 u1 (t). The solution of (3.4) can be obtained by the
change of variable:
x(t) = x∗ (t) − A−1 γ (3.5)

where x∗ (t) is the solution of the system (3.4) with γ = 0 and −A−1 γ is the steady-state response
of the system to the constant input γ. Substituting (3.5) into (3.4) results in a single-input system
(2.4) for x∗ .
The modification can be used to exclude the constant input from the process and, thereby, reduce
the number of inputs m, thus reducing the computational cost of the ROM creation. The steady-
state response −A−1 γ can be found a posteriori as an FOM solution of the steady-state problem
with the single constant input γ.
Practically, the change of variable method is useful for ROM approximation of heat transfer with
Dirichlet boundary conditions on temperature, for example, for BTMS with prescribed temperature
at the inlet of the cooling channels (see section 4.3). In that case, x∗ (t) is obtained as the ROM
approximation of the solution of (2.4) with zero inlet temperature, while −A−1 γ is the full steady-
state solution obtained with true temperature boundary values and zero internal heating rate, which
needs to be computed only once.

34
 CHAPTER 4

Reduced-Order Modeling for Conjugate Heat

Transfer in a Battery Pack

The work described in this section is presented in detail in [61]. The proposed methods are
applied to a large-scale CFD model of conjugate heat transfer in a battery pack. After an ROM
is derived using Algorithm 1, Algorithm 2, Algorithm 4 and the change of variable method,
the accuracy of its prediction is tested in a series of scenarios of unsteady behavior. In each test,
the model system of equations (2.13) is solved with appropriate input and initial conditions. The
solution xr (t) is converted into a full-order approximation of temperature field using (2.11). The
outcome is compared with the the temperature field obtained in unsteady FOM CFD solution.
A comment is in order concerning the rather high heating rates and, respectively, internal tem-
perature variations observed in the tests. The typical heat generation in an operating automotive
battery cell of the type considered below varies in the range between 1 and 3 W. Stronger heat-
ing, sometimes exceeding 10 W is found in exceptional situations during extreme conditions (e.g.,
charging and discharging of 4 C and higher currents). Such exceptional situations are intention-
ally used for several test cases in our work because they are particularly challenging for accurate
modeling and, therefore, particularly suitable for validation of the proposed ROM method.
The physical model and its CFD discretization are described in section 4.1. The accuracy of
ROM and measures of error are discussed in section 4.2. In section 4.3, the results for single-
input systems are presented using two different input signals. The performance of the ROMs of
multiple-input systems is discussed in section 4.4.

35
4.1 Battery Pack Model

A model battery pack is used to test the method. While retaining the typical geometry and
physical properties of an automotive battery pack, the model does not correspond to any specific
manufactured product. Details of battery design not directly relevant to heat transfer and BTMS
are ignored. In particular, the processes of Joule dissipation, phase change, and mixing are not
explicitly considered. Their cumulative effect on the internal energy balance within the system
(see, e.g., [56]) is simulated as volumetric heat generation treated as an input in our analysis.
As illustrated in Figure 5.5, the pack includes nine battery cells enclosed in aluminum casing
and attached to dual coolant channels. In Figure 5.5, the leftmost casing is displayed as partially
transparent for illustration purposes. The cells are cooled by heat transfer to the cooling liquid
flowing in the channels.

Figure 4.1: Schematic of the battery pack with dual coolant channels.

The parameters of the model are summarized in Table 4.1. Each battery can be viewed as
representing an NMC-graphite cell with an approximate capacity of 10 Ah. The physical properties

36
of the pack materials listed in Table 4.1 correspond to a typical design [62]. Aluminum casing and
water as a coolant are assumed. The interiors of the batteries are assumed to be uniform, with
constant density and specific heat and constant anisotropic thermal conductivity.

Table 4.1: Model specifications.

CFD Model mesh size up to 880,862, 1,071,863 or 1,457,717

Laminar
CFD Model Specifications Segregated Flow
Segregated Fluid Enthalpy

ρ = 2702.0 kg/m3
Solid: Aluminum Cp = 903.0 J/kg-K
κ = 237.0 W/m-K
ρ = 2560.0 kg/m3
Model Components Properties Solid: Cell Bulk Cp = 975 J/kg-K
Anisotropic κ (W/m-K):
0.95 × 30.8 (through/in-plane)
ρ = 997.561 kg/m3
Liquid: H2 O Cp = 4181.72 J/kg-K
κ = 0.620271 W/m-K

Temperature: 15 ◦ C
Inlet
Mass Flow Rate: 0.01 kg/s
Boundary Conditions
Outlet Pressure Outlet: p = 0 Pa
Pack Exterior Wall

The boundary condition on the solid part is zero heat flux (an adiabatic wall). At the inlet of the
cooling channels, uniform velocity corresponding to the mass flow rate of 0.01 kg/s is imposed.
This induces a laminar flow in channels. A constant value of temperature is imposed at the inlet of
the cooling channels. The inlet temperature value of 15 ◦ C is used in the examples below.

37
 The commercial solver STAR-CCM+ [36] is used for the full-order CFD solution. The heat
transfer equations (2.1) and (2.2) are spatially discretized in the entire domain on an unstructured
finite-volume grid adapted to the battery geometry (for example, thin finite-volume cells are used
to resolve the walls of the aluminum casing). Three grids differing from each other by the base
cell size have been tested: with 880,862, 1,071,863, and 1,457,717 cells. A grid sensitivity study
conducted in the form of transient simulations of the response of the pack to constant-rate heating
of all batteries have shown no significant difference between the results obtained with the three
grids. In particular, the grid-related changes of the average temperatures of individual batteries and
the temperatures recorded at four probe points have not exceeded 0.5% of the maximum variation
of temperature within the pack during the entire process. Based on these data and with the goal
of better demonstration of the capabilities of the order reduction method, the FOM on the largest
grid of 1,457,717 cells is used for ROM development and ROM-FOM comparison in the analysis
presented in the rest of this paper.
In order to impose the Dirichlet boundary condition corresponding to the constant cooling tem-
perature at the inlet, the change of variable method described in section 3.5 is used. The vector
−A−1 γ is calculated as the solution at zero internal heat sources and inlet temperature of 15 ◦ C.
Due to the boundary conditions of zero normal temperature gradients applied at all the other bound-
aries of the model, no calculations are, in fact, necessary in our case. The solution is simply the
uniform temperature field T = 15◦ C.
The FOM simulations are performed on a workstation with 20 cores installed on two Intel
Xeon Silver 4114 processors. The ROM solutions are carried out on a personal computer using
two cores.

4.2 Error of ROM Approximation

As discussed in section 3.1, the ROM approximation is derived to match the first r/m moments
of the FOM transfer function. The higher the reduced order r, the more moments are matched,

38
yielding more accurate approximation. The downside of using higher r is the higher computa-
tional costs of development and application of the ROM. The physical nature of the conjugate
heat transfer suggests that the dynamic characteristics of the FOM are dominated by the first few
moments, so large values of r may, in fact, be unnecessary. One anticipates existence of an opti-
mal order r, such that its further increase does not significantly improve the accuracy of the ROM
approximation. The tests presented below fully confirm this expectation.
An important question arising in the tests is that of an appropriate measure of the accuracy of
an ROM approximation. The error of the approximation is defined as the vector

ε = x(t) − V xr (t) (4.1)

that shows the difference between the FOM solution x(t) and the approximate solution V xr (t)
reconstructed from the ROM at every mesh point. The accuracy of the approximation is evaluated
in our work by the root-mean-square and infinity norms of the vector:

 1/2
1 T
∥ε∥2 = ε ε , (4.2)
n
∥ε∥∞ = max |ε| . (4.3)

4.3 Single-Input Cases

A single-input problem appears in the situation when all 9 battery cells generate heat according
to the same schedule. To explore the ROM performance, models of different reduced order r are
developed and applied to two characteristic scenarios: with a step input signal, and with a realistic
rapidly changing input. The initial temperature distribution is uniform at T = 15◦ C.

39
Figure 4.2: Heat input signal for the test case 1 – single input in the form of a stepwise heating
rate.

In the first case, each battery cell produces 10 W at 0 ≤ t ≤ 400s and 20 W after that (see
Figure 4.2). As the first illustration, we show how the order r of the ROM affects the simulated
evolution of temperature at one probed point inside the domain. The point’s location is shown in
Figure 4.3. The results obtained for r = 1, 3, 10, and 15 are presented in Figure 4.4.

Figure 4.3: Location of the point within the battery pack, at which temperature is probed. (a) Front
view; (b) Side view.

40
 We see that the ROM of order one is clearly inaccurate. Much better accuracy is obtained by
the models with r ≥ 3. Some small-amplitude, but still unrealistic fluctuations of temperature
are shown by the ROMs with r = 3 and r = 5 soon after the stepwise change of the input,
approximately at t = 2s and 402s. Very accurate results (no discernable difference with the FOM
curve) are produced by the ROM with r = 15. Evidently, the data at a single probed point cannot
comprehensively reflect the approximation accuracy. Thus, the norms (4.2) and (4.3) of the error
are computed and shown in Table 4.2.

34
FOM
32 ROM 15
ROM 10
30
ROM 3

28 ROM 1
Temperature/C

26

24

22

20

18

16

14
0 100 200 300 400 500 600 700 800
Time/s

Figure 4.4: Comparison between ROM and FOM solutions for the test case 1 (see Figure 4.2).
Transient temperature signals at the probed point obtained in the FOM and the ROMs with different
values of r are shown.

41
 Table 4.2: The norms (in ◦ C) of the error at various time instants for
the test case 1 (see Figures 4.2, 4.4 and 4.5).

402s 410s 800s

r
∥ε∥2 ∥ε∥∞ ∥ε∥2 ∥ε∥∞ ∥ε∥2 ∥ε∥∞

1 0.28265 0.42378 0.29965 0.45800 0.46674 0.66457

2 0.00534 0.01702 0.01679 0.03740 0.03301 0.05755

3 0.00571 0.01564 0.00563 0.01929 0.01045 0.03697

5 0.00382 0.01704 0.00336 0.01823 0.00754 0.03727

8 0.00369 0.01712 0.00400 0.01766 0.00758 0.03726

10 0.00379 0.01684 0.00425 0.01761 0.00757 0.03726

12 0.00391 0.01696 0.00423 0.01756 0.00757 0.03726

14 0.00394 0.01691 0.00421 0.01758 0.00757 0.03726

15 0.00395 0.01691 0.00421 0.01759 0.00757 0.03726

We see that the approximation error decreases with r and plateaus at a low level at fairly small
r, so that further growth of r does not improve the accuracy. As an example, at t = 800s, the error
remains nearly constant with ∥ε∥2 ≈ 0.02◦ C and ∥ε∥∞ ≈ 0.2◦ C at r ≥ 8. At t = 402s and 410s,
the saturation occurs at r ≈ 12. We conclude that the ROM with r = 15 is certainly sufficient for
a reliably good approximation of the FOM in this case.
The temperature distributions in the cross-section of the battery pack at t = 3, 402 and 800s
are presented in Figure 4.5 for the FOM and the ROM of order 15. Despite the fact that small
difference can be observed between the ROM and FOM contours, for example, at the top of the
aluminum casing of the leftmost cell and at the coolant channel near the outlet (see Figure 4.5a and
4.5b), the distributions of temperature computed by the ROM are almost identical to those by the
FOM. This is also true in the distributions at t = 3s, i.e. right after the stepwise heating increase
and at the later stages of the evolution characterized by strong influence of the cooling system.

42
Figure 4.5: Temperature observed in the middle cross-section of battery pack for the test case 1.
The heating input signal is shown in Figure 4.2. The results of ROM with r = 15 and FOM at
different instants are presented. (a) FOM t=3s; (b) ROM t=3s. (c) FOM t=402s; (d) ROM t=402s.
(e) FOM t=800s; (f) ROM t=800s.

An even greater confidence in reliability and accuracy of the ROM model is provided by the
results obtained for a dynamic heat input. The transient heat input signal (see the dashed line in
Figure 4.6) reflects the typical behavior of a battery pack in a driven electric vehicle.
The probed point signals of temperature presented in Figure 4.6 show excellent agreement

43
between the FOM and the ROM with r = 15.
34 70
ROM
FOM
32
InputSignal
60

30

50
28

26

Heating Rate/W
Temperature/C

40

24

30
22

20
20

18

10
16

14 0
0 200 400 600 800 1000 1200 1400 1600
Time/s

Figure 4.6: Results for the test case 2 with a realistic transient single input. The heat input signal
(yellow line, right-hand side axis) and the temperature values at the probed point obtained in the
FOM and the ROM with r = 15 are shown.

For further verification, the norms of the error vector calculated at different time instances are
shown in Table 4.3. We see that both the norms remain small. The largest error ∥ε∥∞ = 0.14175◦ C
is observed at t = 250s immediately after a sharp peak of the heating input.

Table 4.3: The norms (in ◦ C) of the error at r = 15 for the test case 2.

Time/s ∥ε∥2 /◦ C ∥ε∥∞ /◦ C

250 0.01236 0.14175

800 0.01295 0.06311

950 0.03703 0.06773

We also compared ROM and FOM in the frequency domain using the Bode plot. The system’s
Bode diagrams for the signal at the probed point obtained by the FOM and ROMs with different
values of r are shown in Figure 4.7. We see that the ROM can approximate FOM accurately in
the low-frequency interval. When the frequency is increased, the accuracy of the approximation is

44
slightly reduced but remain acceptable. We also see that the accuracy of the ROM approximation
increases with growing r and becomes certainly sufficient at r = 15.

Figure 4.7: The Bode diagrams of the FOM and the ROM at different values of r for a single-input
system. (a) Magnitude plot; (b) Phase plot.

The ROM dramatically saves the computation effort of modeling. The only significant compu-
tational expense is for the construction of ROM. Once this task is completed, the ROM can be used
for multiple simulations of the system’s evolution over long periods of time. As an example, data
for the computational costs in our work are presented in Table 4.4. We see that the FOM transient
simulation requires 18 hours of parallel computation on a workstation. The ROM with r = 15
takes about 40 hours to develop, while the simulation of the same 1500s of evolution in the test
case 1 requires only 20s.

Table 4.4: The computation cost of the ROM creation and transient simulations in the single input
case compared with the cost of the FOM simulations.

ROM FOM

Transient simulation (1500 s) 20 s/2 cores 64509s(18 hours)/20 cores

ROM creation (r = 15) 143043 s (39.7 hours)/20 cores

45
4.4 Multiple-Input Cases

In this chapter, we present the result of the implementation of the multiple-input algorithm to
the model of a battery pack. Two multiple-input systems are considered: with two and nine inde-
pendently controlled heating rates in individual battery cells. The first case is used for illustration
purposes. The second case demonstrates the potential of the approach for practical applications.
ROMs of the orders up to 20, with each vector in the input matrix B ∈ Rn×2 generating up to
10 basis vectors of the Krylov subspace, was developed for the first case. No linear dependency
that would trigger the deflation mechanism was detected during the model generation.
The heating scenario used to test the ROM is illustrated in Figure 4.8. In the input matrix
B ∈ Rn×2 , the input signal b1 controls the signal of one battery cell at the center of the pack. The
input signal b2 controls the heating of the remaining eight cells. In the scenario, all the nine battery
cells generate 10 W per cell uniformly for the first 500 seconds. The signals bifurcate at t = 500s
with input signal b1 jumping to 25W per cell and input signal b2 dropping to 5W per cell for the
remaining time.

Figure 4.8: Heating rates for the test case 3 – multiple inputs with two stepwise heating rates.

The temperature distributions obtained by the FOM and the ROM with r = 20 for different
time moments are nearly identical (see Figure 4.9).

46
Figure 4.9: Comparison between the ROM results with r = 16 and the FOM results for the test
case 3 (see Figure 4.8 for input signals). Temperature distributions in the middle cross-section of
battery pack are shown. (a) FOM t=3s; (b) ROM t=3s. (c) FOM t=510s; (d) ROM t=510s. (e)
FOM t=1000s; (f) ROM t=1000s.

The temperature signals at the probed point (see Figure 4.3) are plotted in Figure 4.10. They
clearly show that the accuracy of the approximation increases with r. The ROM of the cumulative
order 16 or higher accurately reproduces the dynamics of the FOM. Significant deviations between
the ROM and FOM are found at lower reduced orders, such as r = 1, 2, 3 and 5.

47
 24
FOM
ROM20
23 ROM16
ROM5
22 ROM3
ROM2
ROM1
21
Temperature/C

20

19

18

17

16

15
0 100 200 300 400 500 600 700 800 900 1000
Time/s

Figure 4.10: Comparison of ROM and FOM solutions for the test case 3 (see Figures 4.8 and
4.9.) Transient temperature signals at the probed point obtained in the FOM and the ROMs with
different values of r are shown.

The norms of the error, ∥ε∥2 and ∥ε∥∞ , are presented in Table 4.5. The ROMs of order 10 and
higher appear to be sufficient for accurate representation of the FOM. However, in order to reliably
reproduce the fast dynamics of the original system possibly not reflected by our test, ROM of a
higher order, e.g., with r = 16 can be considered for practical use.

Table 4.5: The norms of the error vector (in ◦ C) for the test case 3 (see
Figures 4.8–4.10).

502s 510s 1000s

r
∥ε∥2 ∥ε∥∞ ∥ε∥2 ∥ε∥∞ ∥ε∥2 ∥ε∥∞
1 6.71094 10.04838 6.62344 9.91439 4.40738 6.71387
2 1.35025 1.92200 1.32306 1.95377 0.35767 1.21754
4 0.08832 0.19034 0.12001 0.30500 0.11812 0.37082
6 0.01700 0.09508 0.03376 0.16518 0.01229 0.07855
8 0.01127 0.09550 0.02395 0.11719 0.01000 0.07782
10 0.00683 0.09302 0.00618 0.09133 0.00985 0.07769
12 0.00684 0.09342 0.00576 0.09263 0.00989 0.07771
15 0.00628 0.09384 0.00522 0.09291 0.00989 0.07771
16 0.00544 0.09382 0.00535 0.09305 0.00989 0.07770
20 0.00525 0.09371 0.00531 0.09309 0.00989 0.07770

48
 The computational cost data for the test case 3 are shown in Table 4.6. Generation of the ROM
with r = 20 takes 42.1 hours on 20 CPU cores. Once the ROM is generated, it takes only 26
seconds to compute a transient simulation of 1000s of evolution on a personal computer the task
which would require 20.6 hours on a workstation using the FOM. The computational time for
simulation is reduced by a factor of 103 .

Table 4.6: The computation cost of the ROM creation and transient simulations compared with the
cost of the FOM simulation of transient process in the test case 3 lasting 1000 seconds.

ROM FOM

Transient simulation (1000s) 26s/2 cores 74244s(20.6 hours)/20 cores

ROM creation (r = 20) 151709s (42.1 hours)/20 cores

In the second multiple-input case, we constructed an ROM for the realistic situation, in which
all nine battery cells were controlled independently so B ∈ Rn×9 . We built ROMs of the orders
r ≤ 50. The deflation was not triggered during the execution of multiple-input algorithm.
The model was verified using the input signals shown in Figure 4.11. The heating rate of all the
battery cells was the same 10 W, during the first 400 seconds. After that, the square wave signals
changing every 50s were applied to the battery cells as indicated in Figure 4.11.

Figure 4.11: Heating rate input of the battery cells for the test case 4 – multiple inputs. The
numbers indicate the heating rates applied to individual cells.

49
 We first compare the FOM and ROM temperature distributions in the cross-section of the battery
pack. We see that the ROM accurately calculates the highly variable temperature field at various
stages of the evolution. As an example, the difference between the highest temperature values
predicted by the two models does not exceed 0.007◦ C at t = 400s, 0.015◦ C at t = 505s, and
0.006◦ C at t = 600s.

50
Figure 4.12: Comparison between the ROM results at r = 50 and the FOM results for the test case
4 (see Figure 4.11 for input signals). Temperature distributions in the middle cross-section of the
battery pack are shown. (a) FOM t=400s; (b) ROM t=400s. (c) FOM t=505s; (d) ROM t=505s. (e)
FOM t=600s; (f) ROM t=600s.

Further comparison is made using the temperature signal at the probed point (see Figure 4.13).
The FOM data and the ROM data obtained with r in the range between 10 and 99 are shown in
Figure 4.13. We see that the ROM with r = 10 fails to reproduce the dynamic of the system.

51
Though the ROM with r = 20 greatly improves the approximation accuracy, a deviation from the
ROM signal can still be observed during the various stages of the transient process, e.g., at 0-100
s and 400-600 s. The ROMs with r = 50, 80, and 99 indistinguishably follow the FOM’s transient
dynamics at the probed point.

25
FOM
24 ROM72
ROM60
23
ROM
50
ROM
22 34
ROM10
Temperature/C

21

20

19

18

17

16

15
0 100 200 300 400 500 600
Time/s

Figure 4.13: Comparison between the ROM and FOM results for the test case 4 (see Figures 4.11
and 4.12). Transient temperature signals at the probed point obtained in the FOM and the ROMs
with different values of r are shown.

The norms of the error vectors obtained in the test case 4 are shown in Table 4.7. The data
confirm poor accuracy of the ROM at r = 10. The accuracy, as it is reflected by the norms of the
error, is good at r = 34. Further increase of r does not lead to decrease of the norms. We conclude
that r = 34 is sufficient in this test case, although a higher value, e.g., r = 50 may be necessary as
a safety measure to accurately approximate the features of the stack behavior not reflected in the
test.
The construction of the ROM with r = 50 takes around 150 hours of the 20-CPU-cores par-
allel computation. The computational time required to complete the ROM simulation of transient
behavior, such as that in the test case 4, remain exceptionally small, as in the previous cases con-
sidered in this paper.

52
 Table 4.7: The norms (in ◦ C) of the error vector for the test case 4
(see Figures 4.11, 4.12 and 4.13).

455s 505s 600s

r
∥ε∥2 ∥ε∥∞ ∥ε∥2 ∥ε∥∞ ∥ε∥2 ∥ε∥∞
10 0.93124 1.79201 1.07766 1.98469 0.99904 2.25001
34 0.05344 0.12464 0.06379 0.12266 0.06040 0.14394
50 0.05382 0.12344 0.06181 0.11407 0.06019 0.14342
60 0.05384 0.12458 0.06177 0.11302 0.06019 0.14376
72 0.05444 0.11962 0.06171 0.11485 0.06017 0.14371

4.5 Boundary Condition as Input

In this section, we present the result of the implementation of the boundary condition as an
input. The Dirichlet boundary condition of the ODE system, i.e., the inlet temperature of the
coolant channel in the battery pack, is considered. It is controlled to be one of the time-varying
variables of u(t) in (2.4).
The heat source vector B = γ is generated such that the internal heating represents the Dirichlet
boundary condition within one extruded cell at the coolant inlet (as shown in Figure 4.14). The
implementation of the method does not follow the formula (2.9) directly. The reason for that is the
possibility of deviation of the internal STAR-CCM+ solver from the FV formulas (2.7) and (2.8).
The implementation is based on the linearity of the problem and follows the following approach.
We solve the steady-state heat transfer problem with TBC = 0 and some volumetric heating density
F in the extruded cells, e.g., 109 W/m3 . The solution vector is a uniform field with a temperature
equal to d Celsius degrees. By linearity, the volumetric heating of extruded cells with density F/d
is, then, identical in its effect on the system to the inlet temperature TBC of 1 Celsius degree. Vector
B, which has elements equal to F/d in the extruded cells and zero elsewhere, can be considered
as an input vector in the state-space representations used to develop ROM, i.e., from (2.4).

53
 Figure 4.14: Extruded layer method for BC as input. The extruded layer is marked in pink.

We present a ROM with the boundary condition implemented as an input in a single-input

system. Note that the boundary condition as input can always be implemented for the multiple-
input systems for practical demand.
To explore the performance of the boundary condition as an input method, we present an ROM
of order 20 using single-input algorithm 1 for the validation. It is validated using a transient inlet
temperature as the boundary condition. The inlet temperature starts at temperature of 25 ◦ C for
100 seconds and step-changes to 5 ◦ C until 200 seconds (as shown in Figure 4.15). The initial
temperature distribution is uniform at T = 15◦ C.

Figure 4.15: Inlet temperature signal for boundary condition as input validation.

54
 The temperature signals at the probed point (see Figure 4.3) are plotted in Figure 4.16. They
clearly show that the accuracy of the approximation using the boundary condition as input. The
ROM of order 20 reproduces the dynamics of the FOM. Derivation takes place after the step-
change of the signal, i.e., t = 102s. Other than that, the approximation achieved by the boundary
condition as the input indicates a great accuracy. To be more clear about the approximation of the
method, the norm of the error, ∥ε∥2 and ∥ε∥∞ , are presented shortly in Table 4.8.

19

ROM
18 FOM

17
Temperature/C

16

15

14

13
0 20 40 60 80 100 120 140 160 180 200
Time/s

Figure 4.16: Comparison of ROM and FOM solutions for the boundary condition as input. Tran-
sient temperature signals at the probed point obtained in the FOM and the ROM of order 20 are
shown.

Compared with the results shown in previous sections for single-input or multiple-input cases,
the norms of the error is rather large with respective to the approximation to the system’s dynamics.
We expect the inaccuracy come from accumulated error by imposing the volumetric heat source to
the model. The artificially defined ‘zero’ boundary temperature yield very small error near the inlet.
This error can potentially accumulated as during the ROM iterations. However, the approximation
level is still a relatively good approximation to the FOM with norm of error equals to 0.373319◦ C
while the maximum error among the whole temperature does not exceed to 0.846873◦ C. As the
transient process goes further, the norms of the error vector restores to excellent similarity to the
FOM. We expect the boundary condition as the input method to have overall excellent performance
in applying the reduced-order modeling.

55
Table 4.8: The norms of the error vector (in ◦ C) for the boundary condition as the input (see Figures
4.16).

102s 110s 200s

r
∥ε∥2 ∥ε∥∞ ∥ε∥2 ∥ε∥∞ ∥ε∥2 ∥ε∥∞

20 0.373319 0.846873 0.060261 0.202178 0.007995 0.089852

4.6 Parametric Reduced-Order Modeling

In this section, we present the implementation of the parametric reduced-order modeling. As

illustrated in section 2.2.2, given several local reduced-order models, a parametric reduced-order
model can be calculated by interpolating the system matrices. The new ROM obtained in such
way can be used to predict a dynamic behavior of the system within a certain range of parameters,
e.g., mass flow rates for conjugate heat transfer or flow rate in an inlet of a ventilation system
for airborne transmission. We illustrate the results of the parametric ROM for the conjugate heat
transfer of a battery pack.
We firstly generate three local reduced-order models at different mass flow rates
(0.01, 0.03, 0.05kg/s). With the pre-computed local ROMs, the transformation matrix R is com-
puted by singular value decomposition. Then the local ROMs are transferred into a common re-
duced subspace for the interpolation. The parametric ROM is created according to the Algorithm
4.
The parametric ROM valid for the mass flow rate p in the range from 0.01 to 0.05kg/s is con-
structed by the means of quadratic interpolation. The quadratic interpolation weighted factors are
solved based on ki=1 ϖi (p) = 1 and three local quadratic points, i.e., ϖ1 (0.01) = 1, ϖ2 (0.03) = 1
P

and ϖ3 (0.05) = 1. The weighted coefficients ϖi are:

56
     
2
 ϖ1 (p)   1250 −100 15/8   p 
    
 ϖ (p)  =  −2500 150 −5/4   p  (4.4)
 2    
    
ϖ3 (p) 1250 −50 3/8 1

where p is the mass flow rate for interpolation.

The model is tested for two mass flow rates p = 0.02, 0.04kg/s. In the test case, each battery
cell produces 10W constantly for 400 seconds. The local ROMs of order 15 found in section 4.3
to provide accurate results for the single-input cases are used.
The results obtained at p = 0.02kg/s are shown in Figure 4.17 and 4.18. Similar results are
observed at the mass flow rate of 0.04kg/s. We see that the parametric ROM accurately predicts
the temperature field at various stage of the evolution.

Figure 4.17: Comparison between the parametric ROM results at MFR = 0.02kg/s and the FOM
results. Temperature distributions in the middle cross-section of battery pack are shown. (a) FOM
t=5s; (b) parametric ROM t=5s. (c) FOM t=400s; (d) parametric ROM t=400s.

57
Figure 4.18: Comparison of parametric ROM and FOM solutions at mass flow rate = 0.02kg/s.
Transient temperature signals at the probed point obtained in the FOM and the ROM of order 15
are shown.

To further examine the accuracy of the parametric ROM, the norms of the error are calculated
(see Table 4.9). At t = 5s, the maximum difference between parametric ROM and the FOM ∥ε∥∞
is 0.0146◦ C while the rms difference ∥ε∥2 is 0.002◦ C. The norms of the error between parametric
ROM and FOM indicate the excellent approximation has been achieved using the parametric ROM.

Table 4.9: The norms of the error vector (in ◦ C) for parametric ROM at mass flow rate = 0.02kg/s
(see Figures 4.17 and 4.18).

5s 400s
r
∥ε∥2 ∥ε∥∞ ∥ε∥2 ∥ε∥∞

15 0.002 0.0146 0.0314 0.1233

4.7 Systems With Time-Varying Parameters

The implementation of parametric reduced-order modeling for systems with time-varying pa-
rameters, i.e., time-varying mass flow rates at the inlet, is presented in this section. The appli-
cability of the reduced-order modeling is extended into broader setting. We drop the assumption

58
of a steady-state flow velocity field. The system with time-varying parameters is described using
decoupled momentum equation and heat equation. The energy equation is now a linear equation
with time-varying three-dimensional flow velocity field U(x, t) and constant flow properties. The
system is solved in two steps: (i) solve the momentum equation to obtain the flow velocity field U;
(ii) substitute U into the heat equation and solve it.
Instead of the previously considered ROM with constant system matrix A, the ROM with time-
varying parameters has a time-dependent system matrix A and input matrix B as shown in (2.19).
It contains the discretization coefficients matrix of the ODEs system which express the diffusive
and adjective terms at given time-varying parameter. Until then, the time-varying system matrix
A(t) is determined using interpolation.
We simulate the battery pack with time-varying MFRs and constant heat input by parametric
reduced-order modeling. The parametric ROM is essentially an ODE system. It is found by
interpolating the system matrices of the local reduced-order models obtained at specific parameters.
The approach is tested at the mass flow rate p(t):

p(t) = 0.03 + 0.01 ∗ (sin(2 ∗ pi ∗ t/200) + sin(2 ∗ pi ∗ t/100))(kg/s). (4.5)

The ODEs are interpolated at every time step based on the instantaneous mass flow rate p(t) (see
Figure 4.19). The weighted coefficients for the quadratic interpolation are computed as in (4.4).
Then, the system matrix A(t) and input matrix B(t) are interpolated as a pROM of parameter
p using (2.16). In this way, the FOM that computes partial differential equations (PDE) with
decoupled momentum and heat equations are simplified to an ODE system.
The results are compared with FOM for validation. In the pROM, we use three local ROMs
with r = 15 at MFR= 0.01, 0.03 and 0.05kg/s. The temperature at the inlet and initial condition
at the domain are 15o C. We apply a constant heat input of 10 W for all battery cells. While in the
FOM, we choose implicit unsteady solver with segregated flow. At each time step, the momentum
and energy equations are solved with 400 inner iterations. The residual of energy reaches below
10−7 ; the residual of momentum and continuity reaches below 10−14 .

59
Figure 4.19: Comparison of parametric ROM and FOM solutions under transient mass flow rates.
Temperature signals at the probed point obtained in the FOM and the parametric ROM are shown.

The results show a clear deviation between FOM and ROM when the change of the MFR is
rapid. The ∥ε∥2 and ∥ε∥∞ norms evaluated at different time instants are shown in Table 4.10. It
is observed that the parametric ROM fails to capture the dynamics of the original FOM when the
MFR undergoes rapid changes at t=200 and t=400s.

Table 4.10: The norms of the error vector (in ◦ C) for parametric ROM at time-varying mass flow
rates (see Figures 4.19).

Time/s ∥ε∥2 ∥ε∥∞

200 0.038173 0.177921

250 0.022414 0.096755

300 0.018023 0.046944

350 0.029274 0.115097

400 0.052249 0.244504

60
 The results show us a clear limitation of the current parametric ROM technique. The limitation
comes from the description of system matrix A(t) in (2.19). The parametric ROM uses interpola-
tion between approximations obtained for the system matrices based on steady-state velocity field.
However, the velocity field, by nature, is unsteady at transient MFR signal. The parametric ROM
fails to represent the resulting dynamics of heat transfer.
FOM simulation performed with small time step ∆t = 10−4 s clearly illustrate the difference
between the unsteady and steady-state velocity fields in the cooling channel. Specifically, the
‘reverse flow’ near the wall is clearly observed once the MFR is changed (see Figure 4.20).

Figure 4.20: Reverse flow developing in the cooling channel with time-dependent mass flow rate

The interpolated parametric ROM cannot take this effect into account. The use of a fully de-
veloped flow field by the interpolation leads to ‘over-reaction’ of the heat transfer behavior of the
system. Therefore, it is not surprising that the time-varying MFRs by the current parametric ROM
technique based on interpolation does not show the results as accurate as in the previous sections.

4.8 Conclusion

A method for development of reduced order models (ROMs) of conjugate heat transfer is pre-
sented. The method is based on the Krylov subspace approach and the Arnoldi algorithm, and can

61
be applied to single input and multiple input systems. The method utilizes a full order (FOM) CFD
solution of a steady-state problem, but does not require access to the discretization matrix. This
makes the method especially practical for industrial applications, where commercial CFD software
is used and the discretization matrix is usually inaccessible.
As a demonstration and validation study, the method is applied to conjugate heat transfer in
a simplified model of a typical automotive electric battery pack. The ROMs of various orders
are derived and applied to reproduce transient behaviors in four distinct scenarios characterized
by strong temporal variations and spatial gradients of temperature. The results of the FOM CFD
simulations of the same scenarios based on the computational grid of more than 106 finite-volume
cells are used for validation. The analysis shows that ROMs of small order (r between 5 and
50 depending on the number of independent inputs and the type of transient behavior) accurately
reproduce the evolution of the entire temperature field. It is noted that the physical parameters in
the FOM will still need to be validated by experiments to guarantee the accuracy of the ROM when
compared with the actual system.
Significant, but feasible amount of computations is required to derive an ROM for a given
heat transfer system. Once the ROM is developed, the computational costs of its application to
prediction of transient behaviors is exceptionally low, three or more orders of magnitude lower than
the time required for achieving the same goal using a full CFD solution. This opens the opportunity
of using the ROMs for real-time on-board control of battery thermal management systems and for
the design optimization studies requiring multiple simulations. Another possible extension of the
developed ROM for heat transfer is to couple it with an electrothermal model of battery cells for
real-time simulation of both electric and thermal behaviors of a battery pack including its thermal
management system. In such a coupled simulation, the electrothermal model of battery cells will
calculate the heat generation based on current input and the predicted cell temperature from the
ROM. The ROM, on the other hand, will take the predicted heat generation from the electrothermal
cell model as input.
An attempt to extend the method to the case of time-varying system parameters using matrix

62
interpolation has been also made. The reduced-order modeling approach based on the framework
of pROM is proposed for systems with time-varying parameters. The original decoupled PDE
system is approximated by an ODE system using matrix interpolation. The resulting ROM is
shown to capture the dynamics of the system. However, the quality of the results is, admittedly,
lower than in the case of time-independent parameters. This is attributed to the effect of not fully
developed fluid velocity under time-varying mass flow rates. To the best of our knowledge, the
choice of weighted function for pROM or adapting the reduced-order modeling technique that can
potentially improve the quality of the results. This is left for future investigations.

63
 CHAPTER 5

Reduced-Order Modeling for Airborne Transmission

of Respiratory Infections

This Chapter presents the reduced-order modeling of airborne transmission of respiratory infec-
tions in indoor environments with forced convection. Firstly, a computational experiment compar-
ing the results of the RANS-based Eulerian approach with the RANS-based Lagrangian approach
of [1] is presented in section 5.1. Then, the development of ROM for ventilated indoor environ-
ment is presented in section 5.2. The optimized ROM is applied to assess the risk of infection as
presented in section 5.3.

5.1 Comparison Between Eulerian Approach and Lagrangian

Approach Applied to Aerosol Transport

We investigate aerosol transport to predict airborne transmission in indoor ventilation systems.

The results are compared with previous works analyzing the airborne transmission with the La-
grangian method [1]. The goal is to confirm the reliability of our approach.
The geometry of the classroom model, as well as the iso-contour at time instance of 100s are
shown in Fig 5.1. The classroom model in the numerical study of aerosol transport [1] is applied
in the current work. A typical medium-size classroom is represented by a room of 9m x 9m x 3m.
Desks with simplified human bodies are located as shown in Fig. 5.1. A total of 9 students and
one instructor are located in each of their sites in the face of a lecture on the stage. Each simplified
geometry of the student has a rectangular mouth surface (0.06 x 0.03 m2 ) where the inlet boundary
condition is applied.

64
 Figure 5.1: Model geometry and iso-surface of infectivity field C=10 at t=100s.

The ventilation system follows the standard of ASHRAE 62.1 for recommended indoor air
quality. In the ventilation system, a total of 5 air supply diffusers and 4 air recycling diffusers
are settled on the top of classroom. The Cubic Feet per Minute (CFM) is approximated to be ∼
1230 for adequate ventilation according to the recommended configuration [1]. The inlet boundary
condition for air supply diffusers follows the settings in [1] in order to have an identical base indoor
ventilation flow field. Specifically, the airflow injects with a vertical air velocity of 0.395m/s at an
angle of 37 o from the horizontal surface, and the diffuser area gives 0.294m2 . The model mesh
contains a total of 3,041,512 cells with a minimum cell size of 0.5 cm and maximum cell size of 10
cm with gradual transition, maximum skewness of 0.823 (a mean value of 0.593), and maximum
aspect ratio of 3.21 (a mean value of 1.43) as presented in Fig. 5.2 [1].

65
Figure 5.2: Illustration of (a) the classroom model and (b) the computational mesh used in the CFD
simulations. The imagine from [1] is represented with permission of AIP Publishing.

The indoor ventilation flow field is solved from continuity and momentum equations in the be-
ginning using Reynolds Averaged Navier-Stokes (RANS) incompressible solver of ANSYS FLU-
ENT. The Re-Normalization Group (RNG) k − ϵ model is used to simulate the turbulence flow.
The SIMPLE algorithm is used in computing the flow fields. First order scheme is employed for
pressure interpolation while the convection and viscous terms are discretized using a second-order
scheme. The computed airflow fields (shown in Fig. 5.3) are frozen once the solution is converged.

66
Figure 5.3: (a) Turbulent kinetic energy, (b) velocity magnitude distribution, and (c) velocity vec-
tors across a slice going through students 2, 5, and 8. The imagine from [1] is represented with
permission of AIP Publishing.

The temporal transport behavior of the infectivity field is modeled in an Eulerian framework.
The second order convection scheme is used to converge the transport equation 2.3 for the infec-
tivity field. The Newman boundary conditions (zero flux for C) are applied to all walls, surfaces
of tables, and humans. Constant zero values of C is applied at the inlets.
We study the scenario by assuming student 5 in the center of the classroom carried the infectious
aerosol by exhaling activities such as coughing, and laughing. The transmission mode with the base
of steady-state flow fields is studied numerically through the Eulerian approach. In this way, the
airborne transmission can be assessed using the CFD solver. We first compare the prediction of
aerosol transmission between the Eulerian and Lagrangian approaches.

67
Figure 5.4: Distribution of 1 µm aerosol particles in the classroom at different points in time for
the (a) student 5 source. The imagine from [1] is represented with permission of AIP Publishing.

It has been observed from [1] in simulating the aerosol transport from initial diffusion of the
aerosol particles (t=20s) to the convected transport of aerosol particles (t=100s) (see Fig. 5.4). An
exhaling process is modeled using a single-release impulse source from student 5. The distribution
of the released particles are modeled under the framework of the Lagrangian approach. The result
in [1] predicts the spatial distribution of the aerosol particle of 1 µm followed by the initial one-
single impulse.
For a comparative study, a transient signal for the transport scalar source imitating the La-
grangian particle seeding in [1] is used. The signal for the source intensity is shown in Figure
5.5.

68
 Figure 5.5: Transient Scalar Intensity signal of the aerosol cloud.

The source initiates a cloud of aerosol with the infectious virus in a cubic box in front of student
5 (See Fig 5.6) with an intensity of 1000 for the first 5 seconds (as illustrated in Figure 5.6).

69
Figure 5.6: Initial scalar intensity source cloud (Source is marked in red box with 0.38x0.4 m2 in
area and 0.3m in height).

The results obtained using the Eulerian methods are presented in Figures 5.7, 5.8.
At 20 seconds, the aerosol scalar is still in the initial diffusion stage (See Figure 5.7). The area
outside the aerosol cloud for the scalar source remains uniform 0. In the area of the scalar source,
the aerosol source gradually spreads around due to the concentration gradient of the aerosol scalar.
However, it can be observed that aerosol scalar transport is dominated by diffusion due to the low
degree of airflow in the space where student 5 is located. The results predicted using the Eulerian
approach (see Figure 5.7) are consistent with the results obtained by the Lagrangian approach.

70
 Figure 5.7: The spatial distribution of infectivity field C at t=20s.

From 50 to 100 seconds of the transient process using the Eulerian approach, the propagation
mode of aerosol gradually changes to the co-leading propagation of diffusion and forced convec-
tion modes. As shown in Figure 5.8, the scalar is transported more efficiently with turbulent fluid
motions. With the transport of scalar intensity, the most intense part of the aerosol field gradually
shifts to the top of student 5, and the distribution of scalar is mainly concentrated in the positions
of students 4, 6, and 8. In addition, through comparative study with the result from [1], simulation
of the given transient process using the Eulerian approach shares the same characteristic of aerosol
scalar distribution.

71
 Figure 5.8: Infectivity field C at t=50s (on the left) and 100s (on the right)

It has been observed that a high degree of agreement has been reached between the two methods.
We acquired the particles distribution at t=50s and 100s from [1]. Next, we evaluate the spatial
distribution of the aerosol scalar with respect to the percentage of the aerosol in divided sections
over the total amount of aerosol.
The line plots with respect to the percentage of amount of aerosol particle among total par-
ticles number (Lagrangian method) or passive scalar concentration (Eulerian method) along the
classroom are compared in Fig 5.9. The probed line is picked along with the central direction of
classroom (y=2m, z=4.5m). The intermedium of probed points along the probed line follows x=
0.25 : 0.5 : 8.75 m.

Figure 5.9: Line plot of total percentage of Infectivity Field C/Particle along central direction of
classroom. The plot at t=50s is shown in (a); the plot at t=100s is shown in (b).

Along the selected direction, only one peak is shown in the line plot. Despite the difference in

72
percentage magnitude between Lagrangian and Eulerian approaches, the line plots at different time
instants clearly illustrate the agreement between the two approaches. The peaks of the signal are
found near the x=4.25, 4.75m in Fig 5.9. At t=100s, two sub-peaks at x=1.25 and 7.25 m are found
along the central direction. Note that the difference in line plots magnitude between Lagrangian
and Eulerian approaches is inevitable because the percentage magnitude depends on the size of the
probed boxes, i.e., how many Lagrangian particles contains in a probed box. Again, the probed
plots clearly illustrate a reasonable agreement can be achieved between the two approaches.
In Fig 5.10, we evenly divide the classroom model into 3 levels vertically as upper, middle,
and lower with a height of 1 m. In each zone, the space is divided into 9 sections with a length
and width of 3 m. In general, the entire classroom is divided into 27 sections. By calculating the
percentage of aerosol particles in each section with respect to the total amount of indoor aerosol
particles, aerosol concentration in each section can be evaluated and then compared between the
two approaches. The identified distribution of aerosol spread in the room is matched statistically
between Eulerian and Lagrangian approaches at t=50s and 100s as shown in the Table 5.1 and 5.2,
respectively.

Figure 5.10: Schematic diagram of the partition of the classroom

At t=50s, the area with high aerosol concentration is located in the Middle and Upper zones’
section 5. A reasonable agreement between Eulerian and Lagrangian methods has been achieved at

73
this moment. Since diffusion prior to t=50s dominates the transport of scalar, both the Eulerian and
Lagrangian methods predict that most sections outside section 5 have low aerosol concentration.

Table 5.1: The quantitative study between Eulerian/Lagrangian methods at t=50s

t=50s
Sections Upper:1 Upper:2 Upper:3 Upper:4 Upper:5 Upper:6 Upper:7 Upper:8 Upper:9

Lagrangian Approach/% 0 0 0 0 19 0 0 0 0

Eulerian Approach/% 0 0 0 0 16 0 0 0 0

Sections Middle:1 Middle:2 Middle:3 Middle:4 Middle:5 Middle:6 Middle:7 Middle:8 Middle:9

Lagrangian Approach/% 0 0 0 0 81 0 0 0 0

Eulerian Approach/% 0 0 0 0 80 0 0 1 0

Sections Lower:1 Lower:2 Lower:3 Lower:4 Lower:5 Lower:6 Lower:7 Lower:8 Lower:9

Lagrangian Approach/% 0 0 0 0 0 0 0 0 0

Eulerian Approach/% 0 0 0 0 3 0 0 0 0

When the propagation process reaches a considerable period of time, in the results at t=100s,
we can see that the aerosol particles have spread over the classroom, for which the transport is
dominated by convection.
At t=100s, sections 2, 4, 6, and 8 pose a high risk of infection because these areas have relatively
high aerosol concentrations as shown in Table 5.2. At the same time, in sections 1, 3, 7, and 9,
the risk of infection is relatively low, as aerosols are rarely transmitted to this area predicted by
both Eulerian and Lagrangian methods. Therefore, we can conclude that the aerosol propagation
pattern/distribution predicted by the Eulerian method is highly consistent with the works in [1]
using the Lagrangian method.

74
 Table 5.2: The quantitative study between Eulerian/Lagrangian methods at t=100s

t=100s
Sections Upper:1 Upper:2 Upper:3 Upper:4 Upper:5 Upper:6 Upper:7 Upper:8 Upper:9

Lagrangian Approach/% 0 4 0 5 61 2 0 2 0

Eulerian Approach/% 0 3 0 3 58 3 0 3 0

Sections Middle:1 Middle:2 Middle:3 Middle:4 Middle:5 Middle:6 Middle:7 Middle:8 Middle:9

Lagrangian Approach/% 0 2 0 3 14 1 0 2 0

Eulerian Approach/% 0 1 0 2 18 2 0 4 0

Sections Lower:1 Lower:2 Lower:3 Lower:4 Lower:5 Lower:6 Lower:7 Lower:8 Lower:9

Lagrangian Approach/% 0 1 0 2 0 0 0 0 0

Eulerian Approach/% 0 1 0 1 0 0 0 1 0

In conclusion, we confirm, quantitatively, the similar prediction of the spatial distribution of

aerosol particles by Lagrangian and Eulerian approaches.

5.2 ROM Generation and Optimization for Ventilated Indoor

Environments

5.2.1 Convergence Issues: Regarding the Residual of Transport Equation

Solution

We use commercial CFD solver, STAR-CCM+, to solve the transport equation (2.3) based
on iterative approach. It is very important to have an accurate and reliable CFD solution while
constructing a Krylov-based ROM. In other word, decent convergence is needed. The problem does
not arise in dissuion-dominated processes, such as the heat transfer in a battery pack considered
in Chapter 4. It may become acute, however, in processes with a stronger convection component
such as the aerosol propagation discussed in this chapter.
Relative residuals and absolute residuals can be used as convergence criteria [52]. Both abso-
lute or relative (to the 0th iteration) value of the residual are good measures of accuracy. If the

75
initial guess is inaccurate (e.g., we start with a zero scalar everywhere), relative value is more ap-
propriate. If the initial guess is more accurate, and the residual is small from the beginning, using
the relative value can sometimes overestimates the convergence error, so the absolute residual is a
better choice.
The absolute residual value is a better option to evaluate convergence in our case. Each CFD
solution of the ROM derivation algorithm (see section 3.3) starts with the previous solution as an
initial guess. The initial value of the residual is, therefore, relatively small (∼ 10−4 ). We use
absolute residual in all tests to guarantee an unbiased criterion for the final convergence level of
each iteration of ROM.
STAR-CCM+, or similar commercial CFD software, solve the convection equation (2.2) using
a 2nd order discretization scheme. The under-relaxation factor (URF) can be adjusted to ensure
sufficient convergence. The price to pay by lowing the URF is usually the longer computation time
needed to get a fully converged solution.
Additionally, it should be noted that the convergence is not only be judged by a sufficiently
small residual. One also need to make sure that variables of the solution remain steady for a
sufficiently long period of time.
To summarize, the CFD solution needs to be accurate enough in the ROM calculation process.
The strategies and methods that can be used to guarantee the converged solution are adjusting
the URF and ensuring the high confidence of the CFD solution by observing the stability of both
residual and solution fields as well as the quality of the mesh. The convergence reaches to absolute
residual value of ∼ 10−12 during the ROM generation.

5.2.2 ROM Generation Using Single-Input Algorithm

In this section, we present the results, using the indirect one-sided Arnoldi algorithm 1 to con-
struct ROM. The generated ROMs are assessed to determine the optimal reduced order. We use
the Eulerian approach with the commercial CFD software, STAR-CCM+, to solve the discrete
diffusion convection equation for the infectivity field C. Theoretically, as the order increases, the
accuracy of the ROM increases accordingly due to the moment matching property of the Krylov

76
subspace method [29]. However, in the large-scale CFD models with large number of grid cells,
the error of the CFD solver accumulates in the course of execution of the ROM-generation algo-
rithm. As a result, there is a trade-off in the process of selecting the optimal reduced order for
ROM. We quantify the approximation error by calculating the ∥ε∥2 and ∥ε∥∞ of the difference
between the full-order and reduced-order solutions.
The accuracy tests are performed for the source intensity illustrated in Fig 5.5. The intensity
is constant 1000 for the first 5s within the source zone and zero otherwise. The simulations are
carried out for 100s.
We observe the dynamics of the transient passive scalar signal monitored by a probe point in
the model under different r (shown in Figure 5.11). The probed point is located in front of the
source person, student 5. It can be observed that at low r, the transient signal does not closely
follow the dynamics predicted by the FOM solution. But as r increases, the quality of the ROM
prediction increases. It should be emphasized that the fluctuation of the signal keeps existing along
with increase of reduced orders.

77
 4000

FOM
r=11
3000
r=22
r=33
r=50
2000 r=75
C

1000

-1000
0 20 40 60 80 100
Time/s
Figure 5.11: Transient signal at the probed point in front of student 5

Secondly, we compare the passive scalar line probed plot of FOM and ROM of different r in
50s (r=5, 15, 30, 60, 85). The probed line is picked along with central direction of classroom
(y=2.25m, z=4.5m). As shown in Fig 5.12, the performance of ROM improves at larger r. At
low order r, such as r=5, 15, the line probed signal of ROMs is obviously distorted and deviates
significantly from the FOM line probe signal. As r increases, the line probed signal gradually starts
to capture the correct behavior of the system, but there are still relatively large fluctuations locally.
Taking r=30 as an example, the probed line signal still has a mismatch near the location x=4.5m.
Noticeable derivations are also found at x=1.5∼3.5m. Accuracy is achieved at high values of r,
such as at r=60, 85. The probed line of the ROMs can already match the probed signal of the FOM
very well.

78
Figure 5.12: Distributions of C along the line at y=2.25m, z=4.5m obtained by the FOM and by
SIMO ROMs with various r.

We further compared the fit of the ROMs with the FOM at different r using the contour in the
cross-section at z=4.5m. Their approximation degrees under the high frequency response (t=5s)
and low frequency response (t=50s) of the system, respectively, we calculated the results in the
corresponding ROM through contours with FOM.

79
(a) (b)

(c) (d)

(e) (f)

Figure 5.13: Contour plot at t=5s in the cross-section z=4.5m under the signal of 5.5. ROM of (a)
r=5; (b) r=15; (c) r=30; (d) r=60; (e) r=85; (f) FOM

In Figure 5.13, the low-order ROMs show obvious deviations from the FOM solution. As r
increases to 60 and 85, the ROM becomes more accurate although some unphysical fluctuation in
the upper part of the cross-section are still visible.

80
(a) (b)

(c) (d)

(e) (f)

Figure 5.14: Contour plot at t=50s in the cross-section z=4.5m under the signal of 5.5. ROM of (a)
r=5; (b) r=15; (c) r=30; (d) r=60; (e) r=85; (f) FOM

We also compared the predictions of ROMs and FOM at 50s as shown in Figure 5.14. Good
accuracy is found at high r. In the case of r=5 or 15, the ROM predicts that the C circulates around
the ventilation flow and the convection occurs more violently. In the meantime, obvious distortion
in the field is observed near the ventilation exit and in front of student 5. When r=15, there is
local obvious noise in the transport phenomenon in front of the source person. At r=30, although
the ROM can basically simulate the transport pattern, we still observe that the transport prediction
near the vent and in front of the source person has relatively large interference. Interestingly, ROM
at r=60 has better accuracy than at r=85. To confirm this observation, we further evaluated the
simulation accuracy of ROM in a quantitative manner.
The results are presented in Table 5.3, where we show the ∥ε∥2 and ∥ε∥∞ for the entire domain
for different values of r at 5s and 50s. We observe that the simulation accuracy of ROM gradually
increases with the increase of r. But as r rises to a certain value, the accuracy of ROM begins to

81
decrease. Therefore, we choose r=65 as the optimal ROM in the case of the considered system.

Table 5.3: The quantitative evaluation for ventilation ROM using Algorithm 1.

5s 50s
r ∥ε∥2 ∥ε∥∞ r ∥ε∥2 ∥ε∥∞
1 4101.71614 0.033970722 1 420.0520974 0.004919662
2 3897.204849 0.0327505 2 464.5043782 0.005266173
3 3209.994665 0.027981335 3 261.2120021 0.004700886
4 2643.473463 0.024007687 4 177.4318848 0.00393388
5 2405.838849 0.022142363 5 160.9852151 0.00340241
6 2193.454966 0.020517757 6 231.9732018 0.003135373
7 2059.275084 0.019469939 7 273.1889522 0.003136924
8 1887.51468 0.018219563 8 301.2598654 0.003070044
9 1736.582754 0.017197164 9 337.2806877 0.003132623
10 1373.395574 0.014404866 10 274.3627911 0.002946727
11 1184.091838 0.01257946 11 186.1686172 0.002787503
12 1014.870606 0.010761379 12 136.9093779 0.002324464
13 900.0639051 0.009412445 13 124.8334285 0.001947135
14 809.0717399 0.008383301 14 120.0027065 0.001719401
15 716.5651398 0.007404798 15 109.6250602 0.001522216
16 629.4961751 0.006417915 16 112.3293944 0.001433778
17 481.0320972 0.005281978 17 114.3209319 0.001278743
18 579.3687363 0.005197281 18 99.00432428 0.001233456
19 449.6107714 0.00424985 19 76.66943915 0.001139877
20 373.4839549 0.003678724 20 58.95351351 0.001050186
21 348.0382623 0.003230563 21 50.05238558 0.000971481
22 357.9744149 0.003018883 22 50.64196978 0.00089657
23 365.7603587 0.002872882 23 53.76353273 0.000805993
24 374.760428 0.002792323 24 61.05185464 0.000734782
25 376.8608569 0.002781679 25 60.55166094 0.000702364
26 380.9151918 0.002777583 26 56.73706994 0.000673788
27 385.5845241 0.002791613 27 43.48603165 0.000603633
28 386.9239719 0.002796791 28 33.9478127 0.00056523
29 387.2511066 0.002796745 29 25.7624994 0.000538178
30 381.7347837 0.002751962 30 23.22064723 0.000506549
31 385.5874459 0.002775623 31 23.26875473 0.000473082
32 384.119255 0.002758119 32 20.12503697 0.000432877
33 379.1195754 0.00270728 33 21.05494061 0.000404867
34 378.1193029 0.002701546 34 24.4136724 0.000373739
35 375.1193819 0.002671556 35 24.74476856 0.00035428
36 370.8379458 0.002630417 36 23.76344851 0.000337107

82
37 360.3596259 0.002517002 37 19.89312445 0.000323574
38 361.9317091 0.002510891 38 16.12372054 0.000316971
39 356.3608593 0.002459875 39 14.65690652 0.000305186
40 353.5764519 0.002411337 40 13.65480811 0.000294066
41 349.5181615 0.002360237 41 12.39039554 0.000281039
42 340.9504808 0.002267651 42 10.77884304 0.000263454
43 339.2978647 0.002249151 43 9.566147923 0.000244988
44 336.2699749 0.00221663 44 8.870075914 0.000233505
45 333.1574472 0.002184976 45 9.549938831 0.000223595
46 327.463079 0.002143439 46 10.60079039 0.000214064
47 318.8379016 0.002057057 47 11.14635746 0.00020724
48 318.7863228 0.002071424 48 11.82151129 0.000194305
49 312.2281737 0.002028678 49 12.13707098 0.000189179
50 310.9706287 0.002040752 50 12.40439857 0.000184712
51 305.0965482 0.002014014 51 12.52483672 0.000182363
52 300.8735916 0.001994799 52 12.51498364 0.000178176
53 298.2618394 0.001991562 53 12.25911892 0.000173662
54 294.671046 0.001994129 54 11.81147078 0.000167723
55 291.4015874 0.001990397 55 11.20097912 0.000164592
56 287.254523 0.001982461 56 10.41117734 0.000157704
57 274.9814845 0.001930557 57 9.332362694 0.000163888
58 280.0147708 0.001975191 58 8.450156536 0.000152497
59 277.3655784 0.001972204 59 7.760303013 0.000149295
60 275.1197137 0.001975357 60 7.056458758 0.000150088
61 271.5996527 0.001971519 61 6.938409566 0.000150522
62 269.9444072 0.001969991 62 7.488890788 0.00015198
63 262.8344384 0.001943989 63 8.294343379 0.000158903
64 266.3766401 0.001978931 64 9.428823962 0.000158024
65 264.4725261 0.001980763 65 10.32634107 0.000162373
66 263.2362821 0.001982362 66 11.00605074 0.000165951
67 254.6469817 0.001955302 67 11.997259 0.000175456
68 259.4744452 0.00198699 68 13.04606974 0.000178997
69 258.4825942 0.001990735 69 13.76362921 0.000184953
70 256.2293271 0.001990671 70 14.62744158 0.000193826
71 255.3931855 0.001996734 71 15.55298206 0.000203092
72 243.0935102 0.001971095 72 16.77684116 0.000219302
73 252.3462542 0.002002973 73 17.6671766 0.000226228
74 250.7122946 0.00200422 74 18.70726266 0.000239174
75 245.5665462 0.001993861 75 19.91535501 0.000257106
76 248.6873329 0.002012259 76 21.0015391 0.000272937
77 246.790942 0.002009456 77 22.02492086 0.000287745
78 247.4924163 0.002016787 78 23.04102533 0.000302372
79 247.0724657 0.002017424 79 23.81176658 0.000312108

83
 80 247.1386967 0.002014025 80 22.3660962 0.00029611
81 240.6176409 0.001993884 81 21.50051095 0.000288445
82 240.4322138 0.001993596 82 21.4791231 0.000288299
83 243.8656064 0.001999515 83 20.60844857 0.000280471
84 249.7060725 0.002009539 84 14.17542578 0.000227744
85 250.4318961 0.002007852 85 32.00092188 0.000281084

The ROM of order 65 is validated in predicting the transient dynamics of the ventilation system
under the same transient signal as in Figure 5.5. It can be observed that the ROM can well predict
the distribution of the infectivity field C. As an example, we compare the ROM of with r=65 and
FOM at t=100s in Fig 5.15.

(a) (b)

Figure 5.15: Infectivity field C at t=100s. (a) Optimal ROM, with r=65, using SIMO algorithm,
and (b) FOM.

After the comparative study, ROM of r=65 can be used as the optimal result, which can well
simulate the aerosol distribution of the original large-scale model.

5.2.3 ROM Generation Using the Frequency-Shift Algorithm

In this section, we present the ROMs generated by frequency-shift Arnoldi algorithm, following
with optimization of reduced order. We select a total of 11 sampling points as Sigma set in con-
struction of ROM following the Algorithm 3. As presented in Table 5.4, the Sigma set is picked
from a range of frequency to capture the characteristics of the system among different response
rates, i.e., from low frequency to high frequency. For each of the 11 value Sigma, we develop
ROMs of different orders using the frequency shift Arnoldi approach.

84
 Table 5.4: The 11 Sigma points for frequency-shift Krylov method

1 2 3 4 5 6 7 8 9 10 11

Sigma 0 0.001 0.002154 0.004642 0.01 0.021544 0.046416 0.1 0.215443 0.464159 1

5.2.3.1 Time Derivative Method Without Access to the System Matrix A

We apply a new method to construct the orthogonalization matrix X during the frequency-shift
algorithm. The method discretizes the time derivative of the field using the 2nd order scheme.
Technically, the method is validated to be able to produce an accurate approximation of the field.
In our case, the orthogonalization matrix X is defined as X = (A−σI)−1 V . The method calculates
the time derivative of C to get the vectors of matrix X. Therefore, the vectors in X is defined as
xi = (A − σI)−1 vi .
As mentioned before, commercial CFD software, in general, does not give access to the sys-
tem’s matrix. We have developed an indirect method to get the necessary information. In one-sided
Arnoldi algorithm, the orthogonalization matrix X is defined as AV = A[v1 v2 . . . vr ]. The proposed
method calculates the 2nd order forward time derivative of passive scalar field in order to get xi :

dC 3C0 − 4C0+∆t + C0+2∆t

= (A − σI)C ≈ lim , (5.1)
dt ∆t→0 2∆t

where reasonable approximation of AV can be achieved with substantially small time step ∆t.
One can validate the method by evaluating the b = Av1 . The 2nd norms of cell-by-cell error
vector b − Av1 does not exceed 10−6 . As a conclusion, time-derivative method is a reasonable way
to access to the system matrix when A cannot be accessed directly.

5.2.3.2 ROM Optimization

The optimization of ROM developed using the frequency shift Arnoldi algorithm is presented
in this section. We apply the same step-change signal as the one used for the Arnoldi algorithm
optimization process (See Figure 5.5) to evaluate the ROM approximation.

85
 The computed time-varying signals at the probed point are presented. The ROM of r=11
matches the 0th order moment of the system. The transient signal that matches only one mo-
ment is off-track for completely missing the dynamics of the system after 5 seconds. As r goes
to higher order, the transient signal at r=22, 33 can follow the dynamics from FOM that produce
the high-fidelity result of the system. However, we can observer the fluctuation of the signal at
high-frequency response area. For example, the signal at t from 10s to 20s. As r increase, the
ROM starts to closely follow the temporal dynamics of the system.

4000

3500 FOM
r=99
3000 r=66
r=33
r=22
2500 r=11
C

2000

1500

1000

500

0
0 20 40 60 80 100
Time/s
Figure 5.16: Results of frequency-shift ROMs for the source signal in Fig 5.5

Secondly, we compare the passive scalar line probed plot of FOM and ROM of different r at
50s (r=11, 22, 33, 66, 99). The probed line is picked along with central direction of classroom
(y=2.25m, z=4.5m). As shown in Fig 5.17, the line probe signals of ROM moves closer to that of
FOM at higher r. The accuracy is poor at low values of r (r=11,22, and 33). As r increases, the line

86
probed signal can gradually capture the behavior of the system, but there is still a noticeable error,
i.e., in the range of x=[4, 5] m. The results obtained at r=99 are more accurate than the results of
r=66.

Figure 5.17: Distributions of C along the line at y=2.25m, z=4.5m obtained by the FOM and by
frequency-shift ROMs with various r.

We further evaluate the ROMs at different r comparing with the FOM contour plot of C in the
cross-section at z=4.5m. The contours at t=5s and 50s are plotted in order to assess the high-
frequency and low-frequency response, respectively. At t=5s, ROMs of lower order presents sig-
nificant errors. As shown in Figure 5.18, the ROM of r=11 in (a) illustrate the basic structure of the
field. The contour illustrates the high-concentration cloud of aerosol located in front of index per-
son. However, the contour illustrates obvious deviations around the field. As r increases, the ROM
can be more accurate, in the case of r=66 or 99, to approximate the high-frequency characteristic
of the FOM.

87
(a) (b)

(c) (d)

(e) (f)

Figure 5.18: Contour plot at t=5s in the cross-section z=4.5m under scalar intensity signal 5.5.
Frequency-shift ROMs of (a) r=11; (b) r=22; (c) r=33; (d) r=66; (e) r=99; (f) FOM

The ROMs at t=10s of different r are shown in Figure 5.19. The ROMs of r=66 and 99 shows
a robust approximation of the FOM. As shown in (f) and (g), the distribution of the passive scalar
is almost indistinguishable compared with FOM. The error vectors’ contours between ROMs and
FOM are presented in Figure 5.20. The error mainly locate near the ventilation flow between index
person and ventilation inlets at r=11. As r increase to 99, the fluctuation errors are largely removed
outside the aerosol cloud area.

88
(a) (b)

(c) (d)

(e) (f)

Figure 5.19: Contour plot at t=10s in the cross-section z=4.5m under scalar intensity signal 5.5.
Frequency-shift ROMs of (a) r=11; (b) r=22; (c) r=33; (d) r=66; (e) r=99; (f) FOM

Note that the ROMs using frequency-shift Arnoldi algorithm significantly reduce the error. As
shown in Figure 5.13, the ROMs using conventional Krylov method show strong fluctuations of
C in comparison with the FOM distribution even with optimal reduced order. The fluctuations are
strongly reduced by using the frequency-shift Arnoldi algorithm.

(a) (b)

Figure 5.20: Contour plot of error vector at t=10s. ROM of (a) r=11; (b) r=99.

89
(a) (b)

(c) (d)

(e) (f)

Figure 5.21: Contour plot at t=50s in the cross-section z=4.5m under scalar intensity signal 5.5.
Frequency-shift ROMs of (a) r=11; (b) r=22; (c) r=33; (d) r=66; (e) r=99; (f) FOM

Next, we compare the predictions of ROMs at 50s to evaluate the low-frequency response of the
ROMs. The results are illustrated in Fig 5.20. Large unphysical fluctuations of C are predicted by
the ROMs at lower reduced orders. For example, the ROM of r=11 produces significant error near
the boundary of aerosol cloud. The ROMs of r=22, 33, 66 have obvious errors at the field near the
inlets. It is found that only when r is high enough, the FOM can be accurately approximated (see
Figs 5.21 d-f). It can be seen that when r=99, ROM simulates the slow-dynamic characteristics of
the system accurately. To confirm this observation, we further evaluated the simulation accuracy
of ROM using the norms of error vectors at different time moments.
The computed norms of the error vector at t=5 and 50s are shown in Table 5.5. It is observed
that the ROMs using frequency shift algorithm perform more accurately than the ROMs generated
by the SIMO algorithm. The 2nd and max norms of the error show a compatible approximation at
r=33 with previous ROM of optimal order, i.e., r=65. Furthermore, ROMs produced by frequency

90
shift Arnoldi algorithm improve the approximation of high-frequency response. For example, the
max norm of error at t=5s has been dramatically reduced at r=85 if compared with that using SIMO
algorithm.

Table 5.5: The quantitative evaluation for ventilation ROM using Algorithm 3.

5s 50s
r ∥ε∥∞ ∥ε∥2 r ∥ε∥∞ ∥ε∥2
1 4140.738195 0.03429009 1 424.3104154 0.005002981
2 3872.915672 0.032636572 2 475.6279231 0.005371553
3 3868.398211 0.032625085 3 447.7609663 0.005155719
4 2984.428398 0.026558019 4 209.8513179 0.004580417
5 2055.412017 0.019686727 5 253.1895297 0.00306651
6 2059.551988 0.019695938 6 251.9027246 0.003041048
7 994.3827201 0.010891809 7 190.9258701 0.0030331
8 351.3838783 0.003417481 8 109.8881168 0.001629379
9 286.1558871 0.002476145 9 86.52083028 0.000957381
10 284.6152344 0.002477172 10 86.29178799 0.000957047
11 294.7784753 0.00253217 11 85.37398925 0.000957132
12 290.348287 0.002513407 12 89.04973739 0.00091977
13 280.5593175 0.002462447 13 90.32852677 0.000921084
14 280.717766 0.00245801 14 84.58507311 0.000985052
15 280.6759955 0.002454161 15 85.42174877 0.000978037
16 270.1835864 0.002394936 16 68.75056717 0.001103538
17 269.305744 0.002396156 17 77.85048094 0.000913818
18 322.2060712 0.002235286 18 63.97881905 0.000877883
19 269.1437511 0.001981614 19 48.60275106 0.000814185
20 196.9411118 0.001717461 20 46.11742955 0.000788997
21 87.04535113 0.000496735 21 38.96297114 0.000668758
22 95.21826627 0.0004599 22 36.72728765 0.000621495
23 95.12572062 0.000461065 23 36.92057896 0.000630628
24 94.45730934 0.000462076 24 36.49872509 0.000646658
25 94.50420003 0.000456035 25 36.53172436 0.000608468
26 94.56666133 0.000459529 26 35.60666628 0.000615415
27 95.09070737 0.000458524 27 34.88218678 0.000594917
28 94.85938324 0.000455672 28 47.83768379 0.000509329
29 94.78131777 0.000459864 29 51.20705135 0.000450591
30 92.97670777 0.0004637 30 34.84213546 0.000419225
31 102.7345595 0.000498162 31 29.79555923 0.000407504
32 96.13269183 0.000339518 32 29.63707332 0.000400093

91
33 74.16153765 0.000286972 33 29.94253344 0.000397919
34 74.8185961 0.000286509 34 29.9995258 0.000393284
35 73.18978949 0.000288978 35 29.81938558 0.000378538
36 75.691734 0.00028609 36 30.04034733 0.000368594
37 76.16985833 0.000285139 37 29.4139569 0.000373346
38 75.40586068 0.000282799 38 28.81070929 0.000362738
39 76.38162624 0.00028283 39 25.82486129 0.000355806
40 76.88557145 0.000286754 40 22.15686792 0.000347243
41 70.64468065 0.00028431 41 21.73414025 0.000340462
42 80.3399607 0.000285165 42 21.6523286 0.000339963
43 84.63160314 0.000281628 43 20.91215712 0.000339243
44 84.52769471 0.000282077 44 27.15305415 0.000335716
45 84.48766833 0.000282165 45 27.68647875 0.000329695
46 84.5982654 0.000282049 46 28.28085811 0.000326172
47 84.38433642 0.000282678 47 27.42632961 0.000336297
48 83.13373865 0.000283416 48 26.88153079 0.000331807
49 81.71833426 0.000284112 49 26.58917599 0.000328711
50 82.85945237 0.000280403 50 22.02096767 0.000333279
51 88.49075232 0.000286043 51 18.64539311 0.000303598
52 88.98159797 0.000284216 52 11.53393584 0.000222751
53 84.85617623 0.00025752 53 11.29946056 0.000214516
54 82.10516749 0.000236192 54 11.39499322 0.000211473
55 80.83618406 0.000229284 55 11.09211658 0.000211159
56 80.23495479 0.000229685 56 11.22642284 0.000209509
57 80.99212561 0.000229824 57 11.2933125 0.000207852
58 80.81120688 0.000230307 58 11.53847567 0.000206681
59 80.72245316 0.000230333 59 11.53824361 0.000206755
60 80.38113269 0.000230464 60 11.53868949 0.000206822
61 79.12503821 0.000231323 61 11.53736913 0.000207302
62 79.52041087 0.000231239 62 11.08537666 0.000209503
63 80.60808813 0.000231679 63 9.719092421 0.000190487
64 79.74282466 0.000232522 64 10.45560793 0.000166313
65 80.94356662 0.000226247 65 9.762007456 0.000161283
66 80.21971661 0.000224379 66 9.812793271 0.000160651
67 80.15292254 0.00022388 67 9.858599507 0.000160376
68 80.13184813 0.000223841 68 9.843595986 0.00016099
69 80.09164418 0.000223853 69 9.889235497 0.00016133
70 80.26324389 0.000223825 70 9.890472838 0.000161144
71 80.11301682 0.000223883 71 9.909891519 0.000160963
72 79.84987124 0.000223946 72 9.911772143 0.00016111
73 80.41588967 0.000223536 73 9.993808244 0.000154979
74 79.61406187 0.000223843 74 7.44763887 0.000147302
75 79.74200266 0.000222421 75 8.45632956 0.000136316
76 79.89825842 0.000223552 76 9.003205244 0.000130581

92
 77 79.42457592 0.000218642 77 8.803988357 0.000128499
78 78.81811283 0.000218809 78 8.815514561 0.00012866
79 78.15710558 0.000219241 79 8.806006378 0.00012873
80 79.49446507 0.000218638 80 8.827649788 0.000127879
81 79.39284568 0.000218646 81 8.814389973 0.00012799
82 79.09378749 0.00021875 82 8.814828881 0.000128132
83 78.77479734 0.000218887 83 8.813601796 0.000128404
84 78.51413584 0.000220823 84 8.775443972 0.000123604
85 79.25587066 0.000219975 85 9.097768139 0.000121834
86 77.97167387 0.000219744 86 7.910817006 0.000108163
87 77.47846467 0.000220641 87 7.823645515 0.000107796
88 77.96779923 0.000219313 88 7.829633899 0.000109517
89 78.11194335 0.000219124 89 7.807124303 0.000109407
90 78.12878867 0.000219139 90 7.807935979 0.000109109
91 78.08862327 0.000219205 91 7.799255806 0.000108844
92 78.0487269 0.000219199 92 7.803915837 0.000109723
93 77.79960026 0.000219261 93 7.806329952 0.000109621
94 78.07206798 0.000219172 94 7.806841216 0.000109584
95 78.02977001 0.000219243 95 7.866997447 0.000108631
96 77.20928465 0.000218776 96 7.930122796 0.000108553
97 76.99507008 0.000219701 97 8.273007198 0.000101248
98 76.21877517 0.000220542 98 8.255143466 0.000100166
99 78.14531087 0.000220394 99 8.243749986 9.82E-05

Through Table 5.5, we found that the ROM have optimal accuracy with respect to high-
frequency and low-frequency response at r=88. Thus, the ROM of r=88 is the optimal approxi-
mation of the system. As r rises above 88, the accuracy of ROM remains approximately the same.
All the further Ventilation-ROM cases will be verified and evaluated using the optimal frequency-
shift ROM of r=88.

5.3 ROM Application

In this section, we apply the optimal ROM generated by frequency-shift Arnoldi algorithm to
predict the aerosol transmission in an indoor classroom.

93
5.3.1 Exhalation Cycle

A time-dependent infectious aerosol released from an human source student in the classroom
is the subject of current study. We define the source term S(x, t) in (2.3) as S(x, t) = B(x)u(t),
where B(x) is the localized spatial distribution, and u(t) is the time modulation. The spatial
distribution is limited to the generic conical breathing cloud illustrated in Fig 2.2. The cloud is
0.5m long. The circle radii of the upper and lower bases are 0.2 and 0.04m, respectively. The source
intensity gradually decreases along the central axis of the breathing cone. The source intensity
follows a normal curve distribution within the cross-sectional direction.
The release of the aerosol within the cone is defined as a time-dependent modulation u(t).
The signal considered in our study is the combination of breathing cycle and coughing events of
different magnitude. During breathing activities, the source person regularly exhales a volume of
aerosol cloud from the mouth. The breathing cycle has a period of 4 seconds. According to the
previous study of exhaling activities [48]. The first two seconds of the cycle are the exhalation
phase. Following the exhalation phase, the next two seconds are the inhalation phase. We assume
that the intensity of the source is described by a sinusoidal function during the exhalation phase
and zero during the inhalation phase (see Figure 5.22 (a)).
The coughing cycle has period of 0.6s [48]. The time modulation u(t) during one cycle is taken
from the experiments [48] (see Figure 5.22 (b)). Considering that coughing varies in intensity,
we use the coughing cycles of different peak intensity magnitudes. As shown in Figure 5.22
(b), the coughing intensity peaks are defined as 200, 400, 600 and 1000. The cumulative source
modulation signal used in the computational validation discussed below is shown in Fig. 5.23.In
the most severe coughing situation, the peak value of the breathing cone intensity is ∼ 1000.

94
 Breathing Cycle Coughing Cycle
(a) 60 (b) 1200
Peak of intensity signal
50 1000 1000
800
600
Scalar Intensity

Scalar Intensity
40 800 400
200

30 600

20 400

10 200

0 0
0 1 2 3 4 0 0.1 0.2 0.3 0.4 0.5 0.6
Time/s Time/s

Figure 5.22: Breathing and coughing cycles used in simulations. Source modulations u(t) corre-
sponding to breathing cycles (a) and coughing cycles (b) are shown.

5.3.2 ROM Validation Under the Designed Coughing Cycle

The validation study based on the scenario when the infections aerosol particles are generated
within the cone in front of the student 5 according to the time modulation shown in Fig. 5.23 is
presented here.

95
 Time-varying input signal
1200 400

1000 200
Scalar Intensity

0
0 0.2 0.4 0.6
800

600

400

200

0
0 50 100 150 200 250
Time/s
Figure 5.23: Source modulation signal u(t) used in the numerical experiment.

The infectivity field C is monitored in front of students 4, 5, 6 and 8, which are identified as
‘high risk’ for exposure to aerosol particles. As shown in Figure 5.24, the optimal frequency-shift
ROM closely follows the full-order model solution. For students 4,6 and 8 located at a significant
distance from the source, the infectious aerosol particles starts accumulate after approximately 100
seconds. The data clearly illustrate the potential of infection for these students. We note that small
deviations from the FOM results can be observed during the transient process. The deviations
are caused by different source of errors including, the numerical error of matrix manipulations,
full-order model convergence, and imperfect linear independence of vectors in the Krylov sub-
space. We carefully operate the ROM generation in order to eliminate these factors and produce
the optimal ROM. The results, as oen can see in Fig. 5.24, are reasonably accurate.

96
 Probed point signal in front of student 4 Probed point signal in front of student 5
300
(a) 0.3
(b) ROM with r=88
FOM

C 0.2 200

C
ROM with r=88
0.1 FOM 100

0 0
0 50 100 150 200 250 0 50 100 150 200 250
Time/s Time/s
Probed point signal in front of student 6 Probed point signal in front of student 8
0.3
(c) ROM with r=88
FOM
(d) 0.4

0.2
C

C
0.2
ROM with r=88
0.1 FOM

0
0
0 50 100 150 200 250 0 50 100 150 200 250
Time/s Time/s

Figure 5.24: Transient scalar signals, under the input of 5.23, at the probed points in front of
students whose frontal space are identified as ’high-risk’ area.

The distribution of the infectivity field in the cross-section z = 4.5m at different time moments
are presented in Fig. 5.25. We see that aerosol starts to diffuse around the space for the first 50
seconds. Its transport is dominated by the forced convection by turbulent flow at later time. The
transport of the aerosol particle is accelerated by the ventilation which enable it to travel to a larger
distances (see Fig. .5.25 e).

97
(a) (b)

(c) (d)

(e) (f)

Figure 5.25: Cross section at t=10, 50, 220s of ROM at (a), (c), (e) and FOM at (b), (d), (f).

We use the line probes across the middle of classroom at y=1.3 and 2.5 m to examine the quality
of the ROM approximation. The results are plotted at different times. As shown in Figure 5.26, the
ROM can capture the spatial distribution of the infectivity field when the transport of the aerosol
is dominated by convection. However, at the early stage of transport, e.g., t=10s and 50s as shown
in Figure 5.26 (a) and (b), some inaccuracies can be observed. This can be attributed to very low
(nearly zero) value of C in most of the room at this time. The ROM produces inaccurate results in
such situations, although the absolute value of incorrectly predicted C is typically low, so the error
does not have significant practical consequences.

98
(a)

(b)

(c)

Figure 5.26: Probe line signal of (a) t=10s,(b) 50s and (c) 220s at y=2.5 m (elevation)

99
 The accuracy of the ROM prediction is much better along the probe line at y=1.3m (see Fig.
5.27). At all three time moments, the ROM data are practically indistinguishable from those of
FOM.

100
(a)

(b)

(c)

Figure 5.27: Probe line signal of t=10s (a), 50s (b) and 220s(c) at y=1.3 m (elevation)

101
 Finally, the accuracy of the ROM approximation is evaluated through norms of error vectors at
different time moments as shown in Table 5.6. It is observed that the maximum deviation between
the FOM and ROM fields does not exceed 4 at t=220s. This indicates good accuracy of ROM in
predicting the response to coughing activities (high-frequency response). Note that the maximum
value of C at 220s is 184. Therefore, the ROM poses a relative error of ∼ 2.2%. In this case, the
ROM presents a reasonable approximation with respect to the dynamics of FOM.

Table 5.6: The norms at t=10, 50 and 220s

Norms
Time/s
∥ε∥∞ ∥ε∥2

10 2.5974 1.74E-05

50 3.0363 1.12E-05

220 4.0385 1.99E-05

5.3.3 Time-Varying Process Predicted by ROM

At this point, the accuracy of the ROM has been confirmed by validation test conduction for
the first 250s with ROM with r=88 in the previous section. We now apply the ROM with r=88 to
predict the aerosol transmission in the same classroom after t=250s in order to assess the risk of
infection. This will illustrate the ability of ROM to rapidly predict and quantify the risk level in
each area of the classroom.
We consider the spread of infectivity field in the classroom during a lecture lasting 45 minutes.
We assume the student 5 is the source person while the other persons in the classroom are the
potential recipients of the infection. We apply ROM to predict and analyze the aerosol transmission
for the 45-min transient process in order to determine the area that is highly likely to be exposed
to infectious particles.
We assume that the classroom is ventilated before the class, so value of C is uniformly 0 in the
classroom. The time modulation u(t) of the source shown in Fig. 5.28 is utilized. We assume that

102
the human source has a steady breathing cycle during the trial, with brief intermittent coughing
occurring from time to time. The whole process is a transient process with a duration of 2700s (45
minutes) as shown in Figure 5.28.

A time-varying input signal during class

1200
Red Line

1000

800
400
Scalar Intensity

600 200

0
0 0.2 0.4 0.6
400

200

0
0 200 400 600 800 1000 2600 2800

Time/s

Figure 5.28: Time-varying source modulation signal u(t) in simulating a lecture.

Simulation of the process using the FOM is estimated to have taken 7200 core-hours to com-
plete. Even a high-performance workstation would need at least half a month for parallel com-
puting on 20 cores to complete the task. It would be a time consuming and inefficient process
especially if a large number of parametric studies were needed. We apply the ROM with r=88 to
simulate the same process. The task is finished using a single core on a personal desktop within 90
minutes. This reduces the computation effort by approximately 4800 times. As a result, ROM can
substantially reduce the computation time while maintain the accurate results.

5.3.4 Risk Assessment Using ROM Data

The values at the probed points are shown in Figure 5.29. The value for index patient, student
5, is represented on the left axis, while the data for the other are represented on the right axis.
We see that the infectivity field is firstly transported into the frontal area of student 4,6 and 8. As

103
shown in Figure 5.29, the passive scalar signal in front of student 4, 6 and 8 grows around 100s.
As time goes further beyond 250s, the values for other 5 students start to grow. It is found that the
magnitude of C for students 4, 6 and 8 are greater than for the others. Therefore, we define the
locations of students 4, 6 and 8 as ‘high-risk’ zones where infections are more likely. Interestingly,
the signals at these three points fluctuate some time after the source signal fluctuates. However,
when the peak magnitude of coughing is not severe, e.g., with peak value of 200 at ∼ t=2300s, the
fluctuation is not reflected in the signals in front of students 4, 6 and 8. Compare with that, the
other five signal sources will not fluctuate with the fluctuation of the signal source, and the average
value is relatively small. Therefore, we define the positions of five students other than students 4,
6, and 8 as low-risk locations.

300 1.6
student 5 (Index patient) - row 2 column 2
student 1 - row 1 colunm 1
1.4
student 2 - row 1 column 2
250 student 3 - row 1 column 3
student 4 - row 2 column 1
student 6 - row 2 column 3 1.2
student 7 - row 3 column 1
200 student 8 - row 3 column 2

Passive Scalar
student 9 - row 3 column 3 1
C

150 0.8

0.6
100

0.4

50
0.2

0 0
0 500 1000 1500 2000 2500
Time/s

Figure 5.29: Concentrations of infectivity field C at probe points in front of students under the
input signal of 5.28. Left axis data for student 5; Right axis-data for the other students.

Among the entire classroom, the passive scalar field mainly transported in the positions of
frontal box of student 4 and 8 as indicated by the probed points’ signals. We further illustrate
that in Fig. 5.30. The iso-surface of the passive scalar with iso-value of 0.8 shows that the area
of aerosol cloud distributes along the outflow of ventilation systems. The airflow of ventilation

104
transports the aerosol from student 5 to the location of student 4, 6 and 8 respectively.

(a) (b)

Figure 5.30: At t=2000s, (a) Iso-surface of passive scalar with iso-value of 0.8; (b) Cross-section
of C.

We integrate the passive scalar concentration over volume of box in front of each student at
t=2000s. The box is defined as a rectangular box in front of students as shown in Figure 5.31 in
pink andthought of as an approximation of the zone of air inhaled by each student.

Figure 5.31: Integration area in front of students.

As indicated by the data of integration shown in Table 5.7 and chart in Fig. 5.32, amounts of
infectious aerosol particles enhaled by students 4,6, and 8 are relatively high.

105
Table 5.7: Volume integration of passive scalar concentration over boxes in front of students at
t=2000s

Student 1 2 3 4 5

Volume Integration 6.61E-02 9.38E-02 6.20E-02 1.25E-01 4.25E+00

Student 6 7 8 9

Volume Integration 1.33E-01 5.18E-02 1.50E-01 6.99E-02

Figure 5.32: The data for student 5 (the source) are excluded.

5.4 Conclusion

In the current study, the reduced order modeling using Krylov subspace methods is discussed
which allows us to simulate the transport of aerosol particles as passive scalar in a ventilation
indoor environments. The Arnoldi algorithm-based SIMO algorithm and an Arnoldi based variant,

106
frequency-shift algorithm, are evaluated.
The aerosol infectivity field is simulated using the Eulerian approach as the concentration of
the passive scalar, transport of which is determined by an advection-diffusion problem with con-
stant background turbulent velocity fields. Through comparative study between the Eulerian and
Lagrangian approaches, it has been found that the predictions of aerosol particles’ transmission are
very similar between the two approaches.
To the best of our knowledge, this study is the first, in which the Krylov-based reduced-order
modeling approach is applied to indoor ventilation system. Different Arnoldi-based algorithms
constructing the reduced order modeling are applied and evaluated. The SIMO algorithm that has
record of success in the analysis of the pollutant transmission with external turbulent flow [29] and
conjugate heat transfer in battery thermal management system [61] is evaluated first. We further
develop a frequency-shift Arnoldi algorithm to improve the performance of ROM. We find that the
frequency-shift Arnoldi algorithm substantially improves the accuracy of the approximation. The
frequency-shift Arnoldi algorithm can capture the high and low frequency response of the system
depending on the selection of the frequency points. Additionally, a discretized time derivative
method is introduced to indirectly access the system matrix A which is normally inaccessible from
the commercial CFD software.
According to the comparative analysis and numerical experiments, ROM can greatly save the
computational cost required for the analysis of transient processes. Once the ROM is built, the
user can perform extensive numerical exploration through the ROM, i.e., apply different transient
exhaling scenarios (breathing, coughing, singing etc) to the model. A full-order model requires
a supercomputer to utilize parallel computing for several days of computational tasks, while the
ROM only takes a few seconds to get results on a single core to predict the spread of aerosols. It has
been found that the simulation of aerosol transmission using ROM can accelerate the simulation
speed up to 4800 times in comparison with the FOM simulation.
Admittedly, some unphysical fluctuations of scalar are found in the fields produced by ROMs.
We identify the fluctuations as the limitation of the ROMs based on the existing algorithms. The

107
limitation comes from the FOM system’s characteristics, i.e., indoor ventilation system with cir-
culation flow, where the transport is dominated by convection.
Finally, the quantitative studies with parametric simulation can be achieved by ROM. The time-
varying passive scalar input signals can be presumed and applied to optimized ROM in order to
have large number of quantitative studies within a short period. The proposed approaches can
potentially be applied to a broad range of scenarios, including a cabin of a commercial airplane,
a movie theater, or a library. The efficient yet accurate predictions of aerosol transmission can be
simulated by ROMs. The price one pays is only the computation effort in generating the ROM.
A large number of rapid results from ROMs can effectively help policy makers or designer to
predict/analyze the aerosol transmission of the aerosol particle. The high-risk areas in public places
due to aerosol transmission can be identified. As a result, public places can be arranged and
designed in advance, thereby reducing the risk of aerosol transmission in public places.

108
 CHAPTER 6

Conclusion and Future Works

Large-scale linear systems are often used to predict and simulate the behavior of a system
given various sets of parameters. In this dissertation, the work focus is on developing an efficient
reduced-order model that can approximate the original model and allow one to accurately pre-
dict its dynamics at greatly reduced computation effort. Applications in different areas show the
practical potential of reduced-order modeling. For example, the method can be applied to online
predictive control of the BTMS as presented in Chapter 4 and to prediction of aerosol transmission
in indoor environments under various exhaling behavior for a long period of time as presented
in Chapter 5. Various techniques have been proposed in the study of reduced-order modeling,
which was reviewed in Chapter 1. However, the system matrices are very large and, in general,
unavailable from the commercial CFD software, which makes most of the reduced-order modeling
methodologies inapplicable from the perspectives of unaffordable computation cost or require of
access to the system matrices.
In the course of the dissertation, several algorithms for generation of reduced-order models of
large-scale linear systems are presented in Chapter 3. This includes the single-input algorithm,
multiple-input algorithm, frequency-shift algorithm, and parametric reduced order modeling algo-
rithm. The algorithms are tested in application to modeling conjugate heat transfer in a battery
pack in Chapter 4. Direct ROM-FOM comparison shows high accuracy at the computation costs
reduced by many orders of magnitude.
As presented in Chapter 4, further investigation based on the Arnoldi-type algorithm has been

109
presented. The multi-input multi-output(MIMO) is developed in order to extend the applicability
of the Arnoldi-type algorithm to multi-input scenarios. The parametric reduced-order modeling
algorithm shows the potential of adapting and interpolating the pre-computed local ROMs to new
parametric ROMs that allow the reduction of the large-scale linear system within a range of the
parameter. Last but least, change of variable and boundary conditions as input methods have been
introduced and examined in a battery pack.
As presented in Chapter 5, a comparative study demonstrates the quantitatively similar result
that can be achieved between the Eulerian and Lagrangian frameworks. The single-input Arnoldi
algorithm is examined in the modeling of airborne transmission of respiratory infections. The
frequency-shift Arnoldi algorithm is developed and shows proof of success in accurately simulating
the airborne transmission of respiratory infection.
Our last comment concerns the range of applicability of the proposed method. Its accuracy and
efficiency have been proven in this work only for the cases of conjugate heat transfer in a battery
pack and the prediction of airborne transmission of respiratory infection in indoor environments.
We do not, however, foresee any major obstacles precluding the use of the method for other situ-
ations in a much wide range of diffusive and convective transport phenomena in technology and
nature.
Several topics are found during the course of the dissertation that are worth further investigation
including:

• Investigation of the parametric reduced-order modeling for systems with time-dependent pa-
rameters (such as, e.g., mass flow rate of the inlet for battery thermal management systems);

• Investigation of the stability of the ROM generated by Arnoldi-type Krylov method;

• Development of ROM for indoor environments with different scenarios, including MIMO
system, parametric system and time-varying systems.

110
 BIBLIOGRAPHY

[1] M. Abuhegazy, K. Talaat, O. Anderoglu, and S. V. Poroseva. Numerical investiga-

tion of aerosol transport in a classroom with relevance to covid-19. Physics of Fluids,
32(10):103311, 2020.

[2] T. M. Bandhauer, S. Garimella, and T. F. Fuller. A critical review of thermal issues in lithium-
ion batteries. J. Electrochem. Soc., 158(3):R1–R25, 2011.

[3] S. Al Hallaj, H. Maleki, J. S. Hong, and J. R. Selman. Thermal modeling and design consid-
erations of lithium-ion batteries. J. Power Sources, 83(1-2):1–8, 1999.

[4] L. H. Saw, Y. Ye, A. A. O. Tay, W. T. Chong, S. H. Kuan, and M. C. Yew. Computational

fluid dynamic and thermal analysis of lithium-ion battery pack with air cooling. Appl. Energy,
177:783–792, 2016.

[5] S. Basu, K. S. Hariharan, S. M. Kolake, T. Song, D. K. Sohn, and T. Yeo. Coupled electro-
chemical thermal modelling of a novel li-ion battery pack thermal management system. Appl.
Energy, 181:1–13, 2016.

[6] E. Meyerowitz and A. Richterman. SARS-CoV-2 Transmission and Prevention in the Era of
the Delta Variant. Infect Dis Clin North Am., 2021.

[7] R. Wang, J. Chen, and G.-W. Wei. Mechanisms of SARS-CoV-2 evolution revealing vaccine-
resistant mutations in Europe and America. The journal of physical chemistry letters,
12(49):11850–11857, 2021.

[8] J. Park and G. Kim. Risk of CoViD-19 infection in public transportation: the development of
a model. International journal of environmental research and public health, 18(23):12790,
2021.

[9] J. Pauser, C. Schwarz, J. Morgan, J. Jantsch, and M. Brem. SARS-CoV-2 transmission during
an indoor professional sporting event. Scientific Reports, 11(1):1–5, 2021.

[10] V. Vuorinen, M. Aarnio, M. Alava, V. Alopaeus, N. Atanasova, M. Auvinen, N. Balasub-

ramanian, H. Bordbar, P. Erästö, R. Grande, et al. Modelling aerosol transport and virus
exposure with numerical simulations in relation to SARS-CoV-2 transmission by inhalation
indoors. Saf. Sci., 130:104866, 2020.

[11] W. W. Nazaroff. Indoor aerosol science aspects of SARS-CoV-2 transmission. Indoor air,
32(1):e12970, 2022.

111
[12] M. P. Wan, G. N. Sze To, C. Y. H. Chao, L. Fang, and A. Melikov. Modeling the fate of
expiratory aerosols and the associated infection risk in an aircraft cabin environment. Aerosol
Sci. Technol., 43(4):322–343, 2009.

[13] G. Bagheri, O. Schlenczek, L. Turco, B. Thiede, K. Stieger, J.-M. Kosub, M. L. Pöhlker,

C. Pöhlker, J. Moláček, S. Scheithauer, et al. Exhaled particles from nanometre to millimetre
and their origin in the human respiratory tract. medRxiv, 2021.

[14] Z. Wang, E. R. Galea, A. Grandison, J. Ewer, and F. Jia. A coupled computational fluid
dynamics and wells-riley model to predict covid-19 infection probability for passengers on
long-distance trains. Safety science, 147:105572, 2022.

[15] D. S. Thatiparti, U. Ghia, and K. R. Mead. Computational fluid dynamics study on the
influence of an alternate ventilation configuration on the possible flow path of infectious
cough aerosols in a mock airborne infection isolation room. Sci. Technol. Built Environ.,
23(2):355–366, 2017.

[16] C. Xu, W. Liu, X. Luo, X. Huang, and P. V. Nielsen. Prediction and control of aerosol
transmission of SARS-CoV-2 in ventilated context: from source to receptor. Sustainable
cities and society, 76:103416, 2022.

[17] M. Muratori, M. Canova, Y. Guezennec, and G. Rizzoni. A reduced-order model for the
thermal dynamics of Li-ion battery cells. In Proc. 6th IFAC Symp. Adv. Automotive Control,
volume 43, pages 192–197, 2010.

[18] F. He, A. A. Ams, Y. Roosien, W. Tao, B. Geist, and K. Singh. Reduced-order thermal model-
ing of liquid-cooled lithium-ion battery pack for evs and hevs. In 2017 IEEE Transportation
Electrification Conference and Expo (ITEC), pages 507–511, 2017.

[19] F. He and L. Ma. Thermal management of batteries employing active temperature control
and reciprocating cooling flow. Int. J. Heat Mass Transf., 83:164–172, 2015.

[20] X. Hu, S. Asgari, I. Yavuz, S. Stanton, C. C. Hsu, Z. Shi, B. Wang, and H. K. Chu. A transient
reduced order model for battery thermal management based on singular value decomposition.
In 2014 IEEE Energy Conversion Congress and Exposition (ECCE), pages 3971–3976. IEEE,
2014.

[21] X. Hu, S. Krishnaswamy, S. Asgari, and S. Stanton. An efficient transient thermal model
for electronics thermal management based on singular value decomposition. In 2015 31st
Thermal Measurement, Modeling Management Symposium (SEMI-THERM), pages 280–286.
IEEE, 2015.

[22] A. C. Antoulas and D. C. Sorensen. Approximation of large-scale dynamical systems: an

overview. Int. J. Appl. Math. Comput. Sci., 11(5):1093–1121, 2001.

[23] A. C. Antoulas, D. C. Sorensen, and S. Gugercin. A survey of model reduction methods for
large-scale systems. Contemp. Math., 280:193–219, 2001.

112
[24] U. Baur, P. Benner, and L. Feng. Model order reduction for linear and nonlinear systems: a
system-theoretic perspective. Arch. Comput. Methods Eng., 21(4):331–358, 2014.

[25] P. Benner, V. Mehrmann, and D. C. Sorensen. Dimension reduction of large-scale systems,

volume 45 of Lect. Notes in Comp. Sci. Eng. Springer, 2005.

[26] R. W. Freund. Model reduction methods based on Krylov subspaces. Acta Numerica, 12:267–
319, 2003.

[27] B. Salimbahrami, B. Lohmann, T. Bechtold, and J. G. Korvink. Two-sided arnoldi algorithm

and its application in order reduction of mems. In Proc. 4th MATHMOD, pages 1021–1028,
2003.

[28] Y. Wang, H. Song, and K. Pant. A reduced-order model for whole-chip thermal analysis of
microfluidic lab-on-a-chip systems. Microfluid. Nanofluidics, 16(1-2):369–380, 2014.

[29] L. Vervecken, J. Camps, and J. Meyers. Stable reduced-order models for pollutant dispersion
in the built environment. Build. Environ., 92:360–367, 2015.

[30] R. S. Puri and D. Morrey. A Krylov–Arnoldi reduced order modelling framework for efficient,
fully coupled, structural–acoustic optimization. Struct. Multidiscipl. Optim., 43(4):495–517,
2011.

[31] K. Willcox, J. Peraire, and J. White. An Arnoldi approach for generation of reduced-order
models for turbomachinery. Comput. Fluids, 31(3):369–389, 2002.

[32] C. Lanczos. An iteration method for the solution of the eigenvalue problem of linear differ-
ential and integral operators. J. Res. Nat. Bur. Stand., 45(4):255–282, 1950.

[33] Z. Bai, P. Feldmann, and R. W. Freund. Stable and passive reduced-order models based
on partial Padé approximation via the Lanczos process. Numerical Analysis Manuscript,
97(3):10, 1997.

[34] R. W. Freund. Krylov-subspace methods for reduced-order modeling in circuit simulation. J.

Comput. Appl. Math., 123(1-2):395–421, 2000.

[35] ANSYS. Inc. ANSYS FLUENT User’s Guide. Canonsburg, PA, 17.2 edition, 2016.

[36] Siemens PLM Software. STAR-CCM+ User’s Guide. Plano, Texas, 13.04.011-R8 edition,
2019.

[37] J. Cullum and T. Zhang. Two-sided Arnoldi and nonsymmetric Lanczos algorithms. SIAM J.
Matrix Anal. Appl., 24(2):303–319, 2002.

[38] B. Salimbahrami, B. Lohmann, T. Bechtold, and J. G. Korvink. A two-sided Arnoldi algo-

rithm with stopping criterion and MIMO selection procedure. Math. Comput. Model Dyn.
Syst., 11(1):79–93, 2005.

[39] C. C. Wang, K. A. Prather, J. Sznitman, J. L. Jimenez, S. S. Lakdawala, Z. Tufekci, and L. C.

Marr. Airborne transmission of respiratory viruses. Science, 373(6558):eabd9149, 2021.

113
[40] W. Garcia, S. Mendez, B. Fray, and A. Nicolas. Model-based assessment of the risks of viral
transmission in non-confined crowds. Safety Science, 144:105453, 2021.

[41] P. J. Bueno de Mesquita, W. W. Delp, W. R. Chan, W. P. Bahnfleth, and B. C. Singer. Control

of airborne infectious disease in buildings: Evidence and research priorities. Indoor air,
32(1):e12965, 2022.

[42] C.-H. Chien, M.-D. Cheng, P. Im, K. Nawaz, B. Fricke, and A. Armstrong. Characterization
of the indoor near-field aerosol transmission in a model commercial office building. Interna-
tional Communications in Heat and Mass Transfer, 130:105745, 2022.

[43] S. Asadi, N. Bouvier, A. S. Wexler, and W. D. Ristenpart. The coronavirus pandemic and
aerosols: Does covid-19 transmit via expiratory particles? Aerosol Sci. Technol., 45(6):635–
638, 2020.

[44] J. K. Gupta, C. H. Lin, and Q. Chen. Flow dynamics and characterization of a cough. Indoor
air, 19(6):517–525, 2009.

[45] Z. Y. Han, W. G. Weng, and Q. Y. Huang. Characterizations of particle size distribution of

the droplets exhaled by sneeze. J. R. Soc. Interface, 10(88):20130560, 2013.

[46] W. F. Wells. On air-borne infection: study ii. droplets and droplet nuclei. American journal
of Epidemiology, 20(3):611–618, 1934.

[47] L. Bourouiba, E. Dehandschoewercker, and J. W.M. Bush. Violent expiratory events: on

coughing and sneezing. J. Fluid Mech., 745:537–563, 2014.

[48] G. Buonanno, L. Morawska, and L. Stabile. Quantitative assessment of the risk of airborne
transmission of SARS-CoV-2 infection: prospective and retrospective applications. Environ.
Int., 145:106112, 2020.

[49] N. Van Doremalen, T. Bushmaker, D. H. Morris, M. G. Holbrook, A. Gamble, B. N.

Williamson, A. Tamin, J. L. Harcourt, N. J. Thornburg, S. I. Gerber, et al. Aerosol and surface
stability of SARS-CoV-2 as compared with SARS-CoV-1. N. Engl. J. Med., 382(16):1564–
1567, 2020.

[50] N. L. Ribaric, C. Vincent, G. Jonitz, A. Hellinger, and G. Ribaric. Hidden hazards of SARS-
CoV-2 transmission in hospitals: A systematic review. Indoor air, 32(1):e12968, 2022.

[51] S. Shao, D. Zhou, R. He, J. Li, S. Zou, K. Mallery, S. Kumar, S. Yang, and J. Hong. Risk
assessment of airborne transmission of covid-19 by asymptomatic individuals under different
practical settings. Journal of aerosol science, 151:105661, 2021.

[52] O. Zikanov. Essential Computational Fluid Dynamics. John Wiley & Sons, Hoboken, NJ,
second edition, 2019.

[53] M. Goodson, J. Feaster, A. Jones, G. McGowan, L. Agricola, W. Timms, and M. Uddin.

Modeling Transport of SARS-CoV-2 Inside a Charlotte Area Transit System (CATS) Bus.
Fluids, 7(2):80, 2022.

114
[54] A. Bayram and A. Korobenko. Modelling the transport of expelled cough particles using
an Eulerian approach and the variational multiscale method. Atmospheric Environment,
271:118857, 2022.

[55] E. Lee, C. E. Feigley, and J. Khan. An investigation of air inlet velocity in simulating the
dispersion of indoor contaminants via computational fluid dynamics. Annals of occupational
hygiene, 46(8):701–712, 2002.

[56] D. Bernardi, E. Pawlikowski, and J. Newman. A general energy balance for battery systems.
J. Electrochem. Soc., 132(1):5, 1985.

[57] D. D. Gray and A. Giorgini. The validity of the Boussinesq approximation for liquids and
gases. Int. J. Heat Mass Transf., 19(5):545–551, 1976.

[58] C. Y. H. Chao, M. P. Wan, L. Morawska, G. R. Johnson, Z.D. Ristovski, M. Hargreaves,

K. Mengersen, S. Corbett, Y. Li, X. Xie, et al. Characterization of expiration air jets and
droplet size distributions immediately at the mouth opening. Journal of aerosol science,
40(2):122–133, 2009.

[59] Z. Bai. Motivations and realizations of Krylov subspace methods for large sparse linear
systems. J. Comput. Appl. Math., 283:71–78, 2015.

[60] Peuscher, H. and Mohring, J. and Eid, R. and Lohmann, B. Parametric model order reduction
by matrix interpolation. Automatisierungstechnik, 58:475–484, 08 2010.

[61] L. Xiang, C. W. Lee, O. Zikanov, and C.-C. Hsu. Efficient reduced order model for heat
transfer in a battery pack of an electric vehicle. Appl. Therm. Eng., Submitted, 2020.

[62] H. Lundgren, P. Svens, H. Ekström, C. Tengstedt, J. Lindström, M. Behm, and G. Lindbergh.

Thermal management of large-format prismatic lithium-ion battery in PHEV application. J.
Electrochem. Soc., 163:A309, 2016.

115