by Christopher E. Schneider · 2019-03-25 · Abstract Incipient Behaviour of Flashback in a Lean...

Incipient Behaviour of Flashback in a Lean Premixed Swirl Burner

by

Christopher E. Schneider

A thesis submitted in conformity with the requirementsfor the degree of Masters of Applied Science

Graduate Department of Institute for Aerospace StudiesUniversity of Toronto

c© Copyright 2019 by Christopher E. Schneider

Abstract

Incipient Behaviour of Flashback in a Lean Premixed Swirl Burner

Christopher E. Schneider

Masters of Applied Science

Graduate Department of Institute for Aerospace Studies

University of Toronto

2019

The dynamics of flame flashback were studied in a lean premixed swirl burner with central bluff-body.

A range of conditions with varied flow velocities, inlet temperatures, and hydrogen/methane flow rates

were investigated. Intermittent movement of the flame into the feed tube, was found over a range of

conditions, with consistent trends as the system moved from stable operation to complete flashback.

Statistical analysis of chemiluminescence data showed a strong link between characteristic behaviours

of the system, such as the statistical frequency of upstream propagating flame protrusions, and the

magnitude of the flashback, independent of the inlet conditions. Effects of conditions on the flashback

magnitude and abruptness of transition are described. Existing metrics for predicting flashback were

found to be inadequate for describing the observed dynamics. Laser diagnostics revealed only a slight

statistical drop in axial flow velocity upstream of the flame, which appeared to strengthen as the flame

moved upstream.

ii

Contents

1 Introduction 1

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

1.2 Emissions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.3 Swirl Stabilization in Premixed Combustors . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.4 Flashback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.5 Optical Diagnostic Techniques for Reacting Flows . . . . . . . . . . . . . . . . . . . . . . 10

1.6 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2 Apparatus 15

2.1 Combustor Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.2 Test Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2.3 Chemiluminescence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.4 Laser Diagnostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.5 Flame Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3 Results 25

3.1 Evolution of Flame Shape and Dynamically Stable

Intermittent States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3.2 Statistical Qualities of Intermittent Flashback Events . . . . . . . . . . . . . . . . . . . . . 28

3.3 Effects of Operating Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3.4 Flashback Predictors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

3.5 Observations From Laser Diagnostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

4 Conclusion 46

Bibliography 48

iii

List of Tables

2.1 Chemiluminescence test cases preformed at 720 SLPM air flow rate and room temperature

reactants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

iv

List of Figures

1.1 Stability limits in a premixed combustor [5] . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.2 Relationship between equivalence ratio, emissions, and temperature [6] . . . . . . . . . . . 2

1.3 Typical axisymmetric streamlines for a confined swirl burner featuring a central bluff-

body. Flame location, inner and outer recirculation zones are also indicated. . . . . . . . . 5

1.4 Velocity field and flame front showing combustion induced vortex breakdown [39]. . . . . 6

1.5 Representation flow and flame demonstrating critical gradient theory [3]. . . . . . . . . . . 7

1.6 Plot of flame and flow velocities for critical gradient theory. In condition 1 the velocity

gradient falls below the critical value and flashback occurs, in condition 2 the velocity

gradient is at the critical value and the flame remains at the same location, and in condition

3 the velocity gradient is above the critical value so the flame retreats [3, 40]. . . . . . . . 8

1.7 Computer simulation showing flame induced flow reversal in a channel burner, adapted

from [46]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1.8 Propagation of a flame tongue [49]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.9 Computer simulated PIV image pair and crosscorrelation field. The largest peak at a

shift of 13 pixels in the x and 6.5 pixels in the y corresponds to the particle displacement

between frames. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.1 Combustor design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.2 Arrangement of cameras and laser sheet with respect to combustor . . . . . . . . . . . . . 20

2.3 Demonstration of the thresholding algorithm for the processing of chemiluminescence images 23

2.4 Demonstration of the outline generating algorithm for PILF images . . . . . . . . . . . . . 24

3.1 Sequence of chemiluminescence images showing the progression a flame protrusion near

stable conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3.2 Sequence of chemiluminescence images showing the progression a flame protrusion at

higher intensity conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3.3 Di time signal showing many short duration propagations . . . . . . . . . . . . . . . . . . 28

3.4 Di time signal showing regular long duration propagations that fully exit the upstream

section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

3.5 Di time signal showing conditions just before continuous flashback where the flame rarely

exits the upstream section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.6 Rate of intermittent flashback events (f) versus mean peak depth (Dm) across all 451 cases. 29

3.7 Mean event times . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3.8 f versus Dm simulated from the lines of best fit for tf and tb . . . . . . . . . . . . . . . . 31

v

3.9 Dp percentiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

3.10 Effect of fuel composition on Dm at room temperature and 720 SLPM air . . . . . . . . . 32

3.11 Effect of bulk flow velocity on Dm at room temperature and 6.1% methane . . . . . . . . 33

3.12 Effect of preheat on Dm at 720 SLPM air and 44 SLPM methane . . . . . . . . . . . . . . 34

3.13 Effect of fuel composition on Dm at 500 K and 720 SLPM air . . . . . . . . . . . . . . . . 34

3.14 Effect of composition on tf and tb at room temperature and 765 SLPM Air Flow Rate . . 36

3.15 Dm versus Damkohler number showing the effectiveness of critical gradient theory as a

flashback predictor. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

3.16 Dm versus Damkohler correlation showing the effectiveness of the Damkohler correlation

as a flashback predictor. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

3.17 Dm versus hydrogen content. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

3.18 Typical flame propagation during flashback. The background shows axial velocity for ease

of interpretation. Taken from Case 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

3.19 Flame propagation showing broadening of reduced velocity area. The background shows

axial velocity for ease of interpretation. Taken from the Case 4. . . . . . . . . . . . . . . . 41

3.20 Flame propagation showing axial deflection of upstream flow. The background shows

axial velocity for ease of interpretation. Taken from Case 4. . . . . . . . . . . . . . . . . . 42

3.21 Difference between axial velocity at flame tip and mean axial velocity without the flame

for Case 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

3.22 Effect of ~δf on Uyf |~x(t)+~δf during flashback . . . . . . . . . . . . . . . . . . . . . . . . . . 43

3.23 Difference between axial velocity offset by ~δf = [+2.4 ı − 2.4 ] mm, and mean axial

velocity without the flame for Case 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3.24 Difference between axial velocity offset by ~δf = [+2.4 ı − 2.4 ] mm, and mean axial

velocity without the flame for Case 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3.25 Difference between axial velocity at the flame tip and mean axial velocity without the

flame versus axial position of flame tip for Case 3. . . . . . . . . . . . . . . . . . . . . . . 45

3.26 Difference between axial velocity at the flame tip and mean axial velocity without the

flame versus axial position of flame tip for Case 4. . . . . . . . . . . . . . . . . . . . . . . 45

vi

Nomenclature

Acronyms

CIVB Combustion Induced Vortex Breakdown

FWHM Full width Half Maximum

IRZ Inner Recirculation Zone

LIF Laser Induced Fluorescence

NOx Oxides of Nitrogen

ORZ Outer Recirculation Zone

PIV Particle Image Velocimetry

PLIF Planar Laser Induced Fluorescence

PM Particulate Matter

PVC Precessing Vortex Core

S-PIV Stereo Particle Image Velocimetry

SLPM Standard Liters per Minute

Introduced Symbols

~δf Positional Offset From Flame Tip

~x(t) Position of Flame Tip

Di Instantaneous Upstream Depth of Travel

Dm Mean Peak Upstream Depth of Travel

Dp Peak Upstream Depth of Travel

Dr Upstream Depth of Travel at Maximum Retreat

f Statistical Frequency of Flashback Events

tb Mean Time Between Flashback Propagations

tf Mean Time of Flashback Duration

vii

Uy Axial Flow Velocity

Uyf Axial Flow Velocity Influenced By Flame

∆Uyf Difference in Axial Flow Velocity Due to Influence of Flame

Standard Symbols

αFB Flashback Flame Angle

δp Flame Propagation Distance

δq Flame Quenching Distance

M Momentum Flux

φ Swirler Vane Angle

Da Damkohler Number

gc Critical Velocity Gradient

Le Lewis Number

P Pressure

Pe Peclet Number

R Burner Radius

S Swirl Number

SL Laminar Flame Speed

ST Turbulent Flame Speed

T Temperature

u Flow Velocity

viii

Chapter 1

Introduction

1.1 Motivation

Gas turbines are used extensively throughout the fields of aviation, terrestrial power generation, and

marine power generation. Their compact size, ability to quickly ramp power output up or down, and

lower emissions when compared to coal, means that they complement renewable energy sources well,

and demand is expected to remain strong [1]. Tightening regulatory control and shifts in public opin-

ion, however, mean that all combustion based systems must undergo development aimed at reducing

emissions. Lean premixed combustion, which can be implemented in gas turbine combustors, provides

a means of reducing pollutants such as NOx and particulate matter. High hydrogen content fuels offer

a means of reducing both CO2 emissions as well as the products of incomplete combustion [2, 3].

While lean premixed technology has made its way into many terrestrial power generation applications,

they are susceptible to a range of combustion dynamics that do not affect conventional diffusion flame

combustors [3,4]. These dynamics include thermoacostic instabilities, flashback, and blow off. Flashback

is the propagation of a flame front upstream of its designed burning location and is particularly prominent

in lean premixed combustors running on fuels with high hydrogen contents. Since upstream regions are

not necessarily designed to handle the high heat produced by combustion, flashback can lead to damage

of components [3]. Simply designing these components to withstand these conditions does not provide

a solution since upstream burning results in less mixed or even diffusion flames, negating the reasoning

of using a lean premixed combustor in the first place.

Flashback through the boundary layer defines the upper operating limit for the power band of many

practical combustor designs. Despite recent advances in the understanding of this phenomena, no high

fidelity method or online measuring method for when flashback will occur has been developed. Models

developed so far have been tested in only small ranges of conditions, and require detailed information

not necessarily available during operation in off-design conditions. With the narrow margin between

flashback and blow off for some conditions, a mixture setting that gives blow off at one condition, may

give flashback in another, as can be seen in Fig. 1.1. A more robust means of providing warning of

imminent boundary layer flashback is therefore highly desirable.

1

Chapter 1. Introduction 2

Figure 1.1: Stability limits in a premixed combustor [5]

1.2 Emissions

Emission of pollutants is inherent to the burning of hydrocarbon fuels. Common combustion engines

produce carbon dioxide, carbon monoxide, particulate matter (PM), oxides of nitrogen (NOx), and

unburnt hydrocarbon emissions. With the exception of CO2, the production of these pollutants is

strongly influenced by the local equivalence ratio. As shown in Fig. 1.2, lean equivalence ratios provide

the lowest emissions. Since lean premixed burners offer a uniform equivalence ratio that can be tightly

controlled, they provide a means to achieve significant reductions in these emissions [1, 6].

Figure 1.2: Relationship between equivalence ratio, emissions, and temperature [6]

Reductions in NOx emissions is the primary motivating factor for the development of lean premixed

burners. NOx may be formed in an engine through three mechanisms. Under the thermal or Zel’dovich

mechanism, NOx is formed by the reaction of atmospheric gases at high temperatures. Prompt NOx

is generated as a result of hydrocarbon interaction with atmospheric nitrogen, while the nitrous oxide

mechanism generates NOx via a N2O intermediate [7,8]. For non-premixed flames, most NOx is produced

via the thermal mechanism. Under this mechanism, NO is formed in the high temperature regions in the

combustion chamber, while NO2 is formed afterwards from NO. As the total produced NO at combustion


temperatures is far below the equilibrium value, the formation rate is highly dependant on temperature

in a non-linear manner. The high activation energy associated with NO formation, however, means that

the production of thermal NOx can be reduced to negligible amounts if it can be ensured that no regions

of the flow exceed about 1800 K. Lean premixed combustion therefore has the potential to largely avoid

thermal NOx production. This is because the evenly distributed mixture ensures a consistent, lean burn

at all locations along the flame front, resulting in evenly distributed temperatures. Diffusion flames,

on the other hand, burn at the interface between the fuel and the air, leading to hot flames burning

at near stoichiometric equivalence ratios around the fuel injector, while the remaining air far from the

injectors plays no role in combustion and remains at the combustor inlet temperature until mixing with

the combustion products. [7–9]

Concerns over NOx result from both the direct effects of NOx as well as the products of interactions

it has with atmospheric components. Directly, NO2 is known to cause damage to the respiratory and

cardiovascular systems and compounds the effects of other pollutants [10, 11]. N2O, which is emitted

directly alongside NOx as well as formed later by atmospheric interactions, is a potent greenhouse

gas [8, 11]. NOx emissions also react with water in the atmoshpere to cause acid rain that damages

vegetation, soil, and aquatic ecosystems [12]. Interaction of NOx with the ozone cycle is of major

concern. Near ground level, NOx emissions react to form tropospheric ozone. This ozone is known to

cause increased mortality and harm to the respiratory system [13, 14]. Plants are also affected by this,

with studies showing damage caused to forests [15,16] and tens of billions of dollars in annual reduction

of crop yield [17–20]. At higher altitudes this situation plays out in reverse, with NOx contributing

instead to the destruction of ozone in the ozone layer [11].

It is also worth noting that lean premixed combustion has advantages in regards to the emission

of the incomplete products of combustion. Unburnt hydrocarbons and soot both result from improper

local equivalence ratios, with both being produced in regions of rich burning, while areas of overly lean

burning can result in the production of unburnt hydrocarbons due to partial oxidization of the fuel [8].

By maintaining a lean, even distributed mixture, lean premixed combustors can avoid these problems.

Unburnt hydrocarbons, along with NOx, are the cause of photochemical smog, the exception being

methane, a major component of natural gas, which causes no smog but is a potent greenhouse gas [21].

Soot, a source of particulate matter (PM), has been shown to reduce life expectancy due to damage

caused to the respiratory and cardiovascular systems [22–26].

Interest has been shown in the use of hydrogen containing fuels to help reduce emissions. These

fuels can be generated from traditional coal and hydrocarbon fuel sources by pre-combustion carbon

capture techniques [2, 3, 27]. Since it contains no carbon, the combustion of hydrogen is not associated

with the production of any CO2, CO, or unburnt hydrocarbons regardless of the type of burner used.

However, as the components of NOx are present in the atmosphere, hydrogen combustion does not

remove the thermal path for NOx production. Furthermore, the combustion of high hydrogen fuels

presents increased challenges in some areas such as a notably increased susceptibility to flashback in lean

premixed burners due to their high laminar flame speeds and small quenching distances [28].

1.3 Swirl Stabilization in Premixed Combustors

Due to the nature of premixed combustion, it is necessary to manipulate the flow in some manner at

the designed burning location in order to achieve stable combustion. To prevent flashback, the local


axial flow velocity upstream of the flame location must exceed the local flame speed, while there must

be some region immediately behind the flame where the local flame speed constantly exceeds the local

axial velocity in order to prevent blowoff. Most modern gas turbine combustors employ the use of a swirl

burner, normally combined with a dump chamber to induce flow recirculation via vortex breakdown,

due to the superior flame holding and mixing properties of this configuration [4, 9, 29].

In order to stabilize a flame in the combustor, a zone of recirculation is normally generated. This

is achieved with steps or dump chambers by taking advantage of the flow’s momentum over a sudden

increase in cross-sectional area to create flow separation. In swirl burners, centripetal forces are used

instead to create an adverse pressure gradient via decaying radial pressure gradients, which generates

a central flow reversal. By locating the flame at the boundary of a pocket of recirculation, the balance

between flame speed and flow speed required for stable combustion can be achieved. Hot products

trapped in the recirculation zone are able to provide an ignition source to the reactants, while shear

between the recirculation bubble and the rest of the flow allow enough new reactants to enter the

recirculation bubble so as to allow burning to continue [30]. The flame in a typical swirl burner is

contained in the inner recirculation zone (IRZ, see Fig. 1.3), and can spread to the outer recirculation

zone (ORZ) at high power settings [31].

In swirl combustors, rotation about the direction of bulk flow is used to generate the forces necessary

to form a central recirculation zone via the process of vortex breakdown. Rotation of the flow is induced

either by the use of axial vanes, or by tangential entry of fluid into a mixing chamber upstream of the

burning location. This flow is characterized by the swirl number, S, defined by the equation:

S =Mtangential

MaxialRhydraulic(1.1)

where Mtangential is the tangential momentum flux, Maxial is the axial momentum flux, and Rhydraulic

is the hydraulic radius at the exit plane of the mixing section into the combustion chamber. For the

specific case of axial swirlers, the equation can be written as:

S =2

3

(1− (Ri

Ro)3

1− (Ri

Ro)2

)tan(φ) (1.2)

where Ri is the inner radius of the annulus, Ro is the outer radius, and φ is the angle between the

swirler vanes and the axial direction. The characteristic qualities of a swirl burner typically develop in

the range of 0.4 < S < 0.6, and further strengthen with increasing swirl numbers [4, 32, 33]. Despite

its near universal use, swirl number does not fully describe the rotation of the flow. Particularly, flows

with the same swirl number, but different radial distributions of tangential velocity, such as free vortex

flow (constant tangential flow velocity) and rigid body flow (tangential flow velocity is proportional to

radius), can exhibit different charateristics. This leads to different critical swirl numbers for certain

transitions being given different values for axial vane and tangential entry swirlers [33].

As the flow rotates upstream of the burning location, centripetal forces result in the formation of

a radial pressure gradient. The rotation of the flow decays over the length of the upstream section by

means of one or more of friction with the wall, the use of a diffusor, or the opening of the upstream

section into a dump chamber. As a result of the decaying rotation, the radial pressure gradient also

decays, creating an adverse pressure gradient near the center of the flow. When the adverse pressure


gradient reaches a strength greater than the forward momentum of the flow, the flow along the center of

the upstream section reverses. The phenomena of this central flow reversal is termed vortex breakdown.

One of three shapes for the recirculation bubble may arise: toroidal (axisymmetric), spiral (helical), or

double helix [32,33]. The center of the vortex often does not line up with the centerline of the upstream

section, but rather is offset slightly and the center of the vortex rotates around the centerline of the

upstream section at a particular frequency in what is called the precessing vortex core (PVC). The

addition of a central bluff-body in the upstream section is common in modern gas turbine combustors,

terminating at the entrance to the dump chamber. This helps to increase the size of the recirculation

zone [30,33].

Figure 1.3: Typical axisymmetric streamlines for a confined swirl burner featuring a central bluff-body.Flame location, inner and outer recirculation zones are also indicated.

1.4 Flashback

Since, unlike diffusion flame based combustors, a combustable mixture exists upstream of the burning

point in lean premixed combustors, it is possible for the flame front to advance from the designed burning

location. Four different mechanisms for the propagation of flashback have been recognized [3]. Core flow

flashback is the trivial case that occurs when the laminar or tubulent flame speed in the mixture exceeds

the flow rate, resulting in the upstream propagation of the flame front through the entirety of the flow.

Thermoacustic combustion instabilities may also induce flashback. These instabilities are capable of

generating large pressure waves in the combustor, at times resulting in reduced flow velocities or even

pockets of back-flow upstream of the combustion chamber. When these pockets include flame, flashback

occurs. This was the first form of flashback observed with flow reversals [34]. Combustion induced vortex

breakdown (CIVB) is caused by aggravation of the existing vortex breakdown in swirl burners, resulting

in the upstream movement of bubbles of recirculating flow [35]. Boundary layer flashback occurs when

the flame front propagates upstream in the area of reduced flow velocity that exists along the wall of

the upstream section.

CIVB is typically seen in lean premixed swirl combustors that lack a centerbody, though cases of

CIVB has been observed in combustors with centerbodies [36]. The balance between the vorticity gener-

ated by baroclinic torque and volumetric expansion provides the driving mechanism for this phenomena.


When the flame front approaches the stagnation point, it generates negative azimuthal vorticity through

baroclinic torque. Countering this is positive azimuthal vorticity generated by the volumetric expansion

caused by heat release. If the flame front moves ahead of the stagnation point, the volumetric expansion

will dominate and stabilize the flame. However, if the flame front remains slightly behind the stagnation

point, then baroclinc torque may dominate. Figure 1.4 shows this scenario, where flashback is caused

when the induced vorticity further reduces the axial flow velocity at the center of the vortex bubble,

pulling it upstream. [37,38]

Figure 1.4: Velocity field and flame front showing combustion induced vortex breakdown [39].

The classical mechanism for boundary layer flashback is the critical gradient theory originally devel-

oped by Lewis and von Elbe [40]. A representation of this theory is shown in Fig. 1.5 In its original

form, this theory stated that the occurrence of boundary layer flashback was governed by three factors:

the laminar flame speed, the quenching distance of the flame, and the radial gradient of axial velocity

at the wall. The theory states that, while the bulk flow speed is normally higher than the laminar flame

speed, the local flow velocity reduces toward zero at the wall. Therefore, it may still be possible for the

flame to propagate upstream through the boundary layer where the laminar flame speed does exceed

the local flow velocity.

The trivial case created by this, where flow velocity is always zero at the wall, so the flame speed

should always exceed the local flow velocity at this location, is prevented by quenching. Quenching

occurs as a result of a combination of the cooling effect of the wall on the flame, and the absorption of

the radical species that would normally cause the continuation of the chain reaction by the wall. As a

result, the flame speed begins to drop off at a certain distance from the wall, reaching zero and being

completely unable to propagate before it reaches the wall. Therefore, for flashback to occur, the flame

speed must exceed the local flow velocity at a particular distance from the wall near the flame speed

drop off value as shown in Fig 1.6.

To predict the local flow speed at this location, the assumption was made that velocity increases


linearly with the distance from the wall in this region, allowing the local velocity at this point to be

predicted from the velocity gradient and quenching distance. Hence, flashback will occur if the velocity

gradient drops below a critical value given by

gc =SLδp

(1.3)

where SL is the laminar flame speed and δp is the propagation distance. δp is at the point between the

quenching distance, δq, and the point where laminar flame speed starts to drop off where the gradient

in laminar flame speed is equal to the laminar flame speed divided by the distance to the wall as seen

in Fig. 1.6.

Models exist to easily calculate flame speed and the distances associated with quenching for a given

mixture. Gradients can be easily computed from a combination of volumetric flow rates and either

Poiseuille flow equations for simple cases, or computer simulations for more complex scenarios. As a

result, this formula can be easily implemented. The equation can also be written for several configurations

in non-dimensional form with flame and flow Peclet numbers [41]. While this theory was originally

developed only for laminar flames, it was later applied to turbulent as well [42–44].

Figure 1.5: Representation flow and flame demonstrating critical gradient theory [3].

More recent studies have shown that critical gradient theory fails to accurately describe boundary

layer flashback, due to the lack of consideration for flame flow interactions. Pseudo-2D channel flow

experiments with diffusers [28] and various flame confinement configurations [45] showed significant

differences to the established flashback limits for unconfined tube flames. Due to the nature of critical

gradient theory, boundary layer flashback should depend only on the laminar sublayer. However, these

studies showed that it can be induced by adverse pressure gradients. A study by Eichler and Sattelmayer

[46] using optical diagnostic techniques revealed the existence of a number of flame cusps at the leading

edge of turbulent flashback events. Each flame cusp advanced, maintained position, or retreated and

broke up, with new flame cusps constantly being created, broadening, and then splitting into more

cusps. When enough of the flame cusps were able to advance, the flame front as a whole advanced.

Ahead of the flame cusps, it was seen that the flow actually reversed, with the flame front being carried

forward in a small zone of recirculation similar to flashback induced by CIVB or themoacostics. Laminar

flashback events were observed to occur in a similar manner, however, with much smoother bulges. It

was confirmed numerically using URANS simulations (see Fig. 1.7), that the cause of this reversal was

the influence of the pressure gain across the flame front on the upstream flow [46]. Further study of this

phenomena using DNS showed that the flame cusps in channel flow were formed by a combination of


Figure 1.6: Plot of flame and flow velocities for critical gradient theory. In condition 1 the velocitygradient falls below the critical value and flashback occurs, in condition 2 the velocity gradient is at thecritical value and the flame remains at the same location, and in condition 3 the velocity gradient isabove the critical value so the flame retreats [3, 40].

turbulent streaks in the boundary layer and the Darrieus-Landau instability. However, when the cusps

were fully developed, the role of the Darrieus-Landau instability became negligable [47].

Figure 1.7: Computer simulation showing flame induced flow reversal in a channel burner, adaptedfrom [46].

A study on the initiation of boundary layer flashback in a channel displayed a non-obvious mechanism

[48]. Rather than the flame advancing from the lip of the tunnel, the flame began its advance from a

fold in the flame front further back. The folded flame front then advanced parallel to the main flame

front until it entered the channel, leaving a gap of unburnt fuel between the two flames as it did.

Observations by Ebi and Clemens [49] in a swirl burner showed some differences in the mechanism

when compared to flashback in a channel. Boundary layer flashback occurring along the centerbody

exhibited two different types of protrusions of the flame front. Small protrusions along the windward

side of the flame were labeled as bulges where as the larger protrusions were labeled as flame tongues.


The bulges resulted from small pockets of complete flow reversal and were unable to maintain this

reversal long enough to contribute to the upstream propagation of the flame. Instead, the flame front

was advanced forward by means of the flame tongues. Flow ahead of the flame tongues did not reverse

in the azimuthal direction, but was deflected so as to cause a local region of axial flow reversal (see

Fig. 1.8). This pulled the flame upstream while it rotated around the combustor in the direction of

swirl. Unlike the other observed cases, no recirculation occurred in this mechanism.

Figure 1.8: Propagation of a flame tongue [49].

In another study on flashback in a swirl burner with centerbody by Heeger et al. [50], small protrusions

were noted to continuously propagate and retreat near the lip of the centerbody. Labeled flame tails,

these protrusions are similar in shape to the flame tongues observed by Ebi and Clemens. However, at

conditions closer to flashback these flame tails changed to a flame continuously wrapped around the lip

of the centerbody. They tied this phenomena to the influence of the PVC in the combustor, suggesting

possible connection to the CIVB phenomena.

Further interaction between CIVB and boundary layer flashback can be seen in studies conducted by

Sangl et al. [51] and Mayer et al. [52] where the system exhibited both mechanisms. A swirl burner was

used consisting of a converging upstream section without centerbody, followed by a dump chamber. At

the initiation of flashback they observed CIVB, however once the flame proceeded far enough upstream,

the flame front moved onto the walls of the upstream section and the flashback was completed by the

boundary layer flashback mechanism.

Using the improved understanding of boundary layer flashback, several attempts have recently been

made to create a model for flashback prediction. Kalantari et al. [53] recognized that while critical

gradient theory does not correctly describe the processes occurring in flashback, it does contain many

parameters still relevant to flashback. They proposed a Damkohler number based on propagation depth,

as defined by critical gradient theory, divided by flame thickness as a predictor. Instead of calculating

this directly, a correlation based on Buckingham Pi theory was used and the Damkohler correlation was

given as

DaDC = Const · Le1.68 · Pe1.91f ·(TuT0

)2.57

·(TtipT0

)−0.49

·(PuP0

)−2.1

(1.4)

where Le is the Lewis number, Pef is the flame Peclet number, Tu is the unburnt mixture temperature,

Ttip is the surface temperature at the point of flame anchoring, Pu is the upstream pressure, and T0 and

P0 are the reference temperature and pressure respectively.


For confined flows it has been proposed that flow reversal, and therefore flashback, should occur when

the pressure rise generated across the flame front is high enough to induce boundary layer separation [54].

For unconfined channel flows it was proposed that the initiation mechanism for flashback should be the

limiting factor. Based on the previous study showing that flashback starts at a point back from the

burner’s edge, and that it forms a flame cone around the burner exit [48], it was proposed that flashback

occurs when the local turbulent flame speed exceeds the component of the local flow speed acting

perpendicular to the flame surface [55]. This gives the equation

St(yFB) = u(yFB) sinαFB (1.5)

where St(yFB) is the turbulent flame speed at the location of flashback, u(yFB) is the axial flow velocity

at the location of flashback, and αFB is the angle between flame cone and axial direction at which

flashback occurs. The flame angle is assumed to be equal across the entire flame cone, meaning that

average properties determine the flame angle which is given by

αFB = sin−1

(St

UFB

)(1.6)

A study testing and comparing flame angle theory and Damkohler correlation at different condi-

tions [56] lead to the development of a modified Damkohler correlation. Flame angle theory showed

it could make good predictions, but only as good as the data being fed into it, which can be difficult

to obtain accurately for all conditions. While the Damkohler correlation needed less information to

predict flashback than flame angle theory, it could not create accurate predictions over a wide range of

equivalence ratios due to its failure to account for flame stretching. To compensate for this, the modified

Damkohler correlation introduces an empirically determined factor that must be determined for each

fuel. The modified Damkohler correlation is then given as:

DaMDC = C(φ) · Le · Pe2f ·(TuTref

)0.47

·(PuPref

)−C4

(1.7)

where C(φ) and C4 are empirically determined coefficients, and Tref and Pref are reference temperature

and reference pressure respectively.

1.5 Optical Diagnostic Techniques for Reacting Flows

Understanding the dynamics of reacting flows requires being able to characterize both the flame and

flow simultaneously. Optical diagnostic techniques provide a way to achieve this. Unlike probes, optical

diagnostics are non-intrusive and will not alter how the system behaves. Measurement types and locations

can be easily changed without impacting on the measured system. Furthermore, optical diagnostics can

allow for large amounts of data to be captured concurrently. Many optical diagnostic techniques are

capable of planar or even volumetric sampling, providing the capability to examine multidimensional

phenomena. By filtering for different wavelengths of light, it is possible to overlap different diagnostic

techniques in the same space without interference. Techniques capturing data via lens/sensor systems

can easily scale spatial resolution by changing positioning or the type of lens used in order to achieve

the required resolution. When both high resolution and large areas are required to be captured, the


addition of multiple imaging systems can be used for different areas without interference. Since optical

diagnostics do not have memory effects, high speed measurements can be taken without the effects of

previous conditions limiting temporal resolution.

Particle image velocimetry (PIV) is a means of determining the velocity fields of a fluid. Most of the

information provided in this section about PIV is referenced from Raffel et al. [57] where not otherwise

referenced. In order to be able to visually track the movement of the flow, PIV and several related

techniques, such as particle tracking velocimetry (PTV), seed the flow with tracer particles. So long

as the tracer particles are sufficiently small such that drag forces are capable of changing the particles’

velocity within the length of the flow’s smallest features, then the movement of the particles can be

considered the same as the movement of flow. This is determined by the Stokes number defined as

Stk =τpuflf

(1.8)

where uf is the characteristic flow velocity, lf is the characteristic flow length, and τp is the particle

relaxation time defined by

τp = d2pρp

18µ(1.9)

where dp is the particle diameter, ρp is the particle density, and µ is the fluid’s viscosity. If the Stokes

number is much greater than one, then the particle will pass through the flow structure undisturbed,

while a particle having a Stokes number much less than one will closely follow the movement of the fluid.

Achieving a Stokes number less than 0.1 is generally sufficient to properly resolve the flow behaviour

[57,58].

In PIV, the seed particles are illuminated by a thin plane of light. This is generally created by using

a laser beam that is expanded in one direction, and then passed through a converging lens aligned in the

other direction. The narrowest portion of the laser sheet is called the beam waist and is positioned at the

region of interest. A combination of diffraction and geometric aberrations prevent perfect convergence

of the sheet, resulting in a finite sheet thickness.

Two pulses of light are used in quick succession to illuminate the particles. In older experiments

the resultant signal was often captured on the same frame of film, while in newer experiments high

speed cameras typically record the images on separate frames. Next, the image is divided into a series

of windows. For double exposure of a single frame, an autocorrelation is preformed in each window,

then the window is shifted slightly and an autocorrelation is preformed again. This forms a grid of

autocorrelation values which will feature two peaks, one where the displacement of the frame lines up

with the displacement of the particles, and one where the displacement of the frame lines up with the

reverse of the particles. This leads to directional ambiguity of which way the particles have moved.

Several methods were devised in order to solve this problem, however, the introduction of high speed

digital cameras has largely eliminated their necessity.

With modern digital cameras, the laser pulses can easily be timed for “frame straddling”, where

the first pulse occurs just before the shutter closes and the second pulse occurs just after the shutter

opens, even at kHz frame rates. Since this allows the two particle images to be on different frames,

a cross-correlation is used instead of an autocorrelation, which eliminates directional ambiguity, and

does not require the added complexity of previous directional ambiguity elimination techniques. Finally,

the displacement is divided by the time between pulses, which gives the flow velocity in that window.

Figure 1.9 shows a computer generated example of how this process works.


(a) Frame 1 (b) Frame 2

2030

200

(13,6.5)

100

-10

displacement in the y direction [pixels]

-20-20

displacement in the x direction [pixels]

-30

(c) Crosscorrelation

Figure 1.9: Computer simulated PIV image pair and crosscorrelation field. The largest peak at a shiftof 13 pixels in the x and 6.5 pixels in the y corresponds to the particle displacement between frames.

The timing, velocities, and resolution of the system are all closely tied together in PIV. The longest

time between pulses that can be used is limited by out of plane particle losses. If the displacement

of the particles between pulses in the direction perpendicular to the light sheet exceeds the thickness

of the light sheet, then no particles will appear in both images of the pair, and no PIV data will be

able to be recovered. Therefore, the separation between laser pulses must be small enough that most

particles appear in both images of the pair. Another limit on the length of time between image pairs is

that the distance traveled by the particles between pulses must be small enough to be able to assume

constant linear movement, which means that movements must be smaller in regions of high acceleration.

Finally, the pulses must be far enough apart that the camera can resolve a sufficient movement instance

for the required accuracy. Some extra accuracy can be found by using super resolution PIV. Super

resolution PIV works when the size of the particle’s image on the sensor is greater than the pixel size.

The brightness of each pixel will then be determined by how much of the particle covers it. As a result,

interpolation can be performed, and the image can be shifted by sub-pixel sizes for each cross-correlation,

allowing more accurate determination of velocities.


Spatial resolution is determined by the size of the window used to calculate vectors. Smaller windows

allow more vectors to be measured, but reduce the number of particles per window, and therefore, the

strength of the correlation peak. It is possible to fit in more vectors with a given window size by

overlapping the windows. However, this does not produce higher vector resolution, as the same particle

movements are being measured multiple times. If increased resolution is required, then the camera setup

can be changed to increase magnification, and/or the amount of seeding particles can be increased to

reduce the required window size.

Large local velocities and strong rotational velocity gradients can reduce the signal strength of the

PIV data. This can be remedied by using multi-pass techniques to iteratively compensate for these

factors. The basic reduction in quality occurs when high velocities result in few of the particles that are

in the window during the first frame, are present in the same window during the second frame. Shifting

the location of the window in one or both frames by the originally predicted velocity can increase the

number of particles that match to achieve greater correlation. If a large velocity gradient exists that is

misaligned with the velocity field, such as in rotating flows, then the velocities of some particles will be

different than other particles within the window. As a result, the correlation peak will spread as different

parts of the window correspond to different velocities. By examining the velocity gradients, a skew can

be determined and applied to deform the image within the window between pairs, compensating for

this effect and increasing the signal strength. It can be advantageous to perform the initial passes with

larger window sizes. Larger windows capture more particles and suffer less loss of particles due to high

velocities. Starting with a larger window size can therefore allow more robust determination of velocity,

allowing for the initial pass at the final resolution to be preformed with a predetermined shift that may

even exceed the size of the final window itself.

When using a single camera to capture PIV images, only the velocity components orthogonal to the

camera’s line of site can be observed. This can be solved by stereo particle image velocimetry (S-PIV),

which adds a second camera pointed at the same area from a different angle. Since the cameras see

the particle motion from different angles, all three velocity components can be determined. Another

technique, tomographic particle image velocimetry (T-PIV) is able to extend determination of velocities

to three dimensional volumes by using a thicker region of illumination, additional cameras, and volume

reconstruction algorithms.

Laser induced florescence (LIF) of reacting species for combustion exploits the presence of chemical

intermediates in order to provide information on the reaction. Radical intermediate species are excited

by a laser tuned to their absorption frequency. Excited species then re-emit radiation to return to a more

grounded state. Some of the emitted radiation occurs at a slightly longer wavelength, this radiation is

termed Stokes shifted. It is useful to measure this shifted radiation rather than emissions at the original

frequency as it eliminates the effects of reflection and scattering of the original exciting laser on the

signal. Reemission occurs rapidly, on the order of 10−5 to 10−10 seconds depending on species. This

makes LIF suitable for high speed measurement systems. Many different species can be identified with

LIF including OH, NO, NH, CH2O, CH, and C2. By examining where these species occur, information

about flame structure and topology can be gained. [59]

A variant of LIF, planar laser induced fluorescence (PLIF), performs well with PIV measurements.

In PLIF the laser is formed into a sheet the same way it is in PIV. This allows overlaying the sheets for

simultaneous data gathering in the same space.

Chemiluminescence is a technique using the same basis as LIF, but without external excitation. In


chemiluminescence excitation of radical species is achieved as a natural result of the combustion process.

Emissions can be observed at the same wavelengths as LIF using the same detection equipment, but the

signal is several orders of magnitude weaker, requiring much longer gate times. Since the excitation of

the radical species in chemiluminescence is not limited to locations illuminated by a laser, the resulting

images are line of sight integrated.

1.6 Objectives

The aim of this work is to investigate the phenomena of boundary layer flashback for early warning signs

of the transition from stable to flashed back. To do this, observations will be made at conditions where

the flame enters the upstream section without propagating further upstream. While the phenomena

of the flame entering the upstream section has been noted in even the earliest studies [40], with the

exception of Heeger et al. [50], very little attention has been paid to the system’s behaviour at such

conditions. By studying the flame and flow characteristics in this regime, it is hoped that characteristic

behaviours may be found that precede the full flashing back of the flame. Such information would allow

the creation of safety measures to prevent entering into unstable operating conditions. Particular benefit

from an online measurement system could be had in scenarios where off-design operation is a risk, such as

when fuel can not be guaranteed consistent, or when the combustor must function regardless of changes

to upstream flow such as that caused by fouling or compressor damage.

To achieve these goals, both information about flow velocities and flame locations is required. Chemi-

luminescence was used to gather information on the movement of flames across a large number of con-

ditions, both to provide information on how the system reacts to different inputs in this regime, and

to evaluate the consistency of any characteristic behaviours under different conditions. To evaluate the

underling mechanisms causing the behaviours a smaller number of cases, transiting the regime from just

past the point of stability to just before fully flashing back, will be examined using simultaneous S-PIV

and OH PLIF.

The apparatus, testing conditions, diagnostic setup, and data processing methods are detailed in

Chapter 2. Discussion and analysis of the produced data is covered in Chapter 3, with a summary of

the findings and future areas for study provided in Chapter 4.

Chapter 2

Apparatus

2.1 Combustor Design

6(&7,21�%�%

�

�

�

�

$ $

% %

��

��

��

��

Figure 2.1: Combustor design

A new combustor, shown in Fig. 2.1, was constructed for

the purpose of this study. It consists of three chambers: a

plenum where the fuel air mixture is introduced, a reactant

feed section which simulates the mixing tube in an indus-

trial burner, and a combustion chamber where the flame

burns during stable combustion. Electromechanical mass

flow controllers (Brooks) were used to meter fuel and air.

The air could be preheated using an electric torch (Farnam-

Tutco) just before mixing with the fuel in the feed tubing.

After mixing, the air continued through insulated tubing,

eventually being divided into three streams just before the

plenum. Each stream flowed through equal length tubing

into a manifold that further split each stream into another

three streams. All nine streams then entered the plenum

through equally spaced radial ports. A K-type thermocou-

ple located in one of the manifolds provided the inlet tem-

perature used for setting the preheat.

Flow entering the plenum moved into the reactant feed

section by passing through a perforated conditioning mesh

followed by an eight-bladed axial swirler, which imparted

a swirl number of S = 0.9. The reactant feed section was

an annulus formed by a 50 mm inner diameter quartz tube

and a 25 mm diameter central bluff-body constructed from

316 stainless steel. After 150 mm, the reactant feed section

exited into the combustion chamber with an abrupt step

increasing diameter to 100 mm. The central bluff-body ter-

minated at this location. The combustion chamber’s outer

wall was constructed of quartz tubing with a length of 150 mm, exhausting into the the atmosphere.

15

Chapter 2. Apparatus 16

The use of quartz allowed optical accessibility for both the reactant feed section and the combustion

chamber for visible, near IR, and near UV wavelengths.

The combustor assembly was mounted on a moveable stage with 1 mm per rotation leads operated

by handwheels. This allowed the combustor to be positioned relative to the laser beam paths instead of

adjusting the sheet forming optics required for PIV and PLIF, resulting in easier setup.

The initial set-up of the combustor suffered from diffuse reflections where the laser met the quartz

tube. To solve this problem, the original extruded quartz tubes were flame polished, which reduced the

diffuse reflection. A light shield was used to cover the most serious remaining reflection which occurred

at the beam entry point into the combustor for the back scatter camera. Remaining reflections where

removed by preprocessing in DaVis.

2.2 Test Conditions

It was observed during testing of the combustor that a region of conditions existed where the flame

advanced into the upstream reactant feed section and then retreated back into the combustion chamber

without ever stabilizing in either section. Two sets of experiments were conducted in order to characterize

the flashback behaviour in this region. A series of cases at 451 different conditions using chemilumines-

cence provided information regarding the effects of changing parameters on flame behaviour. A smaller

set of 6 conditions approximately centred within the range of conditions used in the previous set and

covering the full range of flashback intensities were monitored using simultaneous S-PIV/PLIF to provide

information on mechanistic changes seen at different intensities.

In order to collect information about the effects of operating conditions on flashback for the chemi-

luminescence cases, data was taken at various air flow rates, reactant inlet temperatures, equivalence

ratios, and fuel compositions. For each given air flow rate and reactant inlet temperature, experiments

were conducted at six evenly spaced methane flow rates between 5% and 7.8% of the volumetric airflow

rate. At each methane flow rate, hydrogen was added incrementally to the system until the first value at

which flames were observed propagating upstream into the reactant feed section, and data was recorded

for this condition. Then, an additional 2 SLPM of hydrogen was added, data was recorded, and this

process was repeated until flashback reached the point at which the flame burned continuously in the

upstream section without ever retreating to the combustion chamber. Table 2.1 shows this procedure

for the case at room temperature and 720 SLPM air flow. A total of 289 tests were conducted at room

temperature for five different air flow rates ranging from 630 SLPM to 810 SLPM, as well as 162 cases

in three groups of cases at 720 SLPM with the inlet temperature elevated to 400 K, 500 K, and 600 K.

These flow rates and temperatures were chosen based on a combination of equipment limits, maintain-

ing equivalence ratios relevant to industry operation, a desire to achieve sufficient resolution in each

parameter to determine trends in the effects on flashback, and to ensure turbulent conditions.

For the runs to collect data about the flashback mechanism using S-PIV/PLIF, the air flow rate was

set to 720 SLPM, the methane flow rate was set to 48 SLPM, and the reactants were fed into the system

at room temperature. This put the measurements in the middle of the set of conditions used during the

chemiluminescence runs. Hydrogen was run at 20, 24, 28, 30, 32, and 34 SLPM, giving coverage over

the full range of observed flashback intensities with 4 SLPM gaps between the initial cases due to the

limited changes occurring between those conditions.

Initial testing of the limits of flashback for the combustor revealed that heating of the central bluff-


Methane Content 5% 5.6% 6.1% 6.7% 7.2% 7.8%

HydrogenFlow Rate[SLPM]

18 1820 20 20

22 22 22 22 22 2224 24 24 24 24 2426 26 26 26 26 2628 28 28 28 28 2830 30 30 30 30 3032 32 32 32 32 3234 34 34 34 34 3436 36 36 3638 38

Table 2.1: Chemiluminescence test cases preformed at 720 SLPM air flow rate and room temperaturereactants

body could result in the flame propagating into the reactant feed section significantly earlier than when

operating from cold conditions. The most pronounced cases of this occurred just following flashback

events where a relatively high equivalence ratio flame (φ ≈ 0.8) burned continuously in the reactant

feed section. Immediately afterwards, the system would be capable of producing continuous flashback in

conditions where stable burning had previously been observed. In order to ensure test cases produced

consistent results independent of conditions of the previous case, a set of operating procedures were

created to maintain a consistent bluff-body temperature.

Prior to collecting data for any case where a previous case had not been preformed immediately

before, a mock case was run to bring the combustor up to temperature. After each case, fuel was cut

and air flow was maintained for at least three minutes to normalize the temperature on upstream regions

of the central bluff-body. This prevented the possibility of the heated bluff-body causing flashback during

the thermal control procedure. Next, the combustor was warmed for two minutes using a stable flame

at the targeted methane flow rate for the next case and the minimal amount of hydrogen required to

stabilize against blow-off.

To further assist in maintaining consistent temperature on the central bluff-body, an infrared ther-

mometer, designed for through glass measurements of metallic components (Optris CT 3MH1 SF), was

mounted so as to take readings 10 mm below the entrance to the combustion chamber. Temperature

readings were made for all cases at just prior to, and upon completion of data recording. Although

these readings suffered a systematic error from flame interference, they could be compared against each

other for abnormalities. Additional readings were made after the extinguishing of the tested flame and

before the ignition of the warming flame in cases with elevated inlet temperatures. Consistency of these

readings coupled with consistency in the measured results suggest that these procedures were effective

in limiting the influence of bluff-body heating on the results.

2.3 Chemiluminescence

Chemiluminescence was imaged for the OH* radical using a filtered camera/intensifier setup. A high

transmission (>70%), 10 nm FWHM bandpass filter centred at 307 nm spectrally filtered incoming

radiation. The remaining UV light was collected by a 45 mm, f/1.8 Cerco UV lens and focused on

the intensifier. Light from the intensifier was relayed to the high speed camera (Photron SA-Z) via a


macro lens. Since the intensifier photocathode was smaller than the size of the camera’s sensor, the

full resolution of the camera was not available. Data was acquired at a frame rate of 2.5 kHz. This

frame rate was chosen based on the requirements of resolving the smallest scale flame movements, and

reducing statistical noise by capturing the longest possible period of time without significant influence of

centerbody heating. To maximize the signal, gate time was set to 350 µs out of a possible 400 µs. Gain

was selected so as to come as close to saturating the sensor as possible without actually doing so. The

field of view was centred on the upstream section of the combustor, covering the downstream half of the

reactant feed section. Since the edges of the field of view extended horizontally past the edges of the

upstream section, the image was cropped to a 640 × 1024 pixel region. This allowed for the recording

time for each case to be increased to 14 s.

2.4 Laser Diagnostics

Simultaneous measurements of S-PIV and OH PLIF were recorded at 2.5 kHz in the reactant feed section

of the combustor. Laser sheets for both measurements overlapped and traveled tangentially to the central

bluff-body. The top of each laser sheet was positioned just below the plate separating the reactant feed

section and the combustion chamber. The edge of the PIV laser sheet was positioned 0.5 mm away from

the centerbody.

The S-PIV laser sheet was generated from a 532 nm frequency doubled, double pulsed Nd:YAG laser

(Quantronix Hawk-Duo 532-120-M). Both laser oscillators were run at 2.5 kHz with a separation between

pulses of 20 µs. Power was reduced slightly from maximum due to saturation of the camera sensors.

The optical path consisted of a beam expanding telescope followed by two cylindrical lenses to vertically

expand the beam, and finally a cylindrical lens used to converge the beam into a sheet with the beam

waist positioned at the combustor central bluff-body. An iris positioned between the two expanding

cylindrical lenses clipped the top and bottom of the laser sheet in order to provide a more even power

distribution across the height of the laser sheet. Rotational adjustment of the cylindrical lenses ensured

that the vertical axis of the sheet was aligned with the central bluff-body. After the laser sheet was

adjusted for rotation, the moveable stage was used to shift the combustor back 0.5 mm past the point

at which the laser sheet contacts the central bluff-body.

A frequency doubled rhodamine 590 (6G) based dye laser at 283 nm (Sirah Credo Dye) pumped by

a frequency doubled 532 nm Nd:YAG laser was used to provide the PLIF laser sheet. Due to reduced

output and a beam divergence problem with the pump laser, only limited power was available from the

dye laser. In order to compensate, the internal optics of the dye laser were realigned for maximum power

and an evenly distributed beam profile at operating conditions. The beam was expanded vertically using

two cylindrical lenses and converged to a sheet with another. An iris was used to clip the beam between

the two vertically expanding cylindrical lenses. However, due to the reduced power this was more limited

than the clipping done to the S-PIV laser beam. A dichroic mirror was used to overlap and center the

PILF laser sheet onto the S-PIV laser sheet. The two laser sheets were set orthogonal to the optical

axis of the PLIF/chemiluminescence camera/intensifier setup. The dye laser was adjusted through its

frequency spectrum to find the point of maximum flame emissions as seen by the PLIF intensifier/camera

combination. This frequency was found to be at the OH Q1(6) transition of the (1-0) band located at

about 283 nm. Both lasers were synced via a pulse generator (BNC Model 577) and monitored via

photodiodes and an oscilloscope. The PLIF laser was set to fire halfway between the pulses of the S-PIV


laser.

Two high speed CMOS cameras (Photron SA-5) were used to capture S-PIV images, while the image

intensifier/camera combination used for chemiluminescence was carried over to record PLIF images. The

intensifier gate time was 300 ns, synced to the laser pulse using the pulse generator and oscilloscope.

Despite the higher intensity of the PLIF signal compared to chemiluminescence, the shorter exposure

time meant that a high intensifier gain was required. Gain was again selected based on best signal for

each condition. One S-PIV camera was placed on each side of the PLIF camera. Light entering each

S-PIV camera was filtered through a 532 nm bandpass filter, and focused using a macro lens (Tokina,

f = 100 mm, f/# = 2.8). The cameras were angled approximately ∼35 degrees from normal to the

laser plane. In order to align the focus plane with the laser plane, Scheimpflug adapters were used to

tilt the lenses such that the plane of the lens, image sensor, and laser all intersected along a common

verticle line. After tilting, all the camera field of views covered approximately the same region. Since

this arrangement resulted in one camera operating in forward scatter and the other in back scatter, there

were significant differences in signal brightness between the cameras. To compensate, the aperture of the

camera in forward scatter was reduced to f/5.6, while the camera in backscatter was set to f/4. This

allowed both sensors to be operated close to saturation. All cameras were synced to the pulse generator

and monitored on the oscilloscope. The S-PIV cameras were operated at 5 kHz with frame straddling

of the laser pulse, while the PLIF camera/intensifier combination was operated at 2.5 kHz. Each case

spanned slightly over 2 seconds. The experimental setup is shown in Fig. 2.2.

Zirconium dioxide particles of 1 µm diameter were used to seed the flow instead of titanium dioxide

due to its lower tendency to contaminate the tubes. While titanium dioxide is normally preferred to

zirconium dioxide as it has a lower density and therefore produces lower Stokes numbers. As the Stokes

number of these experiments was calculated to be sufficiently low at 0.013 with zirconium dioxide, the

reduced contamination was favoured.

A commercial software package (La Vision DaVis 8.4) was used to process and compute vector fields

from the S-PIV images. The field of view of each S-PIV camera was computed for vector processing

from a reference image of the La Vision 0-58-5 dual plane calibration plate that was positioned on the

laser plane. The plate was also used to align S-PIV images with the PLIF images.

2.5 Flame Processing

In order to quantitatively compare the flashback behaviour between cases, a simplified signal was ex-

tracted from the chemiluminescence data. This signal, Di(t), consisted of the furthest upstream position

of the flame front at each instant in time, determined from binarized versions of the chemiluminescence

images. As flame intensity varied for each condition as a result of changes to fuel composition, equiva-

lence ratio, and reactant temperature, with variation in image intensity further compounded by changes

in gain, an adaptive threshold value had to be determined to binarize each case.

Due to the intermittent nature of flashback in the conditions of interest, the flame was not present

in the region of measurement for each frame, and hence using adaptive thresholding individually on

each frame, which would force some values as white and others as black in each frame regardless of the

presence of a flame, would guarantee errors such as those seen in the center and bottom center images in

Fig. 2.3. Hence, a robust threshold was determined as follows. In any given frame one of two states could

be expected, flame or no flame. Furthermore, in any given recording period one of three combinations of


Figure 2.2: Arrangement of cameras and laser sheet with respect to combustor

states could occur: no flame in any frame, a flame present in every frame, or a mix of frames with flames

and frames without flames. While the conditions examined were specifically chosen because they fit in

the last category, understanding of the algorithm’s reaction to the other two combinations is necessary

as they are approached as limit cases. Finally, for any binarization attempt to produce meaningful data,

there must be no overlap between the image intensity level in places where a flame exists, and the image

intensity level where no flame exists.

Using the above information, the following algorithm was devised. First, a threshold was calculated

for each frame of a case using Otsu’s method [60]. For frames where a flame was present, this gave the

threshold value between the flame and the background such as the top center image in Fig. 2.3, while

for frames without a flame present, it gave a value somewhere within the range of background noise

producing result like the bottom center image in Fig. 2.3. Otsu’s method was then used to calculate an

overall threshold value from the threshold values for each frame. If the original assumptions hold true, the

threshold values when a flame is present should form a closely spaced group, while the threshold values

where no flame is present should form another. As these groups should be separated by a significant

region where the image intensity is higher than the background noise, but lower than the intensity

given by a flame, the threshold of these values will give a value between these two groups that is both

greater than the background noise, but lower than the lowest image intensity given by a point on the

flame. Hence, this algorithm should be able to detect even the smallest flames without causing any false


positives being generated by background noise.

In the limiting case where a flame exists in every single frame, this method may produce a threshold

value that is too high to detect the lowest intensity regions of the lowest intensity flames. This is because

the threshold value for each frame will give a point between background noise and the lowest intensity

flame region. Since there will be no frames with threshold values corresponding solely to background

noise, the overall threshold value will be determined somewhere between the value for the most intense

flames and the least intense flames. This can give false negatives for some pixels on the least intense

flames. Complete failure is not guaranteed by this limit however. Since the flame intensity tends to be

non-uniform, even the loss of detection of some pixels for the least intense flames will not necessarily

result in total non-detection of the entire flame, meaning that, while an error value for Di(t) may occur,

it will not necessarily drop all the way to zero.

Complete avoidance of any error in this limiting case is still possible however. Since the threshold

values will be between the lowest flame intensity and the background noise for each frame, the high

intensity threshold values will not necessarily exceed the lowest flame intensity values of the low intensity

flames. The best conditions to promote this scenario and avoid errors is when a large gap exists between

the background noise and the lowest image intensities from the flame, and a small gap exists between

the lowest image intensities from the flame in the brightest and dullest flames. If the background noise

and lowest intensity flame regions have a clear gap but the spread in flame intensities is too high, it may

be possible to assist the algorithm by applying a gamma correction before processing. This problem was

not found to occur in any of the cases examined, and so no correction attempt was required during this

study.

For the limiting case where no flame enters the upstream section during the test, no binarization will

be possible since the true representation of the flame would be a solely black image. As this limit is

approached, the reduced number of frames containing a flame will result in the threshold of threshold

values eventually picking an overall threshold value within the region of the background noise. A way

to delay the onset of this problem is to suppress the background noise, as the narrower the range of

values found in the background noise, the less attracted the threshold of thresholds will be to it. If the

algorithm must be run in a situation where the possibility exists that no flame ever enters the region of

interest, it maybe possible to still produce an accurate threshold if the field of view is set to include a

small portion of a region where a flame may exist.

The data for all chemiluminescence cases was originally processed as described above. This proved

sufficient for the vast majority of conditions but did struggle with a small set of cases. These cases

consisted of conditions where the flame was present in the upstream section for only a small portion

of the time. While some of the problematic cases were at room temperature, a disproportionally large

number of the problematic cases occurred at high preheat temperatures. The high preheat cases not

only had higher flame luminosity than the rest of the cases, but due to logistical reasons, were conducted

after the PIV cases. As a result, these cases featured a glow around the flame from the flame emissions

being scattered by residual particulate contamination on the quartz tube.

Two adjustments therefore were made to the processing algorithm in order to produce robust results

for all cases. First, the image was cropped to remove a small area at the top where the flame in the

combustion chamber was visible. Allowing the image of the flame to be processed caused the data set to

move into the limiting case of there always being a flame. While the separation between background noise

and flame luminescence was high enough to ensure that the threshold was still good for most conditions,


in those conditions where the flame rarely entered the upstream section, the flame that entered the

upstream section was of notably lower luminous intensity than that in the combustion chamber. Since a

flame did enter the upstream in every measured condition, removing the portion of the image containing

the combustion chamber flame did not move the data to the no flame ever measured limit case.

The second adjustment to the algorithm was to apply a cutoff to reduce background noise. An

average of all pixel intensity levels for all frames in a case was taken. This value was then subtracted

from all values, with any number that would have become negative left at zero. As all cases contained

a large portion of image area where no flame existed, this average value was somewhere between the

average value of the background noise and the image intensity value for the flame. Subtracting the

average therefore did not reduce any areas with flame to zero, but did reduce a substantial portion of

the background to zero. Since the spread of the background values was then reduced, the threshold was

less likely to be selected somewhere in these background values. As an extra precaution against errors,

the binarized images were post-processed with a despeckle filter to remove any areas of less than 20 pixels,

although no cases where this effected the final result were observed in the sample portions reviewed to

check algorithm success. This process was sufficiently robust to accurately perform binarization for all

451 chemiluminescence conditions used in this study.

The value for Di(t) was then found from the binarized images by taking the location of the lowest

white pixel. The position was converted from pixel values to physical space using a scale determined using

a depth gauge created by suspending a ruler into the reactant feed section before the commencement of

experiments at the being of each day. This scale was also used to determine the zero position for Di(t),

which was set at the opening of the upstream section into the combustion chamber.

The right column in Fig. 2.3 shows the results of the adjusted algorithm on a high preheat case. The

processing was able to remove almost all background noise, leaving only a small spot near the base of

the bright flame where it would not affect the detected Di(t) value. While the binarization of the dull

flame exhibits roughness along the edge, the roughness did not produce a notable effect on the depth of

the lowest white pixel, so it would not affect the Di(t) calculation.

Intensity distribution of the PLIF laser in the vertical direction was compensated for based on

a calibration sequence of acetone PLIF images. Flame fronts were originally detected from the PLIF

images using an existing program based on gradients in the image intensity. This program was eventually

determined to have insufficient precision and was not robust enough for the required calculations being

made. A new program derived from the one made for chemiluminescence data was then created. This

program used the same initial thresholding steps as the chemiluminescence version, with an added edge

smoothing step at the end to promote more robust determination of the x-coordinate of the flame tip

that was not required for chemiluminescence. The smoothing was provided by converting the binarized

image to a greyscale format, preforming a gaussian blur, and rebinarizing at a 50% intensity threshold.

A comparison between the two approaches is shown in Fig. 2.4.


Plain Image Individually Group(γ = 0.5) Thresholded Thresholded

Bright Flame

Dull Flame

No Flame

Figure 2.3: Demonstration of the thresholding algorithm for the processing of chemiluminescence images


(a) PLIF Image (b) Gradient Based Outline (c) Thresholding Based Outline

Figure 2.4: Demonstration of the outline generating algorithm for PILF images

Chapter 3

Results

3.1 Evolution of Flame Shape and Dynamically Stable

Intermittent States

As described in the test conditions section, a region of conditions exhibiting incipient flashback behaviour

was found during experimentation. At such conditions, the flame entered and exited the reactant feed

section intermittently through elongated protrusions. This differed from previous descriptions of flash-

back in which no mention of intermittent behaviour was made [40], and with Heeger et al. [50] where the

flame was noted to rotate about the wall at the end of the central bluff body. As hydrogen was added

with all other parameters fixed, the system moved further away from stable operating conditions and

consistent trends could be seen in the change of the flame’s shape and behaviour. At each condition,

the characteristic behaviour of the flame remained fairly constant over time. No memory effects were

observed other than those caused by heating of the central buff-body, with the same flame characteristics

presenting themselves if approached from a stable burning condition or a condition of greater flashback.

Conditions close to stable burning exhibited occasional small sized protrusions into the mixing section,

an example of which is shown in Fig. 3.1. These protrusions traveled only a short distance into the

upstream section. While advancing upstream, the shape of the protrusion was smooth and tapered,

with the tip angled into the swirling flow. When retreating back into the combustion chamber, the flame

often broke up into multiple tips and the shape disintegrated. The upstream propagation rate started

positive and steadily dropped to negative values.

These propagations bear some similarity in appearance to those observed by Heeger et al. [50].

However, these propagations seem to differ in that they lack any small scale bends or folds in the flame

front. Furthermore, the base of these propagations tend to spread out, where as no indication of such

behaviour was described by Heeger et al. In both the work presented here, as well as the study by Heeger

et al., the tip of the flame is shown to be angled in one direction. For the work of this study, it was

angled toward the oncoming flow. No indication of the direction of swirl was provided by Heeger et al.

The shape and size of these protrusions is maintained across a wide range of conditions as the system

moves further from flashback, with an increasing rate in the number of propagations occurring.

Eventually, as hydrogen was further increased, the flame protrusions underwent a marked change in

appearance and behaviour, as shown in Fig. 3.2. At this point, flame protrusions began to propagate

further upstream in the reactant feed section, with the penetration depth following an increasing trend

25

Chapter 3. Results 26

1cm

(a) t = 0 ms

1cm

(b) t = 0.4 ms

1cm

(c) t = 0.8 ms

1cm

(d) t = 1.2 ms

1cm

(e) t = 1.6 ms

1cm

(f) t = 2.0 ms

1cm

(g) t = 2.4 ms

1cm

(h) t = 2.8 ms

1cm

(i) t = 3.2 ms

1cm

(j) t = 3.6 ms

Figure 3.1: Sequence of chemiluminescence images showing the progression a flame protrusion nearstable conditions

1cm

(a) t = 0 ms

1cm

(b) t = 3.6 ms

1cm

(c) t = 5.6 ms

1cm

(d) t = 8.0 ms

1cm

(e) t = 15.2 ms

1cm

(f) t = 20 ms

1cm

(g) t = 23.2 ms

1cm

(h) t = 32 ms

1cm

(i) t = 48.4 ms

1cm

(j) t = 50.8 ms

Figure 3.2: Sequence of chemiluminescence images showing the progression a flame protrusion at higherintensity conditions

as more hydrogen was added. Notable wrinkling became visible during all stages of the propagation.

Flame movement became more disorganized, with protrusions switching between advancing, holding

position, and retreating in a single excursion into the reactant feed section, similar to the way originally

described for channel burners by Eichler and Sattelmayer [28]. Bulges in the flame appeared to ripple

backwards along the sides of the protrusion at times when the flame was advancing or holding position.

Similar to the smaller protrusions, the tip of the flame tended to break apart during retreat.

In all conditions, similar rotational behaviour was observed. When advancing into the upstream

section, the flame tip tended to either hold its azimuthal location, or advance slowly into the direction

of the oncoming swirl flow. When the flame retreated, it rotated approximately a quarter to a half

turn with the rotation of the flow between the point of maximum advance and the point of maximum

retreat. This amount of travel held true for all sizes of propagations, while small propagations moved

quickly in rotation, very large propagations rotated very slowly as they retreated larger distances. Due


to this characteristic and the thickness of the flame propagations in the radial direction, very few false

propagation detections were made.

Plotting Di versus time provided a time signal showing the flame’s movements. For near stable

conditions, each protrusion was readily distinguishable as a separate event with a Di value of zero being

reached between each propagation. Just past the stable state, only a small number of propagations could

be seen in the signal. As hydrogen was added to the system, the number of these propagations increased

while the other characteristics remain constant, leading to a signal showing many quick spikes such

as Fig. 3.3. Further increases in hydrogen content resulted in the length of time for each propagation

increasing as seen in Fig. 3.4. As a result, the time the flame spent in the upstream section came

to exceed the time it spent outside the upstream section. While the flame tip’s movements at such

conditions was still mostly a steady movement from positive propagation rate to negative, the first

significant back and forth movements could be seen. Also present were a small number of anomalously

deep propagations that far exceeded the maximum depth reached by the other propagation events.

As the system approached a state of continuous flashback, such as in Fig. 3.5, it became common for

multiple significant advance/retreat events to occur without fully retreating to the combustion chamber,

with small scale fluctuations within each event.

In order to separate each event for analysis, the peaks were identified using a prominence based

method. The prominence of a peak is defined by measuring the height of the peak over the lowest point

between the measured peak and the next higher peak. This is measured on both sides, with the lower of

the two values being the prominence. By thresholding peaks by prominence, events where large changes

in Di occurred from a low Di starting position could be included, while peaks that resulted from small

fluctuations at a higher Di could be ignored. The Di value of each peak that met the requirements of

the prominence filter was labeled as the peak depth of travel (Dp) value of the event, while the minimum

Di value between filtered peaks was labeled the depth of travel at retreat (Dr) between events. At most

given conditions, Dp values tended to fall within a narrow range. The red band in Fig. 3.3 through 3.5

represents the region between the 10th and 90th percentile values for Dp. An exception to this trend is

in conditions such as Fig. 3.4 where the anomalously high Dp events far exceeded the 90th percentile

value. For conditions near stable, Dr was zero for every event. As conditions reached higher flashback

intensity Dr no longer reached zero every time, however, the Dr value still tended to drop below a

certain threshold; for the conditions in Fig 3.5 this is approximately 40 mm. Neither the Dp or Dr

values of events trended up or down over time at a given operating condition, nor did the number of

events change, indicating that the combustor was dynamically stable in these states.

At all conditions studied, the properties of Di versus time signals consistently followed the same

evolution, moving between states that mirror those presented in Figs. 3.3 through 3.5. In the intermittent

flashback cases examined near stable conditions, the flame protrusion propagated upstream to reach a

Dp within a narrow band, then retreated completely out of the reactant feed section. As conditions

moved towards stronger flashback, the band of possible Dp values eventually began to shift towards

deeper propagation distances. At the same time, the number of events occurring increased, eventually

followed by an increase in the length of each event. Next, the system reached a state where the flame

no longer exited to the combustion chamber between each event, instead it retreated past a certain Dr

value that remained below the band of possible Dp values.


0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

Time, seconds

0

5

10

15

Figure 3.3: Di time signal showing many short duration propagations

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

Time, seconds

0

10

20

30

40

Figure 3.4: Di time signal showing regular long duration propagations that fully exit the upstreamsection

3.2 Statistical Qualities of Intermittent Flashback Events

Quantitative comparison of flashback at different conditions within the intermittent regions necessitates

the use of variables that can describe a given condition over the entire length of the case. In order to

determine the trends in flashback conditions, a variable that acts as a description of magnitude must be

derived. Mean peak depth of travel (Dm), was defined as the mean value of all Dp values in a case; it is

represented as a dashed line in Figs. 3.3 through 3.5. The value of Dm represents the expectation value

for Dp at the given case, covering the most important characteristic for the magnitude of flashback,

viz. how far upstream the flame can reach. This definition is more favourable than definitions such as

absolute maximum Di value in a case, as it is less sensitive to outliers, making comparisons between

conditions more accurate, and it takes into consideration the full length of data.

The frequency with which events occurred at was also of interest. For observations at near stable

burning conditions, the rate at which upstream propagation events occurred was the main change that

resulted from changing conditions. Examination of the Di time signal revealed that the fluctuations were

randomly spaced rather than oscillatory in nature. As a result, unlike the propagations found in Heeger

et al. [50], the system’s state did not possess a harmonic frequency. However, a statistical frequency

could still be described. The statistical frequency used in this study, f , is the mean number of Dp values

per second, providing an expectation value for the number of flashback events over a given time period.

A plot showing f versus Dm at every composition and flow velocity tested at room temperature is

provided in Fig. 3.6a, while the same variables plotted for each composition and temperature at an air

flow rate of 720 SLPM is given in Fig. 3.6b. Together, the two plots contain 451 different conditions

with varying reaction and flow characteristics. Despite this, both plots collapsed well onto a distinctive


0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

Time, seconds

0

20

40

60

80

Figure 3.5: Di time signal showing conditions just before continuous flashback where the flame rarelyexits the upstream section

curve. Extending the curve to a zero f value shows that when the intermittent flashback regime was

entered from stable conditions, it did so at a non-zero Dm value of approximately 7.5 mm. From there, f

underwent a rapid and linear rise over a small change in Dm, until levelling off at a peak at Dm ≈ 10 mm,

after which point f droped off. For the range of tested conditions the f value of this peak was between 80

and 200 s−1. The initial drop off was rapid, with only a slightly shallower slope than the previous linear

rise, with the f value eventually levelling out and beginning to climb again as Dm continued to rise.

This distinctive curve means that, as a system in this regime moves from a state of stable burning to a

state of continuous flashback, the system will show a set pattern with regards to Dm and f , regardless

of the conditions driving that movement through the intermittent flashback regime.

0 10 20 30 40 50 60 70

0

20

40

60

80

100

120

140

(a) Room Temperature Cases

0 10 20 30 40 50 60 70

0

20

40

60

80

100

120

140

160

180

200

295K

400K

500K

600K

(b) Preheat Cases

Figure 3.6: Rate of intermittent flashback events (f) versus mean peak depth (Dm) across all 451 cases.

During conditions where Dr was exclusively equal to zero, that is conditions in which the flame fully

exited the upstream section between each propagation, Di had a value of zero between each propagation

event. This corresponds to the regime of intermittent flashback prior to the beginning of transition to

continuous flashback at values of Dm ≈ 20. Such conditions allowed the separation of the Di(t) time

signal into parts where the flame was in the upstream section, Di 6= 0, and parts where the flame was

not in the upstream section, Di = 0. The mean length of time that the time signal spent at Di = 0

between events was defined as tb, the mean time between flashback events. The mean time that Di > 0

was defined as tf , the mean flashback event duration. For conditions where the Dr = 0 condition was


met, f could be given as a function of event times using the equation

f =1

tb + tf(3.1)

Breaking down the events into tf and tb therefore provided a means of examining the manner by which

the f behaviour of intermittent flashback was reached.

Mean event times tf and tb for the room temperature cases are shown plotted against Dm in Fig. 3.7a.

The value of tf was shown to increase linearly with Dm, while tb decreased in a hyperbolic manner. By

looking at the inverse of tb in Fig. 3.7b, it could be seen that there was a linear relationship between

t−1b and Dm. At the point of transition from stable burning to intermittent flashback tf was zero, while

tb was infinite. The asymptote behaviour of tb also resulted in an infinitely negative slope at this point,

while the slope of tf was finite. As Dm increased, the slope of tb tended to zero while the slope of tf

remained constant. As a result, tb dominated Eq. 3.1 near the transition to stable, reducing it to

f =1

tb(3.2)

which gave the initial linear increase seen in Fig. 3.6. When Dm was large, tf dominated Eq. 3.1 instead,

reducing to

f =1

tf(3.3)

and giving a hyperbolic decrease in the latter part of the f versus Dm curve. Between the two extremes,

the frequency peak occured at the point where the sum of the slopes of tf and tb was equal to zero.

Figure 3.8 displays a simulated f versus Dm plot based on Eqs. 3.1 through 3.3. The plot was created

based on lines of best fit taken from the data displayed in Fig. 3.7, with the experimental f versus Dm

values shown in light grey.

Examining tf and t−1b together revealed additional information about some of the behaviours of a

system in the intermittent flashback regime. While other than some spreading of the peak f value,

the f versus Dm plots in Fig. 3.6 appeared to collapse well to a single curve. However, Fig. 3.7 shows

relatively broad lines that have distinct boundaries, suggesting that this broadening was not caused by

random noise. This means that, while the changes other than the peak f value were small, some effects

of conditions did move the f versus Dm curve.

Another feature seen in Fig. 3.7 is that the values of tf appear to cut-off well above zero, whereas

the t−1b values appear to go right to zero. Furthermore, extending the line of best fit for either to zero

resulted in different x-intercepts. Since the same set of Dm values were plotted for each plot in Fig. 3.7,

continuing the trend lines to a zero value for tf data points would result in the corresponding tb values

being negative. As this is physically impossible, either the trend must deviate sufficiently close to zero,

or a zero value for tf cannot be reached. The idea that a tf value of zero is impossible supports the

minimum Dm that has been previously observed. Since the movement of the flame was restricted by its

flame speed and the flow velocity at its leading edge, a finite amount of time was required for it to travel

the length of Dm and back, therefore, the minimum (zero f) Dm value should impart a minimum tf

value. Figure 3.9 backs this up, showing that during the cases where equation 3.1 held true, a minimum

limit for the Dp value of any given propagation held at the point of the minimum Dm value, and that

as the system neared the edge of the intermittent flashback stable boundary, all propagations converged

to the same Dp value.


6 8 10 12 14 16 180

0.005

0.01

0.015

0.02

0.025

0.03

y=0.00199x-0.0135

(a) tf and tb versus Dm

6 8 10 12 14 16 180

100

200

300

400

500

600

y=86.8x-660.4

(b) Dm versus 1/tb

Figure 3.7: Mean event times

5 10 15 20 25 30 35 40

0

20

40

60

80

100

120

Figure 3.8: f versus Dm simulated from the lines of best fit for tf and tb

3.3 Effects of Operating Conditions

While the statistics showed that the dynamical behaviour of the system in intermittent flashback condi-

tions was dependant on the magnitude of the flashback without regard to the conditions in the combustor,

as demonstrated by the relationship between f and Dm, the magnitude of flashback was still a function

of conditions. It therefore is important to have an understanding of the effects of conditions on the com-

bustor. The conditions studied allowed for direct observation of the effects of changing fuel composition,

equivalence ratio, bulk flow velocity, and reactant temperature. These conditions could also be used to

calculate how the magnitude of flashback relates to non-dimensional quantities. Finally, the magnitude

of flashback could be compared to existing measures for calculating the flashback limits to see how well

they act as predictors.

Figure 3.10a shows Dm for each different fuel composition at 720 SLPM of room temperature airflow.

As hydrogen was added to the flow at a fixed methane flow rate, there was a large range of conditions over

which Dm increased only slightly, followed by a small range with a slightly faster increase in Dm, and

then a sudden jump in Dm as the system transitioned to continuous flashback. Methane flows between

44 SLPM and 52 SLPM shifted the system towards earlier transition to continuous flashback than either


0 10 20 30 40 50 60 70

0

10

20

30

40

50

60

70

(a)

6 8 10 12 14 16 18

5

10

15

20

25

(b) Detail

Figure 3.9: Dp percentiles

higher or lower methane flow rates. Since Dm changed little before the transition to continuous flashback,

it does not serve as a good early warning signal for impending flashback.

Applying the discovery of a rapid, linear increase in f over a small change in Dm made in the

last section, it was possible to more closely examine the initial region where Dm changes very little

by plotting the effect of composition on f in Fig. 3.10b. Similar to the f versus Dm plots, f initially

increased linearly as hydrogen content of the flow was increased before dropping off again. This suggests

that over conditions exhibiting only a slight increases in Dm with increasing hydrogen f increased much

more significantly. The hydrogen flow rate corresponding to the point of maximum f depended on the

conditions, but the peak f value always occurred at slightly lower hydrogen content than the rapid

increase in Dm.

15 20 25 30 35 40

Hydrogen Flow Rate, SLPM

0

10

20

30

40

50

60

70

80

36 SLPM

40 SLPM

44 SLPM

48 SLPM

52 SLPM

56 SLPM

(a) Dm

15 20 25 30 35 40


0

10

20

30

40

50

60

70

80

90

100

110

36 SLPM

40 SLPM

44 SLPM

48 SLPM

52 SLPM

56 SLPM

(b) f

Figure 3.10: Effect of fuel composition on Dm at room temperature and 720 SLPM air

The effects of bulk flow velocity on Dm and f are detailed in Fig. 3.11. The methane flow rate in the

figures has been fixed at 6.1% of the air flow rate, where the transition to continuous flashback occurred

earliest in the compositional data. Since the system was shown to be largely driven by hydrogen content

in the compositional data, it was used here to allow the effects of bulk flow velocity to be seen across


the range of intermittent flashback magnitudes. All cases shown are at room temperature. Figure 3.11a

shows that the transition to continuous flashback occurred earlier at lower bulk flow rates, but that until

the sudden transition, all conditions produced approximately equal Dm values. It also appeared that

higher bulk flow rates showed a larger increase in Dm just before the transition, moving more gradually

into continuous flashback. This more gradual movement towards flashback appeared to create a larger

separation between the point at which the peak f value occurred and when the shift to continuous

flashback occurred. Higher bulk velocities also pushed the system towards higher f , both relative to

hydrogen content, and where the system was positioned relative to the transitions to stable burning and

continuous flashback. As a result, f appeared to give more warning of flashback at higher bulk flow

velocities.

0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055

Volume Fraction of Flow as Hydrogen

0

10

20

30

40

50

60

70

80

630 SLPM

675 SLPM

720 SLPM

765 SLPM

810 SLPM

(a) Dm

0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055


0

20

40

60

80

100

120

630 SLPM

675 SLPM

720 SLPM

765 SLPM

810 SLPM

(b) f

Figure 3.11: Effect of bulk flow velocity on Dm at room temperature and 6.1% methane

Figure 3.12 demonstrates the effects of reactant temperature using the same methods as was used for

Fig. 3.10 and 3.11. It could be seen in Fig. 3.12a that increasing the reactant temperature led to earlier

and more gradual transition to continuous flashback. There was only one point of overlap between the

295 K case and the 400 K, and two points between the 400 K case and the 500 K case, however, it

appeared to show that the lines collapsed back to a single curve near stable conditions. The data set at

600 K did not contain any cases on the stable side of the f peak. Figure 3.12b shows that heating the

reactants resulted in an earlier f peak as well as higher peak values for f .

A curiosity occurred with the effects of composition at elevated temperature. While the trends for

the effects of composition were the same at every bulk flow rate, the trends for bulk flow rate were

the same at every composition, and the trends for increasing reactant temperature were the same at

every composition, the effects of changing composition were significantly different at elevated reactant

temperatures than at room temperature. In Fig. 3.10, it can be seen that a methane flow rate of

56 SLPM, transitions to continuous flashback at higher hydrogen contents than at 44-52 SLPM, despite

the equivalence ratio at transition still being lean at 0.85. Figure 3.13a, shows the effect of different

compositions at a reactant temperature of 500 K. For the elevated temperature cases, the hydrogen

required to transition to continuous flashback continued decrease with increasing methane content for

all tested compositions, as would normally be expected from lean equivalence ratios. The transition

to continuous flashback occurred more gradually at the increased temperature, as can be seen when


0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05


0

10

20

30

40

50

60

70

80

295 K

400 K

500 K

600 K

(a) Dm

0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055


0

20

40

60

80

100

120

295 K

400 K

500 K

600 K

(b) f

Figure 3.12: Effect of preheat on Dm at 720 SLPM air and 44 SLPM methane

compared to Fig. 3.10a. Examining Fig. 3.13b showed that while the peak f reached still decreased with

increasing methane content, the decrease was significantly less than that observed for the same change

at room temperature seen in Fig. 3.10b. Oppositely, while the methane flow rates that transitioned to

continuous flashback at lower hydrogen flow rates also reached their f peak at lower hydrogen flow rates,

both for room temperature and elevated temperatures, the effect was much more pronounced at elevated

temperatures.

5 10 15 20 25 30


0

5

10

15

20

25

30

35

40

45

50

36 SLPM

40 SLPM

44 SLPM

48 SLPM

52 SLPM

56 SLPM

(a) Dm

5 10 15 20 25 30 35


0

20

40

60

80

100

120

36 SLPM

40 SLPM

44 SLPM

48 SLPM

52 SLPM

56 SLPM

(b) f

Figure 3.13: Effect of fuel composition on Dm at 500 K and 720 SLPM air

Other room temperature studies examining flashback in methane and other hydrocarbon based flames

have found that the maximum critical gradient values occurred slightly above stoichiometric [61]. While

this behaviour follows what happened in the elevated temperature case, it does not agree with the

results from the room temperature cases, where despite being at below stoichiometric equivalence ratios,

the highest methane concentrations required more hydrogen to transition to continuous flashback. A

possible explanation for this phenomena is partial quenching of the flame. High hydrogen content fuels

are more susceptible to flashback than pure hydrocarbon fuels due to hydrogen’s high laminar flame

speed and short quenching distances [28]. While studies typically calculate fuel properties based off of


one-dimensional simulations of flames far from walls, this may be a poor assumption for boundary layer

flashback. If the mechanism for boundary layer flashback contains processes that occur between the

distances at which quenching effects start to take place for the hydrocarbon and hydrogen, then it may

be necessary to calculate partially quenched fuel properties to be able to accurately model the flashback

properties of hydrocarbon-hydrogen fuel blends. In Fig. 3.10a it can be seen that the 36 and 56 SLPM

methane cases transitioned to continuous flashback at 38 SLPM hydrogen, and the 44-52 SLPM methane

cases at 34 SLPM hydrogen. Furthermore, all cases reached peak f values in Fig. 3.10b around 30 SLPM

hydrogen. If the methane in the mixture was to be ignored, these values would represent 5.0%, 4.5%,

and 4.0% hydrogen content in the flow respectively. These values are right along the boarder of the

flammability limits for hydrogen in air at room temperature and atmospheric pressure, which range

from 4.2% for upwards propagation and 9.0% for downwards propagation in an open tube, to 9.4% for

side ignition in a closed globe [62]. Being in a situation both close to a wall and propagating downwards,

it is likely that the hydrogen would have been below its flammability limit on its own. The overall

situation could therefore be described as flashback being influenced by a region where methane was at

least partially quenched while hydrogen burned under the influence of both the aggravating effect of the

adjacent unquenched flame and the quenching effect of the wall.

Examination of Dm and f in relation to conditions in Fig. 3.10-3.13 showed that, while conditions

may have minimal affect on the shape of the f versus Dm curve outside of peak f height, they could effect

how the system moved along the curve. Changing room temperature composition and bulk flow velocity,

as shown in Fig. 3.10 and 3.11 respectively, resulted in minimal change in the hydrogen content at which

the peak f occurred with changing conditions. However, more hydrogen was required to transition to

continuous flashback outside of the 44 to 52 SLPM methane range and at high bulk flow rates, and it

can be seen that outside this range the transition occurred more gradually.

Figures 3.12 and 3.13 show the effects of inlet temperature and composition at elevated tempera-

tures respectively. Both increased inlet temperatures and methane contents move both the f peak and

continuous flashback transition to lower hydrogen values. Where as a more gradual transition to con-

tinuous flashback resulted from increasing the inlet temperature, and at elevated reactant temperatures,

lowering the methane content. Bringing this together, it could be seen that a movement from stable

burning to continuous flashback under a certain set of conditions would move along the f versus Dm

curve with conditions controlling the point at which the system transitioned to intermediate flashback,

reached peak f value, and transitioned to continuous flashback, as well as the particular f value reached

at the peak. The condition leading to each of these points, and therefore the separations between them,

could be changed independently of each other or the peak f value. In other words, an earlier transition

did not necessitate a sharper or a more gradual transition, nor did a quick movement from the limit

condition of stable burning to the condition of peak f , necessitate a quick movement from the condition

of peak f to the condition of transition to continuous flashback.

Differences in the peak f value could be explained by examining the effects of conditions on the mean

event times, tf and tb, versus Dm. While Fig. 3.7 appeared to show variation occurring only as shifts in

the tf and t−1b , examination of a plot broken down by conditions, such as Fig. 3.14, revealed that both

shifts and changes in slope of the tf and t−1b lines occurred. Both tf and t−1

b showed higher slopes and a

shift towards slightly higher initial Dm values for lower methane contents, with the change in slope being

more prominent in t−1b . Since the Dm data for each case is the same in each plot, the shift in initial Dm

value is identical for both tf and t−1b . The combination of these two changes helped to hide the slope


change below the f peak, while the lower density of data points past the peak location made it hard to

see in Fig. 3.7. The spreading of tf past the f peak point likely caused the slight broadening of the f

versus Dm curve at high Dm values. While the increased slope of t−1b at low methane contents resulted

in a steeper slope for tb at low Dm values and a shallower slope at high Dm values as it approached the

asymptote, the middle region where the f peak was had little change in slope. Since the Dm value of

the f peak was based on slopes, the small change in slope with conditions in this region, coupled with

the rapid change in slope with Dm for tb, helped to keep the f peak at a constant location as conditions

changed. The actual value of f , however, was dependant on the sum of tf and tb, so the more significant

drop in tb at low methane flow rates caused the higher peak f values at the same conditions. It is notable

that deviation from linear in the t−1b line occurred mostly at higher Dm values where the effect of tb on

f was minimized.

6 8 10 12 14 16 18

0

0.005

0.01

0.015

0.02

0.025

0.03

(a) tf and tb

6 8 10 12 14 16 18

0

100

200

300

400

500

600

5.0%

5.6%

6.1%

6.7%

7.2%

7.8%

(b) 1/tb

Figure 3.14: Effect of composition on tf and tb at room temperature and 765 SLPM Air Flow Rate

3.4 Flashback Predictors

We now seek to evaluate predictive metrics for flashback. Given a reactant composition and temperature,

the laminar flame properties for each predictor were calculated using the Cantera software package [63]

with the GRI 3.0 mechanism [64]. Lewis number was calculated according to the method described

by Bouvet et al. [65]. For a predictor that accurately captures the driving variables, the data should

collapse onto a singular curve when the predictor is plotted against Dm similar to the way it did for

the f versus Dm curve in Fig. 3.6. Failing to meet this condition, a predictor may still produce robust

results if all data intersects at the Dm value where the transition to continuous flashback occurs.

To make a predictor from critical gradient theory, a ratio between the critical gradient under the

inlet conditions of the case, and the actual velocity gradient was required. While the propagation

distance for critical gradient theory could not be calculated itself, it was known that the propagation

distance was proportional to the quenching distance, which was itself proportional to the flame thickness.

Similarly, since the geometry of the combustor was constant, the actual gradient was proportional to

bulk flow velocity. As the chemical time scale is defined by the flame thickness divided by the laminar

flame speed, it is proportional to the inverse of the critical gradient. As a result, the predictor was as


proportional to the Damkohler number, as was used by Kalantari et al. [53] in the design of the Damkohler

correlation. The plot of Dm versus Damkohler number shown in Fig. 3.15 displayed a separation of the

room temperature data into six curves based on the methane concentration, and further separation for

the elevated temperature cases.

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.60

100

200

300

400

500

600

700

800

36 SLPM

40 SLPM

44 SLPM

48 SLPM

52 SLPM

56 SLPM

(a) Room temperature cases

0 0.5 1 1.5 2 2.5

0

100

200

300

400

500

600

700

800

295 K

400 K

500 K

600 K

(b) Elevated temperature cases

Figure 3.15: Dm versus Damkohler number showing the effectiveness of critical gradient theory as aflashback predictor.

While the centerbody tip temperature was not directly measured for the Damkohler correlation, it

was regulated by the thermal procedures, allowing the tip temperature term to be removed. Similarly,

the pressure term could be ignored because pressure was held constant, and as only the data collapse

was of concern, the constant term could be ignored. The resulting equation of

DaDC = Le1.68 · Pe1.91f ·(TuT0

)2.57

(3.4)

is plotted in Fig. 3.16. The same way as the actual Damkohler number, the data for different methane

contents and reactant temperatures did not collapse.

Figure 3.17 shows Dm plotted against hydrogen content. The data showed a much greater collapse for

the room temperature cases for hydrogen than it did for the Damkohler number based plots, although

there was still a large spread present for the elevated temperature cases. The fact that the previous


0 0.5 1 1.5 2 2.5

104

0

100

200

300

400

500

600

700

800

36 SLPM

40 SLPM

44 SLPM

48 SLPM

52 SLPM

56 SLPM


0 2 4 6 8 10 12

104

0

100

200

300

400

500

600

700

800

295 K

400 K

500 K

600 K


Figure 3.16: Dm versus Damkohler correlation showing the effectiveness of the Damkohler correlationas a flashback predictor.

Damkohler number based predictors did not manage to produce a collapse better than the simple hy-

drogen fraction, particularly in regards to temperature and fuel composition, indicates that they are

unsuitable for use with hydrogen/hydrocarbon mixtures. Introducing partially quenched, multicompo-

nent calculations for variables such as laminar flame speed may be an avenue to correct this failure to

collapse.

3.5 Observations From Laser Diagnostics

Of the six S-PIV/PLIF cases recorded, the corresponding chemiluminescence conditions indicated that

the first three cases should have followed the pattern of Fig. 3.3, Cases 4 and 5 Fig. 3.4, and Case 6

Fig. 3.5. Based on the PLIF images however, it would appear that a slight shift towards earlier flashback

occurred in these cases, with Case 3 instead showing behaviour similar to Fig. 3.4, and Case 5 showing

behaviour similar to Fig. 3.5. The reason for this shift is unknown, however, the consistency of results

taken before and after the S-PIV/PLIF cases suggests that no movement of any combustor components

occurred. Cases 1 and 2 showed no detectible changes in the flow field around the flame, however, a

significant portion of the flame in these cases existed in the region blocked by the plate between the


0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055 0.060

100

200

300

400

500

600

700

800

36 SLPM

40 SLPM

44 SLPM

48 SLPM

52 SLPM

56 SLPM


0 0.01 0.02 0.03 0.04 0.05

0

100

200

300

400

500

600

700

800

295 K

400 K

500 K

600 K


Figure 3.17: Dm versus hydrogen content.

upstream section and combustion chamber. Case 6 spent most of recorded time with the tip extending

beyond the measurement region. The data presented, therefore, corresponds to the Cases 3 through 5.

Examination of individual frames of the combined S-PIV/PLIF data did not show a consistent

reduction in axial flow velocity ahead of the flame tip. For a typical flame propagation event such as

the one shown in Fig. 3.18, a strong area of reduced velocity near the center where the laser intersects

the boundary layer could be observed, but no change ahead of the flame tip appeared to exceed the

level of variation present due to turbulence. However, in some flame propagation events at conditions

generating deeper depths of travel, a broadening of the area of reduced flow velocity occurred, such as

seen in Fig. 3.19. Significant deflections away from the axial direction, such as that occurring ahead of

the second deepest flame tip in Fig. 3.19, occurred occasionally either along the tip of a flame or sitting

in the boundary layer away from the flame. A more common occurrence, however, was that the flow

just ahead of the flame would simultaneously deflect towards the axial direction while decelerating, still

producing a net axial deceleration, such as in Fig. 3.20. In general, the flow direction along the lee side

(the side away from the oncoming flow) of the flame matched the overall direction of the flame edge and

commonly ran parallel to small segments of the flame.

The lack of deflection away from the axial direction and a reverse flow region is in contrast to the

findings of Ebi and Clemens [49] in a similar combustor. However, several differences in the operating

conditions may account for this. The first is the measurement system. Ebi and Clemens used a laser

sheet both tangential to the center body as well as a radial slice cutting through it, allowing observation

right up to the surface. The laser sheet used in this experiment ends 0.5 mm above the surface. While

this should have been within range to observe the flow reversal if the it occurred at the same scales as Ebi

and Clemens, it is possible that reversals in our experiments occurred at a smaller scale. Furthermore,

the measurements made by Ebi and Clemens were father upstream in the premixing tube compared to

the measurements in this study. Hence, their observations were of deeper penetrating flame tongues

than were the focus of this study.

The position of the most upstream point on the flame surface at a given time is defined as ~xf (t). The

axial velocity of the flow at the flame tip is then defined as Uyf (t)|~xf (t) = Uy(~xf (t), t). To determine if

there is a statistical effect on the axial velocity of the flow associated with the leading edge of the flame,


Figure 3.18: Typical flame propagation during flashback. The background shows axial velocity for easeof interpretation. Taken from Case 3.

the change in axial velocity at the flame tip is defined by

∆Uyf (t)|~xf (t) = Uyf (t)|~xf (t) − Uy|~xf (t) (3.5)

where Uy|~xf (t) is the time average of the axial velocity at position ~xf (t) in the absence of flashback.

Figure 3.21 shows a plot of ∆Uyf at each position. A slight tendency towards a reduction in axial flow

speed was observed in the center of the flame’s region of travel, while the edges tended towards increased

axial velocity.

While this did not strongly support the idea of axial flow velocity reversal, prior studies such as

Eichler and Sattelmayer [46] had seen large velocity gradients near the flame tip as shown in Fig. 1.7.

By evaluating Uyf (t)|~xf (t)+~δf, where ~δf is some positional displacement, for different values of ~δf , it is

possible to determine the degree to which the flame affects the axial velocity at positions offset from

the tip. Figure 3.22 shows Uyf (t)|~xf (t)+~δfversus [0ı + 1] · ~x(t) for several values of ~δf . An effect

can be observed from a move in either direction. A vertical shift to 5 vectors above the flame tip

(~δf = [+2.4 ] mm) showed a notable increase in Uyf (t)|~xf (t)+~δfas expected due to expansion from

the flame, with a similar increase caused by moving another 5 vectors above that point. The velocity

5 vectors below the flame tip showed a similar drop in Uyf (t)|~xf (t)+~δf, whereas the velocity 10 vectors


Figure 3.19: Flame propagation showing broadening of reduced velocity area. The background showsaxial velocity for ease of interpretation. Taken from the Case 4.

below the flame tip showed only a small further decrease. Shifting to the right or left of the flame

created a similar but smaller change, with Uyf (t)|~xf (t)+~δfincreasing towards the right (windward) side,

and decreasing towards the left (lee) side.

Figures 3.23 and 3.24 shows the values for ∆Uyf |~xf (t)+~δffor ~δf = [+2.4 ı − 2.4 ] mm. At this

offset, a strong tendency towards a drop in axial velocity due to the flame could be observed. When

the change in velocity was compared over depth of travel in Figs. 3.25 and 3.26, it is observed that the

drop in ∆Uyf |~xf (t)+~δfincreased as the flame traveled further down the tube. For ~δf = 0, the drop in

∆Uyf |~xf (t)+~δfwas initially near zero as the flame entered the tube, building up as it progressed further

upstream. When ~δf = [+2.4 ı − 2.4 ] mm, a drop in ∆Uyf |~xf (t)+~δfcould already be detected at the

start of the measurement region, and the build up occurred slightly faster than at the detected flame tip

location.


Figure 3.20: Flame propagation showing axial deflection of upstream flow. The background shows axialvelocity for ease of interpretation. Taken from Case 4.


Figure 3.21: Difference between axial velocity at flame tip and mean axial velocity without the flamefor Case 4.

-40-35-30-25-20-15-10

0

2

4

6

8

10

12

14

16

(a) Vertical offsets

-40-35-30-25-20-15-10

0

2

4

6

8

10

12

14

16

(b) Horizontal offsets

Figure 3.22: Effect of ~δf on Uyf |~x(t)+~δf during flashback


Figure 3.23: Difference between axial velocity offset by ~δf = [+2.4 ı−2.4 ] mm, and mean axial velocitywithout the flame for Case 3.

Figure 3.24: Difference between axial velocity offset by ~δf = [+2.4 ı−2.4 ] mm, and mean axial velocitywithout the flame for Case 4.


-35-30-25-20-15-10

-3

-2

-1

0

1

2

Figure 3.25: Difference between axial velocity at the flame tip and mean axial velocity without the flameversus axial position of flame tip for Case 3.

-40-35-30-25-20-15-10

-3

-2

-1

0

1

2

Figure 3.26: Difference between axial velocity at the flame tip and mean axial velocity without the flameversus axial position of flame tip for Case 4.

Chapter 4

Conclusion

The incipient conditions where the flame enters the upstream reactant feed section before full flashback

have been studied in a lean premixed swirl burner. Intermittent properties not previously reported

were observed. Through the examination of 451 cases using OH* chemiluminescence, characteristic

behaviours that may be used as early warning signs of imminent flashback were found. A further six

cases recorded with simultaneous S-PIV and PLIF provided information regarding the development of

the flow during intermittent flashback.

A consistent pattern of flame propagation properties were found to occur between stable and flashed

back states within the range of conditions tested. Flame propagations into the upstream section near

stable conditions were characterized as being shallow, infrequent, and of short duration. The rate at

which these propagations occurred increased greatly as the system moved further into the intermittent

regime, while the depth they penetrated increased only slightly. Eventually, the duration of the propa-

gations increased, leading to the flame spending more time in the upstream section than not, but still

frequently exiting it. Finally, the flame stopped retreating out of the upstream section between each

propagation, and then at the point of continuous flashback, no longer retreated out of the upstream

section at all.

By studying the instantaneous flame tip depths, Di(t), it was found that the statistical frequency of

flashback events, f , followed a distinct curve in relation to the mean peak depth of travel of flashback

events, Dm. When moving from stable to fully flashed back, the system first entered the intermittent

regime at f = 0 and Dm 6= 0. f then climbed rapidly and linearly over a small change in Dm, before

reaching a peak value and falling off hyperbolically. This curve showed a high degree of collapse regardless

of input conditions, indicating that the dynamical characteristics of intermittent flashback in this regime

depend more on the magnitude of the flashback events occurring than the conditions used to achieve it.

By breaking each propagation down into the time between flames, tb, and the time that the flame was

in the upstream section, tf , it was possible to explain some of this behaviour. f is defined by the sum

of the two event times. When the system first entered the intermittent flashback regime from a stable

state, tb was very high while tf was low. As the operating conditions were adjusted towards continuous

flashback, both tf and t−1b rose linearly with Dm. This gave the f versus Dm curve its characteristic

shape, and allowed the curve to be approximated by determining the lines of best fit for the event times.

Due to the robust nature of both the f versus Dm curve, as well as the event times, they could serve as

early warning signs for imminent flashback even when the inlet conditions were unknown.

46

Chapter 4. Conclusion 47

Examination of the system’s reaction to inlet conditions provided information about how the sys-

tem moved along the f versus Dm curve. As expected from the findings of previous work, the system

would move closer to complete flashback when the flow rate was reduced, inlet temperature increased,

or hydrogen added. Methane content however showed an anomaly; at room temperature the maximum

tendency to flashback was reached at values significantly below stoichiometric, while at elevated tem-

peratures, increased methane brought the system closer to flashback for all tested equivalence ratios.

When the methane characteristics were compared with hydrogen flammability limits, it suggested that

multicomponent effects are playing a significant role in the burning near the wall that is driving flash-

back. Another discovery was that by altering different inlet conditions, the peak f value, how quickly

the system moved through the remaining conditions to reach the peak value, and how quickly the system

moved through those conditions to complete flashback after the peak value was reached, could all be

controlled independently of each other.

Testing of previous simple predictors found them unsuitable for use in hydrogen/hydrocarbon mix-

tures with variations in fuel composition or inlet temperature. The hydrogen content in the flow provided

a better predictor in these cases.

Laser diagnostics showed that only a small reduction in axial flow velocity occurred upstream of the

flame tip in medium intensity intermittent flashback propagations, while no effect on the flow could be

resolved for smaller propagations. In the medium intensity cases, the reduction in velocity was higher

slightly upstream axially and downstream rotationally. The magnitude of the reduction increased as the

flame proceeded farther into the upstream section.

The results of this work suggest several areas for further study involving the incipient flashback

regime. Differences between the velocity fields found in this study and in previous work by Heeger et

al. [50], and Ebi and Clemens [49], suggest the need for more comprehensive measurements including

a larger portion of the burner. The differences with Ebi and Clemens suggest a need to study how

the mechanism of flashback evolves as the flame moves from the combustion chamber all the way to

the bottom of the reactant feed section. Additionally, it should be investigated how the length of

the tube affects flashback properties, and how much of the changes are the result of influence of the

combustion chamber versus how much result from the development of the flow profile. Comparison to

Heeger introduces the question of how much centerbody height affects the characteristics of flashback,

and how it affects the degree to which dynamics in the combustion chamber effect the flow upstream.

More generally, the results also make suggestions for aspects that should be considered during future

flashback research. The room temperature methane anomaly suggests that comparisons of different

fuels, especially hydrogen-hydrocarbon mixtures, must be done at engine relevant conditions in order to

achieve accurate trends. When the anomaly is taken in addition to the known flammability limits for

hydrogen, it becomes apparent that predictors for hydrogen-hydrocarbon fuels require the development

of a means for multicomponent, partially quenched flame property calculation. Finally, having shown

that flashback onset is gradual, it is important that future studies provide a distinct definition as to the

point at which they consider a system flashed back.

Bibliography

[1] Tim Lieuwen, Michael Chang, and Alberto Amato. Stationary gas turbine combustion: Technology

needs and policy considerations. Combustion and Flame, 160(8):1311 – 1314, 2013.

[2] Robert W. Schefer, Christopher White, and Jay Keller. Chapter 8 - lean hydrogen combustion. In

Derek Dunn-Rankin, editor, Lean Combustion, pages 213 – VIII. Academic Press, Burlington, 2008.

[3] Alireza Kalantari and Vincent McDonell. Boundary layer flashback of non-swirling premixed flames:

Mechanisms, fundamental research, and recent advances. Progress in Energy and Combustion Sci-

ence, 61:249 – 292, 2017.

[4] Robert K. Cheng and Howard Levinsky. Chapter 6 - lean premixed burners. In Derek Dunn-Rankin,

editor, Lean Combustion, pages 161 – V. Academic Press, Burlington, 2008.

[5] D.W. Van Krevelen and H.A.G. Chermin. Generalized flame stability diagram for the prediction of

interchangeability of gases. Symposium (International) on Combustion, 7(1):358 – 368, 1958.

[6] Vince McDonell. Chapter 5 - lean combustion in gas turbines. In Derek Dunn-Rankin, editor, Lean

Combustion, pages 121 – IV. Academic Press, Burlington, 2008.

[7] Sanjay M. Correa. A review of nox formation under gas-turbine combustion conditions. Combustion

Science and Technology, 87(1-6):329–362, 1993.

[8] Stephen R. Turns. An Introduction to Combustion Concepts and Applications. McGraw-Hill, 3

edition, 2012.

[9] Ying Huang and Vigor Yang. Dynamics and stability of lean-premixed swirl-stabilized combustion.

Progress in Energy and Combustion Science, 35(4):293 – 364, 2009.

[10] Martin Kraft, Thomas Eikmann, Andreas Kappos, Nino Kunzli, Regula Rapp, Klaus Schneider,

Heike Seitz, Jens-Uwe Voss, and H.-Erich Wichmann. The german view: Effects of nitrogen dioxide

on human health – derivation of health-related short-term and long-term values. International

Journal of Hygiene and Environmental Health, 208(4):305 – 318, 2005.

[11] Kinga Skalska, Jacek S. Miller, and Stanislaw Ledakowicz. Trends in nox abatement: A review.

Science of The Total Environment, 408(19):3976 – 3989, 2010.

[12] Anita Singh and Madhoolika Agrawal. Acid rain and its ecological consequences. J Environ Biol,

29(1):15–24, Jan 2008.

48

Bibliography 49

[13] Richard T. Burnett, Jeffrey R. Brook, Wesley T. Yung, Robert E. Dales, and Daniel Krewski.

Association between ozone and hospitalization for respiratory diseases in 16 canadian cities. Envi-

ronmental Research, 72(1):24 – 31, 1997.

[14] Bell ML, McDermott A, Zeger SL, Samet JM, and Dominici F. Ozone and short-term mortality in

95 us urban communities, 1987-2000. JAMA, 292(19):2372–2378, 2004.

[15] WILLIAM E. HOGSETT, JAMES E. WEBER, DAVID TINGEY, ANDREW HERSTROM,

E. HENRY LEE, and JOHN A. LAURENCE. Environmental auditing: An approach for char-

acterizing tropospheric ozone risk to forests. Environmental Management, 21(1):105–120, Jan 1997.

[16] David F. Karnosky, John M. Skelly, Kevin E. Percy, and Art H. Chappelka. Perspectives regarding

50years of research on effects of tropospheric ozone air pollution on us forests. Environmental

Pollution, 147(3):489 – 506, 2007. Air Pollution and Climate Change: A Global Overview of the

Effects on Forest Vegetation.

[17] G. Mills, A. Buse, B. Gimeno, V. Bermejo, M. Holland, L. Emberson, and H. Pleijel. A synthesis of

aot40-based response functions and critical levels of ozone for agricultural and horticultural crops.

Atmospheric Environment, 41(12):2630 – 2643, 2007.

[18] Rita Van Dingenen, Frank J. Dentener, Frank Raes, Maarten C. Krol, Lisa Emberson, and Janusz

Cofala. The global impact of ozone on agricultural crop yields under current and future air quality

legislation. Atmospheric Environment, 43(3):604 – 618, 2009.

[19] Shiri Avnery, Denise L. Mauzerall, Junfeng Liu, and Larry W. Horowitz. Global crop yield reduc-

tions due to surface ozone exposure: 1. year 2000 crop production losses and economic damage.

Atmospheric Environment, 45(13):2284 – 2296, 2011.

[20] Shiri Avnery, Denise L. Mauzerall, Junfeng Liu, and Larry W. Horowitz. Global crop yield reduc-

tions due to surface ozone exposure: 2. year 2030 potential crop production losses and economic

damage under two scenarios of o3 pollution. Atmospheric Environment, 45(13):2297 – 2309, 2011.

[21] Basil Dimitriades. Effects of hydrocarbon and nitrogen oxides on photochemical smog formation.

Environmental Science & Technology, 6(3):253–260, 1972.

[22] B Brunekreef. Air pollution and life expectancy: is there a relation? Occupational and Environ-

mental Medicine, 54(11):781–784, 11 1997.

[23] C A Pope. Epidemiology of fine particulate air pollution and human health: biologic mechanisms

and who’s at risk? Environmental Health Perspectives, 108(Suppl 4):713–723, 08 2000.

[24] C. Arden Pope, Majid Ezzati, and Douglas W. Dockery. Fine-particulate air pollution and life

expectancy in the united states. New England Journal of Medicine, 360(4):376–386, 2009. PMID:

19164188.

[25] Daniel Krewski. Evaluating the effects of ambient air pollution on life expectancy. New England

Journal of Medicine, 360(4):413–415, 2009. PMID: 19164194.

[26] Ki-Hyun Kim, Ehsanul Kabir, and Shamin Kabir. A review on the human health impact of airborne

particulate matter. Environment International, 74:136 – 143, 2015.

Bibliography 50

[27] Nicholas Syred. The effect of hydrogen containing fuel blends upon flashback in swirl burners.

Applied energy, v. 89(no. 1):pp. 106–110–2012 v.89 no.1, 2012.

[28] Christian Eichler and Thomas Sattelmayer. Experiments on flame flashback in a quasi-2d turbulent

wall boundary layer for premixed methane-hydrogen-air mixtures. Journal of Engineering for Gas

Turbines and Power, 133(1):011503–011503–7, 09 2010.

[29] T.C. Claypole and N. Syred. The effect of swirl burner aerodynamics on nox formation. Symposium

(International) on Combustion, 18(1):81 – 89, 1981. Eighteenth Symposium (International) on

Combustion.

[30] Tim C. Lieuwen. Unsteady Combustor Physics. Cambridge University Press, 2012.

[31] Thibault F. Guiberti, Daniel Durox, Laurent Zimmer, and Thierry Schuller. Analysis of topol-

ogy transitions of swirl flames interacting with the combustor side wall. Combustion and Flame,

162(11):4342 – 4357, 2015.

[32] N. Syred and J. M. Beer. Combustion in swirling flows: A review. Combustion and Flame, 23(2):143–

201, 1974.

[33] David G. Lilley. Swirl flows in combustion: A review. AIAA Journal, 15(8):1063–1078, 1977.

[34] J. 0. Keller, L. Vaneveld, D. Korschelt, G. L. Hubbard, A. F. Ghoniem, J. W. Daily, and A. K.

Oppenheim. Mechanism of instabilities in turbulent combustion leading to flashback. AIAA Journal,

20(2):254–262, 2018/08/15 1982.

[35] J. Fritz, M. Kroner, and T. Sattelmayer. Flashback in a swirl burner with cylindrical premixing

zone. Journal of Engineering for Gas Turbines and Power, 126(2):276–283, 06 2004.

[36] David R. Noble, Qingguo Zhang, Akbar Shareef, Jeremiah Tootle, Andrew Meyers, and Tim

Lieuwen. Syngas mixture composition effects upon flashback and blowout. (42363):357–368, 2006.

[37] F. Kiesewetter, M. Konle, and T. Sattelmayer. Analysis of combustion induced vortex breakdown

driven flame flashback in a premix burner with cylindrical mixing zone. Journal of Engineering for

Gas Turbines and Power, 129(4):929–936, 04 2007.

[38] Marco Konle, Frank Kiesewetter, and Thomas Sattelmayer. Simultaneous high repetition rate

piv–lif-measurements of civb driven flashback. Experiments in Fluids, 44(4):529–538, Apr 2008.

[39] M. Konle and T. Sattelmayer. Time scale model for the prediction of the onset of flame flashback

driven by combustion induced vortex breakdown. Journal of Engineering for Gas Turbines and

Power, 132(4):041503–041503–6, 01 2010.

[40] Bernard Lewis and Guenther von Elbe. Stability and structure of burner flames. The Journal of

Chemical Physics, 11(2):75–97, 1943.

[41] Abbott A. Putnam and Randolph A. Jensen. Application of dimensionless numbers to flash-back and

other combustion phenomena. Symposium on Combustion and Flame, and Explosion Phenomena,

3(1):89 – 98, 1948. Third Symposium on Combustion and Flame and Explosion Phenomena.

Bibliography 51

[42] R. Edse. Studies on burner flames of hydrogen-oxygen mixtures at high pressures. WADC Technical

Report, pages 52–59, 1952. cited By 2.

[43] K. Wohl. Quenching, flash-back, blow-off-theory and experiment. Symposium (International) on

Combustion, 4(1):68–89, 1953.

[44] B. Fine. The flashback of laminar and turbulent burner flames at reduced pressure. Combustion

and Flame, 2(3):253–266, 1958.

[45] Christian Eichler, Georg Baumgartner, and Thomas Sattelmayer. Experimental investigation of

turbulent boundary layer flashback limits for premixed hydrogen-air flames confined in ducts.

(54624):389–398, 2011.

[46] Christian Eichler and Thomas Sattelmayer. Premixed flame flashback in wall boundary layers

studied by long-distance micro-piv. Experiments in Fluids, 52(2):347–360, Feb 2012.

[47] A. Gruber, J. H. Chen, D. Valiev, and C. K. Law. Direct numerical simulation of premixed flame

boundary layer flashback in turbulent channel flow. Journal of Fluid Mechanics, 709:516–542, 2012.

[48] Georg Baumgartner, Lorenz R. Boeck, and Thomas Sattelmayer. Experimental investigation of

the transition mechanism from stable flame to flashback in a generic premixed combustion system

with high-speed micro-particle image velocimetry and micro-plif combined with chemiluminescence

imaging. Journal of Engineering for Gas Turbines and Power, 138(2):021501–021501–10, 09 2015.

[49] Dominik Ebi and Noel T. Clemens. Experimental investigation of upstream flame propagation

during boundary layer flashback of swirl flames. Combustion and Flame, 168:39 – 52, 2016.

[50] C. Heeger, R. L. Gordon, M. J. Tummers, T. Sattelmayer, and A. Dreizler. Experimental analysis

of flashback in lean premixed swirling flames: upstream flame propagation. Experiments in Fluids,

49(4):853–863, Oct 2010.

[51] J. Sangl, C. Mayer, and T. Sattelmayer. Dynamic adaptation of aerodynamic flame stabilization

of a premix swirl burner to fuel reactivity using fuel momentum. Journal of Engineering for Gas

Turbines and Power, 133(7):071501–071501–11, 03 2011.

[52] C. Mayer, J. Sangl, T. Sattelmayer, T. Lachaux, and S. Bernero. Study on the operational window

of a swirl stabilized syngas burner under atmospheric and high pressure conditions. Journal of

Engineering for Gas Turbines and Power, 134(3):031506–031506–11, 01 2012.

[53] Alireza Kalantari, Elliot Sullivan-Lewis, and Vincent McDonell. Flashback propensity of turbulent

hydrogen–air jet flames at gas turbine premixer conditions. Journal of Engineering for Gas Turbines

and Power, 138(6):061506–061506–8, 11 2015.

[54] Vera Hoferichter, Christoph Hirsch, and Thomas Sattelmayer. Prediction of confined flame flashback

limits using boundary layer separation theory. Journal of Engineering for Gas Turbines and Power,

139(2):021505–021505–10, 09 2016.

[55] Vera Hoferichter, Christoph Hirsch, and Thomas Sattelmayer. Analytic prediction of unconfined

boundary layer flashback limits in premixed hydrogen–air flames. Combustion Theory and Modelling,

21(3):382–418, 2017.

Bibliography 52

[56] Vera Hoferichter, Christoph Hirsch, Thomas Sattelmayer, Alireza Kalantari, Elliot Sullivan-Lewis,

and Vincent McDonell. Comparison of two methods to predict boundary layer flashback limits of

turbulent hydrogen-air jet flames. Flow, Turbulence and Combustion, 100(3):849–873, Apr 2018.

[57] Markus Raffel, Christian E. Willert, S. Wereley, and Jurgen Kompenhans. Particle Image Velocime-

try : A Practical Guide. Springer, Berlin, Heidelberg, GERMANY, 2007.

[58] C.T. Crowe, J.N. Chung, and T.R. Troutt. Particle mixing in free shear flows. Progress in Energy

and Combustion Science, 14(3):171 – 194, 1988.

[59] A. C. Eckbreth, P. A. Bonczyk, and J. F. Verdieck. Laser raman and fluorescence techniques for

practical combustion diagnostics. Applied Spectroscopy Reviews, 13(1):15–164, 1977.

[60] Nobuyuki Otsu. A threshold selection method from gray-level histograms. IEEE transactions on

systems, man, and cybernetics, 9(1):62–66, 1979.

[61] Joseph Grumer. Predicting burner performance with interchanged fuel gases. Industrial & Engi-

neering Chemistry, 41(12):2756–2761, 1949.

[62] Norman Cohen. Flammability and explosion limits of h2 and h2/co: a literature review. Technical

report, AEROSPACE CORP EL SEGUNDO CA TECHNOLOGY OPERATIONS, 1992.

[63] David G. Goodwin, Harry K. Moffat, and Raymond L. Speth. Cantera: An object-oriented software

toolkit for chemical kinetics, thermodynamics, and transport processes. http://www.cantera.org,

2018. Version 2.4.0.

[64] Gregory P. Smith, David M. Golden, Michael Frenklach, Nigel W. Moriarty, Boris Eiteneer, Mikhail

Goldenberg, C. Thomas Bowman, Ronald K. Hanson, Soonho Song, Jr. William C. Gardiner,

Vitali V. Lissianski, and Zhiwei Qin. Gri-mech 3.0. http://www.me.berkeley.edu/gri mech/.

[65] Nicolas Bouvet, Fabien Halter, Christian Chauveau, and Youngbin Yoon. On the effective lewis

number formulations for lean hydrogen/hydrocarbon/air mixtures. International Journal of Hydro-

gen Energy, 38(14):5949 – 5960, 2013.

by Christopher E. Schneider · 2019-03-25 · Abstract Incipient Behaviour of Flashback in a Lean...

Documents

Transcript of by Christopher E. Schneider · 2019-03-25 · Abstract Incipient Behaviour of Flashback in a Lean...