International journal of automotive technology, tập 12, số 6, 2011

The PM deposit mass per unit area andeffectiveness drop had maximum values at a coolant temperature of 40oC for every n-dodecane injection rate.. The variation in the deposit mass per un

International Journal of Automotive Technology, Vol 12, No 6, pp 787−794 (2011)

DOI 10.1007/s12239−011−0091−z

Copyright © 2011 KSAE

1229 −9138/2011/061−01

787

COMBUSTION INSTABILITIES AND NANOPARTICLES EMISSION

FLUCTUATIONS IN GDI SPARK IGNITION ENGINE

1)Department of Automotive Technology, Helwan University, Cairo, Egypt

2)School of Technology, Oxford Brookes University, OX33 1HX, UK

(Received 26 October 2010; Revised 16 May 2011)

ABSTRACT−The main challenge facing the concept of gasoline direct injection is the unfavourable physical conditions atwhich the premixed charge is prepared and burned These conditions include the short time available for gasoline to besprayed, evaporated, and homogeneously mixed with air These conditions most probably affect the combustion process andthe cycle-by-cycle variation and may be reflected in overall engine operation The aim of this research is to analyze thecombustion characteristics and cycle-by-cycle variation including engine-out nanoparticulates of a turbocharged, gasolinedirect injected spark ignition (DISI) engine at a wide range of operating conditions Gasoline DISI, turbo-intercooled, 1.6L,

4 cylinder engine has been used in the study In-cylinder pressure has been measured using spark plug mounted piezoelectrictransducer along with a PC based data acquisition A single zone heat release model has been used to analyze the in-cylinderpressure data The analysis of the combustion characteristics includes the flame development (0-10% burned mass fraction)and rapid burn (10-90% burned mass fraction) durations at different engine conditions The cycle-by-cycle variations havebeen characterized by the coefficient of variations (COV) in the peak cylinder pressure, the indicated mean effective pressure(IMEP), burn durations, and particle number density The combustion characteristics and cyclic variability of the DISI engineare compared with data from throttle body injected (TBI) engine and conclusions are developed

KEY WORDS : DISI engine, Cyclic variability, Combustion characteristics, Nanoparticulates

1 INTRODUCTION

The instabilities in the combustion processes were

observed from the very beginning of spark ignition engine

development (Clerk, 1886) These instabilities were

identified as a fundamental combustion problem and it may

cause fluctuations in the flame propagation pattern, the

burned fuel mass, the indicated mean effective pressure,

and consequently the power output in SI engines (Patterson,

1966)

The main factors of combustion instabilities as classified

by Heywood (1988) are aerodynamics in the cylinder

during combustion, amounts of fuel, air, and recycled

exhaust gases supplied to the combustion chamber, and

composition of local mixture near the spark plug

These factors will affect the combustion characteristics

and lead to a significant cycle-by-cycle variation in the

combustion process Furthermore, a disturbing feature of

these combustion instabilities is the unpredictable character

of their occurrence (Sawamoto et al., 1987; Wagner et al.,

1993; Eriksson et al., 1997; Muller et al., 2001; Matsumoto

et al., 2007).

Direct injection spark ignition engine (DISI) technology

has been proved to be a good potential for automobileengines The advantages of DISI engines reflect in theirhigher thermal efficiency due to the higher volumetricefficiency of the unthrottled charge, better potential forreducing the specific fuel consumption, and better control

of injection quantity and timing (Gao et al., 2005).

Significant advancements have been made in recent years inthe development of combustion systems for DISI engineswhich have resulted in larger fuel economy benefits, betterexhaust emissions, and significant power advantagescompared to throttle-body (TB) and port fuel (PF) injected

engines (Kume et al., 1996; Ando; 1996, Jackson et al., 1996; Itoh et al., 1998; Geiger et al., 1999)

DISI engines, however, have their drawbacks due to thetime limitations and gasoline direct injection characteristicswhich may affect the fuel evaporation, charge homogeneity,and the stability of the entire combustion process whichlead to an intense cyclic variability Moreover, it is difficult

to know exactly what type of flame is propagating for anyparticular ignition event in a DISI engine due to the directfuel injection and the highly turbulent motion inside thecombustion chamber The cyclic variability in gasolineengine which were investigated by many researchers affectsthe engine fuel economy and it may decrease the mean

effective pressure by as much as 20% (Litak et al., 2009) Brown et al (1996), concluded on their work that the

*Corresponding author e-mail: geushey@hotmail.com

788 A E HASSANEEN, S SAMUEL and I WHELAN

cycle-by-cycle variation in combustion should be

characterized by the coefficient of variation (COV) of the

indicated mean effective pressure (IMEP) in preference to

the COV of the in-cylinder peak pressure They also added

that cycle-by-cycle variations are lower when the early

combustion is more rapid They also found that the COV of

IMEP is a minimum in the region of MBT ignition timing

Hinze and Cheng (1998) concluded that the variations in

flow field and the inhomogeneous charge in SI engines

contribute to almost 50% of the cycle-by-cycle variations

of IMEP

This paper presents a study of the combustion

characteristics and cycle-by-cycle variations of a

turbo-intercooled gasoline direct injection (DISI) engine at

different engine speeds and loads These characteristics

include the in-cylinder peak pressure, the indicated mean

effective pressure (IMEP), burn durations, and the coefficient

of variation of these parameters and the nanoparticulates

emission from cycle-to-cycle These characteristics will be

compared to those of an older throttle body injected (TBI)

engine

2 EXPERIMENTAL APPARATUS

2.1 Engine

The DISI engine used for this research was a 1.6L, four

cylinders, direct-injection, four-stroke, water cooled,

turbocharged and intercooled engine The TBI engine was

a 1.4L, four cylinders, naturally aspirated, four-stroke, and

water cooled engine A list of the complete engine

specifications for the DISI and TBI engines are provided in

Table 1 and Table 2 respectively A schematic diagram of

the test rig is shown in Figure 1 Both engines were

designed to run at stoichiometric air-to-fuel ratio for the

sake of efficient performance of the catalytic converter

The DISI engine utilises wall guided direct-injection(WGDI) at a pressure of 120bar WGDI engines introducethe fuel into the combustion chamber via a side mountedswirl injector The mixture is then guided toward the sparkplug by the reverse tumble turbulent motion and the bowlshaped pocket in the piston crown At this fuel injectionpressure, the spray velocities are on average an order ofmagnitude faster than piston velocities and spraypenetration distances are of the same order as the stroke Therefore, it is inevitable that the fuel spray will impactupon the piston crown and cylinder walls Any form of fuelimpingement on the walls of the combustion chamber cancause lubricating oil to mix with the bulk gas This canresult in increased PM and HC emissions

Most wall wetting occurs when the fuel injection isadvanced, i.e when the piston is near top dead centre onthe intake stroke The degree of fuel evaporation fromthese surfaces decreases with injection retard With theoptimised flow motion in the combustion chamber ofmodern DISI engines, the amount of wall wetting may bereduced in relation to the first generation DISI engines

2.2 Particulates Measurements

A Differential Mobility Spectrometer, DMS-500, was used

to analyse the exhaust gas sample The DMS-500 provides

a number and size spectrum for particles between 5-1000

nm Particles above 1000 nm are removed by a cycloneseparator upstream to reduce the need for cleaning TheDMS-500 consists of a classifier column (consisting of 22grounded electrometer rings), a high voltage electrode,space charge guide and a conductive tube The classifiercolumn operates at sub-atmospheric pressure, obtainedusing a scroll vacuum pump Within the conductive tube,the particles become ionised, i.e charged The particles aretherefore classified according to their charge to drag ratio

2.3 Test MethodAll the tests have been carried out at steady-stateconditions at fully warmed-up temperatures (coolanttemperature is 83oC, oil temperature is 89oC) Due to thedynamometer loading limitations, the engine operating

Table 1 DISI test engine specifications

COMBUSTION INSTABILITIES AND NANOPARTICLES EMISSION FLUCTUATIONS 789

points used for the statistical analysis were 1600, 2400 and

3200 RPM and a load range of 20-120 Nm (1.57 – 9.42 bar

BMEP) in increments of 20 Nm For the in-cylinder

pressure analysis a single zone heat release model based on

the first law of thermodynamics without heat transfer was

employed in the present work Average cycle obtained over

100 engine cycles were used for the heat release analysis

DISI and GDI refer to the same gasoline direct injected

engine elsewhere in the paper

3 HEAT RELEASE MODEL

The heat release analysis is based on a single-zone model

in which the burned and unburned zones in the combustion

chamber are treated as a single zone The model is based on

the first law of thermodynamics applied to the in-cylinder

control volume as follow (Heywood, 1988):

dQ hr − dQ hl = dU + dW

Where dQ hr represents the fuel chemical energy released,

dQ hl represents the heat loss to the cylinder walls, dU and

dW represent the change in the sensible internal energy and

the work done on the piston respectively Using the

thermodynamic relationships and neglecting the heat losses

term, the above equation of the first law is simplified to the

following form:

Where V and P is the instantaneous in-cylinder volume

and pressure respectively, n is the polytropic exponent

4 RESULTS AND DISCUSSION

4.1 Fuel Economy and Thermal Efficiency

The motivations behind the use of direct injection ofgasoline in spark ignition engines are mainly its better fueleconomy as it appears in Figure 2 where brake specific fuelconsumption and thermal efficiency almost approach thosevalues of the conventional Diesel engines especially at partloads (Heywood, 1988; Roy, 2011)

It is shown from the figures that fuel consumptiondropped from 450 (g/kW.hr) at low loads to around 230 (g/kW.hr) at high loads It seems from the figures that a finercontrol of ignition timing is needed to eliminate thedifferences between the fuel consumption values at differentspeeds especially at the middle load range

4.2 Combustion CharacteristicsThe combustion characteristics include the analysis of in-cylinder peak pressure, indicated mean effective pressure,and burn durations The in-cylinder peak pressure, itslocation and rate of rise at different engine loads and

Figure 2 Fuel economy and brake thermal efficiency (GDI

790 A E HASSANEEN, S SAMUEL and I WHELAN

speeds are shown in Figures 3, 4, and 5

It can be seen from the figures that minimal variation in

the peak pressure and its rate of rise is maintained at most

of the engine speeds except at very high loads This

minimal variation may be attributed to the better control of

the air-to-fuel ratio in the case of GDI engines At high

loads, this better control of air-to-fuel ratio is challenged by

the time limitation for the huge amount of fuel to be

vaporised and well mixed with air

The location of peak pressure is also maintained at the

relatively high engine speed while at low speed it is

retarded as shown in the figure The rate of pressure rise of

the GDI engine was found to be much higher than the rate

of pressure rise of the TBI engine This difference in the

rate of pressure rise may be due to the fact that the mode ofcombustion in GDI engines most likely a mix of stratified,homogeneous, and heterogeneous regimes

In order to achieve a stable operation for a gasolinedirect injection engine, a precise control of ignition timing

is crucial This is evident from the ignition timing chartshown in Figure 4 where a wide range of ignition timing isadopted for better performance

Combustion durations including flame developmentduration (Spark-10% bmf) and rapid burn duration (10-90% bmf) are shown in Figures 4 and 5 The flamedevelopment duration (Spark-10%bmf) vary from 35degree crank angle (oCA) at low loads to 13 (oCA) at highloads These durations are corresponding to 3 ms at 1600rpm at low load and 1.5 ms at high loads The rapid burnFigure 4 Early combustion durations as a function of

COMBUSTION INSTABILITIES AND NANOPARTICLES EMISSION FLUCTUATIONS 791

duration vary from 60 (oCA) at low loads to 20 (oCA) at

high loads In terms of milliseconds, the duration drops

from 6 ms at 1600 rpm and low loads to less than 2 ms at

high loads It can be concluded that increasing the load on

the engine decrease the burn time to 33% from its value at

low loads regardless of engine speed as shown in Figure 5

Comparison between the flame development and rapid

burn duration for the DISI and TBI engines are shown in

Figures 4 and 5 respectively It is shown that the flame

development duration for the DISI engine is longer than

that for the TBI engine at middle to high loads In the very

low and very high range, however, the duration for the

DISI engine is slightly shorter These last findings may be

attributed to the longer time needed for the directly injected

fuel in DISI engine to evaporate and mix with air to form a

homogenous charge

Due to the high throttling and the challenges to control

the A/F ratio at the very low loads in TBI engine, the flame

development duration tends to be larger than DISI engine

The same trend continues to show up at vary high loads

most probably because of the better chance to achieve rich

A/F ratios in DISI engine

4.3 Cyclic Variations in Combustion Parameters

100 engine cycles were acquired and an average cycle was

used for the statistical analysis The coefficient of variation

(COV) in the combustion parameters is shown in Figure 6

and 7

It can be seen in the figure that the COV in peak pressure

is maximum at medium loads at all engine speeds At very

low load, however, COV show some improvement which

may be attributed to the good control of air-to-fuel ratio (A/

F) at low loads in the gasoline direct injected unthrottled

engines

The comparison between DISI and TBI engines show

that COV in peak pressure and IMEP for TBI engine are

generally lower than those of DISI engines as shown in

Figure 6 COV in peak pressure and IMEP of TBI engine at

very low load is higher than DISI engine due to the fact that

A/F ratio in TBI engine is less controlled at low loads

because of the throttle itself and the possible misdistribution

of the charge over the individual cylinder

As the load goes up, the fresh charge is better distributed

over the individual cylinders in TBI engines and throttle

effect becomes smaller which is reflected in an improved

COV In DISI engine, however, as the load goes up, the

amount of fuel injected becomes larger and more time for

its evaporation is needed thus suggesting a pourer mixing

process and most probably inhomogeneous charge is

formed These all may lead to a higher COV in peak

pressure and IMEP for DISI engines at high loads as shown

in Figure 6 and 7 The comparison between the two engines

in terms of the COV in the location where 90% of the mass

is burned show a big difference to the favour of the TBI

engine It reaches a value of 45% for DISI engine and drop

to 15% for TBI engine as can be seen in Figure 7 This

difference may be attributed to the inhomogeneous mixturelocally formed in the DISI engine combustion chamber

4.4 Cyclic Variations in ParticulatesComparisons of the total particle number between the TBIand GDI engines are shown in Figure 8 The total particlenumber density in GDI engine is almost two orders ofFigure 6 COV of combustion parameters as a function ofBMEP

792 A E HASSANEEN, S SAMUEL and I WHELAN

magnitude higher than TBI engine The total particle

number density of GDI engine is very close to the particle

number density of diesel engine

The cycle-by-cycle variation in the main combustion

parameters seems to have a considerable effect on the

fluctuations of the nano-scale particulates as can be seen in

Figures 8 and 9 The variations in the particle number

density for different diameter particles decreases with anincrease in engine load as can be seen in Figure 6 Thevariation in the 10 nm particles was higher than that in the

56 and 100 nm diameter particles at the same loadingconditions Although the COV values of the 10 nmparticles were higher than those for the 56 and 100 nmparticles, its correlation coefficient with the engine brakemean effective pressure (BMEP) was far less than thecorrelations of the other two diameter particles (R2=0.14,

Figure 7 Comparisons of COV in combustion parameters

of the two engines

Figure 8 PM and its COV as a function of MEP

COMBUSTION INSTABILITIES AND NANOPARTICLES EMISSION FLUCTUATIONS 793

0.49, and 0.51 respectively) The same trends were

observed at the various engine speeds as shown in the

figure with less significant correlation coefficients for all

particles

The variations (COV) in the particle number density

with the variation (COV) in the combustion parameters are

presented in Figure 9 The COV of the 10 nm particles

seems to be less correlated to the load condition than the

other two diameter ranges This may be due to the fact that

the higher in-cylinder temperatures at higher loads act as

oxidation factor that compete with the nucleation process

thus producing more stable variation of the 10 nm particles

range

This proposed explanation of the COV in the 10 nm

particles seems to be still valid at different engine speed

and constant load condition which means more stable

nucleation and oxidation process for the whole diameter

ranges That is why COV of the particle number density isless correlated to engine speed than to engine load It isshown that the COV in the particle numbers increases with

an increase in the COV of the combustion parameters withrelatively stronger correlation coefficients with the COV inthe combustion durations of the 10-90% burned massfraction This is again could be attributed to the competitionbetween the nucleation and oxidation processes which areaffected by the fluctuations in the air-to-fuel ratio and in-cylinder temperatures at these different conditions

5 CONCLUSIONS

Two DISI and TBI engines were instrumented and testedfor the combustion characteristics and cycle-by-cyclevariations and conclusions were developed DISI engineachieved the brake thermal efficiency of Diesel engine(36%) The rate of Pressure rise of DISI engine (1.3 – 3.4bar/oCA) is almost double that of TBI engine (1.5 – 2.5 bar/

oCA) The flame development duration (Spark-10% bmf) islonger for DISI engine than TBI engine The rapid burnduration (10-90% bmf) is shorter for DISI engine than TBIengine COV in peak pressure and IMEP for DISI engineare higher than those values for TBI engine at moderateand high loads At low loads, however, COV in bothparameters for DISI engine are lower than the TBI engine.The COV in the particle numbers increases with anincrease in the COV of all these combustion parameterswith relatively stronger correlation coefficients with theCOV in the combustion durations of the 10-90% burnedmass fraction

REFERENCES

Ando, H (1996) Combustion control technologies for

gasoline engines Lean Burn Combustion Engine IMechE

Seminar, Paper S433/001/96, 1996-1920.

Brown, A G., Stone, C R and Beckwith, P (1996) by-cycle variations in spark ignition engine combustion– Part l: Flame speed and combustion measurements and

Cycle-a simplified turbulent combustion model SAE PCycle-aper

Proc IMechE, 219 Part A: J Power and Energy.

Geiger, J., Grigo, M., Lang, O., Wolters, P and Hupperich,

P (1999) Direct injection gasoline engines – Combustion

design SAE Paper No 1999-01-0170.

Hinze, P C and Cheng, W K (1998) Effects of chargeFigure 9 PM and its COV as a function of COV in

combustion parameters

794 A E HASSANEEN, S SAMUEL and I WHELAN

composition on SI engine cyclic variations at idle 4th

Int Symp COMODIA 98.

Itoh, T., Liama, A., Muranaka, S and Tagaki, Y (1998)

Combustion characteristics of a direct-injection stratified

charge S.I engine JSAE Review, 19, 217−222.

J B Heywood (1988) Internal Combustion Engine

Fundamentals McGraw-Hill Int Series, Automotive

Technology Series Boston

Jackson, N S., Stokes, J., Whitaker, P A and Lake, T H

(1996) A direct injection stratified charge gasoline

combustion system for future european passenger cars

1920

Kume, T., Iwamoto, Y., Lida, K., Murakami, M., Akishino,

K and Ando, H (1996) Combustion control

technologies for direct injection SI engines SAE Paper

No 960600.

Litak, G., Kaminski, T., Czarnigowski, J., Sen, A K and

Wendeker, M (2009) Combustion process in a spark

ignition engine: Analysis of cyclic peak pressure and

peak pressure angle oscillations Meccanica, 44, 1−11.

Matsumoto, K., Tsuda, I and Hosoi, Y (2007) Controllingengine system: A low-dimensional dynamics in a spark

ignition engine of a motorcycle Z Naturforsch, 62a,

587−595

Muller, R., Hemberger, H and Baier, K H (2001) Enginecontrol using neural networks; A new method in engine

management systems Meccanica, 32, 423−430

Patterson, D J (1966) Cylinder pressure variations, a

fundamental combustion problem SAE Paper No.

660129

Roy, M R (2011) Performance and emissions of a dieselengine fueled by diesel-biodiesel blends with special

attention to exhaust odor Canadian J Mechanical

Sciences and Engineering 2, 1, 1−10

Sawamoto, K., Kwamura, Y., Kita, T and Matsushita, K.(1987) Individual cylinder knock control by detecting

cylinder pressure SAE Paper No 871911.

Wagner, R M., Daw, C S and Thomas, J F (1993)

Controlling chaos in spark-ignition engines Proc.

Central and Eastern States Joint Technical Meeting of

International Journal of Automotive Technology, Vol 12, No 6, pp 795−812 (2011)

Department of Mechanical Design Engineering, Chonnam National University, Chonnam 550-749, Korea

(Received 24 September 2009; Revised 26 June 2011)

ABSTRACT−Large-Eddy Simulation (LES) was used to perform computations of air entrainment and mixing during diesel

spray combustion The results of this simulation were compared with those of Reynolds Averaged Navier Stokes (RANS)simulations and an experiment The effect of LES on non-vaporizing and vaporizing sprays was evaluated The validity of thegrid size used for the LES analysis was confirmed by determining the subgrid-scale (SGS) filter threshold on the turbulentenergy spectrum plot, which separates a resolved range from a modeled one The results showed that more air was entrainedinto the jet with decreasing ambient gas temperatures The mass of the evaporated fuel increased with increasing ambient gastemperatures, as did the mixture fraction variance, showing a greater spread in the profile at an ambient gas temperature of 920

K than at 820 K Flame lift-off length sensitivity was analyzed based on the location of the flame temperature iso-line Theresults showed that for the flame temperature iso-line of 2000oC, the computed lift-off length values in RANS matched theexperimental values well, whereas in LES, the computed lift-off length was slightly underpredicted The apparent heat releaserate (AHRR) computed by the LES approach showed good agreement with the experiment, and it provided an accurateprediction of the ignition delay; however, the ignition delay computed by the RANS was underpredicted Finally, therelationships between the entrained air quantity and mixture fraction distribution as well as soot formation in the jet wereobserved As more air was entrained into the jet, the amount of air-fuel premixing that occurred prior to the initial combustionzone increased, upstream of the lift-off length, and therefore, the soot formation downstream of the flame decreased

KEY WORDS : Large eddy simulation (LES), Diesel spray, Air entrainment, Diesel combustion, Lift-off length

1 INTRODUCTION

Large-Eddy Simulation (LES) is a relatively new research

field Much research has been carried out over the past

years, but to realize the full predictive potential of LES,

many fundamental questions still have to be addressed In

LES, the major part of the turbulent kinetic energy is

resolved directly, whereas the effects of remaining scales

smaller than the computational grid size are accounted for

in a subgrid-scale (SGS) model (Lesieur and Metais, 1996;

Lesieur, 2005) To compute fluctuating quantities

(temperature, velocity and pressure), the technique consists

of calculating instant fields in a transient calculation, which

solves the Navier-Stokes equations However, flow, being

generally turbulent, cannot be solved explicitly at all scales

Resolving the Kolmogorov scale (size of the smaller eddies)

in three-dimensional calculations is out of the reach of

present computers and will remain so for a long time The

large eddy simulation technique can solve large eddies

explicitly and model smaller eddies Compared to the

well-known Reynolds-Averaged Navier-Stokes (RANS)

approach, the universality of LES is higher because LES

model assumptions are made only about the subgrids,

which occur at energy-negligible scales of the turbulentflow Thus, the LES approach has the advantages of boththe Direct Numerical Simulation (DNS), with respect touniversality and accuracy to a physical experiment, and theRANS, with respect to modeling efficiency and handling ofhigh Reynolds numbers compared with DNS For use in thenear future, LES may be considered the most promisingapproach to provide an accuracy level unattainable byRANS modeling

The LES of chemically reacting turbulent flow hasbecome a topic of much interest in recent years The LES ofturbulent combustion has already been applied to a variety

of combustion problems, including predictions of pollutantsand engine combustion However, much of the necessarytheory for combustion LES has yet to be developed, and thefull predictive potential of combustion LES has not yetbeen reached

In turbulent combustion at high Reynolds and Damkohlernumbers, the essential rate-controlling processes of molecularmixing and chemical reactions occur at the smallest scales

In non-premixed diffusion combustion regimes, for example,these coupled processes occur in the reactive-diffusionlayers, which are much thinner than the resolved scales.Hence, the rate-controlling processes do not occur in theresolved, large scales and have to be modeled However, a

*Corresponding author e-mail: sngkim@chonnam.ac.kr

796 U B AZIMOV and K S KIM

wide range of length and time scales characterizes

non-premixed combustion The length scale varies from the

smallest Kolmogorov scale to the largest integral scale

compatible with the flow geometry (Peters, 2000) The

chemical reaction time scales span a wide range in

particular, when chemical reactions occur in low temperature

conditions (Aceves and Flowers, 2005)

The application of LES to non-premixed combustion is

motivated by a large amount of evidence demonstrating

that mixing rates are controlled by large-scale eddies and

that the scalar mixing process is of paramount importance

to chemical conversion (Poinsot and Veynante, 2001)

Studies on non-reactive and reactive systems showed that

LES predicts the scalar mixing process and dissipation

rates with considerably better accuracy than RANS,

especially for complex flows (Pitsch, 2002; Pitsch and

Steiner, 2000; Raman and Pitsch, 2005) Turbulent mixing

controls most of the global flame properties In LES,

unsteady large scale mixing between the fuel and the

oxidizer in a non-premixed flame is simulated, instead of

being averaged With LES, large structures are explicitly

computed, and instantaneous fresh- and burnt- gas zones

with different turbulence characteristics are clearly

identified This can help to describe some of the properties

of flame/turbulence interactions Additional motivation is

provided by the need to simulate unsteady flows, such as

the spray combustion processes in a diesel engine; LES is

well suited to simulate unsteady combustion problems

because it yields time-accurate information

As was mentioned above, the LES approach resolves all

scales larger than the grid size but models the effects of

scales below grid resolution by imposing a spatial average

on the flow field to yield the subgrid terms that have to be

modeled Various subgrid-scale (SGS) models such as the

Smagorinsky model (Smagorinsky, 1963; Moin et al., 1991;

Germano et al., 1991), one-equation viscosity model

(Menon et al., 1996), subgrid k model (Sone and Menon,

2003; Menon, 2000), scale similarity model (Bardina et al.,

1980) and one-equation dynamic structure model

(Pomraning and Rutland, 2002; Chumakov and Rutland,

2004) have been proposed and tested LES using different

SGS models were applied to study turbulent spray

combustion processes using various combustion models

Pope (2004) raised ten conceptual questions concerning the

LES of turbulent flows These questions have been endorsed

by the research community for further consideration He

addressed such issues as the validity of using LES over the

RANS approach, the tractability of all scales, the computer

power for LES, the dependence of flow statistics on

turbulence-resolution length scale, the relationship between

filtered velocity field and resolved velocity field, the

assessment of different LES models, etc De Villiers et al.

(2004) applied the combined LES with the Volume of Fluid

(VOF) technique to simulate the primary breakup of diesel

sprays The simulations have clearly shown that large-scale

disintegration of the jet results from coherent wave growth,

in conformity with the Kelvin-Helmholtz theory Theyconcluded that this disintegration is associated withperturbation of the jet interface by two effects: turbulenteddies generated in the nozzle and acceleration of theinterface via velocity profile relaxation as the liquid leavesthe nozzle Their combined effect produces three-dimensional waves on the jet surface, which rapidly growdue to aerodynamic interaction with the ambient gas

Kimura et al (2004) performed a large eddy simulation of

turbulent mixing in transient circular gas jets and particleladen jets The mean velocity and turbulent intensity ofsimulated gas jets with different filter widths werecompared to those of experimental gas jets The resultsshowed that the filter around the axis, which was modified

to the radial or axial grid scale level, is effective at reducingthe overestimation of grid scale turbulent diffusion and that

it improves the stability of the calculation around thecentral axis of the jet The calculated distributions ofparticle mass concentrations according to the velocityvector of the gas phase indicated that the LES used in thisstudy can express the accumulation of particles in theperiphery of organized vortices in a particle laden jet Thisfinding suggests that LES has the potential to simulate the

diffusion process of fuel droplets in a diesel spray Hori et

al (2006) performed a three-dimensional LES of

non-evaporative and non-evaporative diesel sprays in a constantvessel using the KIVA-LES code They found that the LESresults depended on the grid size and were in goodagreement with the experiment at only the fine grid Theresults showed that in an evaporative spray simulation byRANS, the distributions of the equivalence ratio and gastemperature in a diesel spray were in layers from the sprayaxis, and the fluctuation of the turbulent flow was notcaptured In contrast, LES could directly capture theturbulent fluctuation in the flow field, and the irregulardistributions of the equivalence ratio and gas temperaturewere obtained Jhavar and Rutland (2006) applied LES toHCCI-type early injection engine simulations using KIVA-3V code Their simulation results for different injectiontimings showed that LES was able to capture more detailedflow structures and more accurately represented localmixing They also showed that with the finer grid, the LESmodels were able to capture more flow structures, unlikethe RANS model The finer grid showed higher rates offuel-air mixing and evaporation LES models have alsobeen coupled with the KIVA-CHEMKIN code.Simulations have shown that the CHEMKIN-LES versiongives results that better match the experimental data thanthe CHEMKIN-RANS version because LES is capable ofrepresenting fuel-air mixing and flow structure more

accurately than RANS Bianchi et al (2007) investigated

the effect of nozzle flow conditions on liquid jet atomizationusing a three-dimensional LES They showed that the finalstage of atomization is the ligament formation Additionally,the Rayleigh-type and base break up mechanisms arisingfrom aerodynamic interactions with surrounding gases were

LARGE-EDDY SIMULATION OF AIR ENTRAINMENT DURING DIESEL SPRAY COMBUSTION 797

predicted These authors found that the characteristics

droplet and ligament size spectra were almost unaffected

by the nozzle flow regime The nozzle flow regime was

found to affect the rate of liquid jet surface breakup

LES was performed on reacting sprays and flames

Hawkes and Cant (2001) applied the flame surface density

(FSD) based LES model of turbulent premixed combustion

to generate results for three-dimensional simulations of a

turbulent propagating flame The results confirmed that the

turbulent motion caused the wrinkling of the resolved

flame surface The resolved strain term was shown to have

a complicated structure, which was dependent on the local

resolved flow features Hawkes and Cant (2001) also stated

that although the effects of resolved strain decreased with

increased turbulent intensity, these effects remained an

important contribution even at higher turbulence

intensities Lee et al (2002) compared the results for the

KIVA-LES scalar flux with experimental data for a

non-reacting jet Good agreement for the statistics of the

mixture fraction was obtained The shape of the sub-grid

scalar dissipation rate was in better agreement with

experimental data than the RANS model In addition, the

LES model was used in conjunction with the PDF

combustion model to simulate the diesel combustion

process The pressure curve accurately matched the

experimental data, indicating the potential of LES for

engine combustion Kaario et al (2003) compared LES

and RANS models in DI diesel engines They used the κ-ε

RNG turbulence model and the LATCT combustion model

and concluded that the LES model gave more flow

structures than the RANS model They stated that

approximately 50 to 60% of the total turbulent kinetic

energy was resolved However, the maximum temperatures

and NOx emissions were not accurately predicted Pitsch

(2006) performed an extensive review of the LES of

turbulent combustion He argued and demonstrated that

LES clearly offers advantages for accurate and predictive

simulations of turbulent combustion He claimed that in the

future, some common practices of combustion LES have to

be revisited He pointed out that for nonpremixed

combustion, the models for the scalar dissipation rate and

the scalar variance as well as the mixing models and

computational efficiency have to be improved He also

recommended that numerical discretization schemes be

revisited and that the effects of explicit versus implicit

filtering be assessed Veynante (2006) reviewed the LES of

turbulent combustion and addressed several issues, such as

filtering and balance equations, unresolved flux modeling,

turbulent premixed modeling, turbulent nonpremixed

modeling, subgrid scale dynamic modeling and numerical

requirements for LES He also briefly presented the

comparison between experimental data and numerical

results in addition to some recent practical examples Hu

and Rutland (2006) performed flamelet modeling with LES

for a diesel engine They implemented a flamelet time scale

combustion model, integrated with dynamic structure LES

models for subgrid stress and scalar mixing and with azero-equation conditional scalar dissipation model, into theKIVA code to simulate the combustion of a turbulentreacting jet and a diesel engine They showed that thedistributions of the key parameters of combustion: meanmixture fraction, mixture fraction variance and scalardissipation rate was reasonable In addition, they developedand applied a new model for analyzing the conditionalscalar dissipation rate The results showed that the shape ofthe conditional scalar dissipation rate in the engine wasdependent on the relevant physical and chemical processes

inside the cylinder Hori et al (2007) applied an LES

model to simulate diesel spray combustion with the Dissipation model using KIVA-LES They found that theresults strongly depended on the grid size and showed thatwith the larger grid size of 2 mm, the temperaturedistribution was unrealistic With the smaller grid size of0.5 mm, the unsteady motion of the diesel spray flame waswell captured; however, the heat release rate for the casewith the grid size of 0.5 mm differed greatly from the

Eddy-experimental data Hori et al (2007) recommended that an

additional grid refinement is needed to examine the gridsensitivity Li and Kong (2008) performed dieselcombustion modeling using an LES model with detailedchemistry They used the LES approach consisting of adynamic structure model for the subgrid scale stress tensorand a gradient model for the subgrid scale scalar flux Theycoupled these two models with detailed chemical kineticsfor diesel spray combustion and emission simulations Thisapproach predicted the overall performance of the engine,including the cylinder pressure history, heat release ratedata, and soot and NOx emission data, with respect to theinjection timing and EGR levels

2 ANALYSIS FORMULATION

Diesel spray combustion is typically controlled by the rate

of mixing The mixing process involved in turbulent jetflow may be separated into a number of processes (Chigier,1981) The first is the large-scale engulfing process, inwhich large vortices entrain the surrounding fluid Once thefluid streams are interflowed on a macro-scale, the smallerscale eddies promote micro-scale mixing, finally resulting

in molecular dissipation Molecular processes are restricted

to small spatial scales, while turbulent mechanisms depend

on the eddy currents arising from the larger scales (Turns,2000)

In turbulent combustion, the rate controlling processes

do not occur in the resolved large scales but instead have to

be modeled Although the essential rate controllingprocesses of molecular mixing and chemical reaction occur

at the smallest scales, and combustion occurs in diffusion layers that are much thinner than the resolvedscales The effect of turbulent mixing must not beunderestimated Mixing in combustion is important notonly to ensure the chemical reactions, but it also affects the

reactive-798 U B AZIMOV and K S KIM

formation of the flame as a whole (Han and Mungal, 2001),

can limit the energy release rate, and impact pollutant

formation (Borman and Ragland, 1998) The largest eddy

extending across the full width of the shear layer has

control over the process and, therefore, is responsible for

the transfer of mass and momentum across and along the

shear flow The longer air-fuel mixing time greatly affects

the chemistry rate in the regions where premixed ignition

takes place As the jet expands to the surrounding ambient

gas, the propagation of the jet and the mixing of air and

fuel are initially controlled by large eddies, which

transform into small eddies Initially the large eddies have

control (prevail) over small eddies, affecting the mixing

pattern, and consequently the progress of combustion In

such a case, using the LES approach would be necessary to

accurately predict flow properties For example, the

formation of soot and NOx depends on the fuel

concentra-tion and temperature in the heterogeneous structure of a

diesel jet To minimize soot formation in diesel jets, the

local equivalence ratio must be maintained below 2 by

increasing the time for mixing the ambient gas with fuel

vapor The longer the time it takes for the fuel jet to

penetrate into the hot ambient gas environment, the wider

its boundary extends and the more diluted it becomes in the

region downstream, around its tip This is because more hot

ambient gas entrains into the jet, increasing the fuel

vaporization rate and mixing with the fuel vapor In the

vicinity of a fuel jet, high local strain rates exist in the flow,

corresponding to high local velocity gradients Associated

with these velocity gradients are high local rates of

inter-diffusion of species and temperature across the reaction

zone of the flame Therefore, LES can predict the unsteady

characteristics of the diesel jet with the anisotropic large

eddy-containing vortex components in the structure of the

flame and capture a three-dimensional and instantaneous

turbulent flow field

The objective of this work is to study the effect of

air-fuel mixing and air entrainment on diesel combustion using

the LES approach This work is divided into two parts In

the first part, instantaneous non-vaporizing and vaporizing

spray structures were analyzed by RANS and LES

approaches, and their results were compared with the results

of experiments The velocity fluctuations and turbulent

kinetic energy at fixed cells during the vaporizing spray

process were computed to demonstrate the validity of the

subgrid-scale filtering concept and its applicability to our

modeling case The air entrainment and equivalence ratio

were determined for the non-reacting vaporizing spray In

addition, the amount of air entrainment was computed for

the reacting fuel spray at the instance before the ignition for

the experimental conditions listed in Engine Combustion

Network (ECN) of Sandia National Laboratories (Sandia,

ECN) In the second part, the effects of the turbulence

model on flame structure and lift-off length were analyzed,

and the heat release rate obtained with the RANS and LES

models at different fuel injection pressures was compared

with the results of the experiments

3 LES GOVERNING EQUATIONS

The LES equations are obtained by the filtering ofcontinuity, momentum, species and energy conservationequations After filtering, the results for LES can be givenas:

(1)

(2)

(3)

(4)The subgrid scale stress τij is defined by:

The subgrid scale eddy viscosity, µt, is computed using ksgs

∆ = V1/3 The subgrid scale stress can then be written as

is obtained by solving itstransport equation:

∂ρ

∂t - ∂

-=

∂ρYs

∂t - ∂

∂xj -(ρujYs)

∂xj - µSc -∂Ys

∂xj -

⎛ ⎞ Ws Wsgs ∂gjsgs

∂xj -–––

=

∂ρh

∂t - ∂

∂xj -(ρhuj)

∂xj - µPr - ∂h

∂xj -

⎛ ⎞ ∂qjsgs

∂xj -–

=

τij=ρuiuj–ρuiuj

τij 13 -τkkδij=–2µtSij

–

Sij 12 - ∂ui

∂xj

- ∂uj

∂xj

+

2 ui 2–

∂ksgs

∂t - ∂ujksgs

∂xj -+ τij∂ui

∂xj - Cεksgs

1 2 ⁄

∆ - ∂

∂xj - µt

σk -∂ksgs

∂xj -

+––

=

LARGE-EDDY SIMULATION OF AIR ENTRAINMENT DURING DIESEL SPRAY COMBUSTION 799

easy to implement and it is available in most commercial

multipurpose CFD packages The EBU model of

Magnussen and Hjertager relates the combustion rate to the

dissipation rate of eddies and expresses the reaction rate of

the reacting species by their mean mass fraction, the

turbulence kinetic energy and the rate of dissipation of this

energy In diesel engines, a significant portion of the

combustion is thought to be mixing-controlled, so

Magnussen’s model should be able to describe the rate of

combustion fairly well However, it has been recognized

that the initiation of combustion relies on laminar chemistry

(Patterson et al., 1994; Heywood, 1989) Combustion is

dominated by turbulent mixing effects in the regions where

the chemistry time scale is much slower than the turbulent

time scale The chemistry time scale, however, is not

negligible near the injector regions in the shear layer

generated by the fuel spray nor when the ambient gas

temperature and oxygen concentration are decreased, as in

case of LTC For this reason, Abraham et al (1985)

suggested replacing the controlling time scale in the

Magnussen model with the slowest time scale of the mixing

time and the chemical time Kong et al (1995) proposed

EBU-LATCT, which is an extended characteristic-time model

of Abraham et al (1985) and accounts for chemical and

turbulent time- scales simultaneously This model was

combined with the Shell ignition model to simulate the overall

combustion processes in a diesel engine In this combined

model, the initiation of the combustion relies on laminar

chemistry, and turbulence starts to have an influence on

combustion only after combustion events have already

Arrhenius-type reaction rate:

(13)The turbulence time scale τt is proportional to the eddyturnover time:

(14)

where C2 = 0.142 for standard k-ε turbulence model (Reitz,

1991; Kong et al., 1992; Reitz and Kuo, 1989), and f is a delay coefficient (Patterson et al., 1994) that simulates the

influence of turbulence on combustion after ignition andwas assumed to be given by:

=

f 1 e

r –

–0.632 -

Table 1 Modeling conditions

Fuel Simulation: CExperiment: Diesel

12H26, C7H16 Tf-350K Simulation: C Experiment: Diesel 7H16, Tf-436KNozzle characteristics Dorifice-0.16 mm, Cd-0.8, Ca-0.85, L/D-5.52, θ/29-0o Dorifece-0.180 mm, Cd-0.77, Ca-0.82, L/D -4.2, θ/2-6.5o

Ambient Gas Content

Ambient gas density 46.8 kg/m3, 17 kg/m3, 16.1 kg/m3, 15.2 kg/m3 14.8 kg/m3

Turbulence model - LES Smagorinsky SGS model- LES k-∆ GGS model

- RANS k-ε high-Reynolds number model

- LES k-∆ SGS model

- RANS k-ε high-Reynolds number model

800 U B AZIMOV and K S KIM

4 ANALYSIS CONDITIONS AND SET-UP

The simulation was conducted using STAR-CD commercial

CFD code in a three-dimensional computation grid under

the conditions mentioned in Table 1 For RANS, three

different hexahedral grid resolutions were used: a coarse

grid, with a 1 mm cell size and a total number of 32,000

cells; a fine grid, with a 0.5 mm cell size and total number

of 256,000 cells; and an ultra-fine grid, with a 0.2 mm cell

size and a total number of 4 million cells For LES, the

computation domain was assumed to have a grid size of 0.2

mm and 4 million total cells In addition, to provide further

validation of the simulation results, experimental conditions

available from ECN (Sandia, ECN), as shown in Table 1,

were also applied to the simulation For RANS with a 1 mm

cell size, the computation domain consisted of 1,259,712

cells and for LES with a 0.2 mm cell size, 5.4 million cells

The discretization of space and time was set according to

the Courant number (Star-CD v3.26) In addition, the

complete spray combustion duration was considered to

match that of the experiment The mesh resolution was set

to achieve good agreement between the simulation results

and experimental results for the penetration of non-reacting

and reacting fuel jets The fuel was injected with spray

characteristics adjusted according to the spray characteristics

assumed in the experiments In the spray model, atomization

proceeded according to the Reitz-Diwakar model, and the fuel

droplets were formed according to the Reitz-Diwakar

breakup model This atomization model assumed that the

liquid was issued from the nozzle as a jet and that the

waves were formed on the jet’s surface; then, the waves

were amplified, and the liquid was eventually broken up

into droplets by aerodynamic forces caused by the high

relative velocities between the liquid and the gas (Reitz,

1987; Reitz and Diwakar, 1987) To apply this model, a

spray semi-cone angle must be known and given as part of

the input data Based on this angle, the initial droplet

velocity is determined This semi-cone angle for both

conditions was determined from experiments (Sandia, ECN;

Jeong, 2003) performed using the same spray characteristics

and ambient gas conditions as those mentioned for

simulation in this paper The autoignition in the present

simulation was controlled by the Shell model for

EBU-LATCT For emission simulation, the 3-step Zeldovich

model and the ERC model were used for NOx and soot

emission calculations, respectively The ambient gas

temperature, ambient gas content, ambient gas pressure,

fuel injection pressure, injection duration and single-hole

injector orifice parameters corresponded to those of the

experiment

5 RESULTS AND DISCUSSION

5.1 Non-vaporizing and Vaporizing Sprays

The comparisons of instantaneous non-vaporizing spray

penetrations in a constant volume chamber for two LES

sub-grid scale (SGS) models (Smagorinsky model and k-∆model) and the RANS model with that of the experimentare presented below Figure 1 shows that the SmagorinskySGS model over-predicted the penetration and that theRANS under-predicted it Consequently, as shown inFigure 2, the SGS Smagorinsky model showed a higher gasvelocity distribution along the axial and the radial directions

of spray penetration than the SGS k-∆ model and the RANSmodel Because the grid size used for all three cases wasthe same, the absolute value of the gas velocity after theinteraction with a liquid drop did not depend on the gridsize; additionally, because the cell type for the entiredomain was cube-hexahedral, the results were not affected

by the grid cell anisotropy In such a case, it is believablethat the centerline velocity distribution and length of theinitial jet region, after which the velocity decays rapidly,depend strongly on the Cs parameter Decreasing the Csvalue will result in a decreased length of the initial velocityregion With a lower Cs value, turbulent mixing appeared

to be more intense, shortening the length of the initialregion, and vice versa A similar trend in the spray tip

penetration was observed by other researches (Hori et al.,

2006) comparing the LES SGS k-∆ model and RANSmodel, when the instantaneous spray penetration with LES

Figure 1 Instantaneous non-vaporizing sprays for RANSand LES; ∆=0.2 mm, Tamb=298 K, ρamb=46.8 kg/m3, Pinj=60MPa, Dorifice=0.163 mm

LARGE-EDDY SIMULATION OF AIR ENTRAINMENT DURING DIESEL SPRAY COMBUSTION 801

showed good agreement with experiments

Figure 3 shows comparisons of the vaporizing spray

penetrations for the SGS k-∆ and RANS models with that

of the experiment (Azimov et al., 2008) At the earlier time

steps, SGS k-∆ overpredicted the spray penetration, but as

time elapsed, the computed spray penetration matched

closely with that of the experiment; on the other hand, the

RANS simulation underpredicted the spray penetration at

the later time steps This under-predicted vapor penetration

is consistent with a well-known shortcoming of eddy

viscosity turbulence models for free shear flow, where the

models over-predict the spreading rates and consequently

under-predict the penetration in round jet flows (Wilcox,

1998) In addition, Magi et al (2001) found that

over-prediction of the spreading rates for round gas jets persists

for CFD simulations based on RANS grids and standard

k-ε turbulence models

Other researchers also reported that the RANS

simulation underpredicted vaporizing spray penetration

before the end of fuel injection (Hori et al., 2006) In

RANS, vaporizing spray represented by a fuel massfraction is symmetrical along the spray axis In contrast,LES shows unsteady spray behavior Vapor phaserepresented asymmetrical spray shape with intermittencyalong the spray boundaries, similar to what was observed

in the experiment

We believe that the overestimated penetration ofvaporizing spray by LES compared with the experiment isdue to two issues:

(1) Spatial discretisation with a convective scheme InLES, higher order spatial and temporal schemes must

be used to accurately resolve the flow, as compared toRANS If LES analysis is performed with lower orderdiscretisation schemes, then the unsteady behavior inthe turbulent flow would not be predicted Numericaldiffusion caused by the computation with convectiveterms has a detrimental effect on the LES results Inother words, numerical diffusion influences the flowfield in addition to the physical and SGS diffusion For

example, Hori et al (2008) performed numerical

analysis with LES to examine the effect of convectiveschemes on the diesel spray mixture formation Theyconducted an LES of diesel sprays with variousdiscretisation schemes and found that some discretisa-tion schemes provided less numerical diffusion butcould be dispersive Others gave good resolution butpoor stability, and vice versa These authors showedthat for most of the schemes used in the analysis,

Figure 2 Comparison of (above) centerline spray velocities

and (below) radial gas velocities for RANS and LES at

802 U B AZIMOV and K S KIM

evaporating spray penetration was overpredicted

compared to the experiment In our simulation, to

optimize the computation in terms of

“accuracy-stability”, we used Monotone Advection

Reconstruc-tion Scheme (MARS) with a blending factor of 0.5

MARS is a multidimensional second-order accurate

differencing scheme It possesses the lowest sensitivity

of solution accuracy to the mesh structure and

skewness

(2) The usage of a collision sub-model in the evaporative

spray simulation This gives the results of overpredicted

vapor penetration when compared to the experiment In

our simulation, to maintain the consistency between

non-evaporating spray and evaporating spray, the

collision sub-model was used in both cases The

collision and coalescence of fuel sprays is of great

interest because it will affect the droplet size, number

density and velocity of the droplet, essentially

influencing the structure of an evaporating spray and

altering engine performance, and as a consequence, fuel

emissions For example, experiments involving

hydrocarbon droplets carried out by Benn and Frohn

(1989) and Jiang et al (1992) demonstrated that

droplet collision and coalescence is directly related to

spray combustion applications

Figure 4 shows the instantaneous distribution of the

equivalence ratio along the centerline The LES shows

strong fluctuations in the equivalence ratio, with the

gradual decrease of its value by the leading part of the

spray These resolved fluctuating values of the equivalence

ratio allow for the estimation of the degree of local air-fuel

mixing, which influences the formation of soot in diesel

combustion In contrast, the RANS simulation showed a

very sharp increase in the equivalence ratio close to the

injector but a rapid decrease in the equivalence ratio toward

the spray tip This profile suggests that with RANS, the

history of the air and fuel distribution in the spray cannot

be recalled and, consequently, local air-fuel mixing cannot

be quantitatively predicted

To determine the spray gas velocity and turbulent kinetic

energy in the non-reacting vaporizing spray jet, velocityfluctuations were monitored during the entire sprayduration at two points: one was close to the injector orificeand spray axis, and the other was at a distance of 40 mmfrom the injector orifice and at the boundary of the jet andsurrounding ambient gas Figure 5 shows the location ofthe monitoring points in the vaporizing jet As shown inthis figure, the major part of the turbulent energy wasresolved and only approximately 10% was modeled at bothpoints At the location further from the injector orifice, themagnitude of the turbulent energy was higher than that atthe location near the injector

This result may support the finding that as the spraypropagated deeper into the ambient gas, its boundariesextended, and the flow was dominated by the large andhigher-energy containing eddies

It should be noted that the atomization and spray modelsapplied to RANS are not quite suited to the LES of sprays.Therefore, to accurately estimate subgrid turbulence in theLES of a two-phase flow with randomly distributeddroplets of various dimensions, a Eulerian description ofthe continuous phase must be adopted and fully coupledwith a Largangian definition of the dispersed phase withthe appropriate stochastic subgrid models However, it isbelieved that the finer the grid, the better the spatialresolution of the gas velocity and the better the prediction

of gas-droplet momentum exchange, which depends on therelative velocity at the drop location If the cell volume islarge, the change in gas velocity due to momentumexchange with liquid drops is small; if the grid resolution ishigh, then the gas velocity increases faster Hence, theabsolute value of the gas velocity after the interaction with

a liquid drop depends on the cell size In LES, the resolvedpart of the instantaneous flow field can readily be

Figure 4 Equivalence ratio distribution along the spray

axis for RANS and LES with ∆=0.2 mm

Figure 5 Turbulent kinetic energy monitoring locations aty=1 mm/z=5 mm and y=5 mm/z=40 mm in the vaporizingjet; ∆=0.2 mm, Tamb=820 K, ρamb=17 kg/m3,Pinj=60 MPa,

Dorifice=0.163 mm

LARGE-EDDY SIMULATION OF AIR ENTRAINMENT DURING DIESEL SPRAY COMBUSTION 803

interpolated to a particle location The major issue is to

determine whether the remaining (subgrid scale) part of the

velocity field flows can have a noticeable influence on the

particle locations We believe that this influence can be

neglected and justified by a low subgrid energy content As

shown in Figure 5, modeled turbulent energy magnitude is

approximately 10% of that of the resolved turbulent energy

Previous studies also followed the same assumption For

example, Wang and Squires (1996) neglected the influence

of the SGS flow velocity field on the disperse phase by

showing that the ratio of the SGS kinetic energy to the

mean velocity squared was roughly 10%

To check the validity of the grid size for LES, the plot of

the power spectrum was generated using the Fast Fourier

Transform (FFT) function Using FFT, the fluctuation in

the turbulent kinetic energy over a certain time period was

decomposed into components of different frequencies, and

the transient energy distribution within a discretized

domain could be monitored According to Kolmogorov’s

hypothesis of similarity (Kolmogorov, 1941a, 1941b,

1941c), turbulence causes the formation of eddies of many

different length scales Most of the kinetic energy of

turbulent motion is contained in large scale structures

Energy cascades from these large scale structures to

smaller scale structures by an inertial and essentially

inviscid mechanism This process continues, creating

smaller and smaller structures, to produce a hierarchy of

eddies Eventually, this process creates structures that are

small enough for molecular diffusion to become important,

and the viscous dissipation of energy finally occurs The

scale at which this occurs is the Kolmogorov length scale

Applying this concept to LES, it became a practice to solve

the flow problem for large eddies explicitly by assuming a

filter grid size that separates the inertial range from the

dissipative range and to model the effect of the more

universal eddies that are smaller than the computational

grid size

To determine the filter grid size, we transferred the time

domain into the space domain by applying Taylor’s

hypothesis Taylor (1938) proposed that for short time

intervals, turbulence can be assumed to be frozen as it

convects past a probe at a fixed point in space Eddies are

fixed or frozen into the mean flow, i.e., they do not change

considerably as they are advected With this assumption,

changes in the measured velocity components, as a

function of time, can be viewed as proportional to their

respective spatial changes This assumption is radical and

is unlikely to hold true much of the time Nevertheless, at a

sufficient distance from the solid boundaries where the

viscous forces are not too high, Taylor’s hypothesis does,

on average, lead to good approximations

(17)

velocityThe relationship between spatial increment, ∆, and the

that energy starts to dissipate at approximately 20 kHz andgradually vanishes in the viscous dissipation subrange,which is the region where eddy sizes are supposed to besmaller than the filter grid size and where the energy ismodeled by the SGS model From this power spectrum, wefound that the energy started to dissipate at ∆=0.25 mm,which is a good prediction, indicating that the major part ofthe computed energy was resolved We can assume thatwith a coarser grid, the filter width location would spreadinto lower frequencies and, therefore, reduce the spectrumrange of the resolved kinetic energy

5.2 Air Entrainment and MixingThe new combustion strategies focused on simultaneousreduction of NOx and soot emissions suggest that ignitiontiming should be controlled to be delayed after the end ofinjection Therefore, air in the combustion chamber can beproperly utilized by optimizing air-fuel mixing in the spray.Furthermore, optimized fuel mixing can be obtained byincreasing the spray momentum and volume by avoidingthe steep gradient in fuel concentrations To achieveoptimized fuel mixing, the spatial fuel concentration andair entrainment should be well characterized The shearforces between the high-velocity spray and the ambient aircause eddies to develop Consequently, air is entrained into

the jet by the roll-up of eddies (Warnatz et al., 2006; Cant

and Mastorakos, 2007) The entrained air moves with thejet and has to be replaced by new air This drives the flowfrom the outer parts of the combustion chamber and thespray tip region toward the injector and the spray periphery.The characteristics of air entrainment into axisymmetricgas jets have been studied extensively since the early work

of Ricou and Spalding (1961) Tomita et al (1995)

measured the ambient air entrained into a free gas jet Theyshowed that the air from the upstream side of the unsteady

804 U B AZIMOV and K S KIM

part of the jet is entrained more than that from the steady

part The ensemble averaged mass of air entrained into the

jet was nearly axisymmetric; however, the entrainment

mass of every jet was not always axisymmetric Sasaki et

al (1998) investigated the surrounding air field using a

particle image velocimetry (PIV) system They found that

rather small amounts of air were entrained near the nozzle

tip and that more air entrainment occurred in the spray

mid-section and the tip Furthermore, they found that increasing

the fuel velocity imposed an almost proportional increase

in the surrounding air velocity in the vicinity of the nozzle

Further downstream, however, this influence weakened As

expected, a smaller hole diameter reduced the surrounding

air velocity because the momentum transferred to the

surrounding air was lower at the lower injection rate

Ishikawa and Zhang (1999) studied air-entrainment with

the air density difference as a tracer of the moving air They

found that mair/ mfuel did not change with injection

velocity, which is in agreement with Sasaki et al (1998)

and the jet theory used by Siebers (1999) Rajalingam and

Farrell (1999) studied air-entrainment using PIV They

found little difference in air entrainment with injection

pressure on the first two-thirds of the spray plume, while

on the last third, there was a major difference Further

studies on non-evaporating diesel sprays by Rhim and

Farrell (2000, 2001) suggested that a significant part of the

overall ambient gas entrained in a spray plume is entrained

from the spray tip This stands in contrast to the common

perspective that most of the gas is entrained in the lateral

sides of the sprays and that the gas near the spray tip is just

pushed aside by the spray tip However, the velocity plot

agrees with the conventional belief about the gas flow

pattern and resulting gas entrainment along the sides of the

spray plume Mohammadi et al (1998) investigated the

droplets and ambient gas interaction in diesel spray They

investigated the disintegration process of non-evaporating

diesel sprays injected into high pressure and room

temperature ambient gases using single and double spark

back light photography methods They found that the

disintegration of diesel spray begins at a very early period

of injection At the early period of the injection, the area

behind the spray tip was quite active in disintegration and

produced a large number of droplets Therefore, these

authors suggested that more air was entrained downstream,

close to the spray tip, and less in the upper part of the spray

Their results agree with the results of Rhim and Farrell

(2000) Tomita et al (1997) investigated the ambient air

entrainment into a transient hydrogen jet and its flame jet

They found that an air mass entrained into the flame jet was

almost the same as that entrained into a jet without

combustion at the same injection rate The rate of the total

air mass entrained into the flame jet per unit area was

smaller than that entrained into a jet without combustion,

but both rates decreased with an increase in penetration

These authors suggested that the transient jet was divided

into two parts, the front part and the side part They showed

that the rate of an air mass entrained into the side part of theflame jet was almost the same as that of an air massentrained into the side part of the cold jet with time Therate of air mass entrained into the front part of the jetdecreased with time in both jets Rhim and Farrell (2002a,2002b) also found that the air-entrainment with respect toburning sprays appeared similar to that with vaporizingsprays

In this study, the total air entrained into the diesel spraywas numerically computed using the relation given below:

(18)

where MA is the mass of air entrained, Vspray is the sprayvolume, Vtotal is the total volume of the chamber, MF is themass of the injected fuel, (A/F)stoich is the stoichiometric airfuel ratio, αe is the entrainment coefficient, which has avalue of 0.25±0.05 (Borman and Ragland, 1998) Asshown in Figure 7, the amount of air entrained increasedwith decreasing ambient gas temperatures On the otherhand, the percentage of evaporated fuel increased withincreasing ambient gas temperatures With the higherambient gas temperature and more intense fuel evaporation,the amount of air entrained due to the higher ambient gasdensity exceeded the amount of air entrained due to moreintense fuel evaporation This trend is reasonable compared

to those resulting from experiments of other researchers.Rhim and Farrell (2000) showed that for a non-evaporatingspray, the accumulated mass of entrained ambient gasacross the conical control surface of the spray increasedwith increasing ambient gas densities Naber and Siebers(1996) showed that for vaporizing sprays, the vaporizationdecreased penetration and dispersion by as much as 20%relative to non-vaporizing sprays; however, the effects ofvaporization decreased with increasing gas densities Theresearch performed by Siebers (1999) on liquid-phase fuelpenetration and vaporization in sprays, using an idealizeddiesel spray model, suggested that the spray- spread angle

is a measure of the growth rate of the spray caused by theentrainment of ambient gas Siebers (1999) showed that asambient gas density increased, the spray-spread angle alsoincreased, which can be an indication of the increasedamount of air entrained into the spray

To validate the air entrainment approach employed inthis study, two cases from ECN were simulated andcompared with the results of air entrainment reported bySiebers and Higgins (2001) As was noted by Siebers andHiggins, changes in lift-off length due to changes ininjection pressure, orifice diameter and ambient gastemperature and density can contribute to changes in theamount of fuel-air premixing, which occurs prior to theinitial combustion zone, upstream of the lift-off length.These changes, in turn, can affect the combustion and soot

ξ %( ) αε

MA Vspray

Vtotal -

⋅

MF AF

⎝ ⎠

⎛ ⎞stoich

⋅ - 100%

=

LARGE-EDDY SIMULATION OF AIR ENTRAINMENT DURING DIESEL SPRAY COMBUSTION 805

formation downstream The quantity of air entrained

upstream of the lift-off length was estimated by Siebers and

Higgins using the expression for the axial variation of the

cross-sectional average equivalence ratio for a non-reacting

fuel jet developed by Naber and Siebers (1996)

Under our study conditions, by estimating the amount of

air entrained into the jet before ignition, we found that the

computed percentage of total air required to burn the

injected fuel was in close agreement with that given by

experimentation In particular, for the experiment with

conditions of O2=21%, ρamb=14.8 kg/m3, ∆Pinj=41 MPa,

Tamb=1000 K, and Dorifice=0.180 mm, 11% of the total air

was entrained (Siebers and Higgins, 2001), whereas by the

computation, 10.8% was entrained The same trend was

observed for similar conditions, but with a different

pressure drop across the injector orifice, ∆Pinj = 62 MPa

The total amount of air entrained as reported by

experimentation was 14% (Siebers and Higgins, 2001), and

13.4% was computed by simulation

Enhanced air entrainment is attributed to better air-fuel

mixing Mixture fraction and mixture fraction variance are

considered as indicators of mixing quality A common

practice in the modeling of nonpremixed combustion is to

relate the various chemical mass fractions to a

mixture-fraction scalar (Bilger, 1976) Chemical reaction rates are

known to be strong functions of the mixture fraction, and

different definitions have been used for the mixture fraction

(Bilger, 1989; Pitsch and Peters, 1998), but essentially, a

mixture fraction is a measure of the local equivalence ratio

Hence, the mixture fraction is a conserved scalar,

independent of the chemistry Considering the simplest case

of infinitely fast chemistry, the mass fractions of all species

and temperatures are a function of the mixture fraction

only Mixture fraction describes the mixing state between

the fuel and oxidizer, while mixture fraction variance

reflects the effect of the non-mixed state The radial

profiles of mixture fraction variance are plotted in Figures

8 and 9 From these predictions, it can be observed that the

radial spreading rate of the mixture fraction, and itscorresponding fluctuations, are reduced as the spraypenetrates deeper into the chamber Figure 8 shows that at alower ambient gas temperature, stoichiometric mixturefraction contours are located closer to the spray axis withless intense fluctuations in the mixture fraction scalar Thisbehavior reflects an improvement in mixing

Figure 9 shows that although the mixture fraction scalardistribution is wider for higher injection pressures, themagnitude of the mixture fraction variance is lower Thisfinding may be explained by the fact that the higher fuelinjection pressure causes more intense fuel atomization andbreakup and wider liquid-vapor fuel distribution within thechamber

The magnitude of air-fuel mixing affects both theautoignition delay time and the duration of the premixedphase The initial increases in temperature and the radical

Figure 7 Air entrained into the spray and evaporated fuel

per mass of total fuel injected at different ambient gas

temperatures Figure 8 Comparison of mixture fraction variance atdifferent ambient gas temperatures; ∆=0.2 mm, Pamb=4

MPa,Pinj=60 MPa, Dorifice=0.163 mm

Figure 9 Comparison of mixture fraction variance atdifferent fuel injection pressures; ∆=0.2 mm, Tamb=1000 K,

ρamb=14.8 kg/m3, Dorifice=0.180 mm

806 U B AZIMOV and K S KIM

buildup take place closer to the oxidizer side With time,

these radicals diffuse toward higher mixture fractions and

ignition eventually occurs at a richer mixture fraction,

denoted in Mastorakos et al (1997) as the most reactive

mixture fraction, which could be richer than the

stoichiometric value (Zst=0.062) Higgins et al (2000)

experimentally showed that ignition occurs in the rich

mixture downstream of the end of the liquid-phase

penetration length and farther upstream as the chamber

temperature and density increase The results of LES can

be seen in Figure 10, in which the initial rise in temperature

is observed, where Z is approximately 0.1 (where the

equivalence ratio is greater than 1), showing that ignition is

predicted to occur in the rich mixture The premixed mode

of diesel combustion has been shown to consist of the

flame propagation phase along the stoichiometric mixture

fraction contour, driven mostly by volumetric expansion

following autoignition A stable flame is established with

the peak temperature occurring close to Zst =0.062

6 DIESEL FLAME STRUCTURE AND

LIFT-OFF LENGTH

Figure 11 shows the temporal change in the diesel flame

structure In this figure, experimental images of spray

combustion were taken by high-speed shadowgraphy, and

temporal temperature images were computed using LES

∆=0.2 mm, RANS ∆=0.5 mm and RANS ∆=1 mm The

temperature scale varies from 920 K (blue) to 2700 K (red)

With LES ∆=0.2 mm, the unsteady nature of diesel flame

was well captured The premixed flame propagated

downstream (0.54-0.6 ms) After the premixed flame

phase, the high temperature region was observed at theperiphery of the diesel jet, indicating that the flame wasestablished along the stoichiometric mixture fractioncontour With RANS ∆=0.5 mm, instantaneous flamepropagation was predicted well; however, information onthe uneven pattern of the structure and irregular distribu-tion of the flame temperature was lost With this structure,the correct distribution of the stoichiometric mixturefraction contour, reaction sites, and the scalar species couldnot be predicted Using an even larger grid size (∆=1 mm),the flame propagation and structure were very muchoverpredicted

Accurate prediction of temperature and OH scalarspecies distribution is very important for estimation of theflame lift-off length The lift-off length plays an importantrole in diesel combustion and emission formation processes,especially in soot formation processes

Here, the analysis was performed on the flame lift-offlength, which defines the amount of air entrained and thefuel-air premixing that occurs upstream of any combustion

in a DI diesel spray LES was performed under theexperimental conditions listed in ECN In this simulation,the lift-off length was determined by selecting three iso-temperature contours of 1600 K, 2000 K and 2200 K,measured from the closest region of the iso-temperaturecontour to the injector orifice Figure 12 shows that the lift-off length did not change with time and stabilized at the

Figure 10 Scatter plot of flame temperatures and mixture

fractions at different ambient gas temperatures; ∆=0.2 mm,

Pamb=4 MPa,Pinj=60 MPa, Dorifice=0.163 mm

Figure 11 Temporal flame distribution for the experiment,RANS and LES; Tamb=920 K, Pamb=4 MPa, Pinj=60 MPa

LARGE-EDDY SIMULATION OF AIR ENTRAINMENT DURING DIESEL SPRAY COMBUSTION 807

same value at 3 ms and 5 ms after the start of injection

Dashed lines on this figure are the experimental values of

lift-off length for two different fuel injection pressures The

results are in good agreement with those from other

researchers who experimentally showed that a flame

remains lifted at a fixed location during the entire

steady-state combustion process (Pickett et al., 2005; Pauls et al.,

2007; Idicheria and Pickett, 2007; Pickett et al., 2009).

However, both LES and RANS underpredicted the lift-off

length at the temperature 1600 K and overpredicted it at the

temperature 2200 K

In fact, there is no any strong recommendation in the

literature on employing a certain fixed value of

iso-temperature to evaluate lift-off length Senecal et al (2003)

used a 2200 K iso-line temperature and compared thenumerical results with Sandia’s experiment whileevaluating lift-off length They showed that the predictedlift-off length values and trends were in good agreementwith the measurements Tap and Veynante (2005)performed a sensitivity analysis of the lift-off length based

on temperature iso-surface They showed that no singleiso-contour yields the best agreement, although they used aflame temperature of 2200 K to evaluate lift-off length

Lehtiniemi et al (2006) used the progress variable

approach, with detailed chemistry, to model diesel sprayignition To assess the applicability of the model, they

Figure 12 Comparison of iso-line temperatures at the flame lift-off length for LES ∆=0.2 mm and RANS ∆=1 mm;

Tamb=1000 K, ρamb=14.8 kg/m3

Figure 13 AHRR for RANS where ∆=1 mm, LES where ∆=0.2 mm and the Sandia experiment

808 U B AZIMOV and K S KIM

evaluated flame lift-off length as a function of nozzle

diameter and injection pressure For comparison, they

chose a threshold temperature of 1600 K to determine the

lift-off length D’Errico et al (2007) compared combustion

and pollutant emission models for DI diesel engines They

simulated spray combustion in a Sandia combustion vessel

and evaluated the lift-off length 3 ms after the start of

injection using an iso-line temperature of 1600 K Campbel

et al (2008) analyzed premixed flame and lift-off in diesel

spray combustion using multi-dimensional CFD They

showed that the grid size may have had an effect on the

lift-off length They used an iso-line temperature of 2200 K,

although past validation studies were performed with an

iso-line temperature of 1600 K

Figure 13 shows a comparison of apparent heat release

rates for the conditions mentioned in Figure 12 The

ignition delays that were obtained from experiments for

two conditions with different pressure drops across the

injector orifice were accurately predicted by LES but were

underpredicted by RANS simulation Nevertheless, the heat

release peaks in the premixed phase were overpredicted at

both injection pressures and underpredicted in the

mixing-controlled phase by both LES and RANS, displaying high

fluctuations in LES

Finally, the interrelation between entrained air and

emission formation is elucidated here Figure 14 shows the

instantaneous images of the experimental spray flame

distributions and the temperatures, mixture fractions, sootand NOx scalars computed by LES The temperature scalewas from a min of 920/820 K (blue) to a max of 2700 K(red), and for the rest of the scalar species the min was 0(blue), with a max of 0.2 (red) for the mixture fraction, amax of 0.08 (red) for soot, and max of -0.1 (red) for NOx.With the ambient gas temperature of 920 K (less airentrainment), the flame luminosity was stronger comparedwith that at 820 K (more air entrainment) We believe thatthis strong luminosity in the flame may be attributed to sootformation because the very intense white region in thediesel jet is recognized as blackbody radiation from soot

particles (Dec, 1997; Flynn et al., 1999; Zhao and

Ladommatos, 1998) The simulated flame temperaturedistribution at Tamb=920 K had larger zones and the highesttemperature value compared to that at 820 K As a result, sootand NOx concentrations were also higher at Tamb=920 K

7 CONCLUSIONS

Large eddy simulation was used to perform computations

of air entrainment and mixing in diesel spray combustion in

a constant volume chamber using multidimensional CFD.The results of this simulation were compared to the resultsfrom experiments and with the results from the spraycombustion simulation that employed the RANS k-εmodel The following conclusions were drawn:

(1) The LES accurately predicted non-vaporizing andvaporizing spray structures and penetrations with theSGS D-k model LES showed strong fluctuations in theequivalence ratio, with the gradual decrease in its valuealong the spray centerline approaching 2.0 at theleading part of the spray The RANS simulationshowed a very sharp increase in the equivalence ratioclose to the injector and the rapid decrease of its value.These resolved fluctuating values of the LES computedequivalence ratio allowed for the estimation of thedegree of local air-fuel mixing However, in the case ofthe RANS, the history of the air and fuel distribution inthe spray could not be recalled, and consequently, localair-fuel mixing could not be quantitatively predicted.(2) It was demonstrated that with LES, the majorcomponent of the computed turbulent energy wasresolved In monitored locations further from theinjector orifice, turbulent energy magnitude was higherthan that of locations near the injector This result maysupport the fact that as the spray propagated deeperinto the ambient gas, its boundaries extended and theflow was dominated by the large and higher-energycontaining eddies

(3) The amount of air entrained increased with decreasingambient gas temperatures On the other hand, thepercentage of evaporated fuel increased withincreasing ambient gas temperatures With higherambient gas temperatures and more intense fuelevaporation, the amount of air entrained due to theFigure 14 Instantaneous jet structure at 1.2 ms; ∆=0.2 mm,

LARGE-EDDY SIMULATION OF AIR ENTRAINMENT DURING DIESEL SPRAY COMBUSTION 809

higher ambient gas density exceeded the amount of air

entrained due to more intense fuel evaporation

(4) From the estimation of the air entrained into the jet

before ignition, it was found that the computed

percentage of the total air required to burn the fuel

being injected was in close agreement with that of the

experiment In particular, for the conditions of

O2=21%, ρamb=14.8 kg/m3, DPinj=41 MPa, Tamb=1000

K, and Dorifice=0.180 mm, the amount of total air

entrained was 11% in the experiment, and 10.8% as

computed by simulation The same trend was observed

with a different pressure drop across the injector

orifice, ∆Pinj=62 MPa The total amounts of air

entrained were 14% and 13.4% by experiment and

simulation, respectively

(5) The results of the LES showed that the initial rise in

temperature occurred where Z was approximately 0.1,

that is, where the equivalence ratio was greater than 1,

which showed that ignition could be predicted to occur

in the rich mixture A stable flame was established

when the peak temperature occurred close to Zst =0.062

(6) The lift off-length was evaluated with iso-temperature

contours of 1600 K, 2000 K and 2200 K, measured

from the closest region of the contour to the injector

orifice The results showed that the lift-off length did

not change as time elapsed and stabilized with the same

value at 3 ms and 5 ms after the start of injection

However, for both the LES and RANS, the lift-off

length was underpredicted at a temperature of 1600 K

and overpredicted at a temperature of 2200 K

(7) The ignition delays obtained from experiments for two

conditions with different pressure drops across the

injector orifice were accurately predicted by the LES

but were underpredicted by the RANS simulation

Nevertheless, the heat release peaks in the premixed

phase were overpredicted for both injection pressures

and underpredicted in the mixing-controlled phase by

both the LES and RANS

(8) The spray flame development computed by the LES

was compared with the spray flame development in

experiments At the ambient gas temperature of 920 K,

the flame luminosity was stronger compared with that

at 820 K This strong luminosity in the flame can be

attributed to soot formation because very intense white

regions in the diesel jet are recognized as blackbody

radiation from soot particles The simulated flame

temperature distribution at Tamb=920 K had larger zones

and the highest temperature value compared with that

at 820 K, and soot and NOx concentrations were also

higher at 920 K

REFERENCES

Abraham, J., Bracco, F V and Reitz, R D (1985)

Comparison of computed and measured premixed

charge engine combustion Combustion and Flame, 60,

conditions Int J Automotive Technology 9, 5, 523−534.

Bardina, J., Ferziger, J H and Reynolds, W C (1980).Improved subgrid model for large eddy simulations

Bianchi, G M., Minelli, F., Scardovelli, R and Zaleski, S.(2007) 3D large scale simulation of the high-speed

liquid jet atomization SAE Paper No 2007-01-0244.

Bilger, R W (1976) The structure of diffusion flames

Combust Sci Technol., 13, 155.

Bilger, R W (1989) Turbulent diffusion flames Ann Rev.

Fluid Mech., 21, 101−135

Borman, G L and Ragland, K W (1998) Combustion

Engineering Int Edn McGraw-Hill New York.

Brenn, G and Frohn, A (1989) Collision and merging of

two equal droplets of propanol Experiments in Fluids 7,

Cant, R S and Mastorakos, E (2007) An Introduction to

Turbulent Reacting Flows Imperial college Press.

London

Chigier, N (1981) Energy, Combustion and Environment.

McGraw-Hill New York

Chumakov, S and Rutland, C J (2004) Dynamicstructure models for scalar flux and dissipation in large

eddy simulation AIAA J., 42, 1132−1139.

Dec, J E (1997) A conceptual model of DI diesel

combustion based on laser-sheet imaging SAE Paper

SAE Paper No 1999-01-0509.

Germano, M., Piomelli, U., Moin, P and Cabot, W (1991)

A dynamic subgrid-scale eddy viscosity model Phys.

Han, D and Mungal, M G (2001) Direct measurement ofentrainment in reacting/nonreacting turbulent jets

Hawkes, E R and Cant, R S (2001) Implication of aflame surface density approach to large eddy simulation

810 U B AZIMOV and K S KIM

of premixed turbulent combustion Combustion and

Heywood, J B (1989) Internal Combustion Engine

Higgins, B S., Siebers, D L and Aradi, A (2000)

Diesel-spray ignition and premixed-burn behavior SAE Paper

No 2000-01-0940.

Hori, T., Senda, J., Kuge, T and Fujimoto, H (2006)

Large eddy simulation of non-evaporative and

evaporative diesel spray in constant volume vessel by

use of KIVALES SAE Paper No 2006-01-3334.

Hori, T., Kuge, T., Senda, J and Fujimoto, H (2007)

Large eddy simulation of diesel spray combustion with

eddy-dissipation model and CIP method by use of

KIVALES SAE Paper No 2007-01-0247.

Hori, T., Kuge, T., Senda, J and Fujimoto, H (2008)

Effect of convective schemes on LES of fuel spray by

use of KIVALES SAE Paper No 2008-01-0930.

Hu, B and Rutland, C J (2006) Flamelet modeling with

LES for diesel engine simulations SAE Paper No

2006-01-0058

Idicheria, C A and Pickett, L M (2007) Effect of EGR

on diesel premixed-burn equivalence ratio Proc.

Combust Inst., 31, 2931−2938

Ishikawa, N and Zhang, L (1999) Characteristics of

air-entrainment in a diesel spray SAE Paper No

1999-01-0522

Jeong, B C (2003) Study on the Spray Characteristics of

Common-rail Injection System M S Thesis Yeosu

National University Korea

Jhavar, R and Rutland, C J (2006) Using large eddy

simulations to study mixing effects in early injection

diesel engine combustion SAE Paper No

2006-01-0871

Jiang, Y., Umemura, A and Law, C K (1992) An

experimental investigation on the collision behaviour of

hydrocarbon droplets J Fluid Mechanics, 234, 171−

190

Kaario, O., Pokela, H., Kjaldman, L., Tiainen, J and

Larmi, M (2003) LES and RNG turbulence modeling

in DI diesel engines SAE Paper No 2003-01-1069.

Kimura, S., Kosaka, H., Matsui, Y and Himeno, R (2004)

A numerical simulation of turbulent mixing in transient

spray by LES (Comparison between numerical and

experimental results of transient particle laden jets) SAE

Paper No 2004-01-2014

Kolmogorov, A N (1941 a) The local structure of

turbulence in incompressible viscous fluid for very large

Reynolds number Dokl Acad Nauk SSSR, 30, 9−13

(Reprinted in Proc R Soc London A 434 1991, 9−13)

Kolmogorov, A N (1941 b) On degradation (decay) of

isotropic turbulence in an incompressible viscous liquid

Kolmogorov, A N (1941 c) Dissipation of energy in

locally isotropic turbulence Dokl Akad Nauk SSSR, 32,

16−18 (Reprinted in Proc R Soc London A 434 1991,

15−17)

Kong, S.-C., Ayoub, N and Reitz, R D (1992) Modelingcombustion in compression ignition homogeneous

charge engines SAE Paper No 920512

Kong, S.-C., Han, Z and Reitz, R D (1995) Thedevelopment and application of a diesel ignition andcombustion model for multidimensional engine

simulation SAE Paper No 950278.

Lee, D., Pomraning, E and Rutland, C J (2002) LES

modeling of diesel engines SAE Paper No

Lesieur, M and Metais, O (1996) New trends in large

eddy simulations of turbulence Ann Rev Fluid

Lesieur, M (2005) Large Eddy Simulations of Turbulence.

Cambridge University Press New-York.?

Li, Y H and Kong, S.-C (2008) Diesel combustionmodeling using LES turbulence model with detailed

chemistry Combustion Theory and Modelling, 12, 205−

219

Magi, V., Iyer, V and Abraham, J (2001) The k-epsilonmodel and computed spreading rates in round and plane

jets Num Heat Transfer, Part A, 40, 317−334.

Magnussen, B F and Hjertager, B H (1976) Onmathematical modeling of turbulent combustion withspecial emphasis on soot formation and combustion

729

Mastorakos, E., Baritaud, T A and Poinsot, T J (1997).Numerical simulations of autoignition in turbulent

mixing flows Combustion and Flame, 109, 198−223

Menon, S., Yeung, P K and Kim, W W (1996) Effects ofsubgrid models on the computed interscale energy

transfer in isotropic turbulence Comput Fluids, 25,

165−180

Menon, S (2000) Subgrid combustion modelling for

large-eddy simulations Int J Engine Res 1, 2, 209−

227

Mohammadi, A., Miwa, K., Ishiyama, T and Abe, M.(1998) Investigation of droplets and ambient gasinteraction in a diesel spray using a nano-spark

photography method SAE Paper No 981073.

Moin, P., Squires, K., Cabot, W and Lee, S (1991) Adynamic subgrid-scale model for compressible

turbulence and scalar transport Phys Fluids, 3, 2746−

2757

Naber, J and Siebers, D L (1996) Effects of gas densityand vaporization on penetration and dispersion of diesel

sprays SAE Paper No 960034.

Patterson, M A., Kong, S.-C., Hampson, G J and Reitz, R

D (1994) Modeling the effects of fuel injectioncharacteristics on diesel engine soot and NOx emissions

LARGE-EDDY SIMULATION OF AIR ENTRAINMENT DURING DIESEL SPRAY COMBUSTION 811

SAE Paper No 940523

Pauls, C., Vogel, S., Grünefeld, G and Peters, N (2007)

Combined simulations and OH-chemiluminescence

measurements of the combustion process using different

fuels under diesel-engine like condition SAE Paper No.

2007-01-0020

Peters, N (2000) Turbulent Combustion Cambridge

University Press New-York

Pickett, L M., Siebers, D L and Idicheria, C A (2005)

Relationship between ignition processes and the lift-off

length of diesel fuel jets SAE Paper No 2005-01-3843.

Pickett, L M., Kook, S., Persson H and Andersson, O

(2009) Diesel fuel jet lift-off stabilization in the presence

of laser-induced plasma ignition Proc Combust Inst., 32,

2793−2800

Pitsch, H (2002) Improved pollutant predictions in

large-eddy simulations of turbulent non-premixed combustion

by considering scalar dissipation rate fluctuations Proc.

Pitsch, H and Peters, N (1998) A consistent flamelet

formulation for non-premixed combustion considering

differential diffusion effects Combustion and Flame,

114, 26−40

Pitsch, H and Steiner, H (2000) Large-eddy simulation of

a turbulent piloted methane/air diffusion flame (Sandia

flame D) Phys Fluids, 12, 2541−2554.

Pitsch, H (2006) Large-eddy simulation of turbulent

combustion Annu Rev Fluid Mech., 38, 453−482.

Poinsot, T and Veynante, D (2001) Theoretical and

Numerical Combustion R T Edwards, Inc.,

Philadelphia

Pomraning, E and Rutland, C J (2002) A dynamic

one-equation non-viscosity LES model AIAA J., 44, 689−

701

Pope, S B (2004) Ten questions concerning the

large-eddy simulation of turbulent flows New J Physics, 6,

35, 1−24

Rajalingam, B V and Farrell, P V (1999) The effect of

injection pressure on air entrainment into transient diesel

sprays SAE Paper No 1999-01-0523.

Raman, V and Pitsch, H (2005) Large-eddy simulation of

a bluff-body stabilized non-premixed flame using a

recursive-refinement procedure Combustion and Flame,

142, 329−347

Reitz, R D (1987) Modeling atomization processes in

high pressure vaporizing sprays Atomization and Spray

Technology, 3, 309.

Reitz, R D and Diwakar, R (1987) Structure of

high-pressure fuel spray SAE Paper No 870598

Reitz, R D (1991) Assessment of wall heat transfer

models for premixed-charge engine combustion

computations SAE Paper No 910267.

Reitz, R D and Kuo, T W (1989) Modeling of HC

emissions due to crevice flows in premixed-charge

engines SAE Paper No 892085

Rhim, D R and Farrell, P V (2000) Characteristics of air

flow surrounding non-evaporating transient diesel

sprays SAE Paper No 2000-01-2789.

Rhim, D R and Farrell, P V (2001) Effect of gas densityand the number of injector holes on the air flowsurrounding non-evaporating transient diesel sprays

SAE Paper No 2001-01-0532.

Rhim, D R and Farrell, P V (2002a) Air flowcharacteristics surrounding evaporating transient diesel

sprays SAE Paper No 2002-01-0499.

Rhim, D R and Farrell, P V (2002b) Air flow

surrounding burning transient diesel sprays SAE Paper

No 2002-01-2668.

Ricou, F P and Spalding, D B (1961) Measurements of

entrainment by axisymmetric turbulent jets J Fluid

Sandia National Laboratories, USA, Engine CombustionNetwork Available at http://www.ca.sandia.gov/ECN.Sasaki, S., Akagawa, H and Tsujimura, K (1998) A study

on surrounding air flow induced by diesel sprays SAE

Paper No 980805.

Senecal, P K., Pomraning, E., Richards, K J., Briggs, T.E., Choi, C Y., McDavid, R M and Patterson, M A.(2003) Multi-dimensional modeling of direct-injectiondiesel spray liquid length and flame lift-off length using

CFD and parallel detailed chemistry SAE Paper No.

2003-01-1043

Siebers, D L (1999) Scaling liquid-phase penetration in

diesel sprays based on mixing-limited vaporization SAE

Paper No 1999-01-0528.

Siebers, D and Higgins, B (2001) Flame lift-off on

direct-injection diesel sprays under quiescent conditions SAE

Paper No 2001-01-0530.

Smagorinsky, J (1963) General circulation experiments

with the primitive equations Monthly Weather Rev., 93,

99−164

Sone, K and Menon, S (2003) Effect of subgrid modeling

on the in-cylinder unsteady mixing process in a direct

injection engine J Eng Gas Turb Power, 125, 435−

Taylor, G I (1938) The spectrum of turbulence Proc R.

Tomita, E., Hamamoto, Y., Tsutsumi, H and Yoshiyama, S.(1995) Measurement of ambient air entrainment into

transient free gas jet by means of flow visualization SAE

Turns, S (2000) An Introduction to Combustion: Concepts

and Applications 2nd Edn McGraw-Hill New-York.

812 U B AZIMOV and K S KIM

Veynante, D (2006) Large eddy simulations of turbulent

combustion Conf Turbulence and Interactions TI 2006,

May 29-June 2, Porquerolles, France

de Villiers, E., Gosman, A D and Weller, H G (2004) Large

eddy simulation of primary diesel spray atomization SAE

Paper No 2004-01-0100.

Wang, Q and Squires, K D (1996) Large eddy simulation

of particle-laden turbulent channel flow Physics of

Warnatz, J., Maas, U and Dibble, R W (2006)

Combustion 4th Edn Springer-Verlag ISBN:

International Journal of Automotive Technology, Vol 12, No 6, pp 813−820 (2011)

DOI 10.1007/s12239−011−0093−x

Copyright © 2011 KSAE 1229−9138/2011/061−03

813

EXPERIMENTAL EVALUATION OF SOF EFFECTS ON EGR

COOLER FOULING UNDER VARIOUS FLOW CONDITIONS

K S HONG1), J S PARK2) and K S LEE3)*

Department of Automotive Engineering, Kyonggi Institute of Technology, Gyeonggi 429-792, Korea

(Received 23 June 2010; Revised 1 April 2011)ABSTRACT−An experiment was conducted to characterize the effects of SOF on EGR cooler fouling A removable single-tube test rig combined with a soot generator was developed to represent an EGR cooler and diesel exhaust gas The use of asoot generator, which controlled the size and concentration of soot particles, enabled independent variables to be completelycontrolled Either n-dodecane or diesel lube oil as substitute SOFs were vaporized and injected into the test rig to evaluate theireffects on the growth of PM deposits and the degradation performance of the EGR cooler Coolant temperature, which seemed

to be associated with SOF content, was chosen as an independent variable, and PM deposit mass per unit area and theeffectiveness drop versus time increased as the coolant temperature decreased The PM deposit mass per unit area andeffectiveness drop had maximum values at a coolant temperature of 40oC for every n-dodecane injection rate For substituteSOFs tested in this experiment, the deposit mass increased when either n-dodecane or diesel lube oil was injected, but theeffect of lube oil was more significant Diesel lube oil seemed to have a stronger effect on the reduction of thermalconductivity by filling pores in the deposits When diesel lube oil was injected, the deposit mass per unit area increased 127%compared to dry soot without injection The effectiveness drop after 10 hours increased only 12.5%

KEY WORDS : Diesel engine, EGR (Exhaust Gas Recirculation), EGR cooler fouling, PM (Particulate Matter), NOx(Nitrogen Oxides)

1 INTRODUCTION

Exhaust gas recirculation (EGR), which has been widely

used in passenger car engines, is an effective strategy for

controlling emissions of nitric oxides (NOx) The key

effects of EGR include lowering the flame temperature and

the oxygen concentration of the working fluid in the

combustion chamber However, mixing the hot exhaust gas

with the inlet air increases the temperature of the inlet

charge, affecting the combustion temperature and the

thermal NOx formation (Ladommatos et al., 1998;

Ladommatos et al., 1998) The increased inlet charge

temperature reduces the mass of air drawn into the cylinder

and lowers the heat capacity, resulting in higher combustion

temperature These detrimental effects can be partially

remediated by cooling the recirculated exhaust gases using

an exhaust gas recirculation cooling device (Zheng et al.,

2004; Abd-Alla, 2002)

However, the high concentration of particulate soot

contained in the EGR gas is likely to deposit on the wall of

the EGR cooler, causing deterioration of the heat transfer

performance and an increased pressure drop Consequently,

the deterioration in cooler performance due to fouling has

an adverse effect on the rate of NOx reduction (Hoard etal., 2008; Zhang and Nieuwstadt, 2008)

Because of the complexity of this phenomenon, manyresearchers have studied this problem to better understandthe physics of soot deposition in the EGR cooler Grillot and Icart (1997) showed that a heat exchangerdisplayed asymptotic fouling resistance behavior andoutlined the role of soot particle thermophoresis and fluidvelocity Charles et al (2005) found that both the exhaustmass flow rate and the coolant temperature had a significantinfluence on the transient performance of an EGR cooler.Lance et al (2009) characterized the deposits in an EGRcooler They found that the main determinant of thethermal conductivity of the deposit is density, which wasmeasured to be just 2% that of the density of the primarysoot particles (or 98% porous)

Other researchers have attempted to find a solution toaddress this phenomenon Ismail et al (Ismail et al., 2005)suggested that short twisted-tape inserts could be added toimprove some current designs of cooling devices Usui(2004) also experimentally demonstrated that depositioncan be reduced by as much as 38% with proper ribletdimensions as compared to flat and smooth designs Zhan

et al (2008) tested the fouling of full-scale EGR coolers

*Corresponding author e-mail: leeks@kinst.ac.kr

814 K S HONG, J S PARK and K S LEE

with various exhaust treatment devices located upstream of

the cooler However, in experiments with real engines, it

was not possible to manipulate the independent variables

that affect EGR cooler fouling

In this study, the use of a soot generator that controlled the

size and concentration of soot particles enabled independent

variables to be completely controlled Recently, with the

development of low combustion engine technology, the

SOF contents in PM increased and became as an important

factor in the operation EGR coolers because soot particles

in the EGR gas with these characteristics are known to

favor the formation of thermophoretic deposits on the wall

of the EGR cooler Therefore, the purpose of this study is to

investigate the effects of SOF on deposit characteristics on

the wall of EGR coolers The primary focus is to evaluate

the thermal performance of the cooler and the mass of

deposits To accomplish these objectives, a removable

single-tube test rig was designed to facilitate the

determination of both the effectiveness of the EGR cooler

and deposit mass data In addition, the effects of cooling

temperature were also evaluated because this factor is

closely related to condensation effects among fouling

phenomenon (William, 1999)

2 FOULING MECHANISM

There are a number of fouling mechanisms that describe

the movement of deposits from the bulk gas flow onto the

cooler surface

2.1 Thermophoresis

Generally, a gas molecule is moved through the thermal

gradient field The gas molecules that influence the particles

on the hot side move faster than those on the colder side As

a result, a net force is generated that causes the particle to

move This is the main mechanism of soot deposition The

thermal force, Fth, on a particle of diameter dp is

Here, T is the absolute temperature of the EGR gas, p is

the EGR gas pressure, ∇T is the temperature gradient, and

λ is the mean free path The minus sign is used to describe

the direction of force and means that temperature decreases

(William, 1999)

2.2 Diffusion with Condensation

SOF, including unburned HCs in the gas flow, will condense

on the cooler wall if the wall temperature is under the dew

point of the HCs at the local pressure In particular, heavy

HCs with high dew points condense easily These

condensa-tion processes induce a local concentracondensa-tion gradient in the

tube, causing diffusion

3 EXPERIMENT

3.1 Experimental SetupOverall, the experimental system was roughly composed ofthree parts (Figure 1) These components were:

Soot generator: In this study, a soot generator miniCAST 5201, JingLTD) was used as a model exhaustgas generator for lab-scale experiments In most of theprevious studies with real engines, handling all of thevariables independently was impossible However, the use

(RSG-of a soot generator, which can control the size andconcentration of soot particles, enabled independentvariables to be completely controlled Thus, it was possible

to control the soot particle size and concentration understable experimental conditions

n-Dodecane and diesel lube oil vaporizer: The sootgenerator is a reasonable apparatus for making model sootbut is limited in terms of controlling the PM components,especially the amount of SOF in the total particulate matter(TPM) To control the SOF contents, a SOF vaporizer wasdesigned to complement the shortcomings of the sootgenerator A single-syringe infusion pump was used to inject

a precise amount of SOF As shown in Figure 1, SOF wasinjected and vaporized simultaneously using a pre-heaterwith P.I.D controller The heated region was 8 m in length,which was long enough to vaporize n-dodecane completely

N2 gas carried the SOFs and prevented their oxidation.EGR cooler fouling test rig device: Simplified shelland tube heat exchangers with a single tube were used inthis experiment Figure 2 shows the schematic of a single-

Fth –pλdp∇T

T

-=

Figure 1 Schematic of the overall experimental apparatus

Figure 2 Schematic diagram of a single-tube EGR coolerfouling test rig

EXPERIMENTAL EVALUATION OF SOF EFFECTS ON EGR COOLER FOULING 815

tube fouling test rig device This device was made of a

straight stainless steel tube; the tube had an inner diameter

of 10 mm and a length of 350 mm, and the tube wall

thickness was 1 mm

3.2 Data Acquisition

To evaluate the fouling characteristics, heat exchanger

effectiveness and deposit masses were measured

3.2.1 Heat exchanger effectiveness

Cooler effectiveness is often used to evaluate the

performance of an EGR cooler The effectiveness for a

typical parallel-counter flow EGR cooler is given in

Equation (2)

(2)

Here, ε is the effectiveness of the fouled EGR cooler,

Tgas,in is the inlet EGR gas temperature, Tgas,out is the EGR

gas temperature after passing through the cooler, and

Tcoolant,in is the coolant inlet temperature

All temperatures were measured using K-type

thermocouples and transmitted to a data acquisition board

(Data translation, DT-9805) that was connected to a PC

During operation, Tgas,in and Tcoolant,in were fixed by the

PID controller Based on Eq 1, variation of the cooler

effectiveness under a fixed EGR flow rate becomes larger

due to the variation of Tgas,out

3.2.2 PM deposition

The deposit mass obtained from each tube was measured in

a weighing chamber system The weighing chamber system

consisted of a micro balance with a 230-g weighing

capacity and 1-mg accuracy and a thermo-hygrostat to

maintain a constant temperature at 25oC and 45% relative

humidity

The deposit mass was obtained by weighing each tube

before and after exposure to exhaust gas Prior to weighing,

each tube was placed in a weighing chamber system with a

constant temperature of 25oC and 45% humidity for 24 h

This procedure was necessary to reduce errors associated

with water on the tube wall

3.3 Experimental Conditions

To evaluate the effects of SOF on EGR cooler fouling,

n-dodecane and diesel lube oil were added to dry soot as

substitute SOFs The SOF injection rate was determined by

referring to the dry soot mass per unit time from the soot

generator Three injection rates were chosen: dry soot

(without injection), 0.2 ml/h and 0.4 ml/h If the soot

generated by the soot generator was supposed to be dry

soot, which meant that the content of SOF was lower than

5% in the TPM, the injection rates of 0.2 ml/h, and 0.4 ml/

h represented approximately 33% and 50%, respectively, of

the SOF proportion To examine the effects of coolant

temperature on EGR cooler fouling, three different coolant

temperatures were studied: 40oC, 60oC and 80oC Thesetemperatures were chosen to approximate the range ofcoolant temperatures that may be present in a 2000ccpassenger car engine The other variables included an EGRgas temperature of 380oC, a mean particle size of 190 nmand a flow rate of 9 sLPM All experiments were carriedout for 10 h The EGR gas conditions, i.e., the flow ratesand gas temperatures, were determined on the basis ofengine conditions with a relatively low load and low speed

in which fouling tends to be more severe in a 2000ccpassenger car engine (Maing et al., 2007)

4 RESULTS AND DISCUSSION

4.1 Effects of Coolant Temperature on Fouling The variation in the PM deposit mass per unit area forcoolant temperatures of 40oC, 60oC and 80oC is shown inFigure 3 Each line represents a different n-dodecaneinjection rate: dry soot (without injection), 0.2 ml/h and 0.4ml/h The other experimental variables included a mean

PM size of 190 nm, a flow rate of 9 sLPM and an EGR gasinlet temperature of 380oC

The PM deposit mass per unit area increased as thecoolant temperature decreased For every n-dodecaneinjection rate, the PM deposit mass per unit area attained amaximum value at a coolant temperature of 40oC The effects of coolant temperature on the growth of PMdeposits were closely related to n-dodecane injection rate.The difference between the deposit masses at coolanttemperatures of 40oC and 80oC was just 7 g/m2 for the drysoot case However, this difference increased to 16 g/m2 at

a n-dodecane injection rate of 0.4 ml/h This resultconfirms the effects of HC condensation In other words, inthe test performed with only dry soot (without injection) thathas a small proportion of SOF, thermophoresis generated bythermal gradients was the dominant deposition mechanism(lower black dotted line) However, in the case of an n-dodecane injection rate of 0.4 ml/h, the PM deposit masswas affected not only by the thermal gradient between thehot gas and cooler wall but also by HCs from the gas flowcondensing on the wall (upper red dotted line)

The effectiveness profiles of the EGR cooler for theselected coolant temperatures are shown in Figures 4-6.Each graph presents the effect of n-dodecane injection ratefor dry soot (without injection), 0.2 ml/h and 0.4 ml/h,respectively

According to most studies on EGR cooler fouling, theeffectiveness of an EGR cooler decreases over a certainperiod of time, appearing to approach an asymptotic value(Teng and Barnard, 2010) In this experiment, in somecases with a high coolant temperature and dry soot thatwere not significantly fouled, the effectiveness profilesexhibited asymptotic characteristics (Figure 4-6) However,under conditions of either low coolant temperature or highinjection rate of n-dodecane, the experiments were not longenough to determine whether the effectiveness profile

ε Tgas in , –Tgas out ,

Tgas in , –Tcoolant in ,

-=

816 K S HONG, J S PARK and K S LEE

would exhibit asymptotic characteristics

The effectiveness drop of the EGR cooler after 10 h,

shown in Figures 4~6, exhibited similar trends to the PM

deposit mass results The effectiveness drop versus timeincreased as the coolant temperature decreased At aninjection rate of 0.4 ml/h of n-dodecane, the effectivenessdrop increased up to 14% at a coolant temperature of 40oC

of after 10 h (Figure 6) Based on Figures 4-6, n-dodecaneinjection increased the effects of coolant temperature onthe effectiveness drop, particularly at a coolant temperature

of 40oC

4.2 Effects of n-dodecane Injection on Fouling

To estimate the SOF derived from diesel fuel, n-dodecanewas used as a substitute n-Dodecane is similar to dieselfuel in terms of its physical and chemical characteristicsand is commonly used as a substitute for diesel fuel inexperiments (Sahetchian et al., 1995)

The variation in the deposit mass per unit area in EGRcooler for the selected n-dodecane injection rates of 0 ml/h(dry soot), 0.2 ml/h, and 0.4 ml/h is shown in Figure 7.Each line represents the deposit mass data obtained at adifferent coolant temperature: 40oC, 60oC and 80oC The

Figure 3 Variation in PM deposit mass per unit area versus

coolant temperature for selected n-dodecane injection rates

The mean PM size was 190 nm, the flow rate was 9 sLPM

and the EGR gas inlet temperature was 380oC

Figure 4 Variation of effectiveness versus time for selected

coolant temperatures without injection The mean PM size

was 190 nm, the flow rate was 9 sLPM, and the EGR gas

inlet temperature was 380oC

Figure 5 Variation of effectiveness versus time for selected

coolant temperatures at an n-dodecane injection rate of 0.2

ml/h The mean PM size was 190 nm, the flow rate was 9

sLPM, and the EGR gas inlet temperature was 380oC

Figure 6 Variation of effectiveness versus time for selectedcoolant temperatures at an n-dodecane injection rate of 0.4ml/h The mean PM size was 190 nm, the flow rate was 9sLPM, and the EGR gas inlet temperature was 380oC

Figure 7 Variation in PM deposit mass per unit area versusn-dodecane injection rate for selected coolant temperatures.The flow rate was 9 sLPM, the mean particle size was 190

nm, and the EGR gas inlet temperature was 380oC

EXPERIMENTAL EVALUATION OF SOF EFFECTS ON EGR COOLER FOULING 817

other experimental conditions included a flow rate of 9

sLPM, a mean particle size of 190 nm, and an EGR gas

inlet temperature of 380oC

The deposit mass per unit area increased as the

n-dodecane injection rate increased The effect of n-n-dodecane

injection rate on the growth of PM deposits was also

associated with a decrease in coolant temperature As

shown in Figure 7, the difference in the deposit mass

between no injection (dry soot) and an injection rate of 0.4

ml/h was just 1 g/m2 at a coolant temperature of 80oC, and

this difference increased to 8 g/m2 at a coolant temperature

of 40oC When the coolant temperature was 80°C, only a

small amount of HCs condensed on the cooler wall, and

these deposits had little effect in terms of increasing the

PM deposit mass In contrast, when the coolant

temperature was 40oC, a large amount of HCs condensed

on the cooler wall, and these deposits had a significant

effect in terms of increasing the PM deposit mass

Figures 8~10 show the effectiveness profiles for the

selected n-dodecane injection rates (dry soot – without

injection, 0.2 ml/h, 0.4 ml/h) with the different coolanttemperature of 40oC, 60oC and 80oC, respectively After operation for the prescribed period of time, theresults observed in the effectiveness profiles were verysimilar to the results obtained for the PM deposit mass perunit area As n-dodecane injection rate increased, thedeposit mass per unit area increased, and thus the overallthermal conductivity of the EGR cooler was reduced Infigure 8, the effectiveness drop reached a maximum value(up to 14%) for an n-dodecane injection rate of 0.4 ml/h at

a coolant temperature of 40oC

4.3 Effects of Diesel Lube Oil Injection on FoulingThe SOF is mainly composed of heavy HCs derived fromlube oil (Cartellieri and Tritthart, 1984) Therefore,studying the influence of lube oil-derived HCs on EGRcooler fouling is important Figure 11 shows the variation

in the EGR cooler deposit mass per unit area for theselected lube oil injection rates of 0 ml/h (no injection), 0.2

Figure 8 Variation of effectiveness versus time for selected

n-dodecane injection rates at a coolant temperature of

40oC The flow rate was 9 sLPM, the mean particle size

was 190 nm, and the EGR gas inlet temperature was 380oC

Figure 9 Variation of effectiveness versus time for selected

n-dodecane injection rates at a coolant temperature of 60oC

The flow rate was 9 sLPM, the mean particle size was 190

nm, and the EGR gas inlet temperature was 380oC

Figure 10 Variation of effectiveness versus time forselected n-dodecane injection rates at a coolant temperature

of 80oC The flow rate was 9 sLPM, the mean particle sizewas 190 nm, and the EGR gas inlet temperature was 380oC

Figure 11 Variation in PM deposit mass per unit area versusdiesel lube oil injection rate for selected coolant temperatures.The flow rate was 9 sLPM, the mean particle size was 190 nm,and the EGR gas inlet temperature was 380oC

818 K S HONG, J S PARK and K S LEE

ml/h and 0.4 ml/h Each line represents data for the selected

coolant temperatures of 40oC, 60oC and 80oC, respectively

The other experimental conditions included a flow rate of 9

sLPM, a mean particle size of 190 nm, and an EGR gas

effectiveness profiles of the EGR cooler for the selectedlube oil injection rates (0 ml/h - without injection, 0.2 ml/hand 0.4 ml/h) at different coolant temperature (40oC, 60oCand 80oC)

Interesting results were revealed when the obtaineddeposit mass data are compared to the effectiveness profiles

As shown in Figure 11, the PM deposit mass per unit areaincreased dramatically as the diesel lube oil injection rateincreased Compared to the dry soot case (without injection),the PM deposit mass per unit area for a lube oil injection rate

of 0.4 ml/h increased by 14 g/m2 at a coolant temperature of

40oC Thus, the HCs derived from lube oil seemed tostrongly accelerate the growth of PM deposits on the EGRcooler wall However, the effectivenes profiles exhibited adifferent trend from that of the obtained deposit mass data

In Figures 12~14, the effectiveness drop at a lube oilinjection rate of 0.2 ml/h was lower than that of the drysoot case, which was expected However, when lube oilwas injected at a rate of 0.4 ml/h, the effectiveness dropwas actually higher than that for the rate of 0.2 ml/h Thisphenomenon was observed on every graph for a coolanttemperature in the range of 40oC~80oC

These results suggest that ‘wet soot’ could be moreeffective in capturing the particles passing through it.However, as the SOF content increased, wet soot begansignificantly contribute to the decrease in overall thermalconductivity of the EGR cooler This phenomenon could beexplained by the physicochemical characteristics of the sootdeposits (Teng and Barnard, 2010) According to Teng’sstudy, when soot particles are covered with a thin film ofSOF, they may become hydrophilic The soot particlescould be wetted with water vapor or condensed water; thissoot hydration disrupts the original structure and results in arestructuring process Consequently, the deposits becomedenser and thus have higher thermal conductivity

Figure 12 Variation of effectiveness versus time for selected

diesel lube oil injection rates at a coolant temperature of

40oC The flow rate was 9 sLPM, the mean particle size was

190 nm, and the EGR gas inlet temperature was 380oC

Figure 13 Variation of effectiveness versus time for selected

diesel lube oil injection rates at a coolant temperature of

60oC The flow rate was 9 sLPM, the mean particle size was

190 nm, and the EGR gas inlet temperature was 380oC

Figure 14 Variation of effectiveness versus time for selected

diesel lube oil injection rates at a coolant temperature of 80oC

The flow rate was 9 sLPM, the mean particle size was 190 nm,

and the EGR gas inlet temperature was 380oC

Figure 15 Comparison of the variations in PM deposit massper unit area between n-dodecane and diesel lube oil based oninjection rate The coolant temperature was 60oC, the flow ratewas 9 sLPM, the mean particle size was 190 nm and the EGRgas inlet temperature was 380oC

EXPERIMENTAL EVALUATION OF SOF EFFECTS ON EGR COOLER FOULING 819

4.4 Comparing the Effects of n-dodecane and Diesel Lube

Oil Injection on Fouling

The effect of SOF on EGR cooler fouling seemed to be

different depending on the HC species A test was carried

out to compare the effects of n-dodecane and diesel lube oil

HCs Figure 15 shows the PM deposit mass per unit area for

n-dodecane and lube oil at varying injection rates The other

experimental conditions included a coolant temperature of

60oC, a flow rate of 9 sLPM, a mean particle size of 190 nm

and an EGR gas inlet temperature of 380oC

The PM deposit mass per unit area with lube oil

injection was greater than that with n-dodecane injection,

as shown in Figure 15 This result indicates that the heavy

HCs from lube oil had a more significant effect on the

growth of PM deposits in the EGR cooler This behavior

arises from the difference in boiling point between the two

SOFs Lube oil is composed of heavy HCs that have a

boiling point range of 340-540oC The normal boiling point

of n-dodecane is about 216.2oC, which is much lower thanthat of lube oil Therefore, lube oil was more likely tocondense in the presence of a colder cooler wall The effectiveness profiles of the EGR cooler, comparingthe effects of n-dodecane and lube oil injection on the decrease

in performance of the EGR cooler, are shown in Figures 16and 17 Figure 16 is based on an injection rate of 0.2 ml/h andFigure 17 is based on an injection rate of 0.4 ml/h

At an injection rate of 0.2 ml/h, the effectiveness profileresults were similar to those for the PM deposit mass data(Figure 16) After 10 h, the effectiveness drop with lube oilinjection was greater than that with n-dodecane injection.However, when the injection rate was increased to 0.4 ml/

h, the effectiveness profile results had a different trendfrom that of the deposit mass data In figure 15, the depositmass for diesel lube oil injection was significantly heavierthan that for n-dodecane However, the effectiveness dropdecreased as compared to the case with n-dodecaneinjection (Figure 17)

As the SOF content increased, two phenomena occurredthat effected the overall thermal conductivity in the EGRcooler First, the deposit with a high SOF content was morelikely to capture the particles that accelerated PM depositgrowth, thus developing a thicker deposit layer that reducedoverall conductivity in the EGR cooler Second, in thephenomenon described by Teng, deposits with a high SOFcontent become easily hydrated This process makes thedeposit more dense and the thermal conductivity subsequentlyincreases

For the substitute SOFs tested in this experiment, thedeposit mass increased when either n-dodecane or diesellube oil was injected, but the lube oil effect was moresignificant Thus, diesel lube oil seemed to have a strongereffect in terms of reducing thermal conductivity by fillingpores in the deposits However, this effect was difficult toinvestigate when n-dodecane was injected

5 CONCLUSION

In this study, EGR cooler fouling characteristics wereinvestigated A single-tube EGR cooler fouling testapparatus using a soot particle generator was developed.The experimental system for evaluating EGR cooler fouling

on the lab-scale allowed independent variables to becompletely controlled Coolant temperature, n-dodecaneand diesel lube oil as substitute SOFs, and the injection rate

of the vaporized substitute SOFs were chosen asindependent variables The conclusions of this study are asfollows

The PM deposit mass per unit area and the effectivenessdrop versus time increased as the coolant temperaturedecreased In addition, this effect of coolant temperature onthe growth of PM deposits increased as the n-dodecaneinjection rate increased, and consequently, the overallthermal conductivity of the EGR cooler was reducedsignificantly A PM deposit mass per unit area of 20 g/m2 and

Figure 16 Comparison of the variations in effectiveness

between n-dodecane and diesel lube oil injection at an

injection rate of 0.2 ml/h The coolant temperature was 60oC,

the flow rate was 9 sLPM, the mean particle size was 190 nm,

and the EGR gas inlet temperature was 380oC

Figure 17 Comparison of the variations in effectiveness

between n-dodecane and diesel lube oil injection at an

injection rate of 0.4 ml/h The coolant temperature was

60oC, the flow rate was 9 sLPM, the mean particle size was

190 nm, and the EGR gas inlet temperature was 380oC

820 K S HONG, J S PARK and K S LEE

an effectiveness drop of 14% were the maximum value

observed for a n-dodecane injection rate of 0.4 ml/h at a

coolant temperature of 40oC The effects of HC condensation

when n-dodecane was injected were apparent when compared

to the case of dry soot without n-dodecane injection This

result confirmed that ‘wet soot’ could be more effective in

capturing the particles passing through it

The HCs derived from lube oil strongly accelerated the

growth of PM deposits on the EGR cooler wall However,

the trend in the effectiveness profile was less similar to that

of the obtained deposit mass data This phenomenon could

be explained by the physicochemical characteristics of the

soot deposits As the SOF content increased, the original

SOF structure was disrupted and restructured Consequently,

the deposits become more dense, and the thermal conductivity

increases

For the substitute SOFs tested in this experiment, the

deposit mass increased when either n-dodecane or diesel

lube oil was injected, but the effect of lube oil was more

significant Thus, diesel lube oil seemed to have a stronger

effect in terms of reducting the thermal conductivity by

filling pores in the deposits However, this effect was

difficult to investigate when n-dodecane was injected

ACKNOWLEDGEMENT−This research was supported by

Basic Science Research Program through the National Research

Foundation of Korea(NRF) funded by the Ministry of Education,

Science and Technology(KRF-2008-313-D00148)

REFERENCES

Abd-Alla, G H (2002) Using exhaust gas recirculation in

internal combustion engines: A Review Energy Conversion

& Management, 43, 1027−1042

Cartellieri, W and Tritthart, P (1984) Particulate analysis

of light-duty diesel engines (IDI & DI) with particular

reference to the lube oil fraction SAE Paper No

840418

Charles, F L R., Ewing, D., Becard, J., Chang, J S and

Cotton, J S (2005) Optimization of the exhaust mass

flow rate and coolant temperature for exhaust gas

recirculation (EGR) cooling devices used in diesel

engines SAE Paper No 2005-01-0654

Grillot, J M and Icart, G (1997) Fouling of a cylindrical

probe and a finned tube bundle in a diesel exhaust

environment Experimental Thermal and Fluid Science,

14, 442−454

Hoard, J., Abarham, M., Styles, D., Giuliano, J M., Sluder,

C S and Storey, J M E (2008) Diesel EGR coolerfouling SAE Paper No 2008-01-2475

Ismail, B., Charles, F., Ewing, D., Cotton, J S and Chang,

J S (2005) Mitigation of the diesel soot depositioneffect on the exhaust gas recirculation (EGR) coolingdevices for diesel engines SAE Paper No 2005-01-0656

Ladommatos, N., Abdelhalim, S M and Zhao, H (1998).Effects of exhaust gas recirculation temperature ondiesel engine combustion and emissions Prog EnergyCombust Sci., 212, 479−500

Ladommatos, N., Abdelhalim, S M., Zhao, H and Hu, Z.(1998) The effects of carbon dioxide in exhaust gasrecirculation on diesel engine emissions Prog EnergyCombust Sci., 212, 25−42

Lance, M J., Sluder, C S., Wang, H., Storey and J M E.(2009) Direct measurement of EGR cooler depositthermal properties for improved understanding of coolerfouling SAE Paper No 2009-01-1461

Maing, S., Lee, K S., Song, S., Chun, K M and Oh, B.(2007) Simulation of the EGR cooler fouling effect onNOx emission of a light duty diesel engine 2007 FallConf Proc., 1, KSAE07-F0035, 214-220

Sahetchian, K., Champoussin, J C., Brun, M., Levy, N.,Blin-Simiand, N., Aligrit, C., Jorand, F., Socoliuc, M.and Heiss, A (1995) Experimental study and modeling

of dodecane ignition in a diesel engine Combustion andFlame, 103, 207−220

Teng, H and Barnard, M (2010) Physicochemicalcharacteristics of soot deposits in EGR coolers SAEPaper No 2009-01-2877

Usui, S., Ito, K and Kato, K (2004) The effect of circular micro riblets on the deposition of diesel exhaustparticulate SAE Paper No 2004-01-0969

semi-William, C H (1999) Aerosol Technology 2nd Edn JohnWiley & Sons New York 171−172

Zhan, R., Eakle, S T., Miller, J W and Anthony, J W.(2008) EGR system fouling control SAE Paper No.2008-01-0066

Zhang, F and Nieuwstadt, M (2008) Adaptive EGRcooler pressure drop estimation SAE Paper No 2008-01-0624

Zheng, M., Reader, G T and Hawley, J G (2004) Dieselengine exhaust gas recirculation – A review on advancedand novel concepts Energy Conversion & Management,

1145, 883−900

International Journal of Automotive Technology, Vol 12, No 6, pp 821−829 (2011)

DOI 10.1007/s12239−011−0094−9

Copyright © 2011 KSAE

1229 −9138/2011/061−04

821

NEURAL-EMPIRICAL TYRE MODEL BASED ON RECURSIVE

LAZY LEARNING UNDER COMBINED LONGITUDINAL AND

LATERAL SLIP CONDITIONS

Automobile Safety Research Institute (ISVA), Mechanical Engineering Department, Carlos III University, Avd de la Universidad 30, Madrid 28911, Spain

(Received 28 November 2008; Revised 28 June 2011)

ABSTRACT−The behaviour of the tyre plays an important role in the vehicle handling An accurate tyre model that estimatesthese forces and moments it is highly essential for the studies of vehicle behaviour For the last ten years neural networks haveattracted a great deal of attention in vehicle dynamics and control Neural networks have been effectively applied to modelcomplex systems due to their good learning capability In this paper a recursive lazy learning method based on neural networks

is considered to model the tyre characteristics under combined braking and cornering The proposed method is validated bycomparison with experimental obtained responses Results show the estimated model correlates very well with the dataobtained experimentally Moreover, the neural model proposed allows to include the asymetric tyre behaviour in the tyremodel without difficulty

KEY WORDS : Tyre modelling, Neural network, Recursive Lazy learning, Combined braking and cornering

1 INTRODUCTION

The behaviour of the tyre plays an important role in the

vehicle handling Thus for the analysis of vehicles and road

safety it is necessary to take into account the forces and

moments generated at contact patch (M'Sirdi et al., 2005).

An accurate tyre model that estimates these forces and

moments it is highly essential for the studies of vehicle

dynamics and control (Guo et al., 2005)

Tyre models may be divided into three basic groups

(Koo et al., 2006): physical, analytical and emprirical

models Physical models, which provide a theorical

expression of the tyre properties based on its deformation,

are usually complex and have poor computational efficiency

if sufficient accuracy is required Complex numerical

approaches, such as finite-element model (Zhang et al.,

2002; Zhang et al., 2004; Bolarinwa and Olatunbosun,

2004) are commonly used to solve equations in these

models Physical models are often too cumbersome for

vehicle dynamics analysis and control Analytical models

calculate the tyre forces and predicts the essential tire

elastic charasteristics by simplifying the physical equations

of a tyre (Goel and Ramji, 2004; Dillinger et al., 2004).

Empirical models, also known as blackbox models, are

obtained through steady-state lab tests or dynamic lab tests

(Bakker et al., 1987; Germann et al., 1994) These models

are widely employed, but they lack predictive power and

theoretical accuracy compared with physical and analitycalmodels

One of the most well known models is the

semi-emprirical tyre model proposed by (Bakker et al., 1987;

Pacejka and Bakker, 1991), also known by the name of

Magic Formula This model uses measured data, obtained

under particular conditions of constant linear and angularvelocity, and fits a predetermined shaped curve byminimizing the sum-squared between the measured dataand the model outputs The general form of the Pacjekamodel is expressed as:

(1)

where:

(2)

(3)

A, B, C, D and E are coeffients for each tyre

characteristic S v and S h are the vertical and horizontal shiftfactors The value of these coefficients is dependent on the

vertical load and camber angle The variable Y represents the longitudinal force, braking and traction force, F long,

when the input variable X represents the slip ratio, i, and the variable represents the lateral or side force, F lat, and the

erepresents the slip angle, α

822 M J L BOADA, B L BOADA, D GARCIA-POZUELO and V DIAZ

Much of the tyre models neglect the coupling of the

forces in different directions They describe the tyre-forces

generated at the pure-slip conditions of braking, driving or

cornering Nevertheless, a steering maneuver during braking,

generally, decreases the braking stiffness, the longitudinal

peak force, and its corresponding slip value A model for the

interaction between the slip in both directions is therefore

inevitable for more advanced vehicle simulations

In 1989, (Pacejka et al., 1989) re-write the Magic

Formula tyre model in order to describe the tyre behaviour

during steady-state braking or driving, cornering and in a

combined slip situation In (Van Oosten and Bakker, 1992)

a regression method to the measurement data in order to

derive the coefficients is used A main drawback of this

procedure is that it requires starting values for the

coefficients to begin the optimization process being specially

difficult for combined slip conditions (Cabrera et al., 2004).

propose a new method based on genetic techniques to

determine these coefficients The main advantages of the

method are its simplicity of implementation and its fast

convergence to optimal solution, without the need of deep

knowledge of the searching space

For the last ten years neural networks have attracted a

great deal of attention in vehicle dynamics and control

(El-Gindy and Palkovics, 1993; Guarneri et al., 2008) Neural

networks have been effectively applied to model complex

systems due to their good learning capability The modelling

is not based on any physical law and does not introduce any

simplifying hypothesis concerning the physics governing

the system (Pracny et al., 2007)

(Palkovics and El-Gindy, 1993) propose a tyre modeling

method in terms of multi-layer perceptron (MLP) neural

network using a backpropagation algorithm This model,

called Neuro-Tyre, is able to predict the combined braking

and cornering characteristics of a tyre with a good accuracy

in a range of vertical loads It predicts any shift and

unsymmetrical curves with the same structure (Kim and

Ro, 1995) propose a tire force model by neural networks

which relates the tire force as function of vertical load, slip

angle and camber angle In all of the operating range, the

neural network tire model predicts tire side force within 3%

error of measured tire force However, MLP neural networks

have shown to be an efficient and reliable method for

characterization of a tyre, the selection of networks structures

and training of samples are often complicated tasks

An alternative representation of the tyre characteristics is

in the form of radial basis function (RBF) networks RBF

networks are powerful computation tools and have been

used extensively in the systems modeling (Matusko et al.,

2008) propose a new scheme for robust tire/road friction

force estimation A RBF neural network is used to

compensate negative influence of the friction model

uncertainties to the accuracy of the force estimation The

main advantages of RBF networks are that they have simple

architecture and learning scheme, fast training speed, and

the possibility of incorporating the qualitative aspects of

human experience in the model selection and training.Nevertheless, they have difficulties in selecting the networkstructure and calculating the model parameters

In this paper, a recursive lazy learning method based onneural networks is proposed in order to model the combinedbraking and cornering characteristics of a tyre

The proposed method is able to predict with a highaccuracy the combined braking and cornering characteristics

of a tyre The errors obtained with the proposed model aremuch smaller than the errors obtained with the most wellknown empirical model: Pacejka model It also predictsany shift and unsymmetrical curves with the same neuralnetwork structure Other advantages that proposed modelpresents in comparison with other empirical tyre modelsbased on neural networks are that it learns quickly and it iseasy both to select the network structure and to calculatethe model parameters Results show the estimated modelcorrelates very well with the data obtained experimentally

2 RECURSIVE LAZY LEARNING

Unlike eager methods which compile input samples and use only the compilations to make decisions, lazy learning

methods perform less precompilation and use the inputsamples to guide decision making (Aha, 1998) Trainingexamples are simply stored for future use rather than used

to construct general and explicit description of the targetfunction This method is also referred to as memory-basedlearning Research works on lazy learning gave a newimpetus to the adoption of local techniques for modeling

and control problems (Bontempi et al., 1998) The

estimation of the value of the unknown function is solvedgiving the whole attention to the region surrounding thepoint where the estimation is required Each time aprediction is required for a specific query point, a set oflocal models is identified The generalization ability ofeach model is then assessed through a local cross-validation procedure Finally, a prediction is obtainedeither combining or selecting the different models on thebasis of some statistic of their cross-validation errors.The main advantages of lazy learning methods are thatthey can respond to unanticipated queries in ways notavailable to all eager learners, they have small trainingcosts, which make lazy algorithms particularly well suitedfor incremental learning tasks, they can store and adaptsolutions for subsequent problems, which can reduceproblem solving effort, they avoid negative interferenceexhibited by other modeling approaches such as multi-layer sigmoidal neural networks and, finally, they cangenerate precedent explanations, which are preferable toabstract explanations for many tasks

In the form of lazy learning, the locally weighted

learning (Atkeson et al., 1997) allows experiences to be

explicitly remembered, and predictions and generalizations

to be performed in real time by building a local model toanswer any particular input for which the function's output

NEURAL-EMPIRICAL TYRE MODEL BASED ON RECURSIVE LAZY LEARNING 823

is desired Locally weighted learning method provides an

approach to learning models of complex phenomena,

dealing with large amounts of data, training quickly, and

avoiding interference between multiple tasks during

control of complex systems (Schaal et al., 2000).

All algorithms of locally weighted learning consider that

the approximation function has the following form:

(1)

n-dimensional output vector and ε is a random variable

such that E[ε] = 0 and E[εi, εj] = 0 for The problem of

local regression can be stated as the problem of estimating

the value that the regression function

assumes for a specific query point x, using information

pertaining only to a neighborhood of x.

The estimation of y, , for a query point x is formed

from the normalized weighted average of the estimation

of all local models:

(2)

where w k are the activation strengths of the corresponding

receptive fields These weights w k are determined from the

size and shape of each receptive field, characterized by a

kernel function often chosen to be a Gaussian:

(3)

corresponds to a positive semi-definite

distance metric that determines the size and shape of region

of validity of the linear model For algorithmic reasons, it is

convenient to generate Dk from an upper triangular matrix

Mk in order to ensures that Dk is positive definite Mk is thedecomposed distance metric

Several authors have used local polynomial of low order

to model the relationship between input and output datawithin each receptive field, particularly linear modelsbecause the achieve a favorable compromise betweencomputational complexity and quality of result (Schaal andAtkeson, 1998):

(4)

linear model, formed by the coefficient vector bk and the

bias b 0,k of the linear model x1= ((x-ck)T, 1)T is a compactform of the center-subtracted, augmented input vector Figure 1 shows the architecture of the neural network forthe implementation of local, receptive field-based learning.The inputs are routed to all receptive fields, each of whichconsists of a linear and a Gaussian unit

(Bontempi et al., 1998) propose a model identification

methodology based on the use of an iterative optimizationprocedure to select the best local model among a set ofdifferent candidates This technique is based on recursiveleast squares methods to compute PRESS in an incrementalway (Schaal and Atkeson, 1998) introduce a modifiedmethod in which the parameters of the locally linear model

as well as the size and shape of the receptive field itself arelearned independently without the need for competition orany kind of communication The recursive lazy algorithmused to learn the learning model for each receptive field isdescribed by the following equations:

824 M J L BOADA, B L BOADA, D GARCIA-POZUELO and V DIAZ

(5)

(6)

(7)

(8)

For non-reliable initialization, it is assumed that

P(0) =λ·I, where I is the identity matrix and λ is a

forgetting factor in order to gradually cancel the

contributions from previous data points For simplicity, it is

assumed that the receptive field centers are uniformly

distributed in the whole input space and their size is given a

priori.

3 NUMERICAL RESULTS

In order to test the applicability of the proposed neural

model to estimate the characteristics of a tyre in combined

slip conditions, two different data sets are used The first

set of measured data used in this analysis is given for the

following tyre type: Michelin XZA 11R22.5, velocity

64.41 km/h, Fz0 = 90 kN These data were obtained from

(Palkovics and El-Gindy, 1993; Cabrera et al., 2004) The

influence of a tyre’s camber angle on its characteristics is

ignored, since the characteristics of the tyre under

examination are measured at zero camber In (Palkovics

and El-Gindy, 1993; Cabrera et al., 2004) the coefficients

for self-aligning torque were not obtained

During the training process, the symmetrical data set

was produced by inverting the original measured data

Table 1 and 2 represent the longitudinal forces and

lateral forces at pure slip, respectively, for different vertical

loads By multiplying the longitudinal and lateral forces at

pure slip with the coefficients given in Tables 3 and 4,respectively, the longitudinal and lateral forces undercombined slip conditions are calculated

For this first set of measured data, two different neuralnetworks are used: the first one for the mapping of:

ω

x+ 1(n 1+ )T⋅P n( ) x⋅ 1(n 1+ ) -–

Table 1 Longitudinal force versus slip ratio for different

vertical loads at pure slip [kN] Tyre Michelin XZA

NEURAL-EMPIRICAL TYRE MODEL BASED ON RECURSIVE LAZY LEARNING 825

(2)

For simplicity, each input (i, α and Fz) is divided into 10

regions uniformly distributed so that the number of

receptive fields for each neural network is 103 = 1000 The

value of the metric distance that determines the size of

region of validity of the linear model is chosen to 2000 for

both neural networks

The results of the training of the longitudinal force

measured data versus slip ratio at different side-slip angles

and three different vertical loads are shown in Figure 2

Figure 3 shows the training of the side force measured data

versus side-slip angle at different slip ratio and threedifferent vertical loads The experimental data are sketches

as circles, while the estimated ones are drawn as crosses.Neural model can accurately describe the measuredcharasteristics of the tyre These results are obtained afterthe experimental data are presented to the neural networkjust once It is only necessary one learning cycle so that thesystem learns quickly The neural network proposed by(Palkovics and El-Gindy, 1993) requires 60000 iterations

to learn properly

In addition to the graphical evidence of the effectiveness ofthe proposed model, the sum-squared errors between thenetwork output and the measured data have been acomplised:

Figure 2 Comparison between the experimentally obtained

and estimated training longitudinal force under longitudinal

and lateral slip conditions for three different vertical loads

Tyre Michelin XZA 11R22.5

Figure 3 Comparison between the experimentally obtainedand estimated training lateral force under longitudinal andlateral slip conditions for three different vertical loads TyreMichelin XZA 11R22.5