You are on page 1of 97

Modeling of a fluid flow in an internal combustion engine

J.J.M. Smits Report number WVT 2006.22


Supervisors: dr. ir. L.M.T. Somers ir. V. Huijnen Eindhoven University of Technology Department of Mechanical Engineering Division Thermo Fluids Engineering Section Combustion Technology

Abstract

The pollution by exhaust gasses produced by the growing number of vehicles on the road increases. A way to reduce the pollution is to change to an alternative fuel. This is a solution for the long-term. A solution for the short-term is to make engines cleaner and more ecient. To accomplish this, detailed information is needed on the processes which take place in an internal combustion engine. Performing experiments on an engine is dicult because of the complexity of this mechanical system, especially when ow measurements are involved. Numerical experiments have the advantage that an expensive and time consuming measurement set-up is not necessary. Because of the increasing power of computers, the processes in an internal combustion engine can be modeled in more and more detail. The main goal of this project is to model the uid ow in an internal combustion engine with the Computational Fluid Dynamic (CFD) package Fluent. Both stationary and dynamic simulations will be performed. In case of the rst variant the cylinder head is blown through and the stationary ow eld will be visualized. Subsequently the results will be compared with Particle Image Velocimetry (PIV) experiments and simulations which are performed with the FASTEST-3D code. The dynamic simulation visualizes the ow eld in case of a running engine. Because of the high velocities of the uid inside the internal combustion engine the ow will certainly be turbulent. Turbulence also plays an important role in the mixing process between air and fuel and the accompanying combustion process. Fluent oers dierent turbulence models and these will be validated on the basis of two simulations which are relevant for the ow inside the combustion chamber. These simulation are the backward-facing step (BFS) experiment and the free air jet. The backward-facing step experiment is an example of a wall bounded ow. The performance of the dierent Reynolds-averaged Navier Stokes (RANS) turbulence models in the neighborhood of walls can be tested. The ow between the valve and its seat is comparable with this geometry. Dierent simulations are performed in which the turbulence models are tested with dierent grid resolutions and wall functions. The grid requirements of the engine simulations make it impossible to solve the complete wall region. Wall functions are semi-empirical functions which solve this viscous dominated region and have the advantage that a more coarse grid is possible which reduces the simulation time. The Realizable k turbulence model in combination with the standard wall functions is the best performing model, taking note of the grid requirements. The second experiment for the validation of the turbulence models is the air jet. This experiment tests the performance of the models in the free space where the inuence of walls can be neglected. The air entering the combustion chamber is comparable with such a ow. Also in this experiment the turbulence models are tested with dierent sorts of grids. Just like in the BFS experiment the Realizable k - model is the better performing model and it will therefore be used in the internal combustion engine simulations. It is dicult to validate the stationary engine simulations because little reference material from the PIV experiments is available. The results of the simulations show agreement with the PIV experiments, but caution must be taken. For a good comparison more experimental PIV data is necessary. Comparison with the Large-Eddy Simulation (LES) results from the FASTEST code is easier because more simulation data is available. The ow eld depicted by the Fluent simulations close to the cylinder head, matches with the FASTEST results. The results further downstream deviate more. This can be attributed to the dierence in near-wall modeling approach. 3

A dynamic grid with a moving piston and valves is constructed to simulate the ow in a running engine. The complex geometry which a simulation of this calibre needs, makes it dicult to validate the ow. This is the reason the turbulence models are validated with simpler experiments. The rotating ow structures in case of the stationary simulations are also visible in the dynamic variant. The jet-like structures between both valves and the combustion chamber are also predicted by the dynamic simulations. An important conclusion is that the compression stroke itself adds no extra energy to the ow. It is therefore important for the mixing of air and fuel and the combustion process to have as much turbulence as possible at the beginning of the compression stroke. The dynamic simulations show realistic results which are useful for a better understanding of the processes in an internal combustion engine. The way is cleared for further research on this subject. The simulations can be extended with other subjects such as injection, combustion and turbo charging to obtain a complete model of an internal combustion engine.

Samenvatting

Door het toenemende aantal voertuigen op de weg neemt de vervuiling door uitlaatgassen sterk toe. Een mogelijkheid om deze vervuiling tegen te gaan is om gebruik te gaan maken van alternatieve brandstoen. Dit is een oplossing voor de lange termijn. Om op korte termijn de uitstoot van voertuigen te verminderen moeten schonere en eci entere motoren gemaakt worden. Hiervoor is een uitgebreide kennis nodig van de processen die plaatsvinden in een interne verbrandingsmotor. Door de complexiteit van een verbrandingsmotor is het lastig experimenten uit te voeren, vooral wanneer het gaat om stromingsmetingen in de motor. Numerieke experimenten hebben het voordeel dat er geen dure testopstelling nodig is. Door de toenemende rekenkracht van computers kunnen de processen in een verbrandingsmotor steeds beter en nauwkeuriger gemodelleerd worden. De doelstelling van dit project is om de gasstromen in een interne verbrandingsmotor te modelleren in het Computational Fluid Dynamics (CFD) pakket Fluent. Het gaat hierbij om stationaire en dynamische simulaties. In het eerste geval wordt de cilinderkop doorgeblazen en het stationaire stromingsveld gevisualiseerd. Vervolgens worden de resultaten vergeleken met Particle Image Velocimetry (PIV) experimenten en simulaties die zijn uitgevoerd met de FASTEST-3D code. Bij de dynamische simulaties wordt het stromingsveld in een draaiende verbrandingsmotor gevisualiseerd. Door de hoge snelheid van het gas in de verbrandingsmotor is deze zeker turbulent. Turbulentie speelt bovendien een zeer belangrijke rol bij de menging van brandstof en lucht en het daarbij behorende verbrandingsproces. Fluent biedt verschillende modellen die turbulentie modelleren en deze zullen worden gevalideerd aan de hand van twee simulaties die relevant zijn voor de stroming in de cilinder, namelijk de backward-facing step (BFS) en de air jet. Het backward-facing step experiment is een voorbeeld van een wand gebonden stroming. Het gedrag van de verschillende Reynolds-averaged Navier Stokes (RANS) turbulentie modellen in de buurt van wanden kan getest worden. De stroming van de lucht tussen de klep en zitting is hiermee te vergelijken. Verschillende simulaties zijn uitgevoerd waarbij de modellen in combinatie met verschillende grids en wandfuncties zijn getest. Door de grid eisen die de motor simulaties stellen, is het niet mogelijk het complete wand gebied op te lossen. Wandfuncties zijn semi-empirische functies die dit viskeus gedomineerde gebied oplossen en hebben als voordeel dat een grover grid mogelijk is wat de simulatietijd verkort. Het Realizable k - turbulentie model in combinatie met de standard wand functies presteert het beste in deze simulatie, de grideisen in acht nemend. Het tweede experiment voor de validatie van de turbulentie modellen is de air jet. Dit experiment test de verschillende modellen in de vrije ruimte waar de invloed van wanden verwaarloosbaar is. De lucht die via de kleppen de verbrandingskamer binnen stroomt is hiermee vergelijkbaar. Ook bij dit experiment zijn de modellen in combinatie met verschillende grids getest. Net zoals bij het BFS experiment is ook hier het Realizable k - model het best presterende turbulentie model en zal daarom worden gebruikt in de motor simulaties. Door gebrek aan vergelijkingsmateriaal van de PIV experimenten, is het moeilijk de stationaire motor simulaties te valideren. De resultaten van de simulaties lijken redelijk overeen te komen met de experimenten maar voorzichtigheid is geboden bij deze uitspraak. Voor een goede vergelijking zijn meer experimentele PIV data nodig. Door het uitgebreidere aanbod aan Large-Eddy Simulation (LES) resultaten van de FASTEST code zijn deze beter te vergelijken met de Fluent resultaten. Het stromingsveld 5

bij de Fluent simulaties dicht bij de cilinderkop komt goed overeen met deze code. Verder stroomafwaarts wordt door de FASTEST code het snelheidsveld overschat. Dit komt deels door het verschil tussen beide simulaties met betrekking tot het oplossen van het wand gebied. Er is een dynamisch grid met een bewegende zuiger en kleppen geconstrueerd om de gasstroom in een draaiende interne verbrandingsmotor te simuleren. Door de complexe geometrie die een simulatie van dit soort kaliber met zich mee brengt, is het moeilijk de stroming te valideren. Dit is mede de reden dat de turbulentie modellen gevalideerd zijn aan de hand van eenvoudigere experimenten. De draaiende structuren die bij de stationaire simulaties zichtbaar zijn, komen terug in de dynamische simulaties. Ook de jet-achtige structuren tussen de beide kleppen en de verbrandingskamer worden in de dynamische simulaties voorspeld. Een belangrijke conclusie is dat de compressieslag zelf geen energie toevoegt aan de lucht in de cilinder. Het is voor een goede menging en verbranding daarom van belang om zoveel mogelijk turbulentie te hebben aan het begin van de compressieslag. De dynamische simulaties laten realistische resultaten zien die van belang kunnen zijn voor het beter begrijpen van de processen in een interne verbrandingsmotor. De weg is vrijgemaakt voor verder onderzoek op dit onderwerp. Zaken als injectie, verbranding en turbo oplading kunnen toegevoegd worden om een zo volledig mogelijk beeld van de motor te krijgen.

List of symbols
Symbol a A B cp Cf D F h k kp K L m M M0 Pr p pw pf Re S Sij t t T T u uref u U U Uc Up va vc x x0 xr x Description acceleration area centreline velocity decay constant heat capacity at constant pressure skin friction coecient diameter force step height kinetic energy per unit of mass kinetic energy at point P diusion coecient length scale mass momentum ux per unit of mass momentum ux per unit of mass at orice Prandtl number absolute pressure static pressure area-averaged static pressure Reynolds number swirl number mean rate of strain tensor time dimensionless time temperature time scale velocity component centerline velocity dimensionless velocity velocity scale mean velocity centreline velocity velocity at point P axial velocity circumferential velocity Cartesian coordinate virtual origin reattachment point dimensionless position [Unit] [m s2 ] [m2 ] [] [J kg 1 K 1 ] [ ] [m] [N ] [m ] [m2 s2 ] [m2 s2 ] [m2 s1 ] [m ] [kg ] [m4 s2 ] [m4 s2 ] [] [P a] [P a ] [P a] [] [] [s1 ] [ s] [] [K ] [ s] [m s1 ] [m s1 ] [] [m s1 ] [m s1 ] [m s1 ] [m s1 ] [m s1 ] [m s1 ] [m ] [m] [m] [ ]

Symbol y yp y+ y

Description Cartesian coordinate wall distance at point P dimensionless wall distance dimensionless wall distance

[Unit] [m] [m] [] []

Greek symbols
Symbol ij t w i ij Description half channel height delta function dissipation rate per unit of mass dimensionless position angle molecular diusion coecient thermal conductivity coecient dynamic viscosity turbulent viscosity kinematic viscosity modied turbulent kinematic viscosity density dimensionless time scale shear stress dimensionless velocity scale specic dissipation rate vorticity mean rate of rotation tensor [Unit] [m] [] [m2 s3 ] [] [rad] [mol m2 s1 ] [W m2 K ] [ N s m 2 ] [kg m3 s1 ] [m2 s1 ] [m2 s1 ] [kg m3 ] [ ] [N/m2 ] [ ] [s1 ] [s1 ] [s1 ]

Contents
Abstract Samenvatting List of symbols 1 Introduction 2 Turbulence 2.1 Turbulence in general . . . . . . . 2.1.1 Governing equations . . . . 2.1.2 Burgers equation . . . . . . 2.1.3 RANS turbulence modeling 2.1.4 Kinetic energy . . . . . . . 2.1.5 Vorticity . . . . . . . . . . . 2.2 Turbulence modeling . . . . . . . . 2.2.1 Laminar . . . . . . . . . . . 2.2.2 Spalart-Allmaras model . . 2.2.3 Standard k - model . . . . 2.2.4 RNG k - model . . . . . . 2.2.5 Realizable k - model . . . 2.2.6 Standard k - model . . . . 2.3 Wall-bounded ows . . . . . . . . . 2.3.1 Wall functions . . . . . . . 2.3.2 Near-wall modeling . . . . . 2.3.3 Mesh guidelines . . . . . . . 2.4 Conclusion . . . . . . . . . . . . . 3 Backward-facing step 3.1 Theory . . . . . . . . . . . . . . . . 3.2 Performance of turbulence models 3.2.1 Laminar . . . . . . . . . . . 3.2.2 Spalart-Allmaras model . . 3.2.3 Standard k - model . . . . 3.2.4 RNG k - model . . . . . . 3.2.5 Realizable k - model . . . 3.2.6 k - model . . . . . . . . . 3.3 Model comparison . . . . . . . . . 3.4 Conclusion . . . . . . . . . . . . . 3 5 7 11 13 13 14 15 19 20 21 23 23 23 24 24 25 26 28 28 30 30 31 33 33 35 35 35 37 38 40 41 43 44

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . . 9

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

CONTENTS

10

4 Air jet 4.1 Theory . . . . . . . . . . . . . . . . 4.2 Performance of turbulence models 4.2.1 Spalart-Allmaras model . . 4.2.2 Standard k - model . . . . 4.2.3 RNG k - model . . . . . . 4.2.4 Realizable k - model . . . 4.2.5 k - model . . . . . . . . . 4.3 Kinetic energy . . . . . . . . . . . 4.4 Conclusion . . . . . . . . . . . . . 5 Engine simulations 5.1 Stationary engine simulations 5.1.1 Simulation description 5.1.2 Simulation results . . 5.2 Dynamic engine simulations . 5.2.1 Simulation description 5.2.2 Simulation results . . 5.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

47 47 49 50 51 52 53 53 54 58 61 61 61 63 68 68 69 75 79 79 80 83 85 85 87 91 91 91 92 95 95 97

6 Conclusion and recommendations 6.1 Final conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2 Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bibliography A Contour plots air jet A.1 Structured mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.2 Unstructured mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B In-cylinder model set-up B.1 2D model set-up . . . . . . . . . B.1.1 Geometry set-up Gambit B.1.2 Simulation set-up Fluent . B.2 3D model set-up . . . . . . . . . B.2.1 Geometry set-up Gambit B.2.2 Simulation set-up Fluent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Chapter 1

Introduction
The roads are getting more and more crowded everywhere. As a consequence, the pollution by exhaust gasses produced by the growing number of vehicles also increases. People are searching for ways to reduce the emission of vehicles. One possibility is to look for alternative fuels like hydrogen. It will take a long time to switch a whole economy from oil to an alternative. A solution for the short-term is to make engines cleaner and more ecient. Therefore a better understanding of the processes in an internal combustion engine is necessary. Performing experiments on an internal combustion engine is dicult. It is a complex mechanical system and therefore measurements are hard to carry out, especially ow measurements inside the cylinder. Numerical experiments have the advantage that an expensive and time consuming set-up is not necessary. Because of the increasing power of computers, the processes in an internal combustion engine can be modeled in more and more detail. The main goal of this project is to model a single cylinder of a heavy duty DAF 12.9 liter Diesel engine. The intake stroke with a moving piston and valves will be modeled in the Computational Fluid Dynamics (CFD) package Fluent. The main problem with in-cylinder simulations is that the mesh, where the geometry is build out of, is moving and deforming. This requires a special treatment in the simulation set-up. Because the velocities inside the engine are high, the ow will certainly be turbulent. Is is not possible to model all the turbulent ow characteristics because this requires too much computational time. There are several methods available which describe the processes in a turbulent ow, but not one of them is generally accepted. Validation of the dierent modeling methods is therefore necessary. In the second chapter of this thesis the essence of turbulence will be treated. The dierent models which are available in Fluent will be discussed and the mathematical basis of turbulence in general is presented. Also the governing equations, which describe the motion of the ow, will be briey treated. To choose the best turbulence model for the internal combustion engine simulations, the dierent models will be validated on the basis of two simulations. The rst one is the Backward-Facing Step (BFS) and will be treated in chapter three. This is an example of a wall bounded shear ow and the model behavior near a wall can be validated. The second simulation is the air jet. This is an example of a free wall-shear ow. With this experiment the behavior of the ow, when no wall-shear is involved, can be validated. This will be discussed in chapter four. The outcome of these two simulations will be compared with experimental data and result in a suitable turbulence model for the nal engine simulations. The fth chapter of this thesis deals with the simulations of the internal combustion engine. First a stationary simulation will be discussed. In this simulation the cylinder head is blown through and the ow prole will be compared with results found with experiments using Particle Image Velocimetry (PIV) on a stationary ow bench. Next the unsteady simulations with the moving piston and valves will be treated. Finally a general discussion will be presented. The best performing turbulence model will be chosen. The results of the engine simulation will be treated and compared with the experimental data. The last 11

12

section also includes the nal conclusion and the recommendations for further research.

Chapter 2

Turbulence
When modeling the air ow in an internal combustion engine, extreme uid velocities are involved. Because these high velocities, the Reynolds number is also substantial which indicates the presence of turbulence. Turbulent ows are characterized by uctuating velocity elds. These uctuations mix transported quantities such as momentum, energy and species concentration. Turbulence also plays an important role when modeling the combustion process in the combustion chamber. Increasing turbulence results, in case of non-premixed combustion, in a better mixing process of air and fuel. Modeling of turbulence is still an obstruction when solving practical ow problems. In practice most ows are turbulent. It requires too much computational eort to solve the governing equations exact to the smallest scales. Therefore these equations will be Reynolds-averaged. This greatly reduces the required computational eorts but introduces additional terms that need to be modeled in order to achieve a closure for the unknowns. There are a lot of turbulence models available. Not one model completely succeeded in describing the processes involved in a turbulent ow. This is because information is lost due to the averaging procedure and therefore turbulence is still an unsolved physical and mathematical problem. In this chapter dierent turbulence models will be discussed. Not the complete mathematical basis of the models will be treated but merely the properties and the dierences between them. Before the description of these models, turbulence in general will be dealt with to get a better understanding of this phenomenon. In this section the basic characteristics and equations are taken into consideration. Turbulent ows near a wall, the so called wall bounded ows, need special attention. Therefore specic boundary conditions need to be applied. The type of boundary condition to be used, depends on various properties. This will be treated in the second last part of this chapter. At the end of this chapter a sort classications will be made of the dierent models. One or more models, which are best suitable for the simulations of the internal combustion engine on the basis of the theory, will be chosen and founded.

2.1

Turbulence in general

In this section the phenomenon of turbulence will be discussed in more detail. Turbulence consists of swirl-like structures with varying dimensions, named eddies. Turbulence distinguishes itself by a chaotic, but not completely random, ow. The velocity measured in two points in the ow is correlated not only by place, but also as a function of time. A similar correlation can not be found when the ow is completely random. Certainly not all chaotic ows are turbulent, waves on the sea surface are a good example of a non-turbulent, chaotic ow. In this part dierent properties of turbulence will be discussed. First the governing equations which describe a viscous ow will be treated. This set of equations include the continuity, momentum and energy equation. This section also includes the Reynolds-averaging method which is used to reduce the computational cost of the solution of the governing equations. Secondly the Burgers equation will be treated. With this relatively simple equation the basics of turbulence can be solved analytically and give 13

2.1. TURBULENCE IN GENERAL

14

a better insight in this chaotic phenomenon. Kinetic energy is also an important quantity in turbulence. This will also be treated in this section. Vorticity is with respect to the dynamics the most important property. A compact denition of turbulence can be Chaotic vorticity. This property will be discussed in the last section.

2.1.1

Governing equations

The dynamics of a ow can be described by the governing equations. This description is based on three fundamental physical principles. These principles are: Conservation of mass Conservation of momentum (Newtons second law) Conservation of energy The rst principle results in the continuity equation (equation 2.1). This equation represents the conservation of mass in a control volume for a compressible ow ui + = 0. t xi (2.1)

In here is the density and ui the velocity in direction i. The symbols t and xi represent respectively the time and the position in direction i. The momentum equation is derived using Newtons second law. This equation is better known as the Navier-Stokes equation ui ui 1 p 2 ui , + uj = + t xj xi x2 j (2.2)

where p represents the pressure and the kinematic viscosity. The second term on the left-hand side is the convective term and the second term on the right-hand side is the diusion term. It says that the net force on a uid element equals the mass times the acceleration of the element. The forces acting on the uid element can be divided into Body and Surface forces. The Body forces act directly on the mass of the uid element. An example is the force as a result of gravitation (dropped in equation 2.2). The Surface forces act directly on the surface of the uid element. Two sources account for this force, namely the pressure distribution imposed by the outside uid surrounding the uid element and the shear and normal stress distributions acting on the surface by means of friction. The last fundamental physical principle which describes the dynamics of the ow, results in the energy equation (equation 2.3), here written in terms of the temperature cp T ui 2T + cp T = 2 , t xj xj (2.3)

where T is the temperature, cp the heat capacity coecient at constant pressure and the thermal conductivity coecient. The physical principle stated here is nothing more than the rst law of thermodynamics. When applied to a ow element, it states that the rate of change of energy inside the uid element is equal to the net ux heat into that element.

Chapter 2. Turbulence

15

2.1.2

Burgers equation

Non-linearity plays a fundamental role in turbulence. Because of this non-linearity the governing equations can not be solved exactly. To gain a better insight into turbulence, the Burgers equation (2.4) will be treated [20]. This model is analytical manageable and contains the essential ingredients of a turbulent ow. It can be written as u u 2u +u = 2 , t x x (2.4)

The second expression on the left-hand side of the Burgers equation is the non-linear term and is called the convective term. It can be compared with the convection term in the Navier-Stokes equation (2.2). The expression on the right-hand side can be interpreted as friction. Both terms will be discussed separately in the following. First, only the friction term will be taken into account. The resulting equation is better known as the diusion equation (equation 2.5). u 2u = 2 t x (2.5)

To come to a solution, a problem with the following begin- and boundary conditions will be considered t=0 x , , u = I (x) u=0 t,

where (x) the Dirac-delta-function is and I a constant which represents the magnitude of the velocity. The solution of this problem reads as follows
2 I ex /(4t) . u= 2 t

(2.6)

The solution is displayed in gure 2.1 for dierent values of t. For an increasing value of t, the gradient u/x decreases. It can be concluded that friction is gradient-weakening. This is called stabilizing. Subsequently the friction term will be omitted in equation 2.4. The resulting equation is called the non-linear advection equation u u +u = 0. t x (2.7)

As can be seen, the transport velocity u is only a function of position x and time t. The general solution of this equation reads u = f (x u t), (2.8)

where f is an arbitrary, dierentiable equation. This equation shows that the velocity u in case of t = 0 travels unchanged in the x, t-plane along the characteristic: x u t = constant. The slope of this characteristic is x/t = u. This means that the slope is determined by the solution itself. This phenomena is represented in gure 2.2. It can be seen from the gure that the gradient u/x gets sharper as a function of the time. After a certain period the solution consists of more than one value. This is of course physically not possible. It can be concluded that the non-linear term is gradient-sharpening and this is called destabilizing.

2.1. TURBULENCE IN GENERAL

16

1.4 t = 1.0 t = 0.25 t = 0.0625

1.2

velocity u [m/s]

0.8

0.6

0.4

0.2

0 5

xcoordinate [m]

Figure 2.1: Solution of the diusion equation at dierent periods in time

Figure 2.2: Solution of the non-linear advection equation. Initially a triangular shape exists

The behavior of equation 2.7 can best be compared with the analogy of the shallow water waves. Waves at the beach break because the non-linear eects make the waves more steep until their knock over and dissipate in the surf. In the previous part two eect are treated that inuence the solution of the Burgers equation. These effects counteract. The rst term is non-linear, gradient-sharpening and destabilizing. The second viscous term is gradient-weakening and stabilizing. The ratio between these two eects determines the solution of the Burgers equation. To characterize the ratio, this equation will be rewritten in dimensionless form. Therefore a set of dimensionless variables will be dened u = x = = t u U x L U t L

(2.9)

are the dimensionless velocity, position and time. U and L are the velocity- and length where u , x and t scales. When substituted in equation 2.4 this results in the dimensionless Burgers equation

Chapter 2. Turbulence

17

u u 1 2u , +u = x Re x 2 t

(2.10)

where Re = UL/ is the Reynolds number. This number represents the ratio between the non-linear advection term and the viscous term. When Re < 1 the viscous term dominates and the ow is characterized as stable or laminar. When Re 1 the non-linear advection term dominates and the ow is unstable. In the last situation the ow is characterized as turbulent. To better understand what is happening at the dierent Reynolds numbers, the exact solutions of the Burgers equation will be treated. For a Reynolds number Re 1 the term on the right-hand side of equation 2.10 is negligible and the solution is given by u= where L x L and with U= 4U 0 . 1 + 2 U 0 t/L U 2 tanh Ux 4 + x L , (2.11)

In gure 2.3 the solution of the Burgers equation is plotted. With length scales in the order of L the solution is dominated by the solution of the non-linear equation 2.7. In this case the exact solution can be approximated to U 2 U 2 x +1 L x 1 L

u u

= =

for all x < 0 for all x > 0 .

2 1.5 1 Velocity scale U [m/s] 0.5 0 0.5 1 1.5 2 0.1 Exact solution Nonlinear term, x < 0 Nonlinear term, x > 0 Viscous term

0.05

0 Length scale L [m]

0.05

0.1

Figure 2.3: Exact solution of the Burgers equation for Re

This solution is also plotted in gure 2.3. Because of the relative large length scales, this part of the solution is called the macrostructure. The non-linear processes dominate in the macrostructure and friction can be neglected. There is a small region where the gradient is that large that the friction term can not be neglected.

2.1. TURBULENCE IN GENERAL

18

This region is called the microstructure and is dominated by friction. The involved length scales are in the order of x = / U and this results in a local Reynolds number Re < 1. In this case the solution can be approximated by (viscous term in gure 2.3) u= U tanh 2 Ux 4 .

The Burgers equation has led to an important insight in turbulence. For large Reynolds numbers two different structures can be found in the solution of the Burgers equation, namely the micro- and macrostructures. In the following these two structures and the properties will be discussed in more detail.

Macrostructure
The structure with the largest dimensions is called the macrostructure. This structure can be described by a length scale L and a velocity scale U and is dominated by non-linear processes. As mentioned before a Reynolds number Re 1 is involved and the viscous eects are negligible. Essentially the macrostructure is described by the limit Re and as a consequence, is independent of the Reynolds number. The only medium parameter is the kinematic viscosity and can be found in the Reynolds number. Because the macrostructures are independent of the Reynolds number, it is also independent of the kinematic viscosity. This is called the Reynoldss similarity and can be expressed as: Turbulence is a property of the ow, not the medium. A turbulent ow is dissipative. This means the ow loses its kinetic energy and as a result fades out if there is no extra energy input. The kinetic energy per unit of mass k scales with k U 2 . For the kinetic energy k the following equation holds (this will be derived later) dk = dt where is the dissipation. One of the most important results from the turbulence theory is [20] U3 , L (2.13) (2.12)

which means that dissipation scales with the macrostructure. This is better known as the Kolmogorov relation. This relation can be explained as: a turbulent eddy with energy U 2 loses in one timescale T L/U its energy, that is in period T the eddy breaks up because of instability and disintegrates into smaller eddies. This also follows from the properties of equation 2.7. Turbulence is also diusive. It is known that a turbulent ow is capable of ecient mixing. The larger eddies stir the ow round a scale L and ow parameters will likely be transported and then mixed across this distance. The diusion equation for a turbulent ow can be written as 2X X =K 2 , t x (2.14)

for the concentration X . K is the diusioncoecient, also known as the turbulent diusioncoecient, and can be scaled with K U L. The turbulent diusioncoecient K is much larger than the moleculair diusioncoecient . In case of a Prandtl-number P r / 1 the Reynolds number Re ( K/) 1, which proves that diusion belongs to the macrostructure.

Chapter 2. Turbulence

19

Microstructure
In the previous section the macrostructure is discussed. This structure loses its kinetic energy by an instability in a typical time scale. The kinetic energy can only be dissipated, that is converted into heat, by viscosity. This viscous dissipation happens in the microstructure because the velocity gradient is that large that viscosity cannot be neglected. Viscosity is the rst scale parameter of importance in the microstructure. The macro scales U and L do not play a role in the microstructures. When larger eddies break up into smaller ones, this information gets lost. The only information the microstructure receives from the macrostructure is the amount of energy that has to be dissipated. This is characterized by . The scale parameters for the microstructure are therefore and . The following scales for length, time, and velocity can be derived 3
1 2 1 4

= =

(2.15)

= ( ) 4 .

These are better known as the Kolmogorov -scales. This results in a Reynolds number for the microstructure of Re = / = 1 , which indicates that the microstructures are dominated by friction as determined by the diusion limit of the Burgers equation (equation 2.5).

Summarized: The macrostructure is fed with energy from the main ow by instability processes. The large, unstable, energy rich eddies break up in smaller ones. This process is called the cascade-process. This process is repeated until the microstructure is reached. At this point the kinetic energy is dissipated into heat by viscosity.

2.1.3

RANS turbulence modeling

When modeling a turbulent ow through a complex geometry, the governing equations can not be solved all the way down to the Kolmogorov -scales. This will cost to much computational eort. To still model turbulence in a ow, the small-scale turbulent uctuations have to be omitted. A way to do this is Reynolds-averaging. A disadvantage of this method is that it brings an additional term in the governing equations that needs to be modeled in order to achieve a closure for the unknowns. The Reynoldsaveraged Navier Stokes (RANS) equations determine the transport of the averaged ow quantities, with the complete range of turbulent scales being modeled as mean ow properties which do not change extremely over place and time. This approach greatly reduces the computational eort and resources and is therefore used in many practical engineering applications. When using the RANS-based modeling approach the variables are split up into a mean and a uctuating component. This is often called the Reynolds decomposition. For the velocity this results in ui = ui + ui with i = 1, 2, 3

and for the other quantities like pressure and energy this results in =+ . These expressions can be substituted in the continuity and momentum equation. After taking the time

2.1. TURBULENCE IN GENERAL

20

average this results in the time-averaged continuity and momentum equations for a compressible ow (equation 2.16 and 2.17). + (ui ) = t xi (ui ) + (ui uj ) = t xj ui uj 2 ui + ij + xj xi 3 xi

0 p + xi (ui uj ) xj

(2.16)

xj

(2.17)

Where ij is the delta function. The above set of equations is called the Reynolds-averaged Navier-Stokes (RANS) equations. The variables now represent the time-averaged values. In the momentum equation (equation 2.17) an additional term has appeared ui uj which represent the eects of turbulence. This term can be interpreted as the transport in the j -direction of momentum in the i -direction. This kind of momentum transport has the same eect as a stress on a surface and is therefore also called the Reynolds stress. It must be modeled in order to close the momentum equation. The closure model, which is used in all the turbulence models described in chapter 2.2, employs the Boussinesq hypothesis [9] to relate the Reynolds stresses to the mean velocity gradients. It is given by ui uj = t ui uj + xj xi 2 3 k + t ui xi ij , (2.18)

where t is the turbulent viscosity and k is the kinetic energy. The advantage of the Boussinesq approximation (or K -theory) is the relative low computational cost associated with the computation of the turbulent viscosity. A disadvantage is that the turbulent viscosity is assumed to be an isotropic scalar quantity and this is not strictly true. In the above, only the RANS-based modeling is discussed. Fluent oers another type called Large-Eddy Simulation (LES). This modeling approach computes the large eddies explicitly in a time-dependent simulation using the ltered Navier-Stokes equations. Filtering is essentially a mathematical manipulation of the exact Navier-Stokes equations to remove the eddies that are smaller than the size of the lter. This method requires a signicant amount of computational time for complex and high Reynolds number ows. It also needs a highly accurate discretization scheme. Because these reasons the LES approach is not suitable for the engine simulations and will therefore not be discussed any further.

2.1.4

Kinetic energy

The microstructures in a turbulent ow dissipate the kinetic energy into heat. This is already discussed in the previous section. In the following section this pronouncement will be founded on the basic of the kinetic energy equation. First the velocity will be split up into a mean and a uctuation ui = ui + ui , which is already called Reynolds decomposition. The kinetic energy per unit of mass of the turbulent velocity uctuations is 2 represented by the variance k = 1 2 ui . The equation of the kinetic energy k is given by equation 2.19. This equation is found by multiplying the momentum equation (2.2) with ui . k k Dk + uj = Pk + Tk + k + Dk Dt t xj where

(2.19)

Chapter 2. Turbulence

21

Pk Tk + k + Dk

= =

ui uj xj

ui xj 1 k p uj + 0 xj

uj k
2

ui xj

Density and temperature eects are neglected. The dierent terms on the right side of equation 2.19 will be discussed separately in the following. Production The rst term Pk represents the production of kinetic energy in the turbulent velocity uctuations. This term is always positive. The turbulence receives energy from the mean ow via this deformation work. Instability mechanisms like the Kelvin-Helmholtz instability take care of this energy transport. Both the velocity gradient ui /xj and the Reynoldsstress ui uj scale with the macro scales U and L, which result in a scale U 3 /L for the production term. Transport The three middle terms Tk + k + Dk form together the transport contribution; the terms describe the redistribution of the kinetic energy in the domain. It is built out of three parts. The rst part is the transport by velocity uctuations (Tk ) which scales with U 3 /L. The second expression is the transport by pressure uctuations (k ) which scales in the same way as the previous term. The last expression is the transport by viscosity (Dk ). It scales, in contrast to the previous two terms, with U 2 /L 2 . This indicates that for large Reynolds numbers the viscous contribution can be neglected with respect to the other terms. In most ows the complete transport term can be neglected because the transport of kinetic energy is minimal. Dissipation The last term which contributes to the transport expression is the dissipation . As mentioned before, dissipation converts the kinetic energy in the microstructures into heat. Because of this, the term is always negative. Equation 2.19 is a balans equation. When turbulence is in equilibrium the left-hand side of equation 2.19 is equal to zero. For most ows the transport term can be neglected, which means that the dissipation term scales with the production term U 3 /L. This is the Kolmogorov relation and is already mentioned as one of the main laws of turbulence.

2.1.5

Vorticity

Vorticity is a natural way to describe turbulence. In the previous section turbulence is characterized as Chaotic vorticity. It can be said that a ow without uctuating vorticity is not turbulent. Vorticity is dened as the rotation of the velocity eld (equation 2.20) and can be interpreted as the turning of a uid element. w =u (2.20)

When the line elements stretch, the vorticity increases. This is called vorticity-line-stretching. This phenomenon causes a vortex to rotate faster. The required energy will be provided by the deformation eld. Because of instabilities in the ow, large eddies will be formed which determine the macrostructure. The large eddies will break up into smaller eddies which results in the structure of the largest eddies being highly directional, or anisotropic, and ow dependent. This proces is already called the cascade -proces. The vorticity-line-stretching mechanism accounts for this cascade -proces. The larger eddies deform the

2.1. TURBULENCE IN GENERAL

22

smaller eddies and because of this the vorticity of the smaller eddies increases. Energy is also transported from the larger to the smaller eddies. The largest vorticity can therefore be found in the microstructures. In the following an equation for the vorticity uctuations will be derived. First the vorticity will be 2 split up into a mean and a uctuation i = i + i . The variance 1 2 i is dened as the enstropy. For homogeneous turbulence equation 2.21 holds exactly. = i2 . (2.21)

This equation is also a good approach for ows with a large Reynolds number. In this case turbulence can locally be considered homogeneous. Enstropy will be dominated by the microstructure where the viscous dissipation takes place. Kinetic energy k gives information about the macrostructures and enstropy information about the microstructures. The equation for the enstropy is D 1 2 1 2 1 2 i i + uj = P + T + D + S Dt 2 t 2 xj 2 i where i xj 1 uj i 2 + 2 xj 1 2 2 i (2.22)

P T + D S

= uj i = xj

= i j sij + i j sij + i j sij = i xj

The left-hand side of 2.22 describes the enstropy change of a point that travels with the average velocity of the ow. The four dierent terms which describe this change will be discussed next. Production The term P is considered as the gradient production of enstropy. This expression can be compared with the gradient production term in the equation for the kinetic energy (2.19). Via this term enstropy is exchanged between the mean and the uctuating vorticity eld. Transport The redistribution of the enstropy is taken care by the transport terms T + D . The rst term T is the transport by velocity uctuations and the second term D is the transport by viscosity. Stretching This is the combination of terms S which accounts for the proces of vorticity-line-stretching. There is contribution of both the mean deformation eld sij as the uctuation deformation eld sij . Moleculair destruction This loss term describes the destruction of enstropy by the moleculair diusion. Therefore this term is always negative.

Chapter 2. Turbulence

23

2.2

Turbulence modeling

As mentioned before, turbulence is still an unsolved problem. There is not one model which completely describes the proces of this natural phenomenon. Over the last decades several models have been introduced. Every model has dierent conditions for which it performs best and has its advantages and disadvantages. The CFD package Fluent, which is used to set-up the simulations, also provides several turbulence models. In order to make a general decision which model to choose, the dierent models will be discussed briey. The pros and cons will be treated as well as the limitations and the basic equations. First the case when no turbulence is modeled, will be considered. After this the actual turbulence models will pass in review.

2.2.1

Laminar

The rst case that is mentioned, is the case when no turbulence model is used. The governing equations are solved and no extra equations for the kinetic energy and dissipation will be involved. When this method is used in case of a turbulent ow, the results deviate from the reality. It is possible to get a correct simulation but extreme requirements are necessary with regards to constructing the grid. The grid resolution must be in the order of the smallest length scales and a Direct Numerical Solution (DNS) study is performed. A grid with such a high resolution is nearly impossible to construct in case of a complex geometry and when constructed it requires far too much computational eort. Therefore this option is not feasible for a complex engine simulation.

2.2.2

Spalart-Allmaras model

The rst model that will be discussed is the Spalart-Allmaras turbulence model. This is the only oneequation model that will be treated. It is a relative simple low Reynolds number model that solves a modeled transport equation for the kinetic eddy viscosity. It was rst designed for aerospace applications involving wall-bounded ows and has shown good results for boundary layers subjected to adverse pressure gradients [6]. In this model is the transported variable. It is similar to the turbulent kinematic viscosity except near the walls. The governing equation for this transport variable is ( ) + (ui ) = G + t xi 1 A xj ( + ) xj + Cb2 xj
2

Y .

(2.23)

The rst term on the left-hand side of equation 2.23 is the increase of turbulent viscosity. The second term represents the convective transport by the mean ow. On the right-hand side of 2.23 the terms represent respectively the production of turbulent viscosity from the mean ow gradients (G ), the transport of the viscosity due to both molecular and turbulent viscosity, the dissipation of turbulent viscosity by the small scales and last the destruction of turbulent viscosity in the near-wall regions due to viscous damping (Y ). The turbulent viscosity t is computed from t = f , where the viscous damping function f is given by f = 3 3 3 + C 1 and = .

2.2. TURBULENCE MODELING

24

Because the Spalart-Allmaras model is relatively new, its applicability in complex ows is uncertain. One-equation models can not rapidly accommodate changes in the length scales. This is the case when a ow changes suddenly from a wall-bounded ow to a shear ow. When simulating the ow in an internal combustion engine, this is certainly the case. The air ows for example through the narrow opening between valve and valve seat into the combustion chamber.

2.2.3

Standard k - model

Another class of turbulence models are the two-equation models. The simplest one is the standard k model, which is proposed by Launder and Spalding (Launder [16]). It is widely used in turbulence simulations because of its general applicability, robustness and economy. The two transport equations for the kinetic energy and dissipation rate are solved to form a characteristic scale for both turbulent velocity and length. These scales represent the turbulent viscosity. The equations for the kinetic energy (2.24) and dissipation rate (2.25) are given below. (k ) + (kui ) = t xi xj ( ) + ( ui ) = t xi xj + t t k k + Gk xj

(2.24) (2.25)

2 + C1 Gk C2 xj k k

The turbulent viscosity is calculated using the following equation t = C k2 , (2.26)

where C is a constant. It is not dependent on the direction and is therefore isotropic. The term Gk in 2.24 and 2.25 represents the generation of turbulence kinetic energy due to the mean velocity gradients. This energy is fed to the small scales by the large scales via the vortex stretching mechanism. The small scales dissipate this energy into heat when it reaches the Kolmogorov scales. Gk is calculated using equation 2.27. The term in front of the velocity gradient uj /xi is the Reynolds stress which is mentioned at the beginning of this chapter (equation 2.18). Gk = ui uj uj . xi (2.27)

The rst term on the right-hand side of both equation 2.24 and 2.25, is the transport due to molecular and turbulent viscosity. The second term on the right-hand side of the dissipation equation (2.25) represents the rate at which the large scales supply energy to the smaller scales. The last term of equation 2.25 represents the dissipation at the small scales. In this semi-empirical model the equation for the kinetic energy is derived mathematically, while the equation for the dissipation is derived using physical reasoning. Further, in the derivation of the standard k - model, it is assumed that the ow is fully turbulent and the eects of molecular viscosity are negligible. Therefore this model is only valid for fully turbulent core ows. This is generally not an issue in case of the simulation of the internal combustion engine.

2.2.4

RNG k - model

The RNG k - turbulence model is derived from the Navier-Stokes equations using a mathematical technique called renormalization group (RNG) methods [2]. The analytical derivation results in a model with dierent constants than those in the standard k - model. Also additional terms and functions will appear in the transport equations for the kinetic energy and the dissipation which will be discussed in the following.

Chapter 2. Turbulence

25

In equation 2.28 and 2.29 the transport equations of the RNG k similar form as the standard k - model with a few dierences. (k ) + (kui ) = t xi xj ( ) + ( ui ) = t xi xj ef f xj k xj

model are shown. They have a

k ef f + C1

+ Gk
2

(2.28) (2.29) equation (2.29). This

G k C2

The main dierence with the standard k - model is an additional term R in the term is given by equation 2.30 R = C 3 (1 /0 ) 2 1 + 3 k
u

(2.30)

ui j where Sk/ , S = 2Sij Sij , Sij = 1 the mean rate of strain tensor and 0 and are 2 xj + xi constants. The last two terms of equation 2.29 can be merged. This results in a new term 2.31.

C2

where

C2 + C2

C 3 (1 /0 ) 1 + 3

(2.31)

becomes In regions of relative low strain rates ( < 0 ) the R term makes a positive contribution and C2 larger than C2 . As a result, for weakly to moderately strained ows, the RNG model tends to give results largely comparable to the standard k - model. On the other hand, for large strain rates ( > 0 ), the becomes less than C2 . In comparison with R term makes a negative contribution. The value of C2 the standard k - model the smaller destruction of increases , reducing k and eventually the eective viscosity. As a result, in rapidly strained ows, the RNG model yields a lower turbulent viscosity than the standard k - model. Thus the RNG k - model is more responsive to the eect of rapid strain and streamline curvature than the standard k - model, which explains the better performance of the RNG model for a wider class of ows.

Another dierence with the standard k - model is the calculation of the turbulent viscosity. In case of the RNG model the viscosity is calculated by the following equation d 2 k = 1.72 d , 3 1 + C (2.32)

where = ef f /. Equation 2.32 is integrated to obtain an accurate description of how the eective turbulent transport varies with the eective Reynolds number, allowing the model to better handle lowReynolds number and wall-bounded ows. In case of high Reynolds numbers equation 2.32 reduces to t = C k2 , (2.33)

which is the same formulation as in the standard k - model.

2.2.5

Realizable k - model

The last member of the k - family is the Realizable k - model [21]. It is a relative new turbulence model. In contrast to the other two models, this model is Realizable. This means the model satises certain mathematical constraints on the Reynolds stresses which are consistent with the physics of turbulent ows. It diers from the standard k - model in two major ways. The rst dierence is the formulation

2.2. TURBULENCE MODELING

26

of the turbulent viscosity and the second dierence is the new transport equation for the dissipation rate. The transport equation has been derived from an exact equation for the transport of the mean-square vorticity uctuation. In 2.34 and 2.35 the transport equations for both the kinetic energy and the dissipation rate are given. (k ) + (kui ) = t xi xj ( ) + ( ui ) = t xi xj where C1 = max 0.43, , +5 k =S , S= 2Sij Sij + t t k k + Gk xj

(2.34) (2.35)

2 + C1 S C2 xj k+

The equation for the kinetic energy has the same expression as the standard and RNG k - models. Only the equation for the dissipation rate is quite dierent. The production term in the equation (all but last term on right-hand side of equation 2.35) does not involve the production of k . This in contrast with the other two family members. It is believed that the present form better represents the spectral energy transfer [21]. The last term in equation 2.35 is the destruction term. A desirable feature is that this term does not have any singularity. The denominator never vanish, even if k vanishes or becomes smaller than zero. Like the other k - models, the turbulent viscosity is calculated using equation 2.26. The only dierence is the constant C . In case of the Realizable k - model this term is no longer a constant. It is computed from C = 1 , A0 + AS kU / (2.36)

where U ij ij , Sij Sij + ij = ij =

1 2

uj ui xi xj

where A0 and AS are constants and ij is the mean rate-of-rotation tensor. It can be seen that C is a function of the mean strain and rotation rates, the angular velocity of the systems rotation and the turbulence elds (k and ). The turbulent viscosity is dependent of the direction and is therefore anisotropic The Realizable k - model has shown substantial improvements over the standard k - model [21]. Especially when the ow features include strong streamline curvature, vortices and rotation. Since the model is relatively new, it is not clear in exact which aspects the Realizable k - model outperforms the RNG model. However, initial studies have shown that the Realizable model provides the best performance of all the k - models for several validations of separated ows and ows with complex secondary ow features. These properties make this turbulence model very suitable for the engine simulations.

2.2.6

Standard k - model

The last turbulence model that will be discussed, is the k - model. This model is based on the Wilcox k - model [24] [20]. It incorporates modications for low-Reynolds number eects, compressibility and shear ow spreading. The Wilcox model predicts free shear ow spreading rates that are in close agreement with measurements for far wakes, mixing layers and dierent types of jets [6]. Therefore this model

Chapter 2. Turbulence

27

is applicable to both wall-bounded ows and free shear ows. The k - model is a two equation semi-empirical turbulence model. The transport equation for the kinetic energy is comparable with the previous k - models. The equation for the dissipation on the other hand, is dierent. Instead of the dissipation per unit of mass , the specic dissipation rate is used. This quantity can be seen as the ratio of to k . The two transport equations for the kinetic energy and specic dissipation rate are represented by (k ) + (kui ) = t xi xj ( ) + (ui ) = t xi xj t k k + Gk Yk xj + G Y , xj

+ +

(2.37) (2.38)

where Gk represents the generation of kinetic energy and G the generation of the specic dissipation rate. The dissipation of k and due to turbulence are represented by Yk and Y respectively. The turbulent viscosity t is calculated using t = k , (2.39)

where is the low-Reynolds correction factor. This coecient is given by


= + Ret /Rk 0 1 + Ret /Rk

where Ret =

k ,

(2.40)

and Rk is a constant. As can be seen, for large Reynolds numbers this coecient is equal to one. The generation of kinetic energy Gk is calculated using equation 2.27. The generation of the specic dissipation on the other hand, is computed by G = Gk , k (2.41)

where is again the low-Reynolds correction factor. It is calculated in a slightly dierent way, namely = 0 + Ret /R 1 + Ret /R , (2.42)

where R is a constant. It can again be seen that for large Reynolds numbers this coecient is equal to one. For the calculation of the dissipation of k and the reader is referred to the Fluent users guide [6]. As mentioned earlier in this section, the k - turbulence model gives good results for both wall bounded ows and free shear ows because the low Reynolds number correction factors. This property makes the k - model also very suitable for the engine simulations. Fluent oers a modied version of the k - model, named the Shear-Stress Transport (SST) k - model. This model diers from the standard k - model in one major way. This main dierence is the gradual change from the standard k - model in the inner region of the boundary layer to a high-Reynolds number version of the k - model in the outer part of the boundary layer. This ensures that the model equations behave appropriately in both the near-wall and far-eld zones. The SST k - model will not be discussed any further, for more information the reader is referred to the Fluent users guide [6].

2.3. WALL-BOUNDED FLOWS

28

Turbulence model constants In tabel 2.1 the constants which are used in the dierent turbulence models are grouped. Table 2.1: Turbulence Cb1 = 0.1355 Cb2 = 0.622 Cw1 = 3.266 Cw3 = 2.0 C1 = 1.44 C2 = 1.92 C1 = 1.42 C2 = 1.68 = 0.012 C1 = 1.44 C2 = 1.9 =1 = 0.52 R = 8 Rk = 6 k = 2.0 = 2.0 k,1 = 1.176 ,1 = 2.0 i,2 = 0.0828
model constants

Spalart-Allmaras Standard k RNG k Realizable k Standard k -

A = 2/3 0.4187 C = 0.09 C = 0.0845 k = 1.0 0 = 1/9 R = 2.95 k,2 = 1.0

Cv1 = 7.1 k = 1.0 0 = 1.0 = 1 .2 = 0.09 = 1.5 1 = 0.31

Cw2 = 0.3 = 1 .3 0 = 4.38 i = 0.072 Mt0 = 0.25 i,1 = 0.075

SST k -

2.3

Wall-bounded ows

Turbulent ows need special attention near walls. This is because turbulent ows are signicantly aected by the presence of walls. In the rst place the mean velocity eld is aected through the no-slip condition that has to be satised at the wall. Another issue is that close to the wall the viscous damping reduces the tangential velocity uctuations, while kinematic blocking reduces the normal uctuations. Toward the outer part of the near-wall region however, the turbulence is rapidly increased by the production of turbulence kinetic energy due to the large gradients in the mean velocity. The near-wall modeling signicantly aects the reliability of the solution because the walls are the main source of vorticity and turbulence. After all, it is in the near-wall region that the solution variables have large gradients. Therefore an accurate representation of the ow near walls is necessary for successful predictions of wall-bounded ows. Experiments have shown that the near-wall region can be divided into three layers. The innermost region is called the viscous sublayer. In this region the molecular viscosity plays an important role in momentum and mass transfer. On the other side of the near-wall region the fully turbulent or log-law region can be found. Like the name says, this region is dominated by turbulence. Between these two layers an interim region can be found. This region is called the buer layer or blending region. In gure 2.4 the dierent layers are displayed. There are two ways to solve the near-wall region. The rst one does not solve the viscous aected region, that is the buer layer and viscous sublayer. Instead it uses semi-empirical formulas to bridge the region between the wall and the fully turbulent ow. These formulas are called wall functions. The second approach modies the turbulence models to enable the viscosity aected region to be resolved all the way to the wall. In this case the mesh must be ne enough. This method is called near-wall modeling. Both approaches are presented in gure 2.5 and will be discussed in the following sections.

2.3.1

Wall functions

As mentioned before, the region close to the wall can be solved in two dierent ways. The rst approach uses semi-empirical formulas called wall functions and is used in the dierent k - models. These functions are used to link the solution variables at the near-wall cells and the corresponding quantities on the wall. The viscosity aected region does not have to be resolved which saves computational resources. This is why the wall function approach is very popular in high Reynolds number ows and is therefore often

Chapter 2. Turbulence

29

eperimental data
U* 1

inner layer

ln ( Ey * )

Dimensionless velocity U *

U * y*

viscous sublayer

buffer layer
y 5
y 60

log-law region

outer layer

Dimensionless wall distance y

(log)

Figure 2.4: Subdivisions of the near-wall region. Dimensionless velocity U as a function of the dimensionless
wall distance y +

turbulent core flow

semi-empirical formulas

buffer layer and viscous sublayer


Near-wall modelling approach

Wall function approach

Figure 2.5: Dierent near-wall treatments

used with the k - turbulence models. Other properties that make this solution popular are robustness and reasonable accuracy. The wall functions include the law-of-the-wall for the mean velocity (equation 2.43) and temperature and formulas for the near-wall turbulent quantities. Two dierent wall functions are available in Fluent, namely the standard [17] and the non-equilibrium wall functions [14]. Both versions will briey be discussed. Standard wall functions The law-of-the-wall (or log-law) for the mean velocity in case of the standard wall functions is given by, U = where U Up C kp w /
1 /4 1 /2

1 ln(Ey )

(2.43)

and

C kp yp ,

1 /4 1 /2

and E is a constant. This logarithmic law for the mean velocity is known to be valid for 30 < y < 300. The dimensionless wall unit y is the distance from the wall boundary to the cells centroid. In Fluent

2.3. WALL-BOUNDED FLOWS

30

this law is used for y > 11.225. Closer to the wall, where y < 11.225, Fluent applies the laminar stress-strain relationship which is given by U = y , (2.44)

which also meets the no-slip boundary condition at the wall. Other quantities like temperature, species concentration, kinetic energy and dissipation are also taken care o by the wall functions but will not be discussed further [6]. The standard wall functions work reasonably well for a broad range of wall-bounded ows. However, when the ow characteristics dier too much from the ideal conditions, these functions become less reliable. The non-equilibrium wall functions can improve the results in such situations. Non-equilibrium wall functions In addition to the standard wall functions, a two-layer-based, non-equilibrium wall function is also available in Fluent. In this approach the log-law for the mean velocity is sensitized to pressure-gradient eects. The log-law is of the same form as equation 2.43. The only dierence is that the velocity Up is replaced by U =U where yv is the physical viscous sublayer thickness. Another dierence with the standard version is the two-layer-based concept, which is adopted to compute the budget of turbulence kinetic energy in the wall-neighboring cells. For more detail, the reader is referred to the Fluent users guide [6]. Because of the capability to partly account for the eects of pressure gradients, the non-equilibrium wall functions are recommended for use in complex ows involving separation, reattachment and impingement where the mean ow and turbulence are subjected to severe pressure gradients and change rapidly. Enhanced wall treatment This last wall function option will be mentioned briey. Fluent can combine the two-layer model with the enhanced wall functions. In this case the near-wall model approach will have the accuracy of the standard two-layer approach for ne near-wall meshes and at the same time will not signicantly reduce the accuracy for wall function meshes.
yv 1 /4 1 /2 C kp

1 dp yv ln 2 dx k

y yv

y yv y2 + v , k

(2.45)

2.3.2

Near-wall modeling

In case of the near-wall modeling approach, the viscosity aected region will be resolved all the way to the wall. Therefore the turbulence models have to be modied. The near-wall approach is also valid for low-Reynolds number ows. The only requirement is that the near-wall mesh must be ne enough to be able to resolve the laminar sublayer. This method is used in the Spalart-Allmaras and k - turbulence models.

2.3.3

Mesh guidelines

In this last section the mesh guidelines for the dierent turbulence models will be discussed. When more than one near-wall treatment can be used, each guideline will be given. In case of the wall function

Chapter 2. Turbulence

31

approach the dimensionless distance y (equation 2.43) is often used. This parameter represents the distance from the wall to the centroid of the cell. Another wall unit which is often used is y + . This dimensionless parameter is given by y+ = u y with u = ui u i uj y (2.46)

When the rst cell is placed in the log-layer (y + > 60) y is approximately equal to y + . These quantities are also equal in equilibrium turbulent boundary layers. The wall unit y is used for the mean velocity in the law-of-the-walls. In the other cases y + is used. In the following the near-wall mesh guidelines for the dierent turbulence models will be treated. It should be noted that the wall units y and y + are not xed geometrical properties, but actually solution dependent. Mesh guidelines for the Spalart-Allmaras model The Spalart-Allmaras turbulence model is a low-Reynolds number model. Therefore the mesh should be ne enough to properly resolve the viscous-aected region. However, the boundary conditions for this model have been implemented so that the model will work on coarser meshes. If the mesh is ne enough, a mesh spacing of y + = 1 is required. When using coarser meshes a spacing of at least y + = 30 is necessary. Mesh guidelines for the k - models The standard wall functions are the default option in Fluent for the k - turbulence models. For this type of wall function and the non-equilibrium variant, each wall-adjacent cells centroid should be located within the log-law, i.e. 30 < y + < 200. A value closer to the lower bound is most desirable. Although Fluent employs the linear laminar law when y < 11.225, using an excessively ne mesh near the walls should be avoided. This because the wall function cease to be valid in the viscous sublayer. It is possible to work with ner meshes but then the wall-adjacent cells should be placed in the buer layer (y + < 5). When using the enhanced wall treatment, mesh guidelines are dierent. This wall function is designed to extend the validity of the near-wall modeling beyond the viscous sublayer. Therefore a mesh must be constructed that will fully resolve the viscosity-aected region. This results in a wall-adjacent cell with a y + in the order of 1. However, a higher value is possible as long as it is well inside the viscous sublayer (i.e. y + < 4 5). Another criterion is that there are at least 10 cells within the viscosity-aected region (Rey < 200). This is necessary to resolve the mean velocity and turbulence quantities. To avoid an excessively ne mesh, the value for y should be around 11. Mesh guidelines for the k - model The last mesh guidelines that will be discussed, are those of the k - model. The k - model is available in both high and low Reynolds number variants. If the Transitional Flow option is enabled in Fluent, the low Reynolds number model will be used and the mesh guidelines should be the same as for the enhanced wall treatment (i.e. y + < 4 to 5). When the high Reynolds version is used, the guidelines should be the same as for the standard wall functions (i.e. 30 < y + < 200).

2.4

Conclusion

In this last section a classication will be made of the dierent discussed turbulence models. The model, which is best suitable for the internal engine simulations with respect to the mesh requirements and performance, will be chosen. This classication is not based on actual simulation results. These will be discussed in the next two chapters. The Spalart-Allmaras model The rst model is the Spalart-Allmaras turbulence model. The advantage of this model is that it is

2.4. CONCLUSION

32

relative simple and does not cost much extra computational eort. Also the grid requirements are no obstacle. A coarser mesh with y + = 30 is possible in the simulations. But this counts for little compared to the disadvantages. The models performance in complex ows is uncertain and it is not capable to rapidly accommodate changes in length scale which is the case in the engine simulations. These disadvantages make the Spalart-Allmaras model not suitable for the engine simulations.

The k - models The advantages of the standard k - turbulence model is its general applicability, robustness and economy. Also the grid requirements (30 < y + < 200) will not give any problems. A possible disadvantage is that the model is semi-empirical. The equation for the dissipation is derived using physical reasoning. The simulations in the following chapters may or may not found this. A last remark is that the ow must be fully turbulent and therefore the standard k - model is not appropriate for wall-bounded ows. A family member of the standard k - model is the RNG variant. This model has the same grid requirements. The dierence can be found in the calculation of the viscosity. For low Reynolds numbers the viscosity is calculated in a dierent way which make this model suitable for low-Reynolds numbers and near-wall ows. Because the engine simulation will be a high Reynolds number ow, this model is probably not more appropriate than its standard counterpart. The last member of the k - family is the Realizable k - model. This model is very suitable for ows involving strong streamline curvature, vortices and rotation. The mesh guidelines are the same as the previous two models and do not form an obstacle. The turbulent viscosity in this model is anisotropic, this in contrast to the standard k - model. These properties will make this model probably the best choice in the k - family. The k - model This last turbulence model is suitable for low Reynolds number ows as well as the prediction of free shear ow spreading rates. Therefore this model is applicable to wall-bounded and free shear ows. These properties make this model also very appropriate for the engine simulations. Because this model is available in both high (30 < y + < 200) and low (y + < 4 to 5) Reynolds variants, the mesh guidelines will not give any problems.

The dierent turbulence models are briey judged in the previous. From the pros and cons it can be seen that the Realizable k - model and the k - model are best suitable for the internal engine simulations. This conclusion will be founded by two simulations, namely the backward-facing step (wall-bounded ow) and the free air jet (free shear ow). This will be treated in the next two chapters.

Chapter 3

Backward-facing step
In the previous chapter the dierent turbulence models are discussed. A classication has been made to check which model, on the basis of the theory, is best suitable for the internal combustion engine simulations. In this chapter this conclusion will be founded on the basis of the backward-facing step (BFS) experiment. This experiment is an example of a wall-bounded ow. It involves reattachment of separated turbulent shear layers and is an established benchmark in Computational Fluid Dynamics (CFD). What has the BFS experiment to do with the engine simulations? With this experiment the behavior of the turbulence models near walls can be validated in general. Moreover, during the intake stroke the air ows through the intake duct via the runners and valves in the combustion chamber. This is roughly comparable with the geometry of the BFS experiment and therefore a representative benchmark for the simulation. The rst section of this chapter deals with the theory of the BFS experiment. The geometry and the corresponding parameters will be treated. The second part contains the performance of each model family in the BFS experiment. The results will be compared with the results found by Driver and Seegmiller [5]. After this, the gird dependency of the dierent near-wall treatments under each model family will be compared. The nal part of this chapter contains the conclusion of which turbulence model and which near-wall treatment is best suitable for the internal combustion engine simulations according to the BFS experiment.

3.1

Theory

Before the discussion about the performance of the dierent turbulent models, the backward-facing step experiment in general will be discussed. Driver and Seegmiller did research into this experiment and obtained data. In the following section the turbulence models will be compared with these data which include the reattachment point and the skin friction coecient. For an accurate comparison the exact geometry and parameters, used by Driver and Seegmiller, will be employed. For an extensive description about the experiment the reader is referred to [5]. The geometry of the BFS experiment consists of two parallel walls with a widening (gure 3.1). The separated ow is generated as the uid passes over the backward-facing step. The step height h is 1.27 cm and the inlet width 10.16 cm, which result in an expansion ratio of 1.125. When the uid passes over the backward-facing step, a recirculation zone appears. The point where the streamline touches the bottom wall is called the reattachment point xr . It is expressed in terms of the step height, in other words the absolute x-coordinate divided by the step height. This point is dened as the point where the skin friction coecient Cf is zero. Driver and Seegmiller found a value of 6.26 for the dimensionless reattachment length. A laser interferometer skin-friction apparatus was used to provide the skin-friction measurements along the step-side wall. The equation for the skin friction coecient is 33

3.1. THEORY

34

One-seventh-power law velocity profile

2
Corner Eddy Recirculation zone Reattachment point

y x
h
xr

Figure 3.1: Geometry of the BFS experiment

given by Cf = w , (1/2)u2 ref (3.1)

where w is the shear stress on the bottom wall and uref is the centreline velocity. In order to compare the numerical solution with the experimental data a set of boundary conditions need to be provided which are as close as possible to the conditions in the original test case. For the inlet velocity prole the so-called one-seventh-power law is used. It is an approximate prole for a fully-developed turbulent ow. The equation for this prole is given by u(y ) = uref y
1 /7

(3.2)

where is one half of the channel height and 0 < y < . The centreline velocity uref at the inlet is set to 44.2 m/s. The Reynolds number is, with respect to the step height, Re = 37, 400. It is important to specify proles for the turbulence quantities k , and at the inlet. Otherwise, the turbulence quantities will not be in equilibrium and the simulation results will be severely over-predicted which slow down the convergence. The proles can be found by running a two dimensional simulation of a turbulent channel ow with periodic boundary conditions at both ends. From the simulation the distribution of the dierent turbulence quantities can be determined and used as inlet boundary conditions in the BFS simulations.
2.5

Dimensionless wall distance y+

1.5

0.5

10

15

20 x/h

25

30

35

40

Figure 3.2: Dimensionless wall distance y + as a function of the horizontal coordinate x/h

Chapter 3. Backward-facing step

35

A last remark before the turbulence models will be discussed, is about the dimensionless wall distance y + (equation 2.46). Because this parameter is dependent on the velocity, it is not constant along the wall. If the dimensionless wall distance is plotted against the dimensionless position x/h, a graph like gure 3.2 will appear. From the text it will be clear which value is meant.

3.2

Performance of turbulence models

The dierent turbulence models will be validated on the basis of the backward-facing step experiment. Every model will be checked in view of the reattachment point and friction coecient. First the laminar case will be treated. Secondly the Spalart-Allmaras will be dealt with. After that, the k - family will pass in review and last but not least the k - turbulence model. At the end of this section the dierent models are compared with each other and the best performing model or models in the BFS experiment will be chosen. For a proper comparison the parameters in the dierent simulations will be held the same and are similar to the ones used by Driver and Seegmiller. All the simulations are performed with a second-order accurate scheme. The segregated implicit solver is used to obtain a steady solution. For the pressurevelocity coupling the SIMPLEC algorithm is used. The uid properties like and are held constant. This for better comparison with the data from Driver and Seegmiller.

3.2.1

Laminar

this is a low-Reynolds number turbulence model. It is designed to be applied throughout the boundary layer provided that the near-wall mesh resolution is sucient. However, the boundary conditions have been modied so that the model will work on coarser or high-Reynolds meshes. Therefore both the ne Contours of1) Stream Function (kg/s)(y + 30) will be treated to validate the model. Mar 14, 2006 and the coarser variant mesh (y + The results of the skin friction coecient Cf as a function of x/h for both mesh variants are displayed in gure 3.4. The mesh with 88200 cells is equally divided. The mesh with 48000 cells on the other hand,

4.21e+00 The rst case that will briey be treated, is the case when no turbulence model is activated. The ow 4.05e+00 will be treated as a laminar ow. This simulation will dier substantially from the data found by Driver 3.88e+00 and Seegmiller 3.71e+00because the resolution of the grid is much larger than the Kolmogorov -scales. This can be seen from gure 3.3. The streamlines show separation at more than one location down of the step. The 3.54e+00 simulation shows unsteady features which are not desired. Therefore this solution method is far from 3.37e+00 suitable for the engine simulation. This is the reason that the laminar case wil not be treated in the 3.20e+00 3.03e+00 remainder of this thesis. 2.87e+00 2.70e+00 2.53e+00 2.36e+00 2.19e+00 2.02e+00 1.85e+00 1.69e+00 1.52e+00 1.35e+00 1.18e+00 Figure 3.3: Streamlines of BFS simulation in the laminar case 1.01e+00 8.43e-01 6.74e-01 5.06e-01 3.37e-01 3.2.2 Spalart-Allmaras model 1.69e-01 The rst turbulence model that will be treated, is the Spalart-Allmaras model. As mentioned before, 0.00e+00

FLUENT 6.1 (2d, segregated, lam)

3.2. PERFORMANCE OF TURBULENCE MODELS

36

is provided with boundary layers which means that the mesh resolution closer to the wall is rened. An example of both grids is displayed in gure 3.5.
SpalartAllmaras turbulence model 2.5 2 Skin friction coefficient C (x1000) 1.5 1 0.5 0 0.5 Driver & Seegmiller 1 1.5 88200 y+ = 2.5 35 48000 BL y = 1 7 0 5 10 15 20 x/h 25 30 35 40
+

Figure 3.4: Skin friction coecient Cf as a function of horizontal coordinate x/h with the Spalart-Allmaras
model. The point x/h = 0 is the location of the step

Figure 3.5: Equally divided mesh (left) and boundary layer mesh (right) The skin friction coecient is underestimated in the low-Reynolds case in both the recirculation zone and further downstream. The low-Reynolds model predicts the reattachment point closer to the experiments, xr = 5.81 versus xr = 5.41 for the high-Reynolds variant. A noticeable dierence between the two models is that the friction coecient just downstream of the step is of opposite sign. The experiments show a positive coecient which indicates the presence of a secondary circulation zone also known as a corner eddy. In gure 3.6 the stream-lines of the low-Reynolds model are displayed. The secondary recirculation region is clearly visible.

Figure 3.6: Stream-lines with the low-Reynolds Spalart-Allmaras model It can be concluded that the low-Reynolds Spalart-Allmaras variant performs better in the BFS simulations. This model variant solves the viscous aected layer instead of using semi-empirical formulas and

0.00e+00

Chapter 3. Backward-facing step

37

therefore gives results closer to the experiments from Driver and Seegmiller. The reattachment length diers 7.2 percent from the value of the experiments. The high-Reynolds model on the other hand, diers 13.6 percent and does not show the second circulation zone. The friction coecient up- and downstream of the reattachment point is under predicted by both models. Among other things, this is caused by the inability to rapidly accommodate changes in length scales.

3.2.3

Standard k - model

The standard k - turbulence model is the rst two-equation model that will be discussed. Again it is available in a low and high-Reynolds variant. Originally the standard k - model is a high-Reynolds number model which is only valid for fully turbulent ows. In this case the standard and non-equilibrium wall functions can be used and an y + > 30 is required. The low-Reynolds variant uses the enhanced wall functions. Because this type of near-wall treatment solves the viscous aected layer, a mesh resolution y + in the order of one is required. A higher value is acceptable as long as it is inside the viscous sublayer. Figure 3.7 displays the results for the standard k - model for the various near-wall mesh methods. The low-Reynolds variant makes use of a boundary layer (BL) to resolve the viscous aected region. The standard and the non-equilibrium wall functions use equally divided meshes.
Standard k turbulence model 2.5 2 Skin friction coefficient Cf (x1000) 1.5 1 0.5 0 0.5 1 1.5 Measurements Driver & Seegmiller 88200 standard y+ = 4 38 88200 nonequilibrium y+ = 3 38 48000 BL enhanced y+ = 1 9 0 5 10 15 20 x/h 25 30 35 40

Figure 3.7: Skin friction coecient with standard k - model for dierent near-wall modeling approaches The reattachment point is best predicted by the enhanced wall treatment, namely xr = 5.67. It diers 9.4 percent with respect to the experiments. The standard and non-equilibrium wall functions give values for xr of 5.05 and 5.47 respectively. The enhanced variant is the only model which shows the second recirculation zone. This can again be attributed to the fact that it is the only near-wall treatment which solves the viscous aected region. Upstream of the reattachment point Cf is under predicted by all models. Downstream of xr the high-Reynolds variants give similar results. Only the low-Reynolds model underestimates the friction coecient more, but this dierence becomes marginal further downstream. In the engine simulations a grid with boundary layers is impossible. It costs too much computational eort to solve the complete viscous aected region. In case of a deforming grid, remeshing is necessary and this is not possible with boundary layers. Moreover a deforming grid is only applicable when it is unstructured and boundary layers are not suitable with this type of grid. Because of the above reasons the near-wall meshes, which make use of boundary layers, are not usable in the transient internal engine simulations. Therefore, only both high-Reynolds models will be further discussed. The dierences between the two high-Reynolds models are marginal downstream of

3.2. PERFORMANCE OF TURBULENCE MODELS

38

the reattachment point. Upstream, the non-equilibrium wall functions predict the friction coecient slightly better. It is dicult to say with complete certainty which model performs better. Therefore the performance of both models with dierent grid resolutions will be considered to see the grid dependency. In gure 3.8(a) the results of the standard wall functions for dierent number of grid cells are displayed. Downstream of the reattachment point the results are nearly similar. The dierences in the prediction of the reattachment point increase with a decreasing number of cells. With a grid resolution of y + = 100 the dierence becomes substantial. Upstream, the simulation with the nest mesh resolution is closest to the experimental data. When using the standard k - model in combination with the standard wall function for the internal engine simulations, a mesh resolution close to y + = 30 is desired for the most accurate results.
Standard k model with standard wall functions 2.5 2 Skin friction coefficient Cf (x1000) 1.5 1 0.5 0 Driver & Seegmiller 0.5 1 1.5 9800 y+ = 100 25000 y+ = 60 39200 y+ = 50 88200 y+ = 30 0 5 10 15 20 x/h 25 30 35 40

Standard k model with nonequilibrium wall functions 2.5 2 Skin friction coefficient C (x1000) 1.5 1 0.5 0 0.5 1 1.5 2 Driver & Seegmiller 9800 nonequilibrium y+ = 110 25000 nonequilibrium y+ = 70 39200 nonequilibrium y = 55 88200 nonequilibrium y = 35 0 5 10 15 20 x/h 25 30 35 40
+ +

(a) Standard wall functions

(b) Non-equilibrium wall functions

Figure 3.8: Inuence grid resolution on skin friction coecient with the standard k - model using dierent wall
functions

Next the non-equilibrium wall functions will be treated. In gure 3.8(b) the results are displayed. The inuence of the grid resolution on this kind of wall function is larger than the standard variant. The friction coecient is severely under-predicted upstream of the reattachment point, although the reattachment point itself is, compared with the standard wall functions, closer to the experimental data. Downstream of xr the dierences become less severe but the inuence of the grid resolution can still clearly be seen. From the above observations it can be concluded that in case of the standard k - model, the standard wall functions are best suitable for the engine simulations. The dierences with the non-equilibrium variant in predicting the reattachment point are small. The friction coecient on the other hand is, especially on coarser meshes, better predicted with the standard wall functions.

3.2.4

RNG k - model

As mentioned before, the RNG k - model handles low-Reynolds numbers and near-wall ows better than the standard variant because of a dierent calculation of the turbulent viscosity. In gure 3.9 the simulation results are presented. The dierences between the models downstream of the reattachment point are, just like the previous model, small. Only the enhanced wall treatment shows a dierence after the reattachment point. Upstream of xr the dierences are somewhat larger. The RNG k - model in combination with the enhanced wall treatment predicts the ow upstream of xr closest to the experimental data. This combination also shows the second recirculation zone in contrast to the other two. The reattachment point itself is best predicted by the non-equilibrium wall function. A value of 6.16 is found which results in a dierence of 1.6 percent with the experiments. For the enhanced wall treatment a value of 6.51 is found which

Chapter 3. Backward-facing step

39

RNG k turbulence model 2.5 2 Skin friction coefficient Cf (x1000) 1.5 1 0.5 0 0.5 1 1.5 Measurements Driver & Seegmiller 88200 standard y+ = 3 37 88200 nonequilibrium y+ = 3 37 48000 BL enhanced y+ = 1 8 0 5 10 15 20 x/h 25 30 35 40

Figure 3.9: Skin friction coecient with RNG k - model for dierent near-wall modeling approaches

deviates 3.8 percent. The standard variant performs least accurate. A value for xr of 5.92 is found. In this case the dierence with the experimental data is 5.4 percent. Although this variant predicts the ow downstream of the reattachment point slightly better. Again because of the grid requirements of the engine simulation, only the two high-Reynolds variants are usable. Again the standard wall functions and the non-equilibrium variant are nearly similar. The standard variant is slightly better in predicting the friction coecient downstream of the reattachment point. The non-equilibrium variant is on the other hand better in predicting the reattachment point and the ow upstream. Because the marginal dierences both functions will be tested with dierent grid resolutions. In gure 3.10(a) the results for the standard wall functions with dierent numbers of grid cells is presented. The grid dependence downstream of the reattachment point is negligible. Upstream, the dierences are marginal, although with a mesh resolution of y + = 100 the dierences become just visible. A mesh resolution with an y + > 100 should therefore be avoided. One last remark is that the RNG k - model with standard wall functions is less sensitive to coarse grids than the standard k - variant, which makes this model more reliable.

The results of the grid dependency in case of the non-equilibrium wall function, are presented in gure 3.10(b). The dierences between the various mesh resolutions is less than with the standard k - model with non-equilibrium wall functions. The prediction of the reattachment point is nearly independent of the grid resolution. Downstream of the reattachment point the inuence of the dierent grids is larger than with standard wall functions. Upstream of xr the dierences are small, only the y + = 100 mesh diers substantial. Therefore, in this case a mesh closer to y + = 30 should be constructed for accurate results. Which wall function can best be used for the engine simulation? The friction coecient is predicted nearly similar by both high-Reynolds wall functions. Just as with the previous model, the grid dependence with the non-equilibrium wall function is larger than the standard variant. The reattachment point is better predicted by the rst one. But again the grid dependency is, because the complex geometry and changing grid resolution, the deciding part. So, also in case of the RNG k - model the standard wall functions can best be used.

3.2. PERFORMANCE OF TURBULENCE MODELS

40

RNG k model with standard wall functions 2.5 2 Skin friction coefficient Cf (x1000) 1.5 1 0.5 0 Driver & Seegmiller 0.5 1 1.5 9800 y+ = 10 110 25000 y+ = 8 70 39200 y+ = 5 55 88200 y+ = 3 37 0 5 10 15 20 x/h 25 30 35 40

RNG k model with nonequilibrium wall functions 2.5 2 Skin friction coefficient C (x1000) 1.5 1 0.5 0 0.5 1 1.5 2 Driver & Seegmiller 9800 nonequilibrium y+ = 105 25000 nonequilibrium y+ = 70 39200 nonequilibrium y+ = 55 88200 nonequilibrium y+ = 35 0 5 10 15 20 x/h 25 30 35 40

(a) Standard wall functions

(b) Non-equilibrium wall functions

Figure 3.10: Inuence grid resolution on skin friction coecient with the RNG k - model using dierent wall
functions

3.2.5

Realizable k - model

The last member of the k - family that will be discussed, is the Realizable k - model. This model has again an other formulation for the turbulent viscosity than both k - models. The viscosity is anisotropic and therefore this model is suitable for ows with strong streamline curvature and rotation.

Realizable k turbulence model 2.5 2 Skin friction coefficient Cf (x1000) 1.5 1 0.5 0 0.5 1 1.5 Measurements Driver & Seegmiller 88200 standard y+ = 2 38 88200 nonequilibrium y+ = 2 38 48000 BL enhanced y+ = 1 8 0 5 10 15 20 x/h 25 30 35 40

Figure 3.11: Skin friction coecient with Realizable k - model

In gure 3.11 the results of the BFS simulations with the Realizable k - model are displayed for the dierent near-wall modeling approaches. The non-equilibrium wall function predict the reattachment point most accurate compared to the experimental data, closely followed by the enhanced wall functions. The rst gives a value of 6.22 for xr which deviates less than one percent from the experiments. The second model gives a value of 6.61 which results in an error of 5.6 percent. The enhanced wall treatment predicts the friction coecient upstream of xr closest to the experimental data and shows the second recirculation zone. The non-equilibrium wall functions also predict Cf accurate, but do not display the second recirculation zone. Both high-Reynolds wall functions predict Cf somewhat downstream of xr better than the low-Reynolds variant but, just like the previous models, this dierence becomes marginal further downstream.

Chapter 3. Backward-facing step

41

Because the grid requirements the low-Reynolds model is unsuitable. To see the inuence of the grid resolution on the simulation results of the high-Reynolds Realizable k - models, the friction coecient for the dierent grid sizes is plotted for both the standard (gure 3.12(a)) and non-equilibrium wall functions (gure 3.12(b)).
Realizable k model with standard wall functions 2.5 2 Skin friction coefficient Cf (x1000)

Realizable k model with nonequilibrium wall functions 2.5 2 Skin friction coefficient C (x1000) 1.5 1 0.5 0 Driver & Seegmiller 0.5 1 1.5 9800 standard y+ = 110 25000 nonequilibrium y+ = 70 39200 nonequilibrium y+ = 55 88200 nonequilibrium y+ = 35
40

1.5 1 0.5 0 Driver & Seegmiller 0.5 1 1.5 9800 standard y = 100 25000 standard y+ = 60 39200 standard y+ = 50 88200 standard y+ = 30 0 5 10 15 20 x/h 25 30 35
+

10

15

20 x/h

25

30

35

40

(a) Standard wall functions

(b) Non-equilibrium wall functions

Figure 3.12: Inuence grid resolution on skin friction coecient with the Realizable k - model using dierent
wall functions

The inuence of the grid resolution on the standard wall functions is minimal. Upstream as well as downstream the dierences are negligible. The non-equilibrium variant on the other hand, shows larger dierences. Especially with the coarser mesh the friction coecient is considerably underestimated upstream of xr . The prediction of the reattachment point is more accurate compared with the standard wall functions. Of all k - models the Realizable variant is in combination with both wall functions, the least dependent of the grid resolution and therefore the most reliable turbulence model of the k - family. It is dicult to conclude which wall function performs better. The standard variant is the least sensitive to dierent grid resolutions. The non-equilibrium variant on the other hand, is better in predicting the reattachment point. Downstream of xr the standard wall functions predict the friction coecient slightly better. Upstream of xr the non-equilibrium variant predicts Cf closer to the experimental data but the dependency of grid resolution is substantial. Therefore, despite the under-prediction of the reattachment point, the standard wall function is the best choice for the engine simulations because the independency of the grid resolution.

3.2.6

k - model

The k - turbulence model is available in two variants. Both the standard and the shear-stress transport (SST) model will be treated. Each model is available is a low and high-Reynolds variant which will both be discussed. In contrast to the k - family, the choice between the high or low-Reynolds model is not made by the used near-wall treatment, but by enabling or disabling the Transitional Flow option in Fluent which is a low-Reynolds correction factor. In gure 3.13 the skin friction coecient for both models is displayed. First the two high-Reynolds models will be treated. Both the standard and the SST variant predict the reattachment point very precise. The rst one predicts xr exact, that is 6.26, and the second model gives a value of 6.16 which deviates 1.6 percent from the experiments. Upstream of the reattachment point both k - models give nearly the same results. Just like all high-Reynolds variants they underestimate Cf just after the step. Downstream of the reattachment point the dierences increase. Both models under predict the skin friction coecient. The standard k - model under predicts Cf by 25 percent and

3.2. PERFORMANCE OF TURBULENCE MODELS

42

Standard and SST k turbulence model 2.5 2 Skin friction coefficient Cf (x1000) 1.5 1 0.5 0 Measurements Driver & Seegmiller 0.5 1 1.5 88200 standard y+ = 3 38 48000 BL transitional y+ = 1 9 88200 SST y+ = 2 35 48000 BL transitional SST y+ = 1 9 0 5 10 15 20 x/h 25 30 35 40

Figure 3.13: Skin friction coecient with the standard and SST k - model

the dierence with SST variant is 10 percent larger. The two low-Reynolds models dier somewhat more. The standard k - model over predicts the reattachment point by 19.8 percent, namely 7.5. The SST variant is more accurate in predicting xr . A value of 6.76 is found which results in a dierence of 8.0 percent. The friction coecient upstream of the reattachment point is also better predicted by this model. Downstream of xr the standard k - model is closer to the experimental data. It diers 12.5 percent against 22.5 percent for the SST model. As mentioned earlier, meshes with boundary layers are not applicable in simulations with moving and deforming geometries. Therefore the low-Reynolds models are once more not suitable. Of the high-Reynolds turbulence models, the standard k - model is performing better downstream of the reattachment point. Upstream the dierences between the two models is negligible. For this reason the standard variant is the best usable turbulence model of the k - family for the engine simulations.

Standard k model 2.5 2 Skin friction coefficient Cf (x1000) 1.5 1 0.5 0 Measurements Driver & Seegmiller 0.5 1 1.5 9800 standard y+ = 100 25000 standard y+ = 60 39200 standard y+ = 50 88200 standard y+ = 30 0 5 10 15 20 x/h 25 30 35 40

Figure 3.14: Skin friction coecient with the k - model

Chapter 3. Backward-facing step

43

In gure 3.14 the simulation results with dierent grid resolutions for the standard k - model are displayed. The dierences are marginal. The mesh with 88200 cells performs best. It predicts the reattachment point exact and is slightly better in predicting the friction coecient further downstream. Again it can be seen that a coarser grid gives less accurate results compared to a ner mesh. Therefore a too coarse mesh should be avoided. In case of the grid with 9800 cells, the dierences become substantial. A grid resolution closer to the lower bound (y + = 30) is therefore most desirable.

3.3

Model comparison

This section deals with the comparison of the dierent models. The grid quality is of great importance for the turbulence models. The models which solve the viscous aected region, perform better compared with the models which use semi-empirical wall functions. It is impossible, especially in the unsteady in-cylinder simulation, to construct a mesh which has a ne enough resolution. Moreover, the unsteady simulations use a dynamic and unstructured mesh which has to be remeshed. It is impossible to construct a mesh of this type with boundary layers that can be properly remeshed. Another issue is the computational time that is required. A ne mesh results in more grid cells which costs more time to compute. Especially in the unsteady simulations this will result in a high computational eort. Because the problems with a ne mesh in the unsteady in-cylinder simulations, the coarser variants will only be compared with each other. For the k - family the standard wall functions will be compared and for the k - model the high-Reynolds standard model. Because the large dierences with the experimental data, the Spalart-Allmaras model will not be treated. The only one-equation turbulence model is not able to rapidly accommodate changes in length scale which will be the case in the engine simulation. In gure 3.15(a) the skin friction coecient is plotted for the dierent turbulence models. The used mesh is the same for all models, namely 88200 equally divided cells with y + 30. In all cases the standard wall functions are applied. Downstream of the reattachment point the standard k - model performs best in predicting the friction coecient but the reattachment point itself is considerably underestimated compared to the other models. The standard k - is the worst predictor of the downstream friction coecient. Both the RNG and Realizable k - model perform similar.
Turbulence models with y+ = 35 2.5 2 Skin friction coefficient C (x1000) 1.5 1 0.5 0 0.5 1 1.5 Driver & Seegmiller Standard k RNG k Realizable k Standard k 0 5 10 15 20 x/h 25 30 35

Turbulence models with y+ = 30 2.5 2 Skin friction coefficient Cf (x1000) 1.5 1 0.5 0 0.5 1 1.5 Driver & Seegmiller 88200 standard k 88200 RNG k 88200 Realizable k 88200 standard k 0 5 10 15 20 x/h 25 30 35 40

(a) Structured grid (88.200 cells)

(b) Unstructured grid (85.000 cells)

Figure 3.15: Performance of dierent turbulence models on dierent grid types Upstream of the reattachment point the standard k - model under-predicts Cf the most. The Realizable model is closest to the experimental data of Driver and Seegmiller although the dierences with the RNG variant are small. The prediction of the reattachment point by the k - model is closest to the experiments. The RNG and Realizable k - model dier slightly more but are still in close agreement with the experiments.

3.4. CONCLUSION

44

In gure 3.15(b) the performance of the dierent turbulence models with an unstructured mesh is depicted. As mentioned, the in-cylinder modeling approach requires a partly unstructured grid. Therefore this mesh type will also be discussed. The results with the k - model in case of an unstructured mesh dier substantially from the results with the structured mesh (gure 3.15(a)). The friction coecient downstream of the reattachment point is severely under predicted and the reattachment point itself severely over predicted. The dierences between the grid types in case of the k - models are negligible. It is dicult to say, on the basis of the skin friction coecient, which model is best suitable for the in-cylinder simulations. The standard k - model can be dropped out because the strong dependency of the grid resolution on the results and the severe under-prediction of the reattachment point. The k - model predicts the reattachment point closest to the value found by Driver and Seegmiller but the grid resolution dependency is a disadvantage. Also the dierences between the structured and unstructured mesh variant make this turbulence model less suitable. Therefore this model is also dropped out. The diculty will be the choice between the two remaining turbulence models. The RNG and Realizable k - model perform downstream of the reattachment point nearly similar. Also the prediction of the reattachment point is the same. Upstream of xr it is dicult to say which model performs better. The Realizable variant is closest to the experiments just after the step, but the RNG model predicts the maximum friction coecient in the recirculation zone better. To still make a choice, the velocity proles at dierent positions along the bottom wall will be observed. In gure 3.16 the velocity proles for the RNG and Realizable k - model are displayed. The dierences between the two models are minimal. In the recirculation zone a variance of a few percent is just visible. The dierences with the experiments of Driver and Seegmiller are somewhat larger. After the step both turbulence models under predict the velocity. From the velocity proles, the choice between the RNG and Realizable k - model can not be made. Another experiment has to be performed to nd the possible dierences and to choose the best model. There is one remark. The velocity proles for both turbulence models are scaled. This because the mass ow of the simulations using the one-seventh-power law deviated by a small amount compared to the experiments. The centreline velocity in case of the simulations is increased by ve percent to have the same mass ow rate. This had little eect on the skin friction coecient and the reattachment point and therefore has no inuence on the comparison between the dierent models.

3.4

Conclusion

The dierent turbulence models are validated by means of the backward-facing step experiment carried out by Driver and Seegmiller. This experiment is an example of a wall bounded ow and therefore the behavior of the models close to the wall can be tested, which is crucial for the quality of the turbulence models. The low-Reynolds turbulence models outperform the high-Reynolds variants. This is not surprising because the low-Reynolds models solve the viscous aected region all the way down to the walls which will give more accurate results. This type also shows the second recirculation zone just after the step in contrast to the high-Reynolds variants. The high-Reynolds models make use of wall functions which describe the ow in the viscous aected region instead of solving it. Despite the better performance the low-Reynolds models can not be used in the engine simulations because of the grid requirements. Therefore the comparison consists only of the high-Reynolds variants. As discussed in the previous section, the choice of the best performing model in the BFS experiment is not completely clear. In case of the standard k - and k - model, the dependency of the grid resolution is substantial. This in contrast to the RNG and Realizable k - model. Because the complex geometry of the internal combustion engine, the dimensionless wall distance y + can not be guaranteed

Chapter 3. Backward-facing step

45

1.5
RNG k -
Realizable k - Driver & Seegmiller

0.5
y/h

-0.5

-1

0
0 0.5

2
1

6
x/h

10

12

14

U / U ref

Figure 3.16: Velocity proles for the dierent turbulence models

everywhere. A model which is less dependent of this parameter will give better and more reliable results. The Spalart-Allmaras is not incorporated in further studies because its poor performance on this aspect. Two models remain, namely the RNG and Realizable k - model. The dierences are very small and on the basic of the BFS simulations it can not be concluded which model is better. Also the ow proles, which are displayed in the previous section, can not give an answer to that question. The transport equation for the dissipation rate in case of the Realizable model, is derived from an exact equation. The model also satises certain mathematical constraints on the Reynolds stresses consistent with the physics of turbulent ows. This in contrast to the RNG turbulence model. Therefore the Realizable k - model is better mathematically founded. Also the turbulent viscosity is in contrast to the RNG model anisotropic. This describes the ow in reality more natural. The above arguments point the Realizable k - model out as the better model. But this in not based on actual simulation results. In the following chapter the dierent models will be validated on the basis of the free air jet. This is an example of a free shear ow. The behavior of the dierent models, when no wall-shear is involved, can be validated. This may give decisive information about the better performing turbulence model. By depicting the velocity proles it turned out that the mass ow of the simulations and experiments did not match. The centreline velocities are equal but the velocity proles dier which results in an unequal mass ow. To compensate for this error the centreline velocity of the one-seventh-power law, which is used in the simulations, is raised by 5 percent. The increase in the centreline velocity also has inuence on the skin-friction coecient Cf . The results for Cf presented in the preceding are somewhat overestimated, in the order of 10 percent. This is the case with all turbulence models and will therefore not have any inuence on the discussion of the best performing turbulence model.

3.4. CONCLUSION

46

Chapter 4

Air jet
The fourth chapter of this report is called the air jet. The axisymmetric jet represents a benchmark for the research into the physics of turbulent uid ows. The jet is an example of a free shear ow. In contrast to the backward-facing step simulations, this experiment can validate the dierent turbulence models when no wall-shear is involved. That is the case when the air enters the combustion chamber. The ow changes from a wall bounded ow in the narrow opening between valve and vale seat to a free shear ow in the combustion chamber. Among others Hussein et al. [12] performed velocity measurements in a high-Reynolds number air jet. The results from the simulations will be validated with these data. The rst section of this chapter describes the theory involving the air jet. The parameters of interest and the geometrical set-up will be discussed. Secondly the performance of the dierent turbulence models in the air jet simulations will be dealt with. The results of the dierent model families will be compared with the experiments, which involves the velocity proles, velocity decay rate and virtual origin. After that the models are compared with one another and a conclusion will be drawn on which model outperforms the others.

4.1

Theory

In this section the theory of the air jet will be discussed. Hussein et al. [12] investigated the velocity proles and turbulent intensities of a self-preserving turbulent jet. The turbulent jet ow results from a velocity distribution with a top-hat prole exhausting into a large room (see gure 4.1). The jet can be divided into four zones. The core zone, where the centreline velocity is equal to the orice outlet velocity, is the rst one (red zone in gure 4.1). The second zone is called the transition zone. In this region the centreline velocity starts to decrease. In the third zone the transverse velocity proles at dierent distances are similar. In the last zone, called the termination zone, the centreline velocities rapidly decrease. The diameter of the orice is 2.54 cm and the ow rate is 0.0287 m3 /s. The mean velocity at the orice is 56.2 m/s. The Reynolds number based on the exit conditions is Re = 95, 500. In gure 4.2 the geometry of the air jet is depicted. The measurements are taken with two dierent measurement techniques. The rst one is the stationary hot-wire anemometer (SHW). This technique uses the change in wire resistance to determine the ow velocity. As the uid velocity increases, the rate of heat ow from the heated wire to the ow stream increases. The current which is needed to maintain a constant wire temperature and thus a constant resistance, is a measure for the velocity. The second measurement apparatus is the burst-mode laser-Doppler anemometer (LDA). This technique is based on the Doppler eect. As a seeding particle entrained in a uid passes through the intersection of two laser beams where a fringe pattern is presented. The scattered light received from the particles uctuates in intensity due to this fringe pattern. The frequency of this uctuation is equivalent to the fringe spacing divided by the velocity of the particle perpendicular to the fringe pattern. For more detail about the two measurement techniques, the reader is referred to [12].

47

4.1. THEORY

48

6.48e+01 6.16e+01 5.83e+01 5.51e+01 5.19e+01 4.86e+01 4.54e+01 4.21e+01 3.89e+01 3.57e+01 3.24e+01 2.92e+01 2.59e+01 2.27e+01 1.94e+01 1.62e+01 1.30e+01 9.72e+00 6.48e+00 3.24e+00 Y 0.00e+00

Z X

Contours of Velocity Magnitude (m/s)

Sep 01, 2006 FLUENT 6.2 (3d, segregated, ske)

Figure 4.1: Velocity magnitude contours of air jet with the standard k - turbulence model

x0
D


x

Uc

U0

Figure 4.2: Geometry of the air jet simulations

To compare the simulation results with the results found by Hussein et al., several parameters have to be explained. Two signicant parameters are the velocity decay rate and the virtual origin of the jet. For a self-preserving jet, the centreline velocity Uc is given by [12]
0. 5 Uc = BM0 /(x x0 ) ,

(4.1)

where is B the centreline velocity decay rate constant, M0 the momentum ux at the orice and x0 the virtual origin of the jet. The last parameter will be discussed later in this section. For a top-hat velocity prole, the momentum ux at the orice is given by M0 = 1 U 2 D2 , 4 0 (4.2)

where U0 is the mean velocity at the orice and D the diameter of the orice. Combining 4.1 and 4.2 results in the centreline velocity variation which is given by 1 x x0 U0 = , Uc Bu D D (4.3)

where Bu = 0.5 0.5 B . From the simulations the centreline velocity is known. When a line is tted through these points, the virtual origin x0 and the centreline velocity decay rate B can be determined from equation 4.3 and compared with the results from Hussein et al.. They found a value of x0 = 4.0 D for the virtual origin and a value of B = 6.5 for the decay rate. The results of each turbulence model with respect to this parameter will be treated in the next section.

Chapter 4. Air jet

49

The origin of the jet is located at the orice, that is x = 0 (gure 4.2). Because the jet issues from a source with nite area instead of an idealized point source, a correction is needed. This correction relates the position of the actual area source to a point source. The point source is located at the virtual origin x0 . There are several ways to compute the virtual origin which are described by Kaye and Hunt [13]. One way to determine the virtual origin is by the line dened by the jet envelope. This is simple to determine from a photo or an animation by drawing a line along the edges of the turbulent region. From the total angle enclosed by the two lines, the virtual origin can be estimated with x0 1 = , D 2 tan( 1 2 ) (4.4)

where is the total angle enclosed by the edges of the turbulent region. This can be validated by the experimental results. When the centreline velocity is plotted as a function of the axial location (equation 4.3) a line will appear which represents the centreline velocity decay. The virtual origin is the distance between the origin and the x-intercept of the velocity decay line. These graphs will be discussed in the following sections where the dierent turbulence models will be treated. A last remark before the actual simulations get a chance, is whether the ow satises the momentum integral. The presence of walls at a nite radius causes the entrainment to be fed by a reverse ow outside the jet. The return ow steals momentum from the actual jet and thereby progressively modies it from the jet which would be observed in an innite environment. Hussein et al. investigated this eect of connement on the momentum conservation and formulated a momentum constraint which is given by M 16 = 1+ M0 B 2 x D
2

A0 AR

(4.5)

where A0 is the area of the orice and AR the cross sectional area of the enclosure parallel to the outlet. With equation 4.5 the room size required to return a given level of momentum in the jet, can be estimated. The decay rate could vary per turbulence model and as a consequence change the momentum constraint. To get an indication of the dimensions of the enclosure, the experiment of Hussein et al. will be used. They found a decay constant of B = 6.5 and all measurements were taken at x/D = 70. For a value of M/M0 = 0.99 the experiments should closely resemble a jet in an innite environment according to this criteria [12]. This results in a cross sectional area of the enclosure of 23.9 m2 .

4.2

Performance of turbulence models

The turbulence models are already validated in case of a wall bounded ow. Next, the same models will be validated when no walls are involved, the so called free shear ow. In the following section the three dierent turbulence model families will be discussed and compared with each other. In the previous chapter the laminar simulation showed unrealistic ow properties. The ow will be treated as a laminar ow which is not the case. Because the ow is fully turbulent this results is major errors due to the fact that the mesh resolution is not sucient to resolve the smallest turbulence scales. This is also the case in the free air jet simulations and therefore the case will not be treated. The parameters for the dierent simulations will be held constant, this for better comparison. The segregated implicit solver is used with the SIMPLEC algorithm as pressure-velocity coupling. The geometry of the nal engine simulations consists of both structured and unstructured grids. Therefore both types will be used with every turbulence model. The structured grid consists out of 765,000 cells and the unstructured variant out of 395,000 cells (see gure 4.3). The dierence in the number of cells can be attributed to the fact that a structured grid is much more dicult to construct. To prevent the creation of poor cells, the number of cells have to be increased. The cross sectional area of the enclosure is 23.76 m2 which satises the momentum constraint (equation 4.5). The uid properties like and are held constant. Because the Mach number M a < 0.3 the ow can be considered incompressible [15]. For

4.2. PERFORMANCE OF TURBULENCE MODELS

50

the inlet the mass-ow-inlet boundary condition is chosen. With a ow rate of 0.0287 m3 /s and a constant density of 1.25 kg/m3 this results in a mass ow of 0.0359 kg/s. For the outlet the pressure-outlet boundary condition is chosen. At rst the rst-order schemes will be used to speed up the convergence of the simulations and reduce the computational costs.

(a) Structured mesh (395,000 cells)

(b) Unstructured mesh (765,000 cells)

Figure 4.3: Cross section of mesh parallel to the centerline of the air jet. Orange line indicates x/D = 70 and
blue line x/D = 45

4.2.1

Spalart-Allmaras model

The Spalart-Allmaras turbulence model is the only one-equation turbulence model available in Fluent. As already mentioned, it cannot rapidly accommodate changes in the length scales. This is the case when the uid passes through the orice and enters the enclosure. Therefore the performance of this model is uncertain and the results will likely dier from the experiments. In spite of this the Spalart-Allmaras model will be treated rst. First the virtual origin of the jet must be determined to create the velocity proles. To achieve this, the velocity will be plotted as a function of the axial location. Both scales will be normalized, the velocity with the exit velocity U0 and the axial coordinate with the exit diameter D. Using the graph and equation 4.3 the virtual origin can be determined. In gure 4.4(a) the results for the Spalart-Allmaras model with both the structured and the unstructured grid are displayed together with the experimental results by Hussein et al.. Both grid types deviate substantially from the experimental results found by Hussein et al.. The velocity decay constant for the unstructured grid is B = 4.5 and the virtual origin is x0 = 4.13 D. Hussein et al. found values of B = 6.5 and x0 = 4.0 D. Remarkable is the dierence in sign for the virtual origin. The simulation results for the structured mesh show a positive value for the virtual origin, that is x0 = 10.0 D. Just after the origin of the jet in the core zone, the results match the experimental results better but the decay constant decreases too rapidly. The disparities between the experiments and the simulations can also be seen back in the axial velocity proles in gure 4.4(b). All velocity proles are taken at a position of 70 diameters downstream from the origin (x/D = 70 in gure 4.3). The mean axial velocity U is normalized with the centreline velocity Uc and is plotted versus the non-dimensional coordinate = r/(x x0 ). The curve ts from the data from the experiments of both the stationary hot-wire anemometer (SHW) and the laser-Doppler anemometer (LDA) are also displayed. Using values of > 0.25 is dangerous as unphysical sign changes can be encountered. Both simulations show a prole which is far too wide due to the underestimated decay constant. A possible cause is the inability of the Spalart-Allmaras turbulence model to rapidly

Chapter 4. Air jet

51

SpalartAllmaras turbulence model 35 SpalartAllmaras unstr. SpalartAllmaras str. Curve fit unstr. Curve fit str. Hussein et al.
1.2

SpalartAllmaras turbulence model LDA Data SHW Data SpalartAllmaras unstr. SpalartAllmaras str.

30

25

0.8

U0 /Uc

15

U/Uc
0 20 40 60 80 100 120

20

0.6

0.4

10

0.2

0.2

0.05

0.1

0.15

0.2

0.25

x/D

= r/(x x0 )

(a) Centreline velocity variation

(b) Mean axial velocity proles with LDA/SHW data

Figure 4.4: Air jet simulation results with Spalart-Allmaras model accommodate changes in length scale.

4.2.2

Standard k - model

The second turbulence model that will be discussed is the standard k - model. This is the rst model out of the k - family. In the inlet the ow is aected by the presence of walls. In all simulations the dimensionless wall distance in the inlet was approximately y + 40 60. In this case both the standard and the non-equilibrium wall functions can be used. As mentioned earlier the standard wall functions showed the best results and therefore this type will be used in the following simulations. Just like in the previous case, the virtual origin has to be found rst. In gure 4.5(a) the normalized velocity is plotted as a function of the normalized axial coordinate for the standard k - model with the standard wall function. Again both mesh variants as well as the results found by Hussein et al. are displayed in the gures below.

Standard k turbulence model with standard wall functions 30 Standard k unstr. Standard k str. Curve fit unstr. Curve fit str. Hussein et al.
1.2

Standard k turbulence model with standard wall functions Experiments LDA Experiments SHW Standard k unstr. Standard k str.

25

0.8

20

U0 /Uc

15

U/Uc
0 20 40 60 80 100 120

0.6

0.4

10
0.2

0.2

0.05

0.1

0.15

0.2

0.25

x/D

= r/(x x0 )

(a) Centreline velocity variation

(b) Mean axial velocity proles with LDA/SHW data

Figure 4.5: Air jet simulation results with standard k - model The dierences between the structured and unstructured grids are, just like the Spalart-Allmaras model, considerable. The unstructured grid fails to reach a consistent centreline velocity decay rate. Therefore the curve t exists of two separate lines which intersect at x/D = 50. In case of x/D < 50 the virtual origin is located at x0 = 2.63 D and the velocity decay constant is B = 4.3. The structured mesh on

4.2. PERFORMANCE OF TURBULENCE MODELS

52

the other hand matches the experimental results better. The decay constant is B = 6.5 which is equal to the value found by Hussein. For the virtual origin a value of x0 = 0.67 D is found which deviates considerably from the experimental results (x0 = 4.0 D). It is clearly visible that the core zone of the jet, where the velocity is equal to the exit velocity (U0 /Uc = 1), is too short. The mean velocity proles for the standard k - model for both the structured and unstructured mesh are displayed in gure 4.5(b). The structured mesh variant closely resembles the experimental results. For 0.10 the simulation results match the experimental SHW data. For > 0.10 the dierences between the structured mesh and experiments become clearly visible. The unstructured mesh deviates substantially on the whole domain. Although for values of close to the end of the domain, the simulation results tend to converge to the experimental results.

4.2.3

RNG k - model

The second member of the k - family is the RNG model. It is similar in form to the standard k - model, but includes several renements. Therefore the RNG model should be more accurate and reliable for a wider class of ows than the standard variant [6]. This will be veried in the following. In gure 4.6(a) the location of the virtual origin and the velocity decay rate are displayed. Again the dierences between the two mesh types are considerable. The unstructured variant fails to reach a constant velocity decay rate. When a curve is tted for x/D < 50 the decay constant is B = 4.8 and the virtual origin is located at x0 = 3.25 D. The structured grid underestimates the decay constant considerably. A steeper curve means a lower decay constant (equation 4.3). This can also be seen in gure 4.6(b) where the axial velocity proles are plotted. The prole for the structured mesh is too wide which is caused by the underestimated decay constant. The core zone on the other hand, matches the experimental results by Hussein et al.. This in contrast to the standard k - model where the core zone was underestimated and the decay constant matched the experiments. When a curve is tted through the simulation data of the RNG k - model, a value of x0 = 8.68 D is found for the virtual origin and a centreline velocity decay constant of B = 4.4.

RNG k turbulence model with standard wall functions 30 RNG k unstr. RNG k str. Curve fit unstr. Curve fit str. Hussein et al.
1.2

RNG k turbulence model with standard wall functions Experiments LDA Experiments SHW RNG k unstr. RNG k str.

25

0.8

20

U0 /Uc

15

U/Uc
0 20 40 60 80 100 120

0.6

0.4

10
0.2

0.2

0.05

0.1

0.15

0.2

0.25

x/D

= r/(x x0 )

(a) Centreline velocity variation

(b) Mean velocity proles with LDA/SHW data

Figure 4.6: Air jet simulation results with RNG k - model The velocity proles for the dierent mesh types are depicted in gure 4.6(b). Because the underestimated decay constant in case of the structured mesh, the velocity prole strongly deviates from the experiments. The unstructured mesh is in closer agreement with the experiments. But caution must be taken when using the unstructured simulation results. The velocity decay is, just like the standard k - model, not constant. The twist in the tted curve is larger than the standard k - variant which makes the results less reliable.

Chapter 4. Air jet

53

4.2.4

Realizable k - model

The last member of the k - family is the realizable k - model. The dierent formulation for the turbulent viscosity and the transport equation for the dissipation rate make this model more accurate in predicting spreading rates of both planar and round jets [6]. In gure 4.7(a) the inverse centreline velocity is plotted which conrms this pronouncement for the structured mesh. The dierences between the meshes are, just like the other k - models, substantial. The unstructured mesh is not able to nd a constant decay rate. The twist in the tted curve can be found at x/D = 48. For values lower than this value the centreline velocity decay constant is B = 4.8 and the virtual origin is located at x0 = 4.35 D. The structured variant on the other hand matches the experimental results very precise. The virtual origin as well as the decay constant are very close to the results from Hussein et al.. For the virtual origin a value of x0 = 3.75 D is found and for the centreline velocity decay constant a value of B = 6.6.
Realizable k turbulence model with standard wall functions 25 Realizable k unstr. Realizable k str. Curve fit unstr. Hussein et al.
1.2 Experiments LDA Experiments SHW Realizable k unstr. Realizable k str. Realizable k turbulence model with standard wall functions

20

0.8

15

U0 /Uc

U/Uc
10 5 0 0 20 40 60 80 100 120

0.6

0.4

0.2

0.2

0.05

0.1

0.15

0.2

0.25

x/D

= r/(x x0 )

(a) Centreline velocity variation

(b) Mean axial velocity proles with LDA/SHW data

Figure 4.7: Air jet simulation results with Realizable k - model The dierences in the mean velocity proles between the dierent mesh variants are, compared to the other k - models, smaller (gure 4.7(b)). The results from the structured mesh match the data from the stationary hot-wire anemometer (SHW) measuring method. The unstructured variant matches the experimental results reasonably well, but again caution must be taken when interpreting the data. The velocity decay is not constant which make the data unreliable.

4.2.5

k - model

Last but not least, the k - turbulence model will be treated. This model should, just like the Realizable k - model, predict the shear ow spreading rates of round and radial jets in close agreement with the measurements [6]. Therefore the k - model is applicable to free shear ows. The shear-stress transport (SST) k - model, which is a combination of the k - model in the near-wall region and the standard k - model in the far eld, will also be treated. In gure 4.8 the simulation results for the standard k - model are displayed. Also in this case the dierences between the two grid types are substantial. The unstructured mesh fails to reach a constant velocity decay rate (gure 4.8(a)). The rst part of the curve t (x/D < 45) results in a decay constant of B = 3.7 and a virtual origin of x0 = 3.27 D. For x/D > 45 a value of B = 5.6 is found for the velocity decay constant which better matches the experimental results of Hussein et al.. The decay constant in case of the structured mesh is overestimated. A value of B = 7.8 is found in contrast to B = 6.5 for the experiments. For the virtual origin in case of the structured grid a value of x0 = 4.1 D is found.

4.3. KINETIC ENERGY

54

Standard k turbulence model 30 Standard k unstr. Standard k str. Curve fit unstr. Curve fit str. Hussein et al.
1.2

Standard k turbulence model Experiments LDA Experiments SHW Standard k unstr. Standard k str.

25

0.8

20

U0 /Uc

15

U/Uc
0 20 40 60 80 100 120

0.6

0.4

10
0.2

0.2

0.05

0.1

0.15

0.2

0.25

x/D

= r/(x x0 )

(a) Centreline velocity variation

(b) Mean axial velocity proles with LDA/SHW data

Figure 4.8: Air jet simulation results with standard k - model


SST k turbulence model 30 SST k unstr. SST k str. Curve fit unstr. Curve fit str. Hussein et al.
1.2 Experiments LDA Experiments SHW SST k unstr. SST k str. SST k turbulence model

25

0.8

20

U0 /Uc

15

U/Uc
0 20 40 60 80 100 120

0.6

0.4

10
0.2

0.2

0.05

0.1

0.15

0.2

0.25

x/D

= r/(x x0 )

(a) Centreline velocity variation

(b) Mean axial velocity proles with LDA/SHW data

Figure 4.9: Air jet simulation results with SST k - model In gure 4.8(b) the velocity proles for both mesh types are plotted. The overrated decay constant in case of the structured mesh, results in a slightly more narrow velocity prole than the experiments. The unstructured variant on the other hand underestimates the decay constant to a greater extent which results in a velocity prole which is too wide. In the next section this phenomenon will be treated further and a possible solution will be found. The unstructured mesh variant of the shear stress transport (SST) k - model performs nearly the same as its standard counterpart (gure 4.9). For the velocity decay constant a value of B = 3.9 is found for x/D < 45 and a value of B = 5.3 for x/D > 45. The location of the virtual origin for x/D < 45 is found at x0 = 3.48 D. The velocity prole is slightly wider than the standard counterpart, but the difference is marginal. For the structured mesh the disparity is somewhat larger. Instead of overestimating the decay constant in case of the standard k - model, the structured SST k - model underestimates this constant. A value of B = 5.7 is found for the decay constant and the virtual origin is located at x0 = 2.28 D.

4.3

Kinetic energy

What causes the dierent turbulence models to deviate from the experimental ow results? An attempt will be made to answer this question by looking at the turbulent kinetic energy k of the ow. This is one of the most important parameters when modeling turbulence. It accounts for the mixing of species

Chapter 4. Air jet

55

and drives the decay rate in case of the air jet. In a turbulent ow the kinetic energy withdrawn from the main ow is dissipated into turbulence and nally into heat by viscosity. Hussein et al. [12] measured the three components of the turbulent kinetic energy, namely the axial u2 , radial v 2 and azimuthal component w2 . Together with the turbulent shear stress uv , these four parameters form the Reynolds stresses. The axisymmetry of the ow requires the radial and azimuthal component to be equal. The simulation results for the turbulent kinetic energy will be compared and validated with these data. There is only one remark. In Fluent it is not possible to select the individual Reynolds stress components when using RANS turbulence models. Instead the average kinetic energy k = 1 2 ui ui is calculated, which is an average of these Reynolds stress components. The Spalart-Allmaras turbulent model will be omitted in the comparison of the turbulent kinetic energy. This one-equation model solves a modeled transport equation for the kinetic eddy viscosity. This transport variable is not comparable with the turbulent kinetic energy measured by Hussein et al.. Therefore a direct comparison is not possible and the Spalart-Allmaras model will not be treated any further in this section. Large dierences exist between the structured and unstructured mesh. Therefore the two variants will be treated separately. In gure 4.10 the results for the turbulent kinetic energy are displayed for the structured mesh for both the k - and the k - model families. The kinetic energy k , which is normalized 2 with the square of centreline velocity Uc , is plotted versus the non-dimensional radial coordinate .
0.18 0.16 0.14 0.12
2 k/Uc
2 k/Uc

0.18

Experiments LDA Experiments SHW Standard k str. Realizable k str. RNG k str.

0.16 0.14 0.12 0.1 0.08 0.06 0.04 0.02 0

Experiments LDA Experiments SHW Standard k str. SST k str.

0.1 0.08 0.06 0.04 0.02 0

0.05

0.1

0.15

0.2

0.05

0.1

0.15

0.2

= r/(x x0 )

= r/(x x0 )

(a) k - model family

(b) k - model family

Figure 4.10: Turbulent kinetic energy for the structured mesh at 70 D downstream of the orice The Realizable k - model in combination with the structured mesh gives the best results (gure 4.10(a)). The simulation matches the results found with the LDA measuring technique closely. This can also be seen in gure 4.7 where both the centreline velocity variation and the velocity prole match the experimental results in case of the structured mesh. The standard k - model slightly over predicts the kinetic energy in case of the structured mesh. This is the cause that the centreline velocity variation (gure 4.5(a)) is somewhat over predicted and the velocity prole (gure 4.5(b)) is slightly wider. The RNG k model is the worst in predicting the normalized turbulent kinetic energy. It should be noted that this is not purely due to the kinetic energy prediction but also the centreline velocity. The kinetic energy is slightly over predicted and the centreline velocity under predicted. This combination results in the large deviation with the experiments. This is also visible in gure 4.6 where the centreline velocity is far over predicted which results in a wider velocity prole than the experiments. When looking at the results of the k - family (gure 4.10(b)), the standard variant matches the experimental SHW results reasonably. The kinetic energy near the axis of the jet is underestimated. The velocity decay is lower than the experiments which results in a slightly more narrow velocity prole (see gure 4.8(b)). The SST k - model is less accurate in predicting the kinetic energy. It overestimates this parameter considerably. This has, just like in the previous cases, eect on the decay rate and velocity

4.3. KINETIC ENERGY

56

prole. The overestimated kinetic energy results in an overestimated turbulent viscosity t and therefore in a quicker decay rate which for his part, results in a more narrow velocity prole compared to the experiments. When the turbulent kinetic energy is overestimated, the air jet will degenerate quickly. The regions with higher kinetic energy diuse quicker with the surrounding, energy poor, air. As a consequence the jet will form a wider cone and the centreline velocity will be lower (mass conservation). The wider cone naturally results in a wider velocity prole. All turbulence models predict the turbulent kinetic energy in a dierent way. Not only the magnitude (gure 4.10) but also the contour plots of the kinetic energy dier considerably. In gure 4.11 the contour plots for the standard and realizable k - model are displayed. Only the area just after the orice is rendered. The production of the kinetic energy Gk (equation 2.27) is determined by the velocity gradients uj /xi . This gradient has the largest value at the side of the jets core zone. The realizable model (gure 4.11(b)) predicts this in a better way than the standard counterpart (gure 4.11(a)). In Appendix A.1 the contour plots of the turbulent kinetic energy and the contour plots of the velocity just after the orice are presented for all the dierent turbulence models with the structured mesh.
2.05e+02 1.95e+02 1.85e+02 1.74e+02 1.64e+02 1.54e+02 1.44e+02 1.33e+02 1.23e+02 1.13e+02 1.03e+02 9.23e+01 8.21e+01 7.18e+01 6.16e+01 5.13e+01 4.10e+01 3.08e+01 2.05e+01 1.03e+01 Y 8.00e-07

Z X

2.49e+02 2.37e+02 2.24e+02 2.12e+02 1.99e+02 1.87e+02 1.75e+02 1.62e+02 1.50e+02 1.37e+02 1.25e+02 1.12e+02 9.97e+01 8.73e+01 7.48e+01 6.23e+01 4.99e+01 3.74e+01 2.49e+01 1.25e+01 Y 2.28e-07

Z X

Contours of Turbulence Kinetic Energy (k) (m2/s2)

Sep 01, 2006 FLUENT 6.2 (3d, segregated, ske)

Contours of Turbulence Kinetic Energy (k) (m2/s2)

Sep 01, 2006 FLUENT 6.2 (3d, segregated, rke)

(a) Standard k - model

(b) Realizable k - model

Figure 4.11: Contours of the turbulent kinetic energy for the structured mesh just after the orice When the dierent models are used in combination with the unstructured mesh variant, the results deviate strongly from the data found by Hussein et al.. In gure 4.12 the results for the two turbulence families are displayed just like the experimental data. In gure 4.12(a) both the Realizable and the RNG k - model show a turbulent kinetic energy which is far too low compared to the experiments. It looks like the air jet has completely degenerated. The standard variant also underestimates the kinetic energy but in a less severe way. The two k - family members predict the kinetic energy at the axis of the jet reasonable. The SST variant over predicts it somewhat, but the standard model matches the experimental results. At a larger distance from the center of the jet ( > 0.07) the results are less accurate. The cone formed from the boundaries of the high kinetic energy regions is far too wide and caution must be taken when interpreting these data. The large dierences between the two grid types can be attributed to numerical or false diusion. The term false diusion is used because it is not a real phenomenon, yet its eect on the ow is analogous to that of increasing the real diusion coecient. The amount of numerical diusion is minimized when the ow is aligned with the gird. Therefore the results obtained with the structured grid, match the experiments better than the unstructured counterpart. In case of an unstructured grid the ow can never be aligned with the grid and therefore these simulations deviate from the experiments. The numerical diusion is clearly visible when looking at the velocity proles for the dierent turbulence models. Nearly every model in combination with an unstructured grid predicts the prole to wide. In Appendix A.2 the contour plots for the kinetic energy as well as the velocity are depicts for the dierent turbulence models with the unstructured mesh. The plots of the turbulent kinetic energy do not resemble the plots with

Chapter 4. Air jet

57

0.1 0.09 0.08 0.07 0.06


2 k/Uc

0.1

Experiments LDA Experiments SHW Standard k unstr. Realizable k unstr. RNG k unstr.

0.09 0.08 0.07 0.06


2 k/Uc

0.05 0.04 0.03 0.02 0.01 0 0 0.05 0.1 0.15 0.2

0.05 0.04 0.03 0.02 0.01 0 0 Experiments LDA Experiments SHW Standard k unstr. SST k unstr. 0.05 0.1 0.15 0.2

= r/(x x0 )

= r/(x x0 )

(a) k - model family

(b) k - model family

Figure 4.12: Turbulent kinetic energy for the unstructured mesh a structured mesh at all. Both the shapes and the magnitudes of the contours are completely dierent. The kinetic energy for the RNG and Realizable model is severely underestimated which explains the path of gure 4.12(a). Possible solutions to reduce the eects of numerical diusion are grid renement and the use of a second-order discretization scheme [6]. The eects of grid renement will be examined with two turbulence models, namely the Realizable k - model and the standard k - model. The original unstructured grid was made up of 395,000 cells. Four additional grids will be tested with respectively 0.6, 0.9, 1.2 and 1.6 million cells. In gure 4.13 the eects of the dierent grid resolutions on the turbulent kinetic energy are displayed. In case of the Realizable k - model (gure 4.13(a)) the renement has a positive, but marginal effect on the results. When the number of grid cells is quadrupled with respect to the original grid, the kinetic energy increases with roughly the same factor but is still far from the experimental results. The numerical diusion is, despite the grid renement, still large. The contour plots of the kinetic energy also show a slight increase but far too little compared to the structured results. The k - model (gure 4.13(b)) shows a dierent eect on the grid renement. When the number of grid cells is increased, the normalized kinetic energy decreases. This has a positive eect at larger distances from the jet axis, but close to the axis this results in an underestimated turbulent kinetic energy. The k - model performs, in comparison with the standard k - model, better on the unstructured grid although the results should be critically examined. A second solution to reduce the numerical diusion is to use a second-order discretization scheme for the momentum, kinetic energy and dissipation rate equation. Again this possible solution will be examined with two turbulence models with dierent grid resolutions. In gure 4.14 the simulation results for the Realizable k - and the standard k - model are depicted. The eects of the second-order discretization scheme on the results of the Realizable k - model (gure 4.14(a)) are substantial. Using the second-order discretization scheme in combination with a ne grid resolution reduces the numerical diusion to a considerable extent. In this case the simulation results match the experimental results. The hump in the graph of the velocity decay is still visible. One remark must be made. The unstructured grid needs twice as many cells compared to its structured variant to come to the same simulation results. The standard k - model (gure 4.14(b)) shows a dierent grid dependency. The kinetic energy found with the second-order scheme decreases with increasing number of grid cells, just like the rst-order scheme. The dierences between the two discretization schemes is less obvious compared to the Realizable k - model. The centreline value for the kinetic energy is closer to the experimental results. Also the

4.4. CONCLUSION

58

0.1 0.09 0.08 0.07 0.06


2 k/Uc

0.09

Experiments LDA Experiments SHW 395,000 cells 667,000 cells 886,000 cells 1,256,000 cells 1,651,000 cells
2 k/Uc

0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0 Experiments LDA Experiments SHW 395,000 cells 667,000 cells 886,000 cells 1,256,000 cells 1,651,000 cells 0 0.05 0.1 0.15 0.2

0.05 0.04 0.03 0.02 0.01 0 0 0.05 0.1 0.15 0.2

= r/(x x0 )

= r/(x x0 )

(a) Realizable k - model

(b) Standard k - model

Figure 4.13: Inuence of grid resolution on the kinetic energy at 70 D downstream of the orice
0.09 0.08 0.07 0.06
2 k/Uc

0.09
Experiments LDA Experiments SHW 395,000 cells 886,000 cells 1,651,000 cells

0.08 0.07 0.06


2 k/Uc

0.05 0.04 0.03 0.02 0.01 0

0.05 0.04 0.03 0.02 0.01 0 Experiments LDA Experiments SHW 395,000 cells 886,000 cells 1,651,000 cells 0 0.05 0.1 0.15 0.2

0.05

0.1

0.15

0.2

= r/(x x0 )

= r/(x x0 )

(a) Realizable k - model

(b) Standard k - model

Figure 4.14: Results using second-order discretization scheme with dierent grid resolutions at 70 D downstream
of the orice

results further away from the jet axis match the experiments better compared to the rst-order counterpart but the dierences are less severe compared to the k - model. Both turbulence models give in combination with a second-order discretization scheme and a higher grid resolution better results. The Realizable model predicts the kinetic energy closest to the experimental results. Also the grid renement shows a more desirable dependency. When the grid resolution is raised, the numerical diusion decreases and the results will match the experiments better. The results found with the k - model also approach the experiments closer compared to the rst-order scheme. The results found with a ne grid further from the jet axis match the experiments but the centreline kinetic energy is underestimated. Also the grid dependency shows a less reliable dependency.

4.4

Conclusion

In the preceding section the turbulence models have been validated on the basis of the free air jet. This is an example of a free shear ow. The behavior of the dierent models, when no wall-shear is involved, can be validated. The backward-facing step experiment pointed the Realizable k - model out as the best performing model in case of a wall-bounded ow but this was not completely based on simulation results. The air jet simulation should result in a decisive conclusion concerning the best performing turbulence model.

Chapter 4. Air jet

59

First a few general remarks will be made about the dierences between the two mesh types. The dierences between the structured and unstructured mesh are considerable. In case of the unstructured grid the ow is not aligned with the grid which causes numerical diusion. Parameters like the velocity and kinetic energy spread out too quickly due to numerical diusion. This results, especially in case of the air jet, in an inaccurate prediction of the ow eld. The inaccuracy is clearly visible when the results of the two mesh types are compared. Because the numerical diusion, all unstructured mesh solutions fail to reach a constant decay rate. Therefore caution must be taken when interpreting these data. Two solutions may reduce the eect of numerical diusion, namely grid renement and the use of a second-order discretization scheme instead of a rst-order scheme. The eects of these adjustments will be treated later, rst the results on the original grids with the rst-order discretization scheme will be discussed. In tabel 4.1 the performance of the turbulence models are presented. The classication is made according to four parameters, namely the virtual origin x0 , the velocity decay constant B , the velocity prole and last but not least the turbulent kinetic energy prole. Table 4.1: Performance of turbulence models Structured mesh Unstructured mesh x0 B v prof 1 k prof 2 x0 B v prof 1 k prof 2 ---n/a n/a ++ ++ +/+ +/+/+/+/++ -++ ++ ++ ++ +/++ -+ ++ + +/+/+ +/+/+/-

Model: Spalart-Allmaras Standard k RNG k Realizable k Standard k - SST k -


1

Velocity prole

Turbulent kinetic energy prole

The only one-equation model, the Spallart-Allmaras model, performs on both mesh types poor. This model solves a transport equation for the kinetic eddy viscosity and therefore the comparison with the turbulence kinetic energy could not be made. In the backward-facing step experiment, this model also showed a strong deviation with the experiments and therefore this model is with certainty not suitable for the internal combustion engine simulations. Of the k - family the Realizable model performs best. In combination with the structured mesh it outperforms all other models also of the k - family. The combination with the unstructured mesh gives less reliable results but it is still one of the better performing models. The RNG model is in combination with the structured grid the worst performing k - family member. The velocity and kinetic energy proles are over predicted which results in inaccurate results. The performance with the unstructured variant is comparable with the Realizable k - model. But just like all other models the reliability in case of an unstructured grid is subject of discussion because of the numerical diusion and the eect on the parameters. The last member is the standard k - model. It performs above average compared to the other members on an unstructured mesh. The structured mesh gives results which place this model just after the Realizable model. The standard k - model performs average. The structured mesh gives better results compared to the unstructured variant. In [6] it was explicitly stated that the k - model should predict the spreading rates of jets in closer agreement with the measurements compared to other models. It predicts the proles reasonably but several models outperform the k - model. The SST model performs slightly below the standard variant in case of both mesh types. As mentioned the results found with the unstructured mesh variant deviate considerably with the experimental results. This is caused by the phenomenon of numerical diusion. There are two ways to reduce this, namely grid renement and the use of a higher order discretization scheme. The eects of these solutions are tested with the Realizable k - model and the standard k - model.

4.4. CONCLUSION

60

The grid renement alone in case of the k - model had a marginal eect on the results. A quadrupled mesh resolution resulted in a slight increase in the kinetic energy, but still far from the experimental data. The use of a second-order discretization scheme in combination with a higher grid resolution gave results which matched the experimental results. Both the grid renement and the higher order discretization scheme had an adverse eect on the k - model. The value for the kinetic energy at the jet axis decreased and diverted from the experimental results. The values further away from the axis on the other hand, approached the experimental data better, but still far from perfect. The plots of centreline velocity variation using the unstructured mesh showed a hump around x/D 40 50 for almost all turbulence models. Even when the grid resolution was doubled and a second-order discretization scheme was used. It is possible that a grid change causes this. In gure 4.3(b) the blue line indicates the position x/D = 45 and no abrupt grid change is visible along the centreline of the jet axis. No explanation has been found yet to explain this change in the velocity decay constant. Final conclusion: The Realizable k - model is in both the backward-facing step and the air jet experiment the better performing turbulence model. Especially in combination with the structured mesh, this model gives accurate results. A ne unstructured mesh in combination with a second-order discretization scheme also gives results which are in close agreement with the experiments. This model will therefore be used in the stationary and dynamic internal combustion engine simulations.

Chapter 5

Engine simulations
The main goal of this thesis is the simulation of the air ow in an internal combustion engine. In this last chapter this subject will be treated. Often the geometry is complex because moving pistons and valves are involved, which makes it dicult to construct a mesh. In Appendix B the grid set-up used for these so-called dynamic engine simulations, will be explained. Two engine simulation cases have been studied. In the rst case the intake valves are constant open and the cylinder head will be blown through. This case is called the stationary engine simulations and is treated in the rst section of this chapter. The resulting ow eld will be examined and compared with the experimental results. V. Huijnen [11] performed Large-Eddy Simulations (LES) with the FASTEST3D code and R. de Leeuw [18] and E. Doosje [4] performed PIV measurements on the cylinder head. Results from the stationary simulations will be compared with these data. The second case is called the dynamic engine simulations. The gas ow inside a running engine is simulated. After the discussion of the two simulations a conclusion will be drawn on the validity of the simulations.

5.1

Stationary engine simulations

In the previous chapters the dierent turbulence models, available in Fluent, have been validated on the basis of two experiments. The Realizable k - model was the best performing model. In this section the performance of this model in a more complex ow will be discussed and compared with experimental data and other simulation results. Before the comparison the background of the experiment is considered as well as the dierent parameters of interest with special attention to the swirl number.

5.1.1

Simulation description

The rst step in setting up a simulation is to make the geometry suitable for adaptation. The geometry used by the foundry to cast the inlet manifold is available in an IGES format. This .igs le format, used by several software packages to exchange geometries, is imported in the pre-processor Gambit which is part of the CFD package Fluent. The IGES geometry le had to be processed in order to be able to mesh it. The process of making a geometry appropriate for the simulation set-up is treated in Appendix B in more detail. The geometry (gure 5.1) consists of a main inlet manifold, two runners and two inlet valves. A dummy cylinder with a length of 0.5 m is placed on the cylinder head for the visualization of the ow eld. The diameter of the cylinder is 130 mm. The cylinder head which will be used for the simulations, is from a DAF 12.9 liter heavy duty Diesel engine. The unstructured mesh variant is chosen because of the complex geometry. A structured grid is possible but will take a lot more time to construct. Moreover, the unstructured grid in combination with a second-order discretization scheme and an appropriate grid resolution gives similar results in case of the Realizable k - turbulence model. The grid consists of 956,000 tetrahedral cells. After the rst simulation this grid resolution turned out to be sucient for the use of the standard wall functions (30 < y + < 300). For the inlet the mass-ow-inlet boundary condition is used with a value of 204 g/s which is the same 61

5.1. STATIONARY ENGINE SIMULATIONS

62

Figure 5.1: Wireframe view of the geometry of the RS 0.5 cylinder head from the IGES-le (dummy cylinder not
visible)

mass ow as used in [11]. For the outlet a pressure-outlet boundary condition is used. The maximum velocity of 140 m/s is reached between the valve and its seat. This value results in a Mach-number of M a = 0.46 which is too high to consider the ow as incompressible [15]. Therefore the ideal gas law is used to simulate the eects of compressibility. For the momentum equation as well as the turbulent kinetic energy and dissipation rate equations, second-order discretization schemes are used. The eciency of an internal combustion engine depends on the specics of the combustion. The quality of the combustion is directly dependent on the oxygen-fuel mixing characteristic in the cylinder. It is very important that the air-fuel mixture is as homogeneous as possible. The level of homogeneity is directly related to the mean velocity eld and to the turbulence characteristics in the cylinder. The quality of mixing process can be improved by creating a rotational ow. When the ow rotates around the axis of the cylinder it is called a swirling ow. The swirl intensity is characterized by a dimensionless number called the swirl number. The cylinder head used for the simulations is called the RS 0.5. The ow eld which is generated as the air ows through the intake valves, is characterized by a swirl number S = 0.5 at a distance of 1.75 D from the top of the cylinder. This will be validated by means of the simulations. The swirl number can be measured and calculated in dierent ways [1] [3] [8]. By classical denition, the swirl number, which allows the determination of the swirl intensity, is the ratio between the tangential and axial components of momentum [3] (equation 5.1) S= U W rdA M = A , 2 Mx R A U dA (5.1)

where U and W are respectively the axial and tangential components of the velocity and the overline denotes the averaging. In this equation the terms related to the turbulence are neglected which is permitted for many applications [3]. It is extremely dicult to determine the swirl number in an operating engine. Therefore steady ow tests are often used to characterize the swirl and determine the swirl number. A technique which is also used to determine the swirl number of the RS 0.5 cylinder head, makes use of an impuls swirl meter. A honeycomb ow straightener measures the total torque exerted by the swirling ow. This torque equals the ow of angular momentum through the plane coinciding with the ow straightener upstream face.

Chapter 5. Engine simulations

63

A more simple way to calculate the swirl number is to determine the ratio of the circumferential velocity vc and axial velocity va at a surface perpendicular to the cylinder axis (equation 5.2). vc (r, ) va (r, )

S=

(5.2)

This is also used in the engine simulations. At a distance of Z = 1.75 D from the top of the combustion chamber, a cross section is made of the ow eld and the swirl number is calculated at every grid point using equation 5.2. The area-weigthed average is taken which results in a swirl number. For the internal combustion engine simulations using the Realizable k model this resulted in a swirl number of S = 0.52 which is in good agreement with the specications of the cylinder head (S = 0.5). In gure 5.2 the path lines are plotted for the RS 0.5 cylinder head. A mass ow of 204 g/s is used in the simulations. The swirl or rotation of the ow around the axis of the dummy cylinder is clearly visible.

Figure 5.2: Path lines released from


inlet with a mass ow of 204 g/s, colored by velocity magnitude

5.1.2

Simulation results

As mentioned before, the velocity elds at dierent distances from the top of the combustion chamber are collected by de Leeuw [18]. Particle Image Velocimetry (PIV) is used to visualize the vector eld of the ow. The particles (seeding) that are added to the ow, are illuminated with a laser-light sheet. A camera which is placed at right angles to the light sheet, registers the positions of the particles by taking two closely-spaced snapshots in time. Using a correlation technique the velocity eld can be extracted from the sequence of images. For more information about the PIV measurement technique the reader is referred to [18] and [4]. There is one remark with respect to the PIV experiments. In [18] two dierent cylinder heads are used, namely the RS 1.0 and the RS 0.5. It turned out that the mass ow in case of the second head deviated from the original value. In the second case a value of 138 g/s was measured in contrast to 204 g/s in the original case. A possible cause is the contamination of the mass ow meter [18]. After the correction was employed on the simulation, the lower mass ow had no inuence on the pattern of the ow eld, only the velocity magnitude was somewhat smaller. A second way to check the accuracy of the simulation results from Fluent, is by comparing them with the simulation results from [12]. The FASTEST code is used to compute the ow characteristics. A grid of 2,5 million cells is used in combination with the Large-Eddy-Simulation approach to simulate turbulence. The simulation results are compared with the experimental PIV results at two locations in the dummy cylinder, namely Z = 1.25 D and Z = 1.75 D from the top of the cylinder. At these locations a cross section is made of the vector eld which shows the velocity components in the x- and y -direction. In gure 5.3 the results at Z = 1.75 D are depicted for the PIV measurements and the simulations with LES and the Realizable k - turbulence model. The experimental PIV data (gure 5.3(a)) show two circulation zones, a larger one in the fourth quadrant and a smaller one in the second quadrant. The Fluent simulation results (gure 5.3(c)) on the other hand, only show the larger one. The pattern of the vector eld, except for the second circulation zone, corresponds reasonably well with the experiments. The results from the FASTEST code (gure 5.3(b)) dier more compared to the Fluent simulations and the experiments. Only one circulation zone is visible but slightly higher located than in the Fluent simu-

5.1. STATIONARY ENGINE SIMULATIONS

64

lation. The ow pattern in the second and third quadrant diers in a more severe way. The ow in case of the FASTEST code turns counterclockwise in the second quadrant, this in contrast to the clockwise direction of the Fluent simulations. In the third quadrant the ow direction diers 90 degrees when both simulations are compared.

60

60

60

40

40

40

20 y [mm]

20 y [mm]

20 y [mm]

20

20

20

40 10 m/s 60 40 20 0 x [mm] 20 40 60

40 10 m/s 60 40 20 0 x [mm] 20 40 60

40 10 m/s 60 40 20 0 x [mm] 20 40 60

60

60

60

(a) PIV measurements with RS 0.5 (b) LES simulation results with the (c) Simulation results with Realizcylinder head (mass ow of 138 g/s) FASTEST code able k - turbulence model in Fluent

Figure 5.3: Vector eld at Z = 1.75 D from cylinder top with 9 mm valve lift and a mass ow of 204 g/s. The
cylinder head is upside down, the two circles represent the location of the valves.

In gure 5.4 the results at a distance of Z = 1.25 D are plotted for both simulations and the experimental case. There are no PIV measurements available at this distance from the RS 0.5 cylinder head with 9 mm valve lift. Therefore the experimental results with 15 mm valve lift and the RS 1.0 cylinder head are used. In this simulation a mass ow of 160 g/s is used instead of the 204 g/s in the simulations. Caution must be taken when comparing the results because the dierent parameters. The vector elds will dier, but the rough structures should be the same.

60

60

60

40

40

40

20 y [mm]

20 y [mm]

20 y [mm]

20

20

20

40 10 m/s 60 40 20 0 x [mm] 20 40 60

40 10 m/s 60 40 20 0 x [mm] 20 40 60

40 10 m/s 60 40 20 0 x [mm] 20 40 60

60

60

60

(a) PIV measurements with RS 1.0 (b) LES simulation results with the (c) Simulation results with Realizcylinder head, 15 mm valve lift FASTEST code, 9 mm valve lift able k - turbulence model, 9 mm valve lift

Figure 5.4: Vector eld at Z = 1.25 D from cylinder top with a mass ow of 204 g/s The dierences between the experiments and simulations are substantial. Both simulations (gure 5.4(b) and 5.4(c)) show two circulation zones in contrast to the experiment (gure 5.4(a)) which shows only one circulation zone. Also the magnitude of the velocity deviates which can be contributed to lower mass ow in the experiments. The simulations show a velocity eld with much higher velocities. The experiments on the other hand show a more moderate eld. To give a more quantitative comparison between the simulations and the experiments, the spanwise velocity magnitude is depicted along the line indicated in gure 5.4(c) at Z = 0.25 D (gure 5.5(a)) and Z = 1.75 D (gure 5.5(b)). PIV data are only available at Z = 1.75 D which explains the absence of PIV

Chapter 5. Engine simulations

65

results in the rst gure.

40 35 30 25 20 Fluent FASTEST

20 Fluent PIV FASTEST

15

10

v [m/s]

10 5 0 -5

v [m/s]

15

-5

-10

-10 -15 -6 -4 -2 0 r [cm] 2 4 6


-15 -6 -4 -2 0 r [cm] 2 4 6

(a) Results at Z = 0.25 D

(b) Results at Z = 1.75 D

Figure 5.5: Mean spanwise velocity magnitude as a function of the radius r along the line indicated in gure
5.4(c)

The results at Z = 0.25 D of both simulations show some agreement. The shapes of the curves match reasonable, the velocity magnitude on the other hand diers on most of the domain. The results at Z = 1.75 D dier more. The dierence in sign of the velocity magnitude of the simulations in the left part of the gure, is caused by the counter rotating vector elds of both simulations (gure 5.3(b) and 5.3(c)). The PIV results are in between the FASTEST and Fluent results. In the right part of the gure the dierences are smaller. This is because the simulations and the experiment predict the vortex in this region. Both simulations predict the vortex more or less equal, the experiments show a smaller vortex (gure 5.4(a)) which results in a less high peak. In gures 5.6 - 5.8 the z -components of the velocity are compared using contour plots. Only the different simulations will be compared. This because the current PIV set-up is not able to measure velocity components in the z -direction. In gure 5.6 the results at Z = 1.75 D for both simulations are depicted.

60

60

40

40

20
20 y [mm] y [mm] 20

0
1 0

10

10

20
20

20

40

10
20

40

20

60 60 40 20

40
0 x [mm] 20 40 60

60 60 40 20 0 x [mm] 20 40 60

(a) LES simulation results with the FASTEST (b) Simulation results with Realizable k - turcode bulence model

Figure 5.6: Contour plots of z -velocity at Z = 1.75 D from cylinder top with a mass ow of 204 g/s and a valve
lift of 9 mm. Contour values are in [m/s]

The two simulations dier substantially at this distance from the cylinder head. The FASTEST simulation (gure 5.6(a)) shows more variation in the z -velocity compared to the Fluent results (gure 5.6(b)). The Fluent simulation show a more even velocity distribution. The dierences become smaller further

5.1. STATIONARY ENGINE SIMULATIONS

66

upstream of the ow, closer to the cylinder head. Figure 5.7 depicts the contour plots at Z = 1.25 D.

60

60

40

40

40

10
20 y [mm] y [mm] 20

10

40

0
0

0
0

2 0
1 0

10

20

20

20
20
40

10
20
40

40

60 60 40 20

40
0 x [mm] 20 40 60

60 60 40 20

0 x [mm]

20

40

60

(a) LES simulation results with the FASTEST (b) Simulation results with Realizable k - turcode bulence model

Figure 5.7: Contour plots of z -velocity at Z = 1.25 D from cylinder top with a mass ow of 204 g/s and a valve
lift of 9 mm. Contour values are in [m/s]

The pattern of the contours as well as the magnitude of the velocity of the LES simulation (gure 5.7(a)) are in close agreement with the contours of the k - simulation (gure 5.7(b)). The contours of the LES simulation are a bit more curved but other than that the results match. The dierences between the simulations with changing height could be caused by the dierences in solving the near-wall region for the turbulence models. This is will be discussed later.

60

60

40

40

2 0

40

20 y [mm]

20 y [mm]

2 0

0 4
0

20

0 6
40

0 8
60 40 20 0 x [mm]

60 20 40 60

20

40

40 60
60 40 20 0 x [mm] 20 40 60

60

(a) LES simulation results with the FASTEST (b) Simulation results with Realizable k - turcode bulence model

Figure 5.8: Contour plots of z -velocity at Z = 0.25 D from cylinder top with a mass ow of 204 g/s and a valve
lift of 9 mm. Contour values are in [m/s]

In gure 5.8 the contour plots at Z = 0.25 D are displayed. Just like the contour plots at Z = 1.25 D, the FASTEST (gure 5.8(a)) and Fluent (gure 5.8(b)) results match reasonably. Again the contours of the FASTEST code are a bit more curved but are still in close agreement with the results from Fluent. In the previous plots only the cross sections perpendicular to the cylinder axis are considered. In the last comparison the cross section parallel to the cylinder axis will be treated. The plane passes through the centers of both intake valves. The cross sections with the z -velocity contours for both simulations are depict in gure 5.9. Again no PIV data are available in this direction. The overall shapes of the contours match reasonably for both simulations. The jet between the two inlet valves is more intens in the LES simulation. Also the jets between the cylinder wall and the valves are larger. This can partly be contributed to the used method of solving the near-wall region. All RANS

2 0

20

0 4

20

Chapter 5. Engine simulations

67

-40 -4 -6 -60 -6 -40 00 -20 0

0 0

20

-6 0

-2

-60
20
-4 0

0.42

0 -6 -8-4 00 -40

20

40 0 1

-80 -60

0.4

4 0
20

-40

-40

-40

-20

-20 0

0.3

0.28

0.26 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0.11 0.12

(a) LES simulation results with the (b) Simulation results with Realizable k turbulence model FASTEST code

Figure 5.9: z -component velocity at the plane parallel to the cylinder axis in [m/s]

300

900

10

100

30

10

100

0.42

100

500

0.4

700 0 60 500

200

30 300 0

0.38

400

30

0.36

0
30 0

0.34

100

100

0.3

0.28

0.26 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0.11

(a) LES simulation results with the (b) Fluent simulation results with Realizable k - model FASTEST code

Figure 5.10: Turbulent kinetic energy k at the plane parallel to the cylinder axis in [m2 /s2 ]

based turbulence models make use of wall functions to solve the viscous aected wall region (see chapter 2). These functions satisfy the no-slip condition at the walls. Because the limitations of the computational resources the coarse LES approach is used. The course LES approach uses viscous shear to model the near-wall region. The wall in not actually solved, only the viscous shear is considered. The total shear consists not only of viscous shear, shear caused by turbulence plays a larger part but has been omitted in this simulation. The total shear is therefore considerably underestimated [23]. This is one of the major diculties using the LES approach in case of wall-bounded ows and results in higher velocity magnitudes at the walls and more intense jet structures. In gure 5.10 the contour plots for the turbulent kinetic energy are depicted. Just like the velocity contours, the overall shape and magnitude of the kinetic energy contours match reasonably for both simulations. The more intense jet in gure 5.9(a) is also visible in the contour plot for the kinetic energy using the LES model (gure 5.10(a)). The regions of high kinetic energy are somewhat larger compared to the Realizable k - model (gure 5.10(b)). The high levels of turbulence and therefore also the high levels of turbulent kinetic energy are found in regions where the velocity gradients are large. This is the case where the air enters the combustion chamber. The turbulent kinetic energy production Pk (equation 2.19) is dependent on the velocity gradient. The jet-like structures have relative high velocities compared with the air in the combustion chamber which results in high velocity gradients.

100
20 0

0.32

300

300

30 0 20 0

100

0.32

100

-60

0 0

10

0.38

0
10

-60

-40
-40

0.36

0
-20

0.34

100
30 0

0 00 0 0055 30 10 90 0 50

300

100

100

5.2. DYNAMIC ENGINE SIMULATIONS

68

5.2

Dynamic engine simulations

The second simulation class that will be considered, is the dynamic engine simulation. This type of simulation is called dynamic because piston and valves are moving. This needs a special treatment concerning the building of the grid. The building of the grid will be briey discussed in this section, for a more detailed description the reader is referred to Appendix B. The dicultly is to validate the simulation results. Experimental data and other simulations are not available. Nevertheless the ow eld is analyzed at dierent crank angles and the global ow features are, as far as possible, compared with the stationary simulations.

5.2.1

Simulation description

The simulation set-up will be treated rst. In Appendix B.1 the two dimensional case is dealt with. The 2D grid set-up is relative simple compared to the 3D case (Appendix B.2). Importing the geometry in the pre-processor and making it suitable for the simulations is complicated. The basic features of the set-up are comparable for both simulations but the 3D case is much more sophisticated. It is recommended to read Appendix B prior to the following section to get familiar with the general simulation set-up. Importing the geometry of the DAF engine (gure 5.11(a)) into the pre-processor Gambit and making it suitable for the simulation is the rst hurdle to take. Gambit oers dierent strategies which also include several geometry heal options [7]. When a geometry is imported, dierent errors can arise and Gambit can automatically repair these. In Appendix B.2 the way used in the simulations is discussed but there are of course other strategies possible.
Mass flow inlet

Inlet runner Valve stem zone Valve layering zone Combustion chamber

(a) Wireframe view of the geometry

(b) Exploded wireframe view

Figure 5.11: Adopted RS 0.5 cylinder head geometry used for the dynamic engine simulations In the two dimensional case the grid is divided in four dierent mesh zones (gure B.1(a)). The three dimensional simulation uses one extra zone to build the dynamic grid around the valve stem. This zone is called the Valve stem zone. In gure 5.11(b) an exploded view of the dierent mesh zones is depicted. The inlet manifold with the two runners is, because of its complex geometry, build out of an unstructured grid. The inlet itself is modeled as a pressure-inlet. For the rst simulation atmospheric pressure is used. The real engine is turbo charged but for simplicity of the simulation this is not modeled. The two valve stem zones consist of a structured grid. The relative simple geometry of the zone makes this possible. The valve layering zones are also provided with a structured grid for the same reason. The combustion chamber consists of both a structured and unstructured grid. The unstructured mesh is only necessary for the opening and closing of the valves because the grid has to be remeshed to avoid degenerated grid cells. Because a structured grid saves computational resources the combustion cham-

Chapter 5. Engine simulations

69

ber will, at a certain crank angle, be build out of structured grid cells instead of unstructured. This is accomplished by inserting a piston layering zone. In gure 5.12 the grid set-up in the combustion chamber is depicted at dierent crank angles. In Appendix B.2 the dierent zones are discussed in more details. The dierent zones are connected via grid interfaces or sliding meshes. A grid interface is a connection between two mesh surfaces that are not really connected in the geometry. When an interface is not dened between the two surfaces, Fluent treats the intersection plane as a wall and no uid can pass. The use of an interface makes uid ow possible. Another advantage of using an interface is that dierent mesh types can exist next to each other (structured versus unstructured). A sliding interface is an interface between two surfaces which are moving and/or deforming. This is the case between the combustion chamber and the valve layering zone. Both zones consist of a dierent mesh type and the mesh is changing when the valves and piston start to move. A sliding interface makes the mesh change possible. A more detailed description can be found in Appendix B.

(a) 0 degrees CA

(b) 45 degrees CA, the structured piston layering zone is visible just like dividing surface (blue)

Figure 5.12: Mesh illustration of the combustion chamber wall (black) and valve sliding interface (red) at dierent
crank angles

At zero degrees crank angle, the gird is build out of 696,000 cells. At 180 degrees crank angle this is increased to 866,000 cells. The segregated implicit unsteady solver is used in combination with the PISO pressure-velocity coupling. This coupling method is recommended for all transient ow problems and allows a larger time step [6]. For the momentum equation as well as the turbulent kinetic energy and dissipation rate equations, second-order discretization schemes are used. The ideal gas law is used to calculate the eects of compressibility. The Realizable k - turbulence model in combination with the standard wall functions accounts for the turbulence eects. The simulation is performed with a crank shaft speed of 1500 rpm and a piston stroke of 0.15 m. The motion of the valves is linear. The maximum valve lift of 9.0 mm is reached at 90 degrees crank angle.

5.2.2

Simulation results

This section deals with the discussion of the dynamic simulation results. In contrast to the stationary simulations, the dynamic variant can not be compared with experimental results, nor with other simulations. Experiments are performed but with dierent geometries [8]. The dynamic simulation results will be analyzed by looking at the ow features. Also the rough ow structures which where found with the stationary simulations, should return in the dynamic simulations. The simulation will be split up in two parts, namely the intake and compression stroke. Intake stroke In gure 5.13 the so-called path lines are depicted for three dierent crank angles (CA) during the intake stroke. The path lines display the path which is followed by a particle released from a certain point or surface (in this case the valve/combustion chamber interface). The swirling ow is, particularly

5.2. DYNAMIC ENGINE SIMULATIONS

70

1.60e+02 1.52e+02 1.44e+02 1.36e+02 1.28e+02 1.20e+02 1.12e+02 1.04e+02 9.60e+01 8.80e+01 8.00e+01 7.20e+01 6.40e+01 5.60e+01 4.80e+01 4.00e+01 3.20e+01 2.40e+01 1.60e+01 8.00e+00 0.00e+00

(a) 45 degrees crank angle

(b) 90 degrees crank angle

(c) 135 degrees crank angle

Figure 5.13: Path lines inside combustion chamber for dierent crank angles, colored by velocity magnitude [m/s]. Particles are released from surface between combustion chamber and valve layering zone

-60

10
0

0 -4
-20

-10

-1

(a) 45 degrees CA, 4.5 mm valve lift

-20 0

20

(b) 90 degrees CA, 9.0 mm valve lift

-40

10

10

20

10

(c) 135 degrees CA, 4.5 mm valve lift

Figure 5.14: z -contours of velocity during intake stroke at dierent crank angles in [m/s] in gure 5.13(c), clearly visible. The particles which form the path lines are released from the interface between the combustion chamber and the valve layering zone. This to keep the gures at reasonable dimensions. The path lines in the inlet manifold are similar to the ones from the stationary case (gure 5.2).

-20

-60 -20

20

-2 0

5
0
-5

-5
10
-10
-10

-1

-5

-5

(d) 180 degrees CA, valves closed

Chapter 5. Engine simulations

71

60

40

20 y [mm]

20

40 50 m/s 60 40 20 0 x [mm] 20 40 60

60

(a) 45 degrees CA at z1

60

60

40

40

20 y [mm]

20 y [mm]

20

20

40 50 m/s 60 40 20 0 x [mm] 20 40 60

40 50 m/s 60 40 20 0 x [mm] 20 40 60

60

60

(b) 90 degrees CA at z1

(c) 90 degrees CA at z2

60

60

60

40

40

40

20 y [mm]

20 y [mm]

20 y [mm]

20

20

20

40 20 m/s 60 40 20 0 x [mm] 20 40 60

40 20 m/s 60 40 20 0 x [mm] 20 40 60

40 20 m/s 60 40 20 0 x [mm] 20 40 60

60

60

60

(d) 135 degrees CA at z1

(e) 135 degrees CA at z2

(f) 135 degrees CA at z3

60

60

60

40

40

40

20 y [mm]

20 y [mm]

20 y [mm]

20

20

20

40 20 m/s 60 40 20 0 x [mm] 20 40 60

40 20 m/s 60 40 20 0 x [mm] 20 40 60

40 20 m/s 60 40 20 0 x [mm] 20 40 60

60

60

60

(g) 180 degrees CA at z1

(h) 180 degrees CA at z2

(i) 180 degrees CA at z3

Figure 5.15: Vector elds at dierent crank angles during the intake stroke at positions indicated in gure 5.16

To give a more detailed plot of the ow features, the z -contours of the plane parallel to the cylinder axis at dierent crank angles are depicted in gure 5.14. The plane passes through both centers of the valves. The jet, which is formed when the air ows between the two intake valves, is clearly visible. This ow structure is also visible in the stationary simulations (gure 5.9(b)). The two vortices under the valves

5.2. DYNAMIC ENGINE SIMULATIONS

72

are of equal size at 45 degrees crank angle. During the intake stroke the left vortex increases in size due to the shape of the inlet runners. This is also visible in gure 5.8 where the vortices dier in size in the plane perpendicular to the cylinder axis. In gure 5.14(d) both intake valves are completely closed and the jet is extinguished. The global ow features of the stationary and dynamic simulations match reasonably. The jets between and on the anks of the intake valves are predicted in the same way. To visualize the ow features perpendicular to the cylinder axis, cross sections are made at the different crank angles to depict the vector elds. Four dierent crank angles are chosen, namely 45, 90, 135 and 180 degrees crank angle. The cross sections are also made at various distances from the cylinder head as indicated in gure 5.16. It should be notes that the vector scales in the rst two cases are dierent from the other two. In gure 5.15 the cross sections are displayed.

z0
0 / 360 CA

z1
45 / 315 CA

z2
90 / 270 CA

z3
135 / 225 CA

180 CA

Figure 5.16: Measuring position (dashed lines) in combustion chamber for dierent crank angles. The solid lines
indicate the top piston location at the given crank angle

At 45 degrees crank angle the ow eld is nearly symmetric. A total of eight vortical structures can be distinguished. At 90 degrees CA at the same height (gure 5.15(b)) the vector eld shows less large structures. Two larger vortices are just visible. At a larger distance from the cylinder head (gure 5.15(c)) these two vortices have grown in size. This ow structure is also visible in the stationary engine simulations at Z = 1.25 D (gure 5.4(c)). The mass ow in the stationary simulation is set equal to the maximum ow during the intake stroke of the operating engine. This maximum ow is reached at about 90o CA where the piston velocity is maximal. At three-quarter of the inlet stroke (135 degrees) the ow eld just under the valves is a bit chaotic. Several smaller vortices can be seen. Halfway the combustion chamber (gure 5.15(e)) two larger and one smaller vortex are visible. When gure 5.15(c) and 5.15(e) are compared, the development of the ow is recognizable. The vortex in the upper left corner in gure 5.15(c) is more intense and moved to the center in gure 5.15(e). The right vortex at 90 degrees CA is pressed further against the cylinder wall. At the bottom of the combustion chamber (gure 5.15(f)) a large counter clockwise rotating vortex appears. At 180o CA the intake valves are completely closed which is clearly visible at the vector eld in the upper part of the combustion chamber (gure 5.15(g)). A few mildly swirling ow structures can be seen and the magnitude of the velocity is decreased compared to the previous vector elds. This is also the case further downstream of the cylinder. At the bottom higher velocities can be found in a counter clockwise rotating swirl. Compression stroke The intake stroke of the previous section is followed by the compression stroke. In case of this stroke, not only the velocity elds will be considered but also the kinetic energy elds. The kinetic energy is directly associated with turbulence and therefore also the mixing process of air and fuel just after the end of the

Chapter 5. Engine simulations

73

compression stroke. In gure 5.17 the vector elds at dierent crank angles are depicted.

60

60

60

40

40

40

20 y [mm]

20 y [mm]

20 y [mm]

20

20

20

40 10 m/s 60 40 20 0 x [mm] 20 40 60

40 10 m/s 60 40 20 0 x [mm] 20 40 60

40 10 m/s 60 40 20 0 x [mm] 20 40 60

60

60

60

(a) 225 degrees CA at z1

(b) 225 degrees CA at z2

(c) 225 degrees CA at z3

60

60

60

40

40

40

20 y [mm]

20 y [mm]

20 y [mm]

20

20

20

40 10 m/s 60 40 20 0 x [mm] 20 40 60

40 10 m/s 60 40 20 0 x [mm] 20 40 60

40 5 m/s 60 40 20 0 x [mm] 20 40 60

60

60

60

(d) 270 degrees CA at z1

(e) 270 degrees CA at z2

(f) 360 degrees CA with dierent vector scale at z0

Figure 5.17: Vector elds at dierent crank angles during the compression stroke at positions indicated in gure
5.16

60 90o CA 135o CA 50 180 CA


o

20 180o CA 18 16 14 12 215o CA 270 CA 315o CA 360o CA


o

40 v [m/s]
v [m/s]

30

10 8

20

6 4 2

10

6 D [cm]

10

12

6 D [cm]

10

12

(a) Intake stroke

(b) Compression stroke

Figure 5.18: Velocity magnitude along the line indicated in gure 5.4(c) perpendicular to the cylinder axis in
[m/s]. Results at 180 degrees CA is presented in both gures to give a reference

The vector elds are, compared with the intake stroke, more homogeneous. There are a few vortices visible, but of less intensity. A large laevorotatory swirl exists at the beginning of the compression stroke at the bottom of the cylinder (gure 5.17(c)). During the compression cyclus the magnitude of the swirl decreases. This is visible in gure 5.17(f) at the top dead center (TDC) of the piston where the magnitude of the velocity has been halved. The compression stroke itself does not make a contribution to the intensity of the ow. Therefore the velocity magnitude during the compression stroke decreases due to

5.2. DYNAMIC ENGINE SIMULATIONS

74

5 -2.
0

10
7.5

2.5

2. 5

0
2.5

7.5

Figure 5.19: z -contours of velocity during compression stroke at dierent crank angles in [m/s]

friction. This is also proven by Song [22]. To give a quantitative comparison between the degree of rotation, the velocity magnitudes at dierent crank angles along the line indicated in gure 5.4(c) are depicted. In gure 5.18 only the velocities in x and y direction are considered and the vertical position of the line is in all cases halfway the stroke at that specic crank angle. The results are divided into the intake 5.18(a) and compression stroke 5.18(b). The maximum ow velocity is reached around the maximum valve lift (90o CA). After that there is a gradual decrease during the intake stroke and further during the compression stroke. The ow velocities and turbulence generated due to purely the compression stroke are much smaller than those generated during the intake stroke and can therefore be neglected [22]. The velocity magnitudes decrease because of friction. Furthermore the turbulence dissipates during the compression stroke because of the decrease of kinematic viscosity due to the increase of temperature and gas pressure. This is also visible in gure 5.20 where the kinetic energy is plotted at dierent crank angles. It is therefore necessary to have high turbulence intensity at the beginning of the compression stroke for a good mixing process of air and fuel at the end of the compression stroke. To analyse the ow parallel to the cylinder axis the contours of the z -velocity are depicted in gure 5.19. At 225 degrees CA the ow is still inuenced by the ow features of the intake stroke. Halfway the compression stroke the ow is largely driven by the piston movement. Because there is no extra energy input by the compression, the z -velocity contours are attened. This will continue to the top dead center of the piston. As mentioned the combustion process is inuenced by the amount of turbulence and thus inextricably connected to the turbulent kinetic energy k . In gure 5.20 the contours for the turbulent kinetic energy are depicted at dierent crank angles at positions indicated in gure 5.16. The kinetic energy is directly related to the velocity components. It is therefore not surprising that the kinetic energy, as well as the velocity, decreases with increasing crank angle. This decrease of kinetic energy commences very rapidly. At the top dead center (gure 5.20(i)) about ten percent of the maximum kinetic energy is left compared to the bottom dead center (BDC) (gure 5.20(c)). Figure 5.19(b) shows a velocity peak at the left side of the cylinder. This is also visible is kinetic energy plots. The below gures show a higher degree of turbulent kinetic energy is the same area.

2.5
10

(a) 225 degrees CA

(b) 270 degrees CA

Chapter 5. Engine simulations

75

A last parameter of interest is the swirl number S . As already mentioned this number is the ratio of the circumferential velocity and axial velocity at a surface perpendicular to the cylinder axis (equation 5.2). The number is used to dene the degree of rotation in the cylinder. After determining the swirl number at dierent positions and at dierent crank angles, the data proved to be not suitable. The swirl number is dependent on the axial velocity. At the TDC and the BDC the velocity of the piston and, to a smaller extent, the uid itself is close to zero. Because the axial velocity appears in the denominator the swirl number becomes very large and is not representative for the ow. The calculation of the swirl number using equation 5.2 is not suitable for dynamic engine simulations and will therefore not be treated any further.

60 40 20 y [mm]

60
20
30

60

30
40

20

40
50

30
40

40 20 y [mm] 0 20

60
50
80

40

20 y [mm]

80

50

50

40

20 40

50
60

20
60

30

60

100

10

40 60

40 60
40 60

40
60 60 40 20 0 x [mm] 20 40 60

40
60 40 20 0 x [mm]

50
20

60

40

20

0 x [mm]

20

40

60

(a) 180 degrees CA at z1


60

(b) 180 degrees CA at z2


60

(c) 180 degrees CA at z3


60 40

15
40

30
20 y [mm]

25

20

40 20 y [mm]
30
15

30
35

25

y [mm]

40

20

35

45

20

40
25
20
0 20 40 60 60 40 20 0 x [mm] 20 40 60
35

0 20 40 60 60

0 20 40 60

45

40

50
20 0 x [mm] 20 40

50

60

60

40

20

0 x [mm]

20

40

60

(d) 225 degrees CA at z1


60

(e) 225 degrees CA at z2


60

(f) 225 degrees CA at z3


60

10

15
40
15

10

40

20

40

20

20 y [mm]

20

20
25

y [mm]

y [mm]

25

.5 12

20

20

20
10

40

40

40

5
60 60 40 20 0 x [mm] 20 40 60

60 60 40 20 0 x [mm] 20 40 60

60 60 40 20 0 x [mm] 20 40 60

(g) 270 degrees CA at z1

(h) 270 degrees CA at z2

(i) 360 degrees CA at z0

Figure 5.20: Contours of the kinetic energy k at dierent crank angles during the compression stroke at positions
indicated in gure 5.16 in [m2 /s2 ]

5.3

Conclusion

In this chapter the results of the stationary and dynamic engine simulations are discussed. The results of the stationary variant are compared with both experimental and other simulation data. The dynamic

5.3. CONCLUSION

76

simulations on the other hand, are only analyzed because reference work is not available. Stationary simulations The RS 0.5 cylinder head is blown through to get the stationary Fluent simulation data. From the previous chapters the Realizable k - turbulence model turned out to be the best performing model and therefore it is used in these simulations. The simulation data are compared with experimental PIV results from [18] and simulation results from [11] with the FASTEST-3D code. The comparison between the Fluent simulations and the experiments with the RS 0.5 cylinder head is dicult. Most experiments are performed with a dierent cylinder head, namely the RS 1.0. This cylinder head has a higher swirl number and will therefore not give suitable material for comparison. The little experiments with the RS 0.5 cylinder head are performed at one specic height, namely at Z = 1.75 D. At this distance the PIV data show two vortices in contrast to one vortex in the simulations. The pattern of the vector eld, except the missing vortex, corresponds reasonably well with the experiments (gure 5.3). There is a remark concerning the experiments. It turned out that the mass ow meter, which is used to measure the air ow through the cylinder head, had a deviation. This was possibly caused by contamination of the apparatus by seeding particles. This error in combination with the little experimental results make the comparison between the Fluent simulations and the PIV experiments not reliable. Reference material obtained with the FASTEST code is available to a greater extent. Not only the vector elds perpendicular but also parallel to the cylinder are available. The vector eld at Z = 1.75 D shows just like the Fluent simulation, one vortex but the left half of the cross section diers more (gure 5.3). The ow elds at Z = 1.25 D (gure 5.4) match. In both simulation cases two vortices are visible at nearly the same location. The contour plots of the z -velocity dierence in the same way as the vector plots. At Z = 1.75 D the contour plots for both simulations show considerable disparities while at Z = 1.25 D and Z = 0.25 D the contours of the z -velocity match reasonably (gures 5.6, 5.7 and 5.8). The contour plot of the z -velocity parallel to the cylinder axis also shows similarities with the FASTEST code (gures 5.9). The jet between the two inlet valves as well as the two jets between valve and combustion wall are somewhat overestimated by the FASTEST code. This dierence can be ascribed to the dierence in solving the near-wall region. The RANS turbulence modeling approach uses the no-slip condition while the LES method uses viscous shear and ignores the larger turbulent shear. This causes the ow velocity near the walls to be overrated. The last parameter of interest is the swirl number S . The swirl number is calculated at a distance of Z = 1.75 D from the cylinder head. A value of 0.52 is found with the simulations which corresponds with the value specied for the cylinder head, namely 0.5. The way used in this chapter to determine the swirl number, is a good approximation for the reality. Dynamic simulations It is more dicult to draw a conclusion with regards to the dynamic engine simulations because the reference material is thinly sown. The path lines in the combustion chamber of the dynamic (during intake stroke) and stationary simulations show resemblance. The swirling ow around the cylinder axis of the dynamic simulations is, especially at larger crank angles (gure 5.13(c)), clearly visible and comparable with the stationary variant (gure 5.2). Also the z -contours of the velocity in the plane parallel to the cylinder axis (gures 5.9 and 5.14) show resemblance. The jet between both intake valves and between the valves and combustion wall are predicted in both cases. The ow eld perpendicular to the cylinder axis is more dicult to compare because the ow is changing rapidly with changing height. At Z = 1.25 D the stationary simulation shows two counter rotating vortices (gure 5.4). This is also the case at 90o CA at the lower half of the combustion chamber in the dynamic simulations (gure 5.15(c)). The mass ow in the stationary simulations is chosen in a way that it represents the maximum mass ow in a running engine. This maximum mass ow is reached at 90o CA where the piston velocity is maximal. At larger crank angles

Chapter 5. Engine simulations

77

one large vortex arises in the dynamic simulations (gure 5.15(i)). The stationary variant also shows one vortex but with a dierent shape (gure 5.3). During the compression stroke the velocity magnitudes perpendicular (gure 5.17) as well as parallel (gure 5.19) to the cylinder axis decrease quickly. There is little energy input due to purely the compression stroke [22] and therefore the velocities decrease because of friction. The turbulent kinetic energy (gure 5.20) also decreases because of the decrease of kinematic viscosity due to the increase of temperature and gas pressure. To have a high intensity of turbulence at the end of the compression stroke for the mixing of air and fuel, it is therefore necessary to have a high turbulence intensity at the beginning of the compression stroke.

5.3. CONCLUSION

78

Chapter 6

Conclusion and recommendations


The main goal of this research is to model the air ow in an internal combustion engine using the CFD package Fluent. Validation of the dierent RANS turbulence models available in this package is also part of the assignment. The very last section of this thesis deals with the nal conclusion and recommendations concerning these subjects of research. Both goals are treated separately. First the turbulence simulations will be discussed followed by the engine simulations.

6.1

Final conclusion

Modeling validation With the backward-facing step (BFS) experiment the performance of the dierent turbulence models in the near-wall region are validated. Because the limitations of the grid resolution in case of the internal combustion engine, only the high-Reynolds turbulence models are appropriate. These models use wall functions to solve the viscous aected region instead of solving it completely. Coarser grids, especially in the near-wall region are possible. On the basis of the BFS simulation results and the mathematical foundation, the Realizable k - model turns out to be the better performing model. Also the inuence of the grid resolution aects the results marginal which is desirable because the grid resolution in a complex geometry can not be guaranteed everywhere. The validation of the turbulence models in case of a free wall-shear ow is done with the air jet simulation. The performance of the dierent models when no walls are involved, can be veried. Because a three dimensional grid is more dicult to construct both structured and unstructured grid types are used. The dierences between the two grid types are substantial. Numerical diusion occurs when using the unstructured grid. Because of this so called false diusion the prediction of the ow is less accurate. Grid renement in combination with the use of second-order discretization schemes reduces this eect substantially. In case of the structured mesh the Realizable k - model is with a lead the best performing model. Both the virtual origin and the velocity decay constant are predicted very close to the experiments. The unstructured mesh variant performs less compared to its structured counterpart, but it is still one of the better performing models in case of this mesh type. In both simulations the standard wall functions are used to solve the near-wall regions. These semiempirical function bridge the region between the wall and fully turbulent ow. Solving the near-wall region gives more accurate results but requires a very ne grid which is not applicable because the complex geometry. The standard wall functions show good results and little grid dependency.

79

6.2. RECOMMENDATIONS

80

Engine simulations Two types of engine simulations are performed, namely a stationary and a dynamic variant. In case of the rst type the cylinder head of a DAF 12.9 liter heavy duty Diesel engine is blown through to get a visualization of the ow eld. The second variant uses a moving and deforming mesh to simulate the air ow in a running engine using the same cylinder head. The stationary results are compared with experimental results found with the Particle Image Velocimetry (PIV) measuring technique and simulation results found with the FASTEST-3D code. The vector eld perpendicular to the cylinder axis found with the stationary Fluent simulations shows similarities with the PIV results. The major part of the vector directions correspond but the magnitude of the velocity is somewhat larger in the simulations. The dierence in velocity magnitude can be ascribed to the inaccurate mass ow measurements in the experiments. Moreover the experimental data are collected at one specic distance from the cylinder head and therefore a comparison is unreliable. Also the number of measurements is not sucient for a reliable conclusion. The simulation results found with the FASTEST code are compared at dierent locations along the cylinder axis. Closer to the cylinder head the results of both simulations match. Not only the vector eld perpendicular to the axis but also parallel to the axis agree. Further from the cylinder head the results start to deviate. This can partly be ascribed to the dierence in the near-wall solution method. The RANS turbulence modeling approach uses the no-slip condition while the LES method used in the FASTEST simulations uses viscous shear. This causes the ow velocity near the walls to be overrated. A mesh with a moving piston and intake valves is realized to simulate the gas ow in the internal combustion engine. The ow proles perpendicular to the cylinder axis during the intake stroke show resemblance with the stationary simulations at some points. It is dicult to compare the dierent simulations because the proles change rapidly with changing height. Parallel to the cylinder axis a comparison is better possible. The dynamic simulations predict a similar jet structure and strength between the valves and cylinder wall as found in the stationary variant. During the compression stroke the vector eld in both directions become attened. The eect of purely the compression stroke on the ow eld is minimal and as a result the velocity magnitudes and turbulent intensities decrease. It is therefore necessary to have a high turbulence intensity at the beginning of the compression stroke for the mixing process of air and fuel at the end of the compression stroke. The dynamic engine simulations show realistic ow features which are useful for a better understanding of the processes in an internal combustion engine. The simulations can be extended with other subjects such as injection, combustion and turbo charging to obtain a complete model of an engine.

6.2

Recommendations

As a result of this nal project several recommendation are made with regard to further research. Modeling validation To investigate the inuence of the grid resolution on the results of the backward-facing step and air jet simulations, a more thorough grid dependency study in necessary. Especially the grid resolution from which the results will not change any further are of importance. Engine simulations For better comparison of the stationary engine simulations with the PIV experiments, more PIV measurements with the RS 0.5 cylinder head are needed and at dierent locations along the cylinder axis. Also the vector eld parallel to the cylinder axis should be determined experimentally. In the dynamic simulations not the exact geometry of the DAF engine is used. The inlet manifold and valves are exactly modeled but the piston shape, stroke and the length of the connecting rod

Chapter 6. Conclusion and recommendations

81

are arbitrarily chosen. The exact geometry should be retrieved and implemented in the dynamic simulations for an even more realistic representation of the reality. The convergence problems during the compression stroke of the dynamic simulations due to low velocities in the inlet manifold should be corrected. A possibility is to stop the calculations of the ow in the inlet manifold as soon as the valves are completely closed. The calculation of the swirl number S during the dynamic simulation gives unsuitable results. A dierent way of calculating the swirl number should be found to give quantitative results for the degree of rotation in the running engine. The possibilities with the CFD package Fluent are extensive. The simulations can be extended to get a more complete view of the internal combustion engine. Subjects like injection, turbo charging and combustion are points of interest for further research.

6.2. RECOMMENDATIONS

82

Bibliography
[1] Boemer, A. and Artiaga-Hahn, S. 2002. Stationary and transient investigation of swirling In-cylinder ow. Fluent Inc. Technical Note TN-192 [2] Choudhury, D. 1993. Introduction to the renormalization group method and turbulence modeling. Fluent Inc. Technical Memorandum TM-107 [3] Crnojevic, C., Decool, F. and Florent, P. 1999. Swirl measurements in a motor cylinder. Experiments in Fluids, vol. 26, 542-548 [4] Doosje, E. 2001. PIV-karakterisering van de uitgaande swirlstroming bij een stationair doorgeblazen Heavy-uty cilinderkop. Technische Universiteit Eindhoven, Eindhoven, WVM 2001.09 [5] Driver, D.M. and Seegmiller, H.L. 1985. Features of a reattaching turbulent shear layer in divergent channel ow. AIAA Journal vol. 23,NO. 2, NASA AMES Research Center, Moett Field, California [6] Fluent 6.2 Users guide. Fluent Inc. [7] Gambit 2.2 Users guide. Fluent Inc. [8] Heuvel, S.L. van den 1998. In-cylinder ow analysis for production-type internal-combustion engines. Technische Universiteit Eindhoven, Eindhoven [9] Hinze, J.O. 1975. Turbulence. McGraw-Hill, London [10] Hoogendoorn, C.J. and Meer, T.H. van der 1991. Fysische transportverschijnselen II. Delftse Uitgevers Maatschappij, Delft [11] Huijnen, V. 2005. Study of turbulent ow structures of a practical steady engine head ow using Large Eddy Simulations. Technische Universiteit Eindhoven, Eindhoven [12] Hussein, H.J., Capp, S.P. and George, W.K. 1994. Velocity measurements in a high-Reynolds-number, momentum-conserving, axisymmetric, turbulent jet. Journal of Fluid Mechanics, vol. 258, 31-75 [13] Kaye, N.G. and Hunt, G.R. 2001. Virtual origin correction for lazy turbulent plumes. Journal of Fluid Mechanics, vol. 435, 377-396 [14] Kim, S.-E. and Choudhury, D. 1995. A near-wall treatment using wall functions sensitized to pressure gradient. In. ASME FED vol. 217, Separated and Complex Flows. ASME [15] Kuerten, J.G.M. 2002. Modelling of physical phenomena. Technische Universiteit Eindhoven, Eindhoven [16] Launder, B.E. and Spalding, D.B. 1972. Mathematical models of turbulence. Academic Press, London and New York [17] Launder, B.E. and Spalding, D.B. 1974. The numerical computation of turbulent ows. Computer Methods in Applied Mechanics and Engineering, vol. 3, 269-289 [18] Leeuw, R. de 2005. Vergelijking van PIV en LDA bij een stationair doorgeblazen heavy-duty cilinderkop. Technische Universiteit Eindhoven, Eindhoven, WVT 2005.02 83

BIBLIOGRAPHY

84

[19] Morton, B.R., Taylor, Georey and Turner, J.S. 1956. Turbulent gravitational convection from maintained and instantaneous sources. Proceedings of the Royal Society of London, vol. 234, 1-23 [20] Nieuwstadt, F.T.M. 1998. Turbulentie, theorie en toepassingen van turbulente stromingen. Epsilon Uitgaven, Utrecht [21] Shih, T.-H., Liou, W.W., Shabbir, A., Yang, Z. and Zhu, J. 1995. A new k - eddy-viscosity model for high reynolds number turbulent ows - Model development and validation, Computers and Fluids, vol. 24, 227-238, Utrecht [22] Song, Y.S., Hong, J.W. and Lee, J.T. 2000. The turbulence measurement during the intake and compression process for high-turbulence generation around spark timing, Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering, vol. 215, 493-501 [23] Temmerman, L., Had ziabdi c, M., Leschziner, M.A. and Hanjali c, K. 2005. A hybrid two-layer URANS-LES approach for large eddy simulation at high Reynolds numbers, Int. J. Heat and Fluid Flow, vol. 26, 173-190 [24] Wilcox, D.C. 1993. Turbulence modeling for CFD. Grin Printing, Glendale

Appendix A

Contour plots air jet


A.1 Structured mesh

This appendix depicts the contour plots of the turbulent kinetic energy and the velocity magnitude for the dierent turbulence models in case of the structured mesh. With this overview the dierences between the models, especially in the eld of the turbulent kinetic energy, will be showed. Remark: The scales next to the plots dier per turbulence model because of the varying range.
2.05e+02 1.95e+02 1.85e+02 1.74e+02 1.64e+02 1.54e+02 1.44e+02 1.33e+02 1.23e+02 1.13e+02 1.03e+02 9.23e+01 8.21e+01 7.18e+01 6.16e+01 5.13e+01 4.10e+01 3.08e+01 2.05e+01 1.03e+01 Y 8.00e-07

Z X

6.48e+01 6.16e+01 5.83e+01 5.51e+01 5.19e+01 4.86e+01 4.54e+01 4.21e+01 3.89e+01 3.57e+01 3.24e+01 2.92e+01 2.59e+01 2.27e+01 1.94e+01 1.62e+01 1.30e+01 9.72e+00 6.48e+00 3.24e+00 Y 0.00e+00

Z X

Contours of Turbulence Kinetic Energy (k) (m2/s2)

Sep 01, 2006 FLUENT 6.2 (3d, segregated, ske)

Contours of Velocity Magnitude (m/s)

Sep 01, 2006 FLUENT 6.2 (3d, segregated, ske)

(a) Contours turbulent kinetic energy

(b) Contours velocity magnitude

Figure A.1: Contours of turbulent kinetic energy and velocity magnitude with Standard k structured mesh

model using the

2.49e+02 2.37e+02 2.24e+02 2.12e+02 1.99e+02 1.87e+02 1.75e+02 1.62e+02 1.50e+02 1.37e+02 1.25e+02 1.12e+02 9.97e+01 8.73e+01 7.48e+01 6.23e+01 4.99e+01 3.74e+01 2.49e+01 1.25e+01 Y 2.28e-07

Z X

6.08e+01 5.78e+01 5.47e+01 5.17e+01 4.87e+01 4.56e+01 4.26e+01 3.95e+01 3.65e+01 3.34e+01 3.04e+01 2.74e+01 2.43e+01 2.13e+01 1.82e+01 1.52e+01 1.22e+01 9.12e+00 6.08e+00 3.04e+00 Y 0.00e+00

Z X

Contours of Turbulence Kinetic Energy (k) (m2/s2)

Sep 01, 2006 FLUENT 6.2 (3d, segregated, rke)

Contours of Velocity Magnitude (m/s)

Sep 01, 2006 FLUENT 6.2 (3d, segregated, rke)

(a) Contours turbulent kinetic energy

(b) Contours velocity magnitude

Figure A.2: Contours of turbulent kinetic energy and velocity magnitude with Realizable k structured mesh 85

model using the

A.1. STRUCTURED MESH

86

1.30e+02 1.23e+02 1.17e+02 1.10e+02 1.04e+02 9.72e+01 9.08e+01 8.43e+01 7.78e+01 7.13e+01 6.48e+01 5.83e+01 5.19e+01 4.54e+01 3.89e+01 3.24e+01 2.59e+01 1.94e+01 1.30e+01 6.48e+00 Y 1.50e-06

Z X

6.19e+01 5.88e+01 5.57e+01 5.26e+01 4.95e+01 4.64e+01 4.33e+01 4.02e+01 3.71e+01 3.40e+01 3.10e+01 2.79e+01 2.48e+01 2.17e+01 1.86e+01 1.55e+01 1.24e+01 9.29e+00 6.19e+00 3.10e+00 Y 0.00e+00

Z X

Contours of Turbulence Kinetic Energy (k) (m2/s2)

Sep 01, 2006 FLUENT 6.2 (3d, segregated, rngke)

Contours of Velocity Magnitude (m/s)

Sep 01, 2006 FLUENT 6.2 (3d, segregated, rngke)

(a) Contours turbulent kinetic energy

(b) Contours velocity magnitude

Figure A.3: Contours of turbulent kinetic energy and velocity magnitude with RNG k - model using the structured
mesh
5.01e+02 4.76e+02 4.51e+02 4.25e+02 4.00e+02 3.75e+02 3.50e+02 3.25e+02 3.00e+02 2.75e+02 2.50e+02 2.25e+02 2.00e+02 1.75e+02 1.50e+02 1.25e+02 1.00e+02 7.51e+01 5.01e+01 2.50e+01 Y 2.65e-10

Z X

6.12e+01 5.81e+01 5.51e+01 5.20e+01 4.90e+01 4.59e+01 4.28e+01 3.98e+01 3.67e+01 3.37e+01 3.06e+01 2.75e+01 2.45e+01 2.14e+01 1.84e+01 1.53e+01 1.22e+01 9.18e+00 6.12e+00 3.06e+00 Y 0.00e+00

Z X

Contours of Turbulence Kinetic Energy (k) (m2/s2)

Sep 01, 2006 FLUENT 6.2 (3d, segregated, skw)

Contours of Velocity Magnitude (m/s)

Sep 01, 2006 FLUENT 6.2 (3d, segregated, skw)

(a) Contours turbulent kinetic energy

(b) Contours velocity magnitude

Figure A.4: Contours of turbulent kinetic energy and velocity magnitude with Standard k - model using the
structured mesh
2.27e+02 2.15e+02 2.04e+02 1.93e+02 1.81e+02 1.70e+02 1.59e+02 1.47e+02 1.36e+02 1.25e+02 1.13e+02 1.02e+02 9.07e+01 7.93e+01 6.80e+01 5.67e+01 4.53e+01 3.40e+01 2.27e+01 1.13e+01 Y 1.74e-10

Z X

6.09e+01 5.78e+01 5.48e+01 5.17e+01 4.87e+01 4.56e+01 4.26e+01 3.96e+01 3.65e+01 3.35e+01 3.04e+01 2.74e+01 2.43e+01 2.13e+01 1.83e+01 1.52e+01 1.22e+01 9.13e+00 6.09e+00 3.04e+00 Y 0.00e+00

Z X

Contours of Turbulence Kinetic Energy (k) (m2/s2)

Sep 01, 2006 FLUENT 6.2 (3d, segregated, sstkw)

Contours of Velocity Magnitude (m/s)

Sep 01, 2006 FLUENT 6.2 (3d, segregated, sstkw)

(a) Contours turbulent kinetic energy

(b) Contours velocity magnitude

Figure A.5: Contours of turbulent kinetic energy and velocity magnitude with SST k - model using the structured
mesh

The dierences between the velocity magnitude contours are marginal. The core zone of the jet for the dierent models (the red peak) where the velocity is equal to the exit velocity, is dierent in length. This can also be seen in the centreline velocity variation plots in chapter 4. Also the width of the cone, formed by the boundaries of the high velocity region, changes marginal with changing turbulence model. The models which treat the turbulent viscosity anisotropic (RNG and Realizable k - model) form a more

Chapter A. Contour plots air jet

87

narrow cone. This in contrast to the isotropic models. The contours of the turbulent kinetic energy show larger dierences. The Standard k - model as well as the Standard k - model predict production of kinetic energy upstream of the orice. The other models on the other hand, predict a more physical correct prole. The regions at the outer boundary of the core zones have the largest velocity gradients which result, according to equation 2.27, in the largest production of k . The Standard k - model over predicts this high energy region. The Standard k - does not show this region at all.

A.2

Unstructured mesh

The previous section treated the results of structured mesh. This part shows the unstructured simulation results for the dierent models. The plots of the velocity contours do not dier a lot from each other. Only the width of the cone varies with dierent turbulence models. Just like the structured mesh variant, this mesh also shows a more narrow cone in case of anisotropic turbulent viscosity (RNG and Realizable k - model). It is clearly visible that the contour plots for both the kinetic energy and the velocity magnitude are more fancifully than the structured counterparts. Remark: The scales next to the plots dier per turbulence model because of the varying range.

1.11e+02 1.05e+02 9.96e+01 9.40e+01 8.85e+01 8.30e+01 7.74e+01 7.19e+01 6.64e+01 6.08e+01 5.53e+01 4.98e+01 4.43e+01 3.87e+01 3.32e+01 2.77e+01 2.21e+01 1.66e+01 1.11e+01 5.53e+00 Y 1.41e-06

Z X

6.43e+01 6.10e+01 5.78e+01 5.46e+01 5.14e+01 4.82e+01 4.50e+01 4.18e+01 3.86e+01 3.53e+01 3.21e+01 2.89e+01 2.57e+01 2.25e+01 1.93e+01 1.61e+01 1.29e+01 9.64e+00 6.43e+00 3.21e+00 Y 0.00e+00

Z X

Contours of Turbulence Kinetic Energy (k) (m2/s2)

Sep 03, 2006 FLUENT 6.2 (3d, segregated, ske)

Contours of Velocity Magnitude (m/s)

Sep 03, 2006 FLUENT 6.2 (3d, segregated, ske)

(a) Contours turbulent kinetic energy

(b) Contours velocity magnitude

Figure A.6: Contours of turbulent kinetic energy and velocity magnitude with Standard k unstructured mesh

model using the

2.29e+01 2.17e+01 2.06e+01 1.94e+01 1.83e+01 1.71e+01 1.60e+01 1.49e+01 1.37e+01 1.26e+01 1.14e+01 1.03e+01 9.15e+00 8.00e+00 6.86e+00 5.72e+00 4.57e+00 3.43e+00 2.29e+00 1.14e+00 Y 2.40e-07

Z X

6.46e+01 6.14e+01 5.82e+01 5.49e+01 5.17e+01 4.85e+01 4.52e+01 4.20e+01 3.88e+01 3.55e+01 3.23e+01 2.91e+01 2.58e+01 2.26e+01 1.94e+01 1.62e+01 1.29e+01 9.69e+00 6.46e+00 3.23e+00 Y 0.00e+00

Z X

Contours of Turbulence Kinetic Energy (k) (m2/s2)

Sep 03, 2006 FLUENT 6.2 (3d, segregated, rke)

Contours of Velocity Magnitude (m/s)

Sep 03, 2006 FLUENT 6.2 (3d, segregated, rke)

(a) Contours turbulent kinetic energy

(b) Contours velocity magnitude

Figure A.7: Contours of turbulent kinetic energy and velocity magnitude with Realizable k unstructured mesh

model using the

A.2. UNSTRUCTURED MESH

88

6.44e+00 6.12e+00 5.80e+00 5.47e+00 5.15e+00 4.83e+00 4.51e+00 4.19e+00 3.86e+00 3.54e+00 3.22e+00 2.90e+00 2.58e+00 2.25e+00 1.93e+00 1.61e+00 1.29e+00 9.66e-01 6.44e-01 3.22e-01Y 5.89e-07

Z X

6.46e+01 6.14e+01 5.82e+01 5.49e+01 5.17e+01 4.85e+01 4.52e+01 4.20e+01 3.88e+01 3.55e+01 3.23e+01 2.91e+01 2.59e+01 2.26e+01 1.94e+01 1.62e+01 1.29e+01 9.70e+00 6.46e+00 3.23e+00 Y 0.00e+00

Z X

Contours of Turbulence Kinetic Energy (k) (m2/s2)

Sep 03, 2006 FLUENT 6.2 (3d, segregated, rngke)

Contours of Velocity Magnitude (m/s)

Sep 03, 2006 FLUENT 6.2 (3d, segregated, rngke)

(a) Contours turbulent kinetic energy

(b) Contours velocity magnitude

Figure A.8: Contours of turbulent kinetic energy and velocity magnitude with RNG k - model using the unstructured mesh
3.44e+02 3.27e+02 3.10e+02 2.93e+02 2.76e+02 2.58e+02 2.41e+02 2.24e+02 2.07e+02 1.89e+02 1.72e+02 1.55e+02 1.38e+02 1.21e+02 1.03e+02 8.61e+01 6.89e+01 5.17e+01 3.44e+01 1.72e+01 Y 2.02e-07

Z X

6.42e+01 6.10e+01 5.78e+01 5.46e+01 5.14e+01 4.82e+01 4.49e+01 4.17e+01 3.85e+01 3.53e+01 3.21e+01 2.89e+01 2.57e+01 2.25e+01 1.93e+01 1.61e+01 1.28e+01 9.63e+00 6.42e+00 3.21e+00 Y 0.00e+00

Z X

Contours of Turbulence Kinetic Energy (k) (m2/s2)

Sep 03, 2006 FLUENT 6.2 (3d, segregated, skw)

Contours of Velocity Magnitude (m/s)

Sep 03, 2006 FLUENT 6.2 (3d, segregated, skw)

(a) Contours turbulent kinetic energy

(b) Contours velocity magnitude

Figure A.9: Contours of turbulent kinetic energy and velocity magnitude with Standard k - model using the
unstructured mesh
6.34e+01 6.03e+01 5.71e+01 5.39e+01 5.07e+01 4.76e+01 4.44e+01 4.12e+01 3.81e+01 3.49e+01 3.17e+01 2.85e+01 2.54e+01 2.22e+01 1.90e+01 1.59e+01 1.27e+01 9.52e+00 6.34e+00 3.17e+00 Y 1.73e-08

Z X

6.45e+01 6.13e+01 5.81e+01 5.48e+01 5.16e+01 4.84e+01 4.52e+01 4.19e+01 3.87e+01 3.55e+01 3.23e+01 2.90e+01 2.58e+01 2.26e+01 1.94e+01 1.61e+01 1.29e+01 9.68e+00 6.45e+00 3.23e+00 Y 0.00e+00

Z X

Contours of Turbulence Kinetic Energy (k) (m2/s2)

Sep 03, 2006 FLUENT 6.2 (3d, segregated, sstkw)

Contours of Velocity Magnitude (m/s)

Sep 03, 2006 FLUENT 6.2 (3d, segregated, sstkw)

(a) Contours turbulent kinetic energy

(b) Contours velocity magnitude

Figure A.10: Contours of turbulent kinetic energy and velocity magnitude with SST k - model using the unstructured mesh

The plots for the kinetic energy on the other hand, show large dierences. Especially between the models with anisotropic and isotropic turbulent viscosity. The anisotropic models (RNG and Realizable k model) under predict the kinetic energy by a considerable amount (keep in mind the dierent scales). In case of the other models the results found with the unstructured mesh are also completely dierent compared to the structured results. The numerical diusion strongly inuences the results in case of an

Chapter A. Contour plots air jet

89

unstructured mesh. Caution must be taken when interpreting the simulation results using this kind of unstructured grids.

A.2. UNSTRUCTURED MESH

90

Appendix B

In-cylinder model set-up


In this appendix the set-up of an In-cylinder model in the CFD package Fluent will be illustrated. The In-cylinder option in Fluent allows users to dene various parameters that are of importance when setting up a simulation of an internal combustion engine. Two cases will be treated. First a two dimensional model will be discussed and secondly a three dimensional problem. The geometry and mesh of the two simulations are generated in the pre-processor Gambit. In this processor important parameters are set which are necessary for a proper operation of the model in Fluent. It is assumed that the reader is familiar with the interface of both Fluent and Gambit. Only the guidelines will be treated here, not every step in detail. To get familiar the tutorials of both programs come in handy.

B.1

2D model set-up

To get familiar with the basic steps of the In-cylinder model option in Fluent, the two dimensional set-up will rst be treated. It is split up in two parts. First the geometry will be build in the pre-processor Gambit and after that the simulation set-up in Fluent will be treated.

B.1.1

Geometry set-up Gambit

In this example the intake stroke of a 4-stroke internal combustion engine will be modeled. For simplicity only the intake valve is modeled. First the geometry will be made and meshed in Gambit. In the next section the model set up in Fluent will be treated. The two dimensional geometry in Gambit (gure B.1(a)) can be realized by using the standard procedure for making geometries (see Gambit tutorials [7]). There is only one remark. The opening and closing of the valve needs a special procedure in Fluent. The setup requires the use of an (sliding) interface between the valve layer and the combustion chamber. This will be considered later in detail. In 2D an interface consists of four edges. Two edges belong to the combustion chamber zone and the others are part of the valve-layering zone (gure B.1(b)). Notice that these edges are not connected in any way, therefore the vertices are also not shared. The two edges, which are part of the combustion chamber, will together be dened as walls in the Boundary condition menu in Gambit. This also counts for the two edges that are part of the valve-layering zone. Other edges that have to be dened in the Boundary condition menu are the vertical combustion chamber walls, the edges that form the valve and the edge that represents the piston. All edges can be dened as walls. Moving and/or deforming walls will be specied in Fluent. The inlet port of the inlet manifold can be dened as an Inlet-vent. Other possibilities are also conceivable. When the geometry is dened, the meshing of the faces can begin. As shown in gure B.1(a) the two dimensional (2D) geometry consists of four zones, namely inlet, combustion chamber, valve layering zone 91

B.1. 2D MODEL SET-UP

92

(a) Dierent geometrical zones

(b) Interface between valve layer and combustion chamber

Figure B.1: Grid details for the 2D dynamic engine simulations and the piston layering zone. Each zone has dierent requirements with respect to boundary conditions and meshing. There are two major types of zones in Fluent, namely the stationary and the dynamic. In the second type moving objects are involved. These objects deform the grid which have to be remeshed to avoid degenerated grid cells. This requires a typical model approach in Gambit. Inlet: The inlet is an example of a stationary zone. This requires no special treatment when meshing this region. It is recommended to use a structured grid because this method requires fewer cells which speeds up the calculations. However, this option requires more experience from the user with respect to creating grids when more complicated geometries are involved. Combustion chamber: This is a dynamic zone. Because the valve is moving the grid cells will be compressed (gure B.1(b)). As a consequence the geometry has to be remeshed. The remeshing is done in Fluent but the grid has to satisfy several constraints. The remesh option is only possible with an unstructured, triangular mesh. The second requirement for the grid is the cell size. The height of the cells near the valve must be larger than the valve movement per time step. When this requirement is not met, the volume of the cells becomes negative which results in an abortion of the simulation. Valve layer: This is a thin zone between the valve and valve seat and represents the minimum valve lift. It is an example of a rigid dynamic zone. This will be explained later when the model set-up in Fluent is discussed. The thin layer represents the minimum valve lift (in this case 0.5 mm). When the valve opens, this zone expands by use of the layering option and a structured grid is possible. Piston layer: The unstructured mesh in the combustion chamber is only necessary near the valves to accommodate the valve movement. It is not necessary and time consuming to use the unstructured mesh for the complete cylinder during the intake stroke. For this a layering zone is introduced. The zone is inserted at a specied crank angle (angle where maximum valve lift is reached). At this angle the mesh, that is created by the moving piston, will be build of layers with a constant specied height instead of the unstructured grid cells. This method requires fewer cells and is computationally more ecient. Now the grid is dened in Gambit and can be used in Fluent for the internal engine simulations.

B.1.2

Simulation set-up Fluent

In this section the model set up in Fluent will be discussed. The applied approach in this work will be treated in detail, but there are dierent approaches possible.

Chapter B. In-cylinder model set-up

93

When the mesh made in Gambit is imported, the model set-up can start. It is always useful to check the grid whether there are negative volumes and/or right-handed cells. This option can be found in the Grid menu. When these errors occur, the simulation is aborted. With the Smooth/Swap option the grid quality can be increased. In case of a deforming mesh, when simulation times are substantial, it is useful to preview the mesh motion and check for errors. This can be done in the Solve menu after the complete simulation has been set up. Another point of interest is to scale the grid to its original dimensions. This option can be found under the Grid menu. Next the actual model set up will be considered. All the steps that are performed in the following part, are found under the Dene menu in Fluent. First the dierent models for Solver, Energy and Viscous (turbulence model) have to be chosen. In this case the Segregated, Implicit, Unsteady solver will be used but of course there are other possibilities. This also holds for the turbulence models. Which model has to be chosen depends on the necessary accuracy, CPU time, memory usage and grid properties. In this case the Standard k - model is used. In the Fluent database dierent materials can be found. The properties of the materials can be adjusted in the Materials menu. Properties can be held constant or, for example, calculated by the ideal-gas law. Again it is up to the user which parameters to chose. The boundary conditions of the model are set up in the Boundary condition menu. For detailed information about the dierent types the Fluent manual can be referred to. The boundaries are already dened in Gambit but can, when necessary, be changed. Deforming or moving zones are dened as walls and specied below. Only the inlet is in this case modeled as an inlet-vent. In the Dynamic Mesh menu the parameters for the deforming zones can be specied under the Parameters menu. Fluent has a special menu where parameters, which are useful for modeling internal combustion engines, can be set, namely the In-Cylinder option menu. The main parameters are engine speed, piston stroke and crank angle step size. In the Dynamic Mesh Parameters menu three mesh parameters can be set, namely Smoothing, Layering and Remeshing. Fluent uses these parameters to adapt the mesh in the dynamic zones. Which option is used for a zone can be specied in the Dynamic Mesh Zone menu. Each of these options will briey be discussed (for more details see [6]). Smoothing: When for a certain dynamic zone the Smoothing option in enabled, Fluent uses a spring analogy, where any two neighboring nodes are connected via a spring. If any boundary deforms, the nodal positions adjust until equilibrium is reached. The spring stiness is represented by the Spring Constant Factor and ranges from 0 to 1. When a value of 1 is chosen, the eect of a moving wall is felt mostly near the moving wall and not at a larger distance. A value of 0 means the stiness is maximal. The eect of the moving wall is felt everywhere in the zone. A value of about 0.5 gave good results but again dierent values are possible. A second parameter for the smoothing option is the Boundary Node Relaxation. This factor controls the extent to which the motion of the adjacent interior nodes aects the position of nodes on the deforming boundary wall. This factor also ranges from 0 to 1. With a value of 1 the interior nodes fully aect the movement of the nodes on the deforming boundary wall. With value of 0 on the other hand, the nodes on the deforming boundary wall will not move at all. This causes degenerated cells next to the deforming boundary, which will inuence the solution. Layering: The option layering is used for purely linear motion. It involves the creation and destruction of layers of cells based on the height of the layer adjacent to the moving surface. The dynamic mesh model in Fluent allows users to specify an ideal layer height on each moving boundary. The layer is split when the following equation is met: hmin > (1 + s ) hideal , (B.1)

where s is the split factor, which can be dened in the Dynamic Mesh menu. A layer is collapsed

B.1. 2D MODEL SET-UP

94

if the following equation holds: hmin < c hideal , (B.2)

where c is the collapse factor, which can also be dened in the Dynamic Mesh menu. The ideal layer height, hideal , can be dened for every dierent layering zone in the Dynamic Zone menu. Remeshing: The last method to adapt the dynamic mesh is the Remesh option. When the boundary displacement is large compared to the local cell size, the cell quality can deteriorate or the cells can become degenerated. This will invalidate the mesh (e.g., result in negative cell volumes) and consequently, will lead to convergence problems when the solution is updated to the next time step. To prevent this, Fluent remeshes those cells that exceed the specied maximum skewness and/or fall outside the specied range of cell volumes. The maximum skewness and maximum and minimum volume can be dened in the Dynamic Mesh menu under Parameters. Now the dierent mesh methods have been described, the dynamic zones in the Dynamic Mesh Zones menu will de dened. In table B.1 the dierent zones with the corresponding mesh methods and types are displayed. Table B.1: Overview of the dynamic zones in the Dynamic zone: Mesh option: Combustion chamber interior Remeshing Piston Layering Piston layer Smoothing Combustion chamber wall Remeshing Interface combustion chamber Remeshing and Smoothing Interface valve layer Remeshing and Smoothing Valve layer interior Smoothing Valve bottom Valve top Layering
2D In-cylinder simulation

Type: Deforming Rigid body Rigid body Deforming Deforming Deforming Deforming Rigid body Rigid body

Movement: Full piston motion Limited piston motion Valve motion Valve motion

There are four types of dynamic zones. In this example two types are used (see table B.1) and will be treated further. The Deforming type is used for all zones where the mesh changes in time. When this type is chosen, the mesh method must be dened in the Mesh Option Panel in the Dynamic Mesh Zone menu. Parameters like maximum and minimum length scales and maximum skewness can be dened here. The piston and valve are of the Rigid Body type. For a rigid body a movement can be prescribed. The movement of the piston is calculated by Fluent itself and can be dened in the Dynamic Mesh Zone menu under Motion Attributes. For the motion of the valve a prole has to be prescribed. This prole can be written in .txt format in any word processor. The extension of the document has to be renamed to .prof and can than be loaded in Fluent under the Prole option in the Dene menu. An example of a prole is shown here. ((valve-movement point 7) (angle 0 10 90 170 180 360 720) (lift 0.0005 0.0005 0.01 0.0005 0.0005 0 .0005 0.0005) ) The lift is dened in meters and the angle in degrees. The minimum valve lift is in this case 0.5 mm. The valve starts to open at 10 degrees and closes at 170 degrees. The movement is linear interpolated between these 7 points. It is also possible to write an User Dened Function (UDF). This is more elaborate because the function has to be written in the C programming language. See the UDF manual in the Fluent documentation [6] for more details. The parameters, which control the dierent mesh methods, have to be optimized for every simulation.

Chapter B. In-cylinder model set-up

95

Therefore they will not be mentioned. It is up to the reader to nd the optimum settings. It is recommended to get familiar with the mesh methods by the use of a simple model. Try dierent settings and observe the dierences. When everything is clear, increase the diculty of the model. Do not bite o more than you can chew! The last detail that has to be treated is the interface between valve-layer and combustion chamber. Fluent assumes that once the mesh topology is set up, it is unchanged throughout the entire simulation. Therefore Fluent does not allow the valve to close completely such that the cells become degenerated (at cells). This requires a minimum opening between valve and valve seat (minimum valve lift). To physically close the valve a wall is formed between combustion chamber and valve layer. When the valve opens, the wall boundary is changed into an interface, which allows uid to pass through. Secondly the interface makes it possible to have dierent grid types next to each other (non-conformal grid interface). This is necessary between the valve-layering zone and the combustion chamber because structured and unstructured meshes exist next to each other. As mentioned in the previous part an interface consists of four edges. Two belong to the combustion chamber and the other two belong to the valve-layer zone. Because a minimum opening has to exist, the interface edges will rst be dened as walls. At a certain crank angle these walls will be transformed into Sliding Interfaces. This can be accomplished in the Dynamic Mesh menu under Events. Under the Dene button select Create Sliding Interface and select the two edge pairs under Interface Zone. When the valve closes, the sliding interface has to be transformed into a wall again. This can be accomplished by dening a Delete Sliding Interface event in the Event menu. When maximum valve lift is reached, it is not necessary to create an unstructured mesh in the complete combustion chamber. Therefore a layer zone will be introduced. This is also possible in the Dynamic mesh menu under Events. Select the crank angle at which the layering has to start. Under type select the Insert interior zone layer option. Finally the base, which is the piston, and the side dynamic zone, which is the combustion chamber wall, have to be selected.

B.2

3D model set-up

The construction of a three dimensional gird is basically the same as a two dimensional grid buildup. In this case the original geometry of an existing internal combustion engine is used. The geometry is complex and the preparation in both Gambit and Fluent are more sophisticated. In this case one cylinder of a 12.9 liter DAF heavy-duty Diesel engine is modeled. The inlet manifold, runners and intake valves are also modeled. Just like the previous case, the three dimensional model set-up is divided into two sections, namely the geometry set-up in Gambit and the simulation set-up in Fluent.

B.2.1

Geometry set-up Gambit

Importing the geometry (gure 5.11(a)) into the pre-processor Gambit and making it suitable for the simulation is the rst hurdle to take. The geometry is available in an IGES-format, which is a kind of exchange format between dierent software packages. Gambit oers dierent import strategies which also include several geometry heal options. When a geometry is imported, dierent errors like short edges can creep in and Gambit can automatically repair these. Because the complexity of the geometry the heal option does not have the desired eect and is turned o. Moreover a part of the geometry was damaged and had to be xed rst. For a detailed description of the import possibilities the reader is referred to [7]. In gure B.2 the adopted geometry of the RS 0.5 cylinder head is depicted. The inlet manifold and both runners are visible, as well as the piston which is at the top dead center (TDC). When the geometry has been imported, it consists of two layers. One layer includes the edges which form the contours of the faces and the second layer consists of faces which contains the shape of the geometry. The dierent faces can be stitched to create a volume. The created volumes (gure B.2(b)) represent the zones described in

B.2. 3D MODEL SET-UP

96

the 2D case.
Mass flow inlet

Inlet runner Valve stem zone Valve layering zone Combustion chamber

(a) Wireframe view of the geometry

(b) Exploded wireframe view

Figure B.2: Adopted RS 0.5 cylinder head geometry used for the dynamic engine simulations Compared to the 2D set-up, the 3D set-up requires one extra zone, namely the Valve stem zone. This cylindrical volume is used to create a layering zone above the curved center of the valve. In gure B.3(a) the dierent zones are schematic depicted. In gure B.3(b) the interfaces and moving zones of the valve are presented.

4
B C D B C

3b 2 3a

(a) The dierent mesh zones

(b) Interfaces (blue), sliding interfaces (red) and moving wall zones (green) of the valve

Figure B.3: Arrangement of the dierent zones and interfaces in the 3D simulations Zone A: The inlet manifold is an example of a stationary zone. Nothing is moving or deforming and therefore no extra requirements are necessary in this zone. Both structured and unstructured mesh types are possible. Three sections have to be dened as interfaces as indicated in gure B.3(b). Zone B: The valve stem zone is the extra zone which is necessary in the 3D simulations for the movement of the center of the valve. To save computational time, the mesh is structured. An useful tool is the Cooper tool to mesh this volume. Two sections have to be dened as interfaces and one for the valve movement as indicated in gure B.3(b). Zone C: The valve layering zone represents the volume between the valve and its seat. As mentioned in the 2D case, this initial volume represents the minimum valve lift. The interfaces with the combustion chamber are dened as walls. Later in the Fluent set-up these will be changed into sliding interfaces.

Chapter B. In-cylinder model set-up

97

Zone D: The combustion chamber has to be remeshed because the opening valves deform the initial mesh. An unstructured mesh is a must because this mesh type can only be remeshed. Two moving zones (piston layer and the bottom of the valve) are present and one sliding interface. Zone E: The last zone is the piston layering zone. This zone will be build out of layers and therefore the mesh must be structured.

B.2.2

Simulation set-up Fluent

The three dimensional simulation set-up in Fluent is almost identical to the two dimensional case. There is one remark concerning the set-up of the dierent interfaces. Every valve has four dierent interfaces (gure B.3(b)). One sliding interface between the valve layering zone and the combustion chamber which is also treated in the 2D case (# 1 in gure B.3(b)) and one normal interface which connects the valve stem zone and the inlet manifold (# 4). The two remaining interfaces are more complicated. The sliding interface between the valve stem zone and the inlet and valve layering zone (# 3a and # 3b) consists of three zones and therefore also of three faces. An interface in Fluent is dened by two faces. Merging the inlet manifold and valve layering zone makes the interface between those two zones unnecessary (# 2) and that side of the interface now also consists of one face. The face between the inlet and valve layering zone must be dened as a stationary zone in the dynamic mesh zones menu. This prevents the movement of the mesh in the inlet and the layering build up starts from that face.

1.60e-01 1.40e-01 1.20e-01 1.00e-01 8.00e-02 6.00e-02 4.00e-02 2.00e-02 0.00e+00

Y Z X

100

200

300

400

500

600

700

800

Crank Angle (deg)

Figure B.4: Motion of valve (red), piston (black) and piston limit (green) as a function of the crank angle In gure B.4 the motion of the valves and piston are depicted. The motion of the piston is split up in the complete stroke (black) and the limited motion (green) for the reason explained in the two dimensional case. The valve motion (red) is in this case linear. The plot is generated in Fluent by the print-plot-lift option. This option can be enabled in the text command interface by specifying the following commands. //dene/models/dynamic-mesh-controls/in-cylinder-parameter> print-plot-lift Lift Prole:(1) [()] Lift Prole:(2) [()] Lift Prole:(3) [()] Lift Prole:(4) [()] Start: [180] 0 End: [900] 720 Increment: [10] 5 Plot lift? [yes] y **piston-full** **piston-limit** valve-movement <Enter>

You might also like