You are on page 1of 18

m v n t kp x #wF2# rk n t kp #ck 1s2#od i11 { n p~#k{rvnt&#kodT1om }t#6iot p1v#qoFy p r t r y|k u {~nk { r ~nk m ~ {pm qk ypm v u t p rk qpn m lk }|Bz#cBxwHovs#BoF#j

eY h ` e` t i@ Fgg f ` 61cdcH4t v d X2ihXp t e Y b ` b ` g Y Vd e V V V pd b g e g Y ` ` p x `d XyF0w vu tHg X%Hp X2ihf2caXXU TRP I S s V S r q V pd b g ed b ` Y W V S Q

G E C A9 5 8 5 3     ) '    !     HFDB@$7$642$#10(&%$#" 

A CFD Investigation of 2D and 3D Viscous Pulse Propagation in Engine Exhaust Systems


Rasmus Hemph
Engineering Mathematics

Abstract The commercial computational uid dynamics code Fluent is validated for viscous pressure pulse propagation using data from a previous experimental setup. The importance of dierent parameters are studied, and the correlation between measurements and simulations are found to be good. Following this validation, a new pulse converted engine exhaust manifold is created, with the goal to decrease the cylinder residual gas level for a ve cylinder turbocharged gasoline engine, currently in development at Volvo Cars. The new manifold is simulated in a 1D-3D coupled CFD simulation, which had to be interrupted prematurely due to excessive convergence times. The results were inconclusive, but indicate poorer performance for the new manifold. Several suggestions to improve the simulations and the manifold are oered.

Sammanfattning Den kommersiella programvaran Fluent valideras fr o visks tryckpulsutbredning mot data fr en tidigare o an experimentuppstllning. Olika parametrar vrderas, a a och korrelansen mellan mtvrden och simuleringar a a bedms vara god. I fljande avsnitt utvecklas ett nytt o o pulskonverterat avgasgrenrr, med avsikt att minska o restgasniv i en femcylindrig turboladdad bensinmoan tor under utveckling vid Volvo personbilar. Det nya grenrret testas i en kopplad 1D-3D CFD simulering, o som m aste avbrytas i frtid p grund av en mycket o a l konvergenstid. Resultaten ar inte entydiga, men ang indikerar att det nya avgasgrenrret ger smre preo a standa an det nuvarande. Flera frslag ges till hur o simuleringar och design kan andras fr att frbttra o o a framtida grenrrssimuleringar. o

Contents Introduction 2

Navier-Stokes equations and Computational Fluid Dynamics Method CFD validation Results Discussion

2 6 6 9

Conclusion Engine charging theory Pulse Converters applied to a 5-cylinder Volvo engine

9 10 12

Results Discussion Conclusion Acknowledgements

12 13 14 15

p. 1

INTRODUCTION In the industry as a whole, and the automotive sector in particular, the ability to track uid ows is essential. To accurately simulate automotive type ows, computational uid dynamics, or CFD, serves as an invaluable tool early in the design process. At Volvo Cars, computational uid dynamics is used both for exterior calculations (aerodynamics) and for the interior (in-cylinder mixing, engine cooling, gas ow/pressure calculations). The process of simulating uid ows by three dimensional CFD is very computing intensive and has previously been viable for stationary ows only. With the increase in computing power over the last decade, it is now feasible also for time dependent ows. However, in computing ows over complex geometries and long time periods, the potential of error largely increases and it is important to be able to validate the results thus obtained. A ve cylinder turbo charged engine, currently under development at Volvo cars, displays unwanted behavior at low engine speeds. The cylinder residual gas level is high, causing power degradation. This is not surprising, as ve cylinder engines, due to valve opening overlaps, are prone to exhaust pulse interference. R.Birmann, 1946, created a modied exhaust manifold, labeled the modular pulse converter in an eort remedy this behavior. In this thesis, a new exhaust manifold, based on modications of Birmanns design, is applied to the 5-cylinder Volvo engine. To simulate the new manifold, unsteady calculations over a full engine cycle is performed in a 1D-3D coupled CFD-model. The pulsating nature of the exhaust ow, with velocities at, or over, the speed of sound, put high demands on the CFD code. Initial transient CFD simulations are performed to validate the performance of the code in viscous pulse propagation. The paper is divided into two parts. The rst begins with a background section, where Navier-Stokes equations, the theory of CFD and the nite volume method is discussed. The part is ended with a comparison between pressure pulse measurements and CFD simulations. The second part of the paper is opened with a theory section on engine charging and exhaust manifold design. In the nal section, the newly developed manifold is simulated and results compared with the the same engine and a conventional manifold. NAVIER-STOKES EQUATIONS AND COMPUTATIONAL FLUID DYNAMICS Looking around us, we can spot ows everywhere: the wind sweeps past us on a gusty day, the water

in a river whirls downstream, the tea-pot boils sending away steam. All these dierent phenomena deal with the transfer of momentum and energy. During the 19th century, the rst model describing the unsteady motions of uid ows was developed. Dierent persons are known to have derived the equations, (most notably the famous physician Poisson) but the French engineer Navier and the British mathematician Stokes (Figure 1) were nally independently credited for the achievement. The model consists of ve cou-

Figure 1: Claude-Louis Navier (17851836) and George Gabriel Stokes (18191903). pled equations describing basic physical properties that a uid must abide. First, an equation which states that the time rate of change of density in a region equals the net ow of mass into the region the continuity equation (1). Second, three equations based on Newtons second law, stating that the change of momentum of a particle is equal to the sum of forces acting on the particle the momentum equation (2). Finally, the energy equation (3), which stems from the rst law of thermodynamics: that the rate of change of energy within a uid region convected with the ow is equal to the rate of energy received by heat and work transfers. In this paper, these ve equations combined will be referred to as the Navier-Stokes equations: (For a thorough derivation, see [1] or [2].) D Dt Dvi Dt DE Dt = i vi = i p + j ji + Fi = pi vi + ji j vi i qi , (1) (2) (3)

where ij is the viscous stress tensor, dened by 2 ij = ij k vk + j vi + i vj 3 . (4)

The equations are written in Cartesian vector notation, where an index runs from one to three, a single index implies a vector component (i.e. vi = v in p. 2

vector notation), two equal indices implies summing a time-averaged quantity. An illustration can be seen over the index (i.e. ii = 11 + 22 + 33 ) and two difin Figure 2. ferent indices denotes a matrix (i.e. aij = A in matrix notation). Furthermore, the derivative symbol D( )/Dt is used in the meaning substantial derivative, (i.e. D( )/Dt = ( )/t+(v )( )), the time derivative u (t) + U of a particle in the ow, looked upon by a stationary bystander. The notation vi in the equations is used for U velocity, E for internal energy, qi for heat ow, for density, p for pressure and for the dynamic viscosity. The term Fi in the momentum equation (2) denotes a body force; a force acting through the center of gravity PSfrag replacements of the uid particle. The dependent variables are vi , p, and E, altogether six variables and ve equations. To close this set, we need to add an equation of state, that is we assume for example that p = RT , meaning t that we are dealing with a perfect gas. The equations above are valid for all ows. They are however impossible to solve analytically, except for in simplied cases with laminar ow. The cause of these diculties is that real ows exhibit complex behavior and often contain turbulent regions. Turbulence can be thought of as an unsteady, irregular, swirly motion that uids exhibit at high velocities (large Reynolds numbers1 ). Turbulence redistribute the energy of the ow, taking energy from the main ow and dispensing it at successively smaller eddies2 in what is known as the energy cascade. At the smallest eddy size, approximately 0.1 1 mm, the turbulent energy is released to the ow in the form of heat. The processes involved in creating and maintaining turbulence are complex and chaotic in nature. Being very challenging to model, turbulence is the main source of diculties in the computational uid dynamics eld. It needs to be modeled in a way consistent with reality, although not too detailed or the computing and storage cost will be overwhelming. This can be achieved by switching our interest to the mean properties of turbulence. For industrial ows, it is rarely necessary to track the smallest eddies or use the smallest time scales. In order to proceed, we smear the short-scale inuences of turbulence by timeaveraging the equation set. First, we split the velocity u(t) into a constant part U and a part which is allowed to vary in time, u (t). The uctuating component, can be thought of as the turbulent part. Furthermore, we consider the uctuations to be truly random, so that the time average of u (t) equals zero. Mathematically we write 1 T u (t) = T 0 u (t) dt = 0, where the overbar is used for
1 The Reynolds number is a parameter closely connected to turbulence. It is dened as Re = U/(L), where U is the velocity, L is a length scale of the ow and is the kinematic viscosity. 2 An eddy is a blob of uid of discernible shape, moving in some way dierent from the main ow

Figure 2: Velocity in a point at turbulent ow. The velocity u(t) is divided into a constant part, U and a uctuating component u (t).

We treat the pressure similarly and substitute u (t)+U for u and p (t) + P for p into (1)(3). Time averaging the set allows us to reformulate the instantaneous continuity and momentum equations as i Ui = i u i = 0 j (Uj Ui ) = i P + j j Ui j (uj ui ). (5) (6)

The reader should note that we for brevity exclude the energy equation and write the equations for an incompressible uid (D/Dt = 0). When we time average the equations, six new terms emerge as uj ui , known as the Reynold stresses. Through these terms, turbulence makes itself known in our equations. The extra stresses are calculated through a turbulence model, which can range from zero equation algebraic approximations to the involved Reynolds stress equation model which adds seven extra equations to the set. The choice of turbulence model is obviously a compromise between computing time and solution detail; the mixing length model for example, an algebraic approximation is completely unable of estimating separating or recirculating ows, but might well be appropriate for simple low-shear cases. For industrial ows, the k- turbulence model is dominating. It is robust and validated for numerous types of industrial ows. It adds two extra PDEs to the set and is as such a moderately expensive method to use. The rst parameter in the k- model denotes kinetic turbulent energy, dened as k = 1 (ui ui ) where we now 2 customary drop the prime sign for uctuating compop. 3

nents. The second parameter, , refers to the dissipation of turbulent kinetic energy per unit mass. It is a PSfrag ow process taking place in the smallest eddies in thereplacements and is described by = 2eij eij , (7)

where is the kinematic viscosity ( = /), and the rate of deformation tensor, eij = (j ui i uj ), has been divided into a uctuating and a constant part (eij = eij + Eij ) and time averaged as before. Transport equations are then solved for k and viz. (k) t + i (kUi ) = i i k + 2t Eij Eij (8) t k and t () + i (Ui ) = i i + t e (9) 2 + C1 2t Eij Eij C2 , k k where t is the eddy viscosity dened by t = C k . The equations contains ve adjustable constants, C , k , , C1 and C2 , which have been found through data tting of a wide range of ows. The k- model is elliptic by nature for steady ows, and must be given a distribution of k and at the inlet and Neumann boundary conditions ( (k,) = const) at an outlet. For n industrial ows, the distribution of k and is seldom known beforehand and is often approximated from formulae connecting inlet velocity and a length scale to turbulence production. [2]
2

Figure 3: Dierent kinds of cells surrounding a grid node. From the left, a prism, a tetraeder and a hexaeder. number of face areas (see Figure 3). For numerical reasons, it is common that some properties are calculated on the volume faces and some in the cell center. Second, the underlying equations are integrated over the CV. The sought parameters are momentarily considered to be known at the neighboring points and are approximated on the cell faces by means of interpolation. This approximation can be done by dierent numerical approaches, e.g. rst and second order accurate. If desired, it is possible to use dierent interpolation schemes for each ow property. Since interpolation always depends on distance, the nite volume method is clearly grid dependent, and the mesh size should be as small as possible to increase accuracy. To save computing time, it is advantageous if the grid can be made locally ner where large gradients are expected. In the third and last step, the control volumes are visited one at a time, an equation system is built and boundary conditions are included. The solution of the, commonly very large, equation system is always done in an iterative manner.

CFD-MODELING In order to model this averaged non-linear set of equations, some sort of numerical scheme of solving and discretizing dierential equations has to be used. In the commercial CFD eld, this is usually done using the nite volume method which has proven robust, and easy to implement and scale. The majority of the commercial codes, such as Fluent, Star-CD, Flow3D, all use this computational means of solving industrial uid ows. The major competition comes from the nite element method, which has attractive mathematical properties, but has yet to nd widespread approval in the commercial uid dynamics world. The nite volume method comprise three main steps: First, the computational domain is divided into small control volumes (CV) in the mesh generating process. Each volume has a central point P where properties such as temperature and velocity are calculated and a

UNSTEADY CFD MODELING Many industrial ows are modeled as being stationary, that is, it is assumed that /t = 0, where is a ow property. In real life ows, this is strictly not true, as stochastic variations make the ow unsteady. However, in cases with time varying boundary conditions, or cyclic ow, the solution will certainly be time dependent, and we need to track the ow in time as well as in space. For marching in time, we can use either explicit or implicit methods. The former category includes the rst order accurate Euler type iteration which is advantageous in that a full equation system does not have to be solved at each time step. The restrictions on the time-step size are rather severe however, and limits its uses for general transient problems. The implicit methods, including the rst accurate order backward Euler, and the second order accurate Crank-Nicholson algorithms, tend to be more stable, and allow a larger time-step but requires additional computational and storing cost. The Courant number, C, is a common p. 4

dimensionless parameter relating space and time steps, dened as ut . (10) C= x It describes the number of cells a uid particle is convected per time step. For accurate solutions, time and grid sizes should be adjusted such that C < 1 (the Courant-Friedrichs-Lewy (CFL) condition). If the unsteady modeling is taking place over large time periods, it is important to emphasize accuracy, as initial errors otherwise tend to grow over time. Tucker [3] generally recommends that higher order discretization schemes, as well double precision arithmetics should be used in order to minimize integration errors. Generally, eorts should be made to make the ow less time dependent. If possible, transformations to the ow and boundary conditions should be made to decrease time dependence.

of the post-processing was done in Fluent, which is capable of iso-surface and vector plotting for many different ow properties. For many 2-D properties, the results were exported to the scientic computational software Matlab, which has excellent plotting capabilities.

INDUSTRIAL CFD-MODELING One of the CFDsoftwares currently used at Volvo is the commercial nite volume code Fluent. Developed by Fluent Incorporated, New Hampshire, it includes modules developed for engine and turbo development. It uses a nite volumes approach and features an easy to use graphical interface and an array of dierent discretization schemes and turbulence models. It also includes a post-processor, capable of displaying a large number of ow properties. The process of CFD includes three main steps: the preprocessing step, the solving and the post-processing. In the pre-processing step, the computational domain is dened, the boundary conditions set and the discretization selected. In this paper the meshes were created in T-grid, a free-standing meshing utility, also from Fluent Inc. It is able to generate meshes consisting of prisms, hexahedral or tetrahedral elements as well as mixes of dierent elements. The grid is generated by means of the Delaunay-Voronay triangulation algorithm, with dierent measures of quality applied in the process. The mesh is subsequently exported to Fluent, where the boundary types (inlet/outlet, wall etc.) and the boundary conditions are set. Finally, the turbulence model, discretization schemes and convergence criteria are dened. When this process is completed, the actual solving is done. The equations are parallelized to t on multiple processors and are allowed to iterate until the convergence criteria is met. In the post-processing step the results are grouped and visualized. To be able to communicate the information from the simulations, it is important that this is done in a consequent and clear manner. In this paper, much p. 5

PSfrag replacements

METHOD CFD VALIDATION


P T Station 1 P 317
+

Volvo Cars have been using CFD to calculate stationary ows for over a decade. Unsteady ows are commonly calculated with the aid of one-dimensional codes, which have been thoroughly tested and give quick results, but by design are not able to simulate complex three dimensional structures or complicated ow paths. With the increase in computing power over the last years, the possibilities of the CFD eld has grown, and it is now viable to do time-dependent calculations for large geometries and relatively long time scales. However, when time-stepping over large periods, the possibility of errors increase. In order to evaluate and verify the CFD calculation process for unsteady automotive type ows, a pressure wave measuring device, labeled the QUB (see Figure 4), was simulated. The purpose of the QUB is to simulate pressure waves of magnitude and composition close to what is normally found in engine inlet and outlet manifolds. The device is suciently long to permit waves traveling without undergoing immediate superposition eects, thereby easing the process of analyzing and comparing results. The device was created by R.K. McMullan at the Queens University of Belfast, and the experiments simulated below was carried out by S.J. Kirkpatrick in order to measure the motion of compression waves in a straight pipe attached to a sudden expansion. Further details regarding the apparatus can be found in [4] or [5]. In Kirkpatricks experiments, the left cylinder is lled with air at a pressure of 1.5 bar. At time t = 0, a valve is opened, releasing a compression pulse to the pipe, lasting approximately 0.008 s. The wave travels to the right at the local velocity of sound, is partly reected at the geometric expansion and fully reected at the open end. Attached to the QUB are three pressure transducers that measures the static pressure. The experiment setup is shown in Figure 4. In order to model the QUB apparatus, a grid was created with specications according to the layout of the device (see Figure 5). The geometry was completed in Ansa, and the meshing was done in T-grid, the Fluent preprocessor. It was initially set up as a 2D-case, assuming radial symmetry in the pipe. The pressure wave was modeled as a half-period sine wave starting in the pipe, immediately after the pneumatic valve in the QUB. As soon as the pressure pulse has passed, the left boundary is changed to a wall. In this setup, initial mesh and time scale sensitivity was investigated, along with dierent turbulence models and discretization schemes. After initial testing, a base congura-

Cylinder

Station 2 P

Station 3 P

3097 3394 3703 6049

Figure 4: The QUB apparatus. At time t = 0, a pneumatic device opens the valve creating a brief connection between the pressurized cylinder and the pipe. Pressure transducers at three locations on the pipe, records the static pressure. The rightmost end is open to the atmosphere. All lengths are in millimeters. tion was found. In the following, unless otherwise mentioned, the grid size is 4 mm, the time-step is 0.1 ms, second order discretization in time and space is used, the pressure-velocity coupling algorithm is PressureImplicit with Splitting of Operators (PISO), and the turbulence model is k- Shear-Stress Transport (SST).

Figure 5: Top: fully meshed QUB geometry. Bottom: detail showing geometry step. In the current gure, 155267 cells were used in the meshing process.

RESULTS First, the sensitivity of the solution with respect to the time step was studied. Four dierent runs were completed, with increasing time step t on a structured grid with 4 mm sides. The Courant number was calculated as ut C= x for each dierent run, to be contrasted with the CFL condition, that C < 1. The results can be seen in Figure 6. p. 6

Static pressure, station 2 1.35 1.3 1.25 1.2 1.15 p/p0 1.1 1.05 1 0.95 0.9 0.85 0 0.01 0.02 0.03 t 0.04 0.05 0.06 Measured t=5e5s, C=0.78 (92 min) t=1e4s, C=1.6 (48 min) t=2e4s, C=3.2 (26 min) t=5e4s, C=7.9 (12 min)

The third parameter investigated is the turbulence model. It has been suggested in several references [3, 6, 7] that the method of choice for industrial applications, the k- model, has drawbacks in unsteady calculations. In order to compare the result based on dierent turbulence models, several runs were completed with the model variants k- Realizable, SpallartAllmaras, k- standard (see Figure 8).
Static pressure, station 3 1.06

1.04

1.02

Figure 6: A comparison between measured values and four dierent time steps and Courant number (C). The values in parentheses is processor time on a 366 MHz HP C360 workstation.

p/p0 0.98 0.96 Measured k Realizable SpalartAllmaras (1 eq) k k standard 0.01 0.02 0.03 t 0.04 0.05 0.06 0.94 0.92 0

It is also possible to alter the Courant number by changing x. In the second run, the solution grid size dependence is investigated. Starting with a 1 mm structured grid, the grid was coarsened to 2, and nally 6 mm (see Figure 7). Also shown in this gure, is the results of one-dimensional simulations of the QUB, done in the one-dimensional CFD-software Wave. It is included here for comparison reasons, as it is the primary software used for unsteady engine calculations at Volvo today.

Figure 8: A comparison between measured values and four dierent turbulence models. Although the results look similar, the dierence between the turbulence models is not negligible, (see Figure 11).

Static pressure, station 1 Measured 1 mm (C=6.30) 2 mm (C=3.15) 6 mm (C=1.05) Wave simulation

1.3

The fourth parameter of interest is the order of discretization. Although suggested by Tucker [3], that all discretization should be at least second order accurate, this might not always be possible due to convergence issues. In this run, we try combinations of rst and second order discretization in time and space, see Figure 9 for results. Finally, the model was set up in three dimensions to study large scale eects. Again, symmetry is assumed and the pipe is modeled in a 90 wedge shape with mirror boundary conditions on the inner sides. The number of cells drastically increases and ranges from 280000 to 300000. In order to accommodate this set, the solving is done on 4 parallel 2.3 GHz processors, running approximately 20 hours. See Figure 10 for results.

1.2

1.1 p/p0

0.9

0.8

0.7 0

0.01

0.02

0.03 t

0.04

0.05

0.06

To better be able to quantify the importance of the dierent parameters, the norm of the error, dened as 1 ||fmeas fCFD || = T
T 1/2 2

Figure 7: A comparison between measured values and four dierent grid sizes. Courant numbers (C) in parentheses.

|fmeas fCFD | dt
0

p. 7

is calculated for every case and monitor point. In Table 1 and Figure 11, the deviation in per cent from the best result is presented.
Static pressure, station 2 1.35

Table 1: The norm of the error is calculated for every case and the deviation in per cent from the result 1.25 most closely matching measurements is calculated at each monitor point. For example, the case with 6mm 1.2 grid size is 5.1 per cent worse than the best case (5e-5s 1.15 time step) at monitor point two. For the discretiza1.1 tion rows, T and S corresponds to time and space; the 1.05 numbers indicate the order of accuracy. Mon 1 Mon 2 Mon 3 1 1 mm 16 20 0.95 2 mm 4.0 9.2 7.6 0.9 6 mm 5.2 5.1 6.3 0.85 Wave 13.9 21.3 0 0.01 0.02 0.03 0.04 0.05 0.06 5 105 s 4.5 0 0 t 1 104 s 5.4 7.2 8.9 2 104 s 14 29 26 Figure 9: Dierent discretization schemes. The second 5 104 s 54 109 88 order discretization is done through QUICK, and rst k Re 0 0.51 3.4 order is done by an upstream weighted algorithm. PSfrag replacements Sp-All 17 23 24 k 5.4 7.2 8.9 k Std 25 25 49 2 T, 2 S 5.4 7.2 8.9 2 T, 1 S 5.2 6.9 5.8 1 T, 1 S 44 55 44 4 mm hex 5.8 9.7 8.6 4 mm tet 2.1 8.3 5.4 6 mm tet 16 34 24
1.3 Measured 2nd order time, 2nd order space nd st 2 order time, 1 order space st st 1 order time, 1 order space p/p0 Static pressure, station 2 1.3 1.25 1.2 1.15 1.1 Measured 2D 4mm hexaeder(240435 cells) 4mm tetraeder(299784 cells) 6mm tetraeder(155267 cells)

3D

Disc.

Turb.

Deviation in per cent from best case


5e-4s
100

Monitor1 Monitor2 Monitor3

p/p0

80
1.05 1

60
0.95 0.9 0.85 0

1T,1S k- Std

40

6mm tet 2e-4s 1mm Wave 1e-4s 5e-5s t Sp-All 4mm tet 2T,2S 4mm hex 2T,1S disc 3D

0.01

0.02

0.03 t

0.04

0.05

0.06

20

Figure 10: A comparison between measured values and four dierent grids.

2mm 6mm
0

k- Re turb

Wave

Figure 11: Bar graph of the deviation from the best case at each monitor point.

p. 8

DISCUSSION The results look promising. The CFD code is able to track the pressure pulsations, and the results closely match the measured values. At high peaks, most clearly seen in Figure 10, we can spot a discrepancy between solution and simulation, but generally all major characteristics of the pulse reections are captured. A possible cause of error is the way the pressure pulse is generated in the code. The left boundary is simulated as a sinusoidal pressure wave, with amplitude matched to t measurements, and period matching the pneumatic device of the QUB. The real pulse, however, might not be so ideally sine shaped, and it is possible that the lag in phase and amplitude seen at the rst pulse in station 1 is due to this simplication. In the time step simulations (Figure 6), we notice the importance of keeping the Courant number low. With t = 5 104 s, C = 7.9, we see a large deviation from measurements. Lowering the Courant number gives increasingly better results, and for C = 0.78, the best overall results are obtained for monitor point 2 and 3. In the grid size simulations, we alter the Courant number by varying the cell size. The time step is kept constant at 0.1 ms, and the Courant number varies from 1 to 6.3 as the grid is made successively ner. As we see results deviate more from measurements with a ner grid. This is unexpected, as all spatial discretization should improve with a smaller grid, but points out the need to match the time step as the grid is made smaller. The results from the turbulence model simulations looked quite similar to the eye, as we see in Figure 8, but the norm calculations show quite large variations. The one-equation Spallart-Allmaras, recommended in [8], for its behavior in wall shear ow, and the standard k- model behaved similarly, deviating 20% from the k- Realizable. The latter being an improved variation of the standard model containing a new formulation for the turbulent viscosity and a new transport equation for the dissipation rate, [9]. The behavior of turbulence models is very ow dependent, and the low turbulence-nature of the current experiment allow for the relatively good result of the oneequation model. It is worth pointing out that several sources disapprove of the k- standard model for unsteady modeling. The results from this case clearly agrees with that conclusion. From the discretization experiment (Figure 9), we notice that going to rst order discretization in space hardly inuence the results. This is possibly due to the

structured grid used in calculations. From the Fluent manual [9] When the ow is aligned with the grid (e.g., laminar ow in a rectangular duct modeled with a quadrilateral or hexahedral grid) the rst-order upwind discretization may be acceptable. When the ow is not aligned with the grid (i.e., when it crosses the grid lines obliquely), however, rst-order convective discretization increases the numerical discretization error (numerical diusion). Switching to rst order discretization in time inuence the results greatly. We see a 37% deviation from the best case, making this one of the most important parameters in this case. This is caused by the high velocity of the pressure pulse. Traveling with the local velocity of sound, second order accuracy is needed to track the quick changes. In the 3D-simulation, we increase the number of elements greatly, and therefore also the risk of numerical diusion. It is often heard that a structured grid minimizes numerical diusion, but in this case, going to a structured grid does not inuence the results. This is consistent with Fluent manual, claiming small dierences between structured and unstructured grids. However, going to a 6 mm tetrahedral mesh, we see deterioration of the results, much more so than in the 2D-case, and it is hard to explain without claiming numerical diusion. In Figure 7, and in Table 1, the results of a Wave simulation are included. It resolved the ow very quickly, and with good accuracy. It should be noted that Wave is developed for these cases, with simple geometry pipe ow, but it is nevertheless assuring to see such good results. CONCLUSION We conclude that the CFD code Wave, used for simulations at Volvo Cars, accurately simulate pressure pulses as tested in the QUB. In doing similar simulations, the following should be kept in mind. 1. The Courant number should be kept close to one. 2. Second order discretization, especially in time, is very important for an accurate result. 3. If possible, use a variation of the k- standard model. In the current study, both k- SST and k- Realizable gave good results.

p. 9

ENGINE CHARGING THEORY In a combustion engine, power is generated by allowing a mixture of air and fuel to be ignited in the cylinder. Optimum mixing conditions for stoichiometric mixing are 14.7 g of air to 1 g of fuel. In a turbocharged engine, a compressor raises the density of the air going into the cylinder, allowing a larger amount of fuel to be burnt at optimum conditions. Since more fuel is burnt at every cycle, the charged engine is not more fuel ecient per se than a suction engine. It is however more ecient per kilogram engine, and we are thus able to get more power from a charged engine than from an equivalently sized naturally aspirated engine.

EXHAUST PULSES AND ENGINE MANIFOLDS The turbine is powered by the pressure dierence between the manifold and the exhaust pipe. This energy potential is used to rotate the turbine and eventually compress the air used in the air intake system of the engine. The demands on the turbo system is high: it should be small, be able to handle a wide mass ow range, withstand high temperatures and at the same time be very reliable [10]. A real life system is always a compromise between these parameters and a turbinecompressor system always exhibits peak performance at a specic engine speed, decided by the characteristics of the turbocharger. A charged engine is sensitive to the layout of the exhaust manifold. The turbine will not function well unless faced with near-constant pressure conditions. To create good working conditions for the turbocharger, the exhaust manifold needs to be well-designed with a minimum of energy loss as the exhaust gases are transported from the cylinders. This transport can be done in dierent ways. Currently there are three dominant setups for exhaust manifolds, presented below.

There are dierent ways of charging an engine. A compressor can be connected directly to the crankshaft, taking power from the engine in order to charge the system. This method is often referred to as supercharging. A more common conguration for automotive engines is to let the compressor be driven by a turbine, which in turn is powered by the exhaust gas from the cylinders. An engine charged this way is what we normally label a turbocharged engine (see Figure 12) and the turbine-compressor system is what we refer to as a turbo. This approach is more energy ecient than a supercharged engine, since it utilizes energy contained in the exhaust gases. However, since the turbine imposes a ow restriction, the manifold pressure is higher than in a suction engine, and more power is needed to expel the exhaust gases. ag replacements
air inlet compressor intake manifold

cylinder ports

exhaust outlet turbine exhaust manifold

Constant pressure turbocharging A turbine reaches its full potential when it is allowed to work under constant pressure conditions. To achieve this, a constant pressure exhaust manifolds is set up as a large volume, connecting the cylinders and the turbine (see Figure 12). The volume functions as a pressure buer, keeping the manifold pressure largely independent of the individual cylinder exhaust pulses. This allows the turbine to work near optimum working conditions and increases the eciency of the engine. However, due to its size, the manifold cannot absorb suddenly applied loads, making it inappropriate for vehicle engines where a quick response is desired. Consequently, it is commonly found in larger engines, such as power supply units or naval engines.

Figure 12: Schematics of a constant pressure turbocharged engine. The turbine, driven by the exhaust ow, is connected to the compressor which supplies the intake with fresh air.

Turbocharging has long been a popular method for increasing power density in automotive type engines. It is a volume- and weight-ecient manner of increasing engine power output. Since it needs high-grade components to function well, it adds extra cost to the engine. Consequently, turbocharging is more commonly found in high-end engines.

Pulse turbocharging In automotive type engines, the pulse turbocharging system is dominating. It aims to utilize more of the energy available in the hot exhaust pressure pulses. Even though the turbine functions better at constant pressure conditions, the increase in available energy when individual pressure pulses are allowed to reach the turbine often compensate for this [11]. The cross-section area of the manifold is kept small to allow for quick pressure build-up and is thus able to handle sudden increases in engine load [12]. Because we use the quick pressure pulses, the potential of exhaust pulses from one cylinder interfering with the

p. 10

scavenging process of another cylinder increases. When this happen, the fresh intake air is mixed with old exhaust gas. Known as backow, this type of interference increases cylinder temperature, impairs combustion and is a big source of system knock [13]. An exhaust opening typically lasts 240 crank angle degrees (CAD.)3 . Since the four-stroke engine completes a full cycle at 720 CAD there are only room for three complete exhaust openings without interference (see Figure 13). To minimize the risk of backow, the manifold should be grouped in 3-cylinder congurations. For a six-cylinder engine, two three-cylinder congurations with a common connection to the turbine are often used. If this type of design is possible, backow is kept low and the pressure pulses arrive at the turbine suciently close to allow for good working conditions.

create a device that preserves the unsteady ow from the cylinders, and yet maintains steady ow at the turbine [10]. Birmanns original design (not shown) was bulky and expensive, and did not meet wide acceptance in the automotive industry. Others continued his work however, and several dierent pulse converter designs were developed and tested with good results over the coming years (see [14, 15, 16]). Curtil and Magnet (1978), and later C.L. Chan (1985) rened and simplied the original design by Birmann, creating a device named the modular pulse converter (M.P.C.) (Figure 14). It is intended to act as a uidic diode, allowing exhaust gas to travel from the cylinders to the turbine while at the same time restricting ow from the manifold to the cylinders. By adding many cylinders to a common manifold, the pressure variation at the turbine entry is reduced, and the eciency of the turbine is improved. To preserve the energy of the exhaust pulses, the manifold area is kept suciently small. By its modular design it is easy to t into an engine bay, and extra cylinders can easily be added without expensive and complex repiping. A manifold thus equipped is therefore also referred to as a compact manifold.

Pulse turbocharging, though very common, is best suited for engines with three-cylinder combinations. On engines with e.g. four, ve or eight cylinders, no ideal congurations can be found and attempts of connecting them in a pulse turbocharge manner are often linked to system knock and residual gases. Also, ordering the cylinders to minimize interference generally eplacements leads to complex manifold arrangements. 1
2

PSfrag replacements 3
Cylinders Turbine 1 2 Cylinders TDC (Cyl.1) BDC TDC BDC Exhaust open Inlet open Cyl.3 Cyl.2 Cyl.1 180 3 Turbine

PSfrag replacements
30 TDC
A1 A2

360 Crank angle

Figure 13: Pulse turbocharger design and approximate exhaust and inlet timings for a three cylinder engine with ring order 132. Illustration from [11].

Pulse converters To overcome the disadvantages of the pulse turbocharger for unfavorable numbers of cylinders, R. Birmann 1946 applied for a patent for a device labeled a pulse converter. His objective was to
3 One crank angle degree is completed when the engine crankshaft has revolved one degree


540 720

From cylinder

Figure 14: Cut-away picture of the modular pulse converter of C.L. Chan (1985). It is intended to allow ow to the turbine and limit inter cylinder ow as much as possible

Pulse converters have been actively discussed since the 1980s and is currently used in GEC-Ruston Diesel engines and in all engines made by SEMT Pielstick, both manufacturers of large engines. Much research has been done in order to improve the original design. Eorts have been made by Basset, Winterbone and Pearson [12, 17] to predict pressure loss coecients and to develop a one-dimensional methods of simulation. The pulsating manner of the ow, and the non-symmetric design makes such predictions dicult however. Leschziner and Dimitriadis [18] modeled a p. 11

3D pulse converter in stationary ow, concluding that this scheme could be used for data collecting for onedimensional simulations. Several practical studies have also been carried out. Winterbone et al. [14] performed an analysis of a M.P.C.-tted turbocharged 20-cylinder engine and found that the pulse converters were eective in propagating the ows toward the turbine while suppressing ow to the cylinders. Winterbone et al. [19] also developed the concept of turbine work function to track the energy propagation from the cylinders to the turbine inlet. By means of this technique, he compared two six-cylinder engines, one with a M.P.C. system, and the other with a common pulse manifold. He showed that the energy reaching the turbine was approximately equal in both systems, and warned that the transient response of the pulse converted system is worse than in the pulse system because the total volume seen by any cylinder is increased. This concern is also brought up by Watson [10], where he claims that the modular pulse converter system might display poor turbocharger acceleration. However, this eect might be balanced by the improved eciency of the turbine, especially in the case of ill-suited engine cylinder congurations (four, ve or eight cylinders). PULSE CONVERTERS APPLIED TO A 5CYLINDER VOLVO ENGINE At Volvo cars, a ve cylinder turbo charged gasoline powered engine is under development. Initial simulation shows that it is suspectible to engine knock, especially at low engine speeds. In todays Volvo engines, knock sensors monitor the cylinders, and reduces engine power should the engine start to knock. It is thus important to minimize knock factors in order to increase the power output. The engine manifold of the ve cylinder engine is of a common pulse-charge type, with all cylinders connected to a common turbine inlet. With ve cylinders connected to one inlet, the potential of exhaust pulse interference is large, and backow is indeed a problem. In an eort to remedy this, a manifold of of pulse-converter type is developed and tested. The design used is based on the modular pulse converter of Chan (see Figure 14). However, initial testing showed that the original design was too restrictive and that the ow exit was poor, creating a very turbulent ow. The new modular pulse converter was developed with a smoother neck, to minimize ow separation. Furthermore, the contraction A2 was enlarged, such that A2 /A1 = 0.5 as suggested in [15], to increase exhaust throughput, see Figure 15. Because of the modular design, ve new modular pulse converters were simply connected to create the complete manifold. In order to

Figure 15: The new modular pulse converter. Based on the model by Chan (Figure 14), it is made smoother and less contracting.

simulate the exhaust ow over a complete engine cycle, the 3D-manifold was connected to the one-dimensional engine simulation program GT-power. The engine, including the turbine and the compressor, is then modeled in one dimension, with the exception for the exhaust manifold which is is simulated in a fully threedimensional Fluent CFD simulation. In Figure 16, we see the CFD-domain included in the GT-power engine layout. In a coupled simulation, the engine is rst simulated as being fully one-dimensional, exchanging the manifold for a one-dimensional volume. After these initial cycles, the 3D CFD-code is connected, and the coupled simulations begin. The boundary conditions are passed between the two programs at each time step until convergence is reached. The manifold was meshed in three dimensions using Ansa and T-grid, with two layer prisms next to the walls and tetrahedral cells in the remainder. The coupled simulation was set up from within Fluent, with the salient parameters shown in Table 2. The simuations were run on a C3600 HP workstation with a 552 MHz RISC processor and 2 GB internal memory. RESULTS Due to several factors, the simulations was more time consuming than expected, and the simulations had to be ended prematurely after 10 coupled cycles, corresponding to a simulation time close to 25 days. Judging from 1D simulations, approximately 80 cycles was needed to reach stable conditions. In order to validate the results obtained in the CFD-simulations, and to compare the 1D and 3D codes, the pulse converted engine manifold was also set up in Wave, a onedimensional code similar to GT-power. To compare the dierence in performance between the two manifolds, the original manifold was simulated in one dimension. p. 12

In Figure 17, we can see a snapshot of the velocity, just after exhaust valve opening. In Figure 18, the residual gas level for the three dierent simulations is shown. In Figure 19, we see the power output at 1800 rpm for the dierent cases.

Figure 17: Visualization of the exhaust process. Figure showing exhaust gas velocity vectors soon after valve one is opened. Figure 16: One-dimensional engine simulation program GT-power showing part of the 5-cylinder engine. The rightmost large gray area is the manifold, included in the model as a CFD-domain. DISCUSSION As mentioned before, the calculations where very time consuming. Usually, CFD simulations are run on eight of more parallel processors, but due to licensing issues this type of setup was not possible in this case. Additionally, the GT-power model was equipped with a
9

Table 2: The most important parameters in the pulse converter simulation. Mesh 3D, 76747 cells 5 mm sides Unsteady formulation 1st-order implicit Discretization 1st-order upwinda Pressure-Velocity coupling PISO Turbulence model k- RNG with viscous heating Time step 4.63 105 sb Boundary conditions mass-ow inlets, no-slip adiabatic walls Engine parameters 1800 rpm, full load
a
b

8 7 Residual gases (%) 6 5 4 3 2 1 0 0 GTstandard WavePC CFDPC 10 20 30 40 50 Engine cycles 60 70 80

Except for pressure, which was 2nd-order accurate. Corresponding to one half engine crank angle at 1800 rpm

Figure 18: The residual gas level of the pulse converter after 25 cycles (10 coupled). Also shown are 1D-simulations of the standard manifold (dashed) and the pulse converter manifold (dashed-dotted). p. 13

400 350 300 Torque (Nm) 250 200 150 100 50 0 10 20 30 40 50 Engine cycles 60 GTstandard WavePC CFDPC 70 80

torque are the same, or worse, at all engine cycles than the standard manifold. Of course, it would be more satisfying to reach a denitive answer. In order to proceed with the simulations, some parameters need to change. The simulations could only be run rst order accurate in time due to convergence issues. From the initial validation (Figure 11), we noted the importance of accurate time discretization. Decreasing the time step, or making the mesh slightly larger could help us decrease the Courant number, and move to a second order scheme. This should be a priority for this high speed, compressible ow. Another concern is the turbulence model. In the rst part of the thesis, we noted dierences in results among the models, indicating the importance of choosing a correct turbulence algorithm. The current case include curved boundaries, diverging sections and highspeed ow, all properties where the k- model show poor performance [2]. To investigate the turbulence model inuence, it is suggested that the full Reynolds stress equation model (RSM) should be used. It is more complex in that all Reynold stresses (ui uj ) are resolved, meaning ve extra equations has to be solved at each iteration, but could prove more accurate in the current case. To further increase accuracy, we suggest that a wall heat loss model should be included. The impact of heat loss to the surroundings is unclear, but should be modeled for comparison in a possible continuation of the thesis. Furthermore, if we study Figure 17, we can see the exhaust gas hitting the top manifold wall. The geometry of the neck should be adjusted slightly, making the ow more directed in the direction of the turbine. In order to decrease the excessive simulation time, care needs to be taken to simplify the one-dimensional engine model. The PID-controller previously mentioned, should be modeled dierently, or removed totally. In doing this, the the number of cycles to convergence could be decreased to between ten and twenty. The most important issue however, is to resolve all licensing problems and run the simulations in parallel. The CFD codes used at Volvo are developed for multiple processors, and simulation time decreases close to linearly up to eight processors, depending on the problem size. CONCLUSION Judging from the results after ten cycles, the pulse converted manifold show poorer performance than the original pulse turbocharged manifold, but no denitive conclusions can be drawn. To improve future simulations, the following suggestions are oered.

Figure 19: The engine torque at 1800 rpm, showing the pulse converter manifold after 25 cycles (10 coupled). Also shown are 1D-simulations of the standard manifold (dashed) and the pulse converter manifold (dasheddotted).

PID-controller, regulating the wastegate. Due to nature of this controller, a larger number of cycles (80), than normal was necessary to reach steady state. This is not a concern in an ordinary 1D simulation, but as each coupled engine cycle nished in approximately 2.5 days, a complete simulation would have taken close to 200 days, clearly outside the time limits set for this thesis project. The simulations were therefore disengaged after ten coupled cycles, where we could hope for an indication of the results. As we see in Figure 18, the 3D-simulation and the 1Dsimulation of the pulse converted engine show similar behavior, both peaking after 18 total cycles (3 coupled). The cylinder residual gas level however, is higher than the original manifold at all cycles. This is discerning, as the purpose of the M.P.C. is to decrease residual gas. Looking at the torque data (Figure 19), we again see poorer performance for the pulse converted manifold, but not a obvious degrading, as the torque actually matches the original manifold at 25 engine cycles where the simulations where nalized. The one-dimensional model of the pulse converter was started with slightly dierent initial conditions (turbine speed 900000 versus 140000 rpm), explaining the initial dierences in results. When we look at the results, we must remember that the results obtained at an individual engine cycle is not directly comparable to between the three simulations, making conclusions dicult. At the least, we can say that the results indicate that the pulse converted manifold display poorer behavior than the standard manifold. Both the residual gas level, and the

p. 14

1. Resolve all license issues to run the simulations in parallel. 2. Simplify one-dimensional model, for faster convergence. 3. Increase the grid size/decrease the time step to allow for second order discretization in time. 4. Include wall heat ow, to track the importance of this energy loss. 5. For comparison, and to overcome the shortcomings of the k- model, simulate the ow with the full RSM turbulence model. ACKNOWLEDGEMENTS I foremost wish to express my gratitude to the entire group at Volvo thermodynamic analysis, all of you have been most helpful! Especially, I wish to thank Christian Gjerlv for answering my Fluent questions, Maro tin Jansson for tips on running coupled simulations, and Mikael Ewaldsson for introducing me to CFD at Volvo Cars. My supervisor, Said Tabar, deserves many thanks for contributing his expertise and good humour when things looked bad. I owe much to Professor Lars Davidsson at the department of thermo and uid dynamics at Chalmers, who generously donated his time in helping me nish this thesis, and to Associate professor Stig Larsson, who gave me valuable advice with the report.

p. 15

REFERENCES [1] R.L. Panton. Incompressible Flow. John Wiley & Sons, Inc, 1984. [2] H.K. Versteeg and W. Malalasekera. An Introduction to Computational Fluid Dynamics the Finite Volume Method. Longman Scientic & Technical, Essex, England, 1995. ISBN 0-582-21884-5. [3] P.G. Tucker. Computation of Unsteady Internal Flows. Kluwer Academic Publishers Group, Dorthdrecht, the Netherlands, 2001. ISBN 0-79237371-5. [4] G.P. Blair. Design and Simulation of Two-Stroke Engines. Society of Automotive Engineers, Inc., Warrendale, Pennsylvania, 2000. ISBN 1-56091685-0. [5] G.P. Blair, S.J. Kirkpatrick, D.O. Mackey, and R. Fleck. Experimental evalutation of a 1D modelling code for a pipe system containing area discontinuities. SAE International Congress and Exposition, Detroit, Michigan, 1995. SAE Paper No. 950276. [6] R. Alsemgeest, C.T. Shaw, S.H. Richardson, and S.Pierson. Modeling the time-dependent ow through a throttle valve. SAE 2000 World Congress, Detroit, Michigan, Mar 2000. SAE Paper No. 2000-01-0659. [7] M.-H. Kim. Three-dimensional numerical study on the pulsating ow inside automotive muer with complicated ow path. SAE 2001 World Congress, Detroit, Michigan, Mar 2001. SAE Paper No. 2001-01-0944. [8] D.C. Wilcox. Turbulence Modeling for CFD. DCW Industries Inc., second edition, 1998. ISBN 09636051-5-1. [9] Fluent 6 Users Guide. Fluent Incorporated, Lebanon, New Hampshire, 2002. [10] N. Watson and M.S. Janota. Turbocharging the Internal Combustion Engine. The MacMillan Press ltd, London, England, 1982. ISBN 0-333-24290-4. [11] B. Challen and R. Baranescu, editors. Diesel Engine Reference Book, second edition. ButterworthHeinemann, Woburn, Massachusetts, 1999. ISBN 0-7506-2176-1. [12] M.D. Bassett, R.J. Pearson, and D.E. Winterbone. Steady-ow loss-coecient estimation for exhaust manifold pulse-converter type junctions.

SAE International Congress and Exposition, Detroit, Michigan, Mar 1999. SAE Paper No. 199901-0213. [13] F. Westin, B. Grandin, and H.-E. ngstrom. A The inuence of residual gases on knock in turbocharged SI-engines. International Fall Fuels and Lubricants Meeting and Exposition, Baltimore, Oct 2000. SAE Paper No. 2000-01-2840. [14] D.E. Winterbone, R.J. Pearson, P.A. Bromnick, and S.K. Sinha. Analysis of a turbocharged intercooled 20-cylinder medium-speed diesel engine. Sixth International Conference on Turbocharging and Air Management, London, 1998. SAE Paper No. 984241. [15] C.L. Chan, D.E. Winterbone, J.R. Nichols, and G.I. Alexander. A detailed study of compact exhaust manifolds applied to automotive diesel engines. IMech E International Conference on Turbocharging, London, May 1986. SAE Paper No. 864113. [16] Z. Yong and L. Bing. Experimental investigatinos of a novel modular pulse converter turbocharging system for vehicle engines. SAE International Congress and Exposition, Detroit, Michigan, Mar 1999. SAE Paper No. 1999-01-1245. [17] M.D. Bassett and D.E. Winterbone. Modelling engines with pulse converted exhaust manifolds using one-dimensional techniques. SAE 2000 World Congress, Detroit, Michigan, Mar 2000. SAE Paper No. 2000-01-0290. [18] M.A. Leschziner and K.P Dimitriadis. Numerical simulation of three-dimensional turbulent ow in exhaust-manifold junctions. In International Conference Computers in Engine Technology, Proceedings of the Institution of Mechanical Engineers, pages 183190, Mar 1987. SAE Paper No. 874009. [19] D.E. Winterbone, J.R. Nichols, and G.I. Alexander. Eciency of the manifolds of tubocharged engines. In Proceedings of the Institution of Mechanical Engineers, volume 199, pages 137149. IMechE, 1985.

p. 16

You might also like