University of the Basque Country - Euskal Herriko UnibertsitateaUniversidad del Paı́s Vasco Faculty of Science and Technology- Zientzia eta Teknologia FakultateaFacultad de Ciencia y Tecnologı́a Máster en Iniciación a la Investigación en Matemáticas Master Thesis AERODYNAMIC ANALYSIS OF AXIAL FAN UNSTEADY SIMULATIONS September 2012 Author Imanol Garcı́a de Beristain Advisors F. Palacios & L. Remaki Agradecimientos (Acknowledgements) Me gustarı́a mostrar mi agradecimiento a todo aquel que haya contribuido a la creación de esta tesis de alguna manera. To Basque Center for Applied Mathematics for the trust they put on me. Specially to Sergey Korotov for letting me join up his group. To Francisco Palacios for his help and patience during my beginnings in the field. At last, thanks also goes to Lakhdar Remaki for helping me finishing this thesis and introducing me into my next steps. Agradezco ası́ mismo al Servicio Técnico de Informática Aplicada a la Investigación de la UPV/EHU, por haberme facilitado el uso de STAR-CCM+ en mi ordenador personal. Por último, mi más sincero agradecimiento para todos aquellos con los que he compartido el dı́a a dı́a de este largo año. Gente que todavı́a sigue ahı́ y gente que ya no esta. Por supuesto, también a mis padres por llevar conmigo esta carga. iii Aerodynamic Analysis of Axial Fan Unsteady Simulations Abstract: The objective of this thesis is to understand turbomachinery unsteady CFD simulation performance depending on the turbulence model selected. For this porpoise, simulations with two in industry extensively used turbulence modes have been carried out: k − ε and k − ω models. Simulations were performed using the commercial software STAR-CCM+. Results have been compared between them and with analytically obtained simplified solutions. This will allow to judge real industrial case simulations. Flow Rate results showed the same mean value downstream the fan for both turbulence models. However, oscillations induced by unsteady condition had different amplitudes. Boundary layers have been studied as well. It wasn’t found any difference among both turbulence models results, but simplified analytical problems solution predicted smaller boundary layers than STAR-CCM+ simulations. We obtained two main conclusions. First, k − ω is overpredicting blade performance compared to k − ε turbulence model because the second is more dissipative. However, if whole fan is considered, the extra blade efficiency on k − ω turbulence model is lost because of bigger recirculation zones. The second conclusion is the need for further understanding on boundary layer simulation, since deviations from expected results by both turbulence models can not be accurately explained by the author. However, it is probably related to surface curvature or blade edge pressure-gradient induced streamline curvature. Keywords: CFD, aerodynamics, axial fan, boundary layer, turbulence. Análisis Aerodinámico de Simulaciones No Estacionarias de Ventiladores Axiales Resumen: El objetivo de esta tesis es comprender la dependencia de los modelo de turbulencia en simulaciones no estacionarias de ventiladores axiales mediante CFD. Con este fin, se han realizado simulaciones empleando dos modelos de turbulencia ampliamente utilizados en la industria: los modelos k − ε y k − ω. Dichas simulaciones se llevaron a cabo empleando el software commcercial STAR-CCM+. Los resultados con los diferentes modelos de turbulencia se han comparado entre si, y con las soluciones analı́ticas de problemas simplificados. Con esto se pretende ser capaz de valorar las simulaciones de casos industriales reales. Los resultados muestran una media de aire igual para ambos modelos de turbulencia. Sin embargo, las oscilaciones alrededor de esta media son diferentes para ambos modelos. Los resultados de las capas lı́mite son independientes respecto del modelo de turbulencia esperado. Sin embargo, los modelos analı́ticos simplificados predicen capas lı́mites más pequeñas que las obtenidas mediante simulaciones. Se han obtenido dos conclusiones principales. Primero, el empleo del modelo k − ω resultará en un rendimiento de álabe mayor que el obtenido mediante k − ε, debido a que el segundo modelo es más disipativo. Por otro lado, si se considera el ventilador en su conjunto, el empleo de k − ω no supondrı́a una mayor eficacia porque también implica reflujos mayores. La segunda conclusión es la necesidad de un mejor entendimiento sobre la simulacion de las capas lı́mite puesto que no se ha logrado explicar de forma convincente la diferencia entre los resultados esperados mediante modelos simplificados y las simulaciones. Lo más probable es que se deba a la curvatura de la superficie o a la curvatura de las lineas de flujo inducidas por gradientes de presión. Palabras clave: CFD, aerodinámica, ventilador axial, capa lı́mite, turbulencia. Contents 1 Introduction 1 2 Background 2.1 Turbomachinery and Centrifugal Fans Essentials . . . . . . 2.2 Fan Aerodynamics . . . . . . . . . . . . . . . . . . . . . . . 2.2.1 Airfoils . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.2 Basic Characteristics of the Fan Flow Aerodynamics 2.3 CFD Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.1 Flow Physics Modeling . . . . . . . . . . . . . . . . . 2.3.2 Turbulence Closure . . . . . . . . . . . . . . . . . . . 2.3.3 Numerical Solution Techniques . . . . . . . . . . . . 3 Turbulence 3.1 Reynolds Equations . . . . . . . . . . . . . . . . . . 3.1.1 Reynolds Stresses . . . . . . . . . . . . . . . . 3.1.2 Anisotropy and Tensor Properties . . . . . . 3.2 Turbulent Boundary Layer . . . . . . . . . . . . . . . 3.2.1 Momentum Equations and Velocity Profiles . 3.2.2 Kinetic Energy and Reynolds-Stress Balances 3.2.3 Theory Limitations . . . . . . . . . . . . . . . 3.2.4 The Mixing Length Theory . . . . . . . . . . 3.3 Eddy Viscosity Based Turbulence Modelling . . . . . 3.3.1 Eddy Viscosity Hypothesis . . . . . . . . . . 3.3.2 Wall Treatment . . . . . . . . . . . . . . . . . 3.3.3 Realizable Two-Layer k − ε . . . . . . . . . . 3.3.4 Menters SST k − ω . . . . . . . . . . . . . . . 4 Simulations 4.1 Set-up . . . . . . . . . . 4.2 Results . . . . . . . . . . 4.2.1 Mass Flow . . . . 4.2.2 Force . . . . . . 4.2.3 Boundary Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 . 3 . 4 . 4 . 6 . 9 . 10 . 11 . 11 . . . . . . . . . . . . . . . . . . . . . . . . . . 13 13 14 15 16 17 21 23 23 24 24 25 25 28 . . . . . 31 31 34 34 35 36 . . . . . 5 Conclusions and future work 39 Bibliography 41 v Chapter 1 Introduction Research and investment in technology is often driven by market or legislation requirements. Turbomachinery is subject to the same conditions and fans are not an exception. For example, noise damping or N Ox reduction in aircraft engines are new research-lines driven by new normative constraints. On the other side, stage efficiency is many times imposed by the manufacturer itself to keep competitiveness in the market. Stage efficiency translates into specific aerodynamic performance requirements. New components will need much greater complexity during its design, including a higher degree of three dimensionality analysis and flow-path configurations. Design tools traditionally available in engineering, such as basic relative velocity triangles, have strong limitations to comply with demanded complex research. The use of advanced aerodynamic tools have to be developed. At this point Computational Fluid Dynamics (CFD) is being extended in the turbomachinery community to fulfil this task. Its ability to simulate flow physics at any point in the domain, with better predicting capabilities every year, is extremely attractive. CFD has already upgraded design of turbomachinery from 2-D inviscid flow models to 3-D, viscous, turbulent analysis [Earl Logan and Roy, 2003]. In this thesis unsteady simulation of an axial fan flow field is studied. Although aerodynamic components such as boundary layers or recirculation zones are considered, boundary layers take most of the weight of the thesis, since blade performance is directly related to phenomena such as boundary layer thickness or boundary layer detachment. Although boundary layers have been target for research for many decades, they are still under study. It is the case of aeroengines, which consider boundary layer induced vortexes because of the impact in aeroelasticity, blade thermal resistance, or noise generation. For this reason it was found appropriate to focus this thesis on a basic and yet extremely valuable boundary layers. Boundary layers modelling is strongly dependant of viscosity an Reynolds Stresses. For this reason, turbulence modelling is included besides boundary layers theory. In view of the extensive area of turbulence in bibliography, it was decided to cover the basics of Reynolds Averaged Navier Stokes (RANS) models, which will be linked to the boundary layer theory during its introduction. This thesis is structured in five main Chapters. Chapter two introduces the basics of turbomachinery aerodynamics and CFD. It also gives an overview of the state-of-the-art in the topic. It covers from airfoil lift explanation to latest trends in CFD simulations. Chapter three is the theoretical core of the thesis. Starting from the Reynolds Equations, boundary layer structure is deduced, to conclude with how turbulence is modelled within STAR-CCM. In Chapter four the Simulation set up and its results are reported such that it can be reproduced by the interested reader. At last, the conclusions are explained in Chapter five. 1 2 CHAPTER 1. INTRODUCTION Chapter 2 Background 2.1 Turbomachinery and Centrifugal Fans Essentials Turbomachinery is found everywhere in the modern world. This big family of machines include pumps, turbines and fans. The essential components are the rotor, which obviously is rotating; a shaft from which the energy is extracted or added to the rotor and a casing where the shaft stems from. Fluid is introduced through pipes in the case. Turbomachinery works transferring energy between fluid and rotor. When energy is extracted from fluid, the machine is called turbine, in the opposite case it will be a pump, fan or compressor. The rotor is mainly made from blades. This blades have a specific shape in order to make the fluid flowing between two blades to execute a specific force on the blades. It is commonly find as well some fixed components in order to drive the fluid smoothly. There are other ways of turbomachinery classification. A common way is the rotor flow exit direction: axial, radial or mixed. • Axial: Fluid flow is parallel to the axis. Ideally there is no radial component of the flow, only axial and tangential • Radial: Fluid flow is orthogonal to the axis, there is no axial velocity of the fluid. • Mixed: The flow has 3 velocity components within the rotor: axial, radial and tangential. Acording to Bleier [1998], there are 4 types of axial flow fans, 1. Propeller fans (PFs) 2. Tubeaxial fans (TAFs) 3. Vaneaxial fans (VAFs) 4. Two-stage axial-flow fans Propeller fan, sometimes called panel fan, is the most commonly used fan in any kind of application or environment. Tubeaxial fans have a cylindrical housing. Gas exhaustion is the most common application for this fans. The main negative outcome is the fast increase in the outlet duct friction losses due to air spin. When venturi inlet is used instead of a duct friction losses are reduced about 10 % of and noise level damped. Vaneaxial fans has a housing, like tubeaxial fans, but they have guided vanes that neutralizes the spinning air, so the unit is usable for blowing and exhausting (exit and inlet ducts). Use of venturi inlet is possible as in tubeaxial fan. Two-stage axial-flow fans are two fans connected in series, so pressure increase add up. It is useful when excessive tip speeds and noise levels are not tolerated. There might be guided vanes between two rotors rotating in the same direction. If counter-rotating rotors are used, guided vanes are unnecessary. 3 4 CHAPTER 2. BACKGROUND L F Suction side of airfoil LE D V TE α Pressure side of airfoil V Relative air velocity D α Angle of attack L Drag LE Lift TE Chord line Leading edge F Trainling edge Resultant Force Figure 2.1: Shape of a typical NACA airfoil. Source: Adapted by the author from [Bleier, 1998] The operating principle of axial-flow fans is simply deflection of air as will be described later in airfoil aerodynamics (section (2.2.1)). After flow passes the blades, the flow pattern has helical shape. Flow can be decomposed into two component: axial velocity and tangential or circumferential velocity. Axial velocity is the desired velocity since it moves air from/to the desired spaces. Tangential velocity is an energy loss in propeller fan or tubeaxial fans. In vaneaxial fans tangential velocity can be converted into static pressure. This makes vaneaxial fans more efficient. For good efficiency on the airflow of an axial fan it is usually demanded evenly distributed flow over the working face of the fan wheel. This means axial velocity should be the same from hub to tip on each blade. However, blade velocity is function of the radial distance. Velocity gradients are then compensated by blade twisting, which outcomes in a smaller blade angle toward the tip. Same hub to tip blade angle results in a loss of fan efficiency since air propulsion will take place mostly on the outer region of the blade. Incorrect blade twist might cause stall in the interior portion of the blade strongly hindering efficiency when working with higher static pressures. 2.2 2.2.1 Fan Aerodynamics Airfoils Airfoils used in fan blades are asymmetric. The best well known airfoils are NACA airfoils, which have been developed by the National Advisory Committee for Aeronautics. A NACA n◦ 6512 is shown in figure (2.1). Some features of this airfoils are acording to Bleier [1998] : • The airfoil has a blunt leading edge which provides robustness to small inlet flow perturbations, and because of structural strength characteristics. The trailing edge is rather sharp. • The chord line is defined as the line that connects the two lowest points of a two dimensional airfoil section when is laid on a flat surface. The airfoil chord, c, is the length of 2.2. FAN AERODYNAMICS 5 the blade profile orthogonal projection onto the chord line. • The airfoil has a convex upper surface, with the maximum at 36 % of the chord, and a maximum section height from the chord of 13.3 %. • A concave lower surface, with maximum distance of 2.5 % of c located at 64 % of the chord from the leading edge. It might happen the lower surface to be flat instead of concave for some applications. • The angle of attack, α, is measured as the angle between the relative air velocity and the base line. • As the airfoil moves through the gas, it produces a positive pressure on the lower surface of the airfoil (or pressure side) and a negative pressure gradient in the upper side (or suction side). Both forcces have approximately the same direction, but suction force is close to twice the positive gradient force. The forces described in the last point define F when they are added. This force can be decomposed into lift and drag forces. The first is perpendicular to the relative air velocity whereas the second is parallel. Lift is the desired component in most engineering applications. Drag is undesired since it causes power-consumption. However, this two objectives are conflicting each other, and a trade off hast to be made while designing to obtain high lift forces, but good lift-drag ratios. As the maximum section height of the airfoil profile increases, lift usually increases but lift-drag ratio tends to worsen. Selection of airfoil shapes is done according to the desired application. For example, for compressors, wide airfoils are used. When efficiency is the important parameter to be considered thinner airfoils are employed. Further, this forces are strongly dependent on the angle of attack, and the overall range of operation has to be considered when choosing the appropriate airfoil shape. Aspect ratio is an important feature. It is the ratio between the total blade height and the chord length. The bigger the aspect ratio, the better lift and lift-drag ratio. This is explained by the trailing vortex phenomena. Trailing vortexes are generated when the fluid flows from the pressure side to the suction surface though the outward space of the airfoil tip (the clearance space in case of turbomachinery with casing). They strongly hint lift force. Big aspect ratio is favorable because trailing vortexes have influence on a certain distance of the total wing. The longer the wing, the smaller the overall efect of the vortex. The use of hubs is common in turbomachinery because reduces turbulence and trailing vortexes at turbomachinery tips, increasing lift forces. Some basic airfoil performance facts are introduced next, • If the airfoil was symmetric, zero lift force would be found at angle of atack 0◦ due to symmetry considerations. • As it can be seen in figure (2.3), as the angle of attack is increased, lift coefficient increases as well. The maximum for this example curve is found at 15◦ . It is the maximum operating point of the airfoil. • The maximum lift-drag ratio is fount at 1◦ . For this example, best operating rate is regarded to be the range [1◦ − 10◦ ], where ratio is still high and airflow is smooth. • Angles of attack from 10◦ to about 15◦ are acceptable. Fluid streamline can still follow the contour of the airfoil. • When the angle is higher than 15◦ , the airfoil stalls. A huge fall in the efficiency occurs driven by the boundary layer detachment. The way airfoils are used in axial fans is shown in figure (2.4). 6 CHAPTER 2. BACKGROUND Figure 2.2: Trailing Vortex generation around the tip of an airfoil 2.2.2 Figure 2.3: NACA 6512 airfoil infinite aspect ratio characteristic curve Basic Characteristics of the Fan Flow Aerodynamics Many different and diverse flow features are involved in turbomachinery which goes from supersonic velocity to rotating flows. It is off special interest complex stress and performance losses that result from viscous flow phenomena, mainly located at the airfoil boundary layers, but also blade to wall transient boundary layers, near-wall flow migration, tip clearance and trailing vortexes , wakes and mixing. Research is also being done in relative end-wall motion and transition between rotation and stationary end walls. Another trend in aerodynamics research is the unsteadiness of the flow for time-varying conditions, vortex shedding from blade trailing edges (and impact in aspect ratio as described in section 2.2.1), flow separation, or intersections between rotating and stationary blow rows, which imparts unsteady loading on the blades and thus the life span of the fan. In centrifugal fans, when high flow is impelled, non-negligible boundary layers are formed in the second half of the impeller flow passages. Consequently, flow separation is encountered in the suction surface, causing wakes region, shifting the flow toward the pressure side. Flow separation diminishes diffusion potential for the impeller and distorted jet/wake structures are found in the discharge. Mixing follows this jet/wake structure and unsteady flow runs into the diffuser. Overall, this wakes reduces the efficiency of the fan. A well-known study of this kind is the one performed by Eckardt [1976] in a radial centrifugal compressor. He obtained detailed measurements of flow velocities and directions at multiple locations in the flow field, from the inducer inlet to the impeller discharge. Eckardt observed 2.2. FAN AERODYNAMICS 7 Trailing edge of airfoil blade Suction side on convex side Rot Hub Whell dia. o.d. Rot D ai efl r fl ec ow te d Incoming air flow Blade angle Leading edge of airfoil blade Pressure side on concave of airfoil blade Figure 2.4: Airfoil shaped blade axial fan [Bleier, 1998] that flow kept undisturbed within the axial inducer and during the first 60 % of the impeller blade chord. At the 60 chord % distance, a flow separation originated in the shroud suctionside corner of the passage. The separation rapidly grew to give raise to a wake. This wake is produced from secondary flows as will be seen later. In short, vortexes near the shroud and the hub/suction-side corners causes detachment in the boundary layers off the channel walls (including blade surfaces) and fed the low-energy fluid into the wake. Additional low-energy passes through the tip clearance space causing the wake to increase through the downstream half of the impeller flow path. The pattern of high and low-energy fluid (jet and wake respectively) is sustained up until the impeller discharge supported by the system rotation and curvature, that causes non-mixing between jet and wake structures. Meridional flow in centrifugal compressors is usually highly non-uniform, dominated by significant jet/wake flow after halve of the chord. Peak meridional velocities are located at the blade hub-pressure sides, due to potential flow at high Reynolds numbers. Jet/wake structure causes the flow to separate at the hub, and peak velocities locate at the blade tip-pressure side. The prediction of this meridional profiles can be done by simple modelling, empirical models or employing Euler flow solvers (for high Reynolds numbers). After meridional flow is obtained, normal and binormal vorticity can be numerically evaluated and applied to vorticity equations to calculate streamwise vorticity. The use of potential flow solvers can be employed to solve passage circulatory secondary flows. n-Euler equation The well known Bernoulli or Euler equations describe the mechanical energy terms (pressure and velocity) along a streamline with constant energy. Whereas the n-Euler equation describes the forces normal to this streamline. Thus, it is the equation explaining why lift occurs in airfoils or why secondary flows/vortexes are created. The equation is obtained balancing the centripetal acceleration of a fluid particle and the net force in the normal direction to the streamline. Consider the mass of a fluid element to be ρR dθ dn. The particles centripetal 8 CHAPTER 2. BACKGROUND Figure 2.5: Separated jet-wake structure from impeller [Tuzson, 2000] acceleration towards the centre of the streamline is proportional to ρv 2 /R. For an inviscid fluid in equilibrium, in absence of significant body forces the equation obtained is: ∂p v2 =ρ ∂n R (2.1) One might read this equation as follows. When a steady streamline is curved there is a centrifugal pressure gradient force acting on the streamline fluid particles. Pressure increases with curvature radius. As we stated before, equation (2.1) contains the principle behind airfoil lift and secondary flows. The concave curvature of the streamline in the pressure side of an airfoil means higher pressure is located below this surface compared to the undisturbed flow in the middle of the pitch (line equidistant between two consecutive blades). Same reasoning allows to demonstrate low pressure is encountered at the suction surface of the airfoil compared to the undisturbed flow. A pressure difference force driven by the airfoil surface is created as sketched in figures (2.6) and (2.7) Using n-Euler equation we will describe the secondary flow formation. Imagine an axial fan blade turning. Employing the n-Euler equation, we deduce a pressure gradient has to be created from the streamline curvature described in the lift force generation. However, further implications have to be considered in the near wall region (usually hub and case), where the flow velocity is smaller although a similar pressure-gradient (because of mechanical stability across the span of the airfoil) exists. From the n-Euler equation, for constant pressure gradient, we deduce ρv 2 /R = constant. If v is decreased due to viscosity effects, R has to decrease as well. The streamlines follow different passage trajectories close to the walls as shown in figure 2.8. The overturning of the fan moves slow velocity fluid close to the wall from the blade pressure surface to the blade suction surface, and the motion is compensated by a return flow near the centre of the passage. The combined effect is that two three-dimensional passage vortexes are set up in the streamwise directions. These vortexes are one of the main sources of secondary flow in blade passages. To completely understand this vortexes full momentum equations have to be 2.3. CFD BASICS Figure 2.6: Lift force description by the pressure gradient of n-Euler equation 9 Figure 2.7: n-Euler equation physical representation Figure 2.8: Secondary flow development in turbomachine blade passage. Cross-stream free flow (A) and the end wall boundary layer streamline (B) (left). Passage vortexes (right) [Japikse and Baines, 1997] employed. This inertia-generated vortexes might be known as circulatory flow. The overall effect of the vortex is a efficiency loose in the turbomachinery. In general, secondary flows are always caused by static pressure and kinetic energy imbalance. The most studied vortex generation mechanism is the horseshoe vortex by a stagnation line: an incoming boundary layer meets a stagnation line and causes a motion of the fluid along the wall, with a subsequent vortex. The strength of the vortex is determined by the starting conditions, and its evolution to the conservation of its angular momentum. The vortex flows are principally generated by the meridional flow field, while the centrifugal and Coriolis forces act on the vortex change of the direction (tilting of the plane) [Brun and Kurz, 2005]. 2.3 CFD Basics This section analyses the state-of-the-art in CFD for both turbomachinery design and complex flow filed analysis. Bear in mind this two objectives may demand diferent computational 10 CHAPTER 2. BACKGROUND resources according to the results desired. For example, in we find two stages in component design: preliminary and detailed design. During preliminary stage many variables are involved and consequently many prototypes are proposed. At this point we expect a tool which allows a selection of the most suitable options as fast as possible. Accurate solutions are not needed. CFD methods are suited to this requirements. We might be looking for efficiency improvements by blade-row spacing analysis or initial blade shape design. Detailed design focuses on a small number of design parameters based on preliminary design analysis. Examples for turbomachinery are tip clearance flows, blade-end wall interactions, flow separation, wakes, etc. Detailed analysis simulations require order of magnitudes longer time because flow physics has to be usually accurately resolved. 2.3.1 Flow Physics Modeling For industrial applications Navier Stokes Equations are not tractable due to computational limitations. Mathematical treatment such as averaging is typically done to obtain the Reynoldsaveraged Navier-Stokes equations (RANS) in the case of Reynolds averaging. Other mathematical treatments produce other models such as Large Eddy Simulation (LES), which is more popular in research. The set of equations including mass, momentum and energy conservation are known as full Navier Stokes Equations. However, use from auxiliary equations is necessary such as the equation of state (usually perfect gas law in fans), the Stokes hypothesis which relates the second coefficient of viscosity (or bulk viscosity) to the molecular viscosity, and Sutherland’s law, which expresses molecular viscosity as function of temperature. Turbulence model for the Reynolds stress closure is also added to the previous system of equations. Use of full RANS equations allows simulation of unsteady, 3D, viscous, turbulent flows in rotation where the primary dependent variables are density, three velocity directions, total energy, pressure, enthalpy, nine components of the turbulent Reynolds stress tensor and three components of turbulent heat-flux vector. Further techniques like the thin layer assumption is widely used to reduce the complexity of the problem. Within this assumption the streamwise diffusion term is neglected. It is used in viscous layers, but must not be used when recirculation zones or viscous structures producing streamwise diffusion is present. Solution of the PDE must be done altogether with appropriate boundary condition. Three type of spatial boundaries may be identified for turbomachinery: (1) wall boundaries, (2) inlet and exit boundaries, and (3) periodic boundaries. Wall boundaries refers to blade surfaces, passage walls, etc. Solid surfaces might be rotating, non-rotating or a combination of them. Zero-relative-velocity (non-slip) conditions should be used. The most natural form of inlet and exit boundary specifications are mass flow rate for inlet and pressure conditions at the outlet. To do this, pressure, temperature, tangential velocity upstream and pressure at downstream is usually specified. Depending on the turbulence model selected, some turbulence properties are required. The inlet/exit boundaries should be placed far enough from the blade, so they are not influenced by its presence. Typical distances to this boundaries are 50 % to 100 % of the blade chord. Distribution of inlet conditions might be included in the spanwise and tangential directions. Periodic boundaries upstream and downstream of the blade are used to model one blade passage to the next, assuming inlet conditions are periodic. Periodicity is forced by setting dependent flow variables equal at equivalent positions on the periodic boundaries. Straight forward initialization of the problem can converge into the solution with no problem. However, an appropriate distribution will fasten the convergence. This might be done by imposing the solution from a 2.3. CFD BASICS 11 preliminary design. For example, unsteady simulation from a previous steady solution, or inviscid flow solution for an viscid simulation. 2.3.2 Turbulence Closure As will be seen in next section, there are a handful of turbulence models providing closure of the Navier Stokes equations. They range from simple algebraic relationship to a set of PDE. The most commonly used models during preliminary design are two-equation models and Full Reynolds stress models. Turbulence tretment such as Large eddy simulation (LES) and full Navier-Stokes equations are to expensive at this stage. n-equation models represent into some extent the real turbulence physics. They use of n partial differential equations which model transport of selected turbulence variables. For example, in two equation models kinetic energy and turbulent energy dissipation are most often selected. When solution of transport equations are computed, they are introduced in algebraic models to obtain what is known as the turbulent viscosity. Full Reynolds stress model is a much more realistic representation of a turbulent flow but is makes use of approximately twice the number of equations. It Solves transport equations for all components of the specific Reynolds stress tensor R = –Tt /ρ ≡ v 0 v 0 . These model naturally account for effects such as anisotropy due to strong swirling motion, streamline curvature, rapid changes in strain rate and secondary flows in ducts. However, this accuracy increases the computational cost. LES and full Navier-Stokes equations are useful in some cases during detailed simulation stage or very sensitive physics simulation. For example, noise creation due to vortex generation in airfoil tips is a good example for use of LES turbulence models. 2.3.3 Numerical Solution Techniques Solution of fluid dynamic equations is a huge area in CFD which basics should be known by the interested people in the field. A basic introduction will be done here. First of all, PDE equations to be solved are discretized. Within fluid simulation, the most used method for discretization is the Finite Volume Method. Finite Difference Method had more users in the past, but requirements of structured grids compared to the flexibility of Finite Volume Method, hinders its selection. Finite Element Method is also very common especially in multiphysics simulations. After discretization, the system of algebraic equations have to be numerically solved. According to Earl Logan and Roy [2003], the most appropriate solution techniques are the time-marching (unsteady) explicit or implicit schemes, although steady Governing-Equations are also possible. Explicit time marching schemes are simpler and use information at grid cells from previous time-intervals. It is less computationally expensive than implicit methods. However, stability is an issue to be considered when setting up the problem: time-step size is dependent on grid size. On the other hand, implicit methods are numerically unconditionally stable to any timestep chosen. In this method, equations are solved all together in a coupled matrix. Obviously, the number of equations increases in the same rate as number of grid points does. Avoidance of matrix inversion by factorization procedures is usually done to reduce the computational cost. Hybrid methods might be used as well, which benefits from both implicit and explicit characteristics. Some examples include the implicit residual smoothing in an explicit RungeKutta technique to relax the stability criterion. Or the two-step explicit one-step implicit Beam 12 CHAPTER 2. BACKGROUND and Warming algorithm. Use of local time stepping is another approach. It is only useful for steady-state simulations. In this case, time-step in time marching methods does not have any physical relevance, and different time steps might be set to different grid points in the mesh. This allows stability to be satisfied while solution convergence is as fast as possible through all the mesh. Non of implicit or explicit methods have demonstrated overriding performance against the other and methods have similar levels of maturity. However, explicit methods are less computationally expensive and are more efficient in parallel-processor and vector computers. Whether implicit/explicit unsteady/steady has been chosen, two type of solvers are available: Coupled or Segregated solvers. The Segregated solver considers the flow equations (one for each component of velocity, and one for pressure) in a segregated, or uncoupled, manner. The linkage between the momentum and continuity equations is achieved with a predictor-corrector approach. This model has its roots in constant-density flows. The Coupled solver on the other side considers the conservation equations for mass and momentum simultaneously when solving using a time- (or pseudo-time-) marching approach. The preconditioned form of the governing equations used by the Coupled Flow model makes it suitable for solving incompressible and isothermal flows. One advantage of this formulation is its robustness for solving flows with dominant source terms, such as rotation. Another advantage of the coupled solver is that CPU time scales linearly with cell count; in other words, the convergence rate does not deteriorate as the mesh is refined. On the negative aspects of the coupled system we find the need of high computational memory resources. In this thesis segregated flow solver has been employed since the number of cells used exceeded the coupled solver capabilities in the computer employed. Formulation of Governing-Equations in STAR-CCM is Z I Z d ρχ d V + ρ v − bmvg · d a = Su d V (2.2) dt V A Z I I I V Z d ρχv d V + ρv ⊗ (v − vg )· da = − pI· d a + T · a + (fr + fω ) d V (2.3) dt V A A A V The terms on the left-hand side of equation (2.3) are the transient term and the convective flux. On the right-hand side are the pressure gradient term, the viscous flux and the body force terms. T is the viscous stress tensor (or Reynolds stress tensor). The body force terms have been simplified to represent exclusively the effects of system rotation. In turbulent flow, the complete stress tensor is given by: 2 T T = µef f ∇v + ∇v − (∇· v) I 3 where the effective viscosity is µ = µl + µt , the sum of the laminar and turbulent viscosities. Chapter 3 Turbulence Turbulence is, of course, an important component to be considered. It must represent the characteristics of typical turbomachinery flow fields, such as flow-path curvature, rotating flow, high-pressure gradients, and separated, recirculating flows. The capability to model unsteady flow and blade-row interaction is also necessary. In order to study turbulence, a simple introduction to the fluid motion equations will be done. They will be applied to Newtonian incompressible flows. 3.1 Reynolds Equations Decomposition of the velocity field U (x, t) into the mean flow U (x, t) and the fluctuation term u(x, t) is a common procedure. u(x, t) = U (x, t) − U (x, t) . (3.1) This decomposition is known as the Reynolds decomposition. The well known continuity equation ∂ρ + ∇ · (ρU ) = 0 ∂t is simplified for incompressible flows to yield the solenoidal or divergence-free equation ∇ · U = 0. (3.2) and applying relationship 3.1, ∇ · hU i = 0 ∇ · u = 0. In the case of the momentum equation some nonlinear terms are created after applying Reynolds averaging. First we write the substantial derivative conservative form. DUj ∂Uj ∂ = + Ui Uj , Dt ∂t ∂xi and calculate the mean value DUj Dt ∂ Uj ∂ = + Ui Uj ∂t ∂xi (3.3) Nonlinear average of Ui Uj is Ui Uj = hUi i + ui Uj + uj D E = hUi i Uj + ui Uj + uj hUi i + ui uj = hUi i Uj + ui uj 13 (3.4) 14 CHAPTER 3. TURBULENCE The velocity covariance ui uj is the so-called Reynolds stress. employing equations (3.2), (3.3) and (3.4) we rewrite the total mass derivative DUj Dt ∂ Uj ∂ = + hUi i Uj + ui uj ∂t ∂xi ∂ Uj ∂ ∂ = Uj + ui uj + hUi i ∂t ∂xi ∂xi (3.5) Further, if we define the mean mass derivative, which represents the rate of change of a point moving with the local mean velocity U (x, t) as D̄ ∂ ≡ + hU i · ∇ ∂t D̄t we obtain the relationship with the mass derivative (3.5) DUj Dt = D̄ ∂ ui uj . Uj + ∂xi D̄t The momentum equation (or Navier Stokes Equation) for an incompressible Newtonian fluid is 1 DU = − ∇p + ν∇2 U . (3.6) Dt ρ And relating with previous deductions, its mean mass derivative or Reynolds equations ∂ ui uj D̄ Uj 1 ∂ hpi 2 = ν∇ Uj − − . ∂xi ρ ∂xj D̄t (3.7) The structure of the Reynolds and the Navier-Stokes equations might look similar, but in the Reynolds equations the Reynolds stresses appear, which give rise to turbulence modelling. 3.1.1 Reynolds Stresses U (x, t) and U (x, t) show very different behaviour due to the Reynolds Stresses, which are one of the big puzzles of the field This stresses are better understood under the form ! D̄ Uj ∂ ∂ hUi i ∂ Uj = µ + − hpi δij − ρ ui uj . ρ ∂xi ∂xj ∂xi D̄t (3.8) The first term between brackets comes from the molecular description of the flow, it is called the Viscous Stress. In Pope [2000] in shown how the Reynolds Stresses are also deducted when calculating the mean momentum transfer. As we saw with the full Navier Stokes equations, a three-dimensional flow is completely described by four equations (assuming incompressible and temperature independent flows): Reynolds equations and the continuity equation. However, in practice, the statistical introduction by Reynolds averaging adds new variables in terms of Reynolds Stresses and the system is underdefined. This problem is known as the turbulence closure problem. 3.1. REYNOLDS EQUATIONS 3.1.2 15 Anisotropy and Tensor Properties In tensor theory the diagonal components u21 = hu1 u1 i , u22 , u23 are called the normal stresses, and the off-diagonal components (ui uj , i 6= j) the shear stresses. When ui uj = ui uj the tensor is called symmetric. Since the stress tensor is coordinate orientation dependant, the principal directions are defined such that the shear stresses are zero. Then, normal stresses coincide with the eigenvalues of the stress tensor matrix, which for physical reasons must be positive (hu1 i ≥ 0). This way we obtain a positive semidefinite tensor. For non-principal directions, isotropic and anisotropic stress definitions are useful. let the turbulent kinetic energy be half of the trace of the Reynolds Stress tensor: k≡ 1 1 hu· ui = hui ui i . 2 2 which represents the mean fluctuating kinetic energy per unit mass. Then, the isotropic stress tensor is obtained as 23 kσij . The difference is the anisotropic part 2 aij ≡ ui uj − kδij . 3 The anisotropic term is many times normalized: ui uj aij 1 = − δij bij = 2k hul ul i 3 we solve for the Reynolds stress tensor 2 ui uj = kδij + aij 3 1 δij + bij . = 2k 3 This last equation helps to understand why the anisotropic component is responsible for momentum transportation. The isotropic part only modifies the pressure term, which is irrelevant for compressible fluids. Application of this ideas is shown in the modified mean pressure equation (3.9). ∂ ui uj ∂aij ∂ hpi ∂ 2 ρ + =ρ + hpi + ρk . (3.9) ∂xi ∂xj ∂xi ∂xj 3 We observe the isotropic 32 k is absorved in the mean pressure term. In irrotational flows the Reynolds Stress tensor produces exclusively a modified pressure. Taking zero mean and fluctuating vorticity, and thus ∂ui /∂xj − ∂uj /∂xi zero, one obtains * !+ ∂uj ∂ui ∂ 1 ∂ ui − = hui ui i − ui uj = 0 ∂xj ∂xi ∂xj 2 ∂xi which leads to the Corrsin-Kistler equation ∂ ∂k ui uj = . ∂xi ∂xj We can appreciate in this equation how the stress tensor has the same effect as the isotropic stress kδij , which can be absorbed into the modified pressure as before. 16 CHAPTER 3. TURBULENCE 3.2 Turbulent Boundary Layer Boundary layers are the location of relevant flow features and are of primary interest in turbomachinery. Study of simplest boundary layers, formed by uniform-velocity flows over a plane plate reveals that statistically can be described by two dimensions. This two-dimensional coordinate system is defined such that the x-coordinate is set in the flow direction, and y-coordinate is perpendicular to the surface. We define, U, V, W as the velocity in the positive x, y, z directions respectively. Boundary layer thickness, δ(x), increases with x, and is generally (but not only) defined as the value of y at which the mean velocity, U (x, y) , equals 99 % of the free-stream velocity, U0 (x). Other definitions are based on integrals, which makes them more reliable for experimental measurements reasons. Some examples are Displacement thickness Z ∞ hU i ∗ 1− δ (x) ≡ dy U0 0 and momentum thickness ∞ Z θ(x) ≡ 0 hU i U0 hU i 1− dy. U0 (3.10) Viscous Scales Viscous scales are some variables deduced in order to accurately describe the boundary layer region of a fluid. We will start describing the total shear stress τ = ρν ∂ hU i − ρ huvi , ∂y (3.11) which can be seen to be made up of a viscous term and the Reynolds Stress tensors. We will further define wall shear stress τw ≡ τ (0). If normalization is applied one arrives to the normalized wall shear stress or skin-friction coefficient. Normalization is made by division with a velocity factor, cf ≡ τ . 1/2ρU02 If non-slip conditions are to be satisfied on the walls (U (x, t) = 0), equation (3.11) simplifies to d hU i τ (y = 0) = τw = ρν . (3.12) dy y=0 For wall shear stress only viscous stress participate. As will be seen through next sections, viscosity plays a central role on near-wall regions and is the reason for viscous scales variables of interest definition. r τw • friction velocity, uτ ≡ ρ r ρ ν = • viscous lengthscale, δν ≡ ν τw uτ • friction Reynolds Number, Reτ ≡ uτ δ δ = ν δν 3.2. TURBULENT BOUNDARY LAYER • viscous lengths or wall units y + ≡ 17 y uτ y = δν ν Notice that y + looks like a Reynolds number, and its magnitude expresses the relative weight of the viscous and turbulent flows. The viscous contribution is, as reasoned before, 100 % at the the wall (y + = 0), 50 % at y + ≈ 10 and less than 10 % when y + = 50. Several layers are defined near the wall. Viscous wall region is defined as y + < 50. Within this region there is a dominant effect of molecular viscosity. When y + > 50 the outer layer starts, where turbulence is the main contribution. Inside the viscous wall region, 3 sublayers can be found: the viscous sublayer for y + < 5, in which the Reynolds shear stress is negligible compared with the viscous stress. The range 5 < y + < 30 is called the buffer layer. And the range 30 < y + < 50 known as the log-law region. As Reynolds number increases, the fraction of the channel dominated by the viscous wall region decreases, since δν /δ varies as Re−1 τ . Boundary layer formation and evolution is described in many fluid mechanics books. When the free fluid stream contacts the surface of the plane plate edge (knonw as the leading edge), a laminar flow region forms at this point and spreads towards the fluid core as the flow extends in the surface. When the Reynolds number (which length parameter must be appropriately defined to account for the laminar layer width) rises up to 106 a transition from laminar to turbulent flow starts in the growing boundary layer. Boundary layer properties strongly change from laminar to turbulent flow. The sketch in figure (3.1) shows this transition. Although a turbulent boundary layer refers to the boundary layer that has undergone this transition, we still find laminar motion in the viscous sublayer. of velocity u against distance y from surface at point x.pdf U0 Boundary layer U=0.99 U0 Transition region δ Turbulent y Laminal Leading edge U Transition point Viscous sublayer x Figure 3.1: Graph of velocity u against distance y from surface at point x. Source: Adapted by the author from Krause et al. [2004] 3.2.1 Momentum Equations and Velocity Profiles Stress and velocity gradients parallel to the wall in boundary layers are much smaller compared to the cross-stream gradients. Considering all velocity terms are zero at the surface the lateral mean momentum equation simplifies to 1 ∂ hpi ∂ v 2 + =0 ρ ∂y ∂y If we integrate this equation between wall and free-stream we obtain D E hpi + ρ v 2 = p0 (x). 18 CHAPTER 3. TURBULENCE 1 1.0 <U>/U0 U/U0 0.8 γ 0.6 τ/τw 0.4 τ/τw 0.2 0 0 0.0 y/δ 1 Figure 3.2: Mean velocity, shear stress and intermittency factor profiles in a zeropressure-gradient boundary layer, Reθ = 8000. Source: Adapted by the author from Klebanoff (1954). 0 1 2 3 y/δx 4 5 6 7 Figure 3.3: Nomalized velocity and shearstress profiles from Blasius solution for the zero-pressure-gradient laminar boundary layer on a flat plate: y is normalized by 1/2 σx ≡ x/Rex = (xvU0 )1/2 . Source: Adapted by the author from Pope [2000] . Since v 2 equals zero at the wall, from previous equation the wall pressure pw (x) equals the free stream pressure p0 (x). A similar deduction for the mean-axial-momentum equation (parallel to the wall) might be done. After corresponding simplifications of the velocity and axial gradient terms, one obtains ∂τ 1 dp0 =− ∂y ρ dx If we consider a zero presure gradient flow (free-stream pressure doesn’t change along x coordinate), we finally get ! ∂ 2 hU i 1 ∂τ =ν = 0. ρ ∂y y=0 ∂y 2 y=0 Which might be integrated from y = 0 to y = ∞. The obtained equation is known as the Kármán’s integral momentum equation, τw = d 2 dθ ρU0 θ = ρU02 . dx dx Where θ is defined by equation (3.10). This relation allows us to understand the influence of the wall shear stress in the momentum thickness. Turbulent experimental results were obtained by Klebanoff [1954] and for laminar flow by Blasius [1908]. Results are shown in figures (3.2) and (3.3). It can be seen turbulent profile is much steeper than laminar mean velocity profile. Lets analyse the turbulent velocity profiles closer. A simple physical description of the flow will consider four velocity laws. Three in the inner layer and one for the outer. In the inner layer we find the viscous sublayer law, the van Driest damping function and the log-law which describes the velocity profile in the viscous sublayer the log-law region and the buffer layer respectively. This three laws together creates what is known as the law of the wall. In the outer layer there is a portion of the log-law starting from the inner layer (generally starts to be applicable for y + > 30), and although is not always mentioned the velocity-defect law is employed for certain big y + values. 3.2. TURBULENT BOUNDARY LAYER 19 We will introduce the universal laws, which means this equations are independent of the flow characteristics such as Reynolds number and, thus, results are limited by the hypothesis imposed. The reader must be aware that more complex descriptions are available in the literature. Nevertheless, this laws provide a good qualitative description, which is the aim of this section. Among the universal-laws mentioned the wake region has been more consistently target for its non-universality. Law of the Wall The law of the wall is the consequence of an adimensional modelling. A flow is completely specified by ρ, ν, δ and uτ parameters. Only two adimensional parameters can be constructed from this variables, so one may write d hU i uτ y y , , = Φ dy y δv δ where Φ is a universal non-dimensional function. The idea behind choosing this adimensional parameters resides in the possibility of neglecting the turbulent δ length scale when considering flows close to the wall (y + < 50) and viscous δv for further regions. As it was mentioned previously: viscosity drives the flow close to the wall, whereas turbulent viscosity is the main source in further regions. In the inner layer (where velocity is determined by viscous scales) function φ(y/δv , y/δ) tends to φ1 y/δv . So, if y + ≡ y/δv and u+ (y + ) ≡ hU i /uτ and for y/δ << 1 then it is posible to express d hU i uτ y = ΦI dy y δv as 1 du+ = + ΦI (y + ) dy + y which after integration computes u+ = fw (y + ) with + Z fw (y ) = 0 y+ (3.13) 1 ΦI y 0 dy 0 . 0 y It has been extensively validated that the function fw is universal for flows with Reynold numbers far from the transition region. • From equation (3.12) and no-slip condition fw (0) = 0 and fw0 (0) = 1. Then, the viscous sublayer is constructed after this results by a Taylor-series expansion. fw (y + ) = y + + O(y +2 ) (3.14) An in detail examination of this Taylor expansion concludes that next non-zero term is of order y +4 . So that equation (3.14) is quite accurate. • As stated before, the log-law region ranges from 30 < y + < 50. In this section of the Inner Layer, equation (3.13) is applicable since the viscous factor y/δv is still dominant. However, ΦI will adopt a constant value, usually expressed as k −1 . the mean velocity gradient is du+ 1 = + + dy ky 20 CHAPTER 3. TURBULENCE which integrates to 1 ln y + + B (3.15) k B is a constant, k is known as the von Kármán constant and, within small variations, typical values are: k = 0.41, B = 5.2. u+ = We have obtained the mean velocity behaviour of the viscous sublayer and the log law region. But the buffer layer remains undetermined. A popular approximation to obtain this region description is the van Driest damping function giving rise to the law with this name. According to the mixing-length hypothesis, which will be introduced in section (3.2.4), the total shear stress is τ (y) ∂ hU i ∂ hU i =ν + νT ρ ∂y ∂y ∂ hU i 2 ∂ hU i 2 . + lm =ν ∂y ∂y + ≡ l /δ Normalizing this equation by the viscous scales and solving for ∂u+ /∂y + , defining lm m v and setting τ /τw unity for the inner layer, the solution yields u+ = f2 (y + ) = Z 0 y+ 2τ /τw h 1+ 1+ + 2 (4τ /τw )(lm ) i1/2 . The mixing length hypothesis is not accurate, but gives an approximation of the real Reynolds + = ky + in both log-law Stresses. According to this hypothesis we might write lm = ky or lm region and the viscous sublayer. However, on the overlapping buffer layer, model consistency + is applied. doesn’t occur unless a damping function to the parameter lm + + + + lm = ky 1 − exp −y /A The term in brackets is the van Driest damping function. A+ = 26 is a standard value. For large y + the damping function tends to unity and the log-law is recovered. For a given k, the specification of A+ determines B. In this case A+ = 26 forces B = 5.3. Last but no least, we find the velocity-defect law. It is applied in the defect layer (outer layer with y/δ > 0.2). In this region flow deviates from the log-law. A second function is slightly then defined, which added to the log-law fw δyν fits the velocity profile in the mentioned layer. This is the wake function w yδ . hU i y Π y = fw + w . (3.16) uτ δν k δ Π is called the wake strength parameter, and its value is flow dependent. The wake function is assumed universal and it is defined to satisfy the normalization condition w(0) = 0 and w(1) = 2. Equation (3.16) is usually expressed as the velocity-defect law, where fw is substituted by the log-law and condition hU iy=δ = U0 is imposed to obtain " # U0 − hU i 1 y y = − ln +Π 2−w uτ k δ δ 3.2. TURBULENT BOUNDARY LAYER 3.2.2 21 Kinetic Energy and Reynolds-Stress Balances The definition of the kinetic energy is 1 E(x, t) ≡ U (x, t)· U (x, t) 2 We can perform a decomposition equivalent to the equation (3.1). E(x, t) = Ē(x, t) + k(x, t). Ē(x, t) is the kinetic energy of the mean flow and K the turbulent kinetic energy. 1 hU i · hU i 2 1 1 k(x, t) ≡ hu· ui = hui ui i 2 2 Ē(x, t) ≡ One might obtain the mean kinetic energy equation from the Reynolds Equation (3.7). D̄Ē + ∇· T̄ = −P − ε̄ D̄t (3.17) where ∂Ui P ≡ − ui uj , ∂xj T̄i ≡ Uj ui uj + hUi i hpi /ρ − 2ν Uj S̄ij ∂ 2 ui uj ε̃ ≡ 2ν S̄ij S̄ij − ν ∂xi ∂xj ! 1 ∂ hUi i ∂ Uj + S̄ij = 2 ∂xj ∂xi (3.18) Similarly, the mean turbulent kinetic energy is obtained after subtracting the Reynolds equations from the Navier Stokes Equation (3.6) D̄k + ∇· T 0 = P − ε D̄t (3.19) where p0 is the fluctuating pressure term Ti0 ≡ 1 ui uj uj + ui p0 /ρ − 2ν uj sij 2 ε ≡ 2ν S̄ij S̄ij where Sij is defined by equation (3.18). This equation can alternatively be written as ∂ 1 D̄k + ui uj uj + ui p0 /ρ = ν∇2 k + P − ε̃. ∂xi 2 D̄t Where p0 = p − hpi. Which in the boundary-layer approximation, the equation reduces to ∂k ∂2k ∂ 1 1 ∂ 0 ∂k + hV i + P − ε̃ + ν 2 − νu· u − νp . 0 = − hU i ∂x ∂y ∂y ∂y 2 ρ ∂y The different terms are, from left to right, the mean flow convection, production, pseudodissipation, viscous diffusion, turbulent convection, and pressure transport. The profiles of the terms are ploted in fig (3.4). 22 CHAPTER 3. TURBULENCE 1.0 gain turbulent convection production 0.5 0.20 viscous diffusion 0.10 viscous diffusion pressure transport 0.00 0.0 turbulent convection pressure transport -0.5 production gain -0.10 mean convection dissipation mean convection dissipation loss loss -0.20 -1.0 0.0 0.2 0.4 0.6 0.8 y/δ 0 1.0 10 20 30 y+ 40 50 Figure 3.4: Turbulent kinetic energy budget in a turbulent boundary layer at Reθ = 1410. Left plot is normalized as function of y. Right plot is normalized by the viscous scales. Source: Adapted by the author from Pope [2000]. According to the figures, the mean-flow convection is negligible in the viscous wall region. In the log-law region, P and ε modules decrease with y. From y + ≈ 40 to y/δ ≈ 0.4 the balance is dominated by production and dissipation. At last, in the outer boundary layer, production becomes small and the balance is between dissipation and the convective transport terms. In a similar fashion to equation (3.19) and (3.17), balance equations for the Reynolds-stresses are obtained from the fluctuating velocity in the Navier Stokes Equations u(x, t). 0=− D̄ ∂ ui uj − ui uj uk + ν∇2 ui uj + Pij + Πij − εij ∂xk D̄t where Pij is the production tensor ∂ hUi i ∂ Uj − uj uk , Pij ≡ − ui uj ∂xk ∂xk (3.20) Πij is the velocity-pressure-gradient tensor 1 Πij ≡ − ρ * ∂p0 ∂p0 ui + uj ∂xj ∂xi + and εij is the dissipation tensor εij ≡ 2ν ∂ui ∂uj ∂xk ∂xk . Data for different terms of the Reynolds Stresses can be found in Pope [2000]. Main properties are reported next. Concerning to the normal-stress balance, since only ∂ hU i /∂y is a significant mean velocity gradient, normal-stress production (equation (3.20)) can be approximated by P11 = 2P = −2 huvi P22 = P33 = 0 ∂ hU i ∂y 3.2. TURBULENT BOUNDARY LAYER 23 Over most boundary layer, P11 is the dominant source term of u2 . Although pressure fluctuation does not play a big role in turbulent kinetic energy balance, it plays a central role in the Reynolds Stress equations. It is responsible for the energy redistribution 2 among the normal 2 turbulent components hu i. Resulting in the dominant sink term for u and source for i 2 2 2 v and w . Last but not least, dissipation will be responsible for the sink of v and w bulk energy. In huvi shear-stress balance, dissipation is negligible. There is an approximate balance between production, P, and the pressure term, Π. Mention that the dissipation term is isotropic in the bulk of the fluid, but anisotropic closer to the wall. 3.2.3 Theory Limitations As early as in the 60s, scientist were interested in the boundary layer development under surface curvature and strong pressure gradient conditions [Patel and Sotiropoulos, 1997]. After few tests it was made clear a strong dependency. Reynold Stresses and boundary layer thickness (δ) are function of curvature radius, Rc . This dependency has deep consequences in turbomachinery applications and consequently several tests were readily performed, specially for airfoil shaped blades. Even today, phenomena such as boundary layer detachment, or boundary layer instability are still under study. Rayleigh criterion is known for boundary layer stabilization under curved surfaces description: in convex curvatures, which has the centre of curvature inside the wall, the angular momentum increases with curvature and an stabilizing process occurs. Opposite solutions are found for concave surfaces, which curvature centre is in the fluid. Production terms are added to Reynolds stresses to account for curvature, but according to Pope [2000], this production terms are an order of magnitude smaller than physical evidences. Curvature of the streamlines produces static pressure gradients through the boundary layer as it was explained in equation (2.1). All this mechanisms for turbulence modelling are still under study. 3.2.4 The Mixing Length Theory As mentioned in section (2.3.2), the mixing length hypothesis is a simple turbulence closure model used in the CFD community. Although its accuracy limitation is a concern in real applications, it gives a rough approximation of the Reynolds Stress values. Lets remember that the idea behind the turbulent viscosity theory is to obtain a model of the form − huvi = vT d hU i dy (3.21) so that the Reynolds Stress in equation (3.8) can be included in the viscosity term. The term νT is called the turbulent viscosity. In the mixing length theory the turbulent viscosity is parametrized by a velocity and a length variables, u∗ and lm respectively. νt = u∗ lm . Next step is to relate this parameters with known flow variables. Since the application of the theory is high Reynolds number flows, and for this flows − huvi ≈ u2τ , from equation (3.21) it is possible to write huvi u2 νt = − dhU i = − dhUτ i . (3.22) dy dy 24 CHAPTER 3. TURBULENCE Further, if this theory will be used within the log-law region where d hU i uτ d y = ky as it was demonstrated in section (3.2.1), we further write u2 νt = − dhUτ i = −uτ ky dy where we identify 1/2 u∗ = uτ = huvi lm = ky Absolute values in the equation above ensures that u∗ is non-negative for all y. This relation is known as Prandtl’s mixing-lenth hypothesis. To sum up ∗ 2 d hU i νt = u lm = lm . dy And in log-law region lm = ky. 3.3 3.3.1 Eddy Viscosity Based Turbulence Modelling Eddy Viscosity Hypothesis According to the Eddy Viscosity Hypothesis, the anisotropy aij ≡ ui uj − 23 kδij is intrinsically determined by the mean velocity gradients ! 2 ∂ hUi i ∂ Uj ui uj − kδij = −νT + 3 ∂xj ∂xi or equivalently aij = −2νT z S̄ij (3.23) where S̄ij is the mean rate-of-strain tensor. In simple shear flows it reduces to huvi = −νT ∂ hU i ∂y as the mixing length theory predicts. Unfortunately, many flows don’t obey this last equation as it has been demonstrated by experimentation in the Axisymmetric contraction test for example. As it was done in section (3.2.4) νT (x, t) is obtained by a velocity, u∗ (x, t), and a length, l∗ (x, t) parameters product νT = u∗ l∗ Based on how the modelling is done different modelling families are found. Algebraic models (such as mixing-length model), which models νT in a simple algebraic way, is called a zero equation model. Where as in two-equation models two transport equations are solved to obtain the turbulent viscous term. Transport equation variables typically are k, ε or ω, giving rise to the k − ε turbulence model or the k − ω turbulence model. It is accepted that the k − ε model performs reasonably well for two-dimensional thin shear flows where small streamline curvature and mean pressure gradients are found. As it could be the case of air surrounding an airfoil. For strong pressure gradients k − ω model gives a more accurate prediction according to bibliography. 3.3. EDDY VISCOSITY BASED TURBULENCE MODELLING 3.3.2 25 Wall Treatment When region in the near-wall is considered, computation of the turbulence models faces some difficulties. Hence, modification of turbulence models is very common. For example, experimental cases have revealed an appropriate Cµ coefficient reduction when y + < 50. In STAR-CCM, this approach is followed and the so called two-layer and wall treatment has been introduced is the code. Different types of wall treatment are available, high, low and ally + wall treatments. The all-wall treatment makes use of both high and low treatments when necessary. This way, high wall treatment is employed in coarse-near wall cells, and low-wall treatment in fine cells. When the wall-cell centroid falls in the buffer region of the boundary layer, the wall treatment is acconditioned so the fluid motion is appropriately resolved. Even though wall formulations is modified when choosing different turbulence models, the standard formulation is as follows: ( + + u+ + lam fory ≤ ym u = (3.24) + + u+ turb fory > ym (3.25) the intersection of the viscous and the fully turbulent regions ym is found by Newtons algorithm. The adimensional values to be employed are yu∗ ν hU i u+ = ∗ u y+ = (3.26) (3.27) although the reference velocity is often related to the wall shear stress u∗ = τw /ρ , in actual practice the reference velocity is derived from a turbulence quantity specific to the particular turbulence model. In STAR-CCM, viscous sublayer is modelled as it was done in section (3.2.1). + u+ lam = y (3.28) On the other hand, the log-law region velocity distribution is: u+ turb = 1 E ln( y + ) k f (3.29) being k = 0.42 and E = 9.0. Roughness function value, f , is unity for smooth walls. Specific modifications to this formulation is explained in each of the turbulence models. 3.3.3 Realizable Two-Layer k − ε The Realizable Two-Layer k − ε model combines the Realizable k − ε model with the two-layer approach. Model coefficients are still the same in all the fluid domain, but enhanced treatment is done near the wall region. In order to understand the Realizable k − ε model, we first first explain the general k − ε modelling. In this model transport equations are solved for turbulent kinetic energy, k, and turbulence dissipation, ε. Specific formulation of parameters is done: for lengthscale (L = 26 CHAPTER 3. TURBULENCE k 3/2 /ε), for timescale (τ = k/ε) and a singular quantity of dimensions νT (k 2 /ε) among other options. With this definitions, lm kind of specifications are not necesary. It is called a complete model. k − ε model evolved for many years, but the roots of the model has always relied on the two transport equations and the algebraic turbulent viscosity model. Two different approaches were taken during transport equations development. In the case of k, the exact transport equation was deduced from the Navier Stokes Equations. In the case of ε viscosity for high Reynolds numbers was neglected. The reason for this last assumption is that the main kinetic energy (flow energy) is coming from the large scale turbulence structures, which are independent of viscosity. It can be said that the equation is rather “empirical”. This empirical development of the equation will induce some constrains in the applicability of the model. The equations reported below are the continuous equations used in STAR-CCM for Realizable k − ε model [CD-ADAPCO]. Basic Transport Equations d dt Z Z ρk(v − vg )· d a = ρk d V + A Z Z h i µt µ+ ∇k· d a + Gk + Gb − ρ (ε − ε0 ) + γ + Sk d V σk A V V (3.30) (3.31) Z A µt µ+ σε d dt Z ∇ε· d a + V Z Z ρε(v − vg )· d a = ρε d V + V A ε ε √ Cε2 ρ (ε − ε0 ) + Sε d V Cε1 Sε + (Cε1 Cε3 Gb ) − k k + νε (3.32) where Sk and Sε are the user-specified source terms, and ε0 is the ambient turbulence value that counteracts turbulence decay. The production Gk is evaluated as: 2 2 Gk = µt S 2 − ρk∇· v − µt (∇· v)2 3√ √3 T S = |S| = 2S : S = 2S : S and S= 1 ∇v + ∇v T 2 (3.33) Two dots operation refers to double dot product of a tensor. The turbulent viscosity is computed as: k2 µt = ρCµ ε although Cµ is not a constant. 1 A0 + As U (∗) kε √ = S :S−W :W Cµ = U (∗) 3.3. EDDY VISCOSITY BASED TURBULENCE MODELLING 27 W is the rotation rate tensor, 1 T W = ∇v − ∇v . 2 Only most important equations are writen here. Other terms and constant values can be found at CD-ADAPCO. wall treatment In k − ε model the wall treatment is employed to set the value u∗ in the law of the wall region and modify Gk and Gε in the near-wall cell. As explained before, the all y + wall treatment tries to mimic the high and low y + formulation. It also gives reasonable results in the buffer layer. For the all-y + formulation, a blending function g is defined in terms of the wall-distance based Reynolds number Rey g = exp − 11 √ ky/v q 1/2 u∗ = gνu/y + (1 − g) Cµ k 2 + 1 ∂u 2 ∗ u Gk = gµt S + (1 − g) ρu + µ u ∂y + Rey = ε= k 3/2 lε (3.34) (3.35) being Rey the mentioned Reynolds number two-layer Realizable k − ε The two-layer Realizable k−ε model always solves the transport equation for k, but algebraically blends ε transport equation with a distance from the wall formulation. More specifically, it parametrizes a length scale function lε = f y, Rey and a turbulent viscosity ratio function, µt /µ = f Rey such that Rey and ε are defined by equations (3.34) and (3.35). The following blending function is used to combine the two-layer formulation with the full two-equation model ! ∗ Rey − Rey 1 λ = 1 + tanh 2 A where Re∗y defines the applicability limit of the two-layer formulation. In STAR-CCM values for this parameters are, Re∗y = 60. Constant A determines the width of the blending function, such that the value λ will be within 1% of the far-field value of a given ∆Rey . Thus, it can be solved ∆Rey A= atanh 0.98 In STAR-CCM+ ∆Rey = 10. Turbulent viscosity from k − ε model is blended in the two-layer model as follows: µt µt = λ µt |k−ε + (1 − λ) µ . µ 2layer 28 CHAPTER 3. TURBULENCE The discretized transport equation for ε is modified and yields ! X X ap n n+1 n n − εp ∆εp + an λ∆εn = λ b − ap εp − an εn + (1 − λ) ap εp ω 2layer n n 3.3.4 Menters SST k − ω k − ω model was developed by Wilcox when he proposed the substitution of the ε equation by a modified variable: ω ≡ ε/k. This leads to an equation similar to ε equation in the standard k − ε model, but an additional term is created (obviously constant values also change). In the differential form of the equation the additional term is 2νT ∇ω· ∇k. (3.36) σω k If this term is neglected we obtain the standard k − ω model. It was found that k − ω model was superior in the viscous near-wall region and in the streamwise pressure gradients. Unfortunately ω treatment in boundaries is problematic: a value for ω has to be given in each boundary, and it turns out the solution is very sensitive to this value. Menter proposed a two-equation model which blends the best properties of both k − ε and k − ω models. This is known as the SST k − ω model. In this model, the omega equation includes a weighting function for the term (3.36). When we consider cells far from the walls the weighting function is set to one, and the k − ε model is employed. When cells are close to the walls, then k − ω model is obtained by setting the weighting function to zero. At last, Menter modified the linear constitutive equation and obtained the model which is employed in STAR-CCM under SST (Menter) model. It is a model widely used in aerospace industry where viscous flows are well resolved and turbulence models in boundary layers extremely important. The basic transport equations from SST k − ω are Basic Transport Equations d dt Z Z ρk d V + ρk v − vg · d a = V Z A Z (µ + σk µt ) ∇k· da + γef f Gk − γ 0 ρβ ∗ fβ (ωk − ω0 k0 ) + Sk d V A V d dt Z Z ρω d V + ρω v − vg · da = V A Z Z 2 2 (µ + σω µt ) ∇ω· da + Gω − ρβfβ ω − ω0 + Dω + Sω d V A V Sk and Sω are the user-specified source terms. k0 and ω0 are the ambient turbulence values in source terms that counteract turbulence decay. γef f is the effective intermittency provided by the Gamma ReTheta Transition model (it is unity if this model is not activated. It basically helps to evaluate the momentum thickness Reynolds number in unstructured messes) and h i γ 0 = min max γef f , 0.1 , 1 . 3.3. EDDY VISCOSITY BASED TURBULENCE MODELLING 29 The production Gω equations is evaluated as " # 2 2 2 2 − ω∇· v Gω = ργ S − ∇· v 3 3 where γ is a blended coefficient of the model and S is the modulus of the mean strain rate tensor, equation (3.33). The cross derivative term Dω , 1 Dω = 2 (1 − F1 ) ρσω2 ∇k· ∇ω. ω Turbulent viscosity is computed as: µt = ρkT Other constants and values might be found at CD-ADAPCO. wall treatment Wall treatment might be employed as in k − ε model. There are high-y + and low-y + wall treatments. In the k − ω model this strategy is used to obtain a reference velocity u∗ , a Gk turbulence production and to compute a specific dissipation ω in wall laws and wall cells. q ∗ u = gνu/y + (1 − g) β ∗ 1/2 k 2 + ∂u 1 ∗ u 2 ρu + Gk = gµt S + (1 − g) µ u ∂y + ∗ 6ν u ω = g 2 + (1 + g) √ ∗ βy β ky with the blending function Rey g = exp − 11 and Rey = √ . ky/ν is the wall-distance Reynolds number introduced before. 30 CHAPTER 3. TURBULENCE Chapter 4 Simulations In this chapter simulation setting up and results are written. In sett up section, information about the geometries, fan meshing, models employed and solver conditions selection is done. In the results section, results obtained after post-processing and data assimilation is done. 4.1 Set-up Geometry Fan surface mesh was available in X B format. The blade has 0.15 m radius, from which approximately 0.12 m belong to the blade and 0.03 m to the hub. The fan has an approximately constant 0.054 m long chord length with a straight blade and an angle of attack close to 9.13 ◦ . The fan speed was set to 300 rpm. Figure 4.1: Axial fan geometry. The black line represents the airfoil plane cross section location Figure 4.2: Blade middle chord length cross section. The simulation domain is defined by a cylinder co-axial to the fan axis. The cylinder radius is of 3 m and the total height of 6 m, centred at the fan. Mesh All simulations have been done under the same meshing conditions. First of all, a surface remeshing was performed, followed by a Prism layer and a Trimmer volume meshing. The surface remeshing basically consists on a re-triangulation of the mesh to make it suitable for Finite Volume Methods. A Prism layer consists on a number of orthogonal cell layers next to walls in order to improve boundary layer resolution. This layers are defined in terms of its thickness, the number of cell layers, the size distribution of the cells and the function used to generate them. It must be combined with another core volume mesh generator such as the Trimmer volume meshing. This is a method employed in STAR-CCM which produces a high quality grid. It combines several mesh attributes, • It is composed mostly by hexahedral meshes with minimal cell skewness 31 32 CHAPTER 4. SIMULATIONS Surface Meshing Property Surface growth rate Fan surface mesh absolute minimum size Fan surface mesh absolute Target size Volume Meshing Property Maximum cell size Number of prism layers Prism layer stretching Prism layer thickness Boundary growth rate Value 1.3 7.5E − 4 m 0.0015 m Value 0.5 5 1.5 0.333 m Very slow Table 4.1: Relevant mesh information in the fan surface and the fluid core • They have a curvature and proximity refinement based upon surface cell size • They are surface quality independent • They can be aligned in any direction based on a user specified cartesian coordinate system. Relevant setting values are gathered in table (4.1). Setting We will expose next the most important modelling settings. In STAR-CCM most settings concerning modelling is under the Continua tab. In table (4.2) the set of models considered under Physics Continua section in STAR-CCM is reported. Let remember in this thesis two simulations have been carried out, one based on k − ε turbulence model, and the other based on k − ω model. k − ε turbulence model k − ω turbulence model Optional Model Cell Quality Remediation Cell Quality Remediation Reynolds-Averaged Navier-Stokes Turbulence Two-layer All y + Wall Treatment All y + Wall Treatment Realizable k − ε two-layer SST (Menter) k − ω Energy Segregated Fluid Temperature Segregated Fluid Temperature Material Ideal Gas Ideal Gas Flow Segregated Flow Segregated Flow Time Implicit Unsteady Implicit Unsteady Table 4.2: Summarise of most relevant Continua Physics section in STAR-CCM simulations Modelling of Region Boundaries allows the specification of the differential equations boundary conditions . We have set four regions: the complete fan, the inlet boundary, the outlet boundary, and the far field that connects inlet and outlet boundaries. 4.1. SET-UP 33 • The Inlet boundary has been set to velocity-inlet type. Other common option is mass flow inlet boundary condition. Velocity normal to the surface is 0.05 m/s. The default Intensity and viscosity ratio was left. • The outlet boundary is a flow split outlet. Which basically will specify all the flow exits through this boundary. It is compatible only with incompressible flows, but for low speed of the flow it is a reasonable assumption. • The far field boundary refers to the cylinder, which is far enough from the fan to have any influence. We set this boundary as a wall type of boundary condition to reduce reversed outflow related errors. Yet, slippery wall conditions are chosen in order to simplify computation effort in this non-relevant area of the field. • The fan is wall type of boundary as well. But in this case non-slippery condition is set. All the system is under constant rotation: Tools Reference framesRotating. A rotating reference frame in STAR-CCM generates a constant flux in de discretized equations. This allows solution for steady-state conditions. Solver and Stopping Criteria Solver conditions are summarised in table (4.3). Time step value has been chosen such that a revolution is done each six time steps. k − ε turbulence model k − ω turbulence model Implicit Unsteady time step=0.033 s time step = 0.033 s 1st order 1st order Wall Distance default default Segregated flow default default Segregated energy fluid und. relax fact. = 0.5 fluid und. relax fact. = 0.5 Turbulence model Under-Relax. fact=0.7 Under-Relax. fact=0.7 Viscosity under-Relax. fact=0.95 Viscosity under-Relax. fact=0.7 Table 4.3: Most relevant Solver values in STAR-CCM simulations Each time step iteration might stop under two criteria: One is the maximum inner iteration number (set to 60) and the second is the error-tolerance criterion which is set to an academical standard value of 10−3 . Some of the parameters are found difficult to converge to this minimum, in which case the value to which converge is set as its minimum. Intermittency and energy variables find the minimum around 0.01 in k − ω model, and in the k − model only energy finds the same difficulty converging asymptotically to 0.005. It has been found easier to work with the k − method rather than the k − ω since last one had convergence difficulties. In both models few iterations were performed with low under-relaxation numbers and smaller time steps since first time steps to overcome convergence errors. 34 CHAPTER 4. SIMULATIONS 4.2 Results The objective behind this simulations is not to make any fan performance assessment, but to compare simulation differences when using two different turbulence models. In order to do this, two averaging have been used: time and space. Scalar fluxes are scalar space averaged values over a constrained plane. This plane is located downstream the fan such that captures velocity and mass flow rate through it. The centre of the plane is at Cartesian coordinates [0, 0, −0.4] with the normal vector parallel to the fan axis. The plane has the same width and height of 0.6 m. Unless explicitly mentioned flow rates (velocity or mass flow rates) are always space averaged over this constrained plane using formulas shown below. Other mean values are obtained after appropriate time averaging over the 30 seconds of flow simulation computed. By appropriate it is meant ignoring initial start up of the flow, where non representative values are found. In order to make the report clear, whenever “mean values” are mentioned during this section, it will refer to the time averaged values of the surface mean variables. When no mean value is mentioned, it will refer to the surface averaged flux for a specific time. • Mass Mass Flow : Mass flow is a type of report system available in STAR-CCM. Mathematically it is measured as follows P f ρf φf v· af P (4.1) ρf vf · af f If mass Mass Flow is chosen, scalar variable φ is substituted with ρf • Velocity Mass Flow : As in the previous case, mass flow averaging (equation 4.1) is used. However, in this case φf is substituted by vf • Force is another type of report available. It is computed as X pressure + ffshear · nf f= ff f where ffpressure and ffshear are the pressure and shear force vectors on the surface face f , and nf is a user-specified direction vector that indicates the direction in which the force should be computed. 4.2.1 Mass Flow Study of Mass flow has been done using equation (4.1). As one could expect, oscillations around a mean value are obtained. This is logical since unsteady simulation are being done. Velocity and mass flow rates are proportional by a constant factor. Conclusions obtained from one flow rate type is directly applicable to the other. A time-averaged mean mass flow rate of 2.15 kg/min, or an equivalent 1.77 m3 /min has been found with k − ε turbulence model. Mean velocity parallel to the fan axis direction is 0.054 m/s. For the k − ω turbulence model, mass flow rate is 2.19 kg/min, and mean velocity 0.055 m/s. We could say equal mean results are obtained with two turbulence models. Oscillations on the other side are different based on which turbulence model is used. When k − ω model is used, maximum mass flow rate maximum oscillation is about 50 %. Whereas in 4.2. RESULTS 35 Figure 4.3: k − ε turbulence model after 20 seconds of flow simulation Figure 4.4: k − ω turbulence model after 20 seconds of flow simulation Figure 4.5: k − ε turbulence model after 30 seconds of flow simulation Figure 4.6: k − ω turbulence model after 30 seconds of flow simulation k − ε turbulence model the maximum oscillation is 44% amplitude. However, both maximum amplitudes coincide in the flow time of the oscillation. Report of the local velocity field for both turbulence models have some differences. In figures (4.3)-(4.6) velocity fields for both turbulence models at two different time intervals are shown. It can be appreciated that for the same maximum and minimum velocity predictions, k − ω model has a wider extension at this values. In another words, velocity distribution variance is bigger for k − ω despite the same maximum / minimum values are found. This might explain why oscillations are bigger in the first model even if mean values are the same. 4.2.2 Force As introduced in this section, force is calculated over the whole fan. In our analysis, only force in the axis direction has been studied. This is, only lift power of the fan is considered, drag force is not included in this thesis. As with flow rate analysis, oscillations over a mean value are found. For k − ε model the 36 CHAPTER 4. SIMULATIONS time averaged force is 0.07 N , and the maximum oscillation is about 28% of the mean value. With k − ω model, average value is the same, 0.07 N , but with a maximum oscillation of 21% amplitude. 4.2.3 Boundary Layer The point of the blade at which boundary layer analysis has been done is shown in figure (4.7). It is located 0.1 m far from the hub centre on the pressure side of the blade: close to the trailing edge and on the chord line centre. This point was chosen because is far from the blade tip, where trailing vortexes exist, and boundary layer is developed as much as possible. Boundary layer velocity is shown in figure (4.8), which is perpendicular to the blade surface. Blade surface normal vector is [0.993, −0.15, 0.0]. We will go through equations in section (3.2.1) doing a comparison of theory and simulations for k − ε model. k − ω model will be studied at the end of this subsection. Data used is summarized below. Wall velocity gradient and the wall stress values have been obtained from simulations. ν = 1.58 10−5 kg/ms ρ = 1.17 kg/m3 τw and dhU i dy τw = 0.07 P a d hU i = 3333.3 m/s. d y y=0 are related by τw = νρ d hU i dy y=0 and from previous values it is computed τw = 0.061 P a, which is close to 0.07. Bear in mind that gradient value has been obtained numerically by the writer from graphical representation and is subject to error. Using τw = 0.070, viscous sublayer ends at y = y + δν = 0.3 mm when y + = 5, which is under the value reported in the simulation (shown in figure (4.8) with a value of 0.45 mm). Hence, simulation and theory show a deviation of 30 % error. In the log-law region similar results are obtained. Solving for hU i in equation (3.15) with y + = 30, 1 + hU i = uτ ln y + 5.2 = 3.3. 0.41 Whereas solution in figure (4.8) shows a value of 2.95 (at a distance y = 2 mm or equivalent y + = 30). The relative error is 13.3 %. Same operations give for y + = 50 a value hU i = 3.65, but 3.15 in STAR-CCM. It is 13.7 % error. It seems analytical boundary layer equations from (3.2) don’t match accurately with predictions from STAR-CCM. Thus, we have tried doing the same calculations with the Standard wall treatment provided by STAR-CCM (equation (3.29) and related). We would expect them to be closer to simulation results. Unexpectedly, results obtained are very similar to mentioned analytical solutions. hU i = 3.33 for y + = 30 and hU i = 3.63 for y + = 50. Measurements at the same blade point at the same simulation time employing k − ω turbulence model gives very close boundary layer values compared to k − ε. For sake of illustration, taking 4.2. RESULTS Figure 4.7: Location of the point used for boundary layer profile study 37 Figure 4.8: Boundary layer velocity magnitude parallel to the wall. It can be appreciated a linear velocity progression up to almost 0.5 mm were the viscous sublayer ends. The right point in the figure is located at y + = 30 point y + = 30 both turbulence models reports 2.95 m/s speed. Moreover, a viscous sublayer size of 0.485 mm is measured. The same as k − ε model. We might summarize that in the boundary layer, there is no difference between using k − ε and k − ω models at the point and properties considered. 38 CHAPTER 4. SIMULATIONS Chapter 5 Conclusions and future work Conclusions will be reported for the three main subsections on (4.2) independently. From Mass Flow subsection we conclude that both k − ε and k − ω turbulence models predict a similar time averaged flow rate. However, for each physical time of the flow, different flow rates are computed. k − ω model seems to overpredict the area extension at which high or low speed flows occur. On the other side k − ε predicts a smoother velocity distribution, but still with similar maximum and minimum values. k − ω also overpredicts the recirculation area behind the fan. An important conclusion from the analysis above is that, if only blade simulation would be done, k − ω would overpredict the blade aerodynamic efficiency (lift capabilities) compared to k − ε. However, if complete fan is taken into account, the gas exhausting power of the fan yields similar results. Mention, understanding the strong oscillations of the flow rates is of big interest. Weighting of the numerical instabilities, simulation setting or fan wrong design contribution to oscillation should be done, despite the physically sensible results obtained in the simulations. Similar conclusions are obtained from Force subsection. While k − ω has strong oscillations in mass flow rate analysis, they are smaller than k − ε for force variable. Both conclusions are not contradictory, they are complementary in fact. As we mentioned already, both turbulence models show similar axial direction velocities in the near-blade (where we find the maximum values). But the bigger force oscillations are bigger in k − ε must strongly damp to obtain a lower flow oscillations. This matches the theory since it is known k − ε is more dissipative, and thus velocity distributions are smoother in the flow field. Most important conclusion from boundary layer subsection is that STAR-CCM overestimates the size of the boundary layer in comparison with the simpler analytical equations reviewed in section (3.2), or even the STAR-CCM standard wall value model. This has been proof by wall distance and mean velocity calculations on the viscous and log law region. This conclusion applies to both turbulence models. Answer to this prediction mismatch is not clear. Differences in simulations and analytical results might be consequence of the different formulations employed in each turbulence model during computation. We saw in section (3.3.2) the normalization velocity is computed from turbulence related variables. Another possibility is the boundary layer model, which was developed for mono-dimensional surface parallel flows, not being representative of our case (we have certain angle of attack altogether with a complex geometry which causes multi-dimensional boundary layers). Another possibility is, according to section (3.2.3), effect of pressure gradients and leading and trailing edge curvatures to affect the streamline curvature and distorts the boundary layer. A pressure plot from the simulations is shown in figure (5.1). Experimental validation should be done to contrast simulations and identify analytical values mismatch reasons. 39 40 CHAPTER 5. CONCLUSIONS AND FUTURE WORK Figure 5.1: Blade pressure side surface pressure distribution. It is appreciated a few Pa overpressure on the leading edge while an under pressure on the trailing edge. Further analysis of the simulation reveals a fast decay in over/underpressure with distance from the blade Bibliography J. Anderson. Computational fluid dynamics: the basics with applications. McGraw-Hill, 1995. H. Blasius. Grenzschichten in Flüssigkeiten mit kleiner Reibung. Annu. Rev. Fluid. Mech, 56: 1–37, 1908. F. Bleier. Fan Handbook: Selection, Application, and Design. McGraw-Hill, 1998. K. Brun and R. Kurz. Analysis of Secondary Flows in Centrifugal Impellers. International Journal of Rotating Machinery, Issue 1:45–52, 2005. CD-ADAPCO. STAR-CCM+ USER GUIDE. STAR-CCM+ Version 6.06.017. J. Earl Logan and R. Roy. Handbook of Turbomachinery. Marcel Dekker, 2003. D. Eckardt. Detailed Flow Investigations Within a High-Speed Centrifugal Compressor Impeller. Journal of Fluids Engineering, Transactions of the ASME, 98 Ser I(3):390–402, 1976. Cited By (since 1996): 94. D. Japikse and N. Baines. Introduction to turbomachinery. Concepts ETI, 1997. P. Klebanoff and U. S. N. A. C. for Aeronautics. Characteristics of Turbulence in a Boundary Layer with Zero Pressure Gradient. Technical note. National Advisory Committee for Aeronautics, 1954. E. Krause, H. Oertel, H. Schlichting, K. Gersten, and C. Mayes. Boundary-Layer Theory. Physics and astronomy online library. Springer, 2004. E. Logan, J. Sherma, and B. Fried. Handbook of Turbomachinery. Chromatographic Science Series. Marcel Dekker Incorporated, 2003. V. Patel and F. Sotiropoulos. Longitudinal curvature effects in turbulent boundary layers. Progress in Aerospace Sciences, 33(1–2):1–70, 1997. S. Pope. Turbulent Flows. Cambridge University Press, 2000. J. Tuzson. Centrifugal Pump Design. A Wiley-Interscience publication. John Wiley & Sons, 2000. 41