Journal of Wind Engineering

and Industrial Aerodynamics 77&78 (1998) 643—651

Computer modelling of flow around bridges

using LES and FEM
R. Panneer Selvam *, Michael J. Tarini , Allan Larsen

BELL 4190, University of Arkansas, Fayetteville, AR 72701, USA

COWI Consulting Engineers and Planners A/S, Parallevej 15, 2800 Lyngby, Denmark

Abstract

An efficient large eddy simulation model (LES) using finite element method (FEM) is
presented. The Navier—Stokes equations are solved using an implicit procedure. The three-
dimensional LES model could capture the drag crisis phenomena for a circular cylinder. Only
2% of the grid point used by Tamura is used to capture the drag crisis. This is possible because
of the use of accurate approximation of convection term using FEM which reduces the
numerical diffusion. This model is used to study the flow over Great Belt East Bridge. The
computed drag and Strouhal number are in reasonable agreement with wind tunnel measure-
ment.  1998 Elsevier Science Ltd. All rights reserved.

Keywords: Computational fluid dynamics; Computational wind engineering; Large eddy

simulation; Turbulence; Bridge loading

1. Introduction

The following issues are to be considered for a safer design of bridges as discussed
by Simiu and Scanlan [1] and Larsen [2]:
Static behaviors: Overturning, excessive lateral deflection, divergence and lateral
buckling.
Dynamic behaviors: vortex shedding excitation, self-exited oscillations or aeroelastic
instability and buffeting by wind turbulence.
Usually the static phenomena are not critical for design of bridges. These issues can
be checked using aerodynamic force components like lift force, drag force and pitching
moment. The dynamic behaviors are critical for design, The vortex shedding excita-
tion may occur at low wind speeds but only occur over a narrow range of wind speeds.

* Corresponding author. E-mail: [email protected].

0167-6105/98/$ — see front matter  1998 Elsevier Science Ltd. All rights reserved.
PII: S 0 1 6 7 - 6 1 0 5 ( 9 8 ) 0 0 1 7 9 - 2
644 R. Panneer Selvam et al./J. Wind Eng. Ind. Aerodyn. 77&78 (1998) 643–651

It can cause oscillations in either vertical flexure or in torsion. It can lead to fatigue
failure. The aeroelastic instability includes phenomenal like vertical instability, tor-
sional instability and coupling of vertical and torsional instability which is called
flutter. If the wind speed is greater than the critical wind speed the aerodynamic
instability develops which leads to failure. Buffeting is caused because of extreme
turbulent wind. This produces critical stresses in the bridge deck.
Currently these issues are considered for bridge design using wind tunnel modelling.
Typically, the wind tunnel test takes 6—8 weeks as discussed by Larsen and Walther
[3]. Due to development in computer technology and numerical modelling it may be
possible to reduce the time using computer modelling. This can be achieved if drag, lift
and moment coefficients are determined for various angles of attack as discussed in
Ref. [3]. Here R , the drag coefficient C , lift coefficient C , moment coefficient C and
  
Strouhal number St are defined as
R "»B/l, C "F /(0.5o»B¼), C "F /(0.5o»B¼),
  V  W
C "M/(0.5o»B¼), St"H/(¹»),

where B is the width, H the height, ¼ the length in the z direction of the bridge, » the
reference velocity, l the kinematic viscosity, F and F the drag and lift forces, M the
V W
moment, ¹ the period of oscillation of the lift forces and o the density. Using C ,

forced horizontal wind response is checked. The bridge cross section is assumed to be
in the xy plane. The flutter derivatives can be derived using these coefficients and from
that the critical wind speed for onset of flutter can be determined [1,3]. The vortex
induced response can be evaluated using St and one of the flutter derivatives.
Onyemelukwe and Bosch [4] used a two dimensional finite difference method and
third-order upwinding without any turbulence to compute the drag and lift coeffi-
cients for a bridge section. Due to lack of computer resources they could not resolve
the region with reasonable grid resolution. Fujiwara et al. [5] used the same technique
for fixed and elastic mounted 1 dof (heave) bridge sections. They used 401;51 grid
and the computed vertical displacements are in agreement for certain sections and for
some the amplitudes are much higher. Lee et al. [6] used k—e turbulence two-
dimensional model to compute C and C . They used QUICK procedure to approx-
 
imate the convective term. Their computed values are in agreement with the measured
values for various angles of attack. Nomura [7] used 2D finite element method and
ALE formulations to compute the forced vibration and the vortex-excited oscillation
cases. They did not compare with any wind tunnel measurements. Bienkiewicz and
Kutz [8] used a two-dimensional inviscid discrete vortex method. Reasonable agree-
ment for the lift and drag coefficients is reported. Larsen and Walther [3] also used
discrete vortex method and computed the flow for fixed and oscillating cylinders.
Their computed C , St and critical velocity for flutter are in reasonable agreement

with wind tunnel measurements. Even though the flow structure is three-dimensional
most of the work is two dimensional. In the discrete vortex method [3,8] the way the
vortex is generated on the boundary needs to be carefully evaluated. A 3D model
takes enormous storage and cpu. The Eulerian grid methods [4—6] suffer from an
error in approximating the convection terms. They some times act as turbulence
 R. Panneer Selvam et al./J. Wind Eng. Ind. Aerodyn. 77&78 (1998) 643–651 645

models and some times dampen the turbulence. Selvam and his group [9,10] used the
finite element procedure to capture the drag crisis for the circular cylinder. They
consider the effect of turbulence using large eddy simulation (LES). When they used
2D models, the predicted C was not in good agreement with wind tunnel measure-

ments. When they used a 3D model they could predict the drag crisis very well [10].
For that work they used 24 860 nodes. This is about 2% of the number of nodes used
by Tamura et al. [11] using finite difference method. This is possible because of the
high accuracy in approximating the convection term. In this work these models will be
used to compute the drag and lift coefficients for bridge sections for an Re of 1;10
and are compared with wind tunnel measurements of Larsen and Walther [3].

2. Computer modelling using large eddy simulation (LES)

2.1. Governing equations

In this work, the LES turbulence model is considered. The two- and three-
dimensional equations for an incompressible fluid using the LES model in general
tensor notation are as follows:

Continuity equation: º "0. (1)

GG
Momentum equation: º #º º "!(p/o#2k/3),
GR H GH G
#[(l#l ) (º #º )], . (2)
 GH HG H
where l "(C h)(S /2) , S "º #º , h"(h h h )  for 3D and (h h )  for
  GH GH GH HG     
2D and k"(l /(C h)).
R I
Empirical constants: C "0.15 for 2D and 0.1 for 3D, and C "0.094,
 I
where º , and p are the mean velocity and pressure, respectively, k is the turbulent
G
kinetic energy, l is the turbulent eddy viscosity, h , h , and h are control volume
   
spacing in the x, y, and z directions and o is the fluid density. Here area or volume of
the element is used for the computation of h. The empirical constants used here are the
values suggested by Murakami et al. [12]. Here a comma represents differentiation,
t represents time and i"1, 2 and 3 mean variables in the x, y and z directions. To
implement higher order approximation of the convection term [9] the following
expression is used in Eq. (2) instead of º º :
H GH
º º !h(º º º ), /2. (3)
H GH H I GH I
Depending upon the values of h different procedures can be implemented. For balance
tensor diffusivity (BTD) scheme h"dt is used; where dt is the time step used in the
integration. For streamline upwind procedure suggested by Brooks and Hughes [13]
h"1/max("º "/dx,"º "/dy,"º "/dz). Here dx, dy and dz are the control volume length
  
in the x, y and z directions. In this computation h"dt is used. This has less numerical
646 R. Panneer Selvam et al./J. Wind Eng. Ind. Aerodyn. 77&78 (1998) 643–651

diffusion than the procedure suggested by Brooks and Hughes [13] compared to
bench mark problems [14].

2.2. Finite element scheme to solve NS equations

The equations are solved using an implicit method suggested in Ref. [9]. The four
step advancement scheme for Eqs. (1) and (2) is as follows:
Step 1: Solve for º from Eq. (2). The diffusion and higher order convection terms
G
are considered implicitly to be in the current time and the first order convection terms
are considered explicitly from the previous time step. The pressure is considered in the
right-hand side of the equation. This set of equations leads to be a symmetric matrix
and the preconditioned conjugate gradient (PCG) procedure is used to solve. The
error in approximating the convection term is almost the same as all the terms are
implicit as reported in Refs. [14,15]. For simplicity here on p/o is considered as p.
Step 2: Get new velocities as º*"º #dt(p, ) where º is not specified.
G G G
Step 3: Solve for pressure from (p, ), "º* /dt.
G GG
Step 4: Correct the velocity for incompressibility: º "º*!dt(p, ) where º is not
G G G G
specified.
Step 2 eliminates the checkerboard pressure field when using equal order interpola-
tion for velocity and pressure in the case of FEM. Implicit treatment of the convective
and diffusive terms eliminates the numerical stability restrictions. In this work the
time step is kept for CFL (Courant-Frederick-Lewis) number less than one. The above
NS equations are approximated by the FEM procedure. The velocity and pressure are
approximated using equal order interpolation. Eight noded brick element is used for
3D and four noded quadrilateral is used for 2D. The details of FEM approximation
are given by Selvam in Ref. [9].

2.3. Solution procedure and convergence criterion

The equations are stored in a compact form as discussed in Ref. [9]. The equations
are solved by preconditioned conjugate gradient (PCG) procedure. To solve the
velocities an underrelaxation factor of 0.7 is used. The iteration is done until the
absolute sum of the residue of the equation reduces to 1;10\ times the number of
nodes for each time step. Usually, the pressure and momentum equations take about
50 and 10 iterations for PCG solution. To run 4950 nodes for a nondimensional time
of 60 takes 21 h in the Sun-Sparcstation 20. For most of the two-dimensional work the
time step may be around 0.002 s for CFL less than one.

2.4. Computational grid, boundary and initial conditions

The computational region and boundary conditions are shown in Fig. 1 for a two
dimensional xy slice. In the z-direction 11 of these slices with total width unity are
considered. The cylinder surface is no slip. Close to the wall the wall boundary
condition is imposed. The surfaces normal to the z-direction free slip boundary
 R. Panneer Selvam et al./J. Wind Eng. Ind. Aerodyn. 77&78 (1998) 643–651 647

Fig. 1. (a) Cross section of the Great Belt East Bridge approach span and (b) solution region and boundary
conditions.

Fig. 2. Grid 1 developed using body-fitted coordinate.

conditions are introduced. The upstream boundary has a uniform velocity of unity in
the x direction and zero in the y and z directions. The outflow boundary is traction
free or normal gradient of the velocities are zero and the side boundaries are slip
boundaries. Two different finite element grids are considered for analysis as shown in
Figs. 2 and 3. The smallest spacing close to the bridge is kept as 0.01B. For the grid in
Fig. 2 a mesh of 91;55 is used. This is formed using body-fitted coordinates. For the
grid in Fig. 3, 2478 nodes and 2352 four-noded quadrilateral elements are used for 2D
and 27 258 nodes and 23 520 eight-noded brick elements are used for 3D. The close up
view is given in Fig. 4. Getting an efficient grid with respect to the flow is important.
This can increase the time step. To start the solution a uniform velocity of one in the
x and zero in the y and z directions are assumed. The time step for computation
ranged around 0.002 s and the computation is done for 25 s. Computation is done for
Re of 10.

3. Results

In this work the Great Belt East Bridge [3] approach span shown in Fig. 1a is
considered for modelling. The variation of drag C and lift C coefficients with time are
 
648 R. Panneer Selvam et al./J. Wind Eng. Ind. Aerodyn. 77&78 (1998) 643–651

Fig. 3. Grid 2 which eliminates the nodes away from the region of bridge flow.

Fig. 4. Close up view of grid 2 in Fig. 3.

plotted in Figs. 5 and 6 for Re of 10 using FEM-2D and the grid in Fig. 3. The data
are plotted using every tenth point because of excessive data (about 12 500). The
computed results using a finite difference method (FDM) and the 2D FEM work are
reported in Table 1. The FDM computation is done using a 91;55 grid. It took
about 84 min in the SUNSparc 20 for 25 s of modelling. The FDM results are in good
agreement with wind tunnel results. Here the convective term is approximated by
combining upwind and central difference. Here 5% of upwind and 95% of central
difference is considered. Hence the error in the approximation has an effect on the
result. For FEM-2D-91;55 grid, h"0.14 is used and for FEM-2D using the grid in
Fig. 3, h"dt is used in Eq. (3). The error in approximating the convection term is less
in FEM-2D-Fig. 3 case than others. Hence the 2D drag is higher than the wind tunnel
work. The computed Strouhal number is in reasonable agreement. Further work is
underway for using the 3D model. When the flow is started as an impulsive flow for
 R. Panneer Selvam et al./J. Wind Eng. Ind. Aerodyn. 77&78 (1998) 643–651 649

Fig. 5. Time variation of drag coefficient for Re of 10 using FEM-2D and grid 2.

Fig. 6. Time variation of lift coefficient for Re of 10 using FEM-2D and grid 2.

Table 1
Comparison of computed results with wind tunnel measure-
ment

Details C St

EB-approach
FEM-2D, Fig. 3 0.348 0.149
FEM-2D, 91;55 0.26 0.129
FDM-2D, 91;55 0.187 0.166—0.199
Wind tunnel 0.19 0.17
650 R. Panneer Selvam et al./J. Wind Eng. Ind. Aerodyn. 77&78 (1998) 643–651

the 3D model the error is very high and the time steps are very small. Work is
underway to take the 2D flow as initial flow and to see the performance.

4. Conclusions

Current status of the computation of flow around bridges is reviewed. The FEM
code using LES turbulence model which could capture the drag crisis for a circular
cylinder [10] is used for 2D and 3D modelling of the flow around the bridges. The
current model used about 24 860 nodes which is about 2% of the nodes used by
Tamura et al. [11] and hence it is an efficient code for computational wind engineer-
ing. Computed drag and lift coefficients using 2D model are reported. The computed
results are in reasonable agreement with wind tunnel results. Further improvements
can be achieved using the 3D model and work is underway. The current modelling
issues are in designing a proper grid which can align with the flow so that a longer
time step can be used.

Acknowledgements

The first author acknowledges the support provided by the Department of Energy
Engineering, Technical University (DTU) of Denmark and the Danish Technical
Research Council under grant no. 5.26.16.31 to perform part of this work as a visiting
professor at the Technical University of Denmark. Help provided by Dr. M. Hansen
from DTU in getting the body fitted grid for the bridge section is greatly appreciated.

References
[1] E. Simiu, R.H. Scanlan, Wind Effects on Structures, 2nd ed., Wiley, NY, 1986.
[2] A. Larsen, Aerodynamics of Large Bridges, Balkema, The Netherlands, 1992.
[3] A. Larsen, J.H. Walther, Aeroelastic analysis of bridge girder sections based on discrete vortex
simulations, J. Wind Eng. Ind. Aerodyn. 67&68 (1997) 253—265.
[4] O. Onyemelukwe, H. Bosch, Numerical simulation of wind flow patterns and wind—induced forces on
bridge deck section models, Proc. 7th US National Wind Engineering Conf., Los Angeles, 27—30 June
1993, pp. 493—501.
[5] A. Fujiwara, H. Kataoka, M. Ito, Numerical simulation of flow field around an oscillating bridge
using finite difference method, J. Wind Eng. Ind. Aerodyn. 46&47 (1993) 567—575.
[6] S. Lee, J.S. Lee, J.D. Kim, Prediction of vortex-induced wind loading on long-span bridges, J. Wind
Eng. Ind. Aerodyn. 67&68 (1997) 267—278.
[7] T. Nomura, A numerical study on vortex-excited oscillations of bluff cylinders, J. Wind Eng. Ind.
Aerodyn. 50 (1993) 75—84.
[8] B. Bienkiewicz, R.F. Kutz, Applying the discrete vortex method to flow about bluff bodies, J. Wind
Eng. Ind. Aerodyn. 36 (1993) 1011—1020.
[9] R.P. Selvam, Finite element modelling of flow around circular cylinder using LES, J. Wind Eng. Ind.
Aerodyn. 67&68 (1997) 129—139.
[10] R.P. Selvam, M.J. Tarini, A. Larsen, Three-dimensional simulation of flow around circular cylinder
using LES and FEM, 2nd European and African Conf. on Wind Engineering, Italy, 1997.
 R. Panneer Selvam et al./J. Wind Eng. Ind. Aerodyn. 77&78 (1998) 643–651 651

[11] T. Tamura et al., On the reliability of two-dimensional simulation for unsteady flows around
a cylinder-type structure, J. Wind Eng. Ind. Aerodyn. 35 (1990) 275—298.
[12] S. Murakami, A. Mochida, On turbulent vortex-shedding flow past 2D square cylinder predicted by
CFD, J. Wind Eng. Ind. Aerodyn. 54&55 (1995) 191—211.
[13] A. Brooks, T.J.R. Hughes, Streamline upwind/Petrov-Galerkin formulations for convection domin-
ated flow with particular emphasis on the incompressible Navier-Stokes equations, Comput. Meth.
Appl. Mech. Eng. 32 (1982) 199—259.
[14] R.P. Selvam, Multidimensional upwinding for finite elements and control volume procedures, Report,
Department of Civil Engineering, University of Arkansas, 1996.
[15] H.M. Leismann, O.E. Frind, A symmetric-matrix time integration scheme for the efficient solution of
advection-dispersion problems, Water Resources Res. 25 (1989) 1133—1139.