RINGSTIFFNESS EVALUATION AND OPTIMIZATION OF STRUCTURED-

WALL PE PIPES

F Fuerle¹, J Sienz¹, M Innocente¹, J.F.T. Pittman¹, V Samaras² and S Thomas²

1
C2EC, School of Engineering, Swansea University, Wales – [email protected]; [email protected]
2
Asset International Ltd., Newport, Wales

Structured-wall high density polyethylene pipes up to 3m diameter are used extensively in civil engineering applications
including storm water attenuation tanks, culverts, surface drainage, inter-process pipe work, sewers etc. The pipes are
manufactured by extruding a hollow box section which is wound in a spiral manner onto a mandrel with successive
turns welded together using PE from an auxiliary extruder. A key quality control measure is the ring stiffness to BS EN
1446: 1996. The ability to predict this accurately as a function of the pipe wall geometry is a pre-requisite for
optimization of the pipe design. The paper describes the development and the validation of finite element modeling of
the ring stiffness test in comparison with experimental results, and considers the relationship between the test results
and in-situ performance. Approaches to optimizing the pipe wall structure are also outlined.

Introduction pipe sample is placed in a compression testing

machine, as shown in Fig. 2.
The structured wall pipes studied within the present
work are produced by winding an extruded high
density polyethylene (HDPE) box profile around a
mandrel and welding each turn to the adjacent one (see
Fig. 1).

Fig. 2 A pipe sample placed in the pipe stiffness

machine.

When compressing the sample between the two

parallel plates, the machine monitors the force (F) that
Fig. 1 The extruded profile is wound around the is necessary to move one of the plates with constant
mandrel and welded to the adjacent one. velocity. According to the standard, the ring flexibility
(SN) is calculated at a vertical deflection equal to 3%
The main objective of the current study is the of the pipe’s inner diameter, as shown in Eq. 1
development of a fast and accurate numerical
 y
F   0.0186  0.025  
simulation of the standardized ring flexibility test

SN   D
according to the standard BS EN 1446: 1996 (for more

yL
information the reader is referred to [1]). This standard Eq. 1
specifies a method to measure the flexibility of a
thermoplastic pipe with a circular cross section. To
conduct such a test, an approximately 1-meter-wide

Proceedings of the Polymer Processing Society 24th Annual Meeting ~ PPS-24 ~ June 15-19, 2008 Salerno (Italy)
where y is the vertical deflection, L is the length of the 4 they are referred to in as “line 1”, “line 3”, and “line
pipe sample, and D is the inner diameter. 2”, respectively.

Future work will look at the improvement of the pipe’s

performance, as well as the minimization of the
expended material. At the present stage the pipe’s
performance testing is restricted to that in the stiffness
test. Thus the pre-requisite for the optimization of the
pipe’s profile is an accurate simulation of that stiffness
test with low computational costs.

The present paper is structured as follows: First, the

developed automated procedure for the generation of
finite element (FE) models from scanned pipe profiles
is discussed. Subsequently, several approaches to FE
simulations are presented by looking at different
element types and various possible boundary
conditions. Those theoretical considerations are then Fig. 3 Rotated profile with highlighted repeating
amended by a convergence study. Furthermore a first pattern.
application of optimization is outlined and finally the
results of the work are presented.

Scan Processing and Automated Finite Element

Model Generation

The theoretical, rectangular profiles of the pipes,

extruded at the very beginning of the production
process – completely defined by the height, width and
wall thicknesses – significantly differ from the cross
sections obtained after production, exhibiting
differences of up to 20%. It should be noticed that the
accurate measurement of the profile’s height is of
utmost importance, since it affects the moment of
inertia with a power of three. Therefore, it was decided
to cut samples of real profiles and scan their precise
cross sections.
The scanning process returns files of discrete data
points, distributed over the entire cross section. These
points have to be processed prior to the creation of the Fig. 4 Surface points separated into 3 bounding lines.
finite element (FE) model. The programs carrying out
that task are written in FORTRAN 77 and the whole The three boundary lines have to be smoothed so that
procedure is controlled by means of a Java tool. The the resulting FE model does not exhibit unreal crooked
latter serves as a Graphical User Interface (GUI) and edges that can cause singularities or stress
simplifies the use of the command-line-type concentrations. To this end, each of the three lines is
FORTRAN programs. Furthermore it stores and replaced by a polynomial, whose parameters are found
manages the entered data. Java was chosen, since it by the minimization of a least square function. That
provides a convenient way for the development of means the difference between existing discrete points
graphics, whereas FORTRAN impresses with its and those from the polynomial is a minimum. The
higher computational speed. arising systems of equations during the determination
The scanning process starts with the rotation of the of the coefficients of the polynomials are solved
profile in a way that all walls are either horizontal or utilizing the open-source collection of mathematical
vertical. After that the repeating pattern can be subroutines LAPACK (see [2]) written in FORTRAN
identified, which is half of one box section. The 77.
identified repeating pattern and the rotated cross The generation of points that lie between the
section are illustrated in Fig. 3. polynomial surface lines, and the definition of their
The next step is the determination which points lie on corresponding thicknesses, are required for the
the boundary of the cross section, since only those will generation of a FE model with shell elements. The
be of use for the model generation. procedure yielding these points starts by dividing the
The surface nodes are now divided into three lines: the profile into an upper and a lower part. Fig. 5 shows an
“top”, the “bottom”, and the “C-shaped” lines. In Fig. example of the procedure for the upper part: First, a

Proceedings of the Polymer Processing Society 24th Annual Meeting ~ PPS-24 ~ June 15-19, 2008 Salerno (Italy)
straight line between points A and B is created. The the GUI. The communication is realized via command
midpoint M is obtained and a straight line files which allow for the execution of all available
perpendicular to segment AB, starting at M, is created. tasks in the program.
The intersections of this line with line 1, line 2 and the The first command file written contains the geometry
left boundary define the points O, P and Q, data and commands for the creation of the material, the
respectively. The shorter of the segments MO and MQ element properties and for the mesh generation.
defines whether the controlling point is O or Q, and the Hypermesh is then called in batch mode, to process the
point N is created as the middle point of the segment command file, and to output a file containing the
defined by the controlling point and P. This last information of the mesh generated. Next, the user
segment defines the thickness corresponding to point developed FORTRAN 77 program, which manages the
N. procedure, reads this output file, applies the boundary
The procedure is repeated using points A and N as well conditions and writes another command file containing
as N and B, and so on, until the desired number of this information. Finally, Hypermesh is called again to
points is generated. process the latter and output the final model to be
solved by Optistruct.
The above procedure is summarized in the flow chart
in Fig. 6. It is characterized by two main features. The
convenience of an automated procedure which allows
for the simulation of the stiffness test of a certain pipe,
is that nothing more than discrete data points of a
scanned cross section and the diameter of the pipe have
to be provided. At the same time the receipt of accurate
geometry data is guaranteed. Summarizing one can say
that a first step towards both accuracy and efficiency of
the simulation has been made.

Fig. 5 Points’ generation and thickness’ determination

for shell model.

An FE model can now be created with solid elements

using the polynomial approximation, or with shell
elements using the points on the profile’s midsurface Fig. 6 Flowchart of the procedure for automated FE
line with the corresponding thicknesses. analyses.
In a first stage of the present work, the commercial
software package Altair Hyperworks (for further In a later stage the entire Hyperworks package has
information see [3]) is used for the remaining pre- been replaced by open-source software or self-
processing, processing and post-processing. The developed programs. FEAP (for further information
engaged tools from Hyperworks are Hypermesh (pre- see [5]) was substituted for Optistruct, Paraview for
processing), Optistruct (for further information see [4]) Hyperview and a self-developed pre-processing tool
(FE solver) and Hyperview (post-processing). for Hypermesh.
Hypermesh incurs the conversion of the point data into
geometry data, as well as the discretization of the Finite Element Analyses
latter.
In order to speed up and automate the entire process, This section discusses the different degrees of accuracy
Hypermesh is accessed in batch mode instead of using and efficiency that can be achieved in the simulation of

Proceedings of the Polymer Processing Society 24th Annual Meeting ~ PPS-24 ~ June 15-19, 2008 Salerno (Italy)
the pipe stiffness test with the FE method. Numerous with higher precision, at the cost of a higher
analyses are conducted in order to discuss and computational effort (refer to Fig. 7).
recommend a model with a suitable combination of Although there are numerous different types of solid
accuracy and efficiency. For the simulations the elements available, only the linear (8 nodes) and
material is modeled with a Young’s Modulus of 1200 quadratic (20 nodes) hexahedral elements and the
MPa and a Poisson ratio of 0.4. linear (6 nodes) and quadratic (15 nodes) pentahedral
The two main aspects to be considered are the type of elements were used in the current work.
element and the use of boundary conditions. The A shell element is usually used for thin structures that
analysis code used within that investigation is present constant thicknesses. Since this is not the case
Optistruct. In a later stage this solver will be partly for the pipe under study, the cross section is divided
replaced by the open-source solver FEAP. Simulations into smaller regions with constant thicknesses
with either one will be of the linear static type. assigned. Fig. 8 shows a shell model where different
Therefore they cannot simulate the moving plate of the colors represent different thicknesses.
pipe stiffness experiment. The constant velocity plate
movement is replaced by a prescribed displacement. On the one hand, the use of shell elements results in
Furthermore, this type of simulation assumes small models with lower numbers of degrees of freedom than
strain theory which is acceptable considering a pipe the use of solid elements. On the other hand, however,
displacement of 3% resulting in strains of at most this introduces further simplifications to the model that
0.5%. might affect the accuracy of the results. While the
accuracy of the global behaviour might not be
Element Types jeopardized by the use of shell elements, the local
behaviour certainly cannot be simulated as detailed as
Two main types of element can be considered to model this is possible with solid elements. This is because the
this problem: “solid elements” and “shell elements”. geometry can not be represented as accurately.
The former are 3D and describe the shape of the profile The shell elements considered in the current work are
more accurately, while the latter are 2½D with a certain the linear triangular (3 nodes) and linear quadrangular
thickness assigned. Once decided upon the use of solid (4 nodes) elements. The corresponding quadratic
or shell elements, different shapes and either linear or elements have 6 and 8 nodes, respectively.
quadratic interpolation are also available. Finally, an
adequate mesh convergence study needs to be carried Boundary Conditions
out.
The definition of the smallest model that can accurately
represent the real problem directly affects the accuracy
and efficiency of the simulation. In this regard, the
exploitation of symmetries and definition of
appropriate boundary conditions are as challenging as
critical.
Clearly, both the geometry and the load state of the
pipe subjected to the ring stiffness test are symmetrical
with respect to all three coordinate planes. Therefore
only one eighth of the whole pipe needs to be modeled,
as shown in Fig. 9.
Fig. 7 Solid model illustration.

Fig. 8 Shell model illustration. Different thicknesses

are defined along the cross section.
Fig. 9 Real repeating pattern (one eighth of the pipe).
A solid element is commonly used for 3D structures
with arbitrary shapes because the latter can be modeled The boundary conditions for this model are quite
straightforward: free surface on the extreme of the pipe
and no displacement on the faces cut by the symmetry

Proceedings of the Polymer Processing Society 24th Annual Meeting ~ PPS-24 ~ June 15-19, 2008 Salerno (Italy)
planes on the directions perpendicular to them or stiffness test. The best that can be aimed for is to
rotations about the axes contained in such planes. develop some artificial boundary conditions for a C-
Finally, the displacements of the nodes in contact with profile – without a real counterpart – that can
the plate must be prescribed. However, in spite of the “represent” rather than “simulate” the real problem.
benefits of profiting from the symmetry planes, the That is to say, this C-profile and its boundary
model is still large and computationally expensive. The conditions would not be simulating any of the C-
geometrically repeating pattern identified in Fig. 3 profiles in Fig. 10. It must be kept in mind that these
makes it interesting to further reduce the model, boundary conditions are unreal.
although the geometrical symmetry is not extended to
the boundary conditions. That is to say, the boundary
conditions are not the same for every isolated
geometrically repeating pattern. The latter, referred to
as the “C profile” from here forth, is shown in Fig. 7.
If the C-profile is to be used to model the ring stiffness
test, the question is what the appropriate boundary
conditions are. A color-map of the displacements in the
“x” direction for the model with the realistic symmetry Fig. 11 Deformed shape of a quarter pipe with
exploitation (i.e. one eighth of the pipe) is shown in horizontal displacements scaled up.
Fig. 10.
It seems self-evident that the ring stiffness of the pipe
in Fig. 9 – with the real boundary conditions as
described before – would be higher than that of the C-
profile marked “A” in Fig. 10 having the restrictive
effect of the rest of the pipe simply removed (i.e.
keeping its left “yz” face unrestrained). Similarly, the
ring stiffness of the whole pipe would be lower than
that of the same C-profile but now having the
restrictive effect of the rest of the pipe replaced by a
perfect restriction (i.e. forcing the x-displacements of
its left “yz” face to zero). Hence the C-profile with the
free border lower-bounds the ring stiffness of the pipe
in Fig. 9, while the C-profile with the restricted border
upper-bounds it. However, the range of ring stiffness
values bounded by these two C-models might be quite
wide. Therefore, some artificial intermediate boundary
condition for the left “yz” face of the C-profile “A”
needs to be conceived if a C-profile is to be used to
model the ring test. The great advantage of this would
be that the C-model is small enough to allow numerous
analyses during the optimization to be carried out in a
Fig. 10 Component “x” of the displacements for the future stage. The model that simulates one eighth of the
real repeating pattern (one eighth of the pipe). pipe could be used to verify the results returned by the
C-model.
As it can be observed, the C-profile closest to the “yz” One way to realize this desired artificial boundary
symmetry plane (marked “A” in Fig. 10) exhibits a condition is to use spring elements as supports at the
behaviour that resembles the one it would have if the x- side, where the higher or lower restrictive effect
displacements of both its faces perpendicular to the “x” desired can be controlled by tuning the springs’
axis were restricted. On the contrary, the C-profile stiffnesses. Thus, the latter can be adjusted in such a
farthest from the “yz” symmetry plane (marked “B” in way that the ring stiffness obtained by using the C-
Fig. 10) effectively has one face unrestricted, and the profile with springs matches that of the full pipe.
overall behaviour resembles the one it would have if its It is important to keep in mind that this model cannot
opposite face was perfectly restricted, except for the be used to analyze the local behaviour of specific C-
additional distortion due to the accumulated profiles, but to efficiently analyze the global behaviour
displacements. The x-displacements scaled up, together such as the calculation of the ring stiffness.
with the corresponding color-map, are shown in Fig. The spring stiffness assigned to every single spring
11. The additional distortion due to the accumulated depends upon the size of the area that every spring
displacements can be clearly seen for the right-most C- stands for. Thus, the concept “lateral pipe stiffness”
profile. (lps) is introduced in Eq. 2, where Sspring is the spring
Therefore, there is no C-profile that can realistically stiffness, nspring is the number of springs, A is the area
model the behaviour of the pipe subjected to the ring that the springs are attached to, Ri and Ro are the inner

Proceedings of the Polymer Processing Society 24th Annual Meeting ~ PPS-24 ~ June 15-19, 2008 Salerno (Italy)
and outer radii of the pipe, and h is its height of the Of course it is to be expected that the most accurate
cross section. simulation would be that of the whole model (i.e. one
This concept allows that, if changes in the geometry or eighth of the pipe). Using this as a reference, the results
discretization are made, the new spring stiffness can be obtained by the C-model with springs approximate the
determined from the constant lps without having to whole model with reasonable accuracy.
tune the former until the C-model results match those Aiming to define the depth of the elements, the

S spring  nspring
of the whole model. analysis of the convergence of the ring stiffness with

lps 
respect to the spin factor – and hence with respect to

Ri  Ro 
Eq. 2 the elements’ depth – was carried out for a 2D element

A
A  h 
size equal to 3 mm.
Eq. 3
2 2 4.00
Convergence curves
Since the pipe should only be stiffened in the x- with respect to element size 3.90
direction, the support of each spring on the extreme 300 spin factor
(element depth = 4.1 mm)
3.80
that is not attached to the pipe should only be restricted

Ring Stiffness [KN/m 2 ]

in that direction. This is a kinematic configuration (i.e. 3.70

not entirely constrained), which cannot be handled by 3.60

linear solvers such as Optistruct. This is overcome by 3.50

restraining the support in all directions, but placing it Whole section

3.40
far away from the pipe so that the components on C section free border

directions other than “x” are comparatively negligible. C section restricted border 3.30
10 9 8 7 6 5 4 3 2 1
Element size [mm]
C section springs border

Convergence Study Fig. 12 Ring stiffness plotted against 2D element size

for various solid models analyzed by Optistruct. The
Two different pipes were analyzed and compared to pipe’s profile is 1500.
experimental data of real stiffness tests. The analyses
included the model consisting of one eighth of the pipe The resulting curves are shown in Fig. 13. Clearly, the
and the C-models with different boundary conditions, ring stiffness is not very sensitive to the spin factor,
as well as different element types and mesh and a value of 200 seems to be sufficiently refined.
refinements. Triangular and tetrahedral elements were Therefore, it is concluded that an element size of 3 mm
avoided. The pipes have diameters of 1500 and 2100 and spin factor of 200 (i.e. an element depth equal to
mm. Most analyses were carried out for the “Pipe 6.14 mm) lead to an appropriate refinement of the
1500”. The “Pipe 2100” was merely used for the mesh for solid and linear elements, considering both
verification of the results. accuracy and efficiency. The obtained simulated ring
stiffness was approximately equal to 3.63 KN/m²,
Due to the axisymmetric geometry of the pipe, a 2D while 3 lab experiments conducted by Asset returned
mesh is generated first, and then the 3D mesh is an average ring stiffness equal to 3.71 KN/m²
created by rotating the 2D mesh 90˚ about the axis of (exhibiting high standard deviation).
symmetry. The number of 3D elements that are created
during the rotation of one 2D element will be called 3.6500

spin factor in following.

The convergence of the ring stiffness with respect to 3.6450

the mesh refinement was studied keeping a spin factor

Ring Stiffness [KN/m 2 ]

equal to 300, which means that 300 3D elements are 3.6400

created during the spin of a single 2D element. This Whole Solid

results in the depth of the elements being equal to 4.1 Springs Solid
3.6350

mm (for the “Pipe 1500”).

The convergence curves are shown in Fig. 12, where it 3.6300

can be observed that the convergence occurs for an

element size of around 2 mm. Nevertheless, since the 3.6250
100 150 200 250 300
Spin factor
use of element sizes of around 3 mm to 4 mm results in
slight variations of the ring stiffness, they are also
acceptable. Fig. 13 Ring stiffness plotted against spin factor for
Note that not only do these curves allow the study of different solid models analyzed by Optistruct. The
the convergence but also the comparison between the pipe’s profile is 1500.
different models. Thus, as anticipated before, the ring
stiffness of the C-model with free boundary lower-
bounds that of the whole model, while the ring stiffness Parallel to using solid elements, simulations using the
of the C-model with restrained boundary upper-bounds more efficient shell elements were also implemented
it. for the whole model, and for the three C-models: free,
restricted, and with springs. The curves of the ring

Proceedings of the Polymer Processing Society 24th Annual Meeting ~ PPS-24 ~ June 15-19, 2008 Salerno (Italy)
stiffness convergence against the element size are interpolation becomes apparent, when applied in
shown in Fig. 14. It appears evident that the returned optimization processes. Here the control points are the
ring stiffness is not sensitive to the refinement of this design variables and not the key nodes. Therefore the
mesh, although the degree, at which the local geometry model is optimized, rather than the FE
behaviour can be investigated, certainly is. That means model. This leads to a smaller number of design
a simulation with a coarser mesh might return the same variables while no accuracy is sacrificed. Furthermore,
ring stiffness as that of a finer mesh, but for instance nodes will not be moved in a way that a different
the stress distribution can not be resolved as accurately. discretization stiffens the system, rather than an
Hence, and given that the computational cost does not improved geometry.
increase significantly for smaller sizes, an element size Before the control values can be calculated the
of 8 mm seems to be a good compromise. segments’ end conditions have to be specified. There
are two possible end conditions. The natural boundary
3.75 condition, which features zero curvature at each end, is
Convergence curves
with respect to element size
3.70
one.
3.65

3.60
Ring Stiffness [KN/m 2 ]

3.55

3.50

Whole section 3.45

C section free border 3.40

C section restricted border
3.35
C section springs border
3.30

3.25
10 9 8 7 6 5 4
Element size [mm]

Fig. 14 Convergence of the ring stiffness with respect

to the element size for different shell models analyzed
by Optistruct. The pipe’s profile is 1500.

It is important to remark that the time required for the Fig. 15 Key points (black) and control points (red).
analysis of a model discretized with shell elements is Two key points and four control points build a sub
only a fraction of the solid counterpart. For instance, segment. The cross section consists of three segments.
the whole model meshed with the recommended shell Each segment is approximated by an individual B-
elements of 8 mm takes less than 15% of the running spline.
time of the same model meshed with solid elements of
3 mm and a spin factor of 200. The advantage of that condition is that no tangent
vectors have to be provided. One should be aware
Optimization though, that this can cause kinks at intersections of
segments. These kinks lead to the so called C(0)
In the optimization procedure, the shell models are connectivity. C(2) connectivity, offering smooth
created by utilizing B-splines (a detailed discussion can transition from one segment to another, can be
be found in [6]). They describe the centerline, as well achieved by choosing the second end condition.
as the varying thickness of the cross section. To this Therefore tangent vectors have to be chosen in such a
end k key points P are generated, as described before, way, that they match at each intersection.
whose x and y coordinates, as well as thickness values Once the control values bi are generated, a B-spline can
are known. be evaluated at an arbitrary position. This position is
The profile is then divided into three segments – two governed by the parameter h. This parameter starts
horizontal ones and one vertical one. Now the variation with 0 at one key point and ends with 1 at the
of the three values (coordinates and thickness) along a following.
segment is described by means of the B-splines, which With the B-spline definition completed, the
are governed by the control values bi. The optimization process can be constructed. The required
corresponding x and y component form a control point elements are a program that can create the control
Bi. Four control points build a sub segment, which is points from discrete key points and thicknesses, as well
bounded by two key points. as evaluate the B-splines to obtain discrete cross
A set of key points (black) and the corresponding sectional points and its corresponding thickness.
control points (red) are depicted in Fig. 15. Additionally, an FE program has to be available. In the
With the control values computed, the splines can be present case this is FEAP, extended by a user
evaluated at arbitrary locations, yielding the developed mesh generation tool. The last part is a
interpolated value for a coordinate component or the software that improves the design variables in order to
thickness. obtain an optimum configuration of those. The
The advantage of the above procedure, in comparison sensititivities of the objective function, in this case the
to using the key points straight away without cross sectional area, and of the constraint, that means

Proceedings of the Polymer Processing Society 24th Annual Meeting ~ PPS-24 ~ June 15-19, 2008 Salerno (Italy)
the pipe stiffness, required for this procedure have to following results within the actual iteration will be
be calculated as well. related to that initial analysis.
The developed optimization procedure is illustrated in To run that initial analysis, the design variables from
Fig. 16 and can be described as follows: In a first step the SQP are read and via the B-spline tool the
the control points of the B-splines are calculated and it geometry is generated, which is converted into the FE
is determined which are used as a design variables. In model. After this is simulation is accomplished, the
case of size optimization only those control values resulting pipe stiffness, as well as the cross sectional
governing the thickness variation are used. In case of area are stored.
shape optimization also some of those describing the x Subsequently nDV analyses are conducted, where nDV is
and y coordinate variation are eligible. the number of design variables. Those analyses
Now the optimization tool, in this case a Sequential distinguish by the design variable which is perturbed at
Quadratic Programming (SQP) tool (the theoretical the time. Comparing the resulting pipe stiffness and
background can be found in [6]), is initialized with the cross sectional area of the actual simulation with the
number of design variables and the number of initial one yields the corresponding sensitivity with
constraints. Furthermore, the initial design variable respect to the active design variable. Naturally after an
values and their boundaries are defined. It returns the analysis with one perturbed design variable is finished,
improved design variables. its pipe stiffness and area of the cross section are stored
and its value is set back to the original one. This is
repeated, until all design variables are processed.
The sensitivity of the objective function and the
constraint with respect to the i-th design variable DVi is
computed according the finite difference method
shown in Eq. 4.

f ( DVi )  f ( DVi    DVi )


DVi
df
dDVi Eq. 4

Here f(DVi) is either the objective function or the

constraint of the initial configuration, whereas f(DVi +
ξ DVi) is the one of a configuration where the i-th
design variable is perturbed by a small factor ξ. This
4
perturbation factor was found to be best as 10 .
Now the values of the objective function and the
constraint as well as the corresponding sensitivities are
written to a file that is read by the SQP. It then returns
a set of improved design variables. The whole
procedure is repeated, until convergence occurs.

So far the above procedure has been applied to the free

C-shell model. The results have been compared to
those obtained by Optistruct which has a built-in
optimization tool. The design variables in either case
were the thicknesses of the shell model. In the case of
FEAP the control points governing the thickness
distribution and in the case of Optistruct the actual
shell element thicknesses.
The resulting optimum thickness variation can be seen
in Fig. 17. It can be seen, that the commercial software
Optistruct and the self developed tool, only consisting
of open-source software, return very similar results.
Note that the minimum thickness was set to 3mm.
Fig. 16 Flowchart of the optimization process. Three
programs, B-spline, FEAP and an SQP optimizer, are Results and Discussion
coupled and can be used for size and shape
optimization. The extensive convergence study of the FE mesh in the
simulation of the stiffness test of the 1500 pipe leads to
Now the FEA part starts. One analysis is conducted the following recommendations. The use of 2D
without changing any of the design variables. All the element sizes of around 3 mm to 4 mm and spin
factors of around 200 (element depth of 6.14 mm) are
recommended for the 1500 diameter pipe. For other

Proceedings of the Polymer Processing Society 24th Annual Meeting ~ PPS-24 ~ June 15-19, 2008 Salerno (Italy)
pipes the spin factor can stay constant, whereas the 2D Test 1 Test 2 Test 3 Mean
element size has to be adjusted in a way that the ratio experiment 3.10 2.62 2.68 2.80
between this and the element depth is around 2.05. shell solid
Thus for the 2100 diameter one, this leads to a value of spring whole spring whole
4.2 mm (element depth of 8.6 mm).
Optistruct 2.58 2.59 2.68 2.58
For shell models, regardless of the pipe’s diameter, an
FEAP 2.58 2.57 n.a. n.a
element size of 8 mm is recommended.
Table 2 Resulting pipe stiffnesses for the 2100
diameter pipe obtained from the solid whole model
9
8 compared with an average pipe stiffness of real tests.
7
thickness [mm]

6
5
4
The element sizes in all listed simulations are the
3 recommended ones. It should be mentioned, that the
2 feap

1 optistruct reason, why the stiffness of the solid-spring model is

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32
higher than its shell counterpart, is that the spring
key node id [-] stiffness was tuned, utilizing the shell model and due to
the significantly coarser mesh in the shell model the
tuning is subjected to a certain inaccuracy. Another
Fig. 17 Optimum thicknesses of the 33 key points,
tuning with the solid model would naturally eliminate
obtained by FEAP (blue) and Optistruct (yellow).
this discrepancy.
Depending on the purpose of a simulation, regarding
The results of a first optimization are illustrated in Fig.
the model type, different alternatives arise from this
18.
study. For instance, if accuracy is being sought and/or
the analysis of local behaviour is of interest, the model
should undoubtedly be the whole profile and the mesh
should contain solid elements.
If efficiency is being quested instead, there are a
number of options that offer different degrees of
efficiency. The C-model with springs is fast and
accurate but requires the tuning of the springs’ stiffness
and it does not represent local behaviour. The whole
model with shell elements is not as efficient as any of
the C-models but it is notably more efficient than the
whole model with solid elements, while it can still
represent local behaviour to some extent (e.g. stresses
in the corners of each box-profile can not be analyzed
accurately, but contrary to the C-models, the entire
pipe is simulated, rather than an artificial substitute that
merely yields the ring stiffness). Finally, the C-model Fig. 18 Shell model. Thickness values of original
with shell elements is the least accurate alternative, but (green) and optimized profile (red).
the time required to obtain the ring stiffness makes it a
candidate, if numerous analyses need to be carried out, The objective was to minimize the area of the pipe’s
for instance for optimization. cross section (leading to less expended material) while
its stiffness remains unchanged. The design variables
The results of the FE analyses using the whole and were the control values of the B-spline that described
spring models, discretized with both, solid and shell the thickness variation. The original thickness values
elements, as well as available experiment data can be are shown in green, while the optimized values are in
found in Table 1 and Table 2 for the 1500 and 2100 red. The blue lines show the true boundary of the
mm pipe, respectively. model. Obviously, it is more efficient to concentrate
the material in the middle of the two horizontal walls,
Test 1 Test 2 Test 3 Mean while the material in the vertical wall does not
experiment 3.74 3.79 3.60 3.71 contribute significantly. This improvement results in a
shell solid reduction of the volume of 15.6%, from 1028.6 cm3 to
867.8 cm3.
spring whole spring whole
The development of the objective function during the
Optistruct 3.57 3.55 3.72 3.63 optimization process, i.e. the area of the cross section
FEAP 3.61 3.62 n.a. n.a of the pipe is depicted in Fig. 19. The development of
Table 1 Resulting pipe stiffnesses for the 1500 the constraint, i.e. the difference between the desired
diameter pipe obtained from the solid whole model pipe stiffness and the computed one can be found in
compared with an average pipe stiffness of real tests. Fig. 20.

Proceedings of the Polymer Processing Society 24th Annual Meeting ~ PPS-24 ~ June 15-19, 2008 Salerno (Italy)
 900
Extensive convergence studies were carried out which
allowed the recommendation of element sizes that
area of cross section [mm2]

850

800
satisfy a sufficient discretization and low
750
computational costs.
700 cr oss sect ar ea

650
Results from the FE analyses could be verified with
600
real pipe stiffness tests. Given the deviation observed
1 2 3 4 5

iteration [-]
6 7 8 9
in those tests, the results can be called sufficiently
accurate. The FE analyses were conducted with
Optistruct and FEAP. Both solvers yielded almost
Fig. 19 Development of the objective function (area of identical solutions.
cross section) with ongoing optimization process. An With those models at hand, an optimization tool, based
optimum solution minimizes the objective function. on the coupling of open-source software, was
developed. Its performance was successfully compared
with the optimization tool built in the commercial
0.03

software Optistruct.
difference between desired and
computed reaction force [kN]

0.025

0.02

0.015 Acknowledgements
0.01

constr ai nt

0.005

0
The first authors gratefully acknowledge the financial
-0.005
1 2 3 4 5 6 7 8 9
support from Asset International Ltd.
iteration [-]

References
Fig. 20 Development of the constraint (difference
between desired pipe stiffness and actual pipe stiffness) 1. British Standards Institution in BS EN 1446 : 1996
with ongoing optimization process. A solution is only –Plastics piping and ducting systems –
accepted to be the optimum, if the constraint is smaller Thermoplastics pipes – Determination of ring
or equal to zero. flexibility; British Standards Institution, London,
1996.
2. Lapack, LAPACK website, http://www.netlib.org/
Conclusions lapack
3. Altair Hyperworks Hypermesh Manual, 2006
FE models with different degrees of accuracy and 4. Altair Hyperworks Optistruct Manual, 2006
efficiency where sought for simulating the pipe 5. Berkeley University, FEAPpv Manual,
stiffness test according to the standard BS EN 1446: http://www.ce.berkeley.edu/~rlt/feappv, 2005
1996. This was realized, and a Java Tool and several 6. E. Hinton; J. Sienz: M. Özakça in Analysis and
FORTRAN 77 programs were developed, that allowed
Optimization of Prismatic and Axisymmetric Shell
for an automated generation of those models from
Structure; Springer-Verlag, London, 2003
scanned pipe profiles, which guaranteed realistic
geometry data.

Proceedings of the Polymer Processing Society 24th Annual Meeting ~ PPS-24 ~ June 15-19, 2008 Salerno (Italy)