You are on page 1of 28

Optimization of Type 4 composite pressure vessels

using genetic algorithms and simulated annealing

V. Alcántara,∗, S.M. Acevesb , E. Ledesmaa , S. Ledesmaa , E. Aguileraa


a Schoolof Engineering Campus Irapuato-Salamanca, University of Guanajuato. Carretera
Salamanca-Valle de Santiago Km 3.5+1.8 Km Comunidad de Palo Blanco, Salamanca,
Gto. ZIP 36885 Mexico.
b Lawrence Livermore National Laboratory, Livermore, CA, USA

Abstract

We propose and demonstrate two methodologies for weight minimization of


Type 4 (plastic lined, fiber wrapped) compressed hydrogen pressure vessels:
genetic algorithms and simulated annealing. We consider 70 MPa vessels with a
safety factor of 2.25, and analyze the vessels with classical laminate theory. We
propose an objective function based on Tsai-Wu criterion, composite thickness,
a safety factor, and a penalization factor. The optimum results are analyzed and
compared by a high resolution finite element model. Computer simulations show
that the proposed methodology produces more efficient designs by reducing the
weight by up to 9.8% and 11.2% when compared to previously published vessel
optimization research.
Keywords: Type 4 vessel, Optimization, Genetic algorithm, Simulated
annealing

Nomenclature

∗ Correspondingauthor
Email address: va.alcantarcamarena@ugto.mx (V. Alcántar )

Preprint submitted to Journal of LATEX Templates March 5, 2017


[A] Extensional stiffness matrix
[B] Coupling stiffness matrix
[D] Bending stiffness matrix
C Constant for Metropolis algorithm
fi , fij Strength parameters
FP Penalization factor
{M } Vector of resultant moments
{N } Vector of resultant forces
n Number of layers
Nx Load per unit of length along x axis
Ny Load per unit of length along y axis
P Internal pressure
Pa Probability of acceptance
[ ]
Q̄ Transformed stiffness matrix
r Outer radius of the vessel boss
R0 Outer radius of the cylindrical section of the vessel
R Outer radius of the head
SV Safety value
T Temperature
tk Thickness of each composite layer
∆E Difference of error
{ 0}
ε Vector of strains in the midplane
{κ} Vector of effects of curvature
ϕ Failure index
θ Winding angle at the cylinder
θ0 Winding angle at the head
σi , σj Stress states
{σ} Vector of stress states

2
1. Introduction

Climate change and environment pollution are two critical problems that can
5 be partially resolved with clean transportation fuels such as hydrogen that can
be produced and consumed without polluting the environment. However, hy-
drogen vehicles demand storage at high pressure (350−700 bar [1]), considerable
increasing cost and technical complexity.
Various recent studies analyze composite pressure vessels for hydrogen stor-
10 age [2–12]. All these recognize the difficulty of reducing the amount of expensive
structural material while maintaining load capacity and safety.
Optimizing a composite pressure vessel demands an objective function. This
function must represent the performance of the composite material, and in some
cases may have some local minima. Optimization methods use this objective
15 function to try to find the best solution. Typically, these methods can be cate-
gorized in two groups: those that use only function evaluations, and those that
additionally require the computation of gradients. Because the mechanical be-
havior of composite materials is very complex, gradient computation may prove
difficult. Consequently, gradient-based methods can be ineffective when multi-
20 ple design variables are involved and may lead to a local minimum making these
methods unreliable. Thus, metaheuristics such as genetic algorithms (GA) or
simulated annealing (SA) may be more efficient for composite pressure vessel
optimization.
Although metaheuristics does not guarantee finding a globally optimum so-
25 lution, they can frequently find better results with less computational effort
than algorithms, iterative methods, or simple heuristics by performing a search
over a large set of possible solutions.
Currently, several intelligent optimization algorithms that simulate the evo-
lution principles of biological computing disciplines, artificial intelligence and
30 immunology have been successfully applied to composite material design. Kim
et al. [13] optimized a Type 3 (aluminum lined, fiber reinforced) vessel sub-
jected to internal pressure using a semi-geodesic path algorithm, a progressive

3
failure analysis, and a modified genetic algorithm. The main objective was re-
ducing vessel weight while avoiding failure. The work was extended in [14] with
35 a combination of a semi-geodesic path algorithm, finite element (FE), and a
genetic algorithm to reduce vessel weight.
Giordano et al. [15] developed an optimization approach using Finite Ele-
ment Method (FEM) to decrease mass while increasing stiffness. Liu et al. [16]
applied the artificial immune system (AIS) method to minimize the weight of a
40 Type 3 vessel used for hydrogen storage. Xu et al. [17] minimized the weight of
a composite hydrogen storage vessel subjected to burst pressure. They proposed
an adaptive genetic algorithm to optimize the vessel. The performance of their
method is also compared with a simple genetic algorithm and a Monte Carlo
optimization method.
45 More recently, Roh et al. [18] addressed the design of Type 4 vessels oper-
ating at 70 MPa. The vessels are wrapped with carbon fiber, and are capable
of storing 1.4 − 5.6 kg of usable hydrogen. Using a safety factor of 2.25, the
vessels were designed for a minimum burst pressure of 158 MPa. The carbon
fiber used in this study was Toray T700S. The fiber-resin composite, with 60%
50 fiber by volume, has a tensile strength of 2550 MPa as specified by the manufac-
turer. The design process starts by determining vessel dimensions, i.e., length,
diameter, and dome shape. Netting analysis is used to estimate the thickness
of the helical and hoop layers in the cylindrical portion of the vessel. Then, 3-D
FE analysis was performed by increasing the thickness of the helical and hoop
55 layers until the fiber stresses in both layers were less than the allowable stresses.
Finally, to reduce carbon fiber utilization, the authors proposed reinforcing the
dome section using doily layers, and a new integrated end-cap design. However,
reducing carbon fiber in the dome section also reduces the amount of material
in the cylindrical section which could result in failure.
60 In this paper we propose the use of genetic algorithm (GA) and simulated
annealing (SA) using the Microsoft Visual Studio Tool Wintempla [19] to opti-
mize the composite weight of two Type 4 pressure vessels. These vessels have
previously been optimized with Netting analysis and FEM [18], however a high

4
resolution finite element model [20, 21] help us to visualize layer by layer be-
65 havior.
In an effort to demonstrate the methodology in a simpler geometry, we start
optimizing a laminated composite under in-plane loads. This composite has a
symmetric configuration where fiber orientations are established along load di-
rections, i.e., 0◦ and 90◦ with respect to the x axis. Two laminate configurations
70 are analyzed, and the results show that GA and SA produce similar solutions.
We later present weight optimization of two configurations of Type 4 pressure
vessels with diameters of 521 mm and 391 mm, and a capacity of 5.6 kg H2 as
reported in [18]. These vessels are designed for 70 MPa working pressure and
2.25 safety factor and have a balanced configuration where fibers follow geodesic
75 paths as in [22].
Composite materials are analyzed with classical laminate theory and an
objective function based on Tsai-Wu failure criterion, layer thicknesses, and
a penalization factor. In order to reduce the weight, the thickness of each
layer is modeled using design variables in a range of 0.127 − 10 mm. Finally,
80 a finite element analysis is performed using the results obtained from vessel
optimization.

2. Methodology

This section describes the optimization methods employed and the param-
eters used. We also establish the variables considered for the structural de-
85 sign, load conditions, and materials. Laminate analysis and Tsai-Wu criterion
are employed to determine the composite structure performance. Constitutive
equations and the objective function are presented.

2.1. Genetic algorithm

A genetic algorithm (GA) is an optimization technique that simulates the


90 phenomenon of natural evolution. In a GA, the design variables of a problem
are encoded as individuals. Each individual is called a chromosome and is

5
represented by a string of symbols, usually (but not necessarily), a binary bit
string, i.e. 01101001. Each position in the string represents a particular feature
of an individual, and the value stored in that position represents how that
95 feature is expressed in the solution [23, 24]. The number of bits in the string is
used to control the resolution of the optimization variables.
GA optimization starts with an initial population of random individuals de-
noting several design vectors. Initial population size is determined by sensitivity
analysis and 200 individuals proved adequate for our case. Larger populations
100 increase computational time without improving performance. Overpopulation
in a GA enables creating a first generation with a large number of solutions,
which are then evaluated and the worst individuals are eliminated. Specifically,
an overpopulation of 50% is used; thus, the actual population is initially 300,
but only 200 survive for the rest of the optimization. Overpopulation improves
105 initial solution and hence reduces execution time. In this paper, the variables
(layer thicknesses) are coded using a 16 bit string resulting in a variable that can
have 65, 536 equally spaced values. Several crossover probabilities were tested
and it was concluded that 0.85 produced accurate results. When this value is
too low, the best individuals are preserved between generations, reducing the
110 creation of new and possibly better individuals. The mutation probability is
0.0001. This parameter should be used moderately as it inserts new genetic
material. Finally, the GA optimization runs for 100 generations.
Each chromosome is subjected to the following operators: selection, repro-
duction and mutation. In the selection process, two parents are chosen to give
115 rise to a new population or set of solutions. A random point called crossover
point is generated; this point creates two sequences of bits in each parent. The
first child is produced by combining the first sequence from one of the parents
and the second sequence from the other. The second child is produced using
the other sequences from the parents that were not used in the first child. From
120 generation to generation, it is very possible that the individuals are very sim-
ilar. To solve this problem, GA introduces mutation which is the flipping of
one random bit (gene). To determine its fitness value, an objective function is

6
used. If the convergence criterion is not satisfied, the algorithm reapplies the
genetic operators creating a new generation. The process continues for several
125 generations until the convergence criterion is satisfied or the maximum number
of generations is reached. Fig. 1 shows the flowchart of the proposed GA.

2.2. Simulated annealing

SA optimization can be compared to the physical process by which a material


changes its state while minimizing its energy. At very high temperature, the
130 atoms of the material have high energy. Slow cooling brings the material to a
highly ordered low energy state. On the other hand, rapid cooling causes defects
inside the material. When an optimization problem is solved with simulated
annealing, the structure of the material represents a codified solution of the
problem, and the temperature is used to determine how and when new solutions
135 are perturbed and accepted.
On the SA algorithm, the substance spends some time at constant temper-
ature while performing a specified number of iterations. Each iteration has two
steps: the solution is perturbed, and the quality of the solution is evaluated.
Once a solution has been perturbed, the new solution must be evaluated to
140 decide whether it is accepted or rejected. In order to perform the error evalu-
ation, an objective function must be defined. In the SA algorithm, the current
solution may be replaced with the new solution with a certain probability called
probability of acceptance. There are no regulations that dictate or limit the im-
plementation of simulated annealing; however, there are some specific criteria
145 that need to be considered for accepting a solution once it has been perturbed.
One criterion is always accepting a solution that has a lower error (objective
function) than the previous solution. It is also possible to accept a new solution
even if it has a larger error than the previous. The Metropolis algorithm (Eq.
(1)), follows the criterion discussed, and is typically used in simulated annealing
150 to compute the probability of acceptance for a perturbed solution [25, 26]. The

7
probability of acceptance is,

 e− C∆E
T ∆E > 0
Pa = , (1)
 1 ∆E ≤ 0

where ∆E is the difference between the solution errors after and before pertur-
bation, T is the current temperature and C is a constant determined using the
method proposed by Ledesma et al. [25].
155 Fig. 2 illustrates the operational sequence of simulated annealing. In this
case, the optimization process begins with a random solution at an initial tem-
perature equal to 100. Then, the simulation iterates (100 times in this case)
at constant temperature by perturbing the current solution and evaluating the
solution error. Computer simulations were performed with different initial tem-
160 perature and numbers of iterations, and the selected values proved be a good
tradeoff between accuracy and computational time. The Metropolis algorithm
is then used to accept or reject the solution. When the number of iterations
is complete, the temperature is reduced, and the simulation iterates at a new
temperature. The simulation stops when the temperature reaches its final value,
165 10−6 .

2.3. Theoretical analysis


The fundamental equation of laminate analysis is
    
 N  A B  ε 0 
=  , (2)
 M  B D  κ 

where {N } and {M } are, respectively, vectors of resultant forces and moments,


{ 0}
ε are deformations in the midplane, {κ} are the effects of curvature, [A] is
170 the extensional stiffness matrix, [D] is the bending stiffness matrix, and [B] is
the coupling stiffness matrix. The presence of [B] induces a coupling between
extension and bending of the laminate. In this study, the coupling matrix
disappears because the laminate has a symmetrical configuration, and the vessel
analysis considers only membrane effects. Therefore, Eq. (2) reduces to,
{ }
{N } = [A] ε0 . (3)

8
175 The extensional stiffness matrix [A] is a function of the total number of layers
[ ]
n, the transformed stiffness matrix Q̄ k , and the thickness of each layer tk ,


n
([ ] )
[A] = Q̄ k tk . (4)
k=1

Finally, using the stress-strain relationships it is possible to calculate the stress


state of each composite layer, {σ}k ,


n
[ ]
{σ}k = Q̄ k {εo }. (5)
k=1

Once the stress states of the composite materials are obtained, it is possible
180 to apply a failure criterion. In this work the Tsai-Wu failure criterion is used
(Eq. (6)), where fi and fij are the strength parameters and the contracted
notation is used where i, j = 1, 2, . . . , 6. In fact, failure of composites occurs
progressively, including matrix cracks, fiber breakage, fiber-matrix de-bonding,
and many more. Much work has been done trying to quantify each of these
185 damage modes, their evolution with respect to load, strain, number of cycles,
etc. But it is difficult to relate all this information quickly to a practical design.
On the other hand, continuum damage mechanics (CDM) simply represents all
these failure modes by the effect they have on behavior of the material, [27].
However, in this work we propose a conservative failure criterion since material
190 performance is not the same once the first failure has been presented.

fi σi + fij σi σj ≤ 1 (6)

This criterion can be written in terms of the failure index, ϕ, which is em-
ployed by several finite element packages [28]. Using the Tsai-Wu criterion,
failure is predicted when,
 √( −1
)2
b b 1
ϕ = − + + ≥ 1, (7)
2a 2a a

where a and b are functions of the stress components and the maximum strength
195 in tension and compression [28].

9
2.4. Objective function
Equation (8), shows the proposed expression for the objective function. Both
optimization approaches (GA and SA) determine optimum values of the design
variables that minimize the value of the objective function. Besides including
200 the thickness tk of the laminate or pressure vessel, the objective function also
includes a safety value SV equal to the desired value of the failure index ϕk
of the composite material layer. SV = 1 for this analysis at the conditions of
maximum load and burst pressure (158 MPa) for the laminate and pressure
vessel optimization, respectively. With this objective function, the minimum
205 composite layer thickness required is calculated on every iteration using the
Tsai-Wu failure criterion [28].

n
objF unction = (|SV − ϕk | + tk + FP ) (8)
k=1

Optimum values of the design variables meet a design constraint: the lam-
inate and vessel must withstand a desired load (minimum force or pressure)
without failure. This constraint enters the objective function through a penalty
210 factor FP . The penalty factor is a criterion that has been inserted into the
optimization algorithms. It functions as a conditional that increases the value
of the objective function when the algorithm determines that the values of the
design variables are not optimal. This factor is initially equal to zero but it in-
creases rapidly as the failure index of any layer exceeds the specified safety value.
215 An increasing penalty factor also increases the objective function, pointing the
optimizer toward composite designs that meet the load demands.

2.5. Parameters of composite structures


We start optimizing a laminated composite subjected to in-plane loads, Nx =
10 MN/m, Ny = 20 MN/m. The laminate under consideration has a symmetric
220 configuration where each ply group above the center line is paired with a similar
ply group below the center line such as [02 /904 /02 ], where the subscript refers
to the number of plies in the ply group. Layer thicknesses are modeled using
design variables in a range from 0.127 to 10 mm, where 0.127 mm is the nominal

10
thickness of a composite fiber strip. Fiber orientations in laminated composites
225 are established along load directions, i.e., 0◦ and 90◦ with respect to the x
axis (Fig. 3). Specifically, two configurations are considered for optimization,
the first has 3 layers and the second has 5. The main objective of analyzing
these two configurations is showing that the two optimization methods produce
similar results.
230 The parameters used for vessel optimization are the same as those reported
in [18]. The vessels are made of a plastic liner wrapped with carbon fiber/epoxy
composite layers. A balanced configuration is considered with the number of
plies in a given θ direction between 0◦ and 90◦ equal to the number of plies in
the −θ direction, assuming that the plies are of the same material and thickness.
235 The cylindrical section of the vessel is reinforced with helical and hoop windings,
but only helical windings extend to the vessel head. The helical winding angle
θ0 at the cylinder and the winding angle θ at the head are calculated as [29],

θ0 = arcsin(r/R0 ), (9)

θ = arcsin(r/R), (10)

where r is the outer radius of the vessel boss, R0 is the outer radius of the
240 cylindrical section of the vessel, and R is the outer radius of the head at any
particular location (Fig. 4). Helical winding at the head follows a geodesic path
along the shortest distance between two points. We assume T700S carbon fiber
for both laminated composites and vessels, and a HDPE liner. Table 1 presents
the material properties.

245 3. Finite element analysis

In this section we describe the finite element analysis of the laminate and
pressure vessel. An ANSYS macro developed in APDL (ANSYS Parametric
Design Language) is created to automatically generate finite element models of
composite pressure vessels.

11
250 Element SOLID191 of ANSYS⃝14.0
R was used to simulate composite mate-
rials. This element is defined by 20 nodes having three degrees of freedom per
node: translation in the nodal x, y and z directions. SOLID191 can be used as
Layered Structural Solid to model layered thick shells or solids. Please refer to
[30] for information on defining layer configuration.
255 Laminate configurations with 3 and 5 layers are analyzed applying equivalent
pressure on side faces to simulate Nx = 10 MN/m and Ny = 20 MN/m loads.
Fiber orientations were established along load directions 0◦ and 90◦ measured
with respect to the x axis, (Fig. 3b). Layer thicknesses were defined according
to the results obtained from the optimization algorithms. The FE models have
260 1600 elements with 5 mm by 5 mm typical element size.
Two composite pressure vessels with an internal volume of 145 L are also
analyzed. These vessels have internal diameters of 521 and 391 mm. Because
hoop layers only apply to the cylindrical section, the vessels are modeled with
a transition zone between cylindrical section and domes, where the laminate
265 has ply drop-off, as shown in Fig. 5. This modeling allows the reduction of
composite thickness. Plies can be gradually terminated from the thick to the
thin part of the element.
We previously developed an ANSYS macro to automatically generate high
resolution finite element composite pressure vessel models [20] based on the num-
270 ber of layers, thickness, orientation, sweep circumferential angle, vessel length,
inside diameter, inlet and rear boss diameter, and number of equal size vol-
umes in the segment. The FE model is generated from a geometric area profile.
Fig. 6 shows the profile and a section of the 3D FE model generated. The
macro calculates the thickness and orientation of the coordinate system of each
275 element. The helical layer orientations at the cylindrical section and domes are
determined from Eqs. (9) and (10), while the hoop layer orientation is 90◦ . As
in the laminate analysis, the thickness of each layer was defined according to
the results obtained from the optimization algorithms. Finally, each pressure
vessel model is solved for three load steps; P1 = 70 MPa, P2 = 115 MPa, and
280 P3 = 158 MPa. Load steps 1, 2 and 3 represent operation at design pressure,

12
at average between minimum burst and design pressure, and at minimum burst
pressure.

4. Results

We show optimal results for the laminates and composite pressure vessels.
285 We also present a profile stress distribution of hoop and helical layer of pressure
vessels, and the failure index for three load steps.

4.1. Laminate optimization

Table 2 shows the results obtained with GA and SA for laminate optimiza-
tion subjected to in-plane loads. Configurations of 3 and 5 layers were analyzed,
290 fiber orientations were set to 0◦ and 90◦ . Due to the manufacturing process,
the layers can be made with strips of pre-impregnated composite with nominal
thickness of 0.127 mm, thus, the results are expressed in terms of the number of
strips required for each composite layer. As can be observed, both algorithms
reach similar solutions. The optimum thickness for 0◦ layers (along x axis) is in
295 the range of 9.4−9.7 mm, and the 90◦ layers have 21.8−22.1 mm optimal thick-
ness. Finite element analysis was performed using the optimal configurations
obtained. Figs. 7 and 8 show the evolution of laminate thickness vs. number of
iterations during genetic algorithm and simulated annealing for 3 and 5 layers,
respectively. As can be seen, both algorithms reach similar solutions, regardless
300 of the number of layers.

4.2. Vessel optimization

Table 3 shows the optimal thicknesses of the helical and hoop layers. It is
observed that the results obtained with GA and SA are very close. Total weight
is compared with two models proposed by Roh [18], and the results indicate
305 that GA and SA produce optimum vessels that are 9.8% (521 mm vessel) and
11.2% (391 mm vessel) lighter that those calculated by Roh [18]. Please notice
that the two models proposed in this paper do not have dome reinforcement.

13
Figures 9 and 10 display the evolution of vessel wall thicknesses vs. number
of iterations during genetic algorithms and simulated annealing for the vessels
310 with 521 mm and 391 mm diameter, respectively. These figures show that
genetic algorithm converges faster than simulated annealing, although SA pro-
duces better (lighter) optimum vessel designs.
We conducted a high resolution finite element analysis of the optimum con-
figuration applying the three load conditions (70, 115, and 158 MPa). Figures
315 11 and 12 show the stress along fiber direction profiles at the helical and hoop
layers determined by FE analysis for load step 3 (158 MPa). As can be seen,
stress decreases from inner to outer hoop layers. These graphs also exhibit that
391 mm diameter vessel has a more uniform distribution of stress in the cylin-
drical section than 521 mm diameter vessel. However, the domes present the
320 same pattern within the design allowable stress.
Figure 13 shows maximum failure indices for each load step of the two com-
posite pressure vessels. This figure shows that both pressure vessels support
the first two load conditions well within the limits established by the Tsai-Wu
failure index. Failure is then approached at the third load step (minimum burst
325 pressure).

5. Conclusions

This work has described the optimization of two laminated composites and
two models of Type 4 (plastic lined) composite pressure vessels based on two
methods: genetic algorithm (GA) and simulated annealing (SA). Optimization
330 uses an objective function based on Tsai-Wu failure criterion, layer thickness,
a penalization factor, and a safety value used to establish the failure index
of the composite layers. The laminate is subjected to in-plane loads, and two
configurations were analyzed. The results indicate that GA and SA reach similar
optimal solutions regardless of the number of layers.
335 Type 4 composite pressure vessels are then optimized. An ANSYS macro was
created for rapid development of composite pressure vessel finite element models.

14
This macro allows defining the number of layers, thicknesses, orientations and
sweep circumferential angle, and it has the capacity to gradually reduce the
thickness of the hoop layers at the transition between the cylindrical and dome
340 sections of the vessel. Two pressure vessels were modeled with 70 MPa design
pressure and 2.25 safety factor. The optimization process was performed using
GA and SA employing laminate analysis. The results show a weight reduction
of up to 11.2% vs. previously published results [18].
It is finally observed that both approaches were successful in producing sim-
345 ilarly high quality optima despite the challenge of composite vessel optimization
characterized by the presence of many local minima. Theoretically, GA and SA
are close relatives, and many of their differences are superficial. Basically, SA
can be seen as GA with a population size of only one. The current solution is
the only individual in the population. Since there is only one individual, there is
350 no crossover, only mutation. While the purpose of this work was demonstrating
the methodologies rather than conducting a detailed performance comparison
between the two, it is observed that GA converges to the optimum with less it-
erations. This is because SA creates new solutions by introducing perturbations
on a single solution, while the genetic algorithm creates multiple solutions by
355 combining individuals. It is also observed that SA reached a better optimum
(lower vessel weight) because the presence of many local minima in composite
optimization affects more the ability of the GA to reach a better optimum, possi-
bly due to the advantage of perturbing the existing solutions (SA) vs. combining
potentially similar individuals (GA). It is, however, possible that a different pa-
360 rameter set (e.g. higher mutation rate in GA) may change the relative quality
of the optima in favor of GA.

References

[1] Hua T, Ahluwalia R, Peng J-K, Kromer M, Lasher S, Mckenney, et


al. Technical assessment of compressed hydrogen stoage tank systems

15
365 for automotive applications. International Journal of Hydrogen Energy
2011;36(4):3037–3049.

[2] Aceves S, Martinez-Frias J, Garcia-Villazana O. Analytical and experimen-


tal evaluation of insulated pressure vessels for cryogenic hydrogen storage.
International Journal of Hydrogen Energy 2000;25(11):1075–1085.

370 [3] Chapelle D, Perreux D. Optimal design of a type 3 hydrogen vessel: Part
I–Analytic modelling of the cylindrical section. International Journal of
Hydrogen Energy 2006;31(5):627–638.

[4] Hu J, Chen J, Sundararaman S, Chandrashekhara K, Chernicoff W. Analy-


sis of composite hydrogen storage cylinders subjected to localized flame im-
375 pingements. International Journal of Hydrogen Energy 2008;33(11):2738–
2746.

[5] Onder A, Sayman O, Dogan T, Tarakcioglu N. Burst failure load of com-


posite pressure vessels. Composite Structures 2009;89(1):159–166.

[6] Xu P, Zheng J, Liu P. Finite element analysis of burst pressure of composite


380 hydrogen storage vessels. Materials & Design 2009;30(7):2295–2301.

[7] Hu J, Chandrashekhara K. Fracture analysis of hydrogen storage composite


cylinders with liner crack accounting for autofrettage effect. International
Journal of Hydrogen Energy 2009;34(8):3425–3435.

[8] Camara S, Bunsell AR, Thionnet A, Allen DH. Determination of lifetime


385 probabilities of carbon fibre composite plates and pressure vessels for hydro-
gen storage. International Journal of Hydrogen Energy 2011;36(10):6031–
6038.

[9] Son DS, Chang SH. Evaluation of modeling techniques for a type III hy-
drogen pressure vessel (70 MPa) made of an aluminum liner and a thick
390 carbon/epoxy composite for fuel cell vehicles. International Journal of Hy-
drogen Energy 2012;37(3):2353–2369.

16
[10] Liu P, Chu J, Hou S, Zheng J. Micromechanical damage modeling and
multiscale progressive failure analysis of composite pressure vessel. Com-
putational Materials Science 2012;60:137–148.

395 [11] Leh D, Saffré P, Francescato P, Arrieux R. Multi-sequence dome lay-up


simulations for hydrogen hyper-bar composite pressure vessels. Composites
Part A: Applied Science and Manufacturing 2013;52:106–117.

[12] Liu, P. F., Chu, J. K., Hou, S. J., Xu, P., Zheng, J. Y. Numerical simulation
and optimal design for composite high-pressure hydrogen storage vessel: A
400 review. Renewable and Sustainable Energy Reviews 2012, 16(4), 1817-1827.

[13] Kim CU, Hong CS, Kim CG, Kim JY. Optimal design of filament wound
type 3 tanks under internal pressure using a modified genetic algorithm.
Composite Structures 2005;71(1):16–25.

[14] Kim CU, Kang JH, Hong CS, Kim CG. Optimal design of filament wound
405 structures under internal pressure based on the semi-geodesic path algo-
rithm. Composite structures 2005;67(4):443–452.

[15] Tomassetti G, Barboni R, Benedetti M. Optimisation methodology for cry-


otanks. Computers & structures 2005;83(28):2293–2305.

[16] Liu P, Xu P, Zheng J. Artificial immune system for optimal design


410 of composite hydrogen storage vessel. Computational Materials Science
2009;47(1):261–267.

[17] Xu P, Zheng J, Chen H, Liu P. Optimal design of high pressure hydrogen


storage vessel using an adaptive genetic algorithm. International Journal
of Hydrogen Energy 2010;35(7):2840–2846.

415 [18] Roh H, Hua T, Ahluwalia R. Optimization of carbon fiber usage in Type 4
hydrogen storage tanks for fuel cell automobiles. International Journal of
Hydrogen Energy 2013;38(29):12795–12802.

17
[19] Microsoft Visual Studio Tool: Wintempla,
https://visualstudiogallery.msdn.microsoft.com/4dfcaf79-9f69-4fea-87b6-
420 ec222f1d0fdd, (last access: October 2014).

[20] Ledesma E, Aceves SM, Espinoza F, Ravani B. Development of a Toolkit


for the Analysis and Design of Composite Pressure Vessels. ASME 2012
International Mechanical Engineering Congress and Exposition 2012:187–
198.

425 [21] Espinosa-Loza F. J. Analysis and Design of Cryogenic Pressure Vessels


for Automotive Hydrogen Storage. PhD Thesis: University of California,
Davis; 2014.

[22] Zickel J. Isotensoid pressure vessels. ARS J 1962;32:950–1.

[23] Jones MT. Artificial intelligence application programming. Dreamtech


430 Press; 2006.

[24] Gen M, Cheng R. Genetic algorithms and engineering optimization. 7th ed.
John Wiley & Sons; 2000.

[25] Ledesma S, Aviña G, Sanchez R. Practical considerations for simulated an-


nealing implementation. In: CT Ming, editor. Simulated Annealing, InTech
435 2008;20:401–420.

[26] Ledesma S, Ruiz J, Garcia G. Simulated Annealing Evolution. Simulated


Annealing–Advances, Applications and Hybridizations, Intech Publishing
2012;201–218.

[27] Liu, P. F., & Zheng, J. Y. Progressive failure analysis of carbon fiber/epoxy
440 composite laminates using continuum damage mechanics. Materials Science
and Engineering 2008: A, 485(1), 711–717.

[28] Barbero EJ. Finite element analysis of composite materials. CRC Press;
2007.

18
[29] Liang CC, Chen HW, Wang CH. Optimum design of dome contour for
445 filament-wound composite pressure vessels based on a shape factor. Com-
posite Structures 2002;58(4):469–482.

[30] ANSYS⃝Academic
R Research, Release 14.0, Help System, Element Refer-
ence, ANSYS, Inc. 2014.

19
Start

Population Initialization
(50% overpopulation)
Each individual represents a
solution for the design variables.

Fitness evaluation
Solution and objective
function evaluation.

Select the best individuals


to create the initial population
Best solutions are selected.

Selection

Modify
composite Crossover
thickness.

Mutation

Fitness evaluation
Solution and objective
function evaluation.

Reached
No
optimal
thickness?
Reached optimal solution
or maximum number
of generations?
Yes

Finish

Figure 1: Flowchart for the genetic algorithm as applied to composite optimization.

20
Start

Create a random
composite thickness Perturb the current solution
T = T0 (modify composite thickness)

Solution and O.F. evaluation Solution and objective


function evaluation
iteration = 1
iteration > max_iter?
iteration + 1 No
Yes
Cool Down No Accept the
T = TC new solution?

Yes
Reached
No optimal thickness The perturbed solution
or final becomes the current solution
temperature?

Yes
End Reached
No
optimal
thickness?

Yes
End

Figure 2: Flowchart for simulated annealing as applied to composite optimization.

Ny

0° Nx
90° Nx

Ny
a) b)
Figure 3: a) Schematic of a symmetric laminated composite of three layers, b) Laminated
composite in a 2-dimensional state of charge with axial loads per unit of length Nx and Ny .

21
Fiber
path θ0

R
θ R0
r

Figure 4: Geodesic fiber path along a pressure vessel surface.

Composite

Transition zone
Liner
Figure 5: Transition zone between cylindrical section and domes where ply drop-off takes
place.

22
Profile

FE Model

Figure 6: Profile area and a section of the 3D FE pressure vessel model.

40.0
GA
SA

37.7
thickness, mm

35.5

33.4

31.5
0 45 90 135 180
Iterations

Figure 7: Laminate thickness vs. number of iterations during genetic algorithm (GA) and
simulated annealing (SA) laminate optimization (3 layers).

23
55.0
GA
SA
thickness, mm 47.7

41.4

36.0

31.2
0 63 125 188 250
Iterations

Figure 8: Laminate thickness vs. number of iterations during genetic algorithm (GA) and
simulated annealing (SA) laminate optimization (5 layers).

46.8
GA
SA

46.0
thickness, mm

45.2

44.5

43.7
0 23 46 69 92
Iterations

Figure 9: Vessel wall thicknesses vs. number of iterations during genetic algorithm (GA) and
simulated annealing (SA) optimization for 521 mm diameter vessel.

24
39.0
GA
SA

37.3
thickness, mm

35.7

34.2

32.7
0 24 48 71 95
Iterations

Figure 10: Vessel wall thicknesses vs. number of iterations during genetic algorithm (GA)
and simulated annealing (SA) optimization for 391 mm diameter vessel.
Stress along fiber direction, GPa

inner layers
2

external layers

0 Hoop Helical
x

300 400 500 600 700 800


Position along the vessel length, mm

Figure 11: Stress along fiber direction in hoop and helical layers of 521 mm diameter vessel
for load step 3 The x-coordinate is measured along the vessel length starting from the middle
of the vessel.

25
Stress along fiber direction, GPa
inner layers
2

1
external layers

0 Hoop Helical
x

700 900 1100 1300


Position along the vessel length, mm

Figure 12: Stress along fiber direction in hoop and helical layers of 391 mm diameter vessel
for load step 3 The x-coordinate is measured along the vessel length starting from the middle
of the vessel.

1.0

0.75
ϕ
Failure Index,

0.5

0.25

0
1 2 3
Load Step

Figure 13: Failure index (Eq. (8)) as a function of load step for both composite pressure
vessels. Load steps 1, 2, 3 represent operation at design pressure (70 MPa), at average between
burst and design pressure (115 MPa), and at minimum burst pressure (158 MPa).

26
Table 1: Material properties for the T700S and HDPE.
Parameters Properties Unit Values
T700S properties
Young moduli E1 , E2 , E3 GPa 135, 9.66, 9.66
Shear moduli G12 , G13 , G23 GPa 5.86, 5.86, 3.46
Poisson’s ratios ν12 , ν13 , ν23 0.25, 0.25, 0.41
Tensile strengths F1t , F1c MPa 2550, 1470
Compressive strengths F2t , F2c MPa 69, 224.1
Shear strength F6 MPa 89.63
HDPE properties
Young modulus E MPa 900
Poisson’s ratio ν 0.46
Yield strength Sy MPa 26

Table 2: Optimum laminate thicknesses obtained from genetic algorithm (GA) and simulated
annealing (SA).
Thickness, mm
Method Configuration
x direction y direction
0◦38 /90◦174 /0◦38 9.7 22.1
GA
0◦12 /90◦87 /0◦50 /90◦87 /0◦12 9.4 22.1
0◦38 /90◦172 /0◦38 9.7 21.8
SA
0◦37 /90◦86 /0◦2 /90◦86 /0◦37 9.7 21.8

27
Table 3: Optimum vessel wall thicknesses obtained from genetic algorithm (GA) and simulated
annealing (SA).
Composite Thickness, mm Weight, kg
ID, mm Method % reduction
Helical Hoop Total Ref. [18] This work Ref. [18]
GA 14.2 29.7 43.9 93.8 9.5
521 45.9 103.6
SA 14.1 29.7 43.8 93.4 9.8
GA 10.9 23.0 33.9 97.6 9.1
391 36.1 107.4 (a)
SA 10.6 22.2 32.8 95.4 11.2
GA 10.9 23.0 33.9 97.6 4.3
391 36.1 102.0 (b)
SA 10.6 22.2 32.8 95.4 6.5

28

You might also like