Professional Documents
Culture Documents
www.elsevier.com/locate/compstruct
Abstract
Moving from a validated finite element model of composite cylindrical absorbers, this work aims to optimise the shape of conical
absorbers with elliptical cross-sections considering simultaneously different impact conditions. Since the use of non-linear finite element
analyses to directly evaluate objectives and constraints during the optimisations would be unaffordable from a computational
standpoint, a global approximation strategy is used. The crash capabilities of the absorbers are approximated with a system of Radial
Basis Functions built by means of a minimum number of finite element analyses. The response surfaces are coupled with Genetic
Algorithms to perform both constrained single- and multi-objective optimisations. The results prove that moderate eccentricity and
conicity lead to high efficiency structures characterised by stable crush fronts and good absorption capabilities with also associated mass
reduction up to the 7% considering vertical impacts and at least of the 20% considering 20° impacts with respect to ideal cylinders.
Ó 2004 Elsevier Ltd. All rights reserved.
This work moves from the consideration that, now- velocity of 8 m/s is used. The vertical acceleration of the
adays, there are numerical tools able to predict with drop mass is measured and the impact loads are evalu-
enough accuracy the crash behaviour of composite ated. The values of the peak force as well as of the mean
materials and aims to develop a flexible design tool able crush force are determined after CFC-180 filtering.
to perform crashworthiness optimisations accounting The cylinders tested have height of 300 mm and a
for the shape of the absorbers. Indeed, Mahdi et al. [14] nominal internal diameter of 70 mm. The material used
in the 2002 experimentally investigated the effects of the is the Cytec Fiberite T410717-950-42%-3KHS-5H-283-
shape of composite absorbers showing that conical 1000 with a volume fraction of the resin of 42%, typical
shapes can guarantee more stable crash without losing in aircraft construction. The stacking sequence consists
absorbed energy per unit mass. Considering metal in four [0°/90°]S oriented layers. The crash behaviour
absorbers [15], similar results were obtained by Chian- experimentally observed is stable and the experimental
dussi and Avalle in 2002. The authors proposed a shape tests reliable and repeatable. Accordingly, the data col-
optimisation of cylindrical absorbers with a conical part lected seem to be suited to develop the numerical model.
in order to reduce acceleration peaks during vertical The final finite element model consists in 4094 shell
crashes. Conical shapes seem to be a good compromise elements having characteristic length of 3 mm. Impact
solution to increase the collapse stability without velocity, impact-mass and boundary conditions are
meaningful penalties on the absorbed energy per unit carefully considered and reproduced according to the
mass. experimental tests. An initial trigger is also realised
In this work, the shape of cylindrical and truncated selectively decreasing of about one-fifth the thickness of
conical absorbers, made of composite fabric, is opti- some elements in the first two shell rings.
mised considering simultaneously different impact con- The material model used, MAT 58 [13,16,17], is
ditions: 0° (vertical impacts), 20° and 30°. The work is especially developed for laminated composite material:
carried out using a validate and reliable numerical an elastic damage model developed around the idea that
model able to correctly predict the crash behaviour of the deformations introduce micro-cracks and cavities
composite energy absorbers. In particular, the numerical into material and these defects cause primarily stiffness
analyses are performed using LSTC LS-Dyna 960 degradation with rather small permanent deformation
[16,17], an explicit Finite Element code, which has been unless material undergoes rather high loading and is not
demonstrated to be a valid tool for developing com- close to deterioration. A non-smooth failure surface is
posite structures. considered and to allow an almost uncoupled failure of
As a matter of fact, the use of non-linear finite ele- an arbitrary composite, all failure criteria are taken to
ment analyses to directly evaluate the objective and the be independent of each other. Three independent dam-
constraint values during an optimisation run would be age parameters are hence defined with regard to the part
unaffordable from a computational point of view. To of the undamaged cross-section that is still able to carry
overcome these difficulties, the optimisation procedure is load. The parameters which characterise the material
here based on a global approximation strategy, where model and the initial/boundary conditions, after being
the crash capabilities of the absorbers are approximated validated with respect to the experimental data, are used
by a system of Radial Basis Functions (RBF) built by in the following optimisation phase.
means of minimum number of non-linear finite element
analyses. The response surfaces are finally coupled with
enhanced Genetic Algorithms (GA) to carry out both
constrained single- and multi-objective optimisations 25
Experimental test
[18,19]. Numerical model
20
Crash load [kN]
Table 1 Table 2
Numerical-experimental correlation Optimisation domain
Experimental Numerical % Error Minimum Maximum
test model value value
Mean crash force (kN) 15.737 16.034 1.88 Lower edge major axis (mm) X (1) 60 110
Absorbed energy (kJ) 3.298 3.294 0.12 Lower edge eccentricity X (2) 0.0 0.7
Taper ratio X (3) 0.5 1.0
work, several optimisation runs considering different the normalized domain. An ad hoc algorithm is devel-
problem formulations and constraints are considered. oped to achieve this objective [25].
Therefore, since the finite element analyses performed to The sample points used to built the response surface
evaluate the crash behaviour of each single absorber are obtained via finite element analyses. The load–time
configuration last more than 10 h on a PC Pentium 4, curves obtained by each analysis are at first approxi-
1.5 GHz 500 Mb RAM memory, the use of GAs seems mated with bi-linear laws having equivalent absorbed
to be unaffordable at least from a computational point energy. Each bi-linear curve is identified with the posi-
of view. tion of the first slope-change point (load and time val-
The computational difficulty is overcome using global ues) and with the load value corresponding to the final
approximation and response surface methodologies. shortening of 280 mm. The adopted bi-linear approxi-
Indeed, during the optimisations, the computational mation is shown in Fig. 3 considering the 20° impact
effort is significantly reduced by replacing the finite conditions for two different absorber shapes.
element analyses with a system of response surfaces built The total number of sample points is limited to 30.
with RBF method in order to evaluate the crash load– The response surfaces are preliminary defined using only
time curves. 20 of the 30 available sample points. In this way, the
The RBF method [23,24] is an interpolating scheme, remaining 10 points are used as verification points to
originally developed by Hardy, used to describe the evaluate the interpolation capabilities and the accuracy
behaviour of non-linear functions once known a set of N level of the response surfaces. Since three different im-
initial sample points with coordinates ~xn and values Fn . pact conditions, vertical impact, 20° impact and 30°
The approximation is provided by a linear combination impact, are considered, 90 finite element analyses are
of radial functions each one centred in one of the initial performed in all. Indeed, a total number of nine different
sample points. Therefore, the distance between the response surfaces are built. In fact, three distinct re-
generic point ~
xn and the sample points ~
xn is computed as: sponse surfaces are required to define each load–time
curve.
r~x~xn ¼ k~
x ~
x n k2 ð5Þ
The selected RBFs, built using 20 sample points, are
and the final expression of the response surface is given summarised in Table 3 together with the maximum
by: percentage error obtained on the 10 remaining verifica-
tion points. The obtained approximations seem quite
X
N
f ð~
xÞ ¼ kn /ðr~x~xn Þ ð6Þ accurate and reliable as shown from the maximum er-
n¼1 rors which remains under the threshold of the 7%. Once
selected the type of Radial Basis Function, i.e. linear,
The following Radial Basis Functions are used in this
cubic or Gaussian, also the other 10 verification points,
work:
used for the preliminary evaluation of the accuracy, are
linear /ðr~x~xn Þ ¼ k r~x~xn with k ¼ 1 ð7:1Þ
3
cubic /ðr~x~xn Þ ¼ ðk þ r~x~xn Þ with k ¼ 0 ð7:2Þ
k r~2 ~
Gaussian /ðr~x~xn Þ ¼ e xxn with k ¼ 1 ð7:3Þ
Once defined the distance matrix:
Ai;k ¼ /ðr~xi ~xk Þ ð8Þ
the values of the coefficients kn are computed as:
ki ¼ Ai;k Fi ð9Þ
Particular attention has to be reached in the settlement
of the initial sample points since their positions directly
reflect on the values of the coefficients kn and, therefore,
on the final accuracy of the approximation.
In order to improve the accuracy of the results, the
input and output data of the response surfaces are linear
scaled between 0 and 1. Consequently, the settlement of
the sample point is obtained starting from an initial
random allocation of points in a normalized domain
where each design variable ranges from 0 to 1. The
underlying idea is to modify the initial positions to ob- Fig. 3. Example of the bi-linear approximated model and of the crash
tain a homogeneous and not systematic allocation inside behaviour of two different absorber shapes.
438 L. Lanzi et al. / Composite Structures 65 (2004) 433–441
Table 3
Radial Basis Functions
RBF function type Maximum % errora
Vertical impact––first point load Linear 4.75
Vertical impact––first point shortening Cubic 6.55
20° impact––second point load Cubic 3.88
20° impact––first point load Gaussian 4.16
20° impact––first point shortening Linear 6.45
20° impact––second point load Gaussian 6.23
30° impact––first point load Linear 2.82
30° impact––first point shortening Linear 6.85
30° impact––second point load Linear 6.21
a
Evaluated during the preliminary training performed using 20 of the 30 available sample points.
5. Optimisation results
Concluding, it is worthy noticing that all the config- then it monotonically increases up to the maximum
urations in the Pareto set exhibit particular high crash value of 1. The eccentricity shows a regular behaviour
forces efficiency: their values of absorbed energy per unit decreasing progressively from 0.55 to 0.
mass range from 32 to 42 kJ/kg.
5.3. Comparison with ideal cylindrical shape
5.2. Pareto set of the energy absorbed during 20° impacts
vs. weight The identified Pareto curves show how the presence
of moderate eccentricity and conicity leads to high effi-
A second multi-objective optimisation is performed ciency structures characterised by stable crush fronts
looking for the absorber shapes which minimise the and good absorption capabilities both considering ideal
weight and maximise the absorption capabilities during vertical and 20° inclined crash conditions.
crashes with 20° impact angle. The number of individ- Two other multi-objective optimisation runs are
uals and generations used for the GA, as in the previous performed in order to evaluate the improving capabili-
optimisation, are 50 and 100 respectively; the cross-over ties of conical absorbers with elliptical section with re-
and mutation probabilities are assumed equal to the spect to ideal cylindrical shells. These new optimisations
0.65 and to the 0.05 respectively. The Pareto set is drawn are performed assuming the same objectives of the
in Fig. 5 together with the values of the energy absorbed previous ones but considering as single design variable
during the corresponding vertical and 30° impact con- the major axis of the lower edge and, therefore, con-
ditions. straining the eccentricity to be 0 and the taper ratio to be
As far as the behaviour of the design variables along 1. In this way, only cylindrical shapes with radius equal
the Pareto curve is concerned, the dimension of the to X (1) are feasible configurations and then considered
major axis of the lower edge increases progressively until in the optimisation process. The Pareto curves are
the maximum value of 110 mm in correspondence of a shown in Fig. 6 together with the ones previously
weight greater than 0.16 kg. The taper ratio remains computed.
equal to 0.55 since the weight reaches a value of 0.16 kg, Considering vertical impact conditions, the Pareto set
of the cylinders is always non-greater than the one of the
conical absorbers with elliptical section. Fixed the ab-
sorbed energy, the maximum improvement in the weight
of the absorbers is within the 7%. The weight reduction
is much more relevant considering 20° impacts. In this
case, fixed the absorbed energy, the saved weight ranges
from the 25% to the 45%.
Similar results are reached with regard to the ab-
sorbed energy per unit mass. Considering vertical im-
pacts, the average absorbed energy per unit mass
computed along the Pareto curve of the conical
absorbers with elliptical sections is 36.5 kJ/kg. It is 32.6
7
Vertical impact
Cylindrical shell - Vertical impact
20° impact
Absorbed energy [kJ]
3
0.08 0.12 0.16 0.2 0.24
Weight [kg]
Fig. 5. Pareto curve of absorbed energy vs. weight under vertical 20° Fig. 6. Comparison between the Pareto curves of absorbed energy vs.
crashes and corresponding absorbed energy under vertical and 30° weight obtained considering conical shapes with elliptical section and
impacts. cylindrical shells.
440 L. Lanzi et al. / Composite Structures 65 (2004) 433–441
Table 4
Optimised configuration
Optimal absorber behaviour Constraints Bi-linear model Complete FE
RBF system FE analysis D% analysis
Absorbed energy under vertical impact (kJ) >5.50 5.55 5.39 2.97 5.42
First peak load under vertical impact (kN) – 16.11 16.10 0.06 16.10
Load corresponding to the second point of the bi- – 20.41 19.92 2.46 –
linear model for the vertical impact (kN)
Absorbed energy under 20° impact (kJ) >4.50 4.52 4.42 2.26 4.39
First peak load under 20° impact (kN) – 13.07 12.99 0.62 12.99
Load corresponding to the second point of the bi- – 18.06 17.62 2.50 –
linear model for the 20° impact (kN)
L. Lanzi et al. / Composite Structures 65 (2004) 433–441 441
giving back maximum errors within the threshold of the [4] McCarthy MA, Harte CG, Wiggenraad JFM, Michielsen ALPJ,
7%. The response surfaces are used with an enhanced Kohlgr^ uber D, Kamoulakos A. Finite element modelling of crash
response of composite aerospace sub-floor structures. Int J
GA in order to carry out both constrained single- and Comput Mech 2000;26:250–8.
multi-objective optimisations. A ranking selection [5] Bisagni C, Lanzi L, Ricci S. Optimization of helicopter subfloor
method is used to identify the Pareto sets. components under crashworthiness requirements using neural
The feasibility and flexibility of the proposed proce- networks. Int J Aircraft 2002;39(2):296–304.
dure is here exploited to formalise different optimisation [6] Delsart D, Delatombe E, Kohlgr^ uber D, Johnson A. Develop-
ment of numerical tools for the crash prediction of composite
problems changing objective and constraints. Indeed, helicopter structures. In: Proceeding of the 56th AHS Forum, 2–4
once built the response surfaces, no further finite ele- May 2000.
ment analyses are required during the optimisation runs. [7] Light Fixed Wing and Rotary Wing Aircraft Crash Resistance,
The optimisation results show how the presence of MIL-STD-1290A (AV), Department of Defence, Washington,
moderate eccentricity and conicity leads to high effi- DC, 1988.
[8] Aircraft Crash Survival Design Guide, USAAVSCOM TR 89-D-
ciency structures characterised by stable crush fronts 22A-E, vol. I–IV, 1989.
and good absorption capabilities both considering ideal [9] Farley L, Jones M. Crushing characteristics of continuous fiber-
vertical and 20° inclined crash conditions. Indeed, con- reinforced composites tubes. J Compos Mater 1992;26(1):37–50.
sidering vertical impact conditions, the Pareto set of the [10] Hull D. A unified approach to progressive crushing of fiber
cylinders is always non-greater than the one of the reinforced composite tubes. Compos Sci Technol 1991;40:377–
421.
conical absorbers with elliptical section. Fixed the ab- [11] Coutelliera D, Rozyckib P. Multi-layered multi-material finite
sorbed energy, the maximum improvement in the weight element for crashworthiness studies. Int J Compos Part A
of the absorbers is within the 7%. The weight reduction 2000;31:841–51.
is much more relevant considering 20° impacts. In this [12] Matzenmiller A, Lubliner J, Taylor RL. A constitutive model for
case, fixed the absorbed energy, the saved weight ranges anisotropic damage in fiber-composite. Int J Mech Mater
1995;20:125–52.
from the 25% to the 45%. [13] Schweizerhof K, Weimar K, Munz T, Rottner T. Crashworthiness
Finally a minimum weight optimisation is performed analysis with enhanced composite material models in LS-
fixing the minimum energy values to be absorbed under DYNA––Merits and limits. In: Proceedings of the 5th Interna-
vertical and 20° impact conditions. The performance of tional LS-DYNA Users Conference, Southfield, 21–22 September
the final configuration, identified using the Radial Basis 1998.
[14] Mahdi E, Hamouda AMS, Ashari BB, Khalid YA. Effect of
Function are validated by finite element analyses with material and geometry on crushing behaviour of laminated
an accuracy within the 3%. The load-shortening curves conical composite shells. Int J Appl Compos Mater 2002;9:265–
and the deformed shape evolutions show a stable crash 90.
front assuring progressive loads and good absorption [15] Chiandussi G, Avalle M. Maximisation of the crushing perfor-
capabilities. mance of a tubular device by shape optimisation. Int J Comput
Struct 2002;80:2425–32.
[16] Hallquist JO. LS-DYNA theoretical manual. Livermore Software
Technology Corporation; 1998.
Acknowledgements [17] Hallquist, JO. LS-DYNA Keyword Users Manual, Version 960.
vol. I and II. Livermore Software Technology Corporation, 2001.
The authors would like to thank Prof. Vittorio [18] Khoo LP, Chen CH. Integration of response methodology with
Giavotto of Politecnico di Milano for his generous ad- genetic algorithms. Int J Adv Manufact Technol 2001;18:483–9.
[19] Lanzi L, Bisagni C. Minimum weight optimization of composite
vises. Appreciations are due to Prof. Rade Vignjevic of stiffened panels using neural networks. In: Proceedings of the 44th
the Cranfield College of Aeronautics for the useful dis- AIAA/ASME/ASCE/AHS Structures, Structural Dynamics, and
cussions on the work. Cecilia Mirandola is greatly Material Conference, Norfolk, 7–10 April 2003.
acknowledged for her contribution in the developing of [20] Ugail H. Optimal design of thin-walled structures by means of
the LS-Dyna numerical model. efficient parameterisation. In: Proceedings of the 9th AIAA/
ISSMO Symposium on Multidisciplinary Analysis and Optimiza-
tion, Atlanta, 4–6 September 2002.
[21] Bloor MIG, Wilson MJ. Spectral approximations to PDE
References surfaces. Int J Comput-Aided Des 1996;28(2):145–52.
[22] Goldberg DE. Genetic algorithms in search, optimization and
[1] Giavotto V. The significance of crashworthiness. In: Proceedings machine learning. London: Addison-Wesley; 1989.
of the CEAS Forum on Crash questions, Capua, Italy, 2000. p. 1– [23] Krishnamurthy T. Response surface approximation with aug-
12. mented and compactly supported radial basis functions. In:
[2] Shin KC, Lee JJ, Kim KH, Song MC, Huh JS. Axial crush and Proceedings of the 44th AIAA/ASME/ASCE/AHS/ASC Confer-
bending collapse of an aluminium/GFRP hybrid square tube and ence, Norfolk, 7–10 April 2003.
its energy absorption capabilities. Int J Compos Struct [24] Hussain MF, Burton RR, Joshi SB. Metamodeling: Radial basis
2002;57:279–87. functions, versus polynomials. Euro J Oper Res 2002;138:142–54.
[3] Bouchet J, Jacquelin E, Hamelin P. Static and dynamic behavior [25] Lanzi L, Bisagni C, Ricci S. Neural network systems to reproduce
of combined composite aluminium tube for automotive applica- the crash behavior of structural components. Int J Comput Struct
tions. Int J Compos Sci Technol 2000;60:1891–900. 2004;82(1):93–108.