Engineering Structures 31 (2009) 1344–1348

Contents lists available at ScienceDirect

Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct

Sensitivity assessment of steel members under compression

Zdeněk Kala
Brno University of Technology, Faculty of Civil Engineering, Department of Structural Mechanics, Veveri Street 95, 602 00 Brno, Czech Republic

article info a b s t r a c t

Article history: The objective of the paper is to analyse the influence of initial imperfections on the behaviour of a steel
Available online 14 May 2008 member under compression. The influence of the variability of initial imperfections on the variability of
the load-carrying capacity studied has been calculated by sensitivity analysis. The advantages of Sobol’s
Keywords:
sensitivity analysis and the most important properties of Sobol’s sensitivity indices are described. The
Sensitivity
Steel
Sobol’s first order sensitivity indices are evaluated in dependence on the nondimensional slenderness.
Stability The Sobol’s sensitivity indices are supplemented with a lucid elaboration based on the Monte Carlo
Imperfections method. Material and geometrical characteristics of a steel member IPE 220 were considered to be
Uncertainty random quantities the histograms of which were obtained from experiments. Imperfections that have
Statistic a dominant influence on the load-carrying capacity are identified.
Simulation © 2008 Elsevier Ltd. All rights reserved.
Random
Variance

1. Introduction mathematical models is studied using sensitivity and uncertainty

analyses [15]. The general definition of sensitivity analysis is the
The properties of structures are influenced by a number of study of the uncertainty of model output (numerical or otherwise)
factors which are of random character (material, geometry, effects arising from varying sources of model input uncertainties [16]. The
of the surrounding environment, load action, etc.). If a structure is uncertainty analysis is aimed rather at the quantification of the un-
to reliably fulfil its function during its service life, it is necessary certainty of model output [2,13]. Sensitivity and uncertainty analy-
to make provision for this during its design. In the general ses are employed as formal methods for the evaluation of data and
classification of initial structural imperfections, three fundamental models, because they both enable the evaluation of uncertainty of
categories of imperfection are considered [8]. They include: output variables and provide information on the importance of in-
1. Geometrical imperfections: initial curvature of member axis, put variables and their influence on the monitored output [15,16].
excentricity of load action, deviation from the theoretical layout Sensitivity and uncertainty analysis methods are divided into
of the cross section (tolerance of dimensions and shape of the two types: (i) deterministic, and (ii) stochastic. Analysis (i) is
cross section), etc. a relatively known method currently used in structure design.
2. Material imperfections: dispersion of the mechanical proper- This analysis accompanies the design procedures for which a
ties of the material (non-homogeneousness of material char- calculation model is applied, e.g. in [10,22]. We usually speak
acterised by the dispersion of the yield strength, ultimate about a parametric study (sometimes called “what-if-study”).
strength, Young’s modulus, etc.), initial stress state (residual When designing a structure, a parameter (e.g. cross section, steel
stress as a consequence of rolling, welding, straightening and grade, weld parameters) gets changed, and the influence on input
other technological manufacturing processes). (e.g. load-carrying capacity, deformations) is studied. However,
3. Structural imperfections: imperfections in the realization of quantified data on the uncertainty and/or sensitivity are not
joints, connections, welds, anchorage and other structural obtained. In the case (ii), we can compute the average output,
details which are apparent in comparison with the theoretical its standard deviation, the quantiles of its distribution, confidence
assumptions introduced in the solution of idealized system, in bounds, plot the distribution itself and so on [16]. Upon obtaining
deviations of the effects of the actual structural system. results of this uncertainty analysis, we can then perform the
sensitivity analysis to determine which of the input parameters
Most initial imperfections arise due to inaccuracy during man- have a more dominant influence on the uncertainty in the model
ufacturing process. The influence of uncertainties of input im- output [16]. For discussions on the existing techniques see, for
perfections on the uncertainty of system response expressed by
example, [4,12]. For application of the fuzzy sets theory in models
with prevailing epistemic uncertainty, see [5,11].
This article is aimed at stochastic sensitivity analysis. The
E-mail address: kala.z@fce.vutbr.cz. stochastic sensitivity analysis will be carried out with the aim
0141-0296/$ – see front matter © 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.engstruct.2008.04.001
 Z. Kala / Engineering Structures 31 (2009) 1344–1348 1345

of assessing the relative sensitivity of the random variability of histogram, etc.) can be performed utilizing the response function
a monitored event to the random variability of individual input (1). For this purpose, it is practical to consider the density function
variables. The random variability influence of a certain input pi (Xi ) within the interval h0; 1i in the simplest possible form with
variable (in comparison with other variables) on the random Rectangular distribution.
variability of a monitored output (e.g. load-carrying capacity, How can the decomposition of the response function (3)
failure probability) will be sought. be utilized for sensitivity analysis? Sensitivity analysis can
The sensitivity analysis, thus, also answers the question as to be carried out by substituting deterministic values into (3)
which variables are dominant and should be considered carefully and subsequently by comparing individual members of the
during: (i) the preparation of input variables; (ii) determination decomposition f0 , fi , fij , . . . , f12...M with the value of output f . Let
and decision making on the improvement of technological us recall that decomposition (3) is not worked out based on the
processes; (iii) conception and organization of control activities [4]. analysis of the response function (1), but by the analysis of change
Ilya M. Sobol’ [19–21], a Russian mathematician, elaborated of output Y arising from quantified changes of input variables given
one of the most coherent sensitivity analyses. In the paper by functions pi (Xi ). Change of function pi (Xi ) during decomposition
presented, the Sobol’s sensitivity analysis will be applied for study leads to change of member f0 and of all members with index i,
of the influence of input random quantities on the load-carrying i.e. fi , fij etc. In practice, this means that the decomposition of the
capacity of a steel grade S235 member. The sensitivity analysis response function (1) into (3) can also be carried out with an
will be carried out using statistical material and geometrical unknown algorithm, for which only input and output are known
characteristics obtained by physical experiment research [9,17]. (black box), whereas the sensitivity analysis can be carried out by
quantifying the influence of change in input variables to change of
output variables.
2. Variance–based sensitivity indices
The change of output variable Y is characterised by standard
deviation σY or variance V (Y ) = σY2 . Since (4) and (5) are valid,
Let us consider a computational model with input variables all members of decomposition (3) are statistically independent
(X1 , X2 , . . . , XM ) of non-zero variance (or uncertainty), and let us random variables and we can write that the variance is equal to
monitor the influence of these variables on the output variable Y the sum of variances of the individual members of decomposition:
utilizing the response function f :
V (Y ) = V (fi (Xi )) + V fij Xi , Xj
X XX


+ ···
Y (X ) = f (X1 , X2 , . . . , Xi , . . . , XM ) . (1) i i j>i

We study the case with statistically fully uncorrelated input + V (f12...M (X1 , X2 , . . . , XM )) (7)
random variables. Let us consider the response function f (1), the where V (fi (Xi )) = V (E (Y |Xi )) etc. If we express the ratio of
integral of which can be performed on its function set Ω M : individual members of the decomposition (7) to the total variance,
Sobol’s first order sensitivity indices may be written in the form:
Ω M = (X |0 ≤ xi ≤ 1; i = 1; . . . , M) . (2)
V (E (Y |Xi ))
Sobol’s concept is based on the hierarchical decomposition of the Si = . (8a)
response function (1) into a form with increasing dimension: V (Y )
X XX For users, the significance of Si is as follows: If the variability
f = f0 + fi + fij + · · · + f12...M (3) Xi is successfully eliminated, the output quantity variance will
i i j>i
decrease by Si · 100% minimum. In [21], Sobol proposed an alternate
where each member is a function
of input variables of given definition Si = corr (Y, E (Y |Xi )) based on the evaluation of the
indices: fi = fi (Xi ), fij = fij Xi , Xj , etc. Each member of the correlation between output random variable Y and the conditional
decomposition of function f should also have an integral on its random arithmetical mean E (Y |Xi ). Analogously as (8a), we can
functional subset. The decomposition (3) is not a series expansion, write the second order sensitivity indices:
because it has 2M (finite number) of members: f0 is a constant,
V E Y |Xi , Xj



number of members fi is M, number of fij members is
M
etc. Each Sij = − Si − Sj . (8b)
2 V (Y )
input variable has a density function pi (Xi ) ≥ 0 defined on interval
Other Sobol’s sensitivity indices enabling the quantification of
h0; 1i, pi (Xi ) = 0 outside this interval. As a result of the Fubini higher order interactions can be expressed similarly. With regard
theorem on the double integral, it holds that if each member of the to (3), the decomposition of Sobol’s sensitivity indices can be
decomposition apart from constant has a zero mean value: written in the form:
Sijk + · · · + S123...M = 1.
Z 1 X XX XXX
Si + Sij + (9)
E (f (xi )) = pi (xi ) · f (xi ) dxi = 0 (4) i i j>i i j>i k>j
0

then all members of decomposition are orthogonal in pairs:

Z 1Z 1 3. Sensitivity analysis of load-carrying capacity
E f (x i ) · f x j = pi (xi ) · pj xj · f (xi ) · f xj dxi dxj = 0,





0 0 Let us now apply the Sobol’s decomposition to the analysis of
i 6= j (5) the load-carrying capacity of a steel strut of profile IPE220, see
Fig. 1.
Members of the decomposition (3) can thus be rewritten utilizing
Let initial deflection of the column be assumed to be half sine
the conditional realization of the response function:
wave with the amplitude e0 , as shown in Fig. 1.
f0 = E (Y ) (6a) π·x
 
fi = E (Y |Xi ) − E (Y ) (6b) y0 = e0 sin . (10)
L
fij = E Y |Xi Xj − fi − fj − E (Y ) .


(6c) The resulting shape of the strut under load action F (see Fig. 1) is
given by the differential equation
The condition that the density pi (Xi ) is zero outside interval h0; 1i
is not limiting, because transformation of each input variable d2 y F (y + y0 )
into another suitable distribution type (Gaussian, lognormal, + =0 (11)
dx 2 E · Iz
1346 Z. Kala / Engineering Structures 31 (2009) 1344–1348

q
A2 · Q 2 + 2 · A · Fcr · Wz· |e0 | · Fcr − fy · Wz + Fcr
2 · W2 − A · Q − F · W


z cr z
R=−
2 · Wz

Box I.

Table 1
Input random quantities
Symbol Density Mean value Standard deviation

1. h Histogram 220.22 mm 0.9746 mm

2. b Histogram 111.48 mm 1.0930 mm
3. t1 Histogram 6.2245 mm 0.2467 mm
4. t2 Histogram 9.1356 mm 0.4214 mm
5. fy Histogram 297.3 MPa 16.8 MPa
6. E Gauss 210 GPa 12.6 GPa
7. e0 Gauss 0 0.76533 L

Fig. 1. Buckling of the member with IPE cross-section.

where Iz is the second moment of area to axis Z (axis perpendicular

to flange around which the section bends during buckling). After
substituting for y0 from (10), and considering the boundary
conditions (x = 0; y + y0 = 0 and x = L; y + y0 = 0), the solution is
obtained:
e0 π·x
 
y= F · sin (12)
Fcr
−1 L

where Fcr is Euler’s critical load. The column deflection at mid-

length x = L/2 [23] is:
e0 e0
e = e0 + F
= F
. (13)
Fcr
−1 1− Fcr
Fig. 2. Sobol’s indices Si versus λ.

The maximum stress σmax due to the combination of the axial

one can enter X1 and X3 , where X3 represents a factor describing
uniform stress and the bending stress is:
noise, and model X2 is a function of X1 and X3 . It is necessary to
R R · |e | note that dependence and correlation are not synonymous. A cor-
σmax = + = fy . (14) relation implies dependence, while the opposite is not true. Depen-
A Wz
dencies described via correlations are useful for practical numeri-
Load-carrying capacity R is maximum load action F of elastic cal computations [16].
member (σmax is equal to yield strength fy ). R can be evaluated from Statistical characteristics h, b, t1 , t2 , fy were considered as his-
(14) with used (13): tograms according to results of physical experimental research [9].
R R · |e 0 | Statistical characteristics of Young’s modulus E were considered
σmax = + = fy ⇒ R (15) according to [1,18]. The standard deviation of the Gaussian distri-
A (1 − R/Fcr ) · Wz bution of density function of the amplitude of initial curvature e0
where A is the sectional area, Wz is the sectional module to axis has been considered based on the assumption that 95% of the re-
Z . The model uncertainty factor as proposed in [27] has been alizations of this random variable are found within the tolerance
neglected. By the solution (15), we can obtain the load-carrying limit ±0.15% L of standard [26], where the strut length L is a com-
capacity R in the form as given in Box I: putational parameter.
The dependence of Sobol’s sensitivity indices Si of the first
Q = |e0 | · Fcr + fy · Wz (16a) order on dimensionless slenderness λ = L/ (iz · 93.9) [25] is
A = 2 · t2 · b + (h − 2t2 ) · t1 (16b) depicted in Fig. 2, where iz = 24.8 mm is the nominal value of
  the radius of gyration of profile IPE220. The sensitivity indices Si
Fcr = π2 EIz / L2 (16c) were evaluated from the basic definition (8a) utilizing the Monte
  Carlo method. The conditional random arithmetical mean E (Y |Xi )
Iz = 2 · t2 · b3 + (h − 2t2 ) · t13 /12 (16d) was evaluated for N = 100 000 simulation runs; the variance
Wz = Iz / (b/2) . (16e) V (E (Y |Xi )) was calculated for N = 100 000 simulation runs, as well,
i.e. the numerically demanding difficulty of the calculation is N2 .
E is the Young’s modulus, L is the strut length, h is the sectional In practice, the procedure is such that indices Si were evaluated for
height, b is the sectional width, t1 is the web thickness and t2 is the strut length L, which was parametrically increased from zero with
flange thickness. Input random variables are listed in Table 1. The the step ∆L = 0.094 m.
input variables are fully uncorrelated. All variables in Table 1 apart from the amplitude of initial
Fully uncorrelated input variables represent one of precondi- curvature e0 with standard deviation σe0 , which was determined
tions for the application of Sobol’s sensitivity analysis. However, according to the tolerance standards [26], were determined
initial imperfections may generally be considered as random fields, experimentally. The dependence of Sobol’s sensitivity indices Si
e.g. in [5,14]. How can we proceed when it is necessary to use the of the first order on the standard deviation σe0 of amplitude
correlated input factors? The useful strategy to circumvent the use e0 of strut length L = 2.35 m is depicted in Fig. 3. The grey
of correlated samples in sensitivity analysis can be illustrated by an background emphasises the interval (±0.1% L; ±0.2% L), where we
example in [16]. Instead of entering X1 and X2 as correlated factors can expect the standard deviation that would be obtained, with
 Z. Kala / Engineering Structures 31 (2009) 1344–1348 1347

indices of variables E, t2 , b is approximately equal to 0.5. In the case

of higher slenderness, variables E, t2 , b have a dominant influence
on the load-carrying capacity.
It is important to note that the influence of residual stresses
was neglected during evaluation. From the differentiation of curves
of normative buckling resistance a, b, c, d of standard EC3 [25],
the curves differ most for λ = 0.83, if the influence of residual
stresses is taken into consideration. A more elaborate description
of the influence of this structural imperfection would require
utilization of thin-walled finite elements to model the strut,
and the evaluation of the load-carrying capacity applying the
geometric non-linear solution; however, it is practically impossible
in connection with the numerically demanding evaluation of
Fig. 3. Sobol’s indices Si versus σe0 . Sobol’s sensitivity indices.
This problem can be solved by applying numerical methods
such as the “response surface” where the approximating function
should be sufficiently detailed and number of approximation
points sufficiently high to enable the description of higher
order interactions. The influence of residual stresses can be
taken into account in (15) by increasing the absolute value
of geometric imperfection of amplitude e0 . This assumption,
however, may not be sufficiently apposite for an elaborate
sensitivity analysis utilizing Sobol’s sensitivity analysis as pointed
out by the sensitivity analysis of the load-carrying capacity
worked out according to (9), where the Spearman correlation
coefficients of residual stress and of amplitude e0 corresponded
only approximately.
Input random imperfections may be divided into two basic
Fig. 4. Sobol’s first order sensitivity indices for λ = 1.0. groups [4]. The first group includes those variables the statistical
characteristics of which can be positively influenced in production
the highest probability, from measurements on real struts. The (yield strength, geometric characteristics, residual stress) and
inequality 0.98 < i Si < 1 indicates the presence of weak higher those that are not sufficiently sensitive to changes in production
P

interactions for all analysed slenderness. technology (e.g. variability of Young’s modulus E). The first group
Sensitivity indices Si for slenderness λ = 1.0 are depicted in of variables may be further divided into two subgroups: (i)
Fig. 4. The sum of all coefficients Si is 100% · i Si ≈ 98.2%, i.e. 1.8% variables for which mean value and standard deviation can be
P

accounts for higher order interactions Sij , Sijk etc. changed by improvement of production quality [4]. Examples
include Young’s modulus; (ii) variables, the mean value of which
4. Conclusion cannot be significantly changed, because it should approximately
correspond to the nominal value (geometric characteristics of
It is evident from results depicted in Fig. 2 that for struts of profile dimensions).
slenderness λ ≤ 0.47, the dominant variable is the yield strength Significant variables in this regard include yield strength,
fy . If the strut slenderness is equal to zero, it presents the case initial axial strut curvature and flange thickness t2 . Decrease in
of simple compression where the load-carrying capacity R = fy · the variability of these variables can be achieved by change in
A is dependent only on the values of the yield strength and of production technology. Decrease of the variability of yield strength
those of sectional area. For λ = 0, the sensitivity coefficient of fy is recommended especially for struts with lower dimensionless
yield strength is Sf y = 0.75, and since the interactions of higher slenderness.
orders are very small the sensitivity index of sectional area may be The sensitivity analysis enables us to identify significant
considered approximately equal to SA ≈ 1 − 0.75 = 0.25, where processes and phenomena which influence the reliability of load-
the flange thickness St2 = 0.185 and the web thickness St1 = 0.055 carrying structures during service life, and therefore it can be
are of dominant influence, see Fig. 2. applied to the development of knowledge of real behaviour and
Initial strut axial curvature (represented by the amplitude e0 ) limit states [3,24]. At present, the Sobol’s sensitivity analysis is
is the dominant variable for slenderness λ = h0.47; 1.67i. The one of the most carefully formulated and most coherent concepts
maximum sensitivity index max Se0 = 0.82 arises for λ = which can be applied to the analysis of the majority of stability
0.88. Young’s modulus E is the dominant variable for slenderness problems [6,7,14]. It is necessary to try and develop experimental
λ > 1.67. Another important variable among the geometric methods aimed at the objectivization of knowledge of the real
characteristics is the flange thickness t2 . The load-carrying capacity structure behaviour, and at the verification of theoretical models
is equal to the Euler critical force in the limit case λ = ∞ (or e0 = in relation to definitions of limit states.
0), i.e. it is dependent only on variables E, t2 , b, which represent the
input variables for the evaluation of stiffness preventing buckling
EIz . Values of sensitivity indices of variables E, t2 , b increase most Acknowledgements
rapidly for λ ≈ 1.0, whereas the sensitivity index of the amplitude
e0 decreases most rapidly. The sensitivity index of yield strength The article was elaborated within the framework of project
decreases practically to zero for λ ≈ 1.2. For λ ≈ 1.3, the GAČR 103/07/1067, GAČR 103/08/0275 and research
sensitivity index of e0 is equal to 0.5, and thus the sum of sensitivity MSM0021630519.
1348 Z. Kala / Engineering Structures 31 (2009) 1344–1348

References Eng 2004;85:11–9.

[14] Sadovský Z, Guedes Soares C, Teixeira AP. Random field of initial deflections
[1] Fukumoto Y, Kajina N, Aoki T. Evaluation of column curves based on and strength of thin rectangular plates. Reliab Eng 2007;92(12):1659–70.
probabilistic concept. In: Proc. of int. conference on stability. Prelim. rep., publ. [15] Saltelli A, Tarantola S, Campolongo F, Ratto M. Sensitivity analysis in practice:
by Gakujutsu Bunken Fukyu. Kai, Tokyo; 1976. p. 1–37. A guide to assessing scientific models. New York: John Wiley and Sons;
[2] Helton JC, Johnson JD, Oberkampf WL. An exploration of alternative 2004.
approaches to the representation of uncertainty in model predictions. Reliab [16] Saltelli A, Ratto M, Andress T, Campolongo F, Cariboni J, Gatelli D, et al. Global
Eng 2004;85(1–3):39–71. sensitivity analysis guiding the worth of scientific models. New York: John
[3] Jatulis D, Kamaitis Z, Juozapaitis A. Static behaviour analysis of masts with Wiley and Sons; 2007.
combined guys. J Civil Eng Management 2007;XIII(3):177–82. [17] Strauss A, Kala Z, Bergmeister K, Hoffmann S, Novák D. Technologische
[4] Kala Z. Sensitivity analysis of the stability problems of thin-walled structures. Eigenschaften von Stählen im europäischen Vergleich. Stahlbau 2006;75(Heft
J Construct Steel Res 2005;61:415–22. 1).
[5] Kala Z. Stability problems of steel structures in the presence of stochastic and [18] Guedes Soares C. Uncertainty modelling in plate buckling. J Structural Safety
fuzzy uncertainty. J Thin-Wall Struct 2007;45/10(11):861–5. 1988;(5):17–34.
[6] Karkauskas R, Norkus A. Truss optimization under stiffness, stability [19] Sobol’ IM. Multidimensional quadrature formulas and haar functions.
constraints and random loading. Mech Res Comm 2006;33(2):177–89. Moscow: Nauka; 1969 [in Russian].
[7] Kotełko M. Load-capacity estimation and collapse analysis of thin-walled [20] Sobol’ IM. Sensitivity estimates for nonlinear mathematical models. Matem-
beams and columns—recent advances. Thin-Wall Struct 2004;42(2):153–75. aticheskoe Modelirovanie 1990 2:112–8. [in Russian, translated into English
[8] Melcher J. Classification of structural steel members initial imperfections. In: in Sobol’ 1993].
Proc. of structural stability research council, Lehigh University, Bethlehem; [21] Sobol’ IM. On freezing unessential variables. Vestnik Mosk Univ Ser Mat 1996;
1980. 6:92–4.
[9] Melcher J, Kala Z, Holický M, Fajkus M, Rozlívka L. Design characteristics [22] Šalna R, Marčiukaitis G. The influence of shear span ratio on load capacity of
of structural steels based on statistical analysis of metallurgical products. fibre reinforced concrete elements with various steel fibre volumes. J Civil Eng
J Construct Steel Res 2004;60:795–808. Management 2007;XIII(3):209–15.
[10] Miedziałowski C, Chyży T, Krętowska. J. Numerical model of three- [23] Timoshenko S, Gere J. Theory of elastic stability. New York: McGraw-Hill;
dimensional coupled wall structures. J Civil Eng Management 2007;XIII(1): 1961.
37–45. [24] Verveckaite N, Amsiejus J, Stragys V. Stress-strain analysis in the soil sample
[11] Möller B, Beer M. Engineering computation under uncertainty—capabilities of during laboratory testing. J Civil Eng Management 2007;XIII(1):63–70.
non-traditional models. Comput Struct 2008;86(10):1024–41. [25] EN 1993-1-1:2005 (E): Eurocode 3: Design of steel structures – Part 1-1:
[12] Novák D, Teplý B, Shiraishi N. Sensitivity analysis of structures—a review. In: General rules and rules for buildings, CEN; 2005.
Proc. of the fifth international conference on civil and structural engineering [26] EN 10034: Structural steel I and H sections. Tolerances on shape and
computing. Edinburgh; 1993. p. 201–7. dimensions; 1995.
[13] Oberkampf WL, Helton JC, Joslyn CA, Wojtkiewicz SF, Ferson S. Challenge [27] JCSS probabilistic model code. Part 3. Resistance models. Static properties of
problems: Uncertainty in system response given uncertain parameters. Reliab structural steel (rolled sections). JCSS Zurich; 2001. http://www.jcss.ethz.ch/.