Professional Documents
Culture Documents
net/publication/338967989
CITATIONS READS
0 93
3 authors, including:
Christoph Eckert
Leibniz Universität Hannover Institute for Risk and Reliability
5 PUBLICATIONS 3 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Christoph Eckert on 20 June 2020.
Isogeometric analysis which extends the finite element method through the usage of B-splines
has become well established in engineering analysis and design procedures. In this paper, this
concept is considered in context with the methodology of polynomial chaos as applied to
computational stochastic mechanics. In this regard it is noted that many random processes used
in several applications can be approximated by the chaos representation by truncating the
associated series expansion. Ordinarily, the basis of these series are orthogonal Hermite
polynomials which are replaced by B-spline basis functions. Further, the convergence of the B-
spline based polynomial chaos is explored and substantiated by numerical results. Furthermore,
it is pointed out‚ that the B-spline chaos is a generalization of the Legendre multi-element
generalized polynomial chaos which is proven by solving a stochastic differential equation.
Keywords: Piecewise polynomial chaos, multi-element generalized polynomial chaos, B-splines,
random variable approximation, stochastic Galerkin.
1
Jie Li, Giovanni Solari and Pol Spanos (Eds)
applied to flow problems (Wan and normal, and gamma distributions are studied
Karniadakis (2006a), Wan and Karniadakis extensively and compared with Hermite and
(2006b)) and others (Wan and Karniadakis Legendre chaos. At last, an ordinary
(2009), Le Meitour (2010), Kewlani et al. differential equation (ODE) is solved by the
(2012), Sarrouy et al. (2013)). More recently, B-spline chaos using a stochastic Galerkin
improvements of the method were proposed scheme and faced with the ME-gPC.
(Gerritsma (2010), Chouvion and Sarrouy
(2016)), but handling long-term integration 2 B-spline based Polynomial Chaos
and stochastic discontinuities remains a
challenging problem. However, in practical Let (Ω,F ,P) be a probability space and X a
applications, one often does not know the real-valued random variable. Denote by
analytical form of the distribution of the input, L2 (Ω,F ,P) the Hilbert space of all random
or, if known, it may not be one of the basic variables with finite second moment.
distributions, e.g. uniform, Gaussian, etc. In Consider the non-Gaussian random variable
this case, the optimal convergence may X as a function of a standard Gaussian
deteriorates (Field and Grigoriu (2004), random variable Z , i.e.
Stefanou (2010)).
In 2005, Hughes et al. (2005) bridged the gap X = g(Z ), (1)
between computer aided design and
engineering by introducing the methodology where g is a deterministic mapping.
of isogeometric analysis (IGA), which
successfully has enhanced many deterministic 2.1 B-spline basis functions
engineering applications (Cottrell et al.
(2006), Bazilevs et al. (2008), Sangalli et al. Next, as a generalization of Bernstein
(2010), Cottrell et al. (2010), Auricchio et al. polynomials B-spline basis functions are
(2012)). In this regard, it is remarkable that characterized by an open non-uniform real
the notion of IGA was not widely adopted knot vector
within stochastic frameworks (Bhardwaj et al.
(2015), Hien and Noh (2017), Li et al. Ξ = [u1 ,…,uN + p+1 ] (2)
(2018)). In the best of the authors’
knowledge, there are two relevant approaches: where p is the polynomial order and N is
Hien and Noh (2017) combined the IGA with the number of basis functions – see e.g.
stochastic perturbation and Li et al. (2018) Hughes et al. (2005). The ui are called knots
numerically solved the Karhunen-Lóeve (KL)
and define element borders if they differ
expansion using isogeometric basis functions.
named knot spans. The B-spline basis
One procedure within the IGA framework is
functions can then be explained by the Cox-de
to use non-uniform rational B-splines
Boor recursion formula for p = 0 :
(NURBS), which are prevalent in engineering
design processes, as a basis for solution fields.
It turned out that B-splines are also beneficial ⎧⎪ 1 if ui ≤ u < ui+1
Bi,0 (u) ! ⎨ ; (3)
in terms of the analysis.
⎪⎩ 0 otherwise
In this paper, the B-spline basis functions are
employed in the truncated PC expansion and
and for p ≥ 1 :
weak convergence for arbitrary input variables
are shown. Further, strong convergence is
achieved for uniform distributed random
inputs. These results supported by numerical
examples where approximations of beta,
2
Proceedings of the Eighth Conference on Computational Stochastic Mechanics
Bi, p (u) !
u−ui
Bi, p−1 (u) this necessitates the introduction of a new
ui+ p −ui
u −u
(4) space for the parameter u of the B-spline
+ u i+ p+1−u Bi+1, p−1 (u). basis functions Bi, p which explicitly depend
i+ p+1 i+1
3
Jie Li, Giovanni Solari and Pol Spanos (Eds)
(
E g(Z )B j, p (FZ (Z )) ) (17) Further, if g(Z ) in Eq. (1) is explicitly
known in terms of Z , L2 -convergence can
= ∫ g (z)B j, p (FZ (z)) f Z (z) dz (18)
Ω be achieved Xiu (2010). However, in most
( ) practical numerical analyses only the PDF of
1
= ∫ g F (u) B j, p (u) du
Z
−1
(19)
0 g(Z ) or even less information is available.
Thus, the matrix A depends only on the But in this case, strong convergence cannot
configuration of the B-spline basis functions, established because of the lack of information
and can be stored before the analysis. Further, concerning g and Z . Nevertheless, the
A is a band matrix, if the knot Vector above theorem ensures weak convergence.
Ξ( p, N ) has inner knots, i.e.
3 Numerical Examples
A ij = 0 for | i − j |> p + 1. (20)
In this section the versatility of the
aforementioned approach is demonstrated, and
Examining the integrals in Eq. (19) and Eq.
(16), it is seen only the iCDF of the describing the convergence results are further
random variable and the mapping g must be substantiated by numerical examples.
known. Thus, this procedure is not limited to Specifically, the probability density function
Gaussian random variables and (PDF) of an uniform, and normal distributed
simultaneously paves the way even for using random variable is approximated by the B-
this method with discrete random variables. spline chaos, respectively. In addition, the
In fact, under the assumption that Z is a results are juxtaposed with Hermite and
uniformly distributed random variable FZ is Legendre chaos - e.g. Xiu and Karniadakis
the identity, and optimal convergence is (2002). All density functions are estimated by
expected in correspondence with the uniform a normal kernel smoothing function available
distribution. in all common statistical toolboxes. The
advantage of the proposed technique lies in
2.3 Convergence
the flexibility of adapting the order, number of
The proposed method is closely related to the elements and continuity over knot spans,
gPC Xiu and Karniadakis (2002) where the which can be quite powerful if the underlying
same mapping property between an uniform
distribution is unknown. See e.g. Xiu (2010)
and arbitrary distribution is utilized. It can be
for other density fittings.
shown, see e.g. Xiu (2010), that the gPC
approximation converges weakly, if the Finally, the ODE from Wan and Karniadakis
random variable to be approximated is square (2005) is solved using a stochastic Galerkin
integrable and the moments in the chaos scheme and compared to the Legendre based
expansion exists. This can be adopted here ME-gPC. Besides, h − p -convergence is
and X N converges in probability and in shown and it turns out, that the B-spline chaos
distribution, i.e. is a generalization of the Legendre multi-
element polynomial chaos, which can
P
XN → X for N → ∞, (21) extensively improved by increasing the
D
continuity over element boundaries.
XN → X for N → ∞. (22)
4
Proceedings of the Eighth Conference on Computational Stochastic Mechanics
0.25
normpdf (y, , )
p=1
p=3
0.2 p=5
p = 10
0.15
0.1
0.05
Figure 1. Approximation of uniform
distribution by Legendre and B-spline chaos.
0
-4 -2 0 2 4 6 8 10
3.1 Uniform distribution Figure 3. Approximation of normal distribution
by B-spline chaos.
Let X be a beta distributed random variable
on [0,1] with PDF B-splines - see figure 2. However, in this case
only a straight line has to be approximated.
1 Therefore, linear B-splines are sufficient.
f X (x) = xα −1 (1− x)β −1 (23)
Β(α , β ) This leads to the conclusion that a
correspondence between the uniform
where α , β > 0 and Β(α , β ) is the beta
distribution and B-splines can be identified.
function. In the case of α = β = 1 , the beta
Remark, for the Hermite chaos more terms are
distribution degenerates to the import uniform
necessary to reach the same accuracy, and
distribution. Figure 1 shows approximations
oscillations are observed at x = 0 and x = 1
of a uniform PDF by B-spline, and Legendre
and is referred to the stochastic Gibbs
chaos for several numbers of basis functions.
phenomenon Xiu (2010).
The B-splines and Legendre polynomials
remains stable and unchanged from the first 3.2 Normal distribution
order on. Neither order elevation nor knot
insertion changes the accuracy. The changing Let X be a normal distributed random
variable with PDF
values of the expansion coefficients are the
main difference. While only the first two 1 −
( x− µ )2
0.2
p = 1, nel = 10
p = 3, nel = 10 y N (t) = ∑ yi (t)Ψ i (Z ), (29)
p = 5, nel = 10 i=1
p = 10, nel = 10
0.15
where Z is the corresponding random
0.1 variable of the polynomials Ψi . { }
Substituting Eq. (28) and (29) in Eq. (25)
0.05
leads to
Figure
0 4. Approximation of normal distribution N
dyi (t)
∑
-4 -2 0 2 4 6 8 10
by B-spline chaos with ten elements. Ψ i (Z ) =
i=1 dt
(30)
distinctly, although moderate oscillations N N
Figure 5. Error of mean and variance of the ODE solved by Hermite, Legendre and B-spline chaos (t = 5).
Proceedings of the Eighth Conference on Computational Stochastic Mechanics
100100
0
102
10
Bernstein N el = 2
Bernstein N el = 3
-2 10-1
10-2
10 Bernstein N el = 4
B-Spline N el = 2 100
B-Spline N el = 3 10-2
-4
Variance
10
10-4 B-Spline N el = 4
Variance
Mean
Mean
-2
Hermite
10-3
10
-6 Legendre
10
10-6 Bernstein Nel = 1
Bernstein Nel = 2
10-4-4
-8 10
10 Bernstein Nel = 3
10-8
Bernstein Nel = 4 10-5
10-10 10-6
10-10 1 2 3 4 10-6 1 2 3 4
4 6 8 10 P 12 14 16 4 6 8 P10 12 14 16
N N
Numerical results of equation (31) are shown for the error of mean and variance of the
in figure 5. Hermite, Legendre and B-spline solution with respect to the number of basis
chaos are opposed against each other. functions is illustrated in figure 6. The dotted
Specifically, the error of mean and variance of lines represent the same results as in figure 6.
the multi-element generalized polynomial The solid lines show the errors of C 3 -B-spline
chaos expansion from Wan and Karniadakis chaos. Exponential h -type convergence for
(2005) are reproduced by using Bernstein both B-spline variants are on hand. Moreover,
polynomials, which are equivalent to C 0 -B- it can clearly be seen that for C 3 -continuity
spline basis functions. Exponential p -type much less basis functions are need to reach
convergence for different stochastic meshes, nearly the same accuracy. The number of
i.e. number of elements in the spectral basis functions N is directly related to the
expansion, are achieved. The Legendre chaos degrees of freedom of the numerical model.
is optimal for the uniform input. Therefore, it Thus, using smooth B-splines over element
generally outperforms the Hermite chaos here, boundaries instead of Legendre polynomials
which error is fluctuating and decreases in each element leads to a drastic reduction of
slower. Furthermore, the Legendre multi- the degrees of freedom and gain of efficiency.
element approach coincides with the C 0 -B- As mentioned above, this improvement is
spline chaos. Increasing the number of predicated on the smoothness over knot spans.
elements validates the results of (Wan and Note that, the exhibit advantages becomes
Karniadakis (2005), Figure 2). This means, even more pronounced if more elements or
through the natural structure of B-splines, that higher degrees are treated.
the performance of Legendre ME-gPC is
inherited by simply using C 0 -B-splines basis
functions in the polynomial chaos expansion. 4 Concluding Remarks
Note in addition, it is much easier to
implement the B-spline basis in an ordinary In this paper the potential of a B-spline based
PC framework then it is the case with ME- polynomial chaos expansion has been
gPC. demonstrated. Weak convergence has been
Further, the capability of C 0 - with C p−1 -B- shown and substantiated by several numerical
splines chaos is compared. The convergence examples. Correspondingly, the B-spline
Figure 6. Error of mean and variance of the ODE solved by Bernstein and B-spline chaos (t = 5).
Jie Li, Giovanni Solari and Pol Spanos (Eds)
approach has been found optimal for uniform 5. Cottrell, J., et al. Isogeometric analysis using
input and generalizes the Legendre multi- T-splines, Comput. Methods in Appl. Mech.
element formulation (ME-gPC). Further, it Eng., 199(5-8), 299-263, Jan, 2010.
has been found that the smoothness property 6. Cottrell, J., Reali, A., and Bazilevs, Y.,
Isogeometric analysis of structural vibrations,
of B-spline basis functions improves Comput. Methods in Appl. Mech. Eng.,
significantly the efficiency when 195(41-43), 5257-5296, Aug, 2006.
decomposing the random space, which comes 7. Chouvion, B., and Sarrouy, E., Development
to a greater extent if the dimensionality is of error criteria for adaptive multi-element
increased. polynomial chaos approaches, Mechanical
As a first example, to show the versatility and Systems and Signal Processing, 66(67), 201-
flexibility of the B-spline chaos a uniform, 222, Jan, 2016.
and normal distributed random variable have 8. Farouki, R.T., Legendre–Bernstein basis
been approximated. Varying the polynomial transformations, Journal of Computational
and Applied Mathematics, 119, 145–160,
degree, number of elements and also the 2000.
continuity over knot spans are powerful tools 9. Field, R.V., and Grigoriu, M., On the
to treat arbitrary non-uniform random inputs. accuracy of the polynomial chaos
Afterwards, a stochastic ODE has been approximation, Probabilistic Engineering
investigated within a Galerkin framework to Mechanics, 19, 65-80, Jan, 2004.
address long-time integration problems. 10. Gerritsma, M., et al. Time-dependent
Exponential h − p -convergence has been generalized polynomial chaos, Journal of
Computational Physics, 229(22), 8333–8363,
achieved for uniform input. Note that Wan 2010.
and Karniadakis (2010) have stated that the 11. Ghanem, R., Stochastic Finite Elements with
efficiency of ME-gPC is reduced significantly Multiple Random Non-Gaussian Properties,
by the rapidly increasing number of random Journal of Engineering Mechanics, 125(1),
elements for high-dimensional problems; the 26–40, Jan, 1999.
results reported herein indicate that the B- 12. Grigoriu, M., Stochastic Calculus:
spline chaos can overcome this drawback of Applications in Science and Engineering,
ME-gPC. Birkhäuser, 2003.
13. Ghanem, R., and Spanos, P., Stochastic
Finite Elements: A Spectral Approach, Dover
References
Publications, New York , 2003.
14. Ghanem R., and Spanos, P., A stochastic
1. Auricchio, F., et al. Isogeometric collocation
Galerkin expansion for nonlinear random
for elastostatics and explicit dynamics,
vibration analysis, Probabilistic Engineering
Comput. Methods in Appl. Mech .Eng., 249,
Mechanics 8(3-4), 255–264, Jan, 1993.
2-14, Dec, 2012.
15. Hughes, T. J., Cottrell, J. A., and Bazilevs,
2. Bazilevs, Y., et al. Isogeometric fluid-
Y., Isogeometric analysis: CAD, finite
structure interaction: Theory, algorithms, and
elements, NURBS, exact geometry and mesh
computations, Comput. Methods in Appl.
refinement, Comput. Methods in Appl. Mech.
Mech. Eng., 43(1), 3-37, 2008.
3. Bhardwaj, G., Singh, I., and Mishra, B., Eng., 194(39-41), 4135-4195, 2005.
Stochastic fatigue crack growth simulation of 16. Hien, T. D., and Noh, H.C., Stochastic
isogeometric analysis of free vibration of
interfacial crack in bi-layered FGMs using
functionally graded plates considering
XIGA, Comput. Methods in Appl. Mech.
Eng., 284, 186-229, Feb, 2015. material randomness, Comput. Methods in
4. Cameron, R. H., and Martin, W. T. The Appl. Mech. Eng., 318 845–863, 2017.
17. Kewlani, G., Crawford, J., and Iagnemma,
Orthogonal Development of Non-Linear
Functionals in Series of Fourier-Hermite K., A polynomial chaos approach to the
Functionals, The Annals of Mathematics, analysis of vehicle dynamics under
uncertainty, Vehicle System Dynamics, 50(5)
48(2), 385, 1947.
749– 774, May, 2012.
8
Proceedings of the Eighth Conference on Computational Stochastic Mechanics
18. Li K., et al. Spectral stochastic isogeometric method for stochastic differential equations,
analysis of linear elasticity, Comput. Methods Journal of Computational Physics, 209(2)
in Appl. Mech. Eng., 332, 157–190, Apr, 617–642, Nov, 2005.
2018. 25. Wan, X., and Karniadakis, G. E., Long-term
19. Le Meitour, J., Lucor, D. and Chassaing, J.C., behavior of polynomial chaos in stochastic
Prediction of stochastic limit cycle flow simulations, Comput. Methods in Appl.
oscillations using an adaptive polynomial Mech. Eng., 195(41-43), 5582–5596, Aug,
chaos method. J. Aero. Struct. Dyn. 2(1), 2006.
2010. 26. Wan, X., and Karniadakis, G. E., Multi-
20. Sangalli, G., et al. Isogeometric Collocation Element Generalized Polynomial Chaos for
Methods, Mathematical Models and Methods Arbitrary Probability Measures, SIAM J on
in Applied Sciences, 20(11), 2075–2107, Scientific. Computing, 28(3), 901-928, 2006.
2010. 27. Wan, X., and Karniadakis, G. E., Solving
21. Sarrouy, E., Dessombz, O. and Sinou, J.-J., elliptic problems with non-Gaussian
Piecewise polynomial chaos expansion with spatially- dependent random coefficients,
an application to brake squeal of a linear Computer Methods in Applied Mechanics and
brake system, Journal of Sound and Engineering, 198(21-26), 1985–1995, May,
Vibration, 332(3), 577–594, Feb. 2013. 2009.
22. Stefanou, G.. The stochastic finite element 28. Xiu, D., Numerical Methods for Stochastic
method: Past, present and future, Computer Computations: A Spectral Method Approach.
Methods in Applied Mechanics and Princeton University Press, 2010.
Engineering, 198(9-12),1031-1051, 2009. 29. Xiu, D., and Karniadakis, G. E., The Wiener–
23. Wiener, N., The homogeneous chaos, Amer. Askey Polynomial Chaos for Stochastic
J. Math, 60897(4), 936, Oct, 1938. Differential Equations, SIAM J. on Scientific
24. Wan, X., and Karniadakis, G. E., An adaptive Computing, 24(2), 619– 644, Jan, 2002.
multi-element generalized polynomial chaos