Beruflich Dokumente
Kultur Dokumente
e-mail: jangin@me.queensu.ca
Il Yong Kim1
e-mail: iykim@me.queensu.ca
Department of Mechanical and Materials
Engineering,
Queens University,
McLaughlin Hall 221,
130 Stuart Street
Kingston, ON, K7L 3N6, Canada
Introduction
elementwise differential equations may be replaced by a domainwide objective function in optimization. A similar relationship is
found in Newtonian mechanics, in which the equilibrium state of
objects is obtained by solving the simultaneous governing equations, and the same equilibrium can also be determined by minimizing the energy state of the objects 8.
In mechanical engineering, there has been a significant progress
in structural optimization, which is often classified into size,
shape, and topology optimization. Among these methods, topology optimization iteratively redistributes material in the domain to
determine an optimal material arrangement 911. Since the
1990s, topology optimization has been successfully applied to the
simulation of bone remodeling. Hollister et al. 12 used the homogenization method to obtain microtrabecular bone architecture.
Fernandes and Rodrigues 13 used a weighted sum of strain energy and mass as the objective function in topology optimization
to simulate bone remodeling around cementless hip stems. Bagge
14 used topology optimization with compliance minimization to
determine the initial material distribution for proximal femur.
Huiskes 15 summarized this observation: It has been shown
that computer algorithms for material design optimization, when
given the opportunity, show the tendency to produce trabecular
patterns, similar to bone. These studies provided insight into
bone as an optimized structure, but they rested on the assumption
without any solid foundation of theoretical proof.
Even though it is well known by experience that the two seemingly different approaches in different fieldsthe elementwise
bone-remodeling algorithms and the structural topology optimization with a domain-wide objective functionproduce similar re-
sults, it has not been thoroughly studied how these two approaches are compared explicitly: In terms of mathematical
formulations, numerical difficulties and their remedies, and behavior of their numerical solutions. Only a few studies tried to
derive the relationships between the two methods. Harrigan and
Hamilton 16 formulated a global functional for the simulation of
bone remodeling as the weighted sum of mass and the total strain
energy. They showed that the role of the functional was equivalent
to that of the remodeling equation developed by Huiskes and coworkers 5,6. Subbaraya and Bartel 17 also suggested the same
global functional for local bone-remodeling phenomenon. In the
aforementioned studies, the proposed global functional was totally
equivalent to the bone remodeling equation by Huiskes et al.
However, the simulations based on this functional are unconstrained optimization tasks, and they are different from the typical
topology optimization with a mass constraint that has been used
most widely in various applications including bone-remodeling
simulation. The analogy between this standard topology optimization with a mass constraint and the stain energy based boneremodeling equations has not been studied, and the relationships
between the topology optimization parameters and the boneremodeling equation variables are not yet identified.
In this paper, we formulated the typical form of topology optimization starting from the global functional by Harrigan and
Hamilton 16 and investigated the similarities and differences
between this topology optimization and the strain energy density
based approach in terms of mathematical formulations and numerical implementations. We then determined the matching relationships among the parameters used in the topology optimization
and bone-remodeling algorithm. Two numerical examples were
solved using the two approaches and the behavior of each numerical convergence was discussed.
Finally, it is also worth noting that the SED-based boneremodeling approaches and the topology optimization methods
have the same numerical difficultiescheckerboard patterns and
solution dependency on mesh resolutionand moreover the remedies used in each area are similar although the numerical techniques were developed independently. Weinans et al. 5 found
checkerboard patterns and mesh dependency in their numerical
bone-remodeling simulations. They used the terms patchwork effect for checkerboard pattern and chaotic phenomena for mesh
dependency. To overcome checkerboard patterns in bone remodeling, Jacobs et al. 18 suggested the use of quadratic elements
over bilinear elements and also the use of node-based approach;
these two remedies were also used in the same way in topology
optimization 19,20. More active schemes were developed to resolve the numerical difficulties: Mullender et al. 6 introduced the
influence area of sensors, and similar techniques were developed in topology optimization with the name of sensitivity filtering 21. Mathematical equations of these techniques are presented in Sec. 3.
Mathematical Formulations
d j
Ua,j
k ,
=B
dt
j
0 j upper
Ua,j =
1
Ui,j
n i=1
where Ui,j is the apparent strain energy density of the jth element
for load case i, and n is the number of load cases. The relation
between Youngs modulus and the density was taken as in Eq. 1.
Usually, in Eq. 1 has a value of 2 or 3. The remodeling process
is considered to converge when d / dt becomes zero in Eq. 2 or
when the density reaches the lower or upper bound.
To make the bone-remodeling process more physiological i.e.,
no checkerboards or no mesh dependency, Mullender et al. 6
introduced a spatial influence function fx to Eq. 2:
d
=
dt
f x
N
Ua,j
j=1
k ,
0 upper
minimize =
where ui is the displacement vector for load case i, n is the number of load cases, K is the global stiffness matrix of the structure,
j is the jth element density, j is the jth element volume, and N
is the total number of elements. We change Eq. 5 to
2u
n
minimize =
T
i
Kui +
i=1
j j M 0
j=1
where M 0 is a constant. Since Eq. 6 is different only by a constant from Eq. 5, they have the same sensitivity and therefore
they will produce the same optimum solutions.
The strain energy and mass together will be minimized as much
as possible in Eq. 6, but if we do not need to reduce mass any
further after we meet a prescribed mass target M 0, we can change
Eq. 6 to
minimize =
i=1
1 T
u Kui + P
2 i
P = max 0,
iv i M 0
i=1
2u
n
minimize =
T
i
Kui
i=1
subject to
v M
j j
j=1
It is important to note that Eq. 8 is the typical standard problem formulation for structural topology optimization with compliance minimization subject to a mass constraint.
From the equivalent form of topology optimization in Eq. 7,
sensitivity equations are derived into two different cases under the
assumption of linear elasticity and isotropy, as follows:
i
Case 1: Nj=1 j j M 0.
The sensitivity of the objective function with respect to
the density of the eth element is
=
e
2u
n
i=1
T K
ui + e
10
where L j is the jth element stiffness matrix with unit density. The sensitivity then becomes
=
e
2
n
2
1
1 T
e u i K eu i
i=1
= ne
1
=
n i=1
= ne
Ua,e
+ e
e
1 T
u K eu i
2 i
e
Ua,e
e
n
2u
n
T K
i=1
+ e
Ua,e =
1
Ui,e
n i=1
ii Case 2: Nj=1 j j M 0.
=
e
u iTL eu i + e
i=1
n
1 T
u Kui + j j
2 i
j=1
i=1
K j = j L j
11
ui = ne
Ua,e
e
12
and
k=
13
g
=
k
1
N
14
j j
j=1
H
j
j=1
= r distk , j,
where g is a functional to be minimized, and H
j
min
j N distk , j rmin, k = 1 , . . . , N.
Incorporating the sensitivity filtering into Eq. 11, we have
=
e
j=1
jH
j
N
j=1
H
j
Ua,j
k =
j
G x,
N
Ua,j
j=1
15
We then realize that the topology optimization sensitivity equation with sensitivity filtering, Eq. 15, is very similar to the boneremodeling equation with a spatial influence function, Eq. 4, in
their form. The main difference is that Gix , in Eq. 15 is a
function of the density as well as the location.
Several researchers have modified the original sensitivity filtering equation, Eq. 14, to improve the performance: By using a
different type of weighting function, by moving a density term in
the denominator inside the summation 23, by dropping the density terms 24, and so on. In this paper, we used an exponential
= edistk,j/D instead of the linear weighting
weighting function H
j
Numerical Examples
Ua,j =
cU
i
i,j
16
i=1
c 2u
3
minimize =
T
i
Kui
i=1
subject to
v M
j j
17
j=1
Fig. 4 Simulation result of structural configuration for a vertebral model: a Strain energy density based bone remodeling;
b topology optimization
Discussion
We showed in Sec. 4 that the SED-based bone-remodeling algorithm and topology optimization produced very similar structures for two case studies. To quantitatively compare the progress
during both simulations, we examined the convergence histories
of mass and total strain energy during the simulations, as shown in
Figs. 5 and 6.
First, in terms of total strain energy, the final values in the 300th
iteration were very close, as shown in Figs. 5 and 6: The differences in the plate model of bone tissue were only 0.42% and
3.35% for Load Case 1 and Load Case 2, respectively, and 5.02%
in a vertebral problem.
Second, based on the experimental data from the literature, we
analyzed the mass convergence histories of the two approaches.
Xinghua et al. 30 examined the functional adaptation in long
bones of young and adult cavia porcellus to study the progress of
the bone-remodeling process quantitatively. They found that the
bone adaptation occurred mainly during the first 20 days: 65.30%
during the first 20 days, 26.30% during the second 20 days, and
8.40% during the third 20 days. Xin et al. 31 also studied the
adaptive bone remodeling in terms of the radius increase for adult
cavia porcellus: The radius increase in percentage was 75% during
the first month, 20% during the second month, and 5% during the
third month. In Figs. 5 and 6, we plotted the experimental data by
Xin et al. 31 for the comparison with numerical simulations in
this study. We find that in terms of the remodeling progress in the
time domain, the SED-based bone-remodeling algorithm produced results that are more consistent with the experimental results in the references; however, topology optimization showed
faster convergences.
The faster convergence of topology optimization in mass may
be explained by the priority of the optimization process: If the
initial design violates the mass constraint, satisfying the constraint
has priority over improving the objective function. Another aspect
JANUARY 2009, Vol. 131 / 011012-5
Fig. 5 Convergence history during simulation of a plate model of bone tissue: a Mass change; b total strain energy
change
Acknowledgment
This work was supported by the Korea Research Foundation
Grant funded by the Korean Government MOEHRD, KRF2006-352-D00004.
References
1 Cowin, S. C., and Hegedus, D. H., 1976, Bone Remodeling I: Theory of
Adaptive Elasticity, J. Elast., 6, pp. 313326.
2 Fyhrie, D. P., and Carter, D. R., 1986, A Unifying Principle Relating Stress to