Professional Documents
Culture Documents
Keywords: A topology optimization method is proposed for transient response problems involving thermoelastic models,
Topology optimization where the objective is to optimize the structure considering time-dependent dynamic and thermal loads.
Multi-objective optimization The proposed method can be applied to problems involving thermal deformation and stress, which damage
Dynamics analysis
heated machine parts. This paper presents a transient response analysis of a thermoelastic model subjected to
Thermoelasticity
mechanical and thermal loads having arbitrary profiles with respect to time. This can be formulated as a multi-
Adjoint sensitivity analysis
Stress minimization
objective optimization problem to obtain a structure that minimizes the deformation and maximum stress. A
sensitivity analysis was conducted by using the adjoint variable method and the discretize-then-differentiate
approach. The proposed method was applied to numerical examples for validation.
∗ Corresponding author.
E-mail address: t.yamada@mech.t.u-tokyo.ac.jp (T. Yamada).
https://doi.org/10.1016/j.finel.2021.103695
Received 7 June 2021; Received in revised form 28 September 2021; Accepted 8 November 2021
Available online 30 December 2021
0168-874X/© 2021 Elsevier B.V. All rights reserved.
S. Ogawa and T. Yamada Finite Elements in Analysis & Design 201 (2022) 103695
thermoelastic problems considering stress constraints. They considered the two approaches for transient response problems. Particularly, they
the maximum stress constraint problem for finding a single material compared the accuracies of the two sensitivity analyses with a one-
configuration assuming a temperature difference from the uniform dimensional model. They showed that the discretize-then-differentiate
stress-free temperature in the design domain. A constraint function for approach is more accurate than the differentiate-then-discretize ap-
stress was defined based on the p-norm concept, and the effectiveness proach because the equations of motion are discretized in the time
of the qp-reduction method [15,16] at stress relaxation to eliminate
domain, and the resulting equations are the same as those solved in
singularities in the no-material domain was demonstrated. Takezawa
the numerical analysis. In particular, it is advantageous in that it can
et al. [17] proposed a method for solving thermo-structural coupled
ensure the accuracy of sensitivity even when the time increment size is
problems considering stress and heat transfer constraints. For the stress
constraint, they applied the same qp-reduction method as in Deaton set large. As a result, the number of analysis steps can be reduced when
et al. However, the treatment of stress as an evaluation function in performing analysis on long time scales, which is highly effective from
the above optimization problems has only been considered for a single the perspective of reducing the computational cost.
material, and no studies have considered multi-materials. Other stress As an application of topology optimization to actual transient re-
relaxation methods have been proposed to eliminate the singularity sponse problems, Zhao et al. [27] discussed minimizing deformation of
phenomenon, such as the epsilon relaxation method [18]. However, a linearly elastic model by using the dynamic compliance and strain
all of these assume a single material, and the stresses evaluated by energy as objective functions. They proposed setting the objective
dynamic structural analysis and stress constraint functions differ in the
function by applying the KS function and p-norm to minimize the
grayscale domain. The case of no stress relaxation is assumed to lead
maximum displacement occurring at all times. However, the previous
to excessive grayscale, so the cause of this problem and its solution
studies focused on displacements, and no studies have considered stress
must still be explored. In topology optimization of density methods,
grayscale issues are well known. Therefore, there are several studies on reduction or thermoelastic behavior. Therefore, topology optimization
grayscale removal methods. Borrvall and Petersson [19] have proposed for transient response problems needs to be expanded in scope to
Explicit penalization. This method considers a penalty function that has reduce both deformation and stress considering thermoelastic behavior.
only grayscale values in the objective function, and removes grayscale In this study, we developed a methodology for topology optimiza-
by including this function and minimizing it. Alternatively, Sigmund tion of a transient response problem considering thermoelastic behav-
and Maute [20] have introduced projection schemes. In addition to fil- ior, which is formulated as a multi-objective optimization problem
tering the design variables, this approach uses the heaviside projection to minimize the deformation and maximum stress. Particularly, we
to sharply represent the optimal configuration and remove grayscale.
developed an objective function for multi-objective optimization and
A new initiative is the study of robust topology optimization in
addressed stress minimization for two-phase materials, which has not
thermoelastic models. Alacoque et al. [21] applied the robust topol-
yet been considered. We conducted a sensitivity analysis to ensure that
ogy optimization to determine the microstructure of a multi-material
problem and showed computational examples such as negative thermal the optimized structure accurately considers the time-varying behav-
expansion materials. Similarly, Li et al. [22] proposed a design method ior of the thermoelastic model. Section 2 discusses the thermoelastic
for meta-materials with negative thermal expansion, considering the constitutive law and governing equations considered in the transient re-
uncertainty of material properties. sponse analysis. Section 3 describes the setup for topology optimization
Compared with static problems, few studies have considered apply- of two-phase materials using the density method. Section 4 defines the
ing topology optimization to thermoelastic models of dynamic prob- evaluation functions and sets up the multi-objective optimization prob-
lems, and no studies have considered stress as an evaluation function. lem of minimizing both deformation and stress. Section 5 describes the
Yang et al. [23,24] dealt with a dynamic compliance minimization sensitivity analysis of the optimization problem defined in Section 4.
problem that considered a two-phase material for the equation of Section 6 validates the sensitivity analysis based on values obtained
motion expanded in the frequency domain. They showed that the
from the finite difference method as a benchmark. Sections 7 and
optimization results for the soft structure differed when the effect
8 presents several numerical examples to demonstrate that the pro-
of loading was considered. However, their study was limited in that
posed method can produce appropriate optimization results. Section 9
it focused on the frequency response and only considered harmonic
vibration loads such as sin(t), so their approach cannot be applied to summarizes the conclusions and future work.
loads with arbitrary waveforms. Particularly, the thermal load in their
numerical example was time-invariant. In actual design problems, loads
and thermal loads often have complex waveforms other than harmonic 2. Governing equations
vibration. Therefore, such methods need to be extended to consider
loads and thermal loads with arbitrary time-dependent profiles.
We considered a homogeneous isotropic thermoelastic material as-
Topology optimization can be applied to transient response prob-
suming a two-dimensional (2-D) plane stress condition. The Cauchy
lems with arbitrary load profiles. This method is highly versatile be-
cause it can be applied to load profiles that cannot be dealt with stress 𝝈 is defined as
by frequency response problems; the equations of motion are directly ( )
𝝈 = 𝑪 𝑚 𝜺 − 𝜺𝑡ℎ , (1)
solved by using the step-by-step integration method. However, this
leads to an optimization problem that depends on all deformation where 𝜺 and 𝜺𝑡ℎ are the strain and thermal strain, respectively. For a
histories considered in the analysis, which complicates the sensitiv-
2-D problem, the thermal strain is defined as
ity analysis required to obtain the structure. Two sensitivity analysis
methods have been proposed: differentiate-then-discretize approach [8] 𝜺𝑡ℎ = 𝛼𝑚 𝛥𝑇 𝑰, (2)
and discretize-then-differentiate approach [25]. Generally, transient
response analysis requires discretization in the time and space domains. where 𝛼𝑚 is the coefficient of thermal expansion, 𝛥𝑇 is the temperature
The differentiate-then-discretize approach applies a sensitivity analysis difference from the reference temperature for the thermal stress, and
on the equations of motion before the discretization in the time do- 𝑰 = (1, 1, 0)T . 𝑪 𝑚 is the elastic stiffness matrix and can be divided into
main. By contrast, the discretize-then-differentiate approach applies the Young’s modulus 𝐸𝑚 and Young’s modulus-independent term 𝑪: ̄
sensitivity analysis on the equations of motion after discretization in
the time domain. Jacob et al. [26] discussed the difference between ̄
𝑪 𝑚 = 𝐸𝑚 𝑪. (3)
2
S. Ogawa and T. Yamada Finite Elements in Analysis & Design 201 (2022) 103695
Eqs. (2) and (3) can be used to rewrite Eq. (1) as follows: unconditional stability condition. By substituting Eqs. (13) and (14)
( ) into Eq. (7), we can express the equations of motion as
𝝈 = 𝐸𝑚 𝑪̄ 𝜺 − 𝛼𝑚 𝛥𝑇 𝑰 ( ) ( )
1 𝛾 ( ) 1 1
= 𝐸𝑚 𝑪𝜺 ̄
̄ − 𝐸𝑚 𝛼𝑚 𝑪𝛥𝑇 𝑰 𝑴+ 𝑪 𝒖𝑖 − 𝒖𝑖−1 + 𝑲𝒖𝑖 − 𝑴𝒗𝑖−1 − − 1 𝑴𝒂𝑖−1
𝛽𝛥𝑡 2 𝛽𝛥𝑡 𝛽𝛥𝑡 2𝛽
̄ − 𝛯𝑚 𝑪𝛥𝑇
= 𝐸𝑚 𝑪𝜺 ̄ 𝑰. (4) ( ) ( )
𝛾 𝛾 ext th
+ 1− 𝑪𝒗𝑖−1 + 1 − 𝛥𝑡𝑪𝒂𝑖−1 = 𝑭 𝑖 + 𝑭 𝑖 . (15)
We can obtain the thermal stress coefficient 𝛯𝑚 [9] by multiplying 𝛽 2𝛽
Young’s modulus 𝐸𝑚 by the thermal expansion coefficient 𝛼𝑚 . The The displacement 𝒖𝑖 of the next time step 𝑖 can be obtained by using
governing equations are the following equations of motion: the variables of the time step 𝑖 − 1.
⎧ −𝜌 𝒂 + ∇ ⋅ 𝝈 = 𝟎 on 𝛺,
⎪ 𝑚
3. Topology optimization
⎨ 𝒖=𝒖 on 𝛤𝑢 , (5)
⎪ 𝒕=𝒕 on 𝛤𝑡 .
⎩ 3.1. Concept of topology optimization considering two-phase materials
𝛺 is the domain occupied by the object, 𝛤𝑢 is the boundary where the
displacement 𝒖 is constrained, and 𝒖̄ is the specified displacement given Topology optimization for two-phase materials can be treated as
as a condition. 𝛤𝑡 represents the boundary where the external load acts, determining the optimal arrangement of two types of materials in a
𝒕 is the surface force vector, and 𝒕̄ is the surface force vector given as a fixed design domain 𝛺𝑑 . The domains occupied by material 1 and
condition. 𝒂 is the acceleration, and 𝜌𝑚 is the mass density. The weak material 2 can be denoted as 𝛺1 and 𝛺2 , respectively. The distribution
form of Eq. (5) can be written as of material 2 at an arbitrary position 𝒙 inside the fixed design domain
𝛺𝑑 is represented by the following characteristic function 𝜒 (𝒙):
𝜌𝑚 𝒂 ⋅ 𝛿𝒖 d𝛺 + 𝝈 ∶ 𝛿𝜺 d𝛺 = 𝒕̄ ⋅ 𝛿𝒖 d𝛤 , (6) {
∫𝛺 ∫𝛺 ∫𝛤𝑡
1 if 𝒙 ∈ 𝛺2
𝜒 (𝒙) = (16)
𝛿𝒖 and 𝛿𝜺 are the virtual displacement and strain, respectively. The 0 if 𝒙 ∈ 𝛺1 ,
constraint 𝛿𝒖 = 0 is satisfied at the displacement-defining boundary 𝛤𝑢 .
where the relation 𝛺𝑑 = 𝛺1 ∪ 𝛺2 is satisfied. The characteristic func-
Note that 𝝈 in Eq. (6) is the Cauchy stress tensor. In finite element
tion 𝜒 (𝒙) has no conditions other than integrability, so the topology
analysis, the equations of motion can be expressed as a matrix by
optimization problem is ill-posed because it allows infinitely small sub-
discretizing Eq. (6) in the spatial domain:
structures. This issue arises from the use of this characteristic function
𝑴𝒂 + 𝑲𝒖 = 𝑭 ext + 𝑭 th , (7) to represent the material distribution, regardless of the objective or
constraint function. The ill-posed problem needs to be replaced with
where 𝑴 is the mass matrix, 𝑭 is the external force vector, and 𝑭 th
ext
a well-posed problem that has a solution.
is the thermal load vector. The shape function 𝑵 and B-matrix 𝑩 used
in the finite element method can be defined as follows:
3.2. RAMP method and filtering method
𝑴= 𝑵 T 𝜌𝑚 𝑵 d𝛺, (8)
∫𝛺
We adopted the density method to replace the ill-posed problem
𝑲= 𝑩 T 𝐸𝑚 𝑪𝑩
̄ d𝛺, (9) with a well-posed problem. Based on the work by Gao et al. [9], we
∫𝛺
used the RAMP method [10] for two-phase materials to specify the ma-
ext
𝑭 (𝑡) = 𝑵 T 𝒕̄ (𝑡) d𝛤 , (10) terial distribution. To determine the arrangement of two materials with
∫𝛤𝑡
different physical properties, we can define the filling rate occupied by
𝑭 th (𝑡) = 𝑩 T 𝛯𝑚 𝑪𝛥𝑇
̄ (𝑡) 𝑰 d𝛺. (11) material 2 at position 𝒙 as the design variable 𝒔 (0 ≤ 𝑠 (𝒙) ≤ 1). In other
∫𝛺 words, if 𝑠 (𝒙) = 0 at position 𝒙, the domain is filled with material 1; if
In transient response analysis, the external force vector 𝑭 ext and ther- 𝑠 (𝒙) = 1, the domain is filled with material 2.
mal load vector 𝑭 th depend on the time 𝑡. Furthermore, if we consider To address the possibility of deriving fine structures including the
the damping force proportional to the velocity 𝒗, the equations of checkerboard phenomenon [29], we adopted the density filter [30].
motion given in Eq. (7) can be extended as follows: In finite element analysis, the design variable 𝑠 (𝒙) can be defined
for each finite element. In this case, we used the design variable
𝑴𝒂 + 𝑪𝒗 + 𝑲𝒖 = 𝑭 ext + 𝑭 th , (12) 𝑠𝑒 (𝑒 = 1, 2, … , 𝑁) of the 𝑒th element discretized by 𝑁 finite elements
where 𝑪 is the damping matrix. In this study, we assumed that there to average the design variable with the surrounding values:
∑
is no damping term for simplicity and set 𝑪 = 𝟎. In the follow- 𝑤𝑗 𝑠𝑗
ing formulations, however, we include the damping matrix 𝑪 for 𝑗∈N𝑒
generality. 𝜌𝑒 (𝒔) = ∑ , (17)
𝑤𝑗
We employed the step-by-step integration method to solve the 𝑗∈N𝑒
equations of motion for our proposed topology optimization method
considering transient response behavior. Therefore, we apply one of the 𝜌𝑒 (𝒔) is the element density depending on the design variable 𝒔, and
implicit approaches, the Newmark 𝛽 method [28], to the discretization N𝑒 is the set of elements that exist in the domain where the distance
in the time direction. The advantage of this method is that it is an between the centers of an element and element 𝑒 is less than the filter
unconditionally stable time integration method and can be analyzed radius 𝑟f il . We define the weight coefficient 𝑤𝑗 as
by giving large time increments. In this case, we assume that the 𝑟f il − 𝑟𝑗
acceleration 𝒂 and velocity 𝒗 are given by 𝑤𝑗 = , (18)
𝑟f il
( )
1 ( ) 1 1 𝑟𝑗 is the distance between the centers of elements 𝑒 and 𝑗. In the RAMP
𝒂𝑖 = 𝒖𝑖 − 𝒖𝑖−1 − 𝒗𝑖−1 − − 1 𝒂𝑖−1 , (13)
𝛽𝛥𝑡 2 𝛽𝛥𝑡 2𝛽 method, Young’s modulus 𝐸𝑚 , thermal stress coefficient 𝛯𝑚 , and mass
( ) ( )
𝛾 ( ) 𝛾 𝛾 density 𝜌𝑚 are required to solve the equations of motion and are defined
𝒗𝑖 = 𝒖 − 𝒖𝑖−1 + 1 − 𝒗𝑖−1 + 1 − 𝛥𝑡𝒂𝑖−1 . (14)
𝛽𝛥𝑡 𝑖 𝛽 2𝛽 as functions of the element density 𝜌𝑒 :
{ ( ( )) ( )
Here, 𝑖 is the discretized time step, and 𝛥𝑡 is the time increment. 𝛽 and 1 − 𝑓𝑅1 𝜌𝑒 𝐸𝑚1 + 𝑓𝑅1 𝜌𝑒 𝐸2 𝐸𝑚1 ≤ 𝐸𝑚2
𝐸𝑚 = ( ) ( ( )) (19)
𝛾 are constants, which we set to 𝛽 = 0.25 and 𝛾 = 0.5 to realize the 𝑓𝑅2 𝜌𝑒 𝐸𝑚1 + 1 − 𝑓𝑅2 𝜌𝑒 𝐸𝑚2 𝐸𝑚1 > 𝐸𝑚2 ,
3
S. Ogawa and T. Yamada Finite Elements in Analysis & Design 201 (2022) 103695
{ ( ( )) ( )
1 − 𝑓𝑅1 𝜌𝑒 𝛯𝑚1 + 𝑓𝑅1 𝜌𝑒 𝛯𝑚2 𝛯𝑚1 ≤ 𝛯𝑚2
𝛯𝑚 = ( ) ( ( )) (20)
𝑓𝑅2 𝜌𝑒 𝛯𝑚1 + 1 − 𝑓𝑅2 𝜌𝑒 𝛯𝑚2 𝛯𝑚1 > 𝛯𝑚2 ,
( )
𝜌𝑚 = 1 − 𝜌𝑒 𝜌𝑚1 + 𝜌𝑒 𝜌𝑚2 , (21)
where (⋅)𝑚1 and (⋅)𝑚2 are material properties specific to material 1 and
material 2, respectively. 𝑓𝑅1 and 𝑓𝑅2 are interpolation functions in the
RAMP method and are defined as follows:
( ) 𝜌𝑒
𝑓𝑅1 𝜌𝑒 = ( ), (22)
1 + 𝑝 R 1 − 𝜌𝑒
( ) 1 − 𝜌𝑒
𝑓𝑅2 𝜌𝑒 = , (23)
1 + 𝑝 R 𝜌𝑒
𝑝𝑅 is the penalty coefficient of the RAMP method and was set to
𝑝𝑅 = 8. By defining the material properties as functions of the element
density 𝜌𝑒 (𝒔), the shape of the object in the fixed design domain can
be reflected in the dynamic analysis. In the topology optimization for
transient response problems, considering a single material, an unstable
phenomenon called spurious mode occurs. Usually, the stiffness is 0
in a void. However, in the fixed design domain, a small stiffness must
be given to the elements in the void domain for the convenience of Fig. 1. Profiles of interpolation functions..
analysis. This minim stiffness results in fictitious vibrations. Since it
affects the optimization results, the interpolation function of the mate-
rial properties should be considered to avoid this issue [25]. However, the p-norm obtained at each time in the time domain, the evaluation
spurious modes are not an issue in this study. The reason for this is that function becomes the sum of the maximum stresses occurring each time
we are dealing with a two-phase material, and even relatively less rigid and maintains the following relationship:
materials have much higher stiffness than a void. 𝑇
𝑣𝑀
𝑝 → ∞ ⇒ 𝑓𝑠 → 𝜎𝑚𝑎𝑥 dt. (29)
4. Multi-objective optimization problem for deformation and max- ∫0
imum stress minimization Minimizing 𝑓𝑠 obtains the optimal structure for minimizing the max-
imum stress for all times covered in the analysis. However, because
4.1. Evaluation function 1: Squared norm of dynamic compliance 𝑝 → ∞ cannot be treated numerically, we used the large value of 𝑝 = 8.
The density method obtains a grayscale (intermediate density do-
We defined the squared norm of dynamic compliance as the eval- main) where the design variable is 0 < 𝑠𝑒 < 1. With the RAMP method,
uation function for the deformation under both an external force 𝑭 ext the dynamic analysis considers Young’s modulus 𝐸𝑚 as a small value in
and thermal load 𝑭 th , the grayscale compared with the case of a linear interpolation function,
[ 𝑇𝑓 ]1∕2 as shown in Fig. 1. Consequently, directly using this stress to evaluate
{ T ( ext )}2
𝑓𝑐 = 𝒖 𝑭 + 𝑭 th dt , (24) the objective function leads to inappropriate results inside the grayscale
∫0
because the stress is considered small in the domain of concern. To
where 𝑡 is the time and 𝑇𝑓 is the last time considered in the analysis. avoid this problem, the qp-reduction method [5] is considered, which
This evaluation function can be minimized to obtain a structure that gives an interpolation function of the following equation:
minimizes the deformation that occurs during the time covered by
the analysis. In the case of transient response problems, the condition 𝑔(𝜌) = 𝜌1∕2 . (30)
( )
𝒖T 𝑭 ext + 𝑭 th > 0 does not always hold. Therefore, we used the
As shown in Fig. 1, regularization of the von Mises stress by the
squared norm to ensure that the evaluation function is a minimization
function 𝑔(𝜌) allows a higher stress in the grayscale than the RAMP
problem.
method, which avoids the above problem. However, the qp-reduction
method is based on the assumption of a single material and cannot
4.2. Evaluation function 2: p-norm for the von Mises stress
be applied to two-phase materials. Furthermore, the stress obtained
by dynamic analysis in the grayscale is theoretically different from
We defined the p-norm with respect to the von Mises stress as the
the stress evaluated by the p-norm. Therefore, a different method is
evaluation function for the maximum stress on an object:
necessary for grayscale elimination.
𝑇𝑓
𝑓𝑠 = 𝜎 𝑃 𝑁 dt, (25)
∫0 4.3. Evaluation function 3: Explicit penalization term
{ }1∕𝑝
𝑃𝑁
( 𝑣𝑀 )𝑝
𝜎 = 𝜎 d𝛺 , (26)
∫𝛺 One of the disadvantages of the density method is the grayscale
remaining in the optimization result. Particularly, the grayscale prob-
𝜎 𝑣𝑀 is the von Mises stress. In the case of a 2-D plane stress condition,
lem has been reported for topology optimization considering thermal
the von Mises stress can be expressed in vector notation as the following
( )T stress [9]. A large grayscale may remain in the solutions to stress min-
components, where the Cauchy stress is denoted by 𝝈 = 𝜎𝑥 , 𝜎𝑦 , 𝜏𝑥𝑦 :
imization problems. We applied the explicit penalization term [19,27]
√
𝜎 𝑣𝑀 = 𝜎𝑥2 + 𝜎𝑦2 − 𝜎𝑥 𝜎𝑦 + 3𝜏𝑥𝑦
2 , (27) to eliminate the grayscale:
( )
∫ 𝜌 − 𝜌𝑚𝑖𝑛 (1 − 𝜌) d𝛺
𝑝 is the exponent of the p-norm, and the following holds [31]: 𝑓𝑝 = 𝛺 , (31)
∫𝛺 d𝛺
𝑝 → ∞ ⇒ 𝜎 𝑃 𝑁 → 𝜎𝑚𝑎𝑥
𝑣𝑀
. (28)
where 𝜌𝑚𝑖𝑛 is the minimum value of the element density. This eval-
In other words, we can apply the p-norm to evaluate the maximum uation function has a value only in the grayscale 𝜌𝑚𝑖𝑛 < 𝜌 < 1. By
stress in the object domain with a single global evaluation function minimizing this function, we can obtain a structure that removes the
rather than local evaluation functions. Furthermore, by integrating grayscale.
4
S. Ogawa and T. Yamada Finite Elements in Analysis & Design 201 (2022) 103695
4.4. Formulation of the optimization problem Based on Eqs. (40)–(42), 𝑹𝑖 = 𝟎, 𝑯 𝑖 = 𝟎, and 𝑳𝑖 = 𝟎. Thus, 𝑓̃ and 𝑓
are equivalent. The objective function can be partially differentiated by
We defined the topology optimization of the transient response the element density 𝜌𝑗 considering the dependency of each variable:
problem through the following equations:
𝜕 𝑓̃ 𝜕𝑓 ∑ 𝜕𝑓 𝜕𝒖𝑖 𝑛
( ) 𝑓 𝑓 𝑓𝑝 = +
min𝑓 𝒔, 𝝆, 𝒖𝑛 , … , 𝒖0 = 𝑤𝑐 𝑐 + 𝑤𝑠 𝑠 + 𝑤𝑝 , (32) 𝜕𝜌𝑗 𝜕𝜌𝑗 𝑖=0 𝜕𝒖𝑖 𝜕𝜌𝑗
𝒔 𝑓𝑐0 𝑓𝑠0 𝑓𝑝0
(
∑𝑛
𝜕𝑹𝑖 𝜕𝑹𝑖 𝜕𝒖𝑖 𝜕𝑹𝑖 𝜕𝒖𝑖−1
s.t.ℎ (𝒔) = 𝜌 d𝛺 − 𝑉lim ≤ 0, (33) + 𝝀T𝑖 + +
∫𝛺𝑑 𝑖=0
𝜕𝜌 𝑗 𝜕𝒖𝑖 𝜕𝜌 𝑗 𝜕𝒖𝑖−1 𝜕𝜌𝑗
( ) )
𝑹𝑖 𝒔, 𝝆, 𝒖𝑖 , 𝒖𝑖−1 , 𝒗𝑖−1 , 𝒂𝑖−1 = 𝟎, (34) 𝜕𝑹𝑖 𝜕𝒗𝑖−1 𝜕𝑹𝑖 𝜕𝒂𝑖−1
+ +
( ) 𝜕𝒗𝑖−1 𝜕𝜌𝑗 𝜕𝒂𝑖−1 𝜕𝜌𝑗
𝑯 𝑖 𝒖𝑖 , 𝒖𝑖−1 , 𝒗𝑖−1 , 𝒂𝑖 , 𝒂𝑖−1 = 𝟎, (35) (
( ) ∑𝑛
𝜕𝑯 𝑖 𝜕𝒖 𝑖 𝜕𝑯 𝑖 𝜕𝒖𝑖−1
𝑳𝑖 𝒖𝑖 , 𝒖𝑖−1 , 𝒗𝑖 , 𝒗𝑖−1 , 𝒂𝑖−1 = 𝟎, (36) + 𝝁T𝑖 +
𝑖=0
𝜕𝒖 𝑖 𝜕𝜌 𝑗 𝜕𝒖 𝑖−1 𝜕𝜌𝑗
0 < 𝑠𝑒 ≤ 1. (37) )
𝜕𝑯 𝑖 𝜕𝒗𝑖−1 𝜕𝑯 𝑖 𝜕𝒂𝑖 𝜕𝑯 𝑖 𝜕𝒂𝑖−1
+ + +
We introduce the weight coefficients 𝑤𝑐 , 𝑤𝑠 , and 𝑤𝑝 and define the 𝜕𝒗𝑖−1 𝜕𝜌𝑗 𝜕𝒂𝑖 𝜕𝜌𝑗 𝜕𝒂𝑖−1 𝜕𝜌𝑗
(
objective function 𝑓 as the weighted sum function of 𝑓𝑐 , 𝑓𝑠 , and 𝑓𝑝 . ∑
𝑛
𝜕𝑳𝑖 𝜕𝒖𝑖 𝜕𝑳𝑖 𝜕𝒖𝑖−1
+ 𝜼T𝑖 +
This allows us to control the optimal structure according to the weight 𝜕𝒖𝑖 𝜕𝜌𝑗 𝜕𝒖𝑖−1 𝜕𝜌𝑗
𝑖=0
coefficients. 𝑓𝑐0 , 𝑓𝑠0 , and 𝑓𝑝0 are the values of the objective function for )
𝜕𝑳𝑖 𝜕𝒗𝑖 𝜕𝑳𝑖 𝜕𝒗𝑖−1 𝜕𝑳𝑖 𝜕𝒂𝑖−1
the initial structure. The evaluation functions 𝑓𝑐 and 𝑓𝑠 are discretized + + + . (46)
𝜕𝒗𝑖 𝜕𝜌𝑗 𝜕𝒗𝑖−1 𝜕𝜌𝑗 𝜕𝒂𝑖−1 𝜕𝜌𝑗
and evaluated in the time domain as follows:
∑
𝑛
𝑓= 𝑓 𝑖 𝑤𝑖 . (38) 𝜕 𝑓̃ 𝜕𝑓 ∑𝑛
𝜕𝑹𝑖
𝑖=0 = + 𝝀T
𝜕𝜌𝑗 𝜕𝜌𝑗 𝑖=0 𝑖 𝜕𝜌𝑗
where 𝑛 is the total number of analysis steps, 𝑓𝑖 is the evaluation ( )
𝜕𝑓 𝜕𝑹𝑛 𝜕𝑯 𝑛 𝜕𝑳𝑛 𝜕𝒖𝑛
function, and 𝑤𝑖 is the weight coefficient of the integration at time + + 𝝀T𝑛 + 𝝁T𝑛 + 𝜼T𝑛
𝜕𝒖𝑛 𝜕𝒖𝑛 𝜕𝒖𝑛 𝜕𝒖𝑛 𝜕𝜌𝑗
step 𝑖. We adopted the highly accurate trapezoidal integration, so
𝑛−1 (
∑
we assumed the time increment 𝛥𝑡 as constant and set the weight 𝜕𝑓 𝜕𝑹 𝑖 𝜕𝑹 𝑖+1
+ + 𝝀T𝑖 + 𝝀T𝑖+1
coefficient 𝑤𝑖 as follows [26,27]: 𝜕𝒖𝑖 𝜕𝒖𝑖 𝜕𝒖𝑖
𝑖=0
)
2𝑤0 = 𝑤1 = ⋯ = 𝑤𝑛−1 = 2𝑤𝑛 = 𝛥𝑡. (39) 𝜕𝑯 𝑖 𝜕𝑯 𝑖+1 𝜕𝑳𝑖 𝜕𝑳𝑖+1 𝜕𝒖𝑖
+𝝁T𝑖 + 𝝁T𝑖+1 + 𝜼T𝑖 + 𝜼T𝑖+1
𝜕𝒖𝑖 𝜕𝒖𝑖 𝜕𝒖𝑖 𝜕𝒖𝑖 𝜕𝜌𝑗
𝐻 is a constraint function for the volume 𝑉lim . 𝑹𝑖 , 𝑯 𝑖 , and 𝑳𝑖 are given 𝜕𝑳𝑛 𝜕𝒗𝑛
by Eqs. (13)–(15) and can be solved with Newmark 𝛽 method at time +𝜼T𝑛
𝜕𝒗𝑛 𝜕𝜌𝑗
step 𝑖: ( )
( ) ∑
𝑛−1
𝜕𝑹𝑖+1 𝜕𝑯 𝑖+1 𝜕𝑳𝑖 𝜕𝑳𝑖+1 𝜕𝒗𝑖
1 𝛾 ( ) + 𝝀T𝑖+1 + 𝝁T𝑖+1 + 𝜼T𝑖 + 𝜼T𝑖+1
𝑹𝑖 = 𝑴+ 𝑪 𝒖𝑖 − 𝒖𝑖−1 + 𝑲𝒖𝑖 − 𝑭 ext 𝑖 − 𝑭𝑖
th
𝜕𝒗𝑖 𝜕𝒗𝑖 𝜕𝒗𝑖 𝜕𝒗𝑖 𝜕𝜌𝑗
𝛽𝛥𝑡2 𝛽𝛥𝑡 𝑖=0
( ) 𝜕𝑯 𝑛 𝜕𝒂𝑛
−
1
𝑴𝒗𝑖−1 −
1
− 1 𝑴𝒂𝑖−1 +𝝁T𝑛
𝛽𝛥𝑡 2𝛽 𝜕𝒂𝑛 𝜕𝜌𝑗
( ) ( ) ( )
𝛾 𝛾 ∑
𝑛−1
𝜕𝑹𝑖+1 𝜕𝑯 𝑖 𝜕𝑯 𝑖+1 𝜕𝑳𝑖+1 𝜕𝒗𝑖
+ 1− 𝑪𝒗𝑖−1 + 1 − 𝛥𝑡𝑪𝒂𝑖−1 , (40) + 𝝀T𝑖+1 + 𝝁T𝑖 + 𝝁T𝑖+1 + 𝜼T𝑖+1 ,(47)
𝛽 2𝛽 𝜕𝒂𝑖 𝜕𝒂𝑖 𝜕𝒂𝑖 𝜕𝒂𝑖 𝜕𝜌𝑗
( ) 𝑖=0
1 ( ) 1 1
𝑯 𝑖 = −𝒂𝑖 + 𝒖𝑖 − 𝒖𝑖−1 − 𝒗 − − 1 𝒂𝑖−1 , (41) where 𝜕𝒖𝑖 ∕𝜕𝜌𝑗 , 𝜕𝒗𝑖 ∕𝜕𝜌𝑗 , and 𝜕𝒂𝑖 ∕𝜕𝜌𝑗 are implicit derivative terms with
𝛽𝛥𝑡2 𝛽𝛥𝑡 𝑖−1 2𝛽
( ) ( ) high computational costs when calculated directly. To avoid direct
𝛾 ( ) 𝛾 𝛾
𝑳𝑖 = −𝒗𝑖 + 𝒖 − 𝒖𝑖−1 + 1 − 𝒗𝑖−1 + 1 − 𝛥𝑡𝒂𝑖−1 . (42) computation, the adjoint equations can be solved to obtain the adjoint
𝛽𝛥𝑡 𝑖 𝛽 2𝛽
variables 𝝀, 𝝁, and 𝜼, where the coefficients of the related terms are
5. Sensitivity analysis zero.
By arranging the coefficients of the implicit derivative terms of
We conducted a sensitivity analysis on the objective function given Eq. (47) for time step 𝑛, we obtain the following adjoint equations:
by Eq. (32). Because the design variable 𝒔 is transformed into the
elemental density by the density filter in Eq. (17), it can be obtained 𝝁𝑛 = 𝟎, (48)
with the chain rule: 𝜼𝑛 = 𝟎, (49)
𝜕𝑓 ∑ 𝜕𝑓 𝜕𝜌𝑗 𝜕𝑹𝑛 𝜕𝑓 𝜕𝑯 𝑛 𝜕𝑳𝑛
= . (43) 𝝀 =− − 𝝁 − 𝜼 . (50)
𝜕𝑠𝑒 𝑗∈M 𝜕𝜌𝑗 𝜕𝑠𝑒
𝑒
𝜕𝒖𝑛 𝑛 𝜕𝒖𝑛 𝜕𝒖𝑛 𝑛 𝜕𝒖𝑛 𝑛
We obtain 𝜕𝜌𝑗 ∕𝜕𝑠𝑒 directly from Eq. (17): Similarly, the coefficients of the implicit derivative terms for steps
𝜕𝜌𝑗 0 ≤ 𝑖 ≤ 𝑛 − 1 can be rearranged to obtain the following equations:
𝑤
= ∑𝑒 . (44)
𝜕𝑠𝑒 𝑤𝑘 𝜕𝑓 𝜕𝑹𝑖 𝜕𝑹𝑖+1 𝜕𝑯 𝑖
+ 𝝀T𝑖 + 𝝀T𝑖+1 + 𝝁T𝑖
𝑘∈M𝑗 𝜕𝒖𝑖 𝜕𝒖𝑖 𝜕𝒖𝑖 𝜕𝒖𝑖
This leaves 𝜕𝑓 ∕𝜕𝜌𝑗 , which we determined by using the adjoint variable T 𝜕𝑯 𝑖+1 T 𝜕𝑳𝑖 T 𝜕𝑳𝑖+1
+𝝁𝑖+1 + 𝜼𝑖 + 𝜼𝑖+1 = 𝟎, (51)
𝜕𝒖𝑖 𝜕𝒖𝑖 𝜕𝒖𝑖
method. We can introduce the three adjoint variables 𝝀, 𝝁, and 𝜼 for
𝜕𝑹𝑖+1 𝜕𝑯 𝑖+1 𝜕𝑳𝑖 𝜕𝑳𝑖+1
the equations of motion and rewrite the objective function as follows: 𝝀T𝑖+1 + 𝝁T𝑖+1 + 𝜼T𝑖 + 𝜼T𝑖+1 = 𝟎, (52)
𝜕𝒗𝑖 𝜕𝒗𝑖 𝜕𝒗𝑖 𝜕𝒗𝑖
∑
𝑛 ∑
𝑛 ∑
𝑛
𝜕𝑹𝑖+1 𝜕𝑯 𝑖 𝜕𝑯 𝑖+1 𝜕𝑳𝑖+1
𝑓̃ = 𝑓 + 𝝀T𝑖 𝑹𝑖 + 𝝁T𝑖 𝑯 𝑖 + 𝜼T𝑖 𝑳𝑖 . (45) 𝝀T𝑖+1 + 𝝁T𝑖 + 𝝁T𝑖+1 + 𝜼T𝑖+1 = 𝟎. (53)
𝑖=0 𝑖=0 𝑖=0 𝜕𝒂𝑖 𝜕𝒂𝑖 𝜕𝒂𝑖 𝜕𝒂𝑖
5
S. Ogawa and T. Yamada Finite Elements in Analysis & Design 201 (2022) 103695
𝜕𝑓𝑝
Based on Eqs. (41) and (42), Eqs. (51)–(53) can be summarized by using = 𝟎. (75)
𝜕𝑳𝑖 ∕𝜕𝒗𝑖 = −1 and 𝜕𝑯 𝑖 ∕𝜕𝒂𝑖 = −1: 𝜕𝒖𝑖
𝜕𝑹𝑖 𝜕𝑴 𝜕𝑪 𝜕𝑲 𝜕𝑭 ext
𝑖 𝜕𝑭 th
𝑖
6.1. Finite difference method
= 𝒂𝑖 + 𝒗𝑖 + 𝒖𝑖 − − . (71)
𝜕𝜌𝑗 𝜕𝜌𝑗 𝜕𝜌𝑗 𝜕𝜌𝑗 𝜕𝜌𝑗 𝜕𝜌𝑗
The finite difference method calculates the sensitivity by using the
We then obtain the remaining terms 𝜕𝑓 ∕𝜕𝜌𝑗 and 𝜕𝑓 ∕𝜕𝒖 as follows. First, central difference of the evaluation function:
𝜕𝑓 ∕𝜕𝒖𝑖 is given by
𝜕𝑓 𝑓 (𝒔 + 𝛥̃𝒔) − 𝑓 (𝒔 − 𝛥̃𝒔)
𝑤𝑝 𝜕𝑓𝑝 = (85)
𝜕𝑓 𝑤 𝜕𝑓𝑐 𝑤 𝜕𝑓𝑠 𝜕𝑠𝑒 2𝛥𝑠𝑒
= 𝑐 + 𝑠 + . (72)
𝜕𝒖𝑖 𝑓𝑐0 𝜕𝒖𝑖 𝑓𝑠0 𝜕𝒖𝑖 𝑓𝑝0 𝜕𝒖𝑖 𝛥𝑠̃𝑒 = 𝛿𝑒𝑗 𝛥𝑠𝑗 (86)
The sensitivity of each evaluation function to the displacement 𝒖𝑖 is
where 𝛿𝑒𝑗 is Kronecker’s delta and 𝛥𝑠𝑒 is a small change in the design
given by
variable. 𝛥̃𝒔 is a vector in which only the e-numbered component has
𝜕𝑓𝑐 1 −1∕2 { T ( ext )} ( ext ) 𝛥𝑠𝑒 and all other components are zero. The design variable set for each
= 𝑓𝑐 𝒖𝑖 𝑭 𝑖 + 𝑭 th
𝑖 𝑭 𝑖 + 𝑭 th
𝑖 𝑤𝑖 , (73)
𝜕𝒖𝑖 2 element is changed sequentially by 𝛥𝑠𝑒 , and the evaluation function is
𝜕𝑓𝑠 ( PN )1−𝑝 ( 𝑣𝑀 )𝑝−1 𝜕𝜎𝑖𝑣𝑀 𝜕𝝈 𝑖 calculated each time. Therefore, as the number of elements increases,
= 𝜎𝑖 𝜎 d𝛺 𝑤𝑖 , (74) the computation cost becomes enormous and impractical. Although the
𝜕𝒖𝑖 ∫𝛺 𝑖 𝜕𝝈 𝑖 𝜕𝒖𝑖
6
S. Ogawa and T. Yamada Finite Elements in Analysis & Design 201 (2022) 103695
Fig. 2. Analysis model: (a) 1/2 sector model used in actual numerical calculations, (b) Entire structure.
Fig. 3. Load profile: (a) transient load, (b) temperature difference from reference temperature for thermal stress that defines the thermal load.
method contains a small numerical error, it is accurate enough to obtain 𝛥𝑇𝑚𝑎𝑥 = 20 K and 𝛥𝑡 = 1.0 × 10−2 s, (b) 𝛥𝑇𝑚𝑎𝑥 = 100 K and 𝛥𝑡 = 1.0 × 10−2
appropriate optimization results. Therefore, it is used as a benchmark s, and (d) 𝛥𝑇𝑚𝑎𝑥 = 20 K and 𝛥𝑡 = 1.0 × 10−2 s. We uniformly assigned se
to validate the proposed method. = 0.5 to the design variables in the fixed design domain.
Figs. 4 and 5 compare the sensitivities obtained by the proposed
6.2. Initial stage condition method and finite difference method. The horizontal axis is the number
of finite elements, and the vertical axis is the sensitivity of each ele-
Fig. 2(a) shows a 2-D rectangular model with dimensions of 800 mm ment. The sensitivities obtained by the finite difference method and the
× 500 mm to represent the fixed design domain. We verified this model proposed method were consistent for both the squared norm of dynamic
under the condition of a small number of elements by dividing it into
compliance and p-norm, regardless of the magnitude of the thermal
8 × 5 = 40 quadrilateral elements. To consider two-phase materials
load and time increments. We confirmed that the proposed method
with different properties, we set the material properties to Young’s
calculates the sensitivity with high accuracy regardless of the analysis
modulus of 𝐸𝑚1 = 70 GPa and 𝐸𝑚2 = 430 GPa, coefficients of thermal
conditions. Therefore, the formulated equations were validated.
expansion of 𝛼𝑚1 = 2.5×10−5 and 𝛼𝑚2 = 3.1×10−6 , and mass densities of
𝜌𝑚1 = 2700 kg∕m3 and 𝜌𝑚2 = 4000 kg∕m3 . Material 1 had a low stiffness
and high thermal expansion, and Material 2 had a high stiffness and low 6.3. Random condition
thermal expansion to simulate aluminum and ceramic, respectively. For
the boundary conditions, 𝛤𝑢 was completely constrained at both ends,
( ) During topology optimization calculations, it is always necessary
and 𝛤𝑡 was subjected to an external load of 𝐹 (𝑡) = 2.0×104 sin3 4𝜋𝑡∕𝑇𝑓
to evaluate the exact sensitivity. In general, as the optimization cal-
𝑁 with a positive downward direction, as shown in Fig. 3(a). Because
culation proceeds, a distribution of the design variables in the fixed
this external load is not a harmonic vibration, it cannot be considered
design domain occurs. Therefore, the validity of the sensitivity, in this
as a frequency response problem and can only be solved as a transient
case, is examined. The geometry, boundary conditions, and material
response problem. We applied a time-varying thermal load to the entire
properties of the analytical model are assumed to be the same as those
analytical model by uniformly applying a time-varying 𝛥𝑇 (𝑡) with the
of the initial stage condition. The time increment is assumed to be
profile shown in Fig. 3(b). This thermal load simulates the operation
of a machine from the initial stage to the steady state. We assumed the 𝛥𝑡 = 1.0 × 10−4 s. The design variables are randomly distributed in the
initial conditions as quiescent, so 𝒂0 = 𝟎, 𝒗0 = 𝟎 and 𝒖0 = 𝟎 at analysis range of 0 ≤ 𝑠𝑒 ≤ 1 to assume general conditions. Fig. 6 shows the
step 𝑖 = 0. model under consideration.
To confirm that the sensitivity of each evaluation function could be Figs. 7 and 8 are the results of each study with the squared norm
calculated correctly regardless of the temperature difference 𝛥𝑇 and of dynamic compliance and p-norm set as the objective function. In all
time increment 𝛥𝑡, we uniformly divided the time domain into n = 50 results, the distribution is the same as the sensitivity obtained from the
steps. Then, we evaluated the squared norm of dynamic compliance finite difference method. This indicates that the proposed method can
(𝑤𝑐 = 1.0, 𝑤𝑠 = 0.0, 𝑤𝑝 = 0.0) and p-norm (𝑤𝑐 = 0.0, 𝑤𝑠 = 1.0, 𝑤𝑝 = 0.0) accurately derive the sensitivity even when the design variables have
under four conditions: (a) 𝛥𝑇𝑚𝑎𝑥 = 20 K and 𝛥𝑡 = 1.0 × 10−2 s, (b) arbitrary distributions, demonstrating the validity of the method.
7
S. Ogawa and T. Yamada Finite Elements in Analysis & Design 201 (2022) 103695
Fig. 4. Initial stage condition: sensitivity validation results of the squared norm of dynamic compliance (𝑤𝑐 = 1.0, 𝑤𝑠 = 0.0, 𝑤𝑝 = 0.0): (a) 𝛥𝑇𝑚𝑎𝑥 = 20 K and 𝛥𝑡 = 1.0 × 10−2 s, (b)
𝛥𝑇𝑚𝑎𝑥 = 20 K and 𝛥𝑡 = 1.0 × 10−4 s, (c) 𝛥𝑇𝑚𝑎𝑥 = 100 K and 𝛥𝑡 = 1.0 × 10−2 s, (d) 𝛥𝑇𝑚𝑎𝑥 = 100 K and 𝛥𝑡 = 1.0 × 10−4 s.
Fig. 5. Initial stage condition: sensitivity validation results of p-norm (𝑤𝑐 = 0.0, 𝑤𝑠 = 1.0, 𝑤𝑝 = 0.0): (a) 𝛥𝑇𝑚𝑎𝑥 = 20 K and 𝛥𝑡 = 1.0 × 10−2 s, (b) 𝛥𝑇𝑚𝑎𝑥 = 20 K and 𝛥𝑡 = 1.0 × 10−4 s,
(c) 𝛥𝑇𝑚𝑎𝑥 = 100 K and 𝛥𝑡 = 1.0 × 10−2 s, (d) 𝛥𝑇𝑚𝑎𝑥 = 100 K and 𝛥𝑡 = 1.0 × 10−4 s.
8
S. Ogawa and T. Yamada Finite Elements in Analysis & Design 201 (2022) 103695
2 was less than 50% of the volume of the fixed design domain. For
the optimization algorithm, we used the globally convergent method
of moving asymptotes (GCMMA) [32] to update the design variables
based on the sensitivity analysis.
7.2. Influence of the magnitude of the thermal load on the optimal structure
Fig. 7. Random condition: sensitivity validation results of the squared norm of dynamic compliance (𝑤𝑐 = 1.0, 𝑤𝑠 = 0.0, 𝑤𝑝 = 0.0): (a) 𝛥𝑇𝑚𝑎𝑥 = 20 K, (b) 𝛥𝑇𝑚𝑎𝑥 = 100 K.
Fig. 8. Random condition: sensitivity validation results of p-norm (𝑤𝑐 = 0.0, 𝑤𝑠 = 1.0, 𝑤𝑝 = 0.0): (a) 𝛥𝑇𝑚𝑎𝑥 = 20 K, (b) 𝛥𝑇𝑚𝑎𝑥 = 100 K.
9
S. Ogawa and T. Yamada Finite Elements in Analysis & Design 201 (2022) 103695
Fig. 9. Optimization results and history of the objective function with the square norm of dynamic compliance: (a) 𝛥𝑇𝑚𝑎𝑥 = 0 K, (b) 𝛥𝑇𝑚𝑎𝑥 = 20 K, (c) 𝛥𝑇𝑚𝑎𝑥 = 50 K, (d) 𝛥𝑇𝑚𝑎𝑥 = 100
K.
Fig. 10. Optimization results and history of the objective function with p-norm: (a) 𝛥𝑇𝑚𝑎𝑥 = 0 K, (b) 𝛥𝑇𝑚𝑎𝑥 = 20 K, (c) 𝛥𝑇𝑚𝑎𝑥 = 50 K, (d) 𝛥𝑇𝑚𝑎𝑥 = 100 K.
(d) is the condition where the effect of thermal load is the largest. It right side shows that the p-norm values converged to a small value rel-
is a complex optimization problem because the thermal load is one of ative to the initial structure, which indicates that the optimal structure
the structure-dependent loads. Therefore, convergence may deteriorate. was derived under the given conditions. These results confirmed that
Improvement of the optimization algorithm may solve this problem, but the stress minimization problem for two-phase materials can be solved
it will be future work. by considering the explicit penalization term in addition to the p-norm
Fig. 10 shows the optimization results when the p-norm (𝑤𝑐 = 0.0, to eliminate grayscale.
𝑤𝑠 = 1.0, 𝑤𝑝 = 0.0) was the objective function. The left side shows
the optimal structure for conditions (a)–(d), and the right side shows 7.3. Effect of the exponent of the p-norm
the transition of the p-norm value at each optimization step under each
condition. The value of the p-norm was smaller than that of the initial As shown in Eq. (28), by setting exponent 𝑝 → ∞, the maximum
structure under all conditions and finally converged. This indicates stress can be evaluated. However, from the viewpoint of stability of
that we performed the optimization calculation properly. However, numerical calculation, it is necessary to set a finite value. The accuracy
the optimization results did not return an appropriate solution for the with which the p-norm can evaluate the maximum stress depends on
composition of two-phase materials because the grayscale (shown in the exponent 𝑝 of the p-norm. Therefore, the effect of this parameter
green) remained over a wide range. With the RAMP method, Young’s on the reproducibility of the maximum stress will be investigated. The
modulus in the grayscale was smaller than the linearly complemented analytical conditions were set to 𝛥𝑇𝑚𝑎𝑥 = 20 K to consider the effects of
element density, so the stress in the grayscale was underestimated by mechanical and thermal loading. Other boundary conditions, material
the p-norm calculation. This resulted in an inappropriate structure for properties, number of analytical steps, and time increments are the
minimizing the maximum stress. same as in Section 7.2. The objective function was set to the p-norm
Fig. 11 shows the optimization results under the same conditions (𝑤𝑐 = 0, 𝑤𝑠 = 1.0, 𝑤𝑝 = 0.25). The p-norm histories calculated from
with 𝑤𝑝 = 0.25 and considering the explicit penalization term to remove the exponents 𝑝 = 4, 𝑝 = 8, 𝑝 = 16, and 𝑝 = 32 are compared with
the grayscale. The results under all conditions (a)–(d) are binarized the maximum von Mises stress obtained directly from the analysis. The
structures of material 1 and material 2 with almost no grayscale. The design variable 𝑠𝑒 = 0.5 is assumed to be a uniformly given condition.
10
S. Ogawa and T. Yamada Finite Elements in Analysis & Design 201 (2022) 103695
Fig. 11. Optimization results and history of the objective function with p-norm (the explicit penalization term 𝑤𝑝 = 0.25): (a) 𝛥𝑇𝑚𝑎𝑥 = 0 K, (b) 𝛥𝑇𝑚𝑎𝑥 = 20 K, (c) 𝛥𝑇𝑚𝑎𝑥 = 50 K, (d)
𝛥𝑇𝑚𝑎𝑥 = 100 K.
7.4. Effect of the number of analysis steps and the time increment size
11
S. Ogawa and T. Yamada Finite Elements in Analysis & Design 201 (2022) 103695
Fig. 13. Left: Optimization results of the objective function with p-norm (the explicit penalization term 𝑤𝑝 = 0.25), Right: time history of maximum von Mises stress: (a) 𝑝 = 8,
(b) 𝑝 = 16.
Fig. 14. Optimization results of the objective function with the square norm of dynamic compliance: (a) 𝑛 = 50 steps, (b) 𝑛 = 100 steps.
Fig. 15. Left: Optimization results of the objective function with p-norm, Right: time history of maximum von Mises stress: (a) 𝑛 = 50 steps, (b) 𝑛 = 100 steps.
𝑤𝑝 = 0.25, (c) 𝑤𝑝 = 0.5, (d) 𝑤𝑝 = 0.75, and (e) 𝑤𝑝 = 1.0. The evaluation 7.6. Multi-objective optimization: Minimizing the deformation and maxi-
functions were the squared norm of dynamic compliance and p-norm. mum stress
Figs. 16 and 17 show the optimization results when the squared
norm of dynamic compliance was the objective function. Figs. 18 and To confirm that the proposed method obtains the optimal structure
19 show the results when the p-norm was the objective function. The that minimizes both the deformation and maximum stress considering
left side of each figure shows the optimization results obtained under thermoelastic behavior, we performed a multi-objective optimization
each condition, and the right side shows the relationship between the with the squared norm of dynamic compliance and p-norm as evalua-
weight coefficient of the explicit penalization term and the calculated tion functions and the weight coefficients 𝑤𝑐 and 𝑤𝑝 as parameters. We
value of the evaluation function for each optimal structure. In this set the maximum temperature difference to two conditions: 𝛥𝑇𝑚𝑎𝑥 = 20
study, we focused on minimizing the objective function, so structures K and 𝛥𝑇𝑚𝑎𝑥 = 100 K. As discussed in Section 7.3, we set the weight
with smaller values were superior. Under all conditions, increasing the coefficient of the explicit penalization term to 𝑤𝑝 = 0.25 to eliminate
weight coefficient of the explicit penalization term increased the value grayscale. We considered five conditions for the weight coefficients 𝑤𝑐
returned by each evaluation, and the performance decreased. In other and 𝑤𝑝 : (a) 𝑤𝑐 = 1.0 and 𝑤𝑠 = 0.0, (b) 𝑤𝑐 = 0.75 and 𝑤𝑠 = 0.25, (c)
words, if the weight of the explicit penalization term is too large, a 𝑤𝑐 = 0.5 and 𝑤𝑠 = 0.5, (d) 𝑤𝑐 = 0.25 and 𝑤𝑠 = 0.75, and (e) 𝑤𝑐 = 0.0
suboptimal structure for other evaluation functions is obtained. Because and 𝑤𝑠 = 1.0.
the purpose of the explicit penalization term is to remove grayscale, the Figs. 20 and 21 show the optimization results under the thermal
minimum weight factor 𝑤𝑝 that ensures no grayscale remains is desired. load conditions 𝛥𝑇𝑚𝑎𝑥 = 20 K and 𝛥𝑇𝑚𝑎𝑥 = 100 K, respectively.
For this example, we found that a weight coefficient of 𝑤𝑝 = 0.25 was The optimal structure is shown on the left of each figure, and
sufficient to remove the grayscale. the relationship between the squared norm of dynamic compliance
12
S. Ogawa and T. Yamada Finite Elements in Analysis & Design 201 (2022) 103695
Fig. 16. Left: Optimization results for minimizing the squared norm of dynamic compliance, Right: Relationship between the weights of the explicit penalization term and the
evaluation function values (𝛥𝑇𝑚𝑎𝑥 = 20 K): (a) 𝑤𝑝 = 0, (b) 𝑤𝑝 = 0.25, (c) 𝑤𝑝 = 0.5, (d) 𝑤𝑝 = 0.75, (e) 𝑤𝑝 = 1.0.
Fig. 17. Left: Optimization results for minimizing the squared norm of dynamic compliance, Right: Relationship between the weights of the explicit penalization term and the
evaluation function values (𝛥𝑇𝑚𝑎𝑥 = 100 K): (a) 𝑤𝑝 = 0, (b) 𝑤𝑝 = 0.25, (c) 𝑤𝑝 = 0.5, (d) 𝑤𝑝 = 0.75, (e) 𝑤𝑝 = 1.0.
and the p-norm after convergence is shown on the right. The opti- 8. Numerical examples 2: L-shape model
mization results confirmed that intermediate optimal structures can be
obtained from minimizing the squared norm of the dynamic compliance 8.1. Analysis model and conditions
(i.e., condition (a)) and p-norm (i.e., condition (e)) depending on the ra-
tio of the two weight coefficients 𝑤𝑐 and 𝑤𝑠 . The results also indicated We examine whether the proposed method can achieve multi-
that the two evaluation functions have a tradeoff relationship. This objective optimization of displacement and maximum stress for dif-
implies that no solution can minimize both the deformation and maxi- ferent models. Fig. 22 shows the geometry and boundary conditions
mum stress simultaneously. Thus, this is a multi-objective optimization of the fixed design domain of the L-shape model. The number of
problem where the optimal solution depends on the priorities of the two elements is 6400. The top surface is assumed to be completely fixed.
objectives. In engineering applications, the desired optimal structure The mechanical load and thermal load profiles are the conditions of
can be obtained by setting the weight coefficients of the objective Fig. 3. The mechanical load is assumed to be positive in the downward
functions according to the required deformation and maximum stress. direction, and the thermal stress is assumed to be the maximum value
Thus, the proposed method can be applied to derive multi-objective of the temperature difference from the reference 𝛥𝑇𝑚𝑎𝑥 = 20 K. The
optimal solutions to dynamic problems considering the behavior of Young’s modulus of the material is assumed to be 𝐸𝑚1 = 2 GPa and
thermoelastic models. 𝐸𝑚2 = 430 GPa. The thermal expansion coefficients are 𝛼𝑚1 = 1.0 × 10−5
13
S. Ogawa and T. Yamada Finite Elements in Analysis & Design 201 (2022) 103695
Fig. 18. Left: Optimization results for minimizing the p-norm, Right: Relationship between the weights of the explicit penalization term and the evaluation function values
(𝛥𝑇𝑚𝑎𝑥 = 20 K): (a) 𝑤𝑝 = 0, (b) 𝑤𝑝 = 0.25, (c) 𝑤𝑝 = 0.5, (d) 𝑤𝑝 = 0.75, (e) 𝑤𝑝 = 1.0.
Fig. 19. Left: Optimization results for minimizing the p-norm, Right: Relationship between the weights of the explicit penalization term and the evaluation function values
(𝛥𝑇𝑚𝑎𝑥 = 100 K): (a) 𝑤𝑝 = 0, (b) 𝑤𝑝 = 0.25, (c) 𝑤𝑝 = 0.5, (d) 𝑤𝑝 = 0.75, (e) 𝑤𝑝 = 1.0.
and 𝛼𝑚2 = 3.1 × 10−6 . The mass densities are set to 𝜌𝑚1 = 20 kg∕m3 and 8.2. Results of multi-objective optimization
𝜌𝑚2 = 4000 kg∕m3 . The total number of analysis steps is set to 𝑛 = 50
steps, and the time increment is set to 𝛥𝑡 = 1.0 × 10−4 s. In this case, Fig. 23 shows the obtained optimal structures and Fig. 24 shows
𝑇𝑓 = 𝛥𝑡 × 𝑛 = 5.0 × 10−3 s. The initial conditions are the stationary the relationship between the square norm of dynamic compliance and
condition, 𝒂0 = 𝟎, 𝒗0 = 𝟎, and 𝒖0 = 𝟎. The objective function fixes the p-norm after computational convergence. In this study, the conditions
Explicit penalization weight as 𝑤𝑝 = 0.5. We vary the squared norm of of small-time increment and high frequency of loading are considered.
Therefore, in (a) minimization of the squared norm of dynamic com-
dynamic compliance weight 𝑤𝑐 and the p-norm weight 𝑤𝑠 to obtain the
pliance, to suppress the vibration displacement around the load-acting
multi-objective optimal solution. We derive the optimal configuration
surface, more rigid materials2 are placed in this domain. On the other
for four conditions: (a) 𝑤𝑐 = 1.0 and 𝑤𝑠 = 0.0, (b) 𝑤𝑐 = 0.75 and
hand, if the weight of the p-norm in the objective function is increased,
𝑤𝑠 = 0.25, (c) 𝑤𝑐 = 0.5 and 𝑤𝑠 = 0.5, (d) 𝑤𝑐 = 0.25 and 𝑤𝑠 = 0.75, and the structure becomes very complicated and it is difficult to consider
(e) 𝑤𝑐 = 0.0 and 𝑤𝑠 = 1.0. The constraints give a volume constraint the features. For this reason, we check Fig. 24 to determine the validity
that material 2 is less than 50% of the fixed design domain. The design of the obtained results. In condition (e), where the effect of p-norm
variables for the initial structure are uniformly set to 𝑠𝑒 = 0.5 for the is large, the value of p-norm is appropriately reduced, indicating that
entire model. The optimization algorithm is GCMMA [32]. it is correctly solved as a multi-objective optimization. In conclusion,
14
S. Ogawa and T. Yamada Finite Elements in Analysis & Design 201 (2022) 103695
Fig. 20. Multi-objective optimization results for the squared norm of dynamic compliance and p-norm (𝛥𝑇 = 20 K): (a) 𝑤𝑐 = 1.0 and 𝑤𝑠 = 0.0, (b) 𝑤𝑐 = 0.75 and 𝑤𝑠 = 0.25, (c)
𝑤𝑐 = 0.5 and 𝑤𝑠 = 0.5, (d) 𝑤𝑐 = 0.25 and 𝑤𝑠 = 0.75, (e) 𝑤𝑐 = 0.0 and 𝑤𝑠 = 1.0.
Fig. 21. Multi-objective optimization results for the squared norm of dynamic compliance and p-norm (𝛥𝑇 = 100 K): (a) 𝑤𝑐 = 1.0 and 𝑤𝑠 = 0.0, (b) 𝑤𝑐 = 0.75 and 𝑤𝑠 = 0.25, (c)
𝑤𝑐 = 0.5 and 𝑤𝑠 = 0.5, (d) 𝑤𝑐 = 0.25 and 𝑤𝑠 = 0.75, (e) 𝑤𝑐 = 0.0 and 𝑤𝑠 = 1.0.
we show that the optimal structure can be derived from the proposed • We introduce an objective function to eliminate the grayscale
method even for complex analytical conditions. domain and to control the deformation and maximum stress
because of external forces and thermal loads. This is a weighted
9. Conclusions sum function comprising three evaluation functions: the squared
norm of dynamic compliance, the p-norm, and the explicit penalty
We proposed a topology optimization method for transient response term. By adjusting the weight coefficients of the squared norm
problems considering the behavior of thermoelastic models to account
of dynamic compliance and p-norm, an optimal structure can
for time-dependent external and thermal loads. We present our findings
be derived for minimizing both the deformation and maximum
and future work below.
stress.
• Because deformation and stress are caused by the different ther- • When the p-norm is the objective function for minimizing the
mal expansion coefficients of multi-materials, topology optimiza- maximum stress, using the RAMP method to regularize Young’s
tion can be applied to finding the optimal configuration of two- modulus retains a large grayscale in the optimization results.
phase materials. This problem can be solved by adding the explicit penalization
15
S. Ogawa and T. Yamada Finite Elements in Analysis & Design 201 (2022) 103695
Fig. 23. Multi-objective optimization results for the squared norm of dynamic compliance and p-norm: (a) 𝑤𝑐 = 1.0 and 𝑤𝑠 = 0.0, (b) 𝑤𝑐 = 0.75 and 𝑤𝑠 = 0.25, (c) 𝑤𝑐 = 0.5 and
𝑤𝑠 = 0.5, (d) 𝑤𝑐 = 0.25 and 𝑤𝑠 = 0.75, (e) 𝑤𝑐 = 0.0 and 𝑤𝑠 = 1.0.
term to the objective function, which can be extended to two- • We presented numerical examples to show that an optimal struc-
phase material problems without any special relaxation of the ture can be derived for both the deformation and maximum
evaluation stress. The grayscale removal method applies to other stress. By controlling the weight coefficients of the objective
topology optimization problems and is highly versatile. Stress functions, the deformation and maximum stress can be priori-
minimization of two-phase materials has not been considered in tized as needed. We conducted a parametric study on the weight
previous studies and was addressed in this study. coefficients of the explicit penalization term, which confirmed
• We conducted a sensitivity analysis to consider the effects of that grayscale can be removed without significantly degrading the
time-dependent external forces and thermal loads. The topology performance regarding deformation and stress.
optimization of transient response problems requires a sensitivity • The scope of this study was limited to problems where the object
analysis that considers the deformation state at all times. We has a uniform temperature. To expand the scope of application,
used the discretize-then-differentiate approach for the sensitivity we need to couple topology optimization with the unsteady heat
analysis, which ensures high accuracy regardless of the time conduction problem. This will enable the optimization to consider
increment. We validated the proposed method validated by com- changes in temperature distribution depending on the structure.
paring its sensitivity values with those obtained from the finite Additionally, We identified an issue where the convergence of the
difference method. optimization becomes worse when the effect of thermal loading
16
S. Ogawa and T. Yamada Finite Elements in Analysis & Design 201 (2022) 103695
17