Accepted Manuscript

Stress-based topology optimization using bi-directional evolutionary

Liang Xia, Li Zhang, Qi Xia, Tielin Shi

PII: S0045-7825(18)30037-9
DOI: https://doi.org/10.1016/j.cma.2018.01.035
Reference: CMA 11755

To appear in: Comput. Methods Appl. Mech. Engrg.

Received date : 27 September 2017

Please cite this article as: L. Xia, L. Zhang, Q. Xia, T. Shi, Stress-based topology optimization
using bi-directional evolutionary structural optimization method, Comput. Methods Appl. Mech.
Engrg. (2018), https://doi.org/10.1016/j.cma.2018.01.035

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form.
Please note that during the production process errors may be discovered which could affect the
content, and all legal disclaimers that apply to the journal pertain.
*Manuscript
Click here to download Manuscript: 2017_CMAME_BESO_Stress.pdf Click here to view linked References

1
2
3
4
Stress-based topology optimization using bi-directional
5
6
evolutionary structural optimization method
7
8 Liang Xia∗, Li Zhang, Qi Xia, Tielin Shi∗
9
10 State Key Laboratory of Digital Manufacturing Equipment and Technology,
11 Huazhong University of Science and Technology, Wuhan 430074, China
12
13
14
15
16
17 Abstract
18
19 This work proposes an evolutionary topology optimization method for stress minimization design
20 using the bi-directional evolutionary structural optimization (BESO) method. The discrete na-
21 ture of the BESO method avoids naturally the well-known “singularity” problem in density-based
22
23 methods with degenerated materials. The p-norm stress aggregation scheme is adopted for the
24 measure of global stress level. A computationally efficient sensitivity number formulation is de-
25 rived from the adjoint sensitivity of the global stress measure. With regard to the highly nonlinear
26 stress behavior, both sensitivity numbers and topology variables are filtered to stabilize the opti-
27
28 mization procedure; meanwhile, the filtered sensitivity numbers are further stabilized with their
29 historical information. The method has been shown efficient, practical and easy-to-implement
30 through a series of 2D and 3D benchmark designs.
31
32 Keywords: Topology optimization, BESO, Stress minimization, Sensitivity analysis
33
34
35
36
37 1. Introduction
38
39 As an advanced and efficient design method, topology optimization has undergone a remark-
40 able development since the seminal paper by Bendsøe and Kikuchi [1] over the past decades
41 in both academic research [2, 3] and industrial applications [4]. Among all research subjects,
42
43 stress-based topology optimization has been recognized as a challenging problem and has been
44 continuously attracting research interests since the pioneering paper by Duysinx and Bendsøe [5].
45 Le et al. [6] has summarized three main challenges of stress-based topology optimization:
46 (i) the “singularity ” problem; (ii) the local nature of stress, and (iii) the highly nonlinear stress
47
48 behavior. The first challenge was reported early for truss layout designs in [7, 8]. In the case of
49 density-based methods, it refers to the fact that elements with low densities can present high stress
50 values, making the optimization algorithm incapable of removing them. Remedy schemes were
51
52
proposed later by relaxing stress constraints such that the element stress and density can decrease
53 simultaneously [5, 9–12]. The second challenge lies in fact that stress is a local quantity, which
54
55
∗
56 Corresponding author
57 Email addresses: [email protected] (Liang Xia), [email protected] (Tielin Shi)
58
59 Preprint submitted to Computer Methods in Applied Mechanics and Engineering January 17, 2018
60
61
62
63
64
65
 1
2
3
4
means that stress at each point of a continuum structure should be controlled in design. The local
5 nature of stress results in a large number of constraints and computationally demanding sensitivity
6 evaluation no matter using the direct method or the adjoint method [5]. A common strategy to
7 this dilemma is to approximate the maximum stress via global measures, such as the p-norm and
8
9 the Kresselmeier-Steinhauser (KS) functions [10, 13]. Global stress measures are computationally
10 efficient for sensitivity evaluation, however at the expense of losing adequate control of local stress
11 behavior. The third challenge is due to that fact that stress levels are highly sensitive to topology
12
changes, which is also a well-known difficulty in shape optimization designs [14]. In the case
13
14 of topology optimization, the highly nonlinear behavior is much more pronounced because stress
15 levels are drastically affected by its neighboring density changes, particularly in critical regions
16 such as sharp and reentrant corners. For this reason, a consistent density filtering versus sensitivity
17
18
filtering was adopted for stabilization consideration by Le et al. [6].
19 The local nature of stress and its highly nonlinear behavior are two common challenges in
20 stress-based topology designs, no matter which particular method is adopted. In contrast to the
21 last two common challenges, the “singularity” problem is however limited only to density-based
22
23 methods with the presence of intermediate densities. The “singularity” problem can in fact be
24 naturally avoided by using discrete topology optimization methods, such as level-set methods and
25 the ESO-type methods [15]. Within the level-set framework, early works accounting for stress
26
27
constraints in topology optimization were given by van Miegroet and Duysinx [16] and Allaire
28 and Jouve [17], followed by a vast number of researches targeted on the same topic using the
29 level-set method (e.g., [18–25]). The evolutionary structural optimization (ESO) method [15]
30 is another important branch of topology optimization. Among all its variants, the convergent
31
32 and mesh-independent bi-directional evolutionary structural optimization (BESO) method [26])
33 allowing both material removal and addition, has been widely adopted for both academic research
34 and engineering applications for its efficiency and robustness [27–30]. Though the ESO method
35 was originally developed upon a stress criterion by gradually removing lowly stressed materials,
36
37 the method or its variant BESO mthod has not yet been extended for stress minimization designs
38 to our best knowledge.
39 This paper aims to propose an efficient, practical and easy-to-implement approach for stress
40
41
minimization topology design through an extension of the BESO method. As compared to con-
42 tinuously defined density-based methods, the BESO method avoids naturally the “singularity”
43 problem for the sake of its discrete nature, resulting in clear physical interpretation and algorith-
44 mic advantages. With regard to the local nature of stress, the conventional p-norm global stress
45
46 measure is adopted for an approximation of the maximum stress. A computationally efficient
47 sensitivity number formulation is derived from the adjoint sensitivity of the global stress mea-
48 sure. With regard to the highly nonlinear stress behavior, both sensitivity numbers and topology
49 variables are filtered to stabilize the optimization procedure; meanwhile, the filtered sensitivity
50
51 numbers are further stabilized with their historical information.
52 This paper is organized as follows. Section 2 presents the stress minimization design model
53 including the p-norm stress measure and the definition of and the optimization problem. Section
54
55
3 gives the detailed sensitivity derivation for the global p-norm stress measure. Section 4 gives
56 the extended BESO variables update scheme. Section 5 validates the proposed method through a
57 series of 2D and 3D benchmark designs. Finally, conclusion and future perspectives are drawn.
58
59 2
60
61
62
63
64
65
 1
2
3
4
2. Stress minimization design
5
6 2.1. Effective stiffness and stress models
7 The design domain is discretized into nel finite elements and each element is assigned with a
8 topology design variable. The nel-dimensional vector containing the design variables is denoted
9
10 x = (x1 , x2 , . . . , xnel )T . Within the framework of the BESO method [27], the design variables take
11 values of either 0 or 1, i.e.
12 xi ∈ {0, 1}, i = 1, 2, . . . , nel. (1)
13
14 The design variables can be interpreted as a generalized material density or an indicator parameter
15 that the value of one corresponds to solid material, whereas zero corresponds to void.
16
17 The element stiffness can be straightforwardly linked to the associated topology design variable
18 in a linear manner that the element effective stiffness matrix is defined as
19
20 Di = xi D0 , (2)
21
22
23 where D0 is the stiffness matrix of solid material. In practice, the void is assigned with an extremely
24 low stiffness to avoid re-meshing and the singularity of the global stiffness.
25 Following [5], the overall stress criterion which aims to control the stress state at the micro-
26 scopic level for rank-2 layered composites is considered. For a power-law type material interpola-
27
28 tion model, the consistent model of the effective stress vector is given by
29
30 σi = (Di Bi ui )/xi = D0 Bi ui , (3)
31
32 where ui is the displacement vector of the i-th element and the matrix Bi relates the displacement
33
34 and the strain. By this definition, it is assumed that the element stress state is independent with
35 its associated design variable. Note that though the Ersatz material scheme is used to avoid re-
36 meshing, stresses are only evaluated for solid materials whilst are set directly to zero for voids.
37 Intuitively, this interpolation model for stress seems redundant for the discrete nature of the vari-
38
39 ables. However, the proper choice of an appropriate interpolation model is essential to sensitivity
40 analysis and consequently to the formulation of an effective sensitivity number.
41
42 2.2. Global p-norm stress measure
43
44 With regard to the local nature of stress, the conventional p-norm global stress measure is
45 adopted for an approximation of the maximum stress
46
47  nel 1/p
X 
=  σ p
48  ,
49 σPN vm,i (4)
50 i=1
51
52 where σvm,i is the von Mises stress at the centroid of the i-th element and p is the stress norm
53 parameter. It is known that the p-norm gives the average stress when p = 1 and approaches on the
54 other hand the maximum stress σmax when p → ∞.
55
56 Intuitively a higher value of p is preferable as it provides a more accurate approximation of the
57 maximum stress, however when p increases the p-norm problem becomes ill-conditioned. The
58
59 3
60
61
62
63
64
65
 1
2
3
4
higher value p takes, the severe oscillation happens during the optimization procedure since only
5 the peak stress is targeted by the algorithm while the others are not. Consequently, the value
6 of the norm parameter p must be appropriately chosen such that the problem smoothness is not
7 deteriorated and meanwhile the maximum stress can be adequately approximated. Please refer to
8
9 [10] for a detailed discussion on the property of p-norm global measure.
10 Without loss of generality, the element von Mises stress is assumed to be computed at the
11 centroid as  1/2
12
σvm = σT Vσ . (5)
13
14 Here, σ denotes the stress in the Voigt notation and V is the stress coefficient matrix
15
16  
17  1 −1/2 0
 
18 V = −1/2 1 0 (6)
 
19 0 0 3
20
21 in the 2D plane stress case and results in the following form
22
23 q
24 σvm = σ2xx + σ2yy − σ xx σyy + 3σ2xy , (7)
25
26
27 in which σ xx , σyy and σ xy are the components of the stress vector σ in the 2D case.
28
29 2.3. Optimization model
30
The mathematical model for the minimization of the p-norm global stress measure subject to
31
32 a material usage constraint is described as follows
33
34 min : σPN
x
35
subject to : KU = F (8)
36 P
37 : V(x) = xi vi = Vreq
38 : xi = 0 or 1, i = 1, . . . , nel,
39
40 Where K is the global stiffness matrix, U and F are the global displacement and force vectors,
41
42 respectively. ve is the volume of the i-th element, V(x) and Vreq are the total and required material
43 volumes, respectively.
44
45
46 3. Sensitivity analysis
47
48 In order to perform the topology optimization, the sensitivity of the global stress measure with
49 respect to the topology design variables needs to be provided. Note that by the BESO method,
50
sensitivities are evaluated only for solid elements whilst are set directly to zero for void elements.
51
52 According to the definition in (4), the derivative of σPN with respect to the j-th design variable
53 equals  nel 
54
∂σPN X p−1 ∂σvm,i 
1−p 
55 = σPN  σvm,i  , j = 1, . . . , nel. (9)
56 ∂x j i=1
∂x j
57
58
59 4
60
61
62
63
64
65
 1
2
3
4
By the definition of von Mises stress in (5), the derivative of element von Mises stress with
5 respect to the stress vector equals
6
∂σvm 1  T −1/2
7
= σ Vσ 2σT V = σ−1 T
vm σ V. (10)
8 ∂σ 2
9
10 With the above at hand, one obtains by the chain rule the derivative of the von Mises stress of the
11 i-th element with respect to the j-th design variable
12
13 ∂σvm,i T ∂σi
14 = σ−1
vm,i σi V . (11)
15 ∂x j ∂x j
16
17 It is known by the effective stress vector definition in (3) that the solid material stiffness ma-
18 trix D0 and the element strain-displacement matrix Bi are independent with the associated design
19 variable, the above derivative can be further written as
20
21 ∂σvm,i ∂U
T
22 = σ−1
vm,i σi VD0 Bi Li , (12)
23 ∂x j ∂x j
24
25 where the matrix Li gathers the nodal displacements of the i-th element from the global displace-
26 ment vector satisfying ui = Li U. It is assumed that the prescribed global force vector is indepen-
27 dent with the topology change. By differentiating both sides of the state balance equation, one
28
29 obtains
∂K ∂U
30 U+K = 0, (13)
31 ∂x j ∂x j
32
33 and therefore (12) can be further written as follows
34
∂σvm,i T −1 ∂K
35 = −σ−1
vm,i σi VD0 Bi Li K U. (14)
36 ∂x j ∂x j
37
38 Substituting (14) into (9), the sensitivity of the p-norm global stress measure can be further
39 expressed as  nel 
40
 X
41 ∂σPN
= −σ1−p  σ p−2 σT VD B L  K−1 ∂K U.

(15)
42 PN  vm,i i 0 i i
∂x j i=1
∂x j
43
44 With the solution of the following adjoint problem
45
46 nel
X
p−2
47 Kλ = σvm,i (D0 Bi Li )T Vσi , (16)
48
i=1
49
50 the sensitivity can be further simplified to
51
52 ∂σPN T ∂K
53 = −σ1−p
PN λ U. (17)
54 ∂x j ∂x j
55
56 The pseudo-load on the right-hand side of the adjoint problem (16) is assembled in analogy to
57 body forces.
58
59 5
60
61
62
63
64
65
 1
2
3
4
Recalling the effective stiffness model (2), the derivative of the global stiffness matrix with
5 respect to the j-th design variable equals
6
nel
7 ∂K X T ki
8 = Li Li = LTj k(0)
j L j, (18)
9 ∂x j i=1
∂x j
10
11
in which k j is the stiffness matrix of the j-th element and k(0)
j is the corresponding element stiffness
12
13 with solid material
14 Z Z
T (0)
15 k j = x j B j D0 B j dΩ j and k j = BTj D0 B j dΩ j (19)
16
17
18 where Ω j denotes the j-th element domain.
19 Substituting (18) into (17), the sensitivity of the p-norm global stress measure can be eventu-
20
21 ally evaluated as
∂σPN
22 = −σ1−p T (0)
PN λ j k j u j , (20)
23 ∂x j
24
25 where λ j is the vector of the adjoint nodal values of the j-th element.
26
27
28 4. Extended BESO variables update scheme
29
30 The standard convergent and mesh-independent BESO method initially developed in [26] and
31 recently summarized in [27] is extended in the following to improve the stability and robustness of
32 the design procedure in view of the highly nonlinear stress behavior. In particular, both sensitivity
33
34 numbers and topology variables are filtered to stabilize the optimization procedure; meanwhile,
35 the filtered sensitivity numbers are further stabilized with their historical information.
36 By the BESO method, the target volume of material usage V (l) at the current design iteration
37 (l-th) is determined by n o
38
39 V (l) = max Vreq , (1 − cer )V (l−1) , (21)
40
41
in which the evolutionary ratio cer determines the percentage of material to be removed from the
42 design of the previous iteration. Once the final required material volume usage Vreq is reached, the
43 optimization algorithm alters only the topology but keeps the material volume constant.
44 At each design iteration, the sensitivity numbers which denote the relative ranking of the el-
45
46 ement sensitivities are used to guide material removal and addition. When uniform meshes are
47 used, the sensitivity number for the global stress measure is defined as following using the ele-
48 ment sensitivity computed from (20)
49
50 ∂σPN
51 αi = −xi = xi σ1−p T (0)
PN λi ki ui , (22)
52 ∂xi
53
54 where xi serves as an indicator to simply ensure that sensitivity numbers are evaluated only for
55 solid elements (xi = 1), while the sensitivity numbers are zero for void elements (xi = 0).
56
57
58
59 6
60
61
62
63
64
65
 1
2
3
4
In order to avoid mesh-dependency and checkerboard patterns, sensitivity numbers are firstly
5 smoothed by means of a filtering scheme [31]
6 Pnel sen
j=1 wi j α j
7
8 αi = Pnel sen , (23)
9 j=1 wi j
10
11 where wsen
i j is a linear weight factor
12
13
14 wsen
i j = max(0, rsen − ∆(i, j)), (24)
15
16 determined according to the prescribed filter radius rsen and the element center-to-center distance
17 ∆(i, j) between the i-th and the j-th elements.
18
Due to the discrete nature of the BESO material model and the highly nonlinear stress behavior,
19
20 the current sensitivity numbers are averaged with their historical information of the last two design
21 iterations to improve the convergence
22
23 α(l) (l−1)
i + αi + α(l−2)
24 αi(l) ← i
, for l > 2. (25)
25 3
26
27 in which the superscript denotes the number of design iteration. Note that the above stabilization
28 scheme is different from the original one in the standard BESO method [26, 27] for a further
29 stabilization of the design procedure.
30
31 All elements are sorted into a vector according to their sensitivity number values. Then, the
32 update of the topology design variables is realized by means of two threshold parameters αthdel and
33 αth
add for material removal and addition, respectively [27, 32]
34
35  (l)
 0 if αi ≤ αth
del and xi = 1,
36 



37 xi(l+1) = 
 1 if αth (l)
add < αi and xi = 0,
(26)


 x (l)
38
i otherwise.
39
40
41 The present scheme indicates that solid elements are removed when their sensitivity numbers
42 are less than αth
del and void elements are recovered when their sensitivity numbers are greater than
43 αth
add . The parameters αth th
del and αadd are obtained from the following iterative algorithm:
44
45 1. Let αth th
add = αdel = αth , where the value αth is determined iteratively such that the required
46
47 material volume usage is met at the current iteration.
48 2. Compute the admission ratio car , which is defined as the volume of the recovered elements
49 divided by the total volume of the current iteration. If car ≤ cmax
ar , the maximum admission
50
51
ratio, skip the next steps; otherwise, αth
del and α th
add are redetermined in the next steps.
th
52 3. Determine αadd iteratively using only the sensitivity numbers of the void elements until the
53 maximum admission ratio is met, i.e., car ≈ cmaxar .
54 th
55 4. Determine αdel iteratively using only the sensitivity numbers of the solid elements until the
56 required material volume usage is met at the current iteration.
57
58
59 7
60
61
62
63
64
65
 1
2
3
4
It has been argued that due to the highly nonlinear stress behavior, it is more preferable to
5 filter densities rather than sensitivities for stress-based topology optimization [6]. However, it
6 needs to be emphasized that in the BESO method the sensitivity filtering scheme (23) serves
7 also for material recovery from void to solid by attributing filtered sensitivity numbers to design
8
9 variables that are associated to voids. Therefore, we propose to perform an additional filtering on
10 the topology variables using the same scheme as above
11
Pnel den
j=1 wi j x j
12
13 xi = Pnel den , (27)
14 j=1 wi j
15
16
17
where the linear weight factor wden
i j is defined as
18
19 wden
i j = max(0, rden − ∆(i, j)). (28)
20
21 with another filter radius rden . The resulting topology with intermediate densities is converted into
22
23 the discrete design through a postprocessing step the same as the above for sensitivity numbers
24 with a determined threshold value xth satisfying the current material volume usage constraint.
25 The filter radii for two filtering steps are defined satisfying rden > rsen > he , where he is the
26
typical element length. The sensitivity filter radius rsen should be set to a small value (2he ∼
27
28 3he ) depending on the mesh discretization resolution to ensure that the merit of material recovery
29 scheme is not suppressed. Meanwhile, the filter radius for topology variables takes a larger value
30 to ensure the convergence and mesh-independency of the design.
31
32
For the discrete nature of the BESO method, an heuristic convergence criterion on the variation
33 of a series of design iterations was conventional used [26, 27]. In the case of the current context,
34 such an heuristic criterion is not appropriate due to the highly nonlinear stress behavior. It is
35 known that the required number of design iterations to reach the target material volume the from
36
37 a full solid design can be estimated upon the evolutionary ratio cer . We propose to stop intuitively
38 the design cycle several more iterations after the target volume Vreq is reached and select one best
39 design among the final designs. The overall design algorithm is summarized as follows:
40
41 1. Define design domain, FE mesh and boundary conditions.
42
43
2. Set the initial topology full of solid material.
44 3. Define the BESO parameters: Vreq , cer , cmax
ar , rsen , rden .
45 4. Estimate the number of design iterations upon Vreq and cer .
46
47 5. Loop over design iterations:
48 (a) determine the target volume of the current iteration as (21)
49 (b) perform FE analysis with the current topology;
50
51 (c) evaluate sensitivity numbers according to (22).
52 (d) filter and average sensitivity numbers according to (23) and (25);
53 (e) update topology variables upon the algorithm in (26);
54
55
(f) filter topology variables upon (27);
56 (g) convert the filtered topology into discrete design;
57 6. End.
58
59 8
60
61
62
63
64
65
 1
2
3
4
5. Numerical examples
5
6
In this section, we show the performance of the proposed method through a series of 2D and
7 3D benchmark tests. In the first two 2D cases, quadrilateral bilinear elements with the plane stress
8 assumption are used. For the last 3D case, eight-node cubic elements is used. In all subsequent
9 cases, the solid material is assumed with Young’s modulus 1.0 MPa and Poisson’s ratio 0.3. All
10
11 parameters involved in the extended BESO method presented in Section 4 hold constant in all
12 following examples. The evolutionary rate cer determining the relative percentage of material to
13 be removed at each design iteration is set to cer = 2%. The maximum admission ratio corre-
14
sponding to the maximum percentage of recovered material that is allowed per iteration is set to
15
16 cmax
ar = 0.5%. The filter radius for sensitivity numbers is set to be twice of the typical finite element
17 size rsen = 2he . The filter radius for topology variables rden is to be precised in each design case de-
18 pending on the mesh resolution. For all following cases, the evolutionary design procedure starts
19
20
from an initial design guess full of solid material. As for the highly nonlinear stress behavior,
21 the convergence in terms of both objective and topology is difficult to guarantee. In all follow-
22 ing examples, we perform 10 additional optimization cycles once the required volume fraction is
23 attained and choose the design with the lowest stress concentration as the design result.
24
25
26 5.1. L-bracket design
27 The problem setting of the well-known L-bracket benchmark is illustrated in Figure 1 with the
28 characteristic dimensions. Square shaped bilinear elements of length he = 1 mm are used for the
29
30 discretization of the L-bracket (200 elements along the left vertical boundary). The top end of
31 the L-bracket is clamped, and the vertical load 4 N is distributed over eight nodes to avoid stress
32 concentration. The volume fraction of solid material to be preserved is set to 40%. The filter radius
33
for topology variables is set to rden = 6he .
34
35
36
37
80
38
39
40
41
42 200
43
44
45
80
46
47
48
49 200
50
51 Figure 1: Illustration of the L-bracket benchmark problem.
52
53 Figure 2(b)-(d) show the designs obtained using the stress norm parameters p = 2, 4, 6, respec-
54
55 tively. It can be observed that the larger value p takes, the lower the maximum von Mises stress
56 is achieved and the stress field becomes smoother. The designed topology for the case with p = 6
57 is in line with the reports for the same benchmark problem upon the literature (e.g., [6, 11, 33]).
58
59 9
60
61
62
63
64
65
 1
2
3
4
For the purpose of comparison, the stiffness maximization design of the L-bracket by the standard
5 BESO method [27, 29] is given in Figure 2(a). It can be noted that the stiffness design is quite
6 similar to the stress design for the case of Figure 2(b) with p = 2.
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27 (a) stiffness design (b) stress design with p = 2 (c) stress design with p = 4 (d) stress design with p = 6
28
29 Figure 2: Comparison of topology designs of the L-bracket for different stress norm parameters (the color
30 scales range from zero to the maximum von Mises stress for each case).
31
32
33
34
35
36
37
38
39
40
41 Iteration 1 Iteration 2 Iteration 5 Iteration 11 Iteration 20
42
43
44
45
46
47
48
49
50 Iteration 25 Iteration 29 Iteration 34 Iteration 41 Iteration 47
51
52 Figure 3: The topological evolution with stress field plot for the case of Figure 2(d) with p = 6 (the color
53 scales range from zero to the maximum stress for each design iteration).
54
55
56 It needs to be emphasized that it required only 47 iterations by the proposed method to achieve
57 the designs in Figure 2(b)-(d) for all three cases, which is much more efficient than the other
58
59 10
60
61
62
63
64
65
 1
2
3
4
methods as reported (up to several hundred or even a thousand iterations required, e.g., [19, 33]) . It
5 has been reported in the literature (e.g.,[6]) that larger p values (up to 12) can be taken when using
6 mathematical programming algorithms provided that many more design iterations are required.
7 By our tests the value of p cannot go higher than 6 for the heuristic update algorithm applied in
8
9 the BESO method, otherwise the design fails due to severe oscillations.
10 The detailed evolution of the topology together with the von Mises stress field for the case of
11 Figure 2(d) is given in Figure 3. As can be observed, the stress concentration appears initially
12
at the reentrant corner when the design domain is full of solid material. The maximum stress
13
14 reduces from the initial 2.09 MPa to 1.44 MPa during the first design iterations through the material
15 removal at the reentrant corner. Along with the evolutionary material removal procedure, the
16 maximum stress holds constant during the intermediate design iterations and rebounds sightly at
17
18
the final phase. The stress design with p = 6 in Figure 2(d) has effectively reduced the maximum
19 stress for 38% from 2.44 MPa to 1.51 MPa compared to the stiffness design in Figure 2(a).
20
21 5.2. MBB beam design with one pre-existing crack notch
22
23
The problem setting of the MBB beam benchmark with one pre-existing crack notch is illus-
24 trated in Figure 4. The beam of ratio 3 by 1 is discretized with 180 by 60 square shaped bilinear
25 elements of length he = 1 mm. The left and right bottom edge corners are constrained and a
26 vertical load 10 N is applied at the middle of the top edge. The vertical load is distributed over
27
28 11 nodes to avoid stress concentration. With the intention to validate the proposed method, the
29 beam is assumed to be with a pre-existing vertical crack notch of length 15 mm at the middle of
30 the bottom edge. The volume fraction of solid material to be preserved is set to 50%. The filter
31
radius for topology variables is set to rden = 4he .
32
33
34
35 10
36
37
60
38
39
40
15
41
42
180
43
44
Figure 4: Illustration of a MBB beam problem with one pre-existing crack notch.
45
46
47 Figure 5(a) gives the stiffness maximization design of the pre-cracked beam. Due to the crack
48 notch, a reentrant corner appears in stiffness maximized topology result, which results in a stress
49
50
concentration of 2.43 MPa. For the purpose of illustration, the stresses at the two supporting points
51 are set to zero. Figure 5(b)-(d) show the designs obtained using the stress norm parameters p =
52 2, 4, 6, respectively. Similar to the previous case, the maximum von Mises stress is lowered and the
53 stress field is more evenly distributed as a result of the evolving typology when a lager p is taken. It
54
55 takes less than 45 iterations to achieve the deigns of Figure 5(b)-(d), which again demonstrates the
56 high efficiency of the proposed method in handling stress-related topology designs. The highest
57 value p can take for this benchmark is 7, which results in a design similar to the case of Figure
58
59 11
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
(a) stiffness design
9
10
11
12
13
14
15
(b) stress design with p = 2
16
17
18
19
20
21
22 (c) stress design with p = 4
23
24
25
26
27
28
29 (d) stress design with p = 6
30
31 Figure 5: Comparison of topology designs of the MBB beam with a pre-existing crack notch for different
32
33 stress norm parameters (the color scales range from zero to the maximum von Mises stress for each case).
34
35
36 5(d) with a slightly lower maximum stress 1.56 MPa. When p takes a higher value surpassing 7,
37 the design fails the same as the L-bracket design case due to severe oscillations.
38 The detailed evolution of the topology together with the von Mises stress field for the case of
39
40 Figure 5(d) is given in Figure 6. The stress concentration appears initially at the crack tip. The
41 maximum stress decreases gradually from the initial 2.20 MPa to the eventual 1.59 MPa, while
42 the normalized stress holds constant initially around 2.90 MPa and grows up during the final stage
43 to 3.46 MPa in the end. The maximum stress has been reduced for almost 35% from 2.43 MPa to
44
45 1.59 MPa by comparing the stress design with p = 6 in Figure 5(d) against the stiffness design in
46 Figure 5(a).
47
48 5.3. Cantilever design
49
50 The problem setting of the cantilever benchmark is illustrated in Figure 7. The dimensions of
51 the cantilever is 200 × 100 mm2 . Square shaped bilinear elements of length he = 1 mm are used for
52 the discretization of the cantilever. The left end of the cantilever is clamped, and the vertical load 5
53
54 N is distributed over 10 nodes at the right-end portion of the top edge to avoid stress concentration.
55 The volume fraction of solid material to be preserved is set to 50%. The filter radius for topology
56 variables is set to rden = 5he .
57
58
59 12
60
61
62
63
64
65
 1
2
3
4
5
6
7
8 Iteration 1 Iteration 2 Iteration 6
9
10
11
12
13
14
15 Iteration 10 Iteration 16 Iteration 20
16
17
18
19
20
21
Iteration 30 Iteration 38 Iteration 42
22
23 Figure 6: The topological evolution with stress field plot for the case of Figure 5(d) with p = 6 (the color
24
25 scales range from zero to the maximum stress for each design iteration).
26
27
28
9
29
30
31 100
32
33
34
35
200
36
37
Figure 7: Illustration of a cantilever problem.
38
39
40
41
42
43
44
45
46
47
48
49
50
51 (a) stiffness design (b) stress design with p = 2 (c) stress design with p = 4 (d) stress design with p = 6
52
53 Figure 8: Comparison of topology designs of the cantilever for different stress norm parameters (the color
54 scales range from zero to the maximum von Mises stress for each case).
55
56
57
58
59 13
60
61
62
63
64
65
 1
2
3
4
5
6
7
8 Iteration 1 Iteration 2 Iteration 6 Iteration 10 Iteration 16
9
10
11
12
13
14
15 Iteration 21 Iteration 25 Iteration 29 Iteration 33 Iteration 36
16
17 Figure 9: The topological evolution with stress field plot for the case of Figure 8(d) with p = 6 (the color
18 scales range from zero to the maximum stress for each design iteration).
19
20
21 Figure 8(b)-(d) show the designs obtained using the stress norm parameters p = 2, 4, 6, re-
22 spectively. Similar to previous cases, the maximum von Mises stress is lowered and the stress
23
24 field is more evenly distributed when a lager p is taken. The stiffness maximization design of the
25 cantilever is also given in Figure 8(a), whose topology is similar to the stress design with p = 2.
26 It takes only 36 iterations to achieve the deigns of Figure 8(b) and (d). For the case of Figure 8(c),
27
28
8 more iterations are used due to a removal of a structural branch at the final design phase. It has
29 also been tested that for the cantilever benchmark design the highest value that p can take is also
30 6, otherwise the design fails the same as previous design cases due to severe oscillations.
31 The detailed evolution of the topology together with the von Mises stress field for the case of
32
33 Figure 8(d) is given in Figure 9. Unlike the L-bracket design case, the stress concentration appears
34 initially at the two corners of the clamped end. Meanwhile, despite the fact that solid material
35 is gradually removed, the maximum stress at the clamped end holds constant during the first 20
36 iterations. Since then, the maximum stress gradually increases along with the further removal of
37
38 solid material until the end. It can be seen from Figure 8(b)-(d) that the supporting topology has
39 an essential effect on the stress concentration at the clamped end. It is more stable and easier to
40 converge for the cantilever problem compared to the L-bracket problem, as the position of stress
41
42
concentration keeps unmoved. The maximum stress has been reduced for almost 30% from 1.50
43 MPa to 1.07 MPa by comparing the stress design with p = 6 in Figure 8(d) against the stiffness
44 design in Figure 8(a).
45
46 5.4. 3D cantilever design
47
48 To further demonstrate the effectiveness of the developed method, a 3D cantilever problem
49 as illustrated in Figure 10 is tested for stress minimization design. The design domain is of the
50
51
dimensions 80 × 40 × 10 mm3 and is discretized into uniformly meshed eight-node cubic elements
52 of length he = 1 mm. Similar to the 2D case, the left surface of the volume is clamped while the
53 vertical load 25 N is distributed on 25 nodes at the right end of the top surface to avoid the stress
54 concentration. The volume fraction of the solid material to be preserved is set to 30%. The filter
55
56 radius for topology variables is set to rden = 2he .
57
58
59 14
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
40
11
12
13 80
14 10
15
16 Figure 10: Illustration of a 3D cantilever problem.
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34 (a) stiffness design (b) stress design with p = 2 (c) stress design with p = 4 (d) stress design with p = 6
35
36 Figure 11: Comparison of topology designs of the 3D cantilever for different stress norm parameters (the
37 color scales range from zero to the maximum von Mises stress for each case).
38
39
40 The designs obtained using the stress norm parameters p = 2, 4, 6 are given in Figure 11(b)-
41
42 (d), respectively. To have a better visualization, the jagged topology designs are postprocessed
43 into smoother representations while the stress distributions are plot on the jagged designs. Similar
44 to the 2D cases, the larger value p takes, the lower the maximum von Mises stress and the more
45 evenly distributed the stress field. The stiffness maximization design given in Figure 2(a) has a
46
47 similar structural topology to the stress design in Figure 2(b) for the case of p = 2. The number
48 of design iterations required is determined by the evolutionary rate (cer = 2% determining the
49 material removal ratio for each design iteration and the total volume to be removed (70% in this
50
51
case). It took 62-64 iterations for all design cases in Figure 11, which again validates the high
52 efficiency of the developed method.
53 Figure 12 shows the detailed evolution of the topology together with the von Mises stress field
54 for the case of Figure 11(d). Compared the 2D cantilever case, the stress concentration initially
55
56 appearing at the two corner edges of the clamped surface hols constant (σmax ) for many more
57 iterations (until the 52-th iteration) along with the material removal. It is until the last iterations
58
59 15
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13 Iteration 1 Iteration 3 Iteration 11 Iteration 20
14
15
16
17
18
19
20
21
22
23
24
25 Iteration 30 Iteration 40 Iteration 52 Iteration 63
26
27 Figure 12: The topological evolution with stress field plot for the case of Figure 11(d) with p = 6 (the
28 color scales range from zero to the maximum stress for each design iteration).
29
30
31 the maximum stress increases due to further removal of solid material. The stress design with
32 p = 6 has reduced the maximum stress for 15% from 1.88 MPa to 1.60 MPa against the stiffness
33
34
design in Figure 8(a). It is expected that the proposed method shall be more efficient for 3D
35 designs if the mesh would be refined yielding more accurate stress estimation. It is seen from
36 Figure 13 that the introduction of the third dimension brings more design freedom and results in
37 hollow region between the front and back surfaces.
38
39
40
41
42
43
44
45
46
47
48
49
50 Figure 13: The topology designs and stress distribution field of Figure 11(d) from alternative view angles.
51
52
53
54 6. Conclusions
55
56 This paper has proposed an efficient, practical and easy-to-implement approach for stress min-
57 imization topology design using an extended BESO method. The approach inherits the merit of
58
59 16
60
61
62
63
64
65
 1
2
3
4
the discrete nature of the BESO method and thus avoids naturally the “singularity” problem in
5 stress-based topology designs. The maximum stress is approximately estimated using the conven-
6 tional p-norm global stress measure. A computationally efficient sensitivity number formulation
7 is particularly developed from the adjoint sensitivity of the global stress measure. With regard to
8
9 the highly nonlinear stress behavior, we proposed to filter both sensitivity numbers and topology
10 variables stabilize the optimization procedure.
11 It has been shown that the proposed method is high efficient in stress-based topology designs
12
in terms of the design iterations required against the other methods reported in the literature. The
13
14 adopted p-norm global stress measure can well restrain the whole structural stress level. The
15 larger value the stress normalization parameter p takes, the lower is the maximum von Mises
16 stress and the smoother is the stress distribution field. Despite the fact that the design fails due
17
18
to severe oscillations when p goes higher than 6 due to the applied heuristic update algorithm, it
19 has been shown that it is sufficient enough to limit the stress level by choosing p within the range
20 of 4 and 6. Thanks for its high efficiency and the simplicity of implementation, the method is
21 expected to be used for practical engineering design problems as demonstrated by the 3D cantilever
22
23 design case. In addition, the method is also expected to be extended and applied for the design
24 of material microstructures to restrain the stress level at the lower scale within the context of
25 multiscale modeling and design.
26
27
28 Acknowledgements
29
30 This work is supported by the National Natural Science Foundation of China (51705165,
31 51405170, 51790171, 5171101743), the Natural Science Foundation for Distinguished Young
32 Scholars of Hubei province of China (2017CFA044), and the National High Technology Research
33
34 and Development Program (“863” Program) of China (2015AA042505).
35
36
References
37
38 [1] M. P. Bendsøe, N. Kikuchi, Generating optimal topologies in structural design using a homogenization method,
39 Computer Methods in Applied Mechanics and Engineering 71 (2) (1988) 197–224.
40 [2] J. D. Deaton, R. V. Grandhi, A survey of structural and multidisciplinary continuum topology optimization: post
41 2000, Structural and Multidisciplinary Optimization 49 (1) (2014) 1–38.
42 [3] Q. Xia, L. Xia, T. Shi, Topology optimization of thermal actuator and its support using the level set based
43
multipletype boundary method and sensitivity analysis based on constrained variational principle, Structural
44
and Multidisciplinary Optimizationdoi:10.1007/s00158-017-1814-1.
45
46 [4] J. Zhu, W. Zhang, L. Xia, Topology optimization in aircraft and aerospace structures design, Archives of Com-
47 putational Methods in Engineering 23 (4) (2015) 595–622.
48 [5] P. Duysinx, M. P. Bendsøe, Topology optimization of continuum structures with local stress constraints, Inter-
49 national Journal for Numerical Methods in Engineering 43 (8) (1998) 1453–1478.
50 [6] C. Le, J. Norato, T. Bruns, C. Ha, D. Tortorelli, Stress-based topology optimization for continua, Structural and
51 Multidisciplinary Optimization 41 (4) (2010) 605–620.
52 [7] U. Kirsch, On singular topologies in optimum structural design, Structural and Multidisciplinary Optimization
53 2 (3) (1990) 133–142.
54 [8] G. Cheng, Z. Jiang, Study on topology optimization with stress constraints, Engineering Optimization 20 (2)
55 (1992) 129–148.
56 [9] G. Cheng, X. Guo, epsilon-relaxed approach in structural topology optimization, Structural and Multidisci-
57 plinary Optimization 13 (4) (1997) 258–266.
58
59 17
60
61
62
63
64
65
 1
2
3
[10] P. Duysinx, O. Sigmund, New development in handling stress constraints in optimal material distribution, Proc.
4
5 7th AIAA/USAF/NASA/ISSMO symposium on multidisciplinary analysis and optimization. A collection of
6 technical papers (held in St. Louis, Missouri) 3 (1998) 1501–1509.
7 [11] M. Bruggi, On an alternative approach to stress constraints relaxation in topology optimization, Structural and
8 Multidisciplinary Optimization 36 (2008) 125–141.
9 [12] M. Bruggi, P. Duysinx, Topology optimization for minimum weight with compliance and stress constraints,
10 Structural and Multidisciplinary Optimization 46 (3) (2012) 369–384.
11 [13] R. Yang, C. Chen, Stress-based topology optimization, Structural and Multidisciplinary Optimization 12 (2)
12 (1996) 98–105.
13 [14] C. Le, T. Bruns, D. Tortorelli, A gradient-based, parameter-free approach to shape optimization, Computer
14 Methods in Applied Mechanics and Engineering 200 (9) (2011) 985–996.
15 [15] Y. M. Xie, G. P. Steven, A simple evolutionary procedure for structural optimization, Computers and Structures
16 49 (5) (1993) 885–896.
17 [16] L. van Miegroet, P. Duysinx, Stress concentration minimization of 2D filets using X-fem and level set descrip-
18 tion, Structural and Multidisciplinary Optimization 33 (4) (2007) 425–438.
19 [17] G. Allaire, F. Jouve, Minimum stress optimal design with the level set method, Engineering Analysis with
20 Boundary Elements 32 (2008) 909–918.
21
[18] X. Guo, W. Zhang, M. Wang, P. Wei, Stress-related topology optimization via level set approach, Computer
22
Methods in Applied Mechanics and Engineering 200 (47-48) (2011) 3439–3452.
23
24 [19] Q. Xia, T. Shi, S. Liu, M. Wang, A level set solution to the stress-based structural shape and topology optimiza-
25 tion, Computers and Structures 90-91 (2012) 55–64.
26 [20] S. Cai, W. Zhang, J. Zhu, T. Gao, Stress constrained shape and topology optimization with fixed mesh: A
27 b-spline finite cell method combined with level set function, Computer Methods in Applied Mechanics and
28 Engineering 278 (2014) 361–387.
29 [21] S. Cai, W. Zhang, Stress constrained topology optimization with free-form design domains, Computer Methods
30 in Applied Mechanics and Engineering 289 (2015) 267–290.
31 [22] H. Emmendoerfer, E. Fancello, A level set approach for topology optimization with local stress constraints,
32 International Journal for Numerical Methods in Engineering 99 (2014) 129–156.
33 [23] H. Emmendoerfer, E. Fancello, A level set approach for topology optimization with local stress constraints,
34 Computer Methods in Applied Mechanics and Engineering 305 (2016) 62–88.
35 [24] A. Sharma, K. Maute, Stress-based topology optimization using spatial gradient stabilized xfem, Structural and
36 Multidisciplinary Optimization (2018) 1–22.
37 [25] R. Picelli, S. Townsend, C. Brampton, J. Norato, H. Kim, Stress-based shape and topology optimization with
38 the level set method, Computer Methods in Applied Mechanics and Engineering 329 (2018) 1–23.
39 [26] X. Huang, Y. M. Xie, Convergent and mesh-independent solutions for the bi-directional evolutionary structural
40
optimization method, Finite Elements in Analysis and Design 43 (14) (2007) 1039–1049.
41
[27] L. Xia, Q. Xia, X. Huang, Y. Xie, Bi-directional evolutionary structural optimization on advanced structures and
42
43 materials: A comprehensive review, Archives of Computational Methods in Engineeringdoi:10.1007/s11831-
44 016-9203-2.
45 [28] F. Fritzen, L. Xia, M. Leuschner, P. Breitkopf, Topology optimization of multiscale elastoviscoplastic structures,
46 International Journal for Numerical Methods in Engineering 106 (6) (2016) 430–453.
47 [29] L. Xia, F. Fritzen, P. Breitkopf, Evolutionary topology optimization of elastoplastic structures, Structural and
48 Multidisciplinary Optimization 55 (2) (2017) 569–581.
49 [30] L. Xia, D. Da, J. Yvonnet, Topology optimization for maximizing the fracture resistance of quasi-brittle com-
50 posites, Computer Methods in Applied Mechanics and Engineering, 332 (2018) 234–254.
51 [31] O. Sigmund, A 99 line topology optimization code written in Matlab, Structural and Multidisciplinary Opti-
52 mization 21 (2) (2001) 120–127.
53 [32] X. Huang, Y. Xie, Topology optimization of nonlinear structures under displacement loading, Engineering Struc-
54 tures 30 (7) (2008) 2057–2068.
55 [33] E. Holmberg, B. Torstenfelt, A. Klarbring, Stress constrained topology optimization, Structural and Multidisci-
56 plinary Optimization 48 (1) (2013) 33–47.
57
58
59 18
60
61
62
63
64
65
 HIGHLIGHTS

 An extended BESO method for stress-based topology optimization is proposed.