Professional Documents
Culture Documents
Additive Manufacturing
journal homepage: www.elsevier.com/locate/addma
A R T I C LE I N FO A B S T R A C T
Keywords: Design for additive manufacturing (AM) requires knowledge of the constraints associated with your targeted AM
Topology optimization process. One important design concern is the unintentional trapping of parasitic mass in occluded void geo-
Parasitic mass metries with either uncured or non-solidified material, or in some cases, sacrificial support material. These
Occluded void elimination occluded features create the need to physically alter the optimal topology to remove the material. In this work, a
Powder bed fusion
projection-based topology optimization design formulation is proposed to eliminate occluded void topological
Vat photopolymerization
features in optimal AM designs. The algorithm is based on the combination and enhancement of two existing
algorithms: a projection-based, overhang-constrained algorithm to design self-supporting structures in AM, and a
void projection algorithm to design topologies through control of the void phase. The combined algorithm
results in topologies with void regions that always possess an exit path to predefined outer surfaces – i.e.
drainage pathways. Solutions are first demonstrated in two dimensions, with increasing design freedom allowed
through algorithm enhancements. The algorithm is then adapted to 3D, adopting a multi-phase TO approach to
not only regain control of the solid phase length scale, but also to drive toward superior performing topologies
with minimal impact on the part performance.
⁎
Corresponding author.
E-mail address: andrew.t.gaynor2.civ@mail.mil (A.T. Gaynor).
https://doi.org/10.1016/j.addma.2020.101149
Received 5 June 2019; Received in revised form 21 February 2020; Accepted 21 February 2020
Available online 27 February 2020
2214-8604/ Published by Elsevier B.V.
A.T. Gaynor and T.E. Johnson Additive Manufacturing 33 (2020) 101149
contributions exploring this design consideration. Langelaar and col- methodology to produce castable designs (design free of concave sur-
leagues contributed a number of design algorithms in this area, most faces and occluded voids) by using a Heaviside design parameterization
notably a novel adjoint sensitivity approach for fast calculation of de- in a specified casting direction. Guo et al. [1] proposed using Moving
rivative information [8,9]. This approach was recently adapted to the Morphable Voids (MMV) to eliminate occluded voids by imposing a
framework proposed by Gaynor and Guest [10,11]. Langelaar also restriction so that printable features exist only at the boundaries of the
harnessed the same underlying approach to simultaneously design design domain. Finally, Zhou and Zhang proposed a side constraint
component and support structures [12], and to design for component scheme within a level set TO framework where voids are essentially tied
and support topology, along with build orientation [9]. An alternative to the edge of the domain – this guarantees a void pathway to any void
approach from the same research group looks at harnessing a front within the physical domain [33].
propagation filter to also guarantee self-supporting structures [13]. In contrast to the above schemes, the approach presented in this
Aside from these two main contributors, a number of other intriguing manuscript is generally less restrictive on the design space and har-
algorithms have been proposed [1,14–24]. nesses a previously developed framework for overhang projection
The overhang algorithm framework harnessed herein builds upon [7,25] with a slight modification for the riddance of occluded voids,
that found in Behrou et al. [10] and Johnson and Gaynor [11]. For and an adjoint speedup for sensitivity calculations [11,10]. As such, the
additional details on the basics of the overhang projection framework, presented approach does not require additional constraints, but does
please see Gaynor and Guest [25]. Behrou et al [10] and Johnson and require the projection scheme to proceed in a directional layer-by-layer
Gaynor [11] provide details on a recent adjouint speedup and adapta- manner. Thus, if the framework for overhang projection is already
tion to the 3D context. implemented [11,10], then the removal of occluded voids can be im-
Occluded voids – the design feature focus of this paper – trap loose plemented within this framework with minimal additional coding ef-
powder in powder-bed fusion processes, or trap liquid photo-polymer in fort.
vat-based steriolithography processes, potentially adding significant The remaining paper is outlined as follows: Sections 1.1 and 1.2 will
parasitic mass to the structure (unless it is somehow removable post- briefly explain the void projection method (VPM) and overhang pro-
print). Interestingly, in an extreme case, Reddy K et al. [26] showed jection method (OPM) schemes, respectively. Section 2 combines the
that trapped material can account for as much as 74% of the total VPM with the OPM to create an algorithm capable of tackling the
material needed to make a part. Fig. 1 illustrates the laser powder-bed
fusion process with trapped powder within a part – a cross-section
shown here. As the fabrication piston moves down, metal powder is
rolled onto the piston head. A laser beam then melts the powder in the
configuration of a prescribed cross-section shape (horizontal plane).
The process repeats until a final structure is produced. In steriolitho-
graphy, a vat-based photo-polymer process, a similar phenomena oc-
curs, where uncured resin is trapped inside solid features. Additionally,
in extrusion-based AM processes, closed cavity geometries that trap
support material can also result in parasitic mass, as support material is
impossible to remove if there is no access from the outside.
Several material density-based TO methods have been proposed for
avoiding occluded voids in design. Liu et al. [27] developed a multi-
physics TO-based methodology to remove occluded voids from struc-
tures. The method, coined the virtual temperature method (VTM), sets
void elements as highly conductive, solid elements as thermal insulators
and the domain boundaries as heat sinks. Li et al. [28] expanded Liu's
[27] methodology to include internal heat sinks and showed that 2D
and 3D TO solutions can be created with occluded voids for passage of
physical structures such as pipes. Harzheim and Graf [29] used Top-
Shape to design topologically optimized cast parts. Their proposed
approach for removing occluded voids eliminates material from the
outside-in. Zhou et al. [30] proposed manufacturing and extrusion
constraints within a TO scheme to remove occluded voids and enforce
constant cross-sections. The method introduces uni- and multi-direc-
tional constraints that reduce element density from the inside-out or
from a so called “growth interface” to the boundaries of the design
domain. By extension, Lu and Chen [31] expanded the methodology
developed by Zhou [30] to search in additional directions (4 or more) in Fig. 1. Diagram of selective laser sintering process showing trapped material in
various orientations. Gersborg and Andreasen [32] developed a a part (2D slice shown here).
2
A.T. Gaynor and T.E. Johnson Additive Manufacturing 33 (2020) 101149
identified occluded void problem. Section 3 briefly describes the adjoint More specifically, in the VPM, μe > 0 produces ρvoid e
= 0 , while
approach employed to calculate sensitivities. Section 4 presents the μe = 0 produces ρvoid
e
equal to one as β goes to ∞. This is the inverse of
optimization problem formulation. Section 5 displays initial results for solid projection logic. Guest [35] showed that applying the VPM alone
single direction two dimensional void growth optimization. Significant often produces solutions with rounded holes instead of sharp corner
enhancements to open up the design space but still eliminate the oc- topologies common from solid projection methods.
currence of occluded voids are seen in Section 6. Extension to 3D is
explored in Section 7, demonstrating how void elimination in a 3D 1.2. Overhang projection method (OPM)
context is a much less restrictive design rule. Finally, there is a dis-
cussion of the pros and cons, algorithmic nuances and some thoughts in The overhang projection method (OPM) guarantees entirely self-
the Discussion Section 8. supporting structures, thus removing the need for sacrificial support
material. The OPM scheme is actualized by three variables: ψ, ρs and ϕ.
1.1. Void Projection Method (VPM) The independent design variable vector, ψ indicates whether material is
desired at a given location, while the dependent variable vector, ϕ –
At a fundamental level, the void projection method (VPM) [3,34,35] referred to here as the “material placement variable” – determines
inverts the design problem such that the algorithm is determining whether material will be projected to the element density space at a
where the void should exist instead of determining where the solid given location, i. ϕi is defined as:
should exist, as is the case in typical solid projection TO – i.e. project
void (ρe = 0) as opposed to solid (ρe = 1). As stated previously, this ϕi = ψiρsi (5)
work uses the Heaviside Projection Method (HPM) [5] to achieve
This dependent variable vector is then either projected to the
minimum length scale control. Fig. 2 illustrates HPM mapping in three
dimensions. In this diagram, the design variables are chosen to exist at
the nodes of finite element mesh. Fig. 2a demonstrates the design
variable projection to elemental density space within a radial distance
of rmin, defining the minimum allowable feature size. Alternatively, one
can view the projection from the elemental perspective, as seen in
Fig. 2b, and define the neighborhood set, denoted as Ne, of design
variables within rmin of the element centroid. Design variables within Ne
either project solid or void information to the element of interest, e.
More specifically, the weighted average, μe, of the magnitudes of the
design variables within Ne is calculated and is either linearly or non-
linearly projected to the element of interest. Linear projection is ef-
fectively the same as linear density filtering and is discussed in Burns
and Tortorelli [36]. In HPM, the weight averaged design variables are
passed through a regularized (continuous) Heaviside function to pro-
duce an element density, ρe. The added non-linearity allows for a crisper
definition of the feature boundary – i.e. less intermediate-density ele-
ments.
Solid projection is first defined for completeness. The relative ele-
ment densities for solid phase projection is defined as:
1, if μe (ϕ) > 0
e
ρsolid = H (μe (ϕ)) = ⎧ e
⎨
⎩ 0, if μ (ϕ) = 0 (1)
where H represents the Heaviside function, μ is the weighted average
e
e e μe (ϕ) −βϕ
ρsolid = 1 − e−βμ (ϕ) + e max
ϕmax (2)
where β is the regularization parameter. Defining the aggressiveness of
the regularized Heaviside function, if β is large, Eq. (2) has a high de-
gree of nonlinearity, producing crisp topological boundaries in the
transition from solid to void (minimal intermediate density elements).
Note that in Eq. (2), ϕmax = 1.
Conversely, in the VPM the design algorithm controls the length
scale of void phase, and the length scale control of the solid phase is
lost. The VPM is defined as:
0, if μe (ϕ) > 0
e
ρvoid = 1 − H (μe (ϕ)) = ⎧ e
⎨
⎩1, if μ (ϕ) = 0 (3)
Thus, the void projection is simply one minus the solid projection
and is represented here:
e e μe (ϕ) −βϕ
ρvoid = e−βμ (ϕ) − e max Fig. 2. Minimum length scale imposed through a projection of design variables
ϕmax (4) to physical space.
3
A.T. Gaynor and T.E. Johnson Additive Manufacturing 33 (2020) 101149
physical space through solid projection (Eq. (2)) or through void pro- examples in this manuscript use quadrilateral elements for 2D problems
jection (Eq.(4)). Ultimately, to convert the independent design variable, and a hexahedral elements for 3D problems – this is done for simplicity
ψ, to the dependent material placement variable, ϕ, one must proceed and because these element types generally result in smoother topologies
in a systematic fashion, calculating ϕ layer-by-layer via Eq. (5). in post-processing efforts.
Further diving into Eq. (5), the variable vector, ρs – designated the The threshold value, T, in Eq. (6) is set depending on number of
“support indicator” – is a function of ϕ in the support neighborhood set, nodes below. For the case seen in Fig. 3, the nominal T value is set to 2
8
NS, and indicates whether a point is adequately supported. NSi for a for internal points. Importantly, the nominal T is then shifted so that ρS
particular ϕi is shown in Fig. 3, indicated as the ϕ captured in the blue equals roughly 0.93 at the threshold (rule of thumb) – see Johnson and
wedge. The support indicator, ρsi , for a particular point, i, is calculated Gaynor [11] for more details. Note that the T for a design point on the
as: edge of the space is different than the T for an internal design point, as a
tanh(βT T ) + tanh(βT (μsi (ϕ ηT ) − T )) varying number of ϕ points are captured.
ρsi = HT (ϕ) =
tanh(βT T ) + tanh(βT (1 − T )) (6) 2. Combined formulation: the void overhang projection method
In Eq. (6), the average of the material placement variables, ϕ, for a (VOPM)
particular point, i, is denoted as μsi , and is defined in the following
equation: Drawing upon the logic of the solid overhang constrained algorithm,
the result of coupling the VPM with the OPM is to produce structures
∑j ∈ N i ϕ jws
μsi = S
, with designed “void pathways” which nucleate at the structure's
∑j ∈ N i ws (7) boundaries and stop within, or penetrate through the structure. The
S
presence of void pathways ensures that unsolidified material will not be
where ws is the uniform support region weighting function (ws = 1), and
trapped by the structure.
j is a counter of ϕ in the support neighborhood set, NSi . This average is
The combined formulation – deemed the void overhang projection
passed through Eq. (6) to calculate the “support indicator,” ρs. The
method (VOPM) – simply involves sequentially implementing the OPM
threshold value, T, is enforced by setting the thresholding Heaviside
and the VPM. First the design variables, ψ are passed through the OPM
parameter, βT, to a sufficiently high value. Additionally, it is found that
to determine the material placement variables, ϕ. Subsequently, the
including a small penalization term, ηT, helps with convergence by
material placement variables are passed through the VPM to project
driving the ϕ in the “wedge” neighborhood set to 0 or 1. ηT = 1.5 for all
void information to the physical density space, ρ, as specified in Eq. (4).
example problems herein. The threshold, T, varies between 0 and 1 and
This can be summed up in the following equation:
is chosen such that the support condition is correctly enforced.
e e
For the OPM, the user must designate the boundary on which the ρvoid = ρvoid (ϕ (ψ)) = 1 − H (μe (ϕ (ψ))) (8)
build plate exists. This equates to setting ρs = 1 along the bottom edge
For solutions to typical VPM and OPM problems, readers are re-
of the design domain. Once this is established, the calculation of ρs and
ferred to previous publications on the matter: VPM see [3,34,35], OPM
ϕ proceed in a layer-by-layer fashion, as previously mentioned. This
see [25,11,10]. In this section, we demonstrate the sequential projec-
layer-by-layer “directionally dependent” approach is fundamental to
tion scheme through a simple example. Fig. 4a presents a test design
the proposed algorithm in Section 2.
variable field, ψ, where red indicates a value of 1 and blue indicates a
value of 0. In this projection example, the design variable field is passed
1.3. Adjusted overhang projection neighborhood set for 2D
through both the VPM and the VOPM to highlight the difference. In the
case of the VPM, the ϕ material placement variables are equal to the ψ
The three dimensional OPM support neighborhood set, NS remains
design variables (Fig. 4a), and the projected topology is seen to have
the same as in the manuscript by Johnson and Gaynor [11]. This ap-
rounded void space corners (Fig. 4d). Alternatively, in the VOPM case
proach for overhang control “looks” one layer of ϕ down in the design
with the OPM projection from bottom to top with a defined angle of
space to determine the support metric, ρs.
45°, the ϕ material placement variables are quite different from the
In two dimensions, a one layer down support neighborhood set (NS)
VPM case. Notably, the two square regions of ψ = 1 are projected (i.e.
only captures three material placement points, ϕ [10], in the neigh-
filtered) to be ϕ = 0, as there is no pathway to the bottom surface,
borhood set. While conceptually simple, the one layer approach in 2D
often results in non-monotonically decreasing optimization progression.
To counteract this phenomena, the authors have gone back to an ap-
proach taken in previous papers [7,25], in which NS included a multi-
layered wedge region below the point of concern. In these papers, the
NS wedge extended at least two layers of ϕ below the design point of
concern. A multi-layered NS will result in overlapping of neighborhood
sets in the build direction (y direction in 2D as seen in Fig. 3), sig-
nificantly reducing the nonlinear behavior and ensuring smooth con-
vergence.
The diagram in Fig. 3 demonstrates how this support neighborhood
set, NS, is defined for one particular material placement variable located
at the yellow dot. The scheme “looks” two material placement layers
below in the blue triangular region to form the NS. To define the
overhang angle, the spacing of the independent design variables ψ and
dependent material placement variables ϕ (locations of ψ and ϕ are
coincident) is varied by adjusting the spacing variables δx and δy. The
overhang angle from horizontal is defined as tan−1 ( ). Fig. 3 demon-
δy
δx
strates the support neighborhood set definition on top of an un- Fig. 3. Two layer overhang projection method mapping scheme. The blue
structured triangular finite element mesh to emphasize that the design Neighborhood set, NSi is illustrated for the yellow design point of concern. (For
space and material placement space are independent from the under- interpretation of the references to color in this figure legend, the reader is re-
lying finite element mesh. It should be noted, however, that all ferred to the web version of this article.)
4
A.T. Gaynor and T.E. Johnson Additive Manufacturing 33 (2020) 101149
Fig. 4. Design variable mapping for standard void projection and for void overhang projection method (VOPM). (For interpretation of the references to color in the
text, the reader is referred to the web version of this article.)
∂ϕ
where all voids must nucleate. Also, the “L-like” ψ region is projected to (previously [25] required storage of the term, which is on the order
∂ψ
create a vertical ψ region with an angled region attached to the side at a of number of the design variables (ndv) by ndv in size).
45° angle, as defined – see Fig. 4c. This ϕ field projects to the physical To obtain the sensitivities of the OPM, one must first calculate the
space as before and creates the topology in Fig. 4e, which allows for sensitivities with respect to the material placement variables, ϕ. The
void pathways to the bottom surface, as desired. Note that in this case, sensitivity of an arbitrary function, C, w.r.t. ϕ is calculated according
the material in the blue void space would be able to flow downward at to:
an angle of 45° or greater from horizontal.
Under the combined VOPM algorithm, instead of a “supported” ∂C ∂C ∂ρ e
∂ϕ
= ∑ ∂ρ e ∂ϕ
solid feature growing from the build plate in the +y, a “supported” void e∈Ω (9)
region must grow from the designated side in the +y direction. This
Next, the calculated derivative w.r.t. the material placement vari-
manufacturing constraint turns out to be rather restrictive, completely
ables are passed through an adjoint approach to obtain the derivatives
eliminating the creation of “holes” or occluded voids seen in the VPM
w.r.t. the design variables, ψ. As demonstrated in [11,10], the calcu-
and OPM. Hence, each void region must have an uninterrupted
lation of the ∂ϕ term is eliminated, replaced instead with an equivalent
pathway to the designated surface. ∂ψ
∂C
In this example, an overhang angle of 45 degrees is chosen for sensitivity through an adjoint approach. This calculation is seen in
∂ψj
convenience. The engineer should choose an angle that corresponds the following equation:
with reliable powder and resin flow under gravitational forces.
Obviously, the closer the OPM angle gets to 0 (horizontal), the more the
design space is relaxed. However, in practice, small angles (less than
∂C
∂ψj
⎛ ∂C
=⎜
∂ϕj
+ λj + 1 ˘ ˘
∂ mj + 1 ⎞ ∂ mj
∂ϕj ⎟ ∂ψj
= λj ˘
∂ mj
∂ψj
1 ⎝ ⎠ (10)
roughly tan−1 ( 3 ) ) can cause issues due to the skewed design point
spacing (δx > > δy in Fig. 3). where the vectors λj are adjoint operators (multipliers). For con-
3. Sensitivity analysis
˘
venience, a new function, m , is defined as:
˘
mj = ϕi = ψiρsi (11)
The objective and constraint function sensitivities are similar to Defining a function in this way also helps draw parallels to the
those found in Johnson and Gaynor's 3D overhang projection paper aforementioned approach by Langelaar [8,37]. The subscript j on the
[11]. A more thorough derivation of the self-supporting adjoint ap-
proach adapted to the HPM context from that seen in Langelaar's papers ˘
function m indicates the “layers” of the design variable Cartesian grid,
where j = 1 is the lowest layer (the build plate for solid OPM). Layer
[8,37] is found in a recent paper [10]. This adjoint approach allows for j = ni is the top layer of the design variable grid. Eq. (10) shows that
order-of-magnitude speedup in the sensitivity calculation by elim- each adjoint operator depends on the adjoint operators associated with
inating the need to directly calculate the ∂ϕ term. It is emphasized that the layer above. Therefore, the sensitivity calculation begins at the top
∂ψ
the adjoint approach also simplifies implementation and significantly layer of the design variable grid and proceeds sequentially downward,
reduces memory requirements, allowing for better scalability until reaching the bottom layer. The following equation provides the
5
A.T. Gaynor and T.E. Johnson Additive Manufacturing 33 (2020) 101149
ρ e
Ke = ⎜⎛ρmin + ⎞ e
⎟K
0
⎝ 1 + η (1 − ρe ) ⎠ (16)
While the occluded void elimination scheme will also work with the
more typical SIMP penalization scheme [38,40], RAMP is found to
produce more stable convergence behavior and superior performing
topologies. Specifically, RAMP produces in non-zero derivative values
at ρe = 0 while SIMP yields zero-valued derivates at ρe = 0 – the non-
zero derivatives help provide constant derivative information, even if
elemental densities goes to zero. Similar benefits are seen in Ha and
Guest [41] and Guest [42]. In Eq. (16), ρmin is set to ρmin = 0.001 to
prevent singularity issues in the system solve. The penalty parameter, η,
is set sufficiently high such that intermediate values of ρe are deterred
through penalized.
The Method of Moving Asymptotes (MMA) optimizer [43], was
employed to tackle the defined problem in Eq. (15). The MMA algo-
rithm is slightly altered such that the asymptote increase parameter is
set to 1.15 instead of the typical 1.2 and the asymptote decrease
parameter was set to 0.6 instead of the typical 0.7. The initial asymp-
tote value was set to 0.5/(max(β, βT) + 1). These parameters produced
smooth convergence behavior.
Fig. 5. Void projection at various angles. Pinned-pinned beam with angled
point load. 5. Two-dimensional algorithm results
multiplier at the top layer: A two-dimensional example is solved here to demonstrate the al-
gorithm and to help define the meaning of the overhang angle in the
∂C void projection context. As defined in Fig. 5a, a pinned-pinned beam
λ nTi =
∂ϕni (12) with an angled load on bottom center is used. The design domain here is
defined by L = 60, H = 10 with 600 by 100 element mesh discretiza-
placement variable, ϕ, is calculated as:
˘
For completeness, the derivative of m with respect to material
tion. In this example, the void grows from the bottom, with the al-
lowable angle of “void growth” varied. In Fig. 5b to d, the angle of void
˘∂ϕ
∂ mj + 1
=ψ
∂HT (ϕ)
∂ϕj
growth decreases from 63 degrees, to 45 degrees, to 27 degrees to
horizontal. As can been seen in the angled member attached to the
j (13)
point load, the allowable void growth becomes less restrictive as the
the form:
˘
Finally, the derivative of m w.r.t. the design variables, ψ, takes on angle of growth decreases. Hence, 27 degrees is much less restrictive on
the design space than 63 degrees.
Inspecting the formulation, the void growth angle is perhaps a bit
˘ =ρ
∂ mj
∂ψj s arbitrary at first glance. However, experienced AM professionals will
(14) attest that when draining material from internal voids, the topology of
Derivatives are not given here for the void projection scheme, as the internal void regions can still trap material, even if there is an exit
they are previously published [35]. Additionally, derivatives for the pathway. This is especially a problem when a winding internal pathway
RAMP material interpolation scheme are straight forward. has no exit on the other end, eliminating the possibility of using pres-
surized air to blow out the material. The proposed algorithm, by
comparison, always guarantees a “downhill” exit pathway for any in-
4. Problem formulation
ternal void regions. This is a major advantage, as a part can simply be
vibrated in a single orientation and the powder or resin should exit to
All design cases in this manuscript solve the well-known minimum
the bottom through gravity alone. An additional advantage comes from
compliance (maximum stiffness) with maximum volume constraint
the designer's ability to specify the minimum diameter of the exit
problem [38], as seen in the following formulation:
pathways (drainage holes). More viscous materials may require larger
min f (ψ) = F Td pathways, while “easy flowing” low viscosity materials may allow for
ψ
much smaller pathways.
subject to: K (ψ) d=F
For the two dimensional VOPM approach, it is useful to define some
∑ ρe (ψ) v e ≤ V parameter values (found to work well). In addition to the parameters
e∈Ω seen below in Table 1, the element type of each beam was a 4-node
0 ≤ ψi ≤ ψmax ∀ i∈Ω (15) quadrilateral. The most critical parameters were β and βT – the values in
T
where the objective function, f, is the structural compliance, F d. The Table 1 were chosen such that OPM logic was correctly imposed and so
compliance and maximum volume constraint, ∑e ∈ Ω ρ e v e ≤ V , are a that convergence was fast and targeted quality local minima.
function of the elemental density vector, ρ, which is ultimately a
function of the independent design variable vector, ψ. Kd = F is an 6. Algorithm extensions and enhancements
equilibrium constraint – where F is the global load vector and K is the
global stiffness matrix. Solving Kd = F yields, d, the nodal displace- The above demonstration produces solutions which adhere to the
ment vector. v is the vector of elemental volumes, and V is the imposed “void growth” constraint. However, the constraint of requiring
6
A.T. Gaynor and T.E. Johnson Additive Manufacturing 33 (2020) 101149
7
A.T. Gaynor and T.E. Johnson Additive Manufacturing 33 (2020) 101149
Fig. 7. Bow tie test case free-form solution, f = 9.63. (For interpretation of the
references to color in the text, the reader is referred to the web version of this
article.)
Fig. 11. Design point setup for void growth from bottom-up and top-down (BU
+TD). Inner circle colored as yellow indicates ρS1 is set to 1 for these design
points. Outer circle colored as red indicates ρS2 is set to 1 for these design
points. (For interpretation of the references to color in this figure legend, the
reader is referred to the web version of this article.)
Fig. 8. Bow tie problem: Optimal topology solution allowing void from bottom
(BU), f = 15.49.
Fig. 9. Design point setup for void growth from bottom-up and sides (BU+s).
Inner circle colored as yellow indicates ρs is set to 1 for these design points. (For
interpretation of the references to color in this figure legend, the reader is re-
ferred to the web version of this article.)
Fig. 10. Bow tie problem: Optimal topology solution allowing void from
bottom and sides (BU+s), f = 13.72.
8
A.T. Gaynor and T.E. Johnson Additive Manufacturing 33 (2020) 101149
perimeter line is 1.0 on each node, except for the corner nodes, where
the magnitude is 0.5.
The free-form reference case is solved (Fig. 18) and the compliance
is normalized to a value of 100 in the penalized compliance plot in
Fig. 22. As can be seen in Fig. 18b, there exists a large ellipsoid-like void
region. This void region of the design is completely closed, meaning
none of the unsolidified material would naturally drain.
To rectify the issue, the algorithm outlined above, and demonstrated
in a 2D sense, is adapted to a 3D context. All concepts of the VOPM are
directly translatable. Additionally, and importantly, for the 3D design
case, a multi-phase TO scheme is employed to design both the void
phase and the solid phase, regaining control of the solid phase
minimum length scale and helping the optimizer obtain better local
minima – the optimization problem's nonlinearity is alleviated through
introduction of a second phase. Here, the minimum length scale of the
void phase guarantees the minimum drainage hole diameter for powder
Fig. 13. Design point setup for void growth from top-down and bottom-up,
or resin. As noted by Guest [35], the multiphase algorithm simply re-
including side growth nucleation (BU+TD+s). Inner circles colored as yellow
indicates ρs1 is set to 1 for these design points. Outer circles colored as red
quires doubling the number of design variables, such that there is a set
of independent design variables associated with solid phase and an-
indicates ρs2 is set to 1 for these design points. (For interpretation of the re-
other set associated with the void phase. In this manuscript, the loca-
ferences to color in this figure legend, the reader is referred to the web version
of this article.) tion of the design points for both phases are coincident. The element
density calculation for the multi-phase formulation takes on the fol-
lowing form:
Applying this scheme to the same bow tie test problem, one obtains
the topology seen in Fig. 14, where the void is allowed to nucleate on 1
ρ= (ρ (ψ ) + ρsolid (ψsolid ))
both sides in addition to the bottom and top, opening up the design 2 void void (19)
space relatively significantly, and allowing the optimization to ap- The calculation of ρsolid proceeds in the standard HPM manner,
proach the free-form solution (Fig. 7). without the intermediate “material placement” dependent variables, ϕ:
Plotting all normalized penalized objective function values for the
e e
bow tie problem, it can be seen how allowing additional void growth ρsolid = ρsolid (ψsolid ) = H (μe (ψsolid )) (20)
directions relaxes the design space and lets the solutions approach the
The void projection scheme proceeds in the same VOPM manner as
free-form TO solution. Here, the most relaxed case (d) “BU+TD+s”
before:
achieved a penalized compliance only 13% higher than the free-form
solution, while the most restricted case (a) “BU” achieved a compliance e
ρvoid e
= ρvoid (ϕ (ψvoid )) = 1 − H (μe (ϕ (ψvoid ))) (21)
61% higher than the free-form solution.
Lastly, a demonstration of the convergence behavior is included in Aside from the multi-phase scheme, the parameter values and op-
Fig. 16 (shown here for the BU+TD+s case). As can be seen in the timization strategy are similar to that seen for 2D and are tabulated in
objective function vs. iteration plot, the convergence is generally very Table 2.
smooth, with only a bit of undesirable oscillatory behavior in the be- The torsion example problem is demonstrated with two levels of
ginning, where the structure attaches and detaches from itself. This restriction on the void growth. For all design cases, a buffer layer of
behavior is seen in the design evolution of the first 10 iterations or so design variables is added outside the bounds of the physical mesh. Here,
(see Fig. 16b to f). Clearly the structure always adheres to the imposed we add four extra design points in the x, y, and z direction (i.e. two
constraint, and quickly converges to a high quality local minima. extra on each side). The extra design points allow relaxation of the
design space, resulting in improved convergence of the algorithm
without any significant computational burden. Physically this allows
7. Three-dimensional algorithm results
for void nucleation away from the boundary of the physical mesh –
lessening the compliance impact of new void nucleation, and improving
Two dimensional design cases are great to demonstrate proof of
optimization behavior.
concept. However, the algorithm restricts the design space rather sig-
In the first instance in Fig. 19, the problem is solved allowing void
nificantly, even when voids are allowed to grow from multiple direc-
growth from bottom-up (BU) only – i.e. the 3D equivalent to the typical
tions and from side regions. Therefore, the VOPM algorithm is de-
2D mapping (ρs on bottom of domain set to 1), except with the afore-
monstrated in 3D, where the added dimension allows for a much closer
mentioned buffer zone of extra design points. The overhang angle for
match between free-form TO solutions and the proposed internal void
the void growth is again set to 27 degrees – this value tends to work
elimination TO scheme. First, it is shown that free-form TO solutions
well, as in 2D.
may exhibit large occluded void topological features, where parasitic
As can be seen in the solution in Fig. 19a, the topology is extremely
mass would be trapped. Second, the 2D algorithm is implemented in 3D
and applied to the same problem. It is shown that visually similar so-
lutions may be obtained with minimal loss of optimality in comparison
to free-form.
While there are a number of potential test problems, we showcase
the algorithm's capabilities on a 3D torsion problem, as defined in
Fig. 17. The design problem here has relative dimensions of L = 64 and
H = 32 discretized with elements nelx = 64, nely = 32, and nelz = 32. A
patch of nodes on the “wall” are fixed in x, y and z: here the patch is size
12 by 12, equating to 13 by 13 patch of fixed nodes centered on the left-
hand face. On a similar patch on the right-hand face, a torsion load is Fig. 14. Bow tie problem: Optimal topology solution allowing void from
applied to the perimeter. The magnitude of the line load on each bottom, top and sides (BU+TD+s), f = 10.91.
9
A.T. Gaynor and T.E. Johnson Additive Manufacturing 33 (2020) 101149
similar to that seen in free-form solution (Fig. 18). The only difference
is two small drainage holes on the bottom face of the design domain, as
seen in the cut view in Fig. 19b. Due to the imposed allowable void
growth angle, an inclined exit path allows for easy material removal. It
is noted that the existence of two holes is a byproduct of void growth
angle – i.e. at a more extreme defined angle, the optimal topology
would likely only require one drainage hole. Looking at the normalized
performance (compliance) of the topologies in plot in Fig. 22, it's noted
that the penalized compliance value of 102.69 is only slightly larger
than that of the free-form solution (f = 100). This 2.7% increase is
minimal in comparison to the drastic performance decrease when im-
posing the internal void elimination scheme to problems in 2D. Clearly,
the constraint is much less restrictive in a 3D context.
As was demonstrated in two dimensions, the introduction of multi-
directional void growth allows for further design relaxation. The torsion
test problem is now solved with the design scheme which allows void
growth from the top and the bottom. As in 2D, the problem employs the
multiple ϕ approach as seen in Eq. (18). The solution in Fig. 20 re-
sembles that of the single direction (Fig. 19), but possesses void exit
pathways to both the bottom and the top. As seen in Fig. 20b, the void Fig. 16. Convergence plot and design evolution for BU+TD+s.
has 2 clear exit pathways to the top and two clear exit pathways to the
bottom. the relaxation of the MMA parameters allowed the solution to progress
While one might expect the addition of more design freedom to to a better local minima. Other design cases may not be as suseptible to
lower the objective function in relation to the single-direction void initial guesses – although in any case, the achieved compliance in all
growth scheme, the opposite is true here – Fig. 22 reports a value of three instances is within 5% of the free-form solution.
f = 104.67, as opposed to the previously achieved f = 102.69. While
this is not expected, it makes logical sense, as the algorithm may not be
able to discern whether or not to use an exit boundary (sensitivity is 8. Discussion
symmetric).
To test the theory that the algorithm found an inferior local min, the The proposed VOPM algorithm for eliminating the occurrence of
initial guess is perturbed slightly and the MMA parameters are relaxed parasitic occluded void material in TO solutions involves combining
to their default values. Thus, the initial ψ is set to ψnew = ψoriginal − δ/ two previously developed algorithms for (1) overhang control in ad-
2 + δR(ndv), where δ is some small value (here δ = 0.08), and R in- ditive manufacturing (OPM) and (2) void design control (VPM). This
dicates a random number vector the length of the number of design algorithm is a clear alternative to the virtual temperature method
variables (ndv). The design from the alternative initial guess (Fig. 21) (VTM). Inspecting the solutions from VTM, it is often seen that there are
yields a normalized performance of f = 102.74, which is almost iden- generally more holes created in the domain, as the algorithm chooses to
tical to the performance achieved in the BU design case. Additionally, it route the “thermally loaded” void region to the closest available heat
should be noted that the design in this case is no longer symmetric. sink border. Hence, it is unlikely that the VTM algorithm would allow
Thus is can be concluded that the slight perturbation in initial guess and for long exit pathways of the void phase. In the proposed VOPM algo-
rithm, that inherent problem is eliminated. Additionally, it is asserted
10
A.T. Gaynor and T.E. Johnson Additive Manufacturing 33 (2020) 101149
Table 2
3D parameter values.
Param Param value
11
A.T. Gaynor and T.E. Johnson Additive Manufacturing 33 (2020) 101149
Fig. 21. 3D torsion problem: void from bottom-up and top-down (BU+TD)
Fig. 20. 3D torsion problem: void from bottom-up and top-down (BU+TD). with alternative initial guess and standard MMA parameters.
Authors’ contributions
Conflicts of interest
12
A.T. Gaynor and T.E. Johnson Additive Manufacturing 33 (2020) 101149
Acknowledgements [22] Y.-H. Kuo, C.-C. Cheng, Self-supporting structure design for additive manufacturing
by using a logistic aggregate function, Struct. Multidiscip. Optim. (2019), https://
doi.org/10.1007/s00158-019-02261-3.
The authors would like to thank Krister Svanberg allowing use of [23] A. Garaigordobil, R. Ansola, E. Veguería, I. Fernandez, Overhang constraint for
the MMA optimizer code. topology optimization of self-supported compliant mechanisms considering ad-
ditive manufacturing, Comput.-Aided Des. 109 (2019) 33–48, https://doi.org/10.
1016/j.cad.2018.12.006.
References [24] C.-J. Thore, H.A. Grundström, B. Torstenfelt, A. Klarbring, Penalty regulation of
overhang in topology optimization for additive manufacturing, Struct. Multidiscip.
[1] X. Guo, J. Zhou, W. Zhang, Z. Du, C. Liu, Y. Liu, Self-supporting structure design in Optim. (2019), https://doi.org/10.1007/s00158-019-02194-x.
additive manufacturing through explicit topology optimization, Comput. Methods [25] A.T. Gaynor, J.K. Guest, Topology optimization considering overhang constraints:
Appl. Mech. Eng. 323 (2017) 27–63, https://doi.org/10.1016/j.cma.2017.05.003. eliminating sacrificial support material in additive manufacturing through design,
[2] O. Sigmund, Design of Material Structures Using Topology Optimization, PhD Struct. Multidiscip. Optim. 54 (5) (2016) 1157–1172, https://doi.org/10.1007/
Dissertation, Department of Solid Mechanics, Technical University of Denmark, s00158-016-1551-x.
1994. [26] V. Maranan, T.W. Simpson, T. Palmer, C.J. Dickman, Application of topology op-
[3] O. Sigmund, Morphology-based black and white filters for topology optimization, timization and design for additive manufacturing guidelines on an automotive
Struct. Multidiscip. Optim. 33 (4) (2007) 401–424, https://doi.org/10.1007/ component, Volume 2A: 42nd Design Automation Conference (2016), https://doi.
s00158-006-0087-x. org/10.1115/detc2016-59719.
[4] T.A. Poulsen, A new scheme for imposing a minimum length scale in topology [27] S. Liu, Q. Li, W. Chen, L. Tong, G. Cheng, An identification method for enclosed
optimization, Int. J. Numer. Methods Eng. 57 (6) (2003) 741–760, https://doi.org/ voids restriction in manufacturability design for additive manufacturing structures,
10.1002/nme.694. Front. Mech. Eng. 10 (2) (2015) 126–137, https://doi.org/10.1007/s11465-015-
[5] J.K. Guest, J.H. Prévost, T. Belytschko, Achieving minimum length scale in topology 0340-3.
optimization using nodal design variables and projection functions, Int. J. Numer. [28] Q. Li, W. Chen, S. Liu, L. Tong, Structural topology optimization considering con-
Methods Eng. 61 (2) (2004) 238–254, https://doi.org/10.1002/nme.1064. nectivity constraint, Struct. Multidiscip. Optim. 54 (4) (2016) 971–984, https://doi.
[6] M. Zhou, B.S. Lazarov, F. Wang, O. Sigmund, Minimum length scale in topology org/10.1007/s00158-016-1459-5.
optimization by geometric constraints, Comput. Methods Appl. Mech. Eng. 293 [29] L. Harzheim, G. Graf, A review of optimization of cast parts using topology opti-
(2015) 266–282, https://doi.org/10.1016/j.cma.2015.05.003. mization, Struct. Multidiscip. Optim. 31 (5) (2006) 388–399, https://doi.org/10.
[7] A.T. Gaynor, J.K. Guest, Topology optimization for additive manufacturing: con- 1007/s00158-005-0554-9.
sidering maximum overhang constraint, 15th AIAA/ISSMO Multidisciplinary [30] M. Zhou, R. Fleury, Y.-K. Shyy, H. Thomas, J. Brennan, Progress in topology opti-
Analysis and Optimization Conference (2014), https://doi.org/10.2514/6.2014- mization with manufacturing constraints, 9th AIAA/ISSMO Symposium on
2036. Multidisciplinary Analysis and Optimization (2002), https://doi.org/10.2514/6.
[8] M. Langelaar, An additive manufacturing filter for topology optimization of print- 2002-5614.
ready designs, Struct. Multidiscip. Optim. 55 (3) (2017) 871–883, https://doi.org/ [31] J. Lu, Y. Chen, Manufacturable mechanical part design with constrained topology
10.1007/s00158-016-1522-2. optimization, Proc. Inst. Mech. Eng. Part B: J. Eng. Manuf. 226 (10) (2012)
[9] M. Langelaar, Combined optimization of part topology, support structure layout 1727–1735, https://doi.org/10.1177/0954405412457643.
and build orientation for additive manufacturing, Struct. Multidiscip. Optim. 57 (5) [32] A.R. Gersborg, C.S. Andreasen, An explicit parameterization for casting constraints
(2018) 1985–2004, https://doi.org/10.1007/s00158-017-1877-z. in gradient driven topology optimization, Struct. Multidiscip. Optim. 44 (6) (2011)
[10] R. Behrou, J.K. Guest, A.T. Gaynor, Projection-Based Overhang Constraints: 875–881, https://doi.org/10.1007/s00158-011-0632-0.
Implementing an Efficient Adjoint Formulation for Sensitivity Analysis, (2018). [33] L. Zhou, W. Zhang, Topology optimization method with elimination of enclosed
[11] T.E. Johnson, A.T. Gaynor, Three-dimensional projection-based topology optimi- voids, Struct. Multidiscip. Optim. (2019), https://doi.org/10.1007/s00158-019-
zation for prescribed-angle self-supporting additively manufactured structures, 02204-y.
Addit. Manuf. (2018), https://doi.org/10.1016/j.addma.2018.06.011. [34] S.R.M. Almeida, G.H. Paulino, E.C.N. Silva, A simple and effective inverse projec-
[12] M. Langelaar, Integrated component-support topology optimization for additive tion scheme for void distribution control in topology optimization, Struct.
manufacturing with post-machining, Rapid Prototyp. J. 25 (2) (2019) 255–265, Multidiscip. Optim. 39 (4) (2009) 359–371, https://doi.org/10.1007/s00158-008-
https://doi.org/10.1108/RPJ-12-2017-0246. 0332-6.
[13] E. van de Ven, R. Maas, C. Ayas, M. Langelaar, F. van Keulen, Continuous front [35] J.K. Guest, Topology optimization with multiple phase projection, Comput.
propagation-based overhang control for topology optimization with additive man- Methods Appl. Mech. Eng. 199 (1) (2009) 123–135, https://doi.org/10.1016/j.cma.
ufacturing, Struct. Multidiscip. Optim. 57 (5) (2018) 2075–2091, https://doi.org/ 2009.09.023.
10.1007/s00158-017-1880-4. [36] T.E. Bruns, D.A. Tortorelli, Topology optimization of non-linear elastic structures
[14] X. Qian, Undercut and overhang angle control in topology optimization: a density and compliant mechanisms, Comput. Methods Appl. Mech. Eng. 190 (26) (2001)
gradient based integral approach, Int. J. Numer. Methods Eng. 111 (3) (2017) 3443–3459, https://doi.org/10.1016/S0045-7825(00)00278-4.
247–272, https://doi.org/10.1002/nme.5461. [37] M. Langelaar, Topology optimization of 3d self-supporting structures for additive
[15] G. Allaire, C. Dapogny, R. Estevez, A. Faure, G. Michailidis, Structural optimization manufacturing, Addit. Manuf. 12 (2016) 60–70, https://doi.org/10.1016/j.addma.
under overhang constraints imposed by additive manufacturing technologies, J. 2016.06.010.
Comput. Phys. 351 (2017) 295–328, https://doi.org/10.1016/j.jcp.2017.09.041. [38] M.P. Bendsœ, Optimal shape design as a material distribution problem, Struct.
[16] Y. Mass, O. Amir, Topology optimization for additive manufacturing: accounting for Optim. 1 (4) (1989) 193–202, https://doi.org/10.1007/bf01650949.
overhang limitations using a virtual skeleton, Addit. Manuf. 18 (2017) 58–73, [39] M. Stolpe, K. Svanberg, An alternative interpolation scheme for minimum com-
https://doi.org/10.1016/j.addma.2017.08.001. pliance topology optimization, Struct. Multidiscip. Optim. 22 (2) (2001) 116–124,
[17] W. Zhang, L. Zhou, Topology optimization of self-supporting structures with https://doi.org/10.1007/s001580100129.
polygon features for additive manufacturing, Comput. Methods Appl. Mech. Eng. [40] M. Zhou, G. Rozvany, The coc algorithm, part ii: topological, geometrical and
334 (2018) 56–78, https://doi.org/10.1016/j.cma.2018.01.037. generalized shape optimization, Comput. Methods Appl. Mech. Eng. 89 (1) (1991)
[18] F. Mezzadri, V. Bouriakov, X. Qian, Topology optimization of self-supporting sup- 309–336, https://doi.org/10.1016/0045-7825(91)90046-9 Second World Congress
port structures for additive manufacturing, Addit. Manuf. 21 (2018) 666–682, on Computational Mechanics.
https://doi.org/10.1016/j.addma.2018.04.016. [41] S.-H. Ha, J.K. Guest, Optimizing inclusion shapes and patterns in periodic materials
[19] Y. Wang, J. Gao, Z. Kang, Level set-based topology optimization with overhang using discrete object projection, Struct. Multidiscip. Optim. 50 (1) (2014) 65–80,
constraint: towards support-free additive manufacturing, Comput. Methods Appl. https://doi.org/10.1007/s00158-013-1026-2.
Mech. Eng. 339 (2018) 591–614, https://doi.org/10.1016/j.cma.2018.04.040. [42] J.K. Guest, Optimizing the layout of discrete objects in structures and materials: a
[20] A. Garaigordobil, R. Ansola, J. Santamaría, I. Fernández de Bustos, A new overhang projection-based topology optimization approach, Comput. Methods Appl. Mech.
constraint for topology optimization of self-supporting structures in additive man- Eng. 283 (2015) 330–351, https://doi.org/10.1016/j.cma.2014.09.006.
ufacturing, Struct. Multidiscip. Optim. 58 (5) (2018) 2003–2017, https://doi.org/ [43] K. Svanberg, The method of moving asymptotes – a new method for structural
10.1007/s00158-018-2010-7. optimization, Int. J. Numer. Methods Eng. 24 (2) (1987) 359–373, https://doi.org/
[21] G. Allaire, C. Dapogny, R. Estevez, A. Faure, G. Michailidis, Structural optimization 10.1002/nme.1620240207.
under overhang constraints imposed by additive manufacturing processes: an [44] M. Langelaar, Topology optimization for multi-axis machining, Comput. Methods
overview of some recent results, Appl. Math. Nonlinear Sci. 2 (2) (2017) 385–402. Appl. Mech. Eng. 351 (2019) 226–252.
13