FINITE ELEMENT ANALYSIS FOR MINIMAL SHAPE

Vinicius Arcaro, Katalin Klinka, Dario Gasparini

This text describes a mathematical model for minimizing

path length, surface area and volume using a line element,
a triangle element and a tetrahedron element respectively.
The elements can also be used together through the unifying
concept of minimizing shape volume. The square of the
relative change in the element’s volume is used to define
an isovolumetric element. A quasi-Newton method is used,
which avoids the evaluation of the Hessian matrix as
required in a Newton method. The source and executable
computer codes of the algorithm are available from the
first author's website.

Keywords: element, line, membrane, minimization, nonlinear,

optimization, tetrahedron, triangle.

1 Introduction
The geometrical shape minimization problem can be
formulated as an unconstrained nonlinear programming
problem, where the objective function is the final total
volume and the displacements of the nodal points are the
unknowns. The shape volume is discretized into line,
triangle and tetrahedron elements. The elements are
interconnected at their nodal points. It is important to
emphasize that by construction, the final element state is
uniquely defined in terms of its nodal displacements.

The following conventions apply unless otherwise specified

or made clear by the context. A Greek letter expresses a
scalar. A lower case letter represents a column vector.

2 Line element definition

Figure 1 shows the geometry of the element. The nodes are
labeled 1 and 2. The nodal displacements transform the
element from its initial state to its final state.
 x2
v

x1
u2

u1

x2
v
1
x

Figure 1

v = v + u2 − u1

v = ( v Tv ) 2
1

The previous expression can be written as:

v = ⎡vTv + 2vT ( u2 − u1 ) + ( u2 − u1 ) ( u2 − u1 ) ⎤ 2
T

⎣ ⎦

2.1 Element final volume

Supposing that the line element has a fixed cross sectional

area α, its final volume can be written as:

φ = α v

2.2 Gradient of the element final volume

The nodal displacements vectors are numbered according to

its node numbers. Their individual components are numbered
as follows:
 ⎡u1 ⎤ ⎡u4 ⎤
u1
= ⎢u2 ⎥ , u = ⎢⎢u5 ⎥⎥
⎢ ⎥ 2

⎢⎣u3 ⎥⎦ ⎢⎣u6 ⎥⎦

The gradient of the element final volume with respect to

the nodal displacements can be written as:

1 ⎡ −v ⎤
∇φ = α
v ⎢⎣ + v ⎥⎦

3 Triangle element definition

Figure 2 shows the geometry of the element. The nodes are
labeled 1, 2 and 3 while traversing the sides in
counterclockwise fashion. Each side is labeled with the
number of its opposite node. The nodal displacements
transform the element from its initial state to its final
state.

x3
v2
x1 v1
3
u
x2
v3
u1
x3
2
v
v1 u2

x1
v3
x2
Figure 2

v1 = v1 + u3 − u2

v2 = v2 + u1 − u3

v3 = v3 + u2 − u1

w = v1 × v2
w = v1 × v2

The previous expression can be written as:

w = w +
+ v1 × u1 + v2 × u2 + v3 × u3 +
+u1 × u2 + u2 × u3 + u3 × u1

The vectors w and w are orthogonal to the element in the

initial state and final state respectively. Notice that
these vectors point toward the observer.

3.1 Element final volume

Supposing that the triangle element has a fixed thickness λ,

its final volume can be written as:

λ
φ = w
2

3.2 Gradient of the element final volume

The derivatives of the element final volume with respect to

the nodal displacements can be written as:

dφ λ ⎛ ∂w1 ∂w2 ∂w3 ⎞

= ⎜ w1 + w2 + w3 ⎟
j
dui 2 w ⎝ ∂ui
j
∂ui
j
∂uji ⎠

The nodal displacements vectors are numbered according to

its node numbers. Their individual components are numbered
as follows:

⎡u1 ⎤ ⎡u4 ⎤ ⎡u7 ⎤

u1
= ⎢u2 ⎥ , u = ⎢u5 ⎥ , u = ⎢⎢u8 ⎥⎥
⎢ ⎥ 2 ⎢ ⎥ 3

⎢⎣u3 ⎥⎦ ⎢⎣u6 ⎥⎦ ⎢⎣u9 ⎥⎦

The gradient of the element final volume can be written as:

 ⎡ w × v1 ⎤
λ ⎢ ⎥
∇φ = ⎢ w × v 2
⎥
2 w
⎢ −w × v1 − w × v2 ⎥
⎣ ⎦

4 Tetrahedron element definition

Figure 3 shows the geometry of the element. The base nodes
are labeled 1, 2 and 3 while traversing the sides in
counterclockwise fashion looking from the apex, which is
labeled 4. The nodal displacements transform the element
from its initial state to its final state.

x4

v1 v3
v2
x1
x3
u4
u1
x2
x4
u2 u3
1
v
v3
v2
x1

x3
x2

Figure 3

v1 = v1 + u1 − u4

v2 = v2 + u2 − u4

v3 = v3 + u3 − u4

4.1 Element final volume

The final volume of the tetrahedron element can be written
as:

φ = − ( v ) ( v × v3 )
1 1 T 2
6

Note that if the apex moves below the base, the volume
becomes negative.

v1 = v1 + u1 − u4 ⎫
⎪
v2 = v2 + u2 − u4 ⎬ ⇒
v3 = v3 + u3 − u4 ⎪⎭

(
1 1
v + u1 − u4 ) ⎡⎣( v2 + u2 − u4 ) × ( v3 + u3 − u4 ) ⎤⎦
T
φ = −
6

di = ui − u4 ⇒

−6φ = ( v1 ) (v × v3 ) +
T 2

+ ( d1 )(v × v3 ) + ( d 2 )
(v × v1 ) + ( d3 ) (v × v2 ) +
T 2 T 3 T 1

+ ( v ) (d
1 T 2
× d ) + ( v ) (d
3 2 T 3
× d1 ) + ( v ) (d
3 T 1
× d2 ) +

+ (d ) (d
1 T 2
× d )3

4.2 Gradient of the element final volume

The nodal displacements vectors are numbered according to

its node numbers. Their individual components are numbered
as follows:

⎡u1 ⎤ ⎡u4 ⎤ ⎡u7 ⎤ ⎡u10 ⎤

u1
= ⎢u2 ⎥ , u = ⎢u5 ⎥ , u = ⎢u8 ⎥ , u = ⎢⎢u11 ⎥⎥
⎢ ⎥ 2 ⎢ ⎥ 3 ⎢ ⎥ 4

⎢⎣u3 ⎥⎦ ⎢⎣u6 ⎥⎦ ⎢⎣u9 ⎥⎦ ⎢⎣u12 ⎥⎦

The gradient of the element final volume can be written as:

 ⎡ v3 × v 2 ⎤
⎢ ⎥
1⎢ v × v
1 3
⎥
∇φ =
6 ⎢ v2 × v1 ⎥
⎢ 3 1⎥
⎣⎢ −v × v − v × v − v × v ⎥⎦
2 1 3 2

5 Isovolumetric element definition

An element that does not change its volume is desirable for
many problems types. The square of the relative change in
the element’s volume can be used to define an isovolumetric
element. Ideally, the volume change should be zero for the
isovolumetric element in its final state.

5.1 Element final volume

Considering φ0 and φ as the initial and final volumes

respectively, an isovolumetric element can be defined
through the following volume function, where η is a penalty
parameter.

2
ηφ0 ⎛ φ − φ0 ⎞
φ = ⎜ ⎟
2 ⎝ φ0 ⎠

5.2 Gradient of the element final volume

Using the chain rule, the gradient of the element final

volume with respect to its nodal displacements can be
written as:

⎛ φ − φ0 ⎞
∇φ = η ⎜ ⎟ ∇φ
⎝ φ0 ⎠

5.3 Severe cancellation

Inaccuracy often results from severe cancellation that

occurs when nearly equal values are subtracted. Notice that
inaccuracy can result from the difference between the final
and initial volumes because they can be arbitrarily close
for the isovolumetric element. However, severe cancellation
can usually be eliminated by algebraic reformulation.

φ
ε = −1
φ0

ηφ0 2
φ = ε
2

∇φ = ηε∇φ

The expression for ε, already reformulated in an attempt to

avoid severe cancellation, can be written for the line,
triangle and tetrahedron elements.

Line element

1
v̂ = v
v

z =
1
v
(u2
− u1 )

ˆTz + zTz
2v
ε =
v̂ + z + 1

Triangle element

w = w
w ˆ

w z =
+ ( v1 − u2 ) × u1 +
+ ( v2 − u3 ) × u2 +
+u3 × ( v1 + v2 + u1 )

ˆTz + zTz
2w
ε =
ŵ + z + 1

Tetrahedron element
 1
v̂i = i
vi
v

zi =
1
vi
(ui
− u4 )

( vˆ ) ( vˆ × vˆ ) ε =
1 T 2 3

+ (v ˆ ) (v
1 T
ˆ + z) ×z
2 2 3
+

+ (v ˆ ) (v
2 T
ˆ + z) ×z
3 3 1
+

+ (v ˆ ) (v
3 T
ˆ + z) ×z
1 1 2
+

+ (z ) (z × z )
1 T 2 3

6 Final shape volume

Considering u as the vector of unknown displacements, the
final shape volume function π and its gradient can be
written as follows:

π (u) = ∑
elements
φ (u)

∇π ( u ) = ∑
elements
∇φ ( u )

In order to find the local minimum points of a nonlinear

multivariate function, the general strategy that can be
used is: Choose a starting point and move in a given
direction such that the function decreases. Find the
minimum point in this direction and use it as a new
starting point. Continue this way until a local minimum
point is reached. In the quasi-Newton method, a descent
direction or a direction such that the function decreases
can be defined with the help of the gradient vector. The
advantages of this approach are: It is not necessary to
derive an expression for the Hessian matrix and it is not
necessary to solve any system of equations. The computer
code uses the limited memory BFGS to tackle large scale
problems as described by [3]. It also employs a line search
procedure with safeguards as described by [1].
7 Examples
The primary colors red, green and blue are used for the
line, triangle and tetrahedron elements respectively. The
secondary colors cyan, magenta and yellow are used for the
isovolumetric line, triangle and tetrahedron elements
respectively.

Example 1: An initially flat circular surface with

thickness = 2, radius r = 1 and the boundary displaced
according to the following hyperbolic paraboloid equation,
where h = 1/2. Figure 4 shows the meshes for the initial
and final surfaces.

Figure 4

⎡⎛ x ⎞2 ⎛y⎞ ⎤
2

z = h ⎢⎜ ⎟ − ⎜ ⎟ ⎥
⎣⎝ r ⎠ ⎝r⎠ ⎦

Example 2: A cylinder with thickness = 1, radius r = 1 and

height h = 1. Figure 5 shows the meshes for the initial and
final surfaces.
 Figure 5

The final surface is symmetrical about the Z axis. The

following analytical solution for the cross-section of the
surface in the YZ plane is described by [2].

⎛z⎞
y (z) = c cosh ⎜ ⎟
⎝c⎠

Note that y(0) = c. The value c is a solution of the

following equation. In this example, c = 0.8483379.

2r 2c ⎛h⎞
= cosh ⎜ ⎟
h h ⎝ 2c ⎠

Table 1 shows the relative error for y(0) with different

initial meshes.
Table 1
Elements y(0) Error
384 0.8470554 0.15 %
1536 0.8480256 0.04 %
6144 0.8482543 0.01 %

Example 3: A frustum cone with thickness = 2, upper radius

= 0.5, lower radius = 1 and height = 0.9. Figure 6 shows
the meshes for the initial and final surfaces.
 Figure 6

Example 4: On the top, Figure 7 shows an initial path

through the corner points of a rectangle with horizontal
dimension equal to 4 and vertical dimension equal to 2. The
line elements have area = 1. On the bottom, Figure 7 shows
the final path.

Figure 7
The general problem of connecting n points by the shortest
path length is called Steiner problem [2]. Its solution
contains straight lines intersecting at 120°. The number of
intersections is between zero and (n - 2).

Example 5: An initially flat square surface with thickness

= 1, side = 1 and two opposite corners displaced by +1/2
while the two other opposite corners displaced by -1/2. The
edges have line elements with area = 5. Figure 8 shows the
meshes for the initial and final surfaces.

Note that minimizing the total volume results in opposite

effects on the lengths of the free edges. It tends to
decrease the surface area defined by the triangle elements
and consequently tends to increase the lengths of the free
edges. It tends to decrease the path lengths defined by the
line elements and consequently tends to decrease the
lengths of the free edges. In this example, the free edges
are curved due to relatively small value for the areas of
the line elements.

Figure 8

Example 6: An initially flat square surface with thickness

= 1, side = 1 and two opposite corners displaced by +1/2
while the two other opposite corners displaced by -1/2. The
edges have line elements with area = 500. Figure 9 shows
the meshes for the initial and final surfaces.

Note that minimizing the total volume results in opposite

effects on the lengths of the free edges. It tends to
decrease the surface area defined by the triangle elements
and consequently tends to increase the lengths of the free
edges. It tends to decrease the path lengths defined by the
line elements and consequently tends to decrease the
lengths of the free edges. In this example, the free edges
are straight due to relatively big value for the areas of
the line elements.

Figure 9

Example 7: A straight prismoid with height = 3. The bottom

and top regular triangles are inscribed in a circle of
radius = 1. It is composed by 3 line elements and 9
isovolumetric line elements. The line elements have area =
1. The penalty parameter = 1.0E+05. The top triangle
rotates 150 degrees clockwise relatively to the bottom
triangle. Figure 10 shows the initial and final shapes.
Appendix 1 presents analytical expressions for this type of
prismoid.
 Figure 10

Example 8: A straight prismoid with height = 2. The bottom

and top regular pentagons are inscribed in a circle of
radius = 0.75 and 0.5 respectively. It is composed by 5
line elements and 15 isovolumetric line elements. The line
elements have area = 1. The penalty parameter = 1.0E+05.
The top pentagon rotates 126 degrees clockwise relatively
to the bottom pentagon. Figure 11 shows the initial and
final shapes. Appendix 1 presents analytical expressions
for this type of prismoid.
 Figure 11

Example 9: A straight prismoid with height = 6.651155. The

projection of the bottom and top closed polygonal chains is
a regular 12-sided polygon inscribed in a circle of radius
= 1. The parameter delta = 0.4277998. The Z coordinates of
the bottom points alternate between +/- delta while the Z
coordinates of the corresponding top points alternate
between the height of the prismoid -/+ delta. The
corresponding bottom and top points are connected by
isovolumetric elements only when its distance is the
smallest between the two possible values. Isovolumetric
elements connect successive top points, and also successive
bottom points, of the polygonal chain. This prevents the
collapse of the prismoid into a line. The prismoid is
composed by 12 line elements and 54 isovolumetric line
elements. The line elements have area = 1. The penalty
parameter = 1.0E+05. The top polygonal chain rotates 120
degrees clockwise relatively to the bottom polygonal chain.
Figure 12 shows the initial and final shapes. Appendix 2
presents analytical expressions for this type of prismoid.
 Figure 12

Example 10: A circular prismoid with axis on a

circumference of radius = 10. The section is defined by a
regular triangle inscribed in a circle of radius = 1. It is
composed by 30 line elements and 60 isovolumetric line
elements. The line elements have area = 1. The penalty
parameter = 1.0E+05. Figure 13 shows the initial and final
shapes.
 Figure 13

Example 11: A square with side = 1 composed by 8

isovolumetric triangle elements with thickness = 10. The
square’s perimeter has line elements with area = 1. The
penalty parameter = 1.0E+05. The triangle elements preserve
the square’s area while the line elements minimize its
perimeter. The square turns into an octagon. Figure 14
shows the initial and final areas.
 Figure 14

Example 12: A cube with side = 2 composed by 24

isovolumetric tetrahedron elements. The cube’s surface has
triangle elements with thickness = 1. The penalty parameter
= 1.0E+05. The tetrahedron elements preserve the cube’s
volume while the triangle elements minimize its surface.
The cube turns into a 24 faces polyhedron. Figure 15 shows
the initial and final volumes.

Figure 15
8 Appendix 1
Figure 16 shows a straight prismoid. The bottom and top
regular polygons are inscribed in circles of different
radius. The sum of the lengths of the red lines is
minimized by rotating the top polygon counterclockwise with
respect to the bottom polygon, while the lengths of the
black lines remain constant.

pi

pi

Figure 16

8.1 Geometry

For n-sided regular polygons, the coordinates of the

vertices can be written as:

2π
γ =
n

⎡r cos ( γi) ⎤
p = ⎢⎢r sin ( γi) ⎥⎥
i

⎢⎣ 0 ⎥⎦

⎡ r cos ( θ + γi) ⎤
p = ⎢⎢r sin ( θ + γi) ⎥⎥
i

⎢⎣ hθ ⎥⎦

bi = pi + 1 − pi
di = pi − pi + 1

vi = pi − pi

8.2 Final height

The square of the norm of vector vi can be written as:

vi = pi − pi ⇒

2
vi = r2 + r2 − 2rr cos θ + h2θ

Since this norm is constant, it leads to the following

expression that relates the initial and final heights.

h2θ = h02 − 2rr (1 − cos θ )

8.3 Rotation angle

The square of the norm of vector di can be written as:

bi + di − vi = 0 ⇒

− 2r2 (1 − cos γ ) − 2rr ⎡⎣cos ( θ − γ ) − cos θ⎤⎦

2 2 2
di = vi + bi

Due to symmetry, minimizing the sum of the norms of all

vectors is equivalent to minimizing the square of the norm
of only one vector.

2
∂ di sin γ
= 0 ⇒ tan θ =
∂θ cos γ − 1

Notice that this expression is valid when the vectors

connect the corresponding bottom and top points in any
symmetric way. Table 2 shows the rotation angle in degrees
and the values for the relation between the initial and
final heights for some n-sided regular polygons.

h02 − h2θ 1
= (2 − 2 cos γ ) 2 + 2
rr
 Table 2
n θ h02 − h2θ
rr
3 150.0 3.732051
4 135.0 3.414214
5 126.0 3.175571
6 120.0 3.000000
7 115.7 2.867767
8 112.5 2.765367
9 110.0 2.684040
∞ 90.0 2.000000

8.4 Minimum initial height

The minimum initial height is given by:

1
h02 ≥ rr (2 − 2 cos γ ) 2 + 2rr ⇒ hθ ≥ 0

9 Appendix 2
Figure 17 shows a straight prismoid. The projection of the
bottom and top closed polygonal chains is a regular polygon
inscribed in a circle of radius = r. The Z coordinates of
the bottom points alternate between ±δ while the Z
coordinates of the corresponding top points alternate
between the height of the prismoid ∓ δ . The corresponding
bottom and top points are connected by elements only when
its distance is the smallest between the two possible
values. The sum of the lengths of the red lines is
minimized by rotating the top polygonal chain
counterclockwise with respect to the bottom polygonal
chain, while the lengths of the black lines remain
constant.
 pi + 2
pi

pi + 2

pi
Figure 17

9.1 Geometry

For n-sided regular polygons, the coordinates of the

vertices can be written as:

2π
γ =
n

⎡r cos ( γi) ⎤
⎢ ⎥
pi = ⎢r sin ( γi) ⎥
⎢ ⎥
⎣ ( −1) δ ⎦
i

⎡r cos ( θ + γi) ⎤
⎢ ⎥
pi = ⎢r sin ( θ + γi) ⎥
⎢ ⎥
⎣ hθ − ( −1) δ ⎦
i

bi = pi + 1 − pi

di = pi − pi + 1

vi = pi − pi

An expression for δ can be written by setting the angle

between the vectors (pi+1 – pi) and (pi+2 – pi) equal to α.
⎛δ⎞ ⎛ 1 + cos γ
2
1⎞
⎜ ⎟ = (1 − cos γ ) ⎜ − ⎟
⎝r⎠ ⎝ 4 cos α 2⎠
2

9.2 Final height

The square of the norm of vector vi can be written as:

vi = pi − pi ⇒

= 2r2 (1 − cos θ ) + h2θ + 4δ2 − ( −1) 4δhθ

2 i
vi

Notice that there are two different values for this norm.
The smallest value, associated with i even, is given by:

= 2r2 (1 − cos θ ) + h2θ + 4δ2 − 4δhθ

2
vi

Since this norm is constant, it leads to the following

expression that relates the initial and final heights.

( hθ − 2δ ) = ( h0 − 2δ ) − 2r2 (1 − cos θ )
2 2

9.3 Rotation angle

The square of the norm of vector di can be written as:

bi + di − vi = 0 ⇒
2 2 2
di = vi + bi +
−2r2 (1 − cos γ ) − 2r2 ⎡⎣cos ( θ − γ ) − cos θ⎤⎦ + 4hθδ − 8δ2

Due to symmetry, minimizing the sum of the norms of all

vectors di is equivalent to minimizing the square of the
norm of only one vector.

2
∂ di
= 2r2φ ( θ ) = 0
∂θ
Notice that this expression is valid when the vectors
connect the corresponding bottom and top points in any
symmetric way.

φ ( θ ) = (cos γ − 1) sin θ − sin γ cos θ +

2δ sin θ
− 1
⎡( h0 − 2δ )2 − 2r2 (1 − cos θ ) ⎤ 2
⎣ ⎦

sin γ
tan θ =
cos γ − 1

The rotation angle is in the following interval.

θ < θ < π

9.4 Minimum initial height

The minimum initial height, which is associated with θ = 180

degrees, is given by:

h0 ≥ 2r + 2δ ⇒ hθ ≥ 2δ

9.5 Special case

The initial height can be set in order that a 2n-sided

prismoid with n vertical elements will rotate the same as
its associated n-sided prismoid with n vertical elements.

sin (2γ )
tan θ = ⇒
cos (2γ ) − 1

2
⎛ h0 δ⎞
2
⎡ 2 cos γ δ⎤
⎜ − 2 ⎟ = ⎢ ⎥ + 2 (1 + sin γ )
⎝r r⎠ ⎣ (1 − cos γ ) r ⎦

hθ 2 δ
=
r (1 − cos γ ) r

Table 3 shows the initial and final heights for α = 60

degrees.
 Table 3
n h0 hθ θ
r r
6 3.808383 2.828427 150.0
8 4.603421 4.060207 135.0
10 5.595633 5.236068 126.0
12 6.651155 6.386283 120.0
14 7.730797 7.522434 115.7

10 Appendix 3

⎡ ∂v1 ∂v1 ∂v1 ⎤

⎢ ⎥
⎢ ∂x1 ∂x2 ∂x3 ⎥
⎡ 0 −a3 a2 ⎤
⎢ ∂v ∂v2 ∂v2 ⎥ ⎢
v = a× x ⇒ ⎢ 2 ⎥ = ⎢ a3 0 −a1 ⎥⎥
⎢ ∂x1 ∂x2 ∂x3 ⎥
⎢⎣ −a2 a1 0 ⎥⎦
⎢ ∂v3 ∂v3 ∂v3 ⎥
⎢ ⎥
⎣ ∂x1 ∂x2 ∂x3 ⎦

w = w +
+ v1 × u1 + v2 × u2 + v3 × u3 +
+u1 × u2 + u2 × u3 + u3 × u1

⎡ ∂w1 ∂w1 ∂w1 ⎤

⎢ j ⎥
⎢ ∂u1 ∂uj2 ∂u3j ⎥
⎡ 0 −vj3 vj2 ⎤
⎢ ∂w2 ∂w2 ∂w2 ⎥ ⎢ ⎥
⎢ j j⎥
= ⎢ vj3 0 −v1j ⎥
⎢ ∂u1 ∂uj2 ∂u3 ⎥
⎢ −v2 v1j
j
0 ⎥⎦
⎢ ∂w3 ∂w3 ∂w3 ⎥ ⎣
⎢ j ⎥
⎣ ∂u1 ∂uj2 ∂u3j ⎦

11 References
[1] Gill, P. E. and Murray, W. Newton type methods for
unconstrained and linearly constrained optimization,
Mathematical Programming 7, 311-350, 1974.

[2] Isenberg, C., The Science of Soap Films and Soap

Bubbles, Dover Publications, 1992.
[3] Nocedal, J. and Wright, S. J., Numerical Optimization,
Springer-Verlag, 1999.