Professional Documents
Culture Documents
Microelectronic Applications
Xin-She Yang, Mehmet Karamanoglu
Simon Fong
I. I NTRODUCTION
In many engineering design and industrial applications, one
of the design objectives is to design optimal geometrical
configurations to suit for a particular purpose. For example,
in structural engineering, engineers often have to design the
best way that the support beams are placed in the right places
so as to carry the maximum loads, while the bending and
shear stresses are within the limits imposed by stringent design
codes. In many cases, design domains are not fixed, and
materials can be placed anywhere so that the final structure can
have an optimal shape. The geometry or shape of the structure
is one of the main design objectives [1], [18]. For example, in
microelectronic applications, heat transfer can be an important
issue. Even with the best designs from the electronic point of
view, the layout of materials for microdevices can be very
challenging so that designs are also optimal for heat management. Miniaturization of electronic devices in communications
makes it more important to find optimal shape and distribution
of materials, or more generally the topology of the designs.
Such shape or topology optimization is the main interest
of this paper. In general, topological shape optimization is
much harder to solve than nonlinear problems in a fixed
domain. The geometrical domain of a topology design problem
can be considered fixed, and they can typically be a regular
region (such as a continuous, rectangular region). However,
their design space or search space of the variables can be
significantly larger. For example, we may want to fill a
rectangular region with two materials, though the geometrical
domain is fixed and can be discretized or subdivided into many
978-1-4673-5861-3/12/$31.00 2012 IEEE
150
(q)d,
(1)
subject to a set of nonlinear constraints or differential equations [1], [14]. The whole domain can be discretized by a
discrete index q {0, 1} which only takes 0 (no material)
and 1 (filled with a given material). This is often true in
151
B
T =1
T =0
structural shape optimization. In many applications, the objective function does not have any explicit dependence on the
design variables, and thus the objective cannot be written in
simple analytical form. In this case, its dependence is implicit
or black-box type. This is especially true for heat management
in microdevices.
Heat management, basically heat transfer modelling and design optimization, is very important for many microelectronic
applications, especially those using large-scale integrated circuits and miniaturization. In fact, nanoscale heat transfer is an
interesting area, and topological optimization for the design
of a nanoscale device is even more challenging [22], [16]. For
example, Evgrafov et al. proposed a topology optimization
benchmark for a nanoscale heat-conducting system with a
size of 150 nm by 150 nm [5]. In real-world applications,
heat transfer can occur at many different scales, and heat
management at smaller scales may be more difficult to control.
Now we extend this benchmark to a standard unit area of 1 mm
by 1 mm, and the aim is to distribute two different materials so
as to maximize the temperature difference |TA TB | at these
two points A and B under the boundary conditions given in
[5] (see Fig. 1).
Two materials used in the design of the unit area have heat
diffussivities of K1 and K2 , respectively. In addition, K1
K2 . For example, Si and Mg2 Si, K1 /K2 10. The domain
is continuous under heat flux conditions at the top and bottom
boudaries and the objective is to distribute the two materials
such that the difference |TA TB | is as large as possible.
In order to formulate this shape optimization properly [7],
[8], [11], [12], we can write
Maximize |TA(q) TB (q)|,
(2)
subject to
hj (x, y, q) = 0, (j = 1, ..., J),
(3)
gk (x, y, q) 0, (k = 1, ..., K),
(4)
where the parameter/index vector q will be determined by
optimization, subject to J equalities and K inequalities. In
= v ti + (xti x )fi ,
v t+1
i
(6)
(7)
where [0, 1] is a vector randomly drawn from a uniform
distribution. After comparing all the solutions among all the
n bats at each iteration t, a current best solution x that
represents the global best solution can be obtained. This
current best solution corresponds to the best value of the
objective at the current iteration t. As the velocity increment
is defined by the product i fi , fi (or i ) is used to refine
the velocity change while keeping the other factor i (or fi )
constant. The configuration can be readjusted depends on the
domain of the problem of interest. The settings of fmin = 0
and fmax = O(1), are used in our experimentation, and again
it depends on the domain size of the problem of interest. A
random value is assigned as the frequency of each bat, which
is taken uniformly from [fmin , fmax].
In the local search, a new solution is drawn locally by using
a random walk around the current best solution, and it is then
tested to see if it is the best among all the solutions. The
random walk is defined as
xnew = xbest + At ,
(8)
where is a number vector randomly drawn from [1, 1].
At =<Ati > is defined as the average loudness of all the bats
at this time step.
The process of the updating the velocities and the locations
of bats are somewhat similar to that of the standard particle
swarm optimization. The pace and range of the movement
are basically controlled by fi , just like the movement of the
swarming particles. To certain extent, the bat algorithm is
deemed to be a balanced combination of swarm optimization
and the intensive local search, that are governed by the
frequency tuning ability and the variables of loudness and
pulse rate.
Nevertheless, these factors that influence the balance, the
loudness Ai and the pulse emission rate ri , are subject to
iterative updates in the optimization process. The value for
loudness can be chosen arbitrarily by the user at will, as long
as the rules are adhered the loudness subsides and the pulse
emission rate escalates, when a bat is approaching near to its
prey. For simplicity sake, we apply A0 = 1 and Amin = 0.
Letting Amin = 0 assumes that a bat has just found the prey
hence temporarily pause on emitting any sound. Now we have
the following parameter control [19]
At+1
= Ati ,
rit = ri0 [1 exp(t)],
(9)
i
where and are constants.
Actually, has the effect of the cooling factor in a cooling
schedule as in the simulated annealing algorithm. For any 0 <
< 1 and > 0, we have
(10)
Ati 0, rit ri0 , as t .
For a simplest solution, = can be used. In our simulations,
we let = = 0.95.
= xti + v t+1
,
xt+1
i
i
g1 (x) = 1
d L
71785w 4
0,
g2 (x) = 1
140.45w
d2 L
0,
g3 (x) =
2(w+d)
3
g4 (x) =
d(4dw)
w 3 (12566dw)
(12)
1 0,
+
1
5108w 2
g3 =
1.93L31
hzd41
1 0,
1 0,
1.93L32
hzd42
1 0,
q
1
1 2
6
g5 = 110d
( 745L
3
hz ) + 16.9 10 1 0,
1
q
1
2 2
6
( 745L
g6 = 85d
3
hz ) + 157.5 10 1 0,
g4 =
(17)
g7 =
hz
40
1 0,
g8 =
5h
b
1 0,
g9 =
b
12h
1 0,
g10 =
1.5d1 +1.9
L1
1 0,
g11 =
1.1d2 +1.9
L2
1 0.
2.6 b 3.6,
(13)
Using BA with n = 25, = = 0.95 and 1000 iterations,
we have obtained the same or slightly better solutions than the
best solution obtained by Cagnina et al. [3]
f = 0.012665
(14)
with
x = (0.051690, 0.356750, 11.287126),
(15)
but bat algorithm uses significantly fewer evaluations.
0.25 d 1.3,
397.5
bh2 z 2
1 0,
g2 =
0.7 h 0.8,
2.0 L 15.0.
17 z 28,
7.3 L1 8.3,
2.9 d1 3.9,
7.8 L2 8.3,
5.0 d2 5.5.
(18)
(19)
(20)
153
1
1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0
0
0.2
0.4
0.6
0.8
0
0
0.2
0.4
0.6
0.8
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
0.8
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
0.8
155