Analog Circuit Design Using Genetic Algorithms

This article proposes that genetic algorithms
(GAs) are viable candidates as tools for design automation of
analog circuits (DAAC). The principles, strengths, and weaknesses
of GAs applied to sizing circuit parameters to meet performance
specifications are discussed and a software implementation is
introduced. Designs of BJT and CMOS operational amplifiers, a CMOS
operational transconductor, and a matching network are presented as

The design of the analog portion of a mixed-signal integrated circuit often requires a large fraction of the overall design time, despite the fact that the analog circuit is often a relatively small portion of the overall circuit. Interestingly, much of the analog design time occurs after a candidate topology, having the
potential to meet required specifications, has been selected as a
building block for the system. This is due to the need to manually
iterate or “tweak” circuit parameters (such as, component values
and transistor sizes) to meet specifications. This iteration
process is particularly challenging when specifications must
include temperature and process corners. Moreover, designs using
short-channel MOSFET's further complicate the design process due to
a lack of manageable equations relating circuit parameters to
specifications that result. If design automation of analog circuits
(DAAC) is to be achieved, a solution to automating this phase of
the design process must be developed. The problem may be classified
as a global optimization problem.

While there have been efforts to develop optimization tools for
analog circuit design, they have generally been based on gradient
search methods. Unfortunately, these methods require initial
guesses, tend to get stuck in local minima, and have difficulty
with discrete variables and non-linear constraints. This paper
asserts genetic algorithms (GAs) may be a better alternative for global
optimization tools for DAAC.

GAs are based on the Darwinian principle of natural selection
and the concepts of natural genetics. GAs have many desirable
characteristics and offer significant advantages over traditional
methods. They are inherently robust and have been shown to
efficiently search large solution spaces containing discrete or
discontinuous parameters and non-linear constraints, without being
trapped in local minima. GAs do not require initial guesses or
derivative information and have often found non-intuitive solutions
to engineering problems.

While GAs have been applied extensively to many areas of
engineering, there are relatively few references in the analog
circuit literature that report the application of genetic
algorithms. In most instances only linear circuits were considered
(a relatively easy problem compared to non-linear circuits) or the
optimizations were based on derived circuit equations. Koza has performed non-linear circuit designs using
genetic programming, but emphasized topology generation. While his
results are very exciting, he has not yet demonstrated their
effectiveness in generating circuits suitable for production in an
industrial environment.

This article illustrates the application of GAs to select
component values and transistor sizes for a given circuit topology
to provide practical design solutions for a given set of
performance specifications. Since this phase of analog circuit
design is time intensive, significant time savings can be achieved
if a software tool can perform this task. The software discussed in
this article represents an important step in that direction.

The Genetic Algorithm

This section introduces some of the concepts of genetic
algorithms and their implementation in GA-Spice. It is provided
only as a guide and is not intended as a substitute for more
detailed expositions on GAs available in the literature.

Central to all genetic algorithms is the concept of the
chromosome. The chromosome contains all information necessary to
describe an individual. In nature, chromosomes are composed of DNA.
In a computer, a long binary or character string is used.
Chromosomes are composed of genes for the various characteristics
to be optimized and can be any length depending on the number of
parameters to be optimized.

Encoding defines the way genes are stored in the chromosome and
translated to actual problem parameters. A possible encoding scheme
for a hypothetical circuit using a 16-bit binary chromosome is
shown below:

Figure 1:   Encoding Example

Fitness is a single numerical quantity describing how well an
individual meets predefined design objectives and constraints.
Fitness can be computed based on the outputs of multiple analyses
using a weighted sum. The definition of good fitness functions is
highly problem dependent. The GA-SPICE program has several basic
fitness metrics that can be applied to any Spice output variable.
The basic metrics are described in terms of the output variables
c i , user specified
response c i , and user
specified point-by-point weighting values s i . All of the basic metrics are
defined such that the superior individuals have the lowest fitness
values. Using these definitions, the raw and standard fitness
defined by Koza are identical.

A raw fitness metric for minimizing an output variable ci
computed at N points can be defined as

Other metrics can also be defined to maximize an output variable
or to measure of the quality of a match of the calculated responses
to a specified response on either a relative or absolute basis.
Constraints are implemented by imposing a large penalty whenever
the constraint is violated. You can combine metrics for various
functions to yield a combined fitness for different output
variables, analysis types or circuit configurations. The total raw
fitness F is then calculated using

where W m is the
weighting applied to each of the basic fitness metrics. The
adjusted fitness, A, is defined by Koza as

and is used to rescale the fitness to the range 0 to 1 with
better individuals having the larger values. Subsequently, the
adjusted fitness of each individual is normalized by the total
adjusted fitness of the M members of the population.

The normalized fitness is used in determining the probability of
selecting a particular individual for mating.

Crossover is a method of exchanging genetic material between two
parents to produce offspring. The operation is simple, a crossover
point is chosen at random and the genetic information to the left
of the cross-over point in parent A is combined with the genetic
information to the right of the cross-over point in parent B to
produce new offspring. A second offspring can be created by using
the information from the right of parent A and the left of parent
B. The operation is illustrated with the example below:

Figure 2:   A single point crossover is widely used
to simulate the sharing of genetic information that occurs during

A simple GA is described in this section. While more
sophisticated forms of the GA exist, they are all extensions of
this basic algorithm. The simple GA provides excellent performance
for a wide range of problems.

The simple GA begins with the creation of an initial random
population of individuals. The size of the population depends on
the problem size. Population size ranges from as few as 20 to tens
of thousands. Values of several hundred are commonly used.

The next step is to evaluate the fitness of each individual
according to the predefined criteria. After the evaluation, the
individuals are ranked according to fitness. Individuals are
selected for mating based on fitness. Fitter individuals have a
higher probability of mating and passing on genetic information to
subsequent generations while less fit individuals have a non-zero
probability of mating to preserve diversity.

Mating is simulated by applying the crossover operation to the
chromosomes of two individuals selected based their fitness.
Mutation is simulated by randomly changing a few bits in the
chromosome of the offspring. Mutation provides a mechanism for
exploring new regions of the solution space and prevents premature
convergence to local minima. Finally, the fitness of the new
generation is evaluated and the process is repeated for a specified
number of generations or until a desired fitness is attained. A
block diagram of the simple GA is shown below:

Figure 3:   . The simple GA is adequate for many
engineering problems.

The GA-Spice Program

As part of this research, the GA-Spice program was developed.
The program allows the encoding of any parameter in a Spice netlist
as an optimization variable. This includes component values, node
numbers, model types, process parameters and other Spice
parameters. Fitness functions can be constructed using any output
variable available from Spice. Analysis types can include AC, DC,
transient, and noise. The designer can construct optimization
variables and fitness functions easily via a point & click
interface. The program is compatible with existing Spice tools for
schematic capture and output visualization. The general nature of
GA-Spice means the simulator is not limited to specific circuit
topologies like equation-based methods.

GA-Spice runs under the Microsoft Windows 9x/NT operating
systems. Since the GA is inherently a coarse-grain parallel
operation it is possible to reduce optimization time by employing
parallel computer architectures. The current version of GA-Spice
can use up to 4 processors in an symmetric multiprocessing system.
A more powerful multi-processor version based on the Parallel
Virtual Machine is under development.

Practical Applications of GA-Spice

The application of GA-Spice for the solution of several
practical analog circuit applications is presented in this

BJT Operational Amplifier
The specifications for the BJT op-amp depicted in Figure 5 were
a differential gain of at least 600V/V, an input offset voltage of
0V, and an output signal swing of at least ±13V. Several
circuit designs produced by GA-Spice met the specifications. One of
the circuits was built and tested in the laboratory to confirm that
the design was practical and met specifications.

Interestingly, this circuit was also given as a design problem
to junior electronics students. It required at least two weeks
(three weeks or more for many) for the students to obtain a design
that met specifications. Using GA-Spice, the design be performed in
under 30 minutes with a 200 MHz Pentium notebook computer.

Figure 5:   BJT Operational Amplifier

CMOS Operational Amplifier
The specifications for the CMOS Op-amp depicted in Figure 6 were
to maximize gain and minimize the input offset voltage. The results
from GA-Spice gave designs that had gains in excess of 2kV/V and
input offset voltages of 0V. Transistor widths were fixed at 5
microns (5µm) except for M 6
and M 7 which were fixed at
2um. GA-Spice required one hour on a 200 MHz Pentium notebook
computer to complete the design.

Figure 6:   CMOS Operational Amplifier

Operational Transconductance Amplifier (OTA) Design
The topology of the OTA is shown in Figure 7 . GA-Spice was used
to determine the transistor widths and resistor value for R 1 required to properly bias the
circuit and provide a specified transconductance. For this problem,
the transistor widths were fixed at 2 um. The chromosome length was
76 bits, yielding 7.5×1022 possible designs. Desipite
this enormous solution space, the program located an acceptable
solution in approximately 15 minutes on 200 MHz Pentium notebook
computer. The following table summarizes the design specifications
and the resulting circuit performance.

Figure 7:   Schematic diagram of OTA optimized by

W1=W2=6.96um, W3=W4=101.5um, W5=W6=4.6um, W7=W8=27.4um,
W9=112.6um, W10=3.23um, W11=375.8um, W12=7.56um, and R1
= 90.374k. The spice model for the NMOS used VTO=0.7V,
Kp=156uA/V, and lambda=30m/V and for the PMOS was VTO=-0.95V,
Kp=48uA/V, and lambda=4.2/V

Input Impedance Matching Network for an LNA
Figure 8 shows the schematic for a model of an input stage of a
low-noise amplifier. C1 and
C2 tune the input impedance, Z
in , to 50 at 2.45GHz. L BOND models bond-wire inductance, C
pad models pad and trace
capacitance, C par models trace
capacitance, R 3 and C 3 model the input impedance of a
biasing circuit, R in and C
in model the impedance looking
into the gate of a MOSFET, R S1 , R S2 , C S1 , C S2 , and R lin models parasitics of an
integrated square spiral inductor.

Figure 8:   Schematic for a model for the input stage
of a low-noise amplifier.

The results from GA-Spice provided Z in = 50.75 – j 3.5 at
2.45GHz. GA-Spice provided a solution after running about 15
minutes on a notebook computer with a 200MHz Pentium processor.

The values for the components are L BOND = 3nH, C pad = 500fF, C par = 12.56fF, R 3 = 1.796k, C 3 =
72.22fF, R in = 107.5, C in =
265.7fF, R S1 = 52.7, R S2 =
79.6, C S1
=113fF, C S2 =93.8fF, and R
lin = 21.9. C 1 and C
2 were selected by GA-Spice to
be 306fF and 20fF.


Section 4 described four circuits with component values selected
using a genetic algorithm. Each of the these circuits met a set of
simple specifications in a very reasonable amount of computer

The input matching circuit was an actual design problem in an
industrial setting. Realistic values for the components were used.
Linear circuits appear to be very easy for GA-Spice to

For the BJT and CMOS operational amplifiers and the CMOS
operational transconductor amplifier, initial designs provided by
GA-Spice proved unrealistic. For example, we found that GA-Spice
had biased several of the Gate-Source voltages (VGS's) of several
MOSFET's at levels of mere millivolts above the threshold voltage.
This is very impractical. Thus, the fitness functions for the
transistor circuits must be defined to prevent this. For example,
we specified the overdrive voltages of all MOSFET's had to be
between 0.2V and 0.5V for threshold voltages of 0.7V. The DC
biasing for the BJT op-amp required similar constraints.

The user also must recognize when transistors in a design must
be matched and properly link the parameters of the second
transistor to the first using a symbolic link option provided by
the program. For example, in the CMOS OTA, M3 and M4, M5 and M6,
and M7 and M8, are MOSFET pairs using this symbolic-link encoding
scheme to provide proper OTA operation. An attempt to use a genetic
algorithm without proper linkage produced a design that met
specifications, however it resulted in a poor overall design due to
the circuit symetry being destroyed.

The CMOS op-amp and OTA used a basic square-law model and not
the current industry-standard BSIM3v3 model. This helped speed up
simulation times and still demonstrates the ability of the program
to design analog circuits where equations are not available.

One measurement used in this paper is the simulation time of the
GA-Spice. It has been stated that for simple circuits, this time is
relatively small even on a slow notebook computer and could be
comparable to the design time of an engineer. However, as GA-Spice
is applied to larger circuits, perhaps the measurement will be the
amount of time saved by the engineer. Even if a GA cannot design a
circuit as quickly as an engineer, if it saves an engineer hundreds
of man-hours, it is still an effective tool.

In an industrial environment, specifications and the MOSFET
models will not be so simple. An effort to show the successful
application of GAs to industrial-strength analog circuits is
presently underway. Future research also includes determining the
types of genetic algorithms most suitable for analog circuit
design, determining the best types of encoding schemes, and
developing optimal fitness functions for the different circuit


This article demonstrates that the applications of GAs are viable
candidates for use in DAAC by successfully designing several
circuits. The methods used are general and can be applied to all
analog circuits. It has been observed that GA-Spice provides a
number of different solutions to a problem in a much shorter time
than trial and error. GA-Spice can help provide new insights into
circuit operation and produce novel designs. It appears GAs can
play a role for future tools for DAAC. At a minimum, they can speed
up or eliminate a considerable amount of the trial and error
engineering being done by many organizations today.

1 comment on “Analog Circuit Design Using Genetic Algorithms

  1. ljldjfsl
    April 17, 2014

    To find a good printer in Canada. I would suggest using the following companies: Toronto Printing Companies | Calgary Printing Companies | Winnipeg Printing Companies | Montreal Printing Companies | Regina Printing Companies | Saskatoon Printing Companies | Quebec Printing Companies | Ottawa Printing Companies | Edmonton Printing Companies | Vancouver Printing Companies | These companies provide the best and most affordable price in whole Canada. Other printing companies I suggest would be Print Quote Services | Printing Company | Best Local Printer | Some other local cities have great printers as well such as the following: Philadelphia Printing | Memphis Printing | San Francisco Printing

    If you are looking to print custom stickers. The following is my best choice: Label and sticker printing | Print Custom Sticker | Print Stickers | No only they offer the best sticker printing quality but also unbeatable price.

    Nowadays, it is hard to find quality print trader. This is my recommendation for best print trader in the united states Best Print Trader | Lowest Price Printing Wholesale | Print Outsourcing | Print Broker Trade

    Lastly, all print should comes with a state of the art Print Management Software. I would highly recommend this print software tool.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.