^
8th Seminar on Neural Network Applications in Electrical Engineering, NEUREL-2006 /m
~~~Faculty of Electrical Engineering, University of Belgrade, Serbia, September 25-27, 2006
+~~~~~it-1/er1ef.b. ,
tp>wwehie.r/e//ofrne.tI
Bias-Dependent Model of Microwave FET
S-parameters Based on Prior Knowledge ANNs
Zlatica D. Marinkovic, Student Member, IEEE, Olivera R. Pronic, Member, IEEE,
Vera V. Markovic, Member, IEEE
Abstract - The applications of artificial neural networks
(ANNs) in bias-dependent modeling of S-parameters of
microwave FETs have been proposed earlier. Here, a model
based on an ANN with additional prior knowledge at its
inputs (PKI ANN) is introduced. S-parameters of the device
that belongs to the same class as the modeled device are used
as the prior knowledge. The PKI concept allows ANN model
to be developed with less training data, which is very
advantageous when training data is expensive or time
consuming to obtain. The proposed modeling concept is
illustrated by an appropriate modeling example.
Keywords - Artificial neural network, microwave FETs,
prior knowledge input model, S-parameters.
I. INTRODUCTION
OWING to generalization capability, i.e. the capability
of providing the correct response even for the input
values not used during the training process, artificial
neural networks (ANNs) have appeared as a good
alternative to the conventional modeling in the field of
microwaves.
Unlike complex and time-consuming electromagnetic
models, once developed model based on ANNs give
responses almost instantaneously, because response
providing is based on performing basic mathematical
operations and calculating elementary mathematic
functions (such as an exponential or hyperbolic tangent
function). ANNs have the capability of approximating
any nonlinear function. Therefore, it is possible to
develop neural model from source-response data points.
ANNs have been applied to a wide range of modeling
problems in the area of microwaves, [1]-[11]. There are
ANN applications, either to passive devices, to active
devices, or to whole systems.
This paper deals with ANN modeling of small-signal
scattering parameters (S-parameters) of microwave FET
This work was supported by Ministry of Science and Environmental
Protection of Republic of Serbia, under the project 1351.
Z. D. Marinkovic is with the Faculty of Electronic Engineering,
University of Nis, Aleksandra Medvedeva 14, 18000 Nis, Serbia, (phone
+381-18-529-303; fax: +381-18-588-399; e-mail:
zlaticaAelfak.ni.ac.yu).
0. R. Proni6 is with the Faculty of Electronic Engineering, University
of Nis, Serbia (e-mail: oljapAelfak.ni.ac.yu).
V. V. Markovic is with the Faculty of Electronic Engineering,
University of Nis, Serbia (e-mail: veraAelfak.ni.ac.yu).
1-4244-0433-9/06/$20.00 (©2006 IEEE.
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 30,2023 at 13:30:52 UTC from IEEE Xplore. Restrictions apply.
WE WM EE
20
IEEE 106
(Field-Effect Transistors) transistors. Almost all of the
standard FET transistors, like MESFET (MEtal
Semiconductor FET) and HEMT (High Electron Mobility
Transistor), have a very low-noise level, and therefore
they are suitable to be applied to the low-noise
microwave active circuits.
When operating under small-signal conditions, the
microwave FETs can be characterized by the scattering
parameters. Processes of design, optimization and
analyses of the microwave active circuits require accurate
and reliable models of S-parameters of the microwave
transistors. Most of the existing small-signal models are
based on an equivalent circuit representation of the
device. The equivalent circuit elements are extracted
usually from the measured S-parameters within a
microwave circuit simulator.
Although the device S-parameters are bias-dependent,
most of the small-signal models are valid for only one
bias. For any other bias, it is necessary to repeat
extraction of the equivalent circuit parameters from the
measured S-parameters. That procedure can be very time-
consuming in the case when models for various bias
conditions are required.
Recently, a bias-dependent model of the FET S-
parameters has been proposed. It is a completely black-
box model consisting of a single multilayer ANN that
models S-parameters dependence on bias conditions and
frequency, [3], [4] and [9]. Once developed, such model
enables prediction of S-parameters in the whole device
operating range without additional optimizations or
additional measurements.
Here, a prior knowledge input (PKI) model is
introduced. It is an extension of the basic model - its
inputs are not only bias conditions and frequency, but
also prior knowledge about S-parameters. The main aim
of this model is to increase the modeling accuracy and to
achieve suitable modeling with as less training samples as
possible. In the proposed case, the ANN additional inputs
are S-parameters of the device that belongs to the same
class as the modeled device.
Validity of the proposed model is verified by the
appropriate modeling examples.
II. MICROWAVE TRANSISTOR S-PARAMETERS
Microwave FETs operating under small-signal
conditions can be characterized by the scattering
185
 ~Amn
parameters (S-parameters) which relate the voltage waves
incident on the ports to those reflected from the ports
(Fig. 1).
The scattering matrix, or [5] matrix, is defined in
relation to these incident and reflected voltage waves as,
[12]:
Vr iSInI S12 noVt+ati
or in matrix notation
[V- I = [S][V+ I
Fig. 1. Incident and reflected waves in a two-port network
An element of the [S] matrix can be determined as
U v Vk=O,zakj
i1k
Sii is the reflection coefficient seen looking into the
port i when all other ports are terminated in load matches.
S is the transmission coefficient from port j to port i
when all ports are terminated in matched loads.
The S-parameters of microwave transistors are
frequency-, temperature- and bias-dependent.
III. BIAS-DEPENDENT ANN MODEL
ANN has been applied for modeling bias- and
frequency- dependence of microwave transistor S-
parameters, [3], [4] and [9]. The model consists of an
MLP ANN trained to predict transistor S-parameters for
given bias voltage, bias current and frequency at its
inputs, Fig. 2. Therefore, the network has three input
neurons corresponding to:
* bias dc drain-to-source voltage, Vds,
* bias dc drain-to-source current, Ids'
* frequency f.
WI g(i )
Ang(S )
*ng(Sl)
ANN
I
Mag($s2
1 Ang\(S2.)
Fig.2. Bias-dependent ANN model
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 30,2023 at 13:30:52 UTC from IEEE Xplore. Restrictions apply.
The output layer consists of eight neurons
corresponding to magnitudes and angles of the scattering
parameters.
Number of the hidden layers can be one or two. The
network is trained using S- parameters' data for certain
number of bias points in the operating frequency range.
Generally, neural networks with different number of
hidden neurons are trained, tested and after their
comparison, the network with the best testing results is
chosen to be the bias-dependent model of S-parameters of
the modeled device.
(2) Once the model is developed, its structure remains
unchanged, and S-parameters for the given bias and
frequency are obtained by calculating ANN response to
the given input values. It should be emphasized that
measured values of the device S-parameters are required
only for the model development.
IV. BIAS-DEPENDENT PKI ANN MODEL
Prior knowledge (or existing models) can be used for
reducing the complexity of the input/output relationships
that an ANN has to learn, as it was proposed in [5]. This
reduction of input/output complexity allows an accurate
(3)
ANN model to be developed with less training data,
which is very advantageous when training data is
expensive or time consuming to obtain. Since the S-
parameters' measurements require complex equipment,
not available to most of the microwave designers, and
manufacturer given data are often not sufficient for
simple black-box ANN model development, a PKI ANN
model is introduced, Fig. 3.
The inputs of a PKI ANN model (ANN Devicel in Fig.
3) are not only bias conditions and frequency (as in the
basic ANN model), but also S-parameters corresponding
to these input values. We propose that the approximate
values are obtained by a basic ANN model of S-
parameters (ANN Device 0 in Fig. 3) developed earlier
for a device from the same class as the modeled device
(reference device).
In this way, as it will be presented in the following
example, a sufficiently accurate model can be developed
with reduced number of training data.
)
Fig.3. Bias-dependent PKI ANN model
186
 V. PKI ANN MODELING EXAMPLE acceptable.
The proposed PKI ANN model was developed for TABLE 1. TESTING RESULTS FOR THE BIAS POINTS USED IN THE
TRAINING PROCESS
ATF34143 pHEMT device (Hewlett Packard). The
measured S-parameter data for this device available on
the manufacturer web-site, [13], refer only to four bias IS,,1 1.899 5.874 0.99646
points (Vds Ids) in (0.5-18) GHz frequency range (23 Ang(Sll) 0.811 2.943 0.99968
frequency points per one bias point). Although, it is
possible to train an ANN to learn these data accurately, it IS211 1.546 5.077 0.99796
is not possible to achieve any acceptable generalization. Ang(S21) 0.741 2.495 0.99969
Therefore, the PKI ANN model was developed. The prior
knowledge input values were those obtained by the basic
IS121 1.228 6.717 0.99864
Ang(S12) 0.708 2.772 0.99979
ANN model (ANNO) developed for ATF35143 pHEMT
device, which belongs to the same class, [9]. The ANN IS221 1.565 6.311 0.99862
model of ATF35143 device has 10 neurons in each of two Ang(S22) 0.662 4.038 0.99963
hidden layers.
The available S-parameters' data for the modeled TABLE 2. TESTING RESULTS FOR THE BIAS POINTS NOT USED IN
device were divided in two subsets: the data referring to THE TRAINING PROCESS
three bias points were used as the ANN (ANNI) training
data and the data for the remaining bias point were used
as the validation test data. The angles of S-parameters IS,,1 14.201 39.758 0.89106
were expressed in one of (0°.360°) or (-180°.180°) Ang(S1l) 2.825 7.588 0.99901
ranges. The criterion for the range selection, for the angle
of each S-parameter, is the angle frequency dependence
IS211 8.014 18.392 0.98909
without sharp changes, in order to help ANN to increase Ang(S2l) 4.223 10.642 0.99942
modeling accuracy.
ANNs with 11 input, 8 output and different number of
IS121 3.957 14.370 0.98386
hidden neurons were trained. The prediction of S- Ang(S12) 7.353 17.555 0.99775
parameters by all of the obtained ANNs was tested for the IS221 8.672 18.155 0.97255
corresponding training and test sets. After comparison of |Ang(S22) 6.191 12.696 0.99835
the test results, the ANN with two hidden layers, with 10
neurons in each hidden layer, was chosen as the best
model. This can be confirmed from results in Fig 4, where
For the purpose of quantifying accuracy of the ANN prediction of the S-parameters for the remaining bias
model, average test error (ATE [0Io]), worst-case error point is shown. The PKI ANN obtained S-parameters are
(WCE [0/O]), and correlation coefficient, r, between the represented by lines and the reference (measured) values
referent and the modeled data are calculated, [1]. are represented by symbols. It can be observed that
The Pearson Product-Moment correlation coefficient r modeled values match well to the reference ones.
is defined by:

-E(xi -X)(Yi Y)y)

Zx
-
r -)(y_
ZI(Xi _x)2Z1(yi )2
where xi is referent value, Yi is the neural network
computed value, x- is the referent sample mean, and y is
the neural network sample mean. The correlation
coefficient indicates how well the modeled values match
the referent values. A correlation coefficient near one
indicates an excellent predictive ability, while a
coefficient near zero indicates poor predictive ability.
In Table 1 there are statistical results of prediction of
the S-parameters by the developed PKI ANN model for
bias points used for the ANN training. Low values of the
ATE and WCE and correlation coefficients very close to
one indicate that the network learnt the training data very
well. In Table 2 there are statistical results of prediction
of the S-parameters by the developed PKI ANN model
for the remaining bias point not used for the ANN
training. The ATE and WCE values are not as low as in
the case for the training data prediction, but they are
VI. CONCLUSION
(4)
Bias-dependent ANN model of S-parameter of
microwave FETs has been proposed earlier. The model
consists of a single ANN that models S-parameters
dependence on bias conditions (bias current and bias
voltage) and frequency. It is able to predict, with a good
accuracy, S-parameters for any given bias and frequency
point from the device operating range, either for those
used for the network training or for those presented to the
network for the first time. Once the model is developed,
its structure remains unchanged. Moreover, the measured
data are required only for the model training procedure.
However, since the S-parameters' measurements
require complex equipment not available to most of the
microwave designers and manufacturer given data are
often not sufficient for simple black-box ANN model
development, a PKI ANN model is introduced. PKI
model, as an extension of the above-mentioned basic
ANN model is introduced. The PKI ANN model includes
prior knowledge as the additional ANN inputs, which can

187
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 30,2023 at 13:30:52 UTC from IEEE Xplore. Restrictions apply.
 help ANN to extract input-output dependences.
0.9 - 0.16 - 120
360
0.8 - 0.14 - -Q
0.7 -
9: , ** 300 --------------------------------------- 60
240 0.12 -
0.6 - O 0
180 Z
0.5- ----------------- ---- ------- -------------- ---------------------- ----- -----------
-----
--,.- - - - -.., .;-... ... .... ... ... ... ... ... 0.10- '''''''''''/------- ----------- o-------,. . . . . . . . .
120 c -60 t
0.4 - '@./ *~~~~~~~~ " 0.08- =1
60 < --,,S,,,,,.,,,,.,,,,,,----'----------- '''''-'''''''''''''
0------'---- ---------------------- - -I- -& ------ ''
0.3 - ---
-
0 -120
0.06 -
0.2 - -60
symbols - reference values - 0.04 - ----------- symbols - reference values|--------' -180
0.1 -
lie neural
-,erl
li-nes model
o e ------------- --------------------------------- --------.. -120
lines - neural model
0.0 - - Is(t 02- , . . . . . "-
-240
2 4 6 8 10 12 14 16 18 0 2 4 6 8 10 12 14 16 18
f [GHz] f [GHz]
a) b)
11 - 180 0.81 360
I
10 - ---------------- symbols - reference values 120
lines neural model 300
9- -
.(

8-
.,,.,,,,,,.,,,,,,,,,,,,---O- ----- -----,------------ ----------------- -----,---------------''.'''''''''''''''','
60 0.6- o ,0 / 240
7-
-----------------

0 Z 180 _
6-
''----------- --------- -------------- ---------- ---,............. -----------'''' '''''''''''''''''''''''''
120
0~~~~~~~~~~~ -60 c 0.4-
5- 0~~~~~~~~~~
60 =
4- -120 <
0
*
.,,.,,.,,.,,,,,,,,,,, -----n-----------.
--------------
-.,o
3- -180 0.2- -60
2- \ o ~ S lines@ nera model
1- -240 -120
U .-JUU 0.0A
2 4 6 8 10 12 14 16 18 2 4 6 8 10 12 14
1

16
.,
-l18-180
f [GHz] f[GHz]
C) d)
Fig. 4. Scattering parameters for bias points not used for the training process: a)SI1 b) S12 c) S21 d) S22

S-parameters of the device belonging to the same class as [6] M. Vai, S. Prasad, Neural networks in microwave circuit design-
the modeled device (reference device) are used as prior beyond black box models, Int. J. RF and Microwave Computer-
Aided Eng., Special issue on Applications of Artificial Neural
knowledge. In that way, in the case when there are no networks to RF and Microwave Design, 1999, pp. 187-197.
enough training data, applying the PKI concept, it is [7] F. Giannini, G. Leuzzi, G. Orengo, M. Albertini, "Artificial neural
possible to achieve sufficiently good generalization of the network approach for MMIC passive and active device
ANN model. This is due to reducing the complexity of characterization" Proc. of Gallium Arsenide applications
the input/output relationships that an ANN has to learn symposium GAAS 2000, 2000, Paris.
-

[8] D. Schreurs, S. Vandenberghe, H. Taher, B. Nauwelaers, "ANN

caused by prior knowledge introduced to the ANN. Model For SiGe HBTs Constructed From Time-Domain Large-
Signal Measurements" Proc. of Gallium Arsenide applications
REFERENCES symposium GAAS 2002, 2002, Milano.
-

[9] Z.Marinkovic, V.Markovic, "Prediction of HEMT's Scattering and

[1] Q. J. Zhang, K. C. Gupta, Neural Networks for RF and Microwave Noise Parameters Using Neural Networks", Microwave Review,
Design, Artech House, 2000. Vol. 8, No. 1, 2002, pp. 28-31
[2] P. M. Watson, K. C. Gupta, "EM-ANN Models for Microstrip Vias [10] V.Markovic, Z.Marinkovic, "HEMT Noise Neural Model Based on
and Intreconnects in Dataset Circuits", IEEE Trans., Microwave Bias Conditions", The Int. Journal for Computation and
Theory Tech., Vol. 44, No-12, 1996, pp. 2495-2503. Mathematics in Electrical and Electronic Engineering COMPEL, -

[3] F.Gunes, H.Torpi, F.Gurgen, "Multidimensional signal-noise Vol. 23 No.2, 2004, pp.426-435.
neural network model", Circuits, Devices and Systems, IEE [11] V. Markovic, 0. Pronic, Z. Marinkovic, "Noise Wave Modeling of
Proceedings, Vol.145, Iss.2, Apr 1998, pp. 111-1 17. Microwave Transistors Based on Neural Networks", Microwave
[4] P.M. Watson, M. Weatherspoon, L. Dunleavy, G.L. Creech, and Optical Technology Letters, Vol. 41 No. 4, 2004, pp. 294-297.
"Accurate and Efficient Small-Signal Modelling of Active Devices [12] D. Pozar, Microwave Engineering, John Wiley & Sons, Inc., New
Using Artificial Neural Networks", GAAS IC Symp.Technical York, 1998.
Digest, 1998, pp 95-98.
[5] P. M. Watson, K. C. Gupta, R. L. Majahan, "Development of [13]
Knowledge Based Artificial Neural Network Models for
Microwave Compnents"IEEE Trans. Microwave Symp. Digest,
Baltimore, MD, vol. 44, pp. 9-12, Dec. 1998.

188Downloaded on May 30,2023 at 13:30:52 UTC from IEEE Xplore. Restrictions apply.
Authorized licensed use limited to: University of Electronic Science and Tech of China.