1752 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO.
6, JUNE 2016
A Novel 4-D Artificial-Neural-Network-Based
Hybrid Large-Signal Model of GaAs pHEMTs
Yunshen Long, Zheng Zhong, Member, IEEE, and Yong-Xin Guo, Senior Member, IEEE
Abstract— A novel hybrid large-signal model of GaAs
pseudomorphic HEMTs (pHEMTs) is proposed for monolithic
microwave integrated circuit design. This new model is
based upon accurate electromagnetic (EM) description and
creative multi-path artificial neural network (ANN) optimization.
To precisely describe the EM effect in the high-frequency range,
the extrinsic part of this model includes both lumped and
distributed components. In order to re-grid the discrete data,
the bias-dependent intrinsic elements are determined by ANNs
rather than traditional interpolations. The dispersion effect is
represented by nonlinear sources with the multi-path-dependent
integration technique, which is described by processed multi-bias
S-parameters. This proposed approach can be applicable to
different bias conditions, which is also verified by different types
of GaAs pHEMTs with good agreement. In addition, a class-AB
Ka-band power amplifier and a Ka-band switch using a 0.15-µm
GaAs pHEMT process were designed based on the novel hybrid
model for further practical verification.
Index Terms— Artificial neural network (ANN)
three-dimensional nonlinear function, hybrid large-signal
model, inconsistency between RF and dc current, integration
path independence.
I. I NTRODUCTION
I N RECENT years, GaAs pseudomorphic
HEMTs (pHEMTs) are widely used in advanced nonlinear
microwave circuits. As it is well known, computer-aided
design (CAD) can assist to predict the behavior of circuits
and especially where accurate large-signal models play an
important role in active large-signal circuit designs, such as
power amplifiers and mixers.
Much work has been undertaken to characterize the devices
based on a broad range of compact models. In general,
empirical-function-based models primarily dominate in the
research and industry fields [1]–[4], [31]. However, empirical
models heavily rely on the detailed expression of the
nonlinear functions. These functions are associated with many
Manuscript received January 14, 2015; revised May 13, 2015,
September 19, 2015, and February 5, 2016; accepted April 14, 2016. Date
of publication April 28, 2016; date of current version June 2, 2016. This
work was supported in part by the Singapore Ministry of Education Academic
Research Fund Tier 2 under Grant MOE 2014-T2-2-15, in part by the Jiangsu
Province Funding for Enterprise-Academic-Research Innovation Platform
under Grant BY2012229, in part by the National Natural Science Foundation
of China under Grant 61401296, and in part by the Natural Science Foundation
for Youths of Jiangsu Province, China, under Grant BK20130375.
The authors are with the Department of Electrical and Computer
Engineering, National University of Singapore, Singapore 117583, and
also with the National University of Singapore Suzhou Research Institute,
Suzhou 215123, China (e-mail: eleguoyx@nus.edu.sg).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TMTT.2016.2555948
0018-9480 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 29,2023 at 13:42:38 UTC from IEEE Xplore. Restrictions apply.
parameters, which are very ambiguous to extract. Alternatively
a table-based model can be employed to represent the
nonlinear I–V and Q–V relationships. It requires no detailed
nonlinear function expressions, which significantly lightens the
load of extracting parameters [5]. However, it must be pointed
out that table-based models still have some limitations. As the
data within the box are discrete, some table-based models are
inaccurate for the high-order distortion simulation when input
signal magnitudes are small. This limitation originates from
the algorithm of interpolation and extrapolation when discrete
data are presented. Moreover, sometimes the convergence
cannot be guaranteed in a large power environment where high
nonlinearity is presented [7], [8]. In addition, the computation
speed in table-based models is very slow, as look-up tables
cover rather broad working ranges. Some researchers consider
physical-based models to be a good candidate to characterize
the devices. In common, physical-based models are very
accurate, especially compared to table-based or empirical
models when the condition is beyond measurement range [10].
Nevertheless, detailed fabrication knowledge is usually not
available to the model builder and its compatibility to popular
CAD software still needs to be further improved. In order
to better characterize the active devices, artificial neural
networks (ANNs) are increasingly recognized as a useful tool,
especially when the device-under-test (DUT) is of complex
nonlinearity and its internal working principle is not well
known to the model builder [11]–[14], [28], [32], [33].
However, it is still difficult for some existing ANN-based
small-signal models relying on traditional measurements to be
directly upgraded to large-signal models, as the consistency
between the small-signal model and the large-signal model
is not guaranteed [12], [32], [33]. Note that some good
ANN-based models [28] require nonlinear vector network
analyzer (NVNA), which may not be available to most
companies or research laboratories.
The dispersion effect can usually compromise the accuracy
of the large-signal model. The non-conservative data extracted
inside the intrinsic part violate the condition for integration.
Hence, many researchers employ pulse measurements to
characterize the thermal and trapping effects [23]–[25].
However, inconsistent S-parameter performance between the
small-signal model and the large-signal model on different
bias conditions still remains a problem for flexible active
circuit designs. The generated nonlinear model’s small-signal
responses under different bias conditions rather than one
specific bias point are crucial to circuit designers, which can
also be used to check the flexibility and practicability of
the model.
 LONG et al.: NOVEL 4-D ANN-BASED HYBRID LARGE-SIGNAL MODEL OF GaAs pHEMTs
Fig. 1. Equivalent representation of this proposed hybrid modeling method.
In this paper, a new ANN-based large-signal model is
presented. One advantage of the proposed model is that it only
requires traditional measurement setups including static dc
measurement and bias-dependent S-parameters measurements.
The process of the model is conducted in a linear environment.
The proposed multi-path-dependent integration technique not
only accurately describes the dispersion effect at microwave
frequencies ensuring the consistency between the small-signal
model and the large-signal model in all operation regions, but
also guarantees the consistency between dc and RF currents
of the generated model. The dispersion effects is mainly
due to thermal and trapping effects, which happen in
S-parameters on a switch-on condition. Therefore, rather
than pulse measurements, multi-bias S-parameters covering
all working ranges are well processed in this study to
characterize the trapping and thermal effect. The large-signal
model is consistently generated from the extracted small-signal
models under all measured bias points, which is flexible and
practical for active circuit designs. As shown in Fig. 1(a),
this proposed ANN-based hybrid model includes a distributed
extrinsic passive part and an intrinsic active part. The
extrinsic part includes all the field-effect transistor (FET)’s
metallic structures, while the intrinsic part is located under
the metallic fingers as an active semiconductor channel.
This passive/active hybrid modeling method can precisely
describe the coupling effect caused by metallic structures
such as stripes and grounded via-holes, especially in the
high-frequency domain. The data for model formulation
requires dc measurements and bias-dependent S-parameter
measurements. The detailed model building procedure will
be presented in Section II. In Section III, different types
of GaAs pHEMTs are investigated to verify the proposed
method. Moreover, in Section IV, a Ka-band power amplifier
and switch monolithic microwave integrated circuits (MMICs)
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 29,2023 at 13:42:38 UTC from IEEE Xplore. Restrictions apply.
1753
have been designed based on the proposed modeling method
for further verification. The conclusion will be given
in Section V.
II. L ARGE -S IGNAL M ODEL G ENERATION
Generally, the large-signal model of a GaAs pHEMT is
gradually built up based on its small-signal model under
multiple biases. Thus, the accuracy of the small-signal
model largely determines the success of the corresponding
large-signal model. The small-signal model is composed of the
extrinsic part and the intrinsic part. In general, the extrinsic
part cannot be separately and independently extracted. The
values of the extrinsic part are determined as a whole through
a special bias condition [16]. However, when the size of
devices becomes larger and the working frequency becomes
higher, the traditional extrinsic lumped components need to
be modified [6], [22], [26], [27]. In this paper, to describe
electromagnetic (EM) effects at high frequencies, the extrinsic
part of the presented model will be composed of lumped and
distributed components. The distributed part of the extrinsic
parts is rebuilt in EM simulation software exactly with their
own physical dimensions [22]. Hence, the parasitic effect and
material losses caused by the extrinsic region can be strictly
calculated by full-wave EM simulation. The EM simulated
results of the distributed extrinsic parts are saved as multi-port
scattering matrices to represent the distributed effects. The
rest of the lumped extrinsic components are extracted through
global optimization after de-embedding the inserted distributed
components. The initial values are based on the “cold-FET”
method [16]. As the values of intrinsic components are
usually bias dependent, the intrinsic part of the small-signal
model is extracted based on global optimization targeting
to fit S-parameters under different biases conditions. The
small-signal equivalent model is presented in Fig. 1(b), where
the intrinsic part of the model is indicated inside the
dashed box.
As mentioned above, the components of the extrinsic part
do not depend on the external voltage. Thus, due to the
existence of voltage drop across the resistors in the extrinsic
part, the dependent voltages of the intrinsic elements must
be recalculated. Consequently, the intrinsic voltages can be
modified from the extrinsic voltages as follows [29]:
Vgs = VGS − Igs (Rg + Rs ) − Ids Rs (1)
Vds = VDS − Ids (Rd + Rs ) − Igs Rs (2)
where Vgs and Vds are the recalculated intrinsic voltages
and VGS and VGD are the extrinsic voltages. Usually,
the devices are measured based on equidistant external
voltages for convenience, which means that the intrinsic
voltages Vgs and Vds are no longer equidistant because
of (1) and (2). Nevertheless, it is not easy to generate nonlinear
current and charge sources based on non-equidistant data.
At the same time, it is difficult to predict the values of
extrinsic resistors before measurement. Therefore, it is very
troublesome to set up suitable measurement grid individually
beforehand to obtain equidistant intrinsic voltages.
 1754 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 6, JUNE 2016

Fig. 2. Three-layer ANN for redistributing values of intrinsic components.

Traditionally, equidistant data can be generated by

interpolation. However, the main limitation of interpolation is Fig. 3. Dots represent the original Cgs with respect to intrinsic voltages.
that it may sometimes produce discontinuities. Moreover, the The surface represents the uniformly redistributed data by ANN.
commonly used cubic interpolation technique will even not
produce a valid result if the points are outside of the convex
hull. To overcome this limitation, ANNs will be employed to
uniformly redistribute the values of intrinsic part with respect
to the intrinsic voltages.
ANNs are constructed by simple cells operating in
parallel with different weights. These cells are stimulated
like biological nervous systems. Commonly, the collections
between these cells determine the function of the neural
networks. Generally speaking, the ANN can be regarded as
a very powerful continuous function, which can approximate
any nonlinear function after the collections, or weights, are
well adjusted. The weights along the cells, or neurons, can
be determined by suitable training algorithm and input–output Fig. 4. Dots represent the original G m with respect to intrinsic voltages.
data. One of the advantages is that ANNs are very smooth The surface represents the uniformly redistributed data by ANN.
as they have theoretically infinite order of derivatives.
A three-layer neural network is shown in Fig. 2 to demonstrate
the working principal of the ANN. The first layer is for input Cgs and G m are shown in Figs. 3 and 4, which demonstrates
variables. In this proposed pHEMTs modeling case, the input the consistency between redistributed data and original data.
can be different sets of intrinsic voltages. The second layer Based on the uniformly redistributed elements of the
contains a number of neurons with different weights waiting intrinsic part, the large-signal model can be constructed by
for training from measurement data. different sorts of nonlinear sources. Within the presented
The hyperbolic tangent (tansig) function is selected as model, Igs and Igd represent conduction currents. The charges
a transfer function to produce the nonlinearity within the are described by Q gs and Q gd . The non-quasi-static effect
network. As shown in (3), it is mathematically equivalent is described by Rgs and Rgd , which are located in series
to the tanh function, but has a fast running speed in the with charge sources [17]–[20]. The differential parameters in
implementation within MATLAB, the small-signal model can be accordingly integrated based
on (4)–(8) without considering the dispersion effect [15], [30].
2 The topology of the generated dispersionless large-signal
tansig(x) = − 1. (3)
1 + e−2x model is shown in Fig. 1,
The third layer is the output for post-processing, which Igs (Vgs , Vds ) = Igs (Vgs0 , Vds0 )
will produce sincere values of the intrinsic elements. It is  Vgs
worth noting that all the measurement data are preprocessed + G gs (v, Vds0 ) · dv (4)
with minimum and maximum mapping to the range from Vgs0

−1 to 1, which can accelerate the learning procedures of Vgs
proposed neural networks. During the training course, the Igd (Vgs , Vds ) = Igd (Vgs0 , Vds0 ) + G gd (v, Vds0 ) · dv
Vgs0
weights are being adjusted. Different sets of the raw intrinsic  Vds
voltages are put into the first layer of ANNs and the − G gd (Vgs , v) · dv (5)
related raw intrinsic element values are put into the last Vds0
 Vgs
layer of ANNs. After training is finished, targeted equidistant
voltages are put into the well-trained ANNs to uniformly Q gs (Vgs , Vds ) = Cgs (v, Vds0 ) · dv
Vgs0
redistribute the intrinsic elements. Finally, these intrinsic  Vds
elements are reversely mapped from −1 to 1 through the same + Cds (Vgs , v) · dv (6)
algorithm reversely adopted in preprocessing. The recalculated Vds0

Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 29,2023 at 13:42:38 UTC from IEEE Xplore. Restrictions apply.
 LONG et al.: NOVEL 4-D ANN-BASED HYBRID LARGE-SIGNAL MODEL OF GaAs pHEMTs
Fig. 5. Traditional method for integration can compromise the accuracy of
the model under the bias at (Vds2 , Vgs2 ) and (Vds3 , Vgs3 ) due to the loss of
partial-derivative information about gm .
 Vgs
Q gd (Vgs , Vds ) = Cgd (v, Vds0 ) · dv
Vgs0
 Vds
− (Cds (Vgs , v) + Cgd (Vgs , v)) · dv
Vds0
 Vgs
IRF (Vgs , Vds ) = IRF (Vgs0 , Vds0 ) + gm (v, Vds0 ) · dv

Vgs0
Vds
+ gds (Vgs , v) · dv.
Vds0
However, due to self-heating and trapping effects of GaAs
pHEMT devices, the existence of the dispersion effect can
compromise the consistency between the small-signal model
and the large-signal model, which can lead to degenerated
performance of the large-signal model. One of the problems
caused by the dispersion effect in the large-signal model is
the violation of integration path independence among intrinsic
elements, which means that the intrinsic elements, especially
gm and gds , do not satisfy the conservation condition. The
accuracy of the model will be compromised when improper
integration is enforced [5].
In this paper, to mathematically handle the dispersion effect,
the detailed mechanism of non-conservative data is worthy of
investigation. It is common to see an accurate small-signal
model, but the transformation to the large-signal model is
vulnerable to naive integration. If traditional integration is
improperly enforced, it is found that the accuracy of the
model will be badly compromised in one direction (it depends
on the details of the choices about the direction and the
path for integration), where the information of the partial
derivatives along this direction will be lost. Nevertheless, it is
interesting that the accuracy along the other specific integration
path can be guaranteed without considering the conservation
condition by executing loop integration. For example, if the
integration is executed as indicated in (8), the information of
partial derivative gm will be compromised if the position is
away from Vds0 . The high accuracy of gm can be ensured
in the whole region only when the partial differential data is
strictly conservative. On the other hand, the information of
gds is automatically reserved in the whole region regardless
that the data is conservative or not if the integration is
executed as indicated by (8). To further demonstrate this
phenomenon, Fig. 5 is shown to illustrate the inaccuracy
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 29,2023 at 13:42:38 UTC from IEEE Xplore. Restrictions apply.
1755
caused by naive integration in the process of traditional
multi-bias S-parameters measurements data [8], [30] when
the dispersion effect is not considered. Based on (8), the
value of the current source at bias point (Vds2 , Vgs2 ) follows
the integration path as AB D E from (Vds0 , Vgs0 ). Similarly,
the value at bias point (Vds3 , Vgs3 ) follows the integration
path as ABC. It can be found that only the information of
gds is strictly conserved at (Vds2 , Vgs2 ) and (Vds3 , Vgs3 ). The
information of gm on (Vds2 , Vgs2 ) and (Vds3 , Vgs3 ) is provided
by integration executed along AB, which is highly relying on
the assumption on that the device is not dispersive. However,
this is not true for most power devices, as the violation
of conservation can be easily observed when the traditional
integration is enforced.
To deal with this issue, a suitable integration path can
be treated as a good solution to the problem. However, the
model in [21] with only one specific integration path cannot
effectively support the large-signal simulation covering all
(7) working bias ranges. To make the solution more flexible and
practical in broad working conditions, variable integration
paths are introduced to fulfill the accuracy requirement.
In Fig. 5, it can be found that only the bias point (Vds1 , Vgs1 )
conserves both partial-derivative information, as both
(8)
directions are crossing point B when integration is executed.
Therefore, a three-dimensional nonlinear function can
properly handle the dispersion effect where the newly added
dimension is used to properly describe the voltage-dependent
integration path. The average port voltages are chosen as
variable integration paths in this paper. The following is the
partial derivatives described by multi-path integration:
∂ IRF (Vgs , Vds , Vgs0 )
gm = (9)
∂ Vgs
∂ IRF (Vgs , Vds , Vgs0 )
gds = (10)
∂ Vds
where Vgs0 is chosen as the average voltage to indicate
the different integration paths. In this way, the accuracy
of gm and gds can be ensured in different application
environments, which can be verified by broadband
S-parameters. It is needed to point out that the small-signal
response is very crucial to the practical success of the
large-signal model, which is sometimes ignored. In general,
the design of active circuits is associated with a selection
of suitable bias condition and good matching with passive
circuits. A good small-signal performance of a power
amplifier is the foundation for large-signal applications. Thus,
an accurate and practical large-signal model should perform
well under both small- and large-signal excitations.
In addition, another problem caused by the dispersion effect
is the inconsistency between the measured RF and static
currents. The RF current is obtained by the integration from
the differential parameters in the small-signal model. The
static current is directly obtained from the dc measurement.
An accurate dc model is important for active circuit designs
in bias adjustment. In this paper, to eliminate the discrepancy
between RF and static current, the average voltage is novelly
used to describe both RF and dc currents. As indicated by
(9) and (10), one of the average voltages has been chosen
 1756
to indicate integration paths to describe the dispersion effect
at RF frequencies. The other average voltage Vds0 can be
included to consistently characterize the dc current. To not
only guarantee the consistency between dc and RF currents,
but also reserve the information of partial derivatives, the
following are the conditions that Ids must satisfy:
∂ Ids (Vgs , Vds , Vgs0 , Vds0 )
gm (Vgs, Vds , Vgs0 , Vds0 ) =
∂ Vgs
∂ Ids (Vgs , Vds , Vgs0 , Vds0 )
gds (Vgs, Vds , Vgs0 , Vds0 ) =
∂ Vds
Ids (Vgs , Vds , Vgs0 = Vgs , Vds0 = Vds ) = Ids_DC
Ids (Vgs, Vds , Vgs0 , Vds0 ) = Ids_DC (Vgs0 , Vds0 )
 Vgs
+ gm (v, Vds0 ) · dv
Vgs0
 Vds
+ gds (Vgs , v) · dv.
Vds0
Ids_DC is the measured dc performance. The difference
between (9) and (11) is the introduction of Vds0 , which
plays an important role in characterizing the dc performance.
Equations (11) and (12) only contain the perturbation
information, which guarantees the accuracy of the partial
derivatives without considering the constant part in (8). At the
same time, (13) can provide the detailed information about
the constant term in the primitive function. To obtain Ids ,
it is required to integrate the differential parameters at
each bias point. The integration path is exactly indicated
by the specific bias point. During multi-bias S-parameter
measurements, constant or average bias points are equal to the
“switch-on bias” points. Each average bias point (Vgs0 , Vds0 ) is
associated with a traditional two-dimensional current source,
which is indicated by (13). Besides, to combine (8) and (13),
we can have (14), where two additional dimensions
Vgs0 and Vds0 are introduced to handle non-conservative
partial-derivative information and eliminate the inconsistency
between dc and RF currents. Thus, this newly introduced
four-dimensional nonlinear current sources can be understood
as a two-dimensional traditional current source mapped
into another two-dimensional average voltage bias condition
function.
In this way, consistency is achieved among the small-signal
model, the large-signal model, and dc performance. When
the FET is switched on, the channel current will cause
thermal effects. In various bias conditions, different channel
currents will cause different thermal effects while the
trapping effects will also exist in this switch-on condition
measurement. Therefore, the thermal and trapping phenomena
are captured and described through switch-on S-parameters.
Thus, it is meaningful to extract dispersion effects by
processing multi-bias S-parameter modeling, which is usually
ignored by the traditional methods [8], [30]. In [7], [8]
is compared with the work of [7] to demonstrate that the
ignorance of dispersion effect will cause the inaccuracy.
Fig. 6 is presented as the necessary steps of the model
construction.
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 29,2023 at 13:42:38 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 6, JUNE 2016
(11)
(12)
(13)
(14)
Fig. 6. Necessary steps of the model construction.
In large-signal simulations, the harmonic-balance simulation
mainly relies on numerical computations, whose linear part
is calculated in the frequency domain and the nonlinear part
is calculated in the time domain. Numbers of iterations are
executed to minimize the error function of the sum of the
current at all nodes based on Kirchoff’s current law (KCL).
When convergence is achieved, which means that the error
function is driven to a target small value, the resulting voltages
approximate a steady-state solution. However, sometimes
convergence cannot be guaranteed. One reason for the
divergence is the unsuitable topology of the large-signal
model. In [7] and [8], too many different orders of differential
equations are involved, which makes the computation
vulnerable to divergence when high nonlinearity is presented.
On the other hand, the equivalent circuit model in Fig. 1 can
semi-physically represent the devices, which is more robust
for convergence in numerical computation. Another cause of
divergence is the unreasonable values obtained through an
improper extrapolation algorithm. Considering that ANNs are
very smooth and capable of approximating any nonlinear
functions, it is reasonable to use the ANN to represent
the nonlinear multi-path-integration function. In addition, the
ANN-based model can provide more robust convergence
in dc, transient, and harmonic-balance simulations compared
to the table-based model [9]. The training data originates
from the current generated along each specific bias point
indicated by (8) and (13). Fig. 7 describes the topology of
the represented neural network. Compared with Fig. 2, it is
clear that two additional dimensions Vgs0 and Vds0 have been
integrated by the ANN in this proposed model to ensure the
accuracy of the small-signal performance.
Compared with traditional static current expression, it is
worth noting that the parameters Vgs0 and Vds0 are introduced
to address the dispersion effect. As mentioned above,
rather than performing traditional pulse measurements, these
two parameters are directly employed through the processing
 LONG et al.: NOVEL 4-D ANN-BASED HYBRID LARGE-SIGNAL MODEL OF GaAs pHEMTs
Fig. 7. Three-layer ANN representing the integration-path-dependent
nonlinear RF current source.
Fig. 8. (a) On-wafer measurement probe station setup and micrographs of:
(b) 4 × 75 μm, (c) 4 × 150 μm, and (d) 8 × 100 μm GaAs pHEMTs devices.
of multi-bias S-parameters. As directly originated from the
small-signal models, the generated large-signal model is very
practical for active circuit designs. The presented model can
be easily implemented in the Advanced System Design (ADS),
compatible to various sorts of simulations, such as dc, ac, and
harmonic-balance simulations.
III. L ARGE -S IGNAL M ODEL V ERIFICATION AND
E XPERIMENT R ESULTS
To verify the accuracy of the presented ANN-based hybrid
large-signal modeling approach, several types of on-wafer
GaAs pHEMTs (4 × 75 μm, 4 × 150 μm, 8 × 100 μm) are
investigated on the Cascade Summit 9000 probe station with
an Agilent 8510C vector network analyzer. The measurement
setup is presented in Fig. 8(a) and the micrographs of the
4 × 75 μm, 4 × 150 μm, 8 × 100 μm GaAs pHEMTs are
shown in Fig. 8(b)–(d), respectively. A variety of precise
measurements, which include the I–V curve, S-parameters,
Gain/power-added efficiency (PAE), Pin/Pout, third-order
intermodulation distortion (IMD3) testing, etc., have been
performed and compared with their corresponding modeled
performances to reveal the accuracy of this proposed model.
First of all, the on-wafer voltage-current measurement is
performed to get the static I–V characteristics. S-parameters
are then measured under different bias conditions. The
frequency range is from 2 to 40 GHz, which can be applicable
to Ka-band circuit designs afterwards.
To build the hybrid large-signal model of GaAs pHEMTs,
the components of their extrinsic part are extracted first. After
de-embedding the distributed components, the remainder of
the lumped components is extracted through optimizations,
where initial values are obtained from the “cold-FET” [16].
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 29,2023 at 13:42:38 UTC from IEEE Xplore. Restrictions apply.
1757
Fig. 9. Comparison of measured and simulated static I–V characteristics of
a 4 × 150 μm GaAs device. The simulations are generated from ANN after
training. The dots are measured results.
Fig. 10. Comparison of measured and simulated static I–V characteristics of
a 8 × 100 μm GaAs device. The simulation results are generated from ANN
after training. The dots are measured results.
The intrinsic part is obtained after de-embedding all the
extrinsic components under different working biases. The
intrinsic elements are then uniformly redistributed by ANNs.
The path-dependent integration at each bias point is then
executed to generate the nonlinear sources represented by
the ANN in the large-signal model. In the process of ANN
building, 60% data are used for training and the rest is for
validation. There are three layers in the ANN network and
ten neurons in the hidden layer. The Levenberg–Marquardt
backpropagation algorithm in MATLAB is selected for
optimization.
Figs. 9 and 10 show the dc performance of the 4 × 150 μm
and the 8 × 100 μm GaAs pHEMTs. The static
I–V characteristics are described by the well-trained
ANN with the introduction of average voltages. It can be seen
clearly in this figure that every measured I–V characteristic
point is precisely represented by the proposed hybrid
large-signal model and its overall agreement is very good.
It is well known that the accuracy of the small-signal
responses from the nonlinear model under different bias
conditions is the basis of the success of the large-signal model.
The performances of the small-signal responses under different
bias conditions of GaAs pHEMTs are shown in Figs. 11–14.
To further exhibit the advantage of this proposed model, the
commercial large-signal model EE-HEMT model [28] is used
to model the same 8 × 100 μm GaAs pHEMT for comparison
in Figs. 12–14. To show the flexibility of the presented model,
 1758 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 6, JUNE 2016
Fig. 11. Simulated (solid line) 4 × 150 μm GaAs pHEMT ANN model and
measured S-parameter (circles) under bias Vgs = −0.5 V and Vds = 5 V.
Fig. 12. Simulated (solid line) 8 × 100 μm GaAs pHEMT ANN model and
measured S-parameter (circles) under bias Vgs = −0.5 V and Vds = 5 V.
The dashed line is simulated by the EE-HEMT model for comparison.
Fig. 13. Simulated (solid line) 8 × 100 μm GaAs pHEMT ANN model and
measured S-parameter (circles) under bias V gs = −1 V and V ds = 3 V. The
dashed line is simulated by the EE-HEMT model for comparison.
Figs. 12 and 13 show the model’s performance in the
saturation while Fig. 14 demonstrates its performance in the
cutoff region. It is clear that the proposed model consistently
keeps its accuracy under different operation conditions while
the EE-HEMT model can only provide accurate S-parameter
modeling in the saturation region. The modeling performance
degenerates in linear and pinch-off regions. Therefore,
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 29,2023 at 13:42:38 UTC from IEEE Xplore. Restrictions apply.
Fig. 14. Simulated (solid line) 8 × 100 μm GaAs pHEMT ANN model and
measured S-parameter (circles) under bias Vgs = −2 V and Vds = 3 V. The
dash line is simulated by the EE-HEMT model for comparison.
Fig. 15. Comparison of simulated (solid line) and measured (circles)
gain/PAE performances of 8 × 100 μm GaAs pHEMT. The working frequency
is 10 GHz.
Fig. 16. Comparison of simulated (solid line) and measured (circles)
gain/PAE performances of 4 × 75 μm GaAs pHEMT. The working frequency
is 35 GHz.
as mentioned before, such inconsistent S-parameter modeling
performance of the traditional large-signal model makes it
unsuitable for the related applications under different operation
regions.
Figs. 15 and 16 show the gain and PAE versus the input
power of different transistors (8 × 100 μm, 4 × 75 μm)
at 10 and 35 GHz in a 50- environment. Furthermore,
a consistent agreement between measurement and modeled
responses at several harmonics can be found in Fig. 17,
 LONG et al.: NOVEL 4-D ANN-BASED HYBRID LARGE-SIGNAL MODEL OF GaAs pHEMTs
Fig. 17. Simulated (solid line) result from 8 × 100 μm GaAs pHEMT ANN
model and measured (circles) output power up to third order. The working
frequency is 10 GHz.
Fig. 18. Contours of normalized measured (circles) and simulated (solid line)
delivered power from 4 × 75 μm GaAs pHEMT. The working frequency
is 10 GHz. The input power is 10 dBm. The maximum power delivered
is 22 dBm. Contours are in 1-dB step.
where 10 GHz is chosen as the sample operating frequency
for testing. It is easy to estimate that the third-order intercept
point (IP3) of this 8 × 100 μm device is around 12 dBm from
the accurate simulated and measured results in Fig. 17. The
comparisons from Figs. 15–17 prove that both the stability and
the nonlinear modeling performances of the proposed model
are very satisfactory. In Fig. 18, the measured and modeled
contours of load–pull testing are compared to validate its
accurate large-signal modeling performance. The source is
fixed to 50  and the maximum delivered power is 22 dBm
with input power as 10 dBm at 10 GHz. As IMD3 is more
and more important in modern nonlinearity characterization,
Fig. 19 shows the simulated and measured IMD3 lower side
performance when the center frequency is 5 GHz and space
frequency is 1 MHz.
IV. R ELATED MMIC D ESIGNS AND V ERIFICATION
To further verify this presented approach, a Ka-band
power amplifier MMIC using 0.15-μm GaAs pHEMT
technology has been designed. With the help of the
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 29,2023 at 13:42:38 UTC from IEEE Xplore. Restrictions apply.
1759
Fig. 19. Simulated (red solid line) result from 8 × 100 μm GaAs pHEMT
ANN model and measured (blue circles) result of the lower side of IMD3.
The center frequency is 5 GHz. The space frequency is 1 MHz.
Fig. 20. Micrograph of a fabricated Ka-band MMIC power amplifier with
8 × 100 μm and 4 × 75 μm GaAs pHEMTs.
proposed ANN-based models, two kinds of GaAs pHEMTs
(4 × 75 μm and 8 × 100 μm) were adopted in this MMIC
design. Fig. 20 shows the actual layout photograph of the
fabricated MMIC chip with a size of 2.8 mm × 0.9 mm.
The left side is input and right side is output, respectively.
A three-stage topology has been adopted in this power
amplifier. The first two stages perform as the drive amplifier
stages and the last stage acts as the power amplifier stage.
At the input port, RF power is first amplified through
a 4 × 75 μm and an 8 × 100 μm GaAs pHEMT devices. Due
to the large drain current capability, four 4 × 75 μm GaAs
pHEMTs are adopted parallel in the third stage to meet
the output power requirement. Finally, the output power is
combined at the output port of this MMIC PA. This PA is
designed to operate in class-AB mode with a 28-dBm output
power and a small-signal gain of 18 dB (±1 dB flatness) from
32 to 37 GHz. Its detailed performances are presented and
discussed in the following sections.
Fig. 21 is its measured and simulated gain versus frequency,
where a good agreement can be found. Fig. 22 shows the
input return loss of this power amplifier while Fig. 23 is
its output power against input power at 35 GHz. All these
performances show a good consistency between the measured
and simulated results generated by the proposed ANN-based
models.
Furthermore, using the same 0.15-μm GaAs pHEMT
technology, a Ka-band single-pole double-throw (SPDT)
switch has been designed to validate the flexibility of the
 1760
Fig. 21. Simulated (solid line) and measured (circles) gain of the power
amplifier versus frequency.
Fig. 22. Simulated (solid line) and measured (circles) return loss of the
power amplifier.
Fig. 23. Simulated (solid line) and measured (circles) gain against input
power at 35 GHz of the power amplifier.
Fig. 24. Micrograph of a fabricated Ka-band MMIC SPDT with 4 ∗ 75 μm
GaAs pHEMTs.
proposed model. 4 × 75 μm GaAs pHEMT has been adopted
in this MMIC design. Fig. 24 is the layout of the fabricated
circuit, where port 1 is the switch input and ports 2 and 3 are
the switch outputs. Fig. 25 is the performance of the switch
compared with the simulation results from the commercial
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 29,2023 at 13:42:38 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 6, JUNE 2016
Fig. 25. Simulated (solid line) and measured (circles) performances of
the Ka-band switch. Dashed line is simulated by the EE-HEMT model for
comparison.
EE-HEMT model and proposed model. It is clear that the
proposed model can provide a much better simulation result
compared with the commercial one.
V. C ONCLUSION
In this paper, an ANN-based hybrid large-signal modeling
method for GaAs pHEMTs has been presented. The
extrinsic part of the presented model has been extracted
based on both lumped and distributed components, which
are more consistent to handle the EM effects in high
frequency. The values of bias-dependent intrinsic elements
have been uniformly redistributed by the ANN rather than
traditional interpolations, which is more effective considering
its smoothness and the powerful capability of nonlinear
approximation. The dispersion effect has been considered
by a multi-path integration technique implemented by a
four-dimensional ANN. The thermal and trapping effects
have been directly described by multi-bias S-parameters.
In this approach, the generated large-signal model is accurately
consistent with small-signal models. At the same time, the
discrepancy between RF and static currents has also been
eliminated by introducing average voltages, which considers
the dispersion effect as a whole. Different experiments results
such as dc, S-parameters, and different harmonics of output
power have been compared with simulated performance from
the proposed large-signal model and it is shown that the
accuracy is satisfactory. Different sizes of GaAs pHEMTs
have been investigated for verification. Finally, a Ka-band
power amplifier and a Ka-band switch have been designed with
the proposed model for further nonlinear verification, which
shows good agreement. Besides its accuracy, this novel hybrid
large-signal model can also be easily implemented in CAD
software and could be very useful GaAs pHEMT technology
based nonlinear microwave circuit design in the future.
R EFERENCES
[1] W. R. Curtice, “A MESFET model for use in the design of GaAs
integrated circuits,” IEEE Trans. Microw. Theory Techn., vol. MTT-28,
no. 5, pp. 448–455, May 1980.
[2] W. R. Curtice and M. Ettenberg, “A non-linear GaAs FET model for
use in the design of output circuits for power amplifier,” IEEE Trans.
Microw. Theory Techn., vol. MTT-33, no. 12, pp. 1383–1394, Dec. 1985.
 LONG et al.: NOVEL 4-D ANN-BASED HYBRID LARGE-SIGNAL MODEL OF GaAs pHEMTs
[3] I. Angelov, L. Bengtsson, and M. Garcia, “Extension of the Chalmer’s
HEMT and MESFET model,” IEEE Trans. Microw. Theory Techn.,
vol. 44, no. 10, pp. 1664–1674, Oct. 1996.
[4] Z. Zhong, Y.-X. Guo, and M. S. Leong, “A consistent charge model of
GaAs MESFETs for Ku-band power amplifiers,” IEEE Trans. Microw.
Theory Techn., vol. 59, no. 9, pp. 2246–2253, Sep. 2011.
[5] D. E. Root, S. Fan, and J. Myer, “Technology independent non
quasi static FET models by direct construction from automatically
characterized device data,” in 21st Eur. Microw. Conf., Stuttgart,
Germany, Sep. 1991, pp. 927–932.
[6] K. S. Yuk, G. R. Branner, and D. J. McQuate, “A wideband
multiharmonic empirical large signal model for high-power GaN
HEMTs with self-heating and charge-trapping effects,” IEEE Trans.
Microw. Theory Techn., vol. 57, no. 12, pp. 3322–3332, Dec. 2009.
[7] Y. Long, Y.-X. Guo, and Z. Zhong, “A 3-D table based method for
non-quasi-static microwave FET device modeling,” IEEE Trans. Microw.
Theory Techn., vol. 60, no. 10, pp. 3088–3095, Oct. 2012.
[8] S. M. Homayouni, D. Schreurs, G. Crupi, and B. Nauwelaers,
“Technology-independent non-quasi-static table based nonlinear model
generation,” IEEE Trans. Microw. Theory Techn., vol. 57, no. 12,
pp. 2845–2852, Dec. 2009.
[9] D. E. Root, “Future device modeling trends,” IEEE Microw. Mag.,
vol. 13, no. 7, pp. 45–59, Nov./Dec. 2012.
[10] S. Khandelwal and T. A. Fjeldly, “Analysis drain–current nonlinearity
using surface-potential-based model in GaAs pHEMT,” IEEE Trans.
Microw. Theory Techn., vol. 61, no. 9, pp. 3265–3270, Sep. 2013.
[11] H. Zaabab, Q. J. Zhang, and M. S. Nakhla, “A neural network modeling
approach to circuit optimization and statistical design,” IEEE Trans.
Microw. Theory Techn., vol. 43, no. 6, pp. 1349–1358, Jun. 1995.
[12] Q. J. Zhang, K. C. Gupta, and V. K. Devabhaktuni, “Artificial neural
networks for RF and microwave design-From theory to practice,” IEEE
Trans. Microw. Theory Techn., vol. 51, no. 4, pp. 1339–1350, Apr. 2003.
[13] Y. Cao, X. Chen, and G. Wang, “Dynamic behavioral modeling of
nonlinear microwave devices using real-time recurrent neural network,”
IEEE Trans. Microw. Theory Techn., vol. 56, no. 5, pp. 1020–1026,
May 2009.
[14] P. Roblin, D. E. Root, J. Verspecht, Y. Ko, and J. P. Teyssier, “New
trends for the nonlinear measurement and modeling of high-power RF
transistors and amplifiers with memory effects,” in 21st Eur. Microw.
Conf., Stuttgart, Germany, Sep. 1991, pp. 927–932.
[15] I. Schmale and G. Kompa, “An improved physics-based nonquasi-static
FET-model,” in 27th Eur. Microw. Conf., 1997, pp. 328–330.
[16] G. Dambrine, A. Cappy, F. Heliodore, and E. Playez, “A new method
for determining the FET small-signal equivalent circuit,” IEEE Trans.
Microw. Theory Techn., vol. 36, no. 7, pp. 1151–1157, Jul. 1988.
[17] A. T. Yang, C. H. Chan, J. T. Yao, R. R. Daniels, and J. P. Harrang,
“Modeling and simulation of high-frequency integrated circuits based
on scattering parameters,” in 28th ACM/IEEE Design Automat. Conf.,
San Francisco, CA, USA, Jun. 1991, pp. 752–757.
[18] R. R. Daniels, A. T. Yang, and J. P. Harrang, “A universal
large/small signal 3-terminal FET model using a nonquasi-static charge
based approach,” IEEE Trans. Electron Devices., vol. 40, no. 10,
pp. 1723–1729, Oct. 1993.
[19] E. S. Mengistu and G. Kompa, “A large signal model of GaN HEMTs
for linear high power amplifier design,” in 26th Eur. Microw. Conf.,
Oct. 1994, pp. 854–859.
[20] A. Orzati et al., “A 110-GHz large-signal lookup-table model for
InP HEMTs including impact ionization effects,” IEEE Trans. Microw.
Theory Techn., vol. 51, no. 2, pp. 468–474, Feb. 2003.
[21] V. Cuoco, M. P. v. d. Heijden, and L. C. N. de Vreede, “The ‘smoothie’
data base model for the correct modeling of non-linear distortion in FET
devices,” in IEEE MTT-S Int. Microw. Symp. Dig., vol. 3, pp. 2149–2152.
[22] B. L. Ooi, Z. Zhong, Y. Wang, and X. C. Shan, “A distributed
millimeter-wave small-signal HBT model based on electromagnetic
simulation,” IEEE Trans. Veh. Technol., vol. 57, no. 5, pp. 2667–2674,
Sep. 2008.
[23] A. Jarndal, A. Z. Markos, and G. Kompa, “Improved modeling of GaN
HEMTs on Si substrate for design of RF power amplifiers,” IEEE Trans.
Microw. Theory Techn., vol. 59, no. 3, pp. 644–651, Mar. 2011.
[24] G. Zucchelli, A. Santarelli, A. Raffo, G. V. Annini, and F. Filicori,
“Influence of dispersive effects on large-signal models based on
differential parameter integration,” in Eur. Gallium Arsenide and Other
Compound Semicond. Appl. Symp. Dig., 2003, pp. 349–352.
[25] A. Raffo et al., “Nonlinear dispersive modeling of electron devices
oriented to GaN power amplifier design,” IEEE Trans. Microw. Theory
Techn., vol. 58, no. 4, pp. 710–718, Apr. 2010.
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 29,2023 at 13:42:38 UTC from IEEE Xplore. Restrictions apply.
1761
[26] G. Crupi et al., “Accurate multibias equivalent-circuit extraction for
GaN HEMTs,” IEEE Trans. Microw. Theory Techn., vol. 54, no. 10,
pp. 3616–3622, Oct. 2006.
[27] G. Crupi, D. Schreurs, A. Raffo, A. Caddemi, and G. Vannini, “A
new millimeter-wave small-signal modeling approach for pHEMTs
accounting for the output conductance time delay,” IEEE Trans. Microw.
Theory Techn., vol. 56, no. 4, pp. 741–746, Apr. 2008.
[28] J. Xu, R. Jones, S. A. Harris, T. Nielsen, and D. E. Root, “Dynamic
FET model—DynaFET—for GaN transistors from NVNA active source
injection measurements,” in IEEE MTT-S Int. Microw. Symp. Dig.,
Tampa, FL, USA, Jun. 2014.
[29] A. Jarndal and G. Kompa, “Large-signal model for AlGaN HEMT
accurately predicts trapping and self-heating induced dispersion and
intermodulation distortion,” IEEE Trans. Microw. Theory Techn., vol. 54,
no. 11, pp. 2830–2836, Dec. 2007.
[30] G. Crupi et al., “Nonlinear modeling of GaAs pHEMTs for
millimeter-wave mixer design,” Solid State Electron., vol. 104, no. 2,
pp. 25–32, Feb. 2015.
[31] E. Sijercic and B. Pejcinovic, “Comparison of non-linear MESFET
models,” in 9th IEEE Int. Electron., Circuits, Syst. Conf., Dubrovnik,
Croatia, Sep. 2002, vol. III, pp. 1187–1190.
[32] Z. D. Marinkovic, O. R. Pronic, and V. V. Markovic, “Bias-dependent
model of microwave FET S-parameters based on prior knowledge
ANNs,” in 8th Neural Network Appl. Electr. Eng. Seminar, 2006,
pp. 185–188.
[33] H. Taher, D. Schreurs, and B. Nauwelaers, “Extraction of small signal
equivalent circuit model parameters for statistical modeling of HBT
using artificial neural,” in Eur. Gallium Arsenide and Other Semicond.
Appl. Symp., 2005, pp. 213–216.
Yunshen Long received the B.Eng. degree from the
University of Electronic Science and Technology
of China (UESTC), Chengdu, China, in 2010, and
the Ph.D. degree from the National University of
Singapore (NUS), Singapore, in 2014.
His research interests include microwave
semiconductor devices, sub-system modeling, and
monolithic microwave integrated circuit (MMIC)
design.
Zheng Zhong (S’08–M’12) received the B.Eng.
and M.E. degrees from the University of Science
and Technology of China (USTC), Hefei,
China, in 2003 and 2006, respectively, and the
Ph.D. degree in microwave engineering from the
National University of Singapore (NUS), Singapore,
in 2010.
Since 2010, he has been a Research Fellow
with the Department of Electrical and Computer
Engineering, NUS. His main research interests
include RF/microwave semiconductor device
modeling and characterization, microwave and millimeter-wave microwave
integrated circuit (MIC)/monolithic microwave integrated circuit (MMIC)
design, and RF energy harvesting.
 1762 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 6, JUNE 2016
Yong-Xin Guo (SM’05) received the B.Eng. and
M. Eng. degrees in electronic engineering from
the Nanjing University of Science and Technology,
Nanjing, China, in 1992 and 1995, respectively,
and the Ph.D. degree in electronic engineering from
the City University of Hong Kong, Hong Kong,
in 2001.
From September 2001 to January 2009,
he was with the Institute for Infocomm
Research, Singapore, as a Research Scientist.
In February 2009, he joined the Department
of Electrical and Computer Engineering, National University of
Singapore (NUS), as an Assistant Professor and, in January 2013,
became a tenured Associate Professor. He is currently the Director of the
Center for Microwave and Radio Frequency, Department of Electrical and
Computer Engineering, NUS. He is concurrently a Senior Investigator with
the National University of Singapore Suzhou Research Institute (NUSRI),
Suzhou, China, and the Director of the Center of Advanced Microelectronic
Devices, NUSRI. He has authored or coauthored 176 international journal
papers and 187 international conference papers. Thus far, his publications
have been cited more than 2086 times and his H-index is 28 (source:
Scopus). He holds sevcen granted/filed U.S. or Chinese patents. He has
graduated seven Ph.D. students at NUS. His current research interests include
monolithic microwave integrated circuit (MMIC) modeling and design,
RF energy harvesting and wireless power for biomedical applications and
Internet of Things (IoT), microstrip antennas for wireless communications,
implantable/wearable antennas, on-chip antennas, and antennas in package.
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 29,2023 at 13:42:38 UTC from IEEE Xplore. Restrictions apply.
Dr. Guo is the General Chair of the 2017 International Applied
Computational Electromagnetics Society (ACES) Symposium,
August 1–4, 2017, Suzhou, China. He was the General Chair of the
2015 IEEE Microwave Theory and Techniques Society (IEEE MTT-S)
International Microwave Workshop Series on Advanced Materials and
Processes for RF and THz Applications (IMWS-AMP 2015), Suzhou,
China, and the IEEE MTT-S International Microwave Workshop Series
2013 (IMWS2013) on “RF and Wireless Technologies for biomedical and
Healthcare Applications” in Singapore. He served as a Technical Program
Committee (TPC) Co-Chair for the IEEE International Symposium on Radio
Frequency Integration Technology (RFIT2009). He has been a TPC Member
and Session Chair for numerous conferences and workshops. He serves
as Associate Editor for IEEE A NTENNAS AND W IRELESS P ROPAGATION
L ETTERS , IET Microwaves, Antennas & Propagation, and Electronics
Letters. He was a recipient of the 2009 Young Investigator Award, National
University of Singapore, the 2013 Raj Mittra Travel Grant Senior Researcher
Award, and the Best Poster Award of the 2014 International Conference
on Wearable & Implantable Body Sensor Networks (BSN 2014), Zurich,
Switzerland. He was a corecipient of the Design Contest Award of the 20th
International Symposium on Low Power Electronics and design (ISLPED),
Rome, Italy, 2015. His Ph.D. students have been the recipients of Best
Student Paper Awards of the 2015 IEEE MTT-S IMWS-Bio, Taipei,
Taiwan, 2013 IEEE iWEM, Hong Kong, 2011 National Microwave and
Millimeter-Wave Conference, Qingdao, China, and the 2010 IEEE ICMMT,
Chengdu, China.