Neural Computing and Applications (2020) 32:7311–7320

https://doi.org/10.1007/s00521-019-04252-3 (0123456789().,-volV)(0123456789().
,- volV)

ORIGINAL ARTICLE

Crosstalk modeling in high-speed transmission lines by multilayer

perceptron neural networks
Kai Siang Ooi1 • Chun Lei Kong2 • Chan Hong Goay2 • Nur Syazreen Ahmad2 • Patrick Goh2

Received: 29 October 2018 / Accepted: 9 May 2019 / Published online: 17 May 2019
Ó Springer-Verlag London Ltd., part of Springer Nature 2019

Abstract
Signal degradation due to crosstalk-related issues has become increasingly important particularly in high-speed signal
transmissions. Conventional analysis of crosstalk requires a full electromagnetic modeling of the signal transmission path
along with a time-domain transient simulation which is computationally demanding. In this work, we apply a multilayer
perceptron neural network for crosstalk prediction in coupled transmission lines. The well-trained neural networks can be
used to predict the time-domain crosstalk directly, thereby replacing complex circuit simulations. Numerical results show a
high degree of generalization of the neural networks, which are able to produce accurate results and can be trained to
include effects such as reflections and input mismatches.

Keywords Crosstalk  Microstrip  Multilayer perceptron  Stripline

1 Introduction Cadence Allegro and Ansys High-Frequency Structure

Simulator (HFSS) that attempt to simulate the actual
Signal integrity in electronic system design has gained behavior of these devices based on theoretical mathemat-
traction as a topic of interest among designers and ical models. Various methods have also been studied and
researchers, especially in high-speed interconnection sys- established to date to predict crosstalk on different medi-
tems with data transmission rates in the gigahertz (GHz) ums, and these methods have been proven to provide
range. With the further trend toward device miniaturiza- results with high accuracies if the right parameters and
tion, high-frequency effects such as ringing, crosstalk, information were available and used while building up the
ground bounce and reflection noise play a crucial role in simulation model [3–6].
determining successful information transmission to meet Although the conventional ECAD and EM simulators
the designated low bit-error-rate requirements in modern are traditionally accurate, with very close correlations
systems [1]. when compared to measured results, these methods require
In many emerging system-in-package (SiP) and system- high amounts of computational resources due to the com-
on-chip (SoC) applications such as high-speed field-pro- plicated nature of the calculations, resulting in long com-
grammable gate array (FPGA) transceivers, one of the putational times for each simulation. In addition, the whole
major contributions to the loss in signal integrity is cross- simulation cycle has to be repeated with a change in any
talk [2]. Crosstalk prediction and modeling are conven- design parameter, and this tends to be time-consuming for
tionally done using electronic computer-aided design designs with a large number of dimensions. This limitation
(ECAD) or full electromagnetic (EM) simulators such as becomes more visible and can lead to simulation bottle-
necks, for example in Monte Carlo type simulations to
analyze the variations in the parameters and when solving
& Patrick Goh
eepatrick@usm.my optimization problems that have large numbers of changing
variables.
1
Keysight Technologies, 11900 Bayan Lepas, Penang, Machine learning methods have garnered an increasing
Malaysia amount of attention recently, especially in regression- and
2
School of Electrical and Electronic Engineering, Universiti classification-type problems. Methods based on support
Sains Malaysia, 14300 Nibong Tebal, Penang, Malaysia

123
7312 Neural Computing and Applications (2020) 32:7311–7320

vector machine (SVM) and Gaussian process regression Sect. 4, and finally, a brief conclusion is some future work
(GPR) have been successfully applied to solve difficult is proffered in Sect. 6.
engineering problems [7, 8]. However, these methods are
often nonparametric and scale poorly with dimension. In
addition, the selection of a proper kernel is not easy and 2 Crosstalk in transmission lines
often requires a deeper understanding of the problem.
Artificial neural networks (ANNs) have shown promising Figure 1 illustrates the lumped equivalent circuit of a
results as a fast and accurate solution when applied in segment of a coupled transmission line pair where Lm and
crosstalk prediction and modeling of interconnects. For Ls represent the mutual and self-inductances per unit
example, in [9], an ANN trained with the Levenberg– length, and Cm and Cs represent the mutual and self-ca-
Marquardt algorithm was used to model the frequency-de- pacitances per unit length of the parallel line. The current
pendent crosstalk in dB of non-uniform cable bundles using and voltages in the two lines are related by (1), (2) and (3)
data from lumped-element models generated using the using the inductance and capacitance matrices of the
random displacement spline interpolation method. From the respective lines in a time-domain analysis [6].
study, it was observed that the difference between the neural      
d v1 Ls L m d i 1
network predicted outputs and the lumped-element simu- ¼ ð1Þ
dz v2 Lm Ls dt i2
lation for varying frequency and height above the ground      
plane is less than 1–2dB. However, due to the reliance on d i1 Ct Cm d v1
¼ ð2Þ
lumped-element equivalence, this result is only valid at low dz i2 Cm Ct dt v2
frequencies when the circuit is electrically small. In [10], a
similar concept was applied to model crosstalk in multi- Ct ¼ Cs þ Cm : ð3Þ
conductor transmission lines by using neural networks to In the equations, v1 and v2 represent the voltages of line
learn the relationship between the cross-sectional geometry 1 and line 2, while i1 and i2 represent the current of line 1
and the per-unit-length parameters. Results show a good and line 2. z is the longitudinal coordinate. These equations
level of accuracy between the ANN predictions and the form coupled differential equations which govern the
finite element method (FEM) for both the capacitance and propagation of signal on the transmission line pair. From
inductance matrices. In [11], ANN was also applied to (1) and (2), it is observed that a change in current on the
model the relation between the geometrical and material active line (line 1) will cause an inductive-coupled voltage
characteristics of interconnects to their respective Simula- drop and a change in voltage on the active line (line 1) will
tion Program with Integrated Circuit Emphasis (SPICE) cause a capacitive-coupled current change on the quiet line
netlist values that were used to analyze the parasitic (line 2).
capacitances. Here, the computational expenses in terms of To observe the effects of crosstalk coupling on the quiet
the simulation time were compared, and ANN was shown to line (line 2) from the signals in the active line (line 1), the
be able to provide up to 88% in computational savings. setup depicted in Fig. 2 is used where the two transmission
However, all of the prior work focuses on extracting lines form a pair of coupled lines. Vin ðtÞ is the input signal
either a frequency-domain model or an equivalent circuit on the active line, while Vout ðtÞ is the output signal on the
model of the physical layout of the interconnects. In active line. Z0 represents the characteristic impedance of
modern high-speed digital designs, analyses are predomi- the transmission lines and l represents the length of the
nantly done in the time domain and it is often desirable to lines. Here, the ends of the parallel pair are assumed to be
have fast and accurate predictions of the actual voltage terminated or matched to Z0 . VNEXT ðtÞ and VFEXT ðtÞ
versus time relationships of the crosstalk to estimate its
eventual effect on the eye diagram of the victim lines. In
this work, we apply ANNs based on the multilayer per- v1 i1 Ls
ceptron (MLP) structure to model time-domain crosstalk
voltages in coupled transmission lines. The method is Cs
general in nature and can be trained to include the effects Lm
i2 Cm Ls
of terminations such as reflections and input mismatches. v2
The remainder of the paper is organized as follows:
Cs
Section 2 presents a brief review of the fundamentals of
crosstalk in transmission lines and its effect in the time
Δz
domain. Section 3 presents the neural network approach to
model the time-domain crosstalk and the underlying Fig. 1 Lumped equivalent circuit of a coupled parallel transmission
training algorithms. Numerical results are presented in line pair

123
Neural Computing and Applications (2020) 32:7311–7320 7313

Vin (t) is governed by the strengths of the connections from the

Z0 Z0 response to the input. An ANN is formed by a number of
+ processing nodes, which are analogous to neurons in the
Z0 Z0 Vout (t)
Vsource (t) – human brain, that are connected by weighted connections,
+ l + which are analogous to synapses in the brain. The output of
Z0 VNEXT (t) Z0 VFEXT (t) each processing node only depends directly on the infor-
– – mation available locally at the node. However, this infor-
mation may originate from external factors or aggregation
Fig. 2 Setup for crosstalk measurement
of inputs from other neurons connected via weighted
represent the near-end and far-end crosstalk, respectively, functions through different layers of the neural network.
that are present in the parallel transmission line pairs. Complex problems or functions can be modeled by con-
Figure 3 shows an example plot of the voltages at dif- necting multiple processing nodes with appropriate weights.
ferent ends of the line. In this case, the effect of the active One popular ANN structure that is widely used and
line on the quiet line during the rising and falling edges of known is the multilayer perceptron (MLP) neural network
the signal can clearly be seen in the form of the induced illustrated in Fig. 4. In this structure, the neurons are sep-
near- and far-end crosstalk. Returning to (1) and (2), it can arated into L layers. The first layer, 1, and the last layer, L,
be observed that VNEXT and VFEXT are dependent on the are the input and output layers, while a number of layers
mutual and self-capacitances and inductances. This trans- known as hidden layers are sandwiched in between. The
lates to a dependence on the physical and geometrical hidden layers are denoted from layers 2 to L  1. nl rep-
parameters. Various studies in the past have been con- resents the number of neurons in the lth layer where
ducted to understand the impact of the changes in the l ¼ 1; 2; :::; L. xn represents the nth input, while ym repre-
physical/geometrical parameters to the crosstalk behavior sents the mth final output where n and m are real numbers.
of the parallel transmission line, and the readers are The output (or activation in neural network terms) of the ith
referred to references [3–6, 12, 13] for a more rigorous neuron in the lth layer is denoted by ali . The weight of the
analysis on the fundamentals of crosstalk. link between the jth neuron of the l  1th layer and the ith
neuron of the lth layer is represented by wlij where i ¼
1; 2; :::; nl while j ¼ 0; 1; :::; nl1 .
3 The neural network approach In an MLP network, the output of each neuron is gen-
erated by processing the inputs from each of the neurons
ANNs have shown promising results in modeling and connected to it via an activation function. This is per-
solving various problems that are arbitrarily nonlinear in formed in two steps. First, the product of the multiplication
nature from RF and microwave design applications of each input from the neurons of the ðl  1Þth layer and
[14–16], to electronic fault diagnosis [17, 18] and to elec- their corresponding weight parameter are summed to pro-
tromagnetic structure modeling and behavioral predictions duce a weighted sum, cli given by
[19–22]. ANNs are mapping systems that have their struc- X
nl1
ture loosely defined based on the biological neural system cli ¼ wlij al1
j þ bli ð4Þ
of the brain. The neurons interact with one another by j¼1

collecting signals from other neurons in a limited area and

where the parameter bli is an additional bias parameter for
generate a response signal which distributes these responses
each neuron. The weighted sum cli is then used in the
via contacts with dendritic trees of other neurons. The
response is related to its inputs via a nonlinear function that
w11l a 1l a 1l w11L
x1 y1
3 V in a 2l a 2l
x2
voltage, V

V out
2 ym
V N EXT
1 V F EXT Input a3l a 3l Output
L
wm n l
0 xn

-1 w nl nl anll anll
0 1 2 3 4 5 6 7 8 9 10
time, ns Hidden layer

Fig. 3 Near- and far-end crosstalk in a coupled transmission line pair Fig. 4 MLP neural network structure

123
7314 Neural Computing and Applications (2020) 32:7311–7320

second step along with an activation function, rðÞ to where the gradient rf ðwÞ and Hessian H are given by
generate the final output of the selected neuron at the lth 2 3
of
layer, given by 6 ow1 7
  6 7
ali ¼ r cli : ð5Þ 6 of 7
6 7
6 7
Consequently, this generated output is then taken as the rf ðwÞ ¼6 2 7
6 ow
7 ð8Þ
6 .. 7
stimulus or input for the neurons in the ðl þ 1Þth layer. The 6 . 7
6 7
activation function’s purpose is to decide whether a neuron 4 of 5
should be activated or not. Good activation functions often own
have certain desirable properties such as being smooth, 2 2 3
bounded, monotonic and continuously differentiable. Some o f o2 f o2 f
6 ow2 
applicable functions are the sigmoid, arc-tangent (tan1 x) 6 1 ow1 ow2 ow1 own 7
7
6 7
and hyperbolic-tangent (tanh x) function given by 6 o2 f o2 f 2
o f 7
6  7
6 ow22 ow2 own 7
2 H ¼6 ow2 ow1 7: ð9Þ
rðxÞ ¼  1: ð6Þ 6 7
1 þ e2x 6 .. .. .. .. 7
6 . . . . 7
6 7
For the input layer, the neurons utilize a relay activation 4 o2 f o2 f o f 5
2

function to simply relay the external stimuli to the input 

own ow1 own ow2 ow2n
layer neurons where a1i ¼ xi , i ¼ 1; 2; :::; n.
A three-layer MLP neural network is considered a uni- However, the computation of the Hessian matrix and its
versal approximator, in that it is capable of approximating inverse is computationally expensive as it requires the
any continuous function as given by the universal computation of the second-order derivatives. To alleviate
approximation theorem [23, 24]. However, before an ANN this, a class of methods which attempts to approximate the
can be applied to model or solve complex problems, it has inverse Hessian matrix instead of using a full computation
to first be trained to do so. During the training, the weights has been proposed, which are collectively known as quasi-
and biases of each neuron are updated iteratively to mini- Newton algorithms. One of the most popular methods to do
mize the errors between the targeted output and the output this is by using the Broyden–Fletcher–Goldfarb–Shanno
from the neural networks. To date, there are various (BFGS) formula which approximates the inverse Hessian
algorithms developed to train an ANN [25], most of which through a one-dimensional line search using the gradient
are based on the backpropagation algorithm where the information. The update rule is given as [26]:
errors are propagated back from the final layer to the first  T !  T !T
 1 kþ1 dk c k  1 k dk c k
and iterative steps are taken to minimize it through the use H ¼ I   T H I   T
of a gradient descent-like algorithm. Two backpropagation dk c k dk c k
training algorithms, namely the quasi-Newton and the  T
dk dk
Levenberg–Marquardt algorithms, are discussed briefly þ  T
here. dk c k
ð10Þ
3.1 Quasi-Newton training algorithm
where
The quasi-Newton algorithm is a root finding method that dk ¼ wkþ1  wk ð11Þ
can be used in optimization tasks for finding local maxima    
or minima of functions. The method is based on the ck ¼ rf wkþ1  rf wk : ð12Þ
Newton method that assumes that the error surface of a
convex function near a minimum is quadratic in nature.
3.2 Levenberg–Marquardt training algorithm
Therefore, the minimization or optimization can be
achieved in one step by taking the second-order partial
Another well-known training algorithm to train ANNs is
derivatives of the error surface with respect to the weights.
the Levenberg–Marquardt algorithm [27], which is often
The second-order derivatives are compiled in matrix form
viewed as an interpolation between the Newton and the
to form the Hessian matrix. A batch update of weights is
gradient descent methods. In the Levenberg–Marquardt
performed at the end of each epoch as illustrated in (7)
algorithm, the Hessian matrix, H, is approximated as a
below
product of a Jacobian matrix, J and its transpose
Dw ¼ H 1 rf ðwÞ ð7Þ

123
Neural Computing and Applications (2020) 32:7311–7320 7315

H  JT J ð13Þ W
S
S H T
where the Jacobian is given by W W W εr
2 3
of1 of1 of1 L
6 ow1 ow2    own 7
6 7 Top view Side view
6 of 7
6 2 of2    of2 7
6 7 Fig. 5 Geometrical structure of a coupled stripline
J¼6 ow ow2
6 1
own 7:
7 ð14Þ
6 .. .. .. .. 7
6 . . . . 7 greatest effect on the crosstalk. A full-factorial 3K DoE
6 7
4 ofN ofN ofN 5
... with four parameters is performed using 34 ¼ 81 analyses
ow1 ow2 own covering the entire design space. From the result of the
The gradient, rf ðwÞ is then calculated using the trans- DoE, a total of 168 training samples are then generated by
pose of the Jacobian matrix and the vector containing the focusing on the main parameters that affect the crosstalk
network errors, e which are the spacing between conductors, S, the substrate
height, H, conductor width, W and conductor thickness,
rf ðwÞ ¼ J T e: ð15Þ T. The remaining parameters including the length of the
The formula for updating the weight values in the neural line, L, dielectric constant, r , conductor conductivity, r
network used in this training algorithm is then given by and dielectric loss tangent, d, are fixed constant. The final
 1 design parameters are summarized in Table 1.
Dw ¼  J T J þ lI J T e ð16Þ Keysight Momentum simulator is used to simulate the
EM response of the circuit in the form of the 4-port S-
where I is the identity matrix and l is a control parameter.
parameters. An example of S-parameter plot for one of the
It can be observed that if the control parameter is set to 0,
designs is shown in Fig. 6. Note that since the circuit is D2 -
then it becomes the same equation used in the Newton
symmetric, S11 ¼ S22 ¼ S33 ¼ S44 , etc. and since it is also
method for weight updates. It is to be noted that the
reciprocal, S12 ¼ S21 ¼ S34 ¼ S43 , etc. The S-parameters
computation of the Jacobian matrix is done through a
generated from the EM simulations were then used in a
standard backpropagation technique that is less compli-
transient analysis to simulate the crosstalk as shown in
cated compared to computing the Hessian matrix.
Fig. 7. A 1.0-V step voltage with a rise time of 1 ns, Vsource
was supplied at the source and all the ports were terminated
to match the characteristic impedance of the line. Since the
4 Numerical examples
waves in the stripline propagate in a homogeneous med-
ium, the far-end crosstalk is zero and only the near-end
In this section, two numerical examples are presented, each
crosstalk is obtained.
with a varying degree of complexity. The first example
A multilayer perceptron neural network is then used to
presents a crosstalk prediction using ANN for a coupled
learn to relation between the physical parameters and the
stripline in a matched terminated system. Due to the
near-end crosstalk from the generated training data. The
homogeneous nature of the wave propagation in a stripline,
only the near-end crosstalk is present. In the second
example, ANNs are used to predict the near- and far-end
Table 1 Summary of design parameters for the stripline
crosstalk in a coupled microstrip transmission line, which
has a non-homogeneous medium. In addition, the effects of Design parameters Values
reflections and impedance mismatches are introduced in Substrate height, H 20 \H\ 60 mils
the modeling by applying a constant termination of 50 X at Conductor thickness, T 0.7 \T\ 1.1 mils
the terminals.
Ratio of conductor width, W to H 0:5 \W=H\ 1:5
Ratio of conductor spacing, S to H 0:275 \S=H\ 3:525
4.1 Stripline
Ratio of conductor spacing, S to W 0:55 \S=W\ 2:35
Length, L 600 mils
In this example, the ANN is used to predict the crosstalk in
Dielectric relative permittivity, r 4.6
a pair of coupled stripline transmission lines. The structure
Dielectric relative permeability, lr 1
of the stripline is shown in Fig. 5.
Dielectric loss tangent, tan d 0.01
First, a design of experiments (DoE) method is used to
Conductor conductivity, r 5:8  107 S/m
generate a reduced set of training data by identifying the
various design parameters in the stripline which have the

123
7316 Neural Computing and Applications (2020) 32:7311–7320

models was evaluated based on the mean squared error

(mse) of the training set calculated as follows
P No P Ni  2
i¼1 j¼1 tij  yij
mse ¼ P No ð17Þ
i¼1 Ni

where No is the number of output neurons, Ni is the number

of samples for the ith output, tij is the target for the ith
output and jth sample and yij is the output from the neural
model for the ith output and jth sample. Note that in this
example, No ¼ 1. The best-performing neural model,
which is the neural model trained using the Levenberg–
Marquardt algorithm with 9 hidden neurons, is selected as
the final design.
Figures 8, 9 and 10 show the training performance
comparison of the neural network target from Keysight
Momentum and the output from the ANN simulation for
the near-end crosstalk with varying substrate height, con-
Fig. 6 Simulated magnitude and phase plots of the S-parameters of a
ductor thickness and conductor width, respectively, versus
stripline with H ¼ 40 mil, T ¼ 0:7 mil, W ¼ 20 mil and S ¼ 17 mil the spacing between conductors. It can be seen that the
in Keysight Momentum EM simulator output from the ANN is able to model the varying crosstalk
accurately.
Z Vin Vout In order to have an unbiased evaluation on the perfor-
Stripline
mance of the neural network, a set of 25 randomly gen-
S-parameters Z erated input data that was not part of the training set was
Vsource VNEXT VFEXT
used to test the neural network. The result is shown in
Z Z Fig. 11 where a good correlation can be observed. The
neural network’s performance on the testing set was cal-
Fig. 7 Schematic diagram of a transient simulation to simulate the culated using the coefficient of correlation, Ri for each
near- and far-end crosstalk output i, where
P Ni   
j¼1 tij  ti yij  yi
Ri ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P Ni  2 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P Ni  2ffi ð18Þ
inputs of the neural network are the four design parameters,
j¼1 ij t  t i j¼1 ij y  y i
S, H, W and T, while the output is the near-end crosstalk
voltage. A single hidden layer is used. The MLP weights where ti and yi are the mean values for the ith target and
were initialized with small random values, and the inputs neural model outputs, respectively, defined as
were normalized to a range of ½1 1. The ANN was
trained using five different training algorithms while 1X Ni
ti ¼ tij ð19Þ
varying the number of hidden neurons and the results are Ni j¼1
summarized in Table 2. The performance of the neural

Table 2 Comparison of the neural network performance using different training algorithms
Training algorithms
BFGS quasi- Levenberg– Scaled conjugate Powell/Beale conjugate Resilient
Newton Marquardt gradient gradient backpropagation

mse of training set 0.0016 6:85  106 0.0037 0.0019 0.0068

Optimal number of hidden 7 9 5 3 6
neurons
Training time (s) 0.125 0.281 0.068 0.086 0.111

123
Neural Computing and Applications (2020) 32:7311–7320 7317

9 4.2 Microstrip

near-end crosstalk, mV
7 Expected result
ANN result
5
In this example, the ANN is used to predict the crosstalk in
H = 60 mil
a pair of coupled microstrip transmission lines. The
3 H = 30 mil
structure of the microstrip line is shown in Fig. 12. A
1 similar simulation flow is used as in Sect. 4.1 to generate
H = 20 mil
-1 the training data. A total of six physical parameters are
10 20 30 40 50 60 70 80
varied, and Table 3 shows the range of parameters that are
spacing between conductor, S, mil
investigated in this example.
Fig. 8 Training result of the ANN for crosstalk in striplines with Since the fields propagate in a non-homogeneous med-
varying H and S ium for a microstrip line, it exhibits both a near-end and
far-end crosstalk. In addition, to emulate the modeling of a
10
near-end crosstalk, mV

T = 1.1 mil real system, the terminations are fixed at 50 X for all ends
8 Expected result of the transmission lines. This results in a more complex
ANN result
6 modeling as the effect of voltage division at the input and
4
T = 0.7 mil reflections due to impedance mismatches is also modeled
along with the couplings to give the final measured
2
crosstalk. The effect on the near- and far-end crosstalk is
0 illustrated in Figs. 13 and 14, respectively, for the case of
5 10 15 20 25 30 35 40
spacing between conductor, S, mil varying line spacing. In addition, the effect of line lengths
is also modeled which has the overall effect of lengthening
Fig. 9 Training result of the ANN for crosstalk in striplines with the duration of the crosstalk as seen in Fig. 15 for the near-
varying T and S
end crosstalk. A similar effect is observed for the far-end
crosstalk as shown in Fig. 16.
9
near-end crosstalk, mV

W = 15 mil The source signal in this example is a voltage pulse with

7 Expected result
ANN result
an amplitude of 5V, rise and fall times of 0.5ns, and a pulse
5 width of 3ns. ANN was used to model the largest positive
W = 30 mil
3 near- and far-end crosstalk and their time of occurrences
W = 45 mil
1 due to the rising and falling edges of the signal. An MLP
with six inputs and four outputs, and a single hidden layer
-1
0 20 40 60 80 100 120 with varying number of hidden neurons is trained using the
spacing between conductor, S, mil Levenberg–Marquardt algorithm to model the system. The
best performance is recorded with 65 neurons in the hidden
Fig. 10 Training result of the ANN for crosstalk in striplines with
layer which converged with an mse of 0.060 after 12
varying W and S
epochs.
To perform an unbiased evaluation on the neural net-
7
near-end crosstalk, mV

work, a set of 50 randomly generated data, which are not

Expected result
5 ANN result
part of the training set, is used to test the system. The
comparison between the outputs from the neural model and
3
the expected outputs is shown in Figs. 17 and 18 for the
1 near-end crosstalk and in Figs. 19 and 20 for the far-end
crosstalk. The neural network’s performance on the testing
-1
0 5 10 15 20 25 30 set in terms of the coefficient of correlation, R, is calculated
test case using (18) for each output, and the results are tabulated in

Fig. 11 Testing result of the ANN for crosstalk in striplines

W air
1X Ni
S
S T
yi ¼ yij : ð20Þ
Ni j¼1 W H W W εr

The neural network recorded a performance of R ¼ L

0:9975 which shows that the trained neural model pos- Top view Side view
sesses a strong generalization capability.
Fig. 12 Geometrical structure of a coupled microstrip line

123
7318 Neural Computing and Applications (2020) 32:7311–7320

Table 3 Summary of design parameters for the microstrip 0.6

far-end crosstalk, V
0.4
Design parameters Values
increasing
0.2
length
Substrate height, H 15 \H\ 105 mils 0
Conductor thickness, T 0.3 \T\ 2.4 mils
-0.2
Conductor width, W 25 \W\ 325 mils
-0.4
Conductor spacing, S 25 \S\ 350 mils
-0.6
Length, L 2000 \L\ 11000 mils 0 1 2 3 4 5 6 7 8 9 10
Dielectric relative permittivity, r 2.2, 4.6, 9.6 time, ns
Dielectric relative permeability, lr 1
Fig. 16 Transient plots showing the impact of the parallel microstrip
Dielectric loss tangent, tan d 0.01
trace lengths on the far-end crosstalk
Conductor conductivity, r 5:8  107 S/m

0.4 Expected result ANN result

near-end crosstalk, V
0.3
0.6
near-end crosstalk, V

increasing spacing 0.2

0.4
0.2 0.1
0
0
-0.2
-0.4
0 5 10 15 20 25 30 35 40 45 50
test case
-0.6
0 1 2 3 4 5 6 7 8 9 10
Fig. 17 Comparison of maximum near-end crosstalk voltage between
time, ns the ANN model outputs and Keysight Momentum simulation results
for the test data
Fig. 13 Transient plots showing the impact of the parallel microstrip
trace spacing on the near-end crosstalk
near-end crosstalk time, ns

10 Expected result ANN result

0.6 8
far-end crosstalk, V

0.4 increasing 6
spacing
0.2
4
0
-0.2 2
-0.4 0
0 5 10 15 20 25 30 35 40 45 50
-0.6
0 1 2 3 4 5 6 7 8 9 10 test case
time, ns
Fig. 18 Comparison of maximum near-end crosstalk occurrence time
between the ANN model outputs and Keysight Momentum simulation
Fig. 14 Transient plots showing the impact of the parallel microstrip
results for the test data
trace spacing on the far-end crosstalk

0.3 Expected result ANN result

near-end crosstalk, V

0.6
far-end crosstalk, V

0.2
increasing
0.1 length 0.4
0
0.2
-0.1
-0.2 0
-0.3
0 1 2 3 4 5 6 7 8 9 10 0 5 10 15 20 25 30 35 40 45 50
time, ns test case

Fig. 15 Transient plots showing the impact of the parallel microstrip Fig. 19 Comparison of maximum far-end crosstalk voltage between
trace lengths on the near-end crosstalk the ANN model outputs and Keysight Momentum simulation results
for the test data

123
Neural Computing and Applications (2020) 32:7311–7320 7319

8 choice of activation functions can also affect the perfor-

far-end crosstalk time, ns

Expected result ANN result
mance of the neural networks. In this work, the hyperbolic-
6 tangent (tanh) activation function was used as it had the
best performance across all networks tested. As a com-
4
parison, for the Levenberg–Marquardt trained neural net-
2 work with nine hidden neurons used in Sect. 4.1, tanh
recorded a testing performance of R ¼ 0:9975, while the
0 sigmoid, rectified linear unit (ReLU) and pure linear
0 5 10 15 20 25 30 35 40 45 50
test case
(purelin) activation functions recorded testing perfor-
mances of R ¼ 0:9962, 0.9943 and 0.5021, respectively.
Fig. 20 Comparison of maximum far-end crosstalk occurrence time This is consistent with other published works, for example,
between the ANN model outputs and Keysight Momentum simulation in [7] where tanh often outperforms other activation
results for the test data
functions for neural networks with a single hidden layer,
while ReLU is often used in deep neural networks to
Table 4 Coefficient of correlation, R, for the near- and far-end overcome the vanishing gradient problem.
crosstalk voltage and occurrence time of the ANN versus Keysight
momentum An interesting extension to this work would predict the
transient response of the crosstalk voltages and not just the
Parameter R
maximum values and occurrence times. However, regular
Near-end crosstalk voltage 0.9424 feedforward ANNs might not be as effective for this.
Near-end crosstalk occurrence time 0.9330 Recent progress has been made in using long short-term
Far-end crosstalk voltage 0.9524 memory (LSTM) neural networks, which form a type of
Far-end crosstalk occurrence time 0.8896 recurrent neural network (RNN) architecture in deep
learning to predict the transient responses of high-speed
channels [29]. However, the method still requires training
data from a circuit simulator over a certain number of time
Table 4. It can be seen that the neural model recorded a steps and the trained neural network cannot be extended to
good performance of between 95.2 and 89.0% despite the a different design. In addition, certain cases such as long,
additional introduced complexity of this six-variable low-loss interconnects might require very long memory
system. lengths and training data in order to capture all the reflected
transients in the line.

5 Discussions
6 Conclusion
This section presents some further discussions on the
developed method. In this work, the outputs of the hidden The advancements in machine learning methods have
neurons are arbitrary and do not represent any physical opened up a wide array of opportunities for signal integ-
quantity. This is typical in scenarios where the neural rity-based analysis. In this paper, we present an application
networks are used as a blackbox input–output representa- of the multilayer perceptron neural network for crosstalk
tion of the system. However, in certain cases, there have modeling in coupled transmission lines. A design of
been progresses in work done to have the hidden neurons experiments method is used to generate a reduced training
represent physically meaningful quantities, such as the set by identifying the main design parameters which have
work in [28] where separate pole and residue learning the greatest impact on the crosstalk. Different training
blocks are incorporated in the neural networks. However, algorithms are implemented and compared, and the best-
this results in added complexity to the final network and is performing neural network is selected for each design.
more appropriate in cases where the neural networks are Numerical examples are presented which show that the
learning from the frequency-domain transfer functions such neural networks are able to model the crosstalk accurately
as the S-parameters. In this work, the networks are trained even in the presence of reflections and input mismatches.
directly from the design parameters, and this allows the The well-trained network can then be used to predict the
final trained networks to bypass the frequency-domain crosstalk in subsequent design and simulations, thereby
simulations completely when determining the crosstalk. bypassing the computationally expensive EM and transient
In this work, the performance of different training simulations. This will be especially useful when analyzing
algorithms was compared and the best one was selected as complex multi-variable designs with stringent time to
the final design. In addition to the training algorithms, the market requirements.

123
7320 Neural Computing and Applications (2020) 32:7311–7320

The initial component-level analysis performed in this 11. Ilumoka AA (2001) Efficient and accurate crosstalk prediction
study can be extended to a system-level modeling using via neural net-based topological decomposition of 3-D intercon-
nect. IEEE Trans Adv Packag 24(3):268–276
neural networks. Capturing the mutual interactions 12. Mbairi FD, Siebert WP, Hesselbom H (2008) High-frequency
between parameters in different components, while transmission lines crosstalk reduction using spacing rules. IEEE
undoubtedly complex, would seem to be a natural fit for Trans Compon Packag Technol 31(3):601–610
artificial intelligence-based methods. The recent advance- 13. Wu S, Fan J (2013) Modeling of crosstalk between two nonpar-
allel striplines on adjacent layers. IEEE Trans Electromagn
ments in related areas, for example in knowledge-based Compat 55(6):1302–1310
networks and hybrid learning methods, could ease the 14. Zhang QJ, Gupta KC, Devabhaktuni VK (2003) Artificial neural
burden of training such systems. In addition, sampling networks for RF and microwave design-from theory to practice.
algorithms to circumvent the large amount of data that is IEEE Trans Microw Theory Tech 51(4):1339–1350
15. Rayas-Sanchez JE (2004) EM-based optimization of microwave
required to train multi-dimensional systems with large circuits using artificial neural networks: the state-of-the-art. IEEE
number of variables presents an interesting future work to Trans Microw Theory Tech 52(1):420–435
further advance the development in this field. 16. Ozkaya U, Seyfi L (2018) A comparative study on parameters of
leaf-shaped patch antenna using hybrid artificial intelligence
Acknowledgements This work is supported by Universiti Sains network models. Neural Comput Appl 29(8):35–45
Malaysia under the Research University (RUI) Grant (1001/PELECT/ 17. Ye F, Zhang Z, Chakrabarty K, Gu X (2013) Board-level func-
8014011). tional fault diagnosis using artificial neural networks, support-
vector machines, and weighted-majority voting. IEEE Trans
Comput Des Integr Circuits Syst 32(5):723–736
Compliance with ethical standards 18. Sheikhan M, Sha’bani AA (2013) PSO-optimized modular neural
network trained by OWO-HWO algorithm for fault location in
Conflict of interest The authors declare that they have no conflict of analog circuits. Neural Comput Appl 23(2):519–530
interest. 19. Beyene WT (2007) Application of artificial neural networks to
statistical analysis and nonlinear modeling of high-speed inter-
connect systems. IEEE Trans Comput Des Integr Circuits Syst
26(1):166–176
References 20. Hsu KT, Guo WD, Shiue GH, Lin CM, Huang TW, Wu RB
(2008) Design of reflectionless vias using neural network-based
1. Fan J, Ye X, Kim J, Archambeault B, Orlandi A (2010) Signal approach. IEEE Trans Adv Packag 31(1):211–218
integrity design for high-speed digital circuits: progress and 21. Kim H, Sui C, Cai K, Sen B, Fan J (2017) Fast and precise high-speed
directions. IEEE Trans Electromagn Compat 52(2):392–400 channel modeling and optimization technique based on machine
2. Deo M, Enabling next-generation platforms using Intel’s 3D learning. IEEE Trans Electromagn Compat 60(6):2049–2052
system-in-package technology [White paper]. Retrieved from 22. Raja MA, Mehmood A, Niazi SA, Shah SM (2018) Computa-
http://www.intel.com tional intelligence methodology for the analysis of RC circuit
3. Hill DA, Cavcey KH, Johnk RT (1994) Crosstalk between modelled with nonlinear differential order system. Neural Com-
microstrip transmission lines. IEEE Trans Electromagn Compat put Appl 30(6):1905–1924
36(4):314–321 23. Hornik K, Stinchcombe M, White H (1989) Multilayer feedfor-
4. Rizvi M, LoVetri J (1996) Modeling and reduction of crosstalk ward networks are universal approximators. Neural Netw
on coupled microstrip line structures and multichip modules: an 2(5):359–366
FDTD approach. Int J Microw Millim-Wave Comput Eng 24. Hornik K (1991) Approximation capabilities of multilayer feed-
6(1):58–68 forward networks. Neural Netw 4(2):251–257
5. Xiao FC, Liu WK, Kami Y (2001) Analysis of crosstalk between 25. Hagan MT, Demuth HB, Beale MH, Jesús OD (2014) Neural
finite-length microstrip lines: FDTD approach and circuit-concept network design. Martin Hagan
modeling. IEEE Trans Electromagn Compat 43(4):573–578 26. Setiono R, Hui L (1995) Use of a quasi-Newton method in a
6. Sohn YS, Lee JC, Park HJ, Cho SI (2001) Empirical equations on feedforward neural network construction algorithm. IEEE Trans
electrical parameters of coupled microstrip lines for crosstalk Neural Netw 6(1):273–277
estimation in printed circuit board. IEEE Trans Adv Packag 27. Hagan MT, Menhaj MB (1994) Training feedforward networks
24(2):521–527 with the Marquardt algorithm. IEEE Trans Neural Netw
7. Lu T, Sun J, Wu K, Yang Z (2018) High-speed channel modeling 5(6):989–993
with machine learning methods for signal integrity analysis. IEEE 28. Nguyen T, Schutt-Aine JE (2018) A pseudo-supervised machine
Trans Electromagn Compat 60(6):1957–1964 learning approach to broadband LTI macro-modeling. In: 2018
8. Das D, Lee CSG (2018) Cross-scene trajectory level intention IEEE international symposium on electromagnetic compatibility
inference using Gaussian process regression and naive registra- and 2018 IEEE Asia-Pacific symposium on electromagnetic
tion. Department of Electrical and Computer Engineering Tech- compatibility. IEEE, Singapore, pp 1018–1021
nical Reports. Paper 491. Purdue University 29. Nguyen T, Lu T, Sun J, Le Q, We K, Schut-Aine J (2018)
9. Dai F, Bao G, Su DL (2010) Crosstalk prediction in non-uniform Transient simulation for high-speed channels with recurrent
cable bundles based on neural network. In: Proceedings of the 9th neural network. In: IEEE 27th conference on electrical perfor-
international symposium on antennas, propagation and EM the- mance of electronic packaging and systems. IEEE, San Jose,
ory. IEEE, Guangzhou, pp 1043–1046 pp 303–305
10. Cannas B, Fanni A, Maradei F (2002) A neural network approach
to predict the crosstalk in non-uniform multiconductor trans-
Publisher’s Note Springer Nature remains neutral with regard to
mission lines. In: 2002 IEEE International Symposium on Cir-
jurisdictional claims in published maps and institutional affiliations.
cuits and Systems. IEEE, Phoenix, pp 573–576

123