A Practical Approach Based on Machine Learning

to Support Signal Integrity Design

Abstract—A PCB design system enhanced with AI/ML mod- an AI/ML-based system, the knowledge gaps in the develop-
ules can support the optimal use of microelectronic components in ment process can be closed by systematically capturing the
the development process. To do this, the PCB and circuit designer domain knowledge of the engineers involved and thus restoring
must be provided with AI-based suggestions for SI-compliant
interconnection of components in the early design phase. AI- the link to the overall concept of the system design. This
based modules can also serve as a reference for engineers working maintains the necessary flexibility in each design phase, but
in the selection of interconnect structures in the pre-, concurrent-, significantly reduces the complexity of dealing with possible
and post-layout analysis phases but having little or no experience parasitic couplings.
with signal integrity (SI). This paper shows from a practical point
of view how to create ML modules for SI analysis. Selected ML II. P ROCESS M ODEL TO I NTEGRATE SI-C ONSTRAINTS
modules (k-Nearest Neighbor (kNN) + Neural Network (NN - INTO E LECTRONIC S YSTEM D ESIGN
Keras) + Support Vector Regression (SVR)) for predicting design
relevant SI parameters for PCB subnetworks are presented. The motivation to consider SI constraints during PCB design
Index Terms—Pre layout analysis, post layout analysis, con- is first of all to ensure a proper operation of the system within
current analysis, signal integrity, high speed systems, automotive given ambient conditions, but as well as a compliance with all
control units, AI/Ml models, AI/ML SI-modules, neural network,
relevant EMC constraints. A model based on the characteristic
NN Keras model, k-Nearest Neighbor (kNN), Support Vector
Regression (SVR) design stages has been developed in order to obtain a general
approach for describing the integration of EMC constraints
I. I NTRODUCTION into the development process, as shown in Figure 1 [9]. In this
The development of today’s electronic systems requires a context, the effective handling of SI constraints is an essential
separation of the respective development steps into different part of ensuring compliance with EMC design constraints. For
tasks. For example, the various development steps |circuit the implementation of AI/ML models respectively modules, an
and layout design| of a printed circuit board and system extension of the process given in [9] was made with respect to
integration as well as the handling of physical couplings AI/ML relevant data sources and objects based on an already
(SI/PDN/EMI/Thermal/...) are often performed by different existing generalized design process for electronic systems.
engineers. This separation often leads to the fact that the
reference to the circuit design can be lost. This is a major
reason why several design cycles often have to be performed
during the PCB layout process.
Despite an established and feature-rich tool landscape to
support signal integrity during PCB design, designers of
complex circuits, who often have less experience with SI, face
challenges such as |I don’t know which parameters to select for
the PCB trace configuration|, or |What are the critical SI rules
for a particular component type|, or |I don’t know who to ask|,
and |I don’t know where to find all the relevant information|.
This problem can be addressed by a design methodology
that takes into account the relevant parasitic physical couplings Fig. 1. Generalized PCB design process and extension regarding AI/ML
and is supported by domain-oriented AI/ML systems. By using relevant data sources and objects
 III. AI/ML M ETHODS D. Neural Networks
The area of Neural Network (NN) algorithms is an es-
A. AI/ML Systems and Frameworks
tablished branch of ML and has been successfully applied
AI methods (machine learning, genetic algorithms, etc.) to solve problems of various disciplines. The highly parallel
have found their way into various engineering domains in architecture enables the execution of complex tasks. Similarly,
recent years. Numerous approaches have been established in Artificial Neural Network (ANN) algorithms mimic this struc-
the field of AI systems and frameworks. Most of todays ture of linked neurons by grouping them into separate layers,
systems rely on deep learning methods; one well known namely the input, output and hidden layers. The common
example is Google’s Tensor Flow tool. In addition, a large designation of Deep Neural Network (DNN) is given to NN
number of other machine learning algorithms exist, such as architectures with two or more hidden layers. Each neuron
Support Vector Machines, Random Forest, Markov Chains, holds an adjustable weight, which is utilized to minimize the
etc. The available OpenSource tools concentrate on the provi- loss (usually MSE) between true and predicted value during
sion of individual algorithms. In the following, Support Vector training. This is achieved by means of backpropagation and
Regression (SVR), k-Nearest Neighbor (kNN), and Neural automatic differentiation. Therefore, NNs provide predictive
Network [5, 7, 8] methods used in this report to generate modelling for classification and regression tasks, whereat the
AI/ML models are briefly described. latter is relevant for the proposed approach.
B. Single-Target Support Vector Regression IV. A PPROACH FOR H ANDLING THE SO - CALLED I NVERSE
P ROBLEM IN THE PCB D ESIGN A PPLICATION
Multi-output regression methods are inevitable with respect
In the following, a solution approach for handling the so-
to real data applications. According to the categorization given
called inverse problem in the PCB design process is presented
in [2], one can distinguish between problem transformation
(Figure 2). The use of AI/ML methods to address SI problems
and algorithm adaption methods for multi-output regression.
like as described is not aiming to replace conventional SI
For simplicity, the single-target method [3] is used here, which
simulations or simulation environments [10]. Instead, PCB
is based on the problem transformation method. In this case,
developers should be enabled to select a proper parameter set
one independent model is used for each target, so that they
suitable for a net topology from a series of suitable predictions
cannot be correlated directly. The individual predictions are
without having to perform time-consuming parameter sweeps.
done by a support vector regression (SVR). The aim of the
SVR algorithm is to find an estimate for a continuous-valued
function y = f (x). For learning this function, training data
(x1 , y1 ), . . . , (xl , yl ) ⊂ Rd × R is used, where d is the
dimension of the input [1]. When the estimated function fˆ(x)
is nonlinear, the vectors xi must be mapped into a higher
dimensional feature space [4]. This transformation is done with
a kernel function. In this work, a Gaussian kernel as described
in [4] was used.

C. k-Nearest Neighbor (kNN)

The k-Nearest Neighbor algorithm is a basic unsupervised Fig. 2. Workflow for the inverse problem
method, whose underlying assumption states that the distance
of similar examples within the data is small. By definition, Therefore, design decisions can be made more quickly. Es-
this proximity-based approach searches k samples, which are pecially since the interpretation of the resulting voltage curves
distance-wise the closest to a given example. There are various from such corner cases analysis requires detailed knowledge
metrics with respect to distance, but most commonly the of the relevant domain.
Euclidean distance is utilized, while the Manhattan, Minowski V. E XAMPLE FOR W ORKING OUT THE K NOWLEDGE
and Mahalanobis distances provide alternatives. To calculate D OMAIN PCB D ESIGN UNDER SI C ONSTRAINTS
the distance-wise closest samples different algorithmic choices
For the illustration of the ML-based approach a simple 3-
are available. The brute force method simply calculates the
point PCB network was used. For this purpose, a schematic
distance to each sample of the dataset individually, while with
generated with [10] (Figure 3) was transformed into an
tree-based algorithms like KD tree the distance information is
equivalent circuit based on linear circuit elements [11]. The
aggregated for clusters of points in close proximity to each
starting point was data generation by means of simulation
other to increase the computational efficiency. Therefore, in
through variation of selected design parameters (characteristic
terms of tuneable hyperparameters there is only the distance
impedances; line lengths; ...). The subsequent training was
metric, the algorithm and the value for k, which makes the
realized with simple AI models (Support Vector Regression;
method quite robust and simple.
k-Nearest Neighbor; Neural Networks). Further models were
also implemented with extended data sets.
 VI. I MPLEMENTATION /R ESULTS
A. Generation of Training Data
For the training of the AI/ML models a reduced parameter
setup for net 1.7 was used for simplicity. The data set
was produced by varying the values for the resistors RS ,
RL2 − RL4 (changed simultaneously) and the impedances of
the transmission lines L1, L2 -L4 (changed simultaneously).
For this purpose, a transmission line routing on inner layers
(symmetric stripline) was assumed.
Thus, 3087 parameter + features constellations (3087
Fig. 3. Test setup (PCB net 1.7) based on linear circuit elements for data rows x 7 columns) with corresponding simulation results were
generation (tr = 500 ps; tf = 500 ps; U 01 = 3.3 [V])
obtained. Subsequently, the features energy, entropy and rise
A. Introduction Feature Analysis: Energy/Entropy/Rise Time - time for the training process were extracted.
Comparison based on Voltage Waveforms
B. Implementation of k-Nearest Neighbor (kNN) Modell
The features used for the selected AI/ML models were
An essential rule in the implementation of AI/ML mod-
energy, entropy, and rise time. These were extracted from the
els is to start with the simplest algorithms. Therefore, for
waveforms shown in Figure 4. In Table I an overview of the
comparison reasons, a kNN model was implemented as an
used features is given. For the PCB net 1.7 (Figure 3) the
unsupervised learning algorithm. The open-source library [5]
following three signal cases were considered.
was also used to analyse the multi parametric data and to
Case#1: Input signal U01; evaluate results. A neural network with an input and output
Case#2: Signal with overshoot (URE2)
[Rs = 1Ω/RL2-4=1Ω/Zs = 47Ω/ZL2-4 = 52Ω]; layer and 2 hidden layers with 32 neurons each was also
Case#3: Signal with low rise time (URE2) developed (Figure 5) [7, 8].
[Rs = 44Ω/RL2-4 = 33Ω/Zs = 31Ω/ZL2-4 = 29Ω].

TABLE I
F EATURES FOR THE IMPLEMENTATION OF AI/ML
∞
|x(n)|2
P
Energy E = ∆x ·
−∞
∞
P
Shannon Entropy Entropy = − pi · log2 (pi )
−∞
pi : Probability density of the amplitudes
dV
Rise Time dT max
(Range: 10% and 90% of
the maximum voltage)

Table II shows the feature settings for cases 1 - 3. The

v(t) waveforms corresponding to the given cases are shown
Fig. 5. Schematic representation of the neural network architecture to be
in Figure 4. considered
TABLE II
F EATURES SETTINGS FOR CASES 1 - 3 The related training and validation loss (MSE) as a function
CASE1 CASE2 CASE3 of the number of epochs (StdSetup) is given in Figure 6 (max.
Energy [nW] 112.28 86.15 123.55 epochs 25,000; runtime: ~ 4893 s).
Entropy [sh] 0.97 4.07 4.19
Rise Time [V/ns] 6.58 1.85 5.67

Fig. 6. Training and validation loss (MSE)

The learning rate was set to 0.01 and for the StdSetup
Fig. 4. Waveforms for feature analysis time NN model for the parameter DropOut the factor 0.25 was
chosen additionally. The separation of the generated dataset depicted in Table IV were selected. In the first configuration,
into training and test set was done according to 75% and 25%, denoted as |worst-case|, a parameter prediction leading to
respectively. The following plots of the training and validation an unacceptable signal shape with low slew rate has been
loss indicate that the generalization of the NN model is well expected. For the |best-case| configuration, a waveform has
adjusted. The validation loss becomes minimal and thus an been expected which shows a high correlation with the input
optimal model capacity can be achieved. An excessive over- signal.
fitting has not been observed. Setting the parameter DropOut
to a factor of 0.25 prevents the model from overfitting and
penalizes large weights with infinitesimal contribution in terms
of training loss reduction.
C. Prediction by using a k-Nearest Neighbor Model and
respectively a NN Model
A k-Nearest Neighbour approach was first used (scikit-learn
library; Euclidean distance) to determine the nearest neigh- Fig. 8. Parameter prediction with multi-output SVR
bours based on predefined feature configurations (see Table
III). To realize the search, the BruteForce implementation Subsequently, we simulated the star point net again based on
was chosen and the k value was set to 20. The choice of the predicted parameters (see Table IV) in order to validate the
features in the |best case| leads to an adequate adaptation results of the implemented multi-output SVR. Figure 9 shows
of the subnetwork from Figure 3. In the |worst case|, the that the outcome for the worst-case scenario is as expected. It
configuration should generate a signal with overshoot. can be seen that the signal has inadequate slew rate value. The
simulation result for the best case parameter configuration is
TABLE III
F EATURE CONFIGURATIONS FOR DETERMINING NEAREST NEIGHBOURS
shown in Figure 9 (right). The signal has a high slew rate with
negligible overshoot, the trained multi-output SVR provides
|worst case| |best case|
Energy [nW] 130 100
reasonable prediction results.
Entropy [sh] 4 3
Rise Time [V/ns] 2.8 2.2 TABLE IV
F EATURE CONFIGURATIONS FOR SVR VALIDATION

Figure 7 shows that the proposed parameters result in the ex- |worst case| |best case|
Energy [nW] 22 22
pected voltage waveforms under both the |worst case| and |best xi Entropy [sh] 3 3
case| conditions. After determination of the parameters of the Rise Time [V/ns] 0.3 3
subnetwork by the nearest-neighbour algorithm, a simulation
prediction

Rs 43Ω 11Ω
with LTspice was performed for validation of the results. The RL2...RL4 36Ω 77Ω
quality of a neural network for predicting the parameters from Z (of L1) 32Ω 42Ω
Figure 3 was also investigated. The architecture of the neural Z (of L2...L4) 30Ω 54Ω
network is identical to Figure 5. Analogous to the procedure
in the nearest-neighbour algorithm, the feature configuration
from Table III was used after training the neural network to
obtain a prediction. The predictions by the neural network
also properly reproduce the two predefined cases similar to
the nearest-neighbour algorithm, except minor deviations in
the voltage response are visible.
Fig. 9. Waveform for worst-case (left) and best-case (right) SVR prediction

VII. A PPLICATION TO P RACTICAL SI P ROBLEMS

A. Best case analysis using kNN and NN model
Results for a selected set of parameters from the kNN (true)
and NN results (predicted) are presented below. The kNN data
for net 1.7 were determined using distances in feature space
Fig. 7. Voltage waveform in |worst case|- (left) and |best case|-case (right) from each other. To do this, the k = 20 nearest neighbours were
with prediction by k-nearest neighbour algorithm determined based on the given feature set (energy = 100 nW;
entropy = 3 Sh; slope = 2.2 V/ns). For the selected feature
D. Implementation and Parameter Prediction with Single- set (Table V) - as one of the 20 nearest neighbours - the
Target SVR voltage waveforms at RE2 were then determined by simulation
Figure 8 illustrates the structure of the implemented single- for the parameters shown in Table V. The NN data set was
target SVR. For testing the SVR, two feature configurations as obtained for the same feature set using the previously trained
neural network (see Figure 5). The similarity of the voltage further improved the results. Again, no excessive overfitting is
waveforms for the parameter set found by means of kNN and observed. However, it must be noted that bad results are not
by means of trained NN confirm the quality of the methods improved. A data selection for good data quality is therefore
(see Figure 10). mandatory (see example: restriction for R5 to a maximum
value of 120 Ω; Figure 12 and 13).
TABLE V
PARAMETER CONFIGURATIONS FOR K NN UND NN VALIDATION

Feature Set [ID14]

Energy [nW] 96.9 TRUE PREDICTION
Entropy [sh] 3.9 kNN NN
Rise Time [V/ns] 1.9 [Ω] [Ω]
Rs 22.0 26.5
RL2...RL3 15.0 13.2
Z (of L1) 40.0 46.0
Z (of L2...L4) 55.0 59.4
Fig. 12. Comparison of prediction quality for a limited data set
(R5|MAX = 120 Ω)

Fig. 10. Voltage for the selected |best-case| case ID14 for the kNN parameters
(TRUE) and the NN parameters (PREDICTION)
Fig. 13. Training and Validation Loss (MSE) as a function of epoch count;
B. Application to SI Problems of a Non-Linear 2 Point Net- max. epochs 10000.
work
C. Comparison of Linear and Nonlinear Buffer Models
Based on the above, an extensive data set for a 2 point
Based on the subnet 1.7, linear LTspice and nonlinear AC86
network (Figure 11) with a non-linear transmitter (AC86) and
IBIS models for transmitter and receiver have been compared
a 1 MΩ termination (this represents the high impedance AC86
regarding the training process and prediction of ML modules
input) was generated to train an another neural network. The
with respect to the inverse problem. Again, the kNN-NN
features used this time were overshoot and rise time of the
approach with the same hyperparameters has been deployed
voltage waveform at R2. The length for TL1 and the resistance
for both datasets. Naturally, the loss functions for training of
R5 (E12: 1Ω to 4.7kΩ) were used for data generation. With
the NNs (shown in Figure 14) indicate some discrepancies
the simulated training data a neural network with one input
and deviations in the range below 1000 epochs, but generally
and one output layer and as well as 2 hidden layers with 512
a similar behaviour and virtually convergence to the same loss
neurons each was built (activation function: ReLU; optimizer:
after the maximum 5000 epochs is observed. In the same way,
Adam; learning rate: 0.01).
the kNN algorithm, which has been deployed for worst and
best case based on the parameters shown in Table VI, also
generates similar results for the linear (based on an adapted
LTspice model for the AC86) and nonlinear (based on AC86
IBIS model using [10] for simulation) dataset of 20 nearest
neighbours for subnet 1.7 using LTspice simulation in Figure
16 (fCLOCK = 50 MHz).

TABLE VI
F EATURE CONFIGURATIONS FOR DETERMINING NEAREST NEIGHBOUR
|worst case| |best case|
Energy [nW] 97.5 93.25
Maximum [V] 5 3.3
Fig. 11. 2 Point PCB network with non-linear transmitter (AC86 IBIS model)
and high impedance termination Rise Time [V/ns] 0.98 0.81

Additionally, a hyperparameter optimization using Bayesian In addition, the prediction results of the NN for the non-
Optimization (utilizing 50 iterations) was performed, which linear AC86 IBIS model as transmitter and receiver of worst
and best case behaviour (shown in Table VII and Figure 16) In addition, the integration of domain knowledge into the
using the simulator from [10], validate the performance of the process of developing AI/ML models must be significantly
proposed method. improved. For this purpose, additional EDA knowledge (tacit
knowledge) must also be incorporated into the implementation
of such AI/ML models. For additional signal topologies (Far
End Cluster; Point2Point; Daisy Chain), extensive learning
data must be generated and mapped into the corresponding
AI/ML modules (training). For the training of new AI/ML
models, it will also be necessary to use appropriate SI
constraints of component technologies such as DDR3 and
DDR4. For this purpose, data generation by simulation must
be significantly improved and simplified.
Additional AI models based on neural networks and multi-
Fig. 14. Training and validation Loss (MSE) for linear (left) and nonlinear
(right) termination; max. epochs 5000.
output support vector regression methods also need to be
studied in more detail for the nonlinear case. This requires
substantial analysis to investigate the consequences of the
nonlinear behavior of fast components for data generation.
To increase the accuracy of the predictions, it should also be
possible to integrate additional constraints. For this purpose,
e.g., different activation functions have to be considered. In the
future, it will also be necessary to develop a general method
for an optimal tuning of hyperparameters for each new AI/ML
model.
Fig. 15. 20 nearest neighbours determined by the kNN algorithm for best-case In order to address SI issues effectively in existing PCB
(left) and worst-case (right) behaviour.
design environments, it will also become important to provide
TABLE VII
AI platforms for the integration of tested AI/ML modules
W ORST- AND BEST- CASE PREDICTIONS OF THE NN ([Link]
RS R2−4 Z1 Z2−4 L1 L2−4
[Ω] [Ω] [Ω] [Ω] [mm] [mm]
Worst 0 11 62 59 101 39
Best 25 52 52 52 100 39

R EFERENCES
[1] Smola, Alex J.; Schölkopf, Bernhard (2004): A tutorial on support vector
regression. Statistics and Computing 14, 199-222.
[2] Borchani, Hanen; Varando, Gherardo; Bielza, Concha; Larranaga, Pedro
(2015): A survey on multi-output regression. Wiley Interdisciplinary
Reviews: Data Mining and Knowledge Discovery, vol. 5, issue 5, pp.
216-233.
[3] Spyromitros-Xioufs, E; Groves, W.; Tsoumakas, G.; Vlahavas, I.
Fig. 16. Voltage curve at RE2 based on the parameters predicted by the NN (2012): Multi-label classification methods for multi-target regression.
trained with the nonlinear AC86 dataset (fCLOCK = 10 MHz). arXiv:1211.6581, Cornell University Library, 1159-1168.
[4] Lu, T.; Sun, J.; Wu, K.; Yang, Z. (2018): High-Speed Channel Modeling
VIII. S UMMARY /F URTHER W ORK With Machine Learning Methods for Signal Integrity Analysis. In: IEEE
Transactions on Electromagnetic Compatibility, vol. 60, no. 6, pp. 1957-
Both, the presented method for data generation and the im- 1964.
[5] Scikit-learn: Machine learning in Python. [Link]
plemented ML models form a solid basis for successful AI/ML Feb 2022.
support of SI engineers in the design process. Nevertheless, [6] [Link]
the development of ML models needs to be examined in a nearest-neighbors
[7] Tensorflow. [Link] Feb 2022.
more differentiated way. For example, optimization must be [8] Keras. [Link] Feb 2022
performed with respect to hyperparameters: Number of hidden [9] W. John; EMC of Printed Circuit Boards and Microelectronic Engineer-
layers; learning rate; number of neurons; activation functions; ing Techniques; lnvited Paper - 13th International Wroclaw Symposium
and Exhibition on Electromagnetic Compatibility. June 1996, Wroclaw,
kernel, number of neighbors k; etc. Special attention must be Poland
paid to the development of metrics in order to evaluate the [10] CR-5000 Lightning 21.0 Training Manual (TM-HT-969-E-10); Zuken
quality of the training data. GmbH/Zuken Ltd; 2010 2021
[11] LTspice. [Link] Feb 2022.