Artificial Neural Networks for RF and Microwave Design: From Theory to Practice Qi-Jun Zhang + Kuldip C. Gupta* and Vijay K. Devabhaktuni + +Department of Electronics, Carleton University, Ottawa, ON, Canada *Department of ECE, University of Colorado, Boulder, CO, USA
55
Embed
Artificial Neural Networks for RF and Microwave Design: From Theory to Practice Qi-Jun Zhang + Kuldip C. Gupta* and Vijay K. Devabhaktuni + +Department.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Artificial Neural Networksfor RF and Microwave Design:
From Theory to Practice
Qi-Jun Zhang+
Kuldip C. Gupta*and
Vijay K. Devabhaktuni+
+Department of Electronics, Carleton University, Ottawa, ON, Canada
*Department of ECE, University of Colorado, Boulder, CO, USA
• Introduction and overview• Neural network structures• Neural network model development process• RF/Microwave component modeling using neural
networks• High-frequency circuit optimization using neural
network models• Conclusions
Outline
• Accurate RF/Microwave design is crucial for the current upsurge in VLSI, telecommunication and wireless technologies
• Design at microwave frequencies is significantly different from low-frequency and digital designs
• Substantial development in RF/microwave CAD techniques have been made during the last decade
• Further advances in CAD are needed to address new design challenges, e.g., 3D-EM optimization, CPW and multi-layered circuits, IC antenna modules, etc
• Fast and accurate models are key to efficient CAD
• Neural network based modeling and design could
significantly impact high-frequency CAD
Introduction
A Quick Illustration Example:
Neural Network Model for Delay Estimation in a High-Speed
Interconnect Network
High-Speed VLSI Interconnect Network
Driver 1
Driver 2
Receiver 1
Driver 3
Receiver 2
Receiver 3
Receiver 4
Circuit Representation of the Interconnect Network
C1R1 R2 C2
C4R4
C3R3
L1 L2L3
L4
1 2
3
4
SourceVp, Tr
Rs
• A PCB contains large number of interconnect networks, each with different interconnect lengths, terminations, and topology, leading to need of massive analysis of interconnect networks
• During PCB design/optimization, the interconnect networks need to be adjusted in terms of interconnect lengths, receiver-pin load characteristics, etc, leading to need of repetitive analysis of interconnect networks
• This necessitates fast and accurate interconnect network models and neural network model is a good candidate
Simulation Time for 20,000Interconnect Configurations
Method CPU
Circuit Simulator (NILT) 34.43 hours
AWE 9.56 hours
Neural Network Approach 6.67 minutes
• Neural networks have the ability to model multi-dimensional nonlinear relationships
• Neural models are simple and the model computation is fast
• Neural networks can learn and generalize from available data thus making model development possible even when component formulae are unavailable
• Neural network approach is generic, i.e., the same modeling technique can be re-used for passive/active devices/circuits
• It is easier to update neural models whenever device or component technology changes
Important Features of Neural Networks
• Neural models are efficient alternatives to closed-form expressions, equivalent circuit models and look-up tables
• Neural network models can be developed from measured or simulated data
• Neural models can also be used to update or improve the accuracy of already existing models
• Neural network models have been developed for active devices, passive components and interconnect networks
• These models have been used in circuit simulators for circuit-level simulation, design and optimization
Neural Networks for RF/Microwave Applications: Overview
Neural Network Structures
• A neural network contains• neurons (processing elements) • connections (links between neurons)
• A neural network structure defines • how information is processed inside a neuron • how the neurons are connected
• Examples of neural network structures• multi-layer perceptrons (MLP)• radial basis function (RBF) networks• wavelet networks• recurrent neural networks• knowledge based neural networks
• MLP is the basic and most frequently used structure
Neural Network Structures
MLP Structure
(Output) Layer L
(Hidden) Layer L-1
1 2 NL
NL-1321
. . . .
. . . . . . .
. . . .
(Hidden) Layer 2
(Input) Layer 1
321 N2
N131 2
x1 x2 x3
. . . .
. . . .
xn
y1 y2 ym
Information Processing In a Neuron
12lz
liz
(.)
….
liw 0
10lz
liw 1 l
iw 2
li
liN l
w1
11lz
1
1
lN l
z
• Input layer neurons simply relay the external inputs to the neural network
• Hidden layer neurons have smooth switch-type activation functions
• Output layer neurons can have simple linear activation functions
Neuron Activation Functions
0
0.5
1
1.5
-25 -20 -15 -10 -5 0 5 10 15 20 25
Sigmoid
()=1/(1 +e-)
Activation Functions for Hidden Neurons
-2
-1
0
1
2
-10 -8 -6 -4 -2 0 2 4 6 8 10
Arc-tangent
()=(2/)arctan()
-2
-1
0
1
2
-10 -8 -6 -4 -2 0 2 4 6 8 10
Hyperbolic-tangent
()=(e+ -e-)/(e+ +e-)
MLP Structure
(Output) Layer L
(Hidden) Layer L-1
1 2 NL
NL-1321
. . . .
. . . . . . .
. . . .
(Hidden) Layer 2
(Input) Layer 1
321 N2
N131 2
x1 x2 x3
. . . .
. . . .
xn
y1 y2 ym
x
z(1)
z(l -1)
z(2)
z(L)
y
Z1 Z2 Z3 Z4
y1 y2
x1 x2 x3
W ’jk
Wki
Outputs
yj =W ’jkZk
k
3 Layer MLP: Feedforward Computation
Inputs
Zk = tanh(Wki xi )
Hidden Neuron Values
i
How can ANN represent an arbitrary nonlinear input-output relationship?
Universal Approximation Theorem(Cybenko, 1989, Hornik, StinchCombe and White, 1989)
In plain words:
Given enough hidden layer neurons, a 3-layer MLP neural network can approximate an arbitrary continuous multidimensional function to any desired accuracy
• The number of hidden neurons depends upon the degree of non-linearity, and dimension of the original problem
• Highly nonlinear problems and high dimensional problems need more neurons while smoother problems and small dimensional problems need fewer neurons
• To determine number of hidden neurons• experience• empirical criteria• adaptive schemes• software tool internal estimation
How many hidden neurons are needed?
Development of Neural Network Models
Notation
y = y(x, w): ANN model x: inputs of given modeling problem or ANN
y: outputs of given modeling problem or ANN
w: weight parameters in ANN
d : data of y from simulation or measurement
Define Model Input-Output and Generate Data
Define model input-output (x, y), for example,
x: physical/geometrical parameters of the component y: S-parameters of the component
Generate (x, y) samples (xk, dk) , k = 1, 2, …, P, such that the data set sufficiently represent the behavior of the given x-y problem
Types of Data Generator: simulation and measurement
The orders of magnitude of various x and d values in microwave applications can be very different from one another.
Scaling of training data is desirable for efficient neural network training
The data can be scaled such that various x (or d ) have similar order of magnitude
Training, Validation and Test Data Sets
The overall data should be divided into 3 sets, training, validation and test.
Notation:Tr - Index set of training dataV - Index set of validation dataTe - Index set of test data
In RF/microwave cases where overall data is limited, validation and test (or training and validation) data can be shared.
Error Definitions
Training error:
Validation and test errors EV and ETe can be similarly defined.
Training Objective: Adjust w to minimize EV , but the update of w is carried out using the information and
At end of training, the quality of the neural model can be tested using test error ETe
)(wErT
w
ErT
q1
Tk
m
1j
q
jmin,jmax,
jkkj
rT
r
r dd
d)w,x(y
m)size(T
1)w(E
• Sample-by-sample (or online) training: ANN weights are updated every time a training sample is presented to the network, i.e., weight update is based on training error from that sample
• Batch-mode (or offline) training: ANN weights are updated after each epoch, i.e., weight update is based on training error from all the samples in training data set
• An epoch is defined as a stage of ANN training that involves presentation of all the samples in the training data set to the neural network once for the purpose of learning
Types of Training
Neural Network Training
The error between training data and neural network outputs is feedback to the neural network to guide the internal weight update of the network
x
Neural Network W
y
Training Data
d
Training Error
-
Typical Training Process
Step 1: w = initial guess, set epoch = 0
Step 2: If (EV(epoch) < required_accuracy) or if (epoch > max_epoch)
then STOP
Step 3: Compute (or and ) using all samples
in training data set (i.e., batch-mode training)
Step 4: Use optimization algorithm to find and update the weights
Step 5: Set epoch = epoch + 1 and GO TO Step 2
)(wErT w
wErT
)(
wwww
)(wErT
Gradient-based Training Algorithms
where h is the direction of the update of w is the step size
Gradient-based methods use information of and to determine update direction h Step size is determined in one of the following ways:
Small value either fixed or adaptive during training Line minimization to find best value of
Examples of algorithms: backpropagation, conjugate gradient, and quasi-Newton
hw
)(wErT w
wErT
)(
Example: Backpropagation (BP) Training(Rumelhart, Hinton, Williams 1986)
In the gradient algorithm,
Let the update direction h be the negative gradient direction, then:
or
where is called learning rate is called momentum factor
hw
w
wEww
rT
)(
1|)(
epoch
T
ww
wEww
r
Desired accuracy achieved?
Desired accuracy achieved?
Yes
Flow-chart Showing Neural Network Training, Neural Model Testing, and Use of Training, Validation and Test Data Sets in ANN Modeling
Evaluate
validation error
Perform feedforward computation for all
samples in validation set
Assign random initial values for all the weight
parameters
Select a neural network structure, e.g., MLP
STOPTraining
Evaluate test error as an independent quality
measure for ANN model reliability
Perform feedforward computation for all samples in test set
START
Perform feedforward computation for all
samples in training set
Compute derivatives of training error w.r.t. ANN
weights
Update neural network weight parameters using a gradient-based algorithm (e.g., BP, quasi-Newton) Evaluate
training error
Evaluate validation error
Perform feedforward computation for all
samples in validation set
Assign random initial values for all the weight
parameters
Select a neural network structure, e.g., MLP
Evaluate test error as an independent quality
measure for ANN model reliability
Perform feedforward computation for all samples in test set
STOPTraining
Perform feedforward computation for all
samples in training set
Compute derivatives of training error w.r.t. ANN
weights
Update neural network weight parameters using a gradient-based algorithm (e.g., backpropagation) Evaluate
training error
START
No
Yes
Desired accuracy achieved?
Desired accuracy achieved?
Evaluate validation error
Perform feedforward computation for all
samples in validation set
Desired accuracy achieved?
Desired accuracy achieved?
Example:EM-ANN Models for CPW
Circuit Design and Optimization
Example: CPW Symmetric T-junction
Range of input parameters of CPW T-junction model
Input Parameter Minimum Value Maximum Value
Frequency 1 GHz 50 GHz
Win 20 m 120 m
Gin 20 m 60 m
Wout 20 m 120 m
Gout 20 m 60 m
Substrate: H=25 mil
r=12.9 tan = 0.0005
Strip: tmetal = 3 m
Wa = 40 m
CPW T-junction Geometry
Strip
Gout
Gout
Wout
Wa
Win GinGin
Error Comparison Between EM-ANN Model and EM Simulations for the CPW Symmetric T-Junction
|S11|S11 |S13|
S13 |S23|S23 |S33|
S33
Training DataAverage ErrorStd. Deviation
0.001500.00128
0.7540.696
0.000710.00058
0.1760.172
0.000840.00097
0.2460.237
0.001060.00109
0.6330.546
Test DataAverage ErrorStd. Deviation
0.003450.00337
0.7820.674
0.000880.00085
0.1410.125
0.001260.00105
0.1770.129
0.000830.00068
0.8380.717
Example:ANN Based Design of
a CPW Folded Double-Stub Filter
• CPW Transmission line
• CPW Bend
• CPW Short-circuit
• CPW Open-circuit
• CPW Step-in-width
• CPW Symmetric T-junction
List of EM-ANN Models Trained from
Detailed Electromagnetic (EM) Data
CPW Folded Double-Stub Filter Designed Using EM-ANN Models of Circuit Components
ANN Based Optimization:• Goal: Resonant at 26 GHz
• Optimize lstub and lmid
• Required 7 overall circuit
simulations• CPU-Time: 3 minutes
Simulation of the optimized filter circuit using EM-ANN models:
•30 seconds•100 frequency points
Full-wave EM simulation of the optimized filter circuit:
•14 hours •17 frequency points
CPW folded double-stub filter geometry
Comparison of Optimized Circuit Responses From EM-ANN Based Simulations and Full-Wave EM Simulations
Example:FET Modeling Using
Neural Networks
Source
L
W
a
Nd
Drain
Gate
Neural Model for MESFET Modeling
….
Id qg qd qs
L W a Nd Vgs Vds
Vd (V)
FET I-V curves: Neural model vs FET test data
Vg = 0V
-1V
-2V
-3V
-4V
-5V
I d (m
A)
FET S-parameters: Neural model vs FET test data
Frequency (GHz)
S21
S11
S22
S12S -
para
met
ers
(dB
)
SIMULATOR
Harmonic Balance
Equation Solver
Linear subnet
Nonlinearsubnet
YV I, QV V
Incorporating Large-Signal FET Neural Network Model into HB Circuit Simulator
I, Q
y
V
x
Neural Model
Example:
Yield Optimization of
a 3-Stage MMIC Amplifier
Using Neural Network Models
Three-Stage X-Band Amplifier CircuitWith 3 FETs Represented By ANN Models
Vd1
Vd3Vd2
Vg1
Vg3Vg2
Input
Output
Variable Mean Deviation (%) Variable Mean Deviation (%) Nd (1/m3) 1023 7.0 d (m) 0.1 4.0 L (m) 1.0 3.5 SC1 (m2) 326.8 3.5 a (m) 0.3 3.5 SC2 (m2) 2022.4 3.5 W (m) 300 2.0 SC3 (m2) 218.2 3.5 WL (m) 20 3.0 SC4 (m2) 352.2 3.5 SL (m) 10 3.0
Distributions for Statistical Variables
3-Stage Amplifier: Before and After ANN Based Yield Optimization500 Monte Carlo Simulations are Shown
• ANNs can be trained to learn RF/Microwave data and the resulting ANN model is fast and can accurately represent the corresponding input-output relationship
• ANN development is a computer-based training process as opposed to human-based trial-and-error process in developing empirical/equivalent circuit models
• Neural network modeling approach is generic, i.e., the same ANN method can be used to model passive or active components, for devices or circuits
• Neural models can be used in place of CPU-intensive detailed models to enhance high-level CAD operations such as circuit simulation and yield optimization
• Neural network based modeling and design promises to address ever increasing demand for efficient CAD
Conclusions
• Advanced methods of combining neural networks with RF/Microwave knowledge for efficient modeling
• Towards full automation of model generation process using neural networks featuring online data generation and ANN training
• Modeling nonlinear dynamic behaviors of devices and circuits employing neural network techniques
• 3D-EM modeling of passive components using ANNs
• Combining neural networks with other advanced CAD concepts such as the space mapping (SM) technique
• Application of artificial neural networks for RF and microwave measurements
NeuroModeler is a software for developing neural network models for passive and active components/circuits for high-frequency circuit design.
Here is a Web demonstration version with which you can train a MLP neural model, test it with test data, and see the neural model reproducing the input-output relationship it learnt.
Basic Steps of Running NeuroModeler
The demonstration version is self-explanatory, where the basic steps are: Press the “New Neural Model” button to define the neural model structure and number of input, output and hidden neurons
Press the “Train Neural Model” button to train the neural model
Press the “Test Neural Model” button to test the quality of the model
Press the “Display Model Input-Output” button to see the input-output relationship reproduced by the neural model
Start NeuroModeler
Make sure your computer is connected to the Internet.
Now you can click NeuroModeler here to start the Web demonstration version of the program.