Build Neural Models for RF/microwave Design
Artificial Neural Networks are emerging as a powerful
technology for RF and microwave characterization, modeling, and design.
NeuroModeler is the first software in the industry that embraces this technology
with complete RF and microwave orientation. It features fully integrated
RF and microwave knowledge based neural network architectures. It helps
you to immediately start developing neural models for RF/Microwave components
and circuits and helps to provide neural models for your simulators.
Recently, artificial neural networks are introduced to
the microwave community, opening the door for an unconventional approach
to microwave computer-aided design. Here is a brief outline of the what's,
why's, when's and how's, intended to give the RF/microwave engineers the
shortest cut to gain from this new technology.
Training Window of a FET Model, with main menu on background
What: Artificial Neural Networks are information
processing systems inspired by the ability of human brain to learn from
observations and to generalize by abstraction. The fact that neural networks
can learn totally different things led to their use in diverse fields such
as pattern recognition, speech processing, control, medical applications
and more. Recently, microwave researchers developed techniques in which
neural networks are trained from microwave data, and then used to enhance
Why: The microwave industry's drive for
manufacturability-driven design and time-to-market, demands efficient and
reliable CAD tools. However, manufacturability-driven design, e.g., statistical
design with accurate models such as EM models are time-consuming. Simpler
models are fast but are often under limited assumptions or mismatch may
occur between computer solutions and hardware measurements. Accuracy, speed
and flexibility have been for most of the time contradictory, until recently
when neural models for microwave components are introduced. Neural models
can be much faster than original detailed models, more accurate and flexible
than empirical models, and easier to develop when a new device/new technology
When: Neural-based modeling is a generic
technique, and shines especially in challenging microwave design situations
where conventional techniques balk. Examples are, EM-level modeling but
not the expensive EM computations, modeling new components when component
formulae are not available, supply user-defined components to simulators
when simulator does not easily support user-plugs, etc.
How: NeuroModeler is a RF/microwave oriented
software tool, to help you to quickly develop neural models for active
and passive components; at both device and circuit-levels; and for linear
or nonlinear simulations. The software is written for typical RF/Microwave
designers who are not neural network experts but would like to immediately
get started with this new technology.
What's Special: NeuroModeler is the first and only software
that allows your RF/microwave knowledge to be integrated with neural network
learning. You can supply your knowledge through symbolic expressions, or
through a circuit netlist with our build-in circuit simulator, or through
your own simulator which NeuroModeler can drive, or through a C program source
code to be linked with NeuroModeler.
A Microstripline Neural Model defined by Visual Editor
Benefit from an Emerging Technology Right Away:
Neural model contains a set of neurons and connections
between neurons. Each neuron has an activation function processing the
incoming information from other neurons. Take a neural model for microwave
transmission line as an example, the transmission line geometrical parameters
(say x) will be model inputs sent to some neurons called input neurons.
After internal processing of all neurons, the neural model will produce
electrical quantities (say y) of the transmission line at some neurons
called output neurons. In model development stage, samples of x-y data
are generated (e.g., from EM simulation or measurement). The model is then
trained to learn from the data. Training is similar to an optimization
process where internal parameters of the neural model are adjusted such
that modeled solution best fits training data. A trained neural model can
then be used online during microwave design stage providing fast model
evaluation replacing original slow EM simulators. Since neural model is
trained directly from data, the model can be developed even if original
problem formulae do not exist.
Highlights of Product Features:
Model Creation and Editing: Use NeuroModeler to easily
create a neural network model. The built-in default automatically defines
a model structure and number of neurons for you. You can choose from a
variety of templates or define customer structures of neural models.
Attractive visual editor let you define model structures graphically.
Data Processing: Here you dictate what the neural network
should learn from your data. NeuroModeler automatically performs basic
data preprocessing for you. NeuroModeler can even help you to generate
data using its Simulation Driver feature.
Training: NeuroModeler automatically checks your data and
suggests a training technique. The program also has an Auto-Pilot
training method, which intelligently adjusts the neural model size and
training methods to achieve the required model accuracy. User can modify
any training defaults and suggestions.
Test: The performance of neural model can be verified using
an independent set of data, either with a simple error criteria, or by a
variety of detailed plots. You can also evaluate interpolation and
extrapolation capabilities of your model.
- Export: You and your work are not locked to NeuroModeler
format. You can export the trained neural model to your own user-environment,
be it a spreadsheet, or a computer-source code, or a simulator.
Multilayer Perceptrons, Radial Basis Functions,
Wavelet Networks, Knowledge-Based Neural Nets, Space Mapped Nerral Nets,
Prior Knowledge Input Networks, Hierarchical Neural Nets, User-defined
|Sigmoid, arctangent, hyperbolic tangent, Gaussian, linear,
quadratic, polynomial, rational, log, exponential, normalized, multisigmoid,
time or frequency domain approximations, inductance/capacitance functions,
symbolic expressions, internal or external simulators, user-defined functions.
Adaptive Back-propagation, Sparse optimization, Simplex,
Conjugate Gradient, Quasi-Newton, Huber-Quasi-Newton, Auto-Pilot, Genetic
Algorithm, Simulated Annealing
Relevance to the RF/MicrowaveCommunity:
One of the keys to achieve manufacturability-driven design
and time-to-market is to use efficient and reliable CAD tools. Components,
circuits and systems must be represented by models, as a pre-requisite
of any computer-aided design. The availability, accuracy, flexibility of
models, and the speed of model computation affect largely the effectiveness
at all stages of computer-aided design.
NeuroModeler: Build and use neural network models
for RF and microwave modeling, simulation and optimization, an unconventional
technology, with surprising answers to some of the toughest problems in
RF and microwave computer-aided engineering. Get speed AND accuracy AND
flexibility of neural model. Pursue re-usability of the modeling technology
for today's AND tomorrow's devices and circuits.
Use NeuroModeler to finish your design sooner than otherwise,
to let you focus on immediately getting a model rather than learning component
theory, to let you start working with any model you need rather than having
to wait and beg for model availability from tools/vendors/manufacturers.
Use NeuroModeler to enhance the "bottom line" (i.e., models)
of all CAD tools to help reducing microwave design cycle and time-to-market.
Developed by one of the world's leaders in this new technology,
NeuroModeler is the first and a unique software tool of its kind for RF
and microwave designers. With NeuroModeler, you can uncover the myth, and
bypass the hardships that a non-neural net expert typically encounters
when attempting this new technology. Neural Model Technology has never
been so reachable, and so connected to RF/microwave designers until today.
For further information on NeuroModeler and NeuroADS, please contact:
Professor Q.J. Zhang
Department of Electronics, Carleton University, Ottawa, Ontario, Canada K1S 5B6