授課教師 : 陳文山學生 : 蘇修賢. neural network theory and eneralities neural network...

授課教師 : 陳文山學生 : 蘇修賢

Neural Network Theory and eneralities

Neural Network Applications in Microwave Modeling and Design

Conclusion

Neural Network Theory and Generalities

Basics of Neural Networks

General Neural Network Modeling Approach

Neural Networks for Inverse Modeling Problem

1. A typical neural network structure comprises two types of basic components,the processing elements and the interconnections between them.

2. The processing elements are called neurons and the connections between the neurons are known as links or synapses.

3. Every link has a corresponding weight parameter associated with it. Each neuron receives stimuli from other neurons connected to it, processes the information, and produces an output.

4. Multilayer perceptron (MLP) is a popularly used neural network structure. In the MLP neural network,


11 22 mm

11 3322 qq

11 3322 nn

x1 x2 x3 xn

…

…

…

y1 y2 ym

Layer 3(Output Layer)

Layer 2(Hidden Layer)

Layer 1(Input Layer)

Figure 1. A three-layer perceptron neural network structure with an input layer, a hidden layer, and an output layer.

1. To develop a neural network model, we first define the input and output variables of a device or a structure. We then generate IO data using EM simulation, physics-based simulation, or measurement.

2. The generated data, training data, are used to train the neural network. Once the model is trained, it can be incorporated into a circuit simulator for fast and accurate simulation and optimization.

3. This allows circuit-level simulation speed with EM-level accuracy. This process is illustrated in Figure 2. For a given component,


資料來源 : H. Kabir, Z. Lei, Y. Ming, P. Aaen, J. Wood, and Q. J. Zhang, “Smart Modeling of Microwave” Devices IEEE Microwave Magazine , Vol. 11 (2009 ) 105-108

Figure 2. Illustration of fast circuit optimization where

a spiral inductor component is represented by a neural network model, avoiding repeated EM simulation when geometrical parameters are changed.

1. To train the neural network inverse model, we swap the generated data so that electrical parameters become training data for neural network inputs and geometric parameters become training data for neural network outputs.

2. Using these data, the trained neural network becomes a direct inverse model. The model is then used to obtain values of geometric design variables from an electrical parameter in a single model evaluation.

3. Unlike the forward model in which the input to output mapping (from geometric parameter to electrical parameter) is usually a one-to-one mapping, the inverse model often encounters the problem of nonunique solutions.

4. This problem also causes difficulties during training, because the same input values to the inverse model will lead to different values at the output (multivalued solutions).


Development of Neural Network Inverse Models for Waveguide Filter

Waveguide Filter Design Using Inverse Models

Neural Network for Parametric Modeling of a Complete Microwave Filter

Neural Network for Correction of Nonlinear Device Models




Conclusion

1.We develop neural network inverse models for waveguide filter. These inverse models will be used to design filters with the inverse approach.

2. The filter is decomposed into three different modules, each representing a separate filter junction. The three models are the IO iris, the internal coupling iris, and the tuning screws.

3.Figure 3 shows a diagram of a waveguide filter revealing various dimensions of the models. Symbol M12 represents the coupling term between cavity 1 and cavity 2. ther coupling terms are also defined similarly.


Figure 3. Diagram of a circular waveguide filter showing

arious geometrical variables. M12 represents the coupling term between cavity 1 and cavity 2.


CD 0 R

Lr Pv Ph Pin

(a) (b) (c)

Figure 4. Neural network inverse models representing junctions of a waveguide filter. (a) Input-output iris model, (b) internal coupling iris model, and (c) tuning screw model. Symbols CD and vo represent cavity diameter and center frequency, respectively.

1. The filter’s center frequency is 12.155 GHz, bandwidth is 64 MHz and cavity diameter is chosen to be 1.072 in. The normalized ideal coupling values are obtained from coupling matrix synthesis, as shown in Figure 5.

2. Figure 6 presents the response of the tuned filter and compares with the ideal one showing a perfect match between each other.


Neural Inverse Model for Filter

Coupling Matrix Synthesis

Filter Specification

Geometrical Dimensions of Filter

Ideal Coupling Values

Figure 4. Design approach using advanced neural network inverse models.

Figure 5. Comparison of the 6-pole filter response with ideal filter response.

The filter was designed, fabricated, tuned and then measured to obtain the dimensions.


1.A parametric model for a complete filter requires that the model has many geometric variables. In the conventional approach, change of a geometric variable requires EM resimulation of the whole filter.

2.Here, we develop a fast neural network model for this purpose. The geometric variables will be formulated as neural network input neurons.


This type of filter offers significant performance improvement and finds its application in the satellite multiplexers with extremely stringent mass, size, and thermal requirements.

Figure 6. Diagram of a side-coupled filter. (a) Side view. (b) Top view. From


Figure 7. Comparisons of reflection coefficients of a sidecoupled circular waveguide dual-mode filter obtained using the neural network model and the EM model.


1.Another useful application of neural network is to map or repair an existing model to match a new device. This process is called Neuro–Space Mapping (Neuro-SM).

2.The starting point for the Neuro-SM is when the existing/available device model (coarse model) cannot match the data of a new device (fine model).


Figure 8. (a) The physical structure of a HEMT device used for the physics-based device simulator (the fine model). (b) The Neuro-SM HEMT intrinsic nonlinear model.

The coarse model is an existing/available equivalent circuit model. The neural network mapping is incorporated as the controlling functions of the controlled sources.


Figure 9. Comparison between the HEMT device data from the fine model (MINIMOS), the existing model (without mapping), and the Neuro-SM model in the HEMT example . (a) dc and (b) S-parameters at four different bias combinations of gate and drain voltages.


Figure 10. Doherty amplifier module used to generate training data for neural networks for power amplifier behavioral modeling: the transistor package contains two separate transistors, which are configured as a Doherty amplifier .




Conclusion

Neural networks are fast to evaluate, and the neural network formulas are easy to implement into microwave CAD.

The simplicity of adding input neurons or hidden neurons makes neural network flexible in handling functions of different dimensions and of different degree of nonlinearity.

Neural networks are helpful in developing parametric or scalable models for passive and active microwave devices.

Conclusion

關於此濾波器讓我學系到了類神經網路的相關知識。

此濾波器在類神經網路之應用，有著不錯的增益。

心得

授課教師 : 陳文山 學生 : 蘇修賢. neural network theory and eneralities neural network...

Documents

授課教師 : 陳文山學生 : 蘇修賢. neural network theory and eneralities neural network...