CN111695230B - A Neural Network Space Mapping Multiphysics Modeling Method for Microwave Passive Devices - Google Patents

Abstract

The invention belongs to the field of modeling of microwave circuits and devices, and provides a neural network space mapping multi-physical modeling method for microwave passive devices. The method adds a mapping network at the input end and the output end of an empirical model, the multi-physical domain input variable of the passive device is mapped to an electromagnetic domain coarse model after being input into a neural network, and the output of the electromagnetic domain coarse model is matched with device modeling data after being further optimized through output mapping. The invention improves modeling precision, reduces modeling data volume, shortens design period, and provides efficient and accurate prediction for multiple physical characteristics of devices.

Description

A neural network spatial mapping multi-physics modeling method for microwave passive devices

Technical Field

The invention belongs to the field of microwave circuit and device modeling, and relates to a neural network space mapping multi-physics modeling method for microwave passive devices.

Background Art

At present, the market has higher and higher requirements for the performance of passive devices, and the application scenarios are becoming more and more complex. In addition to being affected by their own structure and material parameters, the performance of passive devices will also be affected by multiple physical parameters such as temperature, humidity, and stress changes. These changes in multiple physical parameters will cause the electromagnetic parameters of the device to drift and deviate, resulting in reduced device efficiency and poor working stability, affecting the performance of the device. The interaction between multiple physical domains is crucial for accurate system analysis. Passive device design involves electromagnetic analysis and the influence of multiple physical domains such as thermodynamics and electrostatic fields. Introducing multiple physical parameters when modeling device performance and establishing a multi-physical model of the device can accurately describe the multi-physical performance of microwave devices.

Neural network space mapping technology is a kind of knowledge-based neural network, which can improve the accuracy of existing empirical analysis models. The neural network space mapping model not only has the advantages of fast operation speed and good compatibility of the coarse model, but also improves the accuracy of the coarse model. At present, the research results of multi-physics modeling methods for microwave passive devices have application limitations when the multi-physics characteristics of the device are complex or the coarse model is not accurate enough. These results cannot be directly applied to the modeling of passive devices containing multi-physics characteristics. Although the existing device modeling technology is relatively mature, many modeling methods cannot meet the requirements of high accuracy and high speed at the same time, and there is still a certain gap between the designed model and the actual measurement value. The research on multi-physics modeling methods of passive devices still needs to be studied.

Therefore, the purpose of the present invention is to improve the modeling accuracy, reduce the amount of modeling data, shorten the design cycle, and provide efficient and accurate prediction of the multi-physical characteristics of the device by proposing a neural network space mapping multi-physics modeling method for microwave passive devices.

Summary of the invention

The purpose of the present invention is to overcome the shortcomings of the prior art and propose a neural network space mapping multi-physics modeling method for microwave passive devices. The method adds a mapping network to the input and output ends of the coarse model, and the multi-physics domain input variables of the passive device are applied to the electromagnetic domain coarse model after input mapping, and the output of the electromagnetic domain coarse model is further optimized after output mapping to match the multi-physics output of the device.

A neural network spatial mapping multi-physics modeling method for microwave passive devices comprises the following steps:

Step 1: Select and define the data range of the geometric parameters x _g of the multi-physics model, the multi-physics parameters x _m of the multi-physics model and the frequency f of the multi-physics model, and perform multi-physics simulation to generate training and test samples of the multi-physics model;

Step 2: Determine and define the data range of the geometric parameters _xgc and the frequency _fc of the electromagnetic domain coarse model according to step 1, and perform electromagnetic simulation to generate training and test samples of the coarse model. In order to ensure the accuracy of the overall multi-physics model, the range of the geometric parameters _xgc of the electromagnetic domain coarse model should be slightly larger than the geometric parameters _xg in the overall multi-physics model;

Step 3: Establish a coarse model, use the neural network technology to train the coarse model with the training data obtained in step 2, and test the coarse model with the test data obtained in step 2, until the coarse model training error and test accuracy meet the requirements, and the internal weight value w ^* of the coarse model is fixed;

优化输出映射的内部权重变量w₂至

实现x_gc＝x_g、f_c＝f和y＝y_c，保证加入映射网络后不降低整个模型的精度；Step 4: Initialize the input mapping network and the output mapping network to train and optimize the internal weight variables _w1 to w2 of the input mapping.

Optimize the internal weight variable _w2 of the output mapping to

Realize _xgc = _xg , _fc =f and y= _yc , and ensure that the accuracy of the whole model is not reduced after adding the mapping network;

至

实现多物理域输入信号和电磁域输入信号之间的映射；Step 5: Use the coarse model obtained in step 2 and the input mapping unit mapping obtained in step 4 to build a preliminary neural network space mapping multi-physics model, use the multi-physics training data obtained in step 1 to train the preliminary neural network space mapping model, and optimize the weight parameters of the input mapping module

to

Realize the mapping between multiple physical domain input signals and electromagnetic domain input signals;

用步骤1所得的多物理训练数据训练输出映射模块的单位映射，优化输出映射模块的权重参数

至

使粗模型的输出y_c与被建模器件的响应y一致；Step 6: Add the output mapping unit mapping obtained in step 4 to the preliminary neural network space mapping multi-physics model built in step 5, and fix

Use the multi-physics training data obtained in step 1 to train the unit mapping of the output mapping module and optimize the weight parameters of the output mapping module.

to

Make the output _yc of the coarse model consistent with the response y of the device being modeled;

至

调整输出神经网络的权重值

至

直至建立的多物理模型能准确表示无源器件的特性。Step 7: Use the multi-physics training data obtained in step 1 to further train the entire multi-physics model and adjust the weight values of the input neural network.

to

Adjust the weight values of the output neural network

to

Until the established multi-physics model can accurately represent the characteristics of passive devices.

In step 3 of the invention, the rough model formula after training is

y _c =g _ANN (x _gc , f _c , w ^* ) (1)

Where g _ANN (·) is the coarse model neural network mapping formula, and w ^* represents the vector of all weight parameters contained in the coarse model network.

In step 5 of the present invention, the trained input mapping network formula is

表示在该映射网络中包含的所有权重参数的向量。Where f _ANN1 (·) is the input neural network mapping formula, x _g and x _m are the inputs of the input mapping, x _gc and f _c are the outputs of the input mapping,

A vector representing all weight parameters contained in this mapping network.

In step 6 of the present invention, the trained output mapping network formula is

表示该映射神经网络所有内部权重变量。Where f _ANN2 (·) is the output mapping neural network formula, which represents the relationship between y and y _c .

Represents all internal weight variables of the mapping neural network.

The neural network spatial mapping multi-physics modeling method proposed in the present invention not only does not require the internal structure information of microwave passive devices, but also has high model accuracy, strong robustness, less modeling data required, and short modeling time. When the multi-physics characteristics of microwave devices are complex or the coarse model has low accuracy, the two mapping networks adjust each other, and the model can accurately reflect the multi-physics characteristics of the device.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 is a structural block diagram of the present invention;

FIG2 is a flow chart of multi-physics modeling of microwave passive devices according to an embodiment of the present invention;

FIG3 is a comparison diagram of sample data and model output characteristic curves of an embodiment of the invention;

FIG. 4 is a comparison diagram of sample data and model output characteristic curves of an embodiment of the invention.

DETAILED DESCRIPTION

To make the objectives, technical solutions and advantages of the present invention more clear, embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

As shown in FIG2 , in a neural network spatial mapping multi-physics modeling method for microwave passive devices of the present invention, sample data for model training must first be obtained. For a coarse model, let x _gc represent the geometric parameters of the coarse model (electromagnetic domain), f _c represent the frequency of the coarse model, and the input sample of the coarse model is x _c = [x _gc , f _c ] ^T . Let x _g represent the geometric parameters of the multi-physics model. For multi-physics problems, not only geometric parameters x _g are included, but also other physical parameters, x _m is defined as other physical domain parameters, and frequency parameters are defined as f. The input sample of the multi-physics model is x = [x _g , x _m , f] ^T , and the output sample data is S parameters, that is, x _c = [x _gc , f _c ] ^T . The sample data can be obtained by actual measurement of the device or simulation software.

As shown in Figure 2, the model adopts a three-stage training method. The first stage trains the electromagnetic domain coarse model. After the electromagnetic domain coarse model training is completed, the weights in the coarse model are fixed and can be used to represent the electromagnetic response of the device and are ready to be used as prior knowledge for the development of multi-physics models. The second stage initializes the training of the two mapping networks. The purpose of initializing the unit mapping is to provide good initial values for the mapping neural networks before training them. The third stage is multi-physics domain training. The training data in this step are the input and output samples of the multi-physics model. After the three-stage training process is completed, the established multi-physics model can accurately represent the characteristics of the microwave device.

The structure of the present invention mainly consists of three parts: input mapping, output mapping and coarse model. The model shown in Figure 1 is constructed, and the coarse model in Figure 1 is trained with electromagnetic domain coarse model sample data, and the weight parameter w ^* of the coarse model is adjusted until the coarse model training error and test accuracy meet the requirements and w ^* is fixed.

使[x_g，f]^T＝[x_gc，f_c]^T；调整输出映射网络中的权重值w₂至

使y＝y_c。Build the neural network structure shown in Figure 1 and set the input and output mapping network structures. The input mapping network uses a 3-layer perceptron structure, the input signal is [x _g , x _m , f] ^T , and the output signal is [x _gc , f _c ] ^T . The output mapping network uses a 3-layer perceptron structure, the input signal is y _c , and the output signal is y . To ensure that loading the input mapping network does not reduce the accuracy of the coarse model, the input mapping network and the output mapping network are normalized. Adjust the weight value w ₁ in the input mapping network to

Let [x _g , f] ^T = [x _gc , f _c ] ^T ; adjust the weight value w ₂ in the output mapping network to

Let y = y _c .

至

将多物理域参数映射到电磁域，多物理建模问题转化为电磁场建模问题。若测试误差不满足精度要求，则继续用训练数据训练或者调整输入映射网络结构，改变隐含层神经元的个数，重新训练。若测试误差满足精度要求，则停止训练，否则进行输出映射网络调整。The rough model and input mapping unit mapping build a preliminary neural network space mapping multi-physics model, use multi-physics sample data to train the unit mapping of the input mapping module, and optimize the weight parameters of the input mapping module

to

Map the multi-physics domain parameters to the electromagnetic domain, and transform the multi-physics modeling problem into the electromagnetic field modeling problem. If the test error does not meet the accuracy requirements, continue to train with the training data or adjust the input mapping network structure, change the number of hidden layer neurons, and retrain. If the test error meets the accuracy requirements, stop training, otherwise adjust the output mapping network.

用多物理样本数据训练输出映射模块的单位映射，优化输出映射模块的权重参数

至

对粗模型输出信号做调整，缩小模型最终输出与被建模器件输出特性差异。若测试误差不满足精度要求，则继续用训练数据训练或者调整输出映射网络结构，改变隐含层神经元的个数，重新训练。若测试误差满足精度要求，则停止训练，否则微调

至

微调

至

直至满足误差要求。Added output mapping unit mapping on preliminary neural network space mapping multi-physics model, fixed

Use multi-physics sample data to train the unit mapping of the output mapping module and optimize the weight parameters of the output mapping module

to

Adjust the output signal of the coarse model to reduce the difference between the final output of the model and the output characteristics of the modeled device. If the test error does not meet the accuracy requirements, continue to train with training data or adjust the output mapping network structure, change the number of neurons in the hidden layer, and retrain. If the test error meets the accuracy requirements, stop training, otherwise fine-tune

to

Fine-tuning

to

Until the error requirement is met.

3 and 4 are comparison diagrams of the model output characteristic curve established by the modeling method of the present invention and the sample data. It can be seen that the output curve of the model is consistent with the sample data.

Claims (4)

1. A neural network space mapping multi-physical modeling method for microwave passive devices comprises the following steps:

step 1: selecting and defining geometric parameters x of multiple physical models _g Multiple physical parameters x of multiple physical models _m And the data range of the frequency f of the multiple physical models, and performing multiple physical simulations to generate training and testing samples of the multiple physical models;

step 2: determining and defining the geometric parameter x of the electromagnetic domain coarse model according to step 1 _gc And electromagnetic domain coarse model frequency f _c Training and testing samples of the electromagnetic domain coarse model are generated by electromagnetic simulation, and in order to ensure the accuracy of the whole multi-physical model, the geometric parameter x of the electromagnetic domain coarse model is calculated _gc The range is larger than the geometric parameter x in the integral multi-physical model _g ；

Step 3: establishing a coarse model, training the coarse model by using the training data obtained in the step 2 by using a neural network technology, and testing the coarse model by using the test data obtained in the step 2 until the training error and the test precision of the coarse model meet the requirements, wherein the weight value w in the coarse model is the weight value w in the coarse model ^* Fixing;

step 4: initializing the input mapping network and the output mapping network, and optimizing the internal weight variable w of the input mapping ₁ To the point of

Optimizing the internal weight variable w of the output map ₂ To->

Realization of x _gc ＝x _g 、f _c =f and y=y _c The accuracy of the whole model is not reduced after the mapping network is added;

step 5: building a preliminary neural network space mapping multi-physical model by using the rough model obtained in the step 2 and the input mapping unit mapping obtained in the step 4, training the preliminary neural network space mapping model by using the multi-physical training data obtained in the step 1, and optimizing weight parameters of the input mapping module

To->

Mapping between the multi-physical domain input signals and the electromagnetic domain input signals is realized;

step 6: adding the output mapping unit mapping obtained in the step 4 on the preliminary neural network space mapping multi-physical model built in the step 5, and fixing the mapping unit mapping

Training the unit mapping of the output mapping module by using the multi-physical training data obtained in the step 1, and optimizing the weight parameter of the output mapping module>

To->

Let the output y of the coarse model _c Consistent with the response y of the modeled device;

step (a)7: further training the whole multi-physical model by using the multi-physical training data obtained in the step 1, and simultaneously adjusting the weight value of the input neural network

To->

Adjusting the weight value of the output neural network +.>

To->

Until the built multi-physical model can accurately represent the characteristics of the passive device.

2. The neural network spatial mapping multi-physical modeling method for microwave passive devices according to claim 1, wherein in step 3, the trained coarse model formula is:

y _c ＝g _ANN (x _gc ，f _c ，w ^* ) (1)

wherein g _ANN (. Cndot.) is a coarse model neural network mapping formula, w ^* A vector representing all weight parameters contained in the coarse model network.

3. A neural network spatial mapping multi-physical modeling method for microwave passive devices according to claim 1, wherein in step 5, the neural network is used to describe a nonlinear relationship between signals of the coarse model and the modeled device:

wherein f _ANN1 (. Cndot.) is the input neural network mapping formula, x _g And x _m Is the input of the input map, x _gc And f _c Is the output of the input map and,

a vector representing all weight parameters contained in the mapping network.

4. The method of claim 1, wherein in step 6, the neural network is used to describe nonlinear relationships between the output signals of the coarse model and the output signals of the modeled device, respectively:

wherein f _ANN2 (. Cndot.) is the output mapped neural network formula, representing y and y _c The relationship between the two is that,

representing all internal weight variables of the mapped neural network. />

Images