阅读说明：本技术 一种用于设计高性能闪耀光栅结构的预测网络的构建方法 (Construction method of prediction network for designing high-performance blazed grating structure ) 是由 何宗罡 骆文瑾 蒯犇 夏炜炜 曾祥华 于 2021-07-14 设计创作，主要内容包括：本案涉及一种用于设计高性能闪耀光栅结构的预测网络的构建方法,其步骤为：利用时域有限差分法计算出目标闪耀光栅远场模式,在模式的中间切割一定像素作为训练数据,并放入一个由编码器和解码器两部分组成的自动编码器,经过数次迭代训练后,获得训练完成的隐向量和解码器；再将光栅结构参数和隐向量放入一个前馈神经网络中,经过数次迭代训练,训练完成后将之前得到的解码器接在该神经网络之后,即可作为最终的预测网络使用。本发明提出的预测网络模型在实际测试中有95％以上的精度,同时将闪耀光栅耦合器仿真速度缩小到毫秒量级,相比传统计算方法提高了数万倍,可以极大地提高相关设计的效率,因而具有广泛的应用价值和应用前景。(The scheme relates to a construction method of a prediction network for designing a high-performance blazed grating structure, which comprises the following steps: calculating a far-field pattern of a target blazed grating by using a time domain finite difference method, cutting a certain pixel in the middle of the pattern to be used as training data, putting the training data into an automatic encoder consisting of an encoder and a decoder, and performing iterative training for a plurality of times to obtain a trained hidden vector and the decoder; and then, placing the grating structure parameters and the hidden vectors into a feedforward neural network, performing iterative training for a plurality of times, and connecting the decoder obtained before after the training is finished after the neural network, so that the decoder can be used as a final prediction network. The prediction network model provided by the invention has the precision of more than 95% in actual test, and simultaneously reduces the simulation speed of the blazed grating coupler to millisecond magnitude, which is improved by tens of thousands of times compared with the traditional calculation method, and can greatly improve the efficiency of related design, thereby having wide application value and application prospect.)

1. A method of constructing a prediction network for designing a high performance blazed grating structure, comprising the steps of:

1) calculating the blazed grating far-field modes of 768 different etching parameters by using a time domain finite difference method;

2) cutting 51 × 200 pixels in the middle of the blazed grating far-field pattern, and unfolding the pixels into original data with 10200 dimensions;

3) using the original data as an input vector y_iPutting into an automatic encoder, performing 1000 times of iterative training to obtain output vectorInfinite proximity input vector y_iObtaining a final hidden vector z and a trained decoder;

4) the structural parameters alpha, l, d and w of the grating₀、w₁、w₂、w₃8 parameters of incident light wavelength lambda are used as input vectors and put into a feedforward neural network, the final implicit vector z corresponding to the same far-field mode obtained in the step 3) is used as a target, 1000 times of iterative training are carried out, the output vector y is infinitely close to the final implicit vector z, and the trained neural network is obtained;

5) connecting the decoder trained in 3) to the neural network trained in 4), and obtaining a new neural network which is the final prediction network.

2. A method of constructing a predictive network for designing a high performance blazed grating structure as claimed in claim 1, wherein said automatic encoder is composed of two parts, an encoder and an initial decoder; the encoder is used for inputting a vector y_iEncoding to obtain a 100-dimensional hidden vector z, and reconstructing the hidden vector z into an output vector by the initial decoderMake the output vectorInfinite proximity input vector y_iAnd obtaining the final hidden vector z.

3. A method of constructing a predictive network for designing a high performance blazed grating structure as claimed in claim 1, wherein the loss function expression in the training process in the automatic encoder is:

where m represents the vector dimension, the MSE in the above equation is infinitely close to 0.

4. A method of constructing a prediction network for designing a high performance blazed grating structure as defined in claim 1, wherein an input layer, an output layer and a plurality of hidden layers are fixed in each of the automatic encoder and the feedforward neural network; each layer consists of minimum unit neurons; the parameters of the neuron are: input vector x ═ { x₁，x₂，...，x_nThe weight vector a ═ a }₁，a₂，...，a_n-bias amount b, -output value y, -activation function σ (y), expressed as follows:

where e is a natural constant.

5. A method of constructing a prediction network for designing a high performance blazed grating structure according to claim 4, wherein the number of neurons of the input layer of the automatic encoder is 10200; the number of neurons in the output layer is 10200; the number of hidden layers is 5, and the number of neurons in each layer is 2000, 2000, 1000, 2000 and 2000 respectively.

6. A method of constructing a prediction network for designing a high performance blazed grating structure as defined in claim 4, wherein the number of neurons of the input layer of the feedforward neural network is 8; the number of neurons in the output layer was 1000; the number of hidden layers is 3, and the number of neurons in each layer is 100, 100 and 100 respectively.

7. A method for constructing a prediction network for designing a high performance blazed grating structure as defined in claim 6, wherein a gradient descent algorithm is used in the training process of the feedforward neural network, a parameter θ of each layer is updated according to the gradient of the loss function L, and the formula of the loss function and the parameter θ is as follows:

wherein, the superscript n represents the last layer of the feedforward neural network, namely the nth layer, the superscript i represents the ith layer of the feedforward neural network, and the value of the learning rate eta is 0.01.

Technical Field

The invention belongs to the field of computer aided design, and particularly relates to a neural network for designing a high-performance blazed grating structure.

Background

The grating coupler is one of the most concerned coupling device types at present, and has been widely applied to the fields of microelectronic circuits, optical networks, photoelectric devices and the like. Although the conventional periodic grating structure has been widely applied to photonic design, there is still a great limitation in practical applications, for example, in the process of multi-slit diffraction, the light intensity of the periodic grating is mainly concentrated in the zero order, and most of the diffracted light in the zero order direction cannot satisfy the transmission condition of the guided mode. The blazed grating coupler simulates a stepped blazed grating by two or more periodic non-etched parts, can effectively overcome a plurality of problems encountered by the traditional periodic grating structure, and has the potential of obtaining high light intensity at a plurality of angles. However, the blazed grating coupler design needs to go through a difficult modeling and electromagnetic simulation process, and the time consumption for predicting the far-field result is huge.

Various solutions have been proposed for various deficiencies in the design of blazed grating couplers, such as calculation using a time-domain finite difference method, use of a machine learning method, and the like; however, little attention has been paid to the use of automatic encoders to train neural networks as an important way to improve the efficiency of grating coupler designs.

Disclosure of Invention

The invention aims to provide a prediction network which predicts a far-field spectrum by using a neural network, synthesizes a far-field mode, effectively predicts a parameter value corresponding to the far-field mode and reduces the electromagnetic simulation speed of an arc grating coupler to millisecond order.

In order to achieve the purpose, the invention provides the following technical scheme:

a construction method of a prediction network for designing a high-performance blazed grating structure comprises the following steps:

1) calculating the blazed grating far-field modes of 768 different etching parameters by using a time domain finite difference method;

2) cutting 51 × 200 pixels in the middle of the blazed grating far-field pattern, and unfolding the pixels into original data with 10200 dimensions;

3) using the original data as an input vector y_iPutting into an automatic encoder, performing 1000 times of iterative training to obtain output vectorInfinite proximity input vector y_iObtaining a final hidden vector z and a trained decoder;

4) the structural parameters alpha, l, d and w of the grating₀、w₁、w₂、w₃8 parameters of incident light wavelength lambda are used as input vectors and put into a feedforward neural network, the final implicit vector z corresponding to the same far-field mode obtained in the step 3) is used as a target, 1000 times of iterative training are carried out, the output vector y is infinitely close to the final implicit vector z, and the trained neural network is obtained;

5) connecting the decoder trained in 3) to the neural network trained in 4), and obtaining a new neural network which is the final prediction network.

In the above technical solution, the automatic Encoder is composed of an Encoder (Encoder) and an initial Decoder (Decoder); the encoder is used for inputting a vector y_iEncoding to obtain a 100-dimensional hidden vector z, and reconstructing the hidden vector z into an output vector by the initial decoderMake the output vectorInfinite proximity input vector y_iAnd obtaining the final hidden vector z. In the technical schemeIn the above, the loss function expression in the training process in the automatic encoder is:

where m represents the vector dimension, the MSE in the above equation is infinitely close to 0.

In the above technical solution, an input layer, an output layer and a plurality of hidden layers are fixed in the automatic encoder and the feedforward neural network; each layer consists of the smallest unit neurons. The parameters of the neuron are: input vector x ═ { x₁，x₂，...，x_nThe weight vector a ═ a }₁，a₂，...，a_n-bias amount b, -output value y, -activation function σ (y), the expression of which is as follows:

where e is a natural constant.

In the above technical solution, the number of neurons in the input layer of the automatic encoder is 10200; the number of neurons in the output layer is 10200; the number of hidden layers is 5, and the number of neurons in each layer is 2000, 2000, 1000, 2000 and 2000 respectively.

In the above technical solution, the number of neurons in the input layer of the feedforward neural network is 8; the number of neurons in the output layer was 1000; the number of hidden layers is 3, and the number of neurons in each layer is 100, 100 and 100 respectively.

In the above technical solution, a gradient descent algorithm is used in the feedforward neural network training process, and the parameter θ of each layer is updated according to the gradient of the loss function L, where the formula of the loss function and the parameter θ is:

wherein, the superscript n represents the last layer of the feedforward neural network, namely the nth layer, the superscript i represents the ith layer of the feedforward neural network, and the value of the learning rate eta is 0.01.

Compared with the prior art, the invention has the beneficial effects that:

compared with the traditional electromagnetic simulation process, the neural network for predicting the far-field mode has the precision of more than 95% in the actual test, simultaneously reduces the simulation speed of the blazed grating coupler to millisecond magnitude, improves the simulation speed by tens of thousands of times compared with the traditional calculation method, can greatly improve the efficiency of related reverse design, and has wide application value and application prospect.

Drawings

Fig. 1 is a structural design diagram of a blazed grating according to embodiment 1 of the present invention.

Figure 2 is a graph of the far field intensity of the gratings before and after optimization according to example 1 of the present invention.

Fig. 3 is a schematic diagram of an automatic encoder according to embodiment 1 of the present invention.

Fig. 4 is a schematic diagram of the construction of a prediction network according to embodiment 1 of the present invention.

Fig. 5 is a comparison graph of the prediction of the predicted network and the actual results of embodiment 1 of the present invention.

Detailed Description

The technical solutions of the present invention will be described clearly and completely with reference to the accompanying drawings, and it should be understood that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

In addition, the technical features involved in the different embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

Example 1:

step 1, calculating the far-field pattern of the blazed grating in fig. 1 using FDTD software from the company Lumerical, and varying the wavelengths of light λ and w₀、w₁、w₂、w₃Four structural parameters collect training data, wherein:

lambda is [1.50,1.55,1.60] micron

w₀＝[0.2,0.3,0.4,0.5]Micron meter

w₁＝[0.5,0.6,0.7,0.8]Micron meter

w₂＝[0.2,0.3,0.4,0.5]Micron meter

w₃＝[0.5,0.6,0.7,0.8]And (3) micron.

In FIG. 1, incident represents incident light, λ represents wavelength of light, α represents fan angle, l represents length of grating before groove etching, d represents etching depth, and w represents etching depth₀、w₁、w₂、w₃Representing the groove width.

Step 2, cutting 51x200 pixels in the middle of the blazed grating far-field pattern, and unfolding the pixels into original data of 10200 dimensions; as shown in FIG. 3, the raw data is taken as an input vector y_iPutting into an automatic encoder, performing 1000 times of iterative training to obtain output vectorInfinite proximity input vector y_iAnd obtaining a final hidden vector z (latent vector) and a trained decoder.

The automatic Encoder consists of two parts, namely an Encoder (Encoder) and an initial Decoder (Decode), and original data (input vector y)_i) After passing through the encoder, the encoded vector is a 100-dimensional hidden vector z, which is reconstructed by an initial decoder into training data (output vector)) (ii) a Training data (output vector)) Infinite approximation to the original data (input vector y)_i) And obtaining the final hidden vector z. The automatic encoder has been trainedThe loss function in the equation is expressed as:

where m represents the vector dimension, the MSE in the above equation is infinitely close to 0.

Step 3, the structural parameters alpha, l, d and w of the grating₀、w₁、w₂、w₃And 8 parameters of the wavelength lambda of the incident light are used as input vectorsPutting the neural network into a feedforward neural network, taking the final implicit vector z of the same far-field mode obtained in the step 2 as a target, and performing 1000 times of iterative training to enable the output vector y to be infinitely close to the final implicit vector z to obtain the trained neural network; in the feedforward neural network training process, a gradient descent algorithm is used, the parameter theta of each layer is updated according to the gradient of the loss function L, and the formula of the loss function and the parameter theta is as follows:

wherein, the superscript n represents the last layer of the feedforward neural network, namely the nth layer, the superscript i represents the ith layer of the feedforward neural network, and the value of the learning rate eta is 0.01.

In the above embodiment, an input layer, an output layer and several hidden layers are fixed in the automatic encoder and the feedforward neural network; each layer consists of minimum unit neurons; the parameters of the neurons were: input vector x ═ { x₁，x₂，...，x_nThe weight vector a ═ a }₁，a₂，...，a_nH, bias b, output value y, activation function σ (y), activation function expressionThe following were used:

where e is a natural constant.

The number of neurons in the input layer of the automatic encoder is 10200; the number of neurons in the output layer is 10200; the number of hidden layers is 5, and the number of neurons in each layer is 2000, 2000, 1000, 2000 and 2000 respectively.

The number of neurons of an input layer of the feedforward neural network is 8; the number of neurons in the output layer was 1000; the number of hidden layers is 3, and the number of neurons in each layer is 100, 100 and 100 respectively.

And 4) connecting the decoder trained in the step 2) to the neural network trained in the step 3), and then forming a new neural network which is the final prediction network (fig. 4).

The principle of the invention is as follows: the characteristic that learning is carried out by training sample iterative parameters by utilizing a neural network and the characteristic that an automatic encoder encodes an input vector into a low-dimensional hidden vector and then reconstructs the hidden vector back to the input vector by a decoder. By carrying out iterative training on the automatic encoder and using the trained decoder part as a prediction network, a novel scheme for quickly obtaining an algorithm of a grating far-field mode is designed, and the time required by single calculation is greatly reduced.

Fig. 2 is a graph comparing the far field intensity of the gratings (a-c) before optimization and the gratings (d-f) after optimization, and it can be seen that the far field intensity of the gratings after optimization is about 20 times higher than that of the gratings before optimization at three wavelengths of 1500nm, 1550nm and 1600 nm.

Fig. 5 is a comparison graph of actual results of the neural network prediction and FDTD calculation according to embodiment 1 of the present invention, and it can be seen from the graph that compared with the actual results, the prediction results achieve a very high accuracy (more than 95%), and the prediction network successfully constructed can improve the single calculation to millisecond level.