A photonic crystal fiber optical property prediction method based on transfer learning

Abstract

The present application is directed to the problems of low efficiency and high labor cost in the field of numerical simulation calculation of optical fiber, and proposes a high-efficiency and accurate photonic crystal fiber optical property prediction method based on transfer learning. The method converts the cross-section structure of the optical fiber into a color picture, and the three channels of the color picture can load more material information (refractive index, absorption rate, nonlinear coefficient, etc.). Convolutional neural network is used to extract features of the picture, and an efficient and accurate mapping network relationship between the two-dimensional structure of the optical fiber and the optical property is established through regression calculation. On this basis, using the transfer learning method, a small amount of different types of photonic crystal fiber data set is used to update part of the network parameters of the original network model, and the mapping network relationship for the new optical fiber structure and optical property is quickly established. The present application has fast calculation speed and high accuracy, and the transfer learning method has wide application range and can be transferred to different types of photonic crystal fibers, requiring small amount of data.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of optical fibers, and relates to a photonic crystal fiber optical property prediction method based on transfer learning. The method performs feature extraction on a photonic crystal fiber cross-section picture through a pre-trained convolutional neural network, and accurately predicts the optical properties of the fiber through regression calculation. Then, the method uses a small amount of optical data of other types of photonic crystal fibers to update the parameters of the trained network model. The updated network can accurately predict the optical properties of new photonic crystal fibers. BACKGROUND

[0002] With the maturity and development of optical fiber sensing and communication technology, more new functional fiber photonic structures and optical devices need to be developed. Currently, the design and development of fiber photonic structures mainly rely on the experience of researchers, who manually modify and adjust fiber structure parameters, and then solve various optical properties of the fiber structure by combining various numerical simulation methods such as dense calculation. Finally, the fiber photonic structure is repeatedly tried and adjusted in order to obtain ideal optical properties. The design and solution process is heavily dependent on the experience of researchers, which is not only troublesome but also inefficient. At the same time, numerical simulation methods (such as plane expansion, finite difference method, finite element method, etc.) have slow convergence, large calculation amount and long calculation time when calculating complex photonic crystal fiber structures, which greatly reduces the progress and efficiency of research.

[0003] In recent years, neural networks and other machine learning technologies have made significant achievements in many fields with their unique advantages and excellent data fitting capabilities, such as medical image recognition, deaf-mute sign language recognition, and multi-language automatic translation. By training a neural network, the potential physical laws between fiber structure and optical characteristics can be found, so that the optical characteristics of the fiber can be quickly and efficiently predicted according to the fiber structure without the need for redundant calculations by traditional numerical methods. However, the current neural network-based fiber optical property prediction network is designed by setting several fixed structure parameter variables, and the fiber structure is determined by optimizing and adjusting the variables, which artificially limits the parameter design space of the fiber structure.

[0004] On the other hand, the trained neural network can only predict the optical properties of a certain type of fiber structure, and it is difficult to accurately predict the optical properties of another type of fiber. Therefore, the currently reported fiber optical property prediction neural network has a very narrow application range and cannot be used to predict the optical properties of different types of fibers. SUMMARY

[0005] The present application aims at the problems existing in the field of optical fiber simulation optimization calculation at present, and proposes a high-efficiency and accurate photonic crystal fiber optical property prediction method based on transfer learning. The method takes the two-dimensional free distribution structure characteristics of the fiber cross section as the parameter space, converts the structure of the fiber cross section into the form of a picture, uses a convolution network to extract the features of the structure picture, and establishes an efficient and accurate mapping network relationship between the two-dimensional structure of the free distribution of the fiber cross section and the optical properties by regression calculation. On this basis, further use the transfer learning method, use a small amount of different types of photonic crystal fiber data set, update part of the network parameters of the original network model, quickly establish the mapping network relationship for the new fiber structure and the optical properties. The transfer learning method has a wide range of application, can be transferred to different types of photonic crystal fibers, has a wide range of adaptation, and needs a small amount of data.

[0006] The purpose of the present application is to accurately predict the optical properties of different photonic crystal fibers by using the trained convolutional neural network, and to realize the prediction of the optical properties of different types of grid arranged photonic crystal fibers by transfer learning. The present application takes the concentric circle grid arranged photonic crystal fiber as an example to establish a prediction neural network, accurately predicts the mode field area, dispersion and effective refractive index of the LP 01 mode in the concentric circle grid fiber by the trained network, and takes the triangular grid arranged photonic crystal fiber as the transfer learning object, updates part of the network parameters by transfer learning using a small amount of data set, and realizes the accurate prediction of the mode field area, dispersion and effective refractive index of the triangular grid fiber mode. The technical scheme is also applicable to the prediction of other optical structure characteristics.

[0007] The technical scheme adopted by the present application includes the following steps:

[0008] 1. Collect the optical properties of the transmission mode in different photonic crystal fiber models of the same grid arrangement, such as mode field area, dispersion and mode effective refractive index, by using the traditional finite element simulation method;

[0009] 2. Select appropriate resolution and picture pixel size, convert the two-dimensional structure of the fiber cross section into the form of a color picture, and the color of different pixel points of the picture represents the different material characteristics (refractive index, absorption rate, etc.) of the corresponding structure, store the structure picture and the corresponding optical properties as a data set;

[0010] 3. Construct a suitable convolutional neural network to extract the features of the input fiber structure picture, and construct a corresponding regression prediction network to predict the optical properties of the fiber picture structure at different wavelengths;

[0011] 4. On the basis of the third step, further construct a suitable regression prediction network to predict the optical properties corresponding to the fiber picture structure at different wavelengths;

[0012] 5. Using the collected data set to train the above two network models and save the model, and then using the test set to test the prediction performance of the two network models, and saving the most suitable prediction network model;

[0013] 6. Based on the trained prediction network model, a small amount of photonic crystal fiber data with different lattices is collected, and the full connection layer parameters in the network model are quickly trained and updated to obtain a new network prediction model, so that the optical properties of photonic crystal fibers with different lattice arrangements can be accurately predicted.

[0014] The application provides a photonic crystal fiber optical property prediction method based on transfer learning, which has the advantages that:

[0015] 1. The input end of the network structure is a two-dimensional structure feature color picture of a free distribution of a fiber cross section, and the photonic crystal fiber structure with arbitrary two-dimensional distribution can be predicted, and the three channels of the color picture can load more material information, such as refractive index, absorption rate, nonlinear coefficient and the like, so that a new idea is provided for more flexible fiber optical property prediction and optimization design.

[0016] 2. The prediction network can not only accurately predict the optical properties of photonic crystal fibers with different structure parameters and the same lattice arrangement, but also can be extended to more different kinds of photonic crystal fiber prediction applications with different lattice arrangements through the method of transfer learning, and has a wider application range. The neural network model and the corresponding transfer learning method only need a small amount of other lattice fiber data, and after updating part of the parameters of the trained prediction network, a prediction network of a new lattice fiber structure can be quickly obtained, the time for collecting data and training a new network is reduced, and the prediction accuracy is also high.

[0017] 3. Compared with the traditional fiber simulation method, the method has an absolute advantage in calculation efficiency and speed. The proposed method only needs 0.02-0.03 seconds to predict the optical properties of a single photonic crystal fiber model, which is improved by orders of magnitude compared with the traditional fiber simulation method. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The application predicts a concentric circular lattice arranged photonic crystal fiber structure, the fiber is composed of pure silicon dioxide, the air holes are arranged in a concentric circular lattice, there is a certain spacing between the air holes in each layer, the parameters such as the number of air hole layers, the diameter, the number and the spacing are random, and the fiber core can support LP01 mode transmission;

[0019] Figure 2The triangular lattice arranged photonic crystal fiber structure schematic diagram involved in the transfer learning method provided by the application. The fiber is composed of pure silicon dioxide, the air holes are arranged in a triangular lattice, there is a certain spacing between the air holes, the air hole layer number, diameter, number, spacing and other parameters are random;

[0020] Figure 3 The two-dimensional structure of the free distribution of the cross section of the photonic crystal fiber is converted into a quarter color picture schematic diagram (wherein the light color part is blue and the dark color part is red);

[0021] Figure 4 The flowchart of the photonic crystal fiber characteristic prediction method based on the neural network provided by the application is shown;

[0022] Figure 5 The convolutional neural network and the full connection neural network structure schematic diagram of the photonic crystal fiber optical characteristic of the concentric circle lattice arrangement of the fixed wavelength which can be predicted by the application (in step 3) are shown, and the network model is marked as A network;

[0023] Figure 6 The full connection neural network structure schematic diagram of the photonic crystal fiber optical characteristic of the concentric circle lattice arrangement of different wavelengths which can be predicted by the application (in step 4) is shown, and the network model is marked as B network;

[0024] Figure 7 The comparison diagram of the predicted value (Predicted Neff) of the mode effective refractive index obtained by testing the A network by the test set and the mode effective refractive index (FEM Neff) collected by the traditional finite element simulation method is shown;

[0025] Figure 8 The comparison diagram of the predicted value (Predicted Dispersion) of the mode dispersion obtained by testing the A network by the test set and the mode dispersion value (FEM Dispersion) collected by the traditional finite element simulation method is shown;

[0026] Figure 9 The comparison diagram of the predicted value (Predicted Aeff) of the mode field area obtained by testing the A network by the test set and the mode field area (FEM Aeff) collected by the traditional finite element simulation method is shown;

[0027] Figure 10 The comparison diagram of the predicted value (Predicted Neff) of the mode effective refractive index Neff at different wavelengths obtained by testing the B network by the test set and the effective refractive index (FEM Neff) collected by the traditional finite element simulation method is shown;

[0028] Figure 11 : The present application uses the test set to test the B network, and a comparison chart of predicted mode dispersion at different wavelengths (Predicted Dispersion) and mode dispersion values (FEM Dispersion) collected by a traditional finite element simulation method is obtained.

[0029] Figure 12 : The present application uses the test set to test the B network, and a comparison chart of predicted mode field area at different wavelengths (Predicted Aeff) and mode field area (FEM Aeff) collected by a traditional finite element simulation method is obtained.

[0030] Figure 13 : The present application uses the transfer learning method to obtain a comparison chart of predicted mode effective refractive index of a triangular lattice arranged photonic crystal fiber at a fixed wavelength (Predicted value) and a true value (here, the data collected by a traditional finite element simulation method is defined as the true value).

[0031] Figure 14 : The present application uses the transfer learning method to obtain a comparison of predicted mode dispersion of a triangular lattice arranged photonic crystal fiber at a fixed wavelength (Predicted value) and a true value.

[0032] Figure 15 : The present application uses the transfer learning method to obtain a comparison of predicted mode field area of a triangular lattice arranged photonic crystal fiber at a fixed wavelength (Predicted value) and a true value.

[0033] Figure 16 : The present application uses the transfer learning method to obtain a comparison of predicted mode effective refractive index of a triangular lattice arranged photonic crystal fiber at different wavelengths (Predicted value) and a true value.

[0034] Figure 17 : The present application uses the transfer learning method to obtain a comparison of predicted mode dispersion of a triangular lattice arranged photonic crystal fiber at different wavelengths (Predicted value) and a true value.

[0035] Figure 18 : The present application uses the transfer learning method to obtain a comparison of predicted mode field area of a triangular lattice arranged photonic crystal fiber at different wavelengths (Predicted value) and a true value. DETAILED DESCRIPTION

[0036] The present application and technical solutions will be further specifically described below in combination with the drawings.

[0037] A photonic crystal fiber optical property prediction method based on transfer learning, taking a concentric circular lattice arranged photonic crystal fiber as an example, a triangular lattice arranged photonic crystal fiber as a transfer object, the mode field area, dispersion and effective refractive index of the base mode of the concentric circular lattice fiber are accurately predicted through the trained network, and the mode field area, dispersion and effective refractive index of the base mode of the triangular lattice fiber are accurately predicted through the proposed transfer learning method.

[0038] The cross-sectional structure of the concentric circular lattice arranged photonic crystal fiber used is shown in Figure 1 The cross-sectional structure of the triangular lattice arranged photonic crystal fiber used as a transfer object is shown in Figure 2 The two types of air hole photonic crystal fibers are composed of pure silica, and the air holes are arranged in concentric circles and triangular lattices, respectively, and there is a certain spacing between the air holes in each layer, and the number of air hole layers, diameter, number, spacing and other parameters are random, and different structural parameters correspond to different optical properties.

[0039] The structural parameters of the concentric circular lattice fiber are randomly changed, and 920 groups of color pictures corresponding to different structural parameters of the concentric circular lattice fiber and the mode field area, dispersion and effective refractive index of the transmission base mode are collected as a data set, wherein each group of data has 9 wavelengths, and the training set and the test set are divided according to 8:2.

[0040] Taking the concentric circular lattice arranged photonic crystal fiber as an example, the two-dimensional structure of the fiber cross section is converted into a color picture, as shown in Figure 3 At this time, red represents pure silica, and blue represents air holes, and the collected picture is 1 / 4 of the original fiber structure, so that the prediction accuracy is not affected, and the training and prediction speed is also accelerated. The size of the example collected picture is 720x720 pixels, Figure 3 Only one of the conversion examples is shown, and the definition and size of the converted color picture can be flexibly adjusted, and the three channels of the color picture can load more material information, such as refractive index, absorption rate, nonlinear coefficient, etc., and the picture pixel size can also be adjusted as needed, providing a new idea for more flexible fiber optical property prediction and optimization design.

[0041] Figure 4 The flowchart of the present application is shown in

[0042] Figure 5 A network structure for predicting optical properties of photonic crystal fiber with concentric circular lattice is shown. The input layer of the convolutional neural network is 720*720*3, corresponding to the size of the collected structure picture. After adjusting the number of convolution kernels, nodes and weights and other hyperparameters of the network for many times, the final convolutional network includes an input layer, 4 convolutional layers, 2 fully connected layers, and an output layer. After all convolutions, a max pooling layer and an activation function are used to extract important structural features, and the training set is used for 500 times of iterative training.

[0043] Figure 6 A B network structure for predicting optical properties of photonic crystal fiber at different wavelengths is shown. It can realize the prediction of optical properties from a specific wavelength to different wavelengths in series with the A network. The B network structure includes an input layer, three hidden layers containing 250, 400, and 50 nodes, and an output layer. The training set is used for 10,000 times of iterative training of the B network. Both networks use the ReLu activation function and the Adam optimizer. After training, the above two network models are saved.

[0044] Figures 7 to 12 The results of testing the two trained neural networks A and B using the test data set are shown. As can be seen from the figure, the difference between the network prediction value and the value calculated by the finite element method is very small, verifying the correctness and accuracy of the prediction network model.

[0045] Next, the transfer learning method is used to update the parameters of the trained A network and B network.

[0046] First, prepare the transfer learning data set. Randomly change the structural parameters of the triangular lattice fiber, collect 160 groups of color pictures corresponding to different structural parameters of the triangular lattice fiber, and support the transmission of the fundamental mode, the mode field area, the dispersion, and the effective refractive index as the data set. Each group of data has 9 wavelengths, and the training set and the test set are divided according to 8:2.

[0047] Next, use the triangular lattice data set to train and update the network parameters. For the A network, keep the convolutional network weights unchanged and update the fully connected layer network weights. After 200 iterations, a new network is obtained for predicting different structural triangular lattice fibers at a fixed wavelength. For the B network, directly use the triangular lattice data set to update the weights of the three fully connected hidden layers. After 200 iterations, a new network is obtained for predicting triangular lattice fibers at different wavelengths.

[0048] Figures 13 to 18The prediction results of the two new networks obtained after transfer learning on the optical characteristics of triangular lattice photonic crystal fibers are shown, and from the figure, it can be seen that the difference between the network prediction value and the numerical value calculated by the finite element method is very small, verifying the correctness and accuracy of the prediction network model, and proving the effectiveness of the transfer learning, and the method can be widely used in various different lattice arranged photonic crystal fibers, and is also suitable for the prediction of other optical structure characteristics.

[0049] The parts of the application not described in detail are common knowledge to those skilled in the art.

Claims

1. A photonic crystal fiber optical property prediction method based on transfer learning, the method comprising the following steps: Step 1: Calculate the optical properties of the transmission mode in different photonic crystal fiber models with the same lattice arrangement using a traditional finite element simulation method; Step 2: Select an appropriate resolution and picture pixel size, convert the two-dimensional structure of the fiber cross-section into a color picture, and store the picture together with the corresponding optical properties as a data set; Step 3: Build a suitable convolutional neural network and fully connected regression prediction neural network A, the convolutional neural network extracts features from the input fiber structure picture, and the fully connected regression prediction neural network predicts the optical properties of different structure fiber pictures at a fixed wavelength; Step 4: On the basis of step 3, further build a suitable regression prediction network B, connect network B and network A in series, and predict the optical properties of fiber picture structures at different wavelengths; Step 5: Use the collected data set to train network A and network B and save the model, then test the prediction performance of the two network models using the test set, and save the most suitable prediction network model; Step 6: Based on the trained prediction network model, use the transfer learning method to collect a small amount of photonic crystal fiber data with different lattices, quickly train and update the fully connected layer parameters in the network model, and obtain a new network prediction model that can accurately predict the optical properties of photonic crystal fibers with different lattice arrangements. 2.The photonic crystal fiber optical property prediction method based on transfer learning according to claim 1, characterized in that: The photonic crystal fibers involved in training the neural network and transfer learning are composed of pure silicon dioxide, and the fiber structure is composed of air holes arranged in different lattices. 3.The photonic crystal fiber optical property prediction method based on transfer learning according to claim 1, characterized in that: In step 2, the two-dimensional structure of the fiber cross-section is converted into a color picture, and the picture color of different pixels represents the different material characteristics of the corresponding structure, and the three channels of the color picture load the material information of the refractive index, absorption rate, and nonlinear coefficient. 4.The method of claim 1, wherein: In step 6, the transfer learning involves updating the fully connected layer parameters of the original network model using a small amount of fiber data set with different lattice arrangements, for network A, the convolutional network weight is kept unchanged, and the fully connected layer network weight is updated, so that it can accurately predict the optical properties of photonic crystal fibers with different lattice arrangements at a fixed wavelength, for network B, the fully connected hidden layer network weight is trained and updated, so that it can accurately predict the optical properties of photonic crystal fibers with different lattice arrangements at different wavelengths.

Images