A method for extracting two-dimensional features of an optical fiber structure based on a convolutional neural network and predicting optical characteristics

Abstract

The present application is aimed at the problems of low efficiency and high labor cost in the current optical fiber structure optimization design method, and proposes a two-dimensional feature extraction and optical property prediction method for optical fiber structure based on convolutional neural network, taking the two-dimensional free distribution structure of the optical fiber cross section as the parameter space, converting the arbitrary free distribution structure of the optical fiber into a two-dimensional numerical matrix, extracting the two-dimensional structure features by using the convolutional neural network, and establishing an efficient and accurate mapping network relationship between the two-dimensional structure of the optical fiber cross section free distribution and the optical properties by regression calculation. The present application takes the multi-core optical fiber as an example to accurately predict the mode field area, dispersion and effective refractive index. The present application has fast calculation speed and high accuracy, avoids the limitations of the existing mapping network based on fixed feature parameters for optical fiber structure optimization design, and can be combined with various optimization algorithms, and is expected to realize the optimization design of high degree of freedom optical fiber structure. The technical scheme is also applicable to the prediction of other optical structures.

Description

Technical Field

[0001] This invention belongs to the field of optical fiber technology and relates to a method for extracting two-dimensional structural features of optical fiber cross-section and predicting optical properties based on convolutional neural networks. The method extracts two-dimensional structural features of optical fiber cross-section through a pre-trained convolutional neural network and accurately predicts the optical properties of optical fiber through regression calculation. Background Technology

[0002] The development of fiber photonics is inseparable from the development and design of various novel functional fiber photonic structures and devices. Currently, the research and development of fiber photonic structures mainly relies on computationally intensive numerical simulations and a trial-and-error framework. This involves using a priori physical mechanisms and various numerical simulations to repeatedly adjust structural characteristic parameters to attempt to obtain ideal optical properties. This structure design optimization based on a trial-and-error framework heavily depends on the researcher's experience and intuition. Because the relationship between optical structure and properties is often very complex and subtle, and various structural parameters can influence and constrain each other, researchers find it difficult to obtain clear optimization and design directions. The design process often involves a large amount of repetitive and redundant work, making the design method time-consuming and laborious. Relying solely on physical intuition and system fine-tuning makes it difficult to achieve significant breakthroughs or improvements, and it is also difficult to find the structural parameters with optimal or limiting optical properties. On the other hand, while computationally intensive numerical simulation methods (such as plane wave expansion, finite difference, and finite element methods) are stable and accurate, they have slow convergence speeds, large computational loads, and low efficiency, further affecting design and development efficiency.

[0003] In recent years, neural networks, as an interdisciplinary field, have demonstrated significant application value in many areas such as semantic recognition, image processing, biochemistry, and micro-nano optics. The powerful analytical and computational capabilities of neural networks can quickly identify complex nonlinear relationships between mapping parameters, providing strong support for the design of complex fiber optic structures requiring multi-parameter optimization. By pre-defining several optical structure feature parameters, neural networks can quickly establish the mapping relationship between optical structure parameters and optical properties, and the calculation speed is several orders of magnitude faster than traditional numerical simulation, which greatly improves the computational efficiency of optical structures [Optics Express, 27(25): 36414-36425(2019); Optics Express, 28(15): 21668-21681(2020); Optics Letters, 46(6): 1454-1457(2021)]. Combined with other optimization algorithms [Optics Express, 28(15): 21971-21981(2020); Photonic Research, 9(6): B247-B252(2021)], it is more conducive to the automatic global optimization and improvement design of optical structures, which is of great significance to the research and development of new optical devices and structures with extreme optical performance.

[0004] However, the reported mapping networks are all based on mapping calculations with fixed feature parameters. Although they can make accurate and fast predictions of optical structures, they have great limitations: they require several feature parameters to be assumed in advance based on intuitive experience (for example, assuming that the optical fiber is a five-ring structure and defining the structural parameters of each ring, but the optimal optical fiber structure may not be a five-ring structure at all). These intuitive assumptions fix the characteristics of the optical fiber structure in advance, which greatly limits the parameter design space of the optical structure and restricts the design ideas, because it is often difficult for us to know the geometric feature parameters of the optimal target structure in advance. Summary of the Invention

[0005] To address the numerous problems existing in the current optical fiber structure optimization design, this invention proposes a two-dimensional feature extraction and optical property prediction method for optical fiber structures based on convolutional neural networks. Using the two-dimensional freely distributed structure of the optical fiber cross-section as the parameter space, the arbitrary freely distributed structure of the optical fiber is converted into two-dimensional matrix parameters. The two-dimensional structural features are extracted using convolutional neural networks, and an efficient and accurate mapping network relationship between the two-dimensional freely distributed structure of the optical fiber cross-section and its optical properties is established through regression calculation.

[0006] The purpose of this invention is to accurately predict the optical properties of different optical fibers using a trained convolutional neural network. Taking a multi-core optical fiber as an example, the cladding material of the multi-core optical fiber is pure silicon dioxide, and the core material is doped silicon dioxide. The core is composed of multiple closely arranged high-refractive-index cores, forming a unified structure that supports supermode transmission with a large mode area. By changing the structural parameters, the optical properties of the different supermodes supported by the core will change. The proposed fiber optic property prediction and analysis method based on convolutional neural networks can quickly and accurately predict and calculate the corresponding fiber optic parameters, including the mode area, dispersion, and effective refractive index of the supported transmission modes, based on the refractive index and geometric size distribution of different materials. This technical solution is also applicable to the prediction of other optical structural properties.

[0007] The technical solution adopted in this invention includes the following steps:

[0008] 1. Calculate the optical characteristics of different modes that can be transmitted in fiber optic models with different structural parameters using traditional finite element simulation methods, such as mode area, dispersion, and effective mode refractive index;

[0009] 2. Select an appropriate sampling interval to convert the free distribution of refractive index of materials in different optical fiber structure cross sections into a two-dimensional data matrix. The value of each point in the data matrix represents the refractive index of the material at the corresponding optical fiber structure.

[0010] 3. Construct a suitable regression prediction network structure using convolutional neural networks;

[0011] 4. Use the collected dataset to train the regression prediction network and save the model;

[0012] 5. Test the performance of the regression prediction network model using a test set;

[0013] 6. Save the most suitable regression prediction network for fast and accurate prediction of the optical properties of different multi-core optical fibers.

[0014] This invention provides a method for extracting two-dimensional structural features and predicting optical properties of optical fiber cross-sections based on convolutional neural networks. Its advantages are:

[0015] 1. In terms of dataset establishment, the two-dimensional structural features of the free distribution of the optical fiber cross section are transformed into a two-dimensional data matrix. The size of the matrix can be flexibly adjusted according to the sampling interval (similar to the pixel and resolution of an image in image processing, which can be freely adjusted as needed). The collected and calculated optical fiber structure is the free distribution of the material refractive index of the entire cross section, which is not limited by fixed feature parameters or size variables, providing ideas and approaches for more flexible optical fiber structure optimization design.

[0016] 2. The numerical definition of the two-dimensional matrix for converting optical fiber cross-sectional structural parameters is very flexible. It can be defined as the two-dimensional distribution of the material refractive index, or as the material absorption rate or other special material properties, which provides convenience for the optimized design of various types of optical fiber structures or special optical fibers.

[0017] 3. Compared with traditional numerical simulation methods, this method has an absolute advantage in computational speed. The time required to calculate the optical characteristics of a single fiber optic structure is between 0.02 and 0.03 seconds, while traditional mathematical physics numerical simulation methods often require several minutes or even hours to complete the calculation for a single fiber optic structure. Therefore, this method offers an order-of-magnitude improvement in speed, making it easier to combine with various intelligent optimization algorithms and facilitating the further utilization of computers for automated optimization design of fiber optic structures.

[0018] 4. Compared with conventional fully connected artificial neural networks for computing optical structures [Optics Express, 27(25): 36414-36425(2019); Optics Express, 28(15): 21668-21681(2020); Optics Letters, 46(6): 1454-1457(2021)], the two-dimensional matrix parameters input by this method represent the two-dimensional structural features of the entire optical fiber cross-section that are freely distributed, effectively avoiding the limitation of pre-assuming size feature parameters on the free design and optimization of optical fiber structures.

[0019] 5. In image feature extraction, the number of feature parameters is reduced by ignoring unimportant features and retaining important parameters. This can greatly reduce the number of training parameters during model training, prevent overfitting, and significantly speed up model training.

[0020] 6. Based on a convolutional neural network, a convolution operation is performed on the two-dimensional numerical matrix of the refractive index distribution of optical fiber material to extract features. Regression calculations are then performed in the final fully connected layer. The predicted optical properties, such as mode area, dispersion, and effective refractive index, show high accuracy compared to traditional numerical simulation methods. The results demonstrate that the present invention is feasible for predicting optical fiber properties based on the two-dimensional freely distributed refractive index of the optical fiber cross-section, and the prediction accuracy is extremely high. Attached Figure Description

[0021] Figure 1 The example optical fiber structure diagram predicted by this invention has multiple fiber cores (5 to 8 fiber cores), each fiber core has a certain spacing between them, each fiber core is at the same distance from the center of the cladding, the cladding is pure silicon dioxide, and multiple fiber cores form a whole, which can support supermode transmission with a large mode field area.

[0022] Figure 2A schematic diagram of the transformation of the refractive index of a two-dimensional structural material with a freely distributed cross-section of an optical fiber into a two-dimensional data matrix;

[0023] Figure 3 A schematic diagram of the process for two-dimensional feature extraction and optical property prediction of optical fiber structures based on convolutional neural networks provided by this invention;

[0024] Figure 4 : A schematic diagram of the convolutional neural network and fully connected regression computation network structure used in this invention;

[0025] Figure 5 A comparison chart of predicted and actual values ​​of field area for six modes that can be transmitted was obtained using a test set.

[0026] Figure 6 : Using the test set, we obtained a comparison chart of the dispersion prediction values ​​and the true values ​​that can support the transmission of 6 modes;

[0027] Figure 7 and Figure 8 This invention uses a test set to perform tests and obtains a comparison chart of the predicted and true values ​​of the effective refractive index for six modes that can support transmission. Detailed Implementation

[0028] The present invention and technical solution will be further described in detail below with reference to the accompanying drawings.

[0029] A two-dimensional feature extraction and optical property prediction method for optical fiber structures based on convolutional neural networks is presented. Taking multi-core optical fiber as an example, the method predicts the mode field area, dispersion, and effective refractive index of the modes that can support the transmission of supermodes. The multi-core optical fiber used has the following cross-sectional structure: Figure 1 As shown, the fiber core consists of multiple (5 to 8 cores) high-refractive-index doped silica cores, each equidistant from the fiber center. By changing fiber structure parameters (e.g., core radius, number, distribution, etc.), different fiber structures can be obtained. Two-dimensional cross-sectional data matrices corresponding to different fiber structures, as well as the mode field area, dispersion, and effective refractive index of the supported supermodes, can be collected. The example collected mode is a HE supermode that can be transmitted. 11 HE 21 HE 31 HE 41 EH 11 EH 21 The mode field area, dispersion, and effective refractive index of these six modes.

[0030] Taking a multi-core optical fiber as an example, the number of cores is set to 5, each with a radius of 4μm. The refractive index of each core is 0.03 higher than that of pure silica, and the distance between the center of each core and the center of the fiber is 7μm. The two-dimensional distribution of the refractive index of the cross-sectional material of this optical fiber is converted into a two-dimensional numerical matrix, and the matrix size is set to 227×227. Under these parameters, the schematic diagram of the two-dimensional data matrix of the refractive index distribution of the cross-sectional material of this multi-core optical fiber model is shown below. Figure 2 As shown. Figure 2 This is just one example of the transformation. The transformation method can be flexibly adjusted. For example, the matrix size can be adjusted, and the meaning of the transformation values ​​(refractive index, material absorptivity, etc.) can be adjusted as needed.

[0031] Figure 3 The flowchart shown is for this invention. First, the invention uses the finite element simulation software COMSOL Multiphysics to collect data such as mode area, dispersion, and effective refractive index corresponding to several modes that can support transmission at different wavelengths under different optical fiber structures. This data serves as the dataset for a convolutional neural network. Simultaneously, the refractive index distribution of the two-dimensional material in the cross-section of the optical fiber structure is converted into a two-dimensional data matrix. Then, a suitable neural network model is built, and the collected training dataset is fed into the built neural network model for training, resulting in a neural network model with low loss and high prediction accuracy. This model is then saved. Finally, the trained network model is tested using a test dataset, and the predicted values ​​are calculated and compared with the actual values ​​in the test dataset to demonstrate the accuracy of the neural network prediction.

[0032] Figure 4 The diagram shows the neural network model proposed in this invention, capable of predicting the optical properties of a two-dimensional freely distributed structure in the cross-section of an optical fiber. The size of the acquired two-dimensional data matrix is ​​227×227. Therefore, the input layer of the convolutional neural network is (227, 227, 1). After multiple adjustments to the network, a total of three convolutional layers are obtained. After all convolutional layers, max pooling layers and ReLU activation functions are used to further reduce the number of weight parameters. Finally, the two-dimensional matrix is ​​processed through the convolutional layers to obtain 179,776 nodes. In the fully connected layer, after multiple adjustments to the network nodes and weights, the final regression network includes an input layer and an output layer. In the input layer, in addition to the 179,776 nodes obtained from the convolutional layers, wavelength nodes are added, and the wavelength values ​​are feature-scaled and normalized to between 0 and 1. Finally, the network uses the Adam optimizer, with an initial learning rate of 0.00005, and undergoes 500 iterations of training.

[0033] The input layer of the finally trained prediction neural network is a two-dimensional data matrix of the cross-sectional structure of the optical fiber, and the output layer is the mode area, dispersion and effective refractive index corresponding to different wavelengths. Figure 5 , Figure 6, Figure 7 and Figure 8 The figure shows the prediction results of the neural network tested using a test dataset, verifying the accuracy of the prediction network model. The figure also shows the prediction of the optical properties of six modes that a multi-core optical fiber can support, including mode field area, dispersion, and effective mode refractive index at different wavelengths. It can be seen that the proposed two-dimensional feature extraction and optical property prediction method for optical fiber structures based on convolutional neural networks can quickly and accurately predict and calculate the corresponding optical properties based on the refractive index and geometric size distribution of arbitrary materials in the two-dimensional cross-section of the optical fiber. Compared with traditional numerical simulation methods, the error is smaller, and the prediction calculation time is improved by an order of magnitude. The steps of this technical solution are also applicable to the prediction of other optical structure properties.

[0034] The parts of this invention not described in detail are common knowledge to those skilled in the art.

Claims

1. A method for two-dimensional feature extraction and optical property prediction of optical fiber structures based on convolutional neural networks, the method comprising the following steps: Step 1: By selecting an appropriate sampling interval, the two-dimensional material refractive index distribution structural characteristics of the optical fiber cross section are converted into a two-dimensional data matrix. The value of each point in the matrix represents the material refractive index characteristics at the corresponding optical fiber structure. Step 2: Use mathematical physics numerical simulation methods to collect data sets of three optical properties of optical fibers: mode field area, dispersion, and effective refractive index. Step 3: Construct the optical fiber structure feature extraction and regression calculation prediction neural network structure. In the process of constructing the prediction network, a convolutional neural network is used to extract features from the two-dimensional data matrix, and then a fully connected neural network is used to perform regression calculation on the extracted features. Step 4: Train the network using the collected dataset and save the model; Step 5: Test the predictive computation performance of the network model using a test set; Step 6: Save the most suitable prediction network for quickly and accurately calculating the three optical properties of different optical fibers: mode area, dispersion, and effective refractive index.

2. The method for two-dimensional feature extraction and optical property prediction of optical fiber structures based on convolutional neural networks as described in claim 1, characterized in that: In step 3, during the construction of the prediction network, after using a convolutional neural network to extract features from the two-dimensional data matrix, the extracted structural features and wavelength nodes are put into the fully connected input layer in the fully connected layer to obtain regression prediction results at different wavelengths.

3. The method for two-dimensional feature extraction and optical property prediction of optical fiber structures based on convolutional neural networks as described in claim 1, characterized in that: In step 4, before training with the collected dataset, the data needs to be separated according to different types of patterns, and then saved separately according to the three optical properties of mode field area, dispersion and effective refractive index, and put into the regression prediction model for training.

Images