Grating coupler broadband coupling efficiency prediction method based on deep learning
Through the wide-band coupling efficiency prediction method of grating coupler based on deep learning, the discrete matrix is processed by using the prediction neural network to solve the problem of inefficient grating coupler design in the prior art, and high-precision wide-band coupling efficiency prediction is achieved, and a rapid design that meets the needs of complex applications is achieved.
Patent Information
- Application Number
- CN202510093289.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-21
- Publication Date
- 2025-05-16
Smart Images
Figure CN120012579A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of grating couplers, and in particular relates to a method for predicting wide-band coupling efficiency of a grating coupler. Background Art
[0002] Grating coupler is a photonic integrated device, which is mainly used to achieve efficient coupling between high-performance integrated waveguide devices and optical fibers or free-space optical devices. Generally, grating couplers are composed of a series of periodic parallel grating grooves and grating teeth. The structural parameters and shapes of these grooves and teeth can be precisely controlled to disperse or control the diffraction angle of the incident light, so that the light from the optical fiber or free space can be effectively coupled to the silicon-on-insulator (SOI) waveguide. With the continuous development of photonic integrated devices, the application requirements of grating couplers have gradually become more complex, such as broadband grating couplers, wavelength beam splitting grating couplers, and polarization beam splitting grating couplers, which puts higher requirements on the design method of grating couplers.
[0003] Traditional design methods are no longer efficient enough to meet the needs of current applications. For a long time, the design of grating couplers and other photonic devices has relied on a combination of experience and theory, that is, by applying known theories (such as Bragg conditions) to calculate and select the initial grating coupler structural parameters, and then through a large number of parameter scans or optimization algorithms to make small adjustments and optimizations to its structural parameters or shape to meet the target application requirements. This type of design method inevitably has the following problems:
[0004] 1. The structural parameter variables are small and cannot meet the complex and diverse application requirements;
[0005] 2. The optimization process requires a complex electromagnetic simulation solver, which is very time-consuming and inefficient;
[0006] 3. The solution process requires meshing and iterative solution of the device structure, and the required computing resources are very huge.
[0007] To solve the above problems, researchers have explored the use of deep learning-based methods to replace traditional electromagnetic simulation solvers in recent years to achieve more efficient grating coupler design. However, the deep learning frameworks and grating structures used in these studies are relatively simple and are generally limited to a single incident angle and a very narrow band range. At present, there is still a lack of a high-precision prediction method for the coupling efficiency of grating couplers with multiple incident angles and wide frequency bands. Summary of the invention
[0008] The invention aims to solve the problems of long design time and low design efficiency in the existing grating coupler design.
[0009] A method for predicting wide-band coupling efficiency of grating coupler based on deep learning, comprising:
[0010] For the grating coupler, a discrete grating coupler model is established in the software, and the grating area in the model is equivalent to an M*1 discrete matrix, where M is the matrix dimension, corresponding to the total number of grating periods of the grating coupler, and the encoding value of each matrix element represents that the position is a grating tooth or a grating groove; the discrete matrix is taken as input, and the coupling efficiency T is predicted using a predictive neural network.
[0011] Furthermore, the prediction neural network includes: a position encoding module, a wavelength encoding module, a local information extraction module, a transposed attention module, and a non-shared parameter fully connected module;
[0012] The position encoding module performs position encoding on the discrete matrix, and the position encoding is spliced with the discrete matrix representing the grating coupler structure, and then passes through the local information extraction module, the transposed attention module and the non-shared parameter fully connected module in sequence; among them, the wavelength encoding module is located in the transposed attention module, and the input of the wavelength encoding module is N natural numbers, corresponding to different wavelengths of the output coupling efficiency vector, and N is the number of coupling efficiencies T;
[0013] The local information extraction module is used for feature extraction. The output of the local information extraction module is spliced with the output of the wavelength encoding module and then input into the transposed attention module. The non-shared parameter fully connected module includes multiple fully connected layers. In each fully connected layer, the parameters are decomposed into N groups of different groups of parameters, and the parameters of the fully connected layer are not shared; the output of the non-shared parameter fully connected module is the coupling efficiency T.
[0014] Furthermore, the position encoding module includes 4 1×1 convolutional layers.
[0015] Furthermore, the wavelength encoding module includes four 1×1 convolutional layers.
[0016] Furthermore, the local information extraction module includes 8 residual blocks, each residual block includes a 1*1 convolutional layer and a 3*3 convolutional layer.
[0017] Furthermore, the transposed attention module includes two Kolmogorov-Arnold Network modules; the output of the local information extraction module is spliced with the output of the wavelength encoding module and then input into the first Kolmogorov-Arnold Network module, and the output of the first Kolmogorov-Arnold Network module is transposed and spliced with the output of the position encoding module and then input into the second Kolmogorov-Arnold Network module;
[0018] The Kolmogorov-Arnold Network module includes a Kolmogorov-Arnold Network layer, a 1*3 convolution layer and a Layer Normalization layer. The output of the Kolmogorov-Arnold Network layer is the input of the convolution layer. The output of the convolution layer is added to the Kolmogorov-Arnold Network layer and then passes through the LayerNormalization layer. The output of the Layer Normalization layer is processed by Self-Attention, and the result is added to the input of the Kolmogorov-Arnold Network layer to obtain the output of the Kolmogorov-Arnold Network module.
[0019] Furthermore, the non-shared parameter fully-connected module includes 4 fully-connected layers.
[0020] Furthermore, the prediction neural network is pre-trained, and the loss function used in the process of training the prediction neural network is as follows:
[0021] The loss function is as follows:
[0022] L=L 1 (Y,lnT)-L cos (Y,lnT)+L 1 (f(Y),f(lnT))
[0023] Where Y is the coupling efficiency label of the grating coupler, L 1 represents the mean absolute error, L cos represents cosine similarity, and f represents the matrix composed of the absolute values of the difference between any two elements in Y, which is expressed as the following formula:
[0024]
[0025] Among them, Y 1 To Y n is the element in Y.
[0026] Furthermore, the grating coupler is based on a silicon-on-insulator substrate material, which includes a three-layer structure, from top to bottom, a crystalline silicon layer, a silicon dioxide buried layer and a silicon substrate layer; the grating region is distributed in the crystalline silicon layer and is composed of a number of random grating teeth and grating grooves arranged in an alternating manner.
[0027] Furthermore, a silicon dioxide cladding layer is epitaxially grown on top of the crystalline silicon layer by chemical vapor deposition.
[0028] The present invention has the following beneficial effects:
[0029] The present invention can accurately predict the coupling efficiency of light with different incident angles after passing through the grating coupler within a wide band, and the average error is less than 0.2%. The present invention has the advantage of high prediction accuracy; at the same time, the present invention can quickly predict the coupling efficiency. Therefore, the present invention can be used for the rapid design of grating couplers that meet complex requirements, shorten the design time, and improve the design efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 The flowchart of the method for predicting the wide-band coupling efficiency of grating couplers is shown;
[0031] Figure 2 This is a schematic diagram of the structure of the prediction neural network;
[0032] Figure 3 is a schematic diagram of the grating model;
[0033] Figure 4 The predicted value when the incident angle is 0° and the output result of the simulation software are shown;
[0034] Figure 5 The predicted value and simulation software output result diagram when the incident angle is 5°;
[0035] Figure 6 The predicted value and simulation software output result diagram when the incident angle is 10°;
[0036] Figure 7 The predicted value and simulation software output result diagram when the incident angle is 15°;
[0037] Figure 8 The predicted value and simulation software output result diagram when the incident angle is 20°;
[0038] Fig. 9 The figure shows the predicted value and simulation software output result when the incident angle is 25°. DETAILED DESCRIPTION
[0039] Specific implementation method 1: Combination Figure 1 To explain this embodiment,
[0040] The specific implementation method is a method for predicting wide-band coupling efficiency of a grating coupler based on deep learning, comprising the following steps:
[0041] Step 1, prepare a training data set; the training data set includes input data and labels, the input data is a discrete matrix obtained after digital discretization of a randomly generated grating coupler, and the label is the coupling efficiency of the grating coupler within a certain band obtained by using simulation software;
[0042] In this embodiment, the grating coupler is based on a silicon-on-insulator (SOI) substrate material, which includes a three-layer structure, from top to bottom, a crystalline silicon layer, a silicon dioxide buried layer (BOX), and a silicon substrate layer. Figure 3 As shown in the figure, the grating area is distributed in the crystalline silicon layer, which is composed of a number of random grating teeth and grating grooves arranged in an alternating manner. A layer of silicon dioxide (SiO2) is epitaxially grown on the top of the crystalline silicon layer by chemical vapor deposition process. 2 ) cladding to better confine the light field and protect the grating layer.
[0043] The training data set samples are obtained in the following way:
[0044] Step 1. According to the above structure, a discrete grating coupler model is established in the software. The grating area in the model can be equivalent to an M*1 discrete matrix, where M=320 is the matrix dimension, corresponding to the total number of grating periods of the grating coupler; the encoding value of each matrix element is '0' or '1', where 0 indicates that the position is a grating tooth, and 1 indicates that the position is a grating groove, and the sizes of the grating teeth and grating grooves are equal. Finally, the matrix is saved as a data set.
[0045] Step 1 and 2: Run electromagnetic simulation software to obtain the coupling efficiency of the above discrete grating coupler at different incident angles over a wide wavelength (1.35-1.75 μm).
[0046] Step 2: Build a prediction neural network. The input of the prediction neural network is the M*1 matrix mentioned above, which represents the structure of different grating couplers. The output is the coupling efficiency T of the grating coupler within a certain wavelength range, where T is a 201*1 matrix.
[0047] like Figure 2 As shown, the specific structure of the network structure is as follows:
[0048] The prediction neural network includes: a position encoding module, a wavelength encoding module, a local information extraction module, a transposed attention module based on the Kolmogorov-Arnold Network module, and a non-shared parameter fully connected module;
[0049] The position encoding module encodes the position of the discrete matrix, which is concatenated with the discrete matrix representing the grating coupler structure, and then passes through the local information extraction module, the transposed attention module and the non-shared parameter fully connected module in sequence; among them, the wavelength encoding module is located in the transposed attention module.
[0050] The structures of the position coding module and the wavelength coding module can be designed according to actual conditions. In this embodiment, the two include four 1×1 convolutional layers respectively, wherein the features obtained by the first convolutional layer of the position coding module are added to the features obtained by the third convolutional layer, and then connected with the input and sent to the fourth convolutional layer for processing. The structures of the position coding module and the wavelength coding module can be the same or different. In this embodiment, they are set to the same structure; the input of the position coding module is M natural numbers, corresponding to the position of each element of the M*1 matrix; the input of the wavelength coding module is N natural numbers (N is the number of coupling efficiencies T, where N=201), corresponding to different wavelengths of the output coupling efficiency vector.
[0051] The local information extraction module consists of 8 residual blocks, each of which contains a 1*1 convolutional layer and a 3*3 convolutional layer;
[0052] The transposed attention module includes two Kolmogorov-Arnold Network modules; the output of the local information extraction module is spliced with the output of the wavelength encoding module and input into the first Kolmogorov-Arnold Network module, and the output of the first Kolmogorov-Arnold Network module is transposed and spliced with the output of the position encoding module and input into the second Kolmogorov-Arnold Network module.
[0053] The Kolmogorov-Arnold Network module includes a Kolmogorov-Arnold Network layer, a 1*3 convolution layer and a Layer Normalization layer. The output of the Kolmogorov-Arnold Network layer is the input of the convolution layer. The output of the convolution layer is added to the Kolmogorov-Arnold Network layer and then passed through the LayerNormalization layer. The output of the Layer Normalization layer is processed by Self-Attention, and the result is added to the input of the Kolmogorov-Arnold Network layer to obtain the output of the Kolmogorov-Arnold Network module.
[0054] In fact, the prediction neural network described in this embodiment includes 8 transposed attention modules based on the Kolmogorov-ArnoldNetwork module.
[0055] The non-shared parameter fully-connected module includes 4 fully-connected layers. In each fully-connected layer, the parameters are decomposed into N groups of parameters, and the parameters are not shared. The output of the non-shared parameter fully-connected module is the coupling efficiency T.
[0056] Step 3: construct a loss function based on the grating coupler coupling efficiency label Y obtained by simulation and the output of the prediction neural network;
[0057] The loss function is as follows:
[0058] L=L 1 (Y,lnT)-L cos (Y,lnT)+L 1 (f(Y),f(lnT))
[0059] Among them, L 1 represents the mean absolute error, L cos represents cosine similarity, and f represents a matrix consisting of the absolute values of the difference between any two elements in a one-dimensional vector, which can be expressed as the following formula:
[0060]
[0061] Step 4: Use the training data set and the constructed loss function to train the prediction neural network. During the training process, the batch size is set to 600, the ADAM optimization algorithm is used, and 400 epochs are trained.
[0062] Step 5: First, for the grating coupler, use the method in step one to arbitrarily establish a discrete grating coupler model, and obtain the corresponding discrete matrix M according to the structure, and then use the trained prediction neural network to predict its coupling efficiency T.
[0063] According to the above process, the following simulation is performed:
[0064] (1) 20,000 randomly distributed discrete grating couplers are randomly generated, and their coupling efficiencies at different angles are calculated using electromagnetic simulation software.
[0065] (2) After the calculation is completed, the discrete matrices corresponding to different discrete grating couplers (for example, M1 = [0, 1, 1, 0, ..., 0, 1], where M1 is a 320*1 matrix, and M is 320 in all grating couplers) and the completed coupling efficiency T (for example, T1 = [0.05, 0.05, 0.08, ... 0.15, 0.25], where T1 is a 201*1 matrix, and the 201 values correspond to wavelengths in the range of 1.35-1.75 um, and their wavelength interval is 2 nm) are saved in the training data set.
[0066] (3) Build a neural network according to the above content.
[0067] (4) Construct a loss function.
[0068] (5) Use 2 and 4 to train the neural network.
[0069] (6) After the neural network training is completed, the following functions can be realized: a 320*1 matrix is randomly generated and input into the neural network, and the corresponding coupling efficiency T can be predicted within about 0.06s, which is basically consistent with the result of electromagnetic simulation. Figure 4-Figure 9 As shown, Figure 4-Figure 9 Corresponding to the cases of incident angle of 0°, incident angle of 5°, incident angle of 10°, incident angle of 15°, incident angle of 20° and incident angle of 25° respectively
[0070] The above calculation examples of the present invention are only used to explain the calculation model and calculation process of the present invention in detail, and are not intended to limit the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.
Claims
1. A method for predicting wide-band coupling efficiency of grating coupler based on deep learning, characterized in that: include: For the grating coupler, a discrete grating coupler model is established in the software, and the grating area in the model is equivalent to an M*1 discrete matrix, where M is the matrix dimension, corresponding to the total number of grating periods of the grating coupler, and the encoding value of each matrix element represents that the position is a grating tooth or a grating groove; the discrete matrix is taken as input, and the coupling efficiency T is predicted using a predictive neural network.
2. According to the method for predicting wide-band coupling efficiency of grating coupler based on deep learning in claim 1, it is characterized in that: The prediction neural network includes: a position encoding module, a wavelength encoding module, a local information extraction module, a transposition attention module, and a non-shared parameter fully connected module; The position encoding module performs position encoding on the discrete matrix, and the position encoding is spliced with the discrete matrix representing the grating coupler structure, and then passes through the local information extraction module, the transposed attention module and the non-shared parameter fully connected module in sequence; among them, the wavelength encoding module is located in the transposed attention module, and the input of the wavelength encoding module is N natural numbers, corresponding to different wavelengths of the output coupling efficiency vector, and N is the number of coupling efficiencies T; The local information extraction module is used for feature extraction. The output of the local information extraction module is spliced with the output of the wavelength encoding module and then input into the transposed attention module. The non-shared parameter fully connected module includes multiple fully connected layers. In each fully connected layer, the parameters are decomposed into N groups of different groups of parameters, and the parameters of the fully connected layer are not shared; the output of the non-shared parameter fully connected module is the coupling efficiency T.
3. The method for predicting wide-band coupling efficiency of a grating coupler based on deep learning according to claim 2, characterized in that: The position encoding module includes four 1×1 convolutional layers.
4. The method for predicting wide-band coupling efficiency of a grating coupler based on deep learning according to claim 2, characterized in that: The wavelength encoding module includes four 1×1 convolutional layers.
5. The method for predicting wide-band coupling efficiency of a grating coupler based on deep learning according to claim 2, characterized in that: The local information extraction module includes 8 residual blocks, each of which includes a 1*1 convolutional layer and a 3*3 convolutional layer.
6. The method for predicting wide-band coupling efficiency of a grating coupler based on deep learning according to claim 2, characterized in that: The transposed attention module includes two Kolmogorov-Arnold Network modules; the output of the local information extraction module is spliced with the output of the wavelength encoding module and then input into the first Kolmogorov-Arnold Network module, and the output of the first Kolmogorov-Arnold Network module is transposed and spliced with the output of the position encoding module and then input into the second Kolmogorov-Arnold Network module; The Kolmogorov-Arnold Network module includes a Kolmogorov-Arnold Network layer, a 1*3 convolution layer and a Layer Normalization layer. The output of the Kolmogorov-Arnold Network layer is the input of the convolution layer. The output of the convolution layer is added to the Kolmogorov-Arnold Network layer and then passed through the LayerNormalization layer. The output of the LayerNormalization layer is processed by Self-Attention, and the result is added to the input of the Kolmogorov-Arnold Network layer to obtain the output of the Kolmogorov-Arnold Network module.
7. The method for predicting wide-band coupling efficiency of a grating coupler based on deep learning according to claim 2, characterized in that: The non-shared parameter fully connected module includes 4 fully connected layers.
8. The method for predicting wide-band coupling efficiency of a grating coupler based on deep learning according to claim 2, characterized in that: The prediction neural network is pre-trained. During the training of the prediction neural network, the loss function used is as follows: The loss function is as follows: L=L1(Y,ln T)-L cos (Y,ln T)+L1(f(Y),f(ln T)) Where Y is the coupling efficiency label of the grating coupler, L1 represents the mean absolute error, and L cos represents cosine similarity, and f represents the matrix composed of the absolute values of the difference between any two elements in Y, which is expressed as the following formula: Among them, Y1 to Y n is the element in Y.
9. A method for predicting wide-band coupling efficiency of a grating coupler based on deep learning according to any one of claims 1 to 8, characterized in that: The grating coupler is based on a silicon-on-insulator substrate material, which includes a three-layer structure, from top to bottom, a crystalline silicon layer, a silicon dioxide buried layer and a silicon substrate layer; the grating area is distributed in the crystalline silicon layer and is composed of a number of random grating teeth and grating grooves arranged in an alternating manner.
10. The method for predicting wide-band coupling efficiency of a grating coupler based on deep learning according to claim 9, characterized in that: A silicon dioxide cladding layer is also epitaxially grown on top of the crystalline silicon layer using a chemical vapor deposition process.