Design method, system, computer device and storage medium of grating coupler

Abstract

The application relates to the technical field of grating coupler design, and discloses a design method and system of a grating coupler, computer equipment and a storage medium, which comprise the following steps: a topological prediction model is constructed based on a deep neural network, and the topological prediction model is trained by using a preset method; the response spectrum of the grating coupler is input into the trained topological prediction model, so that the structure parameters of the functional area of the grating coupler are obtained; and the topological structure of the functional area of the grating coupler is determined according to the structure parameters, so that the design of the grating coupler is realized. Through the approximate processing of the training data set, the simple network structure is combined with the complex nonlinear algorithm, the model power requirement is reduced, the prediction efficiency and effect of the model are improved, the prediction accuracy of the model is further improved through the training method combining the forward direction and the reverse direction, so that the design of the grating coupler with high coupling efficiency at any angle is realized, and the device performance of the grating coupler is ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of grating coupler design, and in particular to a grating coupler design method and system, computer equipment and a storage medium. BACKGROUND

[0002] With the continuous development of information technology, big data, Internet of Things, cloud computing and other applications lead modern civilization into the era of massive data. In the face of explosive growth of data volume, traditional electrical signal processing methods are gradually no longer applicable, and photonic integrated chips with all-optical signal processing as a typical application have attracted a lot of attention.

[0003] For silicon-based photonic integrated chips, a significant limiting factor is that silicon is a poor light emitter as an indirect bandgap material, so an off-chip light source or a hybrid integrated gain material on-chip needs to be introduced. As a passive device that can couple optical chips and external light sources (optical fibers) to each other, the grating coupler is known for its flexible location, which can be located at any position in the chip, and thus can be used for wafer-level online testing. In addition, its alignment tolerance for coupling is large, which is convenient for packaging and other advantages. However, the grating coupler has the disadvantages of low coupling efficiency and narrow bandwidth, which greatly hinders its practical application.

[0004] In order to improve the performance index of the device, the structure parameters of the device can be optimized. The commonly used structure optimization algorithms at present include genetic algorithm, particle swarm algorithm, topological optimization algorithm, etc., but these algorithms have certain limitations. On the one hand, as a global optimization algorithm, the genetic algorithm is usually suitable for optimization scenarios with fewer parameters, and the optimization result improves as the parameter space increases. However, the cost is the rapid increase of the amount of calculation and the demand for computing resources. Therefore, this algorithm is not suitable for the grating coupler which needs to optimize multiple parameters in parallel. On the other hand, although the particle swarm algorithm can be applied to the design of silicon optical devices, its convergence speed is too slow and is not suitable for complex grating coupler optimization scenarios. In addition, the topological optimization algorithm needs a strict mathematical expression between the optimization target and the structure parameters, and combines the steepest descent algorithm to execute the optimization process. Its process is too complex and the iteration convergence speed is slow, so it is also not suitable for complex grating coupler optimization scenarios. SUMMARY

[0005] In order to solve the above technical problems, the present application provides a grating coupler design method, system, computer equipment and storage medium, which can solve the problems of large computing resource demand, long iteration period and slow convergence speed of the existing design method, and achieve the effect of quickly and efficiently realizing grating coupler design.

[0006] In a first aspect, the present application provides a grating coupler design method, which comprises:

[0007] A topology prediction model is constructed based on a deep neural network, and the topology prediction model is trained using a preset method, the topology prediction model comprising an input layer, an output layer and a plurality of hidden layers, an activation function of the topology prediction model being a Swish function, and a training data set being obtained by approximate processing to adapt the convergence of the Swish function;

[0008] The response spectrum of the grating coupler is input into the trained topology prediction model to obtain the structure parameters of the functional region of the grating coupler;

[0009] According to the structure parameters, the topology structure of the functional region of the grating coupler is determined to realize the design of the grating coupler.

[0010] Further, the step of training the topology prediction model using a preset method comprises:

[0011] Training data is generated by a simulation method, and the training data is approximately processed to obtain a training data set;

[0012] The training data set is used to train the topology prediction model.

[0013] Further, the step of generating training data by a simulation method and approximately processing the training data to obtain a training data set comprises:

[0014] Training data with SOI structure consistency is generated by time-domain finite difference simulation, the training data comprising structure training data and spectrum training data;

[0015] The functional region of the grating coupler corresponding to the structure training data is divided into a plurality of pixel regions according to a preset size;

[0016] The data of each pixel region is converted according to a region state to obtain processed structure training data, the region state comprising an etching state and a non-etching state;

[0017] The processed structure training data and the spectrum training data are combined to form a training data set.

[0018] Further, the step of training the topology prediction model using the training data set comprises:

[0019] A forward training model with the same structure as the topology prediction model is constructed, input data of the forward training model being structure parameters of the functional region of the grating coupler, and output data of the forward training model being a response spectrum of the grating coupler;

[0020] The training data set is used to train the forward training model, and first network parameter configuration of the forward training model is obtained.

[0021] The trained forward training model and the untrained topological prediction model are connected in series to obtain a series network model, and the training data set is used to train the series network model.

[0022] The second network parameter configuration of the topological prediction model is determined from the trained series network model, and the trained topological prediction model is obtained.

[0023] In a second aspect, the present application provides a design system of a grating coupler, the system comprising:

[0024] A model construction module is configured to construct a topological prediction model based on a deep neural network, and train the topological prediction model using a preset method, wherein the topological prediction model comprises an input layer, an output layer and a plurality of hidden layers, the activation function of the topological prediction model is a Swish function, and the training data set is constructed using an approximation process to adapt to the convergence of the Swish function.

[0025] A parameter prediction module is configured to input the response spectrum of the grating coupler into the trained topological prediction model to obtain the structural parameters of the functional region of the grating coupler.

[0026] A grating design module is configured to determine the topological structure of the functional region of the grating coupler according to the structural parameters to realize the design of the grating coupler.

[0027] Further, the model construction module comprises a model training module.

[0028] The model training module is configured to generate training data by a simulation method, and perform an approximation process on the training data to obtain a training data set.

[0029] The training data set is used to train the topological prediction model.

[0030] Further, the model training module further comprises a data set construction module.

[0031] The data set construction module is configured to generate training data with SOI structure consistency by a finite difference time domain simulation, wherein the training data comprises structural parameters and spectral parameters.

[0032] The functional region of the grating coupler corresponding to the structural parameters is divided into a plurality of pixel regions according to a preset size.

[0033] According to the data conversion of each pixel region according to the region state, the processed structure parameter is obtained, and the region state includes an etching state and a non-etching state.

[0034] According to the processed structure parameter and the spectrum parameter, a training data set is composed.

[0035] Further, the model training module is also used to construct a forward training model with the same structure as the topological prediction model, the input data of the forward training model is the structure parameter of the functional region of the grating coupler, and the output data of the forward training model is the response spectrum of the grating coupler.

[0036] The training data set is used to train the forward training model, and the first network parameter configuration of the forward training model is obtained.

[0037] The trained forward training model and the untrained topological prediction model are connected in series to obtain a series network model, and the training data set is used to train the series network model.

[0038] The second network parameter configuration of the topological prediction model is determined from the trained series network model, and the trained topological prediction model is obtained.

[0039] In a third aspect, an embodiment of the present application also provides a computer device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the above method when executing the computer program.

[0040] In a fourth aspect, an embodiment of the present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the above method.

[0041] The present application provides a design method, system, computer device and storage medium of a grating coupler. Through the approximate processing of the training data set, the simple network structure is combined with the complex nonlinear algorithm, the Swish function can converge under the simple network architecture, the algorithm requirement of the model can be reduced, the prediction efficiency and effect of the model can be improved, and the prediction accuracy of the model can be improved through the training method combining the forward and reverse directions, so that the design of the grating coupler with high coupling efficiency at any angle is realized, and the device performance of the grating coupler is ensured. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a flowchart of the design method of the grating coupler in the embodiment of the present application;

[0043] Figure 2is a schematic structural diagram of a grating coupler according to an embodiment of the present invention;

[0044] Figure 3 is a schematic diagram of the structure of a topology prediction model in an embodiment of the present invention;

[0045] Figure 4 Schematic diagram of the structure of the forward training model in an embodiment of the present invention;

[0046] Figure 5 is a structural diagram of a series network model according to an embodiment of the present invention;

[0047] Figure 6 is a schematic structural diagram of a design system for a grating coupler according to an embodiment of the present invention;

[0048] Figure 7 1 is a diagram showing the internal structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0050] See also Figure 1 The first embodiment of the present invention provides a method for designing a grating coupler, which includes steps S10 to S30:

[0051] Step S10: constructing a topology prediction model based on a deep neural network and training the topology prediction model using a preset method. The topology prediction model includes an input layer, an output layer, and several hidden layers. The activation function of the topology prediction model is a Swish function. The training data set is constructed using approximate processing to adapt to the convergence of the Swish function.

[0052] Step S20, inputting the response spectrum of the grating coupler into the trained topology prediction model to obtain the structural parameters of the functional area of ​​the grating coupler;

[0053] Step S30 : determining the topological structure of the functional area of ​​the grating coupler according to the structural parameters to realize the design of the grating coupler.

[0054] In this embodiment, a topology prediction model based on a deep neural network is used to predict the structural parameters of the functional area of ​​the grating coupler. The structure of the grating coupler is as follows: Figure 2As shown, the structure of the high-coupling-efficiency non-periodic grating coupler includes a silicon substrate, a silicon dioxide buried oxygen layer, a silicon structure layer and a silicon dioxide cladding layer, and the topological structure of the functional area responsible for coupling in the silicon structure layer is the part to be designed by the application, which is obtained by etching the silicon structure layer. Therefore, the binary code obtained by the topology prediction model provided by the application represents whether each structure parameter (i.e., pixel area) in the topological structure is etched.

[0055] In the embodiment, the topology prediction model comprises an input layer, an output layer and a plurality of hidden layers, as shown in the figure. Figure 3 In order to distinguish the architecture of the forward training model in the subsequent training process, it is denoted as a first input layer and a first output layer here. In order to reduce the demand for computing resources, shorten the iteration period of the model, improve the convergence speed of the model, and ensure the reliability and stability of the calculation, the hidden layer is set to at least three layers and less than ten layers in the embodiment, and preferably four layers. The input data of the model is the expected response spectrum of the to-be-designed device, and the output is the structure parameter of the functional area of the to-be-designed device. The input layer is used to discretize the input response spectrum and pass it to the hidden layer. The hidden layer is used to extract spectral features from the response spectrum, perform linear or nonlinear processing through an activation function, obtain state probability values of the structure parameters, and pass them to the output layer. That is, the task of the hidden layer is to process the input data passed from the input layer, obtain the pixel point probability of the pixel topological pattern, and pass it to the output layer of the back end. The output layer is used to judge the input structure parameters and output the binary coded structure parameters. For example, when the obtained pixel point probability value is greater than 0.5, the corresponding structure parameter (pixel area) unit is coded as 1, otherwise it is coded as 0.

[0056] The topology prediction model in the embodiment adopts a deep learning neural network architecture with a small scale. For each layer of the neural network model, ι the input vector x (l) needs to be linearly transformed:

[0057] z (l) = W (l) x (l) +b (l)

[0058] where W (l) is the weight vector of the lth layer; b (l) is the bias vector of the lth layer; and z (l) is the linear output of the input vector of the lth layer.

[0059] In order to introduce a nonlinear relationship so that the network can approximate a complex function, a nonlinear activation function is used in each layer. The output after activation is represented as:

[0060] a (l) = F(b (l) )

[0061] where F denotes the activation function, a (l) is the output after the activation function.

[0062] From the input layer to the output layer, linear transformation and activation function are performed layer by layer, and finally the output of the network is obtained:

[0063]

[0064] The goal of deep learning is to minimize the loss function through training. The loss function used in this embodiment is:

[0065]

[0066] where x and y i represent the i-th predicted value and actual value, respectively, and n represents the number of predicted values.

[0067] The gradient of the loss function with respect to the weights and biases is calculated using the backpropagation algorithm through chain rule.

[0068] Assuming the loss function is L, the gradient of the weights and biases for each layer is:

[0069]

[0070] The weights and biases are updated using the gradient descent principle:

[0071]

[0072] where η is the learning efficiency.

[0073] The loss is calculated by forward propagation, and the weights and biases are updated by backpropagation. This process is repeated multiple times (called epoch) until the loss function converges. Deep learning neural networks can be represented as a combination of layer-by-layer linear transformations and nonlinear activations. The output is obtained by forward propagation, and the weights are continuously adjusted by backpropagation and gradient descent to ultimately achieve the desired mapping relationship on a given data set.

[0074] In the above model construction process, the activation function is an important part. Since the design of the grating coupler belongs to a multi-element optimization problem, in order to obtain more accurate optimization results, a more advanced activation function suitable for complex reasoning tasks is needed. For this reason, in this embodiment, the activation function of the topology prediction model uses the Swish function:

[0075]

[0076] where x is the input.

[0077] Swish function is a smooth, non-monotonic activation function suitable for improving model performance and gradient flow in deep neural networks. One of the biggest features of the function is its non-monotonicity, that is, as x increases, the function value may decrease in some intervals, which helps the neural network model learn complex features. Swish function uses self-gating mechanism of input, which makes it perform better than traditional activation functions in many tasks, especially in deep models. The following table 1 compares Swish function with other common functions:

[0078] Table 1 Comparison of Swish function and other common functions

[0079]

[0080] From the above comparison, it can be clearly seen that Swish activation function has excellent performance and is suitable for complex inference tasks, but its cost is high computational complexity and large-scale deep neural network requirement. The following table 2 shows the difference in computational complexity between Swish function and ReLU function:

[0081] Table 2 Computational complexity of Swish function and ReLU function

[0082] Characteristics ReLU Swish Mathematical expression max(0, x) x · σ(x) Computational complexity O(1) O(1) Actual operation steps 1 comparison 1 exponent, 1 division, 1 multiplication, 1 addition Actual time consumption (per neuron) 1 ns 18 ns

[0083] From table 2, it can be seen that Swish function has huge computing power requirement, therefore, Swish function is usually applied in complex architecture network model, which requires several tens of layers or even hundreds of layers to make the model achieve good convergence effect. For the topology prediction model with only a few layers in the embodiment, when Swish function is used as the activation function, the model cannot converge in the training process, that is, Swish function cannot be directly used as the activation function of the topology prediction model. In order to utilize the excellent inference performance of Swish function, the present application preprocesses the training data set and trains the model using the processed data set to solve the mismatch problem between the simple model structure of the topology prediction model and the high computing power requirement of the activation function. The specific steps include:

[0084] generating training data by simulation method and approximately processing the training data to obtain a training data set;

[0085] training the topology prediction model using the training data set.

[0086] In the embodiment, the processing of the data set adopts an approximate processing manner, and the calculation amount required by the algorithm is reduced by adopting a relatively rough fineness, and the steps of the approximate processing include:

[0087] The training data with SOI structure consistency is generated by time domain finite difference simulation, and the training data includes structure training data and spectrum training data;

[0088] The functional area of the grating coupler corresponding to the structure training data is divided into a plurality of pixel areas according to a preset size;

[0089] The data of each pixel area is converted according to a region state, and the processed structure training data is obtained, and the region state includes etching state and non-etching state;

[0090] The training data set is composed according to the processed structure training data and the spectrum training data.

[0091] In the embodiment, the data set for training is generated by time domain finite difference simulation, and the consistency of the SOI structure of the data set is required to guarantee the simplicity of the algorithm in the later stage, and the data set is the structure-performance data set of the photonic device, that is, the structure training data and the corresponding spectrum training data.

[0092] Then the structural training data is approximately processed so that the neural network can converge faster and obtain more stable results. The structural training data is actually the structural parameters of the functional region of the grating coupler. In this embodiment, the functional region of the grating coupler is divided into a plurality of pixel regions according to a preset size. For each pixel region, there are only two states of etching / non-etching, and it is converted into digital 0 or 1 according to the state. For example, the 5um grating coupler functional region is divided into 300 pixel regions with a preset size of 50nm, thereby obtaining binary coding of 0-1. It can be seen that the pixel region division size in this embodiment is large, that is, a lower precision is used in the data processing process. Generally speaking, lower data precision will affect the accuracy of model training and prediction. In this embodiment, the reason for using lower precision to process data is that although the photonic device is theoretically sensitive to the period and duty cycle of the structural parameters, the difference in such subtle structure cannot be accurately reflected in the actual manufacturing process, and the calculation of such subtle structure often consumes a large amount of computing resources. For example, the period of a certain grating coupler is 500nm or 505nm. In theory, there will be a significant difference in performance parameters, but in actual manufacturing, due to the existence of preparation error, the period of the prepared grating coupler may be a value in 495-505nm. Therefore, unlike conventional neural network prediction models, in this embodiment, a relatively rough precision is used to reduce the amount of calculation required by the algorithm, so that it is possible to use a relatively simple neural network architecture to carry complex algorithms, and such approximate processing will not reduce the accuracy of the final prediction result.

[0093] In this embodiment, by using the processed data set, under the premise that the silicon / silicon dioxide layer thickness in the specified SOI structure is unchanged, by constructing a relatively simple network structure, running a complex nonlinear algorithm, the total operation amount is kept within an acceptable range, and the reduction of operation amount means the reduction of demand for computing resources / running time, thereby realizing the rapid convergence of the deep learning network and reducing the time cost and labor cost required in device design. That is, the data set constructed by approximate processing in this embodiment realizes the combination of simple neural network architecture and complex algorithm, thereby reducing the computing power demand of the model and improving the prediction efficiency and effect of the model.

[0094] Based on the training data set, the model is trained, and a trained topological prediction model is obtained. The input of the model is the expected response spectrum of the designed grating coupler, and the output is the structure parameter corresponding to the response spectrum. However, when using a conventional model training method, there may be multiple corresponding output data for an input data, that is, one response spectrum corresponds to several structure parameters, which is not conducive to the fine design of the grating coupler. In order to achieve a one-to-one correspondence between the input data and the output data, a model training method based on a reverse network is preferably used in the embodiment, so that the output structure parameter of the model is the optimal result of the expected response spectrum. The specific steps include:

[0095] A forward training model with the same structure as the topological prediction model is constructed. The input data of the forward training model is the structure parameter of the functional region of the grating coupler, and the output data of the forward training model is the response spectrum of the grating coupler.

[0096] The training data set is used to train the forward training model, and the first network parameter configuration of the forward training model is obtained.

[0097] The trained forward training model and the untrained topological prediction model are connected in series to obtain a series network model, and the training data set is used to train the series network model.

[0098] The second network parameter configuration of the topological prediction model is determined from the trained series network model, and the trained topological prediction model is obtained.

[0099] In the embodiment, a forward training model with the same structure as the topological prediction model is constructed based on the architecture of the topological prediction model. For ease of description, the forward training model is referred to as a forward network, the topological prediction model is referred to as a reverse network, and the input layer and the output layer in the reverse network are represented as a first input layer and a first output layer. The input layer and the output layer in the forward network are represented as a second input layer and a second output layer. As shown in the architecture of the forward network, Figure 4 The forward network includes a second input layer, a plurality of hidden layers and a second output layer. The difference from the reverse network is that the input of the second input layer is the structure parameter of the functional region of the device; the hidden layer is used to assign different weights to the input features; and the second output layer is used to output the spectral response of the device. Preferably, the number of hidden layers of the forward network is four, and each layer has different node numbers.

[0100] For the constructed forward network, the forward network is trained by the training data set to obtain the trained forward network parameter configuration. Then the trained forward network and the untrained reverse network are connected in series and the configuration parameters of the forward network are locked to obtain a series network model as shown in Figure 5 It can be seen that in the series network model, the data input into the reverse network is actually the data output by the forward network, and the input data of the forward network and the output data of the reverse network are both structure parameters. Then the series network model is trained using the training data set to obtain the deployed series network parameter configuration, that is, the parameter configuration of the forward network and the reverse network. Further, the reverse network, that is, the network parameter configuration of the topology prediction model, is obtained, so that the trained topology prediction model is obtained. In this way, the input and output of the topology prediction model can be optimally corresponding, so that the optimal structure parameters corresponding to the expected response spectrum are obtained.

[0101] Through the topology prediction model, 1*N pixel 0-1 codes are obtained to represent the structure parameters of the device functional region, and N represents the number of pixel regions divided in the training data set. Taking the above training data set divided into 300 pixel regions as an example, N is 300 at this time. Then, according to the structure parameter (pixel) unit code obtained by the output layer, it is determined whether to etch the region; wherein when the corresponding pixel point code is 0, the pixel region is etched, and when it is 1, the pixel region is not etched, so as to obtain the structure of the grating coupler as shown in Figure 2 Further, for the obtained structure, finite difference time domain simulation can be performed to verify the performance indicators of the designed device, including diffraction length distribution and response spectrum, so as to determine whether the performance of the device corresponding to the predicted structure parameters meets the requirements. Based on the performance indicators, the structure parameters are fine-tuned to further improve the performance of the device. The specific simulation steps can refer to the conventional steps, which will not be repeated here.

[0102] The design method of the grating coupler provided in the embodiment combines a simple network structure with a complex nonlinear algorithm through approximate processing of the training data set, reduces the algorithm demand of the model, improves the prediction efficiency and effect of the model, and further improves the prediction accuracy of the model through the training method combining the forward and reverse directions, so as to realize the design of the grating coupler with high coupling efficiency at any angle and ensure the device performance of the grating coupler.

[0103] Please refer to Figure 6 Based on the same inventive concept, the design system of the grating coupler provided in the second embodiment of the present application comprises:

[0104] The model construction module 10 is configured to construct a topological prediction model based on a deep neural network, and train the topological prediction model by using a preset method, wherein the topological prediction model comprises an input layer, an output layer and a plurality of hidden layers, an activation function of the topological prediction model is a Swish function, and a training data set is constructed by using an approximation process to adapt to the convergence of the Swish function.

[0105] The parameter prediction module 20 is configured to input a response spectrum of the grating coupler into the trained topological prediction model to obtain a structure parameter of a functional region of the grating coupler.

[0106] The grating design module 30 is configured to determine a topological structure of the functional region of the grating coupler according to the structure parameter to realize the design of the grating coupler.

[0107] Further, the model construction module 10 comprises a model training module.

[0108] The model training module is configured to generate training data by using a simulation method, and perform an approximation process on the training data to obtain a training data set.

[0109] The training data set is used to train the topological prediction model.

[0110] Further, the model training module further comprises a data set construction module.

[0111] The data set construction module is configured to generate training data with SOI structure consistency by using a time domain finite difference simulation, wherein the training data comprises a structure parameter and a spectrum parameter.

[0112] The functional region of the grating coupler corresponding to the structure parameter is divided into a plurality of pixel regions according to a preset size.

[0113] The data of each pixel region is converted according to a region state to obtain the processed structure parameter, wherein the region state comprises an etching state and a non-etching state.

[0114] The training data set is composed according to the processed structure parameter and the spectrum parameter.

[0115] Further, the model training module is further configured to construct a forward training model with the same structure as the topological prediction model, wherein input data of the forward training model is the structure parameter of the functional region of the grating coupler, and output data of the forward training model is a response spectrum of the grating coupler.

[0116] The training data set is used to train the forward training model to obtain a first network parameter configuration of the forward training model.

[0117] The trained forward training model is connected in series with the untrained topological prediction model to obtain a series connection network model, and the series connection network model is trained using the training data set.

[0118] Second network parameter configurations of the topological prediction model are determined from the trained series connection network model to obtain the trained topological prediction model.

[0119] The technical features and technical effects of the design system of the grating coupler proposed in the embodiments of the application are the same as those of the method proposed in the embodiments of the application, and are not repeated here. Each module in the design system of the grating coupler described above can be realized by software, hardware, and combinations thereof, in whole or in part. Each module described above can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0120] In addition, the embodiments of the application also propose a computer device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the above method when executing the computer program.

[0121] Please refer to Figure 7 , the internal structure diagram of the computer device in an embodiment, which can be a terminal or a server. The computer device includes a processor, a memory, a network interface, a display, and an input device connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The network interface of the computer device is used to communicate with external terminals through network connection. The computer program is executed by the processor to implement the design method of the grating coupler. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball, or touchpad provided on the shell of the computer device. It can also be an external keyboard, touchpad, or mouse, etc.

[0122] Those skilled in the art can understand Figure 7 that the structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. A specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have the same component arrangement.

[0123] In addition, the embodiment of the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the above method.

[0124] To sum up, the embodiment of the present application provides a design method and system of a grating coupler, a computer device and a storage medium. The method constructs a topological prediction model based on a deep neural network, and trains the topological prediction model by using a preset method. The topological prediction model includes an input layer, an output layer and a plurality of hidden layers. The activation function of the topological prediction model is a Swish function. The response spectrum of the grating coupler is input into the trained topological prediction model to obtain the structure parameters of the functional region of the grating coupler. The topological structure of the functional region of the grating coupler is determined according to the structure parameters, so as to realize the design of the grating coupler. The present application combines a simple network structure with a complex nonlinear algorithm through the approximate processing of the training data set, reduces the computing power requirement of the model, improves the prediction efficiency and effect of the model, and further improves the prediction accuracy of the model through the forward and reverse combination training method, so as to realize the design of the grating coupler with high coupling efficiency at any angle, and ensure the device performance of the grating coupler.

[0125] Each embodiment in the specification is described in a progressive manner, and the same or similar parts of each embodiment can be referred to each other. Each embodiment focuses on the difference from other embodiments. Especially, the system embodiment is basically similar to the method embodiment, so the description is relatively simple, and the related parts can be referred to the part of the method embodiment. It should be noted that, each technical feature of the above embodiments can be combined arbitrarily, in order to make the description simple, each technical feature of the above embodiments is not described in all possible combinations, however, as long as the combination of technical features does not exist contradictory, it should be considered as the scope of the present application.

[0126] The above-described embodiments only express several preferred embodiments of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be noted that, for ordinary skilled in the art, without departing from the technical principles of the present application, some improvements and replacements can be made, and these improvements and replacements should be considered as the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the protection scope of the claims.

Claims

1. A method of designing a grating coupler, characterized by, The method comprises the following steps: A topology prediction model is constructed based on a deep neural network, and the topology prediction model is trained by using a preset method, the topology prediction model comprises an input layer, an output layer and a plurality of hidden layers, an activation function of the topology prediction model is a Swish function, a training data set is obtained by using approximate processing, so as to adapt the convergence of the Swish function, each layer of the topology prediction model uses a nonlinear activation function, and an output after activation is represented as: In the formula, represents the output of the lth layer after passing through the nonlinear activation function, F represents the nonlinear activation function, represents the bias vector of the lth layer; The response spectrum of the grating coupler is input into the trained topology prediction model, so that the structure parameters of the functional region of the grating coupler are obtained; According to the structure parameters, the topology structure of the functional region of the grating coupler is determined, so that the design of the grating coupler is realized; The step of training the topology prediction model by using the preset method comprises the following steps: Training data is generated by using a simulation method, and the training data is processed by using approximate processing, so that a training data set is obtained; The topology prediction model is trained by using the training data set; The step of generating training data by using a simulation method and processing the training data by using approximate processing to obtain a training data set comprises the following steps: Training data with SOI structure consistency is generated by using a finite-difference time-domain simulation method, the training data comprises structure training data and spectrum training data; The functional region of the grating coupler corresponding to the structure training data is divided into a plurality of pixel regions according to a preset size; The data of each pixel region is converted according to a region state, so that the processed structure training data is obtained, the region state comprises an etching state and a non-etching state; According to the processed structure training data and the spectrum training data, a training data set is formed.

2. The design method of a grating coupler according to claim 1, wherein, The step of training the topology prediction model by using the training data set comprises the following steps: A forward training model with the same structure as the topology prediction model is constructed, input data of the forward training model is the structure parameters of the functional region of the grating coupler, and output data of the forward training model is the response spectrum of the grating coupler; The forward training model is trained by using the training data set, so that a first network parameter configuration of the forward training model is obtained; The trained forward training model and the untrained topology prediction model are connected in series to obtain a series connection network model, and the series connection network model is trained by using the training data set; The second network parameter configuration of the topology prediction model is determined from the trained series connection network model, so that the trained topology prediction model is obtained.

3. A design system for a grating coupler, characterized by The method comprises the following steps: A model construction module is configured to construct a topology prediction model based on a deep neural network, and train the topology prediction model by using a preset method, the topology prediction model comprises an input layer, an output layer and a plurality of hidden layers, an activation function of the topology prediction model is a Swish function, a training data set is obtained by using approximate processing, so as to adapt the convergence of the Swish function, each layer of the topology prediction model uses a nonlinear activation function, and an output after activation is represented as: In the formula, represents the output of the lth layer after the nonlinear activation function, F represents the nonlinear activation function, represents the bias vector of the lth layer; The parameter prediction module is configured to input a response spectrum of the grating coupler into the trained topological prediction model to obtain a structural parameter of a functional region of the grating coupler. The grating design module is configured to determine a topological structure of the functional region of the grating coupler according to the structural parameter, so as to realize design of the grating coupler. The model construction module comprises a model training module. The model training module is configured to generate training data by using a simulation method, and perform approximate processing on the training data to obtain a training data set; and train the topological prediction model by using the training data set. The model training module further comprises a data set construction module. The data set construction module is configured to generate training data with SOI structure consistency by using a time domain finite difference simulation, and the training data comprises a structural parameter and a spectral parameter. The functional region of the grating coupler corresponding to the structural parameter is divided into a plurality of pixel regions according to a preset size. Each pixel region is subjected to data conversion according to a region state to obtain a processed structural parameter, and the region state comprises an etching state and a non-etching state. The training data set is composed of the processed structural parameter and the spectral parameter.

4. The design system of grating coupler according to claim 3, characterized in that, The model training module is further configured to construct a forward training model with the same structure as the topological prediction model, the input data of the forward training model is the structural parameter of the functional region of the grating coupler, and the output data of the forward training model is the response spectrum of the grating coupler. The training data set is used to train the forward training model to obtain a first network parameter configuration of the forward training model. The trained forward training model and the untrained topological prediction model are connected in series to obtain a series network model, and the training data set is used to train the series network model. The second network parameter configuration of the topological prediction model is determined from the trained series network model to obtain the trained topological prediction model.

5. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the steps of the method in any one of claims 1 to 2.

6. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method in any one of claims 1 to 2.

Images