Simulation optical link topological structure implementation method and system based on deep learning

Abstract

The invention belongs to the technical field of analog optical link topological structures, and particularly relates to an analog optical link topological structure implementation method and system based on deep learning, and the method comprises the steps: S1, abstracting different photonic devices or sub-links into sub-modules with specific transmission coefficients, the output response of the total target is regarded as superposition of respective output responses of a plurality of optical paths between an input port and an output port, and each optical path is randomly arranged by different numbers of photonic devices, so that a simulated optical link topological structure is constructed; s2, encoding the topological structure of the analog optical link to obtain a unified parameterized matrix; and S3, a topological structure closest to the target response is searched by using the genetic algorithm, the forward network model is used as a rapid link topological structure output performance evaluation module, and the search efficiency of the optimization algorithm is improved. The method has the advantages that the automatic design of the analog optical link topological structure is realized for the first time under the condition that the physical parameter change of the photonic device is not considered.

Description

Technical Field

[0001] The present application belongs to the technical field of simulated optical link topology structure, and specifically relates to a method and system for realizing simulated optical link topology structure based on deep learning. Background Art

[0002] Most current automatic topology design methods for analog circuits use graph coding or structural coding to abstract devices or subcircuits into nodes or functional modules, and then use optimization algorithms to design topological structures. However, they can only be applied to simple circuit structures such as low-order amplifiers, and it is difficult to expand to more complex circuit designs. In general, most of the research on topological structure generation of analog integrated circuits is to search and select from simple topological structure libraries, and after the structure is abstracted or encoded, it is separated from the whole to the part or integrated from the part to the whole according to the connection rules. The corresponding relationship modeling between structure and output is not implemented, and the design process becomes more complicated as the design scale increases. Summary of the invention

[0003] This application abstracts different photonic devices or sub-links into sub-modules with specific transmission coefficients, and the output response of the total target can be regarded as the superposition of the output responses of multiple optical paths between the input port and the output port. Different numbers of photonic devices are randomly arranged on each optical path. This construction idea is suitable for optical links of different scales. Therefore, we propose a topological structure encoding method based on this idea. After determining the maximum number of devices and the maximum number of optical paths, we unify them into a fixed-size matrix expression, establish a network model between the topological structure parameters and the output response of the simulated optical link, and further use the genetic optimization algorithm to achieve the optimal topological structure generation under a given target response. The technical solution is as follows:

[0004] A method for implementing a simulated optical link topology structure based on deep learning, comprising the following steps:

[0005] S1. Different photonic devices or sub-links are abstracted into sub-modules with specific transmission coefficients, and the output response of the overall target is regarded as the superposition of the output responses of multiple optical paths from the input port to the output port. Different numbers of photonic devices are randomly arranged on each optical path to construct a simulated optical link topology.

[0006] S2. Encode the topological structure of the simulated optical link to obtain a unified parameterized matrix;

[0007] S3. Genetic algorithm is used to search for the topology closest to the target response, and the forward network model is used as a fast link topology output performance evaluation module to improve the search efficiency of the optimization algorithm.

[0008] Preferably, the optical link topology structure includes multiple different photonic devices and multiple branch splitters and combiners. Different photonic devices are replaced by single micro-ring structures with different and fixed physical parameters. The different topological structures formed have two input and output ports as a whole, and the internal parts are free combinations of devices, thereby forming different device connection structures with different output responses.

[0009] Preferably, in step S2, the component devices are treated as a single node according to the graph encoding method, and the parameterization matrix is ​​as follows:

[0010]

[0011] The value of each element in the parameterized matrix is ​​0 or 1, indicating whether there is a unit device represented by the column index on the optical path represented by the row; assuming that all branch devices are evenly divided, any connection structure can ultimately be equivalent to a matrix of size M*N.

[0012] Preferably, the total transmission coefficient of the simulated optical link topology can be expressed as:

[0013]

[0014] In the formula, H n represents the transmission coefficient of a unit device, where only a mn The corresponding matrix elements with a value of ≠0 participate in the formula operation; due to the equivalent expansion, multiple actual connection structures may be generated according to the parameter matrix, but these structures are equivalent and have a unique total transmission coefficient, which does not affect the correspondence of the parameter matrix to structures with different performances.

[0015] Preferably, in step S3,

[0016] S31. Initialize the population size and regard the parameterized matrix p as an individual in the population;

[0017] S32. Set the number of iterations; if the number of iterations reaches the set threshold and no individual that meets the conditions appears, reinitialize the population and search for the global optimal solution again;

[0018] S33. Calculate individual fitness using absolute error;

[0019] S34. Determine the size of the absolute error and the error threshold. If the absolute error is less than the error threshold, output the individual parameters and end this iteration; if the error is not less than the error threshold, use the genetic operator to process, generate a new population, and repeat S32-S34.

[0020] Preferably, in step S32, based on topological coding, a forward network model between the input structure parameters and the corresponding output responses is established, and the network model used is a fully connected network;

[0021] The loss function for network training is the mean square error between the true target response and the response predicted by the forward network:

[0022]

[0023] Among them, y i is the target response for the input, is the predicted response of the output structure, and m is the number of samples. During the network training process, the network will calculate the loss size of the input data sample, and then use the back propagation method to optimize and adjust the weights and biases of each layer in the neural network, and obtain a network model with the minimum error through multiple rounds of iterations.

[0024] A system for realizing a simulated optical link topology structure based on deep learning, comprising a data acquisition module, a data processing module and an output module;

[0025] Data acquisition module: collects data information of optical link topology structure;

[0026] Data processing module: Encode the topology of the simulated optical link to obtain a unified parameterized matrix; use the genetic algorithm to search for the topology closest to the target response, and use the forward network model as a fast link topology output performance evaluation module to improve the search efficiency of the optimization algorithm;

[0027] Output module: output optimal topology parameters.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. Encode the topological structure of the simulated optical link to obtain a unified parameterized expression. Compared with the existing graph encoding method, it reduces the possibility of multiple parameter encodings corresponding to the same topological structure, and can more conveniently convert parameter encodings to analytical formulas and actual topological structures;

[0030] 2. After the topology structure of the simulated optical link is parameterized, a forward network model between different topologies and output responses is established, realizing the mapping expression between topology parameters and output responses. Compared with the process of structure building and then simulation calculation in the simulation software, performance evaluation can be carried out more quickly.

[0031] 3. Use genetic algorithms to search for the topological structure that is closest to the target response, and use the forward network model as a fast performance estimation part of the topological structure to improve the search efficiency of the optimization algorithm;

[0032] 4. The applicability of this method under different numbers of devices has been verified, and the forward network model still has a good response fitting effect for the topological structure with different numbers of optical devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The two-way split simulates the topology of the optical link.

[0034] Figure 2 It is a forward network model.

[0035] Figure 3 is the error change during the training process.

[0036] Figure 4 Comparison between the true response (original) and the model predicted response (ForwardNet).

[0037] Figure 5 For error.

[0038] Figure 6 It is a new topology structure; (a) a four-way topology structure; (b) a three-way topology structure.

[0039] Figure 7 For generalization test, (a) is the equivalent extension (Code1) and (b) is the zero-filling extension (Code2).

[0040] Figure 8 Implement a process for the genetic algorithm. DETAILED DESCRIPTION

[0041] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0042] A method for implementing a simulated optical link topology structure based on deep learning, comprising the following steps:

[0043] S1. Different photonic devices or sub-links are abstracted into sub-modules with specific transmission coefficients, and the output response of the overall target is regarded as the superposition of the output responses of multiple optical paths from the input port to the output port. Different numbers of photonic devices are randomly arranged on each optical path to construct a simulated optical link topology.

[0044] S2. Encode the topological structure of the simulated optical link to obtain a unified parameterized matrix;

[0045] In step S2, the component devices are treated as single nodes according to the graph encoding method, and the parameterization matrix is ​​as follows:

[0046]

[0047] The value of each element in the parameterized matrix is ​​0 or 1, indicating whether there is a unit device represented by the column index on the optical path represented by the row; assuming that all branch devices are evenly divided, any connection structure can ultimately be equivalent to a matrix of size M*N.

[0048] The total transmission coefficient of the simulated optical link topology can be expressed as:

[0049]

[0050] In the formula, H n represents the transmission coefficient of a unit device, where only a mn The corresponding matrix elements that are ≠ 0 participate in the formula operation. Due to the equivalent expansion, multiple actual connection structures may be generated according to the parameter matrix, but these structures are equivalent and have a unique total transmission coefficient, which does not affect the correspondence of the parameter matrix to structures with different performances.

[0051] Based on the significant optimization effect of the forward network model, the inverse design of the simulated optical link structure under specific target response is further realized. Since the parameterization of the topological structure is mainly expressed by 0 and 1. This structural parameter is equivalent to the effect after binary encoding, and the genetic algorithm can be naturally used for global optimization search. The implementation process of this optimization algorithm can skip the encoding step, and its specific implementation steps are as follows: Figure 8 As shown:

[0052] S3. Genetic algorithm is used to search for the topology closest to the target response, and the forward network model is used as a fast link topology output performance evaluation module to improve the search efficiency of the optimization algorithm.

[0053] S31. Initialize the population size and regard the parameterized matrix p as an individual in the population;

[0054] S32. Set the number of iterations; if the number of iterations reaches the set threshold and no individual that meets the conditions appears, reinitialize the population and search for the global optimal solution again;

[0055] S33. Calculate individual fitness using absolute error;

[0056] S34. Determine the size of the absolute error and the error threshold. If the absolute error is less than the error threshold, output the individual parameters and end this iteration; if the error is not less than the error threshold, use the genetic operator to process, generate a new population, and repeat S32-S34.

[0057] In step S33, based on topological coding, a forward network model between input structural parameters and corresponding output responses is established, and the network model adopted is a fully connected network;

[0058] The loss function for network training is the mean square error between the true target response and the response predicted by the forward network:

[0059]

[0060] Among them, y i is the target response for the input, is the predicted response of the output structure.

[0061] During the network training process, the network will calculate the loss size of the input data samples, and then use the backpropagation method to optimize and adjust the weights and biases of each layer in the neural network, and obtain a network model with minimal error through multiple rounds of iterations.

[0062] Here, the trained forward network model replaces the original electromagnetic simulation process, so as to achieve fast individual fitness evaluation and speed up the algorithm operation. As a global optimization algorithm, the genetic algorithm has a slow search speed, especially when the error threshold is low. In order to reduce the difference in iteration rate caused by different initialization populations, we set the population reinitialization operation. In the genetic generation, that is, when the number of cycles reaches 500 times, if no individual that meets the conditions appears, the population is reinitialized and the global optimal solution is searched again. Generally, the setting of the initialization value will affect the global search. Periodically adjusting the initialization value can expand the search range and speed up the global optimal search process. Some hyperparameters of the genetic algorithm need to be set in advance, and the setting values ​​are shown in Table 3.

[0063] Table 3. Genetic algorithm hyperparameter settings

[0064]

[0065] According to the above optimization algorithm, we randomly selected 100 target responses in the test data set and brought them into the genetic algorithm for optimization search. The final results are shown in Table 4.

[0066] Table 4. Genetic algorithm operation results

[0067]

[0068] Here, the response spectra of 100 test samples were selected as the optimization target, and the genetic algorithm optimization process was performed 100 times, which took about 2 hours in total. Among the 100 optimized design structural parameters, 82 are exactly the same as the actual structural parameters, and among the different output results, there are only 2 misalignments at most, but this does not affect the similarity between the output response and the optimization target.

[0069] The optical link topology includes multiple different photonic devices and multiple branch splitters and combiners. Different photonic devices are replaced by single micro-ring structures with different and fixed physical parameters. The different topological structures formed have two input and output ports as a whole, and the internal components are free combinations of components, thus forming different device connection structures with different output responses.

[0070] A system for realizing a simulated optical link topology structure based on deep learning, comprising a data acquisition module, a data processing module and an output module;

[0071] Data acquisition module: collects data information of optical link topology structure;

[0072] Data processing module: Encode the topology of the simulated optical link to obtain a unified parameterized matrix; use the genetic algorithm to search for the topology closest to the target response, and use the forward network model as a fast link topology output performance evaluation module to improve the search efficiency of the optimization algorithm;

[0073] Output module: output optimal topology parameters.

[0074] Embodiment 1:

[0075] The entire link consists of ten different photonic devices and two Y-branch beam splitters and combiners. Different photonic devices are replaced by single micro-ring structures with different and fixed physical parameters. The different topological structures formed have two input and output ports as a whole. The internal components are freely combined, thus forming different device connection structures with different output responses. The Y-branch beam splitters and combiners are 1:1 equally divided, and each photonic device is represented by a different transmission coefficient: H 1 ,H 2 ,H 3 ,H 4 ,H 5 ,H 6 ,H 7 ,H 8 ,H 9 ,H 10 The form of the analog link is as follows: Figure 1 shown. Figure 1 The overall transmission coefficient H of the link structure is:

[0076]

[0077] For the coding of this topological structure, the component devices are regarded as single nodes according to the graph coding method, which can be parameterized into the form of adjacency matrix. Figure 1 The undirected graph adjacency matrix of the structure can be expressed as:

[0078]

[0079] The matrix size of formula (2) is 10×10, such as Figure 1 Medium H 1 and H 4 If the devices are connected, the element at the 4th column and the 1st row in the matrix is ​​1, and 0 indicates no connection. The size of this parameter matrix is ​​related to the number of devices, but it is easy to cause multiple different parameter matrices to correspond to a topological structure, such as Figure 1 Medium H 1 and H 4 , H 3 and H 5 The swapping of positions does not affect the overall structure and output response, but it manifests itself in a different adjacency matrix form. Therefore, we propose a more concise parameterization method that does not affect the position swapping of some devices in series. Figure 1 Taking the structure as an example, its parameterization result can be as follows:

[0080]

[0081] The encoding form in formula (3) is related to the form of formula (1). The two row vectors correspond to the two terms in formula (1) respectively. Each row vector has 10 elements, corresponding to 10 unit devices. The element 1 in the first row indicates that the first item contains the device H corresponding to the column index index. index ,0 means that the device is not included. It can also correspond to the actual topological structure. When the two rows of elements are compared, if both rows in the same column are 1, it means that the device is on the main road, and if only one row is 1, it means that the device is on a branch road. This parameterization method can represent the devices on each path between the input and output ports, regardless of the order of the front and back ports. It solves the problem of position exchange in the adjacency matrix affecting the parameterization results and reduces the multiple values ​​that appear in the parameterization process.

[0082] Based on Figure 1 The link structure is shaped like the topological encoding of formula (3), and a forward network model is established between the input structure parameters and the corresponding output response. The input parameters are the 20 element values ​​after flattening formula (3), and the output is a spectrum discretized into 500 values. The network model used is a commonly used fully connected network, and the model structure is as follows Figure 2 As shown, it is equivalent to a forward simulation operation, which is used to compare the performance of the newly designed structure with the target performance.

[0083] The settings of some hyperparameters of the network model are shown in Table 1:

[0084] Table 1. Model hyperparameter settings

[0085]

[0086] In this supervised learning network model, the data set needs to obtain data samples of features and labels of different optical link topology parameters and corresponding output response components. The topology parameters are randomly generated by the program as shown in formula (3). Then the structure parameters are converted into the expression of the total transmission coefficient of the corresponding link structure, and the corresponding output response spectrum is calculated. The total calculated data set size is 15,000, which is divided into training set, validation set and test set in a ratio of 3:1:1. The training set is used to train the network model, the validation set is used to check whether the model is overfitting during the training process, and determine the model with the lowest validation error. The test set is used to test the performance of the final network model.

[0087] During the network training process, the network will calculate the loss size of the input data sample, and then use the back propagation method to optimize and adjust the weights and biases of each layer in the neural network. Through multiple rounds of iterations, a network model with the minimum error is obtained, such as Figure 3 As shown in the figure, the error in the training process continues to decrease, and the lowest test error is about 4.1e-6.

[0088] After obtaining the network model with the lowest verification error, the topological structure parameters in part of the test set are input into the model for forward testing. The response of the designed structure is compared with the response of the real structure, and the magnitude of the absolute error (MAE) between the two is used to represent the test results of part of the test set samples. Figure 4 shown.

[0089] Figure 4 What is shown is that four target responses (original) are randomly selected from the test set. After the corresponding topological structure parameters are input into the forward network model, the corresponding predicted performance effect will be output. Then these four target responses are compared with the model prediction effect. They have extremely high similarity in shape and error. The model can effectively obtain the mapping relationship between the input topological structure parameters and the output response, and realize the rapid testing of structural performance.

[0090] Although Figure 4The displayed effect is good, but there is a particularity. The network training only considers the link structure that appears when all 10 devices and 2 Y-branch beam splitters are used, and lacks verification of whether the network model is universal when the number of devices in the link structure is different. Therefore, a new data set is added here, which contains link structures composed of different numbers of devices. The data distribution is shown in Table 2.

[0091] Table 2. Distribution of different types of data in the dataset

[0092]

[0093] The data of groups 4, 5, 7, 9, and 10 are mixed and then divided into training and test sets. The data of groups 6 and 8 are not involved in the training and are used as new test data to verify the generalization of the model. The structure and hyperparameter settings of the network remain unchanged. The error changes in the training process and the test results of the test set data are shown in Figure 2. Figure 5 shown.

[0094] Figure 5 (a) The optimal test error after network training is 4.6e-6, which is Figure 3 After the network training is completed, the test set of mixed data and two new sets of data 6 and 8 are input into the network model to verify its performance generalization, as shown in Figure 5 As shown in (b), loss_devices_mix represents the test set divided from the mixed data set, loss_devices_6 and loss_devices_8 are two groups of data that did not participate in the training, each with 2500 data. After being input into the network model, the overall error distribution is obtained. The solid horizontal line in the box represents the median, which is 5.3e-4, 4.9e-4, and 6.1e-4. The dotted horizontal line represents the average, which is 6.8e-4, 5.9e-4, and 7.5e-4. The error distribution of these three groups of test data is very similar, which can show that the network model is universal for simulating different numbers of components in optical links.

[0095] At the same time, considering the combination design of more complex link topology structures, verifying the effectiveness of the forward model in a variety of complex structures, two other types of connection structures are designed here, such as Figure 6 shown.

[0096] The two structures have three and four light paths to synthesize the final output light field. The overall transmission coefficients of the two structures are:

[0097]

[0098] According to the new structural encoding method, the obtained structural parameterization matrix form is:

[0099]

[0100] The two parameter matrices P a , P b The number of rows and the transmission coefficient H a , H b The number of monomials is the same, and the column index of the position of 1 in each row indicates the transmission coefficient of the corresponding device in the corresponding monomial. Corresponding to the actual connection structure, the number of all optical paths from input to output is determined according to the number of rows, and the number and position of 1 in each column determine whether the device is shared by all paths, that is, all 1, or several paths share 1, that is, the device is distributed in the total path after these paths are aggregated, and there is only one 1, that is, the path only passes through this one device from input to output. Although more than one structure that meets the conditions may be generated according to the matrix parameters, the matrix parameters correspond one by one to the overall transmission coefficient of the structure, so these structures must have the same output response and are functionally equivalent.

[0101] Here, the sizes of the two structural parameter matrices are different, which does not meet the requirement of uniform network input parameter size. Therefore, it is necessary to expand the smaller parameter matrix to the maximum matrix size, and the increase from small to large can fully retain the encoded combination structure information. Matrix expansion usually uses the zero padding method, but it is obvious that all zero rows have no practical meaning. Therefore, an equivalent expansion method is proposed here, such as Figure 6 In (b), the three devices on the original single branch can be replaced by the content pointed by the arrow: the original three devices are copied into two identical branches. This method expands the branch without affecting the overall output effect. At this time, the new transmission coefficient is

[0102] Compared with the transmission coefficient in the original formula (6), the 1 / 2 coefficient term is equivalently decomposed into two identical 1 / 4 coefficient terms, which does not change the overall output response. Its parameterized matrix becomes

[0103]

[0104] That is, copy the last row of the matrix to a new row. This extension has a practical structural equivalence and is conducive to the generalization verification of the model on different structures. In order to verify the effectiveness of this method, Figure 1 and Figure 6 The three topological structures are connected and the data sampling of different outputs is carried out. Figure 6 The two structures of the data set are mixed and used to train a new forward network model. Figure 1 After the structure of is equivalently expanded, it is used as new test data to verify the generalization performance of the network model. At the same time, we compare the zero-filling expansion method with the equivalent expansion method. The test structure is as follows:

[0105] like Figure 7 As shown, under the equivalent extension method, Figure 6 The error performance of the model trained with the two mixed data in its own test set (Code1_mixed) is the same as Figure 1 The test errors of the structure (Code1_new) in this model are relatively similar, proving that its model has good generalization, while the model generalization under zero-filling coding is poor. Therefore, we can assume that, given any number N of devices and the maximum number M of possible branching paths of the connection structure, any connection structure under this condition can be represented by a matrix of size M*N.

[0106] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses or adaptations of the present invention that follow the general principles of the present invention and include common knowledge or customary techniques in the art that are not invented by the present invention. The specification and examples are to be considered exemplary only, and the true scope and spirit of the present invention are indicated by the following claims.

[0107] It should be understood that the present invention is not limited to the exact construction that has been described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. A method for implementing a simulated optical link topology structure based on deep learning, characterized in that: The following steps are involved: S1. Different photonic devices or sub-links are abstracted into sub-modules with specific transmission coefficients, and the output response of the overall target is regarded as the superposition of the output responses of multiple optical paths from the input port to the output port. Different numbers of photonic devices are randomly arranged on each optical path to construct a simulated optical link topology. S2. Encode the topological structure of the simulated optical link to obtain a unified parameterized matrix; S3. Genetic algorithm is used to search for the topology closest to the target response, and the forward network model is used as a fast link topology output performance evaluation module to improve the search efficiency of the optimization algorithm.

2. The method for implementing a simulated optical link topology structure based on deep learning according to claim 1, characterized in that: The optical link topology includes multiple different photonic devices and multiple branch splitters and combiners. Different photonic devices are replaced by single micro-ring structures with different and fixed physical parameters. The different topological structures formed have two input and output ports as a whole, and the internal components are free combinations of components, thus forming different device connection structures with different output responses.

3. The method for implementing a simulated optical link topology structure based on deep learning according to claim 1, characterized in that: In step S2, the component devices are treated as single nodes according to the graph encoding method, and the parameterization matrix is ​​as follows: The value of each element in the parameterized matrix is ​​0 or 1, indicating whether there is a unit device represented by the column index on the optical path represented by the row; assuming that all branch devices are evenly divided, any connection structure can ultimately be equivalent to a matrix of size M*N.

4. The method for implementing a simulated optical link topology structure based on deep learning according to claim 3, characterized in that: The total transmission coefficient of the simulated optical link topology can be expressed as: In the formula, H n represents the transmission coefficient of a unit device, where only a mn The corresponding matrix elements with a value of ≠0 participate in the formula operation; due to the equivalent expansion, multiple actual connection structures may be generated according to the parameter matrix, but these structures are equivalent and have a unique total transmission coefficient, which does not affect the correspondence of the parameter matrix to structures with different performances.

5. The method for implementing a simulated optical link topology structure based on deep learning according to claim 1, characterized in that: In step S3, S31. Initialize the population size and regard the parameterized matrix p as an individual in the population; S32. Set the number of iterations; if the number of iterations reaches the set threshold and no individual that meets the conditions appears, reinitialize the population and search for the global optimal solution again; S33. Calculate individual fitness using absolute error; S34. Determine the size of the absolute error and the error threshold. If the absolute error is less than the error threshold, output the individual parameters and end this iteration; if the error is not less than the error threshold, use the genetic operator to process, generate a new population, and repeat S32-S34.

6. The method for implementing a simulated optical link topology structure based on deep learning according to claim 5, characterized in that: In step S32, based on topological coding, a forward network model between input structural parameters and corresponding output responses is established, and the network model adopted is a fully connected network; The loss function for network training is the mean square error between the true target response and the response predicted by the forward network: Where yi is the target response of the input, is the predicted response of the output structure and m is the number of samples. During the network training process, the network will calculate the loss size of the input data samples, and then use the backpropagation method to optimize and adjust the weights and biases of each layer in the neural network, and obtain a network model with minimum error through multiple rounds of iterations.

7. A system for implementing a simulated optical link topology structure based on deep learning, using any of the methods for implementing a simulated optical link topology structure based on deep learning according to any of claims 1 to 6, characterized in that: It includes a data acquisition module, a data processing module and an output module; Data acquisition module: collects data information of optical link topology structure; Data processing module: Encode the topology of the simulated optical link to obtain a unified parameterized matrix; use the genetic algorithm to search for the topology closest to the target response, and use the forward network model as a fast link topology output performance evaluation module to improve the search efficiency of the optimization algorithm; Output module: output optimal topology parameters.

Images