Reverse design method of DFB laser spectrum parameters based on DNN model

Abstract

The present invention relates to the field of semiconductor lasers and deep learning, and discloses a DFB laser spectral parameter reverse design method based on a DNN model, comprising the following steps: given the spectral parameters of a DFB laser, obtaining its threshold spectrum through a transfer matrix method; extracting characteristic parameters of the threshold spectrum; recording the values ​​of the characteristic parameters and the corresponding spectral parameter values ​​of the DFB laser to generate a data set; using a particle swarm algorithm to initialize the weights and back-propagation errors of the DNN model; inputting the data set into the initialized model for training; given a target threshold spectrum, extracting characteristic parameters; inputting the characteristic parameters into the trained model, outputting the spectral parameters of the DFB laser, and realizing the reverse design of the spectral parameters. The method of the present invention can reduce the tedious operation of using a spectrometer to obtain a threshold spectrum multiple times, improve accuracy, reduce the time used and memory usage, and thus optimize its spectral characteristics faster and more accurately.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor lasers and deep learning, and in particular to a DFB laser spectral parameter reverse design method based on a DNN model. Background Art

[0002] In today's optical communications and optical sensing fields, distributed feedback (DFB) lasers play a vital role. The spectral characteristics of the laser it emits, such as central wavelength, linewidth, side mode suppression ratio and other parameters, directly determine the performance of the system. Traditional DFB laser spectral parameter design methods often rely on complex and time-consuming empirical formula derivation and repeated experiments. R&D personnel need to manually adjust structural parameters such as grating period, coupling coefficient, cavity length, etc. based on a lot of theoretical knowledge to approximate the required spectral characteristics. This forward design process not only requires designers to have a deep professional knowledge reserve and a clear understanding of the intricate physical relationship between different parameters, but also because the parameter space is huge and the experimental process is extremely cumbersome, which makes the design cycle long and the manpower and material costs high.

[0003] As optical communication technology develops rapidly towards high speed and large capacity, the requirements for the spectral performance of DFB lasers are becoming more and more stringent, and traditional design methods are increasingly unable to meet the needs of rapid iteration. For example, in ultra-dense wavelength division multiplexing (DWDM) systems, the laser center wavelength is required to be accurate and stable, the line width is extremely narrow, and the side mode suppression ratio is ultra-high. It is almost impossible to find the optimal parameter combination based on existing methods.

[0004] At the same time, deep learning technology, especially deep neural networks (DNN), has demonstrated powerful fitting and prediction capabilities in many scientific and engineering fields. It can automatically learn hidden patterns and laws from massive data, and has the advantages of high efficiency and intelligence compared to traditional physical model-based design. However, in the design of DFB laser spectral parameters, the application of DNN models is still in the exploratory stage, and a mature and systematic reverse design method has not yet been formed. In-depth research is urgently needed to fill this gap, realize the rapid and accurate design of DFB lasers, and promote the rapid development of industries such as optical communications. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a DFB laser spectral parameter reverse design method based on a DNN model, so as to improve the design accuracy, greatly shorten the time used, and reduce the memory usage.

[0006] To achieve the above object, the technical solution of the present invention is as follows:

[0007] A DFB laser spectral parameter reverse design method based on a DNN model comprises the following steps:

[0008] Step 1, given the spectral parameters of the DFB laser, obtaining its threshold spectrum by using a transfer matrix method, wherein the spectral parameters refer to structural parameters related to the spectrum;

[0009] Step 2, extracting characteristic parameters of the threshold spectrum;

[0010] Step 3, recording the value of the characteristic parameter and the spectral parameter value of the DFB laser corresponding thereto to generate a data set;

[0011] Step 4: Use the particle swarm algorithm to initialize the weights and back propagation errors of the DNN model;

[0012] Step 5, input the data set into the initialized model, use the characteristic parameter value as the model input, and use the corresponding DFB laser spectrum parameter value as the model output to train the model;

[0013] Step 6, given a target threshold spectrum, extracting characteristic parameters of the target threshold spectrum, inputting the characteristic parameters into the trained model, and the model outputs the spectrum parameters of the DFB laser, thereby realizing the reverse design of the spectrum parameters of the DFB laser.

[0014] In the above scheme, in step 1, the spectral parameters of the DFB laser include grating period, effective refractive index, cavity length, grating coupling coefficient, end face reflectivity and its corresponding phase value, phase shift size, phase shift position, group refractive index and threshold gain.

[0015] In the above scheme, in step 1, the specific method of obtaining the threshold spectrum by the transfer matrix method is as follows:

[0016] For multi-segment corrugated DFB laser, the cavity length direction is z direction. The transmission matrix of the segment grating is expressed as:

[0017] ;

[0018] ;

[0019] ;

[0020] ;

[0021] ;

[0022] in, The first Two arbitrary length coordinate values ​​of the segment, and satisfy is the average wave number, is the effective refractive index in the cavity, c is the speed of light, is the Bragg wave number, is the grating coupling coefficient, For the The initial phase of the segment grating, is the detuning constant, is the Bragg angular frequency, is the angular frequency, is the group velocity, For the The net field gain of the segment grating;

[0023] The transmission matrix of an n-segment DFB laser can be written as:

[0024] ;

[0025] After adding the reflectivity of the left and right end faces of the laser, the total transmission matrix of the n-segment DFB laser can be written as:

[0026] ;

[0027] in, are the reflectivity of the left and right end faces of the laser, ; , , , is the transmission matrix Elements in

[0028] Transmission coefficient of multi-segment corrugated DFB laser It can be expressed as:

[0029] ;

[0030] Assume that the Bragg wave number of each segment is , group refractive index , gain are all invariant with spatial distribution, and the spectrum with spontaneous radiation can be expressed as:

[0031] ;

[0032] In the formula, is the angular frequency interval, is the reduced Planck constant, , Respectively Segment grating and The starting length coordinate value of the segment grating, is the solution of the Green's function of the k-th grating;

[0033] The normalized threshold spectrum can be expressed as:

[0034] .

[0035] In the above scheme, the specific method of step 2 is as follows: use mathematical analytical methods, numerical methods or computer software methods to extract the main peak and the five peaks on its left and right, the normalized threshold spectral value size and its corresponding wavelength of the first valley on both sides of the main peak, a total of 13 groups of threshold spectral value sizes and corresponding wavelengths.

[0036] In the above scheme, in step 3, the number of spectral parameter values ​​of the DFB laser selected is 6-8, and the data set size is 20000-50000 sets of data.

[0037] In the above scheme, in step 4, the DNN model includes a 7-layer BP neural network, with a total of 25 input neurons, including 13 wavelength sizes after normalization and 12 spectral values ​​​​except the main peak; there are five hidden layers, each with 30 neurons; the output neuron is the spectral parameter value to be predicted, and the number of output neurons is the number of spectral parameters to be predicted; the first six layers of the model select the sigmoid activation function, and the last layer uses a linear activation function.

[0038] In the above scheme, in step 4, the particle swarm algorithm formula is as follows:

[0039] ;

[0040] ;

[0041] Among them, q represents a particle, , Y is the total number of particles in the population, t represents the tth generation, is the particle speed, is the current particle position, is the inertia weight, are two independent random numbers between 0 and 1, , is the acceleration constant, ranging from 0 to 2. is the current local optimal particle, is the global optimal particle;

[0042] The parameters are initialized to , , , , , , , , , particle position is a one-dimensional array whose length is the sum of the number of weights and the number of reverse errors in the neural network;

[0043] At the beginning, the speed and position of each particle are randomly initialized. The change with the number of iterations can be expressed as:

[0044] ;

[0045] Set a mutation function with a probability of 1 / 10 to prevent the particle swarm from falling into the local optimum, and the error function Set to the MSE error of the neural network at the global optimal position of each generation of particle swarm:

[0046] .

[0047] In the above scheme, in step 5, during the model training process, the optimizer is set to the Levenberg-Marquardt optimization algorithm, the learning rate is 0.001, the error is the MSE error, the number of iterations is 10,000, and the error target is 10 -7 , meet the training objectives or stop training when the MSE error is not optimized for 6 consecutive times.

[0048] Through the above technical solution, the DFB laser spectral parameter reverse design method based on the DNN model provided by the present invention has the following beneficial effects:

[0049] 1. The present invention generates the laser threshold spectrum by using the transfer matrix (TMM) method, which reduces the experimental time and saves costs compared to the measurement using a spectrometer;

[0050] 2. The present invention uses a deep neural network (DNN) model to reversely predict the spectral parameters of the laser. Compared with the traditional empirical model, trial and error method and the design optimization model of the optimization algorithm, the memory and running time required by the successfully trained deep neural network model are greatly reduced, and the requirements for computer equipment are lower; the accuracy is greatly improved compared with the accuracy of the traditional optimization algorithm;

[0051] 3. The present invention uses the particle swarm algorithm (PSO) to initialize the weights and errors of the deep neural network model, which provides a better starting point for the deep neural network, allowing the network to converge faster in the early stages of training; in the case of multiple output neurons with different physical meanings, each neuron requires a different learning rate to optimize its weight. The PSO algorithm can find a more suitable weight for each output neuron through its unique search mechanism, thereby balancing the convergence speed of different output neurons to a certain extent, making the prediction of each output neuron more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below.

[0053] Figure 1 A schematic flow chart of a method for reverse designing DFB laser spectral parameters based on a DNN model disclosed in an embodiment of the present invention;

[0054] Figure 2 for Schematic diagram of the phase-shifted DFB laser structure;

[0055] Figure 3 Schematic diagram of comparison between the predicted threshold spectrum and the target threshold spectrum in Example 1 of the present invention;

[0056] Figure 4 It is a schematic diagram of the standard DFB laser structure;

[0057] Figure 5 Schematic diagram of the comparison between the predicted threshold spectrum and the target threshold spectrum in Example 2 of the present invention.

[0058] In the figure, 1, waveguide layer; 2, active area; 3, left end surface; 4, right end surface. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0060] The present invention provides a DFB laser spectral parameter reverse design method based on a DNN model, such as Figure 1 As shown, the following steps are included:

[0061] Step 1: Given the spectral parameters of the DFB laser, its threshold spectrum is obtained by the transfer matrix method.

[0062] The spectral parameters of a DFB laser refer to the structural parameters related to the spectrum, including grating period, effective refractive index, cavity length, grating coupling coefficient, end face reflectivity and its corresponding phase value, phase shift size, phase shift position, group refractive index and threshold gain.

[0063] The specific method of obtaining its threshold spectrum by the transmission matrix method is as follows:

[0064] For multi-segment corrugated DFB laser, the cavity length direction is z direction. The transmission matrix of the segment grating is expressed as:

[0065] ;

[0066] ;

[0067] ;

[0068] ;

[0069] ;

[0070] in, The first Two arbitrary length coordinate values ​​of the segment, and satisfy is the average wave number, is the effective refractive index in the cavity, c is the speed of light, is the Bragg wave number, is the grating coupling coefficient, For the The initial phase of the segment grating, is the detuning constant, is the Bragg angular frequency, is the angular frequency, is the group velocity, For the The net field gain of the segment grating.

[0071] The transmission matrix of an n-segment DFB laser can be written as:

[0072] ;

[0073] After adding the reflectivity of the left and right end faces of the laser, the total transmission matrix of the n-segment DFB laser can be written as:

[0074] ;

[0075] in, are the reflectivity of the left and right end faces of the laser, ; , , , is the transmission matrix Elements in

[0076] Transmission coefficient of multi-segment corrugated DFB laser It can be expressed as:

[0077] .

[0078] Assume that the Bragg wave number of each segment is , group refractive index , gain are all invariant with spatial distribution, and the spectrum with spontaneous radiation can be expressed as:

[0079] ;

[0080] In the formula, is the angular frequency interval, is the reduced Planck constant, , Respectively Segment grating and The starting length coordinate value of the segment grating, is the solution of the Green's function of the kth grating, and its integral can be expressed as;

[0081] ;

[0082] Among them, the subscript r is the real part of the variable, i is the imaginary part of the variable, and * is the complex conjugate operator. is the propagation constant, is the length of the kth grating, is an intermediate variable;

[0083] ;

[0084] ;

[0085] in, , is an intermediate variable.

[0086] The normalized threshold spectrum can be expressed as:

[0087] .

[0088] This method is also applicable to single-section corrugated DFB lasers, in which case n=1.

[0089] Step 2: Extract characteristic parameters of the threshold spectrum.

[0090] The specific method is as follows: extract the threshold spectrum and its wavelength corresponding to the main peak and its nearby peaks and valleys, for example, use mathematical analytical methods, numerical methods or computer software methods to extract the main peak and five peaks on its left and right, the size of the normalized threshold spectrum value of the first valley on both sides of the main peak and its corresponding wavelength, a total of 13 groups of threshold spectrum value sizes and corresponding wavelengths.

[0091] Mathematical analysis method: Assume that the wavelength is continuous , for , find the main peak value The maximum value is the wavelength corresponding to the wavelength value. Finding the peak value allows right Taking the derivative twice, we get ,when , hour, is the peak value, The valley value can be obtained as the minimum value in the wavelength range between the main peak and the left and right peaks, or when , When The two wavelengths closest to the main peak and its corresponding value.

[0092] Numerical method: Assume that the wavelength can be taken as discrete , for , find the main peak value The value when the value is the maximum, and the wavelength is the corresponding wavelength value. ,and hour, is the peak value, The valley value can be obtained as the minimum value in the wavelength range between the main peak and the left and right peaks, or when ,and When The two wavelengths closest to the main peak , and its corresponding value.

[0093] Computer software method: You can use the findpeaks function in matlab software to find the largest peak, and record the index value to derive the peak value and wavelength. Assuming that the index value corresponding to the largest peak is M (the Mth peak is the main peak), then find its left valley value in the spectral data. You can find the minimum value and its corresponding wavelength between the two. You can use the min function to find the minimum value, and the index value can be found by the find function. Similarly, the right valley value can be found in the spectral data. Search between.

[0094] Step 3: Record the values ​​of the characteristic parameters and the corresponding spectral parameter values ​​of the DFB laser to generate a data set.

[0095] Specifically, repeat steps 1 and 2, and record the 13 groups of LTS values ​​and corresponding wavelengths each time; the spectral parameters of the laser include grating period, effective refractive index, cavity length, grating coupling coefficient, end reflectivity and its corresponding phase value, phase shift size, phase shift position, group refractive index and threshold gain and other spectrum-related parameters. In order to avoid multiple solutions, it is not recommended to set all parameters as variable parameters. It is recommended to select 6-8 spectral parameters, and avoid changing parameters that may produce the same results at the same time or the convergence speed of a parameter is much slower than that of other parameters. The size of each training data set is about 20,000-50,000 groups.

[0096] Step 4: Use the particle swarm algorithm to initialize the weights and back propagation errors of the DNN model.

[0097] The DNN model includes a 7-layer BP neural network with a total of 25 input neurons, including 13 normalized wavelength sizes and 12 spectral values ​​​​except the main peak; there are five hidden layers, each with 30 neurons; the output neurons are the spectral parameter values ​​​​to be predicted, and the number of output neurons is the number of spectral parameters that need to be predicted.

[0098] The first six layers of the model use the sigmoid activation function, and the last layer uses the linear activation function. The sigmoid activation function expression is:

[0099] ;

[0100] The expression of the linear activation function is:

[0101] .

[0102] The particle swarm algorithm formula is as follows:

[0103] ;

[0104] ;

[0105] Among them, q represents a particle, , Y is the total number of particles in the population, t represents the tth generation, is the particle speed, is the current particle position, is the inertia weight, are two independent random numbers between 0 and 1, , is the acceleration constant, ranging from 0 to 2. is the current local optimal particle, is the global optimal particle;

[0106] The parameters are initialized to , , , , , , , , , particle position is a one-dimensional array whose length is the sum of the number of weights and the number of reverse errors in the neural network;

[0107] At the beginning, the speed and position of each particle are randomly initialized. The change with the number of iterations can be expressed as:

[0108] ;

[0109] Set a mutation function with a probability of 1 / 10 to prevent the particle swarm from falling into the local optimum, and the error function Set to the MSE error of the neural network at the global optimal position of each generation of particle swarm:

[0110] .

[0111] Step 5: Input the data set into the initialized model, use the characteristic parameter value as the model input, and use the corresponding spectral parameter value of the DFB laser as the model output to train the model.

[0112] During the model training process, the optimizer is set to the Levenberg-Marquardt optimization algorithm, the learning rate is 0.001, the error is the MSE error, the number of iterations is 10,000, and the error target is 10 -7 , meet the training objectives or stop training when the MSE error is not optimized for 6 consecutive times.

[0113] Step 6, given a target threshold spectrum, extracting characteristic parameters of the target threshold spectrum, inputting the characteristic parameters into the trained model, and the model outputs the spectrum parameters of the DFB laser, thereby realizing the reverse design of the spectrum parameters of the DFB laser.

[0114] Application Example 1

[0115] The method provided by the present invention is applied to AR / AR coating In the reverse design of the spectral parameters of the phase-shifted DFB laser, The phase-shifted DFB laser structure is as follows: Figure 2 As shown, The phase-shifted DFB laser has a waveguide layer 1 and an active region 2 inside, and a left end face 3 and a right end face 4 on both sides.

[0116] The phase of the fixed end face reflectivity is 0, the phase shift position is at 1 / 2 of the cavity length, and the preset range of spectral parameters is shown in Table 1:

[0117] Table 1 Preset spectral parameter ranges

[0118]

[0119] Under this parameter setting condition, 20,000 sets of spectra were generated by the TMM method, of which 18,000 sets were used to train the neural network and 2,000 sets were used as test set data; the final neural network mean square error (MSE) was 3.7×10 -5 The training results and the comparison of randomly selected test set data are shown in Table 2.

[0120] Table 2 Comparison of preset spectral parameters and model output spectral parameter values

[0121]

[0122] From the parameter comparison in Table 2, it can be seen that the accuracy of the spectral parameters output by the DNN model in the present invention is very high and close to the preset spectral parameters.

[0123] Next, the model output spectral parameters are used to obtain the predicted threshold spectrum through the TMM method. The results are as follows: Figure 3 As shown, the target threshold spectrum is the solid line, and the predicted threshold spectrum is the dotted line in the figure. It can be seen that the two are very close.

[0124] Application Example 2

[0125] The method provided by the present invention is applied to the reverse design of the spectrum parameters of the standard DFB laser with HR / AR coating. The structure of the standard DFB laser is as follows: Figure 4 As shown, the standard DFB laser has a waveguide layer 1 and an active region 2 inside, and a left end face 3 and a right end face 4 on both sides.

[0126] The reflectivity of the right end face is fixed to 0.005, the phase is 0, the reflectivity of the left end face is fixed to 0.95, and the parameter preset range is shown in Table 3.

[0127] Table 3 Preset spectral parameter ranges

[0128]

[0129] Under this parameter setting condition, 20,000 sets of spectra were generated by the TMM method, of which 18,000 sets were used to train the neural network and 2,000 sets were used as test set data; the final neural network mean square error (MSE) was 2.2×10 -6 The training results and the comparison of randomly selected test set data are shown in Table 4.

[0130] Table 4 Comparison of preset spectral parameters and model output spectral parameter values

[0131]

[0132] From the parameter comparison in Table 4, it can be seen that the accuracy of the spectral parameters output by the DNN model used in the present invention is very high.

[0133] Next, the model output spectral parameters are used to obtain the predicted threshold spectrum through the TMM method. The results are as follows: Figure 5 As shown, the target threshold spectrum is the solid line, and the predicted threshold spectrum is the dotted line in the figure. It can be seen that the two are very close.

[0134] The present invention generates a data set by TMM method, adopts a deep neural network model, and proposes a DFB laser spectral parameter reverse design method based on DNN model. The PSO optimization initialization model is used, and the parameters generated by the model are compared with the preset spectrum by using TMM to generate a threshold spectrum. The advantages of the present invention are low time cost and space cost, fast calculation speed, high calculation accuracy, and low requirements for computer equipment. The present invention can be applied to the design optimization and parameter extraction fields of semiconductor lasers.

[0135] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A DFB laser spectral parameter reverse design method based on a DNN model, characterized in that: The steps include: Step 1, given the spectral parameters of the DFB laser, obtaining its threshold spectrum by using a transfer matrix method, wherein the spectral parameters refer to structural parameters related to the spectrum; Step 2, extracting characteristic parameters of the threshold spectrum; Step 3, recording the value of the characteristic parameter and the spectral parameter value of the DFB laser corresponding thereto to generate a data set; Step 4: Use the particle swarm algorithm to initialize the weights and back propagation errors of the DNN model; Step 5, input the data set into the initialized model, use the characteristic parameter value as the model input, and use the corresponding DFB laser spectrum parameter value as the model output to train the model; Step 6, given a target threshold spectrum, extracting characteristic parameters of the target threshold spectrum, inputting the characteristic parameters into the trained model, and the model outputs the spectrum parameters of the DFB laser, thereby realizing the reverse design of the spectrum parameters of the DFB laser; In step 1, the spectral parameters of the DFB laser include grating period, effective refractive index, cavity length, grating coupling coefficient, end face reflectivity and its corresponding phase value, phase shift magnitude, phase shift position, group refractive index and threshold gain; The specific method of step 2 is as follows: use mathematical analytical methods, numerical methods or computer software methods to extract the main peak and its five peaks on the left and right, the normalized threshold spectral value size and its corresponding wavelength of the first valley on both sides of the main peak, a total of 13 groups of threshold spectral value size and corresponding wavelength.

2. According to claim 1, a DFB laser spectral parameter reverse design method based on a DNN model is characterized in that: In step 1, the specific method of obtaining the threshold spectrum by the transfer matrix method is as follows: For multi-segment corrugated DFB lasers, the cavity length is in the z direction. The transmission matrix of the segment grating is expressed as: ； ； ； ； ； in, The first Two arbitrary length coordinate values ​​of the segment grating, and satisfy is the average wave number, is the effective refractive index in the cavity, c is the speed of light, is the Bragg wave number, is the grating coupling coefficient, For the The initial phase of the segment grating, is the detuning constant, is the Bragg angular frequency, is the angular frequency, is the group velocity, For the The net field gain of the segment grating; The transmission matrix of an n-segment DFB laser can be written as: ； After adding the reflectivity of the left and right end faces of the laser, the total transmission matrix of the n-segment DFB laser can be written as: ； in, are the reflectivity of the left and right end faces of the laser, , ; , , , is the transmission matrix Elements in Transmission coefficient of multi-segment corrugated DFB laser It can be expressed as: ； Assume that the Bragg wave number of each segment is , group refractive index , gain are all invariant with spatial distribution, and the spectrum with spontaneous radiation can be expressed as: ； In the formula, is the angular frequency interval, is the reduced Planck constant, , Respectively Segment grating and The starting length coordinate value of the segment grating, is the solution of the Green's function of the k-th grating; The normalized threshold spectrum can be expressed as: 。 3. The method for reverse designing DFB laser spectrum parameters based on DNN model according to claim 1, characterized in that: In step 3, the number of spectral parameter values ​​of the DFB laser selected is 6-8, and the data set size is 20000-50000 sets of data.

4. The method for reverse designing DFB laser spectrum parameters based on DNN model according to claim 1, characterized in that: In step 4, the DNN model includes a 7-layer BP neural network, with a total of 25 input neurons, including 13 wavelength sizes after normalization and 12 spectral values ​​​​except the main peak; there are five hidden layers, each with 30 neurons; the output neuron is the spectral parameter value to be predicted, and the number of output neurons is the number of spectral parameters to be predicted; the first six layers of the model select the sigmoid activation function, and the last layer uses a linear activation function.

5. The method for reverse designing DFB laser spectrum parameters based on DNN model according to claim 1, characterized in that: In step 4, the particle swarm algorithm formula is as follows: ； ； Among them, q represents a particle, , Y is the total number of particles in the population, t represents the tth generation, is the velocity of the particle, is the current particle position, is the inertia weight, are two independent random numbers between 0 and 1, , is the acceleration constant, ranging from 0 to 2. is the current local optimal particle, is the global optimal particle; The parameters are initialized to , , , , , , , , , particle position is a one-dimensional array whose length is the sum of the number of weights and the number of reverse errors in the neural network; At the beginning, the speed and position of each particle are randomly initialized. The change with the number of iterations can be expressed as: ； Set a mutation function with a probability of 1 / 10 to prevent the particle swarm from falling into the local optimum, and the error function Set to the MSE error of the neural network at the global optimal position of each generation of particle swarm: 。 6. The method for reverse designing DFB laser spectrum parameters based on DNN model according to claim 1, characterized in that: In step 5, during the model training process, the optimizer is set to the Levenberg-Marquardt optimization algorithm, the learning rate is 0.001, the error is the MSE error, the number of iterations is 10,000, and the error target is 10 -7 , meet the training objectives or stop training when the MSE error is not optimized for 6 consecutive times.

Images