A DFB laser parameter extraction method based on multi-objective optimization algorithm

Abstract

The present invention relates to the field of laser parameter extraction and discloses a DFB laser parameter extraction method based on a multi-objective optimization algorithm. The method comprises: constructing and training two deep learning models; using an instrument to test a DFB laser whose parameters are to be extracted, obtaining test spectrum extraction information, a test light-current curve, and a test amplitude modulation response curve as test results; extracting the DFB laser parameters using a multi-objective optimization algorithm, using the root mean square error between the output results of the two deep learning models and the test results as population fitness, retaining the elite of the population based on the population fitness, and outputting the DFB laser parameters when the maximum number of iterations is reached. The method disclosed by the present invention avoids weight adjustment when extracting parameters from different DFB lasers; greatly reduces traditional numerical calculations, thereby reducing computational costs and the time of parameter extraction; and achieves batch extraction of DFB laser parameters in a relatively short time.

Description

Technical Field

[0001] The present invention relates to the field of laser parameter extraction, and in particular to a DFB laser parameter extraction method based on a multi-objective optimization algorithm. Background Art

[0002] DFB lasers have become a commonly used light source in optical communication systems due to their advantages of low threshold current, narrow linewidth, and wide direct modulation bandwidth. During the design and production of DFB lasers, the devices are usually modeled and simulated to reduce costs. After actual production, if there is a significant difference between the test performance of the device and the simulation results, it is necessary to extract parameters based on the test results and compare and analyze them with the design parameters to determine the cause of the performance deviation. In addition, there are complex coupling relationships between some parameters, and the random cleavage at the end during the production process increases the random cavity surface phase relative to the entire optical mode, which will increase the difficulty of accurate parameter extraction.

[0003] Currently, parameter extraction methods typically combine optimization algorithms with numerical calculation methods. The optimization algorithm iteratively modifies the relevant parameters and performs simulations until the error between the laser's output characteristics and the target characteristics reaches the required level or the number of iterations is reached. If the optimization algorithm is run iteratively, and traditional numerical calculation methods consume a large amount of computing resources to obtain the interactions of the DFB laser's multiple physical states and the device's inherent trends, parameter extraction based on this method will be very time-consuming and require a large amount of computational cost, making it difficult to achieve large-scale parameter extraction. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides a DFB laser parameter extraction method based on a multi-objective optimization algorithm to solve the problems of time-consuming and high computational cost of existing parameter extraction methods and reduce the possibility of multiple solutions.

[0005] To achieve the above object, the technical solution of the present invention is as follows:

[0006] A DFB laser parameter extraction method based on a multi-objective optimization algorithm comprises the following steps:

[0007] Step 1: Build and train two deep learning models. The input of deep learning model 1 is the grating-related parameters in the DFB laser parameters, and the output is spectral extraction information; the input of deep learning model 2 is the DFB laser parameters, and the output is the light-current curve and amplitude modulation response curve.

[0008] Step 2: Using an instrument to test the DFB laser whose parameters are to be extracted, obtaining test spectrum extraction information, a test light-current curve, and a test amplitude modulation response curve as test results;

[0009] In step 3, a multi-objective optimization algorithm is used to extract the DFB laser parameters. The root mean square error between the output results of the two deep learning models and the test results is used as the population fitness. The population elite is retained based on the population fitness, and the DFB laser parameters are output when the maximum number of iterations is reached.

[0010] In the above scheme, step 3, extracting the DFB laser parameters using a multi-objective optimization algorithm includes the following process:

[0011] Step 31, initializing a population according to the parameter range of the DFB laser whose parameters are to be extracted, and recording it as the parent population, where a candidate solution of the population corresponds to a set of DFB laser parameters;

[0012] Step 32: Input the parent population into two deep learning models that have passed the test, calculate the root mean square error between the model output and the test result, and record it as the parent population fitness;

[0013] Step 33: Mutate the parent population according to the mutation strategy to obtain a mutant population, cross the parent population and the mutant population according to the crossover strategy to generate a child population, input the child population into two qualified deep learning models, calculate the root mean square error between the model output and the test result, and record it as the fitness of the child population;

[0014] Step 34: Merge the parent population and the child population, divide the merged population into several Pareto fronts according to the fitness of the parent population and the fitness of the child population, and retain the population elite according to the Pareto front level and crowding distance. Retain the candidate solutions and enter the next iteration process. The retained candidate solutions serve as the parent population of the next iteration.

[0015] Step 35 , determining whether the current generation number has reached the maximum number of iterations. If so, output the DFB laser parameters; otherwise, return to step 32 .

[0016] In the above scheme, the DFB laser parameters include optical confinement factor, differential gain, internal loss, gain compression factor, transparent carrier concentration, carrier lifetime, current injection efficiency, linewidth broadening factor and grating-related parameters, and the grating-related parameters include grating period, effective refractive index, coupling coefficient, phase and group refractive index.

[0017] In the above scheme, the deep learning model 1 includes an input layer, 5 hidden layers and an output layer; the input layer contains 5 neurons, corresponding to 5 grating-related parameters respectively; the output layer contains 22 neurons, corresponding to the wavelength coordinates and amplitudes of the spectral peaks and valleys, that is, spectral extraction information; the number of neurons in the 5 hidden layers is 100, 200, 300, 200, and 100 respectively; layer normalization is added before each hidden layer, and the activation function used is Leaky_ReLu.

[0018] The deep learning model 2 includes an input layer, 6 hidden layers and an output layer; the input layer contains 13 neurons, corresponding to 13 DFB laser parameters respectively; the output layer contains 132 neurons, of which 50 neurons correspond to the output power of the DFB laser, that is, the ordinate of the light-current curve, and the corresponding abscissa is the injection current of the DFB laser; the remaining neurons correspond to the amplitude modulation responses under two different currents, that is, the ordinate of the amplitude modulation response curve, and the corresponding abscissa is the modulation frequency; the number of neurons in the 6 hidden layers is 200, 300, 400, 400, 300, and 200 respectively; batch normalization is introduced before each hidden layer to adjust the feature distribution and improve convergence stability, and the activation function used is Swish.

[0019] In the above scheme, training two deep learning models includes the following process:

[0020] (1) Constructing a dataset:

[0021] First, the DFB laser parameter range is given, and then points are evenly selected for each parameter based on the given range. Then, these 13 parameters are randomly combined to obtain several groups of DFB laser parameters.

[0022] For each set of grating-related parameters in the DFB laser parameters, the spectrum of the DFB laser is calculated by the transfer matrix method and the spectrum extraction information is obtained. The grating-related parameters and the spectrum extraction information are recorded as data set 1.

[0023] For each set of DFB laser parameters, the light-current curve and amplitude modulation response curve of the DFB laser are calculated using the traveling wave model, and the DFB laser parameters, photocurrent curve and amplitude modulation response curve are recorded as data set 2;

[0024] Divide dataset 1 and dataset 2 into training set, validation set and test set respectively;

[0025] (2) Model training:

[0026] The training set in Dataset 1 was used to train Deep Learning Model 1, where grating-related parameters were used as model input and spectral extraction information was used as model output. During the training process of Model 1, a loss function was used to quantify the difference between the spectral extraction information predicted by the model and the spectral extraction information in the training set. During the training process, the data in the validation set was used to measure whether the training process was overfitting.

[0027] Model 2 was trained using the training set from Dataset 2, with DFB laser parameters as model input and light-current curves and amplitude modulation response orientation as model outputs. During the training of Model 2, a loss function was used to quantify the difference between the light-current curves and amplitude modulation response curves predicted by the model and the data in the training set. During the training process, the data in the validation set was used to measure whether the training process was overfitting.

[0028] (3) Model testing:

[0029] After the two models are trained, they are tested separately using the data in the two test sets. If they pass the test, they are used to extract the parameters of the multi-objective optimization algorithm. If they fail the test, the model is retrained.

[0030] In the above scheme, in step 2, a spectrometer is used to test the DFB laser whose parameters are to be extracted to obtain test spectrum extraction information; a vector network analyzer, a lightwave component analyzer or an integrating sphere is used to measure the DFB laser whose parameters are to be extracted to obtain a test light-current curve and amplitude modulation response.

[0031] In the above scheme, in step 31, for the 13-dimensional search space, the population is expressed as:

[0032] ;

[0033] in, Is the first in the G generation candidate solutions, each of which contains 13 DFB laser parameters, namely a set of values ​​of optical confinement factor, differential gain, internal loss, gain compression factor, transparent carrier concentration, carrier lifetime, current injection efficiency, linewidth broadening factor, grating period, effective refractive index, coupling coefficient, phase and group refractive index. N is the size of the population. represents the value of the candidate solution in the jth dimension;

[0034] Randomly initialize the initial values ​​of candidate solutions in the search space ,in and are the minimum and maximum values ​​in the search range, respectively, and the seed of the random number generator is used during initialization.

[0035] In the above scheme, in step 33, the mutation strategies include the following three:

[0036] ;

[0037] ;

[0038] ;

[0039] in, It is The mutation vector corresponding to the candidate solution, 、 and are three different candidate solutions in the population, is the individual with the best fitness in the Gth generation, and F is the scaling factor used to control the scale of the difference vector;

[0040] The crossover strategy is a binomial crossover operation, and the offspring population obtained after crossover is expressed as:

[0041] ;

[0042] in, is the crossover probability, is a randomly selected dimension, ensuring that at least one dimension changes; When the evolutionary algebra is G, The candidate solution The value of the dimension, Represents the j-th dimension value of the i-th candidate solution in the G-th generation.

[0043] In the above scheme, in step 34, the method of merging the parent population and the offspring population and dividing the merged population into several Pareto fronts according to the fitness of the parent population and the fitness of the offspring population is as follows:

[0044] Traverse the fitness of the merged population:

[0045] ;

[0046] in, According to the population The root mean square error between the spectrum extraction information predicted by the grating related parameters in the candidate solution and the test spectrum extraction information, According to the population The root mean square error between the light-current curve and amplitude modulation response curve predicted by the candidate solution and the test results;

[0047] The merged population is divided into multiple Pareto levels according to the root mean square error value. If for any two candidate solutions in the population and , define the dominance relationship as follows:

[0048] ;

[0049] For all the error vectors All established, , and there is at least one So that:

[0050] ;

[0051] Then judge Dominate ;

[0052] When performing Pareto frontier hierarchical classification on the merged parent population and child population, the candidate solutions in the first Pareto front are not dominated by any other solutions, the candidate solutions in the second Pareto front are only dominated by the candidate solutions in the first front, and so on, until all candidate solutions are assigned to the Pareto front.

[0053] In the above scheme, in step 34, in the process of retaining candidate solutions according to the Pareto front level and the crowding distance, priority is given to candidate solutions with a high Pareto front level, that is, starting from the first Pareto front, the entire layer of solutions are added to the new population in sequence. If the number of current front solutions is greater than the capacity of the front solutions to be selected, the required number of candidate solutions are selected from high to low according to the crowding distance to increase the diversity of the population. In this way, one iteration process is completed, and the selected candidate solutions will become the parent population in the next iteration process.

[0054] Through the above technical solution, the DFB laser parameter extraction method based on the multi-objective optimization algorithm provided by the present invention has the following beneficial effects:

[0055] 1. The present invention uses two deep learning models to calculate the spectrum extraction information, light-current curve and amplitude modulation response of the DFB laser, avoiding the numerical calculation of traditional transmission matrix and traveling wave model, and reducing the computational cost;

[0056] 2. The present invention extracts parameters through a multi-objective differential evolution algorithm, which does not rely on weight adjustment during the parameter extraction process, thereby reducing the time for parameter extraction of multiple DFB lasers;

[0057] In summary, the multi-objective parameter extraction method provided by the present invention can quickly extract the parameters of the DFB laser, effectively reduce the computational cost of parameter extraction, and realize parameter extraction of batch DFB lasers in a relatively short time. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for describing the embodiments or the prior art.

[0059] Figure 1 A schematic flow chart of a DFB laser parameter extraction method based on a multi-objective optimization algorithm disclosed in an embodiment of the present invention;

[0060] Figure 2 Schematic diagrams of two deep learning models in an embodiment of the present invention; (a) deep learning model 1, (b) deep learning model 2;

[0061] Figure 3 Schematic diagram comparing the spectrum extraction information predicted based on two sets of grating-related parameters in test set 1 and the spectrum extraction information in the test set in an embodiment of the present invention; (a) is sample 1, and (b) is sample 2;

[0062] Figure 4 Schematic diagram comparing the light-current curves predicted based on two sets of laser parameters in test set 2 and the data in the test set in an embodiment of the present invention; (a) is the light-current curve of sample 1, and (b) is the light-current curve of sample 2;

[0063] Figure 5 Schematic diagram comparing the amplitude modulation response curves predicted based on two sets of laser parameters in test set 2 and the data in the test set in an embodiment of the present invention; (a) is the amplitude modulation response curve of sample 1, and (b) is the amplitude modulation response curve of sample 2;

[0064] Figure 6 A schematic diagram of a process for extracting DFB laser parameters using a multi-objective optimization algorithm disclosed in an embodiment of the present invention;

[0065] Figure 7 Comparison between the spectral extraction information, light-current curve, and amplitude modulation response predicted based on the test output characteristic extraction parameters in an embodiment of the present invention and the test results; (a) is the comparison between the spectral extraction information predicted by deep learning model 1 and the test results, (b) is the comparison between the output power predicted by deep learning model 2 and the test results, and (c) is the comparison between the amplitude modulation response predicted by deep learning model 2 and the test results. DETAILED DESCRIPTION

[0066] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0067] The present invention provides a DFB laser parameter extraction method based on a multi-objective optimization algorithm, such as Figure 1 As shown, the following steps are included:

[0068] Step 1: Build and train two deep learning models, where the input of deep learning model 1 is the grating-related parameters in the DFB laser parameters, and the output is spectral extraction information; the input of deep learning model 2 is the DFB laser parameters, and the output is the light-current curve and amplitude modulation response curve.

[0069] (1) Model construction

[0070] like Figure 2 As shown in (a), the deep learning model 1 includes an input layer, 5 hidden layers and an output layer; the input layer contains 5 neurons, corresponding to 5 grating-related parameters respectively; the output layer contains 22 neurons, corresponding to the wavelength coordinates and amplitudes of the spectral peaks and valleys, that is, the spectral extraction information; the number of neurons in the 5 hidden layers are 100, 200, 300, 200, and 100 respectively; in order to improve the training efficiency of the deep learning model, layer normalization is added before each hidden layer, and the activation function used is Leaky_ReLu.

[0071] like Figure 2 As shown in (b), deep learning model 2 includes an input layer, 6 hidden layers and an output layer; the input layer contains 13 neurons, corresponding to 13 DFB laser parameters respectively; the output layer contains 132 neurons, of which 50 neurons correspond to the output power of the DFB laser, that is, the ordinate of the light-current curve, and the corresponding abscissa is the injection current of the DFB laser, and the remaining neurons correspond to the amplitude modulation responses under two different currents, that is, the ordinate of the amplitude modulation response curve, and the corresponding abscissa is the modulation frequency; the number of neurons in the 6 hidden layers are 200, 300, 400, 400, 300, and 200 respectively; batch normalization is introduced before each hidden layer to adjust the feature distribution and improve convergence stability, and the activation function used is Swish.

[0072] (2) Model training and testing

[0073] Training the two deep learning models involves the following process:

[0074] (1) Constructing a dataset:

[0075] First, the parameter range of the DFB laser is given. The DFB laser parameters include the optical confinement factor, differential gain, internal loss, gain compression factor, transparent carrier concentration, carrier lifetime, current injection efficiency, linewidth broadening factor and grating-related parameters. The grating-related parameters include grating period, effective refractive index, coupling coefficient, phase and group refractive index.

[0076] This embodiment is a uniform grating DFB laser. The DFB laser parameter ranges given when generating the data set are shown in Table 1.

[0077] Table 1 Range of DFB laser parameters

[0078]

[0079] Within the parameter range in Table 1, points are uniformly taken for each parameter, and then these 13 parameters are randomly combined to obtain several groups of DFB laser parameters.

[0080] For the five grating-related parameters in each set of DFB laser parameters, the spectrum of the DFB laser is calculated using the transfer matrix method to obtain 78,000 spectral samples. The wavelength coordinates and amplitudes of the peak and valley values ​​of the spectrum are then extracted to obtain spectral extraction information. The grating-related parameters and the corresponding spectral extraction information are then preprocessed respectively; the preprocessed grating-related parameters and spectral extraction information are used to construct dataset one, and dataset one is divided into training set one, validation set one, and test set one; among them, 60,000 samples are used to train deep learning model one, 9,000 samples are used to verify deep learning model one, and 9,000 samples are used to evaluate the generalization ability of deep learning model one.

[0081] For each set of DFB laser parameters, 78,000 DFB laser light-current curves and amplitude modulation response curves were calculated using the traveling wave model. Based on the calculated DFB laser light-current curves and amplitude modulation response curves, the DFB laser parameters and corresponding light-current curves and amplitude modulation response curves were first preprocessed. The preprocessed DFB laser parameters, light-current curves, and amplitude modulation response curves were used to construct Dataset 2, which was then divided into Training Set 2, Validation Set 2, and Test Set 2. Of these, 60,000 samples were used to train Deep Learning Model 2, 9,000 samples were used to validate Deep Learning Model 2, and 9,000 samples were used to evaluate the generalization ability of Deep Learning Model 2.

[0082] The specific method for calculating the spectrum using the transfer matrix method is described in patent CN119358427A and will not be described in detail here.

[0083] The light-current curve and amplitude modulation response curve of the DFB laser calculated by the traveling wave model can be found in the book "Design, Modeling and Simulation of Optoelectronic Devices", which will not be repeated here.

[0084] (2) Model training:

[0085] The training set in dataset one is used to train deep learning model one, where grating-related parameters are used as model input and spectral extraction information is used as model output. During the training process of model one, a loss function is used to quantify the difference between the spectral extraction information predicted by the model and the spectral extraction information in the training set; during the training process, the data in the validation set is used to measure whether the training process is overfitting.

[0086] The deep learning model was trained using 60,000 samples. During the model parameter optimization process, the LBFGH optimizer was used to adjust the neuron weights and biases to minimize the loss function, with a learning rate of 0.1. The loss functions used during optimization were mean squared error and mean absolute error, which together quantify the difference between the wavelength coordinates and amplitudes of the model's predicted spectral peaks and valleys and the training data.

[0087] Model 2 is trained using the training set in dataset 2, with DFB laser parameters as model input and light-current curve and amplitude modulation response orientation as model output. During the training of model 2, a loss function is used to quantify the difference between the light-current curve and amplitude modulation response curve predicted by the model and the data in the training set; during the training process, the data in the validation set is used to measure whether the training process is overfitting.

[0088] The second deep learning model was trained using the same 60,000 samples. The LBFGH optimizer was used to adjust the neuron weights and biases during model parameter optimization, with a learning rate of 0.08. The loss functions used during optimization were mean squared error and mean absolute error, which together quantify the difference between the model's predicted light-current curves and amplitude modulation responses and the training data.

[0089] (3) Model testing:

[0090] After the two models are trained, they are tested separately using the data in the two test sets. If they pass the test, they are used to extract the parameters of the multi-objective optimization algorithm. If they fail the test, the model is retrained.

[0091] Randomly select two groups of grating-related parameter samples from the test set 1 and input them into the constructed deep learning model 1 to predict the wavelength coordinates and amplitudes of the spectral peak and valley values. The results are as follows: Figure 3 As shown in the figure, the spectral information in the test set 1 is the dotted line, and the model prediction information is the solid line. Figure 3 As can be seen in Figures (a) and (b), the spectral extraction information predicted by the constructed model is highly consistent with the data in the dataset in terms of both trends and values. This intuitively demonstrates the reliability of the model and its application for subsequent multi-target parameter extraction. The specific values ​​of the two parameter samples selected from Test Set 1 are shown in Table 2.

[0092] Table 2 Numerical values ​​of two grating-related parameter samples

[0093]

[0094] Two sets of DFB laser parameter samples were randomly selected from the test set 2 and input into the constructed deep learning model 2 to predict the light-current curve and amplitude modulation response of the DFB laser. The results are as follows: Figure 4 and Figure 5 As shown, the output characteristics of the test set are dotted lines, and the prediction information of model 2 is solid line. Figure 4 As can be seen from (a) and (b), the slope efficiency and threshold current of the light-current curve predicted by the constructed model 2 are highly consistent with the data in the test set. Figure 5 As can be seen in Figures (a) and (b), the frequencies of the peak amplitude modulation response at the two bias currents predicted by Model 2 agree well with the results from the test set. These results demonstrate that this model can replace the traveling-wave model for calculating the light-current curve and amplitude modulation response of DFB lasers. The specific values ​​of the two DFB laser parameter samples selected from Test Set 2 are shown in Table 3.

[0095] Table 3 Numerical values ​​of two DFB laser parameter samples

[0096] ;

[0097]

[0098] Step 2: Use the instrument to test the DFB laser whose parameters are to be extracted, and obtain test spectrum extraction information, light-current curve and amplitude modulation response curve as test results.

[0099] Among them, a spectrometer is used to test the DFB laser whose parameters are to be extracted to obtain spectral extraction information, which is used as one target for multi-target parameter extraction; a vector network analyzer, a lightwave component analyzer, an integrating sphere and other devices are used to measure the DFB laser whose parameters are to be extracted to obtain the light-current curve and amplitude modulation response, which is used as another target for multi-target parameter extraction.

[0100] In step 3, a multi-objective optimization algorithm is used to extract the DFB laser parameters. The root mean square error between the output results of the two deep learning models and the test results is used as the population fitness. The population elite is retained based on the population fitness, and the DFB laser parameters are output when the maximum number of iterations is reached.

[0101] like Figure 6 As shown, the specific process includes the following:

[0102] Step 31, initializing a population according to the parameter range of the DFB laser whose parameters are to be extracted, and recording it as the parent population, where a candidate solution of the population corresponds to a set of DFB laser parameters;

[0103] For a 13-dimensional search space, the population is represented as:

[0104] ;

[0105] in, Is the first in the G generation candidate solutions, each of which contains 13 DFB laser parameters, namely a set of values ​​of optical confinement factor, differential gain, internal loss, gain compression factor, transparent carrier concentration, carrier lifetime, current injection efficiency, linewidth broadening factor, grating period, effective refractive index, coupling coefficient, phase and group refractive index. N is the size of the population. represents the value of the candidate solution in the jth dimension;

[0106] Randomly initialize the initial values ​​of candidate solutions in the search space ,in and are the minimum and maximum values ​​in the search range, respectively, and the seed of the random number generator is used during initialization.

[0107] In step 32, the parent population is input into two deep learning models that have passed the test, and the root mean square error between the model output and the test results of the DFB laser whose parameters are to be extracted is calculated, and recorded as the fitness of the parent population. The specific calculation formula is as follows:

[0108]

[0109] in, is the root mean square error, M is the number of sampling points predicted by the deep learning model, is the test value, is the predicted value of the deep learning model.

[0110] Step 33: Mutate the parent population according to the mutation strategy to obtain a mutant population, cross the parent population and the mutant population according to the crossover strategy to generate a child population, input the child population into two qualified deep learning models, calculate the root mean square error between the model output and the test results of the DFB laser whose parameters are to be extracted, and record it as the fitness of the child population;

[0111] The candidate solutions in the parent population select a strategy from the strategy pool according to probability for mutation operation. The mutation strategies include the following three:

[0112] ;

[0113] ;

[0114] ;

[0115] in, It is The mutation vector corresponding to the candidate solution, 、 and are three different candidate solutions in the population, is the individual with the best fitness in the Gth generation, and F is the scaling factor used to control the scale of the difference vector;

[0116] The candidate solution after mutation is subjected to a binomial crossover operation with the candidate solution in the parent population to generate a child population. The child population obtained after crossover is expressed as:

[0117] ;

[0118] in, is the crossover probability, is a randomly selected dimension, ensuring that at least one dimension changes; When the evolutionary algebra is G, The candidate solution The value of the dimension, Represents the j-th dimension value of the i-th candidate solution in the G-th generation.

[0119] Step 34: Merge the parent population and the offspring population, divide the merged population into several Pareto fronts according to the fitness of the parent population and the fitness of the offspring population, and retain candidate solutions according to the Pareto front level and crowding distance to enter the next iteration process. The retained candidate solutions serve as the parent population of the next iteration;

[0120] The method of merging the parent population and the offspring population and dividing the merged population into several Pareto frontiers according to the fitness of the parent population and the fitness of the offspring population is as follows:

[0121] Traverse the fitness of the merged population:

[0122] ;

[0123] in, According to the population The root mean square error between the spectrum extraction information predicted by the grating related parameters in the candidate solution and the test spectrum extraction information, According to the population The root mean square error between the light-current curve and amplitude modulation response curve predicted by the DFB laser parameters in the candidate solution and the test results; the errors of the light-current curve and amplitude modulation response curve are calculated separately and then weighted to obtain a root mean square error.

[0124] The merged population is divided into multiple Pareto levels according to the root mean square error value. If for any two candidate solutions in the population and , define the dominance relationship as follows:

[0125] ;

[0126] For all the error vectors All established, , and there is at least one So that:

[0127] ;

[0128] Then judge Dominate ;

[0129] When performing Pareto frontier hierarchical classification on the merged parent population and child population, the candidate solutions in the first Pareto front are not dominated by any other solutions, the candidate solutions in the second Pareto front are only dominated by the candidate solutions in the first front, and so on, until all candidate solutions are assigned to the Pareto front.

[0130] In the process of retaining candidate solutions according to the Pareto front level and crowding distance (i.e., the elite retention process), priority is given to candidate solutions with a high Pareto front level, that is, starting from the first Pareto front, the entire layer of solutions are added to the new population in sequence. If the number of current front solutions is greater than the capacity of the front solutions to be selected, the required number of candidate solutions are selected from high to low according to the crowding distance to increase the diversity of the population. This completes an iterative process, and the selected candidate solutions will become the parent population in the next iterative process.

[0131] Step 35 , determining whether the current generation number has reached the maximum number of iterations. If so, output the DFB laser parameters; otherwise, return to step 32 .

[0132] Table 4 lists the DFB laser parameters extracted using the multi-objective differential evolution algorithm. Based on these parameters, the wavelength coordinates and amplitude of the spectral peak and valley values, the light-current curve and the amplitude modulation response curve are predicted using the two constructed models and compared with the output characteristics of the test. The results are shown in Table 4. Figure 7 As shown. Figure 7 As can be seen in (a), the wavelength coordinates and amplitudes between the spectrum extraction information predicted by the deep learning model and the test results are highly consistent. Figure 7 As can be seen in Figures (b) and (c), the threshold current and slope efficiency of the light-current curve predicted by deep learning model 2 are highly consistent with the test results. The frequency of the peak amplitude modulation response predicted by the model also agrees well with the test results. These results intuitively demonstrate that the proposed multi-objective parameter extraction method can accurately and efficiently extract DFB laser parameters.

[0133] Table 4 Extraction parameter values

[0134]

[0135] The present invention constructs a deep learning model (one) using grating-related parameters and spectral extraction information to characterize the mapping relationship of the transmission matrix. A deep learning model (two) is constructed using DFB laser parameters, light-current curves, and amplitude modulation responses to map the traveling wave model. The effectiveness of the constructed models is evaluated using two test sets. The multi-objective parameter extraction process begins with population initialization. The initialized DFB laser parameters and grating-related parameters are input into deep learning model two and deep learning model one, respectively, to predict the light-current curve, amplitude modulation response curve, and spectral extraction information, and error analysis is performed with the test results. The population is then mutated and cross-pollinated to generate a child population. The parent and child populations are merged and classified according to the Pareto front level. During elite retention, solutions with high Pareto front levels are prioritized. If the current front number of solutions is greater than the remaining number of candidate solutions to be retained, the required number of solutions are selected from high to low according to the crowding distance to increase the diversity of the population. Error analysis is performed between the deep learning model predictions and the test results. When the error reaches the target value or the number of iterations, the parameter is output.

[0136] This invention extracts DFB laser parameters from the measured output characteristics and uses two trained deep learning models to predict the output characteristics based on the extracted parameters, achieving high consistency with the test results. This invention effectively avoids the extensive numerical calculations required to extract parameters by combining optimization algorithms with traditional numerical methods, significantly reducing time costs and facilitating parameter extraction for batches of DFB lasers.

[0137] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily 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 is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A DFB laser parameter extraction method based on a multi-objective optimization algorithm, characterized in that: The steps include: Step 1: Build and train two deep learning models, where the input of deep learning model 1 is the grating-related parameters in the DFB laser parameters, and the output is spectrum extraction information; The input of deep learning model 2 is the DFB laser parameters, and the output is the light-current curve and amplitude modulation response curve; Step 2: Using an instrument to test the DFB laser whose parameters are to be extracted, obtaining test spectrum extraction information, a test light-current curve, and a test amplitude modulation response curve as test results; Step 3: A multi-objective optimization algorithm is used to extract the DFB laser parameters. The root mean square error between the output results of the two deep learning models and the test results is used as the population fitness. The elite of the population is retained based on the population fitness, and the DFB laser parameters are output when the maximum number of iterations is reached. Step 3, extracting DFB laser parameters using a multi-objective optimization algorithm includes the following process: Step 31, initializing a population according to the parameter range of the DFB laser whose parameters are to be extracted, and recording it as the parent population, where a candidate solution of the population corresponds to a set of DFB laser parameters; Step 32: Input the parent population into two deep learning models that have passed the test, calculate the root mean square error between the model output and the test result, and record it as the parent population fitness; Step 33: Mutate the parent population according to the mutation strategy to obtain a mutant population, cross the parent population and the mutant population according to the crossover strategy to generate a child population, input the child population into two qualified deep learning models, calculate the root mean square error between the model output and the test result, and record it as the fitness of the child population; Step 34: Merge the parent population and the child population, divide the merged population into several Pareto fronts according to the fitness of the parent population and the fitness of the child population, and retain the population elite according to the Pareto front level and crowding distance. Retain the candidate solutions and enter the next iteration process. The retained candidate solutions serve as the parent population of the next iteration. Step 35 , determining whether the current generation number has reached the maximum number of iterations. If so, output the DFB laser parameters; otherwise, return to step 32 .

2. The DFB laser parameter extraction method based on a multi-objective optimization algorithm according to claim 1, characterized in that: The DFB laser parameters include optical confinement factor, differential gain, internal loss, gain compression factor, transparent carrier concentration, carrier lifetime, current injection efficiency, linewidth broadening factor and grating-related parameters, and the grating-related parameters include grating period, effective refractive index, coupling coefficient, phase and group refractive index.

3. The DFB laser parameter extraction method based on a multi-objective optimization algorithm according to claim 2, characterized in that: The deep learning model 1 includes an input layer, five hidden layers, and an output layer; the input layer includes five neurons, each corresponding to five grating-related parameters; the output layer includes 22 neurons, each corresponding to the wavelength coordinates and amplitudes of the spectral peaks and valleys, i.e., spectral extraction information; The number of neurons in the five hidden layers are 100, 200, 300, 200, and 100 respectively; layer normalization is added before each hidden layer, and the activation function used is Leaky_ReLu; The deep learning model 2 includes an input layer, 6 hidden layers and an output layer; the input layer contains 13 neurons, corresponding to 13 DFB laser parameters respectively; the output layer contains 132 neurons, of which 50 neurons correspond to the output power of the DFB laser, that is, the ordinate of the light-current curve, and the corresponding abscissa is the injection current of the DFB laser; the remaining neurons correspond to the amplitude modulation responses under two different currents, that is, the ordinate of the amplitude modulation response curve, and the corresponding abscissa is the modulation frequency; the number of neurons in the 6 hidden layers is 200, 300, 400, 400, 300, and 200 respectively; batch normalization is introduced before each hidden layer to adjust the feature distribution and improve convergence stability, and the activation function used is Swish.

4. The DFB laser parameter extraction method based on a multi-objective optimization algorithm according to claim 3, characterized in that: Training the two deep learning models involves the following process: (1) Constructing a dataset: First, the DFB laser parameter range is given, and then points are evenly selected for each parameter based on the given range. Then, these 13 parameters are randomly combined to obtain several groups of DFB laser parameters. For each set of grating-related parameters in the DFB laser parameters, the spectrum of the DFB laser is calculated by the transfer matrix method and the spectrum extraction information is obtained. The grating-related parameters and the spectrum extraction information are recorded as data set 1. For each set of DFB laser parameters, the light-current curve and amplitude modulation response curve of the DFB laser are calculated using the traveling wave model, and the DFB laser parameters, photocurrent curve and amplitude modulation response curve are recorded as data set 2; Divide dataset 1 and dataset 2 into training set, validation set and test set respectively; (2) Model training: The training set in Dataset 1 was used to train Deep Learning Model 1, where grating-related parameters were used as model input and spectral extraction information was used as model output. During the training process of Model 1, a loss function was used to quantify the difference between the spectral extraction information predicted by the model and the spectral extraction information in the training set. During the training process, the data in the validation set was used to measure whether the training process was overfitting. Model 2 was trained using the training set from Dataset 2, with DFB laser parameters as model input and light-current curves and amplitude modulation response orientation as model outputs. During the training of Model 2, a loss function was used to quantify the difference between the light-current curves and amplitude modulation response curves predicted by the model and the data in the training set. During the training process, the data in the validation set was used to measure whether the training process was overfitting. (3) Model testing: After the two models are trained, they are tested separately using the data in the two test sets. If they pass the test, they are used to extract the parameters of the multi-objective optimization algorithm. If they fail the test, the model is retrained.

5. The DFB laser parameter extraction method based on a multi-objective optimization algorithm according to claim 1, characterized in that: In step 2, the DFB laser whose parameters are to be extracted is tested using a spectrometer to obtain test spectrum extraction information; the DFB laser whose parameters are to be extracted is measured using a vector network analyzer, a lightwave component analyzer or an integrating sphere to obtain a test light-current curve and an amplitude modulation response.

6. The DFB laser parameter extraction method based on a multi-objective optimization algorithm according to claim 1, characterized in that: In step 31, for the 13-dimensional search space, the population is expressed as: ； in, Is the first in the G generation candidate solutions, each of which contains 13 DFB laser parameters, namely a set of values ​​of optical confinement factor, differential gain, internal loss, gain compression factor, transparent carrier concentration, carrier lifetime, current injection efficiency, linewidth broadening factor, grating period, effective refractive index, coupling coefficient, phase and group refractive index. N is the size of the population. represents the value of the candidate solution in the jth dimension; Randomly initialize the initial values ​​of candidate solutions in the search space ,in and are the minimum and maximum values ​​in the search range, respectively, and the seed of the random number generator is used during initialization.

7. The DFB laser parameter extraction method based on a multi-objective optimization algorithm according to claim 1, characterized in that: In step 33, the mutation strategies include the following three: ； ； ； in, It is The mutation vector corresponding to the candidate solution, 、 and are three different candidate solutions in the population, is the individual with the best fitness in the Gth generation, and F is the scaling factor used to control the scale of the difference vector; The crossover strategy is a binomial crossover operation, and the offspring population obtained after crossover is expressed as: ； in, is the crossover probability, is a randomly selected dimension, ensuring that at least one dimension changes; When the evolutionary algebra is G, The candidate solution The value of the dimension, Represents the j-th dimension value of the i-th candidate solution in the G-th generation.

8. The DFB laser parameter extraction method based on a multi-objective optimization algorithm according to claim 1, characterized in that: In step 34, the parent population and the offspring population are merged, and the merged population is divided into several Pareto fronts according to the fitness of the parent population and the fitness of the offspring population. The method is as follows: Traverse the fitness of the merged population: ； in, According to the population The root mean square error between the spectrum extraction information predicted by the grating related parameters in the candidate solution and the test spectrum extraction information, According to the population The root mean square error between the light-current curve and amplitude modulation response curve predicted by the candidate solution and the test results; The merged population is divided into multiple Pareto levels according to the root mean square error value. If for any two candidate solutions in the population and , define the dominance relationship as follows: ； For all the error vectors All established, , and there is at least one So that: ； Then judge Dominate ; When performing Pareto frontier hierarchical classification on the merged parent population and child population, the candidate solutions in the first Pareto front are not dominated by any other solutions, the candidate solutions in the second Pareto front are only dominated by the candidate solutions in the first front, and so on, until all candidate solutions are assigned to the Pareto front.

9. The DFB laser parameter extraction method based on a multi-objective optimization algorithm according to claim 1, characterized in that: In step 34, in the process of retaining candidate solutions according to the Pareto front level and crowding distance, priority is given to candidate solutions with a high Pareto front level, that is, starting from the first Pareto front, the entire layer of solutions are added to the new population in sequence. If the number of current front solutions is greater than the capacity of the front solutions to be selected, the required number of candidate solutions are selected from high to low according to the crowding distance to increase the diversity of the population. In this way, one iteration process is completed, and the selected candidate solutions will become the parent population in the next iteration process.

Images