A method for intelligent optimization design of VCSEL array oxidation aperture

Abstract

The application discloses a VCSEL array oxidation aperture intelligent optimization design method, which is used for improving the temperature distribution uniformity of a VCSEL array, thereby improving the overall performance of the laser. Firstly, the application replaces a thermoelectric feedback model by training a BP neural network model, establishes a mapping relationship between the oxidation aperture and the steady-state junction temperature, then directly selects the oxidation apertures of each unit of the VCSEL array to code the chromosomes, constructs a fitness function representing the temperature uniformity of the VCSEL array to provide an optimization standard, and finally adopts a genetic algorithm to realize the optimal optimization design of the oxidation aperture of the VCSEL array with an arbitrary shape and a unit number. The application effectively shortens the labor cost and time cost in seeking the optimal oxidation aperture design of the VCSEL array. Finally, the oxidation aperture optimization design is carried out by taking a hexagonal VCSEL array with 61 units as an example, and the effectiveness and practicability of the method are proved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semiconductor lasers and intelligent computing, in particular to a VCSEL array oxide aperture intelligent optimization design method. BACKGROUND

[0002] Due to the advantages of low threshold current, high beam quality, single longitudinal mode stable operation and easy two-dimensional array integration, vertical cavity surface emitting lasers (VCSEL) are widely used in communication, laser radar, medical treatment, sensing and other fields.

[0003] Since the output optical power provided by a single-hole VCSEL is very limited, in order to improve the output optical power, a parallel structure of multiple independent small-size VCSEL units, i.e. a two-dimensional VCSEL array, is usually used. However, in the VCSEL array, due to uneven heat dissipation of each unit and thermal coupling effect between them, the VCSEL array junction temperature distribution is uneven, and the bias current with positive temperature coefficient also causes uneven current distribution, forming a thermo-opto-electric feedback, causing the output optical power to be thermally reversed, thereby severely limiting the thermo-opto-electric performance of the VCSEL array.

[0004] Considering that the size of the oxide aperture of the unit in the VCSEL array is one of the important factors that determine the injection current, the oxide aperture design will directly affect the temperature distribution of the VCSEL array, therefore the prior art usually adopts the method of reducing the oxide aperture of the unit in the central region of the VCSEL array to reduce thermal coupling, thereby improving the uniformity of the temperature distribution. However, the above method currently only manually adjusts the oxide aperture of the VCSEL unit to optimize the design of the array. With the increase of the integration degree of the VCSEL array, this method will consume a lot of manpower and time cost, has low accuracy and it is difficult to design the oxide aperture of a large-scale array. In view of the above problems, an effective method for intelligent optimization design of the oxide aperture of the VCSEL array is urgently needed to improve the uniformity of the temperature distribution and improve the overall performance of the device.

[0005] As is known to all, with the rapid development of artificial intelligence and deep learning, neural networks, as a powerful learning model, have been widely used in various complex prediction, classification, regression and other tasks. BP (Back Propagation) neural network, as the most commonly used type of neural network, is trained through multiple layers of nonlinear transformation and back propagation algorithm, so that the network can gradually adjust the parameters to approximate the expected output.

[0006] Meanwhile, as a global optimization algorithm simulating the natural evolution process, the genetic algorithm (GA) is based on the theory of Darwin's evolution and the principle of genetics, and is widely used in various optimization problems by simulating the selection, crossover and mutation of biological chromosomes.

[0007] The genetic algorithm and the BP neural network are combined in the application, and are applied to the VCSEL array oxide aperture optimization design. The main feature is that the fast and accurate BP neural network model is used to replace the complex and time-consuming thermoelectric feedback model to obtain the temperature distribution and calculate the fitness function. Through the genetic algorithm, the oxide aperture of the VCSEL unit can be directly coded and operated, and the limitation of the number of VCSEL array units is avoided. The method does not require the continuity or derivability of the fitness function, and is not constrained by the modeling rules or special requirements of the VCSEL array, and has excellent global optimization capability. Therefore, the BP neural network and the genetic algorithm are very suitable for solving the optimization problem of the VCSEL array oxide aperture design. SUMMARY

[0008] The application aims to solve the problems of huge labor cost and time cost caused by the traditional manual adjustment and design of the oxide aperture of each unit of the VCSEL array, and to provide a VCSEL array oxide aperture intelligent optimization design method.

[0009] To achieve the above-mentioned purpose, the application adopts the following technical solutions:

[0010] The application provides a VCSEL array oxide aperture intelligent optimization design method, and applies the BP neural network and the genetic algorithm to the oxide aperture optimization design of the VCSEL array. The method replaces the thermoelectric feedback model with the BP neural network model to obtain the mapping relationship between the oxide aperture and the steady-state junction temperature, and establishes a fitness function reflecting the temperature uniformity of the array to provide an optimization standard. Then, the genetic algorithm is used to code the oxide aperture of the VCSEL array with any shape and any number of units, and the gene of each chromosome corresponds to the oxide aperture of a VCSEL unit. Each chromosome corresponds to an oxide aperture design method, and finally the optimal oxide aperture optimization design is obtained through iteration. The method improves the optimization efficiency and greatly reduces the labor cost and time cost. The specific steps of the VCSEL array oxide aperture intelligent optimization design method are as follows:

[0011] Step 1: selecting the oxide aperture R corresponding to each unit of the n-unit VCSEL array to be optimized i(i = 1, 2, 3,..., n) as input variables of the BP neural network; and the steady-state junction temperature of each unit in the VCSEL array as output variables of the BP neural network;

[0012] Step 2: Randomly generate N groups of VCSEL array unit oxidation aperture sample values within the allowed variation range (R min ~ R max ) of the oxidation aperture of each unit in the VCSEL array, each group of samples containing n oxidation aperture values R i = R min + rand() · (R max - R min ), where rand() represents a random number between 0 and 1;

[0013] Step 3: Import the n oxidation aperture values R i in each group of samples into the VCSEL array thermoelectric feedback model in turn to simulate temperature distribution and obtain the corresponding steady-state junction temperature T i (i = 1, 2, 3,..., n), thereby generating a data set for the BP neural network;

[0014] Step 4: Train and test the BP neural network using the data set obtained in Step 3, wherein 80% of the samples in the data set are randomly selected as training data for the neural network, and the remaining 20% are used as test data, until the error between the predicted value of the neural network and the sample value is limited within the allowed error range of 1K, and the construction of the BP neural network is completed;

[0015] Step 5: Set the oxidation aperture of the jth VCSEL unit as R j , generate M chromosomes within the allowed variation range (R min ~ R max ) of the oxidation aperture of each unit in the VCSEL array, with a number between 100 and 200, wherein each chromosome contains n oxidation aperture genes satisfying R j = R min + rand() · (R max - R min ), where rand() represents a random number between 0 and 1, and each chromosome represents an oxidation aperture optimization method, forming an initial population;

[0016] Step 6: Import each chromosome in the population into the BP neural network constructed in Step 4 in turn to obtain the steady-state junction temperature T of each unit, and construct a fitness function L p representing the temperature uniformity of the VCSEL array, i.e., using temperature variance as the evaluation standard to calculate the fitness value of the chromosome;

[0017] Step 7: The population is selected based on the fitness value of the reverse ranking selection operation, and the top M chromosomes are selected for genetic;

[0018] Step 8: Determine whether the genetic algorithm reaches the maximum iteration number, if yes, execute step 12, otherwise execute step 9, the iteration number is between 500-1000;

[0019] Step 9: All selected chromosomes in step 7 are subjected to one-point crossover operation, and two chromosomes are selected for crossover according to the crossover probability order, the crossover point is a random integer between 2 and n-1, and the new chromosome generated after crossover is added to the population, and the crossover probability is between 0.4 and 0.6;

[0020] Step 10: All chromosomes in step 9 are subjected to mutation operation based on the oxidation aperture gene point, and one chromosome is selected for random oxidation aperture gene mutation according to the mutation probability, and the new chromosome after mutation is added to the population, and the mutation probability is between 0.01 and 0.1;

[0021] Step 11: The offspring chromosome population is generated, and step 6 is returned to calculate the fitness value of the chromosome in the BP neural network, and the next round of genetic algorithm iteration optimization is started;

[0022] Step 12: The genetic algorithm iteration is completed, and the first ranked chromosome is extracted;

[0023] Step 13: Decode the chromosome to obtain the optimal VCSEL array oxidation aperture size information;

[0024] Step 14: The optimal VCSEL array oxidation aperture size information is introduced into the thermoelectric feedback model to simulate the temperature distribution;

[0025] Step 15: Obtain the optimized VCSEL array steady-state temperature distribution, and the optimization is completed.

[0026] The technical problems to be solved by the present application can also be further realized by the following technical solutions, the step 3 of the thermoelectric feedback model simulating the temperature distribution comprises the following steps:

[0027] (1) Based on the oxidation aperture sample value or the oxidation aperture information corresponding to the optimal chromosome, and the structural parameters including the shape of the VCSEL array, the center-to-center distance of adjacent units, and the substrate size, the ANSYS is used to establish a VCSEL array structure model;

[0028] (2) Set the working conditions of the VCSEL array including the injection current and the ambient temperature;

[0029] (3) Simulate the temperature distribution of the VCSEL array and extract the peak junction temperature T M0;

[0030] (4) Calculate the heat flux density of each unit of the VCSEL array using MATLAB;

[0031] (5) Extract the heat flux density in (4) and load it into the VCSEL array, simulate the temperature distribution of the VCSEL array using ANSYS, and extract the peak junction temperature T M1 ;

[0032] (6) Determine whether |T M1 -T M0 | is less than 1K; if it is less than 1K, perform step (7); otherwise, set T M0 = T M1 , and return to step (4);

[0033] (7) Output the steady-state junction temperature T1, T2, T3, …, T n of each unit in the VCSEL array.

[0034] The step 4 of constructing the BP neural network model comprises the following steps:

[0035] (1) Establish a three-layer BP neural network structure including an input layer, a hidden layer and an output layer, wherein the number of input layer neurons is n, the input variables are R1, R2, R3, …, R n ; the number of output layer neurons is n, and the output variables are The number of neurons in the hidden layer is q is a constant (q ∈ [1, 10]);

[0036] (2) Use tansig function as the activation function of the hidden layer, and purelin linear function as the activation function of the output layer:

[0037]

[0038] purelin(x) = x

[0039] (3) Use the mean square error function to calculate the error of the BP neural network:

[0040]

[0041] Wherein, is the steady-state junction temperature predicted by the BP neural network, T i is the sample steady-state junction temperature corresponding to the test data in the data set;

[0042] (4) Use the gradient descent method trainlm function to train the BP neural network.

[0043] The fitness function L in the step 6P The temperature uniformity of the VCSEL array is selected as the evaluation criterion, and the variance of the VCSEL array temperature is used to calculate the fitness function of the p-th chromosome:

[0044]

[0045] Where n is the number of cells in the VCSEL array. The steady-state junction temperature is the value predicted by the BP neural network for the j-th VCSEL unit.

[0046] In the optimization design of oxide pore size for VCSEL arrays, as the array integration density increases and the number of cells continues to grow, traditional single-tube VCSEL oxide pore size design approaches and methods for manually designing VCSEL cell oxide pore sizes become impractical. This invention employs a BP neural network and a genetic algorithm to automatically design the oxide pore size of the VCSEL array, entrusting the complex and tedious optimization process to computer iteration, greatly saving manpower and time costs, and obtaining the optimal oxide pore size design method. Simultaneously, this invention directly selects the VCSEL array oxide pore size information for chromosome encoding, enabling VCSEL array optimization with any number of cells, solving the problem of difficulty in manual design due to the increasing number of VCSEL cells. Furthermore, during the genetic algorithm iteration process, the fitness of each chromosome is calculated by predicting temperature using a BP neural network model, replacing the lengthy simulation process of the thermoelectric feedback model with a fast and accurate calculation method, making the genetic algorithm iteration and VCSEL optimization results more efficient. Attached Figure Description

[0047] The object and advantages of the invention will be further understood from the following description taken in conjunction with the accompanying drawings. In these drawings:

[0048] Figure 1 A flowchart illustrating the intelligent optimization design method for oxide pore size of VCSEL array proposed in this invention is shown.

[0049] Figure 2 A diagram illustrating the chromosome inheritance process of the genetic algorithm of this invention is provided.

[0050] Figure 3 A schematic diagram of a conventional hexagonal uniformly oxidized pore VCSEL array is shown.

[0051] Figure 4 An array structure diagram with optimized oxidation pore size is shown in the embodiment of the present invention;

[0052] Figure 5 The temperature distribution of a conventional hexagonal uniformly oxidized pore VCSEL array is illustrated.

[0053] Figure 6An example of the array temperature distribution after optimizing the oxidation pore size according to an embodiment of the present invention is shown. Detailed Implementation

[0054] To more clearly illustrate the objectives, technical solutions, and advantages of this invention, a detailed description of the invention will now be provided, in conjunction with the accompanying drawings, using a 61-cell hexagonal VCSEL array as an example. It should be noted that the embodiments described herein are only a part, not all, of this invention. This invention is applicable to VCSEL arrays of any shape and the optimization of oxide pore sizes with any number of cells. Based on the technology of this invention, any other embodiments that can be obtained by those skilled in the art without creative effort, as well as improvements or modifications made without departing from the technical principles of this invention, are all within the protection scope of this invention.

[0055] This invention proposes a smart optimization design method for the oxide pore size of a VCSEL array, the process of which is as follows: Figure 1 As shown, it includes the following steps:

[0056] Step 1: Select the oxide pore sizes R1, R2, R3, ..., R of each cell in the 61-cell hexagonal VCSEL array to be optimized. 61 As input variables to the neural network; the steady-state junction temperature of each cell in the VCSEL array. As the output variable of the BP neural network;

[0057] Step 2: The allowable variation range (R) of the oxide pore size of each unit in the VCSEL array. min ~R max N=500 sets of oxide pore size sample values ​​for VCSEL array cells were randomly generated within the range of 5μm to 10μm. Each set of samples contained 61 oxide pore size values ​​R. i =5 + rand()·5, where rand() represents a random number between 0 and 1;

[0058] Step 3: Sequentially assign the 61 oxide pore size values ​​R to each sample group. i Importing the simulated temperature distribution into the VCSEL array thermoelectric feedback model yields the corresponding steady-state junction temperature T. i (i = 1, 2, 3, ..., 61), thus generating the dataset for the BP neural network;

[0059] Step 4: training and testing the BP neural network using the data set obtained in step 3, wherein 80% of the samples in the data set are randomly selected as training data for the neural network, and the remaining 20% are selected as test data, until the error between the predicted value of the neural network and the sample value is limited within the allowable error range of 1K, and the construction of the BP neural network is completed; in this embodiment, after training the BP neural network, the mean square error MSE is 2.4346e-03, and the training effect is good;

[0060] Step 5: assuming that the oxidation aperture of the jth VCSEL unit is R j In the VCSEL array, M = 100 chromosomes are generated within the allowed variation range of 5 μm to 10 μm of the oxidation aperture of each unit, as shown in Figure 2 (a), and chromosome coding is performed, wherein each chromosome contains 61 oxidation aperture genes satisfying: R j = 5 + rand() · 5, where rand() represents a random number between 0 and 1, and each chromosome represents an oxidation aperture optimization method, forming an initial population;

[0061] Step 6: each chromosome in the population is sequentially introduced into the BP neural network constructed in step 4 to obtain the steady-state junction temperature of each unit A fitness function L representing the temperature uniformity of the VCSEL array is constructed p , that is, the temperature variance is taken as the evaluation standard, and the fitness value of the chromosome is calculated, and the fitness function of the pth chromosome is:

[0062]

[0063] wherein n is the total number of VCSEL array units, and in this embodiment, n is 61, is the steady-state junction temperature of the jth VCSEL unit obtained by the BP neural network prediction.

[0064] Step 7: selecting the population based on the reverse ranking according to the fitness value, when the chromosome fitness value is smaller, it means that the temperature variance is smaller and the uniformity is better, and the top 100 chromosomes are selected for genetic;

[0065] Step 8: determining whether the genetic algorithm reaches the maximum iteration number of 1000 times, if yes, executing step 12, otherwise executing step 9;

[0066] Step 9: performing one-point crossover operation on all the selected chromosomes in step 7, as shown in Figure 2 (b), 2 chromosomes are selected for crossover according to the crossover probability order, the crossover point is a random integer between 2 and 60, and the new chromosomes generated after the crossover are added to the population, and the crossover probability is 0.5.

[0067] Step 10: For all chromosomes in step 9, perform the mutation operation based on the oxidation aperture gene point, as shown in Figure 2 (c), select the chromosome for random oxidation aperture gene mutation according to the mutation probability 0.1, and add the mutated new chromosome to the population;

[0068] Step 11: Generate a population of offspring chromosomes, return to step 6, import the BP neural network, calculate the fitness value, and start the next round of genetic algorithm iteration optimization. The parameters of the example genetic algorithm are shown in the following table:

[0069] Parameter Value Number of unit oxidation pore diameters, i.e. number of genes n 61 Number of chromosomes M 100 Crossing over probability 0.5 Mutation probability 0.1 Minimum pore size of oxidation R min ]]> 5 μm Maximum pore size R max ]]> 10 μm Range of variation of genetic and mutation operations 5 μm to 10 μm Maximum number of iterations 1000

[0070] Step 12: The genetic algorithm iteration is completed, and the first ranked chromosome is extracted;

[0071] Step 13: Decode the chromosome to obtain the optimal VCSEL array oxidation aperture size information, as shown in Figure 4 ;

[0072] Step 14: Import the optimal VCSEL array oxidation aperture size information into the thermoelectric feedback model to simulate the temperature distribution;

[0073] Step 15: Obtain the optimized VCSEL array steady-state temperature distribution as shown in Figure 6 , and the optimization is completed.

[0074] For optimization comparison, the present application adopts a conventional equal-interval non-optimized hexagonal VCSEL array with equal oxidation apertures and an optimized VCSEL array, and keeps the sum of the oxidation aperture areas of each unit of the VCSEL array unchanged before and after optimization, to obtain a non-optimized uniform oxidation aperture of 8.15 μm, as shown in Figure 3 , the center-to-center distance of adjacent units of the conventional hexagonal VCSEL array is 52 μm, the oxidation aperture and light-emitting aperture radii of all units are 8.15 μm, the VCSEL mesa radius is 18.15 μm, and the substrate size is 260 μm. The temperature distribution is simulated by the thermoelectric feedback model. In addition, the neural network training and test data set in step 3 and step 14 in the embodiment of the present application are also calculated by this method, and the steps are as follows:

[0075] (1) Obtain the uniform oxidation aperture information or oxidation aperture sample value of the non-optimized hexagonal VCSEL array and the structure parameters including the VCSEL array shape, the center-to-center distance of adjacent units, and the substrate size, and use ANSYS to establish a VCSEL array structure model.

[0076] (2) Set the injection current of the VCSEL array to 427 mA and the environmental temperature to 300 K;

[0077] (3) Simulate the temperature distribution of the VCSEL array and extract the peak junction temperature T M0 ;

[0078] (4) Calculate the heat flux density of each unit of the VCSEL array using MATLAB;

[0079] (5) Extract the heat flux density in (4) and load it into the VCSEL array, simulate the temperature distribution of the VCSEL array using ANSYS and extract the peak junction temperature T M1 ;

[0080] (6) Determine whether |T M1 -T M0 | is less than 1K; if it is less than 1K, perform step (7); otherwise, make T M0 =T M1 , return to step (4);

[0081] (7) Output the steady-state junction temperature T1, T2, T3, …, T n of each unit in the VCSEL array.

[0082] The step 4 of constructing a BP neural network model in the embodiment of the application includes the following steps:

[0083] (1) Establish a 3-layer BP neural network structure including an input layer, a hidden layer and an output layer, wherein the number of neurons in the input layer is 61, the input variables are R1, R2, R3, …, R 61 , the number of neurons in the output layer is 61, and the output variables are The number of neurons in the hidden layer is 20;

[0084] (2) Use the tansig function as the activation function of the hidden layer and the purelin linear function as the activation function of the output layer:

[0085]

[0086] purelin(x) = x

[0087] (3) Calculate the error of the BP neural network using the mean square error function:

[0088]

[0089] wherein, is the steady-state junction temperature predicted by the BP neural network, and T i is the sample steady-state junction temperature corresponding to the test data in the data set;

[0090] (4) Train the BP neural network using the gradient descent method trainlm function.

[0091] Through the above steps, we get the conventional hexagonal VCSEL array temperature distribution and the temperature distribution of the embodiment of the application after optimization, as shown in Figure 5 and Figure 6 It can be found that the temperature of the central unit of the VCSEL array before optimization is high, and the temperature distribution is uneven. After optimizing the oxidation aperture, the temperature distribution is greatly improved, and the temperature uniformity (variance) before optimization is 22.8374, and the temperature uniformity after optimization is 12.3553. By comparison, the temperature uniformity of the embodiment of the application is improved by 45.90% compared with the conventional array, and the light output power is increased by 2mW. In summary, after the optimization of the application, the thermal problem of the VCSEL array is effectively improved, thereby improving the overall performance of the device, and the effectiveness of the application is illustrated.

Claims

1. A method for intelligent optimization design of oxide pore size in VCSEL arrays, characterized by the following steps: include: Step 1: Select with The oxide pore size of each cell in the VCSEL array to be optimized As the input variable of the BP neural network; the steady-state junction temperature of each cell in the VCSEL array. As the output variable of the BP neural network; Step 2: The allowable variation range of oxide pore size in each unit of the VCSEL array ( Randomly generated within) Group of VCSEL array cell oxide pore size sample values, each group of samples contains Oxidation pore size value ,in, Represented as a random number between 0 and 1; Step 3: Extract samples from each group of samples. Oxidation pore size value By sequentially importing the simulated temperature distribution into the VCSEL array thermoelectric feedback model, the corresponding steady-state junction temperature can be obtained. This generates a dataset for a BP neural network. Step 4: Use the dataset obtained in Step 3 to train and test the BP neural network. Randomly select 80% of the samples in the dataset as training data for the neural network and the remaining 20% ​​as test data until the error between the neural network prediction value and the sample value is limited to within the allowable error range of 1K, thus completing the construction of the BP neural network. Step 5: Let the first... The oxidation pore size of each VCSEL unit is The allowable variation range of oxide pore size in each unit of the VCSEL array Intrinsic generation There are 100 to 200 chromosomes, and each chromosome contains 100 chromosomes. Each oxidation pore size gene satisfies: ,in, Represented as random numbers between 0 and 1, each chromosome represents an oxidation aperture optimization method, forming the initial population; Step 6: Sequentially import each chromosome in the population into the BP neural network constructed in Step 4 to obtain the steady-state junction temperature of each unit. Construct a fitness function that reflects the temperature uniformity of the VCSEL array. That is, using temperature variance as the evaluation standard to calculate the fitness value of chromosomes; Step 7: Perform a reverse-rank selection operation on the population based on fitness values, selecting the top-ranked individuals. Heredity is inherited through chromosomes; Step 8: Determine if the genetic algorithm has reached the maximum number of iterations. If yes, proceed to step 12; otherwise, proceed to step 9. The number of iterations is between 500 and 1000. Step 9: Perform a point-based crossover operation on all chromosomes selected in Step 7, selecting two chromosomes for crossover in order of crossover probability. The crossover point is between 2 and 3. The crossover is a random integer, and the new chromosome generated after the crossover is added to the population. The crossover probability is between 0.4 and 0.

6. Step 10: Perform mutation operations based on the oxidation pore size gene points on all chromosomes in Step 9. Select one chromosome according to the mutation probability to perform random oxidation pore size gene mutation, and add the mutated new chromosome to the population. The mutation probability is between 0.01 and 0.

1. Step 11: Generate the offspring chromosome population, return to step 6, import the chromosomes into the BP neural network to calculate their fitness values, and begin the next round of genetic algorithm iterative optimization; Step 12: After the genetic algorithm completes its iteration, the chromosome ranked first is extracted; Step 13: Decode the chromosome to obtain the optimal VCSEL array oxide pore size information; Step 14: Import the optimal VCSEL array oxide pore size information into the thermoelectric feedback model to simulate the temperature distribution; Step 15: Obtain the optimized steady-state temperature distribution of the VCSEL array; optimization ends. Step 3, the thermoelectric feedback model simulating the temperature distribution, includes the following steps: (1) Based on the sample value of oxide pore size or the oxide pore size information corresponding to the optimal chromosome, as well as the structural parameters including the shape of the VCSEL array, the distance between adjacent units, and the substrate size, a VCSEL array structure model is established using ANSYS. (2) Set the operating conditions of the VCSEL array, including injection current and ambient temperature; (3) Simulate the temperature distribution of the VCSEL array and extract the peak junction temperature T M0 ; (4) Use MATLAB to calculate the heat flux density of each element in the VCSEL array; (5) Extract the heat flux density from (4) and load it into the VCSEL array. Use ANSYS to simulate the temperature distribution of the VCSEL array and extract the peak junction temperature. ; (6) Judgment - Is it less than 1? If less than 1 If so, proceed to step (7); otherwise, make = Return to step (4); (7) Output steady-state junction temperature of each unit in the VCSEL array ; Step 4, constructing the BP neural network model, includes the following steps: (1) Establish a 3-layer BP neural network structure including an input layer, a hidden layer, and an output layer, wherein the number of neurons in the input layer is Input variables are The number of neurons in the output layer is The output variable is The number of neurons in the hidden layer is , A constant ( ); (2) Adopt The function is used as the activation function of the hidden layer. Linear functions as activation functions for the output layer: (3) The error of the BP neural network is calculated using the mean square error function: in, The steady-state junction temperature is predicted by a BP neural network. The steady-state junction temperature of the sample corresponding to the test data within the dataset; (4) Use gradient descent method The function trains the BP neural network.

2. The intelligent optimization design method for oxide pore size of VCSEL array according to claim 1, characterized in that, fitness function in step 6 The temperature uniformity of the VCSEL array was selected as the evaluation criterion, and the variance of the VCSEL array temperature was used to calculate the... The fitness function for each chromosome is: in, The number of cells in the VCSEL array. For the first The steady-state junction temperature of each VCSEL unit is predicted by a BP neural network.

Images