Optical element pollution layer distribution optimization method based on deep learning and differential evolution

By optimizing the processing parameters of optical components using deep learning and differential evolution algorithms, the problem of minimizing the contamination layer in existing technologies is solved, achieving efficient post-processing and cost reduction of optical components.

CN119005010AActive Publication Date: 2024-11-22CHANGCHUN INST OF OPTICS FINE MECHANICS & PHYSICS CHINESE ACAD OF SCI
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202411468534.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-21
Publication Date
2024-11-22
Estimated Expiration
2044-10-21

AI Technical Summary

Technical Problem

Existing methods for selecting manufacturing process parameters for optical components largely rely on experience, which cannot effectively minimize the contamination layer left by the polishing slurry, resulting in complex and costly post-processing.

Method used

A method based on deep learning and differential evolution is adopted. By collecting processing parameter data, the relationship between processing parameters and contamination layer distribution is established. A deep neural network is used to train the model, and the process parameters are optimized by combining differential evolution algorithm to minimize the contamination layer.

Benefits of technology

It significantly reduces the complexity and cost of post-processing of optical components and improves the processing efficiency of high-power laser optical components.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119005010A_ABST
    Figure CN119005010A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of optical element processing, in particular to an optical element pollution layer distribution optimization method based on deep learning and differential evolution. The method comprises the following steps of: firstly, taking processing technology parameters as input, taking pollution layer depth and concentration data acquired by using TOF-SIMS as output data, preprocessing the processing technology parameter data, performing deep learning, constructing and establishing a relationship between the processing technology parameters and pollution layer distribution, and then utilizing a differential evolution algorithm to calculate the processing technology parameters and the pollution layer distribution. And optimizing to obtain process parameters for minimizing the pollution layer. The method has the advantages that the collection amount of the data set in the early stage is large, and once the neural network calculation model is built and trained, the optimization of the pollution layer distribution of the optical element is very rapid, and the complexity of the subsequent pollution removal operation is greatly reduced. Compared with a traditional empirical method for selecting process parameters, the method has obvious advantages in the aspects of improving the post-processing efficiency of the intense laser optical element and reducing the manufacturing cost.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of optical element processing, and in particular to a method for optimizing the distribution of contamination layers of optical elements based on deep learning and differential evolution. Background Art

[0002] During the manufacturing process of optical components, due to the mechanical and chemical effects between the abrasive and the optical components during the polishing process, the polishing liquid will remain on the surface or even in the body of the optical components. These residual contaminations will significantly reduce the laser damage threshold of the optical components in the high-intensity laser system and shorten the service life of the optical components. Therefore, before the optical components are put into use, they must be post-processed to remove the contamination layer formed by the residual polishing liquid. Conventional methods include etching the contamination layer with chemical solvents and etching the contamination layer with plasma. The depth and concentration of the contamination layer determine the etching time and depth. The smaller the contamination layer, the shorter the post-processing cycle will be, and the lower the manufacturing cost of the optical components will be. In order to select the process parameters of the optical component manufacturing process, the depth and concentration of the contamination layer after processing will be minimized and the concentration will be lowest under the existing processing conditions.

[0003] However, most of the existing methods for selecting process parameters for optical component manufacturing are based on experience. Manufacturers use the surface convergence rate as a goal and subjectively determine the process parameters based on the results of each test, without considering the factor of minimizing the contamination layer distribution. This parameter selection method has great uncertainty and is affected by human subjective factors. It is impossible to explore the limit of minimizing the contamination layer, which brings great challenges to the post-processing process of optical components. Summary of the invention

[0004] In order to solve the above problems, the present invention provides a method for optimizing the distribution of contamination layers of optical components based on deep learning and differential evolution.

[0005] The present invention aims to provide a method for optimizing the distribution of contamination layers of optical components based on deep learning and differential evolution, which specifically comprises the following steps: S1. Conduct n Conduct optical component processing experiments under different processing parameters and collect processing parameter data; S2. Extract the distribution data of Ce ion contamination layer of processed optical components under multiple sets of process parameters , La ion contamination layer distribution data , secondary ion concentration data after normalization of Ce ions The secondary ion concentration data normalized by La ions , as training data; S3. Preprocessing the processing parameter data, including maximum and minimum value normalization and B-spline transformation; S4. Input the preprocessed data into the DNN deep neural network for training, output the distribution of Ce ions, the PPM curve of Ce ions, the distribution of La ions and the PPM curve of La ions, and construct a deep learning model; establish the relationship between the processing parameters and the distribution of the contamination layer through the training model; S5. Using differential evolution algorithm to find the combination of processing parameters that minimizes the contamination layer based on the deep learning model; S6. Output the processing parameter combination optimized by the differential evolution algorithm.

[0006] Preferably, the processing parameters include pressure p , speed v , polishing liquid pH , average particle size of abrasive r ; Processing parameter data is .

[0007] Preferably, the Ce ion contamination layer distribution data Specifically expressed as: ; La ion pollution layer distribution data Specifically expressed as: ; Secondary ion concentration data after Ce ion normalization Specifically expressed as: ; Secondary ion concentration data after La ion normalization Specifically expressed as: ; in, w Represents the dimension of the secondary ion concentration vector, indicating that the secondary ion concentration is one The vector of ppm represents the normalized secondary ion concentration.

[0008] Preferably, the loss functions of ion distribution and normalized secondary ion concentration in the DNN deep neural network are: ; ; ; Among them, m represents a matrix with m rows and m columns, , , is the coordinate submatrix predicted by the neural network, , , is the actual coordinate sub-matrix, ppm i represents the actual normalized secondary ion concentration, represents the predicted normalized secondary ion concentration.

[0009] Preferably, the optimization algorithm used for DNN deep neural network training is the Adam algorithm, with a step size of 0.0001 and an iteration number of 100,000 times.

[0010] Preferably, in step S5, the processing parameter boundary limit range is: ; ; ; ; The differential evolution algorithm specifically includes: S51. Initialize population: randomly generate a set of input processing parameter combinations within the processing parameter boundary limit range as the initial population; S52. Define fitness function: Obtain the fitness function by the output value predicted by the deep learning model; S53. Mutation, crossover and selection: First, perform a mutation operation to generate a mutation vector; then perform a crossover operation to generate a test vector by mixing the current individual and the mutation vector; compare the fitness of the current individual and the test vector, and choose to retain the individual with better fitness; iterate t+1 times, enter the next mutation operation, and gradually approach the optimal solution.

[0011] Preferably, in step S53, the mutation vector expression is as follows: ; in, is the mutation vector generated by mutation, F is the scaling factor, are three different individual vectors randomly selected from the population; The test vector expression is as follows: ; in, is the generated trial vector, is the mutation vector The jth component of is the jth component in the current individual, is a uniform random number between 0 and 1. CR is the crossover probability, is a randomly selected index that guarantees that at least one component comes from the mutation vector; The formula chosen is: ; in, is the generated trial vector, represents the i-th individual in the t+1-th iteration.

[0012] Preferably, step S3 specifically includes the following sub-steps: S31. Normalize the maximum and minimum values ​​of the processing parameter data to limit the impact of data features of different units or sizes: ; in, i Indicates i group experimental data, min and max represent the minimum and maximum values ​​in the data respectively. p Indicates pressure, v Indicates the speed, pH Indicates polishing liquid, r Indicates the average particle size of the abrasive. normalization represents the normalized value, initialize Indicates the initial value; S32. After the maximum and minimum value normalization processing is performed, the data is subjected to B-spline transformation. The specific transformation method is: S321. Select a node vector. The node vector is usually expressed as: ; In the formula, m represents the number of nodes; S322. Define the B-spline function: Define the zero-order B-spline function, the recursive formula is as follows: ; in, i Indicates i B-spline, 0 means the order is 0, when input k Located at the node k i and k i+1 The function is 1 when , otherwise it is 0; Define high-order B-spline function, recursive formula as follows: ; in, i Indicates i B-splines, d Indicates the order.

[0013] Preferably, the number of nodes in step S321 is m =3; the order of step S322 d =2.

[0014] Preferably, in step S2, a static time-of-flight secondary ion mass spectrometer is used to extract multiple groups of process parameters.

[0015] Compared with the prior art, the present invention can achieve the following beneficial effects: The present invention proposes a method for optimizing the distribution of the contamination layer of optical components based on deep learning and differential evolution. The processing parameters are first used as input, and the depth and concentration data of the contamination layer collected using TOF-SIMS are used as output data. Deep learning is performed to establish the relationship between the processing parameters and the distribution of the contamination layer. Then, the differential evolution algorithm is used to optimize the process parameters that minimize the contamination layer. The method of the present invention only requires a large amount of data to be collected in the early stage. Once the neural network calculation model is built and trained, the optimization of the contamination layer distribution of optical components based on deep learning and differential evolution will be very rapid and will greatly reduce the complexity of subsequent decontamination operations. Compared with the traditional empirical method for selecting process parameters, the method proposed in this patent has obvious advantages in improving the post-processing efficiency of high-power laser optical components and reducing manufacturing costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 3 is a distribution diagram of the pollution layer of two metal ions, cerium (Ce) and lanthanum (La), provided according to an embodiment of the present invention; (a) represents Ce ions, and (b) represents La ions.

[0017] Figure 2 3 is a curve showing the variation of the secondary ion concentration of Ce and La with depth according to an embodiment of the present invention.

[0018] Figure 3 It is an overall framework diagram of an optical element contamination layer distribution optimization method based on deep learning and differential evolution provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the following description, the same modules are represented by the same reference numerals. In the case of the same reference numerals, their names and functions are also the same. Therefore, the detailed description thereof will not be repeated.

[0020] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not constitute a limitation of the present invention.

[0021] The present invention provides a method for optimizing the distribution of contamination layers of optical components based on deep learning and differential evolution, which specifically comprises the following steps: S1. Conduct nConduct optical component processing experiments under different processing parameters and collect processing parameter data; S2. Use static time-of-flight secondary ion mass spectrometry (TOF-SIMS) technology to extract multiple sets of process parameters (such as pressure p , speed v , polishing liquid pH , average particle size of abrasive r ) The corresponding Ce ion contamination layer distribution data of the processed optical component is shown below , La ion contamination layer distribution data , secondary ion concentration data after normalization of Ce ions The secondary ion concentration data normalized by La ions , as training data; the pollution layer distribution data is as follows Figure 1 As shown in the figure, the normalized secondary concentration data is as follows Figure 2 As shown; Ce ion pollution layer distribution data Specifically expressed as: ; La ion pollution layer distribution data Specifically expressed as: ; Secondary ion concentration data after Ce ion normalization Specifically expressed as: ; Secondary ion concentration data after La ion normalization Specifically expressed as: ; in, w Represents the dimension of the secondary ion concentration vector, indicating that the secondary ion concentration is one The vector of ppm represents the normalized secondary ion concentration.

[0022] Since a large number of experimental studies have shown (such as the literature J. Neauport, L. Lamaignere, H. Bercegol, F. Pilon, J.-C. Birolleau, Polishing-induced contamination of fused silicaoptics and laser induced damage density at 351 nm, Optics Express, 13 (25) (2005) 10163-10171.) that the two metal ions, cerium (Ce) and lanthanum (La), have a major influence on the laser damage threshold of optical components, the present invention mainly performs prediction and optimization on the two examples of Ce and La.

[0023] S3. Preprocessing the processing parameter data, specifically including the following sub-steps: S31. Normalize the maximum and minimum values ​​of the processing parameter data to limit the impact of data features of different units or sizes: ; in, i Indicates i group experimental data, min and max represent the minimum and maximum values ​​in the data respectively. p Indicates pressure, v Indicates the speed, pH Indicates polishing liquid, r Indicates the average particle size of the abrasive. normalization represents the normalized value, initialize Indicates the initial value.

[0024] To reduce the difference in the numerical range between different parameters, a B-spline transformation is performed to convert the input features into a new feature representation.

[0025] S32. In order to reduce the difference in the numerical range between different parameters and avoid missing interactive features, the data is subjected to B-spline transformation after the maximum and minimum value normalization processing. The specific transformation method is as follows: S321. Select the knot vector: The knot vector is the basis of B-spline, which defines the intersection of the polynomial segments; the knot vector is usually expressed as: ; In the present invention, the number of nodes is selected m =3; S322. Define the B-spline function: Define the zero-order B-spline function (the zero-order B-spline function is the most basic form of B-spline, equivalent to a step function), and the recursive formula is as follows: ; in, i Indicates i B-spline, 0 means the order is 0, when input k Located at the node k i and k i+1 When , the function is 1, otherwise it is 0; Define a high-order B-spline function. This recursive formula is used to generate high-order B-splines. as follows: ; in, i Indicates i B-splines, d Indicates the order (here we take d =2); The input features are converted into new feature representations through the B-spline function. Specifically, the input features are linearly combined through the B-spline basis functions to generate a new feature matrix. For example, if the node vector contains multiple nodes, after the conversion, the input features will be mapped to multiple new feature dimensions to form a feature matrix.

[0026] S4. Input the preprocessed data into the DNN deep neural network for training; the DNN deep neural network sets four output terminals, which respectively output the distribution of Ce ions, the PPM curve of Ce ions, the distribution of La ions and the PPM curve of La ions, and construct a deep learning model; through the training model, establish the relationship between the processing parameters and the distribution of the contamination layer; The loss function is defined as the L1 norm of the difference between the actual ion distribution and the predicted ion distribution, and the actual PPM and the predicted PPM. Therefore, the loss function LOSS1 of the ion distribution, the loss function LOSS2 of the PPM, and the overall loss function L are: ; ; ; Where m represents a matrix with m rows and m columns (m is the matrix subscript, and the square root of the sum of the differences between all elements of the predicted and actual matrices is the LOSS expression), w represents the vector subscript, the secondary ion concentration (PPM) is a row vector of 1*W, and the expression is also to find the difference between the corresponding row vector elements). , , is the coordinate submatrix predicted by the neural network, , , is the actual coordinate sub-matrix, ppm i represents the actual normalized secondary ion concentration, represents the predicted normalized secondary ion concentration; Directly use DNN for multiple outputs, that is, train together. Therefore, the loss function here is the sum of the three differences between the prediction and the actual XYZ matrix of the ion distribution, and then optimize it; use the adam optimization algorithm, set the step size to 0.0001, and the number of iterations to 100,000 times; the condition for stopping training is that the overall error L is less than 0.001.

[0027] S5. Using the differential evolution algorithm based on the deep learning model to find the combination of processing parameters that minimizes the contamination layer within the processing parameters boundary limit; The boundary limits of processing parameters are: ; ; ; ; The differential evolution algorithm specifically includes: S51. Initialize population: randomly generate a set of input processing parameter combinations within the processing parameter boundary limit range as the initial population.

[0028] S52. Define fitness function: The fitness function is obtained by the output value predicted by the deep learning model and is defined as: ; in, represents the normalized secondary ion concentration (ppm) of Ce ions predicted based on the input processing parameters, represents the normalized secondary ion concentration (ppm) of La ions predicted based on the input processing parameters, represents the Z value matrix for predicting the Ce ion distribution based on the input processing parameters, represents the Z value matrix for predicting the La ion distribution based on the input processing parameters, q 1. q 2. q 3. q 4 respectively represent their respective weight ratios to the fitness function, which is convenient for calculating the optimal pollution layer input parameters under different influences according to the situation.

[0029] S53. Mutation, crossover and selection: First, perform a mutation operation to generate a mutation vector , the expression is as follows: ; in, is the mutation vector generated by mutation, Fis the scaling factor, usually between 0 and 2. are three different individual vectors randomly selected from the population; Then perform a crossover operation, by mixing the current individual and mutation vector , generate the test vector , the expression is as follows: ; in, is the generated trial vector, is the mutation vector The jth component of is the jth component in the current individual, is a uniform random number between 0 and 1. CR is the crossover probability (usually between 0 and 1, which determines the probability of inheriting a component from the mutation vector), is a randomly selected index that guarantees that at least one component comes from the mutation vector; Selection: The purpose of the selection operation is to compare the current individual and the test vector The fitness of the individuals with better fitness is retained; the selection formula is: ; in, is the generated trial vector, represents the i-th individual in the t+1-th iteration; Iteration: Enter the next mutation operation and gradually approach the optimal solution; the current number of iterations is t+1.

[0030] Principle: Differential Evolution (DE) is a heuristic random search algorithm based on population differences. It generates a new parameter vector by adding the weighted difference vector between two members of the population to the third member. This operation is called "mutation". The parameters of the mutation vector are mixed with the parameters of another predetermined target vector according to certain rules to generate an experimental vector. This operation is called "crossover". If the cost function of the experimental vector is lower than the cost function of the target vector, the experimental vector replaces the target vector in the next generation. This operation is called "selection".

[0031] S6. Output optimized process parameters: Output a combination of processing parameters optimized by the differential evolution algorithm to guide the processing of optical components to minimize the contamination layer.

[0032] Although the amount of data collected in the early stage of this method is relatively large, once the neural network calculation model is built and trained, the optimization of the optical component contamination layer distribution based on deep learning and differential evolution will be very rapid and will greatly reduce the complexity of subsequent decontamination operations. Therefore, compared with traditional methods, the optical component contamination layer distribution optimization based on deep learning and differential evolution proposed in the present invention has the characteristics of low cost, greatly reducing the difficulty of subsequent decontamination operations, and accelerating the processing cycle of high-power laser components.

[0033] The present invention is not limited to the selection of process parameters for the contamination layer of a certain processing method. The prediction of the contamination layer distribution under technologies such as small grinding head polishing, magnetorheological polishing, and elastic adaptive polishing should all be within the protection scope of the present invention; it is not limited to deep learning and differential evolution algorithms. Other optimization algorithms, such as genetic algorithms and particle swarm algorithms, are used to optimize the processing parameters for minimizing the contamination layer, which should all be within the protection scope of the present invention.

[0034] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the disclosure of the present invention can be performed in parallel, sequentially or in different orders, as long as the desired results of the technical solution disclosed in the present invention can be achieved, and this document does not limit this.

[0035] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for optimizing the distribution of contamination layers of optical components based on deep learning and differential evolution, characterized in that: The specific steps include: S1. Conduct n Conduct optical component processing experiments under different processing parameters and collect processing parameter data; S2. Extract the distribution data of Ce ion contamination layer of processed optical components under multiple sets of process parameters , La ion contamination layer distribution data , secondary ion concentration data after normalization of Ce ions The secondary ion concentration data normalized by La ions , as training data; S3. Preprocessing the processing parameter data, including maximum and minimum value normalization and B-spline transformation; S4. Input the preprocessed data into the DNN deep neural network for training, output the distribution of Ce ions, the PPM curve of Ce ions, the distribution of La ions and the PPM curve of La ions, and construct a deep learning model; establish the relationship between the processing parameters and the distribution of the contamination layer through the training model; S5. Using differential evolution algorithm to find the combination of processing parameters that minimizes the contamination layer based on the deep learning model; S6. Output the processing parameter combination optimized by the differential evolution algorithm.

2. The method for optimizing the distribution of contamination layers of optical components based on deep learning and differential evolution according to claim 1, characterized in that: Processing parameters include pressure p , speed v , polishing liquid pH , average particle size of abrasive r ; Processing parameter data is .

3. The method for optimizing the distribution of contamination layers of optical components based on deep learning and differential evolution according to claim 1 or 2, characterized in that: Ce ion pollution layer distribution data Specifically expressed as: ; La ion pollution layer distribution data Specifically expressed as: ; Secondary ion concentration data after Ce ion normalization Specifically expressed as: ; Secondary ion concentration data after La ion normalization Specifically expressed as: ; in, w Represents the dimension of the secondary ion concentration vector, indicating that the secondary ion concentration is one The vector of ppm represents the normalized secondary ion concentration.

4. The method for optimizing the distribution of contamination layers of optical components based on deep learning and differential evolution according to claim 3, characterized in that: The loss functions of ion distribution and normalized secondary ion concentration in the DNN deep neural network are: ; ; ; Among them, m represents a matrix with m rows and m columns. , , is the coordinate submatrix predicted by the neural network, , , is the actual coordinate sub-matrix, ppm i represents the actual normalized secondary ion concentration, represents the predicted normalized secondary ion concentration.

5. The method for optimizing the distribution of contamination layers of optical components based on deep learning and differential evolution according to claim 4, characterized in that: The optimization algorithm used in the DNN deep neural network training is the Adam algorithm, with a step size of 0.0001 and 100,000 iterations.

6. The method for optimizing the distribution of contamination layers of optical components based on deep learning and differential evolution according to claim 1, characterized in that: In step S5, the processing parameter boundary limit range is: ; ; ; ; The differential evolution algorithm specifically includes: S51. Initialize population: randomly generate a set of input processing parameter combinations within the processing parameter boundary limit range as the initial population; S52. Define fitness function: Obtain the fitness function by the output value predicted by the deep learning model; S53. Mutation, crossover and selection: First, perform a mutation operation to generate a mutation vector; then perform a crossover operation to generate a test vector by mixing the current individual and the mutation vector; compare the fitness of the current individual and the test vector, and choose to retain the individual with better fitness; iterate t+1 times, enter the next mutation operation, and gradually approach the optimal solution.

7. The method for optimizing the distribution of contamination layers of optical components based on deep learning and differential evolution according to claim 6, characterized in that: In step S53, the mutation vector expression is as follows: ; in, is the mutation vector generated by mutation, F is the scaling factor, are three different individual vectors randomly selected from the population; The test vector expression is as follows: ; in, is the generated trial vector, is the mutation vector The jth component of is the jth component in the current individual, is a uniform random number between 0 and 1. CR is the crossover probability, is a randomly selected index that guarantees that at least one component comes from the mutation vector; The formula chosen is: ; in, is the generated trial vector, represents the i-th individual in the t+1-th iteration.

8. The method for optimizing the distribution of contamination layers of optical components based on deep learning and differential evolution according to claim 7, characterized in that: The step S3 specifically includes the following sub-steps: S31. Normalize the maximum and minimum values ​​of the processing parameter data to limit the impact of data features of different units or sizes: ; in, i Indicates i group experimental data, min and max represent the minimum and maximum values ​​in the data respectively. p Indicates pressure, v Indicates the speed, pH Indicates polishing liquid, r Indicates the average particle size of the abrasive. normalization represents the normalized value, initialize Indicates the initial value; S32. After the maximum and minimum value normalization processing is performed, the data is subjected to B-spline transformation. The specific transformation method is: S321. Select a node vector. The node vector is usually expressed as: ; In the formula, m represents the number of nodes; S322. Define the B-spline function: Define the zero-order B-spline function, the recursive formula is as follows: ; in, i Indicates i B-spline, 0 means the order is 0, when input k Located at the node k i and k i+1 The function is 1 when , otherwise it is 0; Define high-order B-spline function, recursive formula as follows: ; in, i Indicates i B-splines, d Indicates the order.

9. The method for optimizing the distribution of contamination layers of optical components based on deep learning and differential evolution according to claim 8, characterized in that: The number of nodes in step S321 m =3; the order of step S322 d =2.

10. The method for optimizing the distribution of contamination layers of optical components based on deep learning and differential evolution according to claim 9, characterized in that: In step S2, a static time-of-flight secondary ion mass spectrometer is used to extract multiple groups of process parameters.

Citation Information

Patent Citations

  • Additive manufacturing size prediction and process optimization method and system based on machine learning

    CN113569352A

  • Brain MRI image classification method based on residual neural network and evolutionary algorithm

    CN116524256A

  • Waste gas pollution prediction method for optimizing Elman neural network based on improved dung beetle algorithm

    CN116910542A

  • Water pollutant prediction method based on combination of sample expansion and improved particle swarm optimization

    CN118396180A

  • Method for predicting remaining service life of lithium-ion battery employing WDE-optimized LSTM network

    WO2020191800A1