Glass component design method based on generative adversarial network model

Abstract

The invention discloses a glass component design method based on a generative adversarial network model, and belongs to the technical field of glass.The method comprises the steps that glass data are collected, an initial training sample set is constructed, and the initial training sample set comprises glass components with a mapping relation and performance values corresponding to the glass components; sequentially carrying out data cleaning and feature normalization on the initial training sample set; constructing a network structure of a generator and a discriminator in the generative adversarial network model, performing iterative training of the generative adversarial network model by using the initial training sample set after feature normalization, and taking a generator model as a glass component design model after the iterative training is finished; and generating glass components meeting the target performance requirements of the glass by using the glass component design model. The generative adversarial network model obtained by the method has excellent generation capability and wide applicability, can break through the dilemma of glass component-performance design, and greatly improves the development efficiency of novel glass materials.

Description

Technical Field

[0001] The present invention relates to the field of glass technology, and in particular to a glass component design method based on a generative adversarial network model. Background Art

[0002] Glass is playing an increasingly important role in many fields of today's society. Whether it is architectural decoration in daily life or cutting-edge high-tech fields, the performance of glass itself has a crucial impact on the performance of related devices and its own service. Therefore, in order to meet the different needs in various application scenarios, it is critical to develop glass materials with specific properties. As an amorphous material, the performance of glass strongly depends on its components, but this relationship is not yet fully understood. Therefore, in the process of research and development, researchers have to rely on a large number of experimental attempts with different components to explore in order to obtain glass materials that meet the required performance requirements. Considering the extremely large space for glass component selection, under this research model, it is an extremely time-consuming and labor-intensive process to determine the glass formula that meets the design requirements. It is urgent to introduce faster and more efficient methods to change the traditional development and design model.

[0003] In recent years, with the continuous improvement of hardware facilities, the availability of high-performance computing tools and large glass databases, machine learning technology has received unprecedented attention and favor in the glass field. At present, most machine learning models in the glass field focus on the forward process of predicting glass performance values. However, in the glass design and development process, the truly critical and most challenging link lies in reverse design, that is, being able to efficiently search and recommend formulas that meet performance design requirements in a wide range of component spaces. This kind of design is difficult to achieve using only traditional machine learning methods.

[0004] A Chinese patent document with publication number CN117556716A discloses a design method for electromagnetic glass, which includes: obtaining a control matrix corresponding to a semi-free structure; determining a structural matrix of the control matrix, and determining the electromagnetic response of the structural matrix; constructing a target data set through a multi-objective optimization algorithm based on the control matrix, the structural matrix and the electromagnetic response; training a metasurface diffusion probability model based on the target data set to obtain a target metasurface diffusion probability model; inputting a target electromagnetic response and a random variable into the target metasurface diffusion probability model to generate a transmission metasurface structure of the target electromagnetic response, so as to complete the design of the electromagnetic glass based on the transmission metasurface structure. This method can design a high-performance metasurface structure that meets multiple design index requirements at the same time, but does not focus on glass component design.

[0005] In order to meet the different performance requirements of glass in different application scenarios, it is necessary to develop a glass component reverse design method in combination with a deep learning model. Summary of the invention

[0006] The present invention provides a glass component design method based on a generative adversarial network model based on a glass database and a generative adversarial network model framework. By training a generative adversarial network model with a customized generator loss function (component penalty term and performance penalty term), a large number of glass samples that meet the target performance design requirements can be quickly generated. The method has a wide range of applicability and can be used not only for single performance design but also for multi-performance coordination design. It is particularly suitable for large-scale component screening of glass in the initial design stage, which will greatly improve the development efficiency of new glass materials.

[0007] The specific technical solutions adopted are as follows:

[0008] A glass component design method based on a generative adversarial network model comprises the following steps:

[0009] Step 1: Collect glass data and construct an initial training sample set, wherein the initial training sample set includes glass components (i.e., glass components) with a mapping relationship and their corresponding performance values;

[0010] Step 2: performing data cleaning and feature normalization on the initial training sample set in turn;

[0011] Step 3: Construct the network structure of the generator and discriminator in the generative adversarial network model, and use the initial training sample set after feature normalization to iteratively train the generative adversarial network model. During training, the discriminator loss function based on Wasserstein distance and gradient penalty term is used to optimize and update the parameters of the discriminator, and the designed generator loss function based on component penalty term and performance penalty term is used to optimize and update the parameters of the generator. After the iterative training is completed, the generator model is used as the glass component design model; the glass component design model is used to generate glass components that meet the target performance requirements of the glass.

[0012] Specifically, in step 1, the performance values ​​include but are not limited to at least one of density, linear expansion coefficient, elastic modulus, hardness, Young's modulus, shear modulus, Poisson's ratio, glass transition temperature, strain point temperature, annealing point temperature, softening point temperature, liquidus temperature, refractive index, etc.

[0013] Preferably, the data cleaning in step 2 includes invalid sample removal, feature dimension reduction, extreme value sample removal and duplicate sample removal;

[0014] Invalid sample elimination: remove samples containing null values ​​in the glass target performance values ​​(i.e., performance values ​​to be designed) in the initial training sample set, and remove incomplete samples whose sum of glass components is not 100%;

[0015] Feature dimensionality reduction: remove samples containing placeholder elements in the initial training sample set, and remove glass component features and corresponding samples whose number of non-zero glass component feature values ​​is less than the set number;

[0016] Elimination of extreme value samples: Eliminate samples with abnormal performance values ​​according to the "3σ principle";

[0017] Duplicate sample elimination: For samples with repeated glass components, their performance values ​​are replaced by the median performance values ​​of the corresponding repeated samples, and redundant duplicate samples are removed.

[0018] Preferably, the constructed samples are all initial training sample sets of oxide glasses; when invalid samples are eliminated, samples with oxygen content less than 30% also need to be removed; when feature dimensionality reduction is performed, the placeholder elements are H, S, C, N, F, Cl, Br or I elements.

[0019] In the process of feature normalization, the value range of each component feature is set to [min(n i ),max(n i )], the range of each performance value characteristic is [min(p k ), max(p k )], where min(n i ) and max(n i ) represents the minimum and maximum values ​​of the i-th component feature, min(p k ) and max(p k ) represents the minimum and maximum values ​​of the kth performance value characteristic, then 0.8×min(n i ) and 1.2×max(n i ) is the normalization parameter of the component characteristics, with 0.8×min(p k ) and 1.2×max(p k ) is the normalization parameter of the performance value characteristic; where, if there is 1.2×max(n i ) is greater than 100, it is recorded as 100.

[0020] The framework of the generative adversarial network model used in step 3 is based on the Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP), and the generator and discriminator are both fully connected neural network structures with multiple hidden layers.

[0021] The input features of the generative adversarial network model include component features and performance features.

[0022] Furthermore, the discriminator loss function is expressed as:

[0023]

[0024] In the formula, Represents the data generated by the generator, represents the real training data, represents the mixed data of generated data and real training data, represents the distribution of generated data, represents the distribution of real training data, represents the distribution of mixed data, represents the output of the discriminator for real data, represents the output of the discriminator for the generated data, represents the output of the discriminator for mixed data, represents the calculation of mathematical expectation, It is used to represent the Wasserstein distance between the real sample and the generated sample. is the gradient penalty term used to stabilize training, and λ represents the weight parameter of the gradient penalty term.

[0025] Furthermore, the generator loss function is additionally designed with component penalty and performance penalty. The component penalty is used to constrain the sum of the components of the generated samples to meet 100% of the actual requirements, and the performance penalty is used to constrain the generator to generate in the direction that the performance of the generated samples meets the design target range.

[0026] The generator loss function is expressed as:

[0027]

[0028]

[0029] In the formula, represents the component penalty term, n is the dimension of the training data (the sum of the number of components and the number of performances to be designed), m is the number of performances to be designed, represents the generated sample after denormalization, and represents the input feature normalization parameter used when normalizing the training data, α is the weight parameter of the component penalty term, H represents the performance penalty term, in the single performance target design task, m = 1, when the target performance design interval has only one explicit boundary, if the target performance value is required to be less than this boundary value, then If the target performance is required to be greater than this boundary value, When the target performance design interval has two explicit boundaries, in Indicates the generated sample The nth feature (the feature position of the performance feature to be designed), P1, P 1_bottom , P 1_top It represents the boundary value of the glass target performance design interval after normalization under the corresponding performance value characteristic normalization parameter. In the multi-performance target design task, m>1. The same strategy as above is used to design the corresponding performance penalty item for each performance target, and the sum of the designed performance penalty items is taken as the final performance penalty item H; β represents the weight parameter of the performance penalty item.

[0030] In step 3, the network structure parameters include the number of hidden layers, the number of neurons in each hidden layer, and the type of activation function used by the neurons in each hidden layer. The training parameters include the optimizer type, optimizer learning rate, batch size, number of iterations, weight parameters of each penalty item in the generator and discriminator, whether to prioritize training the discriminator or the generator, and the ratio of the number of training times of the discriminator to the generator.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] (1) The generative adversarial network model has a unique data learning method and powerful sample generation capability. The present invention is based on the generative adversarial network model and does not require manual setting of the component space. Compared with the method of relying on prior knowledge to manually select the component space and then generate samples, and then using the prediction model for prediction and screening, the generative adversarial network model can actively and efficiently search and generate samples in a wide area space, greatly reducing the limitations caused by excessive human participation.

[0033] (2) The method proposed in the present invention can generate more reasonable samples based on the real data distribution. The generative adversarial network model can learn the distribution information of the real data through a multi-layer neural network, so that the generated samples follow the same data distribution. Compared with the method with stronger randomness, which adopts the optimization algorithm combined with the performance prediction model, the method proposed in the present invention can better ensure the experimental feasibility and rationality of the generated samples, and provide a reliable sample source for glass component design. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a diagram of the architecture of the generative adversarial network model in the present invention.

[0035] Figure 2 This is how the loss functions of the generator and the discriminator change with the number of iterations during the training of the generative adversarial network model in Example 1.

[0036] Figure 3 The violin plot superimposed scatter plot visualization results of the component characteristics of the samples generated in Example 1, the real training samples, and the samples that meet the requirements in the real training samples.

[0037] Figure 4 This is a scatter plot visualization result of the samples generated in Example 1, the real training samples, and the samples that meet the requirements in the real training samples after being reduced to two dimensions by t-SNE.

[0038] Figure 5 This is how the loss functions of the generator and the discriminator change with the number of iterations during the training of the generative adversarial network model in Example 2.

[0039] Figure 6 The violin plot superimposed scatter plot visualization results of the component characteristics of the samples generated in Example 2, the real training samples, and the samples that meet the requirements in the real training samples.

[0040] Figure 7 This is a scatter plot visualization result of the samples generated in Example 2, the real training samples, and the samples that meet the requirements in the real training samples after being reduced to two dimensions by t-SNE. DETAILED DESCRIPTION

[0041] The present invention will be further described below in conjunction with the examples. It should be understood that these examples are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0042] Example 1

[0043] Taking the requirements for the linear expansion coefficient of glass in the Technical Specifications for Glass Curtain Wall Engineering (JGJ 102-2003) as an example, the design goal is to generate a glass with a linear expansion coefficient of 80.00 to 100.00×10 -7 / ℃ oxide glass formula:

[0044] Step 1: Extract the data of glass component element content and corresponding linear expansion coefficient (CTE) from the SicGlass database to form an initial training sample set. Each sample in the initial training sample set is composed of the molar percentage content and linear expansion coefficient value of 80 elements with a mapping relationship.

[0045] Step 2: Perform data cleaning on the initial training sample set obtained in step 1. Data cleaning includes invalid sample removal, feature dimension reduction, extreme value sample removal, and duplicate sample removal.

[0046] Step 2-1: Invalid sample removal, remove the samples with empty linear expansion coefficient values ​​in the initial training sample set, remove the samples with oxygen content <30% to ensure that the sample data comes from oxide glass, and remove incomplete samples whose sum of components is not 100%;

[0047] Step 2-2: Feature dimensionality reduction, remove samples containing H, S, C, N, F, Cl, Br or I elements in the initial training sample set, because these elements will occupy the position of oxygen in the glass network, and remove features and corresponding samples with the number of non-zero component eigenvalues ​​less than 100;

[0048] Step 2-3: Eliminate extreme value samples. According to the "3σ principle", remove extreme samples (performance data outliers) whose linear expansion coefficient values ​​are outside [μ-3σ,μ+3σ];

[0049] Step 2-4: Duplicate sample removal: For samples with repeated glass components, their linear expansion coefficient values ​​are replaced by the median values ​​of the linear expansion coefficients of the corresponding duplicate samples, and redundant duplicate samples are removed;

[0050] Step 2-5: Rearrange the elemental characteristics of the glass components into oxide characteristics and perform sample normalization, merging the linear expansion coefficient value and the oxide content characteristic value together.

[0051] After data cleaning, the initial training sample set contains 45,702 samples, including 49 oxide features and 1 performance feature.

[0052] Step 3: Use the glass components and performance values ​​in the initial training sample set after data cleaning in step 2 as the input features of the subsequent generative adversarial network model, and perform feature normalization. Assume that the value range of each component feature is [min(n i ),max(n i )], the performance characteristic range is [min(p), max(p)], where min(n i ) and max(n i ) represents the minimum and maximum values ​​of the i-th component characteristics, min(p) and max(p) represent the minimum and maximum values ​​of the linear expansion coefficient, then 0.8×min(n i ) and 1.2×max(n i ) is the normalized parameter of the component characteristics, and 0.8×min(p) and 1.2×max(p) are the normalized parameters of the linear expansion coefficient characteristics. If there is 1.2×max(n i ) is greater than 100, it is recorded as 100;

[0053] Step 4: Based on the Wasserstein generative adversarial network framework with gradient penalty, design the network structure and training parameters of the generator and discriminator in the generative adversarial network model, as shown in Table 1. The loss function of the discriminator is:

[0054]

[0055] In the formula, Represents the data generated by the generator, represents the real training data, represents the mixed data of generated data and real training data, represents the distribution of generated data, represents the distribution of real training data, represents the distribution of mixed data, represents the output of the discriminator for the real training data, represents the output of the discriminator for the generated data, represents the output of the discriminator for mixed data, represents the calculation of mathematical expectation, It is used to represent the Wasserstein distance between the real sample and the generated sample. is the gradient penalty term used to stabilize training, and λ represents the weight parameter of the gradient penalty term.

[0056] According to the linear expansion coefficient, it is distributed in 80.00~100.00×10 -7 / ℃ target requirement, the loss function of the generator is designed as:

[0057]

[0058] In the formula, represents the component penalty term, n is the dimension of the training data, represents the generated sample after denormalization, and represents the input feature normalization parameter used when normalizing the training data, α is the weight parameter of the component penalty term, represents the performance penalty term, where 0.4255 and 0.3468 are the values ​​obtained by normalizing the boundaries of the target performance range 100 and 80 according to the normalization parameters of the linear expansion coefficient characteristics in step 3, and β is the weight parameter of the performance penalty term. The overall network architecture is as follows Figure 1 shown.

[0059] Table 1 Structural parameters and training parameters of the generative adversarial network model

[0060]

[0061] Step 5: Use the normalized initial training sample set obtained in step 3 as the input of the generative adversarial network model constructed in step 4, and perform iterative training of the generator and discriminator of the generative adversarial network. After 30,000 iterations, the loss functions of the generator and discriminator are obtained as follows: Figure 2As shown, it can be seen that the loss function of the discriminator has converged well to near 0, and the loss function of the generator decreases sharply at the beginning of training, and then slowly decreases with a very small slope and finally converges. It can be considered that the generative adversarial network model is usable after 30,000 iterations. The corresponding generator model is saved as a glass component design model for the generation of target samples.

[0062] Step 6: Use the glass component design model obtained in step 5 to generate 1000 samples that meet the design requirements of the linear expansion coefficient, and conduct experimental verification of uniqueness, diversity and accuracy.

[0063] Step 6-1: Uniqueness verification: Check that the generated samples themselves are not repeated and the generated samples are not repeated with the real training samples.

[0064] Step 6-2: Diversity verification: Extract from the real training samples the samples that have met the linear expansion coefficient of 80.00~100.00×10 -7 / ℃ samples, and visualize the component features of generated samples, real training samples, and samples that meet the requirements in the real training samples, such as Figure 3 As shown. The violin plot is superimposed with a scatter plot, where the gray violin plot and blue scatter plot represent real training samples, the orange violin plot and black scatter plot represent samples that meet the requirements in the real training samples, and the yellow violin plot and red scatter plot represent generated samples. It can be seen that the feature distribution violin plot of the generated samples can completely cover the feature distribution violin plot of the samples that meet the requirements in the real samples, indicating that the generated samples have rich patterns. The above three data sets are reduced to two dimensions through t-SNE and visualized, as shown Figure 4 As shown in Figure 2, it can be seen that the generative adversarial network model correctly captures the relationship between components and features, and the generated samples are mainly distributed around the samples that have met the requirements in the real training samples and are widely distributed.

[0065] Step 6-3: Accuracy experimental verification. A sample is randomly selected from the 1000 samples generated above for experimental verification, and the results are shown in Table 2. The experimental linear expansion coefficient of the selected sample meets the design requirements, and the relative deviation between the generated linear expansion coefficient value and the experimental linear expansion coefficient value is only 4.11%.

[0066] Table 2 Experimental verification of generated samples

[0067]

[0068] Example 2

[0069] Taking the requirements for the linear expansion coefficient and elastic modulus of glass in the Technical Specifications for Glass Curtain Wall Engineering (JGJ 102-2003) as an example, the design goal is to generate a glass with a linear expansion coefficient of 80.00 to 100.00×10 -7 Oxide glass formula with elastic modulus ≥ 72.00 GPa:

[0070] Step 1: Extract the data of glass component element content and corresponding linear expansion coefficient (CTE) and elastic modulus (E) from the SicGlass database to form an initial training sample set. Each sample in the initial training sample set is composed of the molar percentage content of 80 elements with a mapping relationship and the linear expansion coefficient value and elastic modulus value.

[0071] Step 2: Perform data cleaning on the initial training sample set obtained in step 1. Data cleaning includes invalid sample removal, feature dimension reduction, extreme value sample removal, and duplicate sample removal.

[0072] Step 2-1: Invalid sample removal: remove samples with null values ​​in the linear expansion coefficient and elastic modulus in the initial training sample set, remove samples with oxygen content less than 30% to ensure that the sample data comes from oxide glass, and remove incomplete samples whose sum of components is not 100%;

[0073] Step 2-2: Feature dimensionality reduction, remove samples containing H, S, C, N, F, Cl, Br or I elements in the initial training sample set, because these elements will occupy the position of oxygen in the glass network, and remove glass component features and corresponding samples with the number of non-zero glass component feature values ​​less than 100;

[0074] Step 2-3: Eliminate extreme value samples. According to the “3σ principle”, remove extreme samples (performance data outliers) whose linear expansion coefficient values ​​and elastic modulus values ​​are outside the corresponding [μ-3σ, μ+3σ];

[0075] Step 2-4: Duplicate sample removal: For samples with repeated glass components, their linear expansion coefficient values ​​and elastic modulus values ​​are replaced by the median values ​​of the linear expansion coefficient values ​​and elastic modulus values ​​of the corresponding duplicate samples, and redundant duplicate samples are removed;

[0076] Step 2-5: Rearrange the elemental characteristics of the glass components into oxide characteristics and perform sample normalization, and merge the linear expansion coefficient value and elastic modulus value together.

[0077] After data cleaning, the sample set contains 5718 samples, including 37 oxide characteristics and 2 performance characteristics.

[0078] Step 3: Use the components and performance values ​​of the oxide glass sample set after data cleaning in step 2 as the input features of the subsequent generative adversarial network model, and perform feature normalization. Assume that the value range of each component feature is [min(n i ),max(n i )], and the value range of each performance characteristic is [min(p k ), max(p k )], where min(n i ) and max(n i ) represents the minimum and maximum values ​​of the i-th component feature, min(p k ) and max(p k ) represents the minimum and maximum values ​​of the kth performance characteristic, then 0.8×min(n i ) and 1.2×max(n i ) is the normalization parameter of the component characteristics, with 0.8×min(p k ) and 1.2×max(p k ) is the normalized parameter of the performance characteristic, where if there is 1.2×max(n i ) is greater than 100, it is recorded as 100;

[0079] Step 4: Based on the Wasserstein generative adversarial network framework with gradient penalty, the network structure and training parameters of the generator and discriminator in the generative adversarial network model are designed, as shown in Table 3. The loss function of the discriminator is:

[0080]

[0081] In the formula, Represents the data generated by the generator, represents the real training data, represents the mixed data of generated data and real training data, represents the distribution of generated data, represents the distribution of real training data, represents the distribution of mixed data, represents the output of the discriminator for real data, represents the output of the discriminator for the generated data, represents the output of the discriminator for mixed data, represents the calculation of mathematical expectation, It is used to represent the Wasserstein distance between the real sample and the generated sample. is the gradient penalty term used to stabilize training, and λ represents the weight parameter of the gradient penalty term. According to the linear expansion coefficient, it is distributed in the range of 80.00~100.00×10 -7 / ℃ and elastic modulus ≥ 72GPa, the loss function of the generator is designed as:

[0082]

[0083] In the formula, represents the component penalty term, n is the dimension of the training data, represents the generated sample after denormalization, and represents the input feature normalization parameter used when normalizing the training data, α is the weight parameter of the component penalty term, represents the performance penalty term, where 0.4972 and 0.3983 are the values ​​obtained by normalizing the boundaries 100 and 80 of the target linear expansion coefficient range according to the normalized parameters of the linear expansion coefficient characteristics in step 3, respectively, and 0.3892 is the value obtained by normalizing the boundary 72 of the target elastic modulus range according to the normalized parameters of the elastic modulus characteristics in step 3, and β is the weight parameter of the performance penalty term.

[0084] Table 3 Structural parameters and training parameters of the generative adversarial network model

[0085]

[0086] Step 5: Use the normalized sample set obtained in step 3 as the input of the generative adversarial network model constructed in step 4, and perform iterative training of the generator and discriminator of the generative adversarial network. After 30,000 iterations, the loss functions of the generator and discriminator are obtained as follows: Figure 5 As shown, it can be seen that the loss function of the discriminator has converged well to near 0, and the loss function of the generator decreases sharply at the beginning of training, and then slowly decreases with a very small slope and finally converges. It can be considered that the generative adversarial network model is usable after 30,000 iterations. The corresponding generator model is saved as a glass component design model for the generation of target samples.

[0087] Step 6: Use the glass component design model obtained in step 5 to generate 1000 samples that meet the design requirements of linear expansion coefficient and elastic modulus, and conduct experiments to verify uniqueness, diversity and accuracy.

[0088] Step 6-1: Uniqueness verification: Check that the generated samples themselves are not repeated and the generated samples are not repeated with the real training samples.

[0089] Step 6-2: Diversity verification: Extract from the real training samples the samples that have met the linear expansion coefficient of 80.00~100.00×10 -7 / ℃ and elastic modulus ≥72GPa, the component characteristics of the generated samples, the real training samples, and the samples that meet the requirements in the real training samples are visualized, such as Figure 6 As shown. The violin plot is superimposed with a scatter plot, where the gray violin plot and blue scatter plot represent real training samples, the orange violin plot and black scatter plot represent samples that meet the requirements in the real training samples, and the yellow violin plot and red scatter plot represent generated samples. It can be seen that the feature distribution violin plot of the generated samples can completely cover the feature distribution violin plot of the samples that meet the requirements in the real samples, indicating that the generated samples have rich patterns. The above three data sets are reduced to two dimensions through t-SNE and visualized, as shown Figure 7 As shown in Figure 2, it can be seen that the generative adversarial network model correctly captures the relationship between components and features, and the generated samples are mainly distributed around the samples that have met the requirements in the real training samples and are widely distributed.

[0090] Step 6-3: Accuracy experimental verification. A sample is randomly selected from the 1000 samples generated above for experimental verification, and the results are shown in Table 4. The experimental linear expansion coefficient and elastic modulus of the selected sample meet the design requirements, and the relative deviation between the generated linear expansion coefficient and the experimental linear expansion coefficient is only 8.36%, and the relative deviation between the generated elastic modulus and the experimental elastic modulus is only -0.35%.

[0091] Table 4 Experimental verification of generated samples

[0092]

[0093] It can be seen from the above two embodiments that the glass component design method based on the generative adversarial network model provided by the present invention has an excellent generation effect and can effectively shorten the design and development cycle of new glass materials.

[0094] The embodiments described above provide a detailed description of the technical solutions of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements or similar substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A glass component design method based on a generative adversarial network model, characterized in that: The following steps are involved: Step 1: Collect glass data and construct an initial training sample set, wherein the initial training sample set includes glass components with a mapping relationship and their corresponding performance values; Step 2: performing data cleaning and feature normalization on the initial training sample set in turn; Step 3: Construct the network structure of the generator and discriminator in the generative adversarial network model, and use the initial training sample set after feature normalization to iteratively train the generative adversarial network model. During training, the discriminator loss function based on the Wasserstein distance and the gradient penalty term is used to optimize and update the parameters of the discriminator, and the designed generator loss function based on the component penalty term and the performance penalty term is used to optimize and update the parameters of the generator. After the iterative training is completed, the generator model is used as the glass component design model; The glass composition design model is used to generate glass compositions that meet the target glass performance requirements.

2. The glass component design method based on the generative adversarial network model according to claim 1, characterized in that: In step 1, the performance value includes at least one of density, linear expansion coefficient, elastic modulus, hardness, Young's modulus, shear modulus, Poisson's ratio, glass transition temperature, strain point temperature, annealing point temperature, softening point temperature, liquidus temperature, and refractive index.

3. The glass component design method based on the generative adversarial network model according to claim 1, characterized in that: Data cleaning in step 2 includes invalid sample removal, feature dimension reduction, extreme value sample removal, and duplicate sample removal; Invalid sample elimination: remove samples containing null values ​​in the glass target performance values ​​in the initial training sample set, and remove incomplete samples whose sum of glass components is not 100%; Feature dimensionality reduction: remove samples containing placeholder elements in the initial training sample set, and remove glass component features and corresponding samples whose number of non-zero glass component feature values ​​is less than the set number; Elimination of extreme value samples: Eliminate samples with abnormal performance values ​​according to the "3σ principle"; Duplicate sample elimination: For samples with repeated glass components, their performance values ​​are replaced by the median performance values ​​of the corresponding repeated samples, and redundant duplicate samples are removed.

4. The glass component design method based on the generative adversarial network model according to claim 3, characterized in that: When constructing an initial training sample set in which all samples are oxide glasses, when removing invalid samples in step 2, it is also necessary to remove samples with an oxygen content of less than 30%; when reducing the feature dimension, the placeholder elements are H, S, C, N, F, Cl, Br or I elements.

5. The glass component design method based on the generative adversarial network model according to claim 1, characterized in that: In the process of feature normalization, the value range of each component feature is set to [min(n i ),max(n i )], the range of each performance value characteristic is [min(p k ), max(p k )], where min(n i ) and max(n i ) represents the minimum and maximum values ​​of the i-th component feature, min(p k ) and max(p k ) represents the minimum and maximum values ​​of the kth performance value characteristic, then 0.8×min(n i ) and 1.2×max(n i ) is the normalization parameter of the component characteristics, with 0.8×min(p k ) and 1.2×max(p k ) is the normalized parameter of the performance value characteristic; where, if there is 1.2×max(n i ) is greater than 100, it is recorded as 100.

6. The glass component design method based on the generative adversarial network model according to claim 1, characterized in that: Specifically, the framework of the generative adversarial network model used in step 3 is based on the Wasserstein generative adversarial network with a gradient penalty term, and the generator and the discriminator are both fully connected neural network structures with multiple hidden layers.

7. The glass component design method based on the generative adversarial network model according to claim 1, characterized in that: The discriminator loss function is expressed as: In the formula, Represents the generated data of the generator, represents the real training data, represents the mixed data of generated data and real training data, represents the distribution of generated data, represents the distribution of real training data, represents the distribution of mixed data, represents the output of the discriminator for the real training data, represents the output of the discriminator for the generated data, represents the output of the discriminator for mixed data, represents the calculation of mathematical expectation, It is used to represent the Wasserstein distance between real samples and generated samples. is the gradient penalty term used to stabilize training, and λ represents the weight parameter of the gradient penalty term.

8. The glass component design method based on the generative adversarial network model according to claim 1, characterized in that: In the generator loss function, the component penalty term is used to constrain the sum of the components of the generated samples to meet 100% of the actual requirements, and the performance penalty term is used to constrain the generator to generate in the direction that the performance of the generated samples meets the design target range.

9. The glass component design method based on the generative adversarial network model according to claim 1, characterized in that: The generator loss function is expressed as: In the formula, represents the component penalty term, n is the dimension of the training data, m is the number of target performances to be designed, represents the generated sample after denormalization, and represents the input feature normalization parameter used when normalizing the training data, α is the weight parameter of the component penalty term, and H represents the performance penalty term; in the single performance target design task, m = 1. When the target performance design interval has only one explicit boundary, if the target performance value is required to be less than this boundary value, then If the target performance is required to be greater than this boundary value, When the target performance design interval has two explicit boundaries, in Indicates the generated sample The nth feature, P1, P 1_bottom , P 1_top It represents the value after the boundary value of the glass target performance design range is normalized under the corresponding performance value characteristic normalization parameter; In the multi-performance target design task, m>1, the same strategy as above is used to design the corresponding performance penalty item for each performance target, and the sum of the designed performance penalty items is taken as the final performance penalty item H; β represents the weight parameter of the performance penalty term.

10. The glass component design method based on the generative adversarial network model according to claim 1, characterized in that: In step 3, the network structure parameters include the number of hidden layers, the number of neurons in each hidden layer, and the type of activation function used by the neurons in each hidden layer. The training parameters include the optimizer type, optimizer learning rate, batch size, number of iterations, weight parameters of each penalty item in the generator and discriminator, whether to prioritize training the discriminator or the generator, and the ratio of the number of training times of the discriminator to the generator.

Images