Alloy component prediction method based on domain adaptive generative adversarial network

Abstract

The invention belongs to the technical field of alloy component prediction, and particularly relates to an alloy component prediction method based on a domain adaptive generative adversarial network, which effectively aligns data distribution of a source domain and a target domain by introducing the domain adaptive generative adversarial network, remarkably improves the alloy component prediction precision, and improves the alloy component prediction accuracy. According to the method, the model is finely adjusted on the target domain data, so that the generalization ability of the model is enhanced, and the model can keep higher prediction performance under the condition of different experimental equipment or simulation methods.

Description

Technical Field

[0001] The present invention relates to the technical field of alloy composition prediction, and in particular to an alloy composition prediction method based on a domain adaptive generative adversarial network. Background Art

[0002] Alloy materials play a key role in modern manufacturing and are widely used in aerospace, automobile manufacturing, electronic equipment and other fields. The performance of alloys directly affects the quality and service life of products, so the demand for the development of high-performance alloy materials is becoming more and more urgent. With the development of computational materials science, data-driven methods have gradually gained attention in alloy composition design, especially the application of machine learning and deep learning technology, which has opened up a new perspective for alloy composition prediction.

[0003] Existing alloy composition prediction models often show poor generalization ability when dealing with inconsistent data distribution generated by different experimental equipment or different simulation methods. This leads to a decrease in the performance of the alloy composition prediction model in the target domain, thus limiting its promotion and use in practical applications. Solving the performance degradation problem caused by inconsistent data distribution has become an important challenge to improve the accuracy of alloy composition prediction.

[0004] Currently, commonly used alloy composition prediction methods mainly rely on traditional machine learning models, which usually require a large amount of high-quality annotated data and have strict assumptions about data distribution. The diversity and complexity of alloy materials make it difficult and costly to obtain sufficient annotated data. When a single model handles tasks with different data distributions, it is prone to overfitting and insufficient generalization, making it difficult to fully utilize cross-domain data for effective prediction. Summary of the invention

[0005] In order to solve the problems existing in the prior art, the main purpose of the present invention is to propose an alloy composition prediction method based on domain adaptive generative adversarial network.

[0006] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:

[0007] A method for predicting alloy composition based on a domain adaptive generative adversarial network comprises the following steps:

[0008] S1. Establish source domain and target domain datasets, and perform data cleaning and preprocessing;

[0009] S2, extract effective features from the data and use feature selection methods to remove redundant or irrelevant features;

[0010] S3. Select a machine learning algorithm, train the initial model on the source domain data, and perform a 10-fold cross-validation to evaluate the performance as the basis for subsequent generative adversarial network training.

[0011] S4. Construct a domain-adaptive generative adversarial network. The generator converts source domain data into target domain feature data. The discriminator distinguishes between real and generated data. The training is repeated and the adversarial loss is optimized to achieve feature alignment between the source domain and the target domain.

[0012] S5. Use the trained generator to convert the source domain data into target domain feature data, generate a target domain extended data set, and fine-tune the initial model to obtain a fine-tuned model adapted to the target domain characteristics;

[0013] S6. Evaluate the performance of the fine-tuned model on the target domain data to determine whether the model meets the required evaluation indicators. If it does, apply the fine-tuned model to the actual alloy composition prediction. If it does not, repeat S3-S6 and retrain the initial model.

[0014] As a preferred solution of the alloy composition prediction method based on domain adaptive generative adversarial network described in the present invention, the alloy includes aluminum alloy, magnesium alloy, copper alloy, titanium alloy and high entropy alloy.

[0015] As a preferred embodiment of the alloy composition prediction method based on domain adaptive generative adversarial network described in the present invention, wherein: in the step S1, the source domain and target domain data sets contain composition data and performance data, the source domain data set helps the model learn richer features during training, and the target domain data set provides the model with features that must be learned to adapt.

[0016] As a preferred embodiment of the alloy composition prediction method based on domain adaptive generative adversarial network described in the present invention, in step S1, data cleaning and preprocessing include processing missing values, outliers, and standardization or normalization.

[0017] As a preferred solution of the alloy composition prediction method based on domain adaptive generative adversarial network described in the present invention, wherein: in the step S1,

[0018] Collect source domain data sets, including alloy composition data and corresponding target performance data;

[0019] Collect target domain data sets, including alloy composition data and corresponding target performance data;

[0020] The source domain dataset helps the model learn richer features during training;

[0021] The target domain dataset provides the features that the model must learn to adapt to;

[0022] Preprocessing of missing values ​​is done by inserting 0;

[0023] Preprocessing of outliers is done by removing them;

[0024] Preprocess the selected features by Z-score standardization or Min-Max normalization;

[0025] The Z-score standardization method preprocesses the data according to the formula

[0026] Among them, z is the standardized value, x is the original data value, μ is the mean of the data, and σ is the standard deviation of the data;

[0027] Min-Max normalization is used to preprocess the data according to the formula

[0028] Among them, x′ is the normalized value, x is the original data value, and x min is the minimum value of the data set, x max is the maximum value of the data set.

[0029] As a preferred solution of the alloy composition prediction method based on domain adaptive generative adversarial network described in the present invention, wherein: in the step S2,

[0030] The extracted effective features include the content of key alloy elements, element ratio, heat treatment time and temperature; when selecting features, elements with more than one-third of the occurrence times and an average content greater than 0.01wt% are used as composition features.

[0031] As a preferred embodiment of the alloy composition prediction method based on domain adaptive generative adversarial network described in the present invention, wherein: in the step S3, the initial model includes a support vector machine, a neural network, and a gradient boosting decision tree; the training set is divided into 10 subsets by a ten-fold cross-validation method, and 9 of the subsets are used in turn as training sets to train the machine learning model, and the remaining 1 subset is used as a test set to find the optimal parameters and calculate the accuracy of the test set, which serves as the basis for subsequent generative adversarial network training.

[0032] As a preferred solution of the alloy composition prediction method based on domain adaptive generative adversarial network described in the present invention, wherein: in the step S4,

[0033] Design a generator network to convert source domain data into data with target domain characteristics through the generator, and learn the distribution difference between the source domain and the target domain;

[0034] The generator network inputs source domain data and outputs pseudo target domain data, generating feature distribution similar to that of the target domain;

[0035] Design a discriminator network. The discriminator network inputs real target domain data and pseudo target domain data generated by the generator to train the ability to distinguish between real data and generated data.

[0036] The generator and discriminator are trained alternately, and the features of the source domain data are aligned with the features of the target domain data by optimizing the adversarial loss function.

[0037] The adversarial loss function uses binary cross entropy loss, and the formula is:

[0038]

[0039] Among them, xq is the real target domain sample, z is a random vector, G q (z) is the generated sample, D q (G q (z)) and D q (x q ) are the predicted probabilities of the discriminator for the generated samples and the real samples respectively;

[0040] The loss function of the generator is expressed as:

[0041]

[0042] The generator parameter update formula is:

[0043]

[0044] Among them, θ G represents the parameters of the generator, η is the learning rate, is the generator loss function L G For its parameter θ G The gradient of

[0045] The loss function of the discriminator is expressed as:

[0046]

[0047] Discriminator parameter update formula:

[0048]

[0049] Among them, θ D represents the parameters of the discriminator, η is the learning rate, is the discriminator loss function L D For its parameter θ D gradient.

[0050] As a preferred solution of the alloy composition prediction method based on domain adaptive generative adversarial network described in the present invention, wherein: in the step S5,

[0051] Using the trained generator, the source domain data is converted into target domain style data, generating synthetic data with similar style to the target domain, expanding the size of the target domain dataset and providing more samples for the model to learn, thus enhancing the performance of the model in the target domain.

[0052] Using the generated target domain style data, fine-tune the initial model to adapt it to the distribution characteristics of the target domain data;

[0053] The fine-tuning process involves taking the generated data as a new training set, retraining the model parameters, and optimizing its prediction performance on the target domain data.

[0054] As a preferred solution of the alloy composition prediction method based on domain adaptive generative adversarial network described in the present invention, wherein: in the step S6,

[0055] Comprehensively evaluate the model’s performance on target domain data;

[0056] The evaluation indicators include mean square error (MSE), mean absolute error (MAE), determination coefficient (R 2 );The mean square error evaluates the model according to the formula

[0057] The mean absolute error is used to evaluate the model according to the formula

[0058] The coefficient of determination evaluates the model according to the formula

[0059] Where n is the total number of samples, y i is the actual value of the ith sample, is the predicted value of the ith sample, is the average of the actual values;

[0060] MAE<0.15, MSE<0.35, R 2 When >0.9, the model is considered to meet the required evaluation indicators.

[0061] The beneficial effects of the present invention are as follows:

[0062] The present invention proposes an alloy composition prediction method based on a domain adaptive generative adversarial network. By introducing a domain adaptive generative adversarial network, the data distribution of the source domain and the target domain is effectively aligned, which significantly improves the accuracy of alloy composition prediction. The method enhances the generalization ability of the model by fine-tuning the model on the target domain data, so that it can maintain a high prediction performance under different experimental equipment or simulation method conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0064] Figure 1 Schematic diagram of the process of the alloy composition prediction method of the present invention.

[0065] Figure 2 Schematic diagram of the evaluation results of the Al-Zn-Mg-Cu alloy composition prediction model.

[0066] Figure 3 Schematic diagram of the comparison between the predicted and actual values ​​of the Al-Zn-Mg-Cu alloy composition.

[0067] The realization of the effects, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0068] The following will be described clearly and completely in conjunction with the technical solutions in the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] The main purpose of the present invention is to propose an alloy composition prediction method based on a domain adaptive generative adversarial network, which achieves the alignment of source domain and target domain data through a domain adaptive generative adversarial network, thereby improving the performance of the alloy composition prediction model in the target domain. This method can also be trained on smaller data sets, reducing the dependence on a large amount of annotated data, while accelerating the research and development process of alloy composition prediction, providing a new solution for alloy design, and helping to promote the optimization and application of alloy materials.

[0070] like Figure 1 As shown, a method for predicting alloy composition based on a domain adaptive generative adversarial network includes the following steps:

[0071] S1. Establish source domain and target domain datasets, and perform data cleaning and preprocessing;

[0072] S2, extract effective features from the data and use feature selection methods to remove redundant or irrelevant features;

[0073] S3. Select a machine learning algorithm, train the initial model on the source domain data, and perform a 10-fold cross-validation to evaluate the performance as the basis for subsequent generative adversarial network training.

[0074] S4. Construct a domain-adaptive generative adversarial network. The generator converts source domain data into target domain feature data. The discriminator distinguishes between real and generated data. The training is repeated and the adversarial loss is optimized to achieve feature alignment between the source domain and the target domain.

[0075] S5. Use the trained generator to convert the source domain data into target domain feature data, generate a target domain extended data set, and fine-tune the initial model to obtain a fine-tuned model adapted to the target domain characteristics;

[0076] S6. Evaluate the performance of the fine-tuned model on the target domain data to determine whether the model meets the required evaluation indicators. If it does, apply the fine-tuned model to the actual alloy composition prediction. If it does not, repeat S3-S6 and retrain the initial model.

[0077] In a specific embodiment of the present invention, the alloy includes aluminum alloy, magnesium alloy, copper alloy, titanium alloy and high entropy alloy.

[0078] In a specific embodiment of the present invention, in step S1, the source domain and target domain datasets contain component data and performance data, the source domain dataset helps the model learn richer features during training, and the target domain dataset provides the model with features that must be learned to adapt.

[0079] In a specific embodiment of the present invention, in step S1, data cleaning and preprocessing include processing missing values, outliers, and standardization or normalization.

[0080] In a specific embodiment of the present invention, in step S1, a source domain data set is collected, including alloy composition data and corresponding target performance data; and a target domain data set is collected, including alloy composition data and corresponding target performance data.

[0081] In a specific embodiment of the present invention, in step S1, the source domain dataset helps the model learn richer features during training, and the target domain dataset provides the model with features that must be learned to adapt.

[0082] In a specific embodiment of the present invention, in step S1, preprocessing of missing values ​​is completed by inserting 0; preprocessing of abnormal values ​​is completed by deleting abnormal values.

[0083] In a specific embodiment of the present invention, in step S1, the screened features are preprocessed by Z-score standardization or Min-Max normalization;

[0084] The Z-score standardization method preprocesses the data according to the formula

[0085] Among them, z is the standardized value, x is the original data value, μ is the mean of the data, and σ is the standard deviation of the data;

[0086] Min-Max normalization is used to preprocess the data according to the formula

[0087] Among them, x′ is the normalized value, x is the original data value, and x min is the minimum value of the data set, x max is the maximum value of the data set.

[0088] In a specific embodiment of the present invention, step S2 is specifically as follows:

[0089] The extracted effective features include the content of key alloy elements, element ratio, heat treatment time and temperature; when selecting features, elements with more than one-third of the occurrence times and an average content greater than 0.01wt% are used as composition features.

[0090] In a specific embodiment of the present invention, in step S3, the initial model includes a support vector machine, a neural network, and a gradient boosting decision tree; the training set is divided into 10 subsets by a ten-fold cross validation method, and 9 of the subsets are used in turn as training sets to train the machine learning model, and the remaining 1 subset is used as a test set to find the optimal parameters and calculate the accuracy of the test set as the basis for subsequent generative adversarial network training. In a specific embodiment of the present invention, in step S4, the goal of the generator is to generate realistic target domain data, and the goal of the discriminator is to correctly distinguish between real and generated data.

[0091] In a specific embodiment of the present invention, in step S4,

[0092] Design a generator network to convert source domain data into data with target domain characteristics through the generator, and learn the distribution difference between the source domain and the target domain;

[0093] The generator network inputs source domain data and outputs pseudo target domain data, generating feature distribution similar to that of the target domain;

[0094] Design a discriminator network. The discriminator network inputs real target domain data and pseudo target domain data generated by the generator to train the ability to distinguish between real data and generated data.

[0095] The generator and discriminator are trained alternately, and the features of the source domain data are aligned with the features of the target domain data by optimizing the adversarial loss function.

[0096] In a specific embodiment of the present invention, in step S4,

[0097] The adversarial loss function uses binary cross entropy loss, and the formula is:

[0098]

[0099] Among them, xq is the real target domain sample, z is a random vector, G q (z) is the generated sample, D q (G q (z)) and D q (x q ) are the predicted probabilities of the discriminator for the generated samples and the real samples respectively;

[0100] The loss function of the generator is expressed as:

[0101]

[0102] The generator parameter update formula is:

[0103]

[0104] Among them, θ G represents the parameters of the generator, η is the learning rate, is the generator loss function L G For its parameter θ G The gradient of

[0105] The loss function of the discriminator is expressed as:

[0106]

[0107] Discriminator parameter update formula:

[0108]

[0109] Among them, θ D represents the parameters of the discriminator, η is the learning rate, is the discriminator loss function L D For its parameter θ D gradient.

[0110] In a specific embodiment of the present invention, in step S5,

[0111] Using the trained generator, the source domain data is converted into target domain style data, generating synthetic data with similar style to the target domain, expanding the size of the target domain dataset and providing more samples for the model to learn, thus enhancing the performance of the model in the target domain.

[0112] Using the generated target domain style data, fine-tune the initial model to adapt it to the distribution characteristics of the target domain data;

[0113] The fine-tuning process involves taking the generated data as a new training set, retraining the model parameters, and optimizing its prediction performance on the target domain data.

[0114] In a specific embodiment of the present invention, in step S6,

[0115] Comprehensively evaluate the model’s performance on target domain data;

[0116] The evaluation indicators include mean square error (MSE), mean absolute error (MAE), determination coefficient (R 2 );The mean square error evaluates the model according to the formula

[0117] The mean absolute error is used to evaluate the model according to the formula

[0118] The coefficient of determination evaluates the model according to the formula

[0119] Where n is the total number of samples, y i is the actual value of the ith sample, is the predicted value of the ith sample, is the average of the actual values;

[0120] MAE<0.15, MSE<0.35, R 2 When >0.9, the model is considered to meet the required evaluation indicators.

[0121] By using the method provided by the present invention, the tensile strength is 650MPa, 700MPa, 750MPa, and the elongation is 8% as the performance target, and the evaluation indicators MAE<0.15, MSE<0.35, R 2 >0.9, three alloy compositions were designed. The design method is described as follows:

[0122] (1) Data collection: The source domain data set and target domain data set of the alloy were established by collecting and organizing the publicly available literature data on the composition, process, tensile strength and elongation of Al-Zn-Mg-Cu alloys. The collected data include the mass fractions of elements such as Zn, Mg, Cu and Zr, as shown in Table 1. The tensile strength and elongation are used as the target performance data, as shown in Table 2.

[0123] Table 1

[0124] element Range (wt.%) Mean Standard Deviation total Zn 2.393-15.24 8.23 1.94 282 Mg 0.651-5 2.37 0.48 282 Cu 0.01-2.6 1.77 0.59 282 Zr 0.011-0.8 0.15 0.06 235 Sc 0.015-0.74 0.17 0.13 41 Yb 0.19-0.3 0.26 0.05 28 Si 0.002-0.26 0.06 0.06 66 Fe 0.002-0.3 0.11 0.08 82 Ti 0.004-0.81 0.11 0.14 64 Mn 0.001-0.4 0.18 0.15 25 Cr 0.001-0.24 0.15 0.06 58 Ce 0.1-0.2 0.15 0.07 2 Y 0.15-0.45 0.28 0.1 6 Sr 0.015-0.12 0.08 0.04 14 Er 0.1-2 0.49 0.55 16 Ni 0.1-0.2 0.14 0.05 3 Pr 0.1-0.21 0.15 0.04 7 Nb 0.05-0.1 0.05 20.03 2 Ag 0.166-0.2 0.3 0.02 2 Sm 0.1-0.5 0.18 0.17 6 La 0.1-0.15 0.15 0.03 2

[0125] Table 2

[0126]

[0127]

[0128] (2) Data cleaning: The collected Al-Zn-Mg-Cu alloy data are classified according to the preparation process (casting, homogenization, extrusion / rolling, solution, aging), and the alloy composition and performance data under the extrusion process are used as the source domain data, and a small amount of data under the rolling process is used as the target domain data. The outliers in the data are eliminated, the noise data is deleted, and 0 is inserted into the vacant values ​​to form the initial data set;

[0129] (3) Data normalization: Normalize the initial data sets of the source domain and the target domain, and use the Min-Max normalization method to process all component data and performance data to the [0, 1] interval;

[0130] (4) Constructing the initial model: Constructing a BP neural network model on the source domain data and using ten-fold cross validation to evaluate its prediction performance as the basis for subsequent generative adversarial network training;

[0131] (5) Constructing a domain-adaptive generative adversarial network model: A generative adversarial network model is established, and the source domain data is converted into a distribution similar to that of the target domain through the generator; the generator network input is the source domain data, and the output is the pseudo target domain data; the generator adopts a 3-layer fully connected neural network, the number of hidden layer nodes is 64, 32, and 16 respectively, and the transfer function is ReLU; the discriminator network input is the target domain data and the pseudo target domain data generated by the generator, and the output is the real or generated data; the discriminator adopts a 2-layer fully connected network, the number of nodes is 64 and 32 respectively, LeakyReLU is used as the activation function, and a Dropout layer is set to prevent overfitting; the feature extractor shares a layer of feature extractor between the source domain and the target domain, inputs the component data, and outputs a unified feature representation, which is used to connect the input of the generator and the discriminator. The feature extractor uses ReLU activation;

[0132] (6) Model training: The generator and discriminator are trained alternately using the adversarial loss function. The Adam optimizer is used with a learning rate of 0.005. The generator and discriminator are trained alternately until the generator outputs alloy composition data with target domain characteristics.

[0133] (7) Alloy performance → composition BP neural network model: Based on the target domain data generated by the domain adaptive generative adversarial network, a BP neural network model that derives composition from target performance is established; the target performance is used as input, the alloy composition data is used as output, the network structure is set to 2 hidden layers, the number of nodes is 64 and 32 respectively, and the activation function is Sigmoid. The loss function is the mean square error (MAE), and the Adam optimizer is used for training;

[0134] (8) Alloy composition design and optimization: Input the target performance (tensile strength of 650 MPa, 700 MPa, 750 MPa and elongation of 8%) into the BP neural network model, output the initial design of the alloy composition combination, and use the target domain data generated by the domain adaptive generative adversarial network as a reference for further fine-tuning;

[0135] (9) Performance evaluation and adjustment: The components output by the BP neural network are input into the domain adaptive generative adversarial network to predict the corresponding target domain performance and compare it with the given target performance. MAE, MSE, R 2 The model is evaluated in the validation set as an evaluation index; the smaller the MAE and MSE indicators are, the better the model performance is. 2 The closer it is to 1, the better the model performance;

[0136] (10) Denormalization: The alloy composition determined in step (8) was denormalized using the same parameters as in step (3) to obtain the final design compositions of the three alloys, as shown in Table 3.

[0137] Table 3

[0138]

[0139] The present invention introduces a domain adaptive generative adversarial network to effectively align the data distribution of the source domain and the target domain, thereby significantly improving the accuracy of alloy composition prediction. The method enhances the generalization ability of the model by fine-tuning the model on the target domain data, so that it can maintain a high prediction performance under different experimental equipment or simulation method conditions.

[0140] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. All equivalent structural changes made by using the contents of the present invention specification under the inventive concept of the present invention, or directly / indirectly applied in other related technical fields are included in the patent protection scope of the present invention.

Claims

1. A method for predicting alloy composition based on domain adaptive generative adversarial network, characterized in that: The steps include: S1. Establish source domain and target domain datasets, and perform data cleaning and preprocessing; S2, extract effective features from the data and use feature selection methods to remove redundant or irrelevant features; S3. Select a machine learning algorithm, train the initial model on the source domain data, and perform a 10-fold cross-validation to evaluate the performance as the basis for subsequent generative adversarial network training. S4. Construct a domain-adaptive generative adversarial network. The generator converts source domain data into target domain feature data. The discriminator distinguishes between real and generated data. The training is repeated and the adversarial loss is optimized to achieve feature alignment between the source domain and the target domain. S5. Use the trained generator to convert the source domain data into target domain feature data, generate a target domain extended data set, and fine-tune the initial model to obtain a fine-tuned model adapted to the target domain characteristics; S6. Evaluate the performance of the fine-tuned model on the target domain data to determine whether the model meets the required evaluation indicators. If it does, apply the fine-tuned model to the actual alloy composition prediction. If it does not, repeat S3-S6 and retrain the initial model.

2. The alloy composition prediction method based on domain adaptive generative adversarial network according to claim 1 is characterized in that: The alloys include aluminum alloys, magnesium alloys, copper alloys, titanium alloys and high entropy alloys.

3. The alloy composition prediction method based on domain adaptive generative adversarial network according to claim 1 is characterized in that: In step S1, the source domain and target domain data sets include component data and performance data.

4. The alloy composition prediction method based on domain adaptive generative adversarial network according to claim 1 is characterized in that: In step S1, data cleaning and preprocessing include processing missing values, outliers, and standardization or normalization.

5. The alloy composition prediction method based on domain adaptive generative adversarial network according to claim 1, characterized in that: In step S2, the extracted effective features include the content of key alloy elements, element ratio, heat treatment time and temperature; When selecting features, elements that appear more than one-third of the time and have an average content greater than 0.01 wt% are used as composition features.

6. The alloy composition prediction method based on domain adaptive generative adversarial network according to claim 1, characterized in that: In step S3, the initial model includes a support vector machine, a neural network, and a gradient boosting decision tree; the training set is divided into 10 subsets through a ten-fold cross-validation method, and 9 of the subsets are used in turn as training sets to train the machine learning model, and the remaining 1 subset is used as a test set to find the optimal parameters and calculate the accuracy of the test set, which serves as the basis for subsequent generative adversarial network training.

7. The alloy composition prediction method based on domain adaptive generative adversarial network according to claim 1 is characterized in that: In the step S4, Design a generator network to convert source domain data into data with target domain characteristics through the generator, and learn the distribution difference between the source domain and the target domain; The generator network inputs source domain data and outputs pseudo target domain data, generating feature distribution similar to that of the target domain; Design a discriminator network. The discriminator network inputs real target domain data and pseudo target domain data generated by the generator to train the ability to distinguish between real data and generated data. The generator and discriminator are trained alternately, and the features of the source domain data are aligned with the features of the target domain data by optimizing the adversarial loss function.

8. The alloy composition prediction method based on domain adaptive generative adversarial network according to claim 7 is characterized in that: In the step S4, The adversarial loss function uses binary cross entropy loss, and the formula is: Among them, xq is the real target domain sample, z is a random vector, G q (z) is the generated sample, D q (G q (z)) and D q (x q ) are the predicted probabilities of the discriminator for the generated samples and the real samples respectively; The loss function of the generator is expressed as: The generator parameter update formula is: Among them, θ G represents the parameters of the generator, η is the learning rate, is the generator loss function L G For its parameter θ G The gradient of The loss function of the discriminator is expressed as: Discriminator parameter update formula: Among them, θ D represents the parameters of the discriminator, η is the learning rate, is the discriminator loss function L D For its parameter θ D gradient.

9. The alloy composition prediction method based on domain adaptive generative adversarial network according to claim 1, characterized in that: In step S5, the source domain data is converted into target domain style data using the trained generator to generate synthetic data similar to the target domain style, expand the target domain data set size and provide more samples for model learning, and enhance the performance of the model in the target domain; the initial model is fine-tuned using the generated target domain style data to adapt it to the distribution characteristics of the target domain data; the fine-tuning process includes using the generated data as a new training set, retraining the model parameters, and optimizing its prediction performance on the target domain data.

10. The alloy composition prediction method based on domain adaptive generative adversarial network according to claim 1, characterized in that: In step S6, the performance of the model is comprehensively evaluated on the target domain data; the evaluation indicators include mean square error MSE, mean absolute error MAE, determination coefficient R 2 , MAE<0.15, MSE<0.35, R 2 When >0.9, the model is considered to meet the required evaluation indicators.

Images