A metal material design method based on deep learning

Abstract

The present invention discloses a metal material design method based on deep learning, comprising: S1, collecting real chemical composition data, forming training samples, performing data preprocessing on the training samples, and increasing the number of training samples through data enhancement technology, and then improving the feature expression ability through a self-attention mechanism; S2, designing a model architecture, the model architecture includes two basic elements: a generator and a discriminator; S3, in the model architecture, generating data similar to the real chemical composition through the generator, and using an objective function to measure the similarity between the generated data and the actual data; S4, training a generative adversarial network to achieve a balance between the generator and the discriminator. The metal material design method based on deep learning described in the present invention has the advantages of convenient operation of the material design process, low cost, and ease of use.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent material design, and in particular to a metal material design method based on deep learning. Background Art

[0002] Materials are one of humanity's most widely used means of production. Their properties are determined by their organization, which in turn is determined by their chemical composition. There are many different chemical elements, and even a slight change in their content can alter a material's properties. The chemical elements in a material's composition, due to differences in their electronic structures and content, can form a variety of complex compounds. The distribution of these compounds within the material further influences its properties.

[0003] At present, the design of new materials is still in the empirical science stage. New materials are designed based on the experience of previous generations in using materials. This design method requires a lot of time and R&D costs, and there is no guarantee that the final material design will be successful.

[0004] Therefore, it is necessary to develop a convenient, low-cost and easy-to-use metal material design method. Summary of the Invention

[0005] The object of the present invention is to provide a convenient, low-cost, and easy-to-use deep learning-based metal material design method to address the above-mentioned problems.

[0006] The technical solution adopted in the present invention is as follows:

[0007] A metal material design method based on deep learning, comprising the steps of:

[0008] S1: Collect real chemical composition data, form training samples, preprocess the training samples, increase the number of training samples through data augmentation technology, and then improve the feature expression ability through the self-attention mechanism;

[0009] S2, design the model architecture, which includes two basic elements: the generator and the discriminator;

[0010] S3, in the model architecture, generates data similar to the real chemical composition through the generator, and uses the objective function to measure the similarity between the generated data and the actual data;

[0011] S4, train the adversarial generative network to achieve a balance between the generator and the discriminator.

[0012] In step S1, a training sample library is constructed by collecting real chemical composition data, using data augmentation techniques to address the scarcity of material data. A self-attention mechanism is introduced to model nonlinear relationships between elements, such as capturing the synergistic effects of trace alloying elements on crystal structure. Discrete chemical element contents are mapped into a high-dimensional feature space, providing an information foundation for subsequent generation. In step S2, an adversarial mechanism between the generator and the discriminator forces the generator to continuously approximate the data distribution of real materials. In step S3, an objective function is used to measure the similarity between the generated data and the real data, ensuring that the generated data not only resembles the real data distribution but also covers unique material combinations in the long-tail distribution. In step S4, the basic distribution of material composition is first learned through pre-training the generator, and then adversarial training is initiated. The resulting model generates compositional design solutions with a gradient structure, breaking through the single-point optimization limitations of traditional trial-and-error methods. At the microscopic level, the self-attention mechanism of this method analyzes quantum-scale interactions between elements. At the mesoscopic level, the adversarial mechanism of GANs simulates the competitive relationship between compound formation energies. At the macroscopic level, the generator outputs a gradient composition distribution that meets the target performance.

[0013] Optionally, step S1 includes:

[0014] S11, data collection, collects a large amount of real chemical composition data of relevant materials in the material direction to be designed and forms training samples;

[0015] S12, data preprocessing, data preprocessing of the chemical components in the training samples to reduce the scale differences of the chemical component contents;

[0016] S13, data enhancement, training at least one of the samples to rearrange the positions of the chemical elements therein to generate new training samples to achieve data enhancement;

[0017] S14, self-attention feature representation, improves the feature expression ability and training sample capacity through the self-attention mechanism.

[0018] S11 builds an initial sample library covering multiple element combinations by widely collecting real chemical composition data in the target material field, solving the data island problem in material research and development and providing sufficient learning materials for deep learning. S12 eliminates the dimensional differences in the content of different chemical elements to prevent the model from over-focusing on specific elements due to numerical magnitude deviations, ensuring that the characteristics of each component are comparable during training. S13 reorganizes the chemical element arrangement of the original sample to generate equivalent variant samples within the range allowed by physical laws, breaking through the limitations of experimental data scarcity and enhancing the model's ability to understand the diversity of component combinations. S14 uses the self-attention mechanism to dynamically capture the nonlinear correlation between elements, and through the adaptive allocation of weights in the feature space, strengthens the characterization ability of key component combinations, while implicitly expanding the coverage of data distribution in high-dimensional space, improving the model's analytical accuracy for complex material systems.

[0019] Alternatively, in step S14,

[0020] The self-attention feature of the selected element is represented as: W1×a+W2×b+W3×c+W4×d+W5×e;

[0021] Among them, W1, W2, W3, W4, and W5 are the weights of the corresponding elements, which are randomly initialized and are trainable parameters; a, b, c, d, and e are the contents of the remaining elements in the training samples.

[0022] In the self-attention feature representation of step S14, the proposed linear combination formula dynamically adjusts the contribution of different chemical elements to the overall material characterization by introducing trainable weight parameters. During training, the model automatically learns the implicit correlation patterns between elements through backpropagation, allowing the final feature representation to quantitatively reflect the synergistic or competitive relationships between elements. The introduction of weight parameters gives the model the ability to reconstruct feature space. Even if the input is the raw element content scalar value (a to e), the combined effect of the linear transformation of the weight matrix and the subsequent nonlinear activation function can map low-dimensional chemical composition data to a high-dimensional space, explicitly encoding the quantum-scale interactions between elements, thereby breaking through the explicit rule limitations of element combination analysis in traditional empirical methods. A data-driven element interaction descriptor is constructed, providing feature input for generative adversarial networks with both physical meaning and generalization capabilities.

[0023] Alternatively, W1, W2, W3, W4, and W5 can be represented by Softmax(a×a, a×b, a×c, a×d, a×e). By calculating the product of element a with itself and other elements, the model converts the strength of pairwise interactions between elements into scalar values. This implicitly captures the synergistic or antagonistic effects between chemical elements and quantifies them into interpretable numerical form. The application of the Softmax function maps the raw product results into a normalized probability distribution, imbuing weights W1 through W5 with competitive attention allocation properties.

[0024] Alternatively, the softmax function is calculated as follows:

[0025] ;

[0026] Among them, i is the attention that needs to be calculated, j is the influencing factor related to i, the denominator is all the attention, n is the total number of attentions that need to be calculated, and e is the base of the natural logarithm.

[0027] It not only solves the balance problem of multi-element weight distribution, but also enhances the sensitivity to rare combinations in the long-tail distribution through the gradient characteristics of the exponential function, and finally forms a weight distribution with physical interpretability. The size of the weight directly reflects the intensity of the influence of the interaction between elements on the material properties, providing a feature encoding basis for the generative adversarial network that conforms to the laws of materials science.

[0028] Optionally, in step S2, the generator and discriminator are respectively selected from one or more of artificial neural networks, convolutional neural networks, recurrent neural networks or long short-term memory. If the data is structured, such as element content vectors, artificial neural networks can effectively model the complex nonlinear relationships between these features and are suitable for generating and discriminating these vector data. If data with spatial structure is processed, convolutional neural networks can capture local patterns, such as lattice arrangements, thereby improving the authenticity of the generated structure. If material design needs to consider the sequential nature of the synthesis process, or the changes in material properties over time, recurrent neural networks or long short-term memory can handle this sequential dependency and generate data with temporal consistency. By fusing the inherent advantages of different networks, multi-level feature fusion from atomic-level composition to macroscopic performance can be achieved, enabling generative adversarial networks to adapt to the diverse data forms and physical constraints in metal material design, thereby breaking through the limitations of a single architecture.

[0029] Optionally, the discriminator comprises a fully connected layer, with a classifier at the end. This design, with fully connected layers and a final classifier, offers a key advantage in distinguishing material composition. The dense connectivity of the fully connected layer globally integrates all dimensional information from the input features, ensuring that elemental interactions in the chemical composition data are fully modeled, avoiding the potential loss of contextual information that might result from extracting local features.

[0030] Alternatively, in step S3, the final objective function of the model is expressed by the following equation:

[0031] ;

[0032] Among them, Ex~P(data)log(D(x)) is the expectation of the distribution of the true data x, D(x) is the discriminator model, G(z) is the generator model, P(data) is the actual data distribution, Ez~P(z)(log(1-D(G(z)))) is the expectation of the generator distribution, z is the noise sampled from the prior distribution P(z), and G(z) is the data generated by the generator.

[0033] The discriminator improves its recognition ability of real material samples by maximizing the log-likelihood, and exposes the defects of generated data by minimizing the discrimination probability of generated samples; while the generator minimizes the discriminator's recognition accuracy of its generated samples, forcing itself to output synthetic data that is infinitely close to the real material distribution.

[0034] Alternatively, in step S4, the specific process of model training is as follows:

[0035] S41, extract m training samples from the training data, with the label as true;

[0036] S42, randomly extract m training samples from any specific distribution of the material, and generate m training samples with false labels through the generator;

[0037] S43, fix the generator weight, calculate the error through the objective function, and update the discriminator weight using the backpropagation algorithm by maximizing the error;

[0038] S44, repeat steps S41 and S42, fix the discriminator weight, calculate the error through the objective function, and use the back propagation algorithm to update the generator weight;

[0039] S45, repeat steps S41 to S44 until the discriminator loss is almost unchanged. It is considered that the generator and discriminator have reached Nash equilibrium, the model has converged, the training is completed, and the training is stopped.

[0040] This approach ensures statistical stability while also avoiding local optima by covering different regions of the material composition space through random sampling. When updating the discriminator with a fixed generator, the discriminator is forced to learn the essential characteristics of materials science by maximizing the error between real and generated samples. When updating the generator with a fixed discriminator, the error backpropagation through the generator reversely decodes the physical laws extracted by the discriminator into a composition generation strategy.

[0041] After model training is complete, a generator model and a discriminator model are generated. When designing materials, the generator model is used to randomly generate n sets of data from a distribution, and chemical composition data is then generated using the generator model. The generator implicitly learns the fundamental laws of materials science during adversarial training, and its output naturally avoids thermodynamically unstable or inefficient combinations, significantly improving the design success rate. Randomly sampling from the latent space to generate a diverse range of candidate components not only covers reasonable combinations of known material systems but also explores new compositional spaces that are difficult to access using traditional empirical methods.

[0042] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0043] 1. The present invention provides a metal material design method based on deep learning. This method learns the chemical composition of a specific material. Although the chemical composition of the material is extremely complex and has a diverse distribution, it only needs to sample a random value from any specific distribution and then use it through a generator to generate the chemical composition of a new material. The material design process is easy to operate, low-cost, and easy to use.

[0044] 2. The metal material design method based on deep learning provided by the present invention has significant advantages over the existing technology, especially in avoiding trial-and-error design, reducing development costs, and addressing the lack and monopoly of domestic material basic databases. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The present invention will now be described by way of example with reference to the accompanying drawings, in which:

[0046] Figure 1 The model architecture and data flow diagram of the metal material design method based on deep learning described in the present invention;

[0047] Figure 2 This is a schematic diagram of the attention feature representation process using carbon as an example. DETAILED DESCRIPTION

[0048] The present invention will be described in detail below with reference to the accompanying drawings.

[0049] All features disclosed in this specification, or all steps in the disclosed methods or processes, except mutually exclusive features and / or steps, can be combined in any manner.

[0050] Any feature disclosed in this specification, unless otherwise stated, may be replaced by other equivalent or similar features. That is, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.

[0051] As some examples of the present invention, the titanium alloy materials described in the present invention include: α titanium alloy, α-β titanium alloy and β titanium alloy, classified by alloy phase composition.

[0052] As some examples of the present invention, classified by material strength, the titanium alloy materials described in the present invention include: low-strength (strength ≤ 500MPa), medium-strength (strength 500 to 900MPa), medium-high strength (strength 900 to 1000MPa) and high-strength (strength>1000MPa) titanium alloys, etc., but are not limited to the titanium alloy material development and chemical composition design specified above.

[0053] The metal material design method based on deep learning of the present invention comprises the following steps:

[0054] S1: Collect a large amount of real chemical composition data of relevant materials in the material direction to be designed to form training samples. Preprocess the training samples to reduce data scale differences, and use data augmentation technology to increase the number of training samples to meet the needs of deep learning. Then, use the self-attention mechanism to improve the feature expression ability.

[0055] S2, design the model architecture, such as Figure 1 As shown in the figure, the model architecture includes two basic elements: the generator and the discriminator, where the generator and the discriminator are neural networks of any type;

[0056] S3, in the model architecture designed in step S2, creating a generator, generating data similar to the real chemical composition through the generator, and using the objective function to measure the similarity between the data generated by the generator and the actual data;

[0057] S4, during the training period, the discriminator wants to maximize the entire output, while the generator wants to minimize the entire output, thereby training the adversarial generative network to balance the generator and the discriminator. When they reach the Nash equilibrium, the model has converged and the training is completed.

[0058] Furthermore, the step S1 includes:

[0059] S11, data collection, collect a large amount of relevant materials in the material direction to be designed. For example, when designing stainless steel materials, you can use various manuals and tools to collect as much real chemical composition data of stainless steel materials as possible and form training samples;

[0060] S12, data preprocessing, performing data preprocessing on the chemical composition in the training samples through normalization, standardization, minimum-maximum normalization, zero-mean normalization, or decimal scaling normalization techniques, but not limited to these techniques, to reduce the scale differences of the chemical composition content, to avoid gradient instability caused by the large content of some elements, resulting in training failure;

[0061] S13, data augmentation. Since deep learning is an over-parameterized machine learning model, it requires a large amount of training data as a basis. However, in the existing material data, the number of data types with known chemical composition is relatively small. To meet the training needs, the positions of the chemical elements in at least one training sample can be rearranged to generate new training samples. This data augmentation method not only increases the training data capacity but also prevents the neural network from memorizing the position information of the elements.

[0062] S14, the self-attention feature indicates that the chemical elements of the material can be coupled to form some structures such as solid solutions or intermetallic compounds. These specific structures are dispersed on the matrix and can improve the performance of certain aspects of the material, but it is difficult for the neural network to notice these features. Therefore, in the present invention, the self-attention mechanism is used to improve the expressive ability of the features and increase the training sample capacity.

[0063] As some examples of the present invention, in step S13, taking a certain α titanium alloy composition as an example, when the positions of elements are arbitrarily exchanged, the material properties do not change, but multiple new training samples can be generated, thereby achieving data enhancement and significantly increasing the capacity of the training data.

[0064] On this basis, when these enhanced data samples are used to train deep learning models, due to the increase in data volume and the removal of element position information, the deep learning model can better learn the relationship between the material properties and chemical composition, rather than simply remembering the element position information.

[0065] Specifically, as some examples of the present invention, in step S14, the self-attention feature representation process of carbon is illustrated by taking the carbon element as an example:

[0066] Taking carbon as an example, Figure 2 As shown, first multiply the carbon content by the content of other elements to obtain formula (1),

[0067] W1,W2,W3,W4,W5=Softmax(a×a,a×b,a×c,a×d,a×e)(1);

[0068] Among them, W1, W2, W3, W4, and W5 are attention weights respectively, and a, b, c, d, and e are the contents of carbon, silicon, manganese, sulfur, and phosphorus respectively.

[0069] The softmax function is calculated as follows:

[0070] (2);

[0071] Among them, i is the attention that needs to be calculated, j is the influencing factor related to i, the denominator is all the attention, n is the total number of attentions that need to be calculated, Ez is the expectation of the distribution of P(z) (log(1-D(G(z))), and Ex is the expectation of the distribution of P(data)log(D(x)).

[0072] On this basis, the following formula (3) is obtained:

[0073] Self-attention feature representation of carbon = W1×a+W2×b+W3×c+W4×d+W5×e (3);

[0074] Then, the self-attention feature representations of the remaining elements are calculated according to the above calculation method. This self-attention feature representation couples the relationship between other elements and a specified element and has stronger expressive power.

[0075] As some examples of the present invention, in step S14, different network methods can be used according to different specific material design methods, such as: fully connected, convolutional, cyclic, attention and other networks, to achieve feature extraction.

[0076] As some examples of the present invention, in step S2, the generator and discriminator can be one or more of an artificial neural network (ANN), a convolutional neural network (CNN), a recurrent neural network (RNN) or a long short-term memory (LSTM), and the generator and the discriminator do not require the same network architecture.

[0077] Furthermore, the discriminator must have a fully connected layer, with the classifier at the end. In a neural network, a fully connected layer means that every neuron in the layer is connected to all neurons in the previous layer. This connection allows the network to integrate all information from the previous layer to make decisions or predictions. The classifier is the part of the neural network used to classify input data. In the discriminator, the classifier is usually located at the end of the network, used to determine whether the input data comes from a real dataset or is generated by a generator.

[0078] Specifically, in steps S3 and S4, the generator is a neural network whose task is to sample a vector from a random distribution (such as a Gaussian distribution or a uniform distribution) and then map this vector to a new data point similar to the real data, thereby generating data with a chemical composition similar to the real data. The goal of the generator is to generate data with a chemical composition as close as possible to the real data. This means that the generator needs to learn the distribution of the real data in order to generate new samples similar to it. The objective function is a function used to measure model performance and defines the goal to be optimized. In a generative adversarial network, the generator and discriminator each have their own objective function. For the generator, the objective function is typically designed to measure the difference or distance between the generated data and the real data. This difference can be measured in various ways, such as mean squared error (MSE) or cross-entropy loss. During training, both the generator and the discriminator attempt to minimize their respective objective functions. For the generator, this means adjusting its parameters to minimize the difference between the generated data and the real data. Simultaneously, the discriminator also attempts to minimize its objective function, namely, to determine as accurately as possible whether the input data is real or generated. Therefore, in a generative adversarial network, the generator and the discriminator are trained in an adversarial manner. The generator attempts to generate increasingly realistic data to deceive the discriminator, while the discriminator attempts to identify the generated data with increasing accuracy. This adversarial process drives both networks to continuously improve until they reach a balance point, known as the Nash equilibrium, at which point the discriminator can no longer distinguish between real data and generated data, and the generator also generates samples that are very similar to real data.

[0079] Preferably, in step S3, the final objective function of the model is expressed by equation (4) as follows:

[0080] (4);

[0081] Among them, Ex~P(data)log(D(x)) is the expectation of the distribution of the true data x, D(x) is the discriminator model, G(z) is the generator model, P(data) is the actual data distribution, Ez~P(z)(log(1-D(G(z)))) is the expectation of the generator distribution, z is the noise sampled from the prior distribution P(z), and G(z) is the data generated by the generator.

[0082] Furthermore, in step S4, the specific process of model training is as follows:

[0083] S41, extract m training samples from the training data, with the label as true (i.e., the label is 1);

[0084] S42, randomly select m training samples from any specific distribution of the material, and generate m training samples with false labels (i.e., labels of 0) through the generator;

[0085] S43, fix the generator weight, calculate the error through the objective function, and update the discriminator weight using the backpropagation algorithm by maximizing the error;

[0086] S44, repeat steps S41 and S42, fix the discriminator weight, calculate the error through the objective function, and use the back propagation algorithm to update the generator weight;

[0087] S45, repeat steps S41 to S44 until the discriminator loss is almost unchanged. It is considered that the generator and discriminator have reached Nash equilibrium, the model has converged, the training is completed, and the training is stopped.

[0088] After the training is completed, we will obtain a generator model and a discriminator model. When designing materials, we only need to use the generator model to randomly generate n groups of data from a certain distribution, and then generate chemical composition data through the generator model.

[0089] In the present invention, the impact of excessive content of a certain chemical element in the material on training is first reduced through data preprocessing, and then the number of samples is increased through data enhancement, and the neural network is prevented from memorizing position information. On this basis, the relationship between other elements and one of the elements is coupled through self-attention feature representation, making the feature have stronger expressive power.

[0090] Afterwards, by constructing a generator and a discriminator, a new set of chemical compositions is generated, and the label corresponding to the new chemical composition is 0. The generated chemical compositions and the chemical compositions of the collected materials are combined into a batch of training data, where the real material label is 1. The objective function is generated through the discriminant network, and then the generator and discriminator are optimized through the objective function until the Nash equilibrium is reached.

[0091] Finally, random sampling is performed on any specific distribution of materials to generate an excitation signal. The generator then generates the material's specific chemical composition data. The discriminator then distinguishes the generated chemical composition from the real data. Through repeated iterations, the discriminator and generator reach a Nash equilibrium, and training concludes.

[0092] In the subsequent stage of designing new materials, the specific process is to randomly sample from a distribution and generate the chemical composition of new titanium alloys through the generator, thereby achieving the purpose of titanium alloy design based on the adversarial generation method.

[0093] Of course, the metal material design method based on deep learning described in the present invention can also be used for the design of other alloy materials.

[0094] The beneficial technical effect of the deep learning-based metal material design method described in the present invention is: this method learns the chemical composition of a certain specific material. Although the chemical composition of the material is extremely complex and has diverse distributions, such as Gaussian distribution, uniform distribution, exponential distribution, etc., it is not limited to these distributions. However, it only needs to sample a random value from any of these specific distributions, and then use a generator to generate the chemical composition of a new material. The material design process is easy to operate, low-cost, and easy to use.

[0095] In summary, the deep learning-based metal material design method described in the present invention has significant advantages over the existing technology, especially in avoiding trial-and-error design, reducing development costs, and addressing the lack and monopoly of domestic material basic databases.

[0096] First, it avoids the trial-and-error approach to material design, reducing the cost of developing new materials: Traditional material design methods often rely on a large number of experimental trial-and-error experiments, which is not only time-consuming and labor-intensive, but also costly. With the continuous development of materials science, the types and properties of materials have become increasingly complex, and the limitations of the trial-and-error method have become increasingly prominent. The deep learning-based metal material design method described in the present invention predicts the properties of materials by training models, thereby avoiding the trial-and-error process. This method not only greatly improves design efficiency, but also significantly reduces the cost of new material development.

[0097] Second, it avoids the lack of a domestic materials database and the monopoly of other design methods on this database. Currently, the development of domestic materials databases is relatively lagging, and the data integrity and accuracy need to be improved. This has, to a certain extent, limited the application of traditional design methods. Although deep learning models require large amounts of high-quality data, they are not entirely dependent on a specific database. Instead, deep learning models can acquire data from multiple sources and improve their generalization capabilities through techniques such as data augmentation and transfer learning. Therefore, they can effectively avoid the monopoly of other design methods on materials databases. This enables more researchers and companies to participate in materials design research, promoting technological innovation and industrial development. Currently, with the continuous development of deep learning technology, more and more researchers and companies are beginning to recognize the importance of data sharing and openness. By establishing a shared data platform and standard data formats, data exchange and collaboration between different research teams and companies can be promoted, further advancing the development of materials science.

[0098] The embodiments of the present application are described above in conjunction with the accompanying drawings. Unless there is a conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of this application, ordinary technicians in this field can also make many forms without departing from the purpose of this application and the scope of protection of the claims, all of which are within the protection of this application.

Claims

1. A metal material design method based on deep learning, characterized in that: Including steps: S1 collects real chemical composition data, forms training samples, preprocesses the training samples, increases the number of training samples through data augmentation technology, and then improves the feature expression ability through the self-attention mechanism. The steps include: S11, data collection, collect a large amount of real chemical composition data of relevant materials in the material direction to be designed and form training samples; S12, data preprocessing, data preprocessing of the chemical components in the training samples to reduce the scale differences of the chemical component content; S13, data enhancement, training at least one of the samples to rearrange the positions of the chemical elements therein to generate new training samples to achieve data enhancement; S14, self-attention feature representation, improves the feature expression ability and training sample capacity through the self-attention mechanism; The self-attention feature of the selected element is represented as: W1×a+W2×b+W3×c+W4×d+W5×e; W1,W2,W3,W4,W5=Softmax(a×a,a×b,a×c,a×d,a×e); Among them, W1, W2, W3, W4, and W5 are the weights of the corresponding elements, which are randomly initialized and are trainable parameters; a is the content of the selected element; b, c, d, and e are the contents of the remaining elements in the training sample respectively; S2, design the model architecture, which includes two basic elements: the generator and the discriminator; S3, in the model architecture, generates data similar to the real chemical composition through the generator, and uses the objective function to measure the similarity between the generated data and the actual data; S4, train the adversarial generative network to achieve a balance between the generator and the discriminator.

2. The metal material design method based on deep learning according to claim 1, characterized in that: The softmax function is calculated as follows: ； Among them, i is the attention that needs to be calculated, j is the influencing factor related to i, the denominator is all the attention, n is the total number of attentions that need to be calculated, and e is the base of the natural logarithm.

3. The metal material design method based on deep learning according to claim 1, characterized in that: In step S2, the generator and the discriminator are respectively selected from one or more of an artificial neural network, a convolutional neural network, a recurrent neural network or a long short-term memory.

4. The metal material design method based on deep learning according to claim 1 or 3, characterized in that: The discriminator has fully connected layers and the last layer of the discriminator is the classifier.

5. The metal material design method based on deep learning according to claim 1, characterized in that: In step S3, the final objective function of the model is expressed by the following equation: ； Among them, Ex~P(data)log(D(x)) is the expectation of the distribution of the true data x, D(x) is the discriminator model, G(z) is the generator model, P(data) is the actual data distribution, Ez~P(z)(log(1-D(G(z)))) is the expectation of the generator distribution, z is the noise sampled from the prior distribution P(z), and G(z) is the data generated by the generator.

6. The metal material design method based on deep learning according to claim 1, characterized in that: In step S4, the specific process of model training is as follows: S41, extract m training samples from the training data, with the label as true; S42, randomly extract m training samples from any specific distribution of the material, and generate m training samples with false labels through the generator; S43, fix the generator weight, calculate the error through the objective function, and update the discriminator weight using the backpropagation algorithm by maximizing the error; S44, repeat steps S41 and S42, fix the discriminator weight, calculate the error through the objective function, and use the back propagation algorithm to update the generator weight; S45, repeat steps S41 to S44 until the discriminator loss is almost unchanged. It is considered that the generator and discriminator have reached Nash equilibrium, the model has converged, the training is completed, and the training is stopped.

7. The metal material design method based on deep learning according to claim 1, characterized in that: After the model training is completed, a generator model and a discriminator model will be obtained; when designing materials, the generator model is used to randomly generate n groups of data from a certain distribution, and the chemical composition data is generated through the generator model.

Images