Metal material design method based on deep learning

Abstract

The invention discloses a metal material design method based on deep learning, and the method comprises the steps: S1, collecting real chemical component data, forming training samples, carrying out the data preprocessing of the training samples, increasing the number of the training samples through a data enhancement technology, and improving the expression capability of features through a self-attention mechanism; s2, designing a model architecture, wherein the model architecture comprises two basic elements, namely a generator and a discriminator; s3, in the model architecture, generating data similar to real chemical components through a generator, and measuring the similarity between the generated data and actual data by using an objective function; and S4, training the confrontation generation network to balance the generator and the discriminator. The metal material design method based on deep learning has the advantages of convenient operation in the material design process, low cost and convenient use.

Description

Technical Field

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

[0002] Materials are one of the most widely used means of production by humans. The properties of materials are determined by the microstructure of the materials, and the microstructure of the materials is in turn determined by the chemical composition of the materials. There are many types of chemical elements. A slight change in the element content will change the properties of the materials. Due to the different electronic structures and content differences of the chemical elements in the material composition, various complex compounds can be formed. The distribution patterns of these compounds in the materials further affect the properties of the materials.

[0003] At present, the design of new materials still remains in the stage of empirical science. New materials are designed based on the experience of predecessors in using materials. This design method requires a large amount of time, R & D costs, and cannot guarantee the success of the final material design.

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

[0005] The object of the present invention is to provide a convenient, low-cost and easy-to-use method for designing metal materials based on deep learning for the above-mentioned existing problems.

[0006] The technical solution adopted by the present invention is as follows. A method for designing metal materials based on deep learning includes the following steps: S1, collect real chemical composition data, form training samples, perform data preprocessing on the training samples, increase the number of training samples through data augmentation technology, and then improve the expression ability of features through the self-attention mechanism; S2, design the model architecture, which includes two basic elements: a generator and a discriminator; S3, in the model architecture, generate data similar to the real chemical composition through the generator, and use the objective function to measure the similarity between the generated data and the actual data; S4, train the generative adversarial network to balance the generator and the discriminator.

[0007] Step S1 constructs a training sample library by collecting real chemical composition data and uses data augmentation techniques to address the scarcity of material data. The self-attention mechanism is introduced to model the non-linear relationships between elements, such as capturing the synergistic effects of trace alloying elements on the crystal structure. The discrete chemical element contents are mapped into a high-dimensional feature space, providing an information basis for subsequent generation. In step S2, the 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 actual data, enabling the generated data to not only be similar to the real data distribution but also cover special material combinations in the long-tail distribution. In step S4, the generator is first pre-trained to learn the basic distribution of material compositions, and then adversarial training is initiated. The final model can generate a composition design scheme with a gradient structure, breaking through the single-point optimization limitation of the traditional trial-and-error method. At the micro level, the self-attention mechanism analyzes the quantum-scale interactions between elements in this method; at the mesoscopic level, the adversarial mechanism of GANs simulates the competitive relationship of compound formation energy; at the macroscopic level, the generator can output a gradient composition distribution that meets the target performance.

[0008] Optionally, step S1 includes: S11, data collection, collecting real chemical composition data of a large number of relevant materials in the material direction to be designed and forming a training sample; S12, data preprocessing, preprocessing the chemical compositions in the training sample to reduce the scale difference of chemical composition contents; S13, data augmentation, training at least one of the samples to rearrange the positions of the chemical elements therein to generate new training samples and achieve data augmentation; S14, self-attention feature representation, enhancing the expression ability of features and the training sample capacity through the self-attention mechanism.

[0009] S11 constructs an initial sample library covering multi-element combinations by widely collecting real chemical composition data in the target material field, solves the data island problem in material R & D, and provides sufficient learning materials for deep learning. S12 eliminates the dimensional difference of different chemical element contents, avoids the model from over-focusing on specific elements due to numerical magnitude deviation, and ensures the comparability of each component feature in training. S13 generates equivalent variant samples within the allowable range of physical laws by reorganizing the chemical element arrangement of the original samples, breaks through the limitation of scarce experimental data, and enhances the model's understanding ability of the diversity of component combinations. S14 uses the self-attention mechanism to dynamically capture the non-linear correlations between elements, strengthens the representation ability of key component combinations through the adaptive allocation of weights in the feature space, and simultaneously implicitly expands the coverage of the data distribution in the high-dimensional space, improving the model's parsing accuracy for complex material systems.

[0010] Optionally, in step S14, The self-attention feature representation of the selected elements is: W1×a + W2×b + W3×c + W4×d + W5×e; Among them, W1, W2, W3, W4, and W5 are the weights of the corresponding elements in sequence, 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 in sequence.

[0011] In the self-attention feature representation of step S14, the proposed linear combination formula dynamically adjusts the contribution degrees of different chemical elements to the overall material feature representation by introducing trainable weight parameters. During the training process of the model, the implicit correlation patterns among the elements are automatically learned through backpropagation, so that the final feature representation can quantitatively reflect the synergistic effect or competitive relationship among the elements. The introduction of the weight parameters endows the model with the ability to reconstruct the feature space - even if the input is the scalar values of the original element contents (a to e), through the linear transformation of the weight matrix and the combined action of the subsequent non-linear activation function, the low-dimensional chemical composition data can be mapped to a high-dimensional space, explicitly encoding the interactions at the quantum scale among the elements, thus breaking through the explicit rule limitations of element combination analysis in traditional empirical methods. A data-driven element interaction descriptor is constructed, providing a feature input with both physical meaning and generalization ability for the generative adversarial network.

[0012] Optionally, W1, W2, W3, W4, W5 = Softmax(a×a, a×b, a×c, a×d, a×e). By calculating the products of element a with itself and other elements, the model converts the pairwise interaction strengths among the elements into scalar values. Implicitly captures the synergistic or antagonistic effects among the chemical elements and quantifies them into an interpretable numerical form. The application of the Softmax function maps the original product results to a normalized probability distribution, making the weights W1 to W5 have the characteristics of competitive attention allocation.

[0013] Optionally, the softmax function is calculated as shown in the following formula: ; where i is the attention to be calculated, j is the influencing factor related to i, the denominator is all the attentions, n is the total number of attentions to be calculated, and e is the base of the natural logarithm.

[0014] It not only solves the balance problem of multi-element weight allocation, but also enhances the sensitivity to rare combinations in the long-tailed distribution through the gradient characteristics of the exponential function, and finally forms a physically interpretable weight distribution. The size of the weight directly reflects the influence intensity of the interaction among the elements on the material properties, providing a feature coding basis that conforms to the laws of materials science for the generative adversarial network.

[0015] Alternatively, in the step S2, the generator and the discriminator are respectively selected from one or more of artificial neural network, convolutional neural network, recurrent neural network or long short-term memory. If the data is structured, such as the elemental content vector, the artificial neural network can effectively model the complex non-linear relationships between these features and is suitable for generating and discriminating such vector data. If dealing with data having a spatial structure, the convolutional neural network can capture local patterns, such as lattice arrangements, thereby improving the authenticity of the generated structure. If the material design needs to consider the sequentiality of the synthesis process or the change of material properties over time, the recurrent neural network or long short-term memory can handle this sequential dependence and generate data with temporal consistency. By integrating the inherent advantages of different networks, multi-level feature fusion from atomic-level composition to macroscopic properties can be achieved, enabling the generative adversarial network to adapt to diverse data forms and physical constraints in metal material design, thereby breaking through the limitations of a single architecture.

[0016] Alternatively, the discriminator has a fully connected layer, and the discriminator is finally a classifier. The design of adopting a fully connected layer and finally connecting a classifier provides key advantages for the discrimination process of material composition. The dense connection characteristic of the fully connected layer can globally integrate all dimensional information of the input features, ensuring that the elemental interaction in the chemical composition data is completely modeled and avoiding the loss of associated information that may be caused by local feature extraction.

[0017] Alternatively, in the step S3, the following equation is used to represent the final objective function of the model: ; where, Ex~P(data)log(D(x)) is the expectation of the true data x distribution, 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.

[0018] The discriminator improves the recognition ability of real material samples by maximizing the log-likelihood, and at the same time exposes the defects of the generated data by minimizing the discrimination probability of the generated samples; while the generator forces itself to output synthetic data that is infinitely close to the real material distribution by minimizing the recognition accuracy of the discriminator for the samples it generates.

[0019] Alternatively, in the step S4, the specific process of model training is as follows: S41, Take out m training samples from the training data, and the label is true; S42. Randomly select m training samples from any specific distribution of materials, and through the generator, generate m training samples with false labels; S43. Fix the weights of the generator, calculate the error through the objective function, and update the weights of the discriminator using the backpropagation algorithm by maximizing the error; S44. Repeat steps S41 and S42, fix the weights of the discriminator, calculate the error through the objective function, and update the weights of the generator using the backpropagation algorithm; S45. Repeat steps S41 - S44 until the discriminator loss is almost unchanged, then consider that the generator and the discriminator reach the Nash equilibrium, the model has converged, the training is completed, and the training stops.

[0020] It not only ensures statistical stability but also covers different regions of the material composition space through random sampling, avoiding the model falling into local optima. When fixing the generator to update the discriminator, by maximizing the discrimination error between real / generated samples, the discriminator is forced to learn the essential features in materials science; while when fixing the discriminator to update the generator, through the error backpropagation of the generator, the physical laws extracted by the discriminator are decoded reversely into the composition generation strategy.

[0021] Optionally, after the model training is completed, a generator model and a discriminator model will be obtained; when designing materials, use the generator model to randomly generate n groups of data from a certain distribution and generate chemical composition data through the generator model. The generator has implicitly learned the basic laws of materials science during the adversarial training, and its output results naturally avoid thermodynamically unstable or performance-inefficient combinations, significantly improving the design success rate. Random sampling from the latent space to generate diverse candidate compositions can not only cover the reasonable combinations of known material systems but also explore new composition spaces that are difficult to reach by traditional empirical methods.

[0022] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present invention are as follows: 1. A metal material design method based on deep learning provided by the present invention. By learning the chemical composition of a specific material, although the chemical composition of the material is extremely complex and diverse in distribution, only by sampling a random value from these arbitrary specific distributions and then through the generator, the chemical composition of a new material can be generated. The material design process is convenient to operate, low in cost, and easy to use.

[0023] 2. A metal material design method based on deep learning provided by the present invention has significant advantages compared with the prior art, especially in avoiding trial-and-error design, reducing development costs, and dealing with the problems of lack and monopoly of domestic material basic databases. Description of the Drawings

[0024] The present invention will be described by way of examples with reference to the accompanying drawings, wherein: Figure 1 is the model architecture and data flow diagram of the deep learning-based metal material design method described in the present invention; Figure 2 is a schematic diagram of the attention feature representation process taking carbon element as an example. Detailed implementation manners

[0025] The present invention will be described in detail below with reference to the accompanying drawings.

[0026] All features disclosed in this specification, or all steps in the disclosed methods or processes, except for mutually exclusive features and / or steps, can be combined in any manner.

[0027] Any feature disclosed in this specification, unless specifically stated, can be replaced by other equivalent or similar-purpose alternative features. That is, unless specifically stated, each feature is only an example in a series of equivalent or similar features.

[0028] As some examples of the present invention, classified by alloy phase composition, the titanium alloy materials described in the present invention include: α-titanium alloy, α-β titanium alloy, and β-titanium alloy.

[0029] As some examples of the present invention, classified by material strength, the titanium alloy materials described in the present invention include: low-strength (strength ≤ 500 MPa), medium-strength (strength 500 to 900 MPa), medium-high strength (strength 900 to 1000 MPa), and high-strength (strength > 1000 MPa) titanium alloys, etc., but not limited to the development of titanium alloy materials and chemical composition design with the above specific requirements.

[0030] The deep learning-based metal material design method described in the present invention includes the following steps: S1. Collect the true chemical composition data of a large number of relevant materials in the material direction to be designed, form training samples, perform data preprocessing on the training samples to reduce data scale differences, and increase the number of training samples through data augmentation technology to meet the requirements of deep learning. Then, improve the expression ability of features through the self-attention mechanism; S2. Design the model architecture, as Figure 1 shown: The model architecture includes two basic elements, a generator and a discriminator, wherein the generator and the discriminator are neural networks of any type; S3. In the model architecture designed in step S2, create a generator, generate data similar to the true chemical composition through this generator, and use the objective function to measure the similarity between the data generated by the generator and the actual data; S4. During the training period, the discriminator attempts to maximize the overall output, while the generator aims to minimize the overall output. Thus, the adversarial generative network is trained to balance the generator and the discriminator. When they reach the Nash equilibrium, the model has converged and the training is completed.

[0031] Further, the step S1 includes: S11. Data collection: Collect a large number of relevant materials in the material direction to be designed. For example, when designing stainless steel materials, various manuals and tools can be used to collect as much real chemical composition data of the same stainless steel material as possible and form training samples. S12. Data preprocessing: Use techniques such as normalization, standardization, minimum - maximum normalization, zero - mean normalization, or decimal scaling normalization, but not limited to these techniques, to preprocess the chemical components in the training samples to reduce the scale difference of chemical component contents and avoid gradient instability caused by some large element contents, resulting in training failure. S13. Data augmentation: Since deep learning is an over - parameterized machine learning model and requires a large amount of training data as a basis, but in existing material data, the number of known chemical composition data types is small. To meet the training needs, the positions of 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 volume but also avoids the neural network from memorizing the position information of elements. S14. Self - attention feature representation: Some structures such as solid solutions or intermetallic compounds can be formed through the coupling between material chemical elements. These specific structures are dispersed on the matrix and can improve certain aspects of the material's performance, but it is difficult for the neural network to notice these features. Therefore, in the present invention, through the self - attention mechanism, the expression ability of features is enhanced and the training sample volume can be increased.

[0032] As some examples of the present invention, in the step S13, taking a certain α - titanium alloy composition as an example, when arbitrarily exchanging the positions of elements, the material properties do not change, but multiple new training samples can be generated, thereby realizing data augmentation and significantly increasing the training data volume.

[0033] On this basis, when using these enhanced data samples to train the deep learning model, 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's performance and chemical composition, rather than simply memorizing the position information of elements.

[0034] Specifically, as some examples of the present invention, in the step S14, taking the carbon element as an example, the process of its self - attention feature representation is illustrated as follows: Taking the carbon element as an example, such asFigure 2 As shown, first multiply the contents of carbon and other elements to obtain Equation (1). W1, W2, W3, W4, W5 = Softmax(a×a, a×b, a×c, a×d, a×e) (1); where W1, W2, W3, W4, W5 are attention weights in sequence, and a, b, c, d, e are the contents of carbon element, silicon element, manganese element, sulfur element and phosphorus element in sequence.

[0035] The softmax function is calculated as shown in the following Equation (2): (2); where i is the attention to be calculated, j is the influencing factor related to i, the denominator is all attentions, n is the total number of attentions to be calculated, Ez is the expectation of the P(z)(log(1 - D(G(z)))) distribution, and Ex is the expectation of the P(data)log(D(x)) distribution.

[0036] On this basis, the following Equation (3) is obtained: Self-attention feature representation of carbon = W1×a + W2×b + W3×c + W4×d + W5×e (3); Afterwards, calculate the self-attention feature representations of the remaining elements respectively according to the above calculation method. This kind of self-attention feature representation couples the relationship between other elements and a specified element and has stronger expressive ability.

[0037] As some examples of the present invention, in the step S14, according to the specific material design method, different network methods can be adopted, such as: fully connected, convolutional, recurrent, attention and other networks to realize feature extraction.

[0038] As some examples of the present invention, in the step S2, the generator and the discriminator can be one or more of artificial neural network (ANN), convolutional neural network (CNN), recurrent neural network (RNN) or long short-term memory (LSTM). The generator and the discriminator do not require the same network architecture.

[0039] Furthermore, the discriminator must have a fully connected layer, and the last part of the discriminator is a classifier. In a neural network, a fully connected layer means that each neuron in this layer is connected to all neurons in the previous layer. This connection method enables the network to comprehensively use all information in the previous layer for decision-making or prediction. And the classifier is the part in the neural network used to classify the input data. In the discriminator, the classifier is usually located at the end of the network and is used to judge whether the input data comes from the real data set or is generated by the generator.

[0040] Specifically, in the steps S3 and S4, the generator is a neural network. Its task is to sample a vector from a certain random distribution (such as Gaussian distribution, uniform distribution, etc.), and then map this vector to a new data point similar to the real data, generating data similar to the real chemical composition. The goal of the generator is to generate data as close as possible to the real chemical composition, which means 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 the performance of the model, which defines the goal that the model needs to optimize. In the generative adversarial network, the generator and the discriminator each have their own objective functions. For the generator, the objective function is usually 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 square error, cross - entropy loss, etc. During the training process, both the generator and the discriminator will try to minimize their respective objective functions. For the generator, this means it needs to adjust its parameters so that the difference between the generated data and the real data is minimized. At the same time, the discriminator will also try to minimize its objective function, that is, to accurately judge whether the input data is real or generated. Therefore, in the generative adversarial network, through the adversarial training of the generator and the discriminator, the generator tries to generate more and more realistic data to deceive the discriminator, while the discriminator tries to more accurately identify the generated data. This adversarial process will drive the two networks to continuously improve until a balance point is reached, that is, the Nash equilibrium. At this time, the discriminator can no longer distinguish between real data and generated data, and the generator has also generated samples very similar to the real data.

[0041] Preferably, in the step S3, the following equation (4) is used to represent the final objective function of the model: (4); Among them, Ex~P(data)log(D(x)) is the expectation of the real data x distribution, 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.

[0042] Furthermore, in the step S4, the specific process of model training is as follows: S41, Take out m training samples from the training data, with the label being true (i.e., the label is 1); S42, Randomly take out m training samples from any specific distribution of materials, and through the generator, generate m training samples with the label being false (i.e., the label is 0); S43. Fix the generator weights, calculate the error through the objective function, maximize the error, and update the discriminator weights using the backpropagation algorithm; S44. Repeat steps S41 and S42. Fix the discriminator weights, calculate the error through the objective function, and update the generator weights using the backpropagation algorithm; S45. Repeat steps S41 - S44 until the discriminator loss is almost unchanged. Then, it is considered that the generator and the discriminator reach the Nash equilibrium, the model has converged, and the training is completed and stopped.

[0043] 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 generate chemical composition data through the generator model.

[0044] In the present invention, first, the influence of excessive content of a certain chemical element in the material on training is reduced through data preprocessing. Then, the number of samples is increased through data augmentation, and the neural network is prevented from memorizing position information. On this basis, through self - attention feature representation, the relationship between other elements and one of the elements is coupled, making the features have stronger expressive ability.

[0045] After that, 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. Among them, the real material label is 1. The discriminator network is used to generate the objective function, and then the generator and the discriminator are optimized through the objective function until the Nash equilibrium is reached.

[0046] Finally, random sampling is performed on any specific distribution of the material to generate an excitation signal, and the specific chemical composition data of the material is generated through the generator. The generated chemical compositions are distinguished between real data and data generated by the generator by the discriminator. Through repeated iteration, the discriminator and the generator reach the Nash equilibrium, and the training ends.

[0047] In the subsequent stage of designing new materials, the specific process is to randomly sample on a distribution and generate the chemical composition of a new titanium alloy through the generator, thereby achieving the purpose of titanium alloy design based on the adversarial generation method.

[0048] 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.

[0049] The beneficial technical effects of the metal material design method based on deep learning described in the present invention are as follows: By learning the chemical composition of a specific material, although the chemical composition of the material is extremely complex and diverse in distribution, such as Gaussian distribution, uniform distribution, exponential distribution, etc., and is not limited to these distributions, only a random value needs to be sampled from these arbitrary specific distributions, and then through a generator, the chemical composition of a new material can be generated. The material design process is convenient to operate, low in cost, and easy to use.

[0050] In summary, compared with the prior art, the metal material design method based on deep learning described in the present invention has significant advantages, especially in avoiding trial-and-error design, reducing development costs, and addressing the lack and monopoly of domestic material basic databases.

[0051] First, it avoids designing materials in a trial-and-error manner and has a low cost for developing new materials: Traditional material design methods often rely on a large number of experimental trials and errors. This method is not only time-consuming and laborious but also costly. With the continuous development of material 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 metal material design method based on deep learning described in the present invention predicts the properties of materials by training a model, thus avoiding the trial-and-error process. This method not only greatly improves the design efficiency but also significantly reduces the cost of developing new materials.

[0052] Second, it avoids the lack of domestic material basic databases and the monopoly of other design methods on material basic databases: At present, the construction of domestic material basic databases lags behind, and the integrity and accuracy of data need to be improved, which to a certain extent restricts the application of traditional design methods. Although deep learning models have a high demand for a large amount of high-quality data, they do not completely rely on a specific database. Instead, deep learning models can obtain data from multiple sources and improve the generalization ability of the models through techniques such as data augmentation and transfer learning. Therefore, it can effectively avoid the monopoly of other design methods on material basic databases. This enables more researchers and enterprises to participate in the research of material design, promoting technological innovation and industrial development. Currently, with the continuous development of deep learning technology, more and more researchers and enterprises have begun to realize the importance of data sharing and openness. By establishing a shared data platform and standard data formats, data exchange and cooperation between different research teams and enterprises can be promoted, further driving the development of material science.

[0053] The embodiments of the present application have been described above in conjunction with the accompanying drawings. Without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The present application is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present application, those of ordinary skill in the art can also make many forms without departing from the purpose of the present application and the scope protected by the claims, and all of them fall within the protection scope of the present application.

Claims

1. A design method of metal materials based on deep learning, characterized in that, Including the steps: S1. Collect real chemical composition data, form training samples, perform data preprocessing on the training samples, and increase the number of training samples through data augmentation techniques. Then, enhance the expression ability of features through the self-attention mechanism; S2. Design the model architecture, which includes two basic elements: a generator and a discriminator; S3. In the model architecture, generate data similar to the real chemical composition through the generator, and use the objective function to measure the similarity between the generated data and the actual data; S4. Train the generative adversarial network to balance the generator and the discriminator.

2. The method for designing metal materials based on deep learning according to claim 1, characterized in that The step S1 includes: S11. Data collection: Collect the real chemical composition data of a large number of relevant materials in the material direction to be designed, and form training samples; S12. Data preprocessing: Perform data preprocessing on the chemical components in the training samples to reduce the scale difference of chemical component contents; S13. Data augmentation: Train at least one sample to rearrange the positions of the chemical elements therein to generate new training samples, realizing data augmentation; S14. Self-attention feature representation: Enhance the expression ability of features and the capacity of training samples through the self-attention mechanism.

3. The method for designing metal materials based on deep learning according to claim 2, wherein In the step S14, The self-attention feature representation of the selected element is: W1×a + W2×b + W3×c + W4×d + W5×e; wherein, W1, W2, W3, W4, and W5 are the weights corresponding to the elements in sequence, which are randomly initialized and are trainable parameters; a, b, c, d, and e are the contents of the remaining elements in the training sample in sequence.

4. The deep learning-based metal material design method according to claim 3, wherein W1, W2, W3, W4, W5 = Softmax(a×a, a×b, a×c, a×d, a×e).

5. The method for designing metal materials based on deep learning according to claim 4, characterized in that The softmax function is calculated as shown in the following formula: ； wherein, i is the attention to be calculated, j is the influencing factor related to i, the denominator is all the attentions, n is the total number of attentions to be calculated, and e is the base of the natural logarithm.

6. The method for designing metal materials based on deep learning according to claim 1, characterized in that In the step S2, the generator and the discriminator are respectively selected from one or more of artificial neural networks, convolutional neural networks, recurrent neural networks, or long short-term memories.

7. The method for designing metal materials based on deep learning according to claim 1 or 6, characterized in that, The discriminator has a fully connected layer, and the last of the discriminator is a classifier.

8. The method for designing metal materials based on deep learning according to claim 1, wherein In the step S3, the following equation is used to represent the final objective function of the model: ； wherein, Ex~P(data)log(D(x)) is the expectation of the real data x distribution, 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.

9. The method for designing metal materials based on deep learning according to claim 1, wherein In the step S4, the specific process of model training is as follows: S41. Take out m training samples from the training data, and the label is true; S42. Randomly take out 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 weights, calculate the error through the objective function, maximize the error, and update the discriminator weights using the backpropagation algorithm; S44. Repeat steps S41 and S42, fix the discriminator weights, calculate the error through the objective function, and update the generator weights using the backpropagation algorithm; S45. Repeat steps S41 - S44 until the discriminator loss is almost unchanged. When the generator and the discriminator reach the Nash equilibrium and the model has converged, the training is completed and the training stops.

10. The method for designing metal materials based on deep learning according to claim 1, wherein After the model training is completed, a generator model and a discriminator model will be obtained; when designing materials, use the generator model to randomly generate n groups of data from a certain distribution and generate chemical composition data through the generator model.

Images