A performance prediction method and system for high-entropy alloys based on adversarial transfer

Abstract

The application discloses a high-entropy alloy performance prediction method and system based on adversarial transfer, and relates to the technical field of high-entropy alloy component process design. In the application, an adversarial transfer network model is trained through image data of a source domain and image data of a target domain, the image data of the source domain and the target domain are learned to sufficiently learn the connection and offset between the image data of the source domain, pre-training of a feature representation network in a high-entropy alloy performance prediction model is realized, and then high-precision training of the high-entropy alloy performance prediction model by using small-sample image data of the target domain is further realized, and the prediction precision of the performance of the high-entropy alloy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of high-entropy alloy composition process design, in particular to a performance prediction method and system for high-entropy alloy based on adversarial transfer. BACKGROUND

[0002] With the extensive research on high-entropy alloys, the excellent performance of this new high-performance metal material, such as high strength, high hardness, wear resistance, corrosion resistance and high-temperature softening resistance, has attracted widespread attention in the scientific research community. However, the design and optimization of high-entropy alloys are still challenging due to their high dimensionality and complex composition space. In this context, data-driven machine learning methods bring new directions to the composition process design of high-entropy alloys. However, different high-entropy alloy datasets are scattered and heterogeneous due to their different compositions, heat treatments and research performances. In addition, current domain adaptation methods often ignore the complex relationships between domain knowledge, relevance and causality between source and target domains, and only rely on a large amount of data to connect the source and target domains, which is particularly difficult on high-entropy alloy datasets with small data volume and high labeling cost. For example, some performances, such as tensile strength and elongation, are difficult to measure, resulting in insufficient experimental data and low prediction accuracy compared to other related datasets. SUMMARY

[0003] The purpose of the present application is to provide a performance prediction method and system for high-entropy alloy based on adversarial transfer, to overcome the low prediction accuracy caused by insufficient data in the existing data-driven machine learning methods, and to improve the prediction accuracy of the performance of high-entropy alloy.

[0004] To achieve the above-mentioned purpose, the present application provides the following scheme:

[0005] The present application provides a performance prediction method for high-entropy alloy based on adversarial transfer, which comprises the following steps:

[0006] Obtain material data of source domain with labels and material data of target domain without labels;

[0007] Perform feature mapping and alignment on the material data of the source domain with labels and the material data of the target domain without labels to generate image data of the source domain with labels and image data of the target domain without labels;

[0008] Train a weak classifier according to the common image data of the source domain and the target domain, use the trained weak classifier to classify the image data of the target domain without labels according to the performance of the source domain, and obtain image data of the target domain with labels;

[0009] The adversarial transfer network model is trained by using image data with labels of a source domain and image data with labels of a target domain; the adversarial transfer network model comprises a feature representation network, a source domain bucketing network, a generation network and a discrimination network;

[0010] A high-entropy alloy performance prediction model comprising the trained feature representation network is constructed; the trained feature representation network is the feature representation network in the trained adversarial transfer network model;

[0011] The high-entropy alloy performance prediction model is trained by using image data with target domain performance labels of the target domain; the trained high-entropy alloy performance prediction model is used to predict the target domain performance of the high-entropy alloy of the target domain.

[0012] Optionally, the adversarial transfer network model is trained by using image data with labels of a source domain and image data with labels of a target domain, and specifically comprises:

[0013] The sample image data is input into the adversarial transfer network model, and a source domain bucketing loss, a generation loss, a discrimination loss and a CORAL loss are calculated; the sample image data is image data with labels of a source domain or image data with labels of a target domain;

[0014] The parameters of the feature representation network and the source domain bucketing network are updated according to the bucketing loss;

[0015] The parameters of the generation network are updated according to the generation loss;

[0016] The parameters of the feature representation network and the discrimination network are updated according to the discrimination loss;

[0017] The parameters of the source domain bucketing network and the discrimination network are updated according to the CORAL loss.

[0018] Optionally, the calculation formula of the bucketing loss is:

[0019]

[0020] wherein, represents the bucketing loss, x s represents image data of a source domain, F(x s ) represents a feature vector of x s generated based on the feature representation network F, C(F(x s )) represents a bucketing label of x s output by the source domain bucketing network C; is a true label of x s , and f L () represents a cross-entropy loss function.

[0021] Optionally, the calculation formula of the generation loss is:

[0022]

[0023] in, Let x represent the generation loss, α represent the proportion of generation loss to total loss, and x represent the generation loss. g Indicates the input data, x g =[F(x),z,l x ], z is noise, l x According to L x We obtain a one-hot vector, where x represents the sample image data, and G(x) g ) indicates that the generator network G is based on x g Output generated image data, D data (G(x g )) represents the output G(x) of the discrimination network D. g Y represents the probability of the true source domain image data. T,x D is the label for whether x is a true source domain image. cls (G(x g )) represents the output G(x) of the discrimination network D. g Category tags, f B () represents the logarithmic loss function.

[0024] Optionally, the formula for calculating the discrimination loss is:

[0025]

[0026] in, Let represent the discrimination loss, β represent the proportion of discrimination loss to the total loss, F(x) represent the representation vector of x generated based on the feature representation network F, and D cls (F(x)) represents the bucket label of the output F(x) of the discrimination network D, where D data (F(x)) represents the probability that the output F(x) of the discrimination network D is real image data from the source domain, Y T,x The label representing whether x is a true source domain image, Y F,x The label indicating whether x is a non-real source domain image, L x This represents the actual label of x.

[0027] Optionally, the formula for calculating CORAL loss is:

[0028]

[0029] Among them, L CORAL Indicates CORAL loss. The covariance matrix of the source domain characteristic matrix is ​​represented by the following: Let represent the covariance matrix of the target domain feature matrix. The source domain feature matrix is ​​constructed from the representation vectors of the image data of each source domain generated based on the feature representation network F. The target domain feature matrix is ​​constructed from the representation vectors of the image data of each target domain generated based on the feature representation network F. d ​​represents the dimension of the features. Let λ represent the Frobenius norm, and λ represent the proportion of CORAL loss to the total loss.

[0030] A performance prediction system for high-entropy alloys based on adversarial migration, the system using the method described above, the system comprising:

[0031] The data acquisition module is used to acquire material data of the source domain with labels and material data of the target domain without labels;

[0032] The alignment module is used to perform feature mapping and alignment on the material data of the labeled source domain and the material data of the unlabeled target domain, generating image data of the labeled source domain and image data of the unlabeled target domain.

[0033] The weak classification module is used to train a weak classifier based on the common image data of the source and target domains. The trained weak classifier is then used to classify the unlabeled target domain image data according to the source domain performance to obtain labeled target domain image data.

[0034] A pre-training module is used to train an adversarial transfer network model using labeled source domain image data and labeled target domain image data; the adversarial transfer network model includes: a feature representation network, a source domain bucketing network, a generator network, and a discriminator network;

[0035] The model building module is used to build a high-entropy alloy performance prediction model that includes a trained feature representation network; the trained feature representation network is the feature representation network in the trained adversarial transfer network model.

[0036] The training module is used to train the high-entropy alloy performance prediction model using image data of the target domain with target domain performance labels; the trained high-entropy alloy performance prediction model is used to predict the target domain performance of high-entropy alloys in the target domain.

[0037] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described above.

[0038] A computer-readable storage medium having a computer program stored thereon, which, when executed, implements the above-described method.

[0039] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0040] This invention provides a method and system for predicting the performance of high-entropy alloys based on adversarial transfer learning. In this invention, an adversarial transfer learning network model is trained using image data from both the source and target domains. By learning from the image data in both domains, the model fully learns the relationships and shifts between the image data in the source domain, achieving pre-training of the feature representation network in the high-entropy alloy performance prediction model. This further enables high-precision training of the high-entropy alloy performance prediction model using a small sample of image data from the target domain, thereby improving the prediction accuracy of high-entropy alloy performance. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 A flowchart of a performance prediction method for high-entropy alloys based on adversarial migration provided in an embodiment of the present invention;

[0043] Figure 2 A schematic diagram illustrating a performance prediction method for high-entropy alloys based on adversarial migration, provided in an embodiment of the present invention.

[0044] Figure 3 A flowchart for data feature alignment and mode conversion provided in an embodiment of the present invention;

[0045] Figure 4 A framework diagram of an adversarial migration network provided in an embodiment of the present invention;

[0046] Figure 5 This is a flowchart of training an adversarial transfer representation network F according to an embodiment of the present invention;

[0047] Figure 6 This is a flowchart illustrating high-precision modeling using representation vectors, provided as an embodiment of the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] The purpose of this invention is to provide a method and system for predicting the performance of high-entropy alloys based on adversarial migration, so as to overcome the shortcomings of existing data-driven machine learning methods that result in low prediction accuracy due to insufficient data, and improve the prediction accuracy of the performance of high-entropy alloys.

[0050] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] There are domain-knowledge-based connections between different properties of high-entropy alloys. It is necessary to integrate domain knowledge into domain adaptive networks to enhance the generalization ability and interpretability of the model.

[0052] Based on this, this invention proposes a novel domain adaptation method to address these issues. This method can improve prediction performance, reduce annotation costs, enhance the model's generalization ability and interpretability, and also serve a universal purpose in different high-entropy alloy performance prediction tasks. This method is a joint adversarial and representation training approach. It uses a generator-discriminator model to pass target distribution information into the learning embedding and, based on domain knowledge, constructs a connection between source domain target values ​​and target domain target values. This allows the network to perform gradient updates through the target domain labels, obtaining a better representation of the high-entropy alloy composition features and improving the representation's performance in the regression model.

[0053] Example 1

[0054] Embodiment 1 of the present invention provides a method for predicting the performance of high-entropy alloys based on adversarial migration, such as... Figure 1 and Figure 2 As shown, the method includes the following steps:

[0055] Step 101: Obtain material data for the source domain with labels and material data for the target domain without labels.

[0056] Step 102: Perform feature mapping and alignment on the material data of the labeled source domain and the material data of the unlabeled target domain to generate image data of the labeled source domain and image data of the unlabeled target domain.

[0057] Step 103: Train a weak classifier based on the common image data of the source and target domains, and use the trained weak classifier to classify the unlabeled target domain image data according to the source domain performance to obtain labeled target domain image data.

[0058] like Figure 2As shown, step A: In the source and target domains, due to differences in data sources and the complexity of data measurement, data features often exhibit bias. To adjust these feature biases, this embodiment of the invention provides a method using Xenonpy, which aligns features by mapping the fundamental compositional features of materials to a unified space. (Refer to...) Figure 3 A specific embodiment of step A includes the following steps:

[0059] Step A1: First, extract multiple elements from the composition of the alloy. For example, if the alloy is composed of 5 elements, the composition formula is A42B25C18D10E5. Next, normalize these 5 percentages to obtain two values, which are the proportions of each element: A-42, B-25, C-18, D-10, and E-5.

[0060] Step A2 involves using XenonPy tools to expand the atomic-level features of elements A through E. This may include physical and chemical properties such as atomic radius, electron cloud density, and electron cloud shape. These features are mapped to a pre-defined multidimensional space that reflects the comprehensive characteristics of the atoms, based on the type and importance of the atoms, resulting in the original high-dimensional features.

[0061] Step A3: Based on the atomic-level features obtained in Step A2, a relevance feature arrangement framework is constructed. This framework arranges features according to certain rules, ensuring that related features are adjacent. For example, atomic radius and electron cloud density, two features that directly affect the mechanical properties of materials, might be arranged adjacently. This allows for better capture of the relationship between these two features during subsequent convolutional neural network processing. Since most materials have compositional features, both the source and target domains can perform feature alignment in this dimension to obtain the mode-transformed and aligned features X. Image data representing the source domain, Image data representing the target domain. Simultaneously, based on a small amount of shared data from both the source and target domains, a weak classifier is built to predict source domain performance bin labels based on target domain performance. The target domain data is also labeled with source domain bin labels for subsequent image generation and bin identification tasks.

[0062] Step 104: Train an adversarial transfer network model using labeled source domain image data and labeled target domain image data; the adversarial transfer network model includes: a feature representation network, a source domain bucketing network, a generator network, and a discriminator network, such as... Figure 2 As shown, it specifically includes:

[0063] Step B: Construct an adversarial transfer network framework with domain adaptability, i.e., an adversarial transfer network model. In this invention, a composite network architecture comprising a feature representation network F, a source domain bucketing network C, a generator network G, and a discriminator network D is designed and constructed. In this architecture, the primary responsibility of the feature representation network F is to learn prior knowledge from the source and target domain data to facilitate understanding and learning of domain shifts. The generator network G and the discriminator network D effectively assist the feature representation network F in extracting more representative embedding vectors during the adversarial process. The network architecture constructed in this implementation scheme refers to... Figure 4 This includes the following steps:

[0064] Step B1: First, process the target characteristic in the source domain, assuming the target characteristic is the material's hardness. Then, bin the hardness according to a certain range, generating a set of labels L = {0, 1, 2... N}. c}, where N c It refers to the number of buckets. For example, if the hardness range is 1000-2000, it can be divided into 5 buckets, each with a range of 200.

[0065] Step B2, next, the present invention constructs a feature representation network F: The goal of this network is to generate a d-dimensional response vector based on the input source and target domain data. During training, network F implicitly captures the source domain data by learning and adjusting its parameters. and target domain Domain offset between.

[0066] Step B3, further, construct a source domain bucket network C: This network is trained using the hard bucket labels generated in step one and the feature representations extracted in step B2. Note that network C can only access the labels of the data sampled from the source distribution, but not the data sampled from the target distribution.

[0067] Step B4, construct the generator network G: And the identification network D: These two networks work against each other to improve the efficiency of the feature representation network F. Specifically, the generator G uses the feature representation generated in step B2, hard bucket labels, and random noise z to generate a new "fake" image X'. The discriminator network D then needs to distinguish whether the input image X is a real image or a fake image generated by the generator, and predict its hard bucket label.

[0068] Step B5, finally, to further improve the algorithm's performance, a residual connection is added between the discrimination network D and the feature representation network F. This allows F to directly learn more information from D during training. This also prepares the ground for subsequent gradient reversal training.

[0069] Step C: Implement domain-adaptive embedding vector acquisition based on adversarial transfer learning. This embodiment of the invention introduces a domain adaptation strategy based on an adversarial generative network framework. This strategy induces a symbiotic relationship between the embedding learning process and the generative adversarial network, thereby making the integrated feature space closer to the distribution of the source and target domains. To address the challenge of small sample sizes, this embodiment employs CORAL Loss to reduce domain bias and uses a gradient inversion layer to train the embedding generation. The entire network framework is jointly trained and continuously optimized iteratively until the model finally converges. (Refer to...) Figure 5 A specific embodiment of step C includes the following steps:

[0070] Step C1: Input the real source image x obtained in step A into F to obtain the embedding F(x). Input F(x) into C to predict the bucket probability distribution C(F(x)) of HV. Calculate the loss between the bucket probability and the source domain label and backpropagate for training.

[0071]

[0072] in, Indicates the loss from bucket division, x s F(x) represents the image data of the source domain. s ) represents x generated based on the feature representation network F. s The representation vector, C(F(x) s )) represents the x output by the source domain bucketing network C. s The bucket labels; For x s The real label, f L () represents the cross-entropy loss function.

[0073] Step C2, the input to the generator network G is x g = [F(x), z, l], where It is obedience noise, The one-hot vector is obtained based on the bucket label L, {N c The +1 dimension is used to label the source of F(x), where 0 represents that it comes from the source domain and 1 represents that it comes from the target domain. G outputs the generated image G(x). g The generated image is input into the discrimination network D, and is trained by receiving the gradient back propagated from D. Note that the goals of the two are opposite.

[0074]

[0075] in, Let x represent the generation loss, α represent the proportion of generation loss to total loss, and x represent the generation loss. gIndicates the input data, x g =[F(x),z,l x ], z is noise, l x According to L x We obtain a one-hot vector, where x represents the sample image data, and G(x) g ) indicates that the generator network G is based on x g Output generated image data, D data (G(x g ) represents the output G(x) of the discrimination network D. g Y represents the probability of the true source domain image data. T,x D is the label for whether x is a true source domain image. cls (G(x g )) represents the output G(x) of the discrimination network D. g The bucket label, f B () represents the logarithmic loss function.

[0076] Step C3, the input of the identification network D has three parts: (1) First, the real image input x. D needs to correctly identify the source of x (if it comes from the source domain, it is identified as real; if it comes from the target domain, it is identified as fake) and give the corresponding D. cls (x); (2) Next is the image G(x) generated by G. g D needs to identify it as fake, because it is a fake image, so the output D is... cls (G(x g In D, gradients do not need to be backpropagated. Note that only the labels of the real images in the source domain are Y. T The source domain generated image and the target domain real or generated image are both labeled Y. F (3) The third part is the output F(x) of F. D needs to correctly identify its domain origin and give the corresponding D. cls It's worth noting that gradient backpropagation here only involves the fully connected layers of the D network; the gradient is not backpropagated to CNN layers. For the first and second parts of image data discrimination, the image discrimination part of the D network can be used. For the third part, F(x), the feature vector discrimination part of the D network can be used. The image discrimination part and the feature vector discrimination part can be completely independent or implemented based on a cascaded network. For example, the D network includes interconnected feature representation modules and feature discrimination modules. When discriminating images, the feature representation module is input first, and its output is then input to the feature discrimination module. When discriminating feature vectors, they are directly input to the feature discrimination module. The loss of these three parts can be expressed in the following mathematical form:

[0077]

[0078] in, Let represent the discrimination loss, β represent the proportion of discrimination loss to the total loss, F(x) represent the representation vector of x generated based on the feature representation network F, and D cls (F(x)) represents the bucket label of the output F(x) of the discrimination network D, where D data (F(x)) represents the probability that the output F(x) of the discrimination network D is real image data from the source domain, Y T,x The label representing whether x is a true source domain image, Y F,x The label indicating whether x is a non-real source domain image, L x This represents the actual label of x.

[0079] Step C4 introduces CORAL Loss, which can intuitively reduce the difference between the two domain embeddings. However, in the early training stage, the goal is still to converge the source domain binning network C and the discriminant network D, and to construct feature embeddings that can be reasonably binned and confused by the domain. Therefore, CORAL Loss needs to be multiplied by a dynamic weight to balance the two tasks of domain offset and label binning in the later stage of training.

[0080]

[0081] Where L CORAL Indicates CORAL loss. The covariance matrix of the source domain characteristic matrix is ​​represented by the following: Let represent the covariance matrix of the target domain feature matrix. The source domain feature matrix is ​​constructed from the representation vectors of the image data in each source domain generated by the feature representation network F. The target domain feature matrix is ​​constructed from the representation vectors of the image data in each target domain generated by the feature representation network F. d ​​represents the dimension of the features, used to adjust the scale of the loss value. Let λ represent the Frobenius norm, and λ represent the proportion of CORAL loss to the total loss.

[0082] Step C5 combines the gradients generated from the previous C, G, and D network training to train the embedding extraction network F, and introduces CORALLoss to ensure the convergence of the model on small samples. Note that the gradient obtained by F from D is the gradient after being flipped by the gradient inversion layer, so the goal of F is also to make the distribution of the source domain and the target domain in the learned feature space closer.

[0083]

[0084]

[0085]

[0086]

[0087] Step C6: Iteratively train the four networks F, C, D, and G until network C converges, then stop training to obtain the jointly trained feature representation network F. During joint training, the representation network F effectively extracts useful features from the rich labeled data in the source domain and transfers them to the target domain with a smaller sample size, thus reducing dependence on a large amount of labeled data in the target domain. This method avoids the risk of overfitting in small-sample training because the additional information provided by the source domain enhances the network's generalization ability. Due to the consistency of certain features between the source and target domains, the domain-adaptive network can identify and learn domain-invariant features, enabling stable performance even on a limited number of samples in the target domain.

[0088] Step 105: Construct a high-entropy alloy performance prediction model that includes a trained feature representation network; the trained feature representation network is the feature representation network in the trained adversarial transfer network model.

[0089] Step 106: Train the high-entropy alloy performance prediction model using image data of the target domain with target domain performance labels; the trained high-entropy alloy performance prediction model is used to predict the target domain performance of high-entropy alloys in the target domain.

[0090] like Figure 2 As shown, steps 104 and 105 specifically include:

[0091] Step D: Using the feature representation network F trained in step 103, effective component embedding vectors can be extracted from datasets containing different target properties, compositions, and processes. These vectors can be used to build high-precision prediction models for high-entropy alloy properties and applied to the reverse design and optimization process of high-entropy alloys. (Refer to...) Figure 6 A specific embodiment of step D includes the following steps:

[0092] Step D1: First, the composition information of the materials is obtained by feature alignment and mode transformation to obtain image data x for each material.

[0093] Step D2: Subsequently, these images are input into the trained feature representation network to obtain a component representation vector F(x).

[0094] Step D3, next, combine this component representation vector F(x) with other parameters (e.g., manufacturing method, temperature, etc.) x process The merging process yielded the final feature X.

[0095] Step D4: After obtaining the final features, the high-entropy alloy performance prediction model can be trained. This can be accomplished in many ways, such as using random forests, gradient boosting trees, or neural networks. We chose the edRVFL network for training, and finally used grid search to find the optimal model parameters to achieve the highest prediction accuracy.

[0096] Step D5, finally, allows the trained high-precision prediction model to be used in the reverse design of high-entropy alloys. For example, given a hardness target, optimization algorithms can be used to find material compositions and manufacturing methods that can achieve this hardness target. The model can also guide experiments, for example, identifying which material compositions or manufacturing methods might improve the hardness of the high-entropy alloy.

[0097] Example 2

[0098] Embodiment 2 of the present invention provides a performance prediction system for high-entropy alloys based on adversarial migration. The system uses the method described above and includes:

[0099] The data acquisition module is used to acquire material data of the source domain with labels and material data of the target domain without labels.

[0100] The alignment module is used to perform feature mapping and alignment on the material data of the labeled source domain and the material data of the unlabeled target domain, generating image data of the labeled source domain and image data of the unlabeled target domain.

[0101] The weak classification module is used to train a weak classifier based on the common image data of the source and target domains. The trained weak classifier is then used to classify the unlabeled target domain image data according to the source domain performance, thus obtaining labeled target domain image data.

[0102] The pre-training module is used to train an adversarial transfer network model using labeled source domain image data and labeled target domain image data; the adversarial transfer network model includes: a feature representation network, a source domain bucketing network, a generator network, and a discriminator network.

[0103] The model building module is used to build a high-entropy alloy performance prediction model that includes a trained feature representation network; the trained feature representation network is the feature representation network in the trained adversarial transfer network model.

[0104] The training module is used to train the high-entropy alloy performance prediction model using image data of the target domain with target domain performance labels; the trained high-entropy alloy performance prediction model is used to predict the target domain performance of high-entropy alloys in the target domain.

[0105] Example 3

[0106] Embodiment 3 of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method.

[0107] Furthermore, when the computer program in the aforementioned memory is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0108] Example 4

[0109] Embodiment 4 of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the above-described method.

[0110] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0111] First, to address the bias in data features between the source and target domains due to differences in data sources and measurement difficulty, Xenonpy is used to obtain the basic composition feature maps of the material and align these features. Then, a domain-adaptive adversarial transfer network model is constructed, consisting of a feature representation network, a source domain binning network, a generator network, and a discriminator network, to learn the domain shifts between the source and target domains. Next, through the framework of the adversarial generator network, a domain-adaptive method is developed that coexists between learning embeddings and the generator adversarial network. This method makes the joint feature space closer to the distribution of the source and target domains, effectively solving the small sample size problem. CORAL Loss is introduced to reduce the domain shift, and a gradient inversion layer is used to train the generation of embeddings. The entire network framework is jointly trained until the model converges. Finally, the obtained domain-adaptive embedding vectors are used for regression modeling to establish a high-precision predictor of the high-entropy alloy performance, and the inverse design and optimization of the high-entropy alloy are performed. This invention introduces a domain adaptive network that couples feature extraction and adversarial generation, making the domain offset learning between high-entropy alloy datasets with different target performance more accurate. Simultaneously, it introduces covariance distance and residual inverse gradient connections, effectively helping the feature representation network obtain better high-entropy alloy composition feature representations from limited data for constructing a regressor. Finally, based on a high-precision performance predictor, it realizes the reverse design of high-entropy alloy composition and heat treatment processes with excellent performance, and assists in experimental verification. This invention not only improves the accuracy of material performance prediction but also optimizes the design and manufacturing process of high-entropy alloys, providing an effective solution to current problems in the industrial production of high-entropy alloys.

[0112] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0113] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting the performance of high-entropy alloys based on adversarial migration, characterized in that, The method includes the following steps: Obtain material data for the labeled source domain and material data for the unlabeled target domain; Feature mapping and alignment are performed on the material data of the labeled source domain and the material data of the unlabeled target domain to generate image data of the labeled source domain and image data of the unlabeled target domain. A weak classifier is trained based on the common image data of the source and target domains. The trained weak classifier is then used to classify the unlabeled target domain image data according to the source domain performance, thus obtaining labeled target domain image data. An adversarial transfer network model is trained using labeled source domain image data and labeled target domain image data; The adversarial transfer network model includes: a feature representation network, a source domain bucketing network, a generator network, and a discrimination network; A high-entropy alloy performance prediction model is constructed, which includes a trained feature representation network; the trained feature representation network is the feature representation network in the trained adversarial transfer network model. A high-entropy alloy performance prediction model is trained using image data of the target domain labeled with target domain performance; the trained high-entropy alloy performance prediction model is then used to predict the target domain performance of high-entropy alloys in the target domain. The adversarial transfer network model is trained using labeled source domain image data and labeled target domain image data, specifically including: The sample image data is input into the adversarial transfer network model to calculate the source domain binning loss, generation loss, discrimination loss, and CORAL loss; the sample image data is either labeled source domain image data or labeled target domain image data. The parameters of the feature characterization network and the source domain bucketing network are updated based on the bucketing loss. Update the parameters of the generator network based on the generation loss; The parameters of the feature representation network and the discrimination network are updated based on the discrimination loss; Update the parameters of the source domain bucketing network and the discrimination network based on the CORAL loss.

2. The performance prediction method for high-entropy alloys based on adversarial migration according to claim 1, characterized in that, The formula for calculating the loss from bucket division is: ； in, Indicates the loss from dividing the container. Represents image data from the source domain. Represents the generation based on the feature representation network F The representation vector, This represents the output of the source domain bucketing network C. The bucket labels; for The true label, This represents the cross-entropy loss function.

3. The performance prediction method for high-entropy alloys based on adversarial migration according to claim 2, characterized in that, The formula for calculating the generation loss is: ； in, Indicates the generation loss. This indicates the proportion of generated loss to total loss. Indicates input data, , For noise, According to Obtain the one-hot vector. express The true label, where x represents the sample image data. The generator network G is based on Output generated image data, This indicates the output of the discrimination network D. The probability of the image data being from the real source domain. Let x be a label indicating whether it is a true source domain image. This indicates the output of the discrimination network D. The bin labels, This represents the logarithmic loss function.

4. The performance prediction method for high-entropy alloys based on adversarial migration according to claim 3, characterized in that, The formula for calculating identification loss is: ； in, Indicates identification loss, This indicates the proportion of identification loss to total loss. Represents the generation based on the feature representation network F The representation vector, This indicates the output of the discrimination network D. The bin labels, This indicates the output of the discrimination network D. The probability of the image data being from the real source domain. express Whether the label is from a real source domain image express Is the label for an image that is not from the real source domain? express The true label.

5. The performance prediction method for high-entropy alloys based on adversarial migration according to claim 4, characterized in that, The formula for calculating CORAL loss is: ； in, Indicates CORAL loss. The covariance matrix of the source domain characteristic matrix is ​​represented by the following expression. The covariance matrix represents the feature matrix of the target domain. The feature matrix of the source domain is constructed from the representation vectors of the image data of each source domain generated based on the feature representation network F. The feature matrix of the target domain is constructed from the representation vectors of the image data of each target domain generated based on the feature representation network F. The dimension representing the feature. Describing the Frobenius norm, This indicates the proportion of CORAL loss to total loss.

6. A performance prediction system for high-entropy alloys based on adversarial migration, characterized in that, The system uses the method according to any one of claims 1-5, the system comprising: The data acquisition module is used to acquire material data of the source domain with labels and material data of the target domain without labels; The alignment module is used to perform feature mapping and alignment on the material data of the labeled source domain and the material data of the unlabeled target domain, generating image data of the labeled source domain and image data of the unlabeled target domain. The weak classification module is used to train a weak classifier based on the common image data of the source and target domains. The trained weak classifier is then used to classify the unlabeled target domain image data according to the source domain performance to obtain labeled target domain image data. A pre-training module is used to train an adversarial transfer network model using labeled source domain image data and labeled target domain image data; the adversarial transfer network model includes: a feature representation network, a source domain bucketing network, a generator network, and a discriminator network; The model building module is used to build a high-entropy alloy performance prediction model that includes a trained feature representation network; the trained feature representation network is the feature representation network in the trained adversarial transfer network model. The training module is used to train the high-entropy alloy performance prediction model using image data of the target domain with target domain performance labels; the trained high-entropy alloy performance prediction model is used to predict the target domain performance of high-entropy alloys in the target domain.

7. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, It stores a computer program thereon, which, when executed, implements the method as described in any one of claims 1 to 5.

Images