An alloy yield strength prediction method and device, computer equipment and medium

Abstract

The application provides an alloy yield strength prediction method and device, computer equipment and medium, and belongs to the metal strength prediction field, and the method comprises the following steps: acquiring chemical composition data and metallographic image data of an alloy, respectively extracting data feature vectors and image feature vectors and performing linear projection transformation mapping to a pre-configured common semantic space with a certain dimension; the projected data feature vectors and the projected image feature vectors are spliced to obtain a joint feature vector; the data modal dynamic weight and the image modal dynamic weight of the current alloy are generated by globally modeling the joint feature vector; the projected data feature vectors and the projected image feature vectors are weighted and spliced respectively to obtain a fusion feature vector; the yield strength prediction value of the alloy is obtained by predicting the mechanical properties based on the fusion feature vector. In this way, the rationality and accuracy of multi-modal information fusion are significantly improved, and the prediction accuracy of the yield strength of the alloy is effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of metal strength prediction, specifically relating to a method, apparatus, computer equipment, and medium for predicting the yield strength of alloys. Background Technology

[0002] The chemical composition and microstructure of an alloy jointly determine its mechanical properties, with yield strength being a key indicator of a material's load-bearing capacity. Accurate prediction of yield strength is crucial for the design of new alloys and process optimization. Traditional prediction methods often rely on empirical formulas relating chemical composition and process parameters, such as Hall-Petch relations and semi-empirical constitutive equations. However, when these models are extended to different alloy systems, they struggle to decouple the independent contributions and interactions of individual alloying elements, resulting in poor universality.

[0003] In recent years, deep learning technology has been widely used in predicting material properties. On the one hand, researchers use numerical data such as chemical composition and process parameters as input to build prediction models through deep neural networks or one-dimensional convolutional neural networks; on the other hand, progress has also been made in using convolutional neural networks to extract microstructural features from metallographic images for prediction. However, most of the above methods rely only on single-modal information, ignoring the synergistic effect between chemical composition and microstructure, resulting in limited prediction accuracy.

[0004] Therefore, multimodal fusion methods have emerged. In existing technologies, some methods simply concatenate chemical composition features with image features before inputting them into a prediction network. However, this approach cannot adaptively adjust the contribution weights of different modalities to the prediction result, limiting the fusion effect. Further, some methods introduce attention mechanisms to generate modal weights. For example, chemical composition features and image features are concatenated and input into an attention network, and the weights of the two modalities are obtained through fully connected layers and Softmax. The original features are then weighted and fused. However, these methods share a common drawback: the physical meanings of chemical composition features and image features are vastly different, making it difficult to match the physical meanings of different modalities. This results in insufficiently precise weight learning, thus limiting the fusion effect and prediction accuracy.

[0005] This shortcoming makes it difficult to reliably assess the model's application value in actual materials research and development. Summary of the Invention

[0006] To address the challenge of achieving accurate semantic alignment of multimodal features to improve fusion performance and prediction accuracy, this invention provides a method, apparatus, computer device, and medium for predicting the yield strength of alloys.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting the yield strength of an alloy, the method comprising: Acquire chemical composition data and metallographic image data of the alloy; Extract the data feature vectors of the chemical composition data and the image feature vectors of the metallographic image data respectively; The data feature vector and the image feature vector are respectively subjected to linear projection transformation and mapped to a pre-defined common semantic space to obtain projected data feature vector and projected image feature vector with the same dimension. The projected data feature vector and the projected image feature vector are concatenated to obtain a joint feature vector; global context modeling is performed on the joint feature vector to generate the dynamic weights of the data modes and the dynamic weights of the image modes of the current alloy. Using the generated data modality dynamic weights and image modality dynamic weights, the projected data feature vector and the projected image feature vector are weighted respectively, and the two weighted feature vectors are concatenated to obtain a fused feature vector; Mechanical properties are predicted based on the fused feature vector, and the predicted yield strength of the alloy is obtained.

[0008] Optionally, the extraction of data feature vectors from the chemical composition data and image feature vectors from the metallographic image data respectively includes: The chemical composition data is regarded as a one-dimensional sequence. The collaborative features between local adjacent elements are extracted sequentially through one-dimensional convolution operation, and then compressed into a fixed-length feature vector through adaptive pooling. This fixed-length feature vector is the data feature vector. The metallographic image data is input into a pre-trained visual Transformer model, which outputs a high-dimensional image feature vector.

[0009] Optionally, in the linear projection transformation, the dimension of the common semantic space is preset to 256; the data feature vector is mapped to the 256-dimensional projected data feature vector through a trainable first linear projection matrix; the image feature vector is mapped to the 256-dimensional projected image feature vector through a trainable second linear projection matrix.

[0010] Optionally, generating the data modal dynamic weights and image modal dynamic weights of the current alloy by performing global context modeling on the joint feature vector includes: The joint feature vector is input into a pre-constructed weight generation network, which consists of a first fully connected layer, a second fully connected layer, and a softmax layer connected sequentially. The first fully connected layer compresses the dimension of the joint feature vector and uses ReLU non-linear activation. The second fully connected layer maps the compressed features to 2D. The softmax layer normalizes the 2D values ​​into a probability distribution that sums to 1. The two components of this probability distribution are the data modality dynamic weights and the image modality dynamic weights.

[0011] Optionally, the step of using the generated data modality dynamic weights and image modality dynamic weights to weight the projected data feature vector and the projected image feature vector respectively, and then concatenating the two weighted feature vectors to obtain the fused feature vector includes: The element-wise product of the dynamic weights of the data modal and the projected data feature vector is calculated to obtain the weighted data feature vector; the element-wise product of the dynamic weights of the image modal and the projected image feature vector is calculated to obtain the weighted image feature vector; the weighted data feature vector and the weighted image feature vector are concatenated along the feature channel dimension to generate the fused feature vector.

[0012] Optionally, based on the fused feature vector, the predicted yield strength of the alloy is obtained by predicting its mechanical properties, including: The fused feature vector is input into a pre-constructed multilayer perceptron regression network to obtain a normalized predicted yield strength value. The multilayer perceptron regression network includes: a first linear layer, which maps the dimension of the fused feature vector to an intermediate dimension and connects it to a Dropout layer; a second linear layer, which further compresses the intermediate dimension and connects it to a Dropout layer; and a third linear layer, which outputs a normalized predicted yield strength value. The normalized predicted yield strength is denormalized to obtain the actual predicted yield strength.

[0013] Optionally, before acquiring the chemical composition data and metallographic image data of the alloy, the method further includes a joint training step: Acquire labeled alloy sample data, divide the training set according to alloy type, and perform joint training. Use mean squared error as the loss function to compare the predicted yield strength value with the actual yield strength value. With minimizing the loss function as the objective, use an adaptive moment estimation optimizer to optimize the parameters. During the training process, the projection matrix in the linear projection transformation, all parameters in the weight generation network, and all parameters in the multilayer perceptron regression network are jointly learned and updated under the supervision of the loss function.

[0014] An alloy yield strength prediction device, the device comprising: The acquisition module is used to acquire the chemical composition data and metallographic image data of the alloy; The extraction module is used to extract the data feature vector of the chemical composition data and the image feature vector of the metallographic image data, respectively. The mapping module is used to perform linear projection transformation on the data feature vector and the image feature vector respectively, mapping them to a pre-defined common semantic space, to obtain projected data feature vector and projected image feature vector with the same dimension. The generation module is used to concatenate the projected data feature vector and the projected image feature vector to obtain a joint feature vector; and to generate the data mode dynamic weights and image mode dynamic weights of the current alloy by performing global context modeling on the joint feature vector. The fusion module is used to use the generated data modality dynamic weights and image modality dynamic weights to weight the projected data feature vector and the projected image feature vector respectively, and then concatenate the two weighted feature vectors to obtain the fused feature vector. The prediction module is used to predict the mechanical properties based on the fused feature vector to obtain the predicted yield strength value of the alloy.

[0015] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for predicting the yield strength of an alloy.

[0016] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned method for predicting the yield strength of an alloy.

[0017] The alloy yield strength prediction method provided by this invention has the following beneficial effects: This invention achieves explicit semantic alignment of heterogeneous features by first linearly projecting chemical composition data features and metallographic image features onto the same common semantic space, eliminating the incomparability of physical meanings between different modalities caused by differences in the original feature spaces. Based on this, global context modeling is performed using the aligned joint features, adaptively generating dynamic weights at the sample level. This allows the model to dynamically adjust the contribution ratio of composition and microstructure information in the prediction based on the specific circumstances of each alloy sample. This two-stage fusion strategy of "alignment first, then dynamic weighting" significantly improves the rationality and accuracy of multimodal information fusion, thereby effectively improving the prediction accuracy of alloy yield strength. Simultaneously, because the feature representations in the common space are more compact and semantically consistent, the model's adaptability to different alloy systems is enhanced, exhibiting good generalization stability when facing new alloys with significant differences in composition and microstructure. In summary, this invention, while ensuring prediction accuracy, enhances the model's cross-type generalization ability for unseen alloy types, providing a more reliable and practical multimodal fusion scheme for material performance prediction. Attached Figure Description

[0018] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. 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.

[0019] Figure 1 This is a flowchart illustrating an alloy yield strength prediction method according to an exemplary embodiment of the present invention.

[0020] Figure 2 This is a radar chart comparing the true and predicted values ​​of a test set according to an exemplary embodiment of the present invention.

[0021] Figure 3 This is a block diagram of an alloy yield strength prediction device provided by the present invention according to an exemplary embodiment. Detailed Implementation

[0022] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0023] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] First, this invention provides a method for predicting the yield strength of an alloy, specifically as follows: Figure 1As shown, it includes the following steps: S101. Obtain the chemical composition data and metallographic image data of the alloy, and extract the data feature vector of the chemical composition data and the image feature vector of the metallographic image data respectively.

[0025] In this step, the chemical composition data and metallographic image data of the alloy are first obtained. For example:

[0026] (1) Chemical composition data: The mass fraction (%) of 28 elements including Fe, C, Cr, Mn, Si, Ni, Co, Mo, W, Nb, Al, P, Cu, Ti, Ta, Hf, Re, V, B, N, O, S, Zr, and Y is stored in tabular form, with each sample being a 28-dimensional vector.

[0027] (2) Metallographic image data: Microscopic images of the alloy taken by an optical microscope or a scanning electron microscope, uniformly saved in JPEG or PNG format, with a resolution of not less than 224×224 pixels.

[0028] Next, preprocessing is performed: (1) Normalize the chemical composition data and use the MinMaxScaler method to scale the content of all elements to the [0,1] range.

[0029] (2) The same MinMax normalization process is applied to the target value of yield strength to eliminate the influence of dimensions.

[0030] (3) Metallographic image preprocessing: Adjust the image size to 224×224 pixels, convert it to RGB three channels, and normalize the pixel values ​​(divide by 255).

[0031] In one embodiment, the chemical composition data is regarded as a one-dimensional sequence. The collaborative features between local adjacent elements are extracted sequentially through one-dimensional convolution operation, and then compressed into a fixed-length feature vector through adaptive pooling. This fixed-length feature vector is the data feature vector. The metallographic image data is input into a pre-trained visual Transformer model, which outputs a high-dimensional image feature vector.

[0032] For example, when the chemical composition data is tabular data containing the mass fractions of multiple elements, a one-dimensional convolutional neural network (1D-CNN) is constructed as a table feature extractor, named TableCNN. The design of this encoder is based on the premise that the arrangement order of alloying elements (sorted according to the periodic table and common alloying rules) implies the proximity relationship of physicochemical properties, and one-dimensional convolution can capture the synergistic effect between adjacent elements.

[0033] The specific structure of TableCNN is shown in Table 1 below: Table 1. Specific Structure of TableCNN The output is a 128-dimensional tabular feature vector.

[0034] Secondly, the pre-trained OpenCLIP visual model (ViT-B-16 architecture) is used as the image feature extractor. OpenCLIP is a visual-language model pre-trained on large-scale image-text pair data (400 million samples) and has strong general visual representation capabilities.

[0035] Feature extraction process: ① Input the preprocessed metallographic image (224×224×3) into the OpenCLIP ViT-B-16 model.

[0036] ②The model output is a 768-dimensional image feature vector.

[0037] ③ Freeze all parameters of the encoder and do not perform gradient updates during subsequent training; use it only as a fixed feature extractor.

[0038] Design considerations: Using a frozen pre-trained large model avoids the overfitting problem that is common when training CNNs on small-scale material image datasets.

[0039] OpenCLIP's multimodal pre-training makes it more sensitive to features such as texture and phase boundaries of microstructures, outperforming randomly initialized CNNs or conventional ImageNet pre-trained models.

[0040] S102. Perform linear projection transformation on the data feature vector and the image feature vector respectively, mapping them to a pre-defined common semantic space, to obtain projected data feature vector and projected image feature vector with the same dimension.

[0041] The core innovation of this invention lies in: firstly, projecting heterogeneous features onto a common semantic space for dimensional alignment, and then achieving adaptive fusion by learning sample-level dynamic weights. Specifically, it is divided into two stages, and this step is the first stage.

[0042] In this linear projection transformation step, the dimension of the common semantic space is preset to 256; the data feature vector is mapped to the 256-dimensional projected data feature vector through a trainable first linear projection matrix; the image feature vector is mapped to the 256-dimensional projected image feature vector through a trainable second linear projection matrix.

[0043] For example, since the table features h_t (128 dimensions) and image features h_i (768 dimensions) have mismatched dimensions and different physical meanings, they cannot be directly compared or weighted. Therefore, they are mapped to a common space of the same dimension (256 dimensions, which was determined to be the optimal value through experiments) through trainable linear projection layers.

[0044] Calculation formula: h̃_t = W_t·h_t + b_t, h̃_i = W_i·h_i + b_i, where W_t∈R^(256×128), W_i∈R^(256×768); Where W_t and W_i are the weight matrices of the table feature projection layer and the image feature projection layer, respectively, namely the first linear projection matrix and the second linear projection matrix, and b_t and b_i are the trainable bias terms of the table feature projection layer and the image feature projection layer, respectively.

[0045] The resulting h̃_t and h̃_i after projection are both 256-dimensional vectors, possessing comparable semantic representations in the common space. This feature alignment step forms the basis for subsequent dynamic weight learning.

[0046] S103. The projected data feature vector and the projected image feature vector are concatenated to obtain a joint feature vector. The data mode dynamic weights and image mode dynamic weights of the current alloy are generated by performing global context modeling on the joint feature vector.

[0047] This step is the second stage of the core innovation of this invention. In this step, the joint feature vector is input into a pre-constructed weight generation network, which is composed of a first fully connected layer, a second fully connected layer, and a softmax layer connected sequentially. The first fully connected layer compresses the dimension of the joint feature vector and uses ReLU nonlinear activation. The second fully connected layer maps the compressed features to 2D. The softmax layer normalizes the 2D values ​​into a probability distribution that sums to 1. The two components of this probability distribution are the dynamic weights of the data modality and the dynamic weights of the image modality.

[0048] For example, concatenating the two projected 256-dimensional feature vectors yields a 512-dimensional joint feature vector: h_joint = concat(hmber_t, hmber_i); The h_joint is input into the Attention Network, which consists of two fully connected layers and a Softmax layer, to generate the dynamic weights a_t and a_i for the two modalities.

[0049] The attention network structure is as follows: First layer: Linear (512→256), ReLU activated.

[0050] Special design: The output dimension (256) of the first layer is consistent with the dimension of the common space, which enables the weight generation network to directly abstract modal importance in the semantic feature space.

[0051] Second layer: Linear(256 → 2), no activation.

[0052] Softmax layer: Normalizes 2-dimensional logits into a probability distribution (ɑ_t, ɑ_i) that satisfies ɑ_t + ɑ_i = 1.

[0053] The generated weights depend entirely on the features of the current sample, so different samples can obtain different (ɑ_t, ɑ_i), thus achieving dynamic gating at the sample level.

[0054] S104. Using the generated dynamic weights of the data modality and the image modality, the projected data feature vector and the projected image feature vector are weighted respectively, and the two weighted feature vectors are concatenated to obtain the fused feature vector.

[0055] In this step, the element-wise product of the dynamic weights of the data modality and the projected data feature vector is calculated to obtain the weighted data feature vector; the element-wise product of the dynamic weights of the image modality and the projected image feature vector is calculated to obtain the weighted image feature vector; the weighted data feature vector and the weighted image feature vector are concatenated along the feature channel dimension to generate the fused feature vector.

[0056] For example, the generated weights are used to weight the feature vectors in the common space: h_t^w = ɑ_t·h̃_t; h_i^w = ɑ_i·h̃_i; The two weighted 256-dimensional vectors are concatenated again to obtain the final 512-dimensional fused feature h_fused: h_fused = concat(h_t^w, h_i^w); Thus, the present invention has a two-stage structure of "projection alignment first, then dynamic gating", and both the projection layer and the attention network are jointly optimized end-to-end under the supervision of the yield strength prediction task.

[0057] S105. Based on the fused feature vector, the mechanical properties are predicted to obtain the predicted yield strength of the alloy.

[0058] In this step, the fused feature vector is input into a pre-constructed multilayer perceptron regression network to obtain the output normalized yield strength prediction value. The multilayer perceptron regression network includes: a first linear layer, used to map the dimension of the fused feature vector to an intermediate dimension and then connect it to a dropout layer; a second linear layer, used to further compress the intermediate dimension and then connect it to a dropout layer; and a third linear layer, used to output the normalized yield strength prediction value. The normalized yield strength prediction value is then denormalized to obtain the actual yield strength prediction value.

[0059] The fused feature h_fused is input into the multilayer perceptron (MLP) regression head, and the normalized yield strength prediction value is output. Then, it is denormalized to obtain the true yield strength (unit: MPa).

[0060] The MLP structure is shown in Table 2: Table 2 MLP Structure Furthermore, before acquiring the chemical composition data and metallographic image data of the alloy to predict the yield strength, this invention also requires end-to-end joint training.

[0061] For example, joint training is performed using labeled sample data, with mean squared error as the loss function, and the predicted yield strength value is compared with the actual yield strength value. The parameters are optimized using an adaptive moment estimator with the goal of minimizing the loss function. During training, the projection matrix in the linear projection transformation, all parameters in the weight generation network, and all parameters in the multilayer perceptron regression network are jointly learned and updated under the supervision of the loss function.

[0062] For example, the hyperparameter settings are as follows: Loss function: Mean squared error (MSE), which compares the normalized predicted value with the true value.

[0063] Optimizer: Adam, initial learning rate = 1e-3.

[0064] Learning rate scheduling: ReduceLROnPlateau (monitor validation set loss, factor=0.5, patience=10).

[0065] Number of training rounds: 200 rounds.

[0066] Batch size: 8.

[0067] Training / Verification / Testing Division: Divided according to the cross-alloy zero-sample evaluation protocol unique to this invention.

[0068] Among them, the cross-alloy zero-shot evaluation protocol is a training method proposed in this invention to improve the accuracy of prediction results. To rigorously verify the model's generalization ability to novel alloy types, this invention does not employ conventional random partitioning or partitioning by sample proportion, but instead partitions the dataset by alloy type:

[0069] Training set: contains 44 alloys, with a total of 334 samples.

[0070] Test set: contains 10 alloy types that are completely different from those in the training set, with a total of 66 samples.

[0071] The alloys in the test set never appeared during the training phase, and their chemical composition and metallographic image features did not overlap with those in the training set.

[0072] This protocol can simulate real-world application scenarios: the model's predicted performance when faced with a new alloy that has never been developed before.

[0073] All comparative experiments (including single-modal models and other fusion methods) used the exact same training / test split to ensure fairness in the comparison.

[0074] By employing the aforementioned method, chemical composition data features and metallographic image features are first linearly projected onto the same common semantic space, achieving explicit semantic alignment of heterogeneous features and eliminating the incomparability of physical meanings between different modalities due to differences in the original feature spaces. Based on this, global context modeling is performed using the aligned joint features, adaptively generating dynamic weights at the sample level. This allows the model to dynamically adjust the contribution ratio of compositional and microstructural information in prediction based on the specific circumstances of each alloy sample. This two-stage fusion strategy of "alignment first, then dynamic weighting" significantly improves the rationality and accuracy of multimodal information fusion, thereby effectively improving the prediction accuracy of alloy yield strength. Simultaneously, because the feature representations in the common space are more compact and semantically consistent, the model's adaptability to different alloy systems is enhanced, exhibiting good generalization stability when facing new alloys with significant differences in composition and microstructure. In summary, this invention, while ensuring prediction accuracy, enhances the model's cross-type generalization ability for unseen alloy types, providing a more reliable and practical multimodal fusion scheme for material performance prediction.

[0075] Based on the above method steps, the present invention also provides embodiments for illustration.

[0076] Example 1: Prediction of yield strength of CPJ-7G alloy.

[0077] The method of this invention is used to predict the yield strength of CPJ-7G alloy. This alloy belongs to a type not seen in the training set (used for the test set). The 28-dimensional chemical composition data and corresponding metallographic images of the alloy are input, and the predicted values ​​are output after processing in steps S1-S5. A total of 8 samples of this alloy are tested, and the predicted values ​​are compared with the experimentally measured yield strengths. Figure 2 As shown.

[0078] Results: The predicted values ​​of all 8 samples fell within ±10% of the measured values, with an average relative error of only +0.14% and a median error of -0.49%. This indicates that the present invention has extremely high prediction accuracy and stability for this alloy.

[0079] Example 2: Prediction of yield strength of CPJ-10 alloy.

[0080] The method of this invention was also used to predict the performance of 8 samples of CPJ-10 alloy (test set, unseen). The results showed that 7 samples were within ±10% error range, with a pass rate of 87.5% and an average relative error of -3.60%. This indicates that the present invention has a good predictive effect on this alloy.

[0081] Example 3: Comparative Experiment – ​​Performance Comparison of Different Fusion Modules

[0082] To highlight the advantages of the dynamic gating fusion module of this invention, while keeping all other conditions (feature extractor, training parameters, data partitioning) completely identical, only the S4 fusion module was replaced with: ConcatBaseline, FixedWeightFusion, LowRankBilinearFusion, and SimpleCrossAttentionFusion. Evaluation was performed on the same zero-sample test set, and the results are shown in Table 3 below:

[0083] Table 3. Test reagent evaluation results Experiments have shown that the method of this invention achieves the best results among all comparative schemes, verifying the effectiveness of the two-stage feature alignment and dynamic gating mechanism.

[0084] Example 4: Cross-alloy generalization stability analysis.

[0085] The predicted R² values ​​were calculated for 10 unseen alloys in the test set. The results showed that 8 alloys had an R² value higher than 0.85, 6 had an R² value higher than 0.90, and the lowest R² value was 0.76. This indicates that the present invention has good generalization ability for different types of new alloys and has practical engineering application value.

[0086] Secondly, the present invention also provides an alloy yield strength prediction device, such as... Figure 3 As shown, it includes: The acquisition module 201 is used to acquire the chemical composition data and metallographic image data of the alloy.

[0087] Extraction module 202 is used to extract the data feature vector of the chemical composition data and the image feature vector of the metallographic image data, respectively.

[0088] The mapping module 203 is used to perform linear projection transformation on the data feature vector and the image feature vector respectively, and map them to a pre-defined common semantic space to obtain projected data feature vector and projected image feature vector with the same dimension.

[0089] The generation module 204 is used to concatenate the projected data feature vector and the projected image feature vector to obtain a joint feature vector; and to generate the data mode dynamic weight and image mode dynamic weight of the current alloy by performing global context modeling on the joint feature vector.

[0090] The fusion module 205 is used to weight the projected data feature vector and the projected image feature vector using the generated data modality dynamic weight and image modality dynamic weight, respectively, and then concatenate the two weighted feature vectors to obtain the fused feature vector.

[0091] The prediction module 206 is used to predict the mechanical properties based on the fused feature vector to obtain the predicted value of the alloy's yield strength.

[0092] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The steps of a method for predicting the yield strength of an alloy are provided.

[0093] This invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for various operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above-mentioned functions. Figure 1 The steps of a method for predicting the yield strength of an alloy are provided.

[0094] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0095] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0096] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0097] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0098] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the patent of the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for predicting the yield strength of an alloy, characterized in that, The method includes: Acquire chemical composition data and metallographic image data of the alloy; Extract the data feature vectors of the chemical composition data and the image feature vectors of the metallographic image data respectively; The data feature vector and the image feature vector are respectively subjected to linear projection transformation and mapped to a pre-defined common semantic space to obtain projected data feature vector and projected image feature vector with the same dimension. The projected data feature vector and the projected image feature vector are concatenated to obtain a joint feature vector; global context modeling is performed on the joint feature vector to generate the dynamic weights of the data modes and the dynamic weights of the image modes of the current alloy. Using the generated data modality dynamic weights and image modality dynamic weights, the projected data feature vector and the projected image feature vector are weighted respectively, and the two weighted feature vectors are concatenated to obtain a fused feature vector; Based on the fused feature vector, mechanical properties are predicted to obtain the predicted yield strength of the alloy.

2. The method according to claim 1, characterized in that, The extraction of data feature vectors from the chemical composition data and image feature vectors from the metallographic image data includes: The chemical composition data is regarded as a one-dimensional sequence. The collaborative features between local adjacent elements are extracted sequentially through one-dimensional convolution operation, and then compressed into a fixed-length feature vector through adaptive pooling. This fixed-length feature vector is the data feature vector. The metallographic image data is input into a pre-trained visual Transformer model, which outputs a high-dimensional image feature vector.

3. The method according to claim 1, characterized in that, In the linear projection transformation, the dimension of the common semantic space is preset to 256; the data feature vector is mapped to the 256-dimensional projected data feature vector through a trainable first linear projection matrix; the image feature vector is mapped to the 256-dimensional projected image feature vector through a trainable second linear projection matrix.

4. The method according to claim 1, characterized in that, The generation of data modal dynamic weights and image modal dynamic weights for the current alloy by performing global context modeling on the joint feature vector includes: The joint feature vector is input into a pre-constructed weight generation network, which consists of a first fully connected layer, a second fully connected layer, and a softmax layer connected sequentially. The first fully connected layer compresses the dimension of the joint feature vector and uses ReLU non-linear activation. The second fully connected layer maps the compressed features to 2D. The softmax layer normalizes the 2D values ​​into a probability distribution that sums to 1. The two components of this probability distribution are the data modality dynamic weight and the image modality dynamic weight.

5. The method according to claim 1, characterized in that, The generated data modality dynamic weights and image modality dynamic weights are used to weight the projected data feature vector and the projected image feature vector, respectively. The two weighted feature vectors are then concatenated to obtain a fused feature vector, which includes: The element-wise product of the dynamic weights of the data modal and the projected data feature vector is calculated to obtain the weighted data feature vector; the element-wise product of the dynamic weights of the image modal and the projected image feature vector is calculated to obtain the weighted image feature vector; the weighted data feature vector and the weighted image feature vector are concatenated along the feature channel dimension to generate the fused feature vector.

6. The method according to claim 1, characterized in that, Based on the fused feature vector, the mechanical properties are predicted, and the predicted yield strength of the alloy is obtained, including: The fused feature vector is input into a pre-constructed multilayer perceptron regression network to obtain a normalized predicted yield strength value. The multilayer perceptron regression network includes: a first linear layer, which maps the dimension of the fused feature vector to an intermediate dimension and connects it to a Dropout layer; a second linear layer, which further compresses the intermediate dimension and connects it to a Dropout layer; and a third linear layer, which outputs a normalized predicted yield strength value. The normalized predicted yield strength is denormalized to obtain the actual predicted yield strength.

7. The method according to claim 6, characterized in that, Before acquiring the chemical composition data and metallographic image data of the alloy, the method further includes a joint training step: Acquire labeled alloy sample data, divide the training set according to alloy type, and perform joint training. Use mean squared error as the loss function to compare the predicted yield strength value with the actual yield strength value. With minimizing the loss function as the objective, use an adaptive moment estimation optimizer to optimize the parameters. During the training process, the projection matrix in the linear projection transformation, all parameters in the weight generation network, and all parameters in the multilayer perceptron regression network are jointly learned and updated under the supervision of the loss function.

8. An alloy yield strength prediction device, characterized in that, The device includes: The acquisition module is used to acquire the chemical composition data and metallographic image data of the alloy; The extraction module is used to extract the data feature vector of the chemical composition data and the image feature vector of the metallographic image data, respectively. The mapping module is used to perform linear projection transformation on the data feature vector and the image feature vector respectively, mapping them to a pre-defined common semantic space, to obtain projected data feature vector and projected image feature vector with the same dimension. The generation module is used to concatenate the projected data feature vector and the projected image feature vector to obtain a joint feature vector; and to generate the data mode dynamic weights and image mode dynamic weights of the current alloy by performing global context modeling on the joint feature vector. The fusion module is used to use the generated data modality dynamic weights and image modality dynamic weights to weight the projected data feature vector and the projected image feature vector respectively, and then concatenate the two weighted feature vectors to obtain the fused feature vector. The prediction module is used to predict the mechanical properties based on the fused feature vector to obtain the predicted yield strength value of the alloy.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 7.

10. A computer device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 7.

Images