Mechanical property prediction deep learning method based on microstructure image representation

Abstract

The invention discloses a mechanical property prediction deep learning method based on microstructure image representation. The method comprises the following steps: 1, collecting a microstructure image and mechanical properties of a sample; 2, performing multi-scale cutting and image enhancement on the microscopic structure image to generate an extended data set; 3, inputting the enhanced image into a variational auto-encoder model for two-stage training; 4, coding the plurality of sub-images into potential vectors through a variational auto-encoder, and splicing the potential vectors to form a large-size microscopic structure image; 5, performing statistical quantization on the potential vectors to obtain statistical characteristics such as a mean value and a variance as input, and learning a mapping relation between the microscopic structure and the mechanical property by adopting a regression model; and 6, analyzing the correlation between the potential vector statistical characteristics and the grain size, orientation and the like, and identifying a key microscopic region. According to the mechanical property prediction deep learning method based on microstructure image representation, the problems that an existing method is prone to over-fitting, lacks interpretability and cannot reveal a microcosmic strengthening mechanism under the condition of data scarcity are solved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of materials science and artificial intelligence, and specifically relates to a deep learning method for predicting mechanical properties based on microstructure image representation. Background Technology

[0002] The microstructure of a material, such as grain size, grain boundaries, texture, precipitate distribution, and dislocation-related fields, determines its mechanical properties. Traditional physical or statistical methods require experts to manually extract these features. When dealing with complex, multi-scale, and multi-phase microstructures, the extraction process is time-consuming and often fails to fully capture the influence of multiple physical mechanisms.

[0003] In recent years, deep learning has shown outstanding performance in microscopic image analysis. Convolutional neural networks and other technologies can automatically extract image features and use them for downstream performance prediction. However, when the amount of experimental data is limited, end-to-end supervised training models can lead to overfitting and the resulting representations lack physical interpretability. To address these issues, this invention proposes a hybrid method combining image enhancement, self-supervised variational autoencoders, and regression models. This method improves the structure-performance prediction capability under small sample conditions by learning stable and physically relevant microstructural representations. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning method for predicting mechanical properties based on microstructure image representation. It proposes a deep learning method that combines image enhancement, self-supervised learning, and regression modeling to extract physically meaningful microstructure representations and establish a high-precision performance prediction model. This addresses the problems of existing methods, such as overfitting under data scarcity, lack of interpretability, and inability to reveal microstructure enhancement mechanisms.

[0005] The technical solution adopted in this invention is a deep learning method for predicting mechanical properties based on microscopic tissue image representation, as detailed below: S1. Collect microstructure images and corresponding mechanical property data of material samples; S2. Perform multi-scale cropping and image enhancement on the original large-size microstructure images of the material samples to generate an extended dataset; S3. Based on the extended dataset, the enhanced image is input into the variational autoencoder model for two-stage training of image reconstruction and restoration. S4. The multiple sub-images obtained by uniform cropping in S2 are encoded into latent vectors by variational autoencoders and then spliced ​​together to form a comprehensive representation of large-size microscopic tissue images. S5. Using statistical characteristics such as mean and variance obtained by statistical quantification of potential vectors as input, a regression model is used to learn the mapping relationship between microstructure and mechanical properties. S6. By analyzing the correlation between the statistical characteristics of latent vectors and microstructural characteristics such as grain size and orientation, and combining the backpropagation activation mapping and anomaly detection methods of latent vectors, key micro-regions related to high performance are identified and the physical mechanisms are verified.

[0006] The invention is further characterized by: The specific method for S1 is as follows: S1.1. The process parameters for heat treatment of alloy materials are controlled by orthogonal experimental methods. The process includes rolling thinning ratio, solution or aging temperature, and holding time. The process parameters include rolling thinning ratio of 0% to 50%, solution or aging temperature of 650℃ to 1050℃, and holding time of 1h to 20h. S1.2. Electron backscatter diffraction (EBSD) was used to obtain microstructure images of the material samples, with a scanning step size of 1 μm, an operating voltage of 15 kV, an image size of 800×1000 pixels, and the original .ang / .ctf or equivalent data, as well as PNG / TIFF format visualization images were saved. S1.3. Perform tensile testing on the samples according to ASTM E8, Standard Test Method for Tensile Testing of Metallic Materials. The gauge length is 205 mm, the width is 50 mm, the thickness is 2–6 mm, and the tensile rate is 0.002 s. -1 The yield strength of the alloy is extracted from the stress-strain curve and matched one-to-one with the microstructure image to form a data pair, thus constructing a material microstructure-property database.

[0007] The specific method for S2 is as follows: S2.1. Randomly crop each original 800×1000 pixel large-size microscopic tissue image into multi-scale small-size images, including 64×64, 128×128, 192×192, 256×256, and 320×320 pixels, and crop multiple small-size images for each size. S2.2. Perform two types of image enhancement on the cropped small-sized images: Perform type A data enhancement on the small-sized images, including horizontal flipping, vertical flipping, and rotation of 90º, 180º, and 270º, to obtain images that retain structural information; then further apply type B data enhancement to these images, including Gaussian noise, geometric distortion, random occlusion of 5% to 25% of the area, erosion, and dilation, to obtain images that lose structural information. The extended dataset is obtained through the above image data enhancement operations.

[0008] In S3, the self-supervised variational autoencoder model is trained in two stages: image reconstruction and restoration, using augmented images from the extended dataset. The first stage is based on reconstruction training using images that preserve structural information, and the second stage is based on restoration training using images that lose structural information.

[0009] The specific method for S3 is as follows: S3.1. A variational autoencoder model is constructed using multi-layer convolution and transposed convolution structures. Its latent dimension is optimized and adjusted between 16 and 512. The loss function... L It consists of the reconstruction error and the Kullback-Leibler divergence term:

[0010] in L MSE To mitigate reconstruction errors, ensure the generated Approximately the original input x, L KL Divergence term, constraining the latent distribution Approximate prior distribution ; S3.2 Image reconstruction training process of variational autoencoder model: The obtained image that preserves structural information is input into the variational autoencoder model for image reconstruction training, of which 80% is used for training and 20% is used for testing. S3.3 Image restoration training process of variational autoencoder model: For all images that have lost structural information, continue to train image restoration to recover structural information based on the model trained in S3.2, so as to enhance the image restoration ability of the model, while maintaining the same training / test set division.

[0011] The specific method for S4 is as follows: S4.1. Each large-size original microscopic tissue image of 800×1000 pixels is first uniformly cropped into 20 smaller images of 200×200 pixels each. The relative position information of each smaller image is recorded. Then, each smaller image is input into a trained variational autoencoder to obtain a low-dimensional latent vector. z 1... z 20 ; S4.2. Based on the recorded location information, sequentially sort all the latent vectors of the small-sized images. z 1... z 20 spliced ​​together z L The long vector is used to obtain a comprehensive characterization of large-size original microscopic tissue images.

[0012] The specific method for S5 is as follows: S5.1 To reduce the complexity of the regression model and improve its robustness, a comprehensive characterization based on large-size original microscopic tissue images is proposed. z L Calculate and use the following statistical characteristics:

[0013] in, µ L and These are the mean and variance of the latent vectors, respectively. µ L and Input into the random forest to predict the yield strength of the material; S5.2 Optimize key hyperparameters using grid search or random search: number of decision trees and maximum depth; use cross-validation or repeated random partitioning to evaluate the stability of model performance; use the coefficient of determination R0. 2 The root mean square error (RMSE) is used as a performance evaluation metric; the same training / test set partition is maintained as that used for the variational autoencoder pre-training to avoid data leakage.

[0014] The specific method for S6 is as follows: S6.1 Calculation µ L and Pearson correlation coefficient with traditional artificially quantified features such as average grain area, grain orientation distribution, and geometrically required dislocation density is used to determine whether a potential vector corresponds to a known mechanism. S6.2, Specific elements of the latent vector z i Calculate the inverse gradient and project it back onto the original image region to generate an activation map, showing different... z The component focuses on the microscopic region, identifying which regions contribute most to the latent representation, and the latent vector. z i The formula for calculating importance weight is:

[0015] in, The first convolutional layer represents the... k Each feature map at location ( u , v The activation value of ) for all positions ( u , v Sum the results and then divide by the number of pixels in the feature map. Z We obtain a global average sensitivity, and based on the calculated back gradient, we calculate the activation map using the following formula:

[0016] Activate each feature map f According to weight A linear combination ultimately yields It is the latent vector z i The spatial importance distribution, once visualized, reveals which regions of the input image the latent vector corresponds to. S6.3. Train the variational autoencoder with low-intensity samples with yield strength <500 MPa to obtain the latent vector. z x mean µ x σ x Then, the latent vectors of high-strength samples with yield strength > 800 MPa are extracted. z y mean µ y σ y Calculate its normal difference distribution:

[0017] In the formula, e and π are both constants; By distribution Medium sampling Z P It then backpropagates to generate a thermal activation map, locating microscopic regions associated with high intensity.

[0018] The beneficial effects of this invention are: (1) The mechanical property prediction deep learning method based on microstructure image representation in this invention significantly improves prediction accuracy: under limited experimental sample conditions, the coefficient of determination R is achieved on the test set. 2 The performance is improved from approximately 0.35 in a typical supervised training convolutional neural network to over 0.76, significantly reducing prediction error and enhancing model generalization performance; (2) The mechanical property prediction deep learning method based on microstructure image representation of the present invention has enhanced physical interpretability: the learned latent vector is significantly correlated with traditional microscopic measures such as grain size and orientation, and the statistics of the latent vector are power-law related to the yield strength, which can be interpreted as Hall-Petch-like reinforcement. (3) The mechanical property prediction deep learning method based on microstructure image representation of the present invention is small sample friendly: through multi-scale cropping and large-scale image enhancement, robust representation can still be obtained without the need for pre-training in large-scale external databases such as ImageNe, which is suitable for research or engineering scenarios where experimental material data is scarce. (4) The mechanical property prediction deep learning method based on microstructure image representation of the present invention is lightweight and highly robust: the combination model of variational autoencoder and random forest has fewer parameters, good training stability, and low computational resource requirements, thus reducing the risk of overfitting; (5) The mechanical property prediction deep learning method based on microstructure image representation of the present invention has scalability and versatility: the framework can be extended to the prediction of other properties such as hardness and fatigue life, and is also applicable to a variety of metal systems, such as nickel-based alloys, titanium alloys, high-entropy alloys, steel, etc. (6) The present invention is based on a deep learning method for predicting mechanical properties based on microstructure image representation, which supports reverse design and accelerates research and development: by optimizing and decoding the potential vector space, microstructure candidates that meet the target performance can be generated in reverse, providing a scientific basis for material and process design. Attached Figure Description

[0019] Figure 1 This is a flowchart of the deep learning method for predicting mechanical properties based on microstructure image representation, as described in this invention. Figure 2 This is a schematic diagram of the data augmentation used in the deep learning method of this invention; Figure 3 This is a graph showing the loss variation during the two-stage training process of the variational autoencoder model of this invention. Figure 4 This invention provides a long vector distribution map obtained by stitching together the distributions of multiple potential vectors of cropped images corresponding to the same sample. Figure 5 This is a graph showing the yield strength prediction results of the present invention based on the mean and variance of the latent vector as features and the random forest model; Figure 6 This is a graph showing the relationship between the potential microstructure representation extracted in this invention and the average grain area of ​​the alloy. Figure 7 This is a graph showing the relationship between the potential microstructure representation extracted in this invention and the yield strength; Figure 8 This is the thermal activation map obtained by backpropagation of latent vectors in this invention; Figure 9 This is a map of key microstructure regions related to high-performance materials obtained through anomaly detection in this invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0021] This invention relates to a deep learning method for predicting mechanical properties based on microstructure image representation. It combines image enhancement, self-supervised learning, and regression modeling to represent microstructure knowledge and predict performance. Figure 1 As shown, the details are as follows: S1. Acquire microscopic images of the material samples and their corresponding mechanical properties; the specific methods are as follows: S1.1. The heat treatment process of alloy materials is controlled by orthogonal experimental method. A dataset containing 25 samples is constructed. The process includes rolling thinning ratio, solution or aging temperature, and holding time. The rolling thinning ratio is 0% to 50%, the solution or aging temperature is 650℃ to 1050℃, and the holding time is 1h to 20h. The optimal value of each process parameter is obtained through orthogonal experiment and then adjusted within the above optimal range.

[0022] S1.2. Electron backscatter diffraction (EBSD) was used to obtain microstructure images of the material samples, with a scanning step size of 1 μm, an operating voltage of 15 kV, an image size of 800×1000 pixels, and the original .ang / .ctf or equivalent data, as well as PNG / TIFF format visualization images were saved. S1.3. Perform tensile testing on the samples according to ASTM E8, Standard Test Method for Tensile Testing of Metallic Materials. The gauge length is 205 mm, the width is 50 mm, the thickness is 2–6 mm, and the tensile rate is 0.002 s. -1 The yield strength of the alloy is extracted from the stress-strain curve and matched one-to-one with the microstructure image to form a data pair, thus constructing a material microstructure-property database.

[0023] S2. Perform multi-scale cropping and image enhancement on the original large-size microscopic images of the material samples to generate an extended dataset; the specific method is as follows: S2.1. Each original 800×1000 pixel large-size microscopic tissue image is randomly cropped into multi-scale small-size images, including 64×64, 128×128, 192×192, 256×256, and 320×320 pixels. 20 images are cropped from each size, and a total of 2500 small-size images are obtained from 25 sets of samples. S2.2. Perform two types of image enhancement on the cropped small-sized images: Type A data enhancement is performed on 2500 small-sized images, including horizontal flipping, vertical flipping, and rotation by 90º, 180º, and 270º, resulting in 12500 images that retain structural information; then, Type B data enhancement is further applied to these 12500 images, including Gaussian noise, geometric distortion, random occlusion of 5%–25% of the area, erosion, and dilation, resulting in 62500 images that lose structural information. The expanded dataset is obtained through these image data enhancement operations. For example... Figure 2 As shown, type A is a transformation that preserves structural information and type B is a transformation that loses structural information; S3. A self-supervised variational autoencoder model is used to perform two-stage image reconstruction and restoration training on the augmented images in the extended dataset. The first stage is based on reconstruction training using images that preserve structural information, and the second stage is based on restoration training using images that lose structural information. The specific method is as follows: S3.1. A variational autoencoder model is constructed using multi-layer convolution and transposed convolution structures. Its latent dimension is optimized and adjusted between 16 and 512. The loss function... L It consists of the reconstruction error and the Kullback-Leibler divergence term:

[0024] in L MSE To mitigate reconstruction errors, ensure the generated Approximately the original input x, L KL Divergence term, constraining the latent distribution Approximate prior distribution ; S3.2 Image reconstruction training process of variational autoencoder model: 12,500 images that retain structural information are input into the variational autoencoder model for image reconstruction training, of which 10,000 images are used for training and 2,500 images are used for testing. S3.3, Image Restoration Training Process of Variational Autoencoder Model: 62,500 images with lost structural information, such as... Figure 3 As shown, image reconstruction training to restore structural information is continued based on the model trained in S3.2 to enhance the model's image reconstruction ability, while maintaining the same training / test set partitioning.

[0025] like Figure 3 As shown, Figure 3 Phase 1 is the process of S3.2, and Phase 2 is the process of S3.3, indicating that the model loss gradually decreases in the two phases, and the model converges.

[0026] S4. The multiple sub-images obtained by uniform cropping in S2 are encoded into latent vectors using a variational autoencoder and then concatenated to form a comprehensive representation of the large-size microscopic tissue image; the specific method is as follows: S4.1. Each large-size original microscopic tissue image of 800×1000 pixels is first uniformly cropped into 20 smaller images of 200×200 pixels each. The relative position information of each smaller image is recorded. Then, each smaller image is input into a trained variational autoencoder to obtain a low-dimensional latent vector. z 1... z 20 ; S4.2. Based on the recorded location information, sequentially sort all the latent vectors of the small-sized images. z 1... z 20 spliced ​​together z L The long vector is used to obtain a comprehensive characterization of large-size original microscopic tissue images.

[0027] S5. Using statistical features such as mean and variance obtained from the statistical quantification of the latent vectors as input, a regression model is used to learn the mapping relationship between microstructure and mechanical properties; the specific method is as follows: S5.1 To reduce the complexity of the regression model and improve its robustness, a comprehensive characterization based on large-size original microscopic tissue images is proposed. z L Calculate and use the following statistical characteristics:

[0028] in, µ L and These are the mean and variance of the latent vectors, respectively. µ L and Input into the random forest to predict the yield strength of the material; like Figure 4 As shown, gray represents the small image vector zi, and black represents the concatenated long vector. z L Distribution, Explanation z i And the spliced z L They all conform to a normal distribution, and their mean and variance can be extracted as inputs to the performance prediction model.

[0029] S5.2 Optimize key hyperparameters using grid search or random search: number of decision trees and maximum depth; use cross-validation or repeated random partitioning to evaluate the stability of model performance; use the coefficient of determination R0. 2 The root mean square error (RMSE) is used as a performance evaluation metric; the same training / test set partition is maintained as that used for the variational autoencoder pre-training to avoid data leakage.

[0030] like Figure 5 The figure shows the performance prediction results and R of the optimal model selected in S5.2. 2 This indicates that the prediction was relatively accurate.

[0031] S6. By analyzing the correlation between latent vector statistical characteristics and microstructural features such as grain size and orientation, and combining latent vector backpropagation activation mapping and anomaly detection methods, key microscopic regions related to high performance are identified and the physical mechanisms are verified. The specific methods are as follows: S6.1 Calculation µ L and Pearson correlation coefficient with traditional artificially quantified features such as average grain area, grain orientation distribution, and geometrically required dislocation density is used to determine whether a potential vector corresponds to a known mechanism. like Figures 6-7 As shown, these are the 25 groups of samples in S6.1. z L The relationship between the vector and yield strength and grain size demonstrates that the characteristics of the potential vector are correlated with the physical characteristics of the material.

[0032] S6.2, Specific elements of the latent vector z i Calculate the inverse gradient and project it back onto the original image region to generate an activation map, showing different... z The component focuses on the microscopic region, identifying which regions contribute most to the latent representation, and the latent vector. z i The formula for calculating importance weight is:

[0033] in, The first convolutional layer represents the... k Each feature map at location ( u , v The activation value of ) for all positions ( u , v Sum the results and then divide by the number of pixels in the feature map. Z We obtain a global average sensitivity, and based on the calculated back gradient, we calculate the activation map using the following formula:

[0034] like Figure 8 As shown, this is the activation mapping region of a single vector dimension in S6.2, which activates each feature map. f According to weight A linear combination ultimately yields It is the latent vector z i The spatial importance distribution, once visualized, reveals which regions of the input image the latent vector corresponds to. S6.3. Train the variational autoencoder with low-intensity samples with yield strength <500 MPa to obtain the latent vector. z x mean µ x σ x Then, the latent vectors of high-strength samples with yield strength > 800 MPa are extracted. z y mean µ y σ y Calculate its normal difference distribution:

[0035] In the formula, e and π are both constants; By distribution Medium sampling Z P And it backpropagates to generate a thermal activation map, locating microscopic regions associated with high intensity. For example... Figure 9 As shown, S6.3 is the result of anomaly detection, and the activated region represents the area that has a significant impact on performance.

[0036] Example 1 The research object is Inconel 625 alloy sheet annealed in the rolled state, and the main components are shown in Table 1.

[0037] Table 1. Elemental composition and content of Inconel 625 alloy

[0038] Based on S1, a database containing a wide range of performance characteristics and microstructural variations was constructed. Processing routes with different process parameters were designed, including hot rolling deformation at 900 ℃, aging temperature, and aging time. The initial hot rolling deformation amounts were 0%, 20%, 30%, 40%, and 50%. Subsequently, the rolled plates underwent aging treatment at temperatures of 650 ℃, 750 ℃, 850 ℃, 950 ℃, and 1050 ℃, with aging times of 1 hour, 5 hours, 10 hours, 15 hours, and 20 hours. Finally, 25 different processing routes were designed using an orthogonal experimental method, as shown in Table 2.

[0039] Table 2. 25 different processing routes and their yield strengths based on orthogonal experiments.

[0040] To obtain comprehensive microstructural information, EBSD characterization was performed using a Gemini 500 scanning electron microscope, generating 1000×800 pix Euler angle diagrams to describe the different grain sizes and orientations. The EBSD scan step size was set to 0.1 µm, consistent across all samples. Tensile properties were tested on an INSTRON universal testing machine. Test samples were prepared in the rolling direction according to ASTM E8 standards, with a scale length of 205 mm, a width of 59 mm, and a thickness of 2–6 mm. Tensile tests were conducted at a strain rate of 0.002 s⁻¹. The testing procedures were performed according to the procedures specified in GB / T 228-2002 and GB / T 4338-2006.

[0041] From 25 original EBSD images of 800×1000 pixels, 20 images of sizes 64×64, 128×128, 192×192, 256×256, and 320×320 were randomly cropped, generating a total of 2500 images. Two different types of data augmentation strategies were then applied to the cropped images, which were then input into a variational autoencoder for two-stage model training. Stage 1: 12500 new images were generated by flipping and rotating, preserving information from the original images. These images were used for reconstruction training, and the model parameters were retained after training. Stage 2: The augmented images from Stage 1 underwent further transformations, such as Gaussian blur, distortion, and the addition of black blocks, resulting in 62500 images. These images suffered information loss and were then input into the variational autoencoder for image restoration training. The model training results based on these two stages are as follows: Figure 3 As shown, each stage divides the dataset into an 80% training set and a 20% test set. The training set is used for model training and hyperparameter optimization, while the test set is used to evaluate the performance of the final model.

[0042] To predict the yield strength of the material, we first divide the original 800×1000 pixel image into 20 smaller 200×200 images. These smaller images are then input into a variational autoencoder model, yielding 20 128-dimensional latent vectors for each original large image. z 1, z 2, ..., z 20 They follow a normal distribution. Then, the 20 latent variables, each with 128 dimensions, are concatenated into a long variable of 2560 (128 × 20) dimensions. z .like Figure 4 As shown, based on the characteristics of normally distributed data, the resulting concatenated vectors also follow a normal distribution, i.e. z ~ N ( µ , σ 2 ),in , We utilize µ and σ 2 These two parameters are used as inputs to a random forest model to predict the yield strength of materials, such as Figure 5 The model accurately predicts the yield strength of materials, and its R-value on the test set is shown. 2 It reached 0.82.

[0043] Interpretability analysis aims to provide a deeper understanding of the model to guide materials design. To further understand the latent variables... µ L and We analyze the Pearson correlation (PCC) between the values ​​and the material's grain size and yield strength. For example... Figures 6-7 As shown, It shows a high positive correlation with grain area (PCC=0.87) and a negative correlation with yield strength (PCC=-0.69). Furthermore, the variance... A clear Hall-Petch relationship exists between the latent variables and the material's yield strength. Furthermore, using interpretable algorithms, we investigated which image information different dimensions of different latent variables can capture.

[0044] Specifically, we examined two dimensions. z 1 and z 2. And their representation in different samples. For example... Figure 8 The three samples shown were randomly selected from the dataset, among which... z The activation region of 1 corresponds to the reddish grains in the microstructure image. z 2 corresponds to a bluish grain. Specifically, z 1 can represent red, orange-red, and pink grain regions, with the pink grain region attracting the most attention. z The blue grain region is of most interest, followed by the green region. This observation indicates that latent vectors of different dimensions capture grain regions with different orientations. Therefore, the entire latent variable can represent the complete information of the microscopic image, demonstrating the good representational ability of the variational autoencoder model.

[0045] Further analysis of key structural regions affecting material properties was conducted, and a variational autoencoder model was trained using microstructure images of alloys with yield strengths below 500 MPa. The encoder extracted latent variables from all training data, and the overall mean and variance of the latent space for the low-performance alloys were calculated. µ x and σ x For microstructure images of alloys with yield strength exceeding 800 MPa, these are considered anomalous data. Their latent vectors are extracted using a trained encoder, and the normal distribution difference is calculated using formula (5). Then, an anomaly detection attention map is generated after resampling and backpropagation. Figure 9 As shown, the prominent activation regions represent the differences between high-strength and low-performance alloy microstructures. These regions can be intuitively understood as key areas that significantly improve material properties. Clearly, the model gives greater attention to the large number of small, disordered grain regions, suggesting that the alloy's high performance is attributed to the presence of more small grains. It can be understood that these grains effectively reduce the average grain size, thereby improving material properties through the Hall-Petch effect.

[0046] Example 2 This invention relates to a deep learning method for predicting mechanical properties based on microstructure image representation. It combines image enhancement, self-supervised learning, and regression modeling to represent microstructure knowledge and predict performance. Figure 1 As shown, the details are as follows: S1. Collect microstructure images and corresponding mechanical property data of material samples; S2. Perform multi-scale cropping and image enhancement on the original large-size microstructure images of the material samples to generate an extended dataset; S3. A self-supervised variational autoencoder model is used to perform two-stage image reconstruction and restoration training on the augmented images in the extended dataset. The first stage is based on reconstruction training based on images that preserve structural information, and the second stage is based on restoration training based on images that lose structural information. S4. The multiple sub-images obtained by uniform cropping in S2 are encoded into latent vectors by variational autoencoders and then spliced ​​together to form a comprehensive representation of large-size microscopic tissue images. S5. Using statistical characteristics such as mean and variance obtained by statistical quantification of potential vectors as input, a regression model is used to learn the mapping relationship between microstructure and mechanical properties. S6. By analyzing the correlation between the statistical characteristics of latent vectors and microstructural characteristics such as grain size and orientation, and combining the backpropagation activation mapping and anomaly detection methods of latent vectors, key micro-regions related to high performance are identified and the physical mechanisms are verified.

[0047] Example 3 This embodiment is based on a deep learning method for predicting mechanical properties using microstructure image representation. Building upon Embodiment 2, the S1 method is further detailed as follows: S1. Acquire microscopic images of the material samples and their corresponding mechanical properties; the specific methods are as follows: S1.1. The heat treatment process of alloy materials is controlled by orthogonal experimental method. A dataset containing 25 samples is constructed. The process includes rolling thinning ratio, solution or aging temperature, and holding time. The rolling thinning ratio is 0% to 50%, the solution or aging temperature is 650℃ to 1050℃, and the holding time is 1h to 20h. The optimal value of each process parameter is obtained through orthogonal experiment and then adjusted within the above optimal range.

[0048] S1.2. Electron backscatter diffraction (EBSD) was used to obtain microstructure images of the material samples, with a scanning step size of 1 μm, an operating voltage of 15 kV, an image size of 800×1000 pixels, and the original .ang / .ctf or equivalent data, as well as PNG / TIFF format visualization images were saved. S1.3. Perform tensile testing on the samples according to ASTM E8, Standard Test Method for Tensile Testing of Metallic Materials. The gauge length is 205 mm, the width is 50 mm, the thickness is 2–6 mm, and the tensile rate is 0.002 s. -1 The yield strength of the alloy is extracted from the stress-strain curve and matched one-to-one with the microstructure image to form a data pair, thus constructing a material microstructure-property database.

[0049] Example 4 This embodiment is based on a deep learning method for predicting mechanical properties using microstructure image representation. Building upon Embodiment 2, the specific method in S2 is as follows: S2.1. Each original 800×1000 pixel large-size microscopic tissue image is randomly cropped into multi-scale small-size images, including 64×64, 128×128, 192×192, 256×256, and 320×320 pixels. 20 images are cropped from each size, and a total of 2500 small-size images are obtained from 25 sets of samples. S2.2. Perform two types of image enhancement on the cropped small-sized images: Type A data enhancement is performed on 2500 small-sized images, including horizontal flipping, vertical flipping, and rotation by 90º, 180º, and 270º, resulting in 12500 images that retain structural information; then, Type B data enhancement is further applied to these 12500 images, including Gaussian noise, geometric distortion, random occlusion of 5%–25% of the area, erosion, and dilation, resulting in 62500 images that lose structural information. The expanded dataset is obtained through these image data enhancement operations. For example... Figure 2 As shown, type A is a transformation that preserves structural information and type B is a transformation that loses structural information.

[0050] Example 5 This embodiment is based on a deep learning method for predicting mechanical properties using microstructure image representation. Building upon Embodiment 2, the specific method in S3 is as follows: S3.1. A variational autoencoder model is constructed using multi-layer convolution and transposed convolution structures. Its latent dimension is optimized and adjusted between 16 and 512. The loss function... L It consists of the reconstruction error and the Kullback-Leibler divergence term:

[0051] in L MSE To mitigate reconstruction errors, ensure the generated Approximately the original input x, L KL Divergence term, constraining the latent distribution Approximate prior distribution ; S3.2 Image reconstruction training process of variational autoencoder model: 12,500 images that retain structural information are input into the variational autoencoder model for image reconstruction training, of which 10,000 images are used for training and 2,500 images are used for testing. S3.3, Image Restoration Training Process of Variational Autoencoder Model: 62,500 images with lost structural information, such as... Figure 3 As shown, image reconstruction training to restore structural information is continued based on the model trained in S3.2 to enhance the model's image reconstruction ability, while maintaining the same training / test set partitioning.

[0052] Example 6 This embodiment is based on a deep learning method for predicting mechanical properties using microstructure image representation. Building upon Embodiment 2, the specific method in S4 is as follows: S4.1. Each large-size original microscopic tissue image of 800×1000 pixels is first uniformly cropped into 20 smaller images of 200×200 pixels each. The relative position information of each smaller image is recorded. Then, each smaller image is input into a trained variational autoencoder to obtain a low-dimensional latent vector. z 1... z 20 ; S4.2. Based on the recorded location information, sequentially sort all the latent vectors of the small-sized images. z 1... z 20 spliced ​​together z L The long vector is used to obtain a comprehensive characterization of large-size original microscopic tissue images.

[0053] Example 7 This embodiment is based on a deep learning method for predicting mechanical properties using microstructure image representation. Building upon Embodiment 2, the specific method in S5 is as follows: S5.1 To reduce the complexity of the regression model and improve its robustness, a comprehensive characterization based on large-size original microscopic tissue images is proposed. z L Calculate and use the following statistical characteristics:

[0054] in, µ L and These are the mean and variance of the latent vectors, respectively. µ L and Input into the random forest to predict the yield strength of the material; S5.2 Optimize key hyperparameters using grid search or random search: number of decision trees and maximum depth; use cross-validation or repeated random partitioning to evaluate the stability of model performance; use the coefficient of determination R0. 2 Root mean square error (RMSE) is used as a performance evaluation metric; the same training / test set partitioning as the variational autoencoder pre-training is maintained to avoid data leakage; Example 8 This embodiment is based on a deep learning method for predicting mechanical properties using microstructure image representation. Building upon Embodiment 2, the specific method in S6 is as follows: S6.1 Calculation µ L and Pearson correlation coefficient with traditional artificially quantified features such as average grain area, grain orientation distribution, and geometrically required dislocation density is used to determine whether a potential vector corresponds to a known mechanism. S6.2, Specific elements of the latent vector z i Calculate the inverse gradient and project it back onto the original image region to generate an activation map, showing different... z The component focuses on the microscopic region, identifying which regions contribute most to the latent representation, and the latent vector. z i The formula for calculating importance weight is:

[0055] in, The first convolutional layer represents the... k Each feature map at location ( u , v The activation value of ) for all positions ( u , v Sum the results and then divide by the number of pixels in the feature map. Z We obtain a global average sensitivity, and based on the calculated back gradient, we calculate the activation map using the following formula:

[0056] S6.3. Train the variational autoencoder with low-intensity samples with yield strength <500 MPa to obtain the latent vector. z x mean µ x σ x Then, the latent vectors of high-strength samples with yield strength > 800 MPa are extracted. z y mean µ y σ y Calculate its normal difference distribution:

[0057] In the formula, e and π are both constants; By distribution Medium sampling Z P It then backpropagates to generate a thermal activation map, locating microscopic regions associated with high intensity.

Claims

1. A deep learning method for predicting mechanical properties based on microstructure image representation, characterized in that, Specifically as follows: S1. Collect microstructure images and corresponding mechanical property data of material samples; S2. Perform multi-scale cropping and image enhancement on the original large-size microstructure images of the material samples to generate an extended dataset; S3. Based on the extended dataset, the enhanced image is input into the variational autoencoder model for two-stage training of image reconstruction and restoration. S4. The multiple sub-images obtained by uniform cropping in S2 are encoded into latent vectors by variational autoencoders and then spliced ​​together to form a comprehensive representation of large-size microscopic tissue images. S5. Using statistical characteristics such as mean and variance obtained by statistical quantification of potential vectors as input, a regression model is used to learn the mapping relationship between microstructure and mechanical properties. S6. By analyzing the correlation between the statistical characteristics of latent vectors and microstructural characteristics such as grain size and orientation, and combining the backpropagation activation mapping and anomaly detection methods of latent vectors, key micro-regions related to high performance are identified and the physical mechanisms are verified.

2. The deep learning method for predicting mechanical properties based on microstructure image representation according to claim 1, characterized in that, The specific method of S1 is as follows: S1.1 The process parameters for heat treatment of alloy materials are controlled by orthogonal experimental methods. The process includes rolling thinning ratio, solution or aging temperature, and holding time. The process parameters include a rolling thinning ratio of 0% to 50%, a solution or aging temperature of 650℃ to 1050℃, and a holding time of 1h to 20h. S1.

2. Electron backscatter diffraction (EBSD) was used to obtain microstructure images of the material samples, with a scanning step size of 1 μm, an operating voltage of 15 kV, an image size of 800×1000 pixels, and the original .ang / .ctf or equivalent data, as well as PNG / TIFF format visualization images were saved. S1.

3. Perform tensile testing on the samples according to ASTM E8, Standard Test Method for Tensile Testing of Metallic Materials. The gauge length is 205 mm, the width is 50 mm, the thickness is 2–6 mm, and the tensile rate is 0.002 s. -1 The yield strength of the alloy is extracted from the stress-strain curve and matched one-to-one with the microstructure image to form a data pair, thus constructing a material microstructure-property database.

3. The deep learning method for predicting mechanical properties based on microstructure image representation according to claim 1, characterized in that, The specific method of S2 is as follows: S2.

1. Randomly crop each original 800×1000 pixel large-size microscopic tissue image into multi-scale small-size images, including 64×64, 128×128, 192×192, 256×256, and 320×320 pixels, and crop multiple small-size images for each size. S2.

2. Perform two types of image enhancement on the cropped small-sized images: Perform type A data enhancement on the small-sized images, including horizontal flipping, vertical flipping, and rotation of 90º, 180º, and 270º, to obtain images that retain structural information; then further apply type B data enhancement to these images, including Gaussian noise, geometric distortion, random occlusion of 5% to 25% of the area, erosion, and dilation, to obtain images that lose structural information. The extended dataset is obtained through the above image data enhancement operations.

4. The deep learning method for predicting mechanical properties based on microstructure image representation according to claim 1, characterized in that, In S3, the self-supervised variational autoencoder model is trained in two stages: image reconstruction and restoration, using enhanced images from the extended dataset. The first stage is based on reconstruction training using images that preserve structural information, and the second stage is based on restoration training using images that lose structural information.

5. The deep learning method for predicting mechanical properties based on microstructure image representation according to claim 4, characterized in that, The specific method of S3 is as follows: S3.

1. A variational autoencoder model is constructed using multi-layer convolution and transposed convolution structures. Its latent dimension is optimized and adjusted between 16 and 512. The loss function... L It consists of the reconstruction error and the Kullback-Leibler divergence term: in L MSE To mitigate reconstruction errors, ensure the generated Approximately the original input x, L KL Divergence term, constraining the latent distribution Approximate prior distribution ; S3.2 Image reconstruction training process of variational autoencoder model: The obtained image that preserves structural information is input into the variational autoencoder model for image reconstruction training, of which 80% is used for training and 20% is used for testing. S3.3 Image restoration training process of variational autoencoder model: For all images that have lost structural information, continue to train image restoration to recover structural information based on the model trained in S3.2, so as to enhance the image restoration ability of the model, while maintaining the same training / test set division.

6. The deep learning method for predicting mechanical properties based on microstructure image representation according to claim 1, characterized in that, The specific method of S4 is as follows: S4.

1. Each large-size original microscopic tissue image of 800×1000 pixels is first uniformly cropped into 20 smaller images of 200×200 pixels each. The relative position information of each smaller image is recorded. Then, each smaller image is input into a trained variational autoencoder to obtain a low-dimensional latent vector. z 1... z 20 ; S4.

2. Based on the recorded location information, sequentially sort all the latent vectors of the small-sized images. z 1... z 20 spliced ​​together z L The long vector is used to obtain a comprehensive characterization of large-size original microscopic tissue images.

7. The deep learning method for predicting mechanical properties based on microstructure image representation according to claim 1, characterized in that, The specific method of S5 is as follows: S5.1 To reduce the complexity of the regression model and improve its robustness, a comprehensive characterization based on large-size original microscopic tissue images is proposed. z L Calculate and use the following statistical characteristics: in, µ L and These are the mean and variance of the latent vectors, respectively. µ L and Input into the random forest to predict the yield strength of the material; S5.2 Optimize key hyperparameters using grid search or random search: number of decision trees and maximum depth; use cross-validation or repeated random partitioning to evaluate the stability of model performance; use the coefficient of determination R0. 2 The root mean square error (RMSE) is used as a performance evaluation metric; the same training / test set partition is maintained as that used for the variational autoencoder pre-training to avoid data leakage.

8. The deep learning method for predicting mechanical properties based on microstructure image representation according to claim 1, characterized in that, The specific method of S6 is as follows: S6.1 Calculation µ L and Pearson correlation coefficient with traditional artificially quantified features such as average grain area, grain orientation distribution, and geometrically required dislocation density is used to determine whether a potential vector corresponds to a known mechanism. S6.2, Specific elements of the latent vector z i Calculate the inverse gradient and project it back onto the original image region to generate an activation map, showing different... z The component focuses on the microscopic region, identifying which regions contribute most to the latent representation, and the latent vector. z i The formula for calculating importance weight is: in, The first convolutional layer represents the... k Each feature map at location ( u , v The activation value of ) for all positions ( u , v Sum the results and then divide by the number of pixels in the feature map. Z We obtain a global average sensitivity, and based on the calculated back gradient, we calculate the activation map using the following formula: Activate each feature map f According to weight A linear combination ultimately yields It is the latent vector z i The spatial importance distribution, once visualized, reveals which regions of the input image the latent vector corresponds to. S6.

3. Train the variational autoencoder with low-intensity samples with yield strength <500 MPa to obtain the latent vector. z x mean µ x σ x Then, the latent vectors of high-strength samples with yield strength > 800 MPa are extracted. z y mean µ y σ y Calculate its normal difference distribution: In the formula, e and π are both constants; By distribution Medium sampling Z P It then backpropagates to generate a thermal activation map, locating microscopic regions associated with high intensity.

Images