A metal mechanical property prediction method based on multi-region metallographic map fusion

Abstract

The application discloses a metal mechanical property prediction method based on multi-region metallographic image fusion, which comprises the following steps: S1, collecting N multi-region metallographic images of a metal sample in the same heat treatment / processing state, collecting along the radius interval from the center of the sample, and synchronously obtaining the real value of the tensile strength and other mechanical properties through tensile testing; S2, using OpenCV to batch grayscale the metallographic images, and then using a blank image with a pixel value of 0 to interval splice to form a fusion image; S3, constructing a convolutional neural network under a PyTorch framework, taking the fusion image as the input and the performance value as the output to divide the data set for training; S4, generating a fusion image for a sample to be tested by repeating the previous steps, inputting the model to obtain a predicted value, and evaluating the performance by a coefficient of determination and a Pearson correlation coefficient. The application comprehensively covers the microstructure distribution of the material, eliminates splicing artifacts, improves the prediction accuracy and stability, standardizes the process, is convenient for industrial deployment, and greatly reduces the material testing cost and the research and development cycle.

Description

Technical Field

[0001] This invention relates to the field of material property prediction technology, and specifically to a method for predicting the mechanical properties of metals based on multi-region metallographic image fusion. Background Technology

[0002] The macroscopic mechanical properties of metallic materials are determined by their microstructure. Establishing a reliable mapping model between metallographic images and mechanical properties is the core of achieving rapid material performance evaluation and intelligent design. Traditional methods rely on empirical formulas (such as the Hall-Petch relationship) and destructive tensile tests, which not only make it difficult to quantify the spatial inhomogeneity of the microstructure, but also make the results susceptible to human experience. Furthermore, these methods are time-consuming, costly, and fail to meet the development needs of high-throughput screening and intelligent manufacturing.

[0003] In recent years, deep learning technology has driven the development of microstructure-property mapping models, and some studies have attempted to use convolutional neural networks (CNNs) to directly predict mechanical properties from a single metallographic image. However, the microstructure of metallic materials generally exhibits spatial heterogeneity (such as grain size gradients and compositional segregation), and the representation range of a single field-of-view image is limited, resulting in poor model generalization ability and insufficient prediction stability.

[0004] To overcome the limitations of single images, some studies have attempted to stack or directly stitch multiple metallographic images before inputting them into the model. However, traditional stitching methods produce edge artifacts—random combinations of pixels from adjacent images form false features. During model training, these artifacts are easily misclassified as valid information related to mechanical properties, leading to the learning of irrelevant noise and further degrading prediction accuracy and stability. Therefore, how to eliminate edge artifact interference through multi-view metallographic image fusion and achieve high-precision, high-stability mechanical property prediction has become an urgent technical problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies, such as incomplete characterization by a single metallographic image and interference from artifacts at the stitching edges in model prediction. It provides a method for predicting the mechanical properties of metals based on multi-region metallographic image fusion, which improves prediction accuracy and stability through clear multi-region acquisition rules, interval stitching strategies, and standardized model training processes.

[0006] To achieve the above objectives, the solution of the present invention is as follows:

[0007] A method for predicting the mechanical properties of metals based on multi-region metallographic image fusion includes the following steps:

[0008] S1, Multi-region metallographic image acquisition: For standard metal samples under the same heat treatment or processing condition, metallographic images of different representative regions are acquired using an optical microscope with varying magnification. N images are acquired for each sample, where N is greater than 3, representing the metallographic regions. Figure 1 Metallography Figure 2 Metallographic image N is acquired by taking one metallographic image at intervals along the radial direction starting from the center region of the sample; at the same time, the mechanical property parameters of the sample are obtained through tensile testing.

[0009] S2, Image Fusion Preprocessing: First, OpenCV is used to perform batch grayscale processing on the acquired metallographic images. Then, for N metallographic images of the same sample, N-1 blank images with the same size as the metallographic images and a pixel value of 0 are created. Figure 1 ... Blank diagram N-1, according to "metallography" Figure 1 +blank Figure 1 Metallography Figure 2 +blank Figure 2 The images are stitched together horizontally in the order of "+......metallographic image N" to form a fused image;

[0010] S3, Building and Training the Prediction Model: Construct a convolutional neural network model based on the PyTorch framework, take the fused image as input and the corresponding mechanical performance parameters as output, form a dataset and divide it into training set and test set according to the proportion, train the model through the training set, and establish the mapping relationship between multi-region tissue features and mechanical performance.

[0011] S4, Performance Prediction Application: For the test sample, obtain multi-region metallographic images according to steps S1-S2, perform grayscale processing and interval stitching to generate a fused image; input the fused image into the trained model, and output the predicted mechanical properties of the test sample; use a test set to evaluate the model performance, and the evaluation indicators include the coefficient of determination and the Pearson correlation coefficient.

[0012] Furthermore, the mechanical property parameters include tensile strength, yield strength, and elongation.

[0013] Furthermore, in step S1, the metallographic structure observation is set to a magnification of 200x, and a metallographic image is acquired every 0.2 mm along the radial direction starting from the center region of the sample.

[0014] Furthermore, in step S1, the tensile test was performed using an Instron 3382 universal testing machine, and the test conditions were the same tensile temperature and strain rate; the metallographic image was acquired using a ZEISS-Image optical microscope.

[0015] Furthermore, in step S3, during training, each original metallographic image before splicing is subjected to real-time random image transformation. The transformation methods include horizontal flipping, vertical flipping, and random rotation with an amplitude of ±10°. The same random seed is used to transform the five metallographic images of the same sample.

[0016] Furthermore, in step S3, the ratio of the training set to the test set is 7:3.

[0017] Furthermore, in step S3, the convolutional neural network model includes a front-end adaptation layer, a backbone network, and a back-end regression head; the convolutional neural network model adapts the aspect ratio of the fused image through the front-end adaptation layer, extracts deep organizational features through the backbone network, and achieves accurate mapping of performance parameters through the back-end regression head; the training strategy improves the model's convergence speed and generalization ability through a combination of loss function, optimizer, and learning rate scheduling.

[0018] Furthermore, in step S3, the backbone network adopts EfficientNet; the backend regression head includes a global pooling layer and multiple fully connected layers.

[0019] Furthermore, in step S3, the model training uses the SmoothL1 Loss loss function and the AdamW optimizer, with an initial learning rate of 3e-4, or 3×10⁻⁴. -4 Where 'e' represents scientific notation, indicating a power of 10, with a weight decay of 1e-4, or 1 × 10⁻⁴. -4 The training process combines learning rate warm-up and cosine annealing scheduling, with a warm-up period of 5 training rounds, and also introduces an early stop mechanism.

[0020] Furthermore, in step S4, the model performance is evaluated using a test set, and evaluation metrics between predicted and actual values ​​are calculated, including the coefficient of determination and the Pearson correlation coefficient.

[0021] After adopting the above scheme, the gain effect of the present invention is as follows:

[0022] First, it fundamentally eliminates technical interference, avoiding the negative impact of edge artifacts generated by traditional stitching methods on machine learning model training, and effectively improves the robustness of the model. Second, it significantly enhances information representation capabilities, providing the model with more comprehensive material microstructure statistical information and spatial distribution information than a single image by fusing images from multiple regions. Third, it achieves structural guidance; the "pattern-blank-pattern" structure formed by interval stitching can provide the model with implicit spatial location encoding, promoting the model's learning of the correlation between the structure of different regions and the overall performance of the material. Fourth, this method is simple to implement and highly reliable, requiring no reliance on complex image fusion algorithms, with low computational overhead, stable and controllable process, and easy to deploy and apply quickly in industrial practice scenarios. Attached Figure Description

[0023] Figure 1 This is a flowchart of the present invention;

[0024] Figure 2 This is a schematic diagram of the multi-region metallographic image fusion of the present invention;

[0025] Figure 3 This is a schematic diagram of the convolutional neural network of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are preferred embodiments of the present invention and should not be considered as excluding other embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0027] This invention provides a method for predicting the mechanical properties of metals based on multi-region metallographic image fusion, such as... Figure 1 As shown, it includes the following steps:

[0028] S1, Multi-region metallographic image acquisition: For standard metal samples under the same heat treatment or processing condition, metallographic images of different representative regions are acquired using an optical microscope with varying magnification. N images are acquired for each sample, where N is greater than 3, representing the metallographic regions. Figure 1 Metallography Figure 2 Metallographic image N is acquired by taking one metallographic image at intervals along the radial direction starting from the center region of the sample; at the same time, the mechanical property parameters of the sample are obtained through tensile testing.

[0029] S2, Image Fusion Preprocessing: First, OpenCV is used to perform batch grayscale processing on the acquired metallographic images. Then, for N metallographic images of the same sample, N-1 blank images with the same size as the metallographic images and a pixel value of 0 are created. Figure 1 ... Blank diagram N-1, according to "metallography" Figure 1 +blank Figure 1 Metallography Figure 2 +blank Figure 2 The images are stitched together horizontally in the order of "+......metallographic image N" to form a fused image;

[0030] S3, Building and Training the Prediction Model: Constructing a convolutional neural network model based on the PyTorch framework, such as... Figure 3 As shown, Figure 3 The diagram below illustrates the convolutional neural network of this invention. The fused image is used as input and the corresponding mechanical performance parameters are used as output to form a dataset, which is divided into a training set and a test set according to a certain ratio. The model is trained through the training set to establish the mapping relationship between multi-regional tissue features and mechanical performance.

[0031] S4, Performance Prediction Application: For the test sample, obtain multi-region metallographic images according to steps S1-S2, perform grayscale processing and interval stitching to generate a fused image; input the fused image into the trained model, and output the predicted mechanical properties of the test sample; use a test set to evaluate the model performance, and the evaluation indicators include the coefficient of determination and the Pearson correlation coefficient. Specific Implementation Example 1:

[0033] S1, Multi-region metallographic image acquisition: Metal materials with the same heat treatment (e.g., solution treatment, aging treatment) or processing state (e.g., forging, rolling) are selected as test objects. In this embodiment, rare earth magnesium alloy is selected, with a mass fraction of Gd 9.55%, Y 3.28%, Zn 1.77%, Zr 0.34%, and Mg balance. The material is processed into a cylindrical standard tensile specimen of Φ10mm×50mm to ensure that the size and processing accuracy of all specimens are consistent, avoiding errors introduced by specimen differences; metallographic structure is observed using a ZEISS-Image optical microscope, with the magnification set to 200x. The acquisition method is to start from the center area of ​​the specimen and collect one metallographic image every 0.2mm along the radial direction, with N=5 images collected for each specimen (i.e., metallographic...). Figure 1 The metallographic figures (Figure 5) are used to avoid errors caused by the inhomogeneity of microstructure in different regions of the sample. They correspond to the center, 0.2 mm, 0.4 mm, 0.6 mm, and 0.8 mm radii, respectively, to fully cover the radial microstructure distribution of the sample. The tensile test was conducted using an Instron 3382 universal testing machine, and the test conditions were set to room temperature (25℃) to ensure that the test environment and loading rate of all samples were consistent. During the test, the tensile strength, yield strength, and elongation of each sample were recorded in real time through the data acquisition system of the testing machine, which served as label data for model training.

[0034] S2, Image Fusion Preprocessing: Here are 5 metallographic images from the same sample, all of the same size, each with dimensions H=512 pixels and W=512 pixels. The OpenCV library is called using Python. The input is a list containing 5 metallographic images [img1,img2,img3,img4,img5], each image being a NumPy array. The acquired color metallographic images are read in BGR mode. To avoid the influence of image color on the model, they are converted to single-channel grayscale images with a shape of (512,512). Finally, they are saved to a specified folder. All images are saved in PNG format to avoid image quality loss due to compression. For each sample's 5 metallographic images, 4 blank images are created using Python (blank). Figure 1(Refer to blank image 4). The size of the metallographic images is uniformly set to 512×512 pixels (this can be adjusted according to the microscope resolution; this embodiment uses this size to balance clarity and computational efficiency). The size of the blank image is exactly the same as that of the metallographic images, and all pixel values ​​are set to 0 to present pure black, ensuring that no additional pixel interference is introduced; Figure 2 As shown, according to "metallography" Figure 1 +blank Figure 1 Metallography Figure 2 +blank Figure 2 Metallography Figure 3 +blank Figure 3 The images are stitched horizontally in the order of "+metallographic image 4+blank image 4+metallographic image 5". The resulting fused image has a size of 512×(512×5+512×4)=512×4608 pixels and is saved for subsequent model training and prediction. This fused image can eliminate the edge interference generated when traditional images are directly stitched together.

[0035] S3, Building and Training the Prediction Model: Collect fused images of 100 standard metal specimens and their corresponding tensile strength, yield strength, and elongation parameters. After one-to-one mapping, form the total dataset. Randomly divide the dataset in a 7:3 ratio. Data from 70 specimens is used as the training set for model parameter learning, and data from 30 specimens is used as the test set for evaluating model generalization ability. During training, each original metallographic image before stitching undergoes real-time random image transformation. Transformation methods include horizontal flipping, vertical flipping, and random rotation within ±10°. Five metallographic images of the same specimen are transformed using the same random seed to ensure consistent transformation of microstructure characteristics in each region. To avoid disrupting the spatial distribution of the sample's microstructure, a front-end adaptation layer with stride-optimized convolutions was added, since the aspect ratio of the fused image was 9:1 (4608 pixels: 512 pixels). This involved using convolutions and pooling with varying strides to progressively compress the width dimension. Specifically, the convolution kernel (7,1) and stride (1,2) of the Convolutional layer (Conv2d) were used, and the pooling kernel (3,3) and stride (1,2) of the MaxPool2d layer were used. A backbone network, EfficientNet, was employed for deep feature extraction. Finally, global pooling and a multi-layer fully connected regression head were used to map the extracted global features into continuous predicted values ​​of mechanical performance parameters. The training process used the SmoothL1Loss loss function and the AdamW optimizer, with an initial learning rate of 3e-4, or 3×10⁻⁴. -4 Where 'e' represents scientific notation, indicating a power of 10, with a weight decay of 1e-4, or 1 × 10⁻⁴. -4 The optimization combines learning rate warmup and cosine annealing scheduling, with a warmup period of 5 training epochs, and introduces an early stopping mechanism to prevent overfitting.

[0036] Among them, learning rate preheating addresses the issues of gradient instability and weight oscillation in the early stages of training, while cosine annealing achieves precise convergence in the later stages by dynamically adjusting the learning rate. The learning rate update formula is as follows:

[0037] ,

[0038] in, η t Let be the learning rate at step t; η max This is the upper limit of the learning rate (initial learning rate); η min The lower bound of the learning rate (usually 1) η max / 100 or 1e-6); T t This represents the current number of training steps. T max This represents the total number of steps in one annealing cycle (the number of steps in one training round).

[0039] S4, Performance Prediction Application: For the test sample, obtain multi-region metallographic images according to the methods in steps S1-S2, and perform grayscale processing and interval stitching to generate a fused image; input the fused image into the trained model, and the model calculates through forward propagation and outputs three predicted values, which correspond to the tensile strength, yield strength and elongation of the test sample, respectively, to complete the performance prediction.

[0040] The model performance was evaluated using a test set. Evaluation metrics included the coefficient of determination (R²) and the Pearson correlation coefficient (R), as shown in the following formulas:

[0041] ;

[0042] Among them, y i This represents the true value of the i-th sample. Let R² represent the predicted value of the i-th sample. The closer R² is to 1, the better the model fit.

[0043] ;

[0044] in, This represents the actual measured value of the mechanical properties of the i-th metal sample. denoted as the average of the actual measured values ​​of the mechanical properties of all metal samples, and yi represents the predicted value of the mechanical properties of the i-th metal sample. R represents the average of the predicted mechanical properties of all metal samples, where R=1 indicates a perfect positive correlation. 1 indicates a completely negative correlation, and R=0 indicates no linear relationship.

[0045] This invention addresses the incomplete characterization of a single metallographic image through multi-region image fusion. Combined with an interval stitching strategy, it suppresses the interference of edge effects on model learning, thereby improving the accuracy and stability of predicting the mechanical properties of metals based on microstructure. Economically, by transforming a time-consuming and costly physical experiment into an instantaneous and low-cost image computation process, it reduces the cost of material preparation and mechanical testing, and shortens the development cycle of new materials. The application of this method can effectively achieve rapid evaluation of the mechanical properties of metallic materials, providing a theoretical basis for the optimized design of material microstructures. It has broad application prospects and significant practical implications in aerospace, automotive manufacturing, and other fields.

[0046] In the claims, description and accompanying drawings of this invention, the terms "comprising," "having," and variations thereof are used to mean "including but not limited to."

[0047] The above description is only a preferred embodiment of the present invention and is not intended to limit the design of this case. All equivalent changes made based on the key design features of this case shall fall within the protection scope of this case.

Claims

1. A method for predicting the mechanical properties of metals based on multi-region metallographic image fusion, comprising the following steps: S1, Multi-region metallographic image acquisition: For standard metal specimens under the same heat treatment or processing condition, metallographic images of different representative regions are acquired using an optical microscope with magnification of 200x. N images are acquired for each specimen, where N is greater than 3, namely metallographic image 1, metallographic image 2, ..., metallographic image N. The acquisition method is to start from the center region of the specimen and acquire one metallographic image every 0.2 mm along the radial direction. At the same time, the mechanical property parameters corresponding to the specimen are obtained through tensile testing. S2, Image Fusion Preprocessing: First, OpenCV is used to batch grayscale the acquired metallographic images. Then, for N metallographic images of the same sample, N-1 blank images with the same size as the metallographic images and a pixel value of 0 are created, namely Blank Image 1, ..., Blank Image N-1. Following the order of "Metallographic Image 1 + Blank Image 1 + Metallographic Image 2 + Blank Image 2 + ... Metallographic Image N", implicit spatial location encoding can be provided for the model. The images are stitched together in the horizontal direction to form a fused image, eliminating the edge interference generated when traditional images are directly stitched together. S3, Building and Training the Prediction Model: A convolutional neural network model based on the PyTorch framework is constructed. The fused image is used as input, and the corresponding mechanical performance parameters are used as output, forming a dataset that is proportionally divided into training and testing sets. The model is trained using the training set to establish a mapping relationship between multi-regional tissue features and mechanical performance. The convolutional neural network model adapts to the aspect ratio of the fused image through a front-end adaptation layer, extracts deep tissue features through the backbone network, and achieves accurate mapping of performance parameters through a back-end regression head. The training strategy improves the model's convergence speed and generalization ability through a combination of loss functions, optimizers, and learning rate scheduling. The convolutional neural network model includes a front-end adaptation layer, a backbone network, and a back-end regression head; the backbone network uses EfficientNet; the back-end regression head includes a global pooling layer and multiple fully connected layers. The model was trained using the SmoothL1Loss loss function and the AdamW optimizer, with an initial learning rate of 3e-4 and a weight decay of 1e-4. The training process combines learning rate warm-up and cosine annealing scheduling, with a warm-up period of 5 epochs, and also introduces an early stop mechanism; S4, Performance Prediction Application: For the test sample, obtain multi-region metallographic images according to steps S1-S2, perform grayscale processing and interval stitching to generate a fused image; input the fused image into the trained model, and output the predicted mechanical properties of the test sample; use a test set to evaluate the model performance, and the evaluation indicators include the coefficient of determination and the Pearson correlation coefficient.

2. The method for predicting the mechanical properties of metals based on multi-region metallographic image fusion according to claim 1, characterized in that, The mechanical properties include tensile strength, yield strength, and elongation.

3. The method for predicting the mechanical properties of metals based on multi-region metallographic image fusion according to claim 1, characterized in that, In step S1, the tensile test was performed using an Instron 3382 universal testing machine, and the test conditions were the same tensile temperature and strain rate; the metallographic images were acquired using a ZEISS-Image optical microscope.

4. The method for predicting the mechanical properties of metals based on multi-region metallographic image fusion according to claim 1, characterized in that, In step S3, during training, each original metallographic image before splicing is subjected to real-time random image transformation. The transformation methods include horizontal flipping, vertical flipping, and random rotation with an amplitude of ±10°. Furthermore, the same random seed is used to transform the five metallographic images of the same sample.

5. The method for predicting the mechanical properties of metals based on multi-region metallographic image fusion according to claim 1, characterized in that, In step S3, the ratio of the training set to the test set is 7:

3.

6. The method for predicting the mechanical properties of metals based on multi-region metallographic image fusion according to claim 1, characterized in that, In step S4, the performance of the final optimized model is evaluated using the test set, and evaluation metrics between predicted and true values ​​are calculated, including the coefficient of determination and the Pearson correlation coefficient.

Images