Amorphous forming ability prediction method based on two-dimensional grayscale image

Abstract

The invention discloses an amorphous forming ability prediction method based on a two-dimensional gray level image. The method comprises the following steps: acquiring a component element proportion of an amorphous alloy of a metal alloy; performing gray scale image processing on the component element proportion to generate a combined two-dimensional gray scale image; an amorphous forming ability prediction model is loaded, the amorphous forming ability prediction model is constructed and generated after being trained by an amorphous forming ability attribute data set, and the amorphous forming ability attribute data set comprises a two-dimensional gray scale image corresponding to the proportion of component elements of the amorphous alloy and a critical casting diameter # imgabs0 #; and outputting the predicted critical casting diameter # imgabs1 # information according to the combined two-dimensional grayscale image. According to the technical scheme, the GFA of the alloy can be accurately predicted without additionally calculating other parameters, the area containing the high GFA alloy and the area containing the low GFA alloy are predicted, dependence on feature engineering is greatly reduced, and the needed time cost and economic cost can be greatly reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of prediction of amorphous forming ability of amorphous alloys, and in particular to a method for predicting amorphous forming ability based on two-dimensional grayscale images. Background Art

[0002] As an advanced metal material system, amorphous alloys have irreplaceable advantages in high-tech industries such as precision instruments and electronic information due to their high strength, high corrosion resistance and unique physical properties. The glass forming ability (GFA) directly determines the feasibility of materials in engineering manufacturing. The stronger the index is, the alloy can form a stable amorphous phase at a lower cooling rate, thereby realizing the preparation of large-scale components. However, existing amorphous alloys generally have the defect of insufficient forming ability, which seriously restricts their large-scale application in industrial scenarios. Accurately evaluating the glass forming ability is not only a key prerequisite for material research and development, but also an important breakthrough in deepening the understanding of the structure of amorphous materials.

[0003] At present, the amorphous forming ability is mainly quantified by the critical cooling rate and the critical casting diameter. However, the critical cooling rate scheme has the technical bottleneck of difficult experimental determination. Although the academic community has proposed a method based on the glass transition temperature ( )、Supercooled liquid region( ) and other thermodynamic parameters, these methods are only applicable to specific alloy systems and have limited prediction accuracy; the existing evaluation model predicts by establishing a correlation between basic physical parameters and amorphous forming ability, but due to the significant differences in elemental composition of different alloy systems, there is obvious heterogeneity in the mapping relationship between physical parameters and amorphous forming ability. In addition, the uneven distribution of experimental data leads to insufficient generalization of the model, making it difficult to achieve universal evaluation of multi-component alloy systems.

[0004] Therefore, it is urgent to build a new evaluation framework that integrates multidimensional physical property parameters, break through the existing technical bottleneck by strengthening feature correlation and model adaptability, and realize the prediction of amorphous forming ability. Summary of the invention

[0005] To achieve the above object, the present application provides a method for predicting amorphous forming ability based on a two-dimensional grayscale image, comprising the following steps: Obtaining the ratio of constituent elements of the metal alloy; the ratio of constituent elements is the atomic percentage of each constituent element in the alloy, and satisfies the constraint that the total is 100%; Performing grayscale image processing on the ratio of the constituent elements to generate a combined two-dimensional grayscale image; the grayscale image processing refers to: converting the one-dimensional vector of the constituent elements into a two-dimensional grayscale image by an outer product conversion method; Loading a glass-forming ability prediction model, the glass-forming ability prediction model is constructed and generated after training with a glass-forming ability attribute data set, the glass-forming ability attribute data set includes a two-dimensional grayscale image corresponding to the proportion of constituent elements of the amorphous alloy and a critical casting diameter ; Based on the combined two-dimensional grayscale image, the predicted critical casting diameter is output information.

[0006] Among them, grayscale image processing includes: Normalize the proportion of the constituent elements so that they are linearly mapped to the interval [0, 1]; The normalized ratio value is expressed as a k-dimensional vector, which is expressed as: ,in, is the normalized value of the composition element ratio, k is a natural number, and x 1 arrive x k The atomic numbers of elements gradually increase; Will and The transposed matrix of Perform outer product operation to generate k×k two-dimensional matrix S, which is expressed as: The two-dimensional matrix S is converted into a grayscale image, and each pixel value of the grayscale image is the product of two element ratio values.

[0007] Converting the two-dimensional matrix S into a grayscale image includes: Normalize the elements of the two-dimensional matrix S; After normalization, each element value is multiplied by 255 to generate a grayscale image.

[0008] Among them, the normalization of the elements of the two-dimensional matrix S is implemented using the Min-Max method, which is expressed as: ,in, is the minimum value among the elements of S, is the maximum value among the elements of S, is the normalized result.

[0009] Furthermore, before loading the amorphous forming ability prediction model, constructing the amorphous forming ability prediction model includes: Prepare a model training data set; wherein the model training data set is composed of an amorphous forming ability attribute data set and an extended data set; the amorphous forming ability attribute data set is a two-dimensional grayscale image generated by the proportion of constituent elements of the amorphous alloy and the corresponding critical casting diameter The amorphous forming ability attribute data set is expanded to simulate the transition state of different alloy components to form an extended data set; Define a deep learning model; the deep learning model is an improved ResNet18 model; The deep learning model is trained and verified through the model training data set to build an amorphous forming ability evaluation model.

[0010] The amorphous forming ability attribute data set is expanded to include: Use Mixup strategy to linearly mix two random sample data , The composition element ratios and their corresponding labels are used to generate new samples simulating the transition states of different alloy compositions. , where m and n are the serial numbers of random samples; When the Gaussian noise strategy is adopted, random noise that obeys the Gaussian distribution is added to the new sample, and the mean value of the noise is 0.

[0011] Among them, the new sample The generation method is: ; in, and are two random samples, is a random number drawn from the Beta distribution Beta(a,a) that satisfies .

[0012] Furthermore, the new sample right The ratio of the constituent elements Perform validity constraints, including: .

[0013] Furthermore, the structure of the deep learning model includes: Initial convolutional layer with a stride of 2 The maximum pooling layer and the global average pooling layer, the InceptionBlock module, the fully connected layer, the ReLU activation and the Dropout layer; and the deep learning model includes four main paths and one bypass. After the input features pass through multiple main paths in parallel, the outputs of each main path are connected in the channel dimension, and the bypass passes the input to the output for residual connection; if the output dimensions of the connected main path and the bypass are not equal, the output of the connected main path is passed through The number of convolution control channels matches the number of output channels of the bypass and then a residual connection is performed; the CNN architecture and convolution kernel size of each main path are different.

[0014] According to the present invention, the element interaction is characterized by outer product imaging, combined with parallel multi-scale convolution and data enhancement strategies, accurate prediction of amorphous forming ability can be achieved without artificial feature engineering; through the prediction model provided by the present invention, areas containing higher GFA alloys and areas containing lower GFA alloys can be predicted, which is convenient for researchers to start experiments from areas where higher GFA alloys may exist and avoid experiments in areas containing lower GFA alloys, and the GFA of the alloy can be predicted without the need for additional calculation of other parameters, which greatly reduces the dependence on feature engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a schematic diagram of the steps of a method for predicting amorphous forming ability based on a two-dimensional grayscale image according to an embodiment of the present invention; Figure 2 is a schematic diagram of grayscale image processing steps provided according to an embodiment of the present invention; Figure 3 is a schematic diagram of the structure of InceptionBlock in the amorphous forming ability prediction model provided in an embodiment of the present invention; Figure 4 is a schematic diagram of the structure of a prediction model for amorphous forming ability provided according to an embodiment of the present invention; Figure 5 is a comparison chart of the prediction results of the amorphous forming ability prediction model provided by an embodiment of the present invention and other schemes; Figure 6 It is a ternary phase diagram predicted for the Mg-Cu-Gd and Zr-Al-Cu systems provided according to the embodiments of the present invention. DETAILED DESCRIPTION

[0016] The present invention aims at the complex relationship between the constituent element properties of amorphous alloys and their critical casting diameter, and constructs an efficient amorphous forming ability prediction framework through multi-model integration and data enhancement strategy. The specific implementation method is described in detail below in conjunction with the accompanying drawings of the specification.

[0017] The method for predicting the amorphous forming ability provided by the present invention is as follows Figure 1 As shown, the following steps are included: Step S100: obtaining the ratio of constituent elements of the metal alloy; In the present invention, the critical casting diameter of the metal alloy can be predicted according to the ratio of the constituent elements of the metal alloy. The effective ratio of the constituent elements is the atomic percentage of each constituent element in the alloy, and satisfies the constraint that the total is 100%.

[0018] Step S110: performing grayscale image processing on the proportion of constituent elements to generate a combined two-dimensional grayscale image; The grayscale image processing refers to: converting a one-dimensional vector composed of elements into a two-dimensional grayscale image by an outer product conversion method; The implementation methods of grayscale image processing include: First, obtain the components (such as ) is taken as the proportion of all elements in the metal alloy as the proportion of the constituent elements, and the proportion of the constituent elements is normalized to make a linear mapping to the interval [0, 1], and ensure that the sum of all element proportions is 1; The normalized ratio value is expressed as a k-dimensional vector, which is expressed as: ,in It is the normalized value of the proportion of all elements in the metal alloy. If the corresponding element is not contained, it will be filled with zero. The atomic number of the corresponding element gradually increases; wherein k represents the number of constituent elements, and k is a natural number, and an embodiment of k=45 is provided in the present invention; Further, and The transposed matrix of Perform outer product operation to generate k×k two-dimensional matrix S, which is expressed as: The two-dimensional matrix S is converted into a grayscale image, where each pixel value of the grayscale image is the product of two element ratio values. The specific processing method includes: The elements of the two-dimensional matrix S are normalized using the Min-Max method. During the normalization process, each element value in S is subtracted from the minimum value in S and then divided by the difference between the maximum and minimum values ​​in S. The formula of the Min-Max method is expressed as: ,in, is the minimum value among the elements of S, is the maximum value among the elements of S, is the normalized result; After normalization, each element value is multiplied by 255 to generate a grayscale image.

[0019] Figure 2 by As an example, the whole process of converting element components into images is shown. In the result of the outer product operation, all data are multiplied by 10. -2 .

[0020] The image constructed based on the outer product transformation method can reflect the relative size of the data variance and covariance; in addition, The value of each position is the product of the ratio of two elements, and this value is used as the pixel value of each point in the image; compared with other methods of constructing images based on element composition, this feature can also reflect the interaction between different elements and reduce the dependence on artificial feature engineering.

[0021] The two-dimensional grayscale image generated in this step can be input into the network model loaded in the subsequent step to output the amorphous forming ability, that is, the critical casting diameter.

[0022] Step S120: Loading the amorphous forming ability prediction model, the amorphous forming ability prediction model is constructed and generated after training the amorphous forming ability attribute data set, the amorphous forming ability attribute data set includes a two-dimensional grayscale image corresponding to the proportion of constituent elements of the amorphous alloy and the corresponding critical casting diameter .

[0023] Before loading the amorphous forming ability prediction model, constructing the amorphous forming ability prediction model includes: 1) Prepare a model training data set; wherein the model training data set consists of an amorphous forming ability property data set and an extended data set; A glass forming ability property dataset. A 2D grayscale image generated from the composition element ratios of a glassy alloy and the corresponding critical casting diameter. Composition, the proportion of constituent elements of amorphous alloys is extracted from metal alloy sample data; The ratio of the constituent elements is the atomic percentage of each element in the alloy, which must satisfy the constraint that the total is 100%. In order to eliminate the impact of different feature dimension differences on model training, all element ratios are normalized and linearly mapped to the interval of 0,1 to ensure the uniformity of data distribution; and samples related to low-frequency elements (such as Yb and Pb) are eliminated to reduce statistical noise.

[0024] When collecting data, the composition element ratios of metal alloy samples and their corresponding critical casting diameters ( ) initial sample, in the embodiment provided by the present invention, the initial sample reaches 7795; then remove the samples with missing critical casting diameter and repeated element composition, remove For extreme values ​​greater than 35 mm, 990 valid samples covering 45 elements are retained, and each sample is represented by a feature vector: ; Overall, the sample data presents a long-tail distribution, with samples ranging from 0.2 to 5 mm accounting for 66.5%.

[0025] The dataset was then randomly divided into a training set (792 samples) and a test set (198 samples) in a ratio of 8:2 to ensure balanced distribution.

[0026] The method of grayscale image processing for the sample is the same as the grayscale image processing in step S110. After generating a 45×45 two-dimensional matrix S, it is converted into a grayscale image. At this time, a grayscale image corresponding to the proportion of the constituent elements and its corresponding critical casting diameter ( ) constitutes a data set of amorphous glass forming ability properties.

[0027] In order to solve the problem of limited data on amorphous alloys, the present invention adopts a data enhancement strategy to improve the generalization ability of the model: specifically, two methods, Mixup and Gaussian noise (GN), are introduced to expand the amorphous forming ability attribute data set, and the transition state of different alloy components is simulated to form an extended data set; The data enhancement strategy includes generating extended data through a linear interpolation formula and working in synergy with the noise injection dual path. While expanding the amount of training data, it forces the model to learn a stable mapping relationship under feature perturbations, and ultimately improves the prediction ability of minority class samples by smoothing the decision boundary. The specific implementation process includes: In order to alleviate the problem of skewed distribution of data sets, the present invention adopts the Mixup strategy to linearly mix two random sample data , The composition element ratios and their corresponding labels are used to generate new samples simulating the transition states of different alloy compositions. , where m and n are the serial numbers of random samples.

[0028] New Sample The generation method is: ; in, and are two random samples, is a random number drawn from the Beta distribution Beta(a,a) that satisfies .

[0029] New Sample right The ratio of the constituent elements The validity constraint is performed (that is, the generated samples must satisfy the constraint that the element ratio is non-negative and the sum is 1), which is expressed as: .

[0030] In this step, the Gaussian noise strategy is used to add noise to the new sample. Random noise that obeys Gaussian distribution (normal distribution) is added to simulate the uncertainty of experimental data, and the noise mean is 0. The parameters after noise perturbation are truncated by standard deviation and adjusted in the experiment to control the intensity of perturbation. Its distribution characteristics make the noise values ​​concentrated near the mean, with a small amount distributed at the tail, which can enhance data diversity and avoid physical properties out of bounds and prevent excessive deviation from the true distribution.

[0031] The amorphous forming ability property dataset and the extended dataset together constitute the model training dataset for training the deep learning model.

[0032] 2) Define a deep learning model; the deep learning model proposed in the present invention is an improved ResNet18 model, which adds an improved parallel multi-scale convolution module to enhance the model's ability to extract features of different scales.

[0033] ResNet18, as a classic lightweight neural network architecture based on residual connections, has broad application prospects and good performance in the field of metal materials. In view of the fact that the key information of the image in the data set is distributed in a discrete dot matrix, the conventional convolutional neural network structure is improved in the present invention, and a multi-scale feature fusion mechanism is added to avoid local or global information loss.

[0034] The improved convolutional neural network structure includes four main paths and one bypass. After the input features pass through multiple main paths in parallel, the outputs of each main path are connected in the channel dimension. The bypass passes the input to the output for residual connection. If the output dimensions of the connected main path and the bypass are not equal, the output of the connected main path is connected through The number of convolution control channels matches the number of output channels of the bypass, and then a residual connection is performed. The CNN architecture and convolution kernel size of each main path are different, ensuring that the parallel structure can focus on information of different scales of the image; and using the improved residual block, the model can ensure the depth of the network while broadening the width of the network, ensuring that the model does not lose detail information and global information.

[0035] Specifically, the structure of the deep learning model includes: Initial convolutional layer, stride 2 Max pooling layer and global average pooling layer, InceptionBlock module, fully connected layer, ReLU activation and Dropout layer; Among them, the structure of the InceptionBlock module as a residual module is as follows Figure 3As shown in the figure: Based on the multi-scale feature extraction, a parallel multi-scale convolution module is used to synchronously process input features through four main paths; the CNN architecture and convolution kernel size of each main path are different, ensuring that the parallel structure can focus on information of different scales of the image; the outputs of each main path are connected in the channel dimension, and the number of channels is adjusted to the specified number through the convolution layer, and the original input is subjected to the same operation to ensure that the two can be residually connected. The improvement of the residual module is implemented in this feature, which not only ensures the depth of the network but also broadens the width of the network, ensuring that the model does not lose detail information and global information. Different from the traditional residual connection of element-by-element addition and connecting features through the channel dimension, the improved residual module enables the independent transmission and fusion of features of different scales, and strengthens the model's ability to analyze the distribution features of discrete lattices. It can be seen that the InceptionBlock module can capture the microscopic details and macroscopic laws of element interaction at the same time, significantly improving the accuracy and robustness of the prediction of amorphous forming ability.

[0036] When the four main paths process the input features synchronously, they form a parallel convolution structure, which leads to channel stability problems. To address this problem, the present invention adopts a unified output channel strategy: all convolutional layers and pooling layers maintain the same number of output channels, eliminating the risk of information loss caused by the increase or decrease of the number of channels in the traditional residual block. The inverted residual structure is used to achieve dynamic adjustment of a fixed number of channels through an expansion factor (such as 6): after the input channel C is expanded to 6C through the main branch convolution layer, the number of channels is kept constant in subsequent operations. Compared with the traditional structure "compression-convolution-expansion" path, this design adopts the "expansion-convolution-compression" process to ensure that the high-dimensional channel space is always maintained in the feature extraction stage, effectively suppressing information confusion.

[0037] The optimized deep learning model architecture is as follows Figure 4 As shown, Conv represents the convolution operation, MaxPooling represents the maximum pooling, AdaptiveAvgPooling represents the adaptive average pooling, and Indicates the size of the convolution kernel and pooling kernel, InceptionBlock has different numbers of channels, BN indicates batch normalization, relu is the activation function, and Dense indicates the fully connected layer. The grayscale image is enlarged to ; and the step size of ResNet18 is 2 Initial convolutional layer, stride 2 The maximum pooling layer and global average pooling improve the original residual block and fully connected layer (the original 8 residual blocks are replaced by the improved InceptionBlock module).

[0038] The traditional ResNet18 model is used for image classification tasks. In the present invention, in addition to the improvement of the residual module, the fully connected layer is also improved to achieve the prediction of the GFA of the alloy as a regression task. The present invention adopts a fully connected layer structure with intermediate dimensions of 768 and 256, which is suitable for the regression task of images constructed based on the outer product method.

[0039] After the last InceptionBlock module, the number of channels of the image feature is 512. After global average pooling, the feature dimension becomes (1,1,512). After global average pooling, the image feature passes through the first linear layer to map the number of channels from 512 to 768. Then, after ReLU activation and Dropout layer, the second linear layer maps the number of channels from 768 to 256. After ReLU activation and Dropout layer again, the third linear layer maps the number of channels from 256 to 1, and the prediction result of the model is obtained. The Dropout layer prevents the model from overfitting by randomly setting the output of some neurons to 0. Its parameter p represents the probability of each neuron in each layer being set to 0. This model sets p to 0.5.

[0040] 3) Train and verify the deep learning model through the model training data set to build an amorphous forming ability evaluation model.

[0041] The present invention uses a stochastic gradient descent optimizer to train the model, with the initial learning rate set to 0.001, the momentum parameter to 0.9, and the loss function to be the mean square error. A dynamic learning rate decay strategy is introduced during the training process. When the validation set loss does not decrease continuously, the learning rate is gradually reduced. In order to evaluate the performance of the model, the root mean square error and the coefficient of determination are used as core indicators. The former quantifies the absolute deviation between the predicted value and the true value, and the latter measures the model's ability to explain data changes. In addition, the data enhancement effect is comprehensively evaluated by the statistical parameters PCD value, skewness and kurtosis to ensure the rationality and diversity of the generated samples. Among them, the PCD value is used to measure the distribution similarity between the original data and the synthetic data, the skewness evaluates the data symmetry, and the kurtosis describes the steepness of the distribution.

[0042] Specifically, the root mean square error, mean absolute error and absolute coefficient calculation formula are: , , ,in, Indicates Observations, represents the corresponding predicted value, is the mean of the observed values, is the sample size. Values ​​range from 0 to 1, with values ​​closer to 1 indicating a better fit of the model to the data.

[0043] The above evaluation indicators can be used to analyze the prediction performance of the model.

[0044] The present invention verifies the effectiveness of the model improvement and data enhancement strategy through comparative experiments: BaseModel is the baseline model in which the fully connected layer of ResNet18 is replaced with the regression head, and Inc-BaseModel is the improved model after the introduction of parallel multi-scale convolution. The experiment shows that when the expansion factor is 1 (i.e., no channel expansion), the Inc-BaseModel is 0.837; after expanding the number of channels to 6 times, It was increased to 0.846, which verified the enhancement effect of the "expansion first and then compression" structure on the prediction ability of GFA. Based on this verification, the expansion factor was set to 6. The prediction effects of the above four groups of models are as follows Figure 5 As shown in the figure, the horizontal and vertical axes correspond to the true value and the predicted value respectively, the color of the point reflects the prediction error (blue → small, red → large), and the red straight line is the linear fitting result. Figure 5 (a) It can be seen that BaseModel It reaches 0.8074, with only two significant outliers, proving that the outer product composition method can effectively capture the interaction between elements; Figure 5 (b) shows that the Inc-BaseModel that introduces parallel multi-scale convolution It is improved to 0.846, indicating that multi-scale feature extraction has a key contribution to model performance; Figure 5 (c)(d) Comparison of the effects of Mixup and Gaussian noise (GN) enhancement strategies: Both alleviate the problem of insufficient data by expanding samples, but Mixup It reaches 0.86 without significant outliers, and its performance is better than GN. It can be seen that the advantage of the amorphous forming ability prediction model structure of the present invention comes from the synthetic samples generated by Mixup, which has stronger distribution generalization, and the weighted label mechanism can guide the model to build a smoother decision boundary.

[0045] After the training is completed, the amorphous forming ability prediction model is built, that is, the amorphous forming ability prediction model can be loaded in the actual prediction scenario to achieve the critical casting diameter prediction.

[0046] Step S130: Output the predicted critical casting diameter according to the combined two-dimensional grayscale image generated in step S110 information.

[0047] In practical applications, some alloys of the Mg-Cu-Gd and Zr-Al-Cu systems that did not appear in the model training data set were selected, and two-dimensional grayscale images were generated according to the proportion of the alloy's constituent elements to predict the amorphous glass forming ability. At the same time, the Inc-BaseModel-Mixup model was used to predict its amorphous glass forming ability. Its ternary phase diagram is shown in Figure 6 As shown: Each axis represents the proportion of the corresponding element, the scattered points represent the true value measured experimentally, and the predicted GFA of the alloy increases from the blue area to the red area. Figure 6 (a) It can be concluded that The alloys in the nearby red area may have stronger GFA, and the alloys in the yellow area outside the red area have stronger GFA, while the star points and triangle points are mainly distributed in the red and yellow areas, which is consistent with the actual data measured in the experiment. The model predicts that the GFA of the alloys in the outer green area is poor, and the distribution of the dots in this area is also consistent with the experimental results; when exploring high GFA alloys in the Mg-Cu-Gd system, we first start from Starting experiments with nearby alloys can shorten the time and reduce the design cost of MGs with high GFA: Figure 6 (b) The model predicts that the alloys with high GFA in the Zr-Al-Cu system are distributed in the red and yellow areas. The distribution of stars in this area proves the accuracy of the model's prediction results and shows its potential in developing new bulk metallic glasses (BMGs). The distribution of dots in the green and blue areas proves that the model has the ability to exclude alloys with poor GFA.

[0048] In general, the present invention uses outer product visualization to characterize element interactions, combined with parallel multi-scale convolution and data enhancement strategies, to achieve accurate prediction of amorphous forming ability without artificial feature engineering. The amorphous forming ability prediction method provided by the present invention can predict areas containing higher GFA alloys and areas containing lower GFA alloys, which is convenient for researchers to start experiments from areas where higher GFA alloys may exist first, and avoid experiments in areas containing lower GFA alloys. The present invention only requires elemental composition as input, and does not require additional calculation of other parameters to predict the GFA of the alloy, which greatly reduces the dependence on feature engineering; compared with traditional design methods and traditional ML models, it simplifies the development process of new BMGs. Using this invention to guide the development process of BMGs can greatly reduce the time and economic costs required and accelerate the development process of BMGs.

[0049] The above disclosures are only several specific embodiments of the present invention; however, the present invention is not limited thereto, and any changes that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for predicting amorphous forming ability based on two-dimensional grayscale images, characterized in that: The following steps are involved: Obtaining the proportion of constituent elements of metal alloys; The ratio of the constituent elements is the atomic percentage of each constituent element in the alloy, and satisfies the constraint that the total is 100%; Performing grayscale image processing on the ratio of the constituent elements to generate a combined two-dimensional grayscale image; the grayscale image processing refers to: converting the one-dimensional vector of the constituent elements into a two-dimensional grayscale image by an outer product conversion method; Loading a glass-forming ability prediction model, the glass-forming ability prediction model is constructed and generated after training with a glass-forming ability attribute data set, the glass-forming ability attribute data set includes a two-dimensional grayscale image corresponding to the proportion of constituent elements of the amorphous alloy and a critical casting diameter ; Based on the combined two-dimensional grayscale image, the predicted critical casting diameter is output information.

2. The method for predicting the amorphous forming ability according to claim 1, characterized in that: The grayscale image processing comprises: Normalize the proportion of the constituent elements so that they are linearly mapped to the interval [0, 1]; The normalized ratio value is expressed as a k-dimensional vector, which is expressed as: ,in, is the normalized value of the composition element ratio, k is a natural number, and The atomic numbers of elements gradually increase; Will and The transposed matrix of Perform outer product operation to generate k×k two-dimensional matrix S, which is expressed as: The two-dimensional matrix S is converted into a grayscale image, and each pixel value of the grayscale image is the product of two element ratio values.

3. The method for predicting the amorphous forming ability according to claim 2, characterized in that: The converting the two-dimensional matrix S into a grayscale image comprises: Normalize the elements of the two-dimensional matrix S; After normalization, each element value is multiplied by 255 to generate a grayscale image.

4. The method for predicting the amorphous forming ability according to claim 2, characterized in that: The normalization of the elements of the two-dimensional matrix S is implemented using the Min-Max method, which is expressed as: ,in, is the minimum value among the elements of S, is the maximum value among the elements of S, is the normalized result.

5. The method for predicting amorphous forming ability according to claim 1, characterized in that: Before loading the amorphous forming ability prediction model, constructing the amorphous forming ability prediction model includes: Prepare a model training data set; wherein the model training data set is composed of an amorphous forming ability attribute data set and an extended data set; the amorphous forming ability attribute data set is a two-dimensional grayscale image generated by the proportion of constituent elements of the amorphous alloy and the corresponding critical casting diameter The amorphous forming ability attribute data set is expanded to simulate the transition state of different alloy components to form an extended data set; Define a deep learning model; the deep learning model is an improved ResNet18 model; The deep learning model is trained and verified through the model training data set to build an amorphous forming ability evaluation model.

6. The method for predicting the amorphous forming ability according to claim 5, characterized in that: The expanding the amorphous forming ability attribute data set comprises: Use Mixup strategy to linearly mix two random sample data , The composition element ratios and their corresponding labels are used to generate new samples simulating the transition states of different alloy compositions. , where m and n are the serial numbers of random samples; When the Gaussian noise strategy is adopted, random noise that obeys the Gaussian distribution is added to the new sample, and the mean value of the noise is 0.

7. The method for predicting the amorphous forming ability according to claim 6, characterized in that: The new sample The generation method is: ； in, and are two random samples, is a random number drawn from the Beta distribution Beta(a,a) that satisfies .

8. The method for predicting the amorphous forming ability according to claim 6, characterized in that: The new sample right The ratio of the constituent elements Perform validity constraints, including: 。 9. The method for predicting the amorphous forming ability according to claim 5, characterized in that: The structure of the deep learning model includes: Initial convolutional layer, stride 2 Max pooling layer and global average pooling layer, InceptionBlock module, fully connected layer, ReLU activation and Dropout layer.

10. The method for predicting the amorphous forming ability according to claim 5, characterized in that: The deep learning model includes four main paths and one bypass. After the input features pass through multiple main paths in parallel, the outputs of each main path are connected in the channel dimension, and the bypass passes the input to the output for residual connection; if the output dimensions of the connected main path and the bypass are not equal, the output of the connected main path is matched with the output channel number of the bypass through 1×1 convolution to perform residual connection; the CNN architecture and convolution kernel size of each main path are different.

Images