A method for predicting amorphous forming ability based on a two-dimensional gray scale image

Abstract

The application discloses a kind of based on two-dimensional gray image's amorphous forming ability prediction method, comprising: obtaining the composition element proportion of metal alloy's amorphous alloy;Composition element proportion is carried out gray image processing, and generates combination two-dimensional gray image;Load amorphous forming ability prediction model, amorphous forming ability prediction model is constructed and generated after training by amorphous forming ability attribute data set, and amorphous forming ability attribute data set includes the two-dimensional gray image and critical casting diameter corresponding to the composition element proportion of amorphous alloy;According to combination two-dimensional gray image, the critical casting diameter information of predicted output is exported.According to the above technical solution, other parameters do not need additional calculation to accurately predict the GFA of alloy, and the area containing higher GFA alloy and the area of lower GFA alloy are predicted, which greatly reduces the dependence on feature engineering, can greatly reduce the time cost and economic cost required.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of amorphous alloy glass forming ability prediction, in particular to a two-dimensional gray image-based amorphous alloy glass forming ability prediction method. BACKGROUND

[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) of amorphous alloys directly determines the feasibility of materials in engineering manufacturing. The stronger the index means that the alloy can form a stable amorphous phase at a lower cooling rate, thereby realizing the preparation of large-size components. However, the existing amorphous alloys generally have the defect of insufficient forming ability, which seriously restricts their large-scale application in industrial scenarios. Accurate evaluation of 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 amorphous material structure.

[0003] Currently, the glass 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 experimental determination difficulty. Although the academic community has proposed evaluation criteria based on thermodynamic parameters such as glass transition temperature (Tg), supercooled liquid phase region (ΔTx), etc., these methods are only applicable to specific alloy systems and have limited prediction accuracy. The existing evaluation models predict by establishing the correlation between basic physical parameters and glass forming ability. However, due to the significant difference in element composition of different alloy systems, the mapping relationship between their physical parameters and glass forming ability has obvious heterogeneity. In addition, the uneven distribution of experimental data leads to insufficient model generalization ability, making it difficult to achieve universal evaluation of multi-component alloy systems. Therefore, it is urgent to build a new evaluation framework that integrates multi-dimensional physical parameters to break through the existing technical bottlenecks by strengthening feature correlation and model adaptability, and realize the prediction of glass forming ability.

[0004] Therefore, it is urgent to build a new evaluation framework that integrates multi-dimensional physical parameters to break through the existing technical bottlenecks by strengthening feature correlation and model adaptability, and realize the prediction of glass forming ability. SUMMARY

[0005] To achieve the above purpose, the present application provides a two-dimensional gray image-based amorphous alloy glass forming ability prediction method, comprising the following steps:

[0006] Obtaining the composition element ratio of the metal alloy; the composition element ratio is the atomic percentage of each composition element in the alloy, and satisfies the constraint condition that the total is 100%;

[0007] Performing gray image processing on the composition element ratio to generate a combined two-dimensional gray image; the gray image processing refers to converting a one-dimensional vector of element composition into a two-dimensional gray image by an outer product conversion method;

[0008] ​Load an amorphous forming ability prediction model, which is generated by training an amorphous forming ability attribute data set, the amorphous forming ability attribute data set including a two-dimensional gray scale image corresponding to a composition element ratio of an amorphous alloy and a critical casting diameter ;

[0009] According to the combined two-dimensional gray scale image, output a predicted critical casting diameter Information.

[0010] Wherein, the gray scale image processing includes:

[0011] Normalizing the composition element ratio, so that the linear mapping is in the interval [0, 1];

[0012] The normalized ratio value is represented as a k-dimensional vector, represented as: , wherein,

[0013] is the normalized value of the composition element ratio, k is a natural number, and x 1 to x k The atomic number of the element gradually increases;

[0014] The outer product operation is performed on the transpose matrix of , to generate a k×k two-dimensional matrix S, represented as:

[0015]

[0016] The two-dimensional matrix S is converted into a gray scale image, and each pixel value of the gray scale image is the product of the two element ratio values.

[0017] Converting the two-dimensional matrix S into a gray scale image includes:

[0018] Normalizing the elements of the two-dimensional matrix S;

[0019] Each normalized element value is multiplied by 255 to generate a gray scale image.

[0020] Wherein, the normalization of the elements of the two-dimensional matrix S is realized by the Min-Max method, represented as:

[0021] , wherein, is the minimum value of the S elements, is the maximum value of the S elements, is the normalized result.

[0022] ​Furthermore, before loading the amorphous forming capability prediction model, the amorphous forming capability prediction model is constructed, including:

[0023] Prepare a model training dataset; wherein, the model training dataset consists of an amorphous forming capability attribute dataset and an extended dataset; the amorphous forming capability attribute dataset consists of a two-dimensional grayscale image generated from the component element ratios of the amorphous alloy and the corresponding critical casting diameter. The dataset is constructed by expanding the amorphous forming ability attribute dataset and simulating the transition states of different alloy compositions.

[0024] Define a deep learning model; the deep learning model is an improved ResNet18 model;

[0025] A model for evaluating amorphous formation capability is constructed by training and validating a deep learning model using a model training dataset.

[0026] Expanding the aforementioned amorphous forming ability attribute dataset includes:

[0027] The Mixup strategy is used to linearly mix two arbitrary sample data. , The compositional 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] The Gaussian noise strategy adds random noise that follows a Gaussian distribution to the new sample, with a noise mean of 0.

[0029] Among them, new samples The generation method is as follows:

[0030]

[0031] ;

[0032] in, and For two random samples, It is a random number drawn from the Beta distribution Beta(a,a) that satisfies .

[0033] Furthermore, new samples right The proportion of constituent elements Validity constraints include: .

[0034] Furthermore, the structure of deep learning models includes: Initial convolutional layer with stride of 2 The maximum pooling layer and the global average pooling layer, the InceptionBlock module, the full connection layer, the ReLU activation and the Dropout layer; and the deep learning model comprises four main paths and a bypass, the input features are transmitted in parallel through the multiple main paths, the outputs of the main paths are connected in the channel dimension, the bypass transmits the input to the output for residual connection; if the output dimensions of the connected main paths and the bypass are not equal, the output of the connected main paths is matched with the output channel number of the bypass through The output of the connected main paths is matched with the output channel number of the bypass through the convolution control channel number, and then residual connection is performed; the CNN architecture and the convolution kernel size of each main path are different.

[0035] According to the present application, the interaction of elements is characterized by the outer product image, combined with the parallel multi-scale convolution and the data enhancement strategy, so that the amorphous forming ability can be accurately predicted without artificial feature engineering; by using the prediction model provided by the present application, the regions containing high GFA alloys and the regions containing low GFA alloys can be predicted, which facilitates researchers to first start experiments from the regions possibly containing high GFA alloys, and avoids experiments in the regions containing low GFA alloys, and the GFA of the alloy can be predicted without additional calculation of other parameters, which greatly reduces the dependence on feature engineering. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 It is a step schematic diagram of the amorphous forming ability prediction method based on a two-dimensional gray image according to an embodiment of the present application;

[0037] Figure 2 It is a step schematic diagram of gray image processing according to an embodiment of the present application;

[0038] Figure 3 It is a structure schematic diagram of InceptionBlock in the amorphous forming ability prediction model according to an embodiment of the present application;

[0039] Figure 4 It is a structure schematic diagram of the amorphous forming ability prediction model according to an embodiment of the present application;

[0040] Figure 5 It is a comparison diagram of the amorphous forming ability prediction model and other scheme prediction results according to an embodiment of the present application;

[0041] Figure 6 It is a ternary phase diagram of Mg-Cu-Gd and Zr-Al-Cu systems predicted according to an embodiment of the present application. DETAILED DESCRIPTION

[0042] The present application is directed to the composition element properties of amorphous alloys and their complex correlation with critical casting diameters, and through multi-model integration and data enhancement strategies, an efficient amorphous forming ability prediction framework is constructed. The specific embodiments will be described in detail below with reference to the accompanying drawings.

[0043] The amorphous forming ability prediction method provided by the present application, as shown in Figure 1 , includes the following steps:

[0044] Step S100: Obtain the composition element proportion of the metal alloy;

[0045] In the present application, the critical casting diameter of the metal alloy can be predicted according to the composition element proportion of the metal alloy. The effective composition element proportion is the atomic percentage of each composition element in the alloy, and satisfies the constraint condition that the sum is 100%.

[0046] Step S110: Perform gray-scale image processing on the composition element proportion to generate a combined two-dimensional gray-scale image;

[0047] The gray-scale image processing refers to converting a one-dimensional vector of element composition into a two-dimensional gray-scale image through an outer product conversion method;

[0048] The implementation method of gray-scale image processing includes:

[0049] First, the proportion of each composition element (for example ) in all elements of the metal alloy is obtained as the composition element proportion, and the composition element proportion is normalized to map linearly to the [0, 1] interval and ensure that the sum of all element proportions is 1;

[0050] The normalized proportion value is represented as a k-dimensional vector, denoted as: , where is the normalized value of the composition element proportion of all elements of the metal alloy, and if it does not contain the corresponding element, it is filled with zero, and The atomic number of the corresponding element gradually increases; wherein k represents the number of composition elements, and k is a natural number, in the present application, an embodiment of k=45 is provided;

[0051] Further, the outer product operation is performed on and the transpose matrix of to generate a k x k two-dimensional matrix S, denoted as:

[0052]

[0053] The two-dimensional matrix S is converted into a gray-scale image, and each pixel value of the gray-scale image is the product of the two element proportion values, and the specific processing method includes:

[0054] The elements of a two-dimensional matrix S are normalized using the Min-Max method. During normalization, each element value in S is subtracted from its minimum value and then divided by the difference between the maximum and minimum values. The formula for the Min-Max method is as follows:

[0055] ,in, The minimum value among the elements of S The maximum value among the elements of S This is the result of normalization;

[0056] Each normalized element value is multiplied by 255 to generate a grayscale image.

[0057] Figure 2 by The example illustrates the entire process of converting elemental components into an image, where all data in the outer product operation are multiplied by 10. -2 .

[0058] Images constructed using the outer product transformation method can reflect the relative magnitudes of data variance and covariance; additionally, The value at each position is the product of the proportions of the two elements, and this value is used as the pixel value of each point in the image. Compared with other image construction methods that use elements to form an image, this feature can also reflect the interaction between different elements, reducing the reliance on manual feature engineering.

[0059] The two-dimensional grayscale image generated in this step can be input into the network model loaded in subsequent steps to output the amorphous forming capability, i.e., the critical casting diameter.

[0060] Step S120: Load the amorphous forming capability prediction model. The amorphous forming capability prediction model is constructed and generated after training on the amorphous forming capability attribute dataset. The amorphous forming capability attribute dataset includes two-dimensional grayscale images corresponding to the component element ratios of amorphous alloys and the corresponding critical casting diameters. .

[0061] Before loading the amorphous forming capability prediction model, the amorphous forming capability prediction model is constructed, including:

[0062] 1) Prepare the model training dataset; the model training dataset consists of an amorphous forming ability attribute dataset and an extended dataset;

[0063] The amorphous forming capability attribute dataset consists of two-dimensional grayscale images generated from the elemental proportions of amorphous alloys and their corresponding critical casting diameters. The composition of amorphous alloys is extracted from metal alloy sample data.

[0064] wherein the proportion of constituent elements is the atomic percentage of each element in the alloy, and the constraint condition of summing to 100% needs to be met. In order to eliminate the influence of the dimensional difference of different characteristic quantities on model training, all element proportions are normalized to linearly map them to the interval of 0, 1, and the uniformity of data distribution is ensured; and the related samples of low-frequency elements (such as Yb, Pb) are removed to reduce statistical noise.

[0065] When collecting data, the proportions of constituent elements of the metal alloy samples and the corresponding critical casting diameters (Dc) are collected ) Initial samples, in the embodiments provided by the present application, the initial samples reach 7795; then the samples with missing critical casting diameters and the samples with repeated element compositions are removed, the extreme values greater than 35 mm are removed, and 990 effective samples covering 45 elements are reserved, each sample is embodied in the form of a feature vector: ; overall, the sample data presents a long-tail distribution, and the sample proportion of 0.2-5 mm accounts for 66.5%.

[0066] Subsequently, the data set is randomly divided into a training set (792 samples) and a test set (198 samples) in a ratio of 8:2, and the distribution uniformity is ensured.

[0067] The method and steps of the gray image processing of the sample are the same as the gray image processing in step S110, and after a 45x45 two-dimensional matrix S is generated, it is converted into a gray image; at this time, an amorphous forming ability attribute data set composed of the gray image corresponding to the proportion of constituent elements and the corresponding critical casting diameter (Dc) is generated.

[0068] In view of the problem of limited data quantity of amorphous alloys, the present application adopts a data enhancement strategy to improve the generalization ability of the model: specifically, two methods of Mixup and Gaussian noise (GN) are introduced to expand the amorphous forming ability attribute data set, simulate the transition state of different alloy compositions to form an expanded data set;

[0069] The data enhancement strategy includes generating expanded data through a linear interpolation formula, and cooperating with noise injection in a double path, expanding the training data quantity while forcing the model to learn the stable mapping relationship under feature disturbance, and finally improving the prediction ability for a small number of class samples through a smooth decision boundary, and the specific implementation process includes:

[0070] In order to alleviate the problem of skew distribution of the data set, the present application adopts a Mixup strategy to linearly mix the proportions of constituent elements of two random sample data 、 and their corresponding labels to generate new samples simulating the transition state of different alloy compositions , wherein m and n are the serial numbers of the random samples. ​​

[0071] New sample The generation method is as follows:

[0072]

[0073] ;

[0074] in, and For two random samples, It is a random number drawn from the Beta distribution Beta(a,a) that satisfies .

[0075] New sample right The proportion of constituent elements Applying validity constraints (i.e., the generated samples must satisfy the constraints that the proportion of elements is non-negative and the sum of their elements is 1), can be expressed as: .

[0076] In this step, Gaussian noise is used again to apply it to the new samples. Random noise following a Gaussian (normal) distribution is added to simulate the uncertainty of experimental data, with a noise mean of 0. The parameters after noise perturbation are truncated by the standard deviation and adjusted in the experiment to control the perturbation intensity. Its distribution characteristics make the noise values ​​concentrated near the mean, with a small number distributed at the tail, which can enhance the diversity of data and avoid physical properties going out of bounds, thus preventing excessive deviation from the true distribution.

[0077] The amorphous forming ability attribute dataset and the extended dataset together constitute the model training dataset for training the deep learning model.

[0078] 2) Define the deep learning model; The deep learning model proposed in this invention is an improved ResNet18 model, which incorporates an improved parallel multi-scale convolution module to enhance the model's ability to extract features at different scales.

[0079] 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 characteristic that the key information of the image in the dataset is distributed in a discrete lattice, this invention improves the conventional convolutional neural network structure and adds a multi-scale feature fusion mechanism to avoid the loss of local or global information.

[0080] The improved convolutional neural network structure includes four main paths and one bypass. Input features pass through multiple main paths in parallel, and the outputs of each main path are concatenated along the channel dimension. The bypass passes the input to the output for residual connections. If the output dimensions of the concatenated main paths and the bypass are not equal, the output of the concatenated main paths is passed through... Residual connections are made after the number of convolutional control channels is matched with the number of output channels in the bypass paths. The CNN architecture and convolutional kernel size are different for each main path to ensure that the parallel structure can focus on information at different scales of the image; and improved residual blocks are used to broaden the network width while ensuring network depth, so that the model does not lose detailed and global information.

[0081] Specifically, the structure of a deep learning model includes: Initial convolutional layer with stride of 2 Max pooling layer and global average pooling layer, InceptionBlock module, fully connected layer, ReLU activation and Dropout layer;

[0082] The InceptionBlock module, as a residual module, has the following structure: Figure 3 As shown: Based on multi-scale feature extraction, a parallel multi-scale convolutional module is adopted to simultaneously process input features through four main paths. Each main path has a different CNN architecture and convolutional kernel size, ensuring that the parallel structure can focus on information at different scales of the image. The outputs of each main path are concatenated along the channel dimension, and then the number of channels is adjusted to a specified number through convolutional layers. The original input undergoes the same operation to ensure residual connections between the two. This feature implements an improved residual module, which broadens the network width while maintaining network depth, ensuring that the model does not lose detailed and global information. Unlike traditional residual connections that add features element-wise and connect them along the channel dimension, the improved residual module allows features at different scales to be passed independently and then fused, enhancing the model's ability to analyze discrete lattice distribution features. It is evident that using the InceptionBlock module as the residual module can simultaneously capture the microscopic details and macroscopic patterns of elemental interactions, significantly improving the accuracy and robustness of amorphous formation capability prediction.

[0083] The simultaneous processing of input features by the four main paths constitutes a parallel convolutional structure, thus introducing channel stability issues. To address this problem, this invention employs a unified output channel strategy: all convolutional and pooling layers maintain the same number of output channels, eliminating the information loss risk caused by changes in the number of channels in traditional residual blocks. By inverting the residual structure and using an expansion factor (e.g., set to 6), a fixed number of channels is dynamically adjusted: after the input channel C is expanded to 6C by the main branch convolutional layer, the number of channels remains constant in subsequent operations. Compared to the traditional "compression-convolution-expansion" path, this design uses an "expansion-convolution-compression" process, ensuring that a high-dimensional channel space is maintained throughout the feature extraction stage, effectively suppressing information confusion.

[0084] Optimized deep learning model architecture, such as Figure 4As shown, Conv represents convolution operation, MaxPooling represents max pooling, and AdaptiveAvgPooling represents adaptive average pooling. and This indicates the size of the convolutional kernel and pooling kernel; InceptionBlock has different numbers of channels; BN represents batch normalization; ReLU is the activation function; and Dense represents a fully connected layer. The image is processed from its original state before input. The grayscale image was magnified to It also retains the stride size of 2 from ResNet18. Initial convolutional layer with stride of 2 The max pooling layer and global average pooling have been improved, along with improvements to the original residual blocks and fully connected layers (the original 8 residual blocks have been replaced with the improved InceptionBlock module).

[0085] Traditional ResNet18 models are used for image classification tasks. In this invention, in addition to improvements to the residual module, the fully connected layers are also improved to achieve GFA (Geometric Array Optimization) for predicting alloys as a regression task. This invention employs a fully connected layer structure with intermediate dimensions of 768 and 256, which is suitable for regression tasks of images constructed based on the outer product method.

[0086] After passing through the last InceptionBlock module, the image feature has 512 channels. After global average pooling, the feature dimension becomes (1, 1, 512). After global average pooling, the image feature passes through the first linear layer, mapping the number of channels from 512 to 768. Then, after ReLU activation and Dropout layers, the number of channels passes through the second linear layer, mapping the number of channels from 768 to 256. After another ReLU activation and Dropout layer, the number of channels passes through the third linear layer, mapping the number of channels from 256 to 1, thus obtaining the model's prediction result. The Dropout layer prevents overfitting by randomly setting the output of a subset of neurons to 0. Its parameter p represents the probability that each neuron in each layer is set to 0, and this model sets p to 0.5.

[0087] 3) Train and validate the deep learning model using the model training dataset to construct an amorphous formation capability evaluation model.

[0088] This invention employs a stochastic gradient descent optimizer to train the model, with an initial learning rate of 0.001, a momentum parameter of 0.9, and a mean squared error (MSE) loss function. A dynamic learning rate decay strategy is introduced during training; the learning rate is gradually reduced when the validation set loss fails to decrease continuously. To evaluate model performance, the root mean square error (RMSE) and coefficient of determination (COP) are used as core metrics. The former quantifies the absolute deviation between predicted and true values, while the latter measures the model's ability to explain data variations. Furthermore, the data augmentation effect is comprehensively evaluated using statistical parameters such as PCD (Plan-Do-D), skewness, and kurtosis to ensure the rationality and diversity of generated samples. Specifically, PCD measures the distribution similarity between the original and synthetic data, skewness assesses data symmetry, and kurtosis describes the steepness of the distribution.

[0089] Specifically, the formulas for calculating the root mean square error, mean absolute error, and absolute coefficient are as follows:

[0090] , , ,in, Indicates the first One observation value, This represents the corresponding predicted value. It is the average of the observed values. It refers to the number of samples. The value ranges from 0 to 1, and the closer the value is to 1, the better the model fits the data.

[0091] The predictive performance of the model can be analyzed using the above evaluation metrics.

[0092] This invention verifies the effectiveness of model improvement and data augmentation strategies through comparative experiments: BaseModel is the baseline model with ResNet18 fully connected layers replaced by regression heads, and Inc-BaseModel is the improved model with the introduction of parallel multi-scale convolutions. Experiments show that when the expansion factor is 1 (i.e., no channel expansion), Inc-BaseModel... The value is 0.837; after increasing the number of channels by 6 times, The value was increased to 0.846, verifying the enhancement effect of the "expansion-compression" structure on the predictive ability of GFA. Based on this verification, the expansion factor was set to 6. The prediction results of the above four models are as follows: Figure 5 As shown, the horizontal and vertical axes correspond to the actual and predicted values, respectively. The color of the points reflects the prediction error (blue → smaller, red → larger), and the red line represents the linear fitting result. Figure 5 (a) It can be seen that BaseModel's The value reached 0.8074, with only two significant outliers, proving that the outer product graphing method can effectively capture the interactions between elements; Figure 5(b) shows that the Inc-BaseModel, which introduces parallel multi-scale convolutions, will... The value increased to 0.846, indicating that multi-scale feature extraction makes a key contribution to the model's performance. Figure 5 (c)(d) Comparison of the effects of Mixup and Gaussian noise (GN) augmentation strategies: Both alleviate the problem of insufficient data by expanding the sample size, but Mixup makes... The performance reached 0.86 with no significant outliers, outperforming GN. This demonstrates that the advantage of the amorphous forming ability prediction model structure of this invention stems from the synthetic samples generated by Mixup, exhibiting stronger distribution generalization, and the weighted labeling mechanism guides the model to construct a smoother decision boundary.

[0093] After training, the amorphous forming capability prediction model is complete. This model can then be loaded into actual prediction scenarios to achieve the critical casting diameter. The prediction.

[0094] Step S130: Based on the combined two-dimensional grayscale image generated in step S110, output the predicted critical casting diameter. information.

[0095] In practical applications, some alloys of the Mg-Cu-Gd and Zr-Al-Cu systems that did not appear in the model training dataset were selected. Two-dimensional grayscale images were generated based on the elemental proportions of the alloys to predict their amorphous forming ability. Simultaneously, the Inc-BaseModel-Mixup model was used to predict their amorphous forming ability, and their ternary phase diagrams are shown below. Figure 6 As shown: each axis represents the proportion of the corresponding element, the scatter plot represents the experimentally measured true value, and the change from the blue area to the red area indicates an increase in the predicted GFA of the alloy. From Figure 6 (a) It can be concluded that in The alloy in the nearby red area may have a strong GFA, and the alloy in the yellow area outside the red area has a strong GFA. The star-shaped and triangular dots are mainly distributed in the red and yellow areas, which matches the actual experimental data. The model predicts that the alloy in the outer green area has a poor GFA, and the distribution of dots in this area also matches the experimental results. When exploring high GFA alloys in the Mg-Cu-Gd system, we first start from... Experiments with nearby alloys can shorten the design time for MGs with high GFA and reduce design costs. Figure 6 (b) indicates that the model predicts that alloys with high GFA in the Zr-Al-Cu system are distributed in the red and yellow regions. The star-shaped distribution in these regions proves the accuracy of the model's predictions and demonstrates its potential in developing novel bulk metallic glasses (BMG). The round dots distributed in the green and blue regions prove that the model has the ability to exclude alloys with poor GFA.

[0096] In summary, this invention characterizes elemental interactions through outlay visualization, combining parallel multi-scale convolution with data augmentation strategies to achieve accurate prediction of amorphous forming ability (BMG) without manual feature engineering. The amorphous forming ability prediction method provided by this invention can predict regions containing high-GFA alloys and low-GFA alloys, allowing researchers to begin experiments in regions likely to contain high-GFA alloys and avoid experiments in regions containing low-GFA alloys. This invention only requires elemental composition as input and does not require additional parameter calculations to predict the GFA of alloys, greatly reducing reliance on feature engineering; compared to traditional design methods and traditional ML models, it simplifies the development process of novel BMGs. Using this invention to guide the BMG development process can significantly reduce the required time and economic costs, accelerating the BMG development process.

[0097] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the present invention is not limited thereto, and any variations 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 formation capability based on two-dimensional grayscale images, characterized in that, Includes the following steps: To obtain the compositional element ratios of a metal alloy; The proportion of the constituent elements is the atomic percentage of each constituent element in the alloy, and satisfies the constraint that the total is 100%. The proportions of the constituent elements are subjected to grayscale image processing to generate a combined two-dimensional grayscale image; the grayscale image processing refers to: converting the one-dimensional vector composed of the elements into a two-dimensional grayscale image through the outer product transformation method; An amorphous forming capability prediction model is loaded. This model is constructed by training a deep learning model using an amorphous forming capability attribute dataset. The amorphous forming capability attribute dataset consists of a two-dimensional grayscale image generated from the component element ratios of the amorphous alloy and the corresponding critical casting diameter D. max The deep learning model is an improved ResNet18 model. The deep learning model includes four main paths and one bypass. The input features pass through multiple main paths in parallel, and the outputs of each main path are connected in the channel dimension. The bypass passes the input to the output for residual connection. Based on the combined two-dimensional grayscale image, the predicted critical casting diameter D is output. max information; The grayscale image processing includes: The proportions of the constituent elements are normalized so that the linear mapping is applied to the interval [0, 1]. The normalized scale values ​​are represented as a k-dimensional vector, as follows: X = (x1, x2, ..., x...). k ), where x1 to x k It is the value after normalizing the proportions of the constituent elements, where k is a natural number, and x1 to x k The atomic number of an element gradually increases; Transform X into X and X's transpose X. T Perform the outer product operation to generate a k×k two-dimensional matrix S, represented as: The two-dimensional matrix S is converted into a grayscale image, where each pixel value of the grayscale image is the product of the ratio values ​​of the two elements.

2. The method for predicting amorphous forming ability according to claim 1, characterized in that, The step of converting the two-dimensional matrix S into a grayscale image includes: Normalize the elements of the two-dimensional matrix S; Each normalized element value is multiplied by 255 to generate a grayscale image.

3. The method for predicting amorphous forming ability according to claim 1, characterized in that, The normalization of the elements of the two-dimensional matrix S is achieved using the Min-Max method, expressed as: Among them, X min The minimum value in S, X max The maximum value among the elements of S, X normalized This is the result of normalization.

4. The method for predicting amorphous forming ability according to claim 1, characterized in that, Before loading the amorphous forming capability prediction model, the amorphous forming capability prediction model is constructed, including: Prepare a model training dataset; wherein the model training dataset consists of an amorphous forming ability attribute dataset and an extended dataset; the amorphous forming ability attribute dataset is extended to simulate the transition states of different alloy compositions to form an extended dataset; Define a deep learning model; A model for evaluating amorphous formation capability is constructed by training and validating a deep learning model using a model training dataset.

5. The method for predicting amorphous forming ability according to claim 4, characterized in that, The expansion of the amorphous forming ability attribute dataset includes: The Mixup strategy is used to linearly mix two arbitrary sample data (x) m y m ), (x n y n The compositional proportions and corresponding labels of the alloys 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; The Gaussian noise strategy adds random noise that follows a Gaussian distribution to the new sample, with a noise mean of 0.

6. The method for predicting amorphous forming ability according to claim 5, characterized in that, The new sample The generation method is as follows: Among them, (x m y m ) and (x n y n Let f(a,a) be two random samples, and let λ be a random number drawn from the Beta distribution Beta(a,a) such that 0 ≤ λ ≤ 1.

7. The method for predicting amorphous forming ability according to claim 5, characterized in that, The new sample right The proportion of the constituent elements X = (x1, x2, ..., x...) i ,...,x k ) Validity constraints are imposed, including:

8. The method for predicting amorphous forming ability according to claim 4, characterized in that, The structure of the deep learning model includes: a 7×7 initial convolutional layer, a 3×3 max pooling layer with a stride of 2 and a global average pooling layer, an InceptionBlock module, a fully connected layer, a ReLU activation layer and a Dropout layer.

9. The method for predicting amorphous forming ability according to claim 1, characterized in that, If the output dimensions of the main path and the side path are not equal after concatenation, the output of the main path after concatenation is matched with the output channels of the side path by a 1×1 convolution and then residual concatenation is performed; the CNN architecture and convolution kernel size of each main path are different.

Images