Amorphous glass forming ability prediction method based on Transformer and table data conversion

Abstract

The present invention discloses a method for predicting amorphous forming ability based on Transformer and table data conversion, comprising the following steps: obtaining basic properties and amorphous forming ability properties of amorphous alloy constituent elements in a metal alloy sample; basic properties refer to element properties based on experiments or density functional theory simulations, including: atomic basic properties, chemical properties and physical properties, and amorphous forming ability properties include critical casting diameter D max ; Generate an amorphous alloy data set; where the data set includes independent variables and response variables, and the independent variables include element composition and mixing enthalpy H mix , the response variable is the critical casting diameter D max ; Define a relational model, which is an integrated model composed of three types of models; Use an amorphous alloy data set to train the integrated model and build an amorphous forming ability evaluation model. According to the above technical solution, multiple relational models for predicting critical casting diameters can be trained, and multiple relational models can be combined to build an optimal amorphous forming ability evaluation model.

Description

Technical Field

[0001] The present invention relates to the field of amorphous forming ability analysis of amorphous alloys, and in particular to an amorphous forming ability prediction method based on Transformer and table data conversion. Background Art

[0002] Amorphous alloys are an important component of modern alloy materials. Due to their excellent mechanical and physical and chemical properties, they are widely used in high-tech fields such as energy, communications, and aerospace. The amorphous forming ability is an important indicator used to characterize the ease with which alloys form amorphous materials. According to the amorphous forming ability of amorphous alloys, their use in high strength, high hardness, good corrosion resistance, and special magnetic environments can be determined. However, the amorphous forming ability of most amorphous alloys is low, which poses a major challenge to engineering manufacturing. Therefore, the evaluation of the amorphous forming ability of amorphous alloys can not only promote the application of amorphous alloys themselves, but also promote the development of condensed matter physics.

[0003] Generally speaking, the glass-forming ability of amorphous alloys can be evaluated by the critical cooling rate or critical casting diameter. The critical cooling rate is difficult to obtain through experiments and thermal analysis calculations. Therefore, in actual research and development, although it can be obtained through methods such as Parameters, glass transition temperature T rg , supercooled liquid region ΔT x There are schemes to evaluate the parameters of amorphous forming ability based on standards, but these standards can only be used for specific alloy components and the prediction accuracy is not high. When using the currently popular relationship model to establish an evaluation model for properties and amorphous forming ability, due to the different constituent elements of different amorphous alloys, the relationship between basic properties and amorphous forming ability is also different. Due to the uneven data density of the samples, there are certain limitations on the amorphous alloys that the evaluation model is suitable for. Therefore, a technical solution is needed to integrate the basic properties of amorphous alloys, expand the relationship between basic properties and amorphous forming ability, and achieve accurate evaluation of amorphous forming ability. Summary of the invention

[0004] To achieve the above objectives, the present application provides a method for predicting amorphous forming ability based on Transformer and table data conversion, comprising the following steps:

[0005] Obtain the basic properties of the amorphous alloy constituent elements and the amorphous forming ability properties in the metal alloy sample; the basic properties refer to the element properties based on experiments or density functional theory simulations, including: atomic basic properties, chemical properties and physical properties, and the amorphous forming ability properties include the critical casting diameter D max ;

[0006] Generate an amorphous alloy data set; the data set includes independent variables and response variables, and the independent variables include element composition EC and mixing enthalpy H mix , the response variable is the critical casting diameter D max ;

[0007] Define the relational model, which is an integrated model composed of three types of models;

[0008] The integrated model is trained using the amorphous alloy dataset to build an amorphous forming ability evaluation model and realize the prediction of amorphous forming ability.

[0009] Among them, the independent variables are generated by calculating the basic properties of the constituent elements of the amorphous alloy; when generating the independent variables, the parameters related to the basic properties are calculated using linear mixing rules, derivative mixing rules, deviations and differential formulas.

[0010] Among them, the basic properties of atoms include: atomic number, atomic mass, group, period; the chemical properties include: Pauling Electronegativity, Mulliken Electronegativity;

[0011] The physical properties include: density, melting point, and specific heat capacity.

[0012] Furthermore, before generating the amorphous alloy data set, the basic properties and amorphous forming ability properties are normalized, and the property value range is controlled to be between 0 and 1, forming a first sample set.

[0013] The first sample set is divided into a training set, including: dividing the first sample set into a training set and a test set according to a set ratio; oversampling the training set to generate synthetic data, and merging the synthetic data with the training set into a second data set; wherein the oversampling processing algorithms include: ADASYN, SMOGN, and GN.

[0014] Furthermore, the ADASYN algorithm is used to assign different weights to a few existing samples and to supplement and add different numbers of samples;

[0015] The GN algorithm generates new samples by adding Gaussian white noise to samples in the rare domain;

[0016] The SMOGN algorithm combines the SMOTER algorithm with the GN algorithm. After making the dense data in the sample sparse through random undersampling, the data in the rare domain becomes dense through SMOTER interpolation or the introduction of Gaussian noise to achieve sample supplementation.

[0017] When the synthetic data and the training set are combined into a second data set, statistical parameters are used to evaluate the quality of the second data set, to judge the improvement effect of the data set distribution after data enhancement processing, and to determine the amorphous alloy data set according to the quality of the second data set; the statistical parameters include: PCD value, skewness and kurtosis.

[0018] When training the integrated model, weights are assigned to each model according to its importance in the prediction, and the prediction results of each model are multiplied by the corresponding weights and then summed to generate the training prediction value.

[0019] Wherein, the integrated model includes a first relationship model, a second relationship model and a third relationship model;

[0020] The first relational model is used to generate prediction results based on the dependencies and interactions of the basic attributes embodied by the independent variables;

[0021] The second relationship model is used to generate prediction results based on the difference between the minimized feature distance matrix and the pixel distance matrix embodied by the independent variables;

[0022] The third relational model is used to process each feature according to the piecewise linear coding embodied by the independent variable to generate a prediction result.

[0023] The present invention aims at the properties of the constituent elements of amorphous alloys and the relationship between each basic property and the critical casting diameter, adds a data enhancement strategy suitable for different intervals, trains and generates a variety of relationship models for predicting the critical casting diameter, and combines the multiple relationship models to construct the optimal amorphous forming ability evaluation model to achieve amorphous forming ability prediction. The method provided by the present invention has fast calculation speed, high prediction accuracy, low cost, and short cycle, and can quickly and accurately predict the critical casting diameter before experimentally preparing amorphous alloys; wherein, the data set covers a wide range of elements, and the model performs well on the data set, so that the prediction method has universality and is applicable to all types of amorphous alloys. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a step diagram of a method for predicting amorphous forming ability according to an embodiment of the present invention;

[0025] Figure 2 is a sample processing step diagram in the process of constructing an amorphous forming ability evaluation model provided by an embodiment of the present invention;

[0026] Figure 3 is a kernel density map of original data and oversampled data provided by an embodiment of the present invention;

[0027] Figure 4 It is a relationship model determination logic diagram provided by an embodiment of the present invention;

[0028] Figure 5 is a comparison chart of model prediction results provided according to an embodiment of the present invention;

[0029] Figure 6 It is a ternary phase diagram predicted for the Zr-Al-Ni system provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The present invention aims at the properties of the constituent elements of the amorphous alloy, constructs different relationship models that establish predictive relationships with the critical casting diameter through different data enhancement strategies, and combines multiple relationship models to achieve the prediction of characteristic parameters corresponding to the amorphous forming ability of the amorphous alloy.

[0031] The specific implementation of the present invention is described in detail below with reference to the accompanying drawings.

[0032] The method for predicting the amorphous forming ability of the present invention is as follows: Figure 1 As shown, the following steps are included:

[0033] Step S100: obtaining basic properties of amorphous alloy constituent elements and amorphous forming ability properties in a metal alloy sample;

[0034] Among them, the basic properties come from the properties of elements based on experiments or density functional theory simulations, including basic atomic properties such as atomic number, atomic mass, and groups in the periodic table; Pauling Electronegativity, Mulliken Chemical properties such as electronegativity; physical properties such as density, melting temperature, specific heat capacity;

[0035] Specifically, the types of basic properties include: topological structure, thermodynamic characteristics, physical property characteristics, etc. The basic property parameters that characterize the topological structure include: atomic radius difference ΔD c , atomic metal radius difference ΔD m , Pauling Poor electronegativity Δ χ p , Mulliken Poor electronegativity Δχ m , Theoretical molar volume V tm Equivalent parameters; Basic property parameters based on thermodynamic parameters: mixing enthalpy H mix , Mixing Entropy S mix , Mismatch Entropy S mis , average fusion entropy S f , thermal diffusion coefficient Ktd , PHS mis , PHSS and PHS mix and other parameters; the basic properties based on physical property characteristics include: density, melting point, specific heat capacity and other parameters.

[0036] This step also captures the glass forming ability properties including the critical casting diameter D max .

[0037] Step S110: Generate an amorphous alloy data set; wherein the data set includes independent variables and response variables, and the independent variables include element composition EC and mixing enthalpy H mix , the response variable is the critical casting diameter D max ;

[0038] The independent variables are formed by processing the basic properties of the constituent elements of the amorphous alloy. When generating independent variables, the parameters related to the basic properties are linearly mixed ( x l ), derivative mixing rule ( x r ),deviation( x d1 ) and the difference formula ( x d2 ) were calculated, and finally 111 characteristic parameters were obtained. These characteristics are more useful in describing alloy properties than the traditional three characteristic temperatures (glass transition temperature T g , crystallization transition temperature T x and liquidus temperature T l ), which can more accurately and comprehensively reflect the characteristics of the constituent elements.

[0039] After obtaining the characteristic parameters, before forming the amorphous alloy data set, the parameters of the basic properties and the amorphous forming ability property parameters are normalized using the feature scaling method to normalize the range of all features, and the parameter values ​​of the basic properties are proportionally scaled to specific values ​​between 0 and 1 to form the first sample set, which can be used for subsequent training of the relationship model.

[0040] In the present invention, the amorphous alloy data set generated at this time needs to be further processed before participating in the model training, as shown in the following example. Figure 2 As shown:

[0041] 1) Division of training set: The first sample set is divided into a training set and a test set according to a set ratio, for example, a ratio of 8:2.

[0042] 2) Oversampling processing:

[0043] In practical applications, due to the different constituent elements of amorphous alloys, the correlation between the critical casting diameter and the basic properties in different intervals is different. To ensure the training results, the present invention further strengthens the training set: the training set is oversampled using a data enhancement algorithm to generate new synthetic data; the synthetic data is merged with the training set into a second data set.

[0044] Among them, in the present invention, according to the characteristics of amorphous alloy data, three data enhancement algorithms are adopted, namely ADASYN, SMOGN and GN. Through the oversampling algorithm, the data shortage and imbalance problem of critical casting diameter in the prediction of amorphous forming ability is alleviated.

[0045] Specifically, ADASYN focuses on synthesis on the boundary of normal domain and rare domain, calculates weights according to the distance between minority samples and their k-nearest neighbors, and then assigns different weights to minority samples to generate different numbers of samples, thereby achieving sample supplementation and data enhancement; GN generates new samples by adding Gaussian white noise to samples in rare domain; SMOGN combines SMOTER with GN, makes dense data sparse through random undersampling, and then makes the data in rare domain dense through SMOTER interpolation or introduction of Gaussian noise, thereby achieving sample supplementation and data enhancement.

[0046] 3) Data quality judgment: Furthermore, this step also uses statistical parameters to evaluate the quality of the second data set, judges the improvement effect of the data set distribution after data enhancement processing, and determines the amorphous alloy data set according to the quality of the second data set.

[0047] The statistical parameters specifically include PCD value, skewness and kurtosis; the PCD value is the difference between the correlation matrix between the real data and the synthetic data calculated based on the Frobenius norm. The smaller the value, the better the similarity between the generated data and the original data; the skewness is used to measure the direction and degree of deviation of the data distribution. When the skewness is zero, it means that the distribution is symmetrical, that is, normal distribution; a negative skewness indicates left skewness with a longer right tail, while a positive skewness is the opposite; the kurtosis describes the steepness of the distribution of the data set. A positive kurtosis value indicates a distribution that is steeper than a normal distribution, and a negative kurtosis value indicates a flatter distribution.

[0048] The theoretical calculation formulas for statistical parameters are:

[0049]

[0050] in, is the PCD value, Skewness is the skewness, Kurtosis is the kurtosis, XR is the matrix of real data, X S Generates the matrix of oversampling data. corr is the correlation matrix, is the i-th original data, is the mean of the original data, is the sample serial number, is the sample size, F is the Frobenius norm.

[0051] 4) After data enhancement, the dataset and the test set together constitute the training samples of the model.

[0052] The present invention provides a specific embodiment to illustrate the method of constructing an amorphous forming ability evaluation model.

[0053] In this step, we first import the amorphous alloy dataset from the database. The dataset contains 840 samples, each with 112 attributes, one of which is the critical casting diameter. D max , which quantitatively describes the amorphous forming ability of amorphous alloys; the other 111 properties are elemental properties based on experiments or density functional theory (DFT) simulations, including basic atomic properties (such as atomic number, atomic mass, groups in the periodic table), chemical properties (such as Pauling Electronegativity, Mulliken electronegativity) and physical properties (density, melting temperature, specific heat capacity), etc.

[0054] Secondly, the amorphous alloy data set is divided into two parts, one is the training set (accounting for 80% of the total sample size), and the other is the test set (20%). Due to the large difference in the order of magnitude between various features, the attributes with larger values ​​will have a greater impact on the model than the attributes with smaller values. Therefore, the data is standardized to eliminate the impact of dimension and order of magnitude: the minimum value of each attribute is subtracted from the attribute set and divided by the difference between the maximum and minimum values ​​in the attribute set, so that each attribute value ranges between 0 and 1, which is conducive to improving the accuracy of model prediction.

[0055] After data preprocessing, three data enhancement algorithms, ADASYN, GN, and SMOGN, are used to enhance the data of the training set. D max Data greater than 5 are oversampled to generate new data to alleviate the impact of data shortage and skew problems on GFA prediction.

[0056] In a specific embodiment of the present invention, the kernel density of the oversampled data is compared with that of the original data. Figure 3 As shown: It can be seen that these three oversampling algorithms are Dmax The sample gaps in some intervals are supplemented, for example: ADASYN generates more D max For samples between 5 and 7.5, the density curve of GN is the lowest within this range, but D max The GN curve of the larger part is relatively high, which means that the algorithm generates more large MG samples; in the embodiment of the present invention, the original data set D max The maximum value is only 25, and the SMOGN algorithm generates multiple D max The number of new samples is 27, which helps to improve the prediction performance of the model for bulk MG.

[0057] It can be seen that the distribution of the oversampled data set is more uniform and more conducive to modeling.

[0058] After the data set is completed, the amorphous alloy data set is divided into a training set and a test set according to a set ratio, for example, a ratio of 8:2.

[0059] Step S120: defining a relationship model, wherein the relationship model is an integrated model composed of three types of models;

[0060] The integrated model includes a first relationship model, a second relationship model and a third relationship model;

[0061] The first relational model is used to generate prediction results according to the dependency and interaction of some parameters of basic attributes embodied by independent variables, and specifically includes three GBDT type models, such as XGBoost, LightGBM, and CatBoost;

[0062] The second relationship model is used to generate a prediction result by minimizing the difference between the feature distance matrix and the pixel distance matrix according to the basic attribute part parameters embodied by the independent variable, and specifically includes a CNN based on a data conversion algorithm (T2I, REFINED, IGTD);

[0063] The third relational model is used to process each feature according to the piecewise linear coding of the basic attribute part parameters embodied by the independent variable to generate a prediction result, specifically including three models based on the Transformer architecture, such as: TabNet, FT-Transformer-Linear, and FT-Transformer-Periodic.

[0064] Select 3 relationship models from each of the above three relationship models, combine them with the oversampling method, input independent variables and response variables, and implement training in each relationship model; for example, put TabNet, T2I, REFINED, IGTD, FT-linear, and FT-Periodic models into three data sets composed of three enhancement algorithms ADASYN, GN, and SMOGN, and get XGBoost, LightGBM, CatBoost, TabNet-OR, TabNet-GN, TabNet-SMOGN, TabNet-ADASYN, T2I-OR, T2I-GN, T2I-SMOGN, T2I-ADASYN, REFINED-OR, REFINED-GN, REFINED-SMOGN, REFINED-ADASYN, IGTD-OR, IGTD-GN, IGTD-SMOGN, IGTD-ADASYN, FT-l inear-OR, FT-linear-GN, FT-linear-SMOGN, FT-linear-ADASYN, FT-Periodic-OR, FT-Periodic-GN, FT-Periodic-SMOGN, FT-Periodic- ADASYN has a total of 27 models.

[0065] Next, we compare the performance of different models on different data sets on the test set, and integrate the models with better performance on the test set to obtain the structure of the final relational model. The processing flow is as follows: Figure 4 shown.

[0066] The present invention provides a method for model evaluation:

[0067] 1. Calculate the coefficient of determination R of the relationship model 2 And the root mean square error RMSE:

[0068] Among them, R 2 It is used to measure the degree of fit of the model. The closer it is to 1, the better the fit of the model. 1 indicates complete fit. The closer RMSE is to 0, the lower the prediction error of the model. The calculation formula is:

[0069] ,

[0070] ,

[0071] in, for D max The predicted value of As its average value, for Dmax The true value of .

[0072] 2. From all the relationship models, extract the three relationship models with the best model performance, use the weighted integration algorithm based on Dirichlet distribution, assign appropriate weights to the base models according to the importance of each model in the prediction, multiply the prediction results of various deep neural network models by the corresponding weights and then sum them. The specific process of the weighted integration algorithm based on Dirichlet distribution is as follows;

[0073] ,

[0074] in, Refers to the prediction result after weighted integration: W i are the weights of the three relationship models respectively. These are the prediction results of the three relationship models.

[0075] In this step, based on the weighted integration algorithm of Dirichlet distribution, the prediction results of the base learners are integrated as the training data of the meta-learner to improve the generalization ability of the model.

[0076] Compared with the traditional GBDT model, the FT-Transformer that introduces piecewise linear coding and periodic activation function and the CNN based on data conversion method have higher prediction accuracy. FT-Periodic performs best on the original dataset. Among all models, IGTD-SMOGN has the highest R 2 The value is the highest, followed by FT-linear-ADASYN.

[0077] In the embodiment provided by the present invention, according to the R 2 The values ​​are sorted and the three models with the best performance are selected as the meta-models of integrated deep learning. In this example, the base models used are: REFINED-ADASY, IGTD-SMOGN, and FT-linear-ADASYN:

[0078] 1) The core idea of ​​REFINED is to convert high-dimensional vectors into images with spatial neighborhood dependency so that they can be trained using the traditional CNN architecture: first, the Euclidean distance between the input features is calculated as a distance matrix; then the Bayesian version of the multidimensional scaling (BMDS) algorithm is applied to the distance matrix to create an initial image that retains the feature distance in 2D space at a low computational cost; then BMDS is used to estimate the position of the features in a finite field and limit each pixel to contain at most one feature; finally, the feature arrangement scheme is optimized using a hill climbing algorithm, which uses the absolute difference between the new feature position and the estimated true distance as the loss function. For each feature, different feature arrangements are tried by constantly exchanging the positions of the central feature and its neighboring features, and the arrangement scheme that minimizes the loss function is selected. Once the loss function converges, unique coordinates are generated for each feature in two-dimensional space. Based on these coordinates, the features are mapped into two-dimensional space to create an image with spatial domain dependency for each sample. In this embodiment, an image is generated for each amorphous alloy sample according to the REFINED algorithm and input into CNN to predict the corresponding D max .

[0079] 2) IGTD calculates the Euclidean distance between all features and sorts them to generate a feature distance matrix R; then connects the features row by row as the initial image, calculates the Euclidean or Manhattan distance between all pixels in the image and sorts them to generate a pixel distance matrix Q; then calculates the difference between the matrix R and Q based on the error function, and assigns each feature to a suitable pixel position by constantly exchanging feature positions to minimize the error function, so that similar features are close to each other in the image. In each iteration, the algorithm will consider the feature that has not been exchanged for the longest time and look for another feature to replace it until the error function converges; if it does not converge, the algorithm will repeat to the set maximum number of times to reduce the error. The iterative process is reflected as follows:

[0080] ,

[0081] Among them, R and Q are feature distance matrix and pixel distance matrix respectively. r i,j and q i,j are the feature distance elements and pixel distance matrix elements in the feature distance matrix, respectively, i and j are the horizontal and vertical coordinate values ​​of the elements, respectively;

[0082] In the generated images, features with strong correlations and interactions are adjacent to each other.

[0083] In this embodiment, an image is generated for each amorphous alloy sample according to the IGTD algorithm and input into CNN to predict the corresponding Dmax .

[0084] 3) FT-Transformer adapts the Transformer architecture to better apply to tabular data. The model consists of a feature marker component and a Transformer component: the feature marker converts all features (categorical and numerical features) into embeddings, and then the Transformer component uses these embeddings to analyze the data. Therefore, each transformer layer operates on the feature level of a single object. And use piecewise linear encoding PLE for numerical features, such as feature i-th, to divide its value into disjoint T i Interval set (B1 i , …, B T i ) to construct bins: B t i =[b t-1 i , b t i ), once the i-th box is determined, it is encoded according to the following formula:

[0085] ,

[0086] in, PLE(X) is the piecewise linear encoding value, e 1 to e T For each box’s specific value, R T is a set of box values, b t is the right boundary of the tth interval, t is the interval number, is the characteristic value, e t is the specific value of the box where the eigenvalue is located.

[0087] The components corresponding to the boxes whose right borders are lower than the given eigenvalue are set to 1, and the components corresponding to the boxes whose left borders are higher than the given eigenvalue are set to 0. The specific values ​​of the boxes where the eigenvalues ​​are located are given by In this embodiment, the piecewise linear encoding (PLE) generates different initial representations for each numerical feature to replace the original scalar value. The 111 element features of the amorphous alloy are input into FT-Transformer and each feature is processed using piecewise linear encoding to predict the corresponding D max .

[0088] The results of CNN show that there are correlations and interactions between the characteristics of various elements in amorphous alloys. IGTD generates images that better retain the characteristic neighborhood structure by minimizing the difference between the feature distance matrix and the pixel distance matrix. The CNN effect based on this algorithm is the best among the three data conversion algorithms.

[0089] Step S130: using the amorphous alloy data set to train the integrated model and construct an amorphous forming ability evaluation model.

[0090] When training the ensemble model, weights are assigned to each model according to its importance in the prediction, and the prediction results of each model are multiplied by the corresponding weights and then summed to generate the training prediction value. The weighted average model is the Ensemble model, which is also an integrated deep learning model for predicting amorphous forming ability. The prediction results of the base learners are integrated to build an ensemble model to improve the generalization ability of the model.

[0091] The formula is as follows:

[0092] .

[0093] The performance and robustness of the model are improved after the weighted integration based on Dirichlet distribution. The main reason is that the prediction result of the integrated model is the weighted average of the prediction results of the other three models. The weight given to the model algorithm with good prediction effect is also larger, that is, the effect of the strong model is enhanced, and the effect of the weak model is weakened, so that the overall performance is optimized.

[0094] Figure 5 This is a scatter plot of the four models used in this specific example on the test set. The x-axis represents the actual measured D max The y-axis represents the model prediction D max The red line in the figure is the linear fitting curve of the model, and the black diagonal line represents the best fitting curve. The color of the data point reflects the size of the prediction error. The darker the blue, the smaller the error, and the darker the red, the larger the error. R 2 The prediction error of most data points is controlled within 2.5, and only three dark red data points have large prediction errors. It can be seen that the robustness of the model has been improved after the weighted integration based on Dirichlet distribution, which is in line with the expectations of this study.

[0095] The weighted integration algorithm provided by the present invention integrates the prediction results of multiple models, thereby improving the accuracy of the integrated model. Taking the Zr-Al-Ni alloy system as an example, the experimental values ​​of the Zr-Al-Ni alloy system and the prediction values ​​of the amorphous forming ability evaluation model are placed in a ternary diagram to determine whether the change trend of the prediction value is consistent with the experimental results. Figure 6 As shown, the three axes in the figure represent the proportion of each element, and the scattered points are the experimental values ​​of commonly used technologies. The colors from blue to red in the figure indicate that the predicted GFA gradually increases, and it is more likely to produce large amorphous particles. It can be observed from the figure that the changing trend of the predicted values ​​is consistent with the experimental results. In the experiment, most of the alloys with high GFA values ​​are located in the yellow and red areas, the stars are located in the middle part of the red area, and the triangles are mostly located in the red and yellow areas. The white points corresponding to the alloys with lower experimental values ​​are all on the dividing line between green and blue. This proves that the integrated model established by the present invention has good generalization ability for GFA prediction. The prediction results based on the model are searched in areas where large amorphous particles can be produced, which provides a data basis for the discovery of new alloys and greatly reduces the time and material consumption of traditional trial and error methods in finding large amorphous particles.

[0096] The amorphous forming ability prediction method provided by the present invention has fast calculation speed, high prediction accuracy, low cost and short cycle, and can quickly and accurately predict the critical casting diameter before experimental preparation of amorphous alloys; during the construction process, the constructed data set covers a wide range of elements, and the model performs excellently on the data set, and is applicable to various types of amorphous alloys.

[0097] 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 Transformer and table data conversion, characterized in that: The following steps are involved: Obtain the basic properties of the amorphous alloy constituent elements and the amorphous forming ability properties in the metal alloy sample; The basic properties refer to the element properties based on experiments or density functional theory simulations, including: atomic basic properties, chemical properties and physical properties. The amorphous forming ability properties include critical casting diameter D max ; The basic properties and amorphous forming ability properties are normalized, and the control property value range is between 0 and 1 to form a first sample set; the first sample set is divided into a training set, including: dividing the first sample set into a training set and a test set according to a set ratio; oversampling the training set to generate synthetic data, and merging the synthetic data with the training set into a second data set; wherein the oversampling processing algorithm includes: ADASYN, SMOGN, GN; when the synthetic data and the training set are merged into the second data set, statistical parameters are used to evaluate the quality of the second data set, and the improvement effect of the data set distribution after data enhancement processing is judged, and the amorphous alloy data set is determined according to the quality of the second data set; the statistical parameters include: PCD value, skewness and kurtosis; Among them, the ADASYN algorithm is used to assign different weights to a small number of existing samples, and to supplement and add different numbers of samples; the GN algorithm generates new samples by adding Gaussian white noise to samples in the rare domain; the SMOGN algorithm combines the SMOTER algorithm with the GN algorithm, and after making the dense data in the sample sparse through random undersampling, the data in the rare domain becomes dense through SMOTER interpolation or the introduction of Gaussian noise, thereby achieving sample supplementation; Generate an amorphous alloy data set; wherein the data set includes independent variables and response variables, and the independent variables include elemental composition EC and mixing enthalpy H mix , the response variable is the critical casting diameter D max ; A relationship model is defined, wherein the relationship model is an integrated model composed of three types of models; the integrated model includes a first relationship model, a second relationship model and a third relationship model; the first relationship model is used to generate a prediction result according to the dependency and interaction of basic attributes embodied by the independent variable; the second relationship model is used to generate a prediction result according to the difference between the minimized feature distance matrix and the pixel distance matrix embodied by the independent variable; the third relationship model is used to process each feature according to the piecewise linear coding embodied by the independent variable to generate a prediction result; The integrated model is trained using the amorphous alloy data set to construct an amorphous forming ability evaluation model to achieve amorphous forming ability prediction; when the integrated model is trained, weights are assigned according to the importance of each model in the prediction, and the prediction results of each model are multiplied by the corresponding weight and then summed to generate a training prediction value.

2. The method for predicting the amorphous forming ability according to claim 1, characterized in that: The independent variables are generated by calculating the basic properties of the constituent elements of the amorphous alloy; when generating the independent variables, the parameters related to the basic properties are calculated using linear mixing rules, derivative mixing rules, deviations and differential formulas.

3. The method for predicting the amorphous forming ability according to claim 2, characterized in that: The basic atomic properties include: atomic number, atomic mass, group, period; the chemical properties include: Pauling Electronegativity, Mulliken Electronegativity; the physical properties include: density, melting point, specific heat capacity.

Images