Method and system for determining composition and process parameters of an alloy to be designed, and computer storage medium

Abstract

Embodiments of the present application relate to the technical field of computer materials science, in particular to a method and system for determining the composition and process parameters of an alloy to be designed, and a computer storage medium. The method trains an attention network model with input being alloy feature data and output being alloy performance data, and sets a storage node in the attention network model to record the attention weight matrix of all attention layers and all attention heads corresponding to each set of sample alloy data in the prediction process of the model, to learn the attention degree of the model to different input features when generating performance prediction results, and to determine the influence mechanism of the alloy features on the performance of the alloy accordingly, to realize the essential interpretability of the model, and to efficiently and accurately determine the composition and process parameters of the alloy meeting the target performance and to design the alloy meeting the target performance.

Description

Technical Field

[0001] Embodiments of this application relate to the field of computer materials science and technology, specifically to a method and system for determining the composition and process parameters of an alloy to be designed, as well as a computer storage medium. Background Technology

[0002] The statements herein are provided merely as background information in connection with this application and do not necessarily constitute prior art.

[0003] The composition and processing parameters of an alloy directly affect its properties. Therefore, establishing a quantitative relationship between the composition and processing parameters of an alloy and its properties is crucial for designing alloys with target properties. Traditional trial-and-error experimental development is time-consuming, costly, and inefficient in exploring a wide range of compositional dimensions. Summary of the Invention

[0004] A brief overview of this application is provided below to offer a basic understanding of certain aspects thereof. It should be understood that this overview is not an exhaustive summary of the application. It is not intended to identify key or essential parts of the application, nor is it intended to limit its scope. Its purpose is merely to present certain concepts in a simplified form as a prelude to the more detailed description that follows.

[0005] In a first aspect, embodiments of this application provide a method for determining the composition and process parameters of an alloy to be designed, comprising the following steps: S10: Determine the training dataset and the test dataset. Both the training dataset and the test dataset include multiple sets of sample alloy data. Each set of sample alloy data includes the feature data and performance data of the sample alloy. S20: Based on the training dataset determined in step S10, train an attention network model based on a multi-head attention mechanism. The feature data of the sample alloy is used as the input of the attention network model, and the performance data of the sample alloy is used as the output of the attention network model to obtain the trained attention network model. The attention network model includes storage nodes to record the attention weight matrices of all attention layers and all attention heads, as well as the information entropy and attention concentration of the attention weight matrices. S30: Using the trained attention network model, predict the performance of the sample alloys in the test dataset and determine the predicted performance values ​​corresponding to the feature data of the sample alloys in the test dataset. During the prediction process, the attention network model records the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset, as well as the information entropy and attention concentration of the attention weight matrix. S40: Based on the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset recorded in step S30, as well as the information entropy and attention concentration of the attention weight matrix, determine the influence mechanism of the alloy's characteristics on its performance. S50: Based on the influence mechanism of the alloy's characteristics on its properties, and the target properties of the alloy to be designed, determine the composition and process parameters of the alloy to be designed.

[0006] The method for determining the composition and process parameters of an alloy to be designed, provided in the embodiments of this application, can obtain the degree of attention the model pays to different input features when generating performance prediction results by recording the attention weight matrix calculated by the model during the prediction process. This allows for the determination of the influence mechanism of alloy features on its performance, efficiently and accurately determining the composition and process parameters of the alloy that meet the target performance. Since the explanatory information of the attention weight matrix is ​​generated synchronously with the prediction process, it achieves the essential interpretability of the model, avoiding the problems of fidelity, stability, and misleadingness in post-interpretation. The influence mechanism of alloy features on its performance determined based on the model's inherent explanatory information is also more accurate and reliable. The information entropy of the attention weight matrix is ​​used to measure the dispersion of the attention distribution, and the attention concentration of the attention weight matrix is ​​defined as the Gini coefficient or maximum weight ratio of the weight distribution, used to measure the degree of inequality in the attention distribution. When determining the influence mechanism of alloy features on its performance, considering the information entropy and attention concentration of the attention weight matrix can determine the training quality of the attention network model, thereby ensuring high training quality, accurate prediction results, and reliable decision-making process.

[0007] Secondly, embodiments of this application also provide a system for determining the composition and process parameters of an alloy to be designed, comprising: Data acquisition module and processor; The data acquisition module is set to acquire training datasets and test datasets. Both training datasets and test datasets include multiple sets of sample alloy data. Each set of sample alloy data includes the feature data and performance data of the sample alloy. Processor settings: Based on the training dataset, an attention network model based on a multi-head attention mechanism is trained. The feature data of the sample alloy is used as the input of the attention network model, and the performance data of the sample alloy is used as the output of the attention network model to obtain the trained attention network model. The attention network model includes storage nodes to record the attention weight matrices of all attention layers and all attention heads, as well as the information entropy and attention concentration of the attention weight matrices. The trained attention network model is used to predict the performance of sample alloys in the test dataset, and the predicted performance values ​​corresponding to the feature data of the sample alloys in the test dataset are determined. During the prediction process, the attention network model records the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset, as well as the information entropy and attention concentration of the attention weight matrix; Based on the attention weight matrices of all attention layers and all attention heads corresponding to each group of sample alloy data in the recorded test dataset, as well as the information entropy and attention concentration of the attention weight matrices, the influence mechanism of the alloy's characteristics on its performance is determined. Based on the influence mechanism of alloy characteristics on its properties, and the target properties of the alloy to be designed, the composition and process parameters of the alloy to be designed are determined.

[0008] Thirdly, embodiments of this application provide a computer storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the method provided in any embodiment of the first aspect of this application. Attached Figure Description

[0009] Other objects and advantages of this application will become apparent from the following description of embodiments of this application with reference to the accompanying drawings, and will help to provide a comprehensive understanding of this application.

[0010] Figure 1 This is a flowchart of a method provided according to an embodiment of this application; Figure 2 This is a bar chart showing the prediction accuracy of various models according to embodiments of this application; Figure 3 This is a scatter plot showing the prediction accuracy and optimization efficiency of various models according to embodiments of this application.

[0011] It should be noted that the accompanying drawings are not necessarily drawn to scale, but are shown only in a schematic manner without affecting the reader's understanding. Detailed Implementation

[0012] Exemplary embodiments of this application will be described below with reference to the accompanying drawings. For clarity and brevity, not all features of actual implementations are described in the specification. However, it should be understood that many implementation-specific decisions must be made in the development of any such actual embodiment to achieve the developer's specific goals, such as complying with constraints related to the system and business, and these constraints may vary depending on the implementation. Furthermore, it should be understood that while development work can be very complex and time-consuming, such development work is merely a routine task for those skilled in the art who benefit from the content of this application.

[0013] It should also be noted that, in order to avoid obscuring this application with unnecessary details, only the equipment structure and / or processing steps closely related to the solution according to this application are shown in the accompanying drawings, while other details that are not closely related to this application are omitted.

[0014] The inventors of this application have discovered that using machine learning models to determine the relationship between the composition and processing parameters of an alloy and its performance is relatively efficient. However, machine learning models suffer from a "black box effect," meaning they only provide output results, while the decision-making process by which the model processes inputs to arrive at the output is unexplainable to the user. Researchers not only need accurate predictions but also need to understand the mechanisms by which the alloy's composition and processing parameters affect its performance to guide experimental research.

[0015] Based on this, embodiments of this application provide a method for determining the composition and process parameters of an alloy to be designed. See also Figure 1 , Figure 1 This is a flowchart of a method provided according to an embodiment of this application, the method comprising the following steps: S10: Determine the training dataset and the test dataset. Both the training dataset and the test dataset include multiple sets of sample alloy data. Each set of sample alloy data includes the feature data and performance data of the sample alloy. S20: Based on the training dataset determined in step S10, train an attention network model based on a multi-head attention mechanism. The feature data of the sample alloy is used as the input of the attention network model, and the performance data of the sample alloy is used as the output of the attention network model to obtain the trained attention network model. The attention network model includes storage nodes to record the attention weight matrices of all attention layers and all attention heads, as well as the information entropy and attention concentration of the attention weight matrices. S30: Using the trained attention network model, predict the performance of the sample alloys in the test dataset and determine the predicted performance values ​​corresponding to the feature data of the sample alloys in the test dataset. During the prediction process, the attention network model records the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset, as well as the information entropy and attention concentration of the attention weight matrix; S40: Based on the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset recorded in step S30, as well as the information entropy and attention concentration of the attention weight matrix, determine the influence mechanism of the alloy's characteristics on its performance. S50: Based on the influence mechanism of the alloy's characteristics on its properties, and the target properties of the alloy to be designed, determine the composition and process parameters of the alloy to be designed.

[0016] The method for determining the composition and process parameters of an alloy to be designed, provided in the embodiments of this application, can obtain the degree of attention the model pays to different input features when generating performance prediction results by recording the attention weight matrix calculated by the model during the prediction process. This allows for the determination of the influence mechanism of alloy features on its performance, efficiently and accurately determining the composition and process parameters of the alloy that meet the target performance. Since the explanatory information of the attention weight matrix is ​​generated synchronously with the prediction process, it achieves the essential interpretability of the model, avoiding the problems of fidelity, stability, and misleadingness in post-interpretation. The influence mechanism of alloy features on its performance determined based on the model's inherent explanatory information is also more accurate and reliable. The information entropy of the attention weight matrix is ​​used to measure the dispersion of the attention distribution, and the attention concentration of the attention weight matrix is ​​defined as the Gini coefficient or maximum weight ratio of the weight distribution, used to measure the degree of inequality in the attention distribution. When determining the influence mechanism of alloy features on its performance, considering the information entropy and attention concentration of the attention weight matrix can determine the training quality of the attention network model, thereby ensuring high training quality, accurate prediction results, and reliable decision-making process.

[0017] In some embodiments, step S10 includes: S11: obtaining an original dataset, the original dataset including multiple sets of sample alloy data, each set of sample alloy data including feature data and performance data of the sample alloy, the feature data including composition feature data and process parameter feature data of the sample alloy; S12: removing abnormal sample alloy data from the original dataset, and processing the sample alloy data after removing abnormal sample alloy data so that the processed sample alloy data can be used for the attention network model; S13: dividing the processed sample data into a training dataset and a test dataset, both the training dataset and the test dataset including multiple sets of sample alloy data, each set of sample alloy data including feature data and performance data of the sample alloy, the feature data including composition feature data and process parameter feature data of the sample alloy.

[0018] In this embodiment, by removing outlier alloy data from the original dataset, the quality of the sample alloy data used for training and testing the attention network model is ensured to be high. As a result, the trained attention network model makes more accurate predictions of alloy performance, and the influence mechanism of alloy characteristics on its performance determined by the attention network model is more realistic and accurate.

[0019] In some embodiments, the original dataset can be obtained through a combination of literature review and supplementary experiments.

[0020] The compositional characteristics of an alloy are the mass fractions of the various chemical elements that make up the alloy. The process parameters of an alloy include annealing temperature, annealing time, equivalent diameter, gauge length, irradiation temperature, irradiation dose, test temperature, and strain rate. The properties of an alloy are mechanical properties, including yield strength and elongation.

[0021] In some embodiments, step S50 includes: S51: determining the mass fraction of various chemical elements constituting the alloy to be designed, and the process parameter characteristics of the alloy to be designed, based on the influence mechanism of the mass fraction of different chemical elements constituting the alloy and the process parameter characteristics of the alloy on the performance of the alloy, and the target performance of the alloy to be designed.

[0022] In such an embodiment, since the compositional characteristics of an alloy are the mass fractions of the various chemical elements that make up the alloy, determining the influence mechanism of the compositional characteristics of the alloy on the alloy's performance can actually determine the influence mechanism of the content of different chemical elements that make up the alloy on the alloy's performance. Similarly, determining the compositional characteristics of the alloy to be designed can actually determine the mass fractions of the various chemical elements that make up the alloy to be designed. When designing an alloy, based on the influence mechanism of the content of different chemical elements that make up the alloy on the alloy's performance, the mass fractions of chemical elements that have a strong influence on the target performance in the alloy can be adjusted in a targeted manner, thereby enabling the design of an alloy that meets the target performance more efficiently and accurately.

[0023] In some embodiments, step S12 includes: S121: determining the optimal number of clusters in the original dataset; S122: dividing the sample alloy data in the original dataset into datasets of unequal sizes according to the optimal number of clusters, and determining the cluster center of each dataset, wherein each dataset includes multiple sets of sample alloy data; S123: identifying and removing outlier sample alloy data from all datasets based on the distance from each set of sample alloy data in each dataset to its cluster center and the size of the dataset; S124: processing the sample alloy data in the dataset after removing outlier sample alloy data so that the processed sample alloy data can be used in the attention network model.

[0024] In such an embodiment, the sample alloy data in the original dataset is divided into a corresponding number of datasets based on the optimal number of clusters in the original dataset. This minimizes the differences between sample alloy data in the same dataset and maximizes the differences between sample alloy data in different datasets, which helps to identify anomalous datasets in each dataset, as well as anomalous sample alloy data in each dataset.

[0025] In some embodiments, in step S121, the optimal number of clusters for the original dataset is determined using the silhouette coefficient. In such embodiments, the silhouette coefficient takes into account both the similarity of data within a dataset and the difference between datasets, so as to maximize the difference between sample alloy data in different datasets divided according to the optimal number of clusters determined using the silhouette coefficient, and minimize the difference between sample alloy data within the same dataset.

[0026] In some embodiments, in step S123, when the distance from a group of sample alloy data in the dataset to its cluster center is greater than a predetermined value, the group of sample alloy data is determined to be abnormal sample alloy data, wherein the predetermined value can be determined by those skilled in the art based on the actual situation; when the dataset is too small, that is, when the number of sample alloy data in the dataset accounts for less than a predetermined proportion of the total number of sample alloy data in all datasets, the dataset is determined to be an abnormal dataset, and all sample alloy data in the dataset are abnormal sample alloy data, wherein the predetermined proportion can be determined by those skilled in the art based on the actual situation.

[0027] In some embodiments, in step S124, the sample alloy data in the dataset after removing abnormal sample alloy data is processed so that the processed sample alloy data can be used for the attention network model, including: performing min-max normalization on the feature data in all sample alloy data so that the value of the feature data is scaled to between [0,1]; performing fractional transformation on the performance data in all sample alloy data so that the performance data is converted into a uniform distribution or a standard Gaussian distribution, wherein the yield strength data is converted into a uniform distribution and the elongation data is converted into a standard Gaussian distribution.

[0028] In this embodiment, min-max normalization of the feature data eliminates dimensional differences between different feature data, making the input feature data more suitable for machine learning models. Fractional transformation of the performance data, converting it to a uniform or standard Gaussian distribution, eliminates the impact of differences in dimensions and magnitudes between the two performance indicators on model training. Furthermore, it reduces the dominance of extreme values ​​on the loss function when validating its prediction accuracy. Since yield strength data is relatively continuous in the dataset and has no significant prior skewness, converting it to a uniform distribution allows the model to learn uniformly across the entire yield strength range. Because elongation data is typically concentrated in the low-value region and contains long-tailed data, converting it to a standard Gaussian distribution compresses extreme values, making the data distribution more symmetrical and helping to ensure the model's regression stability.

[0029] In some embodiments, in step S13, those skilled in the art can determine the proportion of sample alloy data in the divided training dataset and test dataset according to actual needs.

[0030] The characteristics of an alloy also include its physical parameters. Each physical parameter of an alloy reflects a microscopic mechanism that affects the alloy's performance, and each physical parameter has a quantitative relationship with the alloy's composition, which can be determined based on the alloy's composition.

[0031] In some embodiments, step S13 includes: S131: determining the physical parameter characteristic data of all sample alloys based on the compositional characteristic data in the sample alloy data of the processed dataset; S132: dividing the sample alloy data in the processed dataset into a training dataset and a test dataset, wherein both the training dataset and the test dataset include multiple sets of sample alloy data, each set of sample alloy data including the characteristic data and performance data of the sample alloy, and the characteristic data including the compositional characteristic data, physical parameter characteristic data and process parameter characteristic data of the sample alloy.

[0032] In this embodiment, based on the compositional characteristics of the sample alloy, the physical parameter characteristics of the sample alloy are obtained. In subsequent steps, these physical parameter characteristics are also used as input to the attention network model. This allows for the determination of a quantitative relationship between the physical parameter characteristics and the alloy's properties. Furthermore, this relationship can be transformed into a quantitative relationship between the alloy's compositional characteristics and its properties, thus determining the influence mechanism of the alloy's compositional characteristics on its properties from another perspective. Since the compositional characteristics of an alloy are the mass fractions of various chemical elements that make up the alloy, the physical parameter characteristics are actually determined by the mass fractions of different chemical elements that make up the alloy. Therefore, it is also possible to quantify the relationship between the content of different chemical elements that make up the alloy and the alloy's properties from another perspective, determining the influence mechanism of the content of various chemical elements that make up the alloy on the alloy's properties.

[0033] The compositional and process parameter data of the sample alloys, along with their performance data, can be used to train the first attention network model. Conversely, the physical and process parameter data, along with their performance data, can be used to train the second attention network model. These two attention network models determine the influence mechanism of alloy compositional characteristics on alloy performance from different perspectives, resulting in more realistic and reliable results.

[0034] The inventors of this application have discovered that in predicting the elongation of alloys, models that only consider the physical parameters obtained by traditional linear weighted averaging have poor fitting effects on the relationship between alloy characteristics and their properties, resulting in low accuracy. This is because ductility, as a plasticity index, is essentially controlled by the weakest link in the material, rather than the average effect of all components. Local microscopic features such as the embrittlement effect of interstitial elements, grain boundary segregation, and the type and distribution of precipitates play a decisive role in the ductility of alloys, and these effects cannot be characterized by simple linear weighting.

[0035] In some embodiments, the physical parameter characteristics of the alloy include conventional physical parameter characteristics and nonlinear physical parameter characteristics.

[0036] The traditional physical parameters include: basic physical properties of the sample alloy, including valence electron concentration, average melting point, average Young's modulus, average bulk modulus, average shear modulus, and the ratio of bulk modulus to shear modulus; thermodynamic parameters of the sample alloy, including enthalpy of mixing, entropy of mixing, and free energy of mixing; structural stability parameters, including BCC stability parameters, phase stability parameter Omega, and gamma parameter; atomic-scale mismatch parameters, including atomic radius mismatch, electronegativity mismatch, and shear modulus mismatch; and elastic constant C'.

[0037] The nonlinear physical parameters include: interstitial atom embrittlement index, solid solution strengthening-embrittlement ratio, grain refinement factor, ductile-brittle transition temperature shift parameter, and fracture toughness estimate.

[0038] The interstitial atom embrittlement index is defined as (8.5×O + 6.2×N + 2.1×C + 1.3×H)^0.6, which characterizes the embrittlement effect of interstitial elements in the alloy. The coefficient reflects the difference in embrittlement ability of each interstitial element, and the index of 0.6 characterizes the embrittlement saturation effect.

[0039] The solid solution strengthening-embrittlement ratio is defined as 100×(12.0×Re + 8.5×W + 6.5×Ta + 5.2×Nb) / (1+ Re + W + Ta + Nb), which characterizes the solid solution strengthening contribution of each refractory element in the alloy. The denominator is used to ensure saturation behavior, and the coefficient reflects the strengthening efficiency of each element.

[0040] The grain refinement factor is defined as 100×(0.9×Ti + 0.85×Zr + 0.75×Hf + 0.6×V + 0.55×Nb)^0.33, which characterizes the grain refinement effect achieved by titanium group elements in the alloy through precipitation and pinning of relative grain boundaries. Here, the exponent 0.33 is an approximation of 1 / 3, which comes from the empirical scaling relationship between grain size and solute concentration, i.e., grain size ∝ solute concentration^(-1 / 3). The exponent 0.33 here reflects the nonlinear contribution of the solute to the pinning effect on grain boundaries.

[0041] The ductile-brittle transition temperature offset parameter is defined as 45×C + 65×N + 85×O + 15×P + 20×S - 45×Re + 18×W - 35×Ti - 30×Zr - 40×La, which characterizes the effect of alloying on the ductile-brittle transition temperature. Among them, interstitial elements increase the ductile-brittle transition temperature, while reactive elements decrease the ductile-brittle transition temperature.

[0042] The fracture toughness estimate is defined as 50 + 25×Ti + 28×Zr + 22×Hf + 30×La + 15×Al -35×C - 45×O - 40×N, which characterizes the fracture toughness estimate based on compositional characteristics. Among them, nano-precipitated phase elements improve toughness, while interstitial elements reduce toughness.

[0043] In such embodiments, the five nonlinear physical parameters—interstitial atom embrittlement index, solid solution strengthening-embrittlement ratio, grain refinement factor, ductile-brittle transition temperature shift parameter, and fracture toughness estimate—reflect the local microscopic characteristics that determine the ductility of the alloy, such as the embrittlement effect of interstitial elements, grain boundary segregation, and the type and distribution of precipitates. Therefore, these five nonlinear physical parameters are strongly correlated with the elongation of the alloy. Models that consider these five nonlinear physical parameters can more accurately fit the relationship between the alloy characteristics and its elongation, thereby improving the accuracy of the model's prediction of elongation.

[0044] Figure 2 This is a bar chart showing the prediction accuracy of various models according to embodiments of this application. Figure 3 This is a scatter plot showing the prediction accuracy and optimization efficiency of various models according to embodiments of this application. Referring to Figure 2, the determination coefficient R of the SEAM attention network model according to embodiments of this application... ２ The accuracy is significantly higher than other models, indicating that its prediction accuracy is significantly higher than other models. Referring to Figure 3, the SEAM attention network model of this application is located at the highest accuracy end of the efficiency frontier. Its prediction accuracy is the highest among all comparative models, and it surpasses the prediction accuracy of other models with a similar optimization time order of magnitude. The overall performance reaches Pareto optimality.

[0045] In some embodiments, step S40 includes: S41: determining that the training quality of the attention network model meets the requirements based on the information entropy and attention concentration of the attention weight matrices of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset recorded in step S30; S42: determining the complex features, interactive features, and the importance ranking, functional grouping, and linear determination coefficient of the features based on the attention weight matrices of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset recorded in step S30; S43: determining the influence mechanism of the alloy's features on its performance based on the complex features, interactive features, and the importance ranking, functional grouping, and linear determination coefficient of the features.

[0046] In this embodiment, complex features represent features with a high degree of complexity in influencing alloy properties; interactive features represent a group of features with synergistic interactive effects on alloy properties; feature importance ranking represents the ranking of the importance of different features in influencing alloy properties; feature functional grouping represents feature grouping according to the different influence patterns of features on alloy properties; and the linear determination coefficient of a feature represents the individual explanatory power of a feature on alloy properties. Determining complex features, interactive features, feature importance ranking, functional grouping, and linear determination coefficients can reveal the influence mechanism of alloy features on alloy properties from multiple dimensions and in depth. This allows for multi-dimensional and in-depth consideration during alloy design, enabling the determination of the composition and process parameters of the alloy to be designed, thus helping to solve more challenging alloy design problems.

[0047] Pre-determining that the training quality of the attention network model meets the requirements ensures that the model's decision-making process is relatively reliable. The attention weight matrix of all attention layers and attention heads corresponding to each set of sample alloy data saved during the prediction process can more realistically reflect the model's attention to different features. As a result, the complex features, interactive features, and the importance ranking, functional grouping, and linear determination coefficients of features obtained on this basis are also more realistic and reliable.

[0048] In some embodiments, in step S41, when the number of high-concentration, low-entropy attention heads obtained in step S30 is averaged with the number of low-concentration, high-entropy attention heads, the training quality of the attention network model is determined to meet the requirements.

[0049] In this embodiment, the information entropy of the attention weight matrix is ​​used to measure the dispersion of the attention distribution; a lower entropy value indicates that attention is more concentrated on a few features. The attention concentration of the attention weight matrix is ​​defined as the Gini coefficient or maximum weight percentage of the weight distribution, used to measure the inequality of the attention distribution; a higher attention concentration indicates a more unequal attention distribution. High-concentration, low-entropy attention heads tend to focus on a few key features, while low-concentration, high-entropy attention heads consider multiple features comprehensively. When the number of high-concentration, low-entropy attention heads and low-concentration, high-entropy attention heads is average, it indicates that the training quality of the attention network model is good and meets the requirements.

[0050] In some embodiments, in step S42, determining the importance ranking of features based on the attention weight matrices of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset recorded in step S30 includes: S4211: averaging the attention weight matrices of each feature in all layers and all attention heads across all groups of sample alloy data in the test dataset recorded in step S30 to determine the average attention score of each feature; S4212: determining the integral gradient of each feature; S4213: determining the Shapley value of each feature; S4214: normalizing the average attention score, integral gradient, and Shapley value of each feature. The system performs a process to determine the correlation coefficients between the average attention score and the integral gradient, the average attention score and the Shapley value, and the integral gradient and the Shapley value for each feature; S4215: When the correlation coefficients between the average attention score and the integral gradient, the average attention score and the Shapley value, and the integral gradient and the Shapley value for each feature are all greater than or equal to a first preset threshold, the average values ​​of the average attention score, integral gradient, and Shapley value for the feature are determined. Based on the magnitude of the average values, all features for which the three correlation coefficients are greater than or equal to the first preset threshold are sorted to obtain the importance ranking of the features.

[0051] Attention-based intrinsic interpretation, gradient-based sensitivity analysis, and game-theory-based SHAP analysis are three feature attribution methods based on different mathematical principles. The average attention score, integral gradient, and Shapley value obtained from these three methods can all characterize the importance of each feature to alloy properties. Furthermore, the correlation coefficients between the average attention score and integral gradient, between the average attention score and Shapley value, and between the integral gradient and Shapley value can measure the consistency of the three feature attribution methods in judging the importance of the same feature.

[0052] In this embodiment, by ensuring that the correlation coefficients between the average attention score and the integral gradient, the average attention score and the Shapley value, and the integral gradient and the Shapley value are all greater than or equal to a first preset threshold, three feature attribution methods based on different mathematical principles can be used to cross-validate the importance of features in relation to alloy properties. This reduces the influence of single-method bias and ensures that the features have a high degree of importance in relation to alloy properties. Furthermore, by using the average of the feature's average attention score, integral gradient, and Shapley value as the ranking of feature importance, the feature importance scores obtained from the three feature attribution methods are integrated, resulting in a more reliable ranking.

[0053] Normalizing the average attention score, integral gradient, and Shapley value for each feature unifies their dimensions and numerical ranges to the same scale. Only on this basis can the correlation coefficients between the average attention score and integral gradient, the average attention score and Shapley value, and the integral gradient and Shapley value be determined, thus measuring the consistency of the three feature attribution methods in judging feature importance.

[0054] In some embodiments, the integral gradient and Shapley value of each feature are determined by those skilled in the art.

[0055] In some embodiments, each feature refers to each composition feature or each process parameter feature, and each composition feature refers to the mass fraction of chemical elements in each constituent alloy.

[0056] In some embodiments, the correlation coefficient is the Pearson correlation coefficient or the Spearman rank correlation coefficient.

[0057] In some embodiments, the first preset threshold can be determined by those skilled in the art as needed.

[0058] In some embodiments, in step S42, determining the functional grouping of features based on the attention weight matrices of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset recorded in step S30 includes: S4221: determining the average value of the attention weight matrix of each feature on all groups of sample alloy data; S4222: performing nonlinear dimensionality reduction projection on the average value of the attention weight matrix of each feature on all groups of sample alloy data, and using cluster analysis to identify the functional grouping of features.

[0059] In this embodiment, the average value of each feature across all groups of alloy data reflects the average level of attention each feature receives during the model prediction process. By using a nonlinear dimensionality reduction projection method to project the average attention weight matrix of each feature in the test dataset across all groups of alloy data onto a low-dimensional space, the coordinates of each feature in that low-dimensional space can be obtained. Cluster analysis is then used to identify natural clusters in the low-dimensional space, resulting in multiple feature groups. Features within each feature group exhibit similar patterns of attention in the model and have similar impacts on alloy properties.

[0060] Compared to other linear methods, nonlinear dimension reduction projection methods can preserve the local structure and nonlinear relationships in the attention weight matrix space, and can better capture the complex similarities between features that have similar influence on the alloy, thus enabling features that have similar influence on the alloy to be grouped together.

[0061] In some embodiments, in step S4221, the average attention weight matrix of each feature can be taken for any attention layer and any attention head on all groups of sample alloy data, or the average attention weight matrix of each feature can be taken for all attention layers and all attention heads on all groups of sample alloy data.

[0062] In some embodiments, t-SNE projection is used for nonlinear dimensionality reduction projection.

[0063] In some embodiments, in step S42, determining the linear determination coefficient of a feature based on the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset recorded in step S30 includes: S423: performing linear regression fitting on each feature and its corresponding performance prediction value to determine the linear determination coefficient of each feature with respect to alloy performance.

[0064] In such an embodiment, by fitting a linear regression to each feature and its corresponding performance prediction value, the linear determination coefficient of each feature with respect to the alloy properties can be determined. This allows us to determine the individual explanatory power of each feature for the alloy properties. The larger the linear determination coefficient, the stronger the individual explanatory power of the feature for the properties.

[0065] In some embodiments, in step S42, determining complex features based on the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset recorded in step S30 includes: S4241: fitting each feature and its corresponding performance prediction value with quadratic polynomial, cubic polynomial, spline regression, and random forest respectively, and determining the determination coefficient of each feature with respect to alloy performance under quadratic polynomial, cubic polynomial, spline regression, and random forest respectively; S4242: determining the largest determination coefficient of each feature with respect to alloy performance under quadratic polynomial, cubic polynomial, spline regression, and random forest, and determining the difference between the largest determination coefficient and the linear determination coefficient of the feature determined in step S423; S4243: when the difference between the largest determination coefficient of the feature and the linear determination coefficient of the feature determined in step S423 is greater than a second preset threshold, the feature is determined to be a complex feature.

[0066] In this embodiment, by determining the largest determination coefficient for each feature under quadratic polynomial, cubic polynomial, spline regression, and random forest conditions regarding alloy properties, and comparing the difference between the largest determination coefficient and the linear determination coefficient of the feature with a second preset threshold, features with a non-linear relationship to alloy properties—that is, features with a high degree of complexity in their impact on alloy properties—can be identified. Furthermore, when the largest determination coefficient is equal to the determination coefficient of the feature under quadratic polynomial, cubic polynomial, spline regression, or random forest conditions regarding alloy properties, the complexity of the feature's impact on alloy properties increases sequentially.

[0067] Quadratic and cubic polynomial regression are computationally efficient and interpretable, suitable for capturing monotonic or quadratic relationships. Spline regression outperforms quadratic and cubic polynomials in terms of local smoothness, making it suitable for capturing piecewise nonlinear relationships. Random forests can capture arbitrarily complex nonlinear relationships. These four models—quadratic polynomial, cubic polynomial, spline regression, and random forests—cover a wide range of nonlinear relationships with relatively complete and complex spectrums. By determining the determination coefficients of each feature regarding alloy properties under quadratic polynomial, cubic polynomial, spline regression, and random forest models respectively, the specific nonlinear relationships between features and alloy properties can be identified more accurately, improving the robustness of complex feature identification and offering greater interpretability and reliability compared to a single model.

[0068] In some embodiments, the second preset threshold can be determined by those skilled in the art based on the actual situation.

[0069] In some embodiments, in step S42, the interaction features are determined based on the attention weight matrices of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset recorded in step S30, including: S4251: Select the first component feature and the second component feature, and determine the performance prediction value corresponding to each of the first component feature and the second component feature; S4252: Fit a linear regression to the first component features and their corresponding performance predictions to determine the first linear coefficient of determination; and fit a linear regression to the second component features and their corresponding performance predictions to determine the second linear coefficient of determination. S4253: Multiply the value of the first component feature by the value of the second component feature to determine the interaction feature value between the first component feature and the second component feature, and use the attention network model to determine the performance prediction value corresponding to the interaction feature value; S4254: Fit the interaction eigenvalues ​​and their corresponding performance prediction values ​​by linear regression to determine the linear determination coefficients of the interaction eigenvalues ​​with respect to the alloy properties; S4255: When the linear determination coefficients of the interaction eigenvalues ​​with respect to alloy properties are all greater than the first linear determination coefficient and the second linear determination coefficient, the first compositional characteristic and the second compositional characteristic corresponding to the interaction eigenvalues ​​are determined as a set of interaction characteristics. S4256: Repeat steps S4251 to S4255 to identify all interaction features.

[0070] In such an embodiment, when the coefficients of linear determination of the interaction feature values ​​of the first component feature and the second component feature with respect to the alloy properties are both greater than the coefficients of linear determination of the first component feature and the second component feature with respect to the alloy properties, it indicates that the synergistic effect of the first component feature and the second component feature is stronger in explaining the alloy properties than any single component feature. Therefore, it is possible to identify the interaction feature that has a synergistic effect on the alloy properties and further quantify the synergistic effect between component features.

[0071] In some embodiments, the first component feature and the second component feature may be component features of interest to those skilled in the art, and the value of the first or second component feature is the mass fraction of the selected chemical element.

[0072] To further illustrate the effectiveness of the method for determining the composition and process parameters of the alloy to be designed in the embodiments of this application, the following examples are provided.

[0073] Example 1: Prediction of Yield Strength of Molybdenum Alloy In step S10, the raw dataset containing 874 sets of sample molybdenum alloy data is first obtained. Each set of sample molybdenum alloy data includes the mass fraction of 22 elements that make up the sample molybdenum alloy, 8 process parameters, and yield strength value.

[0074] Preprocessing was performed on 874 sets of molybdenum alloy sample data. First, the optimal number of clusters for these 874 sets of molybdenum alloy sample data was determined to be 5, and they were divided into 5 datasets of unequal size. The cluster centers of each dataset were determined. Based on the Mahalanobis distance from each set of sample alloy data in each dataset to the cluster center, 12 outlier sample alloy data sets with a Mahalanobis distance greater than 3 times the standard deviation were removed, retaining 862 sets of valid sample alloy data.

[0075] The mass fractions of alloying elements and process parameters were subjected to min-max normalization to scale them to the [0,1] interval, and the yield strength values ​​were transformed into a uniform distribution using fractional transformation. Furthermore, the 862 sets of valid alloy sample data were divided into training and testing datasets at an 8:2 ratio.

[0076] In step S20, an attention network model based on a multi-head attention mechanism is trained using the training dataset to obtain the trained attention network model. Storage nodes are then set up in the attention network model to record the attention weight matrices of all attention layers and all attention heads, as well as the information entropy and attention concentration of the attention weight matrices.

[0077] In step S30, the trained attention network model is used to predict the yield strength of the molybdenum alloy samples in the test dataset, determining the predicted yield strength value for each molybdenum alloy sample. During the prediction process, the attention network model records the attention weight matrix of all attention layers and all attention heads corresponding to each group of molybdenum alloy samples in the test dataset, as well as the information entropy and attention concentration of the attention weight matrix.

[0078] In step S40, the attention weight matrix of each feature in the test dataset recorded in step S30 is averaged across all layers and attention heads in all groups of sample alloy data to determine the average attention score for each feature. The integral gradient and Shapley value for each feature are also determined. These values ​​are then normalized, and the Pearson correlation coefficients between the average attention score and integral gradient (0.891), the average attention score and Shapley value (0.870), and the integral gradient and Shapley value (0.848) are all found to exceed a first predetermined threshold of 0.7. Based on these values, the average attention score, integral gradient, and Shapley value for each feature are averaged and ranked. The final ranking of the importance of features in influencing the yield strength of the alloy is: Temperature > Re > Mo > Ti > Other elements.

[0079] In step S50, the number of attention heads with high concentration and low entropy is relatively evenly distributed with the number of attention heads with low concentration and high entropy, resulting in good model training quality.

[0080] The average attention weight matrix of each feature across all sample alloys was subjected to t-SNE dimensionality reduction and projection onto a two-dimensional plane. The results showed that transition metal elements such as Re, Ti, and Mo all affect yield strength through solid solution strengthening mechanisms, and their influence patterns on yield strength are similar. Therefore, the model classifies them into the same functional group. Interstitial elements C, N, and O preferentially occupy interstitial sites in the Mo lattice, and their influence on yield strength is synergistic through interstitial solid solution strengthening and dislocation pinning mechanisms. Their influence patterns on yield strength are also similar, and therefore they are classified into the same functional group. The above functional grouping results are highly consistent with known physical metallurgical mechanisms.

[0081] The relationship between fitted temperature and yield strength, and its R-value under the linear regression model. 2 The value is 0.61, R under the quadratic polynomial model. 2 R is 0.75 under the cubic polynomial model. 2 The R-value is 0.76. Compared to the linear regression model, the quadratic polynomial model has a lower R-value of 0.76. 2 The significant increase indicates a quadratic nonlinear relationship between temperature and yield strength. The relationship between Re content and yield strength is fitted to determine its Ri under a linear regression model. 2 The value is 0.42, and the value under the quadratic polynomial model is 0.44. Compared with the linear regression model, the quadratic polynomial model has a lower R-value. 2 The small increase indicates an approximately linear relationship between Re and yield strength.

[0082] Select Re and Mo elements to determine the R of the Re-Mo interaction model. 2 The gain of 0.08 compared to the additive model of Re and Mo indicates that Re and Mo have an interactive effect on the yield strength of the alloy, forming an interactive characteristic. Similarly, Ti and Mo are also identified as an interactive characteristic.

[0083] For strength-driven alloy design goals, it is recommended that within a temperature window of 1200-1400℃, the Re content be controlled at 3-5 wt.%, the Mo matrix content be maintained at 85-90 wt.%, the Ti content be 1-2 wt.%, and the total amount of interstitial elements C, N, and O be strictly controlled to <100 ppm, which can achieve the design goal of yield strength >600 MPa.

[0084] Example 2: Prediction of elongation of molybdenum alloy In step S10, the raw dataset containing 589 sets of sample molybdenum alloy data is first obtained. Each set of sample molybdenum alloy data includes the mass fraction of 22 elements that make up the sample molybdenum alloy, 8 process parameters, and elongation value.

[0085] Preprocessing was performed on 589 sets of molybdenum alloy sample data. First, the optimal number of clusters for these 589 sets of molybdenum alloy sample data was determined to be 5, and they were divided into 5 datasets of unequal size. The cluster centers of each dataset were determined. Based on the Mahalanobis distance from each set of sample alloy data in each dataset to the cluster centers in the dataset, 15 outlier sample alloy data sets with a Mahalanobis distance greater than 3 times the standard deviation were removed, retaining 574 sets of valid sample alloy data sets.

[0086] The mass fractions of alloying elements and process parameters are subjected to minimum-maximum normalization to scale them to the [0,1] interval. The yield strength value is converted to a standard Gaussian distribution by fractional transformation.

[0087] The content of 22 chemical elements was converted into 21 physical parameters, including 16 traditional material parameters and 5 nonlinear physical parameters specifically designed for elongation. Furthermore, the 574 sets of valid sample alloy data were divided into training and testing datasets at an 8:2 ratio.

[0088] In step S20, an attention network model based on a multi-head attention mechanism is trained using the training dataset to obtain the trained attention network model. Storage nodes are then set up in the attention network model to record the attention weight matrices of all attention layers and all attention heads, as well as the information entropy and attention concentration of the attention weight matrices.

[0089] Specifically, the first attention network model is trained using component feature data, process parameter data, and corresponding elongation rates, while the second attention network model is trained using physical parameter features, process parameter data, and corresponding elongation rates.

[0090] Ablation experiments were used to verify that the R-value of the attention network model after introducing five nonlinear physical parameters was improved. 2 The value is 0.8400, compared to an R value of 0.6769 for the attention network model using only 16 traditional physical parameter features. 2 There has been an improvement, with the root mean square error decreasing by 29.7%.

[0091] In step S30, the trained attention network model is used to predict the elongation of the molybdenum alloy samples in the test dataset, determining the predicted elongation value for each molybdenum alloy sample. During the prediction process, the attention network model records the attention weight matrix of all attention layers and attention heads corresponding to each group of molybdenum alloy samples in the test dataset, as well as the information entropy and attention concentration of the attention weight matrix.

[0092] In step S40, the effect of temperature on elongation is determined to be the most important, while the effects of carbon and nitrogen decrease in importance.

[0093] In step S50, the number of attention heads with high concentration and low entropy is relatively evenly distributed with the number of attention heads with low concentration and high entropy, resulting in good model training quality.

[0094] It was determined that there is a quadratic nonlinear relationship between Mo content and elongation, and that Re and Mo have a very strong positive interaction effect.

[0095] For alloys designed to be ductile-dominant, it is recommended that the Mo content be 97%-98%, the Re content be 0.5%-1%, the total C and N content be <0.01%, the O content be <0.002%, and the annealing temperature be 1300-1500℃.

[0096] Embodiments of this application also provide a system for determining the composition and process parameters of an alloy to be designed, comprising: Data acquisition module and processor, The data acquisition module is set to acquire training datasets and test datasets. Both training datasets and test datasets include multiple sets of sample alloy data. Each set of sample alloy data includes the feature data and performance data of the sample alloy. Processor settings: Based on the training dataset, an attention network model based on a multi-head attention mechanism is trained. The feature data of the sample alloy is used as the input of the attention network model, and the performance data of the sample alloy is used as the output of the attention network model to obtain the trained attention network model. The attention network model includes storage nodes to record the attention weight matrices of all attention layers and all attention heads, as well as the information entropy and attention concentration of the attention weight matrices. The trained attention network model is used to predict the performance of sample alloys in the test dataset, and the predicted performance values ​​corresponding to the feature data of the sample alloys in the test dataset are determined. During the prediction process, the attention network model records the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset, as well as the information entropy and attention concentration of the attention weight matrix; Based on the attention weight matrices of all attention layers and all attention heads corresponding to each group of sample alloy data in the recorded test dataset, as well as the information entropy and attention concentration of the attention weight matrices, the influence mechanism of the alloy's characteristics on its performance is determined. Based on the influence mechanism of alloy characteristics on its properties, and the target properties of the alloy to be designed, the composition and process parameters of the alloy to be designed are determined.

[0097] The system for determining the composition and process parameters of an alloy to be designed, provided in the embodiments of this application, can obtain the degree of attention the model pays to different input features when generating performance prediction results by recording the attention weight matrix calculated by the model during the prediction process. This allows for the determination of the influence mechanism of alloy features on its performance, efficiently and accurately determining the composition and process parameters of the alloy that meet the target performance. Since this explanatory information, the attention weight matrix, is generated synchronously with the prediction process, it achieves the essential interpretability of the model, avoiding the problems of fidelity, stability, and misleadingness in post-interpretation. The influence mechanism of alloy features on its performance determined based on the model's inherent explanatory information is also more accurate and reliable. The information entropy of the attention weight matrix is ​​used to measure the dispersion of the attention distribution, while the attention concentration of the attention weight matrix is ​​defined as the Gini coefficient or maximum weight percentage of the weight distribution, used to measure the inequality of the attention distribution. Considering the information entropy and attention concentration of the attention weight matrix when determining the influence mechanism of alloy features on its performance can determine the training quality of the attention network model, thereby ensuring high training quality, accurate prediction results, and reliable decision-making process.

[0098] Embodiments of this application also provide a computer storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the method provided in any embodiment of the first aspect of this application.

[0099] Regarding the embodiments of this application, it should also be noted that, without conflict, the embodiments of this application and the features in the embodiments can be combined with each other to obtain new embodiments.

[0100] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. The scope of protection of this application shall be determined by the scope of the claims.

Claims

1. A method for determining the composition and process parameters of an alloy to be designed, characterized in that, It includes: S10: Determine the training dataset and the test dataset, wherein both the training dataset and the test dataset include multiple sets of sample alloy data, and each set of sample alloy data includes the feature data and performance data of the sample alloy; S20: Based on the training dataset determined in step S10, train an attention network model based on a multi-head attention mechanism. The feature data of the sample alloy is used as the input of the attention network model, and the performance data of the sample alloy is used as the output of the attention network model to obtain the trained attention network model. The attention network model includes storage nodes to record the attention weight matrices of all attention layers and all attention heads, as well as the information entropy and attention concentration of the attention weight matrices. S30: Using the trained attention network model, predict the performance of the sample alloys in the test dataset, and determine the predicted performance values ​​corresponding to the feature data of the sample alloys in the test dataset. During the prediction process, the attention network model records the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset, as well as the information entropy and attention concentration of the attention weight matrix; S40: Based on the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset recorded in step S30, as well as the information entropy and attention concentration of the attention weight matrix, determine the influence mechanism of the alloy's characteristics on its performance. S50: Based on the influence mechanism of the alloy's characteristics on its properties and the target properties of the alloy to be designed, determine the composition and process parameters of the alloy to be designed.

2. The method according to claim 1, characterized in that, Step S10 includes: S11: Obtain the original dataset, which includes the multiple sets of sample alloy data. Each set of sample alloy data includes the characteristic data and performance data of the sample alloy. The characteristic data includes the compositional characteristic data and process parameter characteristic data of the sample alloy. S12: Remove abnormal sample alloy data from the original dataset, and process the sample alloy data after removing abnormal sample alloy data so that the processed sample alloy data can be used in the attention network model. S13: The processed sample alloy data is divided into a training dataset and a test dataset. Both the training dataset and the test dataset include multiple sets of sample alloy data. Each set of sample alloy data includes the characteristic data and performance data of the sample alloy. The characteristic data includes the compositional characteristic data and process parameter characteristic data of the sample alloy.

3. The method according to claim 2, characterized in that, Step S12 includes: S121: Determine the optimal number of clusters for the original dataset; S122: Based on the optimal number of clusters, divide the sample alloy data in the original dataset into datasets of unequal sizes, and determine the cluster center for each dataset. Each of the datasets includes multiple sets of sample alloy data; S123: Based on the distance from each group of sample alloy data in each dataset to its cluster center, and the size of the dataset, identify and remove all abnormal sample alloy data in the dataset; S124: Process the sample alloy data in the dataset after removing abnormal sample alloy data so that the processed sample alloy data can be used in the attention network model.

4. The method according to claim 3, characterized in that, Step S13 includes: S131: Based on the compositional characteristic data of the sample alloy data in the processed dataset, determine the physical parameter characteristic data of all the sample alloys; S132: Divide the sample alloy data in the processed dataset into a training dataset and a test dataset. The training dataset and the test dataset both include multiple sets of sample alloy data. Each set of sample alloy data includes the characteristic data and performance data of the sample alloy. The characteristic data includes the compositional characteristic data, physical parameter characteristic data and process parameter characteristic data of the sample alloy.

5. The method according to claim 3 or 4, characterized in that, Step S40 includes: S41: Based on the information entropy and attention concentration of the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset recorded in step S30, determine that the training quality of the attention network model meets the requirements. S42: Based on the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset recorded in step S30, determine the complex features, interactive features, and the importance ranking, functional grouping, and linear determination coefficient of the features; S43: Determine the influence mechanism of the alloy's characteristics on its performance based on the complex features, the interactive features, the importance ranking of the features, functional grouping, and linear determination coefficients.

6. The method according to claim 5, characterized in that, In step S41, when the number of high-concentration, low-entropy attention heads obtained in step S30 is averaged with the number of low-concentration, high-entropy attention heads, the training quality of the attention network model is determined to meet the requirements.

7. The method according to claim 6, characterized in that, In step S42, determining the importance ranking of the features based on the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset recorded in step S30 includes: S4211: Take the average of the attention weight matrices of each feature in the test dataset recorded in step S30 across all layers and attention heads in all group sample alloy data to determine the average attention score of each feature; S4212: Determine the integral gradient of each feature; S4213: Determine the Shapley value for each of the aforementioned features; S4214: Normalize the average attention score, integral gradient, and Shapley value for each feature, and determine the correlation coefficient between the average attention score and integral gradient, the correlation coefficient between the average attention score and Shapley value, and the correlation coefficient between the integral gradient and Shapley value for each feature. S4215: When the correlation coefficients between the average attention score and the integral gradient, the average attention score and the Shapley value, and the integral gradient and the Shapley value of the feature are all greater than or equal to a first preset threshold, the average value of the average attention score, integral gradient, and Shapley value of the feature is determined. Based on the magnitude of the average value, all features whose three correlation coefficients are greater than or equal to the first preset threshold are sorted to obtain the importance ranking of the feature.

8. The method according to claim 6, characterized in that, In step S42, determining the functional grouping of the feature based on the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset recorded in step S30 includes: S4221: Determine the average value of the attention weight matrix for each feature across all groups of sample alloy data; S4222: Perform nonlinear dimensionality reduction projection on the average value of the attention weight matrix of each feature on all groups of sample alloy data, and use cluster analysis to identify the functional grouping of the feature.

9. The method according to claim 6, characterized in that, In step S42, determining the linear determination coefficient of the feature based on the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset recorded in step S30 includes: S423: Perform linear regression fitting on each feature and its corresponding performance prediction value to determine the linear determination coefficient of each feature with respect to the alloy properties.

10. The method according to claim 9, characterized in that, In step S42, determining the complex features based on the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset recorded in step S30 includes: S4241: Fit each feature and its corresponding performance prediction value using quadratic polynomial, cubic polynomial, spline regression and random forest respectively, and determine the determination coefficient of each feature with respect to alloy performance under quadratic polynomial, cubic polynomial, spline regression and random forest respectively. S4242: Determine the largest coefficient of determination for each feature among the coefficients of determination for alloy properties under quadratic polynomial, cubic polynomial, spline regression and random forest, and determine the difference between the largest coefficient of determination and the linear coefficient of determination of the feature determined in step S423. S4243: When the difference between the maximum coefficient of determination of the feature and the linear coefficient of determination of the feature determined in step S423 is greater than the second preset threshold, the feature is determined to be a complex feature.

11. The method according to claim 6, characterized in that, In step S42, determining the interaction features based on the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset recorded in step S30 includes: S4251: Select the first component feature and the second component feature, and determine the performance prediction value corresponding to the first component feature and the second component feature respectively; S4252: Fit the first component features and their corresponding performance prediction values ​​to a linear regression to determine the first linear determination coefficient, and fit the second component features and their corresponding performance prediction values ​​to a linear regression to determine the second linear determination coefficient. S4253: Multiply the value of the first component feature by the value of the second component feature to determine the interaction feature value between the first component feature and the second component feature, and use the attention network model to determine the performance prediction value corresponding to the interaction feature value; S4254: Perform linear regression fitting on the interactive feature values ​​and their corresponding performance prediction values ​​to determine the linear determination coefficient of the interactive feature values ​​with respect to the alloy properties; S4255: When the linear determination coefficients of the interaction feature values ​​with respect to the alloy properties are all greater than the first linear determination coefficient and the second linear determination coefficient, the first composition feature and the second composition feature corresponding to the interaction feature value are determined as a set of interaction features; S4256: Repeat steps S4251 to S4255 to determine all the said interaction features.

12. A system for determining the composition and process parameters of an alloy to be designed, characterized in that, It includes: Data acquisition module and processor; The data acquisition module is configured to acquire a training dataset and a test dataset. Both the training dataset and the test dataset include multiple sets of sample alloy data. Each set of sample alloy data includes the feature data and performance data of the sample alloy. The processor is configured as follows: Based on the training dataset, an attention network model based on a multi-head attention mechanism is trained. The feature data of the sample alloy is used as the input of the attention network model, and the performance data of the sample alloy is used as the output of the attention network model to obtain the trained attention network model. The attention network model includes storage nodes to record the attention weight matrices of all attention layers and all attention heads, as well as the information entropy and attention concentration of the attention weight matrices. Using the trained attention network model, the performance of the sample alloys in the test dataset is predicted, and the predicted performance values ​​corresponding to the feature data of the sample alloys in the test dataset are determined. During the prediction process, the attention network model records the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the test dataset, as well as the information entropy and attention concentration of the attention weight matrix; Based on the attention weight matrix of all attention layers and all attention heads corresponding to each group of sample alloy data in the recorded test dataset, as well as the information entropy and attention concentration of the attention weight matrix, the influence mechanism of the alloy's characteristics on its performance is determined. Based on the influence mechanism of the alloy's characteristics on its properties, and the target properties of the alloy to be designed, the composition and process parameters of the alloy to be designed are determined.

13. A computer storage medium, characterized in that, It stores a computer program thereon, which is executed by a processor to implement the method as described in any one of claims 1-11.

Images