Alloy component reverse design method and system based on knowledge distillation

Abstract

The invention discloses an alloy component reverse design method and system based on knowledge distillation, and relates to the field of material gene and alloy reverse design, and the method comprises the steps: collecting alloy component data and corresponding performance data, carrying out the arrangement to obtain a data set, and calculating the atomic characteristics of each alloy in the data set; constructing and training a supervised learning model based on the atomic characteristics and the performance data, and establishing a mapping relation from the atomic characteristics of the alloy to the performance; a deep neural network model is trained based on the component data and the performance data, and a mapping relation from alloy components to performance is constructed; in the deep neural network model training process, a back propagation algorithm is adopted to adjust network parameters, and the prediction difference between the deep neural network model and the supervised learning model is minimized; and sampling optimization is carried out, component optimization is carried out through gradient information of the deep neural network model, and the optimal alloy component is searched. According to the method, the components and atomic characteristics of the alloy are combined, the components of the high-performance alloy are analyzed, and customized design and optimization of the material are achieved.

Description

Technical Field

[0001] The present invention relates to the field of material genes and alloy reverse design, and more particularly to a method and system for alloy composition reverse design based on knowledge distillation. Background Art

[0002] Compared with single metals, alloys often exhibit better physical, chemical and mechanical properties, such as high hardness, heat resistance and corrosion resistance. These properties make alloy materials widely used in many industries such as aerospace, energy, automobile manufacturing, and electronic equipment.

[0003] However, traditional alloy design relies on experience and trial and error, which is not only time-consuming and costly, but also difficult to fully explore the complex relationship between material properties and composition. In contrast, machine learning algorithms can analyze data, learn and predict the intrinsic relationship between alloy composition and its properties, thereby efficiently screening alloy materials with desired properties. However, an important challenge facing alloy design is how to reversely infer and solve its composition based on the design target performance of the alloy, which is the so-called "reverse design". When faced with the vast material space of multi-component alloys, parsing the composition of potential high-performance alloys through inverse design methods is crucial to achieving customized design and optimization of materials.

[0004] Therefore, how to develop an alloy composition reverse design method is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides an alloy composition reverse design method and system based on knowledge distillation, aiming to construct a set of efficient alloy performance prediction and composition reverse design methods. By combining knowledge distillation technology with machine learning models, it provides strong technical support for the design of alloy materials and realizes the precise design of high-performance alloy compositions.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] The present invention discloses a method for reverse design of alloy composition based on knowledge distillation, the specific steps of which are as follows:

[0008] Step 1: Collect chemical composition data and corresponding performance data of multi-component alloys, organize them to obtain a multi-component alloy data set, and calculate the atomic characteristics of each alloy in the multi-component alloy data set;

[0009] Step 2: Based on the atomic characteristics and performance data of each alloy, a supervised learning model is constructed and trained to establish a mapping relationship from the atomic characteristics of the alloy to the performance;

[0010] Step 3: Based on the chemical composition data and the performance data of each alloy, a deep neural network model is trained to construct a mapping relationship between alloy composition and performance; during the training process of the deep neural network model, a back propagation algorithm is used to adjust network parameters to minimize the prediction difference between the deep neural network model and the supervised learning model;

[0011] Step 4: Use LangevinDynamics to perform sampling optimization, and use the gradient information of the deep neural network model to optimize the composition to find the optimal alloy composition.

[0012] Furthermore, each alloy in the multi-alloy dataset is represented by X i =[x1,x2,……,x m ], satisfying ∑x j =1; the corresponding performance index is expressed as Y i ; Among them, X i is the composition vector representing the i-th alloy composition in the data set, i=1,2,...,N, N is the total number of alloys in the data set, X i Composed of m elements, x j It represents the molar percentage of the jth element.

[0013] Furthermore, the atomic characteristics include: average electronegativity, average ionization energy and atomic size difference.

[0014] Furthermore, the mapping relationship from the alloy atomic characteristics to the performance is expressed as in represents the predicted value of the supervised learning model, and f(x) represents the supervised learning model;

[0015] The mapping relationship from alloy composition to performance is expressed as in represents the predicted value of the deep neural network model, g(x) represents the deep neural network model; the deep neural network model is a four-layer fully connected network, the input dimension is the number of types of alloy elements in the data set, and the output is the target performance.

[0016] Furthermore, the step 4 specifically includes:

[0017] Step 4.1, determine the sampling starting point according to the alloy composition to be optimized;

[0018] Step 4.2, inputting the composition vector of the alloy to be optimized into the deep neural network model to obtain a predicted value;

[0019] Step 4.3: Calculate the gradient information of the current predicted value and calculate the optimized alloy composition sampling points based on the gradient information. If the number of sampling times is less than the preset threshold, return to step 4.2; otherwise, proceed to step 4.4.

[0020] Step 4.4, calculate the error between the predicted value of the supervised learning model and the predicted value of the deep neural network model for all sampling points; determine whether the stopping condition is met, if so, select the optimal alloy composition from all sampling points, otherwise add the chemical composition data of all sampling points and the corresponding supervised learning model predicted values ​​as new samples to the multivariate alloy data set, and return to step 3.

[0021] Furthermore, the sampling point is represented by X i,k =[x1,x2,…x j …,x m ], where X i,k represents the composition vector of the kth sampling of the i-th alloy in the data set, consisting of m elements, x j represents the molar percentage of the jth element;

[0022] The formula for calculating the optimized alloy composition sampling points is: i,k+1 =X i,k +λ*grad; where λ>0 is the learning rate, grad represents the gradient information, express x j The partial derivative of .

[0023] Furthermore, the error is expressed as in represents the predicted value of the supervised learning model, represents the predicted value of the deep neural network model;

[0024] The stopping condition is: there is a sampling point X i,k+1 ,satisfy and Achieve the target performance requirement or reach the maximum number of iterations, where ε represents the error factor of the supervised learning model and the deep neural network model.

[0025] The present invention also discloses an alloy composition reverse design system based on knowledge distillation, comprising:

[0026] Preprocessing module: collects chemical composition data and corresponding performance data of multi-component alloys, organizes them into a multi-component alloy data set, and calculates the atomic characteristics of each alloy in the multi-component alloy data set;

[0027] Supervised learning module: Based on the atomic characteristics and performance data of each alloy, a supervised learning model is constructed and trained to establish a mapping relationship from the atomic characteristics of the alloy to the performance;

[0028] Deep learning module: Based on the chemical composition data and performance data of each alloy, a deep neural network model is trained to construct a mapping relationship between alloy composition and performance; during the training process of the deep neural network model, a back propagation algorithm is used to adjust network parameters to minimize the prediction difference between the deep neural network model and the supervised learning model;

[0029] Optimization module: Langevin Dynamics is used for sampling optimization, and the gradient information of the deep neural network model is used for composition optimization to find the optimal alloy composition.

[0030] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a method and system for reverse design of alloy components based on knowledge distillation, which uses deep neural networks and gradient optimization algorithms to perform rapid search and optimization in high-dimensional alloy component space, thereby quickly finding alloy components that meet the target performance, and improving the search efficiency compared to high-throughput forward search. Secondly, in the optimization process, knowledge distillation and data enhancement are achieved by continuously adding newly generated alloy components and their predicted performance to the training data set, which not only improves the prediction accuracy and generalization ability of the DNN model, but also makes the composition feature model gradually approach the performance of the atomic parameter feature model, further optimizing the reverse design effect. The present invention combines the composition and atomic parameter characteristics of the alloy at the same time, and uses DNN to approximate the composition feature model with the atomic parameter feature model, parsing the potential component composition of high-performance alloys, and realizing customized design and optimization of materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0032] Figure 1 Schematic diagram of the overall process of an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0034] The embodiment of the present invention discloses a method for reverse design of alloy composition based on knowledge distillation, such as Figure 1 The specific steps are as follows:

[0035] Step 1: Collect chemical composition data and corresponding performance data of multi-component alloys, organize them into a multi-component alloy dataset, and calculate the atomic characteristics of each alloy in the multi-component alloy dataset;

[0036] Step 2: Based on the atomic characteristics and performance data of each alloy, a supervised learning model is constructed and trained to establish a mapping relationship between the alloy atomic characteristics and performance;

[0037] Step 3: Based on the chemical composition and performance data of each alloy, a deep neural network model is trained to construct a mapping relationship between alloy composition and performance. During the deep neural network model training process, a backpropagation algorithm is used to adjust the network parameters to minimize the prediction difference between the deep neural network model and the supervised learning model.

[0038] Step 4: Use LangevinDynamics to perform sampling optimization and optimize the composition through the gradient information of the deep neural network model to find the optimal alloy composition.

[0039] In a specific embodiment, each alloy in the multi-alloy data set is represented by X i =[x1,x2,……,x m ], satisfying ∑x j =1; the corresponding performance index is expressed as Y i ; Among them, X i is the composition vector representing the i-th alloy composition in the data set, i=1,2,...,N, N is the total number of alloys in the data set, X i Composed of m elements, x j It represents the molar percentage of the jth element.

[0040] In a specific embodiment, atomic features include average electronegativity, average ionization energy, and atomic size differences. Because they incorporate prior knowledge about physics and chemistry, atomic features often offer higher reliability than models based solely on composition. However, atomic features cannot be used to reverse-engineer the composition. Knowledge distillation technology is introduced to address this issue.

[0041] In a specific embodiment, the mapping relationship from alloy atomic characteristics to properties is expressed as in Represents the predicted value of the supervised learning model, f(x) represents the supervised learning model, which is the teacher model of the subsequent knowledge distillation model; the supervised learning model is built based on the supervised learning algorithm, such as support vector machine, random forest or gradient boosting tree.

[0042] The mapping relationship from alloy composition to performance is expressed as in represents the predicted value of the deep neural network model, and g(x) represents the deep neural network model; g(x) serves as the student model of knowledge distillation, aiming to imitate the teacher model's ability to predict alloy properties; during training, the backpropagation algorithm is used to adjust the network parameters to minimize the predicted difference between f(x) and g(x); the deep neural network model is a four-layer fully connected network, the input dimension is the number of alloy elements in the data set, and the output is the target performance.

[0043] In a specific embodiment, step 4 specifically includes:

[0044] Step 4.1: Determine the sampling starting point based on the alloy composition to be optimized. The sampling starting point composition vector is represented by X i,0 =[x1,x2,……,x m ], the number of sampling is n;

[0045] Step 4.2: Input the composition vector of the alloy to be optimized into the deep neural network model to obtain the predicted value;

[0046] Step 4.3: Calculate the gradient information of the current predicted value and calculate the optimized alloy composition sampling points based on the gradient information. If the number of sampling times is less than the preset threshold n, return to step 4.2; otherwise, proceed to step 4.4.

[0047] In step 4.4, the error between the predicted values ​​of the supervised learning model and the predicted values ​​of the deep neural network model is calculated for all sampling points. The stopping condition is determined. If so, the optimal alloy composition is selected from all sampling points. Otherwise, the chemical composition data of all sampling points and the corresponding predicted values ​​of the supervised learning model are added as new samples to the multivariate alloy dataset, and the process returns to step 3.

[0048] Specifically, in order to find the optimal alloy composition in the high-dimensional composition space, Langevin Dynamics is used for composition optimization. Initially, an alloy composition X in the data set is selected. i As the sampling starting point, input into the DNN model g(x) to obtain the prediction performance Then, the gradient of the predicted performance with respect to the molar percentage of each component is calculated, that is, the partial derivative of the performance prediction value with respect to each component. Based on the gradient information, the component vector is updated. This process is iterated continuously, and the predicted performance is checked after each iteration. If the difference with the previous performance is less than the preset threshold ε and meets the target performance or reaches the preset maximum number of iterations, it is considered that a potential optimization component is found, otherwise the optimization continues.

[0049] During the optimization process, the newly generated alloy composition and its predicted properties are This process dynamically expands the dataset by adding the new data to the original dataset D. This newly added data helps the DNN model g(x) better simulate the behavior of f(x), gradually improving the model's prediction accuracy and generalization capabilities. This process, known as knowledge distillation, enables the DNN model to not only approach the performance of the base model but also further optimize it, achieving more efficient reverse engineering.

[0050] In a specific embodiment, the sampling point is represented by X i,k =[x1,x2,…x j …,x m ], where X i,k represents the composition vector of the kth sampling of the i-th alloy in the data set, consisting of m elements, x j represents the molar percentage of the jth element;

[0051] The formula for calculating the optimized alloy composition sampling point is: X i,k+1 =X i,k +λ*grad; where λ>0 is the learning rate, grad represents the gradient information, express x j The partial derivative of .

[0052] After n samplings, an optimized trajectory with a total of n component sampling points is obtained.

[0053] In a specific embodiment, the error is expressed as in represents the predicted value of the supervised learning model, Represents the predicted value of the deep neural network model;

[0054] The stopping condition is: there is a sampling point X i,k satisfy and Achieve target performance requirements or reach the maximum number of iterations, where ε represents the error factor of the supervised learning model and the deep neural network model. The target performance requirements include, but are not limited to, alloy predictability and alloy composition mole percentage limits.

[0055] When the stop condition is not met, the alloy composition data of the sampling point of this iteration and the corresponding The values ​​are added to the dataset as new sample data, and the deep neural network model is retrained. The gradually added data enables the deep neural network model g(x) to learn the knowledge in the supervised learning model f(x), gradually become its proxy model, and more accurately calculate the gradient of alloy performance to composition.

[0056] Based on reliable literature data and modeling methods, the alloy melting point prediction model developed in this paper is simple, fast, low-cost, and pollution-free. Through knowledge distillation, the DNN model can flexibly explore the alloy composition space based on gradient information. By learning from the performance advantages and disadvantages of existing alloys, the DNN model can propose novel composition combinations, promoting the rapid development of alloy materials.

[0057] In a specific embodiment, the design requirement is to use an alloy composition with a melting point < 7 as an example for explanation. In step 1, the data comes from 150 low-melting-point alloys with different compositions reported in the literature, and some of the data examples are shown in Table 1:

[0058] Table 1 Low melting point alloy data table

[0059] Chemical formula Melting point Ga71.9In17.9Sn9.0Zn0.13 8.000 Ga78.3In14.9Sn6.8 10.000 In60.1Sn18.8Bi21.1 63.000 Sn56Bi44 139.060 In34.3Sn22.8Bi43.0 80.700 In60.7Sn17.8Bi21.5 60.420 Sn33.1Bi42.0 101.130 Sn33.1Bi42.7Pb24.1Cd0.1 97.060 Sn33.2Bi38.5Ag7.8 96.840 Bi48.0Pb37.8Ag14.2 92.550 Ga54.5Ge11.1Tl34.4 92.970 In30.2Sn16.7Bi38.7Pb14.4 144.990 Sn19.7Bi42.0Pb22.6Sb15.6 59.730

[0060] In step 4, Ga71.9In17.9Sn9.0Zn0.13 was selected as the starting point for the Langevin Dynamics sampling optimization. Table 2 shows an example of the trajectory of the sampling optimization process for this sample. The design sample obtained in this optimization is the sample of the eighth iteration.

[0061] Table 2 Sampling optimization trajectory of Ga-In-Sn-Zn reverse design

[0062]

[0063]

[0064] This example demonstrates the ability of this technology to rapidly determine the composition of a designed low-melting-point alloy based on a model.

[0065] In a specific embodiment, the process is basically the same as the above embodiment, except that the starting point of the optimization is the five-element alloy and the Pb content is set to not exceed 0.15. Table 3 is an example of the trajectory of the sample sampling optimization process.

[0066] Table 3 Ga-In-Sn-Zn-Pb reverse design sampling optimization trajectory

[0067]

[0068] This embodiment is based on the knowledge distillation alloy composition reverse design method, which shows that this method can flexibly set the composition gradient and add composition optimization constraints to meet the design requirements of different alloy material constraints.

[0069] The embodiment of the present invention discloses an alloy composition reverse design system based on knowledge distillation, comprising:

[0070] Preprocessing module: collects chemical composition data and corresponding performance data of multi-component alloys, organizes them into multi-component alloy data sets, and calculates the atomic characteristics of each alloy in the multi-component alloy data sets;

[0071] Supervised learning module: Based on the atomic characteristics and performance data of each alloy, a supervised learning model is constructed and trained to establish a mapping relationship between the alloy atomic characteristics and performance;

[0072] Deep Learning Module: Based on the chemical composition and performance data of each alloy, a deep neural network model is trained to construct a mapping relationship between alloy composition and performance. During the deep neural network model training process, a backpropagation algorithm is used to adjust network parameters to minimize the prediction differences between the deep neural network model and the supervised learning model.

[0073] Optimization module: Use Langevin Dynamics for sampling optimization and use the gradient information of the deep neural network model to optimize the composition and find the optimal alloy composition.

[0074] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for reverse design of alloy composition based on knowledge distillation, characterized in that: The specific steps are as follows: Step 1, collecting chemical composition data and corresponding performance data of multi-element alloys, collating to obtain a multi-element alloy data set, and calculating the atomic characteristics of each alloy in the multi-element alloy data set; Step 2, based on the atomic characteristics and performance data of each alloy, construct and train a supervised learning model to establish a mapping relationship from alloy atomic characteristics to performance; Step 3: Based on the chemical composition data and the performance data of each alloy, a deep neural network model is trained to construct a mapping relationship from alloy composition to performance; during the training process of the deep neural network model, a back propagation algorithm is used to adjust network parameters to minimize the prediction difference between the deep neural network model and the supervised learning model; Step 4: Use LangevinDynamics to perform sampling optimization, and use the gradient information of the deep neural network model to optimize the composition to find the optimal alloy composition.

2. The alloy composition reverse design method based on knowledge distillation according to claim 1, characterized in that: Each alloy in the multi-alloy data set is represented by X i =[x1,x2,……,x m ], satisfying ∑x j =1; the corresponding performance index is expressed as Y i ; Among them, X i is the composition vector representing the i-th alloy composition in the data set, i = 1, 2, ..., N, N is the total number of alloys in the data set, X i It is composed of m elements, x j It represents the molar percentage of the jth element.

3. The alloy composition reverse design method based on knowledge distillation according to claim 1, characterized in that: The atomic characteristics include: average electronegativity, average ionization energy and atomic size differences.

4. The alloy composition reverse design method based on knowledge distillation according to claim 1, characterized in that: The mapping relationship from the alloy atomic characteristics to the properties is expressed as in represents the predicted value of the supervised learning model, and f(x) represents the supervised learning model; The mapping relationship from alloy composition to performance is expressed as in represents the predicted value of the deep neural network model, g(x) represents the deep neural network model; the deep neural network model is a four-layer fully connected network, the input dimension is the number of types of alloy elements in the data set, and the output is the target performance.

5. The alloy composition reverse design method based on knowledge distillation according to claim 1, characterized in that: The step 4 specifically includes: Step 4.1, determine the sampling starting point according to the alloy composition to be optimized; Step 4.2, inputting the composition vector of the alloy to be optimized into the deep neural network model to obtain a predicted value; Step 4.3, calculate the gradient information of the current predicted value, and calculate the optimized alloy composition sampling points according to the gradient information. If the sampling number is less than the preset threshold, return to step 4.2, otherwise go to step 4.4; Step 4.4, calculate the error between the predicted value of the supervised learning model of all sampling points and the predicted value of the deep neural network model; determine whether the stopping condition is met, if so, select the optimal alloy composition from all sampling points, otherwise add the chemical composition data of all sampling points and the corresponding supervised learning model predicted values ​​as new samples to the multivariate alloy data set, and return to step 3.

6. The alloy composition reverse design method based on knowledge distillation according to claim 5, characterized in that: The sampling point is denoted as X i,k =[x1,x2,…x j …,x m ], where X i,k represents the composition vector of the kth sampling of the i-th alloy in the data set, which consists of m elements, x j represents the molar percentage of the jth element; The formula for calculating the optimized alloy composition sampling points is: i,k+1 =X i,k +λ*grad; where λ>0 is the learning rate, and grad represents the gradient information. express x j The partial derivative of .

7. The alloy composition reverse design method based on knowledge distillation according to claim 5, characterized in that: The error is expressed as in represents the predicted value of the supervised learning model, represents the predicted value of the deep neural network model; The stopping condition is: there is a sampling point X i,k satisfy and The target performance requirement is achieved or the maximum number of iterations is reached, where ε represents the error factor of the supervised learning model and the deep neural network model.

8. A knowledge distillation-based alloy composition reverse design system, applying the knowledge distillation-based alloy composition reverse design method according to any one of claims 1 to 7, characterized in that: include: Preprocessing module: collects chemical composition data and corresponding performance data of multi-element alloys, sorts them to obtain multi-element alloy data sets, and calculates atomic characteristics of each alloy in the multi-element alloy data sets; Supervised learning module: based on the atomic characteristics and performance data of each alloy, construct and train a supervised learning model to establish a mapping relationship from the atomic characteristics of the alloy to the performance; Deep learning module: based on the chemical composition data and the performance data of each alloy, a deep neural network model is trained to construct a mapping relationship from alloy composition to performance; during the training process of the deep neural network model, a back propagation algorithm is used to adjust network parameters to minimize the prediction difference between the deep neural network model and the supervised learning model; Optimization module: Langevin Dynamics is used for sampling optimization, and the gradient information of the deep neural network model is used for composition optimization to find the optimal alloy composition.

Images