Method for predicting brazing performance of high-entropy brazing filler metal based on phase diagram calculation and numerical simulation auxiliary machine learning system

Abstract

The invention discloses a method for predicting the brazing performance of high-entropy alloy brazing filler metal based on a phase diagram calculation and numerical simulation auxiliary machine learning system, and relates to the technical field of material performance prediction. Collecting comprehensive performance data of the target high-entropy alloy and soldered joint performance characteristics as an existing database; determining a specified range of high-entropy alloy components, calculating to obtain key performance points by using a phase diagram, and mixing the key performance points with an existing database to form a basic database; a finite element simulation model is established by using Simulant welling software, and the basic database is processed to form a complete database; inputting the complete database into a machine learning system, training, analyzing contribution by using an SHAP algorithm, and performing hyper-parameter and repeated training to obtain a high-entropy alloy solder prediction model; and inputting corresponding data to obtain predicted comprehensive performance and soldered joint performance characteristics. And carrying out statistical comparison on the machine learning prediction result and the data set. The result shows that the method can accurately predict the brazing performance of the high-entropy alloy brazing filler metal.

Description

Technical Field

[0001] The present invention belongs to the technical field of metal materials, specifically to the technical field of material performance prediction, and in particular to a method for predicting the brazing performance of high entropy alloy solder based on phase diagram calculation and numerical simulation assisted machine learning system. Background Art

[0002] In the traditional brazing filler metal design, after the specific composition range of the high entropy alloy filler metal is selected, a large amount of experimental verification and composition regulation are required to obtain a relatively ideal high entropy alloy filler metal composition. In this process, a lot of time and experimental costs are paid. In the continuous integration of machine learning technology and materials science, there are already some methods that can predict the brazing performance of filler metals. For example, in the Chinese patent (CN202310481615.X), a method for simultaneously predicting the wetting performance of filler metals and the strength of brazed joints through a dual-objective performance prediction model is mentioned; in the Chinese patent (CN202210139496.5), a method for predicting the solid-liquid phase line of multi-precious metal alloy filler metals is mentioned using a weighted method of basic physical and chemical parameters. However, at present, there is still a large gap in the prediction of the comprehensive performance of high entropy alloy fillers and the strength of the brazed joints produced by them. The complexity of various parameters is at a high level. There is an urgent need for a more comprehensive method to accurately and quickly predict the brazing performance of high entropy alloy fillers. Summary of the invention

[0003] In order to solve the above technical problems, the present invention provides a method for predicting the brazing performance of high entropy solder based on phase diagram calculation and numerical simulation assisted machine learning system and a drag reduction and flow diversion method, so as to achieve the purpose of quickly predicting the comprehensive brazing performance exhibited by complex high entropy alloy systems when they are studied and used as solders.

[0004] The object of the present invention is achieved in the following manner:

[0005] A method for predicting the brazing performance of high entropy solder based on phase diagram calculation and numerical simulation assisted machine learning system, the specific steps are as follows:

[0006] Step 1: Collect the existing comprehensive performance data of the target high entropy alloy and the brazing joint performance characteristics corresponding to the process parameters of brazing under specific working conditions as an existing database;

[0007] Step 2: According to the existing database, determine the specified range of high entropy alloy components, generally with each element between 5 and 35wt.% as the specified range, use Matlab to export all high entropy alloy components with a step size of about 1wt.%, and screen out key performance points according to a step size between 5% and 10%wt.%, input the above performance points into the phase diagram calculation software to calculate the above comprehensive performance data, and use the key performance points calculated by the phase diagram and the collected high entropy alloy comprehensive performance to form a basic database;

[0008] Step 3: Based on the basic database and according to the specific conditions of the target solder, a finite element simulation model is established using the welding simulation (Simufact welding) software. The macroscopic performance parameters of the materials in the basic database are input into the material library of the software and numbered. The performance characteristics of the brazed joints formed under the same process parameters for each component are obtained in batches, and the performance characteristics of the brazed joints are combined with the comprehensive performance of the high entropy alloy solder to form a complete database.

[0009] Step 4: Input the above complete database into a machine learning system consisting of the SVR (Support Vector Regression) algorithm, the GBDT (Gradient Boosting Decision Tree) algorithm and the SHAP (SHapley Additive exPlanations) algorithm. After the SHAP algorithm analyzes the feature contribution and undergoes hyperparameterization and repeated training, a high entropy alloy solder prediction model is obtained.

[0010] The SVR algorithm is used to train the model's ability to predict the comprehensive performance of high entropy alloy solders through high entropy alloy components, and the GBDT algorithm is used to train the model's ability to predict the performance characteristics of brazed joints through the comprehensive new energy of high entropy alloy solders;

[0011] The kernel function of the SVR algorithm uses the radial basis kernel function (RBFKernel\Gaussian Kernel) which has better effect on nonlinear data, namely:

[0012] K(x,y)=exp(-γ||xy|| 2 )

[0013] Where x, y are input sample vectors, γ is a core parameter that controls the width of the Gaussian function and is used to adjust the strength of the fit when used as a hyperparameter;

[0014] The GBDT algorithm described uses the Huber loss function, which is highly resistant to noise:

[0015]

[0016] The corresponding negative gradient error is:

[0017]

[0018] Where L(y, f(x)) refers to the general expression of the Huber loss function, y is the actual input vector, and f(x) is the predicted value. δ is a hyperparameter that determines the critical point where the error switches from quadratic loss (MSE) to linear loss (MAE).

[0019] r(yi,f(xi)) refers to the negative gradient error function of the Huber loss function, yi is the vector of the ith input, f(xi) is the corresponding output vector, and ign should be processed as δ·sign(yi-f(xi)), where sign(yi-f(xi)) is a sign function. If the residual is positive, it is equal to 1, and if it is negative, it is -1;

[0020] The effectiveness of the SVR algorithm and GBDT algorithm is calculated using the corrected determination coefficient R 2 This coefficient is commonly used in this field to evaluate the accuracy of the model. When this coefficient is higher than 0.95, it means that the corresponding algorithm model is effective and accurate.

[0021] The SHAP algorithm is a means of enhancing the interpretability of machine learning models by inferring the contribution of various features. Its specific working mode is to calculate the contribution of a certain feature j to the prediction results of the machine learning model, that is, using the value function:

[0022]

[0023] Where valx(S) is the value function. The meaning of the entire expression is that under the constraint of feature subset S, the model predicts the value val x (S) , where dP_(x∈S) is the conditional probability measure of the feature vector x on the subset S, E X (f^(X)) is the global expectation term, which represents the unconditional expectation of the model prediction value in the entire output space X, that is, the predicted average value of all samples; the positive or negative value of the value function can determine the positive or negative contribution of a certain feature to the predicted value;

[0024] Step 5. After the model training is completed, the composition of the high entropy alloy solder and the process parameters for brazing under specific working conditions are used as input to output the corresponding comprehensive performance data and the performance characteristics of the brazed joint that it should exhibit when used as a solder.

[0025] The above-mentioned method for predicting the brazing performance of high-entropy solder based on phase diagram calculation and numerical simulation assisted machine learning system, the comprehensive properties of the high-entropy alloy in step one include: electrical conductivity, thermal conductivity, yield strength and tensile strength at room temperature, melting point, solid-liquid phase line temperature, flow viscosity at a specific temperature and one or more of some common thermodynamic characteristics.

[0026] In the above-mentioned method for predicting the brazing performance of high-entropy solder based on phase diagram calculation and numerical simulation-assisted machine learning system, the potential dynamic characteristics include one or more of Gibbs free energy, Helmholtz energy, enthalpy and entropy in the system.

[0027] In the above-mentioned method for predicting the brazing performance of high-entropy solder based on phase diagram calculation and numerical simulation assisted machine learning system, the brazing joint performance characteristics corresponding to the process parameters for brazing under the specific working conditions described in steps one and five include: one or more of the comprehensive properties corresponding to the base material, the wettability of the solder on the base material, the brazing temperature, the temperature curve and the clamping force of the external fixture.

[0028] In the above-mentioned method for predicting the brazing performance of high-entropy solder based on phase diagram calculation and numerical simulation assisted machine learning system, the performance characteristics of the brazed joint in step one and step five include: one or more of the strength of the brazed joint, the overall electrical conductivity, the overall thermal conductivity, and the overall average residual stress level.

[0029] In the above method for predicting the brazing performance of high entropy alloy solder based on phase diagram calculation and numerical simulation assisted machine learning system, the SVR algorithm and GBDT algorithm described in step 4 will train the model twice in the entire system respectively, and after each training is qualified, the currently formed model will be passed to the next algorithm.

[0030] In the above method for predicting the brazing performance of high entropy alloy solder based on phase diagram calculation and numerical simulation assisted machine learning system, the SHAP algorithm is involved after model training. The contribution relationship between the results and features is analyzed by the SHAP algorithm, and the contribution result of each feature to the predicted value can be obtained, which is used to evaluate the interpretability of the model and explore the relationship between relevant features and alloy system performance.

[0031] The above-mentioned method for predicting the brazing performance of high-entropy solder based on phase diagram calculation and numerical simulation assisted machine learning system, the brazing performance of the solder predicted by this method is targeted at the high-entropy alloy system.

[0032] In the above-mentioned method for predicting the brazing performance of high-entropy solder based on phase diagram calculation and numerical simulation-assisted machine learning system, the high-entropy alloy system is a complex system with more than five elements.

[0033] Compared with the prior art, the present invention has the following technical effects:

[0034] The present invention solves the problem of the prediction ability of machine learning methods for extremely complex features and improves the interpretability of machine learning by adding auxiliary calculations of phase diagram calculation and numerical simulation to machine learning. By combining phase diagram calculation and numerical simulation, the system can accurately simulate the microstructure and phase change process of high entropy alloy solder, thereby improving the prediction accuracy of brazing performance. The system can guide and optimize the composition design of high entropy alloy solder according to the prediction results to meet specific brazing performance requirements. Through the rapid iteration and analysis of machine learning algorithms, the system can significantly shorten the research and development cycle of high entropy alloy solder and speed up the listing of new materials. By reducing the number of experiments and material waste, the system helps to reduce costs in the material research and development and production process. The system can adapt to different types of high entropy alloy solders, is suitable for a variety of brazing application scenarios, and has strong adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flowchart of the machine learning system;

[0036] Figure 2 After the assistance of Simufact welding and Thermocalc software, machine learning predicted AlNiTiFeHf as a brazing material for high nitrogen steel and AlCoFeCrNi 2.1 Joint strength of joints formed by high entropy alloys. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0038] Embodiment 1:

[0039] A method for predicting the brazing performance of high entropy alloy solder based on phase diagram calculation and numerical simulation assisted machine learning system, characterized in that: the implementation steps of the method are:

[0040] Step 1: Collect the existing comprehensive performance data of AlNiTiFeHf high entropy alloy and the welding performance of B550 high nitrogen steel and AlCoFeCrNi 2.1 The brazing joint performance characteristics corresponding to the process parameters of brazing at 1400°C as the working condition are used as the existing database;

[0041] Step 2: According to the existing database, determine the specified range of high entropy alloy components, take 5 to 35 wt.% of each element in the AlNiTiFeHf system as the specified range, use Matlab to export all high entropy alloy components with a step size of about 1 wt.%, and screen out key performance points according to a step size of 5 to 10 wt.%, input the above performance points into the phase diagram calculation software to calculate the above comprehensive performance data, and use the key performance points calculated by the phase diagram and the collected high entropy alloy comprehensive performance to form a basic database, a total of 1001 groups;

[0042] Step 3: Based on the basic database and according to the specific conditions of the target solder, a finite element simulation model is established using Simufact welding software. The macroscopic performance parameters of the materials in the basic database are input into the material library of the software and numbered. The performance characteristics of the brazed joints formed by each component under the same process parameters are obtained in batches. The performance characteristics of the brazed joints are combined with the comprehensive properties of the high-entropy alloy solder to form a complete database, totaling 137 groups.

[0043] Step 4: Input the above complete database into a machine learning system consisting of the SVR (Support Vector Regression) algorithm, the GBDT (Gradient Boosting Decision Tree) algorithm and the SHAP (SHapley Additive exPlanations) algorithm, and obtain a high entropy alloy solder prediction and optimization model after hyperparameterization and training;

[0044] The SVR algorithm is used to train the model's ability to predict the comprehensive performance of high entropy alloy solders through high entropy alloy components, and the GBDT algorithm is used to train the model's ability to predict the performance characteristics of brazed joints through the comprehensive new energy of high entropy alloy solders;

[0045] The kernel function of the SVR algorithm uses the radial basis kernel function (RBFKernel\Gaussian Kernel) which has better effect on nonlinear data, namely:

[0046]

[0047] Where x, y are input sample vectors, γ is a core parameter that controls the width of the Gaussian function and is used to adjust the strength of the fit when used as a hyperparameter;

[0048] The GBDT algorithm described uses the Huber loss function, which is highly resistant to noise:

[0049]

[0050] The corresponding negative gradient error is:

[0051]

[0052] Where L(y, f(x)) refers to the general expression of the Huber loss function, y is the actual input vector, and f(x) is the predicted value. δ is a hyperparameter that determines the critical point where the error switches from quadratic loss (MSE) to linear loss (MAE).

[0053] r(yi,f(xi)) refers to the negative gradient error function of the Huber loss function, yi is the vector of the ith input, f(xi) is the corresponding output vector, and ign should be processed as δ·sign(yi-f(xi)), where sign(yi-f(xi)) is a sign function. If the residual is positive, it is equal to 1, and if it is negative, it is -1;

[0054] The effectiveness of the SVR algorithm and GBDT algorithm is calculated using the corrected determination coefficient R 2 Corrected determination coefficient R 2 Is it a specific value, a value you have calculated, or a value commonly used in this field? When this coefficient is higher than 0.95, it means that the corresponding algorithm model is effective and accurate;

[0055] The SHAP algorithm is a means of enhancing the interpretability of machine learning models by inferring the contribution of various features. Its specific working mode is to calculate the contribution of a certain feature j to the prediction results of the machine learning model, that is, using the value function:

[0056]

[0057] Where valx(S) is the value function. The meaning of the entire expression is that under the constraint of feature subset S, the model predicts the value val x (S) , where dP_(x∈S) is the conditional probability measure of the feature vector x on the subset S, E X (f^(X)) is the global expectation term, which represents the unconditional expectation of the model prediction value in the entire output space X, that is, the predicted average value of all samples; the positive or negative value of the value function can determine the positive or negative contribution of a certain feature to the predicted value;

[0058] Step 5. After the model training is completed, the composition of the high entropy alloy solder and related brazing process parameters such as brazing temperature and holding time are used as input to output its corresponding comprehensive performance data and the performance characteristics of the brazed joint that it should exhibit when used as a solder.

[0059] Embodiment 2:

[0060] The present invention is based on a method for predicting the brazing performance of high entropy solder based on phase diagram calculation and numerical simulation assisted machine learning system. The comprehensive performance of the high entropy alloy in step one includes: electrical conductivity, thermal conductivity, yield strength and tensile strength at room temperature, melting point, solid-liquid phase line temperature, flow viscosity at a specific temperature, and one or more of some common thermodynamic characteristics. The thermodynamic characteristics include one or more of Gibbs free energy, Helmholtz energy, enthalpy and entropy in the system. The performance characteristics of the brazed joint corresponding to the process parameters for brazing under the specific working conditions described in steps one and five include: one or more of the comprehensive performance corresponding to the base material, the wettability of the brazing material on the base material, the brazing temperature, the temperature curve and the clamping force of the external fixture. The performance characteristics of the brazed joint in steps one and five include: one or more of the strength of the brazed joint, the overall electrical conductivity, the overall thermal conductivity, and the overall average residual stress level. The SVR algorithm and GBDT algorithm described in step four will train the model twice in the entire system, and the currently formed model will be passed to the next algorithm after each training is qualified. The SHAP algorithm is used after model training. By analyzing the contribution relationship between results and features, the contribution of each feature to the predicted value can be obtained, which is used to evaluate the interpretability of the model and explore the relationship between relevant features and alloy performance.

[0061] The present invention discloses a method for predicting the brazing performance of high entropy solder based on phase diagram calculation and numerical simulation assisted machine learning system. The brazing performance of the solder predicted by this method is targeted at a high entropy alloy system with a complex constituent system of more than five elements.

[0062] The above is only a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several changes and improvements can be made without departing from the overall concept of the present invention, which should also be regarded as the scope of protection of the present invention.

Claims

1. A method for predicting the brazing performance of high entropy solder based on phase diagram calculation and numerical simulation assisted machine learning system, characterized in that: The specific steps are as follows: Step 1: Collect the existing comprehensive performance data of the target high entropy alloy and the brazing joint performance characteristics corresponding to the process parameters of brazing under specific working conditions as an existing database; Step 2: According to the existing database, determine the specified range of high entropy alloy components, generally with each element between 5 and 35wt.% as the specified range, use Matlab to export all high entropy alloy components with a step size of about 1wt.%, and screen out key performance points according to a step size between 5% and 10%wt.%, input the above performance points into the phase diagram calculation software to calculate the above comprehensive performance data, and use the key performance points calculated by the phase diagram and the collected high entropy alloy comprehensive performance to form a basic database; Step 3: Based on the basic database and according to the specific conditions of the target solder, a finite element simulation model is established using the welding simulation software Simufactwelding. The macroscopic performance parameters of the materials in the basic database are input into the material library of the software and numbered. The performance characteristics of the brazed joints formed under the same process parameters for each component are obtained in batches. The performance characteristics of the brazed joints are combined with the comprehensive performance of the high entropy alloy solder to form a complete database. Step 4: Input the above complete database into the machine learning system composed of SVR algorithm, GBDT algorithm and SHAP algorithm. After the SHAP algorithm analyzes the feature contribution, it undergoes hyperparameterization and repeated training to obtain a high entropy alloy solder prediction model. The SVR algorithm is used to train the model's ability to predict the comprehensive performance of high entropy alloy solders through high entropy alloy components, and the GBDT algorithm is used to train the model's ability to predict the performance characteristics of brazed joints through the comprehensive new energy of high entropy alloy solders; The kernel function of the SVR algorithm uses a radial basis kernel function that has a better effect on nonlinear data, namely: K(x,y)=exp(-γ||x-y|| 2 ) Where x, y are input sample vectors, γ is a core parameter that controls the width of the Gaussian function and is used to adjust the strength of the fit when used as a hyperparameter; The GBDT algorithm described uses the Huber loss function, which is highly resistant to noise: The corresponding negative gradient error is: Where L(y, f(x)) refers to the general expression of the Huber loss function, y is the actual input vector, and f(x) is the predicted value. δ is a hyperparameter that determines the critical point where the error switches from quadratic loss MSE to linear loss MAE. r(yi,f(xi)) refers to the negative gradient error function of the Huber loss function, yi is the vector of the ith input, f(xi) is the corresponding output vector, and ign should be processed as δ·sign(yi-f(xi)), where sign(yi-f(xi)) is a sign function. If the residual is positive, it is equal to 1, and if it is negative, it is -1; The effectiveness of the SVR algorithm and GBDT algorithm is calculated using the corrected determination coefficient R 2 When the coefficient is higher than 0.95, it means that the corresponding algorithm model is effective and accurate; The SHAP algorithm is a means of enhancing the interpretability of machine learning models by inferring the contribution of various features. Its specific working mode is to calculate the contribution of a certain feature j to the prediction results of the machine learning model, that is, using the value function: Where valx(S) is the value function. The meaning of the entire expression is that under the constraint of feature subset S, the model predicts the value val x (S) , where dP_(x∈S) is the conditional probability measure of the feature vector x on the subset S, E X (f^(X)) is the global expectation term, which represents the unconditional expectation of the model prediction value in the entire output space X, that is, the predicted average value of all samples; the positive or negative value of the value function can determine the positive or negative contribution of a certain feature to the predicted value; Step 5. After the model training is completed, the composition of the high entropy alloy solder and the process parameters for brazing under specific working conditions are used as input to output the corresponding comprehensive performance data and the performance characteristics of the brazed joint that it should exhibit when used as a solder.

2. The method for predicting the brazing performance of high entropy solder based on phase diagram calculation and numerical simulation assisted machine learning system according to claim 1 is characterized in that: The comprehensive properties of the high entropy alloy in step 1 include: electrical conductivity, thermal conductivity, yield strength and tensile strength at room temperature, melting point, solidus temperature, flow viscosity at a specific temperature and one or more of some common thermodynamic characteristics.

3. The method for predicting the brazing performance of high entropy solder based on phase diagram calculation and numerical simulation assisted machine learning system according to claim 2 is characterized in that: The dynamic characteristics include one or more of Gibbs free energy, Helmholtz energy, enthalpy and entropy in the system.

4. The method for predicting the brazing performance of high entropy solder based on phase diagram calculation and numerical simulation assisted machine learning system according to claim 1, characterized in that: The brazing joint performance characteristics corresponding to the process parameters for brazing under the specific working conditions described in step 1 and step 5 include: one or more of the comprehensive properties of the base material, the wettability of the brazing material on the base material, the brazing temperature, the temperature curve and the clamping force of the external fixture.

5. The method for predicting the brazing performance of high entropy solder based on phase diagram calculation and numerical simulation assisted machine learning system according to claim 1, characterized in that: The performance characteristics of the brazed joint in step 1 and step 5 include: one or more of the strength, overall electrical conductivity, overall thermal conductivity, and overall average residual stress level of the brazed joint.

6. The method for predicting the brazing performance of high entropy alloy solder based on phase diagram calculation and numerical simulation assisted machine learning system according to claim 1 is characterized in that: The SVR algorithm and GBDT algorithm described in step 4 will train the model twice in the entire system, and the currently formed model will be passed to the next algorithm after each training is qualified.

7. The method for predicting the brazing performance of high entropy alloy solder based on phase diagram calculation and numerical simulation assisted machine learning system according to claim 1, characterized in that: The SHAP algorithm is used after model training. By analyzing the contribution relationship between results and features, the contribution of each feature to the predicted value can be obtained, which is used to evaluate the interpretability of the model and explore the relationship between relevant features and alloy performance.

8. The method for predicting the brazing performance of high entropy solder based on phase diagram calculation and numerical simulation assisted machine learning system according to claim 1, characterized in that: The brazing performance of solder predicted by this method is targeted at high entropy alloy systems.

9. The method for predicting the brazing performance of high entropy solder based on phase diagram calculation and numerical simulation assisted machine learning system according to claim 1, characterized in that: The high entropy alloy system is a complex system with five or more elements.

Images