A machine learning-based tensile strength prediction method for SLM alloy TC4

Abstract

The present invention discloses a method for predicting the tensile strength of Ti-6Al-4V (TC4) SLM alloy based on machine learning. The method comprises the following steps: obtaining a training set of feature data for predicting the tensile strength of the TC4 alloy; establishing a machine learning model database, screening the feature data, and obtaining an optimal feature combination; selecting a machine learning model through a ten-fold cross-validation method; using the selected machine learning model and inputting the optimal feature combination for model training; after the model training is completed, inputting unknown TC4 alloy composition and processing conditions to obtain output data, namely the tensile strength. The present invention utilizes machine learning to identify factors that determine the high tensile strength of the TC4 alloy, promotes the exploration of factors that enhance the tensile strength of the TC4 alloy, establishes an effective tensile strength model, and improves the accuracy of tensile strength prediction. This method is of great significance for promoting the performance optimization of the TC4 alloy and accelerating the composition design optimization of the TC4 alloy.

Description

Technical Field

[0001] The present invention belongs to the technical field of tensile strength prediction of metal materials, and specifically relates to a tensile strength prediction method for SLM alloy TC4 based on machine learning. Background Art

[0002] Selective laser melting (SLM) is a mainstream additive manufacturing technology for metal materials. It uses a high-energy laser beam to selectively melt thin layers of pre-applied metal powder, allowing them to solidify and form. Layer by layer, SLM creates high-density, high-precision three-dimensional metal parts. It is widely used in a variety of manufacturing industries, including aviation, aerospace, shipbuilding, automotive, chemical, and medical.

[0003] The manufacturing process of selective laser melting is as follows: first, a three-dimensional model of the part is designed using computer software, which is converted into a slice file to obtain the contour data of each section. Then, a scanning path is generated based on the contour data; then, a powder spreading device is used to evenly spread a layer of powder on the surface of the substrate; then, a high-energy laser beam is controlled to scan along the planned path, melting the metal powder and waiting for it to solidify, processing the current layer; then, the substrate is moved down one layer, and a new round of powder spreading and scanning begins, and this process is repeated layer by layer until the entire part is manufactured. It should be noted that the entire manufacturing process must be carried out in a processing chamber with a rare gas (inert gas) to prevent the metal material from oxidizing at high temperatures.

[0004] Titanium alloys, with their high specific strength and low density, are widely used in aerospace, weaponry, and other fields. Laser additive manufacturing (LAM) of titanium alloys uses powder or wire as raw material, rapidly solidifying and depositing the material layer by layer through in-situ metallurgical melting with a high-power laser. This method offers advantages such as short processing cycles, high raw material utilization, and high flexibility. Additive Manufacturing (AM) technology has been around since the 1920s and, after years of exploration and development, has now become capable of forming high-performance metal components. Selective laser melting (SLM), a type of AM technology, can directly form the desired component without the aid of molds and fixtures. This relatively simple forming process, high material utilization, and the potential for significantly reducing component manufacturing costs are key. Laser selective melting (SLM) of Ti-6Al-4V (TC4) alloy, due to its innovative structural design and potential for weight reduction, has become a hot topic in research on new materials and structures for civil aircraft. While meeting the lightweighting requirements of aerospace vehicles, titanium alloy parts must also meet long life and high reliability requirements, which necessitates excellent fatigue resistance. With the rapid development of my country's aerospace, national defense and military industries, the demand for the manufacture of titanium alloys is increasing. Therefore, studying the properties of laser selective melting titanium alloys is of great significance for their further widespread application in these fields.

[0005] Traditional methods for searching for Ti-6Al-4V alloys for selective laser melting often rely on experimental, theoretical, or computational characterization of material properties. Compared to machine learning prediction, these methods are time-consuming, labor-intensive, costly, and difficult to perform high-throughput material property characterization. With the development of artificial intelligence and big data, the use of machine learning to predict Ti-6Al-4V alloys with superior performance is becoming a trend.

[0006] Therefore, in order to solve the above problems, this paper proposes a tensile strength prediction method for SLM alloy TC4 based on machine learning. Summary of the Invention

[0007] In order to solve the above technical problems, the present invention designs a tensile strength prediction method for SLM alloy TC4 based on machine learning. Machine learning is used to find the factors that determine the high tensile strength of Ti-6Al-4V alloy, promote the exploration of factors that enhance the tensile strength of Ti-6Al-4V alloy, establish an effective tensile strength model, and improve the prediction accuracy of tensile strength.

[0008] In order to achieve the above technical effects, the present invention is implemented by the following technical solutions: a method for predicting the tensile strength of SLM alloy TC4 based on machine learning, characterized by comprising:

[0009] Step 1, obtain the feature data training set used to predict the tensile strength of TC4 alloy;

[0010] Step 2: Establish a machine learning model database, filter feature data, and obtain the optimal feature combination;

[0011] Step 3: Select the machine learning model through ten-fold cross-validation method;

[0012] Step 4: Use the selected machine learning model and input the optimal feature combination for model training;

[0013] Step 5. After the model training is completed, the unknown TC4 alloy composition and processing conditions are input to obtain the output data, which is the tensile strength.

[0014] Furthermore, the feature data training set in Step 1 includes: capturing the results of temperature, holding time, cooling method, alloy composition and tensile strength of laser selective melting TC4 from public literature data; and saving the captured experimental data and results in a database as the training data.

[0015] Furthermore, the machine learning algorithm includes one of a support vector regression algorithm, a linear regression algorithm and a neural network algorithm.

[0016] Furthermore, the specific steps for selecting the machine learning model in Step 3 are as follows:

[0017] The sample size is divided into several groups, one group of data is selected, and the remaining groups of data are used to build models to obtain the model of the combination and its test value. This cycle can be repeated several times, and the optimal model is selected as the support vector regression model.

[0018] Furthermore, the specific steps of establishing the support vector regression training model are:

[0019] 1) Let the sample set be (x1, y1),…,(xm, ym), and find a nonlinear mapping from the input space to the output space;

[0020] 2) Perform linear regression in the feature space F using the following function: f(x) = w·Φ(x) + b, Φ: Rn→F, w∈F;

[0021] 3) Where b is the threshold and w is the regression coefficient vector;

[0022] 4) Factors affecting w include: the sum of empirical risks and ‖w‖ which makes it flat in high-dimensional space 2

[0023] Right now,

[0024] Where m represents the number of training samples, Lε(f(xi)-yi) is the loss function, C is the penalty factor, and ε is the error;

[0025] 5) Introduce the relaxation factor, establish the Lagrange equation, partial derivative, dual optimization, and finally obtain the linear regression function

[0026] 6) k(x, xi) = Φ(x)·Φ(xi) is a kernel function; different SVR regression models can be generated by selecting different kernel functions.

[0027] Furthermore, in Step 4, when using the machine learning algorithm for simulation training and prediction, the linear constraint of the correlation function is weakened, that is, the configuration is allowed to have a nonlinear dependence on the correlation function.

[0028] Furthermore, the specific steps in Step 5 are:

[0029] Step 5.1: Apply SVR combined with LOOCV method to conduct modeling training and prediction research on samples in the dataset;

[0030] Step 5.2: Using temperature, holding conditions, and cooling method as inputs and the tensile strength of the alloy as the output variable, we sequentially take out one sample at a time as a test sample, and the remaining samples as training samples. We perform SVR modeling and prediction 25 times in this way.

[0031] The beneficial effects of the present invention are:

[0032] Using machine learning to find the factors that determine the high tensile strength of TC4 alloy, promote the exploration of factors that enhance the tensile strength of TC4 alloy, establish an effective tensile strength model, and improve the prediction accuracy of tensile strength are of great significance to promoting the performance optimization of TC4 alloy and accelerating the composition design optimization of TC4 alloy. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0034] Figure 1 This is a comparison chart of the tensile strength predicted by the support vector regression learning model after 25 iterations of Example 1 of the present invention and the experimentally obtained tensile strength;

[0035] Figure 2 This is a comparison chart of the tensile strength predicted by the linear regression learning model after 25 iterations of Example 2 of the present invention and the experimentally obtained tensile strength;

[0036] Figure 3 This is a comparison chart of the tensile strength predicted by the neural network learning model after 25 iterations and the experimental results in Example 3 of the present invention;

[0037] Figure 4 is the relationship between the number of iterations and the error;

[0038] Figure 5 This is a comparison chart of the mean error and determination coefficient of three machine learning models;

[0039] Figure 6 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0040] 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 some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0041] Example 1

[0042] The heat treatment process of the Ti-6Al-4V alloy includes first keeping the alloy at a specific temperature, and then sequentially performing process treatments including solid solution, normalizing, quenching, and tempering.

[0043] Establish a machine learning model database and further divide it into training set and test set.

[0044] The machine learning model is selected through the ten-fold cross-validation method. The steps are as follows: the 100 samples are divided into 10 groups, one group of data is selected, and the remaining nine groups of data are used to build a model to obtain the model of the combination and its test value. This cycle can be repeated ten times, and the optimal model is selected as the support vector regression model.

[0045] The specific steps for establishing the support vector regression training model are:

[0046] 1) Let the sample set be (x1, y1),…,(xm, ym), and find a nonlinear mapping from the input space to the output space;

[0047] 2) And perform linear regression in the feature space F using the following function: f(x) = w·

[0048] Φ(x)+b,Φ:Rn→F,w∈F

[0049] 3) Where b is the threshold and w is the regression coefficient vector.

[0050] 4) Factors affecting w include: the sum of empirical risks and ‖w‖ which makes it flat in high-dimensional space 2 ;

[0051] Right now;

[0052] 5) Where m represents the number of training samples, Lε(f(xi)-yi) is the loss function, C is the penalty factor, and ε is the error;

[0053] 6) Introduce the relaxation factor, establish the Lagrange equation, partial derivative, dual optimization, and finally obtain the linear regression function

[0054] 7) k(x, xi) = Φ(x)·Φ(xi) is the kernel function. Different kernel functions can generate different SVR regression models.

[0055] After the support vector regression model is trained and saved, the specific steps for predicting the tensile strength of titanium alloy are as follows:

[0056] 1) This paper applies SVR combined with LOOCV method to conduct modeling training and prediction research on 25 samples in the data set.

[0057] 2) Using temperature, holding conditions, and cooling method as inputs and the tensile strength of the alloy as the output variable, one sample was taken each time as a test sample, and the remaining samples were used as training samples. SVR modeling and prediction were performed 25 times in this way.

[0058] 3) Figure 1 This is a comparison chart of the prediction results of 25 iterations of the support vector regression learning model in Example 1 of the present invention and the tensile strength obtained by experiment.

[0059] It can be seen that the tensile strength predicted by the model for the screened alloys is consistent with the experimental results in terms of the changing trend with the increase in the number of iterations, which also reflects the good predictive ability of the machine learning model.

[0060] Example 2

[0061] The heat treatment process of Ti-6Al-4V alloy includes holding at a specific temperature, and then sequentially performing process treatments including solid solution, normalizing, quenching, and tempering;

[0062] Establish a machine learning model database and further divide it into training set and test set;

[0063] The machine learning model is selected through the ten-fold cross-validation method. The steps are as follows: the 100 samples are divided into 10 groups, one group of data is selected, and the remaining nine groups of data are used to build a model to obtain the model of the combination and its test value. This cycle can be repeated ten times, and the optimal model is selected as the linear regression model.

[0064] The specific steps of the linear regression training model are:

[0065] 1) Establish a linear regression model of mechanical properties using regression analysis based on the tensile properties listed in the database;

[0066] 2) The multiple linear regression model of mechanical properties is:

[0067] Pc=a0+a1x2+a2x4+a3x1x4+a4x2 2 +a5x3 2 +a6x4 2 +a7x1x2x4

[0068] Wherein, Pc is the tensile property value of titanium alloy; x1 is the percentage of Ti; x2 is the percentage of Al; x3 is the percentage of V, and x4 is the percentage of C.

[0069] Let z1=x2, z2=x4, z3=x1x4, z4=x2 2 , z5=x3 2 , z6=x4 2, z7=x1x2x4. The non-multiple linear regression model can be transformed into a multiple linear regression model.

[0070] P c =a0+a1z1+a2z2+a3z3+a4z4+a5z5+a6z6+a7z7

[0071] Finally, the tensile properties regression equation of the experimental titanium alloy is:

[0072] P c =312710.36-79804.12W Al +23214356.84W C -247164.73W Ti W C +11

[0073] 572.89W Al 2 -78930.32W V 2 +71602.72W C 2 -1788.67W Ti W Al W C ;

[0074] The regression equation shows that the mechanical properties of the titanium alloy tested in the present invention are linearly related to the contents of Al and C, the squares of the contents of Al, V and C, the interaction between Ti and C, and Ti, Al and C.

[0075] Using the mechanical property regression equation, under the same process conditions as this experiment, if the composition of the aluminum alloy is known, the tensile properties of the alloy can be calculated; conversely, if an alloy with specific mechanical properties is required, the composition of the alloy can also be determined through this regression equation.

[0076] Iterate the model through regression equation;

[0077] Figure 2 This is a comparison chart of the tensile strength predicted by the linear regression learning model after 25 iterations of Example 2 of the present invention and the experimentally obtained tensile strength.

[0078] The predicted results are in high agreement with the experimental data.

[0079] Example 3

[0080] The heat treatment process of the Ti-6Al-4V alloy includes first keeping the alloy at a specific temperature, and then sequentially performing process treatments including solid solution, normalizing, quenching, and tempering.

[0081] Establish a deep learning model database and further divide it into training set and test set.

[0082] The machine learning model is selected through the ten-fold cross-validation method. The steps are as follows: 100 samples are divided into 10 groups, one group of data is selected, and the remaining nine groups of data are used to build a model to obtain the model of the combination and its test value. This cycle can be repeated ten times, and the optimal model is selected as the neural network learning model.

[0083] Input the titanium alloy composition and processing conditions to be predicted into the model to obtain the output data, namely tensile strength. Selecting the titanium alloy tensile strength data set includes: viewing the webpage source code to find the pattern of the data to be selected, then writing the web address of the selected content and using regular expressions to extract the required data;

[0084] In data preprocessing, different feature values ​​have different ranges. Feature values ​​with large scales will play a decisive role, while feature values ​​with small scales will be ignored. In order to eliminate the influence of feature interval scale differences and treat each dimension of features equally, the features need to be normalized:

[0085]

[0086] Where Y is the normalized data; X is the input parameter; Xmin is the minimum value of the input parameter, and Xmax is the maximum value;

[0087] The selected data contains a total of 8 parameters, including 7 process parameters and 1 output parameter, namely tensile strength.

[0088] In order to obtain a high-precision prediction model, the contribution weight method is used to calculate the contribution weight C of each input parameter. j Perform calculations.

[0089]

[0090] Where: k Indicates the overall contribution of the k-th layer network structure to the output result; w kj Represents the contribution of the j-th input variable to the k-th layer network structure.

[0091] Among the seven groups of parameters that affect mechanical properties collected, temperature, holding time, cooling method, Ti, Al, V, and C content all have certain weights.

[0092] Network structure determination:

[0093] 1) Activation function

[0094] 2) If saturation occurs, the Relu activation function can be used:

[0095]

[0096] 3) For the prediction performance of the neural network model, the root mean square error (RMSE) and mean absolute percentage error (MAPE) were selected as the evaluation criteria for the model accuracy.

[0097]

[0098] The model is iterated through the adaptive matrix.

[0099] Figure 3 This is a comparison chart of the tensile strength predicted by the neural network learning model after 25 iterations in Example 3 of the present invention and the tensile strength obtained by experiment.

[0100] In general, as the iterations proceed, the consistency between the experimental values ​​of tensile strength and the model prediction results continues to improve.

[0101] The present invention is described using three examples and three different machine learning methods, all of which verify that the search method for exploring the specific composition of high-performance materials in a vast virtual space can successfully predict the tensile strength of titanium alloys. After screening specific combinations, it saves time, effort, and resources, and provides new ideas and methods for exploring more combinations.

[0102] Figure 4 This is a graph showing the relationship between the number of iterations and the error. As can be seen, the error decreases as the number of iterations increases. This confirms that the machine learning model used to predict data is reasonable and tends to approach the true value.

[0103] Figure 5 This is a comparison of the mean error and coefficient of determination of the three machine learning models. It can be seen that the support vector regression learning model has the smallest mean error, the largest coefficient of determination, and the best fitting effect.

[0104] The beneficial effect of the present invention is to propose the combination of big data and machine learning, and to use powerful data analysis capabilities to deeply explore the complex relationships in production big data, thereby achieving accurate prediction of the tensile properties of titanium alloys. According to this idea, the production big data is first processed by removing outliers, clustering, balancing data, and normalization to ensure the validity of the data. On this basis, a machine learning model is constructed and optimized, and the prediction effect of the model is tested and discussed. Through the comparison of the three learning models, the support vector regression learning model finally has the smallest average error, the largest determination coefficient, and the best fitting effect.

Claims

1. A method for predicting the tensile strength of SLM alloy TC4 based on machine learning, characterized in that: include: Step 1, obtain the feature data training set used to predict the tensile strength of TC4 alloy; Step 2: Establish a machine learning model database, filter feature data, and obtain the optimal feature combination; Step 3: Select a machine learning model using the ten-fold cross-validation method. Specifically, the sample size is divided into several groups. One group of data is selected. Models are built on the remaining groups of data to obtain the combined model and its test value. This cycle can be repeated several times. The optimal model is selected as the support vector regression model. Step 4: Use the selected machine learning model and input the optimal feature combination for model training; Step 5. After the model training is completed, input the unknown TC4 alloy composition and processing conditions to obtain the output data, which is the tensile strength. The specific steps are as follows: Step 5.1: Apply SVR combined with LOOCV method to conduct modeling training and prediction research on samples in the dataset; Step 5.2: Using the temperature, holding conditions, and cooling method as inputs and the tensile strength of the alloy as the output variable, one sample is taken each time as a test sample, and the remaining samples are used as training samples. This method is repeated 25 times for SVR modeling and prediction. The specific steps of establishing the support vector regression training model are: 1) Let the sample set be (x1, y1),…,(xm, ym), and find a nonlinear mapping from the input space to the output space; 2) Perform linear regression in the feature space F using the following function: f(x) = w·Φ(x) + b, Φ: Rn→F, w∈F; 3) Where b is the threshold and w is the regression coefficient vector; 4) Factors affecting w include: the sum of empirical risks and ‖w‖ which makes it flat in high-dimensional space 2 Right now, Where m represents the number of training samples, Lε(f(xi)-yi) is the loss function, C is the penalty factor, and ε is the error; 5) Introduce the relaxation factor, establish the Lagrange equation, perform partial derivative and dual optimization, and finally obtain the linear regression function 6) k(x, xi) = Φ(x)·Φ(xi) is a kernel function; different SVR regression models can be generated by selecting different kernel functions.

2. The method for predicting the tensile strength of SLM alloy TC4 based on machine learning according to claim 1, characterized in that: The feature data training set in Step 1 includes: capturing the temperature, holding time, cooling method, alloy composition and tensile strength results of laser selective melting TC4 from public literature data; and saving the captured experimental data and results in a database as the training data.

3. The method for predicting the tensile strength of SLM alloy TC4 based on machine learning according to claim 1, characterized in that: The machine learning algorithm includes one of a support vector regression algorithm, a linear regression algorithm or a neural network algorithm.

4. The method for predicting the tensile strength of SLM alloy TC4 based on machine learning according to claim 1, characterized in that: In the Step 4, when using the machine learning algorithm for simulation training and prediction, the linear constraint of the correlation function is weakened, that is, the configuration is allowed to have a nonlinear dependence on the correlation function.

Images