Physical information driving-based near-beta titanium alloy multi-performance prediction method

Abstract

The invention provides a near-beta titanium alloy multi-performance prediction method based on physical information driving, and relates to the technical field of material informatics. Retrieving related literatures of the near-beta titanium alloy, and establishing an original data set; converting components and process features in the data set, embedding element physical attributes and phase change dynamics into feature engineering, performing data preprocessing and standardization, and dividing a test set and a training set by using stratified sampling; performing parameter tuning on the XGBoost model by adopting an Optuna hyper-parameter optimization framework in combination with five-fold cross validation, training and verifying the model by using a training set and a test set, and constructing a regression prediction machine learning model based on physical information driving; and inputting the physical characteristic parameters of the new material components into the optimized learning model for prediction, and outputting prediction results of the tensile strength and the ductility. The multi-objective performance is collaboratively optimized through physical characteristics, process parameter extrapolation is supported, and a high-precision and low-data-dependence solution is provided for near-beta titanium alloy design.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of material informatics, and particularly relates to a near-beta titanium alloy multi-performance prediction method based on physical information driving. BACKGROUND

[0002] Near-beta titanium alloys are widely used in aerospace and medical fields due to their high strength, low density and good biocompatibility. The mechanical properties of the near-beta titanium alloys are influenced by the complex coupling of composition and multi-step heat treatment process. Traditional machine learning models directly input the composition proportion and original process parameters without associating with the physical mechanism, which leads to an opaque decision-making process and makes it difficult to guide the actual process optimization. In addition, the prediction uncertainty of the model for untrained areas (such as new alloy systems or extreme process parameters) is high, and a large amount of experimental data is needed to support, which significantly prolongs the research and development cycle. Although traditional empirical parameters (such as 、 values) can partially reflect the stability of the alloy, they are not modeled in conjunction with physical characteristics and are difficult to fully characterize the performance correlation.

[0003] Existing machine learning applications in material design mainly focus on single-target performance or simplified process, and have not yet solved the problems of near-beta titanium alloy multi-target performance collaborative optimization and composition-process boundary-free mapping. For example, most methods still use element atomic percentage as input, which cannot be separated from the specific element combination; the process parameter processing ignores the nonlinear characteristics of phase transition dynamics, resulting in insufficient prediction reliability for unexplored process windows. These defects limit the practical application of machine learning in the design of complex alloy systems, and an innovative method that takes into account explainability, boundary-free and high precision is urgently needed. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a near-beta titanium alloy multi-performance prediction method based on physical information driving to solve the problems of the prior art.

[0005] To solve the above technical problems, the technical solution adopted by the present application is:

[0006] A near-beta titanium alloy multi-performance prediction method based on physical information driving, comprising the following steps:

[0007] Step 1: retrieve relevant literature of near-beta titanium alloys, establish an original data set, including alloy composition, process parameters, empirical parameters and performance;

[0008] Step 2: convert the composition and process features in the data set of step 1, embed element physical properties and phase transition dynamics into feature engineering;

[0009] Step 3: Data pre-processing and standardization are performed on the data set after feature engineering in Step 2, and stratified sampling is used to divide it into a test set and a training set;

[0010] Step 4: The Optuna hyperparameter optimization framework is used in combination with 5-fold cross-validation technology to optimize the parameters of the XGBoost model. The training set and test set generated in Step 3 are used to train and validate the model, and a regression prediction machine learning model PI-XGBoost based on physical information is constructed;

[0011] Step 5: The physical characteristic parameters of the new material composition are directly input into the optimized PI-XGBoost model for prediction, and the performance prediction results of tensile strength and elongation are simultaneously output.

[0012] Further, in Step 1, the original data set includes 496 groups of data, all of which are collected from relevant literature on near-beta titanium alloys. The relevant information collected includes:

[0013] Alloy composition: mass percentage of Ti, Al, Mo, V, Cr, Fe, Zr, Sn, and Nb elements;

[0014] Process parameters: solid solution treatment temperature T sol , solid solution treatment time t sol , aging treatment temperature T age , and aging treatment time t age ;

[0015] Empirical parameters: beta phase stability parameter and electron orbital parameter ;

[0016] Alloy performance: tensile strength UTS and elongation El.

[0017] Further, in Step 2, the alloy composition weighted calculation is converted into physical intrinsic properties: equivalent electronegativity , atomic radius difference , and valence electron concentration VEC, which replace the original element proportion; at the same time, explicit association between physical parameters and performance is made to provide a physical explanation path for tracing the influence mechanism of specific elements on alloy performance; details are as follows:

[0018] The equivalent electronegativity is:

[0019] ;

[0020] wherein, is the mole fraction of element , and is the mole fraction of element Pauling electronegativity;

[0021] Atomic radius difference is:

[0022] ;

[0023] wherein, is the number of valence electrons of the element ; is the average atomic radius of all elements in the alloy;

[0024] Valence electron concentration VEC is:

[0025] ;

[0026] wherein, is the number of valence electrons of the element ;

[0027] The process parameters T sol , t sol , T age , t age are converted into nonlinear features by physical models to capture complex physical mechanisms, as follows:

[0028] Temperature feature: the phase transformation behavior in the heat treatment process is described by a Sigmoid function conversion, where θ is the phase transition temperature threshold, and T is T sol or T age ;

[0029] Time feature: based on Johnson-Mehl-Avrami-Kolmogorov (JMAK) phase transformation kinetics, t sol and t age are logarithmically transformed as (t) = ln(t), where t is t sol or t age ;

[0030] After feature engineering, the model input features are: , , VEC, σ(T sol ), (t sol ), σ(T age ), (t age ), , ;

[0031] The output features are: tensile strength UTS and elongation El.

[0032] Further, in step 3, the data set after feature engineering is preprocessed by using RobustScaler and Box-Cox transformation, and the formula is as follows:

[0033] The formula of RobustScaler is:

[0034] ;

[0035] Wherein, x is the input feature value, x scaled is the standardized feature value, median(X) is the median of the data, and IQR(X) is the interquartile range;

[0036] The formula of Box-Cox transformation is:

[0037] ;

[0038] Wherein, y is the original target variable value, y(λ) is the transformed target variable value, and λ is the transformation parameter;

[0039] The stratified sampling strategy is adopted, and the training set and the test set are divided according to the ratio of 8:2.

[0040] Further, the regression prediction machine learning model established in step 4 adopts XGBoost algorithm, and the core hyperparameter settings are as follows: learning rate η [0.001, 0.3], tree depth d [3, 12], regularization coefficient α [0, 10] and λ [0, 10]; the Bayesian optimization algorithm of Optuna framework is used to search in the defined parameter space, the generalization performance of the parameter combination is evaluated by minimizing the negative mean square error (Negative MSE) of 5-fold cross-validation, and the optimal hyperparameter configuration is finally determined; based on the optimized parameters, the training set divided in step 3 is used to train the XGBoost model, and the coefficient of determination (R²) is used as the evaluation index to quantify the regression prediction accuracy of the model on the independent test set; the finally constructed PI-XGBoost model will be used for actual new material prediction.

[0041] The beneficial effects produced by the above technical solutions are that the physical information driven near-beta titanium alloy multi-performance prediction method provided by the application realizes decoupling of composition and process characteristics by converting alloy composition into intrinsic properties, embedding process parameters into a phase change kinetics model, introducing physical information, significantly improves the interpretability and decision transparency of the model, and makes the prediction results related to the mechanism of materials science. The application breaks through the traditional element proportion limit, enhances the reliability of the model under unknown composition and process window, and has stronger generalization ability. Through physical feature collaborative optimization of multi-objective performance, and supporting process parameter extrapolation, a high-precision, low-data-dependent solution is provided for near-beta titanium alloy design. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 A physical information driven near-beta titanium alloy multi-performance prediction method flowchart is provided for the specific embodiments of the application.

[0043] Figure 2 A UTS predicted value-actual value comparison graph of the physical information driven near-beta titanium alloy multi-performance prediction method provided for the second embodiment of the application is provided.

[0044] Figure 3 An El predicted value-actual value comparison graph of the physical information driven near-beta titanium alloy multi-performance prediction method provided for the second embodiment of the application is provided.

[0045] Figure 4 A comparison of the performance precision of the regression prediction model driven by physical information and the traditional regression prediction model for predicting Ti-5Al-5Mo-5V-3Cr alloy is provided for the second embodiment of the application.

[0046] Figure 5 A comparison of the performance precision of the regression prediction model driven by physical information and the traditional regression prediction model for predicting Ti-4Al-7Mo-3Cr-3V alloy is provided for the second embodiment of the application. DETAILED DESCRIPTION

[0047] The specific embodiments of the application will be further described in detail below in combination with the drawings and examples. The following examples are used to illustrate the application, but are not used to limit the scope of the application.

[0048] As shown in Figure 1 , the method of the present embodiment is as follows.

[0049] Step 1: Search for relevant literature on near-beta titanium alloys, establish an original data set, specifically including:

[0050] Alloy composition: mass percentage of Ti, Al, Mo, V, Cr, Fe, Zr, Sn, and Nb elements;

[0051] Process parameters: solution treatment temperature T sol , solution treatment time t sol , aging treatment temperature T age , aging treatment time t age .

[0052] Empirical parameters: beta phase stability parameter value and electron orbital parameter value.

[0053] Alloy properties: tensile strength UTS and elongation El.

[0054] Step 2: Transform the composition and process characteristics in the dataset of Step 1, embed element physical properties, phase transition dynamics into feature engineering.

[0055] Transform the alloy composition weighted calculation into physical intrinsic properties: equivalent electronegativity (χ ), atomic radius difference (ΔR ) and valence electron concentration (VEC), instead of the original element proportion; at the same time, explicitly associate physical parameters with performance, to provide a physical explanation path for tracing the influence mechanism of specific elements on alloy performance. Specifically as follows:

[0056] Equivalent electronegativity is:

[0057] ;

[0058] Wherein, is the mole fraction of element , is the Pauling electronegativity of element .

[0059] Atomic radius difference is:

[0060] ;

[0061] Wherein, is the atomic radius of element , is the average atomic radius of all elements in the alloy.

[0062] Valence electron concentration VEC is:

[0063] ;

[0064] Wherein, is the number of valence electrons of element .

[0065] Transform the process parameters T sol , tsol , T age , t age By converting to nonlinear features through physical models to capture complex physical mechanisms, as follows:

[0066] Temperature feature: through Sigmoid function Converts the phase transition behavior in heat treatment process, where θ is the phase transition temperature threshold, T is T sol or T age .

[0067] Time feature: based on Johnson-Mehl-Avrami-Kolmogorov (JMAK) phase transition kinetics, t sol and t age Take logarithmic transformation to (t) = ln(t), t is t sol or t age .

[0068] After feature engineering, the model input features are: , , VEC, σ(T sol ), (t sol ), σ(T age ), (t age ), , . The output features are: UTS and El.

[0069] Step 3: Data preprocessing and standardization of the data set after feature engineering in step 2, and use stratified sampling to divide into test set and training set.

[0070] In view of the inconsistent dimension of the features in the original data set (such as , , etc.) and the skew distribution of the target variables (UTS and El), robust standardization (RobustScaler) and Box-Cox transformation are used to preprocess the data set after feature engineering to improve the convergence speed and prediction accuracy of the model. The formula is as follows:

[0071] RobustScaler:

[0072] ;

[0073] Where x is the input feature value, x scaled is the standardized feature value, median(X) is the median of the data, and IQR(X) is the interquartile range.

[0074] Box-Cox transformation:

[0075] ;

[0076] where y is the original target variable value, y(λ) is the transformed target variable value, and λ is the transformation parameter.

[0077] To avoid the lack of sample representativeness in some performance intervals of small sample data sets, a stratified sampling strategy is adopted to divide the training set and test set in a ratio of 8:2, ensuring balanced data distribution in different performance intervals and improving the generalization ability of the model.

[0078] Step 4: An XGBoost algorithm is used to establish a regression prediction machine learning model, and the core hyperparameters are set as follows: learning rate η [0.001, 0.3], tree depth d [3, 12], regularization coefficient α [0, 10], and λ [0, 10]. The Bayesian optimization algorithm of the Optuna framework is used to perform efficient search within the defined parameter space. The generalization performance of the parameter combination is evaluated by minimizing the negative mean square error (Negative MSE) of 5-fold cross-validation. The optimal hyperparameter configuration is finally determined. Based on the optimized parameters, the training set divided in step 3 is used to train the XGBoost model, and the coefficient of determination (R²) is used as the main evaluation index to quantify the regression prediction accuracy of the model on the independent test set. The final PI-XGBoost model can directly input the physical feature parameters of new material components for prediction, and simultaneously output the prediction results of tensile strength and elongation.

[0079] Example 1

[0080] The physical information-driven near-beta titanium alloy multi-performance prediction method specifically includes the following steps:

[0081] Step 1: Retrieve relevant literature to establish a 496-group near-beta titanium alloy original data set, including composition, process parameters, empirical parameters, and performance;

[0082] Step 2: Weighted calculation of alloy composition in the established data set to convert to physical intrinsic properties instead of original element proportions, and conversion of process features to nonlinear features through a physical model;

[0083] Step 3: Box-Cox transformation and RobustScaler standardization of the data set after feature engineering processing, and division into test set and training set using stratified sampling;

[0084] Step 4: Optuna framework combined with 5-fold cross-validation technique was used to optimize the XGBoost hyperparameters, and a regression prediction model based on physical information driving (PI-XGBoost) was constructed.

[0085] Step 5: The regression prediction model based on physical information driving (PI-XGBoost) was used to predict the tensile strength and elongation of Ti-5Al-5Mo-5V-3Cr-1Zr alloy after 860℃ / 1.5h solid solution treatment + 620℃ / 6h aging treatment.

[0086] The tensile strength of Ti-5Al-5Mo-5V-3Cr-1Zr alloy after 860℃ / 1.5h solid solution treatment + 620℃ / 6h aging treatment was 1185MPa, and the elongation was 7.0%. The regression prediction model based on physical information driving proposed in this embodiment predicted that the tensile strength was 1205MPa and the elongation was 5.7%. The predicted values and experimental values showed good consistency.

[0087] Example 2

[0088] The method for predicting multiple properties of near-beta titanium alloy based on physical information driving includes the following steps:

[0089] Step 1: Search related literature and establish a dataset of 496 groups of near-beta titanium alloys, including composition, process parameters, empirical parameters and properties;

[0090] Step 2: Weighted calculation of alloy composition in the established dataset to convert to physical intrinsic properties instead of original element proportion, and convert process characteristics to nonlinear features through physical model;

[0091] Step 3: Box-Cox transformation and RobustScaler standardization of the dataset after feature engineering processing, and use stratified sampling to divide into test set and training set;

[0092] Step 4: Optuna framework combined with 5-fold cross-validation technique was used to optimize the XGBoost hyperparameters, and a regression prediction model based on physical information driving (PI-XGBoost) was constructed;

[0093] Step 5: The regression prediction model based on physical information driving (PI-XGBoost) was used to predict the tensile strength and elongation of Ti-5Al-5Mo-5V-3Cr alloy after 810℃ / 0.5h solid solution treatment + 600℃ / 6h aging treatment and Ti-4Al-7Mo-3Cr-3V alloy after 910℃ / 0.5h solid solution treatment + 600℃ / 6h aging treatment.

[0094] The tensile strength of Ti-5Al-5Mo-5V-3Cr alloy and Ti-4Al-7Mo-3Cr-3V alloy after 810 °C / 0.5 h solution treatment + 600 °C / 6 h aging treatment is 1130 MPa and 1276 MPa respectively, and the elongation is 9.2% and 5.5% respectively. The prediction results of the regression prediction model based on physical information driving proposed in the embodiment are: the tensile strength is 1122 MPa and 1275 MPa respectively, the corresponding R 2 is 0.94 and 0.95 respectively; the elongation is 9.8% and 6.8% respectively, and the corresponding R 2 is 0.91 and 0.90 respectively. As shown in Figs. Figure 2 and Figure 3 , the prediction values and the experimental values show good consistency. At the same time, the performance difference mechanism of Ti-5Al-5Mo-5V-3Cr and Ti-4Al-7Mo-3Cr-3V alloys under different heat treatment systems can be analyzed through the converted composition features (x , , VEC) and the converted process features (σ(T sol ), (t sol ), σ(T age ), (t age )).

[0095] The difference between the traditional regression prediction model (XGBoost) and the regression prediction model based on physical information driving (PI-XGBoost) proposed in the embodiment is that the traditional regression prediction model uses the original features of the original data set for regression prediction, that is, the input features include alloy composition (mass percentage of Ti, Al, Mo, V, Cr, Fe, Zr, Sn, Nb elements), process parameters (T sol , t sol , T age , t age ), empirical parameters (x , value); the output features are unchanged.

[0096] The tensile strength and elongation of Ti-5Al-5Mo-5V-3Cr alloy after 810 °C / 0.5 h solution treatment + 600 °C / 6 h aging treatment are predicted by using the traditional regression prediction model. As shown in Fig. Figure 4 , when the tensile strength is regressed, R 2 is 0.80; when the elongation is regressed, R 2The value is 0.65. The prediction accuracy of traditional regression prediction models is significantly lower than that of the physical information-driven regression prediction model proposed in this embodiment.

[0097] The tensile strength and elongation of Ti-4Al-7Mo-3Cr-3V alloy after solution treatment at 910℃ for 0.5h followed by aging treatment at 600℃ for 6h were predicted using a traditional regression prediction model. Figure 5 As shown, when performing regression prediction for tensile strength, R 2 The value is 0.80; when performing regression prediction on elongation, R0 is... 2 The value is 0.61. The prediction accuracy of traditional regression prediction models is significantly lower than that of the physical information-driven regression prediction model proposed in this embodiment.

[0098] Example 3

[0099] A physical information-driven multi-performance optimization design method for near-β titanium alloys includes the following steps:

[0100] Step 1: Search relevant literature and establish a raw dataset containing 496 sets of near-β titanium alloys, including composition, process parameters, empirical parameters, and properties, where the Zr element mass ratio ranges from 0. 13 wt.%;

[0101] Step 2: Weight the alloy composition in the established dataset and convert it into physical intrinsic properties to replace the original element ratios; convert the process characteristics into nonlinear characteristics through the physical model.

[0102] Step 3: Perform Box-Cox transformation and RobustScaler standardization on the feature-engineered dataset, and use stratified sampling to divide the test set and training set;

[0103] Step 4: Optimize the hyperparameters of XGBoost using the Optuna framework combined with 5-fold cross-validation to construct a physics-driven regression prediction model, PI-XGBoost.

[0104] Step 5: The PI-XGBoost regression prediction model driven by physical information is used to predict the tensile strength and elongation of Ti-30Zr-5Mo alloy after solution treatment at 600℃ / 1h.

[0105] The model was never trained on data with a Zr element ratio of 30 wt.%. If a traditional element ratio model were used, several sets of data containing a Zr element ratio of 30 wt.% would need to be collected and trained again. However, the PI-XGBoost regression prediction model based on physical information driven by the physical properties of Zr proposed in this embodiment utilizes these properties. , The regression prediction model based on physical information driving of the application has the characteristics of no boundary, and still has reliable prediction ability under unknown components and process window. The tensile strength of Ti-30Zr-5Mo alloy after 600 DEG C / 1 h solid solution treatment is 1100 MPa, and the elongation is 10.0%. The prediction result of the regression prediction model based on physical information driving proposed in the embodiment is that the tensile strength is 1044.9 MPa, and the elongation is 12.0%. The prediction value and the experimental value present good consistency.

[0106] Finally, it should be pointed out that: the above embodiments are only used to illustrate the technical solutions of the application, but not to limit them; although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the range defined by the application.

Claims

1. A method for predicting multiple properties of near-beta titanium alloys based on physical information driving, characterized in that: The method comprises the following steps: Step 1: search for relevant literatures of near-beta titanium alloys, establish an original data set, including alloy composition, process parameters, empirical parameters and performance; Step 2: convert the composition and process characteristics in the data set of step 1, embed the element physical properties and phase transition dynamics into feature engineering; Step 3: perform data preprocessing and standardization on the data set after feature engineering processing of step 2, and divide into test set and training set using stratified sampling; Step 4: use Optuna hyperparameter optimization framework combined with 5-fold cross-validation technology to optimize the parameters of XGBoost model, use the training set and test set generated in step 3 to train and verify the model, and build a regression prediction machine learning model PI-XGBoost based on physical information driving; Step 5: directly input the physical characteristic parameters of the new material composition into the optimized PI-XGBoost model for prediction, and simultaneously output the performance prediction results of tensile strength and elongation.

2. The physical information driven near-beta titanium alloy multi-property prediction method according to claim 1, characterized in that: In the step 1, the original data set includes 496 groups of data, all of which are collected from relevant literatures of near-beta titanium alloys, and the relevant information collected includes: Alloy composition: mass percentage of Ti, Al, Mo, V, Cr, Fe, Zr, Sn and Nb elements; Process parameters: solution treatment temperature T sol , solution treatment time t sol , aging treatment temperature T age , aging treatment time t age ; Empirical parameter: beta phase stability parameter values and electron orbital parameters values; Alloy performance: tensile strength UTS and elongation El.

3. The physical information driven near-beta titanium alloy multi-property prediction method according to claim 2, characterized in that: In step 2, the alloy composition weighted calculation is converted into physical intrinsic properties: equivalent electronegativity , atomic radius difference , and valence electron concentration VEC, which replace the original element proportion; at the same time, the physical parameters are explicitly associated with the performance, providing a physical explanation path for tracing the influence mechanism of specific elements on alloy performance; as follows: electronegativity is: ； wherein is the mole fraction of the element , is the Pauling electronegativity of the element ; atomic radius difference is: ； wherein is the atomic radius of the element , is the average atomic radius of all elements in the alloy; Valence electron concentration VEC is: ； wherein, is the number of valence electrons of the element ; The process parameters T sol , t sol , T age , t age are converted by a physical model into non-linear features to capture complex physical mechanisms as follows: Temperature feature: by Sigmoid function The conversion describes the phase transition behavior during heat treatment, where θ is the phase transition temperature threshold, T is T sol or T age ; Time characteristics: based on Johnson-Mehl-Avrami-Kolmogorov phase transition kinetics, t sol and t age Logarithmic transformation is used as φ(t) = ln(t), where t is t sol or t age ; After feature engineering, the model input features are: , , VEC, σ(T sol ), φ(t sol ), σ(T age ), φ(t age ), , ; Output features: tensile strength UTS and elongation El.

4. The physical information driven near-beta titanium alloy multi-property prediction method according to claim 1, characterized in that: In the step 3, robust standardization and Box-Cox transformation are used for data preprocessing on the data set after feature engineering processing, and the formulas are as follows: Robust standardization formula is: ； where x is the input feature value, x scaled is the normalized feature value, median(X) is the median of the data, and IQR(X) is the interquartile range; Box-Cox transformation formula is: ； Wherein, y is the original target variable value, y(λ) is the transformed target variable value, and λ is the transformation parameter; Stratified sampling strategy is used to divide the training set and test set according to the ratio of 8:

2.

5. The physical information driven near-beta titanium alloy multi-property prediction method according to claim 1, wherein: The regression prediction machine learning model established in step 4 adopts the XGBoost algorithm, and the core hyperparameter settings are as follows: learning rate η [0.001, 0.3], tree depth d [3, 12], regularization coefficient α [0, 10] and λ [0, 10]; using the Bayesian optimization algorithm of the Optuna framework, search within the defined parameter space, evaluate the generalization performance of the parameter combination by minimizing the negative mean square error of 5-fold cross-validation, and finally determine the optimal hyperparameter configuration; based on the optimized parameters, train the XGBoost model using the training set divided in step 3, and on the independent test set, use the coefficient of determination R² as the evaluation index to quantify the regression prediction accuracy of the model; the finally constructed PI-XGBoost model will be used for actual new material prediction.

Images