Prediction method of phase transition temperature in selective laser melting NiTi alloy based on machine learning

Abstract

The invention discloses a method for predicting phase change temperature of a selective laser melted NiTi alloy based on machine learning, which relates to the technical field of metal material performance prediction, and comprises the following steps: step S1: collecting an initial feature data set of an SLM-NiTi alloy from experiments and literature related to the SLM-NiTi alloy, step S2: performing data preprocessing on the initial feature data set; performing feature screening on a standard feature data set to form a key feature data set; step S3: randomly dividing the key feature data set into a training set and a test set, using a variety of machine learning algorithms, using the screened key features as input features of the model, and using the phase change temperature feature set of the SLM-NiTi alloy as a target feature, constructing multiple SLM-NiTi alloy phase change prediction models, and training them on the training set; step S4: predicting the test set after the training is completed, and obtaining the optimal SLM-NiTi alloy phase change prediction model.

Description

Technical Field

[0001] The present invention relates to the technical field of metal material performance prediction, and in particular to a method for predicting phase transition temperature of a selective laser melted NiTi alloy based on machine learning. Background Art

[0002] The nickel-titanium alloy prepared by SLM (selective laser melting) technology is different from the traditional manufacturing method. It solves the problem of manufacturing high-degree-of-freedom nickel-titanium alloy structural parts, while retaining the inherent superelasticity and shape memory effect, providing it with application prospects in a wide range of fields. Studies have found that the performance exhibited by SLM-NiTi alloy (nickel-titanium alloy prepared by selective laser melting technology) is closely related to its unique phase transition temperature. For every 1at% increase in Ni element in NiTi alloy, the phase transition temperature of NiTi alloy decreases by 10K. Therefore, the alloy composition is considered to be an intrinsic factor affecting the change in the phase transition temperature of NiTi alloy. At present, when studying and analyzing the relationship between the composition or process of NiTi alloy and the phase transition temperature, the experimental trial and error method is usually used.

[0003] However, the phase transition temperature is restricted by many factors such as powder composition, process parameters, and Ni element evaporation. The traditional trial-and-error method is time-consuming and costly to analyze the phase transition temperature of NiTi alloys. It is difficult to achieve precise control and adjustment, and it is difficult to meet the design requirements of high-efficiency and low-cost high-end key metal materials. Therefore, it is still a huge challenge to accurately design and control the phase transition temperature of NiTi alloys in additive manufacturing.

[0004] With the development of machine learning and the rapid increase in data volume, machine learning has become a basic tool for extracting material information. It can infer materials, establish new experiments or simulations, and learn and train these data through relevant algorithms, thereby revealing the nonlinear relationship between the material's "composition-process-structure-performance". Based on this, a method based on machine learning can be invented to predict the phase transition temperature of the selective laser melting NiTi alloy, which can accurately control and adjust the phase transition temperature, solving the problem that the traditional trial and error analysis of the phase transition temperature of the SLM-NiTi alloy is time-consuming, labor-intensive and costly. Summary of the invention

[0005] In view of the above technical problems, the present invention discloses a method for predicting the phase transition temperature of NiTi alloy by selective laser melting based on machine learning, which can save time, effort and cost and improve the problem of difficult adjustment of the phase transition temperature of NiTi alloy prepared by selective laser melting technology.

[0006] The method for predicting phase transition temperature of NiTi alloy by selective laser melting based on machine learning includes the following steps:

[0007] Step S1: Collect the alloy composition feature set, preparation process feature set, powder composition feature set and phase change temperature feature set of SLM-NiTi alloy from SLM-NiTi alloy related experiments and literature as initial samples; the collected set of all initial samples is the initial feature data set, and the phase change temperature feature set is the target feature;

[0008] Step S2: preprocess the initial feature data set; directly delete the initial samples with more missing original data; for the initial samples with less missing original data in the initial feature data set, use the missing features to obtain the mean of the feature set of the category in the initial feature data set to replace the missing feature set of the category, obtain the samples after preprocessing, and then process all the samples with Z-score normalization to obtain the standard feature data set;

[0009] The standard feature data set is screened, and multiple key features that affect the phase transformation behavior of the SLM-NiTi alloy are sorted using a variety of feature screening techniques to form a key feature data set;

[0010] Step S3: randomly dividing the key feature data set into a training set and a test set, using a variety of machine learning algorithms, taking the screened key features as the input features of the model, and the phase change temperature feature set of the SLM-NiTi alloy as the target feature, constructing multiple SLM-NiTi alloy phase change prediction models, and training them on the training set;

[0011] Step S4: After the training is completed, the test set is predicted to obtain the evaluation score of each SLM-NiTi alloy phase change prediction model, and the prediction ability of each SLM-NiTi alloy phase change prediction model is verified. By comparing the evaluation scores, the selected multiple SLM-NiTi alloy phase change prediction models are sorted to obtain the optimal SLM-NiTi alloy phase change prediction model;

[0012] Step S5: Based on the optimal SLM-NiTi alloy phase transformation prediction model, the target NiTi alloy phase transformation temperature characteristics are predicted using the alloy composition characteristic set, the preparation process characteristic set and the powder composition characteristic set of the target NiTi alloy as input values.

[0013] Furthermore, in step S1, the alloy composition feature set only includes the Ni element content feature and the Ti element content feature; the preparation process feature set includes the laser power feature, the scanning speed feature, the scanning spacing feature, the layer thickness feature, the linear energy density feature, the surface energy density feature, and the volume energy density feature; the powder composition feature set includes the powder particle size and its distribution feature; the phase change temperature feature includes M s Features, M f Features, A s Features and Af feature.

[0014] Furthermore, in step S2, the feature screening techniques used include Pearson correlation calculation, grayscale correlation analysis, feature recursive elimination and feature importance ranking.

[0015] Furthermore, in step S2, a Pearson correlation calculation is performed on the standard feature data set, and a Pearson correlation coefficient with an absolute value greater than 0.9 is considered a strong correlation, and features with strong correlation are grouped into the same group; each group is screened out according to the grayscale correlation analysis to select features with a high degree of correlation with the target features to represent the group of features for subsequent screening.

[0016] Furthermore, in step S2, the specific method for feature screening of the standard feature data set is: after performing Pearson correlation calculation and grayscale correlation analysis screening on the standard feature data set, the screened features are recursively eliminated using RF, GBR and Lasso algorithms respectively. Each recursion will select the feature with the largest feature weight, and will sort the feature importance based on the optimal GBR algorithm. Finally, the key features with the greatest importance are screened out to be used as the subsequent training model.

[0017] Furthermore, in step S3, the ratio of the number of samples in the training set and the test set is 80%:20%.

[0018] Furthermore, in step S3, the machine learning algorithm is GRNN, KNN, RF, GBR, SVM and MLP.

[0019] Furthermore, when evaluating and scoring each SLM-NiTi alloy phase transformation prediction model in step S4, the determination coefficient R 2 , root mean square error RMSE and mean absolute error MAE to comprehensively analyze the prediction performance of each SLM-NiTi alloy phase transformation prediction model.

[0020] Beneficial effects of the present invention:

[0021] The present invention does not need to process the target NiTi alloy through experiments, which effectively improves the efficiency of analyzing the performance of the target NiTi alloy and reduces the cost, thereby accelerating the development and design of new SLM-NiTi alloy materials; in addition, the method is expected to be extended to the field of prediction of the performance of other alloy materials and has universal applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a schematic diagram of the result of calculating feature importance when screening features in an embodiment of the present invention;

[0023] Figure 2is the determination coefficient R of machine learning on the training set in the embodiment of the present invention 2 Schematic diagram of comparison results;

[0024] Figure 3 is the determination coefficient R of machine learning on the test set in the embodiment of the present invention 2 Schematic diagram of comparison results;

[0025] Figure 4 is a schematic diagram of an initial feature data set and a method for obtaining the same in an embodiment of the present invention;

[0026] Figure 5 It is a schematic diagram of the phase transition temperature and prediction results of the SLM-NiTi alloy under different process parameters in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] As shown in the accompanying drawings, a method for predicting phase transition temperature of a selective laser melted NiTi alloy based on machine learning includes the following steps:

[0028] Step S1: Collect the alloy composition feature set, preparation process feature set, powder composition feature set and phase change temperature feature set of SLM-NiTi alloy from SLM-NiTi alloy related experiments and literature as initial samples; the collected set of all initial samples is the initial feature data set, and the phase change temperature feature set is the target feature;

[0029] When collecting the initial samples, only NiTi alloys prepared by selective laser melting technology were considered as the initial samples, while laser melt deposition, electron beam melting, arc melting and other methods were not included to avoid excessive features caused by their complex processing steps;

[0030] The alloy composition feature set only includes the Ni element content feature and the Ti element content feature; the preparation process feature set includes the laser power feature, scanning speed feature, scanning spacing feature, layer thickness feature, linear energy density feature, surface energy density feature, and volume energy density feature; the powder composition feature set includes the powder particle size and its distribution feature; the phase change temperature feature includes M s Features, M f Features, A s Features and A f feature.

[0031] The features included in the specific initial sample are as follows: Figure 4 As shown;

[0032] Step S2: preprocess the initial feature data set; directly delete the initial samples with more missing original data; for the initial samples with less missing original data in the initial feature data set, use the missing features to obtain the mean of the feature set of the category in the initial feature data set to replace the missing feature set of the category, and obtain the samples after preprocessing. Then, all samples are processed by Z-score normalization. The processed data conforms to the standard normal distribution, that is, the mean is 0 and the standard deviation is 1, so as to obtain the standard feature data set; the Z-score normalization calculation formula is as follows:

[0033] ;

[0034] in, is the original data value of the feature, is the normalized value, and are the mean and standard deviation of each feature respectively;

[0035] The Pearson correlation calculation was performed on the initial 14 features. The absolute value of the Pearson correlation coefficient was greater than 0.9, which was considered a strong correlation. The features with strong correlation were grouped into the same group. The features with a high degree of correlation with the target features were selected according to the grayscale correlation analysis to represent the features of the group for subsequent screening. The Pearson correlation calculation formula is as follows:

[0036] ;

[0037] in, represents the Pearson correlation coefficient, and Represents the order of any two features in the standard feature dataset Values ​​( =1,2,…,249), and represents the average value of these two features in the standard feature dataset;

[0038] After Pearson correlation calculation and grayscale correlation analysis screening, the selected features are recursively eliminated using RF, GBR and Lasso algorithms. The feature with the largest feature weight is selected each time recursively, and the feature importance is sorted based on the optimal GBR algorithm. Finally, laser power features, scanning speed features, scanning spacing features, layer thickness features, linear energy density features, surface energy density features, volume energy density features, Ni element content features, Ti element content features and powder particle size and its distribution features are selected as key features. These are combined to form a key feature data set.

[0039] Step S3: The key feature data set is randomly divided into 80% training set and 20% test set. A variety of machine learning algorithms are used to select the laser power characteristics, scanning speed characteristics, scanning spacing characteristics, layer thickness characteristics, linear energy density characteristics, surface energy density characteristics, volume energy density characteristics, Ni element content characteristics, Ti element content characteristics and powder particle size and its distribution characteristics as the input characteristics of the model. The M of SLM-NiTi alloy is s Features, M f Features, A s Features and A f The features are used as the prediction target values ​​of the machine learning model, and multiple SLM-NiTi alloy phase transformation prediction models are constructed and trained on the training set;

[0040] The machine learning algorithms are GRNN, KNN, RF, GBR, SVM and MLP; the training set is used for training, and the test set is used for testing and evaluation;

[0041] Step S4: Using the coefficient of determination R 2 , root mean square error RMSE and mean absolute error MAE, etc., to evaluate the SLM-NiTi alloy phase transformation prediction model constructed by six different machine learning algorithms, to comprehensively analyze the accuracy of the machine learning algorithm, so as to obtain the optimal algorithm; the evaluation results are shown in Table 1; it is found that GRNN in the embodiment has the best effect, and its evaluation index results are R 2 =0.97, RMSE=0.182, MAE=0.134.

[0042] Table 1: Evaluation scores of six SLM-NiTi alloy phase transformation prediction models:

[0043] Algorithm Type <![CDATA[R 2 ]]> RMSE MAE GRNN 0.97 0.182 0.134 KNN 0.97 0.237 0.147 RF 0.95 0.517 0.398 GBR 0.56 1.564 1.232 SVM 0.74 0.732 0.611 MLP 0.63 0.654 0.548

[0044] Step S5: Based on the optimal SLM-NiTi alloy phase transformation prediction model constructed by GRNN, the alloy composition feature set, preparation process feature set and powder composition feature set of the target NiTi alloy are used as input values ​​to predict the phase transformation temperature characteristics of the target NiTi alloy.

[0045] This embodiment is for Ni 50.8 Ti 49.2 The actual measured values ​​and predicted values ​​of the phase transition temperature characteristics of the SLM-NiTi alloy under different process parameters when the layer thickness characteristic is 0.03µm are as follows: Figure 5 As shown in Figure 2, it is found that the M of the optimal SLM-NiTi alloy phase transformation prediction model based on GRNN is s The relative error of the feature is 0.022, M f The relative error of the feature is 0.029, As The relative error of the feature is 0.034, A f The average absolute error of the characteristic relative error is 0.022, which is within the acceptable range of 5%, indicating that the proposed optimal SLM-NiTi alloy phase transformation prediction model based on GRNN can accurately predict the SLM-Ni 50.8 Ti 49.2 Phase transition temperature of the alloy.

[0046] See also Figure 5 : The phase transformation temperature and prediction results of SLM-NiTi alloy under different process parameters were obtained.

Claims

1. A method for predicting phase transition temperature of NiTi alloy in selective laser melting based on machine learning, characterized by: The steps include: Step S1: Collect the alloy composition feature set, preparation process feature set, powder composition feature set and phase change temperature feature set of SLM-NiTi alloy from SLM-NiTi alloy related experiments and literature as initial samples; the collected set of all initial samples is the initial feature data set, and the phase change temperature feature set is the target feature; Step S2: preprocess the initial feature data set; directly delete the initial samples with more missing original data; for the initial samples with less missing original data in the initial feature data set, use the missing features to obtain the mean of the feature set of the category in the initial feature data set to replace the missing feature set of the category, obtain the samples after preprocessing, and then process all the samples with Z-score normalization to obtain the standard feature data set; The standard feature data set is screened, and multiple key features that affect the phase transformation behavior of the SLM-NiTi alloy are sorted using a variety of feature screening techniques to form a key feature data set; Step S3: randomly dividing the key feature data set into a training set and a test set, using a variety of machine learning algorithms, taking the screened key features as the input features of the model, and the phase change temperature feature set of the SLM-NiTi alloy as the target feature, constructing multiple SLM-NiTi alloy phase change prediction models, and training them on the training set; Step S4: After the training is completed, the test set is predicted to obtain the evaluation score of each SLM-NiTi alloy phase change prediction model, and the prediction ability of each SLM-NiTi alloy phase change prediction model is verified. By comparing the evaluation scores, the selected multiple SLM-NiTi alloy phase change prediction models are sorted to obtain the optimal SLM-NiTi alloy phase change prediction model; Step S5: Based on the optimal SLM-NiTi alloy phase transformation prediction model, the target NiTi alloy phase transformation temperature characteristics are predicted by taking the alloy composition characteristic set, the preparation process characteristic set and the powder composition characteristic set of the target NiTi alloy as input values; In step S1, the alloy composition feature set only includes the Ni element content feature and the Ti element content feature; the preparation process feature set includes the laser power feature, the scanning speed feature, the scanning spacing feature, the layer thickness feature, the linear energy density feature, the surface energy density feature, and the volume energy density feature; the powder composition feature set includes the powder particle size and its distribution feature; the phase change Temperature characteristics include M s Features, M f Features, A s Features and A f feature; In step S2, the feature screening techniques used include Pearson correlation calculation, grayscale correlation analysis, feature recursive elimination and feature importance ranking; In step S2, the Pearson correlation calculation is performed on the standard feature data set, and the absolute value of the Pearson correlation coefficient greater than 0.9 is considered a strong correlation, and the features with strong correlation are grouped into the same group; each group is screened out according to the grayscale correlation analysis to select the features with a high degree of correlation with the target features to represent the group of features, and then perform subsequent screening; In step S2, the specific method for feature screening of the standard feature data set is: after performing Pearson correlation calculation and grayscale correlation analysis screening on the standard feature data set, the screened features are respectively eliminated by using RF, GBR and Lasso algorithms, and the feature with the largest feature weight is selected for each recursion, and the feature importance is sorted by the optimal GBR algorithm, and finally the key feature with the greatest importance is screened to be used as a subsequent training model; In step S3, the ratio of the number of samples in the training set and the test set is 80%:20%; When evaluating and scoring each SLM-NiTi alloy phase transformation prediction model in step S4, the determination coefficient R 2 , root mean square error RMSE and mean absolute error MAE to comprehensively analyze the prediction performance of each SLM-NiTi alloy phase transformation prediction model.

2. The method for predicting phase transition temperature of NiTi alloy by selective laser melting based on machine learning according to claim 1, characterized in that: In step S3, the machine learning algorithms are GRNN, KNN, RF, GBR, SVM and MLP.

Images