An interpretable method and system for rapidly predicting the ultimate tensile strength of tin-based solders using SHAP package LightGBM

Abstract

The present invention discloses an interpretable method and system for rapidly predicting the ultimate tensile strength of tin-based solder by using SHAP to package LightGBM. The method collects experimental values ​​from literature as data set samples; sorts out the elemental composition of tin-based solder alloys, and calculates the features of samples used for modeling; randomly divides the training set and the test set; based on the training set, uses outlier scaling data for independent variables to establish a preliminary prediction model for the ultimate tensile strength of tin-based solder alloys; according to the established prediction model for the ultimate tensile strength of tin-based solder alloys, uses SHAP to calculate the SHAPvalue of each feature and packages LightGBM for feature screening to select the optimal feature subset; according to the established interpretable prediction model for the ultimate tensile strength of tin-based solder alloys, rapidly predicts the ultimate tensile strength of tin-based solder alloys in independent test sets. The present invention is based on reliable literature data and modeling methods, and the prediction model for the ultimate tensile strength of tin-based solder alloys established by the present invention has the advantages of being simple, fast, low-cost, pollution-free, and interpretable.

Description

Technical Field

[0001] The present invention relates to the field of mechanical properties of tin-based solder alloys, and in particular to an interpretable method and system for rapidly predicting the ultimate tensile strength of tin-based solder alloys using SHAP package LightGBM. Technical Background

[0002] Tin-based solder alloy is the most commonly used solder alloy system for soldering. The earliest Sn-Pb solder has a history of thousands of years. With the development of electronic technology, packaging density has gradually increased, and electronic components and products have also tended to be miniaturized. Correspondingly, the power of electronic devices has increased, the operating temperature has increased, and the thermal stress of devices has increased, which further increases the mechanical strength requirements of solders. The currently used tin-based solder alloys no longer meet the needs. Using the traditional "cooking method" to design tin-based solder alloys with high tensile strength requires a lot of manpower and material resources. Therefore, it is of great significance to use the existing mechanical properties data to study the mechanical strength of tin-based solder alloys and design tin-based solder alloys with better performance.

[0003] SHAP (SHapley Additive exPlanation) is an additive explanatory model inspired by cooperative game theory, which considers all features as "contributors". SHAP can nest machine learning models to explain the output. For each prediction sample, the machine learning model will generate a prediction value for the target value, and SHAP will assign a numerical value to each feature in the sample as the contribution of the feature to the target value.

[0004] LightGBM (Light Gradient Boosting Machine) is an improvement based on GBDT (Gradient Boosting Decision Tree) algorithm. The algorithm automatically adjusts the weight of each round of training samples and amplifies the weight information of poorly fitted samples to improve the stability and robustness of the model, and can effectively improve the execution efficiency of the algorithm while ensuring accuracy. However, how to combine LightGBM and LightGBM for tin-based solder alloy design still needs further exploration and research. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the prior art in the design of tin-based solder alloys, such as high cost of mechanical property testing, low calculation and prediction accuracy, and difficulty in mechanism explanation, and to provide a method and system for quickly predicting the ultimate tensile strength of tin-based solder using SHAP-packaged LightGBM. This method can be a simple, fast, low-cost, highly generalized, and explainable LightGBM model for quickly predicting the ultimate tensile strength of tin-based solder alloys, thereby achieving rapid prediction of the ultimate tensile strength of tin-based solder and optimizing the design of tin-based solder alloys.

[0006] In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical solutions:

[0007] A method for quickly predicting the ultimate tensile strength of tin-based solder using SHAP package LightGBM, comprising the following steps:

[0008] 1) Using a computer system, the element composition, experimental conditions and ultimate tensile strength experimental values ​​of tin-based solder alloys were collected from the literature as data set samples;

[0009] 2) sort out the elemental composition of the tin-based solder alloys, combine the selected physicochemical properties, and calculate the characteristics of each tin-based solder alloy sample for modeling;

[0010] 3) The data set was randomly divided into a training set and a test set with a sample size ratio of 4:1;

[0011] 4) using the ultimate tensile strength experimental value of the tin-based solder alloy collected in step 1) as the target variable, using the features in step 2) as the independent variable, using the training set divided in step 3), using outlier scaling data for the independent variable, and using LightGBM to establish a preliminary prediction model for the ultimate tensile strength of the tin-based solder alloy;

[0012] 5) According to the preliminary prediction model of the ultimate tensile strength of the tin-based solder alloy established in step 4), SHAP is used to calculate the SHAP value of each feature, and LightGBM is packaged for feature screening to select the optimal feature subset for the establishment of the final prediction model;

[0013] 6) Based on the features of the optimal feature subset in step 5), an interpretable prediction model for the ultimate tensile strength of the tin-based solder alloy is established to quickly predict the ultimate tensile strength of the test set samples divided in step 3).

[0014] Preferably, in said step 4), the following steps are specifically included:

[0015] The specific steps of LightGBM are as follows:

[0016] 4-1) Given a data set {x1, ..., x n}, where x i The input space χ s The dimension of the vector is s. In each iteration of gradient boosting, the negative gradient of the loss function output by the model is recorded as {g1,...,g n};

[0017] 4-2) Arrange the training samples in descending order according to the absolute value of their gradients. The first a% of samples with larger gradients are taken as subset A, and the remaining (1-a)% of samples with smaller gradients are taken as subset A. c , and from A c Randomly select b% of samples as subset B;

[0018] 4-3) Calculate the variance gain on A∪B Select To split the sample nodes, The definition is as follows:

[0019]

[0020] Among them, A l ={x i ∈A:x ij ≤d},A r ={x i ∈A:x ij >d},B l ={x i ∈B:x ij ≤d}, B r ={x i ∈B:x ij >d}, coefficient Used to normalize the gradient on B to A c The size of n is the number of samples in the data set, d is the sample node, x i The input space χ s The dimension of the vector is s, g i is the negative gradient of the loss function output by the model; A is the first a% sample subset with larger gradients selected after arranging the training samples in descending order according to the absolute value of their gradients, A c is the remaining (1-a)% small gradient sample subset, B is the sample from A c A subset consisting of b% of samples randomly selected from A, a is the number of samples with larger gradients selected from the first a% of subset A, and b is the number of samples selected from A. c Randomly select b% of the samples; For x i ∈A:x ij The number of sample nodes d ≤ d, For x i ∈A:x ij > the number of sample nodes d of d, x ij is the j-th dimension feature of the i-th sample, i=1,...,n,j=1,...,s;

[0021] 4-4) Select the corresponding sample node and determine the corresponding output value;

[0022] 4-5) Repeat the above steps until the fitting accuracy threshold is reached or the upper limit of the number of trees is reached.

[0023] Preferably, in step 5), the steps of using SHAP to wrap LightGBM for variable screening are as follows:

[0024] 5-1) For the established model, use the following formula to calculate the SHAPvalue of the i-th feature of each sample;

[0025]

[0026] Among them, F is the set of all features, S is the feature subset of F, and f S is the model trained on feature subset S, f S∪{k} is a model trained on the feature subset S∪{k}, F{k} represents the feature subset containing feature k, |S| represents the size of S, and |F| represents the size of F;

[0027] 5-2) For feature i, calculate the absolute value of the SHAP value of each sample and then take the average value as the feature contribution of feature i. Then, sort the features in descending order according to the feature contribution.

[0028] 5-3) Use the previous feature with the largest contribution as a feature subset to build a LightGBM model for the target value, and then use the first two features as feature subsets to build a LightGBM model for the target value, and repeat this process until all features have been modeled;

[0029] 5-4) Select the feature subset with the best modeling effect.

[0030] A SHAP package LightGBM interpretable model system for predicting the ultimate tensile strength of tin-based solder, and a method for quickly predicting the ultimate tensile strength of tin-based solder using SHAP package LightGBM of the present invention, comprising:

[0031] Input module: Find the element composition, experimental conditions and ultimate tensile strength experimental value data of tin-based solder alloy from publicly published literature and use them as input data;

[0032] Data analysis module: using the data obtained from the input module, execute the method of SHAP package LightGBM to quickly predict the ultimate tensile strength of tin-based solder, and quickly predict the ultimate tensile strength of tin-based solder;

[0033] Output module: Output the ultimate tensile strength data of tin-based solder.

[0034] Compared with the prior art, the present invention has the following obvious outstanding substantive features and significant advantages:

[0035] 1. The present invention is simple and quick to predict the ultimate tensile strength of tin-based solder alloys. The collected tin-based solder alloy data is imported into the established model, and the calculation results can be obtained in just a few seconds;

[0036] 2. The present invention has no restrictions on the number of components of the tin-based solder alloy, and the prediction model has universal applicability;

[0037] 3. The feature screening method used in the present invention can calculate the contribution of features to the ultimate tensile strength, help extract important features that affect the structure-property relationship, and explain the influence of important features on the ultimate tensile strength;

[0038] 4. The present invention can effectively improve the design efficiency of tin-based solder alloys with higher ultimate tensile strength; by designing reasonable virtual samples and using the model to predict the virtual samples, the chemical formula of candidate tin-based solder alloys with higher performance can be given, avoiding continuous trial and error, and overcoming the shortcomings of the traditional "cooking method";

[0039] 5. The present invention uses SHAP package LightGBM to predict the ultimate tensile strength of tin-based solder alloys. The entire process does not involve experiments and chemical products, is in line with the concept of green environmental protection, and has low cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a diagram showing the LightGBM preliminary modeling results of the ultimate tensile strength of the tin-based solder alloy of the preferred embodiment of the present invention.

[0041] Figure 2 Iterative diagram for selecting the optimal feature subset for SHAP packaging LightGBM in the preferred embodiment of the present invention.

[0042] Figure 3 This is a graph of the leave-one-out result of the LightGBM interpretable model training set for the ultimate tensile strength of the tin-based solder alloy in the preferred embodiment of the present invention.

[0043] Figure 4 This is a graph of independent test results of the LightGBM interpretable model for the ultimate tensile strength of the tin-based solder alloy of the preferred embodiment of the present invention.

[0044] Figure 5 This is the SHAP value contribution diagram of the LightGBM interpretable model of the ultimate tensile strength of the tin-based solder alloy in the preferred embodiment of the present invention. DETAILED DESCRIPTION

[0045] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] Embodiment 1

[0047] In this embodiment, a method for quickly predicting the ultimate tensile strength of tin-based solder using SHAP package LightGBM includes the following steps:

[0048] 1) Using a computer system, the element composition, experimental conditions and ultimate tensile strength experimental values ​​of tin-based solder alloys were collected from the literature as data set samples;

[0049] 2) sort out the elemental composition of the tin-based solder alloys, combine the selected physicochemical properties, and calculate the characteristics of each tin-based solder alloy sample for modeling;

[0050] 3) The data set is randomly divided into a training set and a test set with a sample size ratio of 4:1;

[0051] 4) using the ultimate tensile strength experimental value of the tin-based solder alloy collected in step 1) as the target variable, using the features in step 2) as the independent variable, using the training set divided in step 3), using outlier scaling data for the independent variable, and using LightGBM to establish a preliminary prediction model for the ultimate tensile strength of the tin-based solder alloy;

[0052] 5) According to the preliminary prediction model of the ultimate tensile strength of the tin-based solder alloy established in step 4), SHAP is used to calculate the SHAPvalue of each feature, and LightGBM is packaged for feature screening to select the optimal feature subset for the establishment of the final prediction model;

[0053] 6) Based on the features of the optimal feature subset in step 5), an interpretable prediction model for the ultimate tensile strength of the tin-based solder alloy is established to quickly predict the ultimate tensile strength of the test set samples divided in step 3).

[0054] In this embodiment, a SHAP package LightGBM interpretable model system for predicting the ultimate tensile strength of tin-based solder is implemented. The method of SHAP package LightGBM for quickly predicting the ultimate tensile strength of tin-based solder described in this embodiment includes:

[0055] Input module: Find the element composition, experimental conditions and ultimate tensile strength experimental value data of tin-based solder alloy from publicly published literature and use them as input data;

[0056] Data analysis module: using the data obtained from the input module, execute the method of SHAP package LightGBM to quickly predict the ultimate tensile strength of tin-based solder, and quickly predict the ultimate tensile strength of tin-based solder;

[0057] Output module: Output the ultimate tensile strength data of tin-based solder.

[0058] The method and system of SHAP packaging LightGBM for quickly predicting the ultimate tensile strength of tin-based solder in this embodiment can be simple, fast, low-cost, and have strong generalization performance. The explainable LightGBM model for quickly predicting the ultimate tensile strength of tin-based solder alloys can achieve rapid prediction of the ultimate tensile strength of tin-based solder and optimize the design of tin-based solder alloys.

[0059] Embodiment 2

[0060] This embodiment is basically the same as the first embodiment, except that:

[0061] In this embodiment, the ultimate tensile strength of tin-based solder alloy is predicted using SHAP package LightGBM, including the following steps:

[0062] (1) The chemical formula, experimental conditions and ultimate tensile strength experimental values ​​of tin-based solder alloys were collected from the literature. A total of 377 samples were used as the data set. The chemical formula, experimental conditions and ultimate tensile strength experimental values ​​of some tin-based solder alloys are shown in Table 1:

[0063] Table 1. Chemical formula, experimental conditions and ultimate tensile strength experimental values ​​of some tin-based solder alloys

[0064]

[0065] (2) Organize the elemental composition of the samples in the data set and convert the mass fractions into atomic fractions; transfer knowledge from related fields and select some physical and chemical properties, such as absolute electronegativity, Pauli electronegativity, melting point, volume model, etc., and use atomic parameters to calculate the corresponding properties of each tin-based solder alloy sample for modeling features;

[0066] (3) The dataset is randomly divided into a training set and a test set in a ratio of 4:1, where the number of samples in the training set and the test set are 301 and 76, respectively;

[0067] (4) Using the ultimate tensile strength experimental value of the tin-based solder alloy collected in step 1) as the target variable, and the feature in step 2) as the independent variable, based on the training set divided in step 3), scaling the data of the independent variable using outliers, and using LightGBM to establish a preliminary prediction model for the ultimate tensile strength of the tin-based solder alloy;

[0068] The specific steps of LightGBM are as follows:

[0069] 4-1) Given a data set {x1, ..., x n}, where x i The input space χ sis a vector of dimension s in the gradient boosting. In each iteration of gradient boosting, the negative gradient of the loss function output by the model is recorded as {g1,...,g n};

[0070] 4-2) Arrange the training samples in descending order according to the absolute value of their gradients. The first a% of samples with larger gradients are taken as subset A, and the remaining (1-a)% of samples with smaller gradients are taken as subset A. c , and from A c Randomly select b% of samples as subset B;

[0071] 4-3) Calculate the variance gain on A∪B Select To split the sample nodes, The definition is as follows:

[0072]

[0073] Among them, A l ={x i ∈A:x ij ≤d},A r ={x i ∈A:x ij >d},B l ={x i ∈B:x ij ≤d}, B r ={x i ∈B:x ij >d}, coefficient Used to normalize the gradient on B to A c The size of n is the number of samples in the data set, d is the sample node, x i The input space χ s The dimension of the vector is s, g i is the negative gradient of the loss function output by the model. A is the first a% sample subset with larger gradients selected after arranging the training samples in descending order according to the absolute value of their gradients. c is the remaining (1-a)% small gradient sample subset, B is the sample from A c A subset consisting of b% of samples randomly selected from A, a is the number of samples with larger gradients selected from the first a% of subset A, and b is the number of samples selected from A. c b% of the samples are randomly selected from . For x i ∈A:x ij The number of sample nodes d ≤ d, For x i ∈A:x ij > the number of sample nodes d of d, x ijis the j-th dimension feature of the i-th sample, i=1,...,n,j=1,...,s.

[0074] 4-4) Select the corresponding sample node and determine the corresponding output value;

[0075] 4-5) Repeat the above steps until the fitting accuracy threshold is reached or the upper limit of the number of trees is reached;

[0076] (5) According to the prediction model of the ultimate tensile strength of the tin-based solder alloy established in step 4), the SHAP value of each feature is calculated using SHAP, and the features are processed in descending order according to the SHAP value, and the features with the highest SHAP value are selected in turn to form a feature subset, and each feature subset is packaged with LightGBM in turn to establish a model, and the feature subset with the best modeling effect is selected as the optimal feature subset;

[0077] The specific steps of SHAP packaging LightGBM for variable screening are as follows:

[0078] 5-1) For the established model, use the following formula to calculate the SHAPvalue of the i-th feature of each sample;

[0079]

[0080] Among them, F is the set of all features, S is the feature subset of F, and f S is the model trained on feature subset S, f S∪{k} is a model trained on the feature subset S∪{k}, F{k} represents the feature subset containing feature k, |S| represents the size of S, and |F| represents the size of F;

[0081] 5-2) For feature i, calculate the absolute value of the SHAP value of each sample and then take the average value as the feature contribution of feature i. Then, sort the features in descending order according to the feature contribution.

[0082] 5-3) Use the previous feature with the largest contribution as a feature subset to build a LightGBM model for the target value, and then use the first two features as feature subsets to build a LightGBM model for the target value, and repeat this process until all features have been modeled;

[0083] 5-4) Select the feature subset with the best modeling effect;

[0084] (6) Based on the optimal feature subset selected in step 5), an interpretable prediction model for the ultimate tensile strength of the tin-based solder alloy is established to quickly predict the ultimate tensile strength of the test set samples in step 3).

[0085] This example is based on 301 tin-based solder alloy samples combined with the preliminary modeling results of the ultimate tensile strength of tin-based solder alloy established by LightGBM, such as Figure 1 shown.

[0086] LightGBM was used to perform regression modeling on 301 tin-based solder alloy sample data, and a preliminary prediction model for the ultimate tensile strength of tin-based solder alloys was established. The correlation coefficient R between the ultimate tensile strength model prediction value of tin-based solder alloys and the experimental value in the literature was 0.976, and the root mean square error RMSE was 3.471.

[0087] Embodiment 3

[0088] This embodiment is basically the same as the above embodiment, with the following features:

[0089] In this example, the optimal feature subset selection iteration results of the LightGBM tin-based solder alloy ultimate tensile strength model are packaged in SHAP, such as Figure 2 shown.

[0090] A preliminary prediction model of the ultimate tensile strength of tin-based solder alloys was established using 301 tin-based solder alloy samples, and the SHAP value feature contribution ranking was given, and the LightGBM tin-based solder alloy ultimate tensile strength model was packaged in turn. When the first 7 features were selected to establish the LIghtGBM model, the LightGBM model had the highest score with the lowest redundancy. The correlation coefficient R between the ultimate tensile strength model prediction value of the tin-based solder alloy and the experimental value in the literature was 0.973, and the root mean square error RMSE was 3.703.

[0091] Embodiment 4

[0092] This embodiment is basically the same as the above embodiment, with the following features:

[0093] In this embodiment, the interpretable prediction model training set leave-one-out method results of the ultimate tensile strength established based on 301 tin-based solder alloy samples and the optimal feature subset combined with LightGBM are as follows: Figure 3 shown.

[0094] The leave-one-out method was used to internally cross-validate the interpretable prediction model of the ultimate tensile strength of tin-based solder alloys established with 301 sample data. The correlation coefficient R between the ultimate tensile strength model prediction value of tin-based solder alloys and the experimental value in the literature was 0.938, and the root mean square error RMSE was 5.32.

[0095] Embodiment 5

[0096] This embodiment is basically the same as the above embodiment, with the following features:

[0097] In this embodiment, the independent test set prediction results of the interpretable prediction model of ultimate tensile strength established based on 301 tin-based solder alloy samples and the optimal feature subset combined with LightGBM are as follows: Figure 4 shown.

[0098] The LightGBM interpretable prediction model of the ultimate tensile strength of tin-based solder alloy was established to predict 76 samples in the independent test set, and good results were obtained. The correlation coefficient R between the ultimate tensile strength model prediction value of tin-based solder alloy and the experimental value in the literature is 0.917, and the root mean square error RMSE is 6.526.

[0099] Embodiment 6

[0100] This embodiment is basically the same as the above embodiment, with the following features:

[0101] In this example, the SHAPvalue contribution graph of 377 tin-based solder alloy samples on the established LightGBM interpretable model of the ultimate tensile strength of tin-based solder alloy is shown in Figure 2. Figure 5 shown.

[0102] The LightGBM interpretable prediction model of the ultimate tensile strength of tin-based solder alloys was used to calculate the SHAP value of all 377 samples, and a very clear feature contribution analysis was given. The strain rate, average atomic weight, average work function and average bulk modulus in the tin-based solder alloy have a positive effect on the ultimate tensile strength, while the test temperature and average atomic environment number have a negative effect on the ultimate tensile strength.

[0103] In summary, the method and system for quickly predicting the ultimate tensile strength of tin-based solder using SHAP packaged LightGBM in the above-mentioned embodiment, and the interpretable model for predicting the ultimate tensile strength of tin-based solder alloys using SHAP packaged LightGBM, include the following steps: (1) Collecting the elemental composition, experimental conditions and ultimate tensile strength experimental values ​​of tin-based solder alloys from the literature as data set samples. (2) Arranging the elemental composition of the tin-based solder alloys, and combining the selected physicochemical properties to calculate the features for modeling of each tin-based solder alloy sample. (3) Randomly dividing the data set into a training set and a test set in a ratio of 4:1. (4) Taking the collected ultimate tensile strength experimental values ​​of the tin-based solder alloys as the target variables, the constructed features as the independent variables, based on the training set, the independent variables are scaled using outliers, and LightGBM is used to establish a preliminary prediction model for the ultimate tensile strength of the tin-based solder alloy. (5) According to the established prediction model for the ultimate tensile strength of tin-based solder alloys, SHAP is used to calculate the SHAP value of each feature and LightGBM is packaged for feature screening, and the optimal feature subset is selected to establish the final prediction model. (6) According to the established interpretable prediction model for the ultimate tensile strength of tin-based solder alloys, the ultimate tensile strength of the tin-based solder alloys in the independent test set is quickly predicted. The above-mentioned embodiments of the present invention are based on reliable literature data and modeling methods, and the prediction model for the ultimate tensile strength of tin-based solder alloys established has the advantages of being simple, fast, low-cost, pollution-free, and interpretable.

[0104] The above describes the embodiments of the present invention in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. Various changes can be made according to the purpose of the invention. Any changes, modifications, substitutions, combinations or simplifications made according to the spirit and principle of the technical solution of the present invention should be equivalent replacement methods. As long as they meet the purpose of the invention and do not deviate from the technical principles and inventive concepts of the present invention, they belong to the protection scope of the present invention.

Claims

1. A method for quickly predicting the ultimate tensile strength of tin-based solder using SHAP package LightGBM. It is characterized in that The following steps are involved: 1) Using a computer system, the element composition, experimental conditions and ultimate tensile strength experimental values ​​of tin-based solder alloys were collected from the literature as data set samples; 2) sort out the elemental composition of the tin-based solder alloys, combine the selected physicochemical properties, and calculate the characteristics of each tin-based solder alloy sample for modeling; 3) The data set is randomly divided into a training set and a test set with a sample size ratio of 4:1; 4) using the ultimate tensile strength experimental value of the tin-based solder alloy collected in step 1) as the target variable, using the features in step 2) as the independent variable, using the training set divided in step 3), using outlier scaling data for the independent variable, and using LightGBM to establish a preliminary prediction model for the ultimate tensile strength of the tin-based solder alloy; 5) According to the preliminary prediction model of the ultimate tensile strength of the tin-based solder alloy established in step 4), SHAP is used to calculate the SHAP value of each feature, and LightGBM is packaged for feature screening to select the optimal feature subset for the establishment of the final prediction model; 6) establishing an interpretable prediction model of the ultimate tensile strength of the tin-based solder alloy based on the features of the optimal feature subset in step 5), and quickly predicting the ultimate tensile strength of the test set samples divided in step 3); In the step 4), the following steps are specifically included: The specific steps of LightGBM are as follows: 4-1) Given a data set {x 1 ,…,x n }, where x i The input space χ s The dimension of the vector is s. In each iteration of gradient boosting, the negative gradient of the loss function output by the model is recorded as {g 1 ,…,g n }; 4-2) Arrange the training samples in descending order according to the absolute value of their gradients. The first a% of samples with larger gradients are taken as subset A, and the remaining (1-a)% of samples with smaller gradients are taken as subset A. c , and from A c Randomly select b% of samples as subset B; 4-3) Calculate the variance gain on A∪B Select To split the sample nodes, The definition is as follows: Among them, A l ={x i ∈A:x ij ≤d},A r ={x i ∈A:x ij >d}, B l ={x i ∈B:x ij ≤d}, B r ={x i ∈B:x ij >d}, coefficient Used to normalize the gradient on B to A c The size of n is the number of samples in the data set, d is the sample node, x i The input space χ s The dimension of the vector is s, g i is the negative gradient of the loss function output by the model; A is the first a% sample subset with larger gradients selected after arranging the training samples in descending order according to the absolute value of their gradients, A c is the remaining (1-a)% small gradient sample subset, B is the sample from A c A subset consisting of b% of samples randomly selected from A, a is the number of samples with larger gradients selected from the first a% of subset A, and b is the number of samples selected from A. c Randomly select b% of the samples; For x i ∈A:x ij The number of sample nodes d ≤ d, For x i ∈A:x ij >The number of sample nodes d of d, x ij is the j-th dimension feature of the i-th sample, i = 1,…,n, j = 1,…,s; 4-4) Select the corresponding sample node and determine the corresponding output value; 4-5) Repeat the above steps until the fitting accuracy threshold is reached or the upper limit of the number of trees is reached.

2. According to the method for quickly predicting the ultimate tensile strength of tin-based solder using SHAP package LightGBM as described in claim 1, in step 5), the steps of using SHAP package LightGBM to perform variable screening are as follows: 5-1) For the established model, use the following formula to calculate the SHAPvalue of the i-th feature of each sample; in, F is the set of all features, S is the feature subset of F, f S is the model trained on feature subset S, f S∪{k} is a model trained on the feature subset S∪{k}, F{k} represents the feature subset containing feature k, |S| represents the size of S, and |F| represents the size of F; 5-2) For feature i, calculate the absolute value of the SHAP value of each sample and then take the average value as the feature contribution of feature i. Then, sort the features in descending order according to the feature contribution. 5-3) Use the previous feature with the largest contribution as a feature subset to build a LightGBM model for the target value, and then use the first two features as feature subsets to build a LightGBM model for the target value, and repeat this process until all features have been modeled; 5-4) Select the feature subset with the best modeling effect.

3. An interpretable model system for predicting the ultimate tensile strength of tin-based solder using SHAP package LightGBM, which implements the method for quickly predicting the ultimate tensile strength of tin-based solder using SHAP package LightGBM as described in any one of claims 1-2, It is characterized in that include: Input module: Find the element composition, experimental conditions and ultimate tensile strength experimental value data of tin-based solder alloy from publicly published literature and use them as input data; Data analysis module: using the data obtained from the input module, execute the method of SHAP package LightGBM to quickly predict the ultimate tensile strength of tin-based solder, and quickly predict the ultimate tensile strength of tin-based solder; Output module: Output the ultimate tensile strength data of tin-based solder.

Images