Wafer yield classification method and system based on high-dimensional alternating feature screening

Abstract

The invention discloses a wafer yield classification method and system based on high-dimensional alternating feature screening, and belongs to the field of integrated circuit process manufacturing. Collecting acceptance test parameters of the wafer samples, and expanding the minority class samples by adopting a synthetic minority class oversampling method to obtain a wafer sample feature set with balanced qualified and unqualified classes; screening features from the wafer sample feature set by using a high-dimensional alternating feature screening method to obtain a feature subset; according to the high-dimensional alternating feature screening method, a fast correlation filtering model serves as a filter and is included in a local effective weight wrapper to achieve high-dimensional alternating feature screening. And training a wafer yield classification model based on XGBoost by using the screened feature subsets and adopting a genetic algorithm, and carrying out wafer yield classification by using the trained model. According to the invention, through comprehensive application of unbalance correction and high-dimensional feature alternate screening technologies, acceptance test parameters of the wafer sample are optimized, feature redundant information is reduced, and a prediction effect is improved.

Description

Technical Field

[0001] The present invention belongs to the field of integrated circuit process manufacturing, and particularly relates to a wafer yield classification method and system based on high-dimensional alternating feature screening. Background Art

[0002] In semiconductor manufacturing, maintaining strict quality control is crucial for cost management and timely product delivery. Wafer yield classification prediction is a means of quality control, aiming to classify wafers into two categories: low yield (unqualified) and high yield (qualified). During the wafer production and testing process, fluctuations in various process parameters related to electrical testing are the key factors causing yield loss. To accelerate the identification of process parameters affecting the yield and effectively control and improve the early detection of potential defective wafers, current research mainly focuses on developing methods for quickly correcting process defects and solving quality control problems, which is crucial for continuously increasing production, saving resources, and improving production efficiency.

[0003] The quality control process in semiconductor manufacturing includes defect detection during the manufacturing process, wafer acceptance test (WAT) after manufacturing, and circuit probing (CP) of each chip on the wafer. WAT is to test the electrical parameters of specific test keys on the wafer using an automatic test equipment (ATE) after wafer processing and before quality inspection, providing process test indicators for each wafer product. These test keys are formed on the scribe lines between wafers and are electrically coupled to the external circuit through metal pads, used to study various electrical characteristics such as breakdown voltage and leakage current by applying a preset bias. The parameters at this stage can reflect abnormal process fluctuations during the wafer manufacturing process to determine whether the wafer meets the electrical specifications of the technology platform. After WAT, engineers use the CP process to obtain wafer yield data and decide whether to advance the wafer to the next production stage or discard it based on the CP results. However, the CP process is usually carried out on expensive equipment and often leads to a delay in defect detection, increasing the cost for semiconductor manufacturers. Therefore, engineers usually use a wafer yield prediction method based on WAT parameters.

[0004] As the size of integrated circuits continues to shrink and the processing technology becomes increasingly complex, the number of process parameters that need to be detected in WAT is also increasing. Traditional wafer yield prediction methods mainly consider the source and number of defects on the wafer, as well as the relationship between the degree of defect aggregation and the yield. By using manual analysis methods such as manual variance analysis, average value analysis, and trend chart comparison, not only does it push up the labor cost, but it also cannot provide dynamic real-time feedback and assist in the adjustment of production processes. In addition, research shows that the distribution of yield data collected usually skews towards the high-yield end, and low-yield wafers are often in the minority, which easily leads to an unbalanced data distribution, making the prediction results tend to favor the majority class. Moreover, due to the complexity of the semiconductor production process, WAT parameters also exhibit intricate interrelationships, accompanied by obvious redundancy problems; at the same time, the correlation between WAT parameters and yield classification labels also presents multi-level complexity. Traditional statistical process control methods (SPC) have limitations for large-scale automatic parameter identification, and it is a challenge for engineers to effectively identify the abnormal causes leading to yield losses in the wafer fab in a short time. To save the corresponding time consumption and testing costs, yield prediction technology is developing towards a more efficient and intelligent direction. Therefore, there is an urgent need for an efficient and highly interpretable wafer yield classification prediction method to accelerate the yield improvement cycle. Summary of the Invention

[0005] In view of the above problems, the present invention proposes a wafer yield classification method and system based on high-dimensional alternating feature screening, and the technical solutions adopted are as follows:

[0006] In the first aspect, the present invention proposes a wafer yield classification method based on high-dimensional alternating feature screening, including:

[0007] Collect the acceptance test parameters of wafer samples, and use the synthetic minority over-sampling technique to expand the minority samples to obtain a balanced wafer sample feature set of qualified and unqualified categories;

[0008] Use the high-dimensional alternating feature screening method to screen features from the wafer sample feature set to obtain a feature subset; the high-dimensional alternating feature screening method uses the fast correlation filtering model as a filter and incorporates it into the local effective weight wrapper to achieve high-dimensional alternating feature screening;

[0009] Use the screened feature subset to train an XGBoost-based wafer yield classification model using the genetic algorithm, and use the trained model for wafer yield classification.

[0010] Further, when using the synthetic minority over-sampling technique to expand the minority samples, select one sample from several nearest neighbor samples of each sample in the minority class, and interpolate to obtain a new sample of the same class:

[0011] Snew =S i +rand(0,1)*|S i -S n |

[0012] Among them, S new is a new synthetic sample, rand(0,1) represents a random number in the interval [0,1], S i is a sample in the minority class, S n is sample S i A nearest neighbor sample of .

[0013] Furthermore, in the high-dimensional alternating feature screening method, the FCBF filter model is embedded into the LVW wrapper model, which is expressed as:

[0014]

[0015] Among them, D T Represents the feature subset finally screened. represents the optimal feature subset obtained by LVW wrapper screening, C represents the category set, and FCBF represents the FCBF filter model.

[0016] Furthermore, the implementation process of the high-dimensional alternating feature screening method is specifically as follows:

[0017] (1) Feature screening based on wrapper

[0018] (1-1) Initialize the maximum number of iterations and error threshold;

[0019] (1-2) In each iteration, randomly generate feature subsets of any size and feature combination, evaluate the error index of the feature subsets, and select the feature subset with the smallest error index as the optimal feature subset;

[0020] (1-3) Repeat the iterative process of step (1-2) until the maximum number of iterations is met or the error index is lower than the error threshold, and the final optimal feature subset is obtained;

[0021] (2) Filter-based feature screening

[0022] (2-1) Calculate the correlation between each feature variable in the optimal feature subset and the target classification;

[0023] (2-2) Only feature variables with correlations greater than a certain threshold are retained to form a candidate list;

[0024] (2-3) Sort the feature variables in the candidate list from large to small according to the correlation value;

[0025] (2-4) Calculate each characteristic variable f in sequencei For each feature variable f that is ranked lower in the candidate list j in terms of relevance, if SU(f i , f j ) is greater than SU(f i , c k ), then delete f from the candidate list Slist j ; where c k represents the target classification, corresponding to qualified or unqualified; SU represents relevance;

[0026] (2 - 5) Repeat steps (2 - 1) to (2 - 4) several times to obtain the final feature subset.

[0027] Furthermore, when training an XGBoost - based wafer yield classification model using a genetic algorithm, the parameter combinations to be optimized include the number of iterations, learning rate, and maximum tree depth of the XGBoost model.

[0028] In a second aspect, the present invention proposes a wafer yield classification system based on high - dimensional alternating feature screening for implementing the above - mentioned wafer yield classification method based on high - dimensional alternating feature screening.

[0029] In a third aspect, the present invention proposes a computer - readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the wafer yield classification method based on high - dimensional alternating feature screening as described above.

[0030] In a fourth aspect, the present invention proposes a computer electronic device, including a memory and a processor;

[0031] The memory is used to store a computer program;

[0032] The processor is used to implement the wafer yield classification method based on high - dimensional alternating feature screening as described above when executing the computer program.

[0033] The beneficial effects of the present invention are:

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] (1) Classification algorithms trained on imbalanced sample distributions often lead to a decline in prediction quality because the model tends to favor the majority class, thus neglecting the minority class that is crucial for many key use cases. Synthetic Minority Over-sampling Technique (SMOTE) adopted in the present invention expands the available minority class samples through an interpolation strategy by synthesizing new data points between real observations. This method functionally fills the scarce areas of minority class samples, effectively reducing the model bias caused by data imbalance and correcting the original uneven distribution of wafer samples.

[0036] (2) Aiming at the problems of high dimensionality, redundancy, and insignificance of key parameters in WAT parameters in the real world, the present invention proposes an improved high-dimensional alternating feature selection method (HFAS). This method combines the FCBF filter embedded in the LVW wrapper. By integrating the FCBF filter into the LVW wrapper, it can more effectively screen out the WAT feature parameters that are more crucial and significant for the correlation with the yield, while reducing the redundant information between parameters, and achieving accurate prediction of wafer yield with fewer key WAT parameter inputs, so as to reduce the test damage to wafer and equipment investment.

[0037] (3) By comprehensively applying imbalance correction (SMOTE) and high-dimensional feature alternating screening (HFAS), the present invention successfully screens 1409 WAT feature parameters. This optimization not only provides a reference for reducing the use of wafer acceptance test equipment, but also effectively reduces the test cost. When the optimized features are input into the XGBoost model optimized by the genetic algorithm, the accuracy rate can reach more than 95%. Compared with a single classifier, the AUC and F1-score are increased by 21% and 19% respectively, indicating that this framework has significant advantages in wafer yield prediction. Description of the Drawings

[0038] Figure 1 The process of obtaining wafer yield data.

[0039] Figure 2 Schematic diagram of the process of using Synthetic Minority Over-sampling Technique (SMOTE) to process imbalanced data.

[0040] Figure 3 WAT feature screening model based on high-dimensional alternating feature selection (HFAS).

[0041] Figure 4 Flow framework diagram of the wafer yield classification method based on high-dimensional alternating feature screening.

[0042] Figure 5 The change of fitness during the optimization process of the genetic algorithm. Detailed Implementation Manner

[0043] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention.

[0044] The accompanying drawings are only schematic diagrams of the present invention and are not necessarily drawn to scale. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor devices and / or microcontroller devices.

[0045] The flowcharts shown in the accompanying drawings are only illustrative and do not necessarily include all steps. For example, some steps can be decomposed, while some steps can be combined or partially combined, so the actual execution order may change according to the actual situation.

[0046] A wafer yield classification method based on high-dimensional alternating feature screening proposed by the present invention uses the Synthetic Minority Over-sampling Technique (SMOTE) to process the unbalanced wafer sample yield distribution; adopts an improved high-dimensional feature alternating screening (HFAS) strategy that embeds a filter into a wrapper to screen key feature subsets; and uses an XGBoost classifier that tunes hyperparameters using a genetic algorithm (GA) to classify wafer yields.

[0047] As Figure 4 shown, it mainly includes the following steps:

[0048] 1. Preview and preprocess the obtained WAT dataset;

[0049] 2. Process the feature distribution characteristics of the WAT dataset;

[0050] 3. Use the improved high-dimensional feature alternating screening algorithm (HFAS) to perform feature selection on the preprocessed dataset to obtain the optimal feature subset;

[0051] 4. Divide the optimal feature subset to obtain a training set and a validation set, input the training set and the validation set into the XGBoost model for pre-training, and use the advantage of the genetic algorithm (GA) that can adjust and optimize hyperparameters according to the performance of the model to optimize the best classifier model.

[0052] 5. Input the test set into the finally trained excellent classifier model to further obtain the predicted classification result.

[0053] The preview and preprocessing of the WAT dataset in step 1 mainly include:

[0054] This embodiment uses the actual manufacturing data of a 55nm wafer production line as sample data to obtain a WAT dataset. As Figure 1 shown, it presents the complete process of a silicon wafer from processing to performance testing. After the silicon wafer undergoes a series of processing steps (lithography, etching, doping, deposition, planarization, etc.), after processing is completed, wafer acceptance testing (WAT) is carried out to obtain electrical parameters, and then circuit probe testing (CP Test) is performed to finally evaluate the wafer yield.

[0055] As shown in Table 1, before data preprocessing, this dataset covers 1657 wafer samples and 1409 WAT parameters. Among them, the WAT parameters, as sample features, mainly include the wafer number, test point coordinates, and their test results; the test results include parameters such as the saturation region threshold voltage (Vts), linear region threshold voltage (Vtl), off-state leakage current (Ioff), saturation current (Isat), power-off consumption (Poff), subthreshold slope, noise figure, and input and output impedance. It can be seen that the sample features have high-dimensional characteristics. Taking the yield label as the sample diagnostic label, since this invention only considers the categories of wafer yield, such labels usually only contain two categories: qualified or unqualified. In addition, as can be seen from Table 1, there are very few unqualified wafers in this wafer manufacturing system, so the sample distribution is generally seriously unbalanced.

[0056] Table 1 Data description and preprocessing

[0057]

[0058] Preprocess the original data in Table 1:

[0059] 1.1 Handling of missing values:

[0060] The following three methods can be selected but are not limited to:

[0061] Delete the sample data corresponding to the missing values;

[0062] Fill the missing values with the mean, mode, or median;

[0063] Build a missing value prediction model based on the existing data to fill the missing values.

[0064] 1.2 Handling of data normalization:

[0065] The following two methods can be selected but are not limited to:

[0066] Standard normalization method: Standardize the data to convert it into a standard normal distribution with a mean of 0 and a standard deviation of 1. The specific process is as follows:

[0067]

[0068] Where x is the original data sample, μ is the mean of all data samples, and σ is the standard deviation of all data samples;

[0069] The maximum - minimum normalization method: Scale the data to a specific range (usually 0 to 1), and the specific process is as follows:

[0070]

[0071] Where x is the original data sample, min(x) and max(x) are the minimum and maximum values of the data sample respectively.

[0072] Step 2 uses the Synthetic Minority Over - sampling Technique (SMOTE) to address the problem of unbalanced data distribution.

[0073] To solve the problem of unbalanced sample distribution in the WAT dataset, the present invention uses the Synthetic Minority Over - sampling Technique to oversample the minority class samples (unqualified wafers). By generating synthetic minority class samples, the yield labels of qualified and unqualified wafers are balanced, enhancing the discrimination effect of the classification model in the imbalanced dataset.

[0074] This embodiment uses the Synthetic Minority Over - sampling Technique (SMOTE), randomly selects a minority class sample S i , and determines its k nearest neighboring minority class samples according to the feature space distance, and then generates new synthetic samples S new by interpolating in turn between the initial sample and the selected k neighbors, so that the distribution of the generated samples is similar to that of the actual samples. The generated synthetic samples are combined with the original data to initially obtain the input data matrix D.

[0075] Specifically, as Figure 2 shown, the specific steps of SMOTE sampling include:

[0076] (1) For each sample S in the minority class of the samples i , using the Euclidean distance as an index, calculate the distance from S i to all samples in the same class sample set to obtain the corresponding k nearest neighbor samples.

[0077] (2) For S i , randomly select several samples from its k nearest neighbor samples. Assume that the selected nearest neighbor is S n . For each selected S i , generate a new sample S new from the original sample, which can be expressed as:

[0078] S new = S i+rand(0,1)*|S i -S n |

[0079] where \(n=(1,2,3,\cdots)\), \(rand(0,1)\) represents a random number in the interval \([0,1]\), that is, the new sample point is randomly selected on the line connecting the original sample point \(S\) i and its neighbor \(S\) n in between.

[0080] (3) According to the imbalance ratio of the dataset, by setting \(N\) to represent the percentage of the newly added minority class samples relative to the number of the original minority class samples (for example, \(N = 100\) means the number of newly added samples is the same as the original number).

[0081] (4) Repeat the above steps until a preset number of minority class sample points are generated.

[0082] (5) Combine the training set and the test set to obtain the matrix \(D\) as the input data.

[0083] In step 3, the HFAS fusion algorithm is used to perform feature selection on the preprocessed dataset to obtain the optimal feature subset, specifically including:

[0084] After establishing the dataset, the problem of uneven data distribution is initially eliminated. However, due to the complexity of semiconductor manufacturing steps, WAT parameters often exhibit complex interactions, and there are also multi-layer and non-linear correlations between WAT parameters and yield classification results. As Figure 3 shown, high-dimensional alternating feature screening proposes an innovative feature screening technology, which models the complex correlation between WAT parameters and yield classification results by incorporating the fast correlation-based filtering (FCBF) model as a filter into the strategy of the locally valid weight (LVW) wrapper. Specifically as follows:

[0085] 3.1 Wrapper-based feature screening:

[0086] Initialization phase: LVW needs to set some key parameters, such as the maximum number of iterations \(T\), the error threshold \(E\), etc.

[0087] Randomly generate a feature subset: In each iteration, LVW randomly generates a feature subset \(A'\) of any size and feature combination from the dataset \(D\).

[0088] Evaluate the feature subset: The generated feature subset will be used to train the learner, and through cross-validation E′ = CrossValidation(ζ(A′)), that is, by dividing the dataset into multiple subsets of the same size, and each time selecting different subsets as the training set and the test set respectively, training and testing the learner repeatedly, and taking the average of the performance results (AUC, accuracy, and F1-score) obtained from each cross-validation as the final error metric E′.

[0089] Record and compare the errors: LVW will record the errors of the feature subsets in each iteration and compare them with the previous errors.

[0090] Select the optimal feature subset: According to the results of the error comparison, LVW will select the feature subset A with the smallest error * as the current optimal feature subset.

[0091] Iteration and termination: LVW will repeat the above steps for multiple iterations until the number of iterations reaches the maximum number of iterations T. Or

[0092] 3.2 Feature screening based on filters

[0093] Calculate the correlation between each feature variable F = [f1, f2, … f i ,..., f m T (i = 1, 2, 3,..., m) and the target variable C = [c1, c2] T The correlation between where I(F, C) is used to measure the mutual correlation between F and C, and H(F) and H(C) represent the information entropy of the feature variable and the target variable respectively. By introducing entropy as a measure of the average amount of information of a random variable, the correlation between variables can be better explained.

[0094] Define the information entropy H(F) of the feature variable F and the conditional entropy H(F|C) of the feature variable F under C as shown below:

[0095]

[0096] p(f i ) is the prior probability when F takes a value, p(c j ) is the prior probability when C takes a value, and p(f i |c j ) is the posterior probability of C given that F takes a value. According to the definition, the correlation measure between the random variables F and C is the information gain I(F, C), as shown below:

[0097] I(F, C) = H(F) - H(F|C)

[0098] Combining the formula of conditional entropy \(H(F|C)\) and the formula of information gain \(I(X,C)\), the expression of uncertainty \(SU\) is obtained:

[0099]

[0100] The value range of \(SU\) is \((0,1)\). As shown in the above formula, the closer the value of \(SU\) is to 1, the stronger the correlation between \(F\) and \(C\).

[0101] In the present invention described above, in order to maximize the correlation between the preliminary screening feature \(f\) i and the target classification \(c\) j while minimizing the redundancy between features,

[0102] 1) Only retain the feature variables \(f\) i , \(c\) j for which \(SU(f\) i , \(c\) list ) is greater than a certain threshold, and form a candidate list \(F\)

[0103] 2) Sort the variables in \(F\) i , \(c\) j ) in descending order according to the value of \(SU(f\) list ;

[0104] 3) Calculate the correlation \(SU\) value \(SU(f\) i , \(f\) i , \(f\) j ) between each feature \(f\)

[0105] 4) If \(SU(f\) i , \(f\) j ) is greater than \(SU(f\) i , \(c\) j ), then delete \(f\) j from the candidate list \(Slist\);

[0106] 5) Repeat the above steps 1)-4) to obtain the final feature subset \(D\) that meets the conditions T .

[0107] Embedding the FCBF filter model into the LVW wrapper model, the specific process can be expressed as:

[0108]

[0109] Steps 4 and 5 mainly further predict and evaluate the classification performance by training the best classifier XGBoost.

[0110] Using the final feature subset \(D\) obtained in step 3 TDivide it into a training set and a test set. Next, input the training set into the XGBoost model for pre-training. During this process, the Genetic Algorithm (GA) is used for parameter optimization.

[0111] The genetic algorithm generates a set of parameter combinations by simulating the process of natural selection, and gradually optimizes these parameters including the number of iterations (estimator), learning rate (the step size when updating weights in each iteration), and maximum tree depth (max depth) through operations such as crossover and mutation to improve the training effect of the model. Finally, evaluate the model performance through the test set to ensure that the selected parameter combination can perform well on unseen data.

[0112] In a specific implementation of the present invention, the data set is divided into a training set (80%) and a test set (20%), and extreme gradient boosting (XGBoost) is used to achieve classification. This algorithm integrates multiple base learners to construct a strong learner. Compared with the traditional Gradient boosting Decision Tree (GBDT) algorithm,

[0113] In the present invention, optimizing the hyperparameters of the XGBoost model is the key to improving the model's prediction ability. To improve the model's performance on a specific data set and effectively manage resources under limited computing resources, the present invention uses the genetic algorithm (GA) to determine the hyperparameters (such as model architecture, regularization, etc.) of the XGBoost model in the prediction stage. The following steps illustrate how the genetic algorithm can be applied as a tool for determining the optimal performance parameter combination by simulating the process of natural evolution to address search and optimization challenges. The following steps explain the entire process.

[0114] (1) Initialization: Generate an initial population, where each individual represents three main parameters of XGBoost, including the number of iterations (estimator), learning rate (the step size when updating weights in each iteration), and maximum tree depth (maxdepth);

[0115] (2) Fitness evaluation: Evaluate the fitness of each individual, that is, calculate its corresponding objective function value:

[0116] f i =fitness(pop i (t))

[0117] (3) Selection: Select a specific group from the population as the parents of the next generation;

[0118] (4) Crossover: Perform a crossover operation on the parental individuals to generate new offspring;

[0119] (5) Mutation: Mutate individuals in the offspring. Introduce random factors to increase population diversity and prevent the algorithm from falling into local optimal solutions.

[0120] (6) Update the population: Combine the parent generation and the offspring to generate a new population;

[0121] (7) Judgment of termination conditions: Judge whether the termination conditions are reached, such as the maximum number of iterations and the objective function reaching a certain threshold.

[0122] (8) Output: Output the objective function corresponding to the optimal solution or the value of the optimal solution.

[0123] The present invention uses a genetic algorithm to optimize three main parameters of XGBoost, including the number of iterations (estimator), the learning rate (the step size when updating weights in each iteration), and the maximum tree depth (max depth). Record the optimal individual with the highest fitness value in each iteration. Then, reproduce offspring through methods such as selection, crossover, and mutation to obtain better genes and avoid falling into local optimal solutions. As Figure 5 shown in the change of fitness during the optimization process of the genetic algorithm, it can be seen that as the number of iterations increases, the fitness value gradually increases. The adjusted recognition accuracy has been significantly improved, indicating that the genetic algorithm plays an optimization role in the model. The two optimal parameter combinations of the current XGBoost model are shown in Table 2.

[0124] Table 2 Optimal hyperparameter combinations

[0125] Parameter Initial value Optimal solution 1 Optimal solution 2 Number of iterations 30 80 110 Learning rate 0.05 0.1 0.8 Maximum tree depth 13 4 2 F1-score 0.91 0.95 0.97

[0126] Table 3 Performance comparison chart of 9 classification prediction models

[0127]

[0128] As shown in Table 2, the following conclusions can be drawn:

[0129] (1) In terms of the selection of the yield classifier for wafers, XGBoost is superior to Adaboost, SVM, and DT.

[0130] (2) In terms of data distribution processing, using SMOTE for imbalance correction is more conducive to correcting the unbalanced performance of wafer sample distribution.

[0131] (3) By comprehensively applying imbalance correction and feature selection, HFAS consists of an FCBF filter embedded in an LVW wrapper and generally performs better than a single FCBF filter. It can be seen that the maximum number of features retained after pre-screening is 80 parameters, and the minimum is 10 parameters. Compared with the original 1409 WAT parameters, irrelevant noise features have been filtered out more fully.

[0132] (4) The accuracy rate can reach over 95%. Compared with a single classifier without sample distribution balance correction and feature screening, the AUC and F1-score of SMOTE-HAFS-XGBoost have increased by 21% and 19% respectively, indicating the advantages of this invention in improving the prediction effect of wafer yield.

[0133] Since the data distribution obtained from a small-batch wafer manufacturing system is highly imbalanced, in an imbalanced dataset, if the model is more inclined to predict majority-class samples, the accuracy rate may be high, but the f1-score and AUC may be low. Therefore, it is necessary to introduce these three indicators simultaneously and then weigh them to avoid contradictions.

[0134] The HAFS model proposed in this invention for the high-dimensional and redundant features presented in the WAT dataset takes into account the influence of individual WAT parameters and combined WAT parameters on wafer yield. It can not only improve the time efficiency of the algorithm but also significantly improve the effect of key parameter identification, solving the challenge that other existing technologies may not be able to effectively capture the key features of minority-class samples. Finally, by optimizing the decision boundary of the classifier, the classification accuracy of different samples is improved. In addition, the proposed technology achieves the best performance in terms of the AUC value. The higher the AUC value, the better the classification performance of the model, that is, it can identify unqualified wafers with a higher true rate while maintaining a lower false positive rate. Through experimental verification, the effectiveness of the prediction framework (SMOTE-HFAS-XGBoost) described in this invention in the accuracy of WAT predicted yield classification is demonstrated.

[0135] This invention also provides a wafer yield classification system based on high-dimensional alternating feature screening, which is used to implement the above embodiments. Terms such as "module" and "unit" used below can be a combination of software and / or hardware that can achieve a predetermined function. Although the system described in the following embodiments is preferably implemented in software, implementation in hardware, or a combination of software and hardware is also possible.

[0136] A wafer yield classification system based on high-dimensional alternating feature screening provided in this embodiment, the system includes:

[0137] A data preprocessing module, which is used to collect the acceptance test parameters of wafer samples, and adopts the synthetic minority over-sampling technique to expand the minority samples, so as to obtain a feature set of wafer samples with balanced qualified and unqualified categories;

[0138] A high-dimensional alternating feature screening module, which is used to screen features from the wafer sample feature set by using the high-dimensional alternating feature screening method to obtain a feature subset; the high-dimensional alternating feature screening method uses the fast correlation filtering model as a filter and incorporates it into the local effective weight wrapper to achieve high-dimensional alternating feature screening;

[0139] A classification model training module, which is used to train an XGBoost-based wafer yield classification model by using the screened feature subset with the genetic algorithm;

[0140] A wafer yield classification module, which is used to classify the wafer yield by using the trained model.

[0141] For the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can be referred to the partial description of the method embodiment, and the implementation method of the module will not be elaborated here. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present invention. Those of ordinary skill in the art can understand and implement it without creative work.

[0142] The embodiments of the system of the present invention can be applied to any device with data processing capabilities, and the any device with data processing capabilities can be a device or apparatus such as a computer. The system embodiment can be implemented by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a logically meaningful device, it is formed by the processor of any device with data processing capabilities reading the corresponding computer program instructions in the non-volatile memory into the memory for operation.

[0143] In addition, it should be noted that the above-mentioned wafer yield classification method based on high-dimensional alternating feature screening can essentially be executed by a computer program. Therefore, similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer electronic device corresponding to the method provided in the above embodiment, which includes a memory and a processor;

[0144] The memory is used to store computer programs;

[0145] The processor is configured to implement the wafer yield classification method based on high-dimensional alternating feature screening in the above embodiments when executing the computer program.

[0146] In addition, when the logical instructions in the above-mentioned memory are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0147] Thus, based on the same inventive concept, in another preferred embodiment of the present invention, there is also provided a computer-readable storage medium corresponding to the method provided in the above embodiment. A computer program is stored on the storage medium, and when the computer program is executed by a processor, it can implement the wafer yield classification method based on high-dimensional alternating feature screening in the above embodiment.

[0148] It can be understood that the above storage medium may include a random access memory (RAM), and may also include a non-volatile memory (NVM), such as at least one disk memory. At the same time, the storage medium may also be various media such as a USB flash drive, a mobile hard disk, a magnetic disk, or an optical disc that can store program codes.

[0149] It can be understood that the above-mentioned processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0150] In addition, it should be noted that those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the system described above can refer to the corresponding process in the foregoing method embodiments, and will not be elaborated here. In the embodiments provided in the present application, the division of steps or modules in the system and method is only a logical function division, and there may be other division methods in actual implementation. For example, multiple modules or steps can be combined or integrated together, and a module or step can also be split.

[0151] The above-described embodiments are only a preferred solution of the present invention, but they are not intended to limit the present invention. Those of ordinary skill in the relevant art can still make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all technical solutions obtained by means of equivalent replacement or equivalent transformation fall within the protection scope of the present invention.

Claims

1. A wafer yield classification method based on high-dimensional alternating feature screening, characterized in that, Including: Collect the acceptance test parameters of wafer samples, and use the Synthetic Minority Over-sampling Technique (SMOTE) to expand the minority class samples to obtain a wafer sample feature set with balanced qualified and unqualified categories; Use the High-dimensional Alternating Feature Screening (HAFS) method to screen features from the wafer sample feature set to obtain a feature subset; in the HAFS method, the Fast Correlation-Based Filter (FCBF) model is used as a filter and incorporated into the Local Validity Weighting (LVW) wrapper to achieve high-dimensional alternating feature screening; Use the screened feature subset to train an XGBoost-based wafer yield classification model using a genetic algorithm, and use the trained model for wafer yield classification.

2. The wafer yield classification method based on high-dimensional alternating feature screening according to claim 1, wherein When using the SMOTE method to expand the minority class samples, select one sample from several nearest neighbor samples of each sample in the minority class and interpolate to obtain a new sample of the same class: S new = S i + rand(0,1) * |S i - S n | Among them, S new is a newly synthesized sample, and rand(0,1) represents a random number in the interval [0,1]. S i is a sample in the minority class, and S n is a sample of S i 's nearest neighbor sample.

3. The wafer yield classification method based on high-dimensional alternating feature screening according to claim 1, characterized in that Before expanding the minority class samples, it also includes handling missing values and normalizing the samples.

4. The wafer yield classification method based on high-dimensional alternating feature screening according to claim 1, characterized in that In the HAFS method, the FCBF filter model is embedded into the LVW wrapper model, expressed as: D T = FCBF(D A* , C) Among them, D T represents the finally selected feature subset, and D A* represents the optimal feature subset selected by the LVW wrapper. C represents the class set, and FCBF represents the FCBF filter model.

5. The wafer yield classification method based on high-dimensional alternating feature screening according to claim 1 or 4, characterized in that The implementation process of the HAFS method is specifically as follows: (1) Feature screening based on the wrapper (1-1) Initialize the maximum number of iterations and the error threshold; (1-2) In each iteration, randomly generate a feature subset of any size and feature combination, evaluate the error index of the feature subset, and select the feature subset with the smallest error index as the optimal feature subset; (1-3) Repeat the iterative process of step (1-2) until the maximum number of iterations is reached or the error index is lower than the error threshold to obtain the final optimal feature subset; (2) Feature screening based on the filter (2-1) Calculate the correlation between each feature variable in the optimal feature subset and the target classification; (2-2) Only retain the feature variables with a correlation greater than a certain threshold to form a candidate list; (2-3) Sort the feature variables in the candidate list from largest to smallest according to the correlation value; (2-4) Calculate each feature variable f in sequence i The correlation with each feature variable f that is ranked lower in the candidate list j , if SU(f i , f j ) is greater than SU(f i , c k ), then delete f from the candidate list Slist j ; where c k represents the target classification, corresponding to qualified or unqualified; SU represents the relevance; (2-5) Repeat steps (2-1) to (2-4) several times to obtain the final feature subset.

6. The wafer yield classification method based on high-dimensional alternating feature screening according to claim 1, characterized in that When training an XGBoost-based wafer yield classification model using a genetic algorithm, the parameter combinations to be optimized include the number of iterations, learning rate, and maximum tree depth of the XGBoost model.

7. A wafer yield classification system based on high-dimensional alternating feature screening for implementing the wafer yield classification method described in claim 1, characterized in that, The system includes: A data preprocessing module, which is used to collect the acceptance test parameters of wafer samples, and use the SMOTE method to expand the minority class samples to obtain a wafer sample feature set with balanced qualified and unqualified categories; A high-dimensional alternating feature screening module, which is used to screen features from the wafer sample feature set using the HAFS method to obtain a feature subset; in the HAFS method, the FCBF model is used as a filter and incorporated into the LVW wrapper to achieve high-dimensional alternating feature screening; A classification model training module, which is used to train an XGBoost-based wafer yield classification model using a genetic algorithm with the screened feature subset; A wafer yield classification module, which is used to perform wafer yield classification using the trained model.

8. The wafer yield classification system based on high-dimensional alternating feature screening according to claim 7, wherein The high-dimensional alternating feature screening module includes: A wrapper-based feature screening unit, which is used to screen out the optimal feature subset by using the LVW wrapper model; A filter-based feature screening unit, which is used to further screen out the final feature subset based on the optimal feature subset screened out by the wrapper-based feature screening unit.

9. A computer-readable storage medium, characterized in that, A computer program is stored on the storage medium, and when the computer program is executed by a processor, the wafer yield classification method based on high-dimensional alternating feature screening as described in any one of claims 1 to 6 is implemented.

10. A computer electronic device, characterized in that, Comprising a memory and a processor; The memory is used to store a computer program; The processor is used to implement the wafer yield classification method based on high-dimensional alternating feature screening as described in any one of claims 1 to 6 when executing the computer program.

Images