Error cause inferring apparatus and inferring method

Abstract

The present invention uses an error cause inference device including: a data preprocessing section that generates training data having a format suitable for input to a machine learning model using processing target data; and a model tree generation section that generates a learning model for detecting an error, i.e., an error detection model, with the training data as input, and generates a model tree that expresses relationships between error detection models using a tree structure with error detection models as nodes. Thereby, even without adding labels to error causes in advance, it is possible to generate learning models that detect errors by various types of errors that occur.

Description

Technical Field

[0001] This invention relates to an apparatus and method for inferring the cause of an error. Background Technology

[0002] Semiconductor metrology and inspection equipment perform inspection and measurement actions on each inspection point on the surface of a semiconductor wafer according to set parameters known as a recipe. In adjusting the recipe parameters, engineers typically optimize each item manually based on the attributes of the object being measured / inspected and the characteristics of the equipment. Therefore, errors can occur during inspection and measurement actions, for example, if an insufficiently adjusted recipe is used or if the characteristics of the equipment change over time. Such errors, caused by the content of the recipe, are called recipe errors.

[0003] When a formula error occurs, maintenance engineers typically analyze internal data from semiconductor measurement and inspection devices to pinpoint the cause. However, with the miniaturization and diversification of semiconductors, the number of formulas and formula setting items have increased, making formula generation more complex. Therefore, determining the cause of formula errors is time-consuming, contributing to reduced equipment uptime.

[0004] Patent Document 1 discloses a technology that aims to reduce the training time by saving the workload of truth value generation of training data and reducing the amount of training data. It provides a pattern inspection system that uses an image of the pattern of the object to be inspected based on an electronic device and data used to manufacture the pattern of the object to be inspected. The system uses a recognizer constructed by machine learning to inspect the image of the pattern of the object to be inspected. The image selection unit (training data selection unit) selects a training pattern image for machine learning from multiple pattern images based on pattern data and pattern images stored in the storage unit. It also clusters data containing multiple position coordinates of the same pattern stored for each pattern into one or more clusters.

[0005] Existing technical documents

[0006] Patent documents

[0007] Patent Document 1: Japanese Patent Application Publication No. 2020-35282 Summary of the Invention

[0008] The technical problem that the invention aims to solve

[0009] The training data selection unit of the pattern inspection system described in Patent Document 1 clusters data from multiple position coordinates based on the same pattern. However, this training data selection unit does not use the contribution of each parameter to the error as a benchmark, so it can be considered that there is room for improvement.

[0010] The purpose of this invention is to generate a learning model that can detect errors of various possible types even without pre-labeling the causes of errors. Furthermore, this invention also aims to reduce the amount of data labeling processing.

[0011] Technical means to solve the problem

[0012] The error cause inference apparatus of the present invention includes: a data preprocessing unit that generates training data with a format suitable for input into a machine learning model using processing object data; and a model tree generation unit that generates a learning model for detecting errors, i.e., an error detection model, using the training data as input, and generates a model tree that uses a tree structure with error detection models as nodes to express the relationship between error detection models.

[0013] Invention Effects

[0014] According to the present invention, a learning model can be generated that can detect errors according to the various types of errors that can occur, even without pre-labeling the causes of errors. Furthermore, according to the present invention, the amount of labeling processing can be reduced. Attached Figure Description

[0015] Figure 1 This is a structural diagram illustrating an example of an information processing system that includes an error cause inference device according to an embodiment.

[0016] Figure 2 It means Figure 1 The structural diagram of the model tree generation part.

[0017] Figure 3 It means Figure 2 The flowchart of the steps in the model tree generation section.

[0018] Figure 4 It means Figure 3 This is a schematic diagram of the steps involved in error data clustering and error detection model generation, which are part of the model tree generation process.

[0019] Figure 5 It means Figure 1 The structural diagram of the error cause inference department.

[0020] Figure 6 This is a flowchart illustrating the steps involved in calculating the probability of an error cause in the error cause inference section.

[0021] Figure 7 This is a schematic diagram illustrating the structure for synthesizing and visualizing the probability of error causes obtained by analyzing multiple error detection models.

[0022] Figure 8 This is a terminal screen showing one example of the error cause probability obtained by analyzing multiple error detection models.

[0023] Figure 9 This is one example of a terminal screen displaying information about an error detection model in a model tree.

[0024] Figure 10 This is a terminal screen showing one example of an operation that replaces an error detection model in the model tree with another error detection model. Detailed Implementation

[0025] The apparatus and method for inferring the cause of errors in the embodiments of this disclosure are summarized and explained.

[0026] The error cause inference device includes a data preprocessing unit and a model tree generation unit.

[0027] Preferably, the model tree generation unit includes: a clustering error detection model generation unit, which takes incompletely clustered data from the training data as input, learns error detection rules based on the different trends of normal data and error data, and generates a clustering error detection model; a model analysis unit, which calculates the sensitivity value representing the degree of contribution of the feature quantity input as training data to the output of the error detection model; a data clustering unit, which clusters the data based on the feature quantity and sensitivity value; a clustering completion judgment unit, which judges whether the clustering is complete for the clustered data; a sub-causal error detection model generation unit, which takes completed clustered data as input, learns error detection rules based on the different trends of normal data and error data, and generates a sub-causal error detection model; and a model connection unit, which connects the clustering error detection model and the sub-causal error detection model based on the data clustering process, thereby generating a model tree composed of a phylogenetic tree with the clustering error detection model and the sub-causal error detection model as nodes.

[0028] Preferably, the clustering error detection model has a simpler model structure compared to the cause-based error detection model.

[0029] Preferably, the error cause inference apparatus further includes an error cause inference unit, which includes: a model evaluation unit that evaluates the performance of a pre-generated sub-cause error detection model (a pre-generated and stored sub-cause error detection model) using error cause inference object data as input, wherein the error cause inference object data is data generated by a data preprocessing unit for the data in the processing object data for which error causes are to be inferred, and has a format suitable for input into a machine learning model; a model selection unit that selects one or more sub-cause error detection models that have been evaluated as having high performance by the model evaluation unit; and a relevant parameter extraction unit that extracts the branch in the model tree where the selected sub-cause error detection model is located, and extracts the relevant parameters of the error detection model included in that branch.

[0030] Preferably, the error cause inference unit further includes an error cause probability calculation unit, which uses the relevant parameters extracted by the relevant parameter extraction unit as input to calculate the probability of candidate error causes.

[0031] Preferably, in the error cause inference apparatus, when multiple sub-cause error detection models are selected by the model selection unit, the probabilities obtained based on the sub-cause error detection models are corrected using the model evaluation value obtained by the model evaluation unit.

[0032] Preferably, the error cause inference apparatus also includes a model database that stores clustering error detection models and sub-causal error detection models in association with information about these models.

[0033] Preferably, the error cause inference device is configured such that when multiple versions of the clustering error detection model and the sub-cause error detection model stored in the model database exist, the user can use a terminal to replace the model in the model tree with other models stored in the model database.

[0034] Preferably, the data to be processed is at least one of the object's set parameters and measurement results.

[0035] Preferably, in the error cause inference device, the probabilities before and after correction are displayed on the terminal.

[0036] Preferably, the error cause inference device is configured such that, by having the user select an error detection model included in the model tree via a terminal, information associated with the selected error detection model is displayed.

[0037] A method for inferring the cause of an error includes: a data preprocessing step, wherein training data with a format suitable for input into a machine learning model is generated using the data to be processed; and a model tree generation step, wherein an error detection model is generated consisting of a learning model that detects errors using the training data as input, and a model tree is generated that uses a tree structure with the error detection models as nodes to express the relationships between the error detection models.

[0038] Preferably, the model tree generation step includes: a clustering error detection model generation step, wherein, using incompletely clustered data from the training data as input, error detection rules are learned based on the different trends of normal data and error data to generate a clustering error detection model; a model analysis step, wherein a sensitivity value representing the degree of contribution of the feature quantity input as training data to the output of the error detection model is calculated; a data clustering step, wherein data is clustered based on the feature quantity and sensitivity value; a clustering completion judgment step, wherein the clustering of the data is judged to determine whether the clustering is complete; a sub-causal error detection model generation step, wherein, using completed clustered data as input, error detection rules are learned based on the different trends of normal data and error data to generate a sub-causal error detection model; and a model connection step, wherein the clustering process based on the data connects the clustering error detection model and the sub-causal error detection model, thereby generating a model tree composed of a phylogenetic tree with the clustering error detection model and the sub-causal error detection model as nodes.

[0039] In the embodiments described below, the “semiconductor inspection apparatus” includes means for measuring the size of a pattern formed on the surface of a semiconductor wafer, means for inspecting whether the pattern formed on the surface of a semiconductor wafer is defective, or means for inspecting whether a bare wafer without a pattern is defective, and also includes a composite apparatus composed of a plurality of such means.

[0040] Additionally, in the embodiments described below, "inspection" is used to mean metrology or inspection, "inspection action" is used to mean measurement or inspection action, and "inspection object" refers to the wafer that is the object of measurement or inspection, or the area of ​​the wafer that is the object of measurement or inspection.

[0041] Example

[0042] Figure 1 This illustrates an example of an information processing system that includes an error cause inference device according to an embodiment.

[0043] In this diagram, the semiconductor inspection device 1 is connected to the database 2 and the error cause inference device 3 (error cause inference device) via network 101. The error cause inference device 3 is connected to the terminal 4. The error cause inference device 3 infers the cause of the error in the inspection operation performed by the semiconductor inspection device 1.

[0044] The data sent from the semiconductor inspection device 1 includes, for example, device data, measurement recipes (hereinafter sometimes simply referred to as "recipes"), measurement results, and error results. Additionally, the recipe may include the number of measurement points, the coordinate information of the measurement points (Evaluation Point: EP), the shooting conditions when capturing images, and the shooting sequence. Furthermore, the recipe may also include, along with the measurement points, the coordinates and shooting conditions of images obtained during the preparation stage for measuring the measurement points.

[0045] The device data includes inherent device parameters, instrument error correction data, and observation condition parameters. Inherent device parameters are correction data used to ensure the semiconductor inspection apparatus 1 operates according to specified specifications. Instrument error correction data are parameters used to correct instrument errors between semiconductor inspection apparatuses. Observation condition parameters, for example, are parameters specifying the observation conditions of a scanning electron microscope (SEM), such as the accelerating voltage of the electron optics system.

[0046] In the formulation, the formulation parameters include wafer pattern, alignment parameters, addressing parameters, and length measurement parameters. The wafer pattern is a coordinate diagram of the semiconductor wafer surface (e.g., the coordinates of a pattern). Alignment parameters are, for example, parameters used to correct for deviations between the coordinate system of the semiconductor wafer surface and the coordinate system inside the semiconductor inspection apparatus 1. Addressing parameters are, for example, information used to determine characteristic patterns located within the inspection area of ​​the pattern formed on the semiconductor wafer surface. Length measurement parameters are parameters describing the conditions for measuring length, such as parameters specifying which part of the pattern's length to measure.

[0047] The measurement results include length measurement results, image data, and an operation log. The length measurement results record the length of the pattern measured on the surface of the semiconductor wafer. The image data are observed images of the semiconductor wafer. The operation log records data on the internal state of the semiconductor inspection device 1 during each action step, such as alignment, positioning, and length measurement. For example, it can list the operating voltage of each component, the coordinates of the field of view, etc.

[0048] Error results are parameters that indicate which step in the process of alignment, positioning, and length measurement the error occurred in, when an error has occurred.

[0049] Data such as device information, formulas, measurement results, and error results are stored in database 2 via network 101. The stored data is analyzed by error cause inference device 3. The analysis results are displayed in a user-readable format via terminal 4 (GUI).

[0050] The error cause inference device 3 includes a data preprocessing unit 11, a model tree generation unit 12, an error cause inference unit 13, a model database 14 (model DB), and a model tree 15.

[0051] In the data preprocessing unit 11, the raw data (processing object data) such as device data, formulas, and measurement results sent from the database 2 are organized into a format suitable for machine learning models and output as training data. Here, data organization includes machine learning preprocessing such as handling missing values, deleting unnecessary variables, data scaling, encoding categorical variables, and generating composite features obtained by combining multiple data such as interaction features.

[0052] In the model tree generation unit 12, a hierarchical tree structure (hereinafter referred to as "model tree") for the learning model used to detect errors is generated using training data as input.

[0053] Figure 2 express Figure 1 The structure of the model tree generation unit 12.

[0054] The model tree generation unit 12 includes a clustering error detection model generation unit 21, a model analysis unit 22, a data clustering unit 23, a clustering completion judgment unit 24, a cause-based error detection model generation unit 25, and a model connection unit 26.

[0055] The data sent from database 2 is processed in data preprocessing unit 11 and output as training data 41. Training data 41 is sent to clustering error detection model generation unit 21. Then, training data 41 is processed in model tree generation unit 12 in the following order: clustering error detection model generation unit 21, model analysis unit 22, data clustering unit 23, clustering completion judgment unit 24, sub-causal error detection model generation unit 25, and model connection unit 26, and output as a model tree.

[0056] During this process, the clustering completion judgment unit 24 determines the clustering status, and depending on the situation, returns the data to the clustering error detection model generation unit 21. The clustering error detection model generation unit 21 and the sub-causal error detection model generation unit 25 then appropriately send the data to the model database 14.

[0057] Figure 3 express Figure 2 The processing steps of the constituent elements in the model tree generation section 12.

[0058] In step S101, in the clustering error detection model generation unit 21, the training data 41 that has not yet been clustered is used as input, and error detection rules are learned based on the different trends of normal data and error data to generate an error detection model for clustering. This error detection model can be generated using any machine learning algorithm such as Random Forest and XGBoost, which are based on decision trees, or Neural Networks.

[0059] In step S102, in the model analysis unit 22, for the error detection model generated by the clustering error detection model generation unit 21, a sensitivity value is calculated indicating the degree to which each feature in the input training data 41 contributes to the model output, i.e., the error prediction result. Regarding this sensitivity, for example, when the error detection model is constructed using a decision tree-based algorithm, it can be evaluated using variable importance calculated based on the number of times each feature appears in the branches within the model and the improvement value of the objective function, and SHAP (SHapley Additive exPlanations) values ​​used to calculate the contribution of each feature value to the model output.

[0060] In step S103, in the data clustering unit 23, the data is clustered into two or more clusters based on the relationship between the feature quantity and its sensitivity to the model. This clustering method can utilize unsupervised learning methods such as k-means and Gaussian mixture models.

[0061] In step S104, the clustering completion determination unit 24 determines whether clustering is complete for the data clustered by the data clustering unit 23. As a method for determining clustering completion, for example, the clustered data can be further divided into training data and test data; if the accuracy of the error detection model generated using the training data is above a threshold for the test data, then clustering is considered complete. For data determined to be clustered complete, a clustering completion flag is added.

[0062] In step S105, it is determined whether all data have been labeled as clustered. If there is data that has not yet been labeled, the processing of steps S101 to S104 is repeated for that data.

[0063] Figure 4 It schematically represents Figure 3 The processing from step S101 to step S104.

[0064] Figure 4 In this process, the data obtained in each step are represented using graphs or other means to make them clear.

[0065] In S101, an error detection model for clustering is generated. Then, in S102, the sensitivity of each feature to the error detection model, such as the SHAP value, is calculated. The relationship between a feature and its sensitivity is as follows: Figure 4 The scatter plot shown in S102 (Figure).

[0066] Unsupervised learning is used based on the relationship between this feature and sensitivity, such as... Figure 4 For example, the data is clustered into group A (Gr.A) and group B (Gr.B). This process is repeated until all data is accompanied by a clustering completion indicator.

[0067] Figure 3 In step S106, in the sub-causal error detection model generation unit 25, for each clustered data, an error detection model that learns error detection rules is generated based on the different trends of normal data and error data. This sub-causal error detection model generation unit 25, like the clustering error detection model generation unit 21 described above, can be generated using any machine learning algorithm such as RandomForest and XGBoost based on decision trees, or Neural Networks. However, the clustering error detection model generation unit 21 and the sub-causal error detection model generation unit 25 can use their own different model parameter values.

[0068] Model parameters include hyperparameters that determine model complexity, such as the number of trees and the depth of trees in a tree-structured model. For example, to facilitate data analysis performed by the model analysis unit 22 and clustering using the analysis results, the clustering error detection model generation unit 21 can set parameters related to model complexity, such as the number of trees and the depth of trees, to smaller values ​​compared to the model generated by the subcausal error detection model generation unit 25. On the other hand, in the subcausal error detection model generation unit 25, to improve the reliability of the generated model, model parameter values ​​adjusted to improve error detection accuracy can be set. In summary, the clustering error detection model has a simpler model structure compared to the subcausal error detection model. Such a model structure... Figure 2 The difference between the clustering error detection model and the cause-based error detection model is represented by the number of branches.

[0069] The error detection models generated by the clustering error detection model generation unit 21 and the sub-causal error detection model generation unit 25, namely the clustering error detection model and the sub-causal error detection model, are stored in the model database 14.

[0070] In step S107, in the model connection unit 26, the models generated by the clustering error detection model generation unit 21 and the sub-causal error detection model generation unit 25 are connected based on the clustering process to generate a phylogenetic tree, i.e., a model tree, with the error detection models as nodes. In other words, the model tree expresses the relationships between the error detection models using a tree structure.

[0071] In this way, a learning model is generated to distinguish between normal and erroneous data, and this model is analyzed to quantify the differences between normal and erroneous data. Data is clustered according to errors exhibiting similar differences, treating them as having the same cause to generate a learning model. Thus, even without pre-labeling the causes of errors, it is possible to generate a learning model that detects various types of errors that can occur.

[0072] In addition, by using model trees to visualize the positional relationships between models, and to make it easier for users to analyze and manage models, such as which model is similar to which model.

[0073] Next, the method by which the error cause inference unit 13 infers the error cause using the generated error detection model and its model tree will be explained.

[0074] Figure 5 It means Figure 1 The structural diagram of the error cause inference section 13.

[0075] In the data preprocessing unit 11, the data from the raw data such as device data, formula, and measurement results sent from the database 2 that are intended to infer the cause of the error are organized into a format suitable for machine learning models and output as error cause inference object data 42.

[0076] In the error cause inference unit 13, the probability of candidate error causes is calculated using the error cause inference object data 42 as input. The error cause inference unit 13 includes a model evaluation unit 31, a model selection unit 32, a relevant parameter extraction unit 33, and an error cause probability calculation unit 34.

[0077] Figure 6 The processing steps within these components of the error cause inference section 13 are indicated.

[0078] In step S201, in the model evaluation unit 31, the performance of the sub-causal error detection model stored in the model database 14 is evaluated for the error cause inference object data 42. Performance evaluation is obtained by comparing the output of the error detection model, i.e., the error prediction result, with the actual errors in the input error data. Performance evaluation metrics can include precision, recall, accuracy, F1 score, and AUC. Here, the F1 score is the harmonic mean of precision and recall. AUC is short for Area Under the Curve. Figure 5 The example shown illustrates the use of accuracy as a performance evaluation metric.

[0079] In step S202, the model selection unit 32 selects one or more sub-causal error detection models with high performance evaluation values ​​given by the model evaluation unit 31 as models suitable for the input error data. As a selection method, methods include selecting models with high evaluation values ​​and high rankings, or models with evaluation values ​​above a predetermined threshold.

[0080] In step S203, in the relevant parameter extraction unit 33, the branch of the model tree in which the selected sub-causal error detection model is located is extracted, and the relevant parameters of the error detection model included in that branch are extracted. As a method for extracting these relevant parameters, the feature quantity with high sensitivity and high ranking calculated by the model analysis unit 22 of the model tree generation unit 12 can be selected. Figure 5 For example, the model with the highest accuracy for input error data is "Sub-cause Error Detection Model B". The branch in the model tree where "Sub-cause Error Detection Model B" is located and the models included in that branch are shown in thick lines. For each model included in that branch, parameters such as "matching score" (the degree of consistency between the registered pattern and the captured image) and related parameters such as "contrast", "pattern detection magnification", and "pattern detection coordinates" are calculated as features due to their high sensitivity.

[0081] In step S204, in the error cause probability calculation unit 34, the probability of each error cause (error cause probability) is calculated using the extracted relevant parameters as input, and the result is displayed to the user as a probability via the terminal 4. Figure 5 Among them, "coordinate deviation" has the highest probability, and users are more likely to judge the cause of the error as "coordinate deviation".

[0082] The probability of this error cause can be calculated using neural network models. In learning this error cause probability model, it is necessary to add labels to the combinations of relevant parameters (labeling). However, this labeling is done on a per-cluster data unit basis, which significantly reduces the amount of processing compared to existing methods that label each error as it occurs.

[0083] If two or more sub-causal error detection models are selected in the model selection unit 32, the error cause probability will be synthesized based on the relevant parameters for each model.

[0084] The method is explained below.

[0085] Figure 7 This represents a structure that synthesizes and visualizes the probability of error causes obtained by analyzing multiple error detection models.

[0086] If, for example, two models, “Sub-cause Error Detection Model A” and “Sub-cause Error Detection Model B”, are selected in the model selection unit 32, the relevant parameter extraction unit 33 extracts the branch of the model tree in which each model is located, and uses it as the relevant parameters of the error detection model included in that branch to extract “Relevant Parameter A” and “Relevant Parameter B” respectively.

[0087] The "relevant parameter A" and "relevant parameter B" are input into the error cause probability calculation unit 34, which calculates the error cause probabilities obtained from them respectively. The calculated error cause probabilities are then corrected using the model evaluation value obtained from the model evaluation unit 31. The simplest method is to multiply the error cause probabilities by their respective model accuracies and normalize them. The corrected error cause probabilities are then summed to represent to the user the error cause probability synthesized from multiple models.

[0088] Figure 8 In the example shown, the error cause probabilities synthesized from multiple models will be displayed on the screen.

[0089] This figure is graphically presented in a way that allows for a comparison of the probability of error causes with respect to parameter A with respect to parameter B.

[0090] Furthermore, for the models generated by the clustering error detection model generation unit 21 and the sub-causal error detection model generation unit 25, the models and information about the models are stored in the model database 14 in association, and information about the models selected by the user via the terminal 4 can be displayed.

[0091] Figure 9 This represents a schematic example of what is being displayed.

[0092] As shown in this figure, information about the model, for example, is obtained through the model analysis unit 22 ( Figure 2 The calculated high-sensitivity, top-ranking features (related parameters), and the formula names, training data collection period, model accuracy, etc., included in the training data.

[0093] Furthermore, if multiple versions of the error detection model stored in the model database 14 exist due to updates to the stored data in the database 2, the user can replace the model in the model tree with the model stored in the model database 14 via the terminal 4.

[0094] Figure 10 This is a diagram illustrating the operation.

[0095] As shown in this figure, for example, by dragging a model stored in the model database 14 on the screen and placing it at the tree position of the model to be replaced, the user can interactively update the model in the model tree.

[0096] Furthermore, this embodiment describes the situation of inferring the cause of errors in a semiconductor inspection device, but the content of this disclosure is not limited to this. It can also be applied to generating parameters regarding the operation of a specified device and models and model trees to determine whether errors occur when these parameters are used. That is, the content of this disclosure can also be applied to devices other than semiconductor inspection devices.

[0097] Explanation of reference numerals in the attached figures

[0098] 1: Semiconductor inspection device; 2: Database; 3: Error cause inference device; 4: Terminal; 11: Data preprocessing unit; 12: Model tree generation unit; 13: Error cause inference unit; 14: Model database; 15: Model tree; 21: Error detection model generation unit for clustering; 22: Model analysis unit; 23: Data clustering unit; 24: Clustering completion judgment unit; 25: Error detection model generation unit for different causes; 26: Model connection unit; 31: Model evaluation unit; 32: Model selection unit; 33: Relevant parameter extraction unit; 34: Error cause probability calculation unit; 41: Training data; 42: Error cause inference object data.

Claims

1. An error cause inference apparatus for inferring the cause of an error in an inspection operation performed by a semiconductor inspection apparatus, comprising: The data preprocessing unit uses the data of the processing object to generate training data in a format suitable for input into a machine learning model; and The model tree generation unit uses the training data as input to generate a learning model, i.e., an error detection model, for detecting errors, and generates a model tree that uses a tree structure with the error detection models as nodes to express the relationships between the error detection models. The processed object data includes the setting parameters and measurement results of the inspection object of the semiconductor inspection device. The model tree generation unit includes: The clustering error detection model generation unit takes the data in the training data that has not been clustered as input, learns error detection rules based on the different trends of normal data and error data, and generates a clustering error detection model. The model analysis unit calculates a sensitivity value that represents the degree to which the features input as the training data contribute to the output of the error detection model. A data clustering unit performs data clustering based on the values ​​of the feature quantity and the sensitivity. The clustering completion determination unit determines whether the clustering of the clustered data is complete. The sub-causal error detection model generation unit takes the clustered data as input, learns error detection rules based on the different trends of normal data and error data, and generates a sub-causal error detection model; and The model connection unit connects the clustering error detection model and the sub-causal error detection model based on the data clustering process, thereby generating the model tree composed of the phylogenetic tree with the clustering error detection model and the sub-causal error detection model as nodes.

2. The error cause inference apparatus as described in claim 1, wherein, The clustering error detection model has a simpler structure compared to the sub-causal error detection model.

3. The error cause inference apparatus as described in claim 1, wherein, It also includes an error cause inference section. The error cause inference section includes: The model evaluation unit evaluates the performance of the generated sub-cause error detection model by taking the error cause inference object data as input. The error cause inference object data is data generated by the data preprocessing unit using the data to infer the error cause from the processing object data, and has a format suitable for input into the machine learning model. The model selection unit selects one or more sub-causal error detection models that have been evaluated as having high performance scores by the model evaluation unit; and The relevant parameter extraction unit extracts the branch in the model tree where the selected sub-cause error detection model is located, and extracts the relevant parameters of the error detection model included in that branch.

4. The error cause inference apparatus as described in claim 3, wherein, The error cause inference unit also includes: The error cause probability calculation unit uses the relevant parameters extracted by the relevant parameter extraction unit as input to calculate the probability of candidate error causes.

5. The error cause inference apparatus as described in claim 4, wherein, When the model selection unit selects multiple sub-cause error detection models, the probabilities obtained based on each sub-cause error detection model are corrected using the model evaluation value obtained by the model evaluation unit.

6. The error cause inference apparatus as described in claim 5, wherein, It also includes a model database that stores the clustering error detection model and the sub-causal error detection model in association with information about these models.

7. The error cause inference apparatus as described in claim 6, wherein the apparatus is configured to: When multiple versions of the clustering error detection model and the sub-causal error detection model stored in the model database exist, the user can use the terminal to replace the model in the model tree with other models stored in the model database.

8. The error cause inference apparatus as described in claim 5, wherein, The probabilities before and after the correction are displayed on the terminal.

9. The error cause inference apparatus as described in claim 6, wherein the apparatus is configured to: Information associated with the selected error detection model is displayed by the user selecting an error detection model from the model tree via a terminal.

10. A method for inferring the cause of an error, comprising inferring the cause of an error in an inspection operation performed by a semiconductor inspection device, including: The data preprocessing step includes generating training data with a format suitable for input into a machine learning model using the data of the object being processed; and The model tree generation step includes generating a learning model, i.e., an error detection model, using the training data as input, and generating a model tree that uses a tree structure with the error detection models as nodes to express the relationships between the error detection models. The processed object data includes the setting parameters and measurement results of the inspection object of the semiconductor inspection device. The model tree generation steps include: The step of generating a clustering error detection model includes taking the data in the training data that has not been clustered as input, learning error detection rules based on the different trends of normal data and error data, and generating a clustering error detection model. The model analysis step involves calculating a sensitivity value that represents the degree to which a feature quantity, as input to the training data, contributes to the output of the error detection model. A data clustering step, wherein data is clustered based on the values ​​of the feature quantity and the sensitivity; The clustering completion judgment step involves determining whether the clustering of the clustered data is complete. The step of generating a sub-causal error detection model includes: using the clustered data as input, learning error detection rules based on the different trends of normal data and error data to generate a sub-causal error detection model; and The model connection step involves connecting the clustering error detection model and the sub-causal error detection model based on data clustering to generate the model tree, which is composed of a phylogenetic tree with the clustering error detection model and the sub-causal error detection model as nodes.

Images