Abnormality cause analysis system and abnormality cause analysis method

Abstract

The objective of the present invention is to efficiently collect data to be used for analyzing the cause of an abnormality of a process device and to improve the accuracy in the analysis of the cause of the abnormality. An abnormality cause analysis system according to the present invention classifies a measurement result into a normal first group and an abnormal second group on the basis of a first standard, also classifies the first group into a normal third group and a fourth group having an abnormality sign on the basis of a second standard, and performs machine learning for predicting the cause of the abnormality by using the second group and the fourth group (see fig. 9).

Description

Anomaly cause analysis system and anomaly cause analysis method

[0001] The present invention relates to a technique for analyzing the cause of an abnormality in a process device.

[0002] Semiconductor processing equipment used in semiconductor manufacturing lines performs various processes on semiconductor wafers during the semiconductor manufacturing process. When an abnormality occurs in a processing equipment, it is important to quickly identify the cause of the abnormality and take measures to improve the operating rate of the manufacturing process.

[0003] In order to reduce the time required for manual anomaly cause analysis of process equipment, Patent Document 1 (JP-A-2005-102666) discloses a two-stage classification analysis that primarily uses Bayesian analysis using a neural network to analyze the cause of anomalies. To analyze the cause of anomalies, Patent Document 1 first acquires defect measurement data, such as defect images, component data, and wafer maps, and then performs defect classification based on these data in the first stage. Furthermore, related data, such as material information on the wafer and past defect data for the process equipment, are added to the defect classification results, and a second stage of classification is performed to identify the cause using these data.

[0004] US11263737B2

[0005] In reality, it is often difficult to continuously collect the data used for cause analysis in Patent Document 1, considering storage capacity and network load. Furthermore, obtaining the defect images and component data used in Patent Document 1 by actual measurement requires a longer measurement time than wafer maps, and therefore is not necessarily performed continuously. In other words, when an abnormality occurs in the process equipment, these data must be acquired again. Even if it were possible to continuously collect the data cited in Patent Document 1, the amount of data would be enormous, making it difficult to identify the circumstances under which each piece of data was acquired, and there is a possibility that the data truly necessary for analyzing the cause of the abnormality would be buried.

[0006] As such, the data used in Patent Document 1 to analyze the cause of an anomaly is data collected only when an anomaly occurs, or data that is not practical to collect reliably at all times. Patent Document 1 suggests that there is room for further study on the extent to which the accuracy of the cause analysis can be ensured using such data. Furthermore, semiconductor devices are not only becoming more miniaturized, but also undergoing significant changes from moment to moment in terms of materials, design, and configuration, which result in a wide variety of anomalies and the emergence of new anomaly modes. Learning about such a wide variety of complex anomalies on a semiconductor production line using only data collected only when an anomaly occurs, as in Patent Document 1, may result in insufficient accuracy in the cause analysis.

[0007] The present invention has been made in view of the above-mentioned problems, and an object of the present invention is to efficiently collect data used for analyzing the cause of an abnormality in a process device and to improve the accuracy of the analysis of the cause of the abnormality.

[0008] The anomaly cause analysis system according to the present invention classifies measurement results into a first normal group and a second abnormal group based on a first criterion, further classifies the first group into a third normal group and a fourth group having signs of an anomaly based on a second criterion, and performs machine learning to predict the cause of an anomaly using the second and fourth groups.

[0009] The anomaly cause analysis system according to the present invention can efficiently collect data used to analyze the cause of an anomaly in a process device, and can also improve the accuracy of the anomaly cause analysis. Problems, configurations, effects, and the like other than those described above will become clear from the following description of the embodiment.

[0010] 1 is a configuration diagram of an anomaly cause analysis system 10 according to a first embodiment. FIG. 1 is an example of a user interface displayed by a user terminal 108. FIG. 2 is an example of a user interface displayed by a user terminal 108. FIG. 3 is a graph illustrating the frequency and timing of viewing the history of anomaly cases on the user terminal 108. FIG. 4 shows a general procedure for detecting an anomaly in a process device in a semiconductor manufacturing line. FIG. 5 shows another procedure for detecting an anomaly in a process device in a semiconductor manufacturing line. FIG. 6 is a flowchart illustrating a procedure for dealing with an anomaly in a process device when an anomaly in a process device is detected. FIG. 7 is an example of a user interface for a user to select data to be sent to an equipment supplier. FIG. 8 is a schematic diagram illustrating how data is sent and received between a user and an equipment supplier. FIG. 9 shows an example configuration of a process device. FIG. 10 shows another example configuration of a measurement device. FIG. 11 is a diagram illustrating a procedure for classifying measurement results by the anomaly cause analysis system 10. FIG. 12 is a schematic diagram illustrating the flow of each data in the anomaly cause analysis system 10. FIG. 13 is a graph illustrating an example of a second criterion. FIG. 14 is a diagram illustrating another example of the second criterion. FIG. 15 is a flowchart illustrating a procedure for classifying data used for learning and analysis by the anomaly cause analysis system 10 according to a second embodiment. FIG. 16 shows an example of subdividing the second and fourth categories. 10 shows another example of subdividing the second and fourth classes. 11 shows another example of subdividing the second and fourth classes. 12 shows another example of subdividing the second and fourth classes. 13 shows an example of a haze map.

[0011] <First Embodiment: System Configuration> FIG. 1 is a configuration diagram of an anomaly cause analysis system 10 according to a first embodiment of the present invention. The anomaly cause analysis system 10 analyzes the causes of anomalies in process equipment in a semiconductor manufacturing process. The anomaly cause analysis system 10 includes process equipment 100-102, measuring equipment 103-105, an AI engine 106 (learning device), a computing device 106a, a database 107 (storage device), a user terminal 108, and an engineer terminal 109. These components are interconnected via a network. The anomaly cause analysis system 10 can also connect to and communicate with other data management systems via the network. The anomaly cause analysis system 10 may also communicate with other devices and systems via the network.

[0012] The process tools 100 to 102 are tools that process semiconductor wafers in the semiconductor manufacturing process. Each process tool may perform a different process, or some of the processes among the process tools may be the same. The process tools 100 to 102 are, for example, an etching tool, a film forming tool, a lithography tool, a CMP (Chemical Mechanical Polishing) tool, etc. The supplier of each process tool may be the same, or the tools may be supplied by different suppliers.

[0013] The measuring tools 103 to 105 are tools that measure samples (semiconductor wafers in this example) processed by the process tools 100 to 102. Each measuring tool may perform a different measurement, or some of the measurements may be the same. The measuring devices 103 to 105 are, for example, an OCD (Optical Critical Dimension) device, an ellipsometry device, a SEM (Scanning Electron Microscope), a CD-SEM (Critical Dimension SEM), an electron beam defect inspection device, a TEM (Transmission Electron Microscope), a SAXS (Small Angle X-ray Scattering) device, an AFM (Atomic Force Microscope), a mass measuring instrument, an XPS (X-ray Photoelectron Spectroscopy) device, an XRD (X-ray Diffraction equipment, etc. When the process equipment itself is equipped with some kind of measuring equipment, data obtained from the measuring equipment is considered to have been acquired from the process equipment.

[0014] The user terminal 108 and the engineer terminal 109 are electronic terminals such as personal computers and tablets. The user terminal 108 can display the status and processing conditions of the devices connected to the anomaly cause analysis system 10. When an abnormality occurs in any of the devices 100 to 105, the user terminal 108 can also send information about the status and information necessary for investigating the cause to the device supplier. The engineer terminal 109 has similar functions. The engineer terminal 109 also has a user interface for displaying data used to analyze the cause of an equipment abnormality and implement countermeasures, and for inputting the details of the countermeasures.

[0015] The AI ​​engine 106 is executed by the computing device 106a. The AI ​​engine 106 performs machine learning to estimate the cause of an abnormality in the process equipment 100-102, and estimates the cause of the abnormality based on the results of the machine learning. Details of the data used for learning will be described later. The database 107 can be configured by a storage device that stores each piece of data acquired from the process equipment 100-102 and the measuring devices 103-105. Details of the data will be described later.

[0016] 2A is an example of a user interface displayed by the user terminal 108. This user interface displays a list of abnormality cases that have occurred in the process equipment 100-102 in the past. On this screen, the user can view the date and time the abnormality occurred, the equipment in which the abnormality occurred, the cause of the abnormality, the countermeasures taken to address the abnormality, and so on. The data can also be sorted or grouped and displayed in chronological order of the abnormality occurrence, abnormalities by production line, and so on.

[0017] Fig. 2B is an example of a user interface displayed by the user terminal 108. When the user selects any of the abnormal cases in Fig. 2A, the user terminal 108 displays the selected abnormal case on the screen of Fig. 2B. These data are stored in the database 107, and the user terminal 108 retrieves and displays these data from the database 107.

[0018] The items to be displayed may be predetermined items, or may be variable items, such as displaying items in order of increasing importance. The importance may be determined by the computing device 106a based on the browsing history of each terminal, for example. The display items may also be set by the user. The user terminal 108 retrieves data of the items specified by the user from the database 107 and displays them.

[0019] 2C is a graph illustrating an example of the frequency and timing of viewing the history of abnormal cases on the user terminal 108. It is often difficult to constantly store data on all past abnormal cases in the database 107 due to capacity constraints. Therefore, the abnormality cause analysis system 10 determines the data to be stored in the database 107 and the storage period according to the following procedure.

[0020] The computing device 106a determines the data to be stored in the database 107 and the retention period based on the abnormal cases viewed on the user terminal 108 and their viewing frequency. For example, for abnormal cases that occurred during a period 201 (from the present to approximately six months ago) with a high viewing frequency, data describing the details of all abnormal cases is stored. For abnormal cases that occurred during a period 202 (six months ago to one year ago) with a significant drop in viewing frequency, for example, fewer items are stored than during the period 201, or only a summary of the data is stored. These methods may be used in combination. For abnormal cases that occurred during a period 203 (more than one year ago) with an even lower viewing frequency, for example, only the headline items are stored.

[0021] In this way, by gradually reducing the data items and their contents to be stored in the database 107 based on the frequency of data browsing, it is possible to achieve both the total amount of data stored in the database 107 and user convenience. When deleting data stored in the database 107 according to the above procedure, the computing device 106a can also inquire of the user on the user terminal 108 whether or not to delete the data.

[0022] The priority of data stored on the database 107 may be determined by the computing device 106a comprehensively evaluating not only the browsing history as shown in Figure 2C, but also the following information, for example: (a) information input from the engineer terminal 109, (b) data transmission and reception history between the user terminal 108 and the engineer terminal 109, and (c) data reference history on the engineer terminal 109.

[0023] FIG. 3A shows a typical procedure for detecting an abnormality in a process tool in a semiconductor manufacturing line. An abnormality in a process tool may first be detected by an alarm function of the process tool itself. After processing wafers flowing through the manufacturing line in the process tool, the wafers are measured by a measurement tool. If the measurement results do not meet the desired state or standard, an abnormality may be detected. Because a manufacturing line involves hundreds of processes, it is difficult to perform measurements using a measurement tool for each process. Therefore, measurements are often performed using a measurement tool after multiple processes have been performed consecutively. The process tool in which the abnormality occurred is identified by tracing the processing history using a system owned by each user.

[0024] 3B shows another procedure for detecting abnormalities in process equipment on a semiconductor manufacturing line. The daily status of the process equipment is checked by passing a quality check wafer (e.g., a wafer without a pattern formed thereon) through the process equipment, comparing the results before and after processing, and checking whether the processing of the process equipment is being performed normally based on whether the difference satisfies a standard. The standard here refers to indicators output by a measuring device measuring the object to be measured, such as the actual shape and dimensions (including depth and thickness) of the circuit pattern, the surface condition (e.g., roughness), the number of defects, defect images, components, electrical characteristics, etc.

[0025] FIG. 4 is a flowchart illustrating the procedure for dealing with a process tool anomaly. When an anomaly is detected based on an alarm or measurement results from the process tool (S1), the suspected tool or chamber is identified. The anomaly is checked for false positives (S2). If false positives are detected, the process is reconfirmed and processing resumed. If the anomaly is truly true, the suspect tool or chamber is shut down to prevent further wafer processing (S3) to prevent further anomalies from occurring. The user investigates the cause of the anomaly and performs recovery work (S4: No). However, in some cases, the anomaly is complex and difficult to resolve by the user alone (S4: Yes). In such cases, the user requests the tool supplier to investigate the cause and perform recovery work (S5). The user may send several pieces of data to the tool supplier to inform them of the anomaly. This data may include, for example, data obtainable from the suspected process tool, the measurement results used to detect the anomaly, and image data of the actual tool. As will be described later, the anomaly cause analysis system 10 can improve the efficiency of the work of sending initial data when a user notifies an equipment supplier of an anomaly.

[0026] When the equipment supplier is notified of the abnormality, the equipment supplier's engineer checks the data and considers the cause and countermeasures (S6). At this time, if work needs to be performed on the actual equipment, the equipment supplier's engineer heads to the site. With the development of AR / VR technology, it is sometimes possible to recreate the state of the actual equipment without going to the site and perform countermeasures remotely from the equipment supplier's base. The equipment supplier's engineer estimates several possible causes of the abnormality and then implements countermeasures to resolve the problem on the process equipment (S7). If the problem is resolved (S8: Yes), the engineer uses a measuring device or the like to confirm that the problem has truly been resolved (S10). If the problem is not resolved (S8: No), the engineer checks for other possible causes and implements countermeasures (S9). This process is repeated until it is confirmed that the abnormality has been resolved. If it is confirmed that the abnormality has been resolved, the process equipment is restored (S11).

[0027] As shown in FIG. 4, the investigation of the cause of an abnormality in equipment and the associated countermeasures and recovery work have traditionally been performed primarily by user engineers and equipment supplier engineers (i.e., humans). However, human work is prone to variability due to differences in experience and misunderstandings between the user and equipment supplier. The abnormality cause analysis system 10 can be used to reduce such variability and improve the efficiency of the investigation of the cause of an abnormality in process equipment and the implementation of countermeasures. The AI ​​engine 106 improves the efficiency of the investigation of the cause of an abnormality in process equipment. It also efficiently performs advance learning to infer the cause of an abnormality. The learning method will be described later.

[0028] In the flowchart of FIG. 4, the efficiency improvement achieved by the anomaly cause analysis system 10 in the step of collecting data to notify the supplier of the suspected anomaly will be described in detail.

[0029] The user decides which data to attach to the equipment supplier. The data required to communicate an abnormality can vary widely, including data related to the process equipment suspected of being abnormal and measurement data from the measuring equipment that detected the abnormality. Some data can be obtained without the user having to enter the clean room where the equipment is located, while other data must be collected directly from the equipment itself. Furthermore, some measurement data from measuring equipment is routinely acquired, while others are acquired only in special circumstances, such as when an abnormality occurs. The process of compiling a set of data to communicate an abnormality varies in time and ease of acquisition for each type of data. Therefore, data acquisition can be a time-consuming task for users. Among the various data representing an abnormality, which data is most important for determining the cause of the abnormality varies based on the experience of both the user and the equipment supplier's engineers. Furthermore, there can be differences in perception between the user and the supplier. When the equipment supplier receives a request from the user and the accompanying data representing the abnormality, they first use the data to confirm the situation. As mentioned above, differences in the time it takes to collect different types of data, and variations in the perception of "necessary data" based on the experience of the engineers sending and receiving the data, affect how accurately the cause of the abnormality can be estimated, which can have a significant impact on the total time required to investigate the cause of the abnormality and take countermeasures.

[0030] FIG. 5 shows an example of a user interface through which a user selects data to be transmitted to an equipment supplier. This user interface (UI) is provided on a user terminal 108. The user selects, on the UI, the process equipment from which the user wishes to acquire data (the equipment suspected of having an abnormality). When making the selection, the UI displays names and other information for identifying the equipment, such as the wafer processing history, the model and serial number of the process equipment, etc. The user can make a selection from this display. Next, the user selects the data to be collected from the data held by the selected equipment. This selection is also made by referring to the display of the processing history, model, etc. of the measurement equipment. These may be selected from options on the UI, or may be manually entered by the user.

[0031] As described in FIG. 1, the user terminal 108 is connected to the database 107, so the user can collect necessary data for a specific process tool or measurement tool via the user terminal 108 without entering the clean room. After selecting data on the UI, the user selects whether or not to send the data to the equipment supplier. A separate approver may be set for each user's selection, and the data may be sent to the equipment supplier only if the approver approves.

[0032] In the UI of FIG. 5, data categories that are particularly necessary for anomaly cause analysis are displayed as selection items. For example, when there is a selection item simply called "Log," even though the word "Log" is used, there are various Logs within a single device. In some cases, the user may not be able to determine which Log is particularly important. In such cases, the options within the selection item can be displayed in advance in descending order of importance. It is also possible to set up the display in advance of Log data that the user wants to refer to frequently. It is also possible for the equipment supplier to set up the display of Log data that is expected to be referred to particularly frequently depending on the process in which the equipment is used.

[0033] FIG. 6 is a schematic diagram showing data transmission and reception between a user and an equipment supplier. When a user requests a process equipment manufacturer to take measures against an abnormality, the user selects and sends data indicating the abnormal condition. The equipment supplier acquires the data and inputs which of the acquired data was referenced and its priority. Experience values, such as whether the inputter is an expert or not, are pre-registered. The importance (display order) of the data in FIG. 5 can be (a) arbitrarily specified by the user or equipment supplier, or (b) estimated by the AI ​​engine 106 (the computing device 106a) based on information entered by an equipment supplier engineer via the engineer terminal 109 regarding the history of data transmission and reception between the user and the equipment supplier in the event of an abnormality and which data actually resolved the abnormality.

[0034] The data to be sent to the supplier may be selected by other means. For example, to identify the device or data category for which the user wants to quote data, the user enters keywords into the user terminal 108 and performs a search. The computing device 106a displays options for the device and data category according to the search criteria and importance.

[0035] The AI ​​engine 106 learns the history of data transmission and reception between the user and the equipment supplier in the event of an anomaly, and information entered on the engineer terminal 109 regarding which data stored in the database 107 was used to resolve the anomaly. Based on this learning, the AI ​​engine 106 identifies data important for investigating the cause of the anomaly as "recommended collection data" and assists the user in selecting that data from the options described above. Furthermore, when the user sends data to the equipment supplier, if there is any missing data in the "recommended collection data" that is highly important, a pop-up message can be displayed on the user terminal 108 to notify the user. The user and the equipment supplier may decide in advance which data items are necessary in the event of an anomaly, and designate them as "recommended collection data." The AI ​​engine 106 (or the computing device 106a) may compare this designation and display a pop-up message if there is a deficiency at the time of actual transmission. Once the device from which data is to be collected and the data to be collected are designated, the user terminal 108 collects the data from the designated device and database 107.

[0036] 6, when data is transmitted and received between the user terminal 108 and the engineer terminal 109, the data may be encrypted. The encryption may be performed by the anomaly cause analysis system 10 (the computing device 106a). Any known technology (e.g., public key encryption) may be used as the encryption method.

[0037] <First Embodiment: Regarding Learning> Next, we will explain how the AI ​​engine 106 performs learning so that it can execute or support anomaly cause analysis tasks that have previously been performed by humans. In anomaly cause analysis, engineers review data such as images / videos showing the actual process equipment and its condition, and results of measurements using a measurement device on wafers processed in the process equipment. The engineers analyze the cause of the anomaly based on this data. Therefore, anomaly cause analysis first requires data indicating the abnormal state when the anomaly is detected. Furthermore, there are often some precursors to an anomaly occurring. Therefore, the anomaly cause analysis system 10 uses the AI ​​engine 106 to learn not only data indicating the state at the time of the anomaly, but also data indicating precursors to an anomaly before it occurs.

[0038] Typically, when and in what type of data a sign of an abnormality appears varies from case to case, making it difficult to uniformly determine which data from which equipment should be acquired and when it can become "data indicating an abnormality" (there are countless options). For example, Patent Document 1 (JP-A-2005-102626) proposes that pre-trained AI shortens the time it takes to identify the cause of an abnormality using process equipment recipes, design data, maintenance history, defect images, composition (EDX), and defect maps. However, constantly collecting data from process equipment and measurement equipment and continuously storing all of it is unrealistic given data capacity constraints. For example, increasing only the types of data used for learning and data analysis shortens the period for accumulating that data due to the limited data capacity available. In particular, defect images and composition data have relatively large data sizes per data point, making it impossible to continuously store all of them due to capacity constraints. Furthermore, more than determining the types of data to be accumulated, efficient learning requires sorting and determining which types of data should be selected (and accumulated) in what situations.

[0039] The following describes how the anomaly cause analysis system 10 efficiently collects data from the process tools 100 to 102 and the measuring tools 103 to 105, and what triggers the data to be sorted. First, an example of the configuration of each tool will be described, and then the procedure for collecting data will be described.

[0040] FIG. 7 shows an example of the configuration of a process apparatus. This process apparatus is a plasma processing apparatus. Examples of plasma processing apparatuses include etching apparatuses and film deposition apparatuses. The plasma processing apparatus includes a vacuum processing chamber 701, a lower electrode (sample stage) 703 provided within the vacuum processing chamber 701, a microwave-transmitting window 704 made of quartz or the like, a waveguide 705 provided above the window, a magnetron (plasma generator) 706, a magnetron drive power supply 713, a solenoid coil 707 provided around the vacuum processing chamber 701, an electrostatic adsorption power supply 708 connected to the lower electrode 703, a substrate bias power supply 709, and a power control unit 714 for controlling the power supplied by the magnetron drive power supply 713 and the substrate bias power supply 709. The lower electrode 703 has a wafer mounting surface for holding a wafer 702. The magnetron drive power supply 713 supplies plasma generation power to the magnetron 706, and the substrate bias power supply 709 supplies substrate bias power to the lower electrode 703. Furthermore, a wafer loading port 710 is provided for loading or unloading a wafer 702 into or from the vacuum processing chamber 701, and a gas supply port 711 is provided for supplying gas to the vacuum processing chamber 701.

[0041] The operation of the plasma processing apparatus configured as described above will now be described. After the pressure inside the vacuum processing chamber 701 is reduced, an etching gas is supplied into the vacuum processing chamber 701 through the gas supply port 711 and adjusted to the desired pressure. Next, a DC voltage of several hundred volts is applied from the electrostatic adsorption power supply 708, thereby electrostatically adsorbing the wafer 702 to the mounting surface above the lower electrode 703. Thereafter, when plasma generation power is supplied from the magnetron driving power supply 713 (when ON), microwaves with a frequency of 2.45 GHz are generated from the magnetron 706. These microwaves are propagated into the vacuum processing chamber 701 through the waveguide 705. When plasma generation power is not supplied (when OFF), the magnetron 706 stops oscillating the microwaves. A magnetic field is generated within the vacuum processing chamber 701 by the solenoid coil 707, and the interaction between this magnetic field and the generated microwaves generates a high-density plasma 112 within the vacuum processing chamber 701. After the plasma 112 is generated, high frequency power is supplied from the substrate bias power supply 709 to the lower electrode 703, and the energy of the ions in the plasma incident on the wafer is controlled, thereby enabling etching of the wafer 702.

[0042] The abnormality cause analysis system 10 can acquire, for example, the following data from the plasma processing apparatus: These data represent the state of the plasma processing apparatus while it is performing processing.

[0043] (1) Sensor Data The controller that controls the operation of the plasma process equipment can create and record the following operation logs at the start and end of an operation command, etc.: Process log: A log related to wafer processing within the vacuum processing chamber 701, e.g., the start / end of processing Transfer log: A log of robot operations related to wafer transport, e.g., the start / end of load / unload operations Function log: A log related to hardware operations, e.g., the start / end of vent / exhaust Lot event log: A log related to lot operations, e.g., the installation / removal of a FOUP, the start / end of a process job (2) In-situ Process Monitor Data For example, the detection values ​​of sensors that monitor the etching process can be acquired via the equipment controller. For example, the detection values ​​of sensors that detect changes in the plasma state or sensors that measure the wafer film thickness can be acquired. - End Point Detection (etching end point monitor): plasma chemistry changes - Film Thickness Monitor (film thickness / depth monitor): wafer thin film, optical interference measurement (3) Data Collection System The equipment controller can record the following processing conditions related to plasma processing: - Equipment data: process condition parameters are collected in 0.1 second cycles (microwaves, bias power, pressure, gas flow rate, chamber temperature, etc.) - Light emission data (4) Equipment specific data: component placement, material information, etc. can be stored in advance in database 107. Database 107 may be updated when equipment is upgraded, etc. (5) Maintenance history of process equipment (6) Processing or cleaning conditions of process equipment

[0044] Fig. 8A shows an example of the configuration of a measurement apparatus. This measurement apparatus is configured as an optical defect inspection apparatus 103A. The measurement object is a mask, a wafer, etc., and the wafer may be patterned or unpatterned. Fig. 8 shows an example of the configuration for measuring defects in a patterned wafer. The optical defect inspection apparatus 103A includes a stage, a light source 120, an optical lens 130, a camera (sensor) 140, an image acquisition unit 150, and a calculation unit 160.

[0045] Chips (dies) are arranged on the wafer 200 in the XY directions. A stage moves the wafer 200 at least in the planar direction (XY directions). A light source 120 irradiates the wafer 200 with light 121 from above or obliquely above. When the light 121 strikes the wafer 200, reflected light 122 and scattered light 123 (both signal light) are generated from the wafer 200. An optical lens 130 directs the reflected light 122 or scattered light 123 toward the imaging surface of a camera (sensor) 140. The camera (sensor) 140 captures the reflected light 122 or scattered light 123. An inspection device that captures and inspects the reflected light 122 is called a bright-field inspection device, and an inspection device that captures and inspects the scattered light 123 is called a dark-field inspection device. If the wafer 200 has a repeating pattern with a fixed period, a spatial filter that cuts light corresponding to the period of the repeating pattern may be provided, and the optical lens 130 may be placed before or after the spatial filter. The image acquisition unit 150 acquires an image of the wafer 200 using the imaging signal acquired by the camera (sensor) 140. This image is compared with a database, a reference image, an adjacent die image, etc. on a chip (die) basis to distinguish between light from patterns and light from defects, thereby detecting defects.

[0046] 8B shows another example of the configuration of the measurement device. This measurement device is configured as an optical defect inspection device 103B and measures defects on an unpatterned wafer. A sample W is placed on a rotary stage ST, and light is irradiated onto the sample A from a light source A. Signal light generated from the sample W is detected by detectors (three detectors B1 to B3 are shown here) installed at different orientations and / or directions relative to the sample W. The detectors output detection signals representing the results of the detected light. A signal processing device D analyzes the detection signals to detect defects on the sample W.

[0047] 9 is a diagram illustrating the procedure for classifying measurement results by the anomaly cause analysis system 10. Each step can be performed by, for example, the computing device 106a. This procedure collects data suitable for use in analyzing the cause of an anomaly by classifying the measurement results into abnormal values ​​and abnormality predictor values.

[0048] The measuring tools 103-105 measure the wafers processed by the process tools 100-102. The arithmetic device 106a determines whether the measurement results satisfy a first criterion. Measurement results that satisfy the first criterion are classified into a first group (normal), and measurement results that do not satisfy the first criterion are classified into a second group (abnormal). The arithmetic device 106a further determines whether the measurement results of the first group satisfy a second criterion. Measurement results that satisfy the second criterion are classified into a third group (normal), and measurement results that do not satisfy the second criterion are classified into a fourth group (signs of abnormality).

[0049] The first criterion may be a management standard determined by the user, or may be a standard independently established by the equipment supplier for that equipment. There are dedicated measuring devices used in semiconductor manufacturing lines for each item to be measured. If the measurement item is a defect, a defect inspection device is used, and if it is a critical dimension, a CD-SEM or OCD is used. For example, if the measurement device is an optical defect inspection device, the measurement item is a defect, and the measurement results are mainly the number of defects and their coordinates. In this case, the first criterion is that the number of defects is less than a predetermined number. If this criterion is not met (if the number of defects is equal to or greater than the predetermined number), the product is classified into the second group.

[0050] When the AI ​​engine 106 investigates the cause of an abnormality in a process tool, the main target of learning and analysis is data from the abnormality. Continuously collecting data output by each tool results in a massive amount of data, making it difficult to select which collected data to use for learning and which to use in data analysis to investigate the cause of the abnormality. When a process tool supplier investigates the cause of an abnormality, the user notifies the supplier of the abnormality (S5 in Figure 4 ) and the supplier analyzes the data using time-series data from the process tool and its associated equipment. When an abnormality is detected by a measurement tool separate from the process tool, measurement data on past wafers processed by the process tool is often unavailable. Even if such data is available, it is often unclear which measurement data is useful for investigating the cause of the abnormality, or it is impossible to extract it. When all data from both the process tool and the measurement tool is available, the data becomes so vast that it is difficult to sort out the data that is most effective for investigating the cause of the abnormality. Furthermore, when only data from a specific process tool or a specific measurement tool is available, the analysis accuracy may be insufficient. Therefore, in the present invention, the measurement results output by the measuring device are not simply sorted into normal and abnormal, but measurement results that indicate signs of abnormality are extracted from the measurement results that have been sorted as normal.

[0051] FIG. 10 is a schematic diagram showing the flow of data in the anomaly cause analysis system 10. The AI ​​engine 106 runs on a server (i.e., on a computing device 106a provided by the server). When a user or a process equipment supplier analyzes the cause of an anomaly in a process equipment, the user or the process equipment supplier acquires the status data of the process equipment at that time (e.g., the data listed in FIG. 7 ) using measurement results classified by a measuring device as data indicating an anomaly or anomaly precursor as a marker / trigger (FIG. 10(1)). The computing device 106a requests the process equipment that processed the wafer to transmit data indicating the status at the time of processing and acquires that data (FIG. 10(2)). The computing device 106a also references the specific data of the process equipment stored in the database 107 (FIGS. 10(4) and 10(5)). If data related to the measurement results exists (e.g., defect images obtained by a review SEM, component analysis data obtained by EDX, etc.), it may also be collected (FIG. 10(3)). The fourth group of data is collected in the same manner as above. The measurement results may be linked to which process in which process equipment by reference to, for example, lot trace data.

[0052] The computing device 106a performs learning and analysis by specializing in data classified into a data group (second or fourth group) indicating an abnormality or an abnormality precursor among the measurement results. This allows prioritization of which data to target in an abnormality cause investigation analysis and the preliminary learning for that purpose. Compared to cases where an abnormality cause is investigated using only process equipment data or only process equipment data and measurement results at the time of abnormality detection, it is possible to narrow down to data indicating an abnormality or an abnormality precursor, link data from both process equipment and measurement equipment, and simultaneously reference these, thereby improving the accuracy and efficiency of both learning and data analysis.

[0053] FIG. 11 is a graph illustrating an example of the second criterion. When the second criterion is based on the amount of change in measurement values ​​over time, if the cumulative amount of change in measurement values ​​over time exceeds a threshold, the measurement value is extracted as data indicating a sign of an abnormality. For example, when the measurement device is an optical defect inspection device, the measurement item is the number of defects. When the number of defects continues to increase over a certain period of time, or when the rate of increase in the number of defects over a certain period of time is generally high despite some fluctuations, the measurement results (i.e., the number of defects) are classified into the fourth group. When the measurement device is a CD measurement device or a film thickness measurement device, the measurement values ​​are CD values ​​or film thickness values, and the classification criteria are similar.

[0054] FIG. 12 shows another example of the second criterion. Here, the measurement device is assumed to be an optical defect inspection device. The optical defect inspection device can output detected defects as a "defect map." A defect map is an image in which defect positions on a wafer are plotted on a planar image of the wafer, showing the in-plane distribution of defects. The defect distribution is useful for understanding the correlation between defects and the distribution of elements that indicate the defect occurrence location on the process equipment side, and the processing conditions during execution, such as the temperature, gas, and plasma density in the processing space. When a difference between a normal defect map and a measured defect map is greater than or equal to a threshold value, or when the measured defect map is close to a predetermined characteristic distribution (the difference between the characteristic distribution and the measured distribution is small), the measurement result is considered to be in the fourth group. The difference can be expressed, for example, by the difference in geometric features of the defect map.

[0055] An example of a normal defect map is when defects are distributed evenly across the wafer and the number of defects is less than a predetermined number. In this case, cases that do not satisfy the second criterion include, for example, the following: (a) when defects are locally concentrated in a specific area (including concentration on the periphery), (b) when the defects (scratches) have a certain level of continuity and length, (c) when the defects appear with directionality or regularity (for example, concentric circles), and (d) when the defect map is predicted to appear in an abnormal state based on the user's empirical defect distribution that is likely to appear in an abnormal state of the process equipment or design information (component layout and gas / heat control mechanism) of the process equipment.

[0056] Summary of First Embodiment The anomaly cause analysis system 10 according to the first embodiment performs a reverse lookup of the state data of the process equipment that processed the wafer for measurement results classified as abnormal (Group 2) based on the first criterion and measurement results classified as indicating anomaly signs (Group 4) based on the second criterion, and links the measurement results to the state data (by assigning a classification or storing these data as a set). This improves learning efficiency and analysis accuracy compared to using only the state data of the process equipment, using only the measurement result data of the measurement equipment, or using only the state data of the process equipment in the abnormal state and the measurement result data in the abnormal state. In other words, the correlation between the cause of an anomaly in the process equipment and the measurement results can be learned and analyzed with high accuracy and efficiency.

[0057] 13 is a flowchart illustrating a procedure for classifying data used for learning and analysis by an anomaly cause analysis system 10 according to a second embodiment of the present invention. This flowchart can be implemented by, for example, the arithmetic device 106a, as in the first embodiment. The configuration of the anomaly cause analysis system 10 is the same as in the first embodiment.

[0058] The first criterion can be, for example, the number of defects. The number of defects can be, for example, the total number of defects, the number of defects of interest (DOIs), the number of specific defect types among the DOIs, or a combination thereof. Since the data belonging to the second group is used to investigate the cause of the abnormality, the status data of the process equipment at the time of the processing is reverse-looked up, linked to the measurement results, and stored in the database 107. The second criterion can be, for example, the result of defect classification. If the number of DOIs is less than a predetermined number, the defect is classified into the third group; if it does not meet the predetermined number, the defect is classified into the fourth group. If the number of DOIs is already used as the first criterion, the criterion can be set so that the number of DOIs in the second criterion is smaller than the number of DOIs in the first criterion. The second criterion can also be set so that the DOI types in the first criterion and the second criterion are different. These criteria are merely examples, and the criteria described in the first embodiment or other criteria may be used.

[0059] Furthermore, to facilitate learning and analysis that more easily correlates abnormal results with their causes, the data in the second and fourth groups are subdivided using the defect classification results. For example, when there are multiple types of DOI, the causes of each are often different. Therefore, the second and fourth groups are classified by DOI type, and the results are combined with processing status data from the process equipment for learning and analysis. In other words, the AI ​​engine 106 performs learning and analysis for each subdivided classification. This leads to improved accuracy in identifying the causes of abnormalities.

[0060] 14A shows an example of subdividing the second and fourth categories. Subdivision by defect type, such as foreign matter, polishing defect, pattern defect, and crystal defect, that is likely to occur in semiconductor processes, is considered. This classification utilizes the differences in the directionality and polarization components of light scattered or reflected from defects depending on the defect type, such as the defect's shape, size, and material. Defect types other than those listed can also be classified as a single defect type by defining their characteristics. When there are multiple defect types in a DOI, subdivision can also be achieved by combining the DOI classification and defect type classification.

[0061] 14B shows another example of further dividing the second and fourth categories. Since the measurement results of the optical defect inspection device also include defect position information, the defects can be classified by their position on the wafer.

[0062] 14C shows another example of further dividing the second and fourth categories. The defect signal intensity of the measurement result represents the defect size. Therefore, the defect size can be estimated based on the defect signal intensity, and the defects can be classified by defect size.

[0063] FIG. 14D shows another example of further subdividing the second and fourth categories. Optical defect inspection systems may have detectors positioned at different azimuth angles relative to the stage holding the wafer. In addition to the sum of signals detected by each detector, it is also possible to extract only signals from specific detectors. This allows for different light directionality, polarization components, intensity, etc., to be obtained for each detector depending on the defect characteristics. Based on this, defects can be classified according to defect type and its components. This allows for more accurate classification than the sum or average of detected signals. Furthermore, optical defect inspection systems may be capable of irradiating a wafer with light and detecting the elastically scattered light and inelastically scattered light obtained therefrom. This inelastically scattered light can also be used to classify defect components (see JP 2012-154946 A). When the measurement system is an electron beam measurement system such as a review SEM, it may be equipped with an EDX. Defect components may also be classified using component information output by the EDX.

[0064] Third Embodiment FIG. 15 shows an example of a haze map. Signals acquired by an optical defect inspection system include not only defect signals used for defect detection but also signals called haze signals. If signals with a fluctuation frequency (time variation of signal intensity) higher than a predetermined value are considered high frequency and signals with a fluctuation frequency lower than a predetermined value are considered low frequency, defect signals correspond to the high-frequency components of signals based on light obtained from the sample. Haze signals correspond to the low-frequency components. For example, signals caused by relatively large irregularities on the sample, such as defects, are likely to be detected as high-frequency components, while signals caused by characteristics of the wafer itself, such as the thickness of a film on the sample or extremely small irregularities (roughness) on the sample surface, are likely to be detected as low-frequency components. Haze signals are signals that are removed in defect inspection and are not generally used. However, haze signals are obtained together with defect signals during defect inspection and are also measurement results obtained together with defect information without increasing labor costs. Therefore, a map (haze map) representing the in-plane distribution of haze signals can be acquired, and based on this, the relationship between the wafer state and the process equipment state that caused that state can be learned. Other configurations are similar to those of the first and second embodiments.

[0065] The left side of Fig. 15 is a haze map of a normal wafer. The center side of Fig. 15 is a haze map indicating a sign of an abnormality. The right side of Fig. 15 is a haze map indicating an abnormal wafer. For example, in Fig. 9 or Fig. 13, the number of defects is used as the first criterion, and as the second criterion, if the difference from the normal haze map is less than a threshold, it can be considered as the third group, and if it is equal to or greater than the threshold, it can be considered as the fourth group.

[0066] The haze map can be obtained simultaneously with the number of defects (first criterion) (at least without performing a new measurement). Therefore, while the number of defects alone can only classify the defects into the first group and the second group, by using the haze map in combination, it is possible to extract abnormality sign data without incurring new measurement man-hours.

[0067] A haze map can also be used as the first standard. When the optical defect inspection device is equipped with multiple detectors arranged at different azimuth angles, a haze map can be created from the detector signals of each detector, for example. The detector used to create the haze map in the first standard may be different from the detector used to create the haze map in the second standard. The normal haze map used in the first standard may be different from the normal haze map used in the second standard (i.e., the threshold value in the first standard may be different from the threshold value in the second standard).

[0068] The way the wafer state appears on the haze map may differ depending on the polarization direction. For example, if an optical defect inspection system can detect signal light for each polarization direction, it can obtain a haze signal for each polarization direction. Using these haze signals makes it possible to detect changes in the wafer state more precisely.

[0069] The computing device 106a can learn a combination of the Haze map and the state data for each orientation of the detector, each polarization direction, or a combination thereof. The Haze map can also be subdivided for each orientation or polarization direction.

[0070] <Regarding Modifications of the Present Invention> The present invention is not limited to the above-described embodiment, and various modifications are included. For example, the above-described embodiment has been described in detail to clearly explain the present invention, and is not necessarily limited to an embodiment including all of the described configurations. Furthermore, it is possible to replace part of the configuration of one embodiment with the configuration of another embodiment, or to add the configuration of another embodiment to the configuration of one embodiment. Furthermore, it is possible to add, delete, or replace part of the configuration of each embodiment with other configurations.

[0071] In the above embodiment, for convenience of description, the process devices 100 to 102 and the measuring devices 103 to 105 have been described as components of the anomaly cause analysis system 10, but some or all of these devices may be configured as devices separate from the anomaly cause analysis system 10, and the anomaly cause analysis system 10 may acquire each piece of data from these devices.

[0072] In the above embodiment, numerical values ​​such as the number of defects are used as the first and second criteria. However, image data such as an ideal image may be used as the criteria. For example, if the measurement device is a device that outputs an image of a measurement object, an image indicating a normal state may be used as the criteria, and the similarity to that image may be used as the first and second criteria.

[0073] In the above embodiments, the first and second standards may be set based on the output of the same type of measuring device, or may be set based on different measuring devices. For example, the first standard may be set based on the output of an optical defect inspection device, and the second standard may be set based on the output of an electron beam inspection and measurement device.

[0074] In the above embodiment, an example of a configuration for analyzing the cause of an abnormality in a process device that processes semiconductor wafer samples has been described. However, it should be noted that the type of sample is not limited to semiconductor wafers, and the process device may be any device that processes samples.

[0075] In the above embodiments, the AI ​​engine 106 and the arithmetic unit 106a can be configured by hardware such as a circuit device that implements its functions, or by software that implements its functions being executed by an arithmetic unit such as a CPU (Central Processing Unit).

[0076] In the above embodiment, examples of the first criterion and the second criterion have been described, but these may be used in combination. For example, the amount of change in FIG. 11 may be used as the second criterion, and the image feature in FIG. 12 may also be used. In this case, it is sufficient to determine whether the second criterion is satisfied using, for example, an evaluation function that comprehensively evaluates these two feature amounts. The same applies to other criterion examples.

[0077] In the above embodiment, the subdivision classifications described with reference to Figures 14A to 14D can also be used in combination. For example, the defects for each planar position shown in Figure 14B can be further classified by defect size. Other subdivision classifications can also be combined in a similar manner.

[0078] 10: Anomaly cause analysis system 100-102: Process equipment 103-105: Measuring equipment 106: AI engine 106a: Computing device 107: Database

Claims

1. An anomaly cause analysis system for analyzing the cause of an anomaly in a process equipment that processes a sample, comprising: a storage device that stores measurement data obtained by a measurement device measuring the sample and status data describing the status of the process equipment; a learning device that analyzes the cause of the anomaly by learning the relationship between the measurement data, the status data, and the cause of the anomaly; and a calculation device that collects data to be learned by the learning device, wherein if the measurement data satisfies a first criterion, the calculation device classifies the measurement data into a first group as a normal measurement result, and if it does not satisfy the first criterion, the calculation device classifies the measurement data belonging to the first group as a normal measurement result in a third group if it satisfies a second criterion different from the first criterion, and if it does not satisfy the first criterion, the calculation device classifies the measurement data belonging to the first group as a normal measurement result in a fourth group as a measurement result that indicates an anomaly; the calculation device collects the measurement data belonging to the second group and the measurement data belonging to the fourth group as learning data for the learning device to perform the learning; and the learning device analyzes the cause of the anomaly by learning the learning data collected by the calculation device.

2. The anomaly cause analysis system according to claim 1, characterized in that the first criterion is configured to distinguish between the first group and the second group by distinguishing whether the measurement data is a normal measurement result or an abnormal measurement result depending on the degree of defect in the sample, and the second criterion is configured to classify the measurement data into the fourth group if it is a normal measurement result but has specific characteristics, and to classify it into the third group if it is not.

3. The anomaly cause analysis system according to claim 2, characterized in that the second criterion is configured to classify the measurement data into the fourth group when the cumulative value of the number of defects in the sample reaches or exceeds a threshold value or the rate of increase over a certain period of time is equal to or exceeds a threshold value, or the second criterion is configured to classify the measurement data into the fourth group when the planar distribution of defects in the sample has a predetermined characteristic.

4. The anomaly cause analysis system according to claim 1, characterized in that the storage device stores data referenced by a user when analyzing the anomaly cause and a level of importance indicating how important that data was in analyzing the anomaly cause, and the learning device analyzes the anomaly cause by learning the data and the level of importance.

5. The anomaly cause analysis system according to claim 1, characterized in that the calculation device does not collect the measurement data belonging to the first group and the measurement data belonging to the third group as the learning data, and the learning device does not use the measurement data belonging to the first group and the measurement data belonging to the third group as the subject of learning.

6. The anomaly cause analysis system according to claim 1, characterized in that the arithmetic device further subdivides the measurement data belonging to the second group or the fourth group according to the characteristics of defects possessed by the specimen, and the learning device learns the measurement data and the condition data for each subdivided classification.

7. The anomaly cause analysis system according to claim 6, characterized in that the computing device performs the subdivision by using at least one of the following as defect features: the type of defect in the sample; the planar position of the defect in the sample; the size of the defect in the sample; and the components of the defect in the sample.

8. The anomaly cause analysis system according to claim 1, characterized in that the measurement data includes the measurement data caused by defects in the sample and haze data derived from the sample, and at least one of the first standard and the second standard is configured to classify the haze data based on a difference between the haze data and the standard.

9. The anomaly cause analysis system according to claim 6, characterized in that the measurement device measures the sample using a plurality of detectors each having a different orientation relative to the sample, the measurement data describes the measurement data for each of the detectors having a different orientation, and the calculation device subdivides the measurement data for each of the detectors having a different orientation.

10. The anomaly cause analysis system according to claim 6, characterized in that the measurement device measures the sample by detecting multiple lights having different polarization directions from the sample, the measurement data describes the measurement data for each polarization direction, and the calculation device subdivides the measurement data for each polarization direction.

11. A method for analyzing the cause of an abnormality in a process equipment that processes a sample, comprising: a step of storing measurement data acquired by a measuring device measuring the sample and status data describing the status of the process equipment in a storage device; a step of collecting learning data used by a learner to learn the relationship between the measurement data, the status data, and the cause of the abnormality; and a step of analyzing the cause of the abnormality by the learner; wherein in the step of collecting learning data, if the measurement data satisfies a first criterion, it is classified into a first group as a normal measurement result, and if it does not satisfy the first criterion, it is classified into a second group as an abnormal measurement result; in the step of collecting learning data, if the measurement data belonging to the first group satisfies a second criterion different from the first criterion, it is classified into a third group as a normal measurement result, and if it does not satisfy the first criterion, it is classified into a fourth group as a measurement result that has a sign of an abnormality; in the step of collecting learning data, the measurement data belonging to the second group and the measurement data belonging to the fourth group are collected by the learner as learning data for performing the learning; and in the step of analyzing the cause of the abnormality, the learner learns the collected learning data to analyze the cause of the abnormality.

Images