Learning method, inference method, and recording medium on which program is recorded

Abstract

To improve the accuracy of recommending a sensor to be analyzed when an abnormality occurs in a substrate processing device.SOLUTION: A learning method causes a computer to execute a process to acquire data related to a sensor when an abnormality occurs in a substrate processing device, input the data related to the sensor when the abnormality occurs into a learning model, and learn the learning model such that the output of the learning model approaches information related to the sensor to be analyzed when the abnormality occurs.SELECTED DRAWING: Figure 4

Description

[Technical field]

[0001] The present disclosure relates to a learning method, an inference method, and a recording medium on which a program is recorded. [Background technology]

[0002] For example, Patent Document 1 proposes a failure detection system that detects failure of a sensor that detects the state of a semiconductor manufacturing device. The failure detection system includes a generation unit that generates time-series data of information related to the detection value of the sensor during a judgment period, a calculation unit that calculates a regression line of the time-series data, and a failure judgment unit that judges whether the sensor is broken or not based on the slope of the regression line. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2021-68831 Summary of the Invention [Problem to be solved by the invention]

[0004] The present disclosure improves the accuracy of recommending sensors to be analyzed when an abnormality occurs in a substrate processing apparatus. [Means for solving the problem]

[0005] According to one aspect of the present disclosure, there is provided a learning method in which a computer performs processing to acquire data related to a sensor when an abnormality occurs in a substrate processing apparatus, input the data related to the sensor when the abnormality occurs into a learning model, and learn the learning model so that the output of the learning model approaches information related to the sensor to be analyzed when the abnormality occurs. Effect of the Invention

[0006] According to one aspect, it is possible to provide a method for improving the accuracy of recommending a sensor to be analyzed when an abnormality occurs in a substrate processing apparatus. [Brief description of the drawings]

[0007] [Figure 1] 1 is a configuration diagram illustrating an example of a substrate processing system according to an embodiment. [Diagram 2] 1 illustrates an example of a hardware configuration of a learning device according to an embodiment. [Diagram 3] 1 illustrates an example of the functional configuration of a learning device and an inference device according to an embodiment. [Figure 4] 1 is a flowchart illustrating an example of a learning method according to an embodiment. [Diagram 5] FIG. 13 is a diagram showing an example of an evaluation value of a sensor according to an embodiment. [Figure 6] FIG. 1 illustrates an example of a sensor and alarm matrix decomposition according to an embodiment. [Figure 7] FIG. 13 is a diagram showing an example of an SPC chart display. [Figure 8] 1 is a flowchart illustrating an example of an inference method according to an embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0008] Hereinafter, an embodiment of the present disclosure will be described with reference to the drawings. In the drawings, the same components are denoted by the same reference numerals, and duplicated descriptions may be omitted.

[0009] When an abnormality occurs in the substrate processing apparatus, sensor data detected by a sensor attached to the substrate processing apparatus is analyzed to determine the cause of the abnormality. When an abnormality occurs in the substrate processing apparatus, log data related to sensors previously analyzed for the same abnormality is acquired, and a statistical process control (SPC) chart is automatically created using the log data. The SPC chart shows time-series changes in statistical values ​​of sensor data (such as the average gas flow rate) (see FIG. 7). From the trends shown in the SPC chart, an expert narrows down the sensors to be analyzed to those that are expected to have a malfunction. Therefore, in this analysis method, an expert who has knowledge of the substrate processing apparatus needs to manually narrow down the sensors to be analyzed to a certain extent, and then analyze the abnormality using the sensor data of the narrowed down sensors.

[0010] On the other hand, it is also possible to analyze an anomaly using the sensor data of all sensors attached to the substrate processing apparatus without narrowing down the sensors to be analyzed. However, in this case, the processing load of the apparatus performing the calculation is high, which is not preferable from the viewpoint of calculation cost. For example, a batch-type substrate processing apparatus has about 1000 sensors attached, and analyzing all of these sensors leads to an increase in the processing load and is not realistic.

[0011] On the other hand, when narrowing down the sensors to be analyzed in consideration of the calculation cost, there is a problem that the narrowing down method is unclear. Even if the narrowing down method is clear, there is a problem that the method depends on the knowledge of experts who have knowledge about the substrate processing apparatus (hereinafter, also referred to as domain knowledge). Note that the domain knowledge refers to knowledge including the experience, intuition, etc. of people who are familiar with the use and processes of the substrate processing apparatus 10.

[0012] Therefore, in the learning method and inference method according to the present embodiment, a method is proposed that allows selection of candidate sensors to be analyzed without relying on domain knowledge when an abnormality occurs in the substrate processing apparatus, thereby making it possible to avoid an increase in analysis time and calculation cost due to an increase in the number of sensors to be analyzed, and a dependency on domain knowledge when narrowing down the number of sensors.

[0013] [Substrate processing system] The substrate processing system 1 will be described with reference to FIG. 1. FIG. 1 is a configuration diagram showing an example of the substrate processing system 1 according to an embodiment. The substrate processing system 1 includes substrate processing apparatuses 10a and 10b, a learning device 20, an inference device 30, a chart creation device 40, and a data storage unit 50, and these devices are connected to each other so as to be able to communicate data with each other via a communication network N such as the Internet or a LAN. The control device 100a is incorporated in the substrate processing apparatus 10a, and the control device 100b is incorporated in the substrate processing apparatus 10b. The substrate processing apparatuses 10a and 10b are also collectively referred to as the substrate processing apparatus 10. The number of substrate processing apparatuses 10 is not limited to two, and may be one or three or more.

[0014] The sensors 110a and 110b are disposed in the substrate processing apparatuses 10a and 10b. For example, about 1000 sensors 110a and 110b are disposed in each of the substrate processing apparatuses 10a and 10b. Examples of the sensors 110a and 110b include, but are not limited to, a film thickness sensor, a temperature sensor, a humidity sensor, a pressure sensor, a vibration sensor, and a distance sensor. The sensors 110a and 110b are sensors that detect the state of the substrate processing apparatuses 10a and 10b. The sensors 110a and 110b are also collectively referred to as sensors 110.

[0015] For example, in the batch-type substrate processing apparatus 10, a boat on which multiple wafers are placed is carried into a processing vessel (loading), and after the temperature inside the processing vessel is stabilized, a film formation process is simultaneously performed on the multiple wafers on the boat. After the film formation process, the boat on which the multiple wafers on which the films have been formed are placed is carried out of the processing vessel (unloading). The process performed by the substrate processing apparatus 10 is not limited to the film formation process, and may be various substrate processes such as an etching process, a sputtering process, etc.

[0016] The sensor 110 measures the state of the substrate processing apparatus 10 for each event, such as loading, temperature stabilization, film formation, and unloading. For example, if the sensor 110 is a pressure sensor, the sensor 110 measures the pressure in the processing vessel for each event. For example, if the sensor 110 is a temperature sensor, the sensor 110 measures the temperature in the processing vessel for each event. The sensor data, such as the temperature and pressure, measured by the sensor 110 is time-series data of the sensor values, and is stored in the data storage unit 50.

[0017] When an abnormality occurs in the substrate processing apparatus 10, the chart creation device 40 acquires necessary data from the data storage unit 50, and automatically creates an SPC chart (see FIG. 7) using the data.

[0018] The learning device 20 learns a learning model capable of selecting a sensor to be analyzed by substituting the learning data and domain knowledge into the learning model, and creates a learned model. The inference device 30 uses the learned model to select a sensor to be analyzed, and provides a user with information on the sensor to be analyzed. This increases the probability that a user can select an appropriate sensor to be analyzed when an abnormality occurs, even if the user does not have domain knowledge. In addition, it is possible to eliminate the need for an expert with domain knowledge to manually narrow down the sensors to be analyzed, thereby reducing the burden on the expert.

[0019] [Hardware configuration of the learning device and inference device] Next, an example of the hardware configuration of the learning device 20 and the inference device 30 will be described with reference to Fig. 2. Fig. 2 is a diagram showing an example of the hardware configuration of the learning device 20 according to an embodiment. Note that since the hardware configurations of the learning device 20 and the inference device 30 may be substantially the same, the hardware configuration of the learning device 20 will be described here, and a description of the hardware configuration of the inference device 30 will be omitted.

[0020] 2, the learning device 20 includes a central processing unit (CPU) 201 and a read only memory (ROM) 202. The learning device 20 also includes a random access memory (RAM) 203 and a graphics processing unit (GPU) 204. The CPU 201, the ROM 202, the RAM 203, and the GPU 204 form a so-called computer.

[0021] Furthermore, learning device 20 has an auxiliary storage device 205, an operation device 206, a display device 207, an I / F (Interface) device 208, and a drive device 209. The various hardware components of learning device 20 are connected to each other via a bus 210.

[0022] The CPU 201 is a computing device that executes various programs installed in the auxiliary storage device 205 (for example, an input data generation program, a learning program, etc.).

[0023] The ROM 202 is a non-volatile memory and functions as a main storage device. The ROM 202 stores various programs, data, etc. required for the CPU 201 to execute various programs installed in the auxiliary storage device 205. Specifically, the ROM 202 stores boot programs such as a Basic Input / Output System (BIOS) and an Extensible Firmware Interface (EFI).

[0024] The RAM 203 is a volatile memory such as a dynamic random access memory (DRAM) or a static random access memory (SRAM), and functions as a main storage device. The RAM 203 provides a working area in which various programs installed in the auxiliary storage device 205 are expanded when the CPU 201 executes them.

[0025] The GPU 204 is a computing device for image processing, and performs image processing calculations when a learning program is executed based on a learning model by the CPU 201. The GPU 204 is equipped with an internal memory (GPU memory) and temporarily stores information required for performing image processing calculations.

[0026] The auxiliary storage device 205 stores learning data and the like to be processed when various programs such as a learning program are executed by the CPU 201. For example, the data storage unit 50 and the memory unit 23 in FIG.

[0027] Operation device 206 is an input device used by an administrator of learning device 20 to input various instructions to learning device 20. Display device 207 is a display device that displays the internal state of learning device 20. I / F device 208 is a connection device for connecting to and communicating with other devices.

[0028] The drive device 209 is a device for setting the recording medium 220. The recording medium 220 here includes media that record information optically, electrically, or magnetically, such as a CD-ROM, a flexible disk, a magneto-optical disk, etc. The recording medium 220 may also include semiconductor memories that record information electrically, such as ROM, flash memory, etc.

[0029] The various programs to be installed in the auxiliary storage device 205 are installed, for example, by setting the distributed recording medium 220 in the drive device 209 and reading out the various programs recorded in the recording medium 220 by the drive device 209. Alternatively, the various programs to be installed in the auxiliary storage device 205 may be installed by being downloaded via a network (not shown).

[0030] [Functional configuration of the learning device and inference device] Next, a functional configuration example of the learning device 20 and the inference device 30 will be described with reference to Fig. 3. Fig. 3 is a diagram showing a functional configuration example of the learning device 20, the inference device 30, etc. according to an embodiment.

[0031] The learning device 20 and the inference device 30 are connected to a chart creation device 40 and a data storage unit 50. The data storage unit 50 accumulates alarm data 52 including the time of abnormality occurrence and the type of alarm when an abnormality occurs during processing (substrate processing) performed by the substrate processing apparatus 10.

[0032] The data storage unit 50 also accumulates sensor data 53 relating to the state of the substrate processing apparatus 10 measured by the sensor 110, and recipe data 54 in which processing conditions for the substrate processing performed by the substrate processing apparatus 10 are set. The sensor data 53 may store analysis results for each abnormality and each sensor (evaluation value of the analyzed sensor, etc.). The recipe data 54 may set processing conditions for each step when the processing performed by the substrate processing apparatus 10 is divided into a plurality of steps. The data storage unit 50 also accumulates maintenance data 55, such as the number of times the sensor 110 has been maintained and the time of the maintenance.

[0033] The learning device 20 includes an acquisition unit 21, a learning unit 22, a storage unit 23, and a display unit 24. The acquisition unit 21 acquires data required as learning data from a data storage unit 50. The learning device 20 stores the acquired data in learning data 25 in the storage unit 23. The storage unit 23 also stores three levels of learning models (first learning model 24a, second learning model 24b, and third learning model 24c). However, the number of learning models stored in the storage unit 23 may be one or more.

[0034] The learning data 25 includes data related to the sensor 110 when an abnormality occurs in the substrate processing apparatus 10. The learning data 25 may include data related to a process performed by the substrate processing apparatus 10 when an abnormality occurs. The data related to the sensor 110 when an abnormality occurs may include sensor data 53 detected by the sensor 110 when the substrate processing apparatus 10 performs substrate processing when the abnormality occurs. The data related to the sensor 110 when an abnormality occurs may include an evaluation value of the sensor 110 analyzed when the abnormality occurs.

[0035] The data related to the process may include recipe data 54 indicating process conditions when the substrate processing apparatus 10 executes the process when the abnormality occurs. Furthermore, the data related to the process may include maintenance data 55 when the substrate processing apparatus 10 executes maintenance before and after the process when the abnormality occurs.

[0036] By inputting the above learning data 25 into the learning model, abnormality factors correlated with each event in the substrate processing can be reflected in the machine learning of the learning model.

[0037] A learning program is installed in the learning device 20, and the learning device 20 functions as a learning unit 22 by executing this learning program.

[0038] The learning unit 22 performs machine learning on the learning model using the learning data 25, and generates a trained model. The trained model generated by the learning unit 22 is output to the inference device 30.

[0039] The learning unit 22 inputs data related to the sensor 110 when an abnormality occurs into the learning model, and executes a learning program to train the learning model so that the output of the learning model approaches information related to the sensor that should be analyzed when an abnormality occurs. The learning unit 22 may further input data related to the process executed by the substrate processing apparatus 10 when an abnormality occurs into the learning model, and train the learning model so that the output of the learning model approaches information related to the sensor that should be analyzed when an abnormality occurs.

[0040] The learning unit 22 selects a learning model to be learned from the learning models (first learning model 24a, second learning model 24b, third learning model 24c) stored in the memory unit 23. The learning unit 22 learns the learning model so that the output of the selected learning model approaches information about a sensor to be analyzed when an abnormality occurs. The learning unit 22 may select multiple learning models. In this case, the learning unit 22 learns each of the multiple learning models so that the output of each of the multiple learning models approaches information about a sensor to be analyzed when an abnormality occurs.

[0041] The learning unit 22 may select the anomaly to be analyzed in accordance with a user operation or automatically. In this case, the learning unit 22 learns the learning model so that the output of the learning model approaches information on the sensor to be analyzed when the selected anomaly occurs.

[0042] In this way, the learning unit 22 learns the learning model so that the output of the learning model selects appropriate sensor information as an analysis target when an abnormality occurs. That is, the learning device 20 learns to improve the accuracy of narrowing down the sensors by inputting the learning data 25 to the learning model so that the candidate analysis target sensor output by the learning model becomes a more appropriate sensor as an analysis target when the selected abnormality occurs. This improves the accuracy of recommending a sensor to be an analysis target when an abnormality occurs in the substrate processing apparatus 10.

[0043] The display unit 24 outputs information about the sensor to be analyzed when an abnormality occurs, which is the learning result of the learning model. For example, the display unit 24 displays an SPC chart of the analysis target sensor output by the learning model.

[0044] For example, it is preferable that a user such as an expert having domain knowledge judges whether a candidate sensor is appropriate as a sensor to be analyzed from the trend shown in the SPC chart, and feeds back the judgment result to the learning device 20. The learning unit 22 further learns the learning model so that the output of the learning model becomes the sensor selected by the user from among the sensor candidates. In this way, the domain knowledge of the expert is learned in the learning model, and the recommendation accuracy of the sensor 110 candidate can be further improved. However, the learning method of the learning model is not limited to this, and for example, it is also possible to automatically judge whether a sensor shown in the SPC chart is appropriate as a sensor to be analyzed from a trend such as the slope of the plot points of the SPC chart, and feed back the automatic judgment result to the learning device 20.

[0045] The learned model generated by performing machine learning in the learning device 20 is loaded onto the inference device 30. The inference device 30 includes an acquisition unit 31, an execution unit 32, a storage unit 33, and a display unit 34. The acquisition unit 31 acquires the learned model learned by the learning device 20 and stores it in the storage unit 33. For example, the acquisition unit 31 may store in the storage unit 33 a first learned model 34a which is a learned model of the first learning model 24a, a second learned model 34b which is a learned model of the second learning model 24b, and a third learned model 34c which is a learned model of the third learning model 24c.

[0046] That is, the inference device 30 has stored in the storage unit 33 trained models (first trained model 34a, second trained model 34b, third trained model 34c) trained to output, when data related to the sensor 110 when an abnormality (hereinafter referred to as a first abnormality) of the substrate processing apparatus 10 is input, closer to information related to the sensor to be analyzed when the first abnormality occurs. The inference device 30 may have stored in the storage unit 33 trained models trained to output, when data related to the sensor 110 when the first abnormality of the substrate processing apparatus 10 is input and data related to a process (hereinafter referred to as a first process) executed by the substrate processing apparatus 10 when the first abnormality occurs, closer to information related to the sensor to be analyzed when the first abnormality occurs.

[0047] An execution program is installed in the inference device 30, and the inference device 30 functions as an execution unit 32 by executing this execution program.

[0048] The execution unit 32 inputs data related to the sensor 110 when an abnormality (hereinafter referred to as a second abnormality) occurs in the substrate processing apparatus 10 into the trained model, and executes an execution program to infer a sensor to be analyzed when the second abnormality occurs. The execution unit 32 may input data related to the sensor 110 when the second abnormality occurs in the substrate processing apparatus 10 and data related to a process (hereinafter referred to as a second process) executed by the substrate processing apparatus 10 when the second abnormality occurs into the trained model, and executes an execution program to infer a sensor to be analyzed when the second abnormality occurs.

[0049] The data related to the sensor 110 when the second abnormality occurs may include sensor data detected by the sensor 110 when the substrate processing apparatus 10 performs the second process when the second abnormality occurs.

[0050] The data relating to the sensor 110 at the time when the second anomaly occurs may include an evaluation value of the sensor 110 analyzed at the time when the second anomaly occurs.

[0051] The data related to the second process may include recipe data indicating process conditions when the substrate processing apparatus 10 performs the second process. The data related to the second process may include maintenance data when the substrate processing apparatus 10 performs maintenance before and after the second process.

[0052] The display unit 34 outputs information about the sensor to be analyzed when the second abnormality occurs, which is the inference result of the trained model. For example, the display unit 24 displays an SPC chart of the analysis target sensor output by the trained model. The SPC chart is created by the chart creation device 40.

[0053] The chart creation device 40 has an acquisition unit 41 and a chart creation unit 42. When an abnormality occurs in the substrate processing apparatus 10, the acquisition unit 41 acquires data related to the abnormality, such as the time of occurrence of the abnormality. When an abnormality occurs in the substrate processing apparatus 10, the acquisition unit 41 acquires alarm data 52 and sensor data 53 from the data storage unit 50.

[0054] The chart creation unit 42 performs statistical processing on the sensor data 53 detected from the substrate processing apparatus 10 around the time when the abnormality occurred, and automatically creates an SPC chart showing the state of the substrate processing apparatus 10. The chart creation unit 42 automatically creates an SPC chart (see FIG. 7) using the alarm data 52 and the sensor data 53 acquired by the acquisition unit 41. The horizontal axis of the SPC chart indicates the date and time (time), and the vertical axis indicates the sensor data (e.g., the average gas flow rate) detected by the sensor to be analyzed. The average gas flow rate is an example of a statistical value, and the sensor data may be, for example, a standard deviation or a statistical value obtained by statistically processing other sensor data.

[0055] [Learning method] Next, a description will be given of a learning method executed by the learning device 20. Fig. 4 is a flowchart showing an example of a learning method according to an embodiment. When an instruction to perform learning is input, the learning device 20 executes the flowchart shown in Fig. 4.

[0056] First, in step S201, the learning unit 22 selects an alarm to be analyzed. However, when the present learning method is performed immediately after an abnormality occurs in the substrate processing apparatus 10, the process of step S201 can be omitted.

[0057] Next, in step S202, the acquisition unit 21 collects learning data.

[0058] Next, in step S203, the learning unit 22 selects a learning model. The learning unit 22 may select one learning model, or may select multiple learning models. When the learning unit 22 selects multiple learning models, the learning unit 22 may execute the processes of steps S204 to S207 in parallel for each learning model.

[0059] Next, in step S204, the learning unit 22 narrows down the sensors to be analyzed using the learning model, that is, selects the sensors to be analyzed when an abnormality occurs.

[0060] Next, in step S205, the learning unit 22 determines whether to further narrow down the analysis target sensors. If the learning unit 22 determines not to narrow down the analysis target sensors, the process proceeds to step S207, the display unit 24 displays the learning result, and the process ends. This allows the user to confirm the learning result. On the other hand, if the learning unit 22 determines to narrow down the analysis target sensors in step S205, the process proceeds to step S206, where an expert with domain knowledge, such as an operator of the substrate processing apparatus 10, checks the information on the analysis target sensors narrowed down in step S204, and further selects candidates for the analysis target sensors to narrow down the information on the analysis target sensors (sensor candidates). The sensor candidates may be further narrowed down by automatically selecting candidates for the analysis target sensors. The narrowed down result is fed back to the learning model as an evaluation value of the sensor, which will be described later.

[0061] After the analysis target sensor is selected, in step S207, the display unit 24 displays the analysis result, and the process ends, allowing the user to check the learning result.

[0062] In step S205, the learning unit 22 determines whether to further narrow down the sensors to be analyzed as a criterion for determining whether to further narrow down the sensors to be analyzed in step S205 when the number of sensors to be analyzed narrowed down in step S204 is greater than a set reference value.

[0063] The more data on the sensor when an abnormality occurs that is input to the learning model, the higher the learning accuracy of the learning model. In other words, in the early stages of the learning stage, there is not much data on the sensor when an abnormality occurs that is input to the learning model, so the accuracy of the learning model is not very high. Therefore, until the accuracy of the learning model increases, it is preferable to determine in step S205 that the sensors to be analyzed should be further narrowed down, and to narrow down the sensors to be analyzed manually or automatically in step S206.

[0064] By repeating the feedback, the accuracy of the sensor recommendation by the learning model increases. Therefore, eventually, manual or automatic narrowing down of the sensors becomes unnecessary. In other words, when the output accuracy of the learning model increases to a certain degree and it becomes possible to appropriately narrow down the analysis target sensors (sensor selection) in step S204, it may be determined in step S205 that further narrowing down of the analysis target sensors is not required.

[0065] (Learning example of the first learning model) In a learning example of the first learning model 24a, the knowledgeable person manually presets a sensor to be analyzed when an abnormality occurs in the substrate processing apparatus 10. At this time, the knowledgeable person manually presets a sensor to be analyzed for each alarm of the substrate processing apparatus 10. The knowledgeable person may also preset a sensor to be analyzed for each event of the substrate processing apparatus 10. The event may correspond to a process or step included in the recipe data.

[0066] The acquiring unit 21 acquires, from the data storage unit 50, an analysis target sensor preset for an alarm identical to the abnormality that has previously occurred in the substrate processing apparatus 10, as data related to the sensor when an abnormality occurs in the substrate processing apparatus 10. The acquiring unit 21 may acquire, as data related to a process to be executed by the substrate processing apparatus 10 when an abnormality occurs, an analysis target sensor preset for an event when an abnormality occurs in the substrate processing apparatus 10.

[0067] The learning unit 22 inputs the acquired learning data into the learning model. The learning data may be only data related to the sensor when an abnormality occurs in the substrate processing apparatus 10. The learning data may be data related to the sensor when an abnormality occurs in the substrate processing apparatus 10 and data related to the process executed by the substrate processing apparatus 10 when an abnormality occurs.

[0068] The learning unit 22 learns the learning model so that the output of the learning model approaches information about the analysis target sensor at the time of abnormality occurrence preset by the expert. The display unit 24 displays information about the analysis target sensor as the learning result. The display unit 24 outputs, for example, an SPC chart as the learning result.

[0069] (Example of learning using the second learning model) A learning example of the second learning model 24b will be described with reference to FIG. 5. FIG. 5 is a diagram showing an example of an evaluation value of a sensor analyzed when an abnormality occurs according to one embodiment. The data shown in FIG. 5(a) is log data that associates an abnormality that has occurred in the substrate processing apparatus 10 in the past with the analyzed evaluation value of the sensor, and is stored in the data storage unit 50. FIG. 5(a) shows the time when an abnormality occurred in the substrate processing apparatus 10, the type of alarm, and the evaluation values ​​of sensors A to X analyzed when each alarm occurred. In this learning example, rule information is defined in advance in the second learning model 24b.

[0070] When an alarm to be analyzed is selected, the acquiring unit 21 collects data related to the sensor at the time of occurrence of an abnormality in the substrate processing apparatus 10 from the data storage unit 50. For example, when "pressure convergence waiting time over" is selected as the alarm to be analyzed, the acquiring unit 21 acquires data 601, 602, 603, and 605 corresponding to the alarm type "pressure convergence waiting time over". Fig. 5(b) shows an example of the collected learning data.

[0071] The acquiring unit 21 may acquire data on an event (process) executed by the substrate processing apparatus 10 when an abnormality occurs. For example, based on the rule information, when the event with the abnormality occurrence time "2023 / 04 / 17 16:59:00" shown in FIG. 5(a) is the event (process) to be analyzed, but the event with the abnormality occurrence time "2023 / 04 / 17 16:58:00" is not the event (process) to be analyzed, the acquiring unit 21 acquires the data 601, 602, and 603, excluding the data 605, based on the rule information. However, the collection of learning data is not limited to these examples.

[0072] In a learning example of the second learning model 24b, the learning unit 22 uses rule information defined by the second learning model 24b to automatically preset a sensor to be analyzed when an abnormality occurs in the substrate processing apparatus 10. At this time, the learning unit 22 automatically presets a sensor to be analyzed for each alarm of the substrate processing apparatus 10 based on the rule information. The learning unit 22 may also automatically preset a sensor to be analyzed for each event of the substrate processing apparatus 10 based on the rule information.

[0073] The learning unit 22 inputs the acquired learning data into the learning model. The learning data may be only data related to the sensor when an abnormality occurs in the substrate processing apparatus 10. The learning data may be data related to the sensor when an abnormality occurs in the substrate processing apparatus 10 and data related to the process executed by the substrate processing apparatus 10 when an abnormality occurs.

[0074] The learning unit 22 learns the learning model so that the output of the learning model approaches information about the analysis target sensor at the time of the occurrence of an abnormality, which is automatically preset based on the rule information. As an example of the rule information, it may be defined that the analysis target sensor is selected based on the evaluation value of the sensor for an alarm identical to the abnormality that occurred in the past in the substrate processing apparatus 10. As another example of the rule information, it may be defined that when one analysis target sensor is selected, a sensor that is often analyzed together with the selected analysis target sensor is selected as the analysis target sensor. As another example of the rule information, it may be defined that the analysis target sensors are selected in order of the oldest installation and replacement based on the period since the sensor was installed and replaced.

[0075] The display unit 24 displays information about the selected analysis target sensor as the learning result, and outputs, for example, an SPC chart as the learning result.

[0076] 5(b) illustrates the sum of the evaluation values ​​of the sensors in the learning data. The learning unit 22 learns a learning model based on the rule information so that a sensor with a high sum of evaluation values ​​calculated from the learning data is selected as a sensor to be analyzed. In FIG. 5(b), the learning unit 22 predicts that sensor A with a high evaluation value, i.e., sensor A with the most number of analysis executions, is highly likely to be the cause of an abnormality. The display unit 24 displays information about sensor A and recommends sensor A as a sensor to be analyzed.

[0077] The evaluation value of a sensor is not limited to the number of times analysis has been performed. For example, a sensor that has not been analyzed may be assigned a value of "0," and a sensor that has been analyzed and is correct as the sensor to be analyzed, i.e., the sensor that is causing the abnormality, may be assigned a value of "1."

[0078] (Example of the third learning model) A learning example of the third learning model 24c will be described with reference to Fig. 6. Fig. 6 is a diagram showing an example of matrix decomposition of sensors and alarms according to an embodiment. In the third learning model 24c, it is possible to narrow down sensors to be analyzed using matrix decomposition of the matrix of sensors and alarms.

[0079] The data shown in Fig. 6(a) is stored in the data storage unit 50 in association with abnormalities that have occurred in the substrate processing apparatus 10 in the past and the analysis results of the sensors. In Fig. 6(a), the vertical axis indicates the type of alarm, and the horizontal axis indicates the number of times that the analysis of the sensor that analyzed the alarm has been performed. The number of times that the analysis has been performed is an example of an evaluation value of the sensor.

[0080] The acquiring unit 21 collects, for example, data related to a sensor when an abnormality occurs in the substrate processing apparatus 10 from the data storage unit 50. Fig. 6(a) shows an example of a result of collection of learning data. The acquiring unit 21 may acquire data related to an event (process) executed by the substrate processing apparatus 10 when an abnormality occurs.

[0081] The learning unit 22 uses the third learning model 24c to decompose this numerical value into latent vectors, and decomposes it into a sensor latent matrix and an alarm latent matrix. The number of elements k of the latent vector is optimized. For example, when the number of elements k of the latent vector is set to 2, the learning unit 22 decomposes it into two latent vectors each. As a result, as shown in FIG. 6(b), it is decomposed into a 2-row, 3-column matrix 704 of sensor latent vectors and a 3-row, 2-column matrix 705 of alarm latent vectors.

[0082] For example, the first row 704a of the matrix 704 is determined to be data related to the temperature of the sensor because Heater Z1 Power is 1, and Vacuum Generator 1 (VG1) and APC1 Pressure (APC1 PRESS) are both 0. The second row 704b is determined to be data related to the pressure of the sensor because Heater Z1 Power is 0, Vacuum Generator 1 is 2, and APC1 Pressure is 1. These determinations are made manually or automatically.

[0083] Similarly, the first column 705a and the second column 705b of the matrix 705 are determined to be data related to the temperature of the alarm, and the second column 705b is determined to be data related to the pressure of the alarm. These determinations are made manually or automatically.

[0084] From the result, the learning unit 22 reconstructs the decomposed latent vector matrices 704, 705 into an evaluation value matrix 706. That is, the reconstructed evaluation value matrix 706 is obtained by multiplying the matrix 704 of the sensor latent vectors and the matrix 705 of the alarm latent vectors.

[0085] In this way, the learning unit 22 inputs the learning data into the learning model, and selects a sensor to be analyzed according to the type of alarm based on the evaluation value matrix 706 reconstructed using the third learning model 24c. In the example of the evaluation value matrix 706 in Fig. 6(b), the learning unit 22 selects the vacuum generator 1, which has the highest sensor evaluation value in the row where the alarm type is "pressure convergence wait time over", as the sensor to be analyzed.

[0086] The display unit 24 displays information about the selected analysis target sensor as the learning result, and outputs, for example, an SPC chart.

[0087] According to the above learning method, the more data related to the sensor when an abnormality occurs, the higher the learning accuracy of the learning model becomes, so that the accuracy of the trained model improves as the substrate processing apparatus 10 is operated.

[0088] In particular, in the learning of the third learning model, an automatically reconstructed evaluation value can be obtained. This makes it possible to discover interactions between latent vectors and feature quantities that are difficult for humans to discover manually. This improves the interpretability of the reasons for selecting a sensor to be analyzed, and increases the possibility of extracting new candidates for a sensor to be analyzed. As a result, it is possible to improve the accuracy of recommending a sensor to be analyzed when an abnormality occurs in the substrate processing apparatus 10.

[0089] In addition, the learning of the third learning model can suppress the processing load and reduce the calculation cost significantly. For example, if there are 1,000 sensors and 100 alarms, the evaluation value matrix will have 100,000 elements in the conventional method. Reconstructing the 100,000 evaluation values ​​increases the processing load and calculation cost, making it unrealistic.

[0090] In contrast, in the learning example of the third learning model 24c, if the number of elements in the latent vector is 3, the sensor latent vector is 1000 × 3, the alarm latent vector is 100 × 3, and the evaluation value matrix can be expressed with 3300 elements. In this way, 100,000 evaluation values ​​can be expressed as 3300 evaluation values ​​by matrix decomposing the latent vectors, which can significantly reduce the processing load and calculation cost.

[0091] The first learning model 24a to the third learning model 24c are examples of learning models, and the learning models used in the learning method are not limited to these. Since the learning models can be realized at various learning levels using various models, it is possible to propose a learning model that matches the user's issues and needs.

[0092] FIG. 7 is a diagram showing a display example of an SPC chart of a sensor. In this way, an SPC chart of a sensor selected as an analysis target may be displayed as an example of a learning result. The SPC chart is an example of information about a sensor that should be analyzed when an abnormality occurs. FIGS. 7(a) to (c) are graphs of the created SPC chart, in which the horizontal axis of the graph indicates the date and time, and the vertical axis plots the statistical values ​​of the sensor data selected as an analysis target. The SPC chart shows, for example, a time series change in the sensor data including before and after the occurrence of an abnormality.

[0093] For example, assume that three sensors are selected as the analysis target, and the SPC charts of the respective sensors are as shown in Figures 7(a) to 7(c). The user may check the SPC charts or the SPC charts may be automatically checked, and the user may manually select or automatically select a sensor that is appropriate as the analysis target sensor based on the variation and gradient of the plot points of the sensor data.

[0094] The user can also select an appropriate sensor based on the slope of the plot points of the sensor data on the SPC chart. In this case, the slope of the line P1 in FIG. 7(a) is greater than the slopes of the lines P2 and P3 in FIG. 7(b) and (c), and based on the variation in the plot points, the user can determine that the SPC chart in FIG. 7(a) has a trend and is appropriate as a sensor to be analyzed. The user can also determine that the SPC charts in FIG. 7(b) and (c) do not have a trend and are inappropriate as a sensor to be analyzed. The learning unit 22 feeds back this determination result to the learning model. This can improve the accuracy of recommending a sensor to be analyzed when an abnormality occurs in the substrate processing apparatus 10.

[0095] These determinations can also be made automatically by the learning unit 22. The learning unit 22 can also automatically determine the recommendation accuracy of a sensor from the output SPC chart based on a numerical value (residual sum of squares) indicating the deviation of the sensor data from the plot points.

[0096] [Inference method] Next, an inference method executed by the inference device 30 will be described with reference to Fig. 8. Fig. 8 is a flowchart showing an example of the inference method according to an embodiment. When an instruction to perform inference is input or an abnormality occurs, the inference device 30 executes the flowchart shown in Fig. 8.

[0097] The memory unit 33 of the inference device 30 has stored therein a trained model that has been trained so that the output when data relating to a sensor at the time an abnormality (first abnormality) occurs in the substrate processing apparatus 10 is input approaches information relating to the sensor that should be analyzed at the time the first abnormality occurs.

[0098] First, in step S301, the learning unit 22 selects an alarm to be analyzed. However, in the case where the present inference method is performed immediately after an abnormality (second abnormality) occurs in the substrate processing apparatus 10, the process of step S301 can be omitted.

[0099] Next, in step S302, the acquiring unit 31 acquires the inference data 35. For example, the acquiring unit 31 acquires data related to a sensor when the second abnormality occurs in the substrate processing apparatus 10 as the inference data 35. The inference data 35 may include sensor data related to the state of the substrate processing apparatus 10 detected by the sensor when the substrate processing apparatus 10 performs the second process when the second abnormality occurs. The inference data 35 may also include an evaluation value of the sensor analyzed when the second abnormality occurs.

[0100] The data for inference 35 may include data on a second process executed by the substrate processing apparatus 10 when the second abnormality occurs. The data on the second process may include recipe data indicating process conditions when the substrate processing apparatus 10 executes the second process. The data on the second process may include maintenance data when the substrate processing apparatus executes maintenance before and after the second process.

[0101] Next, in step S303, the learning unit 22 selects a trained model. The learning unit 22 may select one trained model, or may select multiple trained models. When the learning unit 22 selects multiple trained models, the learning unit 22 can execute the processes of steps S304 to S307 in parallel for each trained model.

[0102] Next, in step S304, the learning unit 22 narrows down the analysis target sensors using the learned model, that is, selects candidates for the analysis target sensors.

[0103] Next, in step S305, the learning unit 22 determines whether to further narrow down the analysis target sensors. If the learning unit 22 determines not to narrow down the analysis target sensors, in step S307, the display unit 34 displays the analysis result. The user checks the analysis result. On the other hand, if the learning unit 22 determines to narrow down the analysis target sensors, in step S306, for example, an expert with domain knowledge checks the analysis target sensors narrowed down in step S304, selects candidate analysis target sensors by the expert, and further narrows down the candidate sensors. Candidate analysis target sensors may be automatically selected, and the candidate sensors may be further narrowed down. The narrowed down result is fed back to the trained model as the evaluation value of the sensor.

[0104] In step S305, the learning unit 22 may use the following criteria to determine whether to further narrow down the sensors to be analyzed: if the number of candidate sensors to be analyzed narrowed down in step S304 is large, the learning unit 22 may determine in step S305 to further narrow down the sensors to be analyzed.

[0105] After the sensor to be analyzed is selected, in step S307, the display unit 34 displays the analysis results. The user checks the analysis results. The display unit 24 outputs, for example, an SPC chart as an example of information related to the sensor to be analyzed when an abnormality occurs. FIG. 7 is a diagram showing an example of a display of an SPC chart of a sensor. In this way, by displaying the SPC chart of the sensor selected as the analysis target, the user can easily determine whether the recommended sensor is appropriate as the analysis target.

[0106] The learning method, the inference method, and the recording medium on which the program is recorded according to the disclosed embodiments should be considered as illustrative and not restrictive in all respects. The embodiments can be modified and improved in various forms without departing from the scope and spirit of the appended claims. The matters described in the above embodiments can be configured in other ways without contradiction, and can be combined without contradiction.

[0107] For example, the collected learning data 25 is not limited to the alarm data 52 and the sensor data 53. When it is desired to obtain information on the sensor to be analyzed when an abnormality occurs according to the recipe used by the substrate processing apparatus 10, the learning data 25 may include the alarm data 52, the recipe data 54, and the sensor data 53. When it is desired to take into account information on maintenance performed before and after the occurrence of an abnormality, the learning data 25 may include the alarm data 52, the maintenance data 55, and the sensor data 53. This makes it possible to realize learning of a learning model based on the maintenance data 55, taking into account that a sensor that has been maintained many times is prone to failure, and that a sensor that has been maintained recently has a low probability of failure.

[0108] The substrate processing system 1 in Fig. 1 is just one example, and it goes without saying that there are various system configuration examples depending on the application and purpose. For example, although the learning device 20 and the inference device 30 are separate information processing devices in Fig. 1, they may be the same information processing device. In addition, although the learning device 20 and the inference device 30 are separate information processing devices, the chart creation device 40 may be realized as a function incorporated in the learning device 20 and the inference device 30.

[0109] The substrate processing apparatus of the present disclosure can be applied to any of single wafer processing apparatuses that process substrates one by one, and batch and semi-batch processing apparatuses that process multiple substrates at once. The substrate processing apparatus can also be applied to any type of substrate processing apparatus, such as a heat treatment apparatus that does not generate plasma. Furthermore, the substrate processing apparatus may be a plasma processing apparatus, such as an atomic layer deposition (ALD) apparatus, a capacitively coupled plasma (CCP), an inductively coupled plasma (ICP), a radial line slot antenna (RLSA), an electron cyclotron resonance plasma (ECR), or a helicon wave plasma (HWP). [Explanation of symbols]

[0110] 1...substrate processing system, 10, 10a, 10b...substrate processing apparatus, 20...learning device, 30...inference device, 40...chart creation device, 50...data storage unit, 52...alarm data, 53...sensor data, 54...recipe data, 110, 110a, 110b...sensor

Claims

1. Acquire data related to sensors when an abnormality occurs in the substrate processing apparatus; inputting data related to the sensor at the time of occurrence of the abnormality into a learning model, and learning the learning model so that an output of the learning model approaches information related to the sensor to be analyzed at the time of occurrence of the abnormality; A learning method in which a computer performs a process.

2. The data related to the sensor at the time of occurrence of the abnormality includes sensor data detected by the sensor when the substrate processing apparatus performed a process at the time of occurrence of the abnormality. The learning method according to claim 1 .

3. The data on the sensor at the time of occurrence of the abnormality includes an evaluation value of the sensor analyzed at the time of occurrence of the abnormality. The learning method according to claim 1 .

4. acquiring data relating to a process performed by the substrate processing apparatus when the abnormality occurs; inputting the data related to the sensor at the time of occurrence of the abnormality and the data related to the processing into a learning model, and learning the learning model so that an output of the learning model approaches information related to the sensor to be analyzed at the time of occurrence of the abnormality; The learning method according to claim 1 .

5. the data relating to the processing includes recipe data indicating processing conditions when the substrate processing apparatus executes the processing when the abnormality occurs; The learning method according to claim 1 .

6. The data relating to the processing includes maintenance data when the substrate processing apparatus performs maintenance before and after the processing when the abnormality occurs. The learning method according to claim 1 .

7. outputting information about a sensor to be analyzed when an abnormality occurs using the learning model; A learning method according to any one of claims 1 to 6.

8. The learning model is trained so as to output information on a sensor to be analyzed when the abnormality occurs, the information being selected by a user or automatically selected from the output information on the sensor. The learning method according to claim 7.

9. The information about the sensor to be output is an SPC chart about the sensor to be analyzed. The learning method according to claim 7.

10. storing a trained model that is trained so that an output when data related to a sensor at the time of occurrence of a first abnormality in the substrate processing apparatus is inputted approaches information related to the sensor to be analyzed at the time of occurrence of the first abnormality; inputting data relating to a sensor at the time of occurrence of a second abnormality in the substrate processing apparatus into the trained model, and inferring a sensor to be analyzed at the time of occurrence of the second abnormality; A method of inference in which processing is carried out by a computer.

11. the data related to the sensor at the time of occurrence of the second abnormality includes sensor data related to a state of the substrate processing apparatus detected by the sensor when the substrate processing apparatus performed a process at the time of occurrence of the second abnormality; The inference method of claim 10.

12. The data related to the sensor at the time of occurrence of the second abnormality includes an evaluation value of the sensor analyzed at the time of occurrence of the second abnormality. The inference method of claim 10.

13. storing the trained model that has been trained so that an output when data on the sensor at the time of occurrence of the first abnormality and data on a first process executed by the substrate processing apparatus at the time of occurrence of the first abnormality are inputted approaches information on the sensor to be analyzed at the time of occurrence of the first abnormality; inputting data on a sensor when a second abnormality occurs in the substrate processing apparatus and data on a second process executed by the substrate processing apparatus when the second abnormality occurs into the trained model, and inferring a sensor to be analyzed when the second abnormality occurs; The inference method of claim 10.

14. the data relating to the second process includes recipe data indicating process conditions when the substrate processing apparatus executes the second process when the abnormality occurs; The method of claim 13.

15. the data relating to the second process includes maintenance data when the substrate processing apparatus performs maintenance before and after the second process when the abnormality occurs. The method of claim 13.

16. outputting information about a sensor to be analyzed when the second abnormality occurs using the trained model; The inference method according to any one of claims 10 to 15.

17. The information about the sensor to be output is an SPC chart about the sensor to be analyzed.

17. The method of inference according to claim 16.

18. storing a trained model that is trained so that an output when data related to a sensor at the time of occurrence of a first abnormality in the substrate processing apparatus is inputted approaches information related to the sensor to be analyzed at the time of occurrence of the first abnormality; inputting data relating to a sensor at the time of occurrence of a second abnormality in the substrate processing apparatus into the trained model, and inferring a sensor to be analyzed at the time of occurrence of the second abnormality; A recording medium on which a program for causing a computer to execute a process is recorded.

Images