Computer program, method for generating trained model, information processing method, and information processing device

Abstract

The present invention provides a computer program, a method for generating a trained model, an information processing method, and an information processing device that can be expected to perform searching of time-series data related to substrate processing.　This computer program causes a computer to execute a process for generating a trained model that receives time-series data related to substrate processing as an input and outputs feature quantities of the time-series data, the computer program causing the computer to execute a process in which: training data in which time-series data related to substrate processing and a plurality of correct answer values related to a process performed on the basis of the feature quantities of the time-series data are associated with each other is acquired; time-series data of the acquired training data is inputted to the trained model, and feature quantities outputted by the trained model are acquired; a plurality of errors are calculated on the basis of the acquired feature quantities, the correct answer values for the training data, and a plurality of loss functions; the total error is calculated on the basis of the calculated plurality of errors; and the parameters of the trained model are updated on the basis of the calculated total error.

Description

Computer program, learning model generation method, information processing method, and information processing device

[0001] The present disclosure relates to a computer program, a method for generating a learning model, an information processing method, and an information processing device.

[0002] Patent Document 1 proposes an anomaly detection device that acquires observation values ​​that are indicators of the operating state of the monitored device at predetermined times during processing that is repeatedly executed in the monitored device, applies statistical modeling to a summary value that summarizes the acquired observation values, infers a state with noise removed from the summary value, generates a predicted value that predicts the summary value one period ahead based on this inference, and detects whether or not there is an abnormality in the monitored device based on the generated predicted value.

[0003] International Publication No. 2018 / 061842

[0004] The present disclosure provides a computer program, a learning model generation method, an information processing method, and an information processing device that are expected to enable searching of time-series data related to substrate processing.

[0005] A computer program according to one embodiment is a computer program that causes a computer to execute a process of receiving time-series data related to substrate processing as input and generating a learning model that outputs feature quantities of the time-series data, acquiring learning data that associates the time-series data related to substrate processing with a plurality of correct answer values ​​related to processing performed based on the feature quantities of the time-series data, inputting the acquired time-series data for learning data into the learning model to acquire feature quantities output by the learning model, calculating a plurality of errors based on the acquired feature quantities, the correct answer values ​​of the learning data, and a plurality of loss functions, calculating an overall error based on the calculated plurality of errors, and updating parameters of the learning model based on the calculated overall error.

[0006] According to the present disclosure, it is expected that search for time-series data related to substrate processing can be realized.

[0007] FIG. 1 is a schematic diagram for explaining an overview of an information processing system according to the present embodiment. FIG. 2 is a block diagram for explaining an example of a configuration of an information processing device according to the present embodiment. FIG. 3 is a schematic diagram for explaining an example of a configuration of a substrate processing DB. FIG. 4 is a schematic diagram for explaining an example of a configuration of a learning model according to the present embodiment. FIG. 5 is a schematic diagram for explaining a machine learning method for a learning model. FIG. 6 is a flowchart showing an example of a procedure for a learning model generation process performed by the information processing device according to the present embodiment. FIG. 7 is a flowchart showing an example of a procedure for a similarity search process performed by the information processing device according to the present embodiment. FIG. 8 is a schematic diagram showing an example of a selection screen for query time series data displayed by a terminal device. FIG. 9 is a schematic diagram showing an example of a selection screen for query time series data displayed by a terminal device. FIG. 10 is a schematic diagram showing an example of a similarity search result display screen. FIG. 11 is a schematic diagram showing another example of information input / output related to similarity search.

[0008] Specific examples of information processing systems according to embodiments of the present disclosure will be described below with reference to the drawings. Note that the present disclosure is not limited to these examples, but is defined by the claims, and is intended to include all modifications within the meaning and scope of the claims.

[0009] <System Overview> FIG. 1 is a schematic diagram illustrating an overview of an information processing system according to this embodiment. The information processing system according to this embodiment is configured to include an information processing apparatus 1, a substrate processing apparatus 3, and a terminal apparatus 5. The substrate processing apparatus 3 is an apparatus that performs various substrate processes, such as CVD (Chemical Vapor Deposition), sputtering, or etching, on semiconductor substrates (wafers). In this embodiment, the substrate processing apparatus 3 includes one or more sensors that measure, for example, the temperature or pressure within a chamber that accommodates the substrate to be processed. The substrate processing apparatus 3 repeatedly acquires (samples) data on the temperature or pressure measured by these sensors during substrate processing, such as etching, at predetermined intervals to acquire time-series data indicating time-series changes in the temperature or pressure values, and transmits the data to the information processing apparatus 1. Note that the exchange of time-series data between the substrate processing apparatus 3 and the information processing apparatus 1 may be performed via a recording medium, such as a memory card or an optical disk, rather than via wired or wireless communication.

[0010] The information processing device 1 is a device that stores and manages time-series data obtained from the substrate processing device 3 in a substrate processing DB (database) 7. To this end, the information processing device 1 performs machine learning using the time-series data obtained from the substrate processing device 3 to generate in advance a learning model 9 that converts the time-series data into feature quantities. The information processing device 1 uses the previously generated learning model 9 to convert the time-series data obtained from the substrate processing device 3 into feature quantities, associates the feature quantities with the time-series data, and stores the feature quantities in the substrate processing DB 7. Note that in this embodiment, the information processing device 1 performs machine learning processing to generate the learning model 9, but this is not limited thereto. A device other than the information processing device 1 may perform machine learning processing to generate the learning model 9 in advance, and the generated learning model 9 may be acquired and used by the information processing device 1.

[0011] The information processing device 1 also performs a similarity search process to search the substrate processing DB 7 for time series data similar to time series data provided by the user. The information processing device 1 accepts a search request from the user via the terminal device 5. In this embodiment, the search request accepted from the user includes at least the time series data to be searched, and may further include information specifying what aspects the search for similar data should focus on. The information processing device 1 uses the learning model 9 to convert the time series data included in the search request acquired from the terminal device 5 into feature quantities, calculates similarities between these feature quantities and the feature quantities of multiple time series data stored in the substrate processing DB 7, and transmits information on one or more time series data with high similarity to the terminal device 5 as similar cases.

[0012] In the information processing system according to this embodiment, a plurality of loss functions are combined to perform machine learning processing of the learning model 9, thereby improving the accuracy of the similarity search processing performed using the learning model 9. Furthermore, in the information processing system according to this embodiment, for the learning data (teacher data) required for the machine learning processing, semi-supervised learning is performed using data to which a designer or developer or the like has assigned a correct answer label (correct answer value) and data to which the information processing device 1 has assigned a pseudo label to data not assigned a correct answer label, thereby increasing the amount of learning data used in the machine learning and improving the accuracy of the learning model 9.

[0013] <Device Configuration> Fig. 2 is a block diagram showing an example configuration of an information processing device 1 according to this embodiment. The information processing device 1 according to this embodiment can be realized by installing a predetermined application program or the like in a general-purpose information processing device such as a personal computer or a server computer. The information processing device 1 according to this embodiment is configured to include a processing unit 11, a storage unit 12, a communication unit 13, etc. Note that in this embodiment, the processing will be described as being performed by a single information processing device 1, but the processing of the information processing device 1 may be distributed among multiple devices.

[0014] The processing unit 11 is configured using an arithmetic processing device such as a CPU (Central Processing Unit), an MPU (Micro-Processing Unit), a GPU (Graphics Processing Unit) or a quantum processor, a ROM (Read Only Memory), a RAM (Random Access Memory), etc. The processing unit 11 reads and executes a program 12a stored in the storage unit 12 to perform various processes, such as generating a learning model 9 by machine learning, which converts time-series data obtained from the substrate processing apparatus 3 into feature quantities, and performing similarity search processing of time-series data using the generated learning model 9.

[0015] The storage unit 12 is configured using a large-capacity storage device such as a hard disk or an SSD (Solid State Drive). The storage unit 12 stores various programs executed by the processing unit 11 and various data required for the processing of the processing unit 11. In the present embodiment, the storage unit 12 stores a program 12a executed by the processing unit 11. The storage unit 12 also includes a model information storage unit 12b that stores information related to the learning model 9 generated by the information processing device 1, and a substrate processing DB 7 that stores information such as time-series data related to substrate processing and feature quantities converted from the data.

[0016] In this embodiment, the program (computer program, program product) 12a is provided in a form recorded on a recording medium 99 such as a memory card or an optical disc, and the information processing device 1 reads the program 12a from the recording medium 99 and stores it in the storage unit 12. However, the program 12a may also be written to the storage unit 12, for example, during the manufacturing stage of the information processing device 1. Alternatively, the program 12a may be distributed by a remote server device or the like and acquired by the information processing device 1 via communication. For example, the program 12a may be read from the recording medium 99 by a writing device and written to the storage unit 12 of the information processing device 1. The program 12a may be provided in a form distributed via a network or in a form recorded on the recording medium 99.

[0017] The model information storage unit 12b stores information about a learning model that has undergone machine learning. The information about the learning model may include, for example, information indicating the configuration of the learning model and information such as values ​​of internal parameters determined by machine learning. In the present embodiment, the model information storage unit 12b stores information about a learning model 9 that converts time-series data into features. The learning model 9 according to the present embodiment is a learning model that has undergone machine learning to receive as input time-series data in a one-dimensional array in which M (M is a natural number) values ​​obtained by periodically sampling measurement results from a sensor of the substrate processing apparatus 3 are arranged in chronological order, and output an N-dimensional (N is a natural number) feature vector. The information processing apparatus 1 generates the learning model 9 by performing machine learning on the time-series data acquired from the substrate processing apparatus 3 using learning data to which correct answer labels have been assigned, and stores information about the generated learning model 9 in the model information storage unit 12b.

[0018] Furthermore, in the present embodiment, the information processing device 1 can use a large language model (LLM) that has undergone machine learning in advance for inputting and outputting information to and from a user of the terminal device 5. In this case, the model information storage unit 12b stores information such as the configuration and internal parameters of the large language model. For example, the large language model may be a learning model such as a Transformer, a Bidirectional Encoder Representations from Transformers (BERT), or a Generative Pre-trained Transformer (GPT), which is a large-scale neural network equipped with an attention mechanism. The large language model used by the information processing device 1 may be a widely available, general-purpose large language model, or may be one that has been trained with information related to substrate processing or the like through fine tuning.

[0019] In this embodiment, information about the learning model is stored in the information processing device 1, and processing using the learning model is performed by the information processing device 1, but this is not limited to this. Information about the learning model may be stored in a device different from the information processing device 1, and this device may perform processing using the learning model, and the information processing device 1 may acquire the processing results from this device. Furthermore, machine learning processing of the learning model may be performed by the information processing device 1 or by a device different from the information processing device 1.

[0020] 3 is a schematic diagram illustrating an example of the configuration of the substrate processing DB 7. The substrate processing DB 7 included in the information processing apparatus 1 according to this embodiment is a database that stores information such as an "equipment ID," a "wafer ID," a "sensor ID," a "timestamp," "time-series data," a "feature amount," and a "document" in association with each other. The "equipment ID" is identification information uniquely assigned to the substrate processing apparatus 3, and stores information such as "equipment A11" and "equipment B23." The "wafer ID" is identification information uniquely assigned to a substrate (wafer) that has undergone substrate processing in the substrate processing apparatus 3, and stores information such as "wafer a03" and "wafer b11." The "sensor ID" is identification information uniquely assigned to a sensor included in the substrate processing apparatus 3, and stores information such as "temperature sensor s4" and "pressure sensor s9." The "timestamp" is information such as the date and time when substrate processing was performed by the substrate processing apparatus 3 or the date and time when the information processing apparatus 1 acquired time-series data. For example, information such as "2024 / 3 / 15, 10:07" and "2024 / 3 / 16, 14:26" is stored. The "time-series data" is time-series data acquired from the substrate processing apparatus 3. For example, one-dimensional array data in which M measurement values ​​from a sensor are arranged in chronological order is stored. The "feature" is time-series data converted into a feature vector by the learning model 9. For example, N-dimensional vector data is stored. The "document" is information such as a sentence written by an administrator or operator of the substrate processing apparatus 3 regarding the results of substrate processing. The information processing apparatus 1 acquires information such as the "equipment ID," "wafer ID," "sensor ID," "timestamp," and "document" from the substrate processing apparatus 3 along with the "time-series data." The information processing apparatus 1 converts the "time-series data" into "feature" using the learning model 9, associates them, and stores them in the substrate processing DB 7.

[0021] The communication unit 13 exchanges data between the substrate processing apparatus 3 and the terminal device 5 via a network N such as the Internet or a LAN (Local Area Network). In this embodiment, the communication unit 13 receives time-series data transmitted from the substrate processing apparatus 3 and provides the data to the processing unit 11. The communication unit 13 also provides information such as a search request transmitted from the terminal device 5 to the processing unit 11, and transmits information such as the results of a similar search transmitted from the processing unit 11 to the terminal device 5.

[0022] The storage unit 12 may be an external storage device connected to the information processing device 1. The information processing device 1 may also be a multi-computer including multiple computers, or may be a virtual machine virtually constructed by software. The information processing device 1 is not limited to the above configuration, and may also include, for example, a reading unit that reads information stored in a portable storage medium, an input unit that accepts operation input, or a display unit that displays images.

[0023] In the information processing device 1 according to this embodiment, the processing unit 11 reads and executes the program 12a stored in the storage unit 12, thereby realizing a time-series data acquisition unit 11a, a learning processing unit 11b, a search processing unit 11c, an input / output processing unit 11d, and the like as software functional units in the processing unit 11. Note that in the figure, functional units related to the process of generating a learning model 9 and the process of similarity search of time-series data are shown as functional units of the processing unit 11, and functional units related to other processes are not shown.

[0024] The time-series data acquisition unit 11a communicates with the substrate processing apparatus 3 via the communication unit 13 to acquire time-series data including measurements such as temperature and pressure repeatedly measured by sensors or the like when the substrate processing apparatus 3 performs substrate processing. The substrate processing apparatus 3 stores sensor measurement data in a memory or the like by repeatedly sampling signals output by the sensors during substrate processing at a predetermined cycle. The substrate processing apparatus 3 transmits the time-series data stored in the memory or the like to the information processing apparatus 1, together with information such as the apparatus ID of the substrate processing apparatus 3 and the substrate ID of the substrate that has undergone substrate processing, at an appropriate timing, for example, after the substrate processing is completed. The time-series data acquisition unit 11a of the information processing apparatus 1 acquires the time-series data transmitted from the substrate processing apparatus 3 after the substrate processing is completed, associates the acquired time-series data with information such as the apparatus ID and substrate ID, and stores the associated data in the substrate processing DB 7. The timing and method, etc., of the time-series data acquisition unit 11a acquiring the time-series data from the substrate processing apparatus 3 may be determined appropriately by the designer of the information processing system.

[0025] If a learning model 9 that converts time series data into features has already been generated, the time series data acquisition unit 11a converts the time series data acquired from the substrate processing apparatus 3 into features and stores the features in the substrate processing DB 7. If a learning model 9 has not yet been generated, the time series data acquisition unit 11a stores the time series data acquired from the substrate processing apparatus 3 in the substrate processing DB 7 without converting it into features, and the features corresponding to this time series data do not need to be stored in the substrate processing DB 7. Time series data for which features are not stored in the substrate processing DB 7 is used in machine learning processing as learning data for the learning model 9 and is discarded after the machine learning is completed, or is converted into features using the learning model 9 generated by machine learning and stored in the substrate processing DB 7. Time series data for which features are stored in the substrate processing DB 7 is used for similarity searches.

[0026] The learning processing unit 11b performs machine learning processing to generate a learning model 9 using the time-series data stored in the substrate processing DB 7. The time-series data used in the machine learning is subjected to a process known as annotation, in which a designer or the like assigns correct labels in advance. In this embodiment, the learning processing unit 11b performs machine learning using multiple loss functions, and therefore multiple correct labels are assigned to one piece of time-series data. The correct labels assigned to the time-series data may include correct labels that classify the state of the substrate processing, such as whether the substrate processing related to the time-series data was completed normally or an abnormality occurred, and, if an abnormality occurred, what type of abnormality it was. The correct labels may also include correct labels that classify the waveform of the time-series data in which an abnormality occurred. The correct labels may also include correct labels (correct values) that serve as prediction results for numerical values ​​related to the substrate processing, such as a numerical value indicating the time timing of the abnormality or a numerical value indicating the severity of the abnormality in the time-series data in which an abnormality occurred. The learning processing unit 11b accepts input of correct labels in advance from a designer or the like for some or all of the time-series data stored in the substrate processing DB 7, and stores the input correct labels in association with the time-series data in the substrate processing DB 7. Note that in this embodiment, since the learning processing unit 11b performs semi-supervised learning, it is not necessary for correct labels to be assigned to all of the time-series data used in machine learning.

[0027] The learning processing unit 11b first performs machine learning processing using time-series data to which correct labels have been assigned. The learning processing unit 11b calculates two errors using two loss functions: one that calculates a center error (center loss) for the correct labels that classify the time-series data, and the other that calculates a cross-entropy error (cross-entropy loss). The learning processing unit 11b calculates one error using a loss function that calculates a mean squared error (MSE) for the correct labels of the values ​​associated with the time-series data. The learning processing unit 11b calculates one or more errors for one correct label assigned to the time-series data, and then weights and sums the multiple errors calculated for the multiple correct labels to obtain a total error. The learning processing unit 11b performs machine learning processing by updating the internal parameters of the learning model 9 using a method such as backpropagation based on the calculated total error.

[0028] After machine learning using time-series data with correct labels has progressed to a certain extent, for example, after the error in the values ​​output by the learning model 9 has become smaller than a predetermined threshold, the learning processing unit 11b performs a process of assigning pseudo-correct labels to time-series data without correct labels. Note that the pseudo labels assigned by the learning processing unit 11b to the time-series data are correct labels related to the classification of the time-series data, and pseudo labels related to numerical prediction are not assigned. The learning processing unit 11b inputs the time-series data without correct labels to the learning model 9 in the middle of learning and obtains the classification result output by the learning model 9. The learning processing unit 11b considers the classification result with the highest confidence as the correct answer, generates a pseudo label with this as the correct answer label, and assigns it to the time-series data.

[0029] The learning processing unit 11b performs machine learning processing on the time-series data with pseudo labels, treating them in the same way as time-series data with correct labels assigned by a designer or the like, and updates the internal parameters of the learning model 9. However, because the pseudo labels are related to classification, the learning processing unit 11b calculates the center error and cross-entropy error for the time-series data with pseudo labels, but does not calculate the mean squared error (for example, the mean squared error is considered to be 0). After updating the internal parameters of the learning model 9, the learning processing unit 11b may update the pseudo labels assigned to the time-series data using the updated learning model 9. The pseudo labels may be updated at predetermined intervals, for example, once when the number of updates of the internal parameters of the learning model 9 reaches a predetermined number. Whether or not to update the pseudo labels, the frequency of updates, etc. can be appropriately determined by the designer or the like of the information processing system according to this embodiment.

[0030] The learning processing unit 11b repeats the above machine learning process until the number of updates to the internal parameters of the learning model 9 exceeds a threshold value or until the error in the value output by the learning model 9 becomes smaller than a threshold value. The learning processing unit 11b stores the final values ​​of the internal parameters of the learning model 9 updated by machine learning in the model information storage unit 12b.

[0031] The search processing unit 11c performs various processes related to similarity searches of the time-series data stored in the substrate processing DB 7. For time-series data that the time-series data acquisition unit 11a has acquired from the substrate processing apparatus 3 and stored in the substrate processing DB 7 and that does not have features stored therein, the search processing unit 11c converts the time-series data into features using the trained learning model 9, the information of which is stored in the model information storage unit 12b, and stores the converted features in the substrate processing DB 7 in association with the original time-series data.

[0032] In the information processing system according to this embodiment, a user can provide time-series data, for example, indicating the presence or absence of an abnormality or the type of abnormality, to the information processing device 1 (hereinafter, the time-series data provided by the user is referred to as query time-series data), and search for time-series data similar to the query time-series data among the time-series data stored in the substrate processing DB 7. The search processing unit 11c acquires the query time-series data in response to a search request from the user provided via the terminal device 5 and converts the acquired query time-series data into features using the trained learning model 9. The search processing unit 11c compares the features of the query time-series data with the features of the time-series data stored in the substrate processing DB 7 and calculates, for example, the Euclidean distance between the two features in a feature space. The smaller the Euclidean distance between two feature values, the higher the similarity between the two corresponding time-series data. The search processing unit 11c can extract, from the substrate processing DB 7, the time-series data with the smallest Euclidean distance, or a predetermined number of time-series data in ascending order of Euclidean distance, as the result of the similarity search. The search processing unit 11c may determine whether or not the time-series data are similar based on criteria such as Manhattan distance, Chebyshev distance, or cosine similarity of the feature amounts, instead of based on the Euclidean distance of the feature amounts.

[0033] Furthermore, in the information processing system according to this embodiment, the user can specify as search conditions what items or features, etc. to focus on in the similarity search to determine similarity. As described above, in the information processing system according to this embodiment, machine learning is performed using multiple loss functions, and the learning data used in the machine learning is time-series data to which multiple types of correct answer labels have been attached. The types of correct answer labels that can be used include, for example, correct answer labels that classify the state of substrate processing, and numerical correct answer labels that indicate the timing of an abnormality occurring in substrate processing. The items that the user can specify as search conditions in the similarity search correspond to the types of correct answer labels attached to this time-series data.

[0034] For example, the learning processing unit 11b can perform machine learning using a technique called DRL (Disentangled Representation Learning) to assign meaning to each element of an N-dimensional vector output as a feature by the learning model 9, and the N elements included in the feature can be assigned meaning as, for example, elements useful for classifying the state of substrate processing or elements useful for predicting the timing of an abnormality occurrence. The search processing unit 11c can realize a similarity search according to the search conditions by extracting feature elements according to the specified search conditions and calculating Euclidean distance, etc., as described above to determine similarity. Note that machine learning using the DRL technique is an existing technology, so a detailed description thereof will be omitted.

[0035] The input / output processing unit 11d performs processing related to the input and output of information with the user by communicating with the terminal device 5 via the communication unit 13. The input / output processing unit 11d, for example, displays an appropriate input screen on the display unit of the terminal device 5 and accepts input operations by the user, such as query time-series data or search conditions. The input / output processing unit 11d also causes the terminal device 5 to display (output) the results of the similarity search, for example, by transmitting (outputting) information related to the results of the similarity search to the terminal device 5.

[0036] <Learning Model Generation Process> FIG. 4 is a schematic diagram illustrating an example of the configuration of the learning model 9 according to this embodiment. The learning model 9 according to this embodiment is, for example, a learning model employing a neural network configuration. The learning model 9 accepts input of a one-dimensional array of time-series data and outputs a vector of features. The learning model 9 includes multiple layers (11 layers in the illustrated example), each of which performs a predetermined calculation. The input time-series data is converted into features by being subjected to calculation processing in each layer in turn. In this example, the time-series data input to the learning model 9 is converted into features by being subjected to calculation processing in the following order: a one-dimensional convolutional layer (Conv1d), an activation function (ReLU), a pooling layer (Pool), a one-dimensional convolutional layer (Conv1d), an activation function (ReLU), a pooling layer (Pool), a one-dimensional convolutional layer (Conv1d), an activation function (ReLU), a pooling layer (Pool), an activation function (ReLU), and a fully connected layer (FC). The layer structure of the learning model 9 shown in the figure is an example and is not limited to this.

[0037] 5 is a schematic diagram illustrating a machine learning method for the learning model 9. The information processing device 1 according to this embodiment uses multiple loss functions to calculate multiple errors based on feature quantities output by the learning model 9. In this example, there are three types of loss functions: a loss function used by the information processing device 1, a loss function that calculates a center error from the feature quantities output by the learning model 9, a loss function that calculates a cross-entropy error, and a loss function that calculates a mean squared error. The center error and cross-entropy error are errors calculated based on correct labels that classify the state of substrate processing and are assigned to the time-series data. The mean squared error is an error calculated based on correct labels that relate to the timing of an abnormality occurrence and are assigned to the time-series data.

[0038] The center error is the error between the feature output by the learning model 9 and the center of the feature classified into the corresponding class in the feature space. The center of the feature for each class is updated at appropriate times during the machine learning process. By using the center error, the learning model 9 can be trained so that the distance between features belonging to the same class is shortened.

[0039] The information processing device 1 classifies the substrate processing state using a plurality of fully connected layers (FC) and a softmax function (Softmax) of neural networks based on the feature quantities output by the learning model 9. The information processing device 1 acquires the classification results output by the softmax function, i.e., the probability of belonging to each class, and calculates the cross-entropy error between the acquired probability of belonging and the correct label for classification assigned to the time-series data.

[0040] Furthermore, the information processing device 1 predicts the timing of an abnormality occurring in the substrate processing by using a plurality of fully connected layers (FC) of neural networks based on the feature amounts output by the learning model 9. The information processing device 1 calculates the mean square error between the predicted value of the timing of an abnormality occurring calculated using the fully connected layers (FC) based on the feature amounts and the correct label of the timing of an abnormality occurring assigned to the time-series data.

[0041] The information processing device 1 calculates an overall error based on the calculated center error, cross entropy error, and mean square error. In this example, the information processing device 1 calculates the total error by adding the center error weighted with a weight λ0 and the mean square error weighted with a weight λ1 to the cross entropy error. The information processing device 1 updates the internal parameters of the learning model 9 using the backpropagation algorithm based on the calculated overall error. At this time, the information processing device 1 may also update the internal parameters of the fully connected layer (FC) for calculating the cross entropy error and mean square error. The information processing device 1 performs machine learning on the learning model 9 by repeatedly calculating the overall error and updating the internal parameters using multiple time series data and correct labels, thereby generating the learning model 9.

[0042] In this example, the information processing device 1 calculates three errors, namely, the center error, the cross-entropy error, and the mean square error, to calculate the overall error. However, this is not limited to this, and four or more errors may be calculated to calculate the overall error. Instead of or in addition to the correct labels classifying the state of substrate processing, correct labels classifying various factors, such as correct labels classifying the quality of processed substrates, may be attached to the time-series data, and the information processing device 1 may calculate the center error and the cross-entropy error based on these correct labels. Instead of or in addition to the correct labels relating to the numerical values ​​of the timing of an abnormality occurrence, correct labels relating to the duration of the abnormality that occurred, the number of times the abnormality occurred, the maximum, minimum, or average value of pressure or temperature given as time-series data, or measurements of the characteristics of processed substrates may be attached to the time-series data, and the information processing device 1 may calculate the mean square error based on these correct labels. For example, the values ​​of the substrate processing settings (recipe) when the time-series data was obtained may be used as the correct label, or various other information may be attached to the time-series data as the correct label. The information processing device 1 can calculate a variety of errors, weight and add the errors to calculate an overall error, and update the internal parameters of the learning model 9 based on the overall error. The information processing device 1 can also calculate various other errors using an appropriate loss function, in addition to the center error, cross entropy error, and mean square error.

[0043] Furthermore, the multiple errors calculated by the information processing device 1 do not need to be a combination of different types of errors, such as a combination of cross-entropy error and mean squared error. The information processing device 1 may, for example, calculate multiple cross-entropy errors for classification items of different content, or may calculate multiple mean squared errors for predictions of different numerical values. The information processing device 1 may calculate a total error based on a combination of multiple errors of the same type. The information processing device 1 according to this embodiment is expected to improve the accuracy of machine learning by calculating a total error by focusing on multiple errors, regardless of whether they are of the same or different types, and performing machine learning based on the total error.

[0044] In the information processing system according to the present embodiment, multiple pieces of time-series data measured during substrate processing performed with various settings (recipes) in the substrate processing apparatus 3 are collected in advance as time-series data to be used in machine learning processing to generate the learning model 9. Furthermore, in the information processing system, multiple pieces of time-series data measured during substrate processing performed with different substrate processing apparatuses 3 are collected in advance. By performing machine learning using time-series data from various settings or apparatuses, the learning model 9 generated by machine learning can output feature quantities corresponding to various settings. Furthermore, the learning model 9 may be configured to accept, as input, not only the time-series data but also setting values ​​related to the substrate processing. The previously collected time-series data is assigned multiple correct answer labels in advance by a designer of the information processing system, and the information processing apparatus 1 performs machine learning processing of the learning model 9 using a set of the time-series data and the multiple correct answer labels as learning data.

[0045] However, in the information processing system according to this embodiment, it is not necessary for a designer or the like to assign correct labels to all of the multiple time-series data collected in advance (correct labels may be assigned to all of the time-series data). The information processing device 1 according to this embodiment repeatedly calculates errors and updates internal parameters using time-series data with correct labels in the early stages of machine learning of the learning model 9. For example, when the machine learning process progresses and the error (total error) of the learning model 9 becomes smaller than a threshold, the information processing device 1 performs a process of assigning pseudo labels to time-series data that do not have correct labels using the learning model 9 in the middle of learning.

[0046] The information processing device 1 inputs time-series data without a correct label to a learning model 9 in the middle of learning, acquires features output by the learning model 9, and acquires information on the attribution probability that will be the classification result based on the acquired features using a fully connected layer (FC) and a softmax function (Softmax) shown in FIG. 5 . The information processing device 1 regards the acquired attribution probability with the highest probability as the correct answer, generates a pseudo-label with this as the correct label, and assigns it to the time-series data. For example, if the state of substrate processing is classified into three states (state 1, state 2, and state 3), and the target time-series data is classified into state 2, a designer or the like assigns a correct label (0, 1, 0) to this time-series data. Furthermore, for example, if the attribute probability output by the fully connected layer (FC) and softmax function (Softmax) shown in FIG. 5 based on the feature converted from the time series data by the learning model 9 is (0.1, 0.7, 0.2), the information processing device 1 regards the one with the largest attribute probability as the classification result and assigns a pseudo label of (0, 1, 0) to the time series data.

[0047] The pseudo labels assigned to the time-series data by the information processing device 1 are correct labels for classifying the time-series data, such as the state of substrate processing, and do not include correct labels for predicting numerical values ​​such as the timing of abnormality occurrence.

[0048] The information processing device 1 performs machine learning of the learning model 9 using time-series data to which correct labels have been assigned by a designer or the like and time-series data to which the above-mentioned pseudo labels have been assigned in a similar procedure. However, because the pseudo labels are only related to classification, the information processing device 1 does not calculate the mean squared error for time-series data to which pseudo labels have been assigned. Instead, it calculates the center error and cross-entropy error based on the pseudo labels, and then calculates the overall error based on the center error and cross-entropy error. Thereafter, the information processing device 1 proceeds with the machine learning process while updating the pseudo labels at appropriate times.

[0049] 6 is a flowchart showing an example of the procedure of a learning model generation process performed by the information processing device 1 according to this embodiment. The time-series data acquisition unit 11a of the processing unit 11 of the information processing device 1 according to this embodiment acquires pairs of time-series data and correct labels stored in advance as learning data, for example, by reading them from the storage unit 12 (step S1). Note that the time-series data acquisition unit 11a reads multiple pairs of time-series data and correct labels in step S1, and subsequent processes (steps S2 to S4) are performed on each of these multiple pairs of time-series data and correct labels.

[0050] The learning processing unit 11b of the processing unit 11 inputs the time-series data acquired in step S1 to the learning model 9, acquires features output by the learning model 9, and calculates multiple errors, such as center error, cross entropy error, and mean square error, using a predetermined loss function based on the acquired features and the correct label acquired in step S1 (step S2). The learning processing unit 11b calculates an overall error by weighting each of the multiple errors calculated in step S2 and calculating a total value (step S3). The learning processing unit 11b updates the internal parameters of the learning model 9, for example, by backpropagation based on the overall error calculated in step S3 (step S4).

[0051] The learning processing unit 11b determines whether a condition for performing semi-supervised learning using unlabeled time-series data is met (step S5). The condition for performing semi-supervised learning may be, for example, when the number of repetitions of supervised learning in steps S1 to S4 exceeds a predetermined number, or when the overall error calculated in step S3 is smaller than a predetermined threshold. This condition is set in advance by a designer of the information processing system according to this embodiment. If the condition is not met (S5: NO), the learning processing unit 11b returns to step S1 and repeatedly calculates the overall error using labeled time-series data and updates the internal parameters.

[0052] If the conditions for semi-supervised learning are met (S5: YES), the time-series data acquisition unit 11a acquires time-series data without labels stored in advance as learning data, for example by reading it from the storage unit 12 (step S6). The time-series data acquisition unit 11a reads multiple pieces of time-series data without labels in step S6. The learning processing unit 11b inputs the time-series data read in step S6 to the learning model 9, acquires features output by the learning model 9, and performs classification based on the acquired features to generate pseudo labels for classification of the time-series data (step S7).

[0053] The learning processing unit 11b inputs the time-series data acquired in step S6 to the learning model 9, acquires features output by the learning model 9, and calculates multiple errors, such as a center error and a cross-entropy error, using a predetermined loss function based on the acquired features and the pseudo labels generated in step S7 (step S8). The learning processing unit 11b calculates an overall error by weighting each of the multiple errors calculated in step S8 and calculating a total value (step S9). The learning processing unit 11b updates the internal parameters of the learning model 9 based on the overall error calculated in step S9, for example, by backpropagation (step S10).

[0054] The learning processing unit 11b determines whether the total error calculated in step S9 is smaller than a predetermined threshold (step S11). The threshold used in step S11 is set in advance by the designer of the information processing system according to this embodiment. If the total error is larger than the threshold (S11: NO), the learning processing unit 11b returns to step S1 and repeatedly calculates the total error and updates the internal parameters. If the total error is smaller than the threshold (S11: YES), the learning processing unit 11b stores information such as the internal parameters of the learning model 9 in the model information storage unit 12b (step S12) and ends the process.

[0055] <Similar Search Processing> In the information processing system according to the present embodiment, the information processing device 1 converts time-series data of substrate processing acquired from the substrate processing device 3 into feature quantities using the learning model 9 generated in advance by the above-described machine learning processing, and stores and accumulates the time-series data and feature quantities in the substrate processing DB 7 while associating them with each other. The information processing device 1 also associates information such as various IDs, timestamps, and documents with the time-series data and feature quantities and stores them in the substrate processing DB 7. After a sufficient amount of data has been accumulated in the substrate processing DB 7, the information processing device 1 provides a service of searching for similar cases from the substrate processing DB 7 in response to a user request.

[0056] For example, when a user acquires new time series data measured when substrate processing is performed in the substrate processing apparatus 3, the user can search for past cases similar to this new time series data from the substrate processing DB 7. The user transmits the new time series data, which is the basis for the similarity search, as query time series data from the terminal device 5 to the information processing apparatus 1 together with a search request. Note that the query time series data may be provided to the information processing apparatus 1, for example, by the user uploading a time series data file from the terminal device 5, or may be provided by the user selecting the query time series data from time series data newly acquired by the information processing apparatus 1 from the substrate processing apparatus 3, or may be provided by the user selecting the query time series data from time series data stored in the substrate processing DB 7, for example.

[0057] The information processing device 1, which has received a search request from the terminal device 5, converts the query time-series data provided together with the search request into feature quantities using the trained learning model 9. The information processing device 1 compares the feature quantities of the query time-series data with the feature quantities of each piece of time-series data stored in the substrate processing DB 7, and calculates the Euclidean distance between the two feature quantities. The information processing device 1 determines that the smaller the Euclidean distance, the higher the similarity between the two pieces of time-series data. For example, the information processing device 1 extracts a predetermined number of pieces of time-series data with high similarity from the top, and transmits information about the extracted time-series data to the terminal device 5 as a similarity search result.

[0058] 7 is a flowchart showing an example of the procedure of a similarity search process performed by the information processing device 1 according to the present embodiment. The search processing unit 11c of the processing unit 11 of the information processing device 1 according to the present embodiment acquires query time-series data that is the basis of the similarity search (step S31). Note that the query time-series data can be acquired by, for example, accepting an upload of a file or the like storing the query time-series data from a user, or by reading the query time-series data based on a user selection from a device or the like that stores time-series data. The search processing unit 11c inputs the query time-series data acquired in step S31 to the trained learning model 9 and acquires features output by the learning model 9, thereby converting the query time-series data into features (step S32).

[0059] The search processor 11c calculates the Euclidean distance between the feature quantities of the time-series data converted in step S32 and the feature quantities of the plurality of time-series data stored in the substrate processing DB 7 (step S33). The search processor 11c compares the plurality of Euclidean distances calculated in step S33 and extracts a predetermined number of time-series data having a small Euclidean distance from the substrate processing DB 7 (step S34). The search processor 11c transmits information about the predetermined number of time-series data calculated in step S34 to the terminal device 5 as a similarity search result (step S35), and ends the process.

[0060] Furthermore, the information processing device 1 according to this embodiment may receive search conditions from the user regarding what items or features, etc., to focus on in the similarity search. In this embodiment, for example, as shown in FIG. 5 , if the learning model 9 performs machine learning based on the total error of the center error and cross entropy error using the correct answer label for the classification of "substrate processing state" and the mean square error using the correct answer label for "timing of abnormality occurrence," the user can specify either "substrate processing state" or "timing of abnormality occurrence" as the search condition. In other words, the user can specify search conditions associated with the types of correct answer labels used in the machine learning and issue a similarity search request to the information processing device 1.

[0061] When the information processing device 1 accepts such search conditions in the similarity search process, it performs machine learning of the learning model using a technique called DRL. This allows the information processing device 1 to assign meaning to each element of the feature vector output by the learning model according to the type of correct label. The information processing device 1 extracts corresponding elements from the feature values ​​output by the learning model 9 based on the type of correct label specified as a search condition by the user. The information processing device 1 can extract time-series data similar to the query time-series data focusing on the search conditions based on the Euclidean distance between the components extracted from the feature values ​​of the query time-series data and the components extracted from each time-series data stored in the substrate processing DB 7.

[0062] In this embodiment, "substrate processing state" and "timing of abnormality occurrence" are listed as search conditions, but the search conditions are not limited to these. By performing machine learning of the learning model 9 by attaching various correct answer labels to the time-series data, various search conditions can be used in similarity search. For example, by attaching values ​​or classifications related to the substrate processing settings (recipe) for the substrate processing apparatus 3 as correct answer labels to the time-series data, a similarity search can be performed using the substrate processing settings as search conditions. Furthermore, by attaching the results of the substrate processing (such as the results of the substrate processing, measurement indexes, pass / fail status, or yield) as correct answer labels to the time-series data, a similarity search can be performed using the results of the substrate processing as search conditions.

[0063] <Information Input / Output Processing> In the information processing system according to the present embodiment, a user can use the terminal device 5 to perform tasks such as issuing a search request to the information processing device 1 or obtaining and viewing search results from the information processing device 1. In order to accept various pieces of information input by the user, the information processing device 1 transmits information necessary for display to the terminal device 5, thereby displaying a screen for information input on the display unit of the terminal device 5. The information processing device 1 also transmits the results of the similarity search process performed in response to the user's search request, i.e., information on time-series data similar to the query time-series data specified by the user, to the terminal device 5, and displays the results of the similarity search on the display unit of the terminal device 5.

[0064] 8 and 9 are schematic diagrams showing an example of a query time-series data selection screen displayed by the terminal device 5. In the information processing system according to this embodiment, for example, time-series data obtained during substrate processing is stored in the storage unit of the substrate processing apparatus 3. The information processing apparatus 1 searches for time-series data that meets user-specified conditions from among a plurality of time-series data that are not stored in the substrate processing DB 7 but are stored in the storage unit of the substrate processing apparatus 3, displays the searched time-series data as a list, and accepts selection of time-series data to be used as query time-series data from the list. Alternatively, for example, the information processing apparatus 1 may search for time-series data that meets user-specified conditions from among a plurality of time-series data stored in the substrate processing DB 7, displays the searched time-series data as a list, and accept selection of time-series data to be used as query time-series data from the list. To accept such selection of query time-series data, the information processing system displays a query time-series data selection screen shown in FIG. 8 on the display unit of the terminal device 5.

[0065] The upper part of the query time-series data selection screen shown in FIG. 8 includes a query condition input area 51 in which multiple input boxes are arranged for inputting information related to the query time-series data, such as a "period," "device name," and a "sensor name." A candidate search execution button labeled "Execute candidate search" is provided in the lower right portion of the query condition input area 51. The user can appropriately input conditions for time-series data that are candidates for the query time-series data into the input boxes and click or touch the candidate search execution button to display a list of time-series data that can be candidates for the query time-series data. When the candidate search execution button is operated, the terminal device 5 acquires the conditions entered in the input boxes of the query condition input area 51 and transmits them to the information processing device 1. The information processing device 1, which has received these conditions from the terminal device 5, extracts one or more time-series data that match the provided conditions from multiple time-series data stored in the storage unit of the substrate processing device 3 or its own substrate processing DB 7, and transmits information about the extracted time-series data (which may not include the time-series data itself) to the terminal device 5.

[0066] 8 , a query candidate display area 52 is provided at the bottom. The query candidate display area 52 displays information about multiple pieces of time-series data, such as "wafer ID," "device name," "sensor name," and "timestamp," in a table format. A selection button labeled "Select" is provided at the bottom right of the query candidate display area 52. Based on the information transmitted from the information processing device 1, the terminal device 5 displays a list of information about time-series data that matches the conditions input in the query condition input area 51 in the query candidate display area 52. The user can select one or more pieces of time-series data from the list and operate the selection button to select query time-series data. When the selection button is operated, the terminal device 5 notifies the information processing device 1 of the selected time-series data. In response, the information processing device 1 transmits the selected time-series data to the terminal device 5. Having received the time-series data from the information processing device 1, the terminal device 5 displays the query time-series data selection screen shown in FIG. 9 .

[0067] 9 is provided at the top of the query time-series data selection screen with a query time-series data display area 53 for displaying information about the query time-series data selected by the user. The query time-series data display area 53 displays information about the selected query time-series data, such as "wafer ID," "device name," "sensor name," and "timestamp," as well as a graph of the query time-series data. The terminal device 5 generates a graph for the query time-series data provided from the information processing device 1, with the horizontal axis representing time and the vertical axis representing values ​​such as temperature or pressure, and displays the graph in the query time-series data display area 53.

[0068] 9 , a similarity search condition selection area 54 is provided at the bottom for selecting similarity search conditions for searching the substrate processing DB 7 for time-series data similar to the query time-series data. A similarity search execution button labeled "Execute Similarity Search" is provided at the bottom right of the similarity search condition selection area 54. In this example, two similarity search conditions, "type of abnormality" and "timing of abnormality occurrence," can be selected. The similarity search condition selection area 54 includes check boxes for selecting these two conditions individually. The user can select similarity search conditions by checking the check boxes, and can instruct execution of the similarity search by operating the similarity search execution button. In this embodiment, the user may select no similarity search conditions, select one, select multiple conditions, or select all conditions. However, in this embodiment, if no conditions are selected, the similarity search is performed assuming that all conditions are selected.

[0069] When the similarity search execution button is operated, the terminal device 5 sends a search request including information about the query time-series data and the selected similarity search criteria to the information processing device 1. In response to the search request from the terminal device 5, the information processing device 1 searches the substrate processing DB 7 for time-series data similar to the query time-series data and extracts a predetermined number of time-series data with the highest similarity. At this time, the information processing device 1 performs the similarity search according to the similarity search criteria specified in the search request. The information processing device 1 transmits to the terminal device 5 one or more time-series data extracted from the substrate processing DB 7 as time-series data similar to the query time-series data and information about the time-series data, and causes the terminal device 5 to display the results of the successive searches. In this embodiment, the information processing device 1 searches for time-series data similar to the query time-series data using a similarity measure such as cosine similarity, and transmits the calculated similarity measure to the terminal device together with the results of the similarity search.

[0070] FIG. 10 is a schematic diagram showing an example of a similarity search result display screen. The terminal device 5, which has received the similarity search results from the information processing device 1, displays a list of a graph of time-series data similar to the query time-series data, information about the time-series data, the similarity to the query time-series data, and document information associated with the time-series data, in association with each other. The graph of the time-series data is, for example, a graph with time on the horizontal axis and a value such as temperature or pressure on the vertical axis. The information about the time-series data is, for example, information such as a wafer ID, an apparatus name, and a timestamp, and is information stored in the substrate processing DB 7. The similarity is, for example, a cosine similarity with the query time-series data calculated during the similarity search. When calculating Euclidean distance during the similarity search, the terminal device 5 may, for example, display the Euclidean distance instead of the similarity, or may calculate the similarity based on the Euclidean distance. For example, the reciprocal of the Euclidean distance may be used as the similarity. The document information is, for example, information such as a sentence written by an administrator or operator of the substrate processing device 3 regarding the time-series data, and is information stored in the substrate processing DB 7. The document information is not limited to text, but may be, for example, an image, a sound, or a video.

[0071] FIG. 11 is a schematic diagram showing another example of information input / output related to similarity search. In the example shown in FIG. 11 , the information processing device 1 exchanges messages with a user via a terminal device 5. The terminal device 5 mediates the message exchange between the user and the information processing device 1 and displays the message exchange screen, or so-called chat screen, shown in FIG. 11 on its display unit. The information processing device 1 inputs a user message provided by the terminal device 5 to the large-scale language model, for example, using a large-scale language model stored in the model information storage unit 12b, and acquires and transmits a response message output by the large-scale language model to the terminal device 5. The information processing device 1 instructs the large-scale language model in advance to proceed with the message exchange, indicating information necessary for the similarity search and requesting the acquisition of this information from the user. When the necessary information is collected, the information processing device 1 performs a similarity search and transmits the search results to the terminal device 5, thereby displaying the search results on the terminal device 5.

[0072] In this example, the information processing device 1 displays a list of information related to time-series data and accepts a user's selection from the list to acquire the query time-series data, but the present invention is not limited to this. For example, if a user has a file of time-series data to be used as query time-series data, the user may upload the file of query time-series data from the terminal device 5 to the information processing device 1, and the information processing device 1 may acquire the file uploaded from the terminal device 5 to acquire the query time-series data.

[0073] 11, messages entered by the user are displayed on the left side, and messages from the system (information processing device 1) are displayed on the right side. For example, a user enters a message such as "Show me a list of wafers for device A01," and in response, the information processing device 1 outputs a query message such as "Please specify a sensor." In response to this query, the user enters the message "temperature sensor," and the information processing device 1 searches for time-series data information related to the temperature sensor of device A01 and displays a list of time-series data together with the message "Displaying a list of wafers." Note that while the display of the wafer list is omitted in FIG. 11, a display similar to the list information displayed in the query candidate display area 52 of FIG. 8 is shown.

[0074] The user selects data to be used as query time-series data from among the multiple time-series data displayed in the wafer list, and in response, the information processing device 1 displays a graph of the selected query time-series data along with a message saying, "The data for the selected wafer is displayed." In this example, the user requests a similarity search by inputting a message saying, "Search for past cases with similar types of abnormalities." In response, the information processing device 1 searches the substrate processing DB 7 for time-series data similar to the selected query time-series data, focusing on the type of abnormality, and displays the similarity search results along with a message saying, "The results of the similarity search are displayed." Note that although the display of the similarity search results is omitted in FIG. 11 , a display similar to that shown in FIG. 10 is shown here, for example.

[0075] <Summary> In the information processing system according to the present embodiment configured as described above, the information processing device 1 receives time-series data related to substrate processing as input and generates a learning model 9 that outputs feature quantities of the time-series data. The information processing device 1 acquires learning data that associates the time-series data related to substrate processing with correct labels (correct values) related to processing such as classification or prediction performed based on the feature quantities of the time-series data. The information processing device 1 inputs the acquired time-series data of the learning data to the learning model 9, acquires the feature quantities output by the learning model 9, and calculates multiple errors, such as a center error, a cross-entropy error, and a mean square error, based on the acquired feature quantities, the correct labels of the learning data, and multiple predetermined loss functions. The information processing device 1 calculates a total error based on the calculated multiple errors, and performs machine learning by updating the internal parameters of the learning model 9 based on the calculated total error, thereby generating the learning model 9. As a result, the information processing system according to this embodiment can generate a learning model 9 that converts time series data into features by taking into account multiple factors, such as classification according to the state of substrate processing and the timing of abnormalities in substrate processing, and it is expected that the features converted by this learning model 9 can be used to realize accurate similarity searches of time series data.

[0076] Furthermore, in the information processing system according to this embodiment, the correct labels assigned to the time-series data as learning data include classification results relating to the substrate processing state (normal, abnormal, type of abnormality, etc.). The information processing device 1 calculates an error, such as a cross-entropy error, using a predetermined loss function based on the classification results obtained by classifying the substrate processing state based on the feature values ​​output by the learning model 9 and the correct labels of the classification results included in the learning data. As a result, the information processing system according to this embodiment can be expected to generate a learning model 9 that outputs feature values ​​taking into account the classification of the substrate processing state.

[0077] Furthermore, in the information processing system according to this embodiment, the information processing device 1 calculates the center error between the feature output by the learning model 9 and the center of the feature group divided based on the correct label of the learning data. As a result, the information processing system according to this embodiment can be expected to generate a learning model 9 that outputs features with higher classification ability.

[0078] Furthermore, in the information processing system according to this embodiment, the correct label attached to the time-series data as learning data includes information regarding the timing at which the substrate processing state changes (such as the timing at which an abnormality occurs). The information processing device 1 calculates an error, such as a mean square error, using a predetermined loss function based on a predicted value of the timing at which the substrate processing state changes, which is calculated based on the feature values ​​output by the learning model 9, and the timing information included in the correct label of the learning data. As a result, the information processing system according to this embodiment can be expected to generate a learning model 9 that outputs feature values ​​that take into account the timing at which the substrate processing state changes.

[0079] Furthermore, in the information processing system according to this embodiment, the information processing device 1 acquires time-series data without a correct label, inputs the acquired time-series data into the learning model 9 to acquire features, and determines a pseudo label for the time-series data based on the acquired features. The information processing device 1 calculates an error based on second learning data in which the time-series data and the pseudo label are associated, and performs machine learning. Furthermore, the information processing device 1 performs machine learning using learning data in which the correct label is assigned to the time-series data from the start of machine learning until a predetermined timing, and after the predetermined timing, performs machine learning using second learning data in which the pseudo label is assigned to the time-series data. As a result, the information processing system according to this embodiment can perform machine learning of the learning model 9 using not only time-series data with a correct label, but also time-series data without a correct label. This makes it possible to easily increase the amount of learning data, which is expected to further improve the accuracy of the learning model 9.

[0080] Furthermore, in the information processing system according to this embodiment, the information processing device 1 performs machine learning of the learning model 9 using multiple pieces of time-series data obtained from substrate processes performed with different settings (recipes). The information processing device 1 may also perform machine learning of the learning model 9 using multiple pieces of time-series data obtained from substrate processes performed in different devices. As a result, the information processing system according to this embodiment is expected to be able to obtain features by using the same learning model 9 to handle time-series data obtained from substrate processes performed with different settings.

[0081] Furthermore, in the information processing system according to this embodiment, the information processing device 1 acquires time-series data related to substrate processing, inputs the acquired time-series data into the learning model 9 generated by the above-described method to acquire feature quantities, and, based on the acquired feature quantities, extracts and outputs one or more time-series data from the substrate processing DB 7 in which the time-series data and feature quantities are stored in association with each other. As a result, the information processing system according to this embodiment is expected to realize highly accurate similarity search of time-series data using the feature quantities converted from the time-series data by the learning model 9.

[0082] In the information processing system according to this embodiment, the information processing device 1 accepts input of at least one search condition from among multiple search conditions associated with multiple correct labels assigned to time-series data in the training data, and extracts time-series data from the substrate processing DB 7 based on the accepted search condition. For example, the information processing device 1 outputs multiple search conditions and accepts selection of at least one search condition from among the multiple search conditions. Furthermore, for example, the information processing device 1 accepts input of a message in natural language and accepts input of search conditions using a large-scale language model based on this message. As a result, the information processing system according to this embodiment is expected to enable a user to appropriately specify search conditions and perform a similarity search of time-series data.

[0083] In the information processing system according to this embodiment, the information processing device 1 outputs information about a plurality of time-series data and accepts selection of time-series data from the output information, and the selected time-series data is used as query time-series data that will be the basis for similarity search. This allows the user to easily determine the query time-series data that will be the basis for similarity search in the information processing system according to this embodiment. Note that if the user has time-series data to be used as query time-series data, the information processing device 1 may acquire the query time-series data by uploading a file of the query time-series data from the terminal device 5 to the information processing device 1, for example.

[0084] In the information processing system according to this embodiment, the information processing device 1 extracts a predetermined number of time-series data items having high similarity in feature quantities from the substrate processing DB 7 based on the feature quantities of the query time-series data, and outputs the plurality of time-series data items in descending order of similarity (smallest Euclidean distance). As a result, the information processing system according to this embodiment is expected to provide the user with a plurality of past cases similar to the given query time-series data in an easy-to-understand manner.

[0085] In the present embodiment, the learning model 9 is configured to accept one piece of time-series data as input and output one feature, but this is not limited thereto. The learning model 9 may be configured to accept multiple pieces of time-series data as input, such as time-series data related to temperature and time-series data related to pressure, and output one feature or multiple feature amounts. Furthermore, for example, if the substrate processing apparatus 3 is equipped with multiple sensors and the information processing apparatus 1 acquires multiple pieces of time-series data, such as temperature and pressure, multiple learning models 9 may be generated and used for each piece of time-series data. Furthermore, the time-series data acquired by the information processing apparatus 1 may be input to the learning model 9 after undergoing appropriate preprocessing, such as normalization.

[0086] The embodiments disclosed herein are to be considered as illustrative in all respects and not restrictive. The scope of the present disclosure is defined by the claims, not by the above meaning, and is intended to include all modifications within the meaning and scope of the claims.

[0087] The matters described in each embodiment can be combined with each other. Furthermore, the independent claims and dependent claims described in the claims can be combined with each other in any and all combinations, regardless of the reference format. Furthermore, the claims use a format in which a claim references two or more other claims (multiple claim format), but this is not limited to this. A multiple claim (multi-multi claim) that references at least one other multiple claim may also be used.

[0088] REFERENCE SIGNS LIST 1 Information processing device (computer) 3 Substrate processing device 5 Terminal device 7 Substrate processing DB 9 Learning model 11 Processing unit 11a Time series data acquisition unit 11b Learning processing unit 11c Search processing unit 11d Input / output processing unit 12 Storage unit 12a Program (computer program) 12b Model information storage unit 13 Communication unit 51 Query condition input area 52 Query candidate display area 53 Query time series data display area 54 Similar search condition selection area 99 Recording medium N Network

Claims

1. A computer program that causes a computer to execute the following processes: receive time-series data related to substrate processing as input, and generate a learning model that outputs feature quantities of the time-series data; acquire learning data that associates the time-series data related to substrate processing with multiple correct answer values ​​related to processing performed based on the feature quantities of the time-series data; input the acquired time-series data of the learning data into the learning model to acquire feature quantities output by the learning model; calculate multiple errors based on the acquired feature quantities, the correct answer values ​​of the learning data, and multiple loss functions; calculate an overall error based on the calculated multiple errors; and update the parameters of the learning model based on the calculated overall error.

2. The computer program of claim 1, wherein the plurality of errors include a center error, a cross-entropy error, and a mean square error.

3. The computer program of claim 1, wherein the correct answer value of the learning data includes a classification result relating to the state of the substrate processing, and the plurality of loss functions include a loss function that calculates an error based on the classification result that classifies the state related to the substrate processing based on the features output by the learning model and the classification result included in the correct answer value of the learning data.

4. The computer program according to claim 3, wherein the plurality of loss functions include a loss function that calculates the error between the feature output by the learning model and the center of a group of feature values ​​divided based on the classification results included in the correct values ​​of the learning data.

5. The computer program of claim 1, wherein the correct answer value of the learning data includes information regarding the timing at which the state of the substrate processing changes in the time series data, and the plurality of loss functions include a loss function that calculates an error based on a prediction result of the timing at which the state related to the substrate processing changes based on the features output by the learning model and information regarding the timing included in the correct answer value of the learning data.

6. The computer program of claim 1, which acquires time series data that has no correct answer value assigned, inputs the acquired time series data into the learning model to acquire features output by the learning model, determines a pseudo-correct answer value for the time series data based on the acquired features, and calculates the overall error based on second learning data in which the determined pseudo-correct answer value is associated with the time series data.

7. The computer program of claim 6, which repeatedly calculates the overall error and updates the parameters of the learning model, calculates the overall error using learning data that associates the time series data with the correct value from the start of the repetition until a predetermined timing, and calculates the overall error after the predetermined timing by further using second learning data that associates the time series data with the pseudo-correct value.

8. The computer program according to claim 1, wherein the learning data includes time-series data relating to substrate processing performed under different settings.

9. A computer program that causes a computer to perform the following processes: acquire time series data related to substrate processing; input the acquired time series data into a learning model that accepts the time series data related to substrate processing as input and outputs feature quantities of the time series data; acquire the feature quantities output by the learning model; extract one or more time series data based on the acquired feature quantities from a database in which time series data and feature quantities are stored in correspondence with each other; and output the extracted time series data.

10. The computer program according to claim 9, wherein the learning model is generated by: acquiring learning data that associates time-series data related to substrate processing with a plurality of correct values ​​related to processing performed based on features of the time-series data; inputting the acquired time-series data of the learning data into the learning model to acquire features output by the learning model; calculating a plurality of errors based on the acquired features, the correct values ​​of the learning data, and a plurality of loss functions; calculating a total error based on the calculated multiple errors; and updating parameters of the learning model based on the calculated total error.

11. The computer program according to claim 10, which accepts input of at least one search condition from among a plurality of search conditions associated with the plurality of correct values, and extracts one or more time series data from a database in which time series data and feature amounts are stored in correspondence with each other, based on the acquired feature amount and the search condition for which the input was accepted.

12. The computer program according to claim 11, further comprising: outputting the plurality of search conditions; and accepting input of the search conditions by accepting selection of at least one search condition from the plurality of search conditions that have been output.

13. The computer program according to claim 11, further comprising: accepting an input of a message in a natural language; and accepting an input of the search condition using a language model based on the accepted message.

14. The computer program according to claim 9, which outputs information relating to a plurality of time series data, accepts selection of time series data based on the output information, and acquires the selected time series data.

15. The computer program according to claim 9, further comprising: extracting a predetermined number of time series data having high similarity in feature quantities from the database based on the acquired feature quantities; and outputting the extracted time series data in descending order of similarity.

16. A method for generating a learning model in which an information processing device receives time-series data related to substrate processing as input and generates a learning model that outputs feature quantities of the time-series data, the method comprising: acquiring learning data that associates the time-series data related to substrate processing with multiple correct answer values ​​related to processing performed based on the feature quantities of the time-series data; inputting the acquired time-series data of the learning data into the learning model to acquire feature quantities output by the learning model; calculating multiple errors based on the acquired feature quantities, the correct answer values ​​of the learning data, and multiple loss functions; calculating a total error based on the calculated multiple errors; and updating parameters of the learning model based on the calculated total error.

17. An information processing method comprising: an information processing device acquiring time series data related to substrate processing; inputting the acquired time series data into a learning model that accepts the time series data related to substrate processing as input and outputs feature quantities of the time series data; acquiring the feature quantities output by the learning model; extracting one or more time series data based on the acquired feature quantities from a database in which time series data and feature quantities are stored in correspondence with each other; and outputting the extracted time series data.

18. An information processing device comprising a processing unit that receives time series data related to substrate processing as input and performs processing to generate a learning model that outputs feature quantities of the time series data, wherein the processing unit: acquires learning data that associates the time series data related to substrate processing with multiple correct answer values ​​related to processing performed based on the feature quantities of the time series data; inputs the acquired time series data of the learning data into the learning model to acquire feature quantities output by the learning model; calculates multiple errors based on the acquired feature quantities, the correct answer values ​​of the learning data, and multiple loss functions; calculates a total error based on the calculated multiple errors; and updates the parameters of the learning model based on the calculated total error.

19. An information processing device comprising a processing unit, wherein the processing unit acquires time series data related to substrate processing, inputs the acquired time series data to a learning model that accepts the time series data related to substrate processing as input and outputs feature quantities of the time series data, acquires the feature quantities output by the learning model, extracts one or more time series data based on the acquired feature quantities from a database in which the time series data and feature quantities are stored in correspondence with each other, and outputs the extracted time series data.

Images