Search methods, search devices, computers, semiconductor device manufacturing systems, search methods, and analysis methods

Abstract

The present invention provides a search method, a search apparatus, a computer, a semiconductor device manufacturing system, a search method, and an analysis method. [Solution] A search method for estimating, by machine learning, when a recipe defined by a recipe file is executed using a semiconductor substrate processing apparatus, wherein the machine learning method takes numerical data describing the recipe as input and estimated data showing the result of estimating the processing of the semiconductor substrate as output, and uses a machine learning model that is trained using training data including the numerical data and data showing the execution result of executing the recipe, and in generating the training data, a dataset including the numerical data and the data showing the execution result is integrated, and further a conversion process is performed to convert the structure of the recipe file into a structure corresponding to the machine learning model.

Description

Technical Field

[0001] The present invention relates to a search method, a search device, a computer, a semiconductor device manufacturing system, a search method, and an analysis method.

Background Art

[0002] Machine learning is performed using AI (artificial intelligence) to autonomously find data laws and patterns from given data in various fields.

[0003] For example, also in the field of semiconductor manufacturing, in general semiconductor processing apparatuses (hereinafter also referred to as "processing apparatuses") (for example, plasma etching apparatuses, film forming apparatuses, CMP (Chemical Mechanical Polishing) apparatuses, etc.), search for recipe parameters (hereinafter referred to as recipe search) is performed using machine learning.

[0004] For example, in Patent Document 1, a parameter compression unit compresses a first input parameter value so that it can be restored by a parameter restoration unit, and generates a first compressed input parameter value with the number of control parameters reduced. A model learning unit learns a prediction model using learning data that is a pair of the first compressed input parameter value and a first output parameter value that is a processing result obtained by giving the first input parameter value to a processing apparatus as a plurality of control parameters. A processing condition search unit discloses a search apparatus that estimates a second compressed input parameter value corresponding to a target output parameter value using the prediction model.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0006] The search device is operated using the following steps: dataset creation (preprocessing), model construction, and model application (recipe extraction). Regarding data preprocessing and the selection of meaningful data as features (explanatory variables), the technique described in Patent Document 1 still has room for improvement. The present invention aims to provide a technology that can optimize the search process. [Means for solving the problem]

[0007] To solve the above problems, one representative search method of the present invention is a search method that uses machine learning to estimate when a recipe defined by a recipe file is executed using a semiconductor substrate processing apparatus, wherein the machine learning method uses a machine learning model in which numerical data describing the recipe is input, estimated data showing the result of the semiconductor substrate processing is output, and training is performed using training data that includes the numerical data and data showing the execution result of executing the recipe, and in generating the training data, a dataset including the numerical data and the data showing the execution result is integrated, and further a conversion process is performed to convert the structure of the recipe file into a structure corresponding to the machine learning model. [Effects of the Invention]

[0008] According to the present invention, it becomes possible to optimize the search process. Other issues, configurations, and effects not mentioned above will be clarified by the description of the embodiments for carrying out the invention below. [Brief explanation of the drawing]

[0009] [Figure 1] Figure 1 shows an example of a recipe in an etching apparatus. [Figure 2] Figure 2 shows an example of a modeling target used for machine learning recipe exploration in etching equipment. [Figure 3] Figure 3 shows an example of a recipe search method using machine learning. [Figure 4] Figure 4 shows the data used in machine learning and the processing performed on that data. [Figure 5] Figure 5 shows an example of a system to which the recipe search method is applied. [Figure 6] Figure 6 shows an example of the configuration of an AP server for a machine learning system. [Figure 7] Figure 7 shows an example of a recipe. [Figure 8] Figure 8 shows the actual processing steps of the search method in Example 1. [Figure 9] Figure 9 shows a flowchart (Figure 9(a)) illustrating the details of the data structure transformation (step S102 in Figure 8) in the search method of Example 1, and a flowchart (Figure 9(b)) illustrating the registration and update process of the data structure transformation method that is performed prior to the definition data reading (step S101 in Figure 8). [Figure 10] Figure 10 shows the first type of data structure transformation. [Figure 11] Figure 11 shows the second type of data structure transformation. [Figure 12] Figure 12 shows the third type of data structure transformation. [Figure 13] Figure 13 shows the fourth type of data structure transformation. [Figure 14] Figure 14 shows the fifth type of data structure transformation. [Figure 15] Figure 15 shows the sixth type of data structure transformation. [Figure 16] Figure 16 shows an example of a data structure transformation. [Figure 17] Figure 17 shows an example of a system to which the recipe search method is applied (Example 2). [Figure 18] Figure 18 shows the pre-processing procedure of the search method in Example 2. [Figure 19] Figure 19 shows an example of the recipe file structure for Example 2. [Figure 20]Figure 20 shows an example of a system to which the recipe search method is applied (Example 3). [Figure 21] Figure 21 shows the pre-treatment procedure of the search method in Example 3. [Figure 22] Figure 22 shows an example of the recipe file structure for Example 3. [Figure 23] Figure 23 shows an example of a system to which the recipe search method is applied (Example 4). [Figure 24] Figure 24 shows an example of the recipe file structure for Example 4. [Figure 25] Figure 25 shows the actual processing procedure of the search method in Example 5. [Figure 26] Figure 26 shows an example of the recipe file structure for Example 5. [Modes for carrying out the invention]

[0010] Embodiments of the present invention will be described below with reference to the drawings. However, the present invention is not limited to these embodiments. Furthermore, in the drawings, identical parts are denoted by the same reference numerals.

[0011] In this disclosure, "recipe" refers to a set of data describing the settings required to operate the processing device. The processing device is composed of components that enable it to perform its functions as a processing device, and the recipe holds the settings required to operate each component. Furthermore, AI refers to information processing technologies such as programs and systems that operate in a manner similar to human thought processes. Furthermore, machine learning refers to AI that implements the learning process equivalent to that performed by humans. It discovers patterns and rules from input data (training data) and, when applied to new data, enables prediction and classification of that new data. Furthermore, features refer to data used as input for a machine learning model. In the context of features, the term also includes parameters that indicate the output of sources such as gas sources, raw material sources, heat sources, electron sources, electromagnetic wave sources, light sources, power sources, and sound sources. Furthermore, although the following explanation uses the term "personal computer (PC)," it is not limited to PC terminals. Virtual areas, servers, and mobile terminals may also be applied.

[0012] [Conventional example] (Example of an etching device recipe) Refer to Figures 1 and 2 to see examples of recipes and models in an etching apparatus. Figure 1 shows an example of a recipe in an etching apparatus. Figure 2 shows an example of a model used for machine learning recipe discovery in an etching apparatus.

[0013] When training an AI for recipe discovery, a dataset consisting of recipe parameters and data showing the execution results of the recipe is input to the AI ​​as training data. Recipe 1, managed in a certain etching apparatus, consists of multiple CSV files, as shown in Figure 1(a). In other words, recipe 1 is defined by multiple recipe files. Each recipe file specifies the operations to be performed by the etching apparatus and includes data such as the vacuum level, exhaust volume, type of gas used, and the amount introduced into the chamber.

[0014] These CSV files are structured according to the hardware configuration of the etching apparatus. Figure 1(b) shows an example of the types of gases included in Recipe 1 and the amounts introduced into the chamber. No. 10, enclosed by a dashed line, indicates the introduction of Ar gas into gas piping (labeled "gas line" in the figure) 1. The flow rate is 100 mL / min in step 1 and 100 mL / min in step 2. No. 13 indicates the introduction of Ar gas into gas piping 4. The flow rate is 5 mL / min in step 1 and 0 mL / min in step 2. Thus, even if the gas type is the same, gases supplied to the chamber from different supply pipes are treated as separate parameters in the recipe, as shown in Figure 1(b).

[0015] When generating a model using machine learning with parameters directly from a recipe, if the same Ar gas is treated as a separate parameter, the learning efficiency cannot be improved, and the predictability of the model cannot be expected to improve. The dataset used for machine learning needs to be structured from the perspective of plasma control. For example, even if the same type of gas is supplied to the chamber from different supply pipes, if it is introduced in the same Step 1, it may be preferable for machine learning to treat it as a common parameter.

[0016] (An example of a model used in machine learning) Refer to Figure 2 to see an example of a modeling target in machine learning. Figure 2 shows an example of a modeling target used for recipe discovery in machine learning for etching equipment.

[0017] The model shown in Figure 2 illustrates the reaction system occurring on a semiconductor substrate within an etching apparatus. The semiconductor substrate (wafer) is placed on a stage within the etching apparatus. The stage is equipped with heating and cooling functions, and the heat generated by these functions is applied to the wafer. Power is also supplied to the stage, generating a bias potential for the plasma generated within the etching apparatus. The wafer within the etching apparatus is affected by ions, radicals, atoms and molecules, and electromagnetic waves. The surface of the wafer is processed under these influences.

[0018] While this disclosure shows an example of a model, it is not limited to this example. The model is determined based on factors such as the performance of the computer processing unit used for machine learning, the type and number of recipes used as training data, and the accuracy required for recipe search. Furthermore, while heat, power, ions, radicals, atoms and molecules, and electromagnetic waves are shown as parameters for the model, this disclosure is not limited to these. For example, it is also possible to use parameters related to the inflow of various particles, such as the amount, frequency, and angle of incidence of ions, as well as parameters related to surface reactions such as reactions and diffusion, and parameters related to outflow, such as the adsorption and desorption of molecules and atoms. Additionally, the model may be constructed from a microscopic perspective, or it may be substituted with macroscopic parameters such as pressure and temperature.

[0019] (The relationship between the modeling target and the recipe (dataset) obtained by the processing unit) Each component within the processing apparatus operates based on the settings defined in the data set. For example, if the processing apparatus is an etching apparatus, it may include components such as a high-frequency power supply that generates electromagnetic waves, a gas flow meter for introducing gases to generate atoms, molecules, ions, and radicals, a power supply for applying electrostatic force to the stage, heaters and coolers for heating the stage, and a vacuum evacuation device for adjusting the pressure in the chamber where the wafer is processed.

[0020] For a processing unit to function correctly, each data point in the data set describing the setting values ​​must be uniquely linked to a component of the processing unit. In other words, each setting value in the data set describing the setting values ​​is stored in a way that links it to a name that uniquely identifies the processing unit.

[0021] When dealing with this type of data in machine learning, the input variables must be items that can be uniquely identified within the processing unit, but this presents several challenges.

[0022] The first challenge is that, since the modeling target in machine learning is the reaction system occurring within the processing chamber, it is preferable that the explanatory variables be variables related to the reaction system. On the other hand, if data obtained directly from the processing device is used, the input variables become the control variables of the processing device. In this case, what is modeled is the processing device, not the reaction system within the processing chamber, which contradicts the original objective of learning about the reaction system. Furthermore, because the model is based on the control variables of the processing device, it becomes difficult to interpret the model generated by machine learning (hereinafter also referred to as the "machine learning model").

[0023] The second challenge is that using variables from the processing unit directly as input variables for a machine learning model may generally result in lower accuracy than using variables from the reaction system as input variables. This is because the conversion rules for converting variables from the processing unit to variables from the reaction system also need to be learned, and when comparing with the same training data, this means that the training data is insufficient.

[0024] The third challenge is that the dimensionality of the training data may increase unnecessarily. Some of the control variables of the processing unit do not necessarily have a one-to-one correspondence with the variables related to the reaction system within the processing unit. For example, in the example shown in Figure 1, the same type of gas piping is not necessarily grouped together in one place within the processing unit, but may be arranged in multiple separate lines from a control perspective, each connected to a processing chamber. In this case, from the control of the processing unit, it is necessary to control the inflow rate of each gas piping, but by the time the incoming gas reaches the semiconductor substrate, it has already been mixed, and the information as to which piping it came from is lost, so it should be treated as a single variable in the machine learning model of the reaction system.

[0025] To address the above challenges, instead of directly using the data sets describing the settings in the processing unit as training data for machine learning, it is necessary to perform a data structure transformation using arithmetic operations, statistical processing, and similar calculations to convert the data into variables that correspond to the reaction system occurring within the processing unit before using it as training data for machine learning.

[0026] An example of data structure transformation will be explained based on the example shown in Figure 1, where multiple identical gas pipes are connected. Gas pipes 1 and 4 are connected to a processing unit. When the processing unit processes a semiconductor substrate, it is considered impossible to distinguish which gas pipe the Ar gas introduced from gas pipe 1 and gas pipe 4 came from when it reached the semiconductor substrate. In this case, the flow rates of the Ar gas flowing in from gas pipe 1 and the Ar gas flowing in from gas pipe 4 are taken as the sum of their flow rates and used as the learning data for the total flow rate of the introduced gas.

[0027] Note that gas pipes 2 and 3, shown in No. 11 and No. 12, supply a different type of gas than Ar gas. The flow rates of these gas pipes are used directly as training data as the gas flow rate if no multiple identical gas pipes are connected. If multiple identical gas pipes are connected, their flow rates are summed up, as in the Ar gas example above, and used as training data as the flow rate of the introduced gas.

[0028] Many other examples of this kind can be cited, including raw material sources, heat sources, electron beam sources, electromagnetic wave sources, light sources, power sources, sound sources, and so on. Furthermore, the way these are handled is not limited to simply dealing with total quantities; some are expressed as differences, such as the relationship between incident and reflected waves from an electromagnetic wave source, where the actual amount introduced into the processing chamber is represented by a difference; others as products or integrals, such as the flow velocity of the raw material fluid and the processing time; and still others as quotients or derivatives, such as the change in heater power for temperature control in the processing chamber. These can be expressed using arithmetic operations and similar methods.

[0029] In this way, the variables in the processing unit are preprocessed to transform the data structure into appropriate variables and their settings, and this transformed data is output as training data.

[0030] [First Embodiment] (Application of a search system in a processing device) Referring to Figure 3, the machine learning processing procedure will be explained. Figure 3 is a diagram showing an example of a recipe search method using machine learning. Depending on the search method, the case in which a recipe defined by a recipe file is executed using a semiconductor substrate processing device is estimated by machine learning. For example, the processing result that is predicted to be obtained if the input recipe is executed is estimated. The machine learning model used in machine learning takes numerical data describing the recipe as input and estimated data showing the estimated result when the aforementioned recipe is executed as output, and training is performed using training data that includes numerical data and data showing the execution result when the recipe is executed. In this disclosure, based on the data obtained from the processing device, the reaction system occurring in the processing device is modeled by machine learning, and a data set describing the setting values ​​during processing is optimized so that the processing result is the desired result. In Figure 3, steps S1 to S5 show the process of training the machine learning model, and steps S6 and S7 show the process of searching for a recipe using the machine learning model.

[0031] First, data containing information necessary for processing in the semiconductor substrate processing unit is obtained (Step S1).

[0032] The acquired data specifically describes various settings required when the processing device is executed, as well as restrictions that must be met during execution. The acquired data is classified into a learning data group and a restriction information group (Step S2). In addition to the numerical data described in the recipe, the data includes data showing the execution results when the recipe is executed and target values ​​for the semiconductor substrate processing. The data showing the execution results here, for example, in the case of an etching device, includes data showing the shape of the semiconductor substrate after processing when the processing device processes the semiconductor substrate according to the recipe, and data showing the amount of processing that was changed by the processing. The target values ​​for the semiconductor substrate processing indicate the target state of the semiconductor substrate that is expected to be achieved when processing is carried out according to the recipe. For example, in the case of an etching device, this is information showing the surface shape profile, such as the depth and width of the trenches formed. The data storage format is set for each processing device and is not limited to a specific storage format.

[0033] Next, when generating training data, a dataset containing numerical data and data showing execution results is integrated, and the structure of the recipe file is further transformed into a structure corresponding to the machine learning model. Details of the data will be described later. Each data is transformed into a data structure according to the reaction system (Step S3). This is the so-called data structure preprocessing. Data acquisition and preprocessing are carried out until all data is ready (Step S4).

[0034] Next, the model used for machine learning is trained (Step S5). The data, after the data structure transformation, is used as training data for the machine learning model. The machine learning model takes the parameters in the reaction system as input variables and the dimensions of the semiconductor substrate after processing by the processing device as output variables. The dimensions of the semiconductor substrate after processing by the processing device are, in other words, estimated data that shows the result of estimating the processing of the semiconductor substrate.

[0035] Subsequently, the trained machine learning model is used to optimize the input variables so that the output predicted post-processing dimensions satisfy the constraint information (step S6). The data input to the machine learning model at this time is the recipe data used when executing the search method of this disclosure (hereinafter also referred to as "trial data"), etc. Care is also taken to ensure that the trial data satisfies the constraint information.

[0036] After optimization is complete, the optimized recipe is generated by converting the optimized input variables back into the processing unit variables (step S7).

[0037] (Data sets used in machine learning) Refer to Figure 4 to explain the data used in machine learning. Figure 4 is a diagram showing the data used in machine learning and the processing performed on the data.

[0038] This includes data obtained from the recipe file, user-side restriction information 11, device-side restriction information 12, device information 13, and other information 14 (corresponding to steps S1 and S2 in Figure 3).

[0039] Data obtained from recipe files, devices, databases, etc., can be classified into two groups: data describing setting values ​​and data describing limiting information. The former includes various setting values ​​required for the operation of the processing device, while the latter includes data describing limits that must be met during the operation of the processing device, as well as target values ​​for semiconductor substrate processing.

[0040] Device information 13 is data that describes the setting values. This data holds the setting values ​​necessary for each component to operate in order to run the processing device. Specifically, examples include, but are not limited to, gas flow rate and temperature. In principle, each component within the processing device operates based on device information 13.

[0041] On the other hand, user-side restriction information 11 and device-side restriction information 12 are classified as data sets that describe restriction information. Furthermore, restriction information is further classified into two types: hardware restrictions and software restrictions.

[0042] Hardware limitations specifically include requirements for the normal operation of the processing unit, such as upper and lower limits on set values ​​and input limits that are dependent on specific set values. Furthermore, upper and lower limits and input limits are not limited to being specified by specific values ​​or thresholds, but may also be specified by specific ranges. In this disclosure, the device-side limitation information 12 corresponds to information indicating hardware limitations. The device-side limitation information 12 includes information that defines the range in which the processing unit will operate normally when the processing unit operates based on the device information 13.

[0043] As illustrated in the example shown in Figure 1(b), each gas pipe from No. 10 to No. 13 is provided with upper and lower limits for the set flow rate. Furthermore, gas pipe 1 (No. 10) and gas pipe 4 (No. 13) share the same Ar gas. When using the flow rate of Ar gas supplied by these gas pipes as a parameter for a machine learning model, the sum of the upper limits for gas pipe 1 and gas pipe 4 is defined as the upper limit. Note that the upper and lower limits defined here can be either the upper and lower limits based on the equipment specifications of each gas pipe, or they can be upper and lower limits that ensure a certain level of accuracy in recipe search.

[0044] Software limitations specifically include post-processing dimensions, yield, and utilization rates that must be satisfied for production management purposes. These values ​​may be specified not only by specific values ​​but also by ranges or thresholds. Post-processing dimensions, which are the target values ​​for semiconductor substrate processing, are also included as software limitations. In this disclosure, user-side limitation information 11 is considered to be information indicating software limitations.

[0045] Other information 14 includes data showing the execution results when a recipe is executed. Other information includes literature values, processing results obtained from other devices, past processing information obtained from databases, virtual data obtained by virtual measurements or physical models, input / output information or inference results from separately created trained machine learning models, etc.

[0046] The classified information undergoes a data structure transformation (step S20) and is divided into constraint conditions 16 and training data 17 (corresponding to step S3 in Figure 3). The constraint conditions 16 include the target value for semiconductor substrate processing included in user-side constraint information 11, and the upper and lower limits of parameters included in device-side constraint information 12.

[0047] Next, the training data 17 is input to the machine learning model 18, and the machine learning model is trained (step S21) (corresponding to step S5 in Figure 3).

[0048] Next, an optimized recipe is searched for (step S22). For example, input data that satisfies the constraint information is generated by the program and input to the trained model 20, and it is determined whether the obtained output result approaches the target value for processing the semiconductor substrate. This result is then reflected in the next input data that satisfies the constraint information, and this method is repeated to optimize the recipe. The optimized recipe is output and provided to the user, for example (step S23).

[0049] (Examples of system application) Below, we describe systems to which the recipe search method is applied, as shown in Examples 1 to 5.

[0050] (Example 1) (Configuration of Example 1) Example 1 will be described with reference to Figures 5 to 14. Figure 5 is a diagram showing Example 1 of a system to which the recipe search method is applied. Example 1 assumes, for example, that the processing device is a single unit. The search system 100 is accessed from user PC 1101 to 110 NThe system includes (N is a positive integer), a machine learning system 120, a manufacturing data system 130, an edge PC 140, and a processing unit 150. A local area network (hereinafter also referred to as "LAN") is formed in the exploration system 100. In the following description, it is assumed that the processing unit 150 is an etching processing unit, but this disclosure is not limited thereto. This disclosure is also applicable to equipment other than etching units. The processing unit 150 is a broad definition of an etching processing unit, which includes not only an etching processing unit in the narrow sense for performing plasma etching, etc., but also equipment related to etching, such as semiconductor substrate inspection equipment and measurement equipment. Furthermore, the exploration system 100 can also be described as a semiconductor device manufacturing system that has a platform on which semiconductor manufacturing equipment is connected via a network, and which has an application implemented to estimate, by machine learning, when a recipe defined by a recipe file is executed using a semiconductor substrate processing unit.

[0051] User PC 1101 to 110 N It accepts requests from users utilizing the search system 100 and presents the results of the search requested by the user to the user. In the following description, when referring to a user PC without specifying it, it is referred to as user PC 110. n (where n is a positive integer less than or equal to N) is used. As will be described later, user PC110 n This is a computer that can communicate with a search device (AP server) that estimates, using machine learning corresponding to a model of the processing performed by a semiconductor substrate processing device, what happens when a recipe defined by a recipe file is executed by the processing device.

[0052] The machine learning system 120 performs machine learning and searches for recipes. The machine learning system 120 includes a web server 121, an application server 122, and a database server 123. Web server 121 is connected to user PC 110 n The system receives requests sent from the user's PC110 and displays the results corresponding to those requests. nThe request is sent to the following server. If the request involves dynamic processing, such as searching for a recipe, the Web server 121 sends the request to the AP server 122. The AP server 122 receives requests sent from the Web server 121, performs a search in response to the request, and sends the found recipe to the Web server 121. Specifically, the AP server 122 is equipped with a learning device in which numerical data defining the recipe is input, and estimated data, which is an estimate of the execution result of the recipe, is output. The DB server 123 receives requests sent from the AP server 122, saves them to the database according to the requests, and extracts necessary data and sends it to the AP server 122. The DB server 123 stores data such as the surface shape and processing amount of semiconductor substrates, which are the results of the recipes used to train machine learning models and the execution results of those recipes.

[0053] The manufacturing data system 130 collects training data and equipment information and transmits it to the machine learning system 120. The manufacturing data system 130 includes a web server 131, an application server 132, and a database server 133. The web server 131 receives requests sent from the AP server 122 of the machine learning system 120 and sends the results corresponding to the requests back to the AP server 122. If the request involves dynamic processing, the web server 131 sends the request to the AP server 132. AP server 132 receives requests sent from Web server 131, processes the requests, and sends the processing results back to Web server 131. The DB server 133 receives requests sent from the AP server 132, saves them to the database according to the requests, and also extracts the necessary data and sends it back to the AP server 132.

[0054] The edge PC 140 processes the data output from the processing unit 150. The processing unit 150 outputs data including the executed recipe and the measurement results of the semiconductor substrate as a result of executing the recipe. The edge PC 140 performs optimization processing on the data output from the processing unit 150 to suppress communication delays when transmitting the data to the manufacturing data system 130.

[0055] The user of the search system 100 is user PC 110 n Recipe exploration is possible either through or directly using the machine learning system 120. Although an example was shown in which the edge PC 140 and the processing unit 150 are connected to the LAN of the exploration system 100, this disclosure can also be applied even if the edge PC 140a and the processing unit 150a are not included in the LAN. Details will be described later.

[0056] (Configuration of the AP server for the machine learning system) Figure 6 shows an example of the configuration of the AP server 122 of the machine learning system 120. The AP server 122 can also be described as a search device that estimates, using machine learning corresponding to a model of processing performed by a semiconductor substrate processing device, what happens when a recipe defined by a recipe file is executed by the processing device. The AP server 122 takes numerical data describing the recipe as input and estimate data showing the estimated results of the semiconductor substrate processing by the processing device 150 as output. It includes a computer on which a learning device is implemented that performs learning using learning data including numerical data and data showing the execution results, and a display that shows a user interface for setting target values ​​for processing. When generating learning data, the computer integrates a dataset including numerical data and data showing the execution results, and further converts the structure of the recipe file into a structure corresponding to the model.

[0057] The AP server 122 includes a CPU (Central Processing Unit) 1220, memory 1221, storage 1222, and interface 1223. The CPU 1220 functions as an arithmetic unit. The memory 1221 stores programs for performing various functions. The memory 1221 includes, for example, a data input / output block 1224, a data structure transformation block 1225, an ML model optimization block 1226, and a recipe generation block 1227. As will be described in detail later, the data input / output block 1224 has a program for acquiring data (step S100) and input variables used in step S104 of Figure 8. The data structure transformation block 1225 has a program for reading definition data (step S101) and transforming the data structure (step S102) of Figure 8. The ML model optimization block 1226 has a program for optimizing the input variables of the machine learning model (step S104). The recipe generation block 1227 has a program for performing the reverse conversion of input variables to processing unit variables (step S105). The storage 1222 is used to store data, for example, when transaction processing is performed in the CPU 1220. The storage 1222 is, for example, an HDD (Hard Disk Drive).

[0058] Interface 1223 accepts requests sent from outside the AP server 122 and also sends the results of the requests to the outside. Input / output device 207 is a device operated by, for example, a user of the search system 100. The user can also operate the machine learning system 120 using, for example, a user interface displayed on the display of input / output device 207. Interface 1223 enables communication between the AP server 122 and the Web server 121, between the AP server 122 and the DB server 123, and between the AP server 122 and the Web server 131 of the manufacturing data system 130.

[0059] The above description is merely an example of the configuration of AP server 122, and this disclosure is not limited thereto. For example, if large-scale computing is required, AP server 122 can also be configured to include a GPU (Graphical Processing Unit) in addition to the CPU 1220.

[0060] (Example recipe) Figure 7 shows an example of a recipe. Recipe 1501, shown in Figure 7(a), is executed, for example, by the processing apparatus 150. Recipe 1501 shows the types of gases and electrode conditions used in processing steps 1 to 3 of the plasma processing. In processing step 1, the types and flow rates of gases are set as follows: Ar gas at 0 mL / min and 100 mL / min, NF3 gas at 10 mL / min, and HBr gas at 20 mL / min. Here, since the Ar gas is controlled by two outputs in the processing apparatus 150, two flow rates are shown. The high-frequency power applied to the wafer is set to 10 W, and the electrode temperature is set to 10 °C. In processing step 2, the Ar gas is set to 20 mL / min and 0 mL / min, NF3 gas at 0 mL / min, and HBr gas at 10 mL / min. The high-frequency power applied to the wafer is set to 0 W, and the electrode temperature is set to 10 °C. In processing step 3, the flow rates of Ar gas are set to 10 mL / min and 0 mL / min, NF3 gas to 20 mL / min, and HBr gas to 0 mL / min. The high-frequency power applied to the wafer is set to 20 W, and the electrode temperature is set to 10 °C.

[0061] Figure 7(b) shows the upper and lower control limits for each item shown in the recipe. The "lower control limit" and "upper control limit" are information that defines the upper and lower limits for the items included in the recipe when performing data structure transformation and recipe search, which will be described later. In addition, "zero input possible" is information that indicates whether or not the flow rate can be controlled to 0 by closing the gas piping. The flow rate of Ar gas in gas line 1 is limited to 5 mL / min or more and 50 mL / min or 0 mL / min. The flow rate of NF3 gas in gas line 2 is limited to 0 mL / min or more and 50 mL / min or 0 mL / min. The flow rate of HBr gas in gas line 3 is limited to 0 mL / min or more and 50 mL / min or 0 mL / min. The flow rate of Ar gas in gas line 4 is limited to 50 mL / min or more and 200 mL / min or 0 mL / min. The upper and lower control limits for gas flow rates are shown here, but the settable range is also defined for other items included in the recipe.

[0062] (Method of Example 1) The search method in Example 1 will be described with reference to Figures 8 and 9.

[0063] Figure 8 shows the actual processing procedure of the search method in Example 1. First, data containing information necessary for processing in the processing unit 150 is obtained from the processing unit 150 (step S100). Numerical data and related information (recipe file name, information indicating the execution result of executing the recipe, etc.) described in the recipe are obtained. The AP server 122 of the machine learning system 120 obtains this data via the edge PC 140 and the manufacturing data system 130. When collecting data sent from the processing unit 150 to the DB server 133, the AP server 122 obtains the data via the DB server 133, AP server 132, and Web server 131.

[0064] Next, the AP server 122 reads the definition data (step S101). The definition data specifies the target and method of data structure transformation. The definition data is stored, for example, in the DB server 123. The registration and updating of the definition data will be described later. The AP server 122 queries the target and method of data structure transformation for the acquired data.

[0065] Next, the AP server 122 performs a data structure transformation on the acquired data based on the target and method of data structure transformation indicated in the definition data (step S102).

[0066] Next, the AP server 122 trains a machine learning model using the data that has undergone data structure transformation (step S103).

[0067] Next, the user inputs the recipe to be implemented and the target values ​​for the process, and then searches for a recipe. The variables (input data) input to the machine learning model are optimized to suit the machine learning model. In addition, the inputs and outputs of the machine learning model are made to satisfy the constraint information classified from the input data (step S104).

[0068] The AP server 122 converts the input variables back into the variables of the processing unit (recipe generation) (step S105). In other words, the searched recipe is expressed in the form of the input variables of the machine learning model. By performing the reverse transformation of the data structure conversion on the searched recipe, a recipe is generated that is converted into a format that describes the recipe of the processing unit 150.

[0069] Figure 9 shows a flowchart (Figure 9(a)) illustrating the details of the data structure transformation (step S102 in Figure 8) in the search method of Example 1, and a flowchart (Figure 9(b)) illustrating the registration and update process of the data structure transformation method that is performed prior to the definition data reading (step S101 in Figure 8).

[0070] Referring to Figure 9(a), the data structure transformation (step S102 in Figure 8) will be explained. The DB server 133 stores data from the processing unit 150, and the AP server 122 reads the data from the DB server 133 (step S1020). For each piece of data read, the AP server 122 determines whether or not the target data requires data structure transformation (step S1021). The AP server 122 performs data structure transformation on all data that requires it (step S1022).

[0071] Referring to Figure 9(b), the registration and updating of the data structure transformation method specified in the definition data (step S101 in Figure 8) will be explained.

[0072] The user is user PC110 n Alternatively, data structure transformation methods are registered and updated via interface 1223 of AP server 122, based on the numerical data described in the machine learning model and recipe. Here, the data structure transformation methods are, for example, those shown in the first to sixth types described later. A definition is registered that associates the data type with the corresponding type.

[0073] The AP server 122 registers the target and method for data structure transformation (step S1001 in Figure 9(b)). Step S1001 is performed for all data (features) obtainable from the recipe. The registered data and method may be stored, for example, in the storage device (memory 1221) of the AP server 122, or they can be stored in the DB server 123.

[0074] (Types of data structure transformations) The data structure transformation will be explained with reference to Figures 10 to 15. The recipe file has a tabular data structure that includes the numerical data and the labels attached to that numerical data. Through the data structure transformation, the numerical data shown in the recipe is transformed into parameters and numerical values ​​suitable for recipe exploration using a machine learning model. The data structure transformation process includes at least one of the following operations: adding, deleting, performing arithmetic operations on values, grouping, splitting, and replacing numerical data.

[0075] Figure 10 shows the first type of data structure transformation. Figure 10(a) schematically illustrates the data structure transformation, and Figure 10(b) shows a specific example of numerical data obtained from a recipe. In the first type, there is no direct input to the data structure transformation, but features are output through indirect input. For example, this could be the case when device-specific corrections are substituted into the numerical data as correction terms. Data 51 before data structure transformation does not contain numerical data. Data 52 after data structure transformation is numerical data that was added without input. As a concrete example, as shown in Figure 10(b), before the data structure transformation, no numerical data is obtained for a given feature 1, and rows #1 to #10 do not contain numerical data. After the data structure transformation, numerical data is added to the given feature 1, and the values ​​20, 30, 100, 70, 30, 90, 10, 30, 80, and 90 are assigned to rows #1 to #10, respectively. These values ​​can also include correction terms that have been registered in advance.

[0076] Figure 11 illustrates the second type of data structure transformation. Figure 11(a) schematically shows the data structure transformation, and Figure 11(b) shows a specific example of numerical data obtained from a recipe. In the second type, there is input for the data structure transformation, but the numerical data is deleted for some reason. For example, this could occur if the recipe contains numerical data unrelated to the modeling target. Data 53 before the data structure transformation contains numerical data, but data 54 after the data structure transformation has the numerical data deleted. As a concrete example, as shown in Figure 11(b), before the data structure transformation, a predetermined feature 1 contains the numerical values ​​20, 30, 100, 70, 30, 90, 10, 30, 80, and 90 in rows #1 through #10, respectively. After the data structure transformation, the predetermined feature 1 is deleted, and the numerical values ​​in rows #1 through #10 are removed. Since feature 1 was not necessary for training the machine learning model, it is deleted through the data structure transformation.

[0077] Figure 12 illustrates the third type of data structure transformation. Figure 12(a) schematically shows a data structure transformation, and Figure 12(b) shows a specific example of numerical data obtained from a recipe. The third type involves establishing a one-to-one correspondence between the numerical data before and after the data structure transformation. For example, this could be done when changing the name of a feature or when converting the units of numerical data. In data 55 before the data structure transformation, the name is feature a, while in data 56 after the data structure transformation, the name is feature 1. As a concrete example, as shown in Figure 12(b), before the data structure transformation, feature a contains the numerical data 20, 30, 100, 70, 30, 90, 10, 30, 80, and 90 in rows #1 through #10, respectively. After the data structure transformation, feature 1 contains the numerical data 0.02, 0.03, 0.1, 0.07, 0.03, 0.09, 0.01, 0.03, 0.08, and 0.09 in rows #1 through #10, respectively. The unit of feature a corresponds to 1000 times the unit of feature 1. For the unit transformation, the numerical data of feature a is multiplied by a constant of 0.001.

[0078] Figure 13 illustrates the fourth type of data structure transformation. Figure 13(a) schematically shows a data structure transformation, and Figure 13(b) shows a specific example of numerical data obtained from a recipe. The fourth type involves the aggregation of two or more numerical data before the data structure transformation into a single numerical data after the transformation (so-called numerical data grouping). For example, this can be seen when dealing with features that are managed as multiple variables by the processing unit, but which are desirable to represent as a single variable from the perspective of the reaction system. Data 57a and data 57b before the data structure transformation are two data points, feature a and feature b, respectively. On the other hand, data 58 after the data structure transformation is a single data point as feature 1. As a concrete example, as shown in Figure 13(b), before the data structure transformation, feature a contains the numerical data 20, 30, 100, 70, 30, 90, 10, 30, 80, and 90 in rows #1 through #10, respectively. Similarly, feature b contains the numerical data 10, 30, 90, 60, 30, 70, 40, 30, 60, and 70 in rows #1 through #10, respectively. After the data structure transformation, feature 1 contains the numerical data 30, 60, 190, 130, 60, 160, 50, 60, 140, and 160 in rows #1 through #10, respectively. Feature a and feature b have the same function in the reaction system and can be treated as synonymous. Therefore, feature a and feature b have the same components and are treated as feature 1.

[0079] Figure 14 illustrates the fifth type of data structure transformation. Figure 14(a) schematically shows the data structure transformation, and Figure 14(b) shows a specific example of numerical data obtained from a recipe. In the fifth type, the data before the data structure transformation was stored in a format that included delimiters, but after the data structure transformation, it is separated into tabular data. For example, this can be considered when dealing with a recipe file in which multiple numerical data items are contained in a single line separated by commas, etc. The data 59 before the data structure transformation is text data with one delimiter per line. On the other hand, the data 60a and 60b after the data structure transformation are tabular data with one numerical value per line, respectively. As a concrete example, as shown in Figure 14(b), before the data structure transformation, feature a contains text data in rows #1 to #10 that consists of two numbers separated by commas, such as "70,100", "10,30", "50,80", "40,70", "70,50", "10,50", "40,50", "40,60", "20,60", and "10,80". After the data structure transformation, feature 1 contains numerical data in rows #1 to #10 that consists of 70, 10, 50, 40, 70, 10, 40, 40, 20, and 10. Also, feature 2 contains numerical data in rows #1 to #10 that consists of 100, 30, 80, 70, 50, 50, 50, 60, 60, and 80. Since feature a was comma-separated text data, the data is split by commas and converted from text data to numerical data, generating feature 1 and feature 2 in tabular format.

[0080] Figure 15 illustrates the sixth type of data structure transformation. Figure 15(a) schematically shows a data structure transformation, and Figure 15(b) shows a specific example of numerical data obtained from a recipe. In the sixth type, two or more data sets before data structure transformation are combined into two or more numerical data sets after data structure transformation, and the numerical data are replaced based on predetermined rules, for example, by a correspondence table. For example, numerical data is replaced with features obtained by referring to past data and corresponding them with current data. Data 61a and 61b before data structure transformation are combined and replaced based on predetermined rules, and data 62a and 62b are generated after data structure transformation. As a concrete example, Figure 15(b) shows a correspondence table of features before data structure transformation. As shown in correspondence table 61X, feature x corresponds to the numerical data 20, 30, 100, 70, 30, 90, 10, 30, 80, and 90 in rows #1 through #10, respectively. As shown in correspondence table 61A, feature a corresponds to the numerical data 40, 10, 30, 0, 20, 10, 60, 60, 70, and 90 when feature x is 10, 20, 30, 40, 50, 60, 70, 80, 90, and 100, respectively. As shown in the correspondence table 61B, feature b corresponds to the numerical data values ​​50, 10, 30, 0, 0, 40, 20, 90, 20, and 90 respectively when feature x is 10, 20, 30, 40, 50, 60, 70, 80, 90, and 100. In correspondence table 61X, the feature x in row #1 is 20. When feature x = 20, feature a is 10 (row 2 of correspondence table 61A) and feature b is 10 (row 2 of correspondence table 61B). Correspondence table 62 shows the data generated after the data structure transformation. As shown in correspondence table 62, in row #1, feature a is 10 and feature b is 10 and feature 2 is 10. Similarly, in correspondence table 61X, the feature x in row #2 is 30. When feature x = 30, feature a is 30 (row 3 of correspondence table 61A) and feature b is 30 (row 3 of correspondence table 61B). In row #2 of correspondence table 62, feature a is 30 and feature b is 30 and feature 2 is 30. When this transformation process is repeated for correspondence tables 61X, 61A, and 61B, as shown in correspondence table 62, feature 1 is associated with 10, 30, 90, 60, 30, 70, 40, 30, 60, and 70 in rows #1 through #10, and feature 2 is associated with 10, 30, 90, 20, 30, 20, 50, 30, 90, and 20 in rows #1 through #10, respectively. In this way, numerical data is combined and replaced through data structure transformation.

[0081] A specific example of the sixth type is the relationship between the power input to the heater inside the vacuum processing apparatus and the temperature of the inner wall of the vacuum processing apparatus. The temperature of the inner wall of the vacuum processing apparatus during operation is one of the factors that affect the processing results. When the vacuum processing apparatus is in operation, the inside is sealed in order to maintain a vacuum state. If we want to know the temperature of the inner wall of the vacuum processing apparatus, it is necessary to attach measuring equipment to the inner wall, but when the vacuum processing apparatus is in operation, it may be difficult to operate it with the measuring equipment attached to the inner wall due to reasons such as the inside of the vacuum processing apparatus being exposed to plasma discharge. Therefore, the relationship between the input power to the heater and the inner wall temperature is obtained in advance while plasma discharge is not being performed. For example, if the input power is feature quantity x, the measured temperature (actual temperature value) at the top of the inner wall of the vacuum processing apparatus when the first heater is operating is feature quantity a, and the measured temperature at the bottom of the inner wall of the vacuum processing apparatus when the second heater is operating is feature quantity b, then the relationship between the input power of the first heater and the measured temperature of the inner wall is shown in correspondence table 61A, and the relationship between the input power of the second heater and the inner wall temperature is shown in correspondence table 61B. In addition, regarding the input power to the first and second heaters during actual operation of the vacuum processing apparatus with plasma discharge, for example, measurements are taken in a time series, and values ​​are obtained as shown in correspondence table 61X. By performing data structure transformation, for example, feature quantity 1 can be generated as the estimated temperature (estimated temperature value) at the top of the inner wall of the vacuum processing apparatus, and feature quantity 2 can be generated as the estimated temperature at the bottom of the inner wall of the vacuum processing apparatus. Feature quantities 1 and 2 can be applied as display values ​​of measuring instruments. As shown in Types 1 through 6, the recipe files are combined when generating training data by performing data structure transformations on numerical data.

[0082] (An example of data structure transformation) FIG. 16 is a diagram showing an example of data structure conversion. FIG. 16 shows a case where a data structure conversion of summing data of the fourth type is performed on the recipe 1501 shown in FIG. 7(a). As shown in FIG. 7(a), in the processing device 150, the Ar gas is managed as two outputs. From the perspective of the reaction system, it is desirable to handle it as a single variable, so the duplicated item Ar is combined into one item and the flow rate values are summed up.

[0083] (Operation and Effect of Example 1) When the numerical data described in the recipe is directly input into the machine learning model, when the machine learning model represents the reaction system in the processing device, the training of the model cannot be efficiently performed, and it is difficult to improve the accuracy of recipe search. In Example 1, since data structure conversion is performed on the information obtained from the recipe according to the machine learning model, the learning efficiency is improved and the accuracy of recipe search can be improved.

[0084] (Modification of Example 1) When performing recipe search for the edge PC 140a and the processing device 150a, which are so-called offline devices not included in the Lan, the learning data is collected by the user PC 110 via the external recording medium 160. n to the user PC 110 n The user PC 110 sends the learning data to the machine learning system 120, and machine learning is performed in the machine learning system 120.

[0085] (Example 2) Referring to FIGS. 17 to 19, Example 2 will be described. Example 2 assumes, for example, that there are a plurality of processing devices, the data specifications in the processing devices are unified, and they are connected to an external Lan. Example 2 is different from Example 1 in that it uses learning data obtained from a plurality of processing devices. In the following description, the same or equivalent components as those in Example 1 described above are denoted by the same reference numerals, and the description thereof is simplified or omitted.

[0086] (Configuration of Example 2) Figure 17 shows an example 2 of a system to which a recipe search method is applied. The search system 200 includes LanA and LanB. A machine learning system 220 and user PCs 2101 to 210 are located in LanB. N These are the machine learning system 120 and user PCs 1101 to 110 in Example 1, respectively. N This corresponds to the manufacturing data system 230 in LanA, which corresponds to the manufacturing data system 130 in Example 1.

[0087] In LANA, from edge PC 2401 to 240 Q Each of (where Q is a positive integer) corresponds to the edge PC140 in Example 1. Also, processing units 2501 to 250 Q Each of these corresponds to the processing unit 150 in Example 1. In the following description, when the edge PC and processing unit are referred to without specification, the edge PC 240 is used. q (q is a positive integer less than or equal to Q) and processing unit 250 q Use this.

[0088] Furthermore, the manufacturing data system 230 on LANA and the machine learning system 220 on LANB are linked via an API (Application Programming Interface). The AP server 222 can utilize services provided by the Web server 231.

[0089] (Method of Example 2) The search method in Example 2 will be described.

[0090] The actual processing procedure is the same as that of Example 1 shown in Figure 8, so we will explain it by referring to Figure 8. First, the processing device 250 q From, processing unit 250 qThe system obtains data containing information necessary for executing the process (corresponding to step S100 in Figure 8). It obtains numerical data and related information described in the recipe (recipe file name, information indicating the execution result of running the recipe, etc.). The AP server 222 of the machine learning system 220 is connected to the edge PC 240. q And these data are acquired via the manufacturing data system 230. Furthermore, the processing unit 250 q When collecting data sent from DB server 233, AP server 222 retrieves the data via DB server 233, AP server 232, and Web server 231. AP server 222 retrieves the data from multiple processing units 2501 to 250. Q It is possible to obtain numerical data and related information.

[0091] Next, the AP server 222 reads the definition data (corresponding to step S101 in Figure 8). The AP server 222 queries the target and method of data structure transformation for the acquired data.

[0092] Next, the AP server 222 performs a data structure transformation on the acquired data based on the target and method of data structure transformation indicated in the definition data (corresponding to step S102 in Figure 8). The data structure transformation is carried out according to the procedure in the flowchart showing the details of the data structure transformation, as shown in Figure 9(a).

[0093] Next, the AP server 222 trains a machine learning model using the data that has undergone data structure transformation (corresponding to step S103 in Figure 8).

[0094] Next, the user inputs the recipe to be implemented and the target values ​​for the process, and then searches for a recipe. The variables (input data) input to the machine learning model are optimized to suit the machine learning model. In addition, the inputs and outputs of the machine learning model are made to satisfy the constraint information classified from the input data (corresponding to step S104 in Figure 8).

[0095] The AP server 222 converts the input variables back into the processing unit's variables (recipe generation) (corresponding to step S105 in Figure 8). In other words, the searched recipe is expressed in the form of the input variables of the machine learning model. The processing unit 250 converts the searched recipe by performing the reverse transformation of the data structure transformation. q A recipe is generated that has been converted into a format for describing recipes.

[0096] Figure 18 shows the preprocessing procedure of the search method in Example 2. Figure 18 shows a flowchart illustrating the registration and update of the data structure transformation method, which is performed prior to the definition data reading (step S101 in Figure 8) as part of the preprocessing procedure. The user is user PC210 n Alternatively, data structure transformation methods are registered and updated via the interface of AP server 222, based on the numerical data described in the machine learning model and recipe.

[0097] The registration and updating of data structure transformation methods are performed by processing units 2501 to 250. Q This is performed on all of the target processing units.

[0098] The AP server 222 registers and updates data structure transformation methods (step S200). The registered methods may be stored in the memory of the AP server 222 or in the DB server 223.

[0099] (Example of recipe file structure for Example 2) Figure 19 shows an example of the recipe file structure for Example 2.

[0100] Recipe information 2400 is data obtained from LanA, for example, edge PC240 qIt can be obtained from [source]. Recipe information 2400 can also be obtained from DB server 233. Recipe information 2400 is the recipe with processing ID A001. Recipe information 2400 specifies the pressure in Pa units, with the pressure in processing step 1 set to 1.0 Pa, the pressure in processing step 2 set to 0.8 Pa, and the pressure in processing step 3 set to 0.8 Pa.

[0101] Recipe information 2230 is data obtained from LanB, and can be obtained, for example, from DB server 223. Note that recipe information 2230 is available to user PCs 2101-210. N It is also possible to input from here. Recipe information 2230 is a recipe with processing ID B001. Recipe information 2230 specifies the pressure in mTorr, with the pressure in processing step 1 set to 7.5 mTorr, the pressure in processing step 2 set to 6 mTorr, and the pressure in processing step 3 set to 4.5 mTorr.

[0102] The merged data 2220 shows the items included in recipe information 2400 and recipe information 2230 listed. Since both recipe information 2400 and recipe information 2230 indicate processes where Pressure is set, "Pressure" is written in the "Item Name" column. In addition, Pa and mTorr are listed together in the "Unit" column, and a numerical value is listed for each process ID. The table shows duplicate entries for the same pressure setting, even though the units are different, indicating that it is not normalized.

[0103] Data 2221 after data structure transformation shows the result of performing data structure transformation on merged data 2220. Only Pa is listed under "Units". In merged data 2220, B001 in processing step 1 was 7.5 mTorr, but in data 2221 after data structure transformation, it has been transformed to 1.0 Pa. Here, a third type of data structure transformation has been performed, specifically, the numerical value in mTorr units is multiplied by the constant 1 / 7.5 for unit conversion. Similar data structure transformations have been performed for processing steps 2 and 3.

[0104] (Effects and Actions of Example 2) In Example 2, since the machine learning model can be trained using data acquired from multiple processing units, the learning efficiency can be improved and the accuracy of recipe search can be enhanced. Furthermore, since the machine learning system 220 can utilize data from a processing unit located on a different LAN than the one where it is deployed, it becomes possible to build exploration systems, for example, between different business locations and between the company's own systems and those of other companies.

[0105] (Example 3) Example 3 will be described with reference to Figures 20 to 22. Example 3 is assumed to include data other than semiconductor manufacturing equipment, such as literature values ​​and databases, in the machine learning model's dataset. Example 3 differs from Example 1 in that it includes data obtained from sources other than semiconductor manufacturing equipment, such as literature and databases, in the machine learning model's dataset. In the following description, components that are the same or equivalent as those in Examples 1 and 2 described above will be denoted by the same reference numerals, and their descriptions will be simplified or omitted.

[0106] (Configuration of Example 3) Figure 20 shows an example of a system to which a recipe search method is applied. The search system 300 includes LanA, LanB, and LanC. A machine learning system 320 and user PCs 3101 to 310 are located on LanC. NThese are the machine learning system 120 and user PCs 1101 to 110 in Example 1, respectively. N This corresponds to the manufacturing data system 330 in LanB, which corresponds to the manufacturing data system 130 in Example 1.

[0107] In LANB, from edge PC 3401 to 340 Q Each of (where Q is a positive integer) corresponds to the edge PC140 in Example 1. Also, processing units 3501 to 350 Q Each of these corresponds to the processing unit 150 in Example 1. In the following description, when the edge PC and processing unit are referred to without specification, the edge PC 340 is used. q (q is a positive integer less than or equal to Q) and processing unit 350 q Use this.

[0108] The bibliographic management system 360 is a system for managing bibliographic information. The web server 361 receives requests from the LanC's AP server 322. The AP server 362 generates data based on the processing requests from the web server 361. The AP server 362 requests data from the DB server 363 if necessary. The DB server 363 stores information such as specifications, patent documents, academic literature, and research records. The DB server 363 provides data to the AP server 362 in response to its data requests.

[0109] The machine learning system 370 consists of processing units 3501 to 350 M Machine learning is performed based on the data obtained from. The web server 371 accepts requests from the LANC AP server 322. The AP server 372 generates data based on the processing requests from the web server 371. The AP server 372 requests data from the DB server 373 if necessary. The DB server 373 receives data from the processing unit 3501. M It stores information related to machine learning (trained machine learning models, etc.) that was performed based on the data obtained from it. DB server 373 provides data to AP server 372 in response to data requests from AP server 372.

[0110] AP servers 322 and 372 have the same configuration as AP server 122 in Example 1. AP servers 322 and 372 are capable of training machine learning models and exploring data.

[0111] Furthermore, the manufacturing data system 330 on LanB, the literature management system 360 and machine learning system 370 on LanA, and the machine learning system 320 on LanC are linked via API (Application Programming Interface). The AP server 322 can utilize the services provided by the web servers 331, 361, and 371.

[0112] (Method of Example 3) The search method in Example 3 will be described.

[0113] The actual processing procedure is the same as that of Example 1 shown in Figure 8, so we will explain it by referring to Figure 8. First, the processing device 350 q From, processing unit 350 q The system obtains data containing information necessary for executing the process (corresponding to step S100 in Figure 8). It obtains numerical data and related information described in the recipe (recipe file name, information indicating the execution result of running the recipe, etc.). The AP server 322 of the machine learning system 320 is connected to the edge PC 340. q And these data are acquired via the manufacturing data system 330. Furthermore, the processing unit 350 q When collecting data sent from DB server 333, AP server 322 retrieves the data via DB server 333, AP server 332, and Web server 331. AP server 322 also retrieves data from multiple processing units 3501 to 350. Q It is possible to obtain numerical data and related information. Furthermore, AP server 322 sends requests to Web server 371 and Web server 361. Web server 371 generates information in response to the request at AP server 372, or retrieves information from DB server 373, and sends it to AP server 322. Also, Web server 361 retrieves information from DB server 363 via AP server 362 and sends it to AP server 322.

[0114] Next, the AP server 322 reads the definition data (corresponding to step S101 in Figure 8). The AP server 322 queries the target and method of data structure transformation for the acquired data.

[0115] Next, the AP server 322 performs a data structure transformation on the acquired data based on the target and method of data structure transformation indicated in the definition data (corresponding to step S102 in Figure 8). The data structure transformation is carried out according to the procedure in the flowchart showing the details of the data structure transformation shown in Figure 9(a).

[0116] Next, the AP server 322 trains a machine learning model using the data that has undergone data structure transformation (corresponding to step S103 in Figure 8).

[0117] Next, the user inputs the recipe to be implemented and the target values ​​for the process, and then searches for a recipe. The variables (input data) input to the machine learning model are optimized to suit the machine learning model. In addition, the inputs and outputs of the machine learning model are made to satisfy the constraint information classified from the input data (corresponding to step S104 in Figure 8).

[0118] The AP server 322 converts the input variables in reverse to the variables of the processing unit (recipe generation) (corresponding to step S105 in Figure 8). In other words, the searched recipe is expressed in the form of input variables for a machine learning model. The processing unit 350 converts the searched recipe by performing the reverse transformation of the data structure transformation. q A recipe is generated that has been converted into a format for describing recipes.

[0119] Figure 21 shows the pre-processing procedure of the search method in Example 3. Figure 21 shows a flowchart illustrating the registration and update process of the data structure transformation method, which is performed prior to the definition data loading (step S101 in Figure 8).

[0120] The user is user PC310 n Alternatively, data structure transformation methods are registered and updated via the interface of AP server 322, based on the numerical data described in the machine learning model and recipe. The registration and updating of data structure transformation methods are performed by processing units 3501 to 350. Q This is performed on all of the target processing units. Furthermore, the system also registers and updates data structure transformations for the literature data stored in the literature management system 360 and the machine learning models in the machine learning system 370.

[0121] The AP server 322 registers the target and method of data structure transformation. Steps S301 to S303 are performed by the processing units 3501 to 350 Q This process is performed on all data (features) obtainable from the recipes handled in the process. Steps S301 to S303 are also performed on all parameters obtained from the literature management system 360 and the machine learning system 370.

[0122] The AP server 322 determines whether the method for data structure transformation is unknown for the data subject to data structure transformation (step S301). If the method for data structure transformation is unknown (Yes in step S301), the user determines the data subject to data structure transformation and the method for data structure transformation (step S302). For example, the AP server 322 presents the user with candidate methods for data structure transformation and allows them to select one. Once a method for data structure transformation has been determined, the AP server 322 registers the data subject to data structure transformation and the method for data structure transformation (step S303). Even for data for which the method for data structure transformation is known in advance (No in step S301), the type of data structure transformation and the method are registered (step S303).

[0123] For example, if the method for data structure transformation is unknown, the AP server 322 will use the AP server interface or the user PC 310. n Through this, it is also possible to notify the user that the target data and the method for data structure transformation are unknown. The AP server 322 can also automatically consider the data structure transformation method used in other data equivalent to the data in question as the data structure transformation method and register it for data whose data structure transformation method is unknown.

[0124] The registered data and methods may be stored in the storage device of the AP server 322, or they may be stored in the DB server 323.

[0125] (Example of recipe file structure for Example 3) Figure 22 shows an example of the recipe file structure for Example 3.

[0126] Recipe information 3630 is data obtained from LanA, and can be obtained from DB server 363, for example. Recipe information 3630 can also be obtained from DB server 373 or DB server 333. Recipe information 3630 is a recipe with processing ID A001. Recipe information 3630 defines the item "Ar," which is set to 100 mL / min in processing step 1, 20 mL / min in processing step 2, and 10 mL / min in processing step 3.

[0127] Recipe information 3400 is data obtained from LanB, for example, edge PC 340 q It can be obtained from [source]. Recipe information 3400 can also be obtained from DB server 333. Recipe information 3400 is a recipe with processing ID B001. Recipe information 3400 defines the item "MFC1(AR)", and is set to 110 mL / min in processing step 1, 10 mL / min in processing step 2, and 20 mL / min in processing step 3.

[0128] The merged data 3220 shows the case where the items included in recipe information 3630 and recipe information 3400 are listed. Since both process ID A001 and process ID B001 indicate processes related to the supply of Ar gas, the item names are listed with both "Ar" and "MFC1(AR)". The table shows numerical values ​​corresponding to different item names for each processing step.

[0129] Data 3221 after data structure transformation shows what happens when data structure transformation is performed on merged data 3220. Only "Ar" is listed under "Item Name". Recipe information 3630 and recipe information 3400 both show the supply amount of Ar gas, although their "Item Names" are different, and the units are the same. Therefore, in data 3221 after data structure transformation, a third type of data structure transformation has been performed, and recipe information 3630 and recipe information 3400 have the common "Item Name" "Ar" and the values ​​are listed as they are.

[0130] (Effects and Actions of Example 3) Multiple processing units 3501 to 350 Q This may include equipment for which an equipment alliance has not been established. Furthermore, in the 360 ​​literature management system, data specifications may not be standardized, and literature values ​​and databases may include data other than semiconductor manufacturing equipment. It is also conceivable that the retrieved data may include not only original data such as recipes, but also data with standardized or dimensionless values, data where only relationships are presented, and data partially provided from the DB server via API. In such cases, as in Example 3, it is possible to define a unique method for transforming the data structure of the target data. Because the data structure transformation can be defined independently, it is possible to train the machine learning model efficiently.

[0131] (Example 4) Example 4 will be described with reference to Figures 23 and 24. Example 4 is envisioned, for example, as a case where data is shared between customers who have entered into an alliance agreement, based on the terms of the agreement. Example 4 differs from Example 1 in that the machine learning model dataset is acquired across different networks. In the following description, the same or equivalent components as those in Examples 1 to 3 described above will be denoted by the same reference numerals, and their descriptions will be simplified or omitted.

[0132] (Configuration of Example 4) Figure 23 shows an example 4 of a system to which a recipe search method is applied. The search system 400 includes LanA to LanC. Example 4 shows a case where LanC includes a machine learning system 420, LanA includes a manufacturing data system 430, and LanB includes a manufacturing data system 460. This is equivalent to adding another Lan containing a manufacturing data system to Example 2. Although Example 4 shows a case where there are two Lans containing manufacturing data systems, this disclosure can also be applied to cases where there are two or more Lans containing manufacturing data systems.

[0133] In LanC, machine learning system 420 and user PCs 4101 to 410 N This refers to the machine learning system 120 and user PCs 1101 to 110 in Example 1. N The configuration of the machine learning system 420 (Web server 421, AP server 422, DB server 423) corresponds to the configuration of the machine learning system 120 in Example 1 (Web server 121, AP server 122, DB server 123).

[0134] In LANA, from edge PC 4401 to 440 P Each of (P is a positive integer) corresponds to the edge PC140 in Example 1. Also, processing units 4501 to 450 P Each of these corresponds to the processing unit 150 in Example 1. In the following description, when the edge PC and processing unit are referred to without specification, the edge PC 440 is used. p (p is a positive integer less than or equal to P) and processing unit 450 p Use this. Furthermore, in LANB, from edge PC 4701 to 470 Q Each of (Q is a positive integer) corresponds to the edge PC140 in Example 1. Also, processing units 4801 to 480 Q Each of these corresponds to the processing unit 150 in Example 1. In the following description, when the edge PC and processing unit are referred to without specification, the edge PC 470 q (q is a positive integer less than or equal to Q) and processing unit 480 q Use this.

[0135] (Method of Example 4) The search method in Example 4 will be described.

[0136] The actual processing procedure is the same as that of Example 1 shown in Figure 8, so we will explain it by referring to Figure 8. First, the processing device 450 p and processing apparatus 480 q From, processing unit 450 p and processing apparatus 480 qThe system obtains data containing information necessary for executing the process (corresponding to step S100 in Figure 8). It obtains numerical data and related information described in the recipe (recipe file name, information indicating the execution result of running the recipe, etc.). The AP server 422 of the machine learning system 420 is connected to the edge PC 440. p and via the manufacturing data system 430, or the edge PC 470 q And these data are acquired via the manufacturing data system 460. Furthermore, the processing unit 450 p When collecting data sent from DB server 433, AP server 422 retrieves the data via DB server 433, AP server 432, and Web server 431. AP server 422 receives data from multiple processing units 4501 to 450 P Numerical data and related information can be obtained. Furthermore, the processing unit 480 q When collecting data sent from DB server 463, AP server 422 retrieves the data via DB server 463, AP server 462, and Web server 461. AP server 422 then processes the data from multiple processing units 4801 to 480. Q It is possible to obtain numerical data and related information.

[0137] Next, the AP server 422 reads the definition data (corresponding to step S101 in Figure 8). The AP server 422 queries the target and method of data structure transformation for the acquired data.

[0138] Next, the AP server 422 performs a data structure transformation on the acquired data based on the target and method of data structure transformation indicated in the definition data (corresponding to step S102 in Figure 8). The data structure transformation is carried out according to the procedure in the flowchart showing the details of the data structure transformation, as shown in Figure 9(a).

[0139] Next, AP server 422 trains a machine learning model using the data that has undergone data structure transformation (corresponding to step S103 in Figure 8).

[0140] Next, the user inputs the recipe to be implemented and the target values ​​for the process, and then searches for a recipe. The variables (input data) input to the machine learning model are optimized to suit the machine learning model. In addition, the inputs and outputs of the machine learning model are made to satisfy the constraint information classified from the input data (corresponding to step S104 in Figure 8).

[0141] The AP server 422 converts the input variables back into variables for the processing unit (recipe generation) (corresponding to step S105 in Figure 8). In other words, the searched recipe is expressed in the form of input variables for a machine learning model. The processing unit 450 converts the searched recipe by performing the reverse transformation of the data structure transformation. p A format for describing the recipe or processing device 480 q A recipe is generated that has been converted into a format for describing recipes.

[0142] Furthermore, the pre-processing procedure in the search method of Example 4 is the same as the pre-processing procedure in the search method of Example 3 shown in Figure 21.

[0143] The user is user PC410 n Alternatively, data structure transformation methods are registered and updated via the interface of AP server 422, based on the numerical data described in the machine learning model and recipe. The registration and updating of data structure transformation methods are performed by processing units 4501 to 450. P and processing units 4801 to 480 Q This is performed on all of the target processing units.

[0144] Referring to Figure 21, the AP server 422 registers the target and method of data structure transformation. Steps S301 to S303 are performed by processing units 4501 to 450 Q and processing units 4801 to 480 Q This process is performed on all data (features) obtainable from the recipes used in the process.

[0145] The AP server 422 determines whether the method for data structure transformation is unknown for the data subject to data structure transformation (step S301). If the method for data structure transformation is unknown (Yes in step S301), the user determines the data subject to data structure transformation and the data structure transformation method (step S302). For example, the AP server 422 presents the user with candidate data structure transformation methods for selection. Once a data structure transformation method has been determined, the AP server 422 registers the data subject to data structure transformation and the data structure transformation method (step S303). Even for data where the data structure transformation method is known in advance (No in step S301), the data structure transformation method and its method are registered (step S303).

[0146] The registered data and methods may be stored in the storage device of the AP server 422, or they may be stored in the DB server 423.

[0147] (Example of recipe file structure for Example 4) Figure 24 shows an example of the recipe file structure for Example 4.

[0148] Recipe information 4400 is data obtained from LanA, for example, edge PC 440 p It can be obtained from [source]. Recipe information 4400 can also be obtained from DB server 433. Recipe information 4400 is a recipe with processing ID A001. Recipe information 4400 defines the item "Ar", and is standardized using a scaling method that sets the mean to 0 and the variance to 1, with values ​​set to -1.197 in processing step 1, 0.473 in processing step 2, and 1.475 in processing step 3.

[0149] Recipe information 4630 is data obtained from LanB, for example, edge PC 470 qIt can be obtained from [source]. Recipe information 4630 can also be obtained from DB server 463. Recipe information 4630 is a recipe with processing ID B001. Recipe information 4630 defines the item "Ar", and standardization, a scaling method with a mean of 0 and a variance of 1, is applied, and is set to -0.841 in processing step 1, 0.060 in processing step 2, and 0.661 in processing step 3.

[0150] The merged data 4220 shows the list of items included in recipe information 4400 and recipe information 4630. Since both relate to gas supply, the category is labeled "Gas". The item name is "Ar". As shown in the LAN section, a numerical value is provided for each LAN from which the recipe information was obtained.

[0151] Data 4221 after data structure transformation shows what happens when data structure transformation is performed on merged data 4220. Regarding features, since the mean and standard deviation used in standardization (a scaling method where the mean is 0 and the variance is 1) are different, classifications have been established for each LAN, namely "Method A" and "Method B," to represent them. Furthermore, new names have been assigned to the features, indicating a third type of data structure transformation.

[0152] (Effects and Actions of Example 4) Example 4 envisions a case where customers who have entered into an alliance agreement share data based on the terms of the agreement. More specifically, it envisions a case where data that has already been filtered, masked, or anonymized based on the agreement (not the original data) is available. In such a case, Example 4 allows customers to share data based on the alliance agreement, making it possible to efficiently train machine learning models. The difference between Example 4 and Example 2 is that Example 2 focuses on a single customer, whereas Example 4 allows for the sharing of necessary information among multiple customers who have entered into an alliance, based on the terms of each contract.

[0153] (Example 5) Example 5 will be described with reference to Figures 25 and 26. Example 5 is an example that utilizes virtual data, namely virtual measurements, digital twins, machine learning model reuse, and physical models. In Examples 1 to 3, existing (or formerly existing) data such as device specifications and operation records are used, but in Example 5, virtually generated data is used, which is not data that actually exists (or existed) as fact. Specific examples of "virtually generated" data include simple interpolation using a single value or mean, interpolation using approximation formulas based on other real data, inference results or simulation results from other machine learning models. Furthermore, while Example 4 focuses on who the data provider is, Example 5 focuses on how the provided data was generated. Therefore, combinations of Example 4 and this example are also possible (e.g., sharing of trained models or physical models based on alliances between customers). In the following description, components that are the same or equivalent as those in Examples 1 to 4 described above will be denoted by the same reference numerals, and their descriptions will be simplified or omitted.

[0154] (Configuration of Example 5) The configuration of Example 5 is the same as that of Example 4 shown in Figure 23, but with the use of virtual data.

[0155] (Method of Example 5) The search method in Example 5 will be described with reference to Figures 25 and 26.

[0156] Figure 25 shows the actual processing steps of the search method in Example 5. In Figure 25, steps S100, S101, and S102 through S105 are the same as the steps included in Examples 1, 2, 3, and 4. The difference is that it includes the generation of virtual data (step S500). As mentioned above, specific examples of virtual data generation include simple interpolation using a single value or mean, interpolation using approximation formulas based on other real data, inference results or simulation results from other machine learning models, etc.

[0157] One example is to establish a digital twin in LanC that shares a common configuration with the manufacturing data system 430. Instead of the machine learning system 420 acquiring recipe information in LanA or LanB, the digital twin can generate information equivalent to the recipe information. Furthermore, the AP server 422 of the machine learning system 420 can perform interpolation, inference, or simulation on the acquired recipe information to supplement it with virtual data.

[0158] The generated virtual data is used to train the machine learning model (step S103).

[0159] The details of the preprocessing procedure and data structure transformation are the same as in Example 4.

[0160] (Example of the recipe file structure for Example 5) Figure 26 shows an example of the recipe file structure for Example 5.

[0161] Recipe information 4700a and acquired data 4700b are both data acquired from LanB. Recipe information 4700a is data acquired from device a, and acquired data 4700b is data acquired from device b. Devices a and b are, for example, DB server 463 and edge PC 4701 to 470 Q Processing unit 4801~480 QIt is one of the following. Recipe information 4700a consists of multiple recipes, including processing IDs B001 and B002. The recipes are categorized as "Gas" and specify the item "Ar". The recipe with processing ID B001 is set to 90 mL / min in processing step 1 and 30 mL / min in processing step 2. The recipe with processing ID B002 is set to 100 mL / min in processing step 1 and 20 mL / min in processing step 2.

[0162] Acquired data 4700b represents estimated data showing the estimated results of processing when a recipe defined by a recipe file, obtained from a semiconductor substrate inspection or measurement device, is executed using a semiconductor substrate processing device. In acquired data 4700b, processing ID B001 is measured at 100 nm at measurement location A and 130 nm at measurement location B. In processing ID B002 is measured at 120 nm at measurement location A and 110 nm at measurement location B.

[0163] The merged data 5220 shows the case where the items included in recipe information 4700a and acquired data 4700b are enumerated. For each processing ID, the data related to gas from recipe information 4700a and the data related to measurement location from acquired data 4700b are enumerated.

[0164] Machine learning model 520 is identified with the ID A001. Machine learning model 520 is trained using data obtained from LanA.

[0165] The transformed data 5221 shows what happens when the merged data 5220 undergoes a data structure transformation. Compared to the merged data 5220, the transformed data 5221 has the measurement location C (estimated value) added. For example, the machine learning model 520 is trained on measurement location C measured in LanA and estimates the measured value of measurement location C for the data shown in the recipe information 4700a obtained from LanB. A new feature called measurement location C (estimated value) is added, and a first type of data structure transformation has been performed.

[0166] (Effects and Actions of Example 5) Example 5 envisions using virtual data to perform virtual measurements when running a physical model within the processing unit, reuse the model (including the machine learning system), and link with the physical model. Since results from both the real and virtual environments can be used for machine learning, it is possible to train the machine learning model efficiently.

[0167] [Second Embodiment] The second embodiment differs from the first embodiment in that it searches for recipes without performing estimation using machine learning. In the following description, components that are the same as or equivalent to those in the first embodiment described above are denoted by the same reference numerals, and their descriptions are simplified or omitted.

[0168] To obtain information for optimizing recipes—a set of data describing the settings used during processing—based on data obtained from the processing unit, we consider searching through past recipe sets to find recipes that satisfy predetermined search conditions.

[0169] First, the past recipes are retrieved. Possible methods of retrieval include obtaining them from a storage medium connected to the processing unit itself, or from an external system that holds information received from the processing unit. While the processing unit is composed of components necessary for its functionality, the recipes hold the setting values ​​required for each component of the processing unit to operate. Specifically, these include, but are not limited to, gas flow rate and temperature.

[0170] For a processing unit to function correctly, each setting value described in the recipe must be uniquely linked to a component of the target processing unit. In other words, each setting value in the recipe is stored in a way that links it to a name that uniquely identifies each component of the processing unit.

[0171] When considering data like this to search for a desired recipe from a set of recipes, the settings used in the search criteria must be items that uniquely identify each component of the processing unit. However, this presents several challenges.

[0172] The first challenge is that, since the search objective is to optimize the recipe so that the processing result is the desired outcome, the settings to be searched should be settings that correspond to the reaction system occurring in the processing room. However, if we use the data obtained directly from the processing device, the settings to be searched will only be the control settings of the processing device.

[0173] The second challenge is that if the configuration of the processing unit is partially changed, the necessary recipes may no longer be found. For example, consider replacing a temperature control component that controls the overall temperature with a temperature control component that can precisely control local temperatures in order to improve the performance of the unit. In this case, even though the location of temperature control within the processing unit remains the same, the settings change before and after the change. As a result, the settings that can be specified as search targets change before and after the configuration change of the processing unit, which prevents consistent searching before and after the configuration change of the processing unit.

[0174] A third challenge is the difficulty in searching across multiple processing devices. Different devices often have different components, and even if the settings are for the exact same type of gas, the settings used in the recipe may differ due to differences in connected piping, for example. In this case, it is extremely difficult to perform a consistent search across multiple processing devices by specifying some search conditions. On the other hand, there is a strong need to optimize by utilizing past processing results between similar processing devices, and there is a strong desire to create a method that can utilize even a part of past information.

[0175] As a means of solving the above problems, the search method of the second embodiment is a search method for searching for a desired recipe from among multiple recipes defined by a recipe file, and when generating the data to be searched, a conversion process is performed to convert some or all of the items in the recipe file into items common to all of the data to be searched. Furthermore, the conversion process includes at least one of the following processes: adding, deleting, performing arithmetic operations on values, grouping, splitting, and replacing the numerical data. Furthermore, the analysis method of the second embodiment is an analysis method for searching for a desired recipe from among multiple recipes defined by a recipe file and analyzing it in comparison with the processing results associated with the recipe, and when generating the data to be searched, a conversion process is performed to convert some or all of the items in the recipe file into items common to all of the data to be searched.

[0176] Specifically, instead of treating the data sets describing the settings in the processing unit as the data to be searched directly, a data structure transformation is performed using arithmetic operations, statistical processing, and similar calculations to convert them into variables that correspond to the reaction system occurring within the processing unit.

[0177] Let's explain using the example mentioned earlier, where multiple identical gas pipes are connected. Suppose gas pipes 1, 2, and 3 are connected to the processing unit, and gas pipes 1 and 2 carry the same type of gas, while gas pipe 3 carries a different type of gas. In this case, when processing a semiconductor substrate with the processing unit, if it is considered impossible to distinguish which gas pipe the gas introduced from gas pipe 1 and gas pipe 2 arrived from when it reached the semiconductor substrate, the flow rates of the gas flowing in from gas pipe 1 and the gas flowing in from gas pipe 2 are taken as the flow rate of the introduced gas and used as the data to be searched. On the other hand, since there are no other gas pipes connected to gas pipe 3 that carry the same type of gas as the gas connected to gas pipe 3, the flow rate of gas pipe 3 is used as the data to be searched.

[0178] Many other examples of this kind can be cited, including raw material sources, heat sources, electron beam sources, electromagnetic wave sources, light sources, power sources, sound sources, and so on. Furthermore, the way these are handled is not limited to simply dealing with total quantities; some are expressed as differences, such as the relationship between incident and reflected waves from an electromagnetic wave source, where the actual amount introduced into the processing chamber is represented by a difference; others as products or integrals, such as the flow velocity of the raw material fluid and the processing time; and still others as quotients or derivatives, such as the change in heater power for temperature control in the processing chamber. These can be expressed using arithmetic operations and similar methods.

[0179] In this way, the variables in the processing unit are preprocessed to transform the data structure into appropriate variables and their settings, and this transformed data is output as the data to be searched.

[0180] By using the outputted search data, you can perform the desired search and extract the necessary recipes. For example, you can search for a group of recipes where only the pressure setting value inside the processing chamber differs, while all other settings are set to the same value.

[0181] With these recipes and their corresponding processing results, it is possible to analyze the responsiveness of the processing results to specific settings. For example, in the example above, by creating a graph with the pressure setting value on the horizontal axis and the processing result value on the vertical axis, the responsiveness of the processing results to pressure can be visualized. The analysis results obtained in this way can be used as information to optimize the recipe so that the processing result yields the desired outcome.

[0182] Furthermore, the search method and analysis method of the second embodiment can be implemented, for example, in Embodiment 1 of Figure 5, by applying a computer system equipped with a CPU and memory instead of the machine learning system 120. In this case, the memory stores a program that causes the CPU to execute these search methods and analysis methods. The computer system can search for recipes from the manufacturing data system 130. The search method and analysis method of the second embodiment are not limited to their application to Embodiment 1, but can be applied in place of the machine learning system in other embodiments.

[0183] Although embodiments of the present invention have been described above, the present invention is not limited to the embodiments described above, and various modifications are possible without departing from the spirit of the present invention.

[0184] The following describes possible embodiments of the present invention. However, it is not limited thereto. (Aspect 1) In a search method that uses machine learning to estimate when a recipe defined by a recipe file is executed using a semiconductor substrate processing device, In this machine learning process, numerical data describing the recipe is taken as input, estimated data showing the estimated results of the semiconductor substrate processing is output, and a machine learning model is used in which learning is performed using training data that includes the numerical data and data showing the execution results of the recipe being executed. A search method characterized by integrating a dataset containing the numerical data and the execution result data when generating the training data, and further performing a transformation process to convert the structure of the recipe file into a structure corresponding to the machine learning model. (Aspect 2) In the search method described in Embodiment 1, The search method is characterized in that the conversion process includes at least one of the following processes: adding, deleting, performing arithmetic operations on values, grouping, splitting, and replacing numerical data. (Aspect 3) In a search device that estimates the processing performed when a recipe defined by a recipe file is executed using a semiconductor substrate processing device, using machine learning corresponding to the processing model performed by the said processing device, A computer equipped with a learning device that takes numerical data describing the recipe and target values ​​for processing as input, outputs estimated data showing the estimated results of processing the semiconductor substrate by the processing device, and performs learning using training data including the numerical data and data showing the execution results of executing the recipe, The system includes a display that shows a user interface for setting the target value, In generating the aforementioned training data, the computer shall A search device characterized by integrating a dataset containing the numerical data and the data showing the execution results, and further performing a conversion process to convert the structure of the recipe file into a structure corresponding to the model. (Aspect 4) In the search device described in Embodiment 3, The search device is characterized in that the conversion process includes at least one of the following processes: adding, deleting, performing arithmetic operations on values, grouping, splitting, and replacing the numerical data. (Appendix 5) In the search device described in embodiment 3 or 4, The computer is a search device characterized by combining the recipe files when generating the learning data. (Aspect 6) In the search device described in any one of embodiments 3 to 5, A search device characterized by comprising a learning device in which the aforementioned target value is input and the aforementioned numerical data is output. (Aspect 7) In a computer that can communicate with a search device that estimates the processing that occurs when a recipe defined in a recipe file is executed using a semiconductor substrate processing device, using machine learning corresponding to the processing model performed by said processing device, The search device takes numerical data describing the recipe as input and outputs estimated data showing the result of estimating the processing of the semiconductor substrate by the processing device. The learning device is equipped with a learning unit that performs learning using learning data including the numerical data and data showing the execution result of executing the recipe. The computer is characterized by comprising a processor that integrates a dataset including the numerical data and the data indicating the execution result, and further performs a conversion process to convert the structure of the recipe file into a structure corresponding to the model. (Pattern 8) In the search device according to any one of embodiments 3 to 6, A search device characterized in that the numerical data describing the recipe during the search for the optimal recipe is generated in such a way that it satisfies control information including the requirements for operating the processing device. (Aspect 9) In the search device according to any one of embodiments 3 to 6 and 8, The search device is characterized in that the recipe file has a tabular data structure that includes the numerical data and the labels attached to the numerical data. (Aspect 10) In the search device described in any one of embodiments 3 to 6 and 8, 9, A search device that receives a recipe file used in a processing device different from the aforementioned processing device via an API-linked server, and performs a data transformation to merge it with the recipe file of the aforementioned processing device. (Aspect 11) In a semiconductor device manufacturing system that includes a platform where semiconductor manufacturing equipment is connected via a network, and an application is implemented to estimate the processing steps when executing a recipe defined by a recipe file using a semiconductor substrate processing device, using machine learning, In this machine learning process, numerical data describing the recipe is input, estimated data showing the estimated results of the semiconductor substrate processing is output, and a machine learning model is used in which learning is performed using training data that includes the numerical data and data showing the execution results of the recipe. In generating the training data, a dataset including the numerical data and the data showing the execution results is integrated. Furthermore, the semiconductor device manufacturing system is characterized in that a conversion process is performed to convert the structure of the recipe file into a structure corresponding to the machine learning model. (Aspect 12) In a search method for finding a desired recipe from among multiple recipes defined by a recipe file, A search method characterized by performing a conversion process to convert some or all of the items in the recipe file into items common to all of the data to be searched, when generating the data to be searched. (Aspect 13) In the search method described in Embodiment 12, The search method is characterized in that the conversion process includes at least one of the following processes: adding, deleting, performing arithmetic operations on values, grouping, splitting, and replacing the data. (Aspect 14) In an analysis method that searches for a desired recipe from among multiple recipes defined by a recipe file and analyzes it by comparing it with the processing results associated with the said recipe, An analysis method characterized by performing a conversion process to convert some or all of the items in the recipe file into items common to all of the data to be searched, when generating the data to be searched. [Explanation of symbols]

[0185] 11: User-side restriction information, 12: Device-side restriction information, 13: Device information, 14: Other information, 16: Constraints, 17: Training data, 18: Machine learning model, 20: Model, 22: Verification data, 51-59, 60a, 60b, 61a, 61b, 62a, 62b: Data, 100, 200, 300, 400: Search system, 1101~110 N , 2101~210 N , 3101~310 N , 4101~410 N : User PC, 120, 220, 320, 370, 420: Machine learning systems, 121, 131, 221, 231, 321, 331, 361, 371, 421, 431, 461: Web servers, 122, 132, 222, 232, 322, 332, 362, 372, 422, 432, 462: AP (Application) Servers 123, 133, 223, 233, 323, 333, 363, 373, 423, 433, 463: DB (Database) servers, 130, 230, 330, 430, 460: Manufacturing data system, 140, 140a, 2401~240 Q , 3401~340 Q , 4401~440 Q , 4701~470 Q Edge PC, 150, 150a, 2501~250 Q , 3501~350 Q , 4501~450 P , 4801~480 Q : Processing equipment, 160: External storage medium, 1220: CPU, 1221: Memory, 1222: Storage, 1223: Interface, 207: Input / Output device, 1224: Data input / output block, 1225: Data structure conversion block, 1226: ML model optimization block, 1227: Recipe generation block, 230, 330, 460: Manufacturing data system, 360: Literature Management System, 2230, 2400, 3230, 3400, 3630, 4400, 4630, 4700a: Recipe Information 4700b: Acquired data 2220, 3220, 4220, 5220: Merged data 2221, 3221, 4221, 5221: Data after data structure transformation

Claims

1. In a search method for finding a desired recipe from among multiple recipes defined by a recipe file, A search method characterized by performing a conversion process to convert some or all of the items in the recipe file into items common to all of the data to be searched, when generating the data to be searched.

2. In the search method described in claim 1, The search method is characterized in that the conversion process includes at least one of the following processes: adding, deleting, performing arithmetic operations on values, grouping, splitting, and replacing the data.

3. In an analysis method that searches for a desired recipe from among multiple recipes defined by a recipe file and analyzes it by comparing it with the processing results associated with the said recipe, An analysis method characterized by performing a conversion process to convert some or all of the items in the recipe file into items common to all of the data to be searched, when generating the data to be searched.

Images