Search device, data compression processing device, semiconductor device manufacturing system, search method, and data compression processing method

Abstract

A purpose of the present invention is to provide a technique that makes it possible to perform shape prediction by compressing sensing data in consideration of a generation process of sensing data.　For this purpose, a search device according to the present invention, which searches for a recipe for obtaining a desired etching shape in a plasma processing device, is characterized in that: a dimension of sensing data that is acquired by the plasma processing device is defined on the basis of the material of a sample to be etched by the plasma processing device, the material quality of a member constituting a processing chamber in which the sample is etched, or the number of types of radicals derived from the recipe; a compressed feature amount is derived using sensing data that is compressed to the defined dimension; and machine learning parameters are adjusted so that the difference between the derived compression feature amount and an intermediate feature amount is reduced, whereby the intermediate feature amount is calculated from the recipe and the etching shape is predicted on the basis of the calculated intermediate feature amount and the recipe.

Description

Searching apparatus, data compression processing apparatus, semiconductor device manufacturing system, searching method, and data compression processing method

[0001] The present invention relates to a search apparatus, a data compression processing apparatus, a semiconductor device manufacturing system, a search method, and a data compression processing method.

[0002] To obtain the desired processed shape in the semiconductor plasma etching process, it is necessary to adjust the recipe, which is the processing parameters of the etching equipment. Generally, recipes include many variables, such as the flow rate of the etching gas and the bias voltage for accelerating ions in the plasma, and adjustments require trial and error based on experience. For this reason, methods that utilize machine learning for recipe optimization are being investigated.

[0003] In a method using machine learning, a machine learning model that predicts a processed shape from a recipe and sensing data is trained, and a recipe that outputs a target shape is searched for. For example, Patent Document 1 (JP-A-2005-102626) discloses a method in which, in addition to a recipe and shape, the intensity of an emission spectrum observed during the etching process is utilized in the machine learning model. Specifically, Patent Document 1 discloses the following as an invention for improving tool performance by correlating tool operation parameters and material measurements with spectroscopic information: "Systems and methods are provided for tuning and analyzing tool performance through the integration of tool operation data and spectral data associated with the tool. Such integration results in aggregate data that allows, to some extent, learning at least one relationship between selected portions of the aggregate data. Tuning the tool performance includes integrating process recipe parameters that rely at least in part on the learned relationship and can adjust the manufacturing process to achieve sufficient tool performance in response to the implementation of the manufacturing process. The process recipe parameters can be generated by solving an inverse problem based on the learned relationship. Analysis of tool performance can include evaluating an integrated performance scenario, identifying spectral conditions that affect performance, and extracting an endpoint based on at least one time-dependent spectral data."

[0004] Special Publication No. 2014-507801

[0005] When using a neural network as a model for predicting shape from a recipe, achieving high prediction accuracy requires extracting intermediate features from the recipe that are useful for prediction. Here, intermediate features are the output of a certain intermediate layer in the neural network and are expressed as multidimensional real vectors. However, the number of etching trials required to achieve a desired shape is generally fewer than in other fields using neural networks, and the amount of training data acquired is also small, making it difficult to extract intermediate features using conventional methods. Therefore, training a model to minimize the discrepancy between sensing data and intermediate features can be considered as a guide to extract sensing data containing information useful for shape prediction from the recipe. However, using sensing data with high dimensionality as is generally necessitates a large number of parameters in the prediction model, which can lead to overfitting when the training data is limited. To avoid this overfitting, compression of the sensing data is necessary. In compressing sensing data, the dimensionality of the compressed features, which are hyperparameters, is important for extracting the information necessary for shape prediction from the sensing data without excess or deficiency. Conventionally, the dimension of compressed features has been determined by methods such as the elbow method, based on a trade-off between the smallness of dimension and the amount of information contained in the features. However, these methods do not take into account the physical generation process of the data. Therefore, the object of this invention is to provide a technology that can compress sensing data and perform shape prediction while taking into account the generation process of the sensing data.

[0006] In order to solve the above-mentioned problems, a representative searching device of the present invention is a searching device that searches for a recipe in a plasma processing device that will obtain a desired etching shape, wherein the dimension of sensing data acquired by the plasma processing device is defined based on the material of a sample to be etched by the plasma processing device, the material of a component constituting a processing chamber in which the sample is etched, or the number of types of radicals derived from the recipe, compressed features are derived using the sensing data compressed to the defined dimension, machine learning parameters are adjusted so that the difference between the derived compressed features and intermediate features is small, thereby calculating the intermediate features from the recipe, and the etching shape is predicted based on the calculated intermediate features and the recipe.

[0007] According to the present invention, it is possible to compress sensing data and perform shape prediction by taking into account the process of generating the sensing data. Problems, configurations, and effects other than those described above will become apparent from the description of the following embodiments of the invention.

[0008] FIG. 1 is a diagram illustrating an example of the configuration of a semiconductor device manufacturing system. FIG. 2 is a diagram illustrating an example of a table for storing various data in a database. FIG. 3 is a flowchart for determining the dimensions of compressed sensing data in a dimension determination unit. FIG. 4 is a diagram illustrating an example of a GUI when searching for dimensions based on user setting information. FIG. 5 is a diagram illustrating an example of a GUI when searching for dimensions without using user setting information. FIG. 6 is a flowchart for processing an etching shape prediction apparatus to adjust parameters of an intermediate feature calculation unit and a shape calculation unit. FIG. 7 is a diagram illustrating an example of the configuration of a semiconductor device manufacturing system. FIG. 8 is a diagram schematically illustrating emission spectrum intensity as a vector. FIG. 9 is a flowchart for verifying abnormalities in training data in an abnormal data detection unit. FIG. 10 is a diagram illustrating an example of a GUI for setting parameters for detecting abnormal data and confirming the results. FIG. 11 is a diagram illustrating an example of the configuration of a semiconductor device manufacturing system. FIG. 12 is a diagram illustrating an example of an emission wavelength-radical correspondence table. FIG. 13 is a diagram illustrating an example of an emission wavelength-radical correspondence table. FIG. 14 is a flowchart for determining the validity of compressed features in a validity evaluation unit. FIG. 15 is a diagram showing an example of a GUI for setting parameters for evaluating the validity of compressed features and confirming the results.

[0009] Hereinafter, examples will be described with reference to the drawings. However, the present invention is not limited to these examples. In addition, in the description of the drawings, the same parts are denoted by the same reference numerals. Furthermore, when there are multiple components having the same or similar functions, they may be described with the same reference numerals but with different subscripts. Furthermore, when it is not necessary to distinguish between these multiple components, the subscripts may be omitted in the description.

[0010] (Semiconductor Device Manufacturing System) A semiconductor device manufacturing system 1 will be described with reference to Fig. 1. Fig. 1 is a diagram showing an example of the configuration of the semiconductor device manufacturing system 1. The semiconductor device manufacturing system 1 includes a plasma processing apparatus 101, a database 106, an input / output device 112, a learning device 120, and an abnormal data detection unit 121.

[0011] (Plasma Processing Apparatus and Database) The plasma processing apparatus 101 performs plasma processing on a wafer (sample). Specifically, the plasma processing apparatus 101 is an apparatus that etches the surface of the wafer (sample) using plasma generated from an etching gas. The plasma processing apparatus 101 is equipped with a sensor 102 that detects various characteristics and measures sensing data 104 during etching. Note that while the present disclosure illustrates a case where there is one sensor 102, the present disclosure is not limited to this case. The sensor 102 may include multiple sensors for each detection target. Furthermore, although etching is described as an example of plasma processing in the present disclosure, the present disclosure is also applicable to processes other than etching.

[0012] Furthermore, information 105 indicating the processed shape of the sample after etching (hereinafter also simply referred to as the "processed shape") is acquired from the plasma processing apparatus 101. The recipe 103, sensing data 104, and processed shape 105 are stored in a database 106. The database 106 also stores information indicating the material of the wafer and information indicating the configuration of the plasma processing apparatus (for example, the material of the sidewall of the chamber that constitutes the plasma processing apparatus 101) as material information 109.

[0013] (Input / Output Device) The input / output device 112 accepts user requests or presents information to the user. The setting / abnormality information 122 includes setting information indicating the user request transmitted to the abnormal data detection unit 121 and abnormality information which is information notified to the user from the abnormal data detection unit 121. The setting / display information 113 includes setting information which is the user request transmitted to the compression processing device (data compression processing device) 107 and display information which is information notified to the user from the compression processing device 107. The input / output device 112 includes, for example, a display device as an output device, and also includes input devices such as a mouse and a keyboard as input devices.

[0014] (Learning Device) The learning device (searching device) 120 acquires sensing data 104 obtained from the plasma processing device 101 that performs plasma processing, and predicts the shape of a sample (wafer) formed by the plasma processing using a neural network. The learning device 120 searches for a recipe that will obtain a desired etching shape in the plasma processing device 101. The learning device 120 includes a compression processing device 107 and an etching shape prediction device 114. The compression processing device 107 includes a dimension determination unit 108 and a dimension compression unit 110, and the etching shape prediction device 114 includes an intermediate feature calculation unit 115, a shape calculation unit 116, and a parameter adjustment unit 117. Note that the compression processing device 107 and the etching shape prediction device 114 are separated components for convenience, and the present disclosure is not limited to this case.

[0015] Alternatively, the learning device 120 may be configured with general hardware including a processor and memory, and a program stored in the memory may be executed by the processor to perform the functions of the compression processing device 107 and the etching shape prediction device 114. The learning device 120 may be configured with dedicated hardware, or may be implemented by a combination of software and hardware.

[0016] (Compression Processing Device) First, the compression processing device 107 will be described. The compression processing device 107 compresses the sensing data 104 stored in the database 106. The compression is performed in two stages: processing in the dimension determination unit 108 and processing in the dimension reduction unit 110. First, the dimension determination unit 108 determines the post-compression dimension, which indicates how many dimensions the sensing data 104 should be compressed into, based on at least one of material information 109 of the wafer and chamber sidewall and the recipe 103. Then, the dimension reduction unit 110 compresses the sensing data 104 into the determined dimension and derives compressed feature quantities 111. In addition, the input / output device 112 transmits and receives setting and display information 113, which is used to determine the post-compression dimension of the sensing data 104, to and from the dimension determination unit 108.

[0017] (Etching Shape Prediction Device) Next, the etching shape prediction device 114 will be described. The compressed feature 111 can be used to improve prediction accuracy in the etching shape prediction device 114, which predicts the processed shape 105 from the recipe 103. The etching shape prediction device 114 includes an intermediate feature calculation unit 115, a shape calculation unit 116, and a parameter adjustment unit 117. The intermediate feature calculation unit 115 extracts intermediate feature 118 from the recipe 103 by machine learning (e.g., a neural network). In learning the neural network, the parameter adjustment unit 117 adjusts parameters of the neural network so as to reduce the difference between the compressed feature 111 and the intermediate feature 118 extracted from the recipe 103.

[0018] Furthermore, the shape calculation unit 116 outputs a predicted shape 119, which is a result of predicting the shape of an etched wafer using the machine learning model based on the recipe 103 and the extracted intermediate feature amount 118. In learning the machine learning model, the parameter adjustment unit 117 adjusts the parameters of the machine learning model so as to reduce the difference between the predicted shape 119 predicted from the recipe 103 and the intermediate feature amount 118 and the actually measured processed shape 105.

[0019] As described above, the compression processing device 107 and the etching shape prediction device 114 predict the shape of the sample after learning the relationships among the sensing data 104, the processed shape 105, the recipe 103, and the material information 109. Therefore, the combination of the compression processing device 107 and the etching shape prediction device 114 can also be referred to as a learning device 120.

[0020] (Abnormal Data Detector) The operation of the etching shape prediction device 114 is checked by the abnormal data detector 121. The abnormal data detector 121 transmits and receives setting and abnormality information 122 to and from the input / output device 112.

[0021] (Data Stored in Database) FIG. 2 is a diagram showing an example of a table for storing various data in the database 106. Table 201 stores learning data and includes five columns: recipe ID, recipe 103, sensing data 104, processed shape 105, and material information 109. An etching process is performed in the plasma processing apparatus 101 based on the recipe 103 corresponding to the recipe ID, and sensing data 104 acquired during the etching process and processed shape 105 obtained by measuring the wafer after the etching process are included in one record of table 201. The recipes 103 may have a common identifier or different identifiers. Table 201 stores data to be tried to process a sample into a single target shape. In FIG. 2, recipe IDs ranging from 1 to N (N is a positive integer) are shown, indicating that N trials have been performed.

[0022] The sensing data 104 is obtained from a sensor 102 mounted on the plasma processing apparatus 101. For example, the spectral intensity of plasma emission (emission spectrum intensity) is used as the sensing data. The sensor 102 may include multiple sensors. For example, an optical emission spectrometer (OES) 601 (described later) separates the emission spectrum and measures the emission spectrum intensity 202 at each wavelength. A mass flow controller (MFC) measures the mass flow rate 203 of the etching gas. Other sensors may also be included, such as sensors related to a voltage system, a pressure system, and a temperature system. Regarding preprocessing of the sensing data 104, if the sensing data 104 is a time series, a time-averaged value may be stored for each recipe ID. Standardization, normalization, or other standardization processes may or may not be performed on the sensing data 104.

[0023] The processed shape 105 is indicated by numerical values ​​such as the depth and width of a trench formed in the wafer. Each numerical value can be obtained, for example, by photographing the cross section of the etched wafer with an electron microscope and measuring the shape in the image. Furthermore, the material information 109 indicates the material of the wafer that has been etched and the material of the sidewall of the chamber of the plasma processing apparatus 101.

[0024] In the table 201, the recipe 103, the sensing data 104, and the processed shape 105 are each represented by a real vector. Generally, the sensing data 104 has a larger dimension than the recipe 103 and the processed shape 105. The dimensions of the recipe 103 and the processed shape 105 are at most several tens. On the other hand, when the emission spectrum intensity 202 is spectrally analyzed in the range from 200 nm to 800 nm with a resolution of 0.3 nm, it is acquired as 2000-dimensional data. Furthermore, the sensing data 104 excluding the emission spectrum intensity 202 also contains data of several hundred dimensions. Furthermore, with regard to the number of data items, the number of rows in the table 201, i.e., the number of experiments performed to process one target shape, is generally at most several tens.

[0025] Generally, in order to handle high-dimensional sensing data 104 using a machine learning model, the number of model parameters needs to be large. However, when a model with a large number of parameters is trained using a small amount of data, the model parameters over-fit to the small amount of data, resulting in overfitting, which reduces generalization performance for test data. Therefore, it is necessary to reduce the number of model parameters by compressing the high-dimensional sensing data 104 into a lower dimension, thereby avoiding overfitting.

[0026] When compressing the sensing data 104, it is reasonable to consider the generation principle of the sensing data 104 in order to determine the number of dimensions to compress the data into. The sensing data 104 is obtained by observing the chemical reaction between the plasma generated from the etching gas and the sample in the chamber of the plasma processing apparatus 101 from the perspective of each sensor. According to the generation principle of the sensing data 104, it is possible to generate data equivalent to the sensing data 104 from the number of radical types, and further, according to the relationship between the number of radical types and the sensing data 104, the sensing data 104 should be compressed based on the number of radical types.

[0027] 3 is a flowchart showing how the dimension determination unit 108 determines the dimension of the compressed sensing data 104. First, the dimension determination unit 108 starts processing by receiving user setting information from the recipe 103, sensing data 104, material information 109, and setting and display information 113 (step 301).

[0028] Next, the dimension of the sensing data acquired by the plasma processing apparatus 101 is determined based on the material of the sample to be etched by the plasma processing apparatus 101, the material of the components constituting the processing chamber in which the sample is etched, or the number of types of radicals derived from the recipe. Specifically, the dimension determination unit 108 specifies the number of types of radicals from at least one of the material information of the wafer, the material information of the chamber sidewall (material information 109), and the recipe 103, and based on the user's setting information (step 302). For example, the etching gas Cl in the recipe 103 ２ , N.F. ３ , N ２ , O ２ , CO ２ When only radicals are used as the causative substances of the emission spectrum, the number of radical types is set to 5. If a search for dimensions is to be performed based on the user's setting information, the process proceeds to step 304; if a search based on the user's setting information is not to be performed, the dimension of the compressed feature is determined to be the specified number of radical types, and the process proceeds to step 308 (step 303).

[0029] When searching for dimensions, the dimension compression unit 110 sets candidate dimensions (hereinafter also referred to as "compressed dimension candidates") after compressing the sensing data 104 based on the set number of radical types (step 304). For example, it is conceivable to set upper and lower limits so as to include the number of radical types, and to set the interval between the upper and lower limits as the compressed dimension candidates.

[0030] Next, the dimension determination unit 108 starts an iterative process for each candidate compression dimension (step 305). In the iterative process, the dimension determination unit 108 attempts to compress the sensing data 104 for each candidate dimension and calculates an index value indicating the quality of compression from the sensing data 104 before compression and the compressed sensing data 104, which are the compressed sensing data (step 306). The dimension reduction unit 110 may use a principal component analysis, autoencoder, nonnegative matrix decomposition, or other dimension reduction techniques during the compression trials. Alternatively, a faster dimension reduction technique may be used for hyperparameter search. Examples of index values ​​indicating the quality of compression include the reconstruction error and singular values ​​corresponding to the principal components in principal component analysis. Here, the reconstruction error can be expressed by the Euclidean distance between the sensing data restored from the compressed features and the sensing data before compression.

[0031] After each iteration, the dimension determination unit 108 determines whether to adopt the candidate dimension based on a condition related to an index value relating to the quality of compression (step 307). For example, if the reconstruction error is equal to or less than a predetermined threshold, the dimension is adopted. If only one dimension is adopted in step 307, the adopted dimension is determined (defined) as the dimension after compression. If multiple dimensions satisfy the condition in step 307, the most suitable one (e.g., the dimension with the smallest reconstruction error) is determined as the dimension after compression.

[0032] Finally, the dimension determination unit 108 transmits the determined dimensions to the dimension compression unit 110, and the process ends (step 308). The dimension compression unit 110 compresses the sensing data 104 into the determined dimensions and derives compressed features 111.

[0033] (GUI Example) Next, a GUI (Graphical User Interface) between the semiconductor device manufacturing system 1 and the user will be described with reference to FIGS. 4 and 5. To determine the dimensions (post-compression dimensions) of the compressed feature 111 using the above method, it is necessary to set the number of radical types, the search range of dimensions, and the conditions for adopting candidate dimensions. To adjust the conditions for adoption, a GUI can be used between the compression processing device 107 and the user. FIG. 4 is a diagram showing an example of a GUI when searching for dimensions based on user setting information.

[0034] 3, for example. The screen 401 is displayed on the display device of the input / output device 112. The screen 401 includes setting and display information 113, and the display information is notified to the user via the screen 401, and the user's setting information is also transmitted to the compression processing device 107. The recipe and material display section 402 displays material information 109 indicating the material of the wafer and the material of the chamber sidewall, and the etching gas included in the recipe 103.

[0035] The user refers to the recipe and material display section 402 and selects a method for specifying the number of radical types in the radical type setting section 403. Examples of methods for specifying the number of radical types include using the number of etching gas types in the recipe as is, including the number of etching gas types and radicals generated by reaction products, and a method set by the user. In Fig. 4, the method for using the number of etching gas types in the recipe as is is selected, and this is transmitted to the compression processing device 107 as user setting information.

[0036] The search execution setting section 404 displays an option for selecting whether or not to perform a dimension search. In Fig. 4, performing a dimension search is selected, and this is transmitted to the compression processing device 107 as user setting information.

[0037] The search range of the dimension is set in the search range setting unit 405 based on the number of radical types specified in the radical type number setting unit 403. In Fig. 4, z (for example, z = 3) is set as the lower limit of the search range of the dimension, and y (for example, y = 7) is set as the upper limit of the search range of the dimension.

[0038] The adoption condition setting unit 406 sets the adoption conditions for adopting dimensions that are candidates for compression dimensions. If it is necessary to set parameters such as thresholds when determining the adoption conditions, the user inputs the parameters in an input form. In Fig. 4, the adoption condition is set to be that the index value relating to the quality of compression is equal to or greater than a threshold value, or equal to or less than a threshold value. Furthermore, a threshold value t is set as the threshold value.

[0039] The trial result display section 407 displays the results of the iterative process that has been tried based on the input values ​​and the conditions used. As an example of how the results may be displayed, a plot is shown with the compressed dimension on the horizontal axis and an index value related to the quality of compression on the vertical axis, but the result display method is not limited to this, and other displays may be used. Figure 4 shows a case where the index value related to the quality of compression is below a threshold value t between dimensions x and y.

[0040] Fig. 5 is a diagram showing an example of a GUI when a dimension search is not performed in the user's setting information. Fig. 5 shows a case where not performing a dimension search is selected in the search execution setting unit 404. In this case, the number of radical types x specified by the user is displayed in the radical type number setting unit 403, and is determined as the dimension of the compressed feature.

[0041] Example 2 describes in detail how the parameters of the neural network are adjusted in the intermediate feature calculation unit 115. Example 2 shows how the machine learning parameters are adjusted so as to reduce the difference between the derived compressed feature 111 and the intermediate feature 118, thereby calculating the intermediate feature from a recipe, and how the etching shape is predicted based on the calculated intermediate feature and the recipe. In the following description, components that are the same as or equivalent to those in Example 1 above are denoted by the same reference numerals, and their description will be simplified or omitted.

[0042] The compressed feature 111 is used to improve prediction accuracy in an etching shape prediction device 114 that predicts a processed shape 105 from a recipe 103. Generally, when a neural network is used as a shape prediction model, it is difficult to automatically learn intermediate feature values ​​of the neural network from a small amount of data. Therefore, by setting targets not only for the output of the shape prediction model but also for the intermediate layer, it is believed that guidance for extracting intermediate feature values ​​can be achieved. The compressed feature 111 obtained by dimensionally compressing the sensing data 104 is suitable as a target for the intermediate layer. This is because the sensing data 104 are observed values ​​in an actual etching process and contain information useful for shape prediction. By guiding the intermediate feature 118 so as to extract compressed feature values ​​111 useful for shape prediction from the recipe 103, it is expected that the accuracy of shape prediction will be improved.

[0043] From the viewpoint of ease of data acquisition, the compressed feature 111 is suitable as a target for the intermediate layer. Furthermore, since the sensing data 104 can be acquired not only from a patterned wafer on which a pattern is arranged but also from a bare wafer in many cases, it is possible to acquire a large amount of data inexpensively. Furthermore, measurement of the sensing data 104 does not require an electron microscope, and much of the sensing data 104 can be easily acquired.

[0044] 6 is a flowchart of a process for adjusting parameters of the intermediate feature value calculation unit 115 and the shape calculation unit 116 in the etching shape prediction device 114. First, the etching shape prediction device 114 receives the recipe 103, the compressed feature value 111, and the processed shape 105 as learning data for parameter adjustment, and starts the process (step 501). The compressed feature value 111 may be standardized before being used in the etching shape prediction device 114. The standardization process is a scaling process in which the average of each dimension in the learning data is set to 0 and the variance is set to 1. The standardization process may be performed in the database 106 or in the parameter adjustment unit 117, for example.

[0045] Next, the parameter adjustment unit 117 adjusts the parameters of the neural network in the intermediate feature calculation unit 115 using the recipe 103, the compressed feature 111, and the intermediate feature 118 (step 502). When adjusting the parameters of the neural network in the intermediate feature calculation unit 115, the compressed feature 111 is used as a target for the intermediate feature 118. Here, using it as a target means that the neural network f, the number of training data N, and the recipe r of the nth training data are used as the target. ｎ (For example, the recipe corresponding to recipe ID=n among the recipes 103), the recipe r among the compressed features 111 ｎ The compressed feature c is a feature corresponding to the sensing data 104 observed from etching using ｎ , the loss function l is set to, the parameter θ of the neural network f is updated so as to reduce the value of (1).

[0046] Thereafter, the shape calculation unit 116 trains the machine learning model in the shape calculation unit 116 to predict the machined shape 105 obtained using the recipe 103, from the intermediate feature 118 obtained by the neural network f whose parameter θ has been adjusted in the intermediate feature calculation unit 115 and the recipe 103 (step 503). The machine learning model in the shape calculation unit 116 is not limited to the neural network, and other machine learning models can be applied.

[0047] Finally, the etching shape prediction device 114 (intermediate feature amount calculation unit 115) holds the adjusted parameter θ and ends the process (step 504).

[0048] After learning, the etching shape prediction device 114 predicts the processed shape from an arbitrary recipe using the intermediate feature calculation unit 115 with the parameter θ adjusted and the shape calculation unit 116. Note that the compressed feature 111 is not necessary at the time of prediction, and the output of the neural network in the intermediate feature calculation unit 115 and the recipe 103 are input to the machine learning model in the shape calculation unit 116.

[0049] In the third embodiment, a case where optical spectrum intensity is used as sensing data will be described in detail. In the following description, the same or equivalent components as those in the first embodiment will be denoted by the same reference numerals, and the description thereof will be simplified or omitted.

[0050] (Semiconductor device manufacturing system) FIG. 7 is a diagram showing an example of the configuration of a semiconductor device manufacturing system 1A. The semiconductor device manufacturing system 1A of Example 3 differs from the semiconductor device manufacturing system 1 of Example 1 in that the plasma processing apparatus 101 has an optical emission spectroscopic sensor 601 and the database 106 stores an optical emission spectrum intensity 202. The optical emission spectroscopic sensor 601 measures the optical emission spectrum intensity 202 during etching processing in the plasma processing apparatus 101. The measured optical emission spectrum intensity 202 is stored in the database 106. Note that, after etching, the processed shape 105 of the sample is also acquired from the plasma processing apparatus 101. The recipe 103, the optical emission spectrum intensity 202, and the processed shape 105 are stored in the database 106.

[0051] The emission spectrum intensity 202 stored in the database 106 is compressed by the compression processing device 107. The compression is performed in two stages: processing by the dimension determination unit 108 and processing by the dimension reduction unit 110. First, the dimension determination unit 108 determines the dimension after compression, indicating the number of dimensions into which the emission spectrum intensity 202 should be compressed, based on at least one of the material information 109 of the wafer and chamber sidewall and the recipe 103, the number of radical types specified by the user's setting information in the setting and display information 113, an index value related to the degree of compression, and the user's setting information. Then, the dimension reduction unit 110 compresses the emission spectrum intensity 202 into the determined dimension and derives the compressed feature 111. In addition, the input / output device 112 exchanges the setting and display information 113 for determining the dimension after compression of the emission spectrum intensity 202 with the dimension determination unit 108.

[0052] (Emission Spectrum Intensity) Considering the principle of emission spectrum generation, the emission spectrum intensity can be derived naturally from the number of radical types. In such cases, a method of compressing the emission spectrum intensity 202 based on the number of radical types is naturally adopted. Referring to FIG. 8 , compressing the emission spectrum intensity 202 to a dimension corresponding to the number of radical types involved in the etching chemical reaction is explained as being rational in that it refers to the physical generation process of the emission spectrum. FIG. 8 is a diagram schematically illustrating the emission spectrum intensity as a vector. The emission spectrum is emitted by the transition of electron orbitals of reactants in the etching chemical reaction. The wavelength of the emitted light (hereinafter simply referred to as the "emission wavelength") is a unique value for each reactant. Therefore, the emission spectrum intensity 202 resulting from each substance can be defined as a vector having the dimension of the spectral resolution.

[0053] For example, consider a case where an emission spectrum intensity 202 indicated by a vector 701 shown in FIG. 8B is observed at a certain time during an etching process. In the vector 701, the component with an emission wavelength of 200 nm is 1, the component with an emission wavelength of 400 nm is 4, and the component with an emission wavelength of 600 nm is 2. In the following description, for ease of understanding, the case where the emission wavelength is three wavelengths of 200 nm, 400 nm, and 600 nm is described. In reality, the emission wavelength is set according to the spectral resolution, and the present disclosure can also be applied to a case where there are more than three emission spectrum intensity components.

[0054] In this case, it is known that the observed emission spectrum intensity 202 is due to substances A, B, and C. Meanwhile, assume that the emission spectrum intensities per unit concentration of substances A, B, and C are known to be 702A, 702B, and 702C, respectively. As shown in FIG. 8( a), the emission spectrum intensity 702A per unit concentration of substance A is 2 for the component with an emission wavelength of 200 nm, 0 for the component with an emission wavelength of 400 nm, and 0 for the component with an emission wavelength of 600 nm. The emission spectrum intensity 702B per unit concentration of substance B is 0 for the component with an emission wavelength of 200 nm, 1 for the component with an emission wavelength of 400 nm, and 1 for the component with an emission wavelength of 600 nm. The emission spectrum intensity 702C per unit concentration of substance C is 0 for the component with an emission wavelength of 200 nm, 4 for the component with an emission wavelength of 400 nm, and 0 for the component with an emission wavelength of 600 nm.

[0055] Next, the relationship between the emission spectrum intensity per unit concentration shown in FIG. 8A and the observed emission spectrum intensity 202 shown in FIG. 8B is derived. The concentration ratios of the substances chemically reacted during etching are measured and determined, and weighting coefficients 703 are set to 0.5 for substance A, 2.0 for substance B, and 0.5 for substance C based on the concentration ratios. As shown in FIG. 8B, vector 701 is derived by multiplying the emission spectrum intensities per unit concentration 702A, 702B, and 702C for each substance by weighting coefficient 703 and taking the sum. This vector 701 corresponds to a vector representation of the observed emission spectrum intensity 202. The weighting coefficient 703, expressed as a vector by arranging the weights, is then regarded as the compressed feature 111 of the emission spectrum intensity 202 shown in vector 701.

[0056] The above explanation is based on the assumption that the emission spectrum intensity per unit concentration for each substance is known. If the emission spectrum intensity per unit concentration for each substance is unknown, the dimension reduction unit 110 must derive it from the training data using a machine learning model. However, machine learning models are not necessarily suitable for accurately determining the emission spectrum intensity per unit concentration for a substance from training data, and compressing the emission spectrum intensity per unit concentration to dimensions corresponding to the number of radicals involved in the etching chemical reaction may not result in optimal results. Therefore, when the emission spectrum intensity per unit concentration for each substance is unknown, a method is adopted in which compression is attempted over a certain range of dimensions, and the number of dimensions to which the emission spectrum intensity 202 should ultimately be compressed is determined based on the results, as shown in Example 1.

[0057] In Example 3, the abnormal data detection unit 121 will be described in detail. In Example 4, it is shown that if the difference between the intermediate feature 118 and the compressed feature 111 is greater than a predetermined threshold, the recipe 103 is determined to be abnormal, or the sensing data 104 is determined to be abnormal. For example, if the difference between the intermediate feature 118 and the compressed feature 111 is greater than a predetermined threshold for the recipe 103 and the sensing data 104 used in the adjustment by the parameter adjustment unit 117, the abnormal data detection unit 121 determines that the recipe 103 or the sensing data 104 is abnormal. A specific description will be given below. In the following description, components that are the same as or equivalent to those in Example 1 above will be denoted by the same reference numerals, and their description will be simplified or omitted.

[0058] (Detection of Abnormal Data) As described above, in semiconductor etching, the number of experiments generally performed to obtain a single target shape is small, resulting in a small amount of training data. Therefore, if there is an abnormality in the training data, it will have a significant impact on prediction accuracy. Therefore, it is desirable for the training device 120 to be equipped with a mechanism for removing training data that includes errors during measurement, etc. Therefore, after adjusting the parameters in the etching shape prediction device 114 (parameter adjustment unit 117), the abnormal data detection unit 121 verifies whether there are any abnormalities in the training data used for the adjustment.

[0059] 9 is a flowchart showing how the abnormal data detection unit 121 verifies abnormalities in training data. First, the abnormal data detection unit 121 receives the intermediate features 118 extracted from the recipe 103 and the compressed features 111 of the sensing data 104 observed using the recipe 103, and starts processing (step 801).

[0060] Next, the abnormal data detection unit 121 calculates the difference between the intermediate feature 118 and the compressed feature 111 (step 802). If the difference is large, there is a possibility that either the recipe 103 or the sensing data 104 is abnormal. Therefore, the abnormal data detection unit 121 determines that a recipe ID for which the difference between the intermediate feature 118 and the compressed feature 111 exceeds a predetermined threshold is abnormal data (step 803). Here, the difference between the intermediate feature 118 and the compressed feature 111 is measured using the loss function l used by the parameter adjustment unit 117 in the neural network training described in Example 2. The threshold value is input from the input / output device 112 as user setting information among the setting / abnormality information 122.

[0061] Finally, the abnormal data detection unit 121 transmits the abnormality information from the setting / abnormality information 122 to the input / output device 112, and the process ends (step 804). The abnormality information includes, for example, a recipe ID, and the values ​​of the intermediate feature 118 and the compressed feature 111. The input / output device 112 notifies the user of the received abnormality information by, for example, displaying it on a GUI.

[0062] (Example of GUI) Fig. 10 is a diagram showing an example of a GUI for setting parameters for detecting abnormal data and confirming the results. A screen 901 is displayed on, for example, the display device of the input / output device 112. In a threshold setting section 902, the user sets a threshold value to be used for abnormality determination. In Fig. 10, 0.5 is set as the threshold value. In an abnormality display section 903, the result of abnormality determination based on the set threshold value is displayed. In Fig. 10, the loss coefficient of recipe ID = xxxx is shown to be 0.8, and the loss coefficient of recipe ID = yyyy is shown to be 0.6. The result may be displayed only for the recipe 103 determined to be abnormal, or for any recipe 103.

[0063] Example 5 describes dimensionality reduction in detail. In Example 5, the spectral intensity (emission spectral intensity 202) is decomposed into a weighted sum, which is the sum of component vectors multiplied by weighting coefficients. The compressed feature is represented as a vector in which the weighting coefficients are arranged. A component vector having a wavelength whose spectral intensity is greater than a predetermined threshold is detected. Based on the correlation between the wavelength of the spectral intensity emitted by the radicals and the type of the radicals, the etching gas and the reaction products generated by etching are associated with the detected component vector. Furthermore, the spectral intensity is decomposed into a weighted sum, which is the sum of component vectors from which the weighting coefficients have been evaporated, by non-negative matrix decomposition. In the following description, components identical or equivalent to those in Example 1 above are denoted by the same reference numerals, and their description will be simplified or omitted.

[0064] (Nonnegative Matrix Decomposition) Nonnegative matrix decomposition is a technique for optimizing W and H so that a matrix (nonnegative matrix) X, which takes nonnegative values, is approximated by the product WH of two nonnegative matrices. In particular, when the emission spectrum intensity 202 is used as the sensing data 104, the emission spectrum intensity 202 takes on 0 or a positive value (nonnegative value) due to the nature of intensity, and therefore nonnegative matrix decomposition can be used as a compression algorithm. In this case, each row of X corresponds to a piece of training data, and each column corresponds to the emission spectrum intensity 202 of each emission wavelength. Each piece of training data may be a statistical value such as a time average for each recipe ID of the recipe 103, or the value at each time may be used as is. Furthermore, W is a matrix in which the compressed feature 111 of each piece of training data is arranged in each row. In nonnegative matrix decomposition, as shown in (2), the i-th row of X, i.e., the i-th training data, is approximated by the sum of each row of H (hereinafter, each row of H is referred to as a component vector, and multiple component vectors corresponding to multiple rows included in H are also referred to as "multiple component vectors 1001") weighted by the compressed feature 111. After H is learned from the training data, the emission spectrum intensity 202 can be approximately restored by multiplying the compressed feature 111 by H from the right.

[0065] (Semiconductor Device Manufacturing System) Figure 11 is a diagram showing an example of the configuration of a semiconductor device manufacturing system 1B. The semiconductor device manufacturing system 1B of Example 5 differs from the semiconductor device manufacturing system 1 of Example 1 in that the plasma processing apparatus 101 has an optical emission spectroscopic sensor 601, the database 106 stores an optical emission spectrum intensity 202 and an optical emission wavelength-radical correspondence table 1002, and the compression processing apparatus 107 has a validity evaluation unit 1003. Note that in Figure 10, in order to explain dimensionality reduction, components corresponding to the etching shape prediction apparatus 114 and the abnormal data detection unit 121 in Figure 1 (Example 1) are omitted.

[0066] The optical emission spectroscopic sensor 601 measures the optical emission spectrum intensity 202 during the etching process in the plasma processing apparatus 101. The measured optical emission spectrum intensity 202 is compressed by the compression processing apparatus 107A.

[0067] The compression processing device 107A includes a dimension determination unit 108, a dimension reduction unit 110, and a validity evaluation unit 1003. The dimension determination unit 108 determines a dimension after compression, which indicates how many dimensions the emission spectrum intensity 202 should be compressed into, based on at least one of material information 109 of the wafer and chamber sidewall and the recipe 103, the number of radical types specified by the user's setting information in the setting and display information 113, the reconstruction error of the emission spectrum intensity 202 by non-negative matrix decomposition, and the setting information by the user.

[0068] The dimension reduction unit 110 compresses the emission spectrum intensity 202 into the determined dimensions by non-negative matrix decomposition, and derives the compressed feature 111. The dimension reduction unit 110 decomposes the emission spectrum intensity 202 into a weighted sum, which is the sum of multiple component vectors multiplied by weighting coefficients (Equation (2)), expresses the compressed feature 111 as a vector in which the weighting coefficients are arranged, detects component vectors 1001 having wavelengths whose spectral intensities are greater than a predetermined threshold, and associates the detected component vectors 1001 with gases used in the plasma processing (etching gases) and reaction products (radicals) generated in the plasma processing based on the relationship between the types of radicals and the wavelengths of the specific spectral intensities emitted by the radicals (emission wavelength-radical correspondence table 1002).

[0069] Specifically, in non-negative matrix decomposition, the emission spectrum intensity 202 is expressed as the sum of a plurality of component vectors 1001 weighted by the compressed feature 111. Then, at least one radical is associated with each of the plurality of component vectors 1001 by using a predetermined emission wavelength-radical correspondence table 1002. By associating a radical with each of the plurality of component vectors 1001, the compressed feature 111 can be regarded as an index indicating the degree of weight with which the component vector resulting from each radical is included in the emission spectrum intensity 202.

[0070] The validity evaluation unit 1003 then evaluates whether the compressed features 111 are valid, and transmits and receives setting and validity information 1004 to and from the input / output device 112. The setting and validity information 1004 includes setting information indicating a user request that is transmitted to the validity evaluation unit 1003, and validity information that is information notified to the user by the validity evaluation unit 1003.

[0071] (Emission wavelength-radical correspondence table) FIGS. 12 and 13 are diagrams showing an example of the emission wavelength-radical correspondence table 1002. FIG. 12 shows the emission wavelength-radical correspondence table 1002A, and FIG. 13 shows the emission wavelength-radical correspondence table 1002B. In the etching reaction, each substance emits light by emitting light with at least one specific wavelength. Also, multiple substances may emit light at the same wavelength. The emission wavelength-radical correspondence table 1002 is created by compiling the correspondence between such substances (radicals) and their emission wavelengths in a table.

[0072] 12 shows the emission wavelength for each substance in the emission wavelength-radical correspondence table 1002A. The substances (radicals) with an emission wavelength of xx nm are substance A and substance B. The substance with an emission wavelength of yy nm is substance C. The substances with an emission wavelength of zz nm are substance A, substance D, and substance E.

[0073] Furthermore, the emission wavelength-radical correspondence table 1002B in Figure 13 shows the emission spectrum intensity per unit concentration for each substance in addition to the emission wavelength. For substance A, the intensity at the emission wavelength xx nm is 2000, the intensity at the emission wavelength zz nm is 500, and the intensities at the emission wavelengths 200 nm, yy nm, and 800 nm are 0. For substance B, the intensity at the emission wavelength xx nm is 300, and the intensities at the emission wavelengths 200 nm, yy nm, zz nm, and 800 nm are 0. For substance C, the intensity at the emission wavelength yy nm is 400, and the intensities at the emission wavelengths 200 nm, xx nm, zz nm, and 800 nm are 0. For substance D, the intensity at the wavelength zz nm is 800, and the intensities at the emission wavelengths 200 nm, xx nm, yy nm, and 800 nm are 0. The emission wavelength-radical correspondence table 1002B includes information useful for determining, for example, the true values ​​of the emission spectrum intensities 702A to 702C per unit concentration.

[0074] 14 is a flowchart showing how the validity evaluation unit 1003 determines the validity of the compressed feature 111. The validity of the compressed feature 111 is evaluated by analyzing the relationship between the types of radicals corresponding to each component of the compressed feature 111, the gas flow rate specified in the recipe 103, and the value of each component of the compressed feature 111. Specifically, the validity evaluation unit 1003 evaluates the validity of the compressed feature 111 by analyzing the gas flow rate in the recipe 103, the value of each component of the compressed feature 111, and the relationship between the types of radicals corresponding to each component of the compressed feature 111. First, the validity evaluation unit 1003 receives a plurality of component vectors 1001, the recipe 103, the compressed feature 111, and the emission wavelength-radical correspondence table 1002, and starts processing (step 1201).

[0075] Next, the validity evaluation unit 1003 detects wavelengths whose values ​​(intensities) are greater than a given threshold value in the plurality of component vectors 1001 (step 1202). The threshold value is input from the input / output device 112 as user setting information in the setting / validity information 1004.

[0076] Thereafter, using a given emission wavelength-radical correspondence table 1002, the radical corresponding to the detected wavelength is associated with the component vector 1001 (step 1203). Here, when associating wavelengths, a range of wavelengths can be set in consideration of measurement errors.

[0077] Next, as an index for determining the validity, the validity evaluation unit 1003 calculates a correlation matrix between the recipe 103 (especially each etching gas flow rate) and the compressed feature 111 (step 1204). Here, the correlation matrix of two multidimensional random variables a and b is a matrix in which the (i, j) components have a i and b j The etching gas flow rate of the recipe 103 corresponds to a, and the compressed feature 111 corresponds to b. However, it is not necessary to calculate correlation coefficients for all components.

[0078] The validity evaluation unit 1003 then determines the validity of the compressed features 111 based on whether the calculated correlation coefficient is greater than a given threshold value (step 1205). The threshold value is input from the input / output device 112 as user setting information.

[0079] Finally, the validity evaluation unit 1003 transmits information related to validity (validity information) to the input / output device 112, and terminates the process (step 1206). By calculating the correlation coefficient, it is possible to confirm whether the emission spectrum intensity 202 of related radicals is strong when the flow rate of a certain etching gas is high. The presence or absence of a correlation is determined, for example, by whether the radicals corresponding to the compressed feature 111 interpreted as the type of etching gas in the recipe 103 contain a common atom. Furthermore, if there is no validity, the dimension determination unit 108 can adjust the dimension again.

[0080] 15 is a diagram showing an example of a GUI for setting parameters for evaluating the validity of the compressed features 111 and confirming the results. A screen 1301 is displayed on, for example, the display device of the input / output device 112. The screen 1301 includes the setting / validity information 1004, and the validity information is notified to the user via the screen 1301, and the user's setting information is also sent to the validity evaluation unit 1003.

[0081] The intensity threshold setting unit 1302 sets a threshold for the emission spectrum intensity 202 in order to extract an emission spectrum intensity that associates a radical with each component vector 1001 from the emission spectrum intensity 202. In FIG. 15 , as a method for setting the threshold, a method for specifying a constant or a method for setting p % of the maximum value in each component vector as the threshold is displayed. In FIG. 15 , the latter method is selected, and p is set to 50. The selected threshold condition is notified to the validity evaluation unit 1003 as setting information.

[0082] The correlation coefficient threshold setting unit 1303 sets a threshold value for the correlation coefficient that serves as a basis for validity evaluation. In Fig. 15, 0.5 is entered in the input form. The entered threshold value is notified to the validity evaluation unit 1003 as setting information.

[0083] The result display unit 1304 displays the evaluation results. For example, the display method may display the recipe items, the radicals corresponding to the compression feature, and the correlation coefficient. In FIG. 15 , the correlation coefficient of radicals for etching gas A, which is indicated by the recipe item "etching gas A" and the compression feature item "etching gas A," is 0.55, and the correlation coefficient of radicals for etching gas B, which is indicated by the recipe item "etching gas B" and the compression feature item "etching gas B," is 0.45.

[0084] (Actions and Effects) As described above, in the present invention, the number of dimensions to which sensing data should be compressed is determined based on the number of types of substances (radicals) generated by etching gases and chemical reactions during etching. In training the shape prediction model, parameters of the shape prediction model are adjusted to reduce the difference between the compressed sensing data and intermediate features. In this way, according to the present invention, sensing data can be compressed and shape prediction can be performed while taking into account the generation process of the sensing data. Furthermore, by compressing sensing data, which is generally high-dimensional, into low dimensions, the number of parameters of the shape prediction model is reduced, making it possible to avoid over-training of the model.

[0085] Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the gist of the present invention.

[0086] The following are examples of aspects that may be included in the present invention, but the present invention is not limited to these. (Aspect 1) A searching device for a plasma processing apparatus that searches for a recipe for obtaining a desired etching shape, wherein the dimension of sensing data acquired by the plasma processing apparatus is defined based on the material of a sample to be etched by the plasma processing apparatus, the material of a component constituting a processing chamber in which the sample is etched, or the number of types of radicals derived from the recipe, wherein compressed features are derived using the sensing data compressed to the defined dimension, wherein machine learning parameters are adjusted so as to reduce the difference between the derived compressed features and intermediate features, thereby calculating the intermediate features from the recipe, and wherein the etching shape is predicted based on the calculated intermediate features and the recipe. (Aspect 2) The searching device described in Aspect 1, wherein the spectral intensity of plasma emission is used as the sensing data. (Aspect 3) The searching device described in Aspect 1 or Aspect 2, wherein the recipe or the sensing data is determined to be abnormal if the difference is greater than a predetermined threshold. (Aspect 4) The searching device according to any one of Aspects 1 to 3, wherein the machine learning is a neural network. (Aspect 5) The searching device according to any one of Aspects 1 to 3, wherein the spectral intensity is decomposed into a weighted sum that is a sum of component vectors multiplied by weighting coefficients, the compressed feature is expressed as a vector in which the weighting coefficients are arranged, component vectors having wavelengths in which the spectral intensity is greater than a predetermined threshold are detected, and the etching gas and reaction products generated by the etching are associated with the detected component vectors based on a correlation between the wavelengths of the spectral intensity emitted by the radicals and the types of the radicals.(Aspect 6) The searching device according to any one of Aspects 1 to 5, wherein the validity of the compressed feature is evaluated by analyzing a relationship between the types of radicals corresponding to each component of the compressed feature, a gas flow rate specified in the recipe, and a value of each component of the compressed feature. (Aspect 7) The searching device according to any one of Aspects 1 to 6, wherein the spectral intensity is decomposed into a weighted sum that is a sum of component vectors multiplied by weighting coefficients by non-negative matrix decomposition. (Aspect 8) A data compression processing device that compresses sensing data acquired from a plasma processing device, wherein a dimension of the sensing data acquired by the plasma processing device is defined based on a material of a sample to be etched by the plasma processing device, a material of a member constituting a processing chamber in which the sample is etched, or the number of types of radicals derived from the recipe, and wherein the compressed feature is derived using the sensing data compressed to the defined dimension. (Aspect 9) A semiconductor device manufacturing system having a platform on which an application for searching for a recipe for a plasma processing apparatus that will obtain a desired etching shape is implemented, wherein the application executes the following steps: defining the dimension of sensing data acquired by the plasma processing apparatus based on the material of a sample to be etched by the plasma processing apparatus, the material of a component constituting a processing chamber in which the sample is etched, or the number of types of radicals derived from the recipe; deriving compressed features using the sensing data compressed to the defined dimension; calculating the intermediate features from the recipe by adjusting machine learning parameters so that the difference between the derived compressed features and intermediate features is small; and predicting the etching shape based on the calculated intermediate features and the recipe.(Aspect 10) A semiconductor device manufacturing system having a platform on which an application for compressing sensing data acquired from a plasma processing apparatus is implemented, characterized in that the semiconductor device manufacturing system comprises: a step of defining the dimension of the sensing data acquired by the plasma processing apparatus based on the material of a sample to be etched by the plasma processing apparatus, the material of a component constituting a processing chamber in which the sample is etched, or the number of types of radicals derived from the recipe; and a step of deriving compressed features using the sensing data compressed to the defined dimension. (Aspect 11) A searching method for searching for a recipe in a plasma processing apparatus that will obtain a desired etching shape, comprising the steps of: defining a dimension of sensing data acquired by the plasma processing apparatus based on a material of a sample to be etched by the plasma processing apparatus, a material of a component constituting a processing chamber in which the sample is etched, or a number of types of radicals derived from the recipe; deriving compressed features using the sensing data compressed to the defined dimension; calculating the intermediate features from the recipe by adjusting machine learning parameters so as to reduce a difference between the derived compressed features and intermediate features; and predicting the etching shape based on the calculated intermediate features and the recipe. (Aspect 12) A data compression processing method for compressing sensing data acquired from a plasma processing apparatus, comprising the steps of: defining a dimension of sensing data acquired by the plasma processing apparatus based on a material of a sample to be etched by the plasma processing apparatus, a material of a component constituting a processing chamber in which the sample is etched, or a number of types of radicals derived from the recipe; and deriving compressed features using the sensing data compressed to the defined dimension.

[0087] 1, 1A, 1B Semiconductor device manufacturing system, 101 Plasma processing apparatus, 102 Sensor, 103 Recipe, 104 Sensing data, 105 Processing shape, 106 Database, 107, 107A Compression processing apparatus, 108 Dimension determination unit, 109 Material information, 110 Dimension compression unit, 111 Compressed feature amount, 112 Input / output device, 113 Display information, 114 Etching shape prediction apparatus, 115 Intermediate feature amount calculation unit, 116 Shape calculation unit, 117 Parameter adjustment unit, 118 Intermediate feature amount, 119 Predicted shape, 120 Learning device, 121 Abnormal data detection unit, 122 Abnormal information, 201 Table, 202 Emission spectrum intensity, 601 Emission spectroscopic sensor, 1001 Component vector, 1002, 1002A, 1002B Emission wavelength-radical correspondence table

Claims

1. A searching device for a plasma processing apparatus that searches for a recipe that will obtain a desired etching shape, wherein the dimension of sensing data acquired by the plasma processing apparatus is defined based on the material of a sample to be etched by the plasma processing apparatus, the material of a component that constitutes a processing chamber in which the sample is etched, or the number of types of radicals derived from the recipe, the sensing data compressed to the defined dimension is used to derive compressed features, machine learning parameters are adjusted so that the difference between the derived compressed features and intermediate features is small, thereby calculating the intermediate features from the recipe, and the etching shape is predicted based on the calculated intermediate features and the recipe.

2. A searching device according to claim 1, wherein the sensing data is the spectral intensity of plasma emission.

3. A searching device according to claim 1, wherein if the difference is greater than a predetermined threshold, the recipe is determined to be abnormal or the sensing data is determined to be abnormal.

4. The search device according to claim 1, wherein the machine learning is a neural network.

5. A searching device according to claim 2, characterized in that the spectral intensity is decomposed into a weighted sum, which is the sum of component vectors multiplied by weighting coefficients, the compressed feature is represented as a vector in which the weighting coefficients are arranged, component vectors having wavelengths in which the spectral intensity is greater than a predetermined threshold are detected, and the etching gas and the reaction products generated by the etching are associated with the detected component vectors based on the correlation between the wavelength of the spectral intensity emitted by the radicals and the type of radical.

6. A searching device according to claim 5, characterized in that the validity of the compression feature is evaluated by analyzing the relationship between the types of radicals corresponding to each component of the compression feature, the gas flow rate specified in the recipe, and the value of each component of the compression feature.

7. A searching device according to claim 5, wherein the spectral intensities are decomposed into weighted sums, which are sums of component vectors multiplied by weighting coefficients, by non-negative matrix decomposition.

8. A data compression processing device that compresses sensing data acquired from a plasma processing device, wherein the dimension of the sensing data acquired by the plasma processing device is defined based on the material of a sample to be etched by the plasma processing device, the material of a component constituting a processing chamber in which the sample is etched, or the number of types of radicals derived from the recipe, and wherein compressed features are derived using the sensing data compressed to the defined dimension.

9. A semiconductor device manufacturing system having a platform on which an application for searching for a recipe for a desired etching shape in a plasma processing apparatus is implemented, wherein the application performs the following steps: defining the dimension of sensing data acquired by the plasma processing apparatus based on the material of a sample to be etched by the plasma processing apparatus, the material of a component constituting a processing chamber in which the sample is etched, or the number of types of radicals derived from the recipe; deriving compressed features using the sensing data compressed to the defined dimension; calculating the intermediate features from the recipe by adjusting machine learning parameters so that the difference between the derived compressed features and intermediate features is small; and predicting the etching shape based on the calculated intermediate features and the recipe.

10. A semiconductor device manufacturing system having a platform on which an application for compressing sensing data acquired from a plasma processing apparatus is implemented, the semiconductor device manufacturing system comprising: a step of defining the dimension of the sensing data acquired by the plasma processing apparatus based on the material of a sample to be etched by the plasma processing apparatus, the material of a component constituting a processing chamber in which the sample is etched, or the number of types of radicals derived from the recipe; and a step of deriving compressed features using the sensing data compressed to the defined dimension.

11. A search method for searching for a recipe in a plasma processing apparatus that will obtain a desired etching shape, comprising: a step of defining the dimension of sensing data acquired by the plasma processing apparatus based on the material of a sample to be etched by the plasma processing apparatus, the material of a component constituting a processing chamber in which the sample is etched, or the number of types of radicals derived from the recipe; a step of deriving compressed features using the sensing data compressed to the defined dimension; a step of calculating the intermediate features from the recipe by adjusting machine learning parameters so that the difference between the derived compressed features and intermediate features is small; and a step of predicting the etching shape based on the calculated intermediate features and the recipe.

12. A data compression processing method for compressing sensing data acquired from a plasma processing apparatus, comprising the steps of: defining the dimensions of the sensing data acquired by the plasma processing apparatus based on the material of a sample to be etched by the plasma processing apparatus, the material of a component constituting a processing chamber in which the sample is etched, or the number of types of radicals derived from the recipe; and deriving compressed features using the sensing data compressed to the defined dimensions.

Images