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

Abstract

The purpose of the present invention is to provide a technique capable of compressing sensing data and performing shape prediction in consideration of a sensing data generation process. Therefore, the search device of the present invention is characterized in that: the search device searches for a recipe for obtaining a desired etching shape in a plasma processing device; the dimensions of the sensing data acquired by the plasma processing apparatus are defined on the basis of the material of a sample etched by the plasma processing apparatus, the material of a member constituting a processing chamber for etching the sample, or the number of types of radicals derived according to a recipe. A compression feature amount is derived using sensing data compressed to a prescribed dimension, a parameter of machine learning is adjusted so that a difference between the derived compression feature amount and an intermediate feature amount becomes small, thereby calculating the intermediate feature amount according to a recipe, and an etching shape is predicted based on the calculated intermediate feature amount and the recipe.

Description

Technical Field

[0001] This invention relates to a search device, a data compression processing device, a semiconductor device manufacturing system, a search method, and a data compression processing method. Background Technology

[0002] In the plasma etching process of semiconductors, the recipe, which is a processing parameter of the etching apparatus, needs to be adjusted to obtain the desired processed shape. Generally, the recipe includes many variables such as the flow rate of the etching gas and the bias voltage used to accelerate ions in the plasma, and this adjustment requires trial and error based on experience. Therefore, this study investigates the application of machine learning methods in recipe optimization.

[0003] In machine learning methods, a machine learning model that predicts the processing shape based on the formula and sensing data is trained to search for a formula that allows the resulting machine learning model to output a target shape. For example, Patent Document 1 illustrates a method that utilizes the intensity of the emission spectrum observed during the etching process in addition to the formula and shape in the machine learning model.

[0004] Specifically, Patent Document 1 discloses an invention that improves tool performance by establishing a correlation between tool operating parameters and material measurement and spectroscopic information. "A system and method are provided for adjusting and analyzing tool performance by integrating tool operating data and spectral data associated with the tool. Such integration results in aggregated data that allows for the learning of at least one relationship between selected portions of the aggregated data. The adjustment of tool performance depends on at least a portion of the learned relationship, including the integration of processing formula parameters that allow for adjustments to the manufacturing process to derive sufficient tool performance in accordance with the implementation of the manufacturing process. The processing formula parameters can be generated by solving an inverse problem based on the learned relationship. The analysis of tool performance can include an evaluation of the integrated performance scenario, confirmation of the spectral state affecting performance, and extraction of at least one endpoint based on time-dependent spectral data."

[0005] Prior art literature

[0006] Patent documents

[0007] Patent Document 1: Japanese Patent Publication No. 2014-507801 Summary of the Invention

[0008] The problem that the invention aims to solve

[0009] When using neural networks as models to predict shape based on a recipe, achieving high prediction accuracy requires extracting intermediate features that are effective for prediction based on the recipe. Here, intermediate features refer to the output of an intermediate layer in the neural network, represented as multi-dimensional real vectors. However, generally, compared to other fields using neural networks, the number of etching trials performed until the desired shape is obtained is less, and the amount of learning data acquired is also less, making it difficult to extract intermediate features using conventional methods. Therefore, for example, one might consider training the model to reduce the difference between the sensed data and the intermediate features, thereby guiding the intermediate features by extracting sensed data that includes information useful for shape prediction based on the recipe. However, generally, when directly using sensed data with high dimensionality, the number of parameters in the prediction model must be increased, leading to overlearning when there is limited learning data. To avoid such overlearning, it is necessary to compress the sensed data.

[0010] In the compression of sensed data, the dimension of the compression feature quantity, as a hyperparameter, is crucial in order to extract the necessary information for shape prediction from the sensed data without over- or under-extracting it. Previously, the dimension of the compression feature quantity was determined using methods such as the elbow rule, based on a trade-off between small dimension and large information content. However, these methods did not consider the physical generation process of the data.

[0011] Therefore, the object of this invention is to provide a technique that can compress sensing data and perform shape prediction by taking into account the generation process of sensing data.

[0012] Methods for solving problems

[0013] To address the aforementioned issues, a representative search device of the present invention is characterized in that, in a search device for searching a formula for obtaining a desired etched shape in a plasma processing apparatus, the dimension of the sensing data acquired by the plasma processing apparatus is defined based on the material of the sample etched by the plasma processing apparatus, the material of the components constituting the processing chamber for etching the sample, or the number of types of free radicals derived according to the formula; a compressed feature quantity is derived using the sensing data compressed to the defined dimension; the parameters of the machine learning are adjusted to reduce the difference between the derived compressed feature quantity and the intermediate feature quantity; the intermediate feature quantity is calculated according to the formula; and the etched shape is predicted based on the calculated intermediate feature quantity and the formula.

[0014] Invention Effects

[0015] According to the present invention, the sensing data can be compressed and shape prediction can be performed by taking into account the generation process of the sensing data.

[0016] Other issues, structures, and effects not described above will become clear through the following detailed description. Attached Figure Description

[0017] Figure 1 This is a diagram illustrating an example of the structure of a semiconductor device manufacturing system.

[0018] Figure 2 This is a diagram illustrating an example of a table used to store various types of data in a database.

[0019] Figure 3 This is a flowchart in the dimension determination section that determines the compressed dimensions of the sensed data.

[0020] Figure 4 This is a diagram illustrating an example of a GUI for performing a dimensional search based on user-defined settings.

[0021] Figure 5 This is a diagram illustrating an example of a GUI where no dimensional search is performed within the user's settings.

[0022] Figure 6 A flowchart illustrating the process of adjusting the parameters of the intermediate feature quantity calculation unit and the shape calculation unit in the etching shape prediction device.

[0023] Figure 7 This is a diagram illustrating an example of the structure of a semiconductor device manufacturing system.

[0024] Figure 8 It is a diagram that schematically illustrates the emission spectral intensity as a vector.

[0025] Figure 9 This is a flowchart for verifying anomalies in learning data within the anomaly detection department.

[0026] Figure 10 This is a diagram illustrating an example of a GUI for setting parameters to detect abnormal data and confirming the results.

[0027] Figure 11 This is a diagram illustrating an example of the structure of a semiconductor device manufacturing system.

[0028] Figure 12 This is a diagram showing an example of a table of emission wavelengths and free radicals.

[0029] Figure 13 This is a diagram showing an example of a table of emission wavelengths and free radicals.

[0030] Figure 14 This is a flowchart for judging the appropriateness of compression features in the appropriateness evaluation department.

[0031] Figure 15 This is a diagram illustrating an example of a GUI for setting parameters to evaluate the appropriateness of compressed feature quantities and verifying the results. Detailed Implementation

[0032] Hereinafter, embodiments will be described using the accompanying drawings. However, the invention is not limited to these embodiments. Furthermore, in the accompanying drawings, the same reference numerals are used to indicate the same parts.

[0033] Furthermore, when there are multiple constituent elements with the same or identical functions, different additional letters are sometimes used to indicate the same reference numerals in the drawings. Additionally, when it is not necessary to distinguish between these multiple constituent elements, the additional letters are sometimes omitted in the description.

[0034]

Example 1

[0035] (Semiconductor device manufacturing system)

[0036] Reference Figure 1 The semiconductor device manufacturing system 1 will be described. Figure 1 This is a diagram showing an example of the structure of a semiconductor device manufacturing system 1. The semiconductor device manufacturing system 1 includes a plasma processing device 101, a database 106, an input / output device 112, a learning device 120, and an abnormal data detection unit 121.

[0037] (Plasma processing equipment and database)

[0038] 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 sensors 102 that detect various characteristics, and the sensed data 104 is measured during etching. Although a single sensor 102 is shown, this disclosure is not limited to this case. Multiple sensors 102 may be included for each detection object. Furthermore, this disclosure describes etching as plasma processing, but this disclosure can also be applied to processes other than etching.

[0039] Furthermore, after etching, information indicating the processing shape of the sample (hereinafter also simply referred to as "processing shape") 105 is obtained from the plasma processing apparatus 101. The formula 103, sensing data 104, and processing shape 105 are stored in the database 106. In addition, information indicating the material of the wafer and information indicating the structure of the plasma processing apparatus (e.g., the material of the sidewalls of the chamber constituting the plasma processing apparatus 101, etc.) are also stored in the database 106 as material information 109.

[0040] (Input / output devices)

[0041] Input / output device 112 receives user requests or provides user prompts. Setting / abnormal information 122 includes setting information indicating user requests sent to abnormal data detection unit 121, and abnormal information as notification from abnormal data detection unit 121 to the user. Setting / display information 113 includes setting information indicating user requests sent to compression processing unit (data compression processing unit) 107, and display information as notification from compression processing unit 107 to the user. Input / output device 112 includes, for example, a display device such as a monitor as an output device, and also includes input devices such as a mouse and keyboard as input devices.

[0042] (Learning device)

[0043] The learning device (search device) 120 acquires sensing data 104 obtained from the plasma processing apparatus 101 performing plasma processing, and uses a neural network to predict the shape of the sample (wafer) formed by plasma processing. In the learning device 120, a formula for obtaining the desired etching shape in the plasma processing apparatus 101 is searched. The learning device 120 includes a compression processing unit 107 and an etching shape prediction unit 114. The compression processing unit 107 includes a dimension determination unit 108 and a dimension compression unit 110, and the etching shape prediction unit 114 includes an intermediate feature calculation unit 115, a shape calculation unit 116, and a parameter adjustment unit 117. Furthermore, the compression processing unit 107 and the etching shape prediction unit 114 are structures created by dividing the constituent parts for convenience, and this disclosure is not limited to this case.

[0044] Furthermore, the learning device 120 can also be constructed from general hardware equipped with a processor and memory, and perform the functions of the compression processing device 107 and the etching shape prediction device 114 by executing the program stored in the memory through the processor. The learning device 120 can also be constructed using dedicated hardware, or it can be installed using a combination of software and hardware.

[0045] (Compression processing device)

[0046] First, the compression processing apparatus 107 will be described. The compression processing apparatus 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 compression unit 110. First, the dimension determination unit 108 determines, based on at least one of the material information 109 of the wafer, the chamber sidewall, and the formula 103, how many dimensions the sensing data 104 should be compressed to. Then, the dimension compression unit 110 compresses the sensing data 104 to the determined dimensions and derives a compression feature quantity 111. Furthermore, the input / output device 112 transmits and receives setting / display information 113 for determining the compressed dimensions of the sensing data 104 with the dimension determination unit 108.

[0047] (Etching shape prediction device)

[0048] Next, the etching shape prediction apparatus 114 will be described. The compressed feature quantity 111 can be used to improve the prediction accuracy in the etching shape prediction apparatus 114, which predicts the processing shape 105 according to the formula 103. The etching shape prediction apparatus 114 includes an intermediate feature quantity calculation unit 115, a shape calculation unit 116, and a parameter adjustment unit 117. The intermediate feature quantity calculation unit 115 extracts intermediate feature quantities 118 from the formula 103 using machine learning (e.g., a neural network). Furthermore, during the learning of the neural network, the parameter adjustment unit 117 adjusts the parameters of the neural network to reduce the difference between the compressed feature quantity 111 and the intermediate feature quantity 118 extracted from the formula 103.

[0049] Furthermore, the shape calculation unit 116 uses a machine learning model to output the predicted shape of the wafer after etching, i.e., the predicted shape 119, based on the formula 103 and the extracted intermediate feature values ​​118. In addition, during the learning of the machine learning model, the parameter adjustment unit 117 adjusts the parameters of the machine learning model so that the difference between the predicted shape 119 predicted based on the formula 103 and the intermediate feature values ​​118 and the actual processed shape 105 becomes smaller.

[0050] As described above, the compression processing device 107 and the etching shape prediction device 114 predict the shape of the sample after learning the relationship between the sensing data 104, the processing shape 105, the formula 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 the learning device 120.

[0051] (Abnormal Data Detection Department)

[0052] The etching shape prediction device 114 confirms its operation through the abnormal data detection unit 121. The abnormal data detection unit 121 sends and receives setting / abnormal information 122 between itself and the input / output device 112.

[0053] (Data stored in the database)

[0054] Figure 2 This diagram illustrates an example of a table used to store various data in database 106. Table 201 stores learning data, including five categories of columns: recipe ID, recipe 103, sensing data 104, processing shape 105, and material information 109. Based on recipe 103 corresponding to recipe ID, an etching process is performed in plasma processing apparatus 101. The sensing data 104 acquired during the etching process and the processing shape 105 obtained by measuring the etched wafer are included in one record of table 201. Recipe 103 can be data with a common identifier or data with different identifiers. Table 201 stores data from trials conducted to process a sample into a target shape. Figure 2 In this context, the recipe ID is represented by numbers from 1 to N (where N is a positive integer), indicating that N trials were performed.

[0055] Sensing data 104 is obtained from sensor 102 mounted on plasma processing apparatus 101. As sensing data, for example, the spectral intensity of plasma emission (emission spectral intensity) is used. Furthermore, sensor 102 can include multiple sensors. For example, an optical emission spectrometer (OES) 601, described later, disperses the emission spectrum and measures the emission spectral intensity 202 at each wavelength. Additionally, a mass flow controller (MFC) measures the mass flow rate 203 of the etching gas. In addition, sensors related to voltage systems, pressure systems, temperature systems, etc., may also be included. Regarding the preprocessing of sensing data 104, if sensing data 104 is a time series, sometimes the average value over time for each recipe ID is saved. Furthermore, the sensing data 104 may or may not undergo normalization or standardization processing.

[0056] The processing shape 105 is represented, for example, by values ​​such as the depth and width of trenches formed on the wafer. These values ​​are obtained, for example, by taking a cross-section of the etched wafer using an electron microscope and measuring the shape within the image. Furthermore, the material information 109 indicates the material of the etched wafer and the material of the sidewalls of the chamber of the plasma processing apparatus 101, etc.

[0057] In Table 201, formula 103, sensing data 104, and processed shape 105 are represented by real vectors. Generally, sensing data 104 has a larger dimension than formula 103 and processed shape 105. Formula 103 and processed shape 105 have at most tens of dimensions. On the other hand, when the light is dispersed at a resolution of 0.3 nm in the range from 200 nm to 800 nm, the emission spectral intensity 202 is acquired as data with 2000 dimensions. In addition, sensing data 104, excluding emission spectral intensity 202, also includes data with several hundred dimensions. Furthermore, regarding the total amount of data, generally, the number of rows in Table 201, i.e., the number of experiments performed to process one target shape, is at most several tens.

[0058] Generally, to process high-dimensional sensing data 104 using a machine learning model, the number of model parameters needs to be increased. However, if a model with a large number of parameters is trained on a small amount of data, the model parameters will overfit to the limited data, resulting in overlearning and decreased generalization performance on the test data. Therefore, by compressing the high-dimensional sensing data 104 into a low-dimensional model, the number of model parameters can be reduced. As a result, overlearning needs to be avoided.

[0059] 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 be compressed. The sensing data 104 is data 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 aspects related to each sensor. According to the generation principle of the sensing data 104, data equivalent to the sensing data 104 can be generated based on the number of types of free radicals. Furthermore, based on the relationship between the number of types of free radicals and the sensing data 104, the sensing data 104 should be compressed based on the number of types of free radicals.

[0060] (The determination of dimensions)

[0061] Figure 3 This is a flowchart of determining the compressed dimension of the sensing data 104 in the dimension determination unit 108. First, the dimension determination unit 108 begins processing by receiving user setting information from the formula 103, sensing data 104, material information 109, and setting / display information 113 (step 301).

[0062] Next, the dimension of the sensing data acquired by the plasma processing apparatus 101 is defined based on the material of the sample etched by the plasma processing apparatus 101, the material of the components constituting the processing chamber for etching the sample, or the number of types of free radicals derived according to the formula. Specifically, the dimension determination unit 108 specifies the number of types of free radicals based on at least one of the wafer material information, the chamber sidewall material information (material information 109), and the formula 103, and based on user settings (step 302). For example, when only the etching gases Cl2, NF3, N2, O2, and CO2 in formula 103 are used as the source substances for the emission spectrum, the number of types of free radicals is determined to be 5. Here, if a dimension search is performed based on user settings, the process proceeds to step 304; if no search is performed based on user settings, the dimension of the compressed feature quantity is determined to be the specified number of types of free radicals, and the process proceeds to step 308 (step 303).

[0063] When performing a dimensional search, the dimensional compression unit 110 sets candidates for compressed dimensions of the sensing data 104 (hereinafter also referred to as "compressed dimension candidates") based on the set number of free radical types (step 304). For example, consider setting an upper limit and a lower limit to include the number of free radical types, and use the interval between the upper limit and the lower limit as compressed dimension candidates.

[0064] Next, the dimension determination unit 108 begins iterative processing for each candidate compression dimension (step 305). During iterative processing, the dimension determination unit 108 attempts to compress the sensed data 104 according to each candidate dimension, and calculates an index value related to the quality of compression based on the sensed data 104 before compression and the compression feature quantity as the compressed sensed data 104 (step 306). During the compression trial, principal component analysis, autoencoders, nonnegative matrix factorization, etc., can be applied as methods for dimension compression used by the dimension compression unit 110. In addition, other, faster dimension compression methods can be used for hyperparameter search. Furthermore, examples of index values ​​related to the quality of compression include reconstruction error and outliers corresponding to principal components in principal component analysis. Here, the reconstruction error can be represented by the Euclidean distance between the sensed data recovered from the compression feature quantity and the sensed data before compression.

[0065] After each iteration, the dimension determination unit 108 determines whether to adopt a candidate dimension based on conditions related to the quality of compression (step 307). For example, if the reconstruction error is below a given threshold, the dimension is adopted. If only one dimension is adopted in step 307, the adopted dimension is determined (defined) as the compressed dimension. If multiple dimensions satisfy the conditions in step 307, the most suitable dimension (e.g., the dimension with the smallest reconstruction error) is determined as the compressed dimension.

[0066] Finally, the dimension determination unit 108 sends the determined dimension to the dimension compression unit 110 and ends the process (step 308). The dimension compression unit 110 compresses the sensed data 104 to the determined dimension and derives the compressed feature quantity 111.

[0067] (Example of a GUI)

[0068] Next, refer to Figure 4 as well as Figure 5 The GUI (Graphical User Interface) between the semiconductor device manufacturing system 1 and the user will be explained. To determine the dimension (compressed dimension) of the compressed feature 111 using the method described above, it is necessary to set the number of free radical types, the search range for the dimension, and the selection criteria for candidate dimensions. The selection criteria can be adjusted using the GUI between the compression processing device 107 and the user. Figure 4 This is a diagram illustrating an example of a GUI for searching dimensions based on user-defined information.

[0069] For example, in screen 401 Figure 3 In step 304, the display device of the input / output device 112 is shown. The screen 401 includes setting / display information 113, which notifies the user of the display information. In addition, the user's setting information is sent to the compression processing device 107. In the recipe and material display unit 402, material information 109 indicating the material of the wafer and the material of the chamber sidewall, as well as the etching gas included in the recipe 103, are displayed.

[0070] Referring to the formula and material display unit 402, the user selects a method for specifying the number of free radical types in the free radical type setting unit 403. Examples of methods for specifying the number of free radical types include, for instance, directly using the number of etching gases in the formula, including free radicals generated from reaction products in addition to the number of etching gases, and a method set by the user. Figure 4 In the process, the method of selecting the number of types of etching gases directly adopted from the formula is used as the user's setting information and sent to the compression processing unit 107.

[0071] In the search execution setting section 404, an option is shown to select whether to perform a dimensional search. Figure 4 In the process, the user selects to perform a dimensional search as the user's setting information and sends it to the compression processing device 107.

[0072] Based on the number of free radical types specified in the free radical type setting unit 403, the search range of the dimension is set in the search range setting unit 405. Figure 4 In this context, z is set as the lower limit of the search range for the dimension (e.g., z = 3), and y is set as the upper limit of the search range for the dimension (e.g., y = 7).

[0073] In the adoption condition setting unit 406, adoption conditions for adopting dimensions that are candidates for compression are set. Furthermore, if setting parameters such as thresholds is required when determining adoption conditions, the user inputs these parameters into an input form. Figure 4 In this context, as an adoption condition, an index value related to the quality of compression is set to be either above or below a threshold. Furthermore, a threshold t is set as the threshold value.

[0074] In the trial result display unit 407, the results of the repeated processing are displayed based on the input values ​​and the conditions applied. As an example of the result display method, a graph is shown with the compressed dimension as the horizontal axis and the index value related to the quality of compression as the vertical axis; however, the result display method is not limited to this, and other displays may be used. Figure 4 The figure shows the case where the index value related to the quality of compression is below the threshold t in the range between dimensions x and y.

[0075] Figure 5 This is a diagram illustrating an example of a GUI where no dimensional search is performed within the user's settings. Figure 5 This illustrates the case where no dimensional search is performed in the search execution setting unit 404. In this case, the number of free radical species x specified by the user is displayed in the free radical species setting unit 403, and it is determined as the dimension of the compressed feature.

[0076]

Example 2

[0077] Example 2 details the adjustment of the neural network parameters in the intermediate feature calculation unit 115. Example 2 illustrates the following scenario: the machine learning parameters are adjusted to minimize the difference between the derived compressed feature 111 and the intermediate feature 118, thereby calculating the intermediate feature based on the formula, and predicting the etching shape based on the calculated intermediate feature and the formula. In the following description, the same reference numerals are used for components identical or equivalent to those in Example 1 described above, and their descriptions are simplified or omitted.

[0078] In the etching shape prediction apparatus 114, which predicts the processing shape 105 according to formula 103, compressed feature quantity 111 is used to improve prediction accuracy. Generally, when using a neural network as the shape prediction model, it is difficult to automatically learn the intermediate feature quantities of the neural network based on a small amount of data. Therefore, it is thought that the extraction of intermediate feature quantities can be guided by determining the target not only for the output of the shape prediction model but also for the intermediate layer. As the target of the intermediate layer, compressed feature quantity 111 obtained by dimensional compression of sensed data 104 is suitable. This is because sensed data 104 is the observation in the actual etching process, which includes information useful for shape prediction. It is expected that by guiding the intermediate feature quantity 118 in a way that extracts the compressed feature quantity 111 useful for shape prediction according to formula 103, the accuracy of shape prediction can be improved.

[0079] Furthermore, from the perspective of ease of data acquisition, compressed feature 111 is also suitable as a target for the intermediate layer. In addition, the sensing data 104 is not limited to being acquired from patterned wafers with patterns; much of it can also be acquired from bare wafers, thus enabling the acquisition of large amounts of data at low cost. Moreover, an electron microscope is not required for the measurement of the sensing data 104, and a large amount of sensing data 104 can be easily acquired.

[0080] (Methods for adjusting parameters)

[0081] Figure 6 This is a flowchart of the process for adjusting the parameters of the intermediate feature quantity calculation unit 115 and the shape calculation unit 116 in the etching shape prediction apparatus 114. First, the etching shape prediction apparatus 114 accepts the recipe 103, the compressed feature quantity 111, and the processed shape 105 as learning data for parameter adjustment and begins processing (step 501). Furthermore, the compressed feature quantity 111 may be standardized before being used in the etching shape prediction apparatus 114. Standardization refers to a scaling process that sets the mean of each dimension in the learning data to 0 and the variance to 1. Standardization can be performed, for example, in the database 106 or in the parameter adjustment unit 117.

[0082] Next, the parameter adjustment unit 117 uses recipe 103, compressed feature quantity 111, and intermediate feature quantity 118 to adjust the parameters of the neural network in the intermediate feature quantity calculation unit 115 (step 502). When adjusting the parameters of the neural network in the intermediate feature quantity calculation unit 115, compressed feature quantity 111 is used as the target of intermediate feature quantity 118. Here, being used as the target means that, given neural network f, the number of learning data N, and recipe r of the nth learning data... n (For example, the recipe in recipe 103 that is equivalent to recipe ID=n), and the recipe r used in compressed feature quantity 111. n The feature quantity corresponding to the sensing data 104 observed in the etching is the compressed feature quantity c. n Given the loss function l, the parameters θ of the neural network f are updated so that the value of (1) becomes smaller.

[0083]

Mathematical Formula 1

[0084]

[0085] Subsequently, the shape calculation unit 116 enables the machine learning model within it to learn, thereby predicting the processed shape 105 produced using the formula 103 based on the intermediate feature quantity 118 obtained by the neural network f, whose parameter θ has been adjusted in the intermediate feature quantity calculation unit 115, and the formula 103 (step 503). The machine learning model in the shape calculation unit 116 is not limited to a neural network; other machine learning models can also be applied.

[0086] Finally, the etching shape prediction device 114 (intermediate feature calculation unit 115) holds the adjusted parameter θ and ends the process (step 504).

[0087] After learning, the etching shape prediction device 114 uses the intermediate feature quantity calculation unit 115 and the shape calculation unit 116, with the parameter θ adjusted, to predict the processing shape based on an arbitrary formula. In addition, during prediction, there is no need to compress the feature quantity 111; the output of the neural network in the intermediate feature quantity calculation unit 115 and the formula 103 are input into the machine learning model in the shape calculation unit 116.

[0088]

Example 3

[0089] Example 3 details the use of spectral intensity from the sensing data. In the following description, the same reference numerals are used for the same or equivalent components as in Example 1 above, and their descriptions are simplified or omitted.

[0090] (Semiconductor device manufacturing system)

[0091] Figure 7This is a diagram illustrating an example of the structure of a semiconductor device manufacturing system 1A. The semiconductor device manufacturing system 1A of Embodiment 3 differs from the semiconductor device manufacturing system 1 of Embodiment 1 in that it has a light emission spectrophotometer 601 in the plasma processing apparatus 101 and a database 106 storing the light emission spectral intensity 202. The light emission spectrophotometer 601 measures the light emission spectral intensity 202 during the etching process in the plasma processing apparatus 101. The measured light emission spectral intensity 202 is stored in the database 106. Additionally, after etching, a processed shape 105 of the sample is obtained from the plasma processing apparatus 101. The formulation 103, the light emission spectral intensity 202, and the processed shape 105 are stored in the database 106.

[0092] The emission spectrum intensity 202 stored in database 106 is compressed by compression processing device 107. Compression is performed in two stages: processing in dimension determination unit 108 and processing in dimension compression unit 110. First, dimension determination unit 108 determines the compressed dimension to which the emission spectrum intensity 202 will be compressed, based on the number of free radical types, an index value related to the quality of compression, and user-defined settings. The number of free radical types is based on at least one of the material information 109 of the wafer, the chamber sidewall, and the formula 103, and is specified by user-defined settings in setting / display information 113. Then, dimension compression unit 110 compresses the emission spectrum intensity 202 to the determined dimension and derives a compression feature value 111. Furthermore, input / output device 112 transmits and receives setting / display information 113 used to determine the compressed dimension of the emission spectrum intensity 202 between itself and dimension determination unit 108.

[0093] (Emission spectral intensity)

[0094] If we consider the principle of emission spectrum generation, the emission spectrum intensity can necessarily be derived from the number of free radical types. In this case, the method of compressing the emission spectrum intensity 202 based on the number of free radical types is naturally adopted. (See reference...) Figure 8 To illustrate, it is reasonable to compress the emission spectrum intensity 202 to a dimension corresponding to the number of types of free radicals involved in the chemical reaction of etching, taking into account the physical generation process of the emission spectrum. Figure 8 This is a diagram schematically showing the emission spectral intensity as a vector. The emission spectrum is emitted during the etching chemical reaction through the migration of electron orbitals of the reactants. Moreover, the emitted wavelength (hereinafter also simply referred to as the "emission wavelength") takes an inherent value for each reactant. Therefore, the emission spectral intensity 202 arising from each substance can be defined as a vector with a spectral resolution dimension.

[0095] For example, consider observing at a certain moment in an etching process. Figure 8 The case of emission spectral intensity 202 shown in vector 701 is illustrated in (b). In vector 701, the component with emission wavelength of 200 nm is 1, the component with emission wavelength of 400 nm is 4, and the component with emission wavelength of 600 nm is 2. Furthermore, in the following description, for ease of understanding, the cases of 200 nm, 400 nm, and 600 nm as emission wavelengths will be explained. In practice, by setting the emission wavelength according to the resolution of the spectral dispersion, this disclosure can also be applied to cases with more than three emission spectral intensity components.

[0096] In this case, it was determined that the observed emission spectral intensity 202 originated from substances A, B, and C. On the other hand, it is assumed that the emission spectral intensities per unit concentration of substances A, B, and C are 702A, 702B, and 702C, respectively. For example... Figure 8 As shown in (a), for substance A, the emission spectral intensity 702A per unit concentration has a component of 200 nm, a component of 400 nm, and a component of 600 nm. For substance B, the emission spectral intensity 702B per unit concentration has a component of 0 Hz, a component of 400 nm, and a component of 600 nm. For substance C, the emission spectral intensity 702C per unit concentration has a component of 0 Hz, a component of 400 nm, and a component of 600 nm.

[0097] Next, export Figure 8 The emission spectral intensity per unit concentration shown in (a) is compared with Figure 8 The relationship between the observed emission spectral intensities 202 is shown in (b). The concentration ratios of the substances that underwent chemical reactions during etching were measured and determined. Based on these concentration ratios, weighting coefficients 703 were set to 0.5 for substance A, 2.0 for substance B, and 0.5 for substance C. (See diagram for example.) Figure 8 As shown in (b), the vector 701 derived by multiplying the emission spectral intensities 702A, 702B, and 702C per unit concentration for various substances by the weighting coefficient 703 and summing them is equivalent to a vector representing the observed emission spectral intensities 202. Moreover, regarding the weighting coefficient 703, which arranges the weights to represent a vector, the vector is regarded as a compressed characteristic 111 of the emission spectral intensities 202 represented by 701.

[0098] The above description assumes that the emission spectral intensity per unit concentration of each substance is known. If the emission spectral intensity per unit concentration of each substance is unknown, it would be necessary to derive it from the learning data using a machine learning model in the dimension compression unit 110. However, machine learning models may not be suitable for accurately determining the emission spectral intensity per unit concentration of the substance based on the learning data, and compressing the emission spectral intensity per unit concentration to a dimension corresponding to the number of free radicals involved in the etching chemical reaction may yield suboptimal results. Therefore, when the emission spectral intensity per unit concentration of each substance is unknown, as shown in Example 1, the following method is used: compression is attempted for a certain range of dimensions, and the final dimension to which the emission spectral intensity 202 is compressed is determined based on the results.

[0099]

Example 4

[0100] Example 3 details the abnormal data detection unit 121. Example 4 shows that when the difference between the intermediate feature quantity 118 and the compressed feature quantity 111 exceeds a given threshold, the formula 103 or the sensed data 104 is determined to be abnormal. For example, the abnormal data detection unit 121 determines that an abnormality exists in either the formula 103 or the sensed data 104 when the difference between the intermediate feature quantity 118 and the compressed feature quantity 111 exceeds a given threshold, based on the formula 103 used for adjustment in the parameter adjustment unit 117 and the sensed data 104. Hereinafter, a detailed description will be provided. Furthermore, in the following description, the same reference numerals are used for components that are the same or equivalent to those in Example 1 described above, and their descriptions are simplified or omitted.

[0101] (Detection of abnormal data)

[0102] As described above, in semiconductor etching, the number of experiments performed to obtain a single target shape is generally small, resulting in limited learning data. Consequently, the presence of anomalies in the learning data significantly impacts prediction accuracy. Therefore, it is desirable to include a mechanism in the learning device 120 for removing learning data, including errors during measurement. Thus, after adjusting the parameters in the etching shape prediction device 114 (parameter adjustment unit 117), the anomaly data detection unit 121 verifies whether there are any anomalies in the learning data used for the adjustment.

[0103] Figure 9 This is a flowchart of verifying anomalies in the learning data in the anomaly data detection unit 121. First, the anomaly data detection unit 121 receives the intermediate feature quantity 118 extracted from the formula 103 and the compressed feature quantity 111 of the sensing data 104 observed through the formula 103, and begins processing (step 801).

[0104] Next, the anomaly detection unit 121 calculates the difference between the intermediate feature quantity 118 and the compressed feature quantity 111 (step 802). If the difference is large, either the recipe 103 or the sensed data 104 may be abnormal. Therefore, the anomaly detection unit 121 determines recipe IDs with a difference between the intermediate feature quantity 118 and the compressed feature quantity 111 exceeding a given threshold as abnormal data (step 803). Here, the difference between the intermediate feature quantity 118 and the compressed feature quantity 111 can be measured using the loss function l used by the parameter adjustment unit 117 in the learning of the neural network shown in Embodiment 2. Furthermore, the threshold is input from the input / output device 112 as user setting information in the setting / anomaly information 122.

[0105] Finally, the anomaly detection unit 121 sends the anomaly information from the setting / anomaly information 122 to the input / output device 112 and ends the process (step 804). Anomaly information may include, for example, the values ​​of the recipe ID, intermediate feature 118, and compression feature 111. The input / output device 112 displays the received anomaly information on a GUI or similar interface to notify the user.

[0106] (Example of a GUI)

[0107] Figure 10 This diagram illustrates an example of a GUI for setting parameters used to detect abnormal data and confirming the results. Screen 901 is shown, for example, on the display device of the input / output device 112. In the threshold setting unit 902, the user sets the threshold used in the anomaly determination. Figure 10 In this context, a threshold of 0.5 is set. The anomaly display unit 903 displays the anomaly determination result based on the set threshold. Figure 10 The table shows that the loss coefficient for recipe ID = xxxx is 0.8, and the loss coefficient for recipe ID = yyyy is 0.6. The results can be displayed only for recipe 103 that is judged as abnormal, or they can be displayed for any recipe 103.

[0108]

Example 5

[0109] Example 5 details dimensionality compression. In Example 5, it is shown that the spectral intensity (emission spectral intensity 202) is decomposed into a weighted sum of component vectors multiplied by weighting coefficients. The compressed feature is represented as a vector obtained by arranging the weighting coefficients. Component vectors with wavelengths having spectral intensities greater than a given threshold are detected. Based on the correlation between the wavelength of the spectral intensity emitted by free radicals and the type of free radical, a correspondence is established between the gas used for etching and the reaction products generated by etching and the detected component vectors. Furthermore, the spectral intensity is decomposed into a weighted sum of component vectors multiplied by weighting coefficients through nonnegative matrix factorization. In the following description, the same reference numerals are used for components that are the same or equivalent to those in Example 1 above, and their descriptions are simplified or omitted.

[0110] (Non-negative matrix factorization)

[0111] Nonnegative matrix factorization is a method of optimizing W and H such that the matrix X, which takes nonnegative values, is approximated by the product of two nonnegative matrices, WH. Specifically, when using emission spectral intensity 202 as sensing data 104, since emission spectral intensity 202 takes values ​​of 0 or positive (nonnegative) in terms of intensity, nonnegative matrix factorization can be used as a compression algorithm. In this case, each row of X corresponds to a learning data point, and each column corresponds to the emission spectral intensity 202 of each emission wavelength. Each learning data point can be a time-averaged statistical value for each formula ID of formula 103, or the value at each time step can be used directly. Furthermore, W is a matrix in which the compressed feature quantities 111 of each learning data point are arranged in rows. In nonnegative matrix factorization, as shown in (2), the i-th row of X, i.e., the i-th learning data, is approximated by the sum of the rows of H (hereinafter, the rows of H are called component vectors, and the multiple component vectors equivalent to the multiple rows included in H are called "multiple component vectors 1001") after being weighted by the compressed feature quantity 111. After learning H based on the learning data, the compressed feature quantity 111 is multiplied by H from the right, thereby approximately restoring the emission spectral intensity 202.

[0112]

Mathematical Formula 2

[0113]

[0114] Here, X i W represents the i-th row of X. i,j H represents the (i, j) component of W. j This represents the j-th row of H.

[0115] (Semiconductor device manufacturing system)

[0116] Figure 11This is a diagram illustrating an example of the structure of a semiconductor device manufacturing system 1B. Regarding the semiconductor device manufacturing system 1B, the semiconductor device manufacturing system 1B of Embodiment 5 differs from the semiconductor device manufacturing system 1 of Embodiment 1 in that the plasma processing apparatus 101 has a light emission spectrophotometer 601, the database 106 stores the light emission spectral intensity 202 and the light emission wavelength-free radical correspondence table 1002, and the compression processing apparatus 107 has a suitability evaluation unit 1003. Furthermore, in... Figure 10 The text explains dimensional compression, thus omitting the equivalent of... Figure 1 The structure of the etching shape prediction device 114 and the abnormal data detection unit 121 in (Example 1).

[0117] The emission spectrophotometer 601 measures the emission spectral intensity 202 during the etching process in the plasma processing apparatus 101. The measured emission spectral intensity 202 is then compressed by the compression processing apparatus 107A.

[0118] The compression processing apparatus 107A includes a dimension determination unit 108, a dimension compression unit 110, and a suitability evaluation unit 1003. The dimension determination unit 108 determines the compressed dimension to which the emission spectrum intensity 202 will be compressed, based on the number of types of free radicals, the reconstruction error of the emission spectrum intensity 202 based on nonnegative matrix decomposition, and setting information provided by the user. The number of types of free radicals is based on at least one of the material information 109 of the wafer and the chamber sidewalls, and the formula 103, and is specified according to the user setting information in the setting / display information 113.

[0119] The dimension compression unit 110 compresses the emission spectral intensity 202 to a determined dimension through non-negative matrix decomposition and derives the compression feature quantity 111. The dimension compression unit 110 decomposes the emission spectral intensity 202 into a weighted sum of multiple component vectors multiplied by weighting coefficients (Equation (2)). The compression feature quantity 111 is expressed as a vector obtained by arranging the weighting coefficients. Among the multiple component vectors 1001, the component vector with a wavelength having a spectral intensity greater than a given threshold is detected. According to the relationship between the type of free radical and the wavelength of the inherent spectral intensity emitted by the free radical (emission wavelength-free radical correspondence table 1002), the gas used in the plasma treatment (etching gas) and the reaction products (free radicals) generated in the plasma treatment are correlated with the detected component vectors 1001.

[0120] Specifically, in the nonnegative matrix factorization, the emission spectral intensity 202 is represented by the sum of multiple component vectors 1001 weighted by a compression feature 111. Furthermore, a correspondence is established between at least one free radical and each component vector of the multiple component vectors 1001 using a given emission wavelength-free radical correspondence table 1002. By establishing a correspondence between the free radical and each component vector of the multiple component vectors 1001, the compression feature 111 can be captured as an index representing the degree of weight given to the component vectors originating from each free radical in the emission spectral intensity 202.

[0121] Furthermore, the appropriateness evaluation unit 1003 evaluates whether the compression feature quantity 111 is appropriate, and transmits and receives setting / appropriateness information 1004 between itself and the input / output device 112. The setting / appropriateness information 1004 includes setting information representing the user's request sent to the appropriateness evaluation unit 1003, and appropriateness information as information notified to the user from the appropriateness evaluation unit 1003.

[0122] (Emission wavelength - free radical correspondence table)

[0123] Figure 12 as well as Figure 13 This is a diagram showing an example of the emission wavelength-free radical correspondence table 1002. Figure 12 Table 1002A shows the correspondence between emission wavelength and free radical. Figure 13 Table 1002B, showing the correspondence between emission wavelength and free radical, illustrates this process. During the etching reaction, each substance emits light with at least one unique wavelength. Furthermore, sometimes multiple substances emit light at the same wavelength. Table 1002 is created by summarizing the correspondence between such substances (free radicals) and their emission wavelengths in a table.

[0124] exist Figure 12 Table 1002A, which corresponds to the emission wavelength of free radicals, shows the emission wavelengths for each substance. Substances (free radicals) with an emission wavelength of xx nm are substance A and substance B. Substance C has an emission wavelength of yy nm. Substances A, D, and E have an emission wavelength of zz nm.

[0125] In addition, Figure 13In the emission wavelength-radical correspondence table 1002B, in addition to the emission wavelength, the emission spectral intensity per unit concentration is also shown for each substance. For substance A, the intensity at emission wavelength xxnm is 2000, the intensity at emission wavelength zznm is 500, and the intensities at emission wavelengths of 200nm, yynm, and 800nm ​​are 0. For substance B, the intensity at emission wavelength xxnm is 300, and the intensities at emission wavelengths of 200nm, yynm, zznm, and 800nm ​​are 0. For substance C, the intensity at emission wavelength yynm is 400, and the intensities at emission wavelengths of 200nm, xxnm, zznm, and 800nm ​​are 0. For substance D, the intensity at wavelength zznm is 800, and the intensities at emission wavelengths of 200nm, xxnm, yynm, and 800nm ​​are 0. The emission wavelength-radical correspondence table 1002B includes, for example, information useful for determining the true values ​​of the emission spectral intensities per unit concentration 702A to 702C.

[0126] (Appropriateness evaluation of compressed features)

[0127] Figure 14 This is a flowchart of the appropriateness assessment of the compressibility of the compressibility feature 111 in the appropriateness assessment unit 1003. The appropriateness of the compressibility feature 111 is assessed by analyzing the relationship between the types of free radicals corresponding to each component of the compressibility feature 111, the gas flow rate specified in the formulation 103, and the values ​​of each component of the compressibility feature 111. Specifically, the appropriateness assessment unit 1003 assesses the appropriateness of the compressibility feature 111 by analyzing the gas flow rate in the formulation 103, the values ​​of each component of the compressibility feature 111, and the relationship between the types of free radicals corresponding to each component of the compressibility feature 111. First, the appropriateness assessment unit 1003 receives multiple component vectors 1001, the formulation 103, the compressibility feature 111, and the emission wavelength-free radical correspondence table 1002, and begins processing (step 1201).

[0128] Next, the appropriateness evaluation unit 1003 detects wavelengths whose values ​​(intensities) are greater than the given threshold in the plurality of component vectors 1001 (step 1202). The threshold is input from the input / output device 112 as user setting information in the setting / appropriateness information 1004.

[0129] Then, using the given emission wavelength-radical correspondence table 1002, a correspondence is established between the free radicals corresponding to the detected wavelengths and the component vectors 1001 (step 1203). Here, when establishing the wavelength correspondence, the measurement error can also be taken into account to give the wavelength a range.

[0130] Next, the appropriateness evaluation unit 1003 calculates the correlation matrix between the formulation 103 (especially the flow rates of each etching gas) and the compression characteristic 111 as an indicator for judging appropriateness (step 1204). Here, the correlation matrix of the two multidimensional probability variables a and b refers to the matrix of a arranged in the (i, j) components. i and b j The matrix is ​​obtained by calculating the correlation coefficients. The etching gas flow rate of formulation 103 corresponds to a, and the compression feature 111 corresponds to b. However, it is not necessary to calculate the correlation coefficients for all components.

[0131] Furthermore, the appropriateness evaluation unit 1003 determines the appropriateness of the compressed feature 111 based on whether the calculated correlation coefficient is greater than a given threshold (step 1205). The threshold is input from the input / output device 112 as user setting information.

[0132] Finally, the appropriateness evaluation unit 1003 sends the appropriateness-related information (appropriateness information) to the input / output device 112 and ends the process (step 1206). By calculating the correlation coefficient, it can be confirmed whether the emission spectrum intensity 202 of the associated free radicals increases when the flow rate of a certain etching gas is high. The presence or absence of correlation is determined, for example, by whether the free radicals corresponding to the compression characteristic amount 111 of the type of etching gas interpreted as formulation 103 contain common atoms. Furthermore, in the absence of appropriateness, the dimension can be readjusted in the dimension determination unit 108.

[0133] (Example of a GUI)

[0134] Figure 15 This is a diagram illustrating an example of a GUI for setting parameters used to evaluate the appropriateness of compression feature 111 and confirming the results. Screen 1301 is shown, for example, on the display device of input / output device 112. Screen 1301 includes setting / appropriateness information 1004, and appropriateness information is notified to the user via screen 1301. In addition, the user's setting information is sent to the appropriateness evaluation unit 1003.

[0135] In the intensity threshold setting unit 1302, in order to extract the emission spectral intensity corresponding to each component vector 1001 of the emission spectral intensity 202, a threshold is set for the emission spectral intensity 202. Figure 15 In the text, methods for setting the threshold are shown, including specifying a constant or setting the threshold as p% of the maximum value in each component vector. Figure 15 In the latter case, p is set to 50. The selected threshold condition is notified to the appropriateness evaluation department 1003 as setting information.

[0136] In the correlation coefficient threshold setting unit 1303, a threshold is set for the correlation coefficient that serves as the benchmark for appropriateness evaluation. Figure 15 In the input form, 0.5 was entered. The entered threshold was notified to the appropriateness evaluation department 1003 as setting information.

[0137] The evaluation results are displayed in the results display unit 1304. Examples of display methods include displaying formulation items, free radicals corresponding to compressibility characteristics, and correlation coefficients. Figure 15 The figure shows that the correlation coefficient of free radicals of etching gas A, whose formulation item is "etching gas A" and whose compressibility characteristic item is represented by "etching gas A", is 0.55, and the correlation coefficient of free radicals of etching gas B, whose formulation item is "etching gas B" and whose compressibility characteristic item is represented by "etching gas B", is 0.45.

[0138] (Function / Effect)

[0139] As described above, in this invention, the number of types of substances (free radicals) generated by the etching gas and the chemical reaction of etching determines the dimensionality to which the sensing data is compressed. During the learning of the shape prediction model, the parameters of the shape prediction model are adjusted to minimize the difference between the compressed sensing data and intermediate features.

[0140] By doing so, according to the present invention, it is possible to compress the sensing data while taking into account the generation process of the sensing data, and to perform shape prediction.

[0141] In addition, generally speaking, by compressing high-dimensional sensing data to low-dimensional data, the number of parameters in the shape prediction model can be reduced, thus avoiding overlearning of the model.

[0142] The embodiments of the present invention have been described above, but the present invention is not limited to the above embodiments, and various modifications can be made without departing from the spirit of the present invention.

[0143] The following description can be, but is not limited to, the content of this invention.

[0144] (Method 1)

[0145] A search device for searching a formula for obtaining a desired etched shape in a plasma processing apparatus, characterized in that the search device performs the following process:

[0146] The dimensions of the sensing data acquired by the plasma processing device are defined based on the material of the sample etched by the plasma processing device, the material of the components constituting the processing chamber for etching the sample, or the number of types of free radicals derived according to the formula.

[0147] The compressed feature quantity is derived using the sensed data compressed to the specified dimension.

[0148] The parameters of the machine learning are adjusted to reduce the difference between the derived compressed feature quantity and the intermediate feature quantity, thereby calculating the intermediate feature quantity according to the recipe;

[0149] The etched shape is predicted based on the calculated intermediate feature values ​​and the formula.

[0150] (Method 2)

[0151] In the search device described in method 1, the characteristic is that,

[0152] The spectral intensity of plasma emission is used as the sensing data.

[0153] (Method 3)

[0154] In the search device described in method 1 or method 2, the characteristic is that,

[0155] If the difference is greater than a given threshold, the recipe or the sensed data is determined to be abnormal.

[0156] (Method 4)

[0157] In the search device described in any one of methods 1 to 3, the characteristic is that,

[0158] The machine learning mentioned is a neural network.

[0159] (Method 5)

[0160] In the search device described in any one of methods 1 to 3, the characteristic is that,

[0161] The spectral intensity is decomposed into a weighted sum of component vectors multiplied by weighting coefficients.

[0162] The compressed feature quantity is represented as a vector obtained by arranging the weight coefficients.

[0163] The component vectors with wavelengths having spectral intensities greater than a given threshold are detected.

[0164] Based on the correlation between the wavelength of the spectral intensity emitted by the free radical and the type of the free radical, a correspondence is established between the gas used for etching and the reaction products generated by the etching and the detected component vector.

[0165] (Method 6)

[0166] In the search device described in any one of methods 1 to 5, the characteristic is that,

[0167] The appropriateness of the compression characteristic quantity is evaluated by analyzing the relationship between the types of free radicals corresponding to each component of the compression characteristic quantity, the gas flow rate specified in the formulation, and the values ​​of each component of the compression characteristic quantity.

[0168] (Method 7)

[0169] In the search device described in any one of methods 1 to 6, the characteristic is that,

[0170] The spectral intensity is decomposed into a weighted sum of component vectors multiplied by weighting coefficients through nonnegative matrix decomposition.

[0171] (Method 8)

[0172] A data compression processing apparatus for compressing sensing data acquired from a plasma processing apparatus, characterized in that the data compression processing apparatus performs the following processing:

[0173] The dimensions of the sensing data acquired by the plasma processing device are defined based on the material of the sample etched by the plasma processing device, the material of the components constituting the processing chamber for etching the sample, or the number of types of free radicals derived according to the formula.

[0174] The compressed feature quantity is derived using the sensed data compressed to the specified dimension.

[0175] (Method 9)

[0176] A semiconductor device manufacturing system includes a platform with an application installed, the application being used to search for recipes in a plasma processing apparatus to obtain desired etched shapes, characterized in that the semiconductor device manufacturing system executes the following via the application:

[0177] The steps for defining the dimensions of the sensing data acquired by the plasma processing device based on the material of the sample etched by the plasma processing device, the material of the components constituting the processing chamber for etching the sample, or the number of types of free radicals derived according to the formula.

[0178] The step of deriving compressed feature quantities using the sensed data compressed to the specified dimensions;

[0179] The steps of adjusting the machine learning parameters to reduce the difference between the derived compressed feature quantity and the intermediate feature quantity, thereby calculating the intermediate feature quantity according to the recipe; and

[0180] The step of predicting the etched shape based on the calculated intermediate feature values ​​and the formula.

[0181] (Method 10)

[0182] A semiconductor device manufacturing system includes a platform with an application program installed for compressing sensing data acquired from a plasma processing device. The semiconductor device manufacturing system performs the following steps:

[0183] The steps for defining the dimensions of the sensing data acquired by the plasma processing device based on the material of the sample etched by the plasma processing apparatus, the material of the components constituting the processing chamber for etching the sample, or the number of types of free radicals derived according to the formula; and

[0184] The step of deriving compressed feature quantities using the sensed data compressed to the specified dimensions.

[0185] (Method 11)

[0186] A search method for searching a formula for obtaining a desired etched shape in a plasma processing apparatus, characterized in that the search method comprises:

[0187] The steps for defining the dimensions of the sensing data acquired by the plasma processing device based on the material of the sample etched by the plasma processing device, the material of the components constituting the processing chamber for etching the sample, or the number of types of free radicals derived according to the formula.

[0188] The step of deriving compressed feature quantities using the sensed data compressed to the specified dimensions;

[0189] The steps of adjusting the machine learning parameters to reduce the difference between the derived compressed feature quantity and the intermediate feature quantity, thereby calculating the intermediate feature quantity according to the recipe; and

[0190] The step of predicting the etched shape based on the calculated intermediate feature values ​​and the formula.

[0191] (Method 12)

[0192] A data compression processing method for compressing sensing data acquired from a plasma processing device, characterized in that the data compression processing method comprises:

[0193] The steps for defining the dimensions of the sensing data acquired by the plasma processing device based on the material of the sample etched by the plasma processing apparatus, the material of the components constituting the processing chamber for etching the sample, or the number of types of free radicals derived according to the formula; and

[0194] The step of deriving compressed feature quantities using the sensed data compressed to the specified dimensions.

[0195] Explanation of reference numerals in the attached figures

[0196] 1, 1A, 1B: Semiconductor device manufacturing system; 101: Plasma processing device; 102: Sensor; 103: Recipe; 104: Sensing data; 105: Processing shape; 106: Database; 107, 107A: Compression processing device; 108: Dimension determination unit; 109: Material information; 110: Dimension compression unit; 111: Compression feature quantity; 112: Input / output device; 113: Display information; 114: Etching shape prediction device; 115: Intermediate feature quantity calculation unit; 116: Shape calculation unit; 117: Parameter adjustment unit; 118: Intermediate feature quantity; 119: Predicted shape; 120: Learning device; 121: Abnormal data detection unit; 122: Abnormal information; 201: Table; 202: Emission spectral intensity; 601: Emission spectrophotometer; 1001: Component vector; 1002, 1002A, 1002B: Emission wavelength-free radical correspondence table.

Claims

1. A search device for searching a formula in a plasma processing apparatus for obtaining a desired etched shape, characterized in that, The search device performs the following processing: The dimensions of the sensing data acquired by the plasma processing device are defined based on the material of the sample etched by the plasma processing device, the material of the components constituting the processing chamber for etching the sample, or the number of types of free radicals derived according to the formula. The compressed feature quantity is derived using the sensed data compressed to the specified dimension. The parameters of the machine learning are adjusted to reduce the difference between the derived compressed feature quantity and the intermediate feature quantity, thereby calculating the intermediate feature quantity according to the recipe; The etched shape is predicted based on the calculated intermediate feature values ​​and the formula.

2. The search device according to claim 1, characterized in that, The spectral intensity of plasma emission is used as the sensing data.

3. The search device according to claim 1, characterized in that, If the difference is greater than a given threshold, the recipe or the sensed data is determined to be abnormal.

4. The search device according to claim 1, characterized in that, The machine learning mentioned is a neural network.

5. The search device according to claim 2, characterized in that, The spectral intensity is decomposed into a weighted sum of component vectors multiplied by weighting coefficients. The compressed feature quantity is represented as a vector obtained by arranging the weight coefficients. The component vectors with wavelengths having spectral intensities greater than a given threshold are detected. Based on the correlation between the wavelength of the spectral intensity emitted by the free radical and the type of the free radical, a correspondence is established between the gas used for etching and the reaction products generated by the etching and the detected component vector.

6. The search device according to claim 5, characterized in that, The appropriateness of the compression characteristic quantity is evaluated by analyzing the relationship between the types of free radicals corresponding to each component of the compression characteristic quantity, the gas flow rate specified in the formulation, and the values ​​of each component of the compression characteristic quantity.

7. The search device according to claim 5, characterized in that, The spectral intensity is decomposed into a weighted sum of component vectors multiplied by weighting coefficients through nonnegative matrix decomposition.

8. A data compression processing apparatus for compressing sensing data acquired from a plasma processing apparatus, characterized in that, The data compression processing device performs the following processing: The dimensions of the sensing data acquired by the plasma processing device are defined based on the material of the sample etched by the plasma processing device, the material of the components constituting the processing chamber for etching the sample, or the number of types of free radicals derived according to the formula. The compressed feature quantity is derived using the sensed data compressed to the specified dimension.

9. A semiconductor device manufacturing system comprising a platform with an application program installed, the application program being used to search for recipes in a plasma processing apparatus for obtaining desired etched shapes, characterized in that, The semiconductor device manufacturing system is executed through the application: The steps for defining the dimensions of the sensing data acquired by the plasma processing device based on the material of the sample etched by the plasma processing device, the material of the components constituting the processing chamber for etching the sample, or the number of types of free radicals derived according to the formula. The step of deriving compressed feature quantities using the sensed data compressed to the specified dimensions; The step of adjusting the parameters of the machine learning to reduce the difference between the derived compressed feature quantity and the intermediate feature quantity, thereby calculating the intermediate feature quantity according to the recipe; and The step of predicting the etched shape based on the calculated intermediate feature values ​​and the formula.

10. A semiconductor device manufacturing system comprising a platform with an application program installed, the application program being used to compress sensing data acquired from a plasma processing device, characterized in that, The semiconductor device manufacturing system performs the following steps: The steps for defining the dimensions of the sensing data acquired by the plasma processing device based on the material of the sample etched by the plasma processing device, the material of the components constituting the processing chamber for etching the sample, or the number of types of free radicals derived according to the formula. and The step of deriving compressed feature quantities using the sensed data compressed to the specified dimensions.

11. A search method for searching a formulation in a plasma processing apparatus for obtaining a desired etched shape, characterized in that, The search method has the following characteristics: The steps for defining the dimensions of the sensing data acquired by the plasma processing device based on the material of the sample etched by the plasma processing device, the material of the components constituting the processing chamber for etching the sample, or the number of types of free radicals derived according to the formula. The step of deriving compressed feature quantities using the sensed data compressed to the specified dimensions; The step of adjusting the parameters of the machine learning to reduce the difference between the derived compressed feature quantity and the intermediate feature quantity, thereby calculating the intermediate feature quantity according to the recipe; and The step of predicting the etched shape based on the calculated intermediate feature values ​​and the formula.

12. A data compression processing method for compressing sensing data acquired from a plasma processing device, characterized in that, The data compression processing method has the following characteristics: The steps for defining the dimensions of the sensing data acquired by the plasma processing device based on the material of the sample etched by the plasma processing device, the material of the components constituting the processing chamber for etching the sample, or the number of types of free radicals derived according to the formula. and The step of deriving compressed feature quantities using the sensed data compressed to the specified dimensions.

Images