Method of generating learning model and apparatus for predicting semiconductor device structure
By refining the similarity of training data, using learning models of non-destructive metrological data and structural data, the accuracy and overfitting problems of semiconductor device structure prediction in the prior art are solved, and the prediction accuracy and data consistency are improved .
Patent Information
- Application Number
- CN202410713503.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-06-05
- Filing Date
- 2024-06-04
- Publication Date
- 2024-12-06
AI Technical Summary
In the existing semiconductor device structure prediction method based on spectral analysis, training structure prediction model needs to be preceded by analyzing data, and overfitting may result in inappropriate data, affecting the prediction accuracy.
By generating a learning model based on training data, the training data is refined using the similarity of non-destructive metrological data and structural data, the low-similarity data are excluded, and the prediction accuracy of the model is improved.
It improves the accuracy of semiconductor device structure prediction, prevents overfitting of the model, and enhances the consistency of the training data.
Smart Images

Figure CN119089945A_ABST
Abstract
Description
[0001] This application is based on and claims the benefit of priority from Korean Patent Application No. 10-2023-0072214, filed on June 5, 2023, in the Korean Intellectual Property Office, the disclosure of which is incorporated herein in its entirety by reference. Technical Field
[0002] The disclosed embodiments described herein relate to a method of generating a learning model predicting a structure of a semiconductor device and an apparatus for predicting a structure of a semiconductor device. Background Art
[0003] Various manufacturing processes are performed to produce semiconductor devices. The shape of the semiconductor device may vary according to process variations applied in the various manufacturing processes.
[0004] As a method of analyzing the structure or shape of a semiconductor device, there is a method of analyzing data measured from the semiconductor device based on spectroscopy, or a method of analyzing an analysis sample such as an electron microscope image.
[0005] The method of analyzing data measured from a semiconductor device based on spectral analysis is a non-destructive method and can be applied during the process. However, in this method, training a structure prediction model based on samples for analysis requires a method prior to analyzing the data. In addition, overfitting may occur due to some inappropriate data. Summary of the invention
[0006] Provided are a method of generating a learning model for predicting a structure of a semiconductor device based on training data with improved consistency, and an apparatus for predicting a structure of a semiconductor device.
[0007] According to one aspect of the disclosure, an apparatus for predicting a structure of a semiconductor device includes: at least one processor; a storage device configured to store a learning model, the learning model configured to predict the structure of the semiconductor device; and a memory configured to store at least one code, and at least one processor operably connected to the memory and configured to execute the at least one code to: input non-destructive metrology data measured from the semiconductor device into the learning model, and predict the structure of the semiconductor device based on the learning model, wherein the learning model is trained using training data, the training data includes first data as non-destructive metrology data and second data as structural metrology data, the second data being reference data for the first data, and wherein the training data is refined based on similarity of the training data in a space having a first axis corresponding to the first data as a reference axis and a second axis corresponding to the second data.
[0008] The first data is metrology data of the semiconductor device based on a spectrum, and the second data is structure data obtained by measuring at least a part of the structure of the semiconductor device.
[0009] According to one disclosed aspect, a method for generating a learning model configured to predict a structure of a semiconductor device includes: identifying training data, the training data including: first data as non-destructive metrology data of the semiconductor device and second data as structural data, the second data being reference data for the first data; refining the training data based on similarity of the training data in a first space having a first axis corresponding to the first data as a first reference axis and a second axis corresponding to the second data; and training the learning model based on the refined training data.
[0010] The first axis is scaled within the range of the second axis.
[0011] The first data is spectrum-based metrology data of the semiconductor device, and the second data is obtained by measuring at least a portion of a structure of the semiconductor device.
[0012] The step of refining the training data includes: determining a search vector in a low-dimensional space in which the first data is reduced in dimension; and refining the training data in a second space represented by a third axis corresponding to the search vector and a fourth axis corresponding to the second data.
[0013] The search vector is a vector having the highest similarity in the second space to the training data among any vectors in the low-dimensional space in which the first data is reduced in dimension.
[0014] A third axis corresponding to the search vector is scaled based on the second data.
[0015] The search vector is a vector in a low-dimensional space in which the first data is reduced in dimension based on at least one of principal component analysis, singular vector decomposition, non-negative matrix decomposition, and partial least squares.
[0016] The step of refining the training data includes visualizing by displaying the training data in a second space represented by a third axis corresponding to the search vector and a fourth axis corresponding to the second data.
[0017] The step of refining the training data includes: receiving user input for the visualization result; and refining the training data based on the user input.
[0018] The similarity is based on at least one of cosine similarity of the training data in the first space, a correlation coefficient, and a residual-based linear regression model evaluation method.
[0019] According to one disclosed aspect, an apparatus for generating a learning model configured to predict a structure of a semiconductor device includes: a memory configured to store at least one code; and at least one processor operably connected to the memory and configured to execute the at least one code to: train the learning model using training data, the training data including: first data as non-destructive metrology data measured from the semiconductor device and second data as structural data, the second data being reference data of the first data, and based on similarity of the training data in a first space having a first axis corresponding to the first data as a first reference axis and a second axis corresponding to the second data, refine the training data.
[0020] The first axis is scaled within the range of the second axis.
[0021] The at least one processor is configured to execute the at least one code to: determine a search vector in a low-dimensional space in which the first data is reduced in dimension, and refine the training data in a second space having a third axis corresponding to the search vector as a second reference axis and a fourth axis corresponding to the second data.
[0022] The search vector is a vector having the highest similarity in the second space to the training data among any vectors in the low-dimensional space in which the first data is reduced in dimension.
[0023] A third axis corresponding to the search vector is scaled based on the second data.
[0024] The at least one processor is configured to execute the at least one code to: visualize by displaying the training data in a second space having a third axis corresponding to the search vector as a second reference axis and a fourth axis corresponding to the second data.
[0025] The at least one processor is configured to execute the at least one code to: receive user input for a visualization result, and refine the training data based on the user input.
[0026] The similarity is based on at least one of cosine similarity of the training data in the first space, a correlation coefficient, and a residual-based linear regression model evaluation method. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] The above and other aspects, features and advantages of the disclosed specific embodiments will become more apparent from the following description taken in conjunction with the accompanying drawings, in which:
[0028] Figure 1 An apparatus for predicting the structure of a semiconductor device according to a disclosed embodiment is shown;
[0029] Figure 2A configuration of an apparatus for predicting a structure of a semiconductor device according to a disclosed embodiment is shown;
[0030] Figure 3 An operating method of an apparatus for predicting a structure of a semiconductor device according to a disclosed embodiment is shown;
[0031] Figure 4 A configuration of an apparatus for generating a learning model for predicting a structure of a semiconductor device according to a disclosed embodiment is shown;
[0032] Figure 5 A method for training a learning model for predicting the structure of a semiconductor device according to a disclosed embodiment is shown;
[0033] Figure 6 A method of refining training data for training a learning model for predicting the structure of a semiconductor device according to a disclosed embodiment is shown;
[0034] FIG. 7A to FIG. 7G A method for refining training data according to a disclosed embodiment is shown;
[0035] Figure 8 A method for determining similarity of training data according to a disclosed embodiment is shown;
[0036] Fig. 9 showing a user interface for refining training data having low similarity according to a disclosed embodiment; and
[0037] Fig. 10A and Fig. 10B Experimental results of a learning model for predicting the structure of a semiconductor device according to the disclosed embodiment are shown. DETAILED DESCRIPTION
[0038] Hereinafter, the disclosed embodiments may be described in detail and clearly to the extent that a person having ordinary skill in the art can easily realize the disclosure.
[0039] The description only illustrates the disclosed principles. Those skilled in the art will be able to design one or more arrangements that embody the disclosed principles, although not explicitly described herein. In addition, all examples stated herein are primarily intended to be used explicitly only for explanatory purposes to help the reader understand the disclosed principles and the concepts contributed by the inventors to advance the art, and should be interpreted as not being limited to such specifically stated examples and conditions. In addition, all descriptions of the disclosed principles, aspects, and embodiments, as well as their specific examples, are intended to cover their equivalents.
[0040] The terms used in the present disclosure are only used to describe specific embodiments, and may not be intended to limit the scope of another embodiment. Unless clearly indicated differently in the context, singular expressions may include plural expressions. The terms used herein (including technical terms or scientific terms) may have the same meaning as the meanings commonly understood by personnel with ordinary knowledge in the technical field described in the present disclosure. The terms defined in the general dictionary among the terms used in the present disclosure can be interpreted using the same or similar meanings as the contextual meanings of the relevant technology, and unless clearly defined in the present disclosure, are not interpreted with ideal or overly formal meanings. In some cases, even the terms defined in the present disclosure cannot be interpreted as excluding embodiments of the present disclosure.
[0041] In one or more embodiments of the disclosure described below, a hardware method is described as an example. However, since one or more embodiments of the disclosure include techniques using both hardware and software, various embodiments of the present disclosure do not exclude software-based methods.
[0042] In addition, in the present disclosure, in order to determine whether a specific condition is satisfied or achieved, an expression of greater than or less than may be used, but this is only a description for expressing an example, and does not exclude a description of greater than or equal to or less than or equal to. A condition described as "greater than or equal to" may be replaced with "greater than", a condition described as "less than or equal to" may be replaced with "less than", and a condition described as "greater than or equal to and less than" may be replaced with "greater than and less than or equal to". In addition, hereinafter, "A" to "B" means at least one of the elements from A (including A) to B (including B).
[0043] The terms "include" and "comprising" and their derivatives mean including but not limited. The term "or" is an inclusive term meaning "and / or". The phrase "associated with..." and its derivatives mean including, included in, interconnected with, contained in, connected to or connected with, combined to or combined with, communicable with, collaborate with, interweave, juxtapose, be close to, bind to or bind with, have, have the property of, have a relationship with, etc. The term "controller" means any device, system or part thereof that controls at least one operation. Such a controller can be implemented in hardware or a combination of hardware and software and / or firmware. Whether local or remote, the functions associated with any particular controller can be centralized or distributed. When used with a list of items, the phrase "at least one of..." means that different combinations of one or more of the listed items can be used, and only one item in the list may be required. For example, "at least one of A, B, and C" includes any of the following combinations: A, B, C, A and B, A and C, B and C, A and B and C, and any variation thereof. The expression "at least one of a, b, or c" may mean only a, only b, only c, both a and b, both a and c, both b and c, all of a, b, and c, or variations thereof. Similarly, the term "set" means one or more. Thus, a set of items may be a single item or a set of two or more items.
[0044] Figure 1 is a diagram illustrating an apparatus 100 for predicting a structure of a semiconductor device and an application environment according to an embodiment of the disclosure.
[0045] According to the disclosed embodiment, the apparatus 100 for predicting the structure of a semiconductor device (also referred to as a "prediction apparatus") is configured to predict the structure of a semiconductor device using a learning model. The training data of the learning model is refined based on the similarity of the training data including first data and second data (the second data is reference data of the first data). Based on the first data (hereinafter referred to as the first data) taking into account the second data 7A to 7FThe first projection data described in the present invention is used to exclude training data with low similarity from a final training data set based on the similarity of the first projection data in a space having a first axis "based on" the first data (corresponding to the first data) and a second axis "based on" the second data (corresponding to the second data) as a reference axis, and the training data can be refined. In the prediction device 100, due to the training data with improved similarity, the prediction accuracy of the structure of the semiconductor device is improved. In one embodiment, the first axis corresponding to the first data may be an axis corresponding to the first projection data. In one embodiment, the similarity of the first projection data may represent the similarity of the training data. In one embodiment, the first projection data is obtained based on the first data and the second data.
[0046] Reference Figure 1 , the prediction device 100 includes a processor 110, a memory 120, and a learning model 131. Throughout the disclosure, the processor 110 represents one or more processors or at least one processor. Figure 1 In the embodiment, the processor 110 may be replaced by a plurality of processors.
[0047] Semiconductor devices such as logic and memory devices are manufactured through a plurality of manufacturing processes applied to a substrate or wafer 300. Various features, structures, and shapes of the semiconductor device are formed based on at least one process among the plurality of manufacturing processes.
[0048] The wafer metrology instrument 200 measures the wafer 300 using various methods such as non-destructive metrology methods such as optical metrology, electromagnetic metrology, and X-ray-based metrology. The wafer metrology instrument 200 transmits at least one of an optical signal, an electromagnetic wave, and an X-ray as a source for metrology to the wafer 300. The wafer metrology instrument 200 observes and analyzes at least one of the optical signal, the electromagnetic wave, and the X-ray obtained based on at least one phenomenon of reflection, transmission, scattering, and diffraction when the emitted metrology source is emitted from the wafer 300 to generate a metrology value. In one embodiment, the wafer metrology instrument 200 includes or corresponds to any one or all of various types of metrology devices that obtain metrology data by non-destructively transmitting at least one of an optical signal, an electromagnetic wave, and an X-ray onto the wafer 300. Therefore, the wafer metrology instrument 200 is not specifically limited to any type of metrology device.
[0049] The wafer metrology instrument 200 may be a spectroscopic ellipsometer or a spectroscopic reflectometer. The wafer metrology instrument 200 may be in the form of (or may involve or perform) scatterometry or reflectometry. The wafer metrology instrument 200 may be in the form of X-ray based scatterometry or X-ray based reflectometry. The wafer metrology instrument 200 may be an optical critical dimension metrology instrument.
[0050] The wafer metrology instrument 200 includes an analysis algorithm for at least one of an optical signal, an electromagnetic wave, and an X-ray emitted together with an optical structure of a scatterometer, a reflectometer, and an ellipsometry meter and an electromagnetic instrument structure. The wafer metrology instrument 200 outputs critical dimensions, film thickness, band gap, composition, overlap, and various structural or morphological parameters of nanoscale structures of semiconductor devices as metrology data. The metrology data of the wafer metrology instrument 200 may be in the form of a spectrum.
[0051] The metrology data may be one of various parameters depending on the type of wafer metrology instrument 200 and metrology source.
[0052] For example, the ellipsometer metrology data may include a plurality of ellipsometric parameters within a spectral range obtained from a metrology region of the wafer 300 .
[0053] The wafer metrology instrument 200 may provide the metrology data (obtained by measuring the wafer 300) to the prediction apparatus 100 via an online (eg, real-time) or offline file. The metrology data is non-destructive metrology data.
[0054] The prediction apparatus 100 according to the disclosed embodiment predicts the structure of a semiconductor device by inputting metrology data provided from the wafer metrology instrument 200 to the learning model 131. In one embodiment, the prediction apparatus 100 may predict a wafer pattern structure.
[0055] The memory 120 of the prediction device 100 is configured to store “code” for causing the processor 110 to perform operations. The “code” refers to a set of instructions loaded into the memory 120 .
[0056] The processor 110 is configured to load the learning model 131 from the storage device 130. The processor 110 is configured to apply various weighting parameters of the learning model 131 to the metering data as input data. The processor 110 is configured to generate output values from values calculated based on the weighting parameters.
[0057] In one embodiment, the learning model 131 may be a deep learning-based model having multiple layers including a neural network. In one embodiment, the learning model 131 may be a machine learning-based model. The training of the learning model 131 may be performed or completed based on the training data.
[0058] According to the disclosed embodiments, the structure of the semiconductor device predicted by the prediction apparatus 100 based on the learning model 131 may be a length, a width, etc. related to the morphology, structure, and geometry of at least a portion of a pattern of the semiconductor device.
[0059] The structure of the semiconductor device predicted by the prediction apparatus 100 may include all numerical values measurable in each of some manufacturing processes of the semiconductor device.
[0060] For example, the structure of a semiconductor device may be a numerical value related to any structural shape (of various types of patterns) (such as film thickness, groove height, pattern thickness or width, profile height, sidewall angle, spacing, intermediate critical dimension (MCD), various critical dimensions (critical dimensions of top, middle, bottom, neck, etc.), channel hole height / radius, mask height, etc.).
[0061] The learning model 131 is trained using training data including metrology data measured by the wafer metrology instrument 200 in each of some manufacturing processes and reference data, which are numerical values related to shape. The reference data may be referred to as "label data" for the metrology data.
[0062] For example, the learning model 131 may be trained using training data including a spectrum measured after performing a thin film deposition process and a thickness of a thin film as reference data.
[0063] For example, the learning model 131 may be trained using training data including a spectrum measured after performing an etching process and an etching depth or width as reference data.
[0064] The structure of the semiconductor device as reference data may be measured using, for example, a transmission electron microscope (TEM), a scanning electron microscope (SEM), an atomic force microscope (AFM) for a sample for analysis after any or each of some manufacturing processes is performed.
[0065] Reference data may be collected from samples for analysis prepared in a destructive manner after each manufacturing process is performed. The reference data may be structural data measured on at least a portion of the area of the wafer 300 from which the metrology data is collected. The sample for analysis may be a sample of at least a portion of the area of the wafer 300 from which the metrology data is collected.
[0066] According to the disclosed embodiments, the training data of the learning model 131 is refined based on the similarity of the training data. The training data is refined based on the similarity of the metrology data (e.g., first projection data) for which the reference data is considered. In one embodiment, the training data is refined in a space having a first axis (based on the reference data) as a reference axis and a second axis (based on the metrology data) (e.g., an axis based on the first projection data).
[0067] The training data may be refined based on, for example, similarities in trends, similarities in distribution directions between metrology data for which reference data is considered.
[0068] The similarity of the meter data for which the reference data is considered may be determined based on a distribution trend of the plurality of pieces of meter data for which the reference data is considered.
[0069] The similarity of the metrology data for which the reference data is considered may be determined based on the similarity between the plurality of pieces of first projection data located in a space having an axis based on the metrology data and an axis based on the reference data as a reference axis of the space.
[0070] The refinement of training data can be referred to Figure 4 The "learning model generation device" described above is implemented.
[0071] For example, when there is a linear trend among the metrology data for which the reference data is considered, the learning model generating apparatus may exclude training data outside the range of a preset linear trend from the final training data set.
[0072] In one embodiment, when there is a linear trend between the metrology data for which the reference data is considered, the learning model generation device may visualize the linear trend. The learning model generation device may receive user input on the visualization result, and may exclude some training data from the final training data set based on the user input.
[0073] In the related art, a learning model for predicting the structure of a semiconductor device is trained by refined training data based on user's empirical knowledge. Alternatively, in the related art, a learning model is trained by refined training data based only on characteristics of metrology data such as spectral data.
[0074] In detail, in the related art, after analyzing metrology data (such as spectral data measured from a semiconductor device), a learning model is trained using training data refined based on the analysis value of the metrology data. Such a technique in the related art is based only on the characteristics of the spectral metrology data in which the reference data is not reflected. The technique in the related art refines the first projection data in a space consisting only of axes based on the metrology data. Therefore, the technique in the related art uses the reference data only in the parameter training process of the learning model. As a result, in the related art, training data (outliers) with low similarity are not excluded from the training of the learning model, and at least a portion of the characteristics of the outliers are reflected in the parameters of the learning model. Therefore, in the related art, the prediction accuracy of the learning model is also reduced.
[0075] In contrast, the learning model 131 of the prediction apparatus 100 according to the disclosed embodiment is trained using refined training data based on the characteristics of the metrology data in which the reference data is reflected. The prediction apparatus 100 predicts the structure of the semiconductor device based on the learning model 131. The training data of the learning model 131 is refined based on the similarity between the first projection data in a space having a first axis based on the metrology data as a reference axis of the space and a second axis based on the reference data.
[0076] Therefore, the learning model 131 of the prediction device 100 according to the disclosed embodiment is trained by the training data in which the outliers are excluded in consideration of the reference data. The training data in which the outliers are excluded in consideration of the reference data have high similarity between the training data. As a result, according to the disclosed embodiment, the characteristics of the outliers are not reflected in the final parameters of the learning model 131. Therefore, even when the learning model 131 is based on a small amount of training data, the learning model 131 is not overfitted, and the prediction accuracy of the structure of the semiconductor device by the learning model 131 is also improved.
[0077] Figure 2 is a block diagram illustrating a configuration of an apparatus 100 for predicting a structure of a semiconductor device according to a disclosed embodiment.
[0078] The prediction apparatus 100 predicts the structure of the semiconductor device based on the learning model 131. The learning model 131 is trained using refined training data reflecting the characteristics of the metrology data and the characteristics of the reference data. The learning model 131 is trained by the refined training data based on the similarity of the metrology data in which the reference data is considered. The training data of the learning model 131 is refined based on the similarity between a plurality of first projection data located in a space having a first axis based on the metrology data as a reference axis and a second axis based on the reference data.
[0079] like Figure 2 As shown in FIG. 1 , the prediction device 100 includes a processor 110, a memory 120, a storage device 130, a display 140, a user interface 150, and a network transceiver 160. Code loaded and temporarily stored in the memory 120 causes the processor 110 to perform operations.
[0080] The processor 110 is configured to load the learning model 131 from the storage device 130 . The processor 110 is configured to temporarily store metrology data received from the wafer metrology instrument 200 online or offline in the memory 120 , and to input the metrology data to the learning model 131 .
[0081] The processor 110 may include an artificial intelligence-based learning processor that accelerates deep learning or machine learning operations. The learning processor may be a processor including a graphics processing unit (GPU), a tensor processor, a neural processing unit (NPU), and a digital signal processor (DSP). The processor 110 may be a processor combined with a memory.
[0082] The processor 110 is configured to apply the weighted parameters of the learning model 131 to the metrology data based on deep learning. The processor 110 is configured to input the output values output from the nodes of each layer of the learning model 131 to the nodes of the subsequent layer based on the neural network structure. The processor 110 is configured to output the structure (such as length and width) of the morphology, structure and geometry of at least a portion of the pattern of the semiconductor device using the learning model 131.
[0083] Processor 110 is configured to input metrology data to learning model 131 based on machine learning. Learning model 131 is configured to predict structure (such as length, width, etc.) regarding morphology, structure, and geometry of at least a portion of a pattern of a semiconductor device.
[0084] For example, in the case of a decision tree-based learning model 131, the processor 110 is configured to input a vector consisting of at least some numerical values of the metrology data to a root node of the decision tree as an input vector. The learning model 131 is configured to output the structure of the semiconductor device based on a branch tree structure according to the determination criteria of each node of the decision tree.
[0085] For example, in the case of a deep learning-based learning model 131, the processor 110 is configured to input a vector consisting of at least some numerical values of the metrology data to each node of the input layer as an input vector. The learning model 131 is configured to output the structure of the semiconductor device based on the network structure and weights of the neural network.
[0086] The memory 120 may be configured to (temporarily) store codes for the operation of the prediction device 100 , data for the operation of the processor 110 , weighting parameters of the learning model 131 , intermediate operation results of the learning model 131 , and the like.
[0087] The storage device 130 is configured to store the learning model 131 trained by the learning model generation device. The storage device 130 may include a computer-readable storage medium. The storage medium includes all types of recording devices storing data that can be read by a computer. The storage medium may include at least one of a hard disk drive (HDD), a solid state disk (SSD), a silicon disk drive (SDD), a read-only memory (ROM), a random access memory (RAM), a compact disk (CD)-ROM, a magnetic tape, a floppy disk, and an optical data storage device.
[0088] According to an embodiment, the storage device 130 may be configured to classify the learning model 131 into a plurality of learning models and store the plurality of learning models according to the learning time, the type of training data, or the structure of the semiconductor device to be predicted. For example, different learning models trained using different training data consisting of different metrology data and reference data of different semiconductor device structures may be stored in the storage device 130. Different learning models 131 may be applied according to the metrology data.
[0089] The learning model 131 may correspond to a deep learning-based model having a plurality of layers with a neural network. The learning model 131 may correspond to a machine learning-based model.
[0090] The neural network of the learning model 131 may include at least one of a convolutional neural network (CNN), a regional convolutional neural network (R-CNN), a region proposal network (RPN), a recursive neural network (RNN), a long short-term memory (LSTM) network, a stacked deep neural network (S-DNN), a state-space dynamic neural network (S-SDNN), a deep belief network (DBN), and a restricted Boltzmann machine (RBM), and other structures of the neural network structure are not excluded.
[0091] The learning model 131 is a model based on machine learning, which may correspond to a model based on a decision tree, association rules, genetic algorithm, inductive learning method, support vector machine (SVM), cluster analysis, Bayesian network, reinforcement learning method, regression model. The model does not exclude other structures based on machine learning structures.
[0092] The learning model 131 may be implemented in hardware, software, or a combination of hardware and software. When part or all of the learning model 131 is implemented using software, one or more commands constituting the learning model may be stored in the storage device 130 .
[0093] According to the disclosed embodiment, the learning model 131 is trained using refined training data reflecting characteristics of metrology data and characteristics of reference data.
[0094] The learning model 131 is trained using training data including metrology data measured by the wafer metrology instrument 200 in each of some manufacturing processes and reference data of the metrology data.
[0095] The learning model 131 is trained using training data including non-destructive metrology data and structural metrology data that is reference data for the non-destructive metrology data.
[0096] In one embodiment, the training data of the learning model 131 includes spectral-based metrology data of the semiconductor device (the metrology data is non-destructive metrology data) and reference data. The reference data can be structural data obtained by directly measuring the length, width, height, etc. of at least a part of the structure of the semiconductor device by SEM, TEM, AFM, etc.
[0097] The training data of the learning model 131 is refined based on the similarity between the non-destructive metrology data in which the structural data as the reference data of the non-destructive metrology data is considered. The training data of the learning model 131 is training data refined in a space having a first axis based on the structural data as a reference axis (the structural data is the reference data of the non-destructive metrology data) and a second axis based on the non-destructive metrology data (for example, an axis based on the first projection data).
[0098] The display 140 is configured to display the result of predicting the structure of the semiconductor device by the prediction apparatus 100 using the learning model 131. The display 140 is not limited to a visual display, and may display the result based on vision, hearing, or touch.
[0099] The user interface 150 includes an interface for receiving a user's input for controlling the prediction apparatus 100. The user interface 150 includes an interface for receiving a user's input for an output displayed on the display 140. The user interface 150 may include a graphical user interface (GUI) displayed on the display 140 and a touch input unit implemented on the display 140.
[0100] In one embodiment, the prediction device 100 may receive the metrology data from the wafer metrology instrument 200 online through the network transceiver 160. The network transceiver 160 may include a mobile communication module based on LTE (Long Term Evolution), LTE-A (Long Term Evolution Advanced), etc., a wireless Internet module based on Wireless Fidelity (Wi-Fi), Wireless Local Area Network (WLAN), etc., and a wireless Internet module based on Bluetooth. TM , RFID (radio frequency identification), infrared communication, UWB (ultra-wideband), ZigBee, NFC (near field communication), etc.
[0101] Figure 3 is a flowchart illustrating an operating method of an apparatus for predicting a structure of a semiconductor device (also referred to as a “prediction apparatus”) according to a disclosed embodiment.
[0102] The prediction device receives the learning model generated by the learning model generation device online or offline.
[0103] The learning model receives metrology data obtained by measuring a wafer on which at least one pattern is generated through a wafer metrology instrument, and predicts a structure of a semiconductor device.
[0104] In one embodiment, the structure of the semiconductor device may be a length, width, height, etc. related to the morphology, structure, and geometry of at least one pattern formed on a wafer by at least one semiconductor process. The structure of the semiconductor device may include all structural values that can be measured in each of some manufacturing processes of the semiconductor device.
[0105] The prediction device receives metrology data measured by a wafer metrology instrument either online or offline.
[0106] In operation S110, the prediction device loads the learning model and input data (e.g., metering data). Loading may mean storing at least a portion of various parameters constituting the learning model (e.g., network structure parameters and weighting parameters of the learning model) and at least a portion of the metering data in a memory (e.g., dynamic random access memory (DRAM)) as a temporary storage device.
[0107] In operation S120 , the prediction apparatus predicts the structure of the semiconductor device by inputting the metrology data into a learning model trained using refined training data reflecting characteristics of the metrology data and characteristics of the reference data.
[0108] According to the disclosed embodiment, the training data of the learning model is refined based on the similarity of the metrology data in which the reference data is considered. The training data of the learning model is refined in a space having a first axis based on the reference data as the reference axis of the space and a second axis based on the metrology data.
[0109] Therefore, according to the disclosed embodiments, the prediction apparatus can accurately predict the structure of a semiconductor device.
[0110] Figure 4 is a block diagram showing a configuration of an apparatus (also referred to as a “learning apparatus”) 400 for generating a learning model for predicting a structure of a semiconductor device according to a disclosed embodiment.
[0111] The learning device 400 according to the disclosed embodiment trains a learning model for predicting the structure of a semiconductor device based on non-destructive metrology data. The learning device 400 refines the training data before training the learning model. The learning device 400 refines the training data based on the similarity of the metrology data in which the reference data is considered. The learning device 400 refines the training data in a space having a first axis based on the reference data as a reference axis and a second axis based on the metrology data.
[0112] Will refer to Figure 4The configuration of the learning device 400 is described in detail. Some of the configuration of the learning device 400 may be similar to the configuration of the prediction device 100. Additional descriptions of components similar to or overlapping with the previous description will be omitted to avoid redundancy.
[0113] The learning device 400 is a variety of devices for training a learning model. In one embodiment, the learning device 400 corresponds to a typical server device or a computer designed separately to perform training.
[0114] The learning device 400 may be implemented using a single computer, a set of multiple computers, a cloud computer, or a combination thereof.
[0115] The plurality of learning devices 400 may include a learning device set (or a cloud computer). At least one learning device 400 included in the learning device set may refine the training data 433 and may train the learning model 431 through distributed processing.
[0116] The learning device 400 includes a processor 410 , a memory 420 , a storage device 430 , a display 440 , a user interface 450 , and a network transceiver 460 .
[0117] The processor 410 is configured to load an untrained learning model 431 from the storage device 430. For example, when based on a neural network, the learning model 431 may be a learning model in which values of at least some parameters including weighting parameters among nodes of each layer are not determined.
[0118] The processor 410 is configured to temporarily store at least a portion of the training data 433 in the memory 420. The processor 410 is configured to input the training data to the learning model 431, and train the learning model 431 by changing the weighted parameter value between each node of the learning model.
[0119] In one embodiment, processor 410 may include an artificial intelligence based learning processor that accelerates deep learning or machine learning operations.
[0120] In one embodiment, processor 410 refines the training data before training the learning model.
[0121] The memory 420 may temporarily store codes for operations of the processor 410 and the learning device 400 , weighted parameters of the learning model 431 , intermediate calculation results during learning of the learning model 431 , and the like.
[0122] The storage device 430 includes a computer-readable storage medium such as an HDD or an SDD, and may store a learning model 431 before training and a learning model after training. The storage device 430 stores training data 433 for training the learning model.
[0123] The learning model 431 may be a deep learning-based learning model or a machine learning-based learning model, wherein the deep learning-based learning model is composed of a plurality of layers including a neural network.
[0124] According to the disclosed embodiment, the display 440 may display the learning result of the learning model 431, the process of refining the training data 433, and the visualization result of the training data being refined. The visualization of the training data being refined may represent displaying the distribution of the training data. The visualization of the training data being refined may represent displaying the distribution of the training data into a low-dimensional space based on the metrological data reduced in dimension during the refinement process. The visualization of the training data being refined may represent displaying the distribution of the first projection data corresponding to the training data in a space having a first axis based on the reference data as a reference axis and a second axis based on the metrological data.
[0125] The display 440 may receive a touch input through the user interface 450. The user interface 450 may include a graphical user interface for user selection. The learning device 400 may receive a user input for a refined distribution visualization of the training data through the user interface 450.
[0126] The learning apparatus 400 may transmit the trained learning model to the apparatus for predicting the structure of a semiconductor device through the network transceiver 460. Alternatively, the learning apparatus 400 may provide the trained learning model to the apparatus for predicting the structure of a semiconductor device in an offline method.
[0127] The processor 410 of the learning device 400 according to the disclosed embodiment refines the training data 433 based on the similarity of the metrology data in which the reference data is considered. The processor 410 refines the training data in a space having a first axis (based on the reference data) as a reference axis and a second axis (based on the metrology data). The refinement of the training data is performed before the training of the learning model 431. The learning model 431 is trained based on the refined training data. The training data 433 includes structure-based non-destructive metrology data and reference data.
[0128] The training data 433 may be refined based on, for example, similarity of trends, similarity of distribution directions between metrology data in which the reference data is considered. The training data 433 may be refined based on, for example, similarity of trends, similarity of distribution directions of first projection data in a space having a first axis based on the reference data as a reference axis and a second axis based on the metrology data.
[0129] When similarities exist between the training data based on the linear trend, the learning device 400 may exclude the training data outside the preset range of the linear trend from the final training data set.
[0130] The measurement data of the training data 433 may be reduced in dimension to a low-dimensional space to determine the similarity of trends and the similarity of distribution directions between the training data. The processor 410 may refine the training data 433 based on the measurement data reduced in dimension to the low-dimensional space. The processor 410 may refine the training data 433 in a space having a first axis (based on the reference data) as a reference axis of the space and a second axis (based on the measurement data reduced in dimension).
[0131] The metrology data may be spectrally based metrology data of the semiconductor device, and the reference data may be structural data obtained by measuring at least a portion of the structure of the semiconductor device. The reference data may be, for example, length, width, height related to the morphology, structure, and geometry of at least one pattern formed in each process. The structural data of the semiconductor device may include all structural values that can be measured in each of some manufacturing processes of the semiconductor device.
[0132] The processor 410 is configured to perform training of the learning model 431 based on the refined training data.
[0133] Therefore, the learning model 431 according to the disclosed embodiment is trained based on the training data from which the training data having low similarity is excluded. Therefore, overfitting of the learning model 431 can be prevented, and the learning model 431 can accurately predict the structure of the semiconductor device.
[0134] Figure 5 is a flowchart illustrating an operation method of an apparatus for generating a learning model for predicting a structure of a semiconductor device (also referred to as a “learning apparatus”) according to a disclosed embodiment.
[0135] A learning device according to a disclosed embodiment refines the training data before training a learning model. The learning device refines the training data based on the similarity of the first data (e.g., the first projection data) in which the second data is considered. In one embodiment, the second data is reference data for the first data. The learning device refines the training data based on the similarity of the first projection data in a space having a first axis (based on the first data) as a reference axis and a second axis (based on the second data).
[0136] In one embodiment, a learning device refines training data including first data and second data, wherein the first data is metrology data of a semiconductor device based on spectroscopy, and the second data is data obtained by measuring at least a portion of a structure of the semiconductor device. The second data may be structural data of the semiconductor device measured in at least a portion of a region where the first data is measured.
[0137] Will refer to Figure 5A method of operating a learning device according to an embodiment of the disclosure is described in detail. Additional descriptions of components similar to or overlapping with those described above will be omitted to avoid redundancy.
[0138] In operation S410, the learning device loads at least a portion of the training data including the first data and the second data into a memory. The learning device loads at least an untrained portion of the learning model into the memory. In one embodiment, the learning device may load the untrained learning model into the memory.
[0139] In operation S420, the learning device according to the disclosed embodiment refines the training data based on the similarity of the first data in which the second data is considered. The learning device refines the training data based on the similarity of the first projection data in a space having a first axis (based on the first data) as a reference axis and a second axis (based on the second data).
[0140] The similarity of the training data can be determined in a space having a first axis as a reference axis and a second axis. The first axis is an axis based on a vector in a low-dimensional space in which the first data is reduced in dimension. The second axis is an axis based on the second data.
[0141] In one embodiment, the similarity of the training data may be determined based on data obtained by projecting the first data onto a search vector in a low-dimensional space in which the first data is reduced in dimension.
[0142] The similarity can be the similarity of trends or the similarity of distribution directions of the training data.
[0143] The learning device determines a final training data set by determining similarities of the training data and excluding training data determined to be outside a preset similarity range.
[0144] In operation S430, the learning apparatus according to the disclosed embodiment trains a learning model using the final training data set (refined training data) from which training data having low similarity (eg, similarity less than a threshold value) is excluded.
[0145] The learning device can provide the trained learning model to the device for predicting the structure of the semiconductor device in an online or offline method.
[0146] Therefore, the learning device according to the disclosed embodiment excludes training data with low similarity in the training process. Therefore, overfitting of the learning model is prevented, and the learning model can accurately predict the structure of the semiconductor device.
[0147] Figure 6is a flowchart illustrating a method of refining training data of an apparatus for generating a learning model for predicting a structure of a semiconductor device (also referred to as a “learning apparatus”) according to a disclosed embodiment. FIG. 7A to FIG. 8 is a diagram describing in detail a method of refining training data.
[0148] Before the learning model is trained, the learning device according to the disclosed embodiment refines the training data based on the similarity of the first data in which the second data is considered. The learning device refines the training data based on the similarity of the first projection data in a space having a first axis (based on the first data) as a reference axis of the space and a second axis (based on the second data). The training data includes the first data and the second data as reference data for the first data.
[0149] The method of refining training data according to the disclosed embodiment is not limited to the spectral-based metrology data and the structural data as its reference data. The following description of the embodiment will be described on the premise that the first data is the spectral-based metrology data of the semiconductor device and the second data is the structural data of the semiconductor device. Additional descriptions of components similar to or overlapping with the above components will be omitted to avoid redundancy.
[0150] In one embodiment, the training data includes first data and second data as reference data for the first data. The first data may be spectrum-based metrology data of a semiconductor device. The second data may be data obtained by measuring at least a portion of a structure of a portion of a wafer for which the spectrum-based metrology data is measured.
[0151] The first data is metrological data based on a spectrum, and may be a vector having a plurality of spectrum parameters as spatial axes.
[0152] The similarity of the first data in which the second data is considered (similarity of the first projection data in a space having a first axis based on the first data and a second axis based on the second data as a reference axis) may be determined based on at least one of cosine similarity, correlation coefficient, and a linear regression model evaluation method based on residuals. The linear regression model evaluation method based on residuals may use a coefficient of determination (R 2 score), mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and residual sum of squares (RSS).
[0153] The similarity of the first data in which the second data is considered may be determined by the similarities between the plurality of pieces of training data in which the second data is considered.
[0154] The similarity of the first data in which the second data is considered may be determined based on a distribution trend of the plurality of pieces of training data in which the second data is considered.
[0155] The similarity of the first data in which the second data is considered may be determined based on the similarity between a plurality of pieces of first projection data located in a space having a first axis (based on the first data) and a second axis (based on the second data) as a reference axis.
[0156] Will refer to Figures 6 to 8 A method for refining training data according to a disclosed embodiment is described in detail.
[0157] In operation S510, the learning apparatus reduces spatial dimensions of a plurality of parameters of first data to a low-dimensional space in terms of dimension.
[0158] For example, refer to Fig. 7A and Figure 7B , the learning device performs principal component analysis (PCA) on the first data 710. The learning device reduces the first data 710 in dimension to a space having some of the plurality of principal components determined in the PCA as basic axes. The learning device converts the first data 710 into first data 720 in a space reduced in dimension (e.g., a low-dimensional space). The space reduced in dimension may be a space having a first principal component PC1 and a second principal component PC2 as basic axes, the first principal component PC1 being a component having the maximum variance of the first data 710, and the second principal component PC2 being orthogonal to the first principal component PC1.
[0159] The following description will be described illustratively on the premise that the first data is reduced to a two-dimensional plane based on PCA. However, it is not excluded that the first data is reduced in dimension by other methods (such as singular vector decomposition (SVD), non-negative matrix decomposition (NMF) and partial least squares (PLS)). In addition, it is not excluded that the first data is reduced in dimension to a space of a dimension higher than two dimensions rather than two dimensions.
[0160] Figure 7B It is shown that the first data 720 is located in a space having the first principal component PC1 and the second principal component PC2 as basic axes.
[0161] The learning apparatus may map the first data 710 to first data 720 in a low-dimensional space. FIG. 7B to FIG. 7F The mapping of the first data 720 to the first projection data 730 is shown as a comparison of the first data 720 and the first projection data 730 .
[0162] In operation S520, the learning apparatus according to the disclosed embodiment determines a search vector SV in the low-dimensional space (or on the reduced dimension of the first data 720). A method of determining the search vector SV will be described in detail below.
[0163] In operation S531 , the learning apparatus according to the disclosed embodiment projects the first data 720 in the low-dimensional space onto the search vector SV.
[0164] Reference Fig.7D , Fig. 7E and Figure 7F , the learning device generates the first projection data 730 by projecting the first data 720 in the low-dimensional space reduced in dimension onto the search vector SV of the low-dimensional space. In one embodiment, the learning device may determine the search vector SV based on the second data, and generate the first projection data 730 in which the second data is considered by projecting the first data 720 onto the search vector SV. In one embodiment, the value corresponding to the first principal component PC1 and the value corresponding to the second principal component PC2 of the first projection data 730 or the data corresponding to the axis based on the first projection data 730 (for example, the axis based on the search vector SV) may be determined based on the data obtained by projecting the first data 720 onto the search vector SV. In one embodiment, the value of the first data 710 corresponding to the axis ST based on the second data may be determined as the value of the first projection data 730 corresponding to the axis ST based on the second data.
[0165] In operation S540, the learning device according to the disclosed embodiment refines the training data based on the similarity of the first projection data 730 in which the second data is considered. In one embodiment, the learning device refines the training data based on the similarity of the first projection data 730 in a space having an axis based on the first projection data 730 (e.g., an axis based on the search vector SV) as a reference axis of the space and an axis based on the second data. The learning device refines the training data based on the similarity of the first projection data 730 in a space having an axis based on the search vector SV as a reference axis of the space and an axis based on the second data.
[0166] For example, refer to Fig. 7E and Figure 7F , the first projection data 730 exists on a search vector SV located on a 2D plane reduced in dimension based on the first data 720. The first projection data 730 in which the second data is considered may be located in a space having an axis based on the search vector SV as a reference axis and an axis ST based on the second data.
[0167] The learning apparatus may determine the similarity of the first projection data 730 in a space having an axis based on the search vector SV as a reference axis and an axis ST based on the second data.
[0168] For example, refer to Figure 7F , among the first projection data 730 located in the space having the axis based on the search vector SV as the reference axis and the axis based on the second data ST, the projection data 731 is determined to have low similarity with the other projection data 732, 733, 734, 735, and 736. Therefore, the training data corresponding to the projection data 731 may be excluded from the final training data set. In one embodiment, the learning device may exclude the first projection data having a similarity less than a threshold value from the training data, wherein the similarity of each first projection data represents the similarity of each first projection data with the other first projection data.
[0169] The similarity of the first projection data may be determined based on at least one of a cosine similarity, a correlation coefficient, and a residual-based linear regression model evaluation method between each data included in the first projection data.
[0170] For example, refer to Figure 8 , the learning device according to the disclosed embodiment may determine the similarity of the first projection data based on the cosine similarity of the data 901, 902, 903, 904, and 905 located in the space represented by the first axis (X axis) and the second axis (Y axis) as described below. In one embodiment, θ may represent the angle between the data 901 and the remaining data 902 relative to the coordinate origin, but is not limited thereto. For example, θ may represent the angle between the data 901 and each of the other remaining data 903, 904, and 905 relative to the coordinate origin.
[0171] The cosine index of the data 901 is calculated as an average of the cosine distances between the data 901 and each of the remaining data 902 , 903 , 904 , and 905 .
[0172] The cosine index of the data 902 is calculated as an average of the cosine distances between the data 902 and each of the remaining data 901 , 903 , 904 , and 905 .
[0173] As in the above description, the cosine indexes of the residual data 903, 904, and 905 are calculated.
[0174] The standard deviation of the cosine indexes of all data 901, 902, 903, 904, and 905 is obtained. A value obtained by dividing the cosine index of each of the data 901, 902, 903, 904, and 905 by the standard deviation of all cosine indexes is determined as the distribution of each of the data 901, 902, 903, 904, and 905. Data having a distribution greater than or equal to a preset standard is determined as data having low similarity. Training data corresponding to data having a distribution greater than or equal to the preset standard may be excluded from the final training data set.
[0175] The similarity of data based on correlation coefficients and residual-based linear regression model estimation methods can also be determined similarly to cosine similarity.
[0176] As an optional embodiment, in operation S532, the learning apparatus according to the disclosed embodiment scales the first projection data 730 based on the second data.
[0177] Reference Figure 7F and Figure 7G , the learning device generates first projection scaled data 740 obtained by scaling the first projection data 730 based on the second data. The learning device may scale the projection distance of the first projection data 730 within the range of the second data. The projection distance is the distance from the reference point of the search vector SV to each of the first projection data 730. Scaling may be performed in various ways, such as minimum to maximum scaling.
[0178] One axis of the first projection scaled data 740 may be obtained by scaling the range of the axis on the search vector SV to the range of the second data. In one embodiment, the axis corresponding to the first projection data 730 (e.g., the axis on the search vector SV or the axis corresponding to the search vector SV) may be scaled based on the second data, or the axis corresponding to the first projection data 730 may be scaled within the range of the second data. Figure 7F , the projection distances of the remaining projection data 731, 733, 734, 735, and 736 are calculated similarly to the projection distance on the search vector SV of the projection data 732. The learning device scales the projection distances on the search vector SV of all the projection data 731, 732, 733, 734, 735, and 736 within the range of the second data, and each of the projection scaled data 741, 742, 743, 744, 745, and 746 may be located as follows Figure 7G In the space shown in with the scaling axis as one of the basic (reference) axes.
[0179] In operation S540, the learning device according to the disclosed embodiment refines the training data based on the similarity of the first projection scaling data 740 in which the second data is considered. In one embodiment, the learning device refines the training data based on the similarity of the first projection scaling data in a space having an axis based on the first projection scaling data 740 as a reference axis of the space and an axis ST based on the second data.
[0180] Detailed reference Figure 7G , the learning device may refine the training data based on the similarity of the projected scaled data 741 , 742 , 743 , 744 , 745 , and 746 scaled based on the second data.
[0181] The learning device may recognize that the other projected scaled data 741, 742, 743, 744, 745, and 745 have similar directivities except for the projected scaled data 746 among the projected scaled data 741, 742, 743, 744, 745, and 746 scaled based on the second data. The learning device may determine the final training data set by excluding the training data corresponding to the projected scaled data 746 having a preset distribution or greater distribution based on the cosine similarity between the projected scaled data 741, 742, 743, 744, 745, and 746 scaled based on the second data.
[0182] Different from the conventional method that only considers the distribution of the first data, Figures 6 to 8 The training data refinement method according to the disclosed embodiment described is based on the similarity of the first projection data in which the reference data of the first data is considered. The refinement of the training data can be performed in a space having a first axis based on the first projection data as a reference axis and a second axis ST based on the second data. The first axis based on the first projection data can be an axis scaled within the range of the second data.
[0183] Therefore, the learning device can exclude training data corresponding to first projection data having low similarity among a plurality of first projection data in a space having a first axis based on the first projection data and a second axis ST based on the second data as a reference axis. As a result, the similarity and consistency of the training data are improved, and the prediction accuracy of the learning model is improved, compared with refining the training data based only on the first data.
[0184] Will refer to Figure 7B and Figure 7C A method of determining the search vector SV is described.
[0185] The learning device sets an arbitrary vector SV_TEMP in a low-dimensional space where the first data 710 is reduced in dimension. The learning device rotates the vector SV_TEMP in all directions of the low-dimensional space. The learning device determines the projection data of the first data 720 projected for each vector SV_TEMP in each rotation direction. The learning device determines the vector SV_TEMP with the highest similarity among the vectors SV_TEMP in each direction as the search vector SV based on the similarity of the projection data in each direction where the second data is considered (for example, the cosine similarity of the projection data in each direction where the second data is considered can be added). In one embodiment, the learning device can set an arbitrary vector in the low-dimensional space, and generate projection data corresponding to each vector by projecting the second data 720 onto each of the arbitrary vectors (the manner of generating the projection data is substantially the same as the manner of generating the first projection data described above, and the projection data can be located in a space having an axis based on each vector as a reference axis and an axis ST based on the second data). The learning device can calculate the similarity of the projection data for each vector, and determine the vector with the highest corresponding similarity among the arbitrary vectors as the search vector SV. In one embodiment, the learning device may calculate the sum of similarities of the projection data corresponding to the arbitrary vector as the similarity corresponding to the arbitrary vector, wherein the similarity of each projection data corresponding to the arbitrary vector represents the similarity of each projection data with other projection data corresponding to the arbitrary vector.
[0186] Fig. 9 A method of refining training data based on user input of a learning device according to a disclosed embodiment is shown.
[0187] The learning device according to the disclosed embodiment visualizes a process of refining training data and receives a user input for the visualization. The learning device refines the training data based on the user input.
[0188] Reference Fig. 9 , the first projection scaling data is displayed on the display. The first projection scaling data is located in a space as a basic space, the basic space having a first axis (based on the scaled projection distance) as a reference axis of the basic space and a second axis (based on the second data as reference data).
[0189] The learning device may receive from the user a touch input 910 for specific training data 914, a setting input 920 for a graphic including the specific training data 914, and a dividing line input 930 between the specific training data 914 and the remaining training data 911, 912, and 913. The learning device excludes the training data 914 specified by the user's inputs 910, 920, and 930 from the final training data set.
[0190] Since the refinement method according to the disclosed embodiment is performed based on the similarity of the first projection data in which the reference data of the first data is considered, the training data with low similarity can be clearly distinguished. Since the refinement method according to the disclosed embodiment is performed based on the similarity of the first projection data in the space having the first axis (based on the first data) and the second axis ST (based on the second data) as the reference axis, the training data with low similarity can be clearly distinguished. Therefore, the user can intuitively recognize the similarity of the training data.
[0191] Fig. 10A and Fig. 10B The prediction accuracy of the recess depth RCSHT of the structure of the wafer according to the learning model trained by the refined training data according to the disclosed embodiment is shown. The combination of "#13", "#14", "#15", "#16", "S1", "S6" and "E12" represents the recognition number of the training data.
[0192] Fig. 10A It is the experimental result of the learning model trained with 12 unrefined training data. Fig. 10B It is an experimental result of a learning model trained using 10 pieces of refined training data according to a disclosed embodiment.
[0193] Reference Fig. 10A and Fig. 10B , it can be seen that, compared with the unrefined case, the predicted value of the depression depth RCSHT of the learning model trained by the refined training data according to the disclosed embodiment is very similar to the value of the depression depth actually confirmed in the TEM image.
[0194] Reference Fig. 10A and Fig. 10B , it can be seen that the RMSE value (0.456), which is the error value associated with the prediction of the depression depth by the learning model trained by the refined training data according to the disclosed embodiment, is much lower than the RMSE value (1.647) of the non-refined case.
[0195] Reference Fig. 10A and Fig. 10B , it can be seen that the determination coefficient (R ) associated with the prediction of the depression depth of the learning model trained by the refined training data according to the disclosed embodiment is 2 ) is closer to "1" (0.9895) than the value (0.8816) of the case without refinement.
[0196] According to the disclosed embodiments, a learning model for predicting a structure of a semiconductor device and an apparatus for predicting a structure of a semiconductor device can improve the prediction accuracy of the structure of a semiconductor device based on metrology data.
[0197] According to the disclosed embodiments, a learning model for predicting the structure of a semiconductor device and an apparatus for predicting the structure of a semiconductor device can improve the consistency of training data by refining training data that is not suitable for training in the process of training a model for predicting the structure of a semiconductor device.
[0198] At the same time, the above description is for carrying out a specific embodiment of the disclosure. Embodiments in which the design is simply changed or easily changed may be included in the disclosure and in the above-described embodiments. In addition, techniques that are easily changed and implemented using the above embodiments may be included in the disclosure. Although the disclosure has been described with reference to the disclosed embodiments, it will be clear to those of ordinary skill in the art that various changes and modifications may be made thereto without departing from the spirit and scope of the disclosure as set forth in the appended claims.
Claims
1. An apparatus for predicting the structure of a semiconductor device, the apparatus include: a storage device configured to store a learning model, the learning model configured to predict a structure of a semiconductor device; memory, configured to store code, and at least one processor operably connected to the memory and configured to execute the code to: input non-destructive metrology data measured from the semiconductor device into the learning model, and predict the structure of the semiconductor device based on the learning model, wherein the learning model is trained using training data, the training data including first data as non-destructive metrology data and second data as structural metrology data, the second data being reference data for the first data, and The training data is refined based on similarity of first projection data in a space having a first axis corresponding to the first data and a second axis corresponding to the second data as a reference axis, wherein the first projection data is obtained based on the first data and the second data.
2. The apparatus for predicting the structure of a semiconductor device according to claim 1, in, The first data is metrology data of the semiconductor device based on a spectrum, and the second data is structure data obtained by measuring at least a part of the structure of the semiconductor device.
3. A method for generating a learning model, the learning model being configured to predict the structure of a semiconductor device, the method include: identifying training data, the training data comprising: first data as non-destructive metrology data of the semiconductor device and second data as structural data, the second data being reference data for the first data; refining the training data based on similarity of first projection data in a first space having a first axis corresponding to the first data and a second axis corresponding to the second data as a first reference axis, wherein the first projection data is obtained based on the first data and the second data; and Train the learning model based on the refined training data.
4. The method according to claim 3, in, The first axis is scaled within the range of the second axis.
5. The method according to claim 3, in, The first data is spectrum-based metrology data of the semiconductor device, and the second data is obtained by measuring at least a portion of a structure of the semiconductor device.
6. The method according to claim 3, in, The steps to refine the training data include: converting the first data into third data in a low-dimensional space in which the first data is reduced in dimension; determining a search vector in a low-dimensional space in which the first data is reduced in dimension; generating first projection data by projecting the third data onto the search vector; First projection data having a similarity less than a threshold value is excluded from the training data, wherein the similarity of each first projection data represents the similarity of each first projection data to other first projection data.
7. The method according to claim 6, in, The steps to determine the search vector include: Set arbitrary vectors in low-dimensional space; generating projection data corresponding to the arbitrary vector by projecting the third data onto the arbitrary vector; Calculating a sum of similarities of the projection data corresponding to the arbitrary vector as the similarity corresponding to the arbitrary vector, wherein the similarity of each projection data corresponding to the arbitrary vector represents the similarity of each projection data to other projection data corresponding to the arbitrary vector; A vector having the highest corresponding similarity among the arbitrary vectors is determined as a search vector.
8. The method according to claim 6, in, A third axis corresponding to the search vector is scaled based on the second data.
9. The method according to claim 6, in, The search vector is a vector in a low-dimensional space in which the first data is reduced in dimension based on at least one of principal component analysis, singular vector decomposition, non-negative matrix decomposition, and partial least squares.
10. The method according to claim 6, in, The step of refining the training data further includes: performing visualization by displaying the first projection data in the first space.
11. The method according to claim 10, in, The steps to refine the training data include: receiving user input regarding the visualization result; and The training data is refined based on user input.
12. The method according to claim 3, in, The similarity is based on at least one of a cosine similarity of the first projection data in the first space, a correlation coefficient, and a residual-based linear regression model evaluation method.
13. A device for generating a learning model, the learning model being configured to predict the structure of a semiconductor device, the device include: a memory configured to store code; as well as at least one processor operably connected to the memory and configured to execute the code to: identifying training data, the training data comprising: first data being non-destructive metrology data measured from a semiconductor device and second data being structural data, the second data being reference data for the first data, refining the training data based on similarity of the first projection data in a first space having a first axis corresponding to the first projection data and a second axis corresponding to the second data as a first reference axis, wherein the first projection data is obtained based on the first data and the second data, and Train the learning model using the refined training data.
14. The device according to claim 13, in, The first axis is scaled within the range of the second axis.
15. The device according to claim 13, in, The at least one processor is configured to execute the code to: converting the first data into third data in a low-dimensional space in which the first data is reduced in dimension; determining a search vector in a low-dimensional space in which the first data is reduced in dimension; generating first projection data by projecting the third data onto the search vector; First projection data having a similarity less than a threshold value is excluded from the training data, wherein the similarity of each first projection data represents the similarity of each first projection data to other first projection data.
16. The device according to claim 15, in, The at least one processor is configured to execute the code to: Set arbitrary vectors in low-dimensional space; generating projection data corresponding to the arbitrary vector by projecting the third data onto the arbitrary vector; Calculating a sum of similarities of the projection data corresponding to the arbitrary vector as the similarity corresponding to the arbitrary vector, wherein the similarity of each projection data corresponding to the arbitrary vector represents the similarity of each projection data to other projection data corresponding to the arbitrary vector; A vector having the highest corresponding similarity among the arbitrary vectors is determined as a search vector.
17. The device according to claim 15, in, A third axis corresponding to the search vector is scaled based on the second data.
18. The device according to claim 15, in, The at least one processor is configured to execute the code to: The visualization is performed by displaying the first projection data in the first space.
19. The device of claim 18, wherein the at least one processor is configured to execute the code to: receiving user input regarding the visualization results, and The training data is refined based on user input.
20. The apparatus according to claim 13, in, The similarity is based on at least one of a cosine similarity of the first projection data in the first space, a correlation coefficient, and a residual-based linear regression model evaluation method.
Citation Information
Patent Citations
System and method for providing artificial intelligence-based positive and negative numerical legal answer services
KR1020230072214A
