Optical metrology in machine learning for characterization

Abstract

The metrology system may include an optical metrology tool configured to generate optical metrology output for one or more features on a processed substrate, and a metrology machine learning model trained using the following training set: (i) topography, critical dimensions, and profiles of a plurality of features, and (ii) the optical metrology outputs of the plurality of features. The metrology machine learning model may be configured to: receive the optical metrology output from the optical metrology tool; and output the topography, critical dimensions, and / or profiles of the one or more features on the processed substrate.

Description

[0001] Incorporated by Reference

[0002] The PCT application form is filed simultaneously with this specification and is made a part of this application. Each application in the PCT application form filed simultaneously with this application that this application claims the benefit or priority of is incorporated herein by reference in its entirety. Background Art

[0003] The performance of semiconductor device manufacturing operations such as plasma-assisted etching processes is often critical to the success of semiconductor device processing workflows. However, optimization or adjustment of manufacturing processes and / or tools associated therewith (e.g., etching reactors, photolithography masks, etc.) can prove to be technically difficult and time-consuming, often involving technicians manually adjusting etching process parameters or tool component designs to produce the desired target feature profile. Metrology tools should accurately assess the performance of etching or deposition processes by measuring features on fabricated or partially fabricated devices. Currently, a variety of metrology tools are used for this purpose.

[0004] The background and context descriptions contained herein are provided only for the purpose of generally presenting the context of the present disclosure. Much of this disclosure presents the work results of the inventors, and merely because such work background is partially described or presented as background elsewhere herein does not constitute an admission that it is prior art. Summary of the invention

[0005] Certain aspects of the present disclosure relate to methods for determining the shape, critical dimension, and / or profile of one or more features on a processed substrate. Such methods may be characterized by the following features: (a) performing optical metrology on one or more features on a processed substrate to produce an optical metrology output; (b) providing the optical metrology output to a metrology machine learning model that has been trained using the following training set: (i) the shape, critical dimension, and / or profile of multiple features, and (ii) the optical metrology output of the multiple features; and (c) receiving the shape, critical dimension, and / or profile of the one or more features on the processed substrate from the metrology machine learning model. In some embodiments, the method further includes training the metrology machine learning model by generating multiple optical metrology output sets, each optical metrology output set generated for a different orientation and / or position of a test feature relative to an optical metrology device for use with the test feature. In some embodiments, the metrology machine learning model is generated using supervised machine learning techniques.

[0006] In some embodiments, the optical metrology is a scatterometry technique. For example, the optical metrology output may include reflectance spectra and / or ellipsometry output data.

[0007] In some embodiments, the method further comprises performing pattern recognition on one or more features on the processed substrate; and determining that the pattern of the one or more features matches the expected pattern. In some embodiments, the method further comprises performing optical metrology on the one or more features on the processed substrate, including raster scanning on the processed substrate.

[0008] In some cases, electron microscopy and / or CD-SAXS are used to obtain the topography, critical dimensions, and / or contours of multiple features in the training set.

[0009] In some embodiments, the method further comprises decomposing the optical metrology output to produce a reduced set of optical metrology values, and providing the reduced set of optical metrology values ​​to a metrology machine learning model. In some cases, decomposing the optical metrology output comprises identifying principal components of the optical metrology output or applying the optical metrology output to an autoencoder.

[0010] Other aspects of the present disclosure relate to a metrology system that can be characterized by the following elements: (a) an optical metrology tool comprising an optical probe source and an optical detector and a processor configured to generate an optical metrology output based on data generated by the optical detector when the optical probe is directed to one or more features on a processed substrate; (b) a metrology machine learning model that has been trained using the following training set: (i) the shape, critical dimension and / or profile of multiple features, and (ii) the optical metrology output of multiple features. In some embodiments, the metrology machine learning model is configured to: receive the optical metrology output from the optical metrology tool; and output the shape, critical dimension and / or profile of one or more features on the processed substrate.

[0011] In some systems, a metrology machine learning model is trained using multiple optical metrology output sets, each optical metrology output set generated for a different orientation and / or position of a test feature relative to a test optical metrology device. In some embodiments, the metrology machine learning model is generated using supervised machine learning techniques.

[0012] In some embodiments, the optical metrology tool is a scatterometer.In some metrology systems, the optical metrology output includes reflectance spectra and / or ellipsometric output data.

[0013] Some metrology systems also include a pattern recognition tool that includes a camera and a processor configured to: perform pattern recognition on one or more features on the processed substrate; and determine that the pattern of the one or more features matches an expected pattern. In some metrology systems, the optical metrology tool is configured to raster scan the processed substrate.

[0014] In some metrology systems, the topography, critical dimensions, and / or contours of multiple features in a training set are obtained using electron microscopy and / or using CD-SAXS.

[0015] In some embodiments, the metrology machine learning model is further configured to: decompose the optical metrology output to produce a reduced set of optical metrology values, and provide the reduced set of optical metrology values ​​to the metrology machine learning model. In some such embodiments, the metrology machine learning model is configured to decompose the optical metrology output by identifying principal components of the optical metrology output or by applying the optical metrology output to an autoencoder.

[0016] These and other aspects are further described below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A flowchart for generating a metrological machine learning model according to certain method embodiments of the present disclosure is shown.

[0018] Figure 2 A flowchart for using a metrological machine learning model according to certain method embodiments of the present disclosure is shown.

[0019] Figure 3 An example computing system that can be used to train and / or run an optical metrology machine learning model is shown.

[0020] Figure 4 A block diagram of a metrology system is shown that includes a model configured to generate a geometric representation of substrate features from optical metrology data, such as would be obtained using an electron microscope. DETAILED DESCRIPTION

[0021] Current methods for evaluating semiconductor wafers and other samples used in semiconductor process development employ techniques such as x-SEM (x-ray scanning electron microscope), STEM (scanning transmission electron microscope) and CD-SAXS (critical dimension small angle x-ray scattering) for cross-sectional image capture. Current methods then use image analysis software to extract the features of interest. Coupons are often used in place of full wafers to study the effects of etching or other manufacturing processes on patterned substrates. Coupons are test samples or substrates that typically have the same or similar materials and layers to those used in semiconductor device manufacturing operations. Typically, coupons are smaller than semiconductor wafers. They can be circular or polygonal (e.g., rectangular) and occupy only a few square centimeters.

[0022] Current electron microscopy-based methods are sometimes particularly subject to the following challenges.

[0023] 1. x-SEM and STEM are expensive, and CD-SAXS is slow;

[0024] 2. The error bar of each x-SEM / STEM image is high due to the roughness of the line edges; to reduce this error, more cross sections are needed, which makes the cost higher.

[0025] 3. Features are extracted manually and are prone to operator error and / or differences between operators.

[0026] 4. STEM and x-SEM are destructive, so all previous manufacturing operations need to be repeated to accurately consider each etching step so that the complete etching effect can be fully understood. However, there is no guarantee that multiple samples taken at different steps are the same. Small differences in these samples, whether from feed conditions or etching conditions, can cause large errors in process development.

[0027] Disclosed herein are methods and systems for supervised machine learning that use spectra collected from specimen samples to generate models that can easily and accurately predict feature representations typically obtained by destructive metrology techniques such as electron microscopy (e.g., x-SEM and STEM) or by other time-consuming processes (e.g., CD-SAXS). During training, many samples can be processed according to a defined process (e.g., etching or deposition) and then laterally sectioned to extract topographic information useful for process development and / or process monitoring. A model trained using this information can receive an optical signal from an optical metrology tool (which tends to operate quickly, cheaply, and non-destructively) and can output a ground truth of a feature established by another metrology tool (which tends to operate slowly, expensively, and often destructively).

[0028] In some cases, the shape information may suffer from inaccuracies due to the extraction of features from fuzzy outlines in micrographs or other images. This method can achieve reliable supervised training if a sufficiently large training set is used (e.g., at least about 20 samples, or at least about 100 samples), and electron microscopy (and / or CD-SAXS) and optical metrology (e.g., performing x-SEM and STEM and measuring spectra) are used for the features of interest.

[0029] In some cases, at least some samples are evaluated at locations where a single type of feature occupies at least a significant portion of the spot produced by the incident light on the sample. Typically, the features of interest are of a few hundred nanometers or less in size, while the spot is about 5-40 microns. In order for a single feature to contribute significantly to the optical signal obtained from the metrology tool, it may be necessary to repeat the feature over the area occupied by the incident spot when the spot size is significantly larger than the feature size. However, this is not a strict requirement for the samples / positions used in the training set. Many machine learning techniques are robust enough that at least some of the training set members (or at least their optical metrology signals) can have multiple types of features in the optical signal. Regardless of the composition of the training set, supervised training can be achieved using a pairing of spectra and CD, shape and / or profile information extracted using a reliable profiling technique (e.g., x-SEM, STEM, CD-SAXS or other destructive metrology techniques or slow metrology techniques).

[0030] Note that training is supervised in the sense that a destructive metrology technique (e.g., x-SEM / STEM) or other slow profile technique provides a known feature representation (e.g., a feature's shape, profile, or critical dimension (CD)) for identifying an optical signal feature that corresponds to the known feature representation. When provided with optical data collected under the conditions used to train the model, the resulting model can accurately predict feature shape information. In the methods and systems described herein, there may be two implementation steps - training and testing. Ultimately, the generated and validated metrology machine learning model can be employed in an optical metrology system to directly provide information about feature shape, CD, and / or profile.

[0031] train

[0032] In some embodiments, training uses a scattering method to capture reflectance spectra using a range of optical information, which can span a range of incidence angles (normal incidence and oblique incidence), including a full 360 degrees of azimuthal direction, wavelengths (e.g., including visible light, infrared and / or ultraviolet light), and polarizations on features of interest. After collecting this optical metrology information, the sample is analyzed by x-SEM, TEM, CD-SAXS, etc. An accumulation of several samples (e.g., greater than about 20 samples or greater than about 100 samples) can be used. Each such sample can generate one or more training set members. For example, a single sample can be analyzed using multiple optical metrology settings (e.g., angles of incidence), each setting producing a separate optical metrology reading and an associated training set member.

[0033] Positioning and / or orienting the specimen in an error-free manner can be a challenge. While metrology tools are designed to fine-tune the position and orientation of wafers having standard sizes, thereby providing relatively low error capture of optical metrology results, these tools are not designed to so carefully control the position and orientation of the specimen. The specimen often has a size and shape that is significantly different from the size and shape of the production wafer. Therefore, measures may need to be taken to account for possible errors in the optical metrology signal generated by the specimen.

[0034] In some embodiments, by translating, rotating and / or otherwise redirecting the sample in different ways under the beam of the optical metrology tool and capturing light signals at all these different positions or orientations, position and / or rotation errors are reduced. For the spot of the beam of the optical metrology tool, the same position can exist for each different position / orientation. In general, these different positions average the errors that may occur in a small part of different translation positions and / or orientations. In order to train, each different optical signal captured at these different positions or orientations is associated with the same x-SEM, STEM, CD-SAXS or other direct measurements. Therefore, a given position on the sample (specimen) may generate a large number of training set members, each member using the same x-SEM, STEM or CD-SAXS measurement value, but using different optical signals obtained at different translation positions or orientations. In general, the number of these training set members is large enough to cover the actual errors of typical experimental conditions.

[0035] Obviously, each sample subjected to optical metrology is then sent to an x-SEM apparatus, TEM apparatus, CDSAXS apparatus or other apparatus to measure the features produced by the process in question.Destructive processes provide information for extracting CD, shape, profile or other feature characterization.

[0036] As mentioned above, the Z-resolved information can be obtained by non-destructive (or less destructive) techniques such as CD-SAXS (critical dimension small angle X-ray scattering). CD-SAXS is an X-ray scattering technique that can be used to reconstruct the in-plane and out-of-plane structure of nanostructured films. The technique involves collecting a series of transmission small angle X-ray scattering images at various sample rotation angles. These (q x ,q y ) images can be combined to reconstruct the 3D reciprocal space (e.g., Fourier transform), in particular to detect (q) containing both in-plane and out-of-plane (membrane normal direction) information x ,q z )slice.

[0037] Training can start with, for example, about 100 pairs of spectra and CD values, profiles and / or contours. Different supervised learning techniques (regression, decision trees, support vector machines, neural networks, etc.) can be used to map the relationship between the spectra and the CD / profile / contour features. Hyperparameters can be optimized to ensure the best validation results before testing. If other information related to sample-to-sample variations is available, such as layer thickness and optical constants, it can also be used as input for supervised learning.

[0038] In one embodiment, the supervised training method utilizes dimensionality reduction techniques / tools, such as principal component analysis (PCA) or autoencoders. Dimensionality reduction techniques are optionally used in conjunction with generating a neural network or other supervised machine learning model. For example, PCA can be used to identify optical signal characteristics or combinations of optical signal characteristics based on optical metrology signals that are relatively strongly affected by target features such as feature critical dimensions, compared to other optical signal characterizations. The resulting principal components (optical signal feature vectors that are strongly affected by the features of interest) are used to train the model and are used as input to the resulting model. More generally, any technique that decomposes or identifies aspects of the optical signal used to train the model (and as input to a subsequently generated model) can be used. Any such technique can receive n spectral intensity values ​​(from optical metrology) that are generated by m substrate feature variation sources and identify approximately m unique spectral features. The value of n will typically be much larger than the value of m, for example, but by at least one order of magnitude.

[0039] In another embodiment, a deep learning neural network is used in a manner that does not require a principal component extraction step. As with the PCA neural network implementation, once trained to produce a deep learning model with sufficient accuracy to predict CD, shape, and / or contour information, it can bypass x-SEM, STEM, CD-SAXS, etc. and directly report the above information (e.g., feature shape or feature geometry).

[0040] Taking advantage of the averaging capabilities of optical techniques over large targets, this method provides average CD / profile / contour information more efficiently than multiple x-SEM, STEM, CD-SAXS, etc. measurements. Periodic structures are most useful due to their translational invariance, but the method is not limited to periodic structures.

[0041] In some embodiments, training can be supplemented with pattern data representing expected patterns at locations on the sample where optical metrology acquires signals and SEM cross-sections are taken. Standard pattern recognition software can be used to take camera images of such locations and determine whether the location has the expected pattern. If the location does not have the expected pattern, optical and microscopy data may not be acquired, or if acquired, it can be eliminated or given a low weight in model training.

[0042] The non-invasive metrological data used to train the model can come from a variety of conventional and non-conventional optical techniques. Such techniques can use, for example, scatterometry and / or ellipsometry to capture reflectivity and / or ellipsometry spectra.

[0043] In some embodiments, training can be supplemented with optical signals obtained by raster scanning a beam from a metrology tool over a specimen or other sample. Although the resulting data is collected over a range of locations on the sample, it can be associated with a characteristic shape (or other geometric feature) at one or more specific locations on the sample.

[0044] Example workflow for training

[0045] Figure 1 An exemplary process flow 101 for generating a model that can be deployed in an optical metrology tool or optical metrology system is shown. As shown, the process begins at operation 103, where a set of substrates (e.g., coupons) are exposed to a variety of different process conditions to produce a variety of different features on the substrates. The goal here is to provide a variety of feature shapes and / or shapes within the prediction range of the model. For example, a first substrate is exposed to a first set of process conditions (plasma power, plasma frequency and other plasma conditions, process gas composition, reactor temperature and pressure, mask geometry, etc.), a second substrate is exposed to a second set of process conditions that are different from the first set of process conditions, a third substrate is exposed to a third set of process conditions that are different from the other two sets of conditions, and so on. At the end of the operation, the plurality of substrates have a plurality of different feature shapes and optionally a plurality of different profiles, critical dimensions, etc. In an alternative approach, the same process conditions (e.g., recorded process conditions) are applied to each substrate, but the inherent variability of the process produces slightly different feature structures on the substrates (as reflected in the shape, profile, and / or CD). These inherent differences can be used as the basis for a training set.

[0046] Next, in operation 105, optical metrology is performed on the features of the test substrate generated in operation 103. In each case, the resulting optical signals can be associated with the substrate or features that generated those signals, although the topography or other geometric details of the features are generally not known. It should be understood that the optical signals may not directly indicate the topography, critical dimensions, contours, or other geometric characteristics of the features that generated those signals.

[0047] Optionally, at process operation 107, the process repeats optical metrology of the feature under different conditions (e.g., different substrate orientations, different probe beam spot sizes, etc.). As explained, multiple forms of optical metrology may be employed, and multiple modes of performing a single metrology may also be employed. Such modes include specific settings of the optical metrology tool, such as the azimuthal position and angle of incidence of the beam, the spot size, the spot location on the substrate, raster scanning, etc. In addition, a pattern recognition pre-processing step may be used to remove certain measurements that do not appear to conform to the expected feature pattern.

[0048] Next, in operation 109, the method performs a form of metrology of the topography or other geometric features of the features generated directly on the test substrate. As noted, such metrology is typically slow and / or destructive. Examples include various forms of electron microscopy, including various forms of TEM (e.g., STEM) and SEM (e.g., x-SEM and CD-SEM) and CD-SAXS. In operation 105 and optionally in operation 107, the resulting geometric representations of the directly measured features are correlated with the corresponding optical signals generated by the features.

[0049] At this point, the complete training set information is available, and as shown in operation 111, the process trains a metrology model using the optical metrology signals generated in operation 105 and (optionally) operation 107 and the associated feature profiles, contours, CD, etc. generated in operation 109. The resulting model is tested and / or validated, as shown in operation 113. The resulting model can be installed in a metrology tool or system.

[0050] The following sequence presents another example of a process flow.

[0051] 1. Provide a specimen having features produced using a process under consideration (e.g., a recording process) and identify one or more locations where the features of interest are located.

[0052] (Optional) Perform pattern recognition at this location.

[0053] 2. From the sample provided in 1, obtain optical metrology data at the location of the feature of interest.

[0054] 3. From the specimen provided in 1, obtain the topography or other geometric characterization of the features at the locations used in 2 using x-SEM, STEM or other destructive or slow techniques that provide direct representation of the geometric characterization (e.g., CD-SAXS).

[0055] 4. Use the data from 2 and 3 to train a model that takes as input the optical metrology data produced in 2 and predicts the geometric representation produced using the process under consideration.

[0056] test

[0057] In testing, inference is performed using the same optical metrology technique (e.g., scatterometry) as was used to capture the raw data during training, which is then fed into a previously built supervised learning model to predict the CD / profile features of interest. If training was performed using metrology performed at multiple translational positions and / or orientations, some or all of these positions / orientations may be employed to capture data for input into the model. Furthermore, if training was performed using pattern recognition via a camera, the same pattern recognition may be employed to filter the data provided to the model. Still further, if training was performed using raster scanning of an optical metrology probe beam, the same raster scanning may be performed during testing and the resulting scan data provided to the model. Of course, the test data should be generated under conditions different from those of the training data. Thus, typically, the test data is generated on substrate features using some manufacturing process different from that used to generate the training data. For example, the test data may be generated using a combination of etching conditions different from any combination used to generate any training data.

[0058] application

[0059] The metrology machine learning models generated as described herein have a variety of applications. The scope of application of any given model depends at least in part on the scope of applicability of the model; some models are more general in the number of processes that can effectively predict feature characterization than other models. For example, a model can be generated for a specific process. For example, an anisotropic etch process that employs a specific chemistry to produce a pattern of a specific design layout.

[0060] After effectively training the model, this approach can significantly reduce the costs associated with performing destructive tests such as STEM and x-SEM. This is useful if the same type of samples are to be used repeatedly (e.g. >100 specimens) to adjust the processing.

[0061] This approach also leads to better sample-to-sample consistency because random errors in CD / profile / contour and LER / LWR (line edge roughness and line width roughness) are spatially averaged by the optical signal. Machine learning is used here to map the complex relationship between the spectrum and the CD / profile / contour of the feature of interest. When errors are well controlled, this workflow may be operator-free. The non-destructive nature of the measurement also allows for multi-etch step consistency of sample results. Each etch step for a single sample can produce multiple inference results and multiple features, maximizing its value.

[0062] The trained metrology model can be used as part of a process or related system for determining the shape, CD, profile, or other geometric characteristics of one or more features on a substrate undergoing a device manufacturing process. The model and related metrology system can be used during process development, process monitoring, and / or process modification. Depending on the application, the substrate undergoing metrology and analysis can be a test coupon or wafer, a production wafer with partially fabricated devices, a production wafer with fully fabricated devices, etc.

[0063] exist Figure 2 An exemplary process flow for using a metrology machine learning model is depicted in FIG. As shown, process 201 begins by processing a substrate under process conditions of a trained metrology machine learning model. See operation 203.

[0064] After processing the substrate to create or modify features on the substrate, optical metrology is performed on one or more of the features. See operation 205. The optical metrology techniques are the same as those used in the training set of the model. In some embodiments, operation 205 is performed by moving a beam or point over multiple locations on the substrate. For example, the beam can be raster scanned over the processed substrate.

[0065] For normal operation, the optical metrology technique generates an optical signal having a characterization determined by features on the substrate. The optical signal may be, for example, a reflectivity value provided as a function of wavelength, polarization state, azimuthal position, etc. Regardless of its information content, the resulting optical signal is provided to a metrology machine learning model, which may be trained as described above. See operation 207. Using the optical signal, the machine learning model outputs information identifying one or more geometric features of the one or more features. See operation 209.

[0066] In some embodiments, operation 205 is performed only after first performing pattern recognition on one or more features on the processed substrate. The pattern recognition process can determine that the pattern of the one or more features matches the expected pattern. If the pattern does not match, the optical metrology operation can be omitted.

[0067] the term

[0068] The terms "semiconductor wafer", "wafer", "substrate", "wafer substrate" and "partially fabricated integrated circuit" may be used interchangeably. It is understood by those of ordinary skill in the art that the term "partially fabricated integrated circuit" may refer to a semiconductor wafer during any of the many stages of integrated circuit fabrication performed thereon. The diameter of the wafers or substrates used in the semiconductor device industry is typically 200 mm or 300 mm or 450 mm. This detailed description assumes that the present embodiment is implemented on a wafer. However, the present disclosure is not limited to this. The workpiece may have a variety of shapes, sizes, and materials. In addition to semiconductor wafers, other workpieces that may utilize embodiments of the present disclosure include, for example, various items such as printed circuit boards, magnetic recording media, magnetic recording sensors, mirrors, optical elements, micromechanical devices, and the like. In addition, the term "specimen" is sometimes used herein generally to describe any substrate, typically including those substrates used for test processing and / or for generating model training set data.

[0069] As used herein, a "semiconductor device manufacturing operation" is a unit operation performed during the manufacture of a semiconductor device. Typically, the entire manufacturing process includes multiple semiconductor device manufacturing operations, each of which is performed in its own semiconductor manufacturing tool, such as an etching and / or deposition plasma reactor, an electroplating cell, a chemical mechanical planarization tool, a wet etching tool, etc. Categories of semiconductor device manufacturing operations include subtractive processes such as etching processes and planarization processes, and material addition processes such as deposition processes. In the context of etching processes, substrate etching processes include processes for etching mask layers, or more generally, processes for etching any material layer previously deposited and / or otherwise residing on the surface of a substrate. Such etching processes can etch stacks of layers in a substrate. The model described herein can be used to replace destructive metrology techniques with non-destructive optical metrology techniques for any type of semiconductor device manufacturing operation.

[0070] As used herein, "metrology results" refers to results generated at least in part by measuring features of a processed substrate. Measurements can be performed while or after semiconductor device manufacturing operations are performed in a reaction chamber operating with a set of process parameter values. In some embodiments, measuring features of a processed substrate generates profile coordinates. In such embodiments, measuring features of a processed substrate may include performing microscopic examination (e.g., SEM, x-SEM, TEM, STEM, CD-SEM, CD-SAXS, REM, AFM). Such techniques may be used, for example, to directly or indirectly generate profile coordinates by characterizing a set of geometric profile parameters of the geometry of features in a processed substrate (e.g., critical dimensions, sidewall angles, depth, etc.).

[0071] Optical metrology results are generated on features of a processed substrate using optical metrology. In some embodiments, optical metrology results are generated by performing reflectometry, dome scattering, angle-resolved scattering, small-angle X-ray scattering, and / or ellipsometry on a processed substrate. Optical metrology outputs can include any of a variety of values ​​obtained from a detector in an optical metrology tool. Examples include reflectivity values, optionally over a range of wavelengths (reflection spectra), polarization information, angular / positional information about detected light signals (e.g., angles relative to the plane of the substrate or detector or angles relative to the angle of incidence). When using optical metrology, the system can obtain shape coordinates by calculating shape coordinates from measured optical metrology signals using the metrology machine learning model described herein.

[0072] In some embodiments, the metrology results are provided as a time series of measured geometrical profiles, reflectivity or ellipsometric data or profile parameters of substrate features. These measured metrology results are generated at different durations of semiconductor device manufacturing operations.

[0073] Geometric information about a feature can take a variety of forms. It can be provided as two-dimensional or three-dimensional shape coordinates and / or more global parameters (e.g., trench depth, sidewall angle, etc.). For repeating structures, one-dimensional information can be provided, such as spacing, critical dimensions, etc. In the case of a two-dimensional representation, the geometric information of a feature can be an outline (viewed as a cross-section of the substrate in the xz plane) or a profile (viewed from above the plane of the substrate in the xy plane).

[0074] Machine learning model-A machine learning model is a trained computational model that can acquire optical metrology output or data (e.g., scatterometry data) and predict the geometric characteristics of one or more features on a sample. Examples of machine learning models include random forest models (including deep random forests), neural networks (including recursive neural networks and convolutional neural networks), restricted Boltzmann machines, recursive tensor networks, and gradient boosting trees. Each of these models has a structure and an associated sequence of logical operations (algorithms) known to those skilled in the art. For example, a neural network includes a plurality of computational nodes organized in layers, which are configured to continuously pass computational results from one layer to the next layer on a node-by-node basis. A given node may have a connection to one or more nodes in an upstream layer and a connection to one or more patterns in a downstream layer. These layers may be organized into an input layer, an output layer, and one or more hidden layers. The input side of the neural network may be constructed in a manner that implements a convolutional neural network. Nodes in a hidden layer or output layer receive input from one or other nodes in an upstream layer. Each node operates on its input through a function (sometimes an activation function) to generate an output to be transmitted to one or more downstream nodes in the next consecutive layer. Training applies weights to each connection between upstream and downstream nodes. The term "classifier" (or classification model) is sometimes used to describe all forms of classification models, including deep learning models (e.g., neural networks with many layers) and random forest models. For more details, see "Hands-On Machine Learning with Scikit-Learn and TensorFlow: Concepts, Tools, and Techniques to Build Intelligent Systems,"1 st Edition, by A. Geron, O'Reilly Media, Inc. (2017); and Deep Learning by Goodfellow et al., MIT Press (2016), the entire contents of which are incorporated herein by reference.

[0075] Deep Learning Model - The deep learning model used in this article is a form of a classification model. It is also a form of a machine learning model. It can be implemented in various forms, such as by a neural network (e.g., a convolutional neural network), etc. Typically, although not necessarily, it includes multiple layers. Each such layer includes multiple processing nodes, and the layers are processed in sequence, where the nodes of the layers closer to the model input layer are processed before the nodes of the layers closer to the model output. In various embodiments, one layer feeds the next layer, and so on. The output layer may include nodes representing various classifications. In some embodiments, a deep learning model is a model that requires only little preprocessing to acquire data.

[0076] In various embodiments, the deep learning model has significant depth (e.g., five or more layers or ten or more layers) and can classify large or heterogeneous collections of optical metrology data. In some cases, the term "deep" means that the model has more than two (or more than three, more than four, or more than five) layers of processing nodes that receive values ​​from previous layers (or as direct inputs) and output values ​​to subsequent layers (or final outputs). Internal nodes are typically "hidden," meaning that their input and output values ​​are not visible outside the model. In various embodiments, the operation of hidden nodes is not monitored or recorded during operation.

[0077] Deep learning models can be trained and retrained without redesigning the number, arrangement, interface with image input, etc. of their nodes and connections.

[0078] As indicated, in various embodiments, the node layers may collectively form a neural network, although many deep learning models have other structures and formats. Some embodiments of deep learning models do not have a hierarchical structure, in which case the above characterization of "depth" having multiple layers is not relevant.

[0079] Context of the Disclosed Computing Implementation

[0080] Some embodiments disclosed herein relate to systems for generating and / or using metrology models disclosed herein. Some embodiments disclosed herein relate to methods for generating and / or using models implemented on such systems. The system for generating the model can be configured to analyze data to calibrate or optimize expressions or relationships for representing the effects of semiconductor device manufacturing operations on a substrate. A programming system for using the model can be configured to (i) receive inputs such as optical metrology data, and (ii) execute instructions for determining the characterization of features on a substrate. A programming system for using the model can be configured to (i) receive optical metrology data obtained for multiple features, (ii) receive SEM, TEM and / or CD-SAXS data for the same multiple features, and (iii) execute instructions for training a metrology machine learning model using the data in (i) and (ii).

[0081] Many types of computing systems having any of a variety of computer architectures can be used as the disclosed system for implementing models, and algorithms for generating, using, and / or optimizing such models. For example, the system may include software components that execute on one or more general-purpose processors or specially designed processors, such as programmable logic devices (e.g., field programmable gate arrays (FPGAs)). In addition, the system can be implemented on a single device or distributed across multiple devices. The functions of the computing elements can be merged with each other or further divided into multiple sub-modules.

[0082] In some embodiments, the code executed during the generation or execution of a model on a suitably programmed system may be embodied in the form of a software element that may be stored on a non-volatile storage medium (e.g., a CD, a flash memory device, a removable hard drive, etc.) and includes a plurality of instructions for manufacturing a computer device (e.g., a personal computer, a server, a network device, etc.).

[0083] At one level, software elements are implemented as command sets prepared by the programmer / developer. However, modular software that can be executed by computer hardware is executable code submitted to memory using "machine code" selected from a specific machine language instruction set or "native instructions" designed into the hardware processor. The machine language instruction set or native instruction set is known to the hardware processor and is essentially built into the hardware processor. This is the "language" in which system and application software communicate with the hardware processor. Each native instruction is a discrete code that is recognizable by the processing architecture and can specify a specific register for an arithmetic, addressing, or control function; a specific storage location or offset; and a specific addressing mode for interpreting operands. More complex operations can be built by combining these simple native instructions, which can be executed sequentially or by control flow instructions.

[0084] The relationship between executable software instructions and hardware processors is structural. In other words, the instructions themselves are a series of symbols or numerical values. They do not inherently convey any information. It is the processor that is pre-configured by design to interpret the symbols / numerical values ​​that give meaning to the instructions.

[0085] The models used in this article can be configured to execute on a single machine at a single location, on multiple machines at a single location, or on multiple machines at multiple locations. When multiple machines are used, each machine can be customized for its specific task. For example, operations that require large code blocks and / or large amounts of processing power can be implemented on large and / or fixed machines.

[0086] In addition, some embodiments relate to a tangible and / or non-temporary computer-readable medium or computer program product including program instructions and / or data (including data structures) for performing various computer-implemented operations. Examples of computer-readable media include, but are not limited to, semiconductor storage devices, phase change devices, magnetic media such as disk drives, tapes, optical media such as CDs, magneto-optical media, and hardware devices specifically configured to store and execute program instructions, such as read-only storage devices (ROM) and random access memories (RAM). Computer-readable media can be directly controlled by end users, or the media can be indirectly controlled by end users. Examples of directly controlled media include media located at user facilities and / or media that are not shared with other entities. Examples of indirectly controlled media include media that users can access indirectly through external networks and / or through services that provide shared resources (e.g., "clouds"). Examples of program instructions include machine code (e.g., machine code generated by a compiler) and files containing higher-level code that can be executed by a computer using an interpreter.

[0087] In various embodiments, data or information used in the disclosed methods and apparatus are provided in an electronic format. Such data or information may include design layouts, fixed parameter values, floating parameter values, feature profiles, metrology results, and the like. In some cases, the set of instructions provided in an electronic format defines an algorithm for executing a metrology machine learning model (e.g., a neural network). As used herein, data or other information provided in an electronic format may be used for storage on a machine and for transmission between machines. Conventionally, data in an electronic format is provided in digital form and may be stored as bits and / or bytes in various data structures, lists, databases, and the like. The data may be embodied electronically, optically, and the like.

[0088] In some embodiments, the models can each be viewed as a form of application software that interfaces with the user and system software. System software typically interfaces with computer hardware and associated memory. In some embodiments, system software includes operating system software and / or firmware, as well as any middleware and drivers installed in the system. System software provides the basic non-task specific functions of the computer. In contrast, modules and other application software are used to perform specific tasks. Each native instruction of a module is stored in a memory device and represented by a numerical value.

[0089] Figure 3800. As shown, computer system 800 includes input / output subsystem 802, which can be used to interact with human users and / or other computer systems according to application implementation. Embodiments of the present invention can be implemented in program code on system 800, wherein I / O subsystem 802 is used to receive input program statements and / or data from human users (e.g., via GUI or keyboard), and display them back to the user. I / O subsystem 802 can include, for example, keyboard, mouse, graphical user interface, touch screen or other interfaces for input, and for example LED or other flat display or other interfaces for output. Other elements of embodiments of the present disclosure, such as the training system for the result model for explaining metering data, can be implemented with a computer system similar to computer system 800, but without I / O.

[0090] The program code may be stored in a non-transitory medium such as permanent storage 810 or storage 808 or both. One or more processors 804 read the program code from one or more non-transitory media and execute the code to enable the computer system to complete the methods performed by the embodiments of the present invention, such as those involving the generation or use of processing simulation models as described herein. It should be understood by those skilled in the art that the processor can accept source code, such as statements for performing training and / or interpretation operations, and interpret or compile the source code into machine code that can be understood at the hardware gate level of the processor. The bus couples the I / O subsystem 802, the processor 804, the peripheral device 806, the memory 808, and the persistent storage 810. In some embodiments, at least some computer system features, such as the processor 804, the memory 808, the storage 810, and / or the storage device 810 are implemented in a cloud or edge network that is physically different from the I / O subsystem (if any). In some embodiments, some computer system features, such as the processor 804, the memory 808, and / or the storage 810, are distributed on multiple physical devices.

[0091] Context of metrology tool implementation

[0092] In various embodiments, the metrology system includes an optical metrology tool and a metrology machine learning model, such as a model generated using the training process described herein. The optical metrology tool may include various configurations, but generally includes a light source and a detector for detecting light reflected, transmitted and / or scattered from features on the substrate being detected. The light source may be monochromatic or polychromatic. The wavelength generated may be in the infrared, visible, ultraviolet or X-ray range. It may be non-polarized, linearly polarized, circularly polarized, etc. The optical metrology tool may additionally include one or more optical elements for directing light onto substrate features and / or collecting and / or directly interacting with substrate features. The optical metrology tool may additionally include one or more processors or other processing elements to capture and / or interpret signals from the detector.

[0093] In some embodiments, the optical metrology tool includes a metrology / reflectance spectrum generator such as the YieldStar 1000 from ASML Netherlands BV of Veldhoven, The Netherlands. TM See, for example, Cramer et al., “High-NA optical CD metrology on small in-cell targets enabling improved higher order dose control and process control for logic,” Proceedings of SPIE, 10145, Metrology, Inspection, and Process Control for Microlithography XXXI, 101451B (28 March 2017), the entire contents of which are incorporated herein by reference.

[0094] Figure 4 A block diagram of an exemplary metrology system 401 is shown. As shown, the system includes an optical metrology tool 407 configured to probe a substrate such as a substrate 403 having a feature 405. The optical metrology tool 407 collects optical information from the feature 405 and generates an optical metrology output (e.g., reflection intensity versus wavelength or optical critical dimension information). A metrology machine learning model 409 receives the optical metrology output and generates geometric information about the feature; for example, feature shape, CD, profile, etc. See output 411. In some embodiments, the machine learning model 409 is implemented as a neural network.

[0095] Systems such as metrology system 401 can use optical information acquired directly from a test sample to determine the etch profile of a feature, such as in the case of OCD measurements, and then process that data through the models or relationships described herein.

[0096] in conclusion

[0097] In the description, many specific details are set forth to provide a thorough understanding of the implementations presented. The disclosed implementations may be practiced without some or all of these specific details. In other cases, well-known processing operations are not described in detail so as not to unnecessarily obscure the disclosed implementations. Although the disclosed implementations are described in conjunction with specific implementations, it should be understood that they are not intended to limit the disclosed implementations.

Claims

1. A method of determining a topography, a critical dimension and / or a profile of one or more features on a processed substrate, the method comprising: performing optical metrology on the one or more features on the processed substrate to produce an optical metrology output; providing the optical metrology output to a metrology machine learning model that has been trained using a training set of: (i) topography, critical dimensions, and / or profiles of a plurality of features, and (ii) the optical metrology outputs of the plurality of features; and A topography, a critical dimension, and / or a profile of the one or more features on the processed substrate is received from the metrology machine learning model.

2. The method according to claim 1 also includes training the metrology machine learning model by generating multiple optical metrology output sets, each optical metrology output set is generated for a different orientation and / or position of the test feature relative to the optical metrology device for the test feature.

3. The method according to claim 1 or 2, wherein: The optical metrology is a scatterometry technique.

4. The method according to claim 1 or 2, wherein: The optical metrology output includes a reflectance spectrum.

5. The method according to claim 1 or 2, wherein: The optical metrology output includes ellipsometric output data.

6. The method according to claim 1 or 2, further comprising: performing pattern recognition on the one or more features on the processed substrate; as well as A determination is made that the pattern of the one or more features matches an expected pattern.

7. The method according to claim 1 or 2, wherein: Performing optical metrology on the one or more features on the processed substrate includes raster scanning across the processed substrate.

8. The method according to claim 1 or 2, wherein: The topography, critical dimensions and / or contours of the plurality of features in the training set are obtained using electron microscopy techniques.

9. The method according to claim 1 or 2, wherein: The shapes, critical dimensions and / or contours of a plurality of features in the training set are obtained using CD-SAXS.

10. The method according to claim 1 or 2, wherein: The quantitative machine learning model is generated using supervised machine learning techniques.

11. The method of claim 1 or 2, further comprising decomposing the optical metrology output to produce a reduced set of optical metrology values, and providing the reduced set of optical metrology values ​​to the metrology machine learning model.

12. The method according to claim 11, wherein: Decomposing the optical metrology output may include identifying principal components of the optical metrology output or applying the optical metrology output to an autoencoder.

13. A metering system comprising: an optical metrology tool comprising an optical probe source and an optical detector and a processor configured to generate an optical metrology output based on data generated by the optical detector when the optical probe is directed over one or more features on a processed substrate; as well as A metrology machine learning model that has been trained using a training set of: (i) shapes, critical dimensions, and / or profiles of a plurality of features, and (ii) optical metrology outputs of the plurality of features, wherein the metrology machine learning model is configured to: receiving the optical metrology output from the optical metrology tool; and The topography, critical dimension and / or contour of the one or more features is output on the processed substrate.

14. The metering system according to claim 13, wherein: The metrology machine learning model is trained using multiple optical metrology output sets, each optical metrology output set generated for a different orientation and / or position of a test feature relative to a test optical metrology device.

15. The metering system according to claim 13 or 14, wherein: The optical metrology tool is a scatterometer.

16. The metering system according to claim 13 or 14, wherein: The optical metrology output includes a reflectance spectrum.

17. The metering system according to claim 13 or 14, wherein: The optical metrology output includes ellipsometric output data.

18. The metrology system of claim 13 or 14, further comprising a pattern recognition tool, the pattern recognition tool comprising a camera and a processor, the camera and the processor being configured to: performing pattern recognition on the one or more features on the processed substrate; and A determination is made that the pattern of the one or more features matches an expected pattern.

19. The metering system according to claim 13 or 14, wherein: The optical metrology tool is configured to raster scan across the substrate being processed.

20. The metering system according to claim 13 or 14, wherein: The topography, critical dimensions and / or contours of a plurality of features in the training set are obtained using electron microscopy techniques.

21. The metering system according to claim 13 or 14, wherein: The shapes, critical dimensions and / or contours of the plurality of features in the training set are obtained using CD-SAXS.

22. The metering system according to claim 13 or 14, wherein: The econometric machine learning model is generated using supervised machine learning techniques.

23. The metering system according to claim 13 or 14, wherein: The metrology machine learning model is also configured to decompose the optical metrology output to produce a reduced set of optical metrology values, and provide the reduced set of optical metrology values ​​to the metrology machine learning model.

24. The metering system of claim 23, wherein: The metrology machine learning model is configured to decompose the optical metrology output by identifying principal components of the optical metrology output or by applying the optical metrology output to an autoencoder.

Images