Combining physical modeling and machine learning

Abstract

A system and method for OCD metrology is provided, including receiving reference parameters, receiving a plurality of sets of measured scatterometry data, and receiving an optical model designed to generate one or more sets of model scatterometry data according to a set of pattern parameters, and training a machine learning model by applying target features including the reference parameters during training and by applying input features including the measured scatterometry data sets and the model scatterometry data sets, such that the trained machine learning model estimates new wafer pattern parameters from a subsequently measured scatterometry data set.

Description

TECHNICAL FIELD

[0001] The present invention relates generally to the field of optical inspection of integrated circuit wafer patterns, and more particularly to algorithms for measuring wafer pattern parameters. BACKGROUND

[0002] Integrated circuits (ICs) are produced on semiconductor wafers through a number of steps of depositing, modifying, and removing thin layers that accumulate in a stack structure on the wafer. These stack structures (or "stacks") are typically formed in a repeating pattern that has optical properties like a diffraction grating. Modern metrology methods for measuring critical dimensions (CDs) and material properties of these patterns exploit these optical properties. Hereinafter, CDs and material properties are also referred to as "pattern parameters", or simply "parameters". These parameters can include height, width, and pitch of the stacks. As described in Dixit et al., "Sensitivity Analysis and Line Edge Roughness Determination of 28 nm Pitched Silicon Fin Using Optical Critical Dimension Metrology Method with Mueller Matrix Spectroscopy Based Ellipsometer", J. Micro / Nanolith. MEMS MOEMS. 14(3), 031208 (2015) (incorporated herein by reference), pattern parameters can also include: side wall angle (SWA), spacer width, spacer pull-down, epi proximity, footing / undercut, over / underfill for 2D (HKMG), 3D profile (FinFET), and line edge roughness (LER) distribution.

[0003] Optical critical dimension (OCD) metrology employs scatterometry to measure scatter data, i.e. reflected optical radiation, which is indicative of the optical properties of the pattern. A measurement set of scatter data (also referred to as a scatter signature) can include data points of reflected zeroth order irradiance versus radiation incidence angle. Alternatively or additionally, scatter data can include a spectrogram, which is a metrology of reflected radiation intensity over a range of wavelengths or frequencies. Other types of scatter data known in the art can also be applied in OCD metrology.

[0004] U.S. Patent 6,476,920 to Scheiner and Machavariani, "Method and Apparatus for Measuring Patterned Structures", incorporated herein by reference, describes the development of an "optical model" (also referred to as a "physical model"). An optical model is a function (i.e. a set of algorithms) that defines the relationship between reflected radiation and the physical structure of the wafer. That is, an optical model is a theoretical model of how light reflects from a pattern with known parameters. This optical model can therefore be applied to estimate scatter data from a set of known pattern parameters, which would be measured during a spectroscopic test. An optical model can also be designed to perform the inverse (or "inverse") function, i.e. to estimate pattern parameters from measured scatter data.

[0005] Optical models are commonly applied in OCD metrology in IC production processes to measure whether the correct parameters were used to manufacture the wafer patterns. Each pattern of a given wafer can be measured to determine how much each pattern’s parameters vary from the design specification or average.

[0006] As an alternative to optical modeling, machine learning (ML) techniques can be used for pattern parameter estimation based on scatterometry data. For example, as described in Rothstein et al. PCT Patent Application WO 2019 / 239380, incorporated herein by reference, a machine learning model can be trained to identify a correspondence between measured scatterometry data and measured reference parameters by the following method. After the ML model is trained to estimate parameters from scatterometry data, it can be applied to make such parameter estimations during IC production.

[0007] Exemplary scatterometry tools for measuring (acquiring) scatterometry data (e.g., spectrograms) can include spectroscopic ellipsometers (SE), spectroscopic reflectometers (SR), polarization spectroscopic reflectometers, and other optical critical dimension (OCD) metrology tools. Such tools have been incorporated into currently available OCD metrology systems. One such OCD metrology system is the NOVA Improved OCD metrology tools, available from Nova Measuring Instruments Ltd. of Ra’anana, Israel, are used to measure pattern parameters that can be at a designated test site or “in-die.” Other methods for measuring critical dimensions (CDs) include interferometry, X-ray Raman spectroscopy (XRS), X-ray diffraction (XRD), and pump probe tools, among others. Some examples of such tools are disclosed in patents WO 2018 / 211505, US 10,161,885, US 10,054,423, US 9,184,102, and US 10,119,925, all assigned to the applicant and incorporated herein by reference in their entirety.

[0008] High precision pattern parameter measurement methods that do not rely on the above optical models include wafer measurements using equipment such as CD scanning electron microscopes (CD-SEMs), atomic force microscopes (AFMs), cross-sectional tunneling electron microscopes (TEMs), or X-ray metrology tools. These methods are generally more expensive and time consuming than optical and ML modeling methods.

[0009] However, both optical and ML modeling have drawbacks. Optical modeling is also time-consuming and prone to errors, especially as the dimensions of the pattern parameters continue to shrink, due to the idealization of the geometric model used for optical modeling of the actual pattern parameters, and due to the difficulty of numerically solving the nonlinear scattering equations. On the other hand, optical modeling is generally reliable due to its basis on physics. Furthermore, the direct link between the physical parameters and the theoretical optical properties means that the results of the optical model are generally easier to interpret than the results of ML. On the other hand, while ML modeling avoids some of the time-consuming hurdles in optical modeling, it generally requires a large set of reference parameters and scattering data for training, and the acquisition of these parameters requires expensive, time-consuming metrology equipment. The embodiments of the invention disclosed below help overcome the drawbacks of both approaches. SUMMARY

[0010] Embodiments of the invention provide a system and method for generating a machine learning model for OCD that utilizes measured scattering data and a known optical model designed according to physical laws. The method provided includes receiving a plurality of sets of reference parameters from a plurality of respective wafer patterns, and receiving a plurality of sets of respective measured scattering data from the plurality of respective wafer patterns; receiving an optical model for calculating a set of model scattering data from pattern parameters provided to the optical model; training a machine learning model by applying target features comprising the reference parameters during training, and by applying input features comprising the measured scattering data sets and the model scattering data sets, such that the trained machine learning model estimates new wafer pattern parameters from subsequently measured scattering data sets.

[0011] In some embodiments, the optical model (OM) can be further designed to perform an inverse function of calculating model pattern parameters from a set of scattering data, and the training of the ML model can include: calculating a plurality of sets of model pattern parameters by applying the optical model to calculate a respective set of model pattern parameters for the plurality of sets of measured scattering data, calculating a plurality of sets of model scattering data by applying the optical model to each respective set of model pattern parameters to generate a respective set of model scattering data, generating a combined feature vector by combining, the plurality of sets of measured scattering data, the measured scattering data, the respective model pattern parameters, and the respective model scattering data, and training the ML model, wherein the reference parameters are the target features for training, and the combined feature vector is the input features.

[0012] In further embodiments, the ML model can be a transfer neural network (NN), the training of the ML model can include training a first NN and a transfer NN, such that training the first NN and the transfer NN includes: generating a plurality of sets of simulated pattern parameters; generating a plurality of sets of model scattering data by applying the optical model for each set of simulated pattern parameters, thereby generating a plurality of sets of model scattering data; training the first NN with a first target set comprising the plurality of sets of simulated pattern parameters and a first input feature comprising the plurality of sets of model scattering data; and training the transfer NN using initial layers transferred from the first NN to train one or more final layers of the transfer NN. The plurality of sets of reference parameters are set as target features and the plurality of sets of measured scattering data are as respective input features.

[0013] In further embodiments, training the ML model can include minimizing a loss function with respect to the reference set of pattern parameters, wherein the loss function is a mean squared error (MSE) function. The reference set of parameters can be measured with high precision metrology by one or more of a CD scanning electron microscope (CD-SEM), an atomic force microscope (AFM), a cross-sectional tunneling electron microscope (TEM), or an X-ray metrology tool, or high precision OCD spectroscopy relying on optical modeling. The plurality of respective wafer patterns can be located on one or more wafers. In some embodiments, the plurality of sets of measured scattering data can be measured by two or more measurement channels.

[0014] By embodiments of the present invention, there is also provided a system for OCD metrology, the system comprising a processor having a non-transitory memory, the memory comprising instructions that, when executed by the processor, cause the processor to implement the steps of: receiving a plurality of sets of reference parameters from a plurality of respective wafer patterns, and receiving a plurality of sets of respective measured scattering data from the plurality of respective wafer patterns; receiving an optical model designed to calculate model scattering data from pattern parameters provided to the model; training a machine learning model by applying target features (comprising the reference parameters) and input features (comprising the sets of measured scattering data and model scattering data) during training, such that the trained machine learning model estimates new wafer pattern parameters from a subsequent set of measured scattering data. BRIEF DESCRIPTION OF DRAWINGS

[0015] For a better understanding of the various embodiments of the present invention and to show how the same can be carried out, reference is made to the accompanying drawings, which are used to illustrate examples of the present invention. The structural details of the present invention are shown in order to provide a thorough understanding of the present invention. The description taken with the drawings will make apparent to those skilled in the art how the several forms of the present invention can be embodied in practice. In the drawings:

[0016] Figure 1 is a schematic illustration of a system for generating a machine learning model for OCD metrology using spectrometer data and an optical OCD model, in accordance with embodiments of the present invention;

[0017] Figure 2 FIG. 1 is a flowchart depicting a procedure for generating a first machine learning model for OCD metrology using spectrometer data and optical OCD models, according to embodiments of the present invention;

[0018] Figure 3 FIG. 2 is a schematic diagram of a neural network implementing the first machine learning model, according to embodiments of the present invention;

[0019] Figure 4 FIG. 3 is a flowchart depicting a procedure for generating a second machine learning model for OCD metrology using spectrometer data and optical OCD models, according to embodiments of the present invention;

[0020] Figure 5A and Figure 5B FIG. 4 is a schematic diagram of a neural network implementing the second machine learning model, according to embodiments of the present invention; and

[0021] Figure 6 and Figure 7 FIG. 5 is a graph depicting accuracy of a machine learning (ML) model trained on scatterometry data (prior art) versus a ML model trained on model scatterometry data generated with measured scatterometry data and optical models, according to embodiments of the present invention. DETAILED DESCRIPTION

[0022] Embodiments of the present invention provide systems and methods for generating machine learning (ML) models for optical critical dimension (OCD) monitoring by utilizing known optical OCD models (referred to herein as “optical models”). By training the ML models using at least some scatterometry data generated by the optical models (i.e., using “model” scatterometry data), rather than training only on actual measurements, the results can be improved and costs can be reduced.

[0023] Figure 1 FIG. 6 is a schematic diagram of a system 10 of embodiments of the present invention that utilizes spectrometer data and optical models to generate machine learning models for OCD metrology.

[0024] System 10 can be run within a production line (not shown) to produce and monitor wafers 12. As shown, wafers 12 include patterns 14. As shown in enlarged pattern 14a, these patterns have parameters such as height (“h”), width (“w”), and pitch (“p”), among other parameters described in the background above. Typically, wafers have multiple regions or segments designed with the same pattern (i.e., all patterns are manufactured using the same pattern design). For each pattern, multiple parameters can be measured. Hereinafter, a set of multiple parameters from a given pattern is also referred to as a pattern vector.

[0025] Manufacturing variations cause slight variations in the pattern parameters from wafer to wafer and on a single wafer, which are represented by variations in the measured scatter data.

[0026] System 10 includes a light source 20 that generates a light beam 22 of a predetermined wavelength range. Light beam 22 is reflected from wafer pattern 14 (shown as reflected or "scattered" light 24) toward a spectrophotometric detector 26. In some configurations, the light source and spectrophotometric detector are included in an OCD metrology system 30 (e.g., an ellipsometer or a spectrophotometer). The construction and operation of metrology system 30 can be of any known type, for example, as disclosed in U.S. Pat. No. 5,517,312, "Thin Film Thickness Measurement Apparatus," by Finarov, incorporated herein by reference. Typically, metrology system 30 includes additional components not shown, such as light guiding optics, which can include beam deflectors with objective lenses, beam splitters, and mirrors. Additional components of such systems can include imaging lenses, polarizing lenses, variable aperture holes, and motors. The operation of these elements is typically performed automatically by a computer controller, which can include I / O devices, and can also be configured to perform data processing tasks, such as generating scatter data 32.

[0027] Scatter data 32 generated by metrology system 30 typically includes various types of plot data 34, which can be represented in vector form (e.g., a spectrogram, whose data points are measurements of reflected light intensity at different wavelengths). As noted above, variations between measured scatter data sets represent different pattern parameters. In typical OCD metrology, the range of light measured can include the visible spectrum, and can also include wavelengths in the ultraviolet and infrared regions. A typical spectrogram output for OCD metrology can have 245 data points, covering a wavelength range of 200 to 970 nm.

[0028] In embodiments of the invention, a computer system including ML tools known in the art (referred to herein as ML modeling system 40) can be configured to train an ML model for OCD metrology. The input training feature vectors used by the ML modeling system can include multiple sets of measured scatter data 34 as well as scatter data generated by optical modeling 42, i.e., "model scatter data." Reference parameters 44 (denoted as "pattern parameter vectors") can be used as the target feature set for ML training. The reference parameters can be acquired from one or more wafer patterns by high precision means known in the art (e.g., CD-SEM, AFM, TEM, X-ray metrology, or high precision OCD spectroscopy relying on optical modeling). After training, the ML model is used to predict pattern parameters based on measured scatter data, which can be applied, for example, in the monitoring of wafer production.

[0029] The ML modeling system 40 can operate independently of the metrology system 30 or can be integrated with the metrology system.

[0030] In the following, ML training procedures are described that utilize training feature vectors based on measured scattering data and model scattering data. By combining the measured data and the knowledge of the optical method, the ML model can become more accurate, despite relying on a relatively small training data set. In particular, two examples of such ML training are described, namely Figure 2 the procedure 200 shown in Figure 3 and Figure 4 the procedure 400 shown in Figure 5A and Figure 5B In the following, the procedure 200 is also referred to as the combined feature vector method, and the procedure 400 is also referred to as the transfer learning neural network method.

[0031] The description of the procedure 200 and the procedure 400 uses the following terminology. The scattering data measured by OCD metrology from a single pattern is referred to as a set of scattering data (i.e., a “single” set), and can be denoted by the measured scattering vector A set of theoretical scattering data calculated by the optical model from a set of pattern parameters (i.e., a “set of model scattering data”) can be denoted by the scattering vector A set of reference parameters, i.e., parameters measured directly from a wafer without the optical model, can be denoted by the reference parameter vector While parameters generated from a set of scattering data according to the optical model can be denoted by the model parameter vector

[0032] The optical model itself can be denoted by a function that can predict (or “generate”) a scattering vector from a set of parameters or perform the inverse function, i.e., predict from a set of scattering data The inverse function of the optical model can also be written in the “pseudo-inverse” form of the function, i.e., the function can be written as f OM The pseudo-inverse form can be written as Thus, mathematically, the above definitions can be written as:

[0033] and

[0034]

[0035] Figure 2is a flowchart depicting a computer-implemented procedure 200 for generating a machine learning model for OCD metrology using scatterometry data and an optical model, according to one embodiment of the present application. The procedure 200 can be implemented by the ML modeling system 40 as described above.

[0036] The procedure 200 begins with a set of steps 210, in which external data and algorithms are provided to the modeling system. The first step S212 includes receiving a plurality of sets of reference parameters, i.e., a plurality of parameter vectors:

[0037] The parameter vectors are measured from respective reference patterns. Each set of reference parameters includes one or more data points (e.g., height, width, pitch, etc.). As described below, each set is subsequently associated with a set of scatterometry data during ML training. The reference wafer patterns are typically a portion of an entire wafer, which is a portion that is repeated on the wafer surface. The reference parameters can be measured with high precision OCD metrology. As described above, high precision OCD metrology can include CD-SEM, AFM, TEM, X-ray metrology, or high precision OCD spectroscopy that relies on optical modeling. As described below, the number of measured reference parameter sets defines the size of the training data set.

[0038] S214 includes receiving a plurality of sets of measured scatterometry data:

[0039] Each set of scatterometry data is measured from a respective wafer pattern from which a respective parameter vector (i.e., a respective “set of parameters”) is also measured.

[0040] In S216, in addition to the measured plurality of sets of scatterometry data and reference parameters, an optical OCD model and its inverse function (i.e., f OM and ) are received. As described above, the optical OCD model is generated from application of optical physics laws. It should be appreciated that the sets of S210 can be performed in any order or in parallel, in particular, the measured scatterometry data can be acquired before the pattern parameters (this option is required when using destructive metrology tools to obtain the pattern parameters).

[0041] Next, in S220, for each set of measured scatterometry data the optical model is applied to calculate a set of model pattern parameters That is,

[0042] The optical model f OM is then applied to each generated set of model pattern parameters to generate a respective set of model scatterometry data, i.e., a model scatterometry vector That is,

[0043] Then, the scattering data of each group of measurements were... (Remove the index k of the scattering dataset used to summarize all measurements) and its corresponding model parameters and its corresponding set of model scattering data The combination is a single combined vector, which is called in this context. Right now

[0044] In the example scenario, vectors and Each includes 245 data points. Vector There are typically 1 to 10 data points. Therefore, the example vector It may include 245 + 245 + 10 = 500 data points.

[0045] In S230, a machine learning model is trained to combine each feature vector. Its corresponding pattern parameters The reference set is relevant. (The input and output features used for training are matched with the same wafer pattern, which is used to combine vectors through the input features.) Measurement output characteristics and measuring the scattering vector )

[0046] Figure 3 A schematic example of an ML model as a neural network (NN) 300 is shown, with the input feature vector represented as an exemplary vector. The output feature vector is represented as The input nodes of the ML model are represented as node 320, and the output nodes are represented as node 322. The number of input nodes is typically adjusted to the size of the input vector, meaning the example above will have 500 input nodes (this can be extrapolated from each). The additional data points are rounded to powers of 2, resulting in 512 nodes. The number of output nodes will correspond to the number of parameters in each set of pattern parameters. As shown in NN300, the neural network can be designed to be fully connected. Depending on the number of measured scattering datasets (i.e., the size of the training dataset), hidden layers 340 can be added to NN 300. (For a finite number of measured scattering datasets, hidden layers may not improve accuracy.) Training is typically performed according to standard ML training methods, which may include, for example, L2 regularization. The correlation between the input and output feature vectors can be defined by the mean squared error (MSE) loss function. Preferably, the validation dataset combining the feature vectors is generated from measured scattering datasets obtained from different wafers (rather than the wafers used in the training dataset).

[0047] Returning to Figure 2 After training the ML model, the ML model can be applied during IC wafer production to monitor the pattern parameters (S240). During the production phase, scatterometry data is measured from wafer patterns. Then, as described above, a combined feature vector is generated from the measured scatterometry data sets. The combined feature vector is then applied to the trained ML model to predict (i.e., estimate) the parameters of the wafer pattern.

[0048] Figure 4 Figure 4 is a flowchart depicting a computer-implemented procedure 400 for generating a machine learning model for OCD metrology using spectral data and an optical OCD model, according to one embodiment of the present application. Similar to procedure 200, procedure 400 can be implemented by ML modeling system 40, as described above Figure 1 with respect to procedure 200. In contrast to procedure 200, procedure 400 includes a two-stage ML training process with transfer learning (i.e., S430, described below).

[0049] Procedure 400 begins with a set of steps 410, in which external data and algorithms are provided to the modeling system. The first step S212 includes receiving a plurality of sets of reference parameters, e.g., parameter vectors These parameters are measured from respective reference patterns. S214 includes receiving a plurality of sets of measured scatterometry data, e.g., Each set of measurements is measured from a corresponding reference pattern from which a corresponding set of reference parameters is also measured.

[0050] In addition to the sets of measured scatterometry data and reference parameters, an optical OCD model f OM is received in S416, defined as Note that the difference between the set of steps S410 and the corresponding steps 210 of procedure 200 is that step 416 does not require obtaining the inverse optical model, in contrast to corresponding step 216, since procedure 400 does not require that inverse function. Note also that the steps of receiving or generating the data required for ML training can be performed in any order. In particular, the optical model can be obtained prior to the other steps, e.g., in a step of initializing the ML modeling system.

[0051] In S420, a plurality of sets of simulated pattern parameters i.e., The simulated parameters can be generated to have a distribution that reflects typical manufacturing variations for such parameters.

[0052] In S422, the optical model is then applied to each set of simulated pattern parameters to generate a set of model scatterometry data In some embodiments, the set of scatterometry data can be a spectrum map, which can have, for example, 245 data points.

[0053] In S430, the neural network is trained by transfer learning, including two training phases, a first phase 432 and a transfer phase 434.

[0054] In the first phase 432, a first neural network (NN) is trained using a form of fully supervised learning, in which the generated model scatterometry data is applied as an input feature vector to the ML model, and each input output feature vector is a corresponding set of simulated pattern parameters

[0055] In the second phase of ML training (S432), the initial layers of the first phase NN (i.e., the activation functions connecting the initial layer nodes) are transferred to the second phase of ML training, which is referred to herein as the “transfer NN” of the “transfer training phase.” The transfer NN is then trained with measured scatterometry data sets as input features, and corresponding reference parameters for each measured scatterometry data as output features. After training and validating the ML model, it can be used in production, as shown in S440, similar to S240 of procedure 200.

[0056] Figure 5A and Figure 5B A schematic example of the two respective phases of the above-described ML training, a first phase 500( Figure 5A ) and a transfer training phase 510( Figure 5B ) is shown. In the exemplary embodiment, in the first phase, for the input features, the vectors generated from the simulated parameter vectors are used to train the input nodes 520, hidden layers 522, and output layer 524, and for the output (“target”) features, the corresponding simulated parameter vectors are used. That is, the training associates each with the corresponding from which it was generated.

[0057] The transfer training phase uses the same activation functions and hidden nodes calculated in the first training phase, while retraining the final activation layer 530 (shown as white nodes). For the input features, the training of the transfer NN uses the vectors For the output features, the corresponding reference parameters acquired in S410 are used. That is, the training associates each with the corresponding measured from the same pattern.

[0058] Generally, the number of input nodes corresponds to the size of the vector In one exemplary embodiment, the two stages of ML modeling can be configured as a fully connected neural network with an input layer of size 490. The input layer of 490 nodes can be used to receive an input feature vector that combines two scatter data vectors of 245 data points each. The two scatter data vectors can be acquired from different "channels," i.e., from different measurement configurations of a given pattern. For example, the angle of incident light or the polarization of light can be changed to create two different scatter data vectors, both of which provide information about the same pattern. The input layer can be followed by two hidden layers of size 16, a second hidden layer followed by a rectified linear activation function (ReLU) activation function, and a linear output layer of size 3, which corresponds to the number of parameters in a set of exemplary pattern parameters.

[0059] Figure 6 A graph showing the accuracy of ML models trained on measured scatter data and model scatter data generated by optical models, according to embodiments of the present application, as opposed to prior art methods tested by the inventors. The prior art methods include ML models trained using scatter data only (see, e.g., PCT patent application WO 2019 / 239380, Rothstein et al., above), and non-ML methods based on estimating pattern parameters using optical models only (see, e.g., patent US 6,476,920, Scheiner and Machavariani, above). As shown in the graph, Figure 6 For the ML models developed by the methods described herein, the standard deviation between the reference parameters and the model-predicted parameters is shown to be lower (i.e., the accuracy is higher), as shown in the graph of The standard deviation (std) of the ML model output

[0060]

[0061] Figure 7 A graph showing the accuracy of the procedures 200 and 400 disclosed herein relative to the above-described prior art ML methods (denoted as "ML Baseline"). The ML Baseline was trained in a single stage with measured scatter data as input features and corresponding reference parameters as output features. Procedure 200 represents the combined feature vector ("Combined FV") method and procedure 400 represents the transfer NN method. Different sizes of training data sets were randomly drawn from a large pool of data, which are represented by error bars marking the graphical results. Due to this resampling, the error bars represented in the graph are 1-sigma uncertainty. As shown in the graph, for training sets of greater than 5 or 10 measured data sets, the methods of the present application show significantly higher accuracy than the Baseline. Dimensionality reduction methods such as Principal Component Analysis (Kernel PCA) do not improve the fit compared to the full set of measured data.

[0062] As shown, both program 200 and program 400 generate a higher accuracy ML model than the ML baseline method.

[0063] It should be understood that, in accordance with conventional techniques, the processing elements shown or described herein are preferably implemented in one or more computers in computer hardware and / or computer software in a non-transitory computer readable medium, such as computer processors coupled by a computer bus or other connection device, memory, I / O devices, and network interfaces. The term "processor" or "device" as used herein is intended to include any processing device, such as devices including central processing units (CPUs) and / or other processing circuitry (e.g., GPUs). Additionally, the term "processor" or "device" can refer to more than one processing device, and various elements associated with the processing device can be shared by other processing devices.

[0064] The term "memory" as used herein is intended to include memory associated with a processor or CPU, such as RAM, ROM, a fixed memory device (e.g., hard drive), a removable memory device (e.g., floppy diskette, Zip® disk, magnetic tape), flash memory, etc. Such a memory can be considered a computer readable storage medium.

[0065] Additionally, "input / output devices" or "I / O devices" as used herein can include one or more input devices (e.g., keyboard, mouse, scanner, HUD, etc.) that input data to a processing unit, and / or one or more output devices (e.g., speaker, display, printer, HUD, AR, VR, etc.) for displaying results associated with the processing unit.

[0066] Embodiments of the application can include systems, methods, and / or computer program products. A computer program product can include a computer readable storage medium (or media) having computer readable program instructions stored therein for causing a processor to carry out operations of the present application.

[0067] The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium can be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. A non-exhaustive list of more specific examples of the computer readable storage medium includes the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a Blu-ray, a memory stick, a floppy disk, a mechanically encoded device such as punch cards or raised structures in grooves of a groove having instructions recorded thereon, and any suitable combination of the foregoing. A computer readable storage medium, as used herein, is not to be construed as being transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission media (e.g., light pulses passing through a fiber-optic cable), or electrical signals transmitted through a wire.

[0068] Computer readable program instructions can be downloaded to respective computing / processing devices from a computer readable storage medium or to an external computer or external storage device via a network, for example, the Internet, a local area network, a wide area network, and / or a wireless network. The network can comprise copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers and / or edge servers. A network adapter card or network interface in each computing / processing device receives computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium within the respective computing / processing device.

[0069] Computer readable program instructions for carrying out operations of the present application can be assembly program instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state-setting data, or any combination of source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C++ or the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer readable program instructions can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate arrays (FPGA), or programmable logic arrays (PLA) can execute the computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry to perform the instructions.

[0070] The computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart diagrams and / or block diagrams.

[0071] The computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart diagrams and / or block diagrams.

[0072] The computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart diagrams and / or block diagrams.

[0073] Any flow diagrams and block diagrams described herein describe possible architectures, functions, and operations for systems, methods, and computer program products according to various embodiments of the present application. In this regard, each block in the flow diagrams or block diagrams can represent a module, segment, or portion of instructions, which includes one or more executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the blocks can occur out of the order noted in the flow diagrams. For example, two blocks shown in succession can in fact be executed substantially concurrently, or the blocks can sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flow diagrams, and combinations of blocks in the block diagrams and / or flow diagrams, can be implemented by special purpose hardware-based systems that perform the specified functions or acts, or combinations of hardware and computer instructions.

[0074] The description of the various embodiments of the present application is merely intended to be illustrative and is not intended to limit or restrict the scope of the embodiments disclosed. Many modifications and variations to the described embodiments are possible and will be apparent to those of ordinary skill in the art. The terminology used herein is intended to be interpreted in a non-limiting manner, with the terms being selected to best explain the principles of the embodiments, the practical application of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for optical critical dimension metrology, comprising: receiving a plurality of sets of reference parameters from a plurality of respective wafer patterns and receiving a respective plurality of sets of measured scatterometry data from the respective plurality of wafer patterns; receiving an optical model designed to calculate model scatterometry data as a function of pattern parameters provided to the optical model, wherein the optical model is designed according to physical laws; and training a machine learning model by training during the training by applying target features comprising the plurality of sets of reference parameters and by applying input features comprising the plurality of sets of measured scatterometry data and the model scatterometry data, such that the trained machine learning model can estimate new wafer pattern parameters from subsequently measured scatterometry data, wherein the optical model is further designed to perform an inverse function of calculating model pattern parameters from a set of scatterometry data, and wherein the training of the machine learning model comprises: calculating a plurality of sets of model pattern parameters by calculating for each set of measured scatterometry data a respective set of model pattern parameters by applying the optical model; calculating a plurality of sets of model scatterometry data to generate a respective set of model scatterometry data by applying the optical model to each respective set of model pattern parameters; generating a combined feature vector for each of the plurality of sets of measured scatterometry data by combining the set of measured scatterometry data, the respective model pattern parameters and the respective model scatterometry data into a combined feature vector; and training the machine learning model with the reference parameters as the target features of the training and the combined feature vector as the input features.

2. The method of claim 1, wherein the machine learning model is a transfer neural network, wherein the training of the machine learning model comprises: training a first neural network and the transfer neural network, and wherein training the first neural network and the transfer neural network comprises: generating a plurality of sets of simulated pattern parameters; generating a plurality of sets of model scatterometry data by generating for each set of simulated pattern parameters a set of model scatterometry data by applying the optical model; training the first neural network using a first target set comprising the plurality of sets of simulated pattern parameters and a first input feature comprising the plurality of sets of model scatterometry data; and training the transfer neural network using initial layers transferred from the first neural network to train one or more final layers of the transfer neural network as a target and the plurality of sets of reference parameters as the target features and the plurality of sets of measured scatterometry data as the respective input features.

3. The method of claim 1, wherein training the machine learning model comprises minimizing a loss function with respect to a set of the reference parameters, and wherein the loss function is a mean squared error (MSE) function.

4. The method of claim 1, wherein the set of reference parameters is measured by one or more of a critical dimension scanning electron microscope (CD-SEM), an atomic force microscope (AFM), a cross-sectional tunneling electron microscope (TEM), or an X-ray metrology tool.

5. The method of claim 1, wherein the plurality of respective wafer patterns are on one or more wafers.

6. The method of claim 1, wherein the plurality of sets of measured scatterometry data are measured by two or more measurement channels.

7. A system for optical critical dimension metrology, comprising a processor having non-transitory memory, the memory including instructions that, when executed by the processor, implement the steps of: receiving a plurality of sets of reference parameters from a plurality of respective wafer patterns, and receiving a plurality of sets of respective measured scatterometry data from the plurality of respective wafer patterns; receiving an optical model designed to compute a set of model scatterometry data from a set of pattern parameters provided to the optical model, wherein the optical model is designed according to physical laws; and training a machine learning model during the training by applying target features including the reference parameters, and by applying input features including the plurality of sets of measured scatterometry data and the plurality of sets of model scatterometry data, such that the trained machine learning model can estimate new wafer pattern parameters from subsequently measured scatterometry data, wherein the optical model is further designed to perform an inverse function of computing model pattern parameters from a set of scatterometry data, and wherein the training of the machine learning model includes: computing a plurality of sets of model pattern parameters by applying the optical model to compute a respective set of model pattern parameters for each set of measured scatterometry data; computing a plurality of sets of model scatterometry data to generate a respective set of model scatterometry data by applying the optical model to each respective set of model pattern parameters; generating a combined feature vector for each of the plurality of sets of measured scatterometry data by combining the set of measured scatterometry data, the respective model pattern parameters, and the respective model scatterometry data into a combined feature vector; and training the machine learning model with the reference parameters as the target features of the training, and the combined feature vectors as the input features. training a first neural network and the transfer neural network, and wherein training the first neural network and the transfer neural network includes: generating a plurality of sets of simulated pattern parameters; generating a plurality of sets of model scatterometry data by applying the optical model to generate a set of model scatterometry data for each set of simulated pattern parameters; training the first neural network using a first target set including the plurality of sets of simulated pattern parameters and a first input feature including the plurality of sets of model scatterometry data; training the transfer neural network using initial layers transferred from the first neural network to train one or more final layers of the transfer neural network as a target, and the plurality of sets of reference parameters as the target features and the plurality of sets of measured scatterometry data as the respective input features.

9. The system of claim 7, wherein training the machine learning model includes minimizing a loss function with respect to a set of the reference parameters, and wherein the loss function is a mean squared error (MSE) function. ​ ​ ​ ​ ​ ​ ​ 8. The system of claim 7, wherein the machine learning model is a transfer neural network, wherein the training of the machine learning model comprises: ​ ​ ​ ​ ​ ​ 10. The system of claim 7, wherein the set of reference parameters are measured by high precision metrology by one or more critical dimension scanning electron microscope (CD-SEM), atomic force microscope (AFM), cross-section tunneling electron microscope (TEM), or X-ray metrology tool.

11. The system of claim 7, wherein the plurality of respective wafer patterns are located on one or more wafers.

12. The system of claim 7, wherein the plurality of sets of measured scattering data are measured by two or more measurement channels.

Images