Composite data for device measurements

Abstract

The machine learning model using composite measurement data determines at least one parameter of the device under test using measured measurement data from the device. The composite measurement data is generated by combining measured measurement data from a reference device with synthetic measurement data calculated from a model of the reference device. The composite measurement data may be generated further based on synthetic measurement data calculated from a model of a modified reference device. The modified reference device may be generated using variation of at least one parameter of the model to expand the parameter space of the training range.

Description

[Technical Field]

[0001] (CROSS-REFERENCE TO RELATED APPLICATIONS) This application claims priority to and the benefit of U.S. Provisional Application No. 63 / 412,339, entitled "ENHANCED MACHINE LEARNING RECIPE," filed September 30, 2022, U.S. Provisional Application No. 63 / 498,474, entitled "COMPOSITE DATA FOR OPTICAL METROLOGY," filed April 26, 2023, and U.S. Non-Provisional Application No. 18 / 339,982, entitled "COMPOSITE DATA FOR DEVICE METROLOGY," filed June 22, 2023, all of which are assigned to the assignee of the present application and are incorporated by reference in their entireties herein.

[0002] The subject matter described herein relates generally to metrology, and more particularly to training and using machine learning models to characterize at least one parameter of a device structure. [Background technology]

[0003] In the semiconductor and other similar industries, optical metrology techniques, such as optical metrology or x-ray metrology, are often used to provide non-contact evaluation of specimens during processing. In optical metrology, for example, the specimen under test is illuminated with light, for example, at a single wavelength or multiple wavelengths. After the light interacts with the specimen, the resulting light is detected and analyzed to determine at least one property of the specimen.

[0004] Various types of metrology, such as optical critical dimension (OCD) of device structures, used for process improvement, monitoring, and control, typically use modeling of the structure under test. Models can be generated, for example, based on the structure's materials and nominal parameters, such as film thickness, line and space widths, etc. Shrinking critical dimensions and tighter tolerances, combined with increasingly complex structures such as 3D NAND, are challenging current modeling capabilities.

[0005] Machine learning-based OCD metrology solutions, and other similar metrology solutions, can be useful when matching against reference metrology data, but are limited to scenarios where inline process conditions are similar to those present in the process conditions used for model training. However, because there is typically only a limited amount of reference metrology data available for training machine learning recipes, robustness is a challenge for machine learning-based metrology solutions. For example, process variations, especially during the early development phase, can impair the usefulness of reference metrology data for training machine learning recipes. This effect is more pronounced when raw spectral data collected from samples after a manufacturing process revision exhibits a certain level of deviation from the data previously collected and used for model training. A possible solution to this problem is to acquire additional reference metrology data after the process revision and retrain the machine learning model. However, acquiring updated reference metrology data for retraining the machine learning model is inefficient because measuring reference metrology data can be prohibitively expensive and retraining the machine learning model can be time-consuming. Furthermore, acquiring additional reference metrology data and retraining the machine learning model in response to process revisions must be performed iteratively without considering further process revisions. Therefore, improvements are desirable. Summary of the Invention

[0006] Measurement of the device structure may be performed using a machine learning model trained using composite measurement data. The composite measurement data may be generated, for example, by integrating measurement data measured from a reference device with synthetic measurement data calculated from a model of the reference device. The composite measurement data may be generated further based on synthetic measurement data calculated from a model of a modified reference device. The modified reference device is generated using a variation in at least one parameter of the model to expand the parameter space of the training range. The composite measurement data may be generated, for example, by modifying the synthetic measurement data calculated from the model of the modified reference device based on a variation, e.g., a mismatch or a spectral transformation, between the measured measurement data from the reference device and the synthetic measurement data calculated from the model of the reference device. In another implementation, the composite measurement data may be generated by modifying the measured measurement data from the reference device based on a variation, e.g., a difference between the synthetic measurement data calculated from the model of the reference device and the synthetic measurement data calculated from the model of the modified reference device. Measurement of the device structure may be performed using a machine learning model trained using a training dataset including at least the composite measurement data.

[0007] In one implementation, a method for characterizing a device on a sample includes obtaining measured metrology data from the device and determining at least one parameter of the device based on the measured metrology data using a machine learning model that uses composite metrology data, where each composite metrology data includes an integration of measured metrology data for a reference device and first composite metrology data for a first model of the reference device.

[0008] In one implementation, a metrology system configured to assist in characterization of devices on a sample includes a radiation source configured to generate radiation incident on a device on the sample, at least one detector configured to detect radiation from the device generated in response to the radiation incident on the device, and at least one processor coupled to the at least one detector. The at least one processor is configured to obtain measured metrology data from the device and determine at least one parameter of the device based on the measured metrology data using a machine learning model that uses composite metrology data. Each composite metrology data includes an integration of metrology data measured for a reference device and first composite metrology data for a first model of the reference device.

[0009] In one implementation, a system configured to assist in characterization of devices on a specimen includes means for obtaining measured metrology data from the devices, and means for determining at least one parameter of the devices based on the measured metrology data using a machine learning model that uses composite metrology data, where each composite metrology data includes an integration of measured metrology data for a reference device and first composite metrology data for a first model of the reference device.

[0010] In one implementation, a method for characterizing a device on a sample includes obtaining measured metrology data for a reference device of the device. The method further includes generating a first set of synthetic metrology data for a first model of the reference device and generating a second set of synthetic metrology data for a second model of the modified reference device that is modified relative to the first model. Composite metrology data is generated for the modified reference device. Each composite metrology data is generated by integrating the measured metrology data, the first synthetic metrology data, and the second synthetic metrology data, and at least the composite metrology data is stored as a training data set.

[0011] In one implementation, a computer system configured to assist in characterization of a device on a sample includes at least one processor configured to acquire measured metrology data for a reference device of the device. The at least one processor is further configured to generate a first set of synthetic metrology data for a first model of the reference device and generate a second set of synthetic metrology data for a second model of the modified reference device that is modified relative to the first model. The at least one processor is further configured to generate composite metrology data for the modified reference device. Each composite metrology data is generated by combining the measured metrology data, the first synthetic metrology data, and the second synthetic metrology data, and store at least the composite metrology data as a training data set.

[0012] In one implementation, a system configured to assist in characterization of a device on a sample includes means for acquiring measured metrology data for a reference device of the device. The system further includes means for generating a first set of synthetic metrology data for a first model of the reference device and means for generating a second set of synthetic metrology data for a second model of the modified reference device that has been modified relative to the first model. Composite metrology data is generated for the modified reference device. Each composite metrology data is generated by integrating the measured metrology data, the first synthetic metrology data, and the second synthetic metrology data, and at least the composite metrology data is stored as a training data set.

[0013] In one implementation, a method for characterizing a device on a sample includes acquiring measured optical metrology data for a reference device of the device and generating composite optical metrology data generated by integrating the measured optical metrology data for the reference device and first composite optical metrology data for a first model of the reference device, respectively. The method further includes training a machine learning model with a training dataset including at least the composite optical metrology data to characterize the device using the measured optical metrology data from the device.

[0014] In one implementation, a computer system configured to assist in characterization of a device on a specimen includes at least one memory configured to store measured optical metrology data and composite optical metrology data, and at least one processor coupled to the at least one memory. The at least one processor is configured to acquire measured optical metrology data for a reference device of the device and generate composite optical metrology data generated by integrating the measured optical metrology data for the reference device and first composite optical metrology data for a first model of the reference device, respectively. The at least one processor is further configured to train a machine learning model using a training dataset including at least the composite optical metrology data to characterize the device using the measured optical metrology data from the device.

[0015] In one implementation, a computer system for characterizing a device on a sample includes means for acquiring measured optical metrology data for a reference device of the device, and means for generating composite optical metrology data generated by integrating the measured optical metrology data for the reference device and first composite optical metrology data for a first model of the reference device, respectively. The computer system further includes means for training a machine learning model with a training dataset including at least the composite optical metrology data to characterize the device using the measured optical metrology data from the device. [Brief explanation of the drawings]

[0016] [Figure 1] FIG. 1 illustrates a schematic diagram of a metrology device that may be used to generate metrology data, as described herein, to generate reference data for training a machine learning model and / or to generate experimental data for characterizing parameters of a device structure under test using a trained machine learning model. [Figure 2] 1 illustrates a workflow for offline recipe creation and inline inference based on a training dataset that includes the composite metrology data discussed herein. [Figure 3] 1 is a graph illustrating parameter space for measurements using machine learning. [Figure 4A] 1 illustrates a workflow for generating composite metrology data based on integrating measured metrology data with synthetic metrology data for inclusion in training data. [Figure 4B] 1 illustrates a workflow for generating composite metrology data by combining synthetic metrology data calculated for a modified reference device with the variation between measured metrology data from the reference device and the synthetic metrology data calculated for the reference device. [Figure 4C]10 illustrates another workflow for generating composite metrology data by combining synthetic metrology data calculated for a modified reference device with the variation between measured metrology data from the reference device and the synthetic metrology data calculated for the reference device. [Figure 4D] 1 illustrates a workflow for generating composite metrology data by combining measured metrology data from a reference device with the variation between synthetic metrology data calculated for the reference device and synthetic metrology data calculated for a modified reference device. [Figure 4E] 1 illustrates another workflow for generating composite metrology data by combining measured metrology data from a reference device with the variation between synthetic metrology data calculated for the reference device and synthetic metrology data calculated for a modified reference device. [Figure 5] 1 illustrates a workflow for generating composite reference data based on integrating parameter values ​​from a reference device and parameter values ​​from models of the reference device and a modified reference device. [Figure 6A] 1 illustrates a graph showing a comparison of composite and measured spectral data for different Mueller matrix elements. [Figure 6B] 1 illustrates a graph showing a comparison of composite and measured spectral data for different Mueller matrix elements. [Figure 7A] 10 is a graph illustrating training and testing of a machine learning model trained using only measured spectra. [Figure 7B] 10 is a graph illustrating training and testing of a machine learning model trained using only measured spectra. [Figure 8A] 10 is a graph illustrating the training and testing of a machine learning model trained using composite measurement spectra and measured spectra. [Figure 8B]10 is a graph illustrating the training and testing of a machine learning model trained using composite measurement spectra and measured spectra. [Figure 9A] Illustrates examples of device structures having non-primary parameters that are strongly correlated with the primary parameters of the device structure being characterized. [Figure 9B] Illustrates examples of device structures having non-primary parameters that are strongly correlated with the primary parameters of the device structure being characterized. [Figure 10A] 1 illustrates a graph showing sample spectra and modeled spectra for several Mueller matrix elements. [Figure 10B] 10 illustrates a graph showing signals from off-diagonal Mueller matrix elements corresponding to asymmetry parameters. [Figure 11] FIG. 1 shows an illustrative flow diagram depicting example operations for assisting in characterizing devices on a specimen according to implementations described herein. [Figure 12] FIG. 1 shows an illustrative flow diagram depicting example operations for assisting in characterizing devices on a specimen according to implementations described herein. [Figure 13] FIG. 1 shows an illustrative flow diagram depicting example operations for assisting in characterizing devices on a specimen according to implementations described herein. DETAILED DESCRIPTION OF THE INVENTION

[0017] During the fabrication of semiconductor and similar devices, it is often necessary to monitor the fabrication process by non-destructively measuring the devices. One type of metrology that can be used for non-destructive measurement of samples during processing is optical metrology, which can use a single wavelength or multiple wavelengths and can include, for example, ellipsometry, reflectometry, Fourier transform infrared spectroscopy (FTIR), etc. Other types of metrology can also be used, including x-ray metrology, photoacoustic metrology, electron beam (E-beam) metrology, etc.

[0018] Optical metrology, such as thin film metrology and optical critical dimension (OCD) metrology, as well as other types of metrology, may use physical modeling of device structures, and may be referred to herein as OCD modeling, OCD modeling, or simply modeling or modeling. Simulated metrology data, e.g., simulated optical data, may be calculated for a physical model using rigorous coupled-wave analysis (RCWA), finite-difference time-domain (FDTD), or finite element method (FEM), or other similar techniques. Variable parameters in the physical model, such as layer thicknesses, line widths, space widths, sidewall angles, material properties, etc., may be adjusted, and simulated data calculated for each variation. Measured data from the device structure may be compared to the simulated data for each parameter variation, e.g., in a nonlinear regression process, until a good fit is achieved, at which point the fitted parameter values ​​are determined to be accurate representations of the device structure's parameters. However, techniques that use physical modeling have a high computational cost due to the calculations required to simulate metrology data, and as device structures become more complex, such techniques become less useful due to the time required to solve and limited model accuracy.

[0019] Machine learning is another technique that can be used to measure various parameters of device structures for process improvement, monitoring, and control. Unlike modeling, machine learning does not use physical models. Instead, training data is obtained from reference samples and used to train a machine learning model. The training data may include, for example, spectral signals and values ​​of structural parameters of interest from the reference samples. The machine learning model is automatically "trained" based on the training data to find relevant spectral features and learn the inherent relationships and connections between input and output features to make decisions and predictions on new data.

[0020] One of the challenges associated with metrology using machine learning is the robustness of machine learning training. For example, the amount of reference data available for training a machine learning model is typically very limited, inherently limiting the predictions that can be generated by the machine learning model. For example, training data can be obtained based on measurement data from at least one reference device. However, the measured data is limited by the number of available reference devices, and therefore there is a practical limit to the parameter variation from the reference device that can be present in the training data thus obtained. Furthermore, manufacturing process revisions can occur over time, for example, while developing a device structure, which can reduce the relevance of training data obtained from previously acquired measurements. Although additional training data can be obtained by generating and measuring additional reference devices based on the process revisions and a machine learning model retrained using the newly acquired training data, this process is inefficient due to the cost of measuring reference metrology data and the retraining time required. Additionally, the process of obtaining new training data and retraining the machine learning model may need to be performed iteratively for each process revision.

[0021] Another challenge is that some key parameters to be measured, such as overlay, may have low sensitivity. Therefore, problems with the training data may occur due to the signal size of the key parameters and / or systematic noise in the training dataset. Furthermore, for example, non-key parameters with stronger signals may be highly correlated with the key parameters. For example, overlay breaks the structural symmetry in the device, resulting in spectral responses in the off-diagonal elements of the Mueller matrix. However, the magnitude of these off-diagonal elements is small. Furthermore, in the presence of structural asymmetry, non-key parameters such as slope may be highly correlated with overlay. Therefore, measuring overlay using traditional spectral fitting with only an OCD model may be difficult. While machine learning can be used for measurement, the robustness of the machine learning recipe is closely related to the design of experiments (DOE) and the quality and quantity of the references for the training dataset. Common failure modes, for example, are related to process variations. If process variations are not included in the DOE and the affected parameters have a strong correlation with overlay, attributing the spectral response of such process variations to overlay will result in poor machine learning recipe behavior, leading to inaccurate predictions and the need for recipe rework.

[0022] Training data can be expanded using synthetic (simulated) data calculated from a physical model. For example, physical modeling techniques such as OCD modeling can be used to generate synthetic metrology data for variations in at least one parameter in the physical model, which can be used to increase the training sample parameter space for better process variation coverage. However, synthetic metrology data has limited usefulness as training data. For example, synthetic training data is not easily adaptable for use in applications with low sensitivity to key parameters, such as overlay measurements. For example, in applications where the key parameter signal is relatively small, the physical model must be extremely accurate to generate useful synthetic metrology data. Furthermore, measured data contains system noise that can vary depending on the system, wafer, environment, and measurement time. It is difficult to determine whether the synthetic metrology data generated by a physical model simulating a key parameter is accurate or overfitting the system noise. Additionally, synthetic metrology data does not include system noise. Therefore, if synthetic metrology data and measurement data are collected separately as training data, the machine learning model may not be able to extract correct measurement information from both and may treat one as an outlier.

[0023] As discussed herein, to overcome the limitations of using only measurement data or a collection of measurement data and separate synthetic measurement data as training data, composite measurement data can be generated and used as training data. The composite measurement data is a mixture of measured measurement data and synthetic measurement data. Thus, rather than simply collecting the measured measurement data and the synthetic measurement data separately in a training data set, the measured measurement data and the synthetic measurement data are integrated together to form the composite measurement data. The composite measurement data can be included in the training data set. In some implementations, the composite measurement data and the measured measurement data can be included in the training data set. For reference, the measurement data used to generate the composite measurement data can be described herein as optical data or, more specifically, spectral data, although it should be understood that other types of measurement data, such as X-ray data, photoacoustic data, and E-beam data, can be used.

[0024] The composite metrology data is generated using measured metrology data from the reference device, a first set of synthetic metrology data from a first model of the reference device (e.g., an accurate model after fitting to the measured metrology data), and a second set of synthetic metrology data from a second model of the modified reference device (e.g., at least one parameter of the first model is varied to generate the modified second model of the reference device). In some implementations, for example, the second set of synthetic metrology data may be modified based on a variation between the measured metrology data and the first set of synthetic metrology data, while in other implementations, the measured metrology data may be modified based on a variation between the first set of synthetic metrology data and the second set of metrology data.

[0025] In some implementations, at least one parameter of the first model that is varied to generate the second model of the modified reference device may include a key parameter that increases the parameter space for the training data set. Composite reference data associated with the composite measurement data may be generated, for example, by combining values ​​of the reference parameter, the fitted parameter, and the altered parameter, and used as a label in the training data set.

[0026] In some implementations, at least one parameter of the first model varied to generate the second model of the modified reference device may be a non-primary parameter. The non-primary parameter may be, for example, a parameter that is strongly correlated with the primary parameter. Thus, composite metrology data may be generated based on the measured metrology data and synthetic metrology data generated from the variation of the highly correlated non-primary parameter. The composite metrology data may be associated with the reference value of the primary parameter as a label in the training dataset.

[0027] Therefore, by using composite metrology data, machine learning recipes can tolerate a wide range of process variations without having to prepare actual wafers with extensive process coverage for machine learning recipe development.

[0028] FIG. 1 illustrates, by way of example, a schematic diagram of a metrology device 100 that may be used to generate metrology data to generate reference data for training a machine learning model and / or to generate experimental data for characterizing parameters of a device structure under test using a trained machine learning model, as described herein. While metrology device 100 is illustrated as an optical metrology device, other types of metrology devices, such as X-ray, photoacoustic, or E-beam metrology devices, may be used. Optical metrology device 100 may be configured to perform, for example, spectroscopic reflectance measurements, spectroscopic ellipsometry (including Mueller matrix ellipsometry), spectroscopic scatterometry, overlay scatterometry, interferometry, or FTIR measurements of a sample 101. Sample 101 may include, for example, at least one device structure to be measured. It should be understood that optical metrology device 100 is illustrated as one example of a metrology device, and other metrology devices, including normal incidence devices, non-polarized light devices, etc., may be used if desired.

[0029] The measurement device 100 includes a source 110 that generates radiation incident on the sample. The measurement device 100 in FIG. 1 is illustrated as an optical measurement device 100, and thus the source is a light source 110 that generates light 102; however, other types of measurement devices may use sources that generate other types of radiation, such as X-rays or E-beams. The light 102 generated by the light source 110 may include ultraviolet-visible light having a variety of wavelengths, i.e., a continuous wavelength range or multiple discrete wavelengths, such as wavelengths between 200 nm and 1000 nm, or may be a single wavelength. The optical measurement device 100 includes collection optics 120 and 130 that collect and receive the light and direct the light so that it is obliquely incident on the top surface of the sample 101. The optics 120 and 130 may be refractive, reflective, or a combination thereof, and may be objective lenses.

[0030] The reflected light may be collected by lens 114 and received by detector 150. Detector 150 may be a conventional charge coupled device (CCD), photodiode array, CMOS, or similar type of detector. Detector 150 may be a spectrometer, for example, if broadband light is used, and detector 150 may generate a spectral signal as a function of wavelength. A spectrometer may be used to disperse the entire spectrum of polarized light into spectral components across an array of detector pixels. One or more polarizing elements may be present in the beam path of optical metrology device 100. For example, optical metrology device 100 may include one or both (or none) of one or more polarizing elements 104 in the beam path before sample 101 and a polarizing element (analyzer) 112 in the beam path after sample 101, and one or more additional elements 105a and 105b, such as compensators or photoelastic modulators, which may be before, after, or both the sample 101 and the analyzer. By using a spectroscopic ellipsometer that uses a double rotating compensator between the polarizing elements 104 and 112 and the sample, the complete Mueller matrix can be measured.

[0031] The optical metrology device 100 further includes one or more computing systems 160 configured to perform measurements of at least one parameter of the sample 101 using the methods described herein. The one or more computing systems 160 are coupled to the detector 150 to receive metrology data acquired by the detector 150 during measurements of the structure of the sample 101. The data acquisition may be to generate reference data from one or more reference devices to train a machine learning model and / or to generate experimental data from a device under test to characterize at least one parameter of the device. The one or more computing systems 160 may be, for example, a workstation, a personal computer, a central processing unit, or other suitable computer system, or multiple systems. The one or more computing systems 160 may be configured to perform optical metrology, for example, based on spectral processing according to the methods described herein, or based on processing of any other desired type of optical metrology data or other metrology data.

[0032] It should be understood that the one or more computing systems 160 may be a single computer system or multiple separate or linked computer systems, and may be referred to interchangeably herein as computing system 160, at least one computing system 160, or one or more computing systems 160. The computing system 160 may be included in, connected to, or otherwise associated with the optical metrology device 100. Different subsystems of the optical metrology device 100 may each include a computing system configured to perform steps associated with the associated subsystem. For example, the computing system 160 may control the positioning of the specimen 101, for example, by controlling the movement of a stage 109 coupled to the chuck. For example, the stage 109 may be capable of horizontal movement in either Cartesian (i.e., X and Y) coordinates, or polar (i.e., R and θ) coordinates, or some combination of the two. The stage may also be capable of vertical movement along a Z coordinate. The computing system 160 may further control the operation of the chuck 108 to retain or release the specimen 101. The computing system 160 may further control or monitor the rotation of one or more polarizing elements 104, 112, or elements 105a and 105b, which may be compensators or photoelastic modulators, or the like.

[0033] Computing system 160 may be communicatively coupled to detector 150 in any manner known in the art. For example, one or more computing systems 160 may be coupled to a separate computing system associated with detector 150. Computing system 160 may be configured to receive and / or acquire metrology data or information from one or more subsystems of optical metrology device 100, such as detector 150 and polarizing elements 104, 112, and controllers of elements 105a, 105b, via a transmission medium that may include wired and / or wireless portions. Thus, the transmission medium may serve as a data link between computing system 160 and other subsystems of optical metrology device 100.

[0034] Computing system 160 includes at least one processor 162 with memory 164, as well as a user interface (UI) 168, communicatively coupled via bus 161. Memory 164 or other non-transitory computer-usable storage medium includes computer-readable program code 166 embodied therein that can be used by one or more computing systems 160 to control optical metrology device 100 and perform functions, as described herein, including any one or more of generating composite metrology data, training a machine learning model using the composite metrology data, or characterizing parameters of a device structure using a machine learning model trained on the composite metrology data. For example, as illustrated, memory 164 can include instructions to cause processor 162 to perform both physical modeling and machine learning (ML) as discussed herein. Data structures and software code for automatically implementing one or more acts described in this detailed description can be implemented by one of ordinary skill in the art in light of this disclosure and stored in a computer-usable storage medium, e.g., memory 164, which can be any device or medium capable of storing code and / or data for use by a computer system such as computing system 160. Computer-usable storage media can include, but are not limited to, read-only memory, random-access memory, magnetic and optical storage devices such as disk drives and magnetic tape. Additionally, functionality described herein can be embodied in whole or in part within the circuitry of an application specific integrated circuit (ASIC) or programmable logic device (PLD), and functionality can be embodied in a computer-understandable descriptor language that can be used to create an ASIC or PLD that operates as described herein.

[0035] The computing system 160 may be configured to acquire optical metrology data from multiple measurement sites on one or more samples, e.g., calibration samples or training samples. Each measurement site may include a device structure, and the optical metrology data may include spectral signals or other types of optical metrology data. The computing system 160 may be configured to generate metrology solutions for optical measurements of device structures as discussed herein, including using modeling and machine learning to generate composite metrology data for training a machine learning model. In some implementations, different computing systems and / or different optical metrology devices may be used to acquire reference data from training samples, generate composite metrology data, train a machine learning model using the composite metrology data, or characterize at least one parameter of a device structure under test using a machine learning model trained with the composite metrology data. The reference data, a training set of data including composite metrology data, or a trained machine learning model may be provided to the computing system 160 via computer-readable program code 166 on a non-transitory computer-usable storage medium, such as memory 164.

[0036] The reference data, a training set of data including composite metrology data, or a trained machine learning model may be stored in memory 164. Results of further analysis of the data, for example, to characterize parameters of the device structure under test, may be reported and stored, for example, in memory 164 associated with the specimen 101 and / or presented to a user via the UI 168, an alarm, or other output device. Additionally, the results of the analysis may be reported and fed back or fed back to process equipment to adjust appropriate manufacturing steps to correct any variations detected in the manufacturing process. The computing system 160 may include a communications port 169, which may be any type of communications connection, such as, for example, to the Internet or any other computer network. The communications port 169 may be used to receive instructions used to program the computing system 160 to perform any one or more of the functions described herein and / or to export measurement results and / or signals with instructions to another system, such as an external process tool, in a feedforward or feedback process to adjust process parameters associated with the specimen fabrication processing steps based on the measurement results.

[0037] FIG. 2 illustrates, by way of example, a workflow 200 for offline recipe creation and inline inference based on a training data set as discussed herein. As illustrated, during the learning phase 210, a training set of metrology data 212 is provided. For example, the training set of metrology data 212 may include composite metrology data generated based on a combination of theoretically calculated metrology data (synthetic metrology data) and experimentally measured metrology data (measured metrology data) acquired from one or more reference structures using the optical metrology device 100, as discussed herein. For example, each composite metrology data is a combination of measured metrology data and synthetic metrology data. The synthetic metrology data may cover a wider range of process condition variations than the measured metrology data. The training set of metrology data 212 may further include experimentally measured data, for example, metrology data measured from one or more reference structures. At least a portion of the training set of metrology data 212 may be labeled based on key parameter reference values.

[0038] As illustrated, training 214 is performed based on a training set of measurement data 212. For example, training may include feature extraction, regression, classification, and other techniques for generating machine learning model 216. For example, training 214 may use any desired machine learning algorithm that may use composite measurement data as at least a portion of the training set of measurement data 212. For example, the machine learning algorithm may be a supervised algorithm or an unsupervised learning algorithm.

[0039] The machine learning model 216 trained using a training dataset including composite metrology data may be used to measure at least one parameter of a device under test, as discussed herein. By using the composite metrology data to train the machine learning model 216, the machine learning model 216 may be better able to tolerate various process variations in the training set of metrology data 212 than would be the case with only measured metrology data, only purely synthetic metrology data, or a combination of measured and purely synthetic metrology data. In some implementations, the at least one parameter that may be measured using the machine learning model 216 may include a key parameter with low sensitivity, such as overlay, because the composite metrology data reduces or eliminates issues due to signal size of the key parameter and system noise in the training dataset.

[0040] During the inference stage 220, test data 222 (e.g., optical metrology data) is acquired from the device structure under test, for example, using the optical metrology device 100 shown in FIG. 1 . The test data 222 can be the same type of metrology data used in the training set of metrology data 212, for example. Using the test data 222 as input data, a trained machine learning model 216 can be used to infer or predict (224) and output (226) values ​​of desired features, e.g., parameters of interest, of the device structure under test. Because the machine learning model 216 is trained on composite metrology data that is subject to wider variations in process conditions, the model predictions may have better metrology performance in the case of a process revision if the composite metrology data better represents the experimental metrology data after the process revision.

[0041] FIG. 3 is a graph 300 illustrating a parameter space for metrology using machine learning. The parameter space illustrated in graph 300 includes two parameters, parameter 1 along the X-axis and parameter 2 along the Y-axis. As illustrated using solid circles in graph 300, a reference target with parameter variations is measured to generate several training samples 312 that define a training range 310 in the parameter space. Measurements of devices under test with parameters within the training range 310 are relatively accurate, but accuracy decreases when parameters fall outside the training range 310. For example, as illustrated using open circles, a test sample 322 generated by a measurement device with parameters outside the training range 310 results in decreased accuracy. For example, devices with parameters outside the training range 310 may be generated due to manufacturing process variations that occur after the initial training sample 312 is obtained.

[0042] To extend the training range, additional training data may be acquired. As illustrated by the gray circle, a composite training sample 332 may be acquired, thereby generating an extended training range 330. As illustrated, the extended training range 330 includes the test sample 322 and is therefore sufficient to generate acceptable accuracy for the test sample 322. Each composite training sample 332 is generated based on a combination of measured metrology data and synthetic metrology data. The synthetic metrology data may be generated, for example, by including some perturbation in the reference data, e.g., the training sample 312. The synthetic metrology data is then combined with the reference data, e.g., the training sample 312, to generate a composite training sample 332 having parameters outside the training range 310. As a result, the parameter space defined by the composite training sample is extended from the initial training range 310.

[0043] The machine learning model may be trained using a training data set that includes composite data, i.e., both the training range 310 and the extended training range 330, and is therefore exposed to a wider variation in process conditions than if it were trained only with the training range 310. Thus, predictions from the training machine learning model may have better metrology performance in the event of a process revision, for example, when the composite metrology data is a better representation of the measurement data after the process revision.

[0044] 4A illustrates a workflow 400 for generating composite metrology data based on integrating measured metrology data and synthetic metrology data for inclusion in training data. The composite metrology data is generated based on a combination of the measured metrology data and the synthetic metrology data. For reference, the metrology data discussed herein, e.g., composite metrology data, measured metrology data, and synthetic metrology data, are sometimes described as optical metrology data, particularly spectral data, although it should be understood that other types of metrology data may be used.

[0045] As illustrated in block 410, measured metrology data is obtained from a reference device having a first set of parameters, which may represent process condition A1 (experimental spectrum A1). The reference device may, for example, be produced using the same manufacturing process (process condition A1) as the device being tested, but at least one parameter may be intentionally altered from its nominal value. The measured metrology data from block 410 may, for example, be obtained from the reference device by optical metrology device 100 shown in FIG. 1.

[0046] In block 420, the measured metrology data (experimental spectrum A1) is fitted to a physical model. For example, a first model, such as an OCD model, can be generated based on the values ​​of at least one parameter of a reference device, which value can be known based on, for example, CD-SEM or other similar types of measurements of the reference structure or other similar reference structures. Calculated metrology data is generated for the physical model using a modeling technique, such as RCWA, FDTD, or FEM, that fits the measured metrology data. For example, various parameters of the physical model can be adjusted, and the calculated metrology data generated for each variation is compared to the measured metrology data, for example, in a nonlinear regression process, until a good fit is achieved. Once a good fit is achieved, the values ​​of the fitted parameters for the physical model are considered to be accurate representations of the parameters of the reference device.

[0047] Calculated metrology data corresponding to the best fit with the measured metrology data from block 420 is generated in block 422 as a first set of synthetic metrology data for a reference device having a first set of parameters representing process condition A1 (e.g., calculated spectrum A1).

[0048] Additionally, a second model is generated that incorporates changes to the first model to simulate the modified reference device, as illustrated in block 430. For example, at least one parameter of the first model from block 420 may be changed to generate the second model. For example, the change to the at least one parameter for the second model may be greater than any of the parameter variations found in the reference device representing process condition A1. In some implementations, the at least one parameter changed for the second model may be a geometric parameter and optical constants that are changed to mimic changes resulting from the expected revised process condition B1, which may be obtained from user input or through process simulation, as illustrated in block 425.

[0049] For example, OCD model parameters that are susceptible to variations due to changing process conditions may be altered by some perturbation, which may be defined for a parameter P as follows: P Perturbed =P OCD_Fit +random_perturbation(P OCD_Fit ) Equation 1

[0050] where P Perturbed is the parameter P after perturbation, and P OCD_Fit is the result of the parameter P after fitting the OCD model to the experimental spectrum A1 (block 420). OCD_Fit The random perturbation as a function of can be defined as follows: random_perturbation(P OCD_Fit )=B+coefficient * std_dev(P OCD_Fit ) * rand_num expression 2

[0051] where B corresponds to a bias term to incorporate design of experiments (DOE) conditions, coefficient is a multiplication coefficient that controls the variance of the random perturbation and can be, for example, 0.1 to 10, std_dev is the standard deviation, and rand_num is a random number that can be, for example, in the range of -0.5 to 0.5, although any desired range can be used, e.g., -1 to 1, 0.1 to 10, etc.

[0052] The one or more parameters that are varied can be primary parameters, non-primary parameters, or both primary and non-primary parameters. In some implementations, the primary parameters of the model can be held static, and only the non-primary parameters in the physical model can be varied. For example, if the primary parameter is overlay, the same overlay value can be used in the first and second models, but other non-primary parameters can be varied. The non-primary parameters can be strongly correlated with the primary parameters. For example, in some device structures, the slope parameter of the bit lines in the structure can be strongly correlated with overlay, e.g., a variation in slope produces a change in the optical metrology data similar to a variation in overlay. Other examples of parameters that can be correlated to overlay can be layer thicknesses or critical dimensions.

[0053] In block 432, calculated metrology data for the second model may be generated as a second set of synthetic metrology data for the modified reference device having a second set of parameters using a modeling technique such as, for example, RCWA, FDTD, FEM, etc., where the second set of parameters may represent process condition B1 (e.g., calculated spectrum B1).

[0054] As illustrated, composite measurement data (composite spectrum) can be generated based on a combination of the measured measurement data (experimental spectrum A1) from block 410, the first set of composite measurement data (calculated spectrum A1) from block 422, and the second set of composite measurement data (calculated spectrum B1) from block 432. As an example, in some implementations, a variation between the measured measurement data (experimental spectrum A1) and the first set of composite measurement data (calculated spectrum A1) can be determined, and then this variation can be combined with the second set of composite measurement data (calculated spectrum B1) to generate composite measurement data (composite spectrum). The variation can be, for example, a discrepancy between the measured measurement data and the composite measurement data, or a spectral transformation between the measured measurement data and the composite measurement data. In other implementations, a variation between the first set of composite measurement data (calculated spectrum A1) and the second set of composite measurement data (calculated spectrum B1) can be determined, and then this variation can be combined with the measured measurement data (experimental spectrum A1) to generate composite measurement data (composite spectrum). The variation between the first set of composite measurement data and the second set of composite measurement data can be, for example, a difference between the composite measurement data.

[0055] Thus, the resulting composite measurement data at block 450 is similar to the measured measurement data from block 410, but includes a change in the parameter space that extends the training range, as illustrated, for example, by the extended training range 330 of FIG. 3.

[0056] The process for generating the composite measurement data illustrated in FIG. 4A can be performed on each individual measured measurement data, for example, the measured measurement data obtained from different reference devices having different parameter values, and can be repeated for different variations of the parameters to generate multiple composite measurement data. The composite measurement data, and in some implementations, the measured measurement data can be combined to form the training set of measurement data 212 (shown in FIG. 2).

[0057] Thus, the composite metrology data is based on measured metrology data, but also includes simulated variations in the metrology data introduced by variations in at least one parameter. Thus, the machine learning model 216 (shown in FIG. 2 ) is trained using the composite metrology data to expand the parameter space, thereby increasing its usefulness. Furthermore, the synthetic information combined with the measured metrology data can be used to help the machine learning model break down metrology data correlations for non-primary parameters, but does not create any new primary parameter data. Thus, the resulting machine learning model 216 trained using the composite metrology data is robust and can tolerate a wide variety of process variations without requiring large amounts of metric data.

[0058] 4B illustrates a workflow 402 for an implementation of generating composite metrology data based on combining measured metrology data and synthetic metrology data to generate training data. Workflow 402 is similar to workflow 400 shown in FIG. 4A, and similarly designated elements are the same.

[0059] 4A , measured metrology data is obtained from a reference device having a first set of parameters that may represent process condition A1 (experimental spectrum A1) in block 410. The measured metrology data (experimental spectrum A1) is fitted to a physical model in block 420, and calculated metrology data corresponding to the best fit with the measured metrology data is generated in block 422 as a first set of synthetic metrology data for the reference device having the first set of parameters that represent process condition A1 (e.g., calculated spectrum A1).

[0060] In block 430, a second model is generated simulating a modified reference device, for example, in which one or more parameters have been changed relative to the parameters used in the first model from block 420, and calculated measurement data is generated for the second model as a second set of synthetic measurement data for the modified reference device having the second set of parameters, which may represent process condition B1 (e.g., calculated spectrum B1) in block 432.

[0061] 4B, the variation between the measured metrology data (experimental spectrum A1) in block 410 and the first set of synthetic metrology data (calculated spectrum A1) from block 422 can be determined as a mismatch between the measured metrology data (experimental spectrum A1) and the first set of synthetic metrology data (calculated spectrum A1). For example, the mismatch spectrum can be determined as follows: Discrepancy spectrum = Experimental spectrum A1 - OCD best fit spectrum (process condition A1) Equation 3

[0062] Here, the OCD best fit spectrum is the calculated spectrum A1.

[0063] The discrepancy from block 430 based on the measured metrology data from block 410 and the first set of composite metrology data from block 422 may be added to the second set of composite metrology data from block 432 in block 450 to generate composite metrology data representative of process condition B1.

[0064] 4C illustrates a workflow 404 for an implementation of generating composite metrology data based on integrating a combination of measured metrology data and synthetic metrology data to generate training data. Workflow 404 is similar to workflow 400 shown in FIG. 4A, and similarly designated elements are the same.

[0065] 4A , measured metrology data is obtained from a reference device having a first set of parameters that may represent process condition A1 (experimental spectrum A1) in block 410. The measured metrology data (experimental spectrum A1) is fitted to a physical model in block 420, and calculated metrology data corresponding to the best fit with the measured metrology data is generated in block 422 as a first set of synthetic metrology data for the reference device having the first set of parameters that represent process condition A1 (e.g., calculated spectrum A1).

[0066] In block 430, a second model is generated simulating a modified reference device, for example, in which one or more parameters have been changed relative to the parameters used in the first model from block 420, and calculated measurement data is generated for the second model as a second set of synthetic measurement data for the modified reference device having the second set of parameters, which may represent process condition B1 (e.g., calculated spectrum B1) in block 432.

[0067] 4C , the variation between the measured metrology data (experimental spectrum A1) in block 410 and the first set of synthetic metrology data (calculated spectrum A1) from block 422 can be determined as a spectral transformation between the measured metrology data (experimental spectrum A1) and the first set of synthetic metrology data (calculated spectrum A1). For example, a mathematical model can be generated to transform the first set of synthetic metrology data (calculated spectrum A1) into the measured metrology data (experimental spectrum A1) for process condition A1.

[0068] The transformation determined in block 442, based on the measured metrology data from block 410 and the first set of composite metrology data from block 422, may be applied to the second set of composite metrology data from block 432 to generate composite metrology data representative of process condition B1 in block 450.

[0069] 4D illustrates a workflow 406 for an implementation of generating composite metrology data based on combining measured metrology data and synthetic metrology data to generate training data. Workflow 406 is similar to workflow 400 shown in FIG. 4A, and similarly designated elements are the same.

[0070] 4A , measured metrology data is obtained from a reference device having a first set of parameters that may represent process condition A1 (experimental spectrum A1) in block 410. The measured metrology data (experimental spectrum A1) is fitted to a physical model in block 420, and calculated metrology data corresponding to the best fit with the measured metrology data is generated in block 422 as a first set of synthetic metrology data for the reference device having the first set of parameters that represent process condition A1 (e.g., calculated spectrum A1).

[0071] In block 430, a second model is generated simulating a modified reference device, for example, in which one or more parameters have been changed relative to the parameters used in the first model from block 420, and calculated measurement data is generated for the second model as a second set of synthetic measurement data for the modified reference device having the second set of parameters, which may represent process condition B1 (e.g., calculated spectrum B1) in block 432.

[0072] 4D , the variation between the first set of synthetic measurement data (calculated spectrum A1) from block 422 and the second set of synthetic measurement data (calculated spectrum B1) from block 432 may be determined as a spectral difference. For example, the spectral difference may be determined as follows: Spectral difference = OCD best fit spectrum (process condition A1) - OCD spectrum (process condition B1) Equation 4

[0073] Here, the OCD best fit spectrum is calculated spectrum A1, and the OCD spectrum is calculated spectrum B1.

[0074] The spectral difference from block 444 based on the first set of synthetic metrology data from block 422 and the second set of synthetic metrology data from block 432 may be added to the measured metrology data from block 410 in block 450 to generate composite metrology data representative of process condition B1.

[0075] FIG. 4E illustrates a more detailed view of the workflow 406 shown in FIG. 4D for an implementation of generating composite metrology data based on integrating measured metrology data and synthetic metrology data to generate training data, where spectral differences between a first set of synthetic metrology data and a second set of synthetic metrology data are used.

[0076] As illustrated in FIG. 4E, in block 422, a first set of synthetic metrology data (calculated metrology data) is obtained. The calculated metrology data from block 422 may be obtained, for example, based on fitting the measured metrology data from block 410 to a first physical model (illustrated in block 420 of FIG. 4D). A physical model, e.g., an OCD model, may be used to generate a physical model based on values ​​of key parameters (and any other parameters) of the reference structure, which values ​​may be known based on CD-SEM or other similar types of measurements of the reference structure. The calculated metrology data is generated for the physical model and fitted to the measured data in a nonlinear regression process until a good fit is achieved, at which point the values ​​of the fitted parameters for the physical model are determined to be accurate representations of the parameters of the reference structure.

[0077] As illustrated, in blocks 432a, 432b, 432c, and 432d (sometimes collectively referred to as block 432), a second set of synthetic metrology data (calculated metrology data) is generated, these data being for variations in one or more non-primary parameters in a second physical model (illustrated in block 430 of FIG. 4D ), i.e., using the same reference values ​​for the primary parameters used in the first model. For example, if the primary parameter is overlay, the same overlay value is used in the first physical model to generate the calculated metrology data in block 422 and in the second physical model to generate the calculated optical metrology data in block 432. The calculated metrology data in block 432 may be generated for the second physical model using the same modeling technique, e.g., RCWA, FDTD, FEM, etc., as used to generate the calculated metrology data in block 422. In some implementations, the non-primary parameters varied to generate the calculated metrology data in block 432 may be strongly correlated with the primary parameters. As in the example discussed above, in some device structures, the slope parameter of the bit lines within the structure may be strongly correlated with overlay; for example, variations in slope will produce similar changes in metrology data as variations in overlay. Other examples of parameters that may be correlated to overlay may be layer thickness or critical dimensions. Figure 4E illustrates, by way of example, calculated metrology data from blocks 432a, 432b, 432c, and 432d determined for four variations of non-primary parameters having values ​​of -0.2, -0.1, +0.1, and +0.2, respectively. It should be understood that additional or fewer variations may be used, and the values ​​of the variations may differ and may depend on the type of non-primary parameters varied and the number of variations.

[0078] Differences between the first set of synthetic metrology data from the first model from block 422 and the second set of synthetic metrology data from block 432 are calculated to generate differences (I, II, III, and IV) in blocks 444a, 444b, 444c, and 444d (sometimes collectively referred to as block 444). The differences determined in block 444 represent changes in the metrology data due to variations in non-primary parameters (the values ​​of the primary parameters are fixed).

[0079] As illustrated, the measured metrology data from block 410 may be modified based on the difference from block 444 to generate composite metrology data for each variation in the non-primary parameters in blocks 450a, 450b, 450c, and 450d (sometimes collectively referred to as block 450). For example, each of the difference from block 444 may be added to the measured metrology data from block 410, which is the measured metrology data used in the fitting to generate the calculated metrology data from block 422, and which has the same reference values ​​for the primary parameters as used in the first set of composite metrology data from block 422 and the second set of composite metrology data from block 432. Thus, the resulting composite metrology data in block 450 is similar to the measured metrology data from block 410 because it has the same reference values ​​for the primary parameters and system noise, but includes changes in the metrology data due to variations in the non-primary parameters.

[0080] 4D and 4E may be performed, for example, for each separate measured metrology data having different primary parameter values ​​to generate multiple composite metrology data for each measured metrology data. The composite metrology data, and in some implementations, the associated measured metrology data, may be combined to form a training set of metrology data 212 (shown in FIG. 2).

[0081] Thus, the composite metrology data is based on measured metrology data but includes simulated differences in the metrology data introduced by variations in non-primary parameters, which, in some implementations, may be highly correlated to the primary parameters. Therefore, the machine learning model 216 (shown in FIG. 2 ) is trained using composite metrology having the same primary parameter reference values ​​as the measured metrology data. Composite metrology data incorporating synthetic information into the measurement information can be used to help the machine learning model overcome metrology data correlations for non-primary parameters, but does not create any new primary parameter data. Therefore, physical models do not need to be fitted to the primary parameters, which reduces or eliminates issues associated with signal sensitivity associated with primary parameters and the presence of system noise in the reference data but not in the synthetic metrology data. Thus, the resulting machine learning model 216 trained using composite data is robust and can tolerate a wide range of process variations without requiring large amounts of metric data.

[0082] 5 illustrates a workflow 500 for an implementation of generating composite reference data based on integrating parameter values ​​from a reference device and parameter values ​​from models of the reference device and a modified reference device. Workflow 500 generates composite reference data that can be included in training data, e.g., the training set of measurement data 212 in FIG. 2, along with composite measurement data, when key parameters are varied in the second model, as discussed with reference to workflows 400, 402, and 404 in FIG. 4A, 4B, and 4C, respectively. The composite reference data generated by workflow 500 may not necessarily be included in training data when key parameters are not varied in the second model, as discussed with reference to workflow 406 in FIG. 4D and 4E.

[0083] 4A, measured metrology data is obtained from a reference device having a first set of reference parameters that may represent process condition A1 (experimental spectrum A1). The reference parameters from the reference device are illustrated in block 512 (reference A1). The values ​​of the reference parameters may be known, for example, based on a CD-SEM or other similar type of measurement of the reference structure or other similar reference structures.

[0084] In block 520, the measured metrology data (experimental spectrum A1) is fitted to a physical model, which may be the same process as performed in block 420 of FIG. 4B, for example, by fitting the measured metrology data to the calculated metrology data and adjusting the parameters of the physical model using a nonlinear regression process until a good fit is achieved. Once a good fit is achieved, the fitted parameter values ​​for the physical model are considered to be an accurate representation of the parameters of the reference device. The resulting parameter values ​​from the fitting process in block 520 are generated in block 522 as a first set of primary parameter values ​​for the first model of the reference device (fitted primary parameter values ​​for A1).

[0085] As illustrated in block 530, a second model is generated with changes to the primary parameters for the first model to simulate the modified reference device by changing the values ​​of one or more primary parameters of the first model from block 520, the changes being based, for example, on application of changes representing estimated process variations for the device, which may be obtained from user input or through process simulation, as illustrated in block 525, which may represent process conditions B1. It is understood that non-primary parameters may be similarly changed for the first model representing possible process variations for the device. The generation of the second model generated in block 530 may be the same process as performed in block 430 of FIG. 4B. In block 532, parameter values ​​for the second model are generated as a second set of primary parameter values ​​for the second model (changed primary parameter values ​​for B1).

[0086] At block 540, a primary parameter bias is determined based on the first set of primary parameter values ​​from block 522 and the second set of primary parameter values ​​from block 532. For example, a parameter bias is the difference between an expected parameter value and an actual parameter value. Thus, at block 540, a primary parameter bias may be determined as the difference between the first set of primary parameter values ​​from block 522 and the second set of primary parameter values ​​from block 532. Next, at block 550, the primary parameter bias is combined with the reference parameters from block 512 to generate a composite reference. The composite reference data may be used as labels for process B1 in the training set of data.

[0087] 6A and 6B are graphs illustrating, by way of example, a comparison of composite and measured spectral data for several Mueller matrix elements (M33 and M34, respectively). The "Experimental Training" curves are experimental spectra generated from the reference under process condition A, and the "Experimental Test" curves are experimental spectra generated from the reference under process condition B. The "Composite" curves are composite spectral data generated after applying changes representing the changes between process condition A and process condition B, as discussed herein.

[0088] FIG. 7A is a graph illustrating, by way of example, the performance during training of a machine learning model using only measured reference optical metrology data (spectra), and illustrating training and testing (validation) of the trained model. FIG. 7B illustrates a blind test of the machine learning model from FIG. 7A. As illustrated, the blind test results in accuracy (R 2 ) is 0.763, the slope is 0.609, and the root means square error (RMSE) is 0.53.

[0089] In contrast, FIG. 8A is a graph illustrating the performance during training of a machine learning model using composite optical metrology data (spectra) and measured reference optical metrology data (spectra), illustrating training and testing (validation) of the trained model. FIG. 8B illustrates a blind test of the machine learning model from FIG. 8A. As illustrated, the blind test yielded higher accuracy than shown in FIG. 7B, with a higher precision (R 2 ) is 0.91, the slope is 0.991, and the root mean square error (RMSE) is 0.33.

[0090] As discussed above, composite measurement data can be useful for measuring key parameters with low sensitivity, such as overlay, because the composite measurement data reduces or eliminates problems caused by the signal size of the key parameters and system noise in the training dataset. Overlay, for example, breaks structural symmetry and results in spectral responses in the off-diagonal elements of the Mueller matrix. However, the magnitude of these off-diagonal elements is small. Furthermore, when structural asymmetry exists, some parameters may be highly correlated with overlay.

[0091] FIG. 9A illustrates, by way of example, an example device structure 900 having key parameters with low sensitivity that can be measured using, for example, the metrology techniques discussed herein. The device structure 900 is illustrated as a simplified DRAM gate including multiple bit lines 902 aligned with an underlying structure 904. For example, FIG. 9A illustrates with dotted lines the process expectations of overlay and tilt for the bit lines 902. FIG. 9B illustrates, by way of example, an example device structure 950 similar to the device structure 900 but with a non-zero overlay (OVL) error and non-zero tilt of the bit lines 952. For example, the overlay error OVL can be a key parameter for characterizing a device structure via optical metrology. The tilt of the bit lines 952, or other parameters such as critical dimensions or thickness, can be non-key parameters but can be strongly correlated with the overlay error OVL. In other words, variations in non-primary parameters such as tilt can strongly affect the same optical metrology data used to determine the overlay error OVL, making it difficult to measure overlay.

[0092] FIG. 10A is a graph illustrating, by way of example, sample and modeling spectra for several Mueller matrix elements, namely, M12, M13, M14, M22, M23, M24, M33, M34, and M44, for which good fits between the sample and model spectra have been achieved using conventional physical modeling techniques for measuring thickness and OCD parameters. In contrast, FIG. 10B illustrates asymmetric signals from off-diagonal Mueller matrix elements corresponding to overlay and tilt. As can be seen in FIG. 10B, the sensitivity of the signal to overlay and tilt is significantly lower than the fit that can be achieved by physical modeling techniques, as illustrated in FIG. 10A.

[0093] 2 may be trained using a training dataset that includes composite metrology data for measurements of key parameters with low sensitivity, such as overlay, because the composite metrology data reduces or eliminates issues due to signal size of the key parameters and system noise in the training dataset, as discussed herein. Thus, by using composite metrology data, the machine learning model 216 can tolerate various process variations better than if purely synthetic metrology data were used along with the measurement data in the training set of metrology data 212.

[0094] 11 shows an illustrative flow diagram depicting example operations 1100 for assisting in characterization of devices on a specimen, according to some implementations. In some implementations, the example operations 1100 may be performed by a metrology device, such as metrology device 100, having one or more processors, such as processor 162 in computing system 160 of FIG.

[0095] One or more processors may acquire 1102 measured metrology data from the device. For example, the means for acquiring measured metrology data from the device may be metrology device 100 and may interface with processor 162 in computing system 160 shown in FIG. 1. The measured metrology data may be, for example, test data 222 shown in FIG. 2.

[0096] The one or more processors may determine at least one parameter of the device based on the measured measurement data using a machine learning model that uses composite measurement data. Each composite measurement data may include, for example, an integration of measurement data measured for a reference device and first synthetic measurement data for a first model of the reference device (1104), as discussed with respect to the generation of training data illustrated in the learning stage 210 and inference stage 220 shown in FIG. 2 and workflows 400, 402, 404, and 406 of FIGS. 4A-4E. A means for determining at least one parameter of the device based on the measured measurement data using a machine learning model that uses composite measurement data, each including an integration of measurement data measured for the reference device and first synthetic measurement data for the first model of the reference device, may be, for example, measurement device 100 shown in FIG. 1, including a computing system 160 having a processor 162 configured to execute a machine learning model according to computer-readable program code 166. In some optional implementations, as illustrated by the dashed lines, each composite measurement data may include an integration of measurement data with first synthetic measurement data, as discussed, for example, with respect to the generation of training data illustrated in the learning stage 210 and inference stage 220 shown in FIG. 2 and workflows 400, 402, 404, and 406 in FIGS. 4A-4E, and may further include second synthetic measurement data relating to a second model of the modified reference device that is altered relative to the first model (1106).

[0097] The one or more processors may provide, e.g., report, at least one parameter of the device to characterize the device on the sample. For example, the means for providing at least one parameter of the device to characterize the device on the sample may be the metrology device 100 and may interface with the processor 162 and memory 164, and the UI 168 in the computing system 160 shown in FIG.

[0098] In some implementations, the second model of the modified reference device has at least one parameter that varies relative to the first model, for example, as discussed in block 430 of FIGS. 4A-4E.

[0099] In some implementations, the machine learning model may further use composite reference parameters for the modified reference device based on a combination of the reference parameters of the reference device, the first set of primary parameter values ​​generated for the first model, and the second set of primary parameter values ​​generated for the second model, for example, as discussed in blocks 540, 512, and 550 of FIG. 5 .

[0100] In some implementations, the first model may be generated by fitting measured metrology data from a reference device to the synthetic optical metrology data of the first model, for example, as discussed in block 420 of Figures 4A-4E.

[0101] In some implementations, a second model of the modified reference device may be generated by changing at least one parameter of the first model, for example, as discussed in block 430 of Figures 4A-4E.

[0102] In some implementations, second synthetic metrology data for a second model of the modified reference device may be generated based on the variation between the measured metrology data from the reference device and the first synthetic metrology data for the first model, for example, as discussed with reference to Figure 4A and illustrated in blocks 440, 442, and 450 of Figures 4B and 4C. The variation between the measured metrology data from the reference device and the first set of synthetic metrology data for the first model may be, for example, a mismatch between the metrology data and the first synthetic metrology data, as discussed in block 440 of Figure 4B, or a spectral transformation between the metrology data and the first synthetic metrology data, as illustrated in block 442 of Figure 4C.

[0103] In some implementations, the composite measurement data may be generated by correcting the measurement data measured from the reference device using a determined difference between the first composite measurement data for the first model of the corrected reference device and the second composite measurement data for the second model, as discussed with reference to FIG. 4A and illustrated in blocks 444 and 450 of FIGS. 4D and 4E.

[0104] In some implementations, the composite metrology data may include multiple sets of composite metrology data for a corresponding plurality of modified reference devices.

[0105] In some implementations, the measured metrology data includes a measured spectrum and the composite metrology data includes a composite spectrum.

[0106] 12 shows an illustrative flow diagram depicting example operations 1200 for assisting in characterization of devices on a specimen, according to some implementations. In some implementations, the example operations 1200 may be performed by a metrology device, such as metrology device 100, having at least one processor, such as processor 162 in computing system 160 of FIG.

[0107] At least one processor may acquire 1202 measured metrology data for a reference device of the device. For example, the means for acquiring measured metrology data for the reference device may be metrology device 100 and may interface with processor 162 in computing system 160 shown in FIG. 1. The measured metrology data may be, for example, reference data having known values ​​of key parameters such as overlay. The measured overlay metrology data may be, for example, measured metrology data from block 410 shown in FIGS. 4A-4E and block 510 shown in FIG. 5.

[0108] The at least one processor may generate 1204 a first set of synthetic metrology data for the first model of the reference device. For example, the first set of synthetic metrology data may be the calculated metrology data shown in block 422 of FIGS. 4A-4E. The means for generating the first set of synthetic metrology data for the first model of the reference device may be metrology device 100 including a computing system 160 having a processor 162 configured by computer-readable program code 166 shown in FIG.

[0109] The at least one processor may generate 1206 a second set of synthetic metrology data for a second model of the modified reference device that is modified relative to the first model. For example, the second set of synthetic metrology data may be the calculated metrology data shown in block 432 of FIGS. 4A-4E. The means for generating the second set of synthetic metrology data for the second model of the modified reference device that is modified relative to the first model may be metrology device 100 including a computing system 160 having a processor 162 configured by computer-readable program code 166 shown in FIG.

[0110] The at least one processor may generate composite metrology data for the modified reference device, each composite metrology data generated by integrating the measured metrology data, the first composite metrology data, and the second composite metrology data (1208). For example, the composite metrology data may be the composite metrology data generated in block 450 of FIGS. 4A-4E. The means for generating composite metrology data for the modified reference device, each composite metrology data generated by integrating the measured metrology data, the first composite metrology data, and the second composite metrology data, may be metrology device 100 including a computing system 160 having a processor 162 configured by computer-readable program code 166 shown in FIG. 1.

[0111] The at least one processor may store 1210 at least the composite measurement data as a training data set, for example, in memory 164 shown in Figure 1. The training data set may be, for example, the training set of measurement data 212 shown in Figure 2 and discussed with reference to Figures 4A-4E. The means for storing at least the composite measurement data as a training data set may be measurement device 100 including a computing system 160 having a processor 162 configured by computer-readable program code 166 and memory 164 within computing system 160 shown in Figure 1.

[0112] In some optional implementations, the at least one processor may further train 1212 a machine learning model using a training data set including at least composite metrology data to characterize the device using measured metrology data from the device, for example, as illustrated by training 214 a machine learning model 216 using a training set of metrology data 212 shown in FIG. 2 .

[0113] In some implementations, the at least one processor may generate the first model by fitting the measured metrology data to a first set of synthetic metrology data for the first model, for example, as discussed in block 420 of Figures 4A-4E.

[0114] In some implementations, the at least one processor may generate a second model of the modified reference device by changing at least one parameter of the first model, for example, as discussed in block 430 of Figures 4A-4E.

[0115] In some implementations, the at least one processor may generate composite measurement data for the corrected reference device by determining a variation between the measured measurement data and the first set of synthetic measurement data, for example, as discussed with reference to FIG. 4A and illustrated at blocks 440 and 442 of FIG. 4B and 4C. The at least one processor may further correct the second set of synthetic measurement data using the variation between the measured measurement data and the first set of synthetic measurement data, for example, as discussed with reference to FIG. 4A and illustrated at blocks 440, 442, and 450 of FIG. 4B and 4C. The variation between the measured measurement data and the first set of synthetic measurement data may be, for example, a mismatch between the measured measurement data and the first set of synthetic measurement data, as discussed with reference to block 440 of FIG. 4B, or a spectral transformation between the measured measurement data and the first set of synthetic measurement data, as illustrated at block 442 of FIG. 4C.

[0116] In some implementations, the at least one processor may generate composite metrology data for the corrected reference device by determining a difference between the first set of synthetic metrology data and the second set of synthetic metrology data, for example, as discussed with reference to FIG. 4A and illustrated at block 444 in FIG. 4D and 4E. The at least one processor may further correct the measured metrology data using the difference between the first set of synthetic metrology data and the second set of synthetic metrology data, as discussed with reference to FIG. 4A and illustrated at blocks 444 and 450 in FIG. 4D and 4E.

[0117] In some implementations, the at least one processor may further generate multiple sets of composite metrology data for the corresponding multiple modified reference devices and store the multiple sets of composite metrology data as a training data set.

[0118] The measured metrology data may be, for example, a measured spectrum, and the first set of synthetic metrology data and the second set of synthetic metrology data may be synthetic spectra.

[0119] In some implementations, the at least one processor may further generate a first set of primary parameter values ​​for the first model of the reference device, e.g., as discussed in block 522 of FIG. 5. The at least one processor may further generate a second set of primary parameter values ​​for the second model, e.g., as discussed in block 532 of FIG. 5. The at least one processor may generate composite reference parameters for the modified reference device by integrating the reference parameters, the first set of primary parameter values, and the second set of primary parameter values, e.g., as discussed in blocks 512, 540, and 550. The at least one processor may store the composite reference parameters together with the composite measurement data as a training data set, e.g., in memory 164 shown in FIG. 1. For example, composite reference parameters for a modified reference device may be generated by determining a primary parameter bias between a first set of primary parameter values ​​and a second set of primary parameter values, e.g., as discussed in block 540 of FIG. 5, and modifying the reference parameters using the primary parameter bias, e.g., as discussed in block 550 of FIG. 5.

[0120] 13 shows an illustrative flow diagram depicting example operations 1300 for assisting in characterization of devices on a specimen, according to some implementations. In some implementations, the example operations 1300 may be performed by a metrology device, such as metrology device 100, having at least one processor, such as processor 162 in computing system 160 of FIG. 1 .

[0121] At least one processor may acquire 1302 measured metrology data for a reference device of the device. For example, the means for acquiring measured metrology data for the reference device may be metrology device 100 and may interface with processor 162 in computing system 160 shown in FIG. 1. The measured metrology data may be, for example, reference data having known values ​​of key parameters such as overlay. The measured overlay metrology data may be, for example, measured metrology data from block 410 shown in FIGS. 4A-4E and block 510 shown in FIG. 5.

[0122] The at least one processor may generate composite metrology data, each composite metrology data generated by integrating the measured metrology data for the reference device with first synthetic metrology data for the first model of the reference device (1304). For example, the composite metrology data may be the generated composite metrology data shown in block 450 of FIGS. 4A-4E. In some implementations, the first set of synthetic metrology data may be the calculated metrology data shown in block 422 of FIGS. 4A-4E. The means for generating composite metrology data, each composite metrology data generated by integrating the measured metrology data for the reference device with the first synthetic metrology data for the first model of the reference device, may be metrology device 100, including a computing system 160 having a processor 162 configured by computer-readable program code 166 shown in FIG. 1. In some optional implementations, as illustrated by the dashed lines, each composite metrology data may include 1306 a combination of the metrology data and the first synthetic metrology data, and further, second synthetic metrology data for a second model of the modified reference device that is altered relative to the first model, where the composite metrology data may be for the modified reference device, for example, as discussed with respect to the generation of training data illustrated in the learning stage 210 and inference stage 220 shown in Figure 2 and workflows 400, 402, 404, and 406 of Figures 4A-4E. In some implementations, the second set of synthetic metrology data may be the calculated metrology data shown in block 432 of Figures 4A-4E.

[0123] The at least one processor may train 1308 a machine learning model using a training dataset including at least the composite metrology data to characterize a device using measured metrology data from the device, as illustrated by, for example, training 214 a machine learning model 216 using a training set of metrology data 212 shown in Figure 2. The means for training a machine learning model using a training dataset including at least the composite metrology data to characterize a device using measured metrology data from the device may be, for example, metrology device 100 including a computing system 160 having a processor 162 configured by computer-readable program code 166 shown in Figure 1.

[0124] In some implementations, the at least one processor may generate the first model by fitting the measured metrology data to a first set of synthetic metrology data for the first model, for example, as discussed in block 420 of Figures 4A-4E.

[0125] In some implementations, the at least one processor may generate a second model of the modified reference device by changing at least one parameter of the first model, for example, as discussed in block 430 of Figures 4A-4E.

[0126] In some implementations, the at least one processor may generate composite measurement data for the corrected reference device by determining a variation between the measured measurement data and the first set of synthetic measurement data, for example, as discussed with reference to FIG. 4A and illustrated at blocks 440 and 442 of FIG. 4B and 4C. The at least one processor may further correct the second set of synthetic measurement data using the variation between the measured measurement data and the first set of synthetic measurement data, for example, as discussed with reference to FIG. 4A and illustrated at blocks 440, 442, and 450 of FIG. 4B and 4C. The variation between the measured measurement data and the first set of synthetic measurement data may be, for example, a mismatch between the measured measurement data and the first set of synthetic measurement data, as discussed with reference to block 440 of FIG. 4B, or a spectral transformation between the measured measurement data and the first set of synthetic measurement data, as illustrated at block 442 of FIG. 4C.

[0127] In some implementations, the at least one processor may generate composite metrology data for the corrected reference device by determining a difference between the first set of synthetic metrology data and the second set of synthetic metrology data, for example, as discussed with reference to FIG. 4A and illustrated at block 444 in FIG. 4D and 4E. The at least one processor may further correct the measured metrology data using the difference between the first set of synthetic metrology data and the second set of synthetic metrology data, as discussed with reference to FIG. 4A and illustrated at blocks 444 and 450 in FIG. 4D and 4E.

[0128] In some implementations, the at least one processor may further generate multiple sets of composite metrology data for the corresponding multiple modified reference devices and store the multiple sets of composite metrology data as a training data set.

[0129] The measured metrology data may be, for example, a measured spectrum, and the first set of synthetic metrology data and the second set of synthetic metrology data may be synthetic spectra.

[0130] In some implementations, the at least one processor may further generate a first set of primary parameter values ​​for the first model of the reference device, e.g., as discussed in block 522 of FIG. 5. The at least one processor may further generate a second set of primary parameter values ​​for the second model (if used), e.g., as discussed in block 532 of FIG. 5. The at least one processor may generate composite reference parameters for the modified reference device by integrating the reference parameters, the first set of primary parameter values, and the second set of primary parameter values ​​(if used), e.g., as discussed in blocks 512, 540, and 550. The at least one processor may store the composite reference parameters together with the composite measurement data as a training data set, e.g., in memory 164 shown in FIG. 1. For example, composite reference parameters for a modified reference device may be generated by determining a primary parameter bias between a first set of primary parameter values ​​and a second set of primary parameter values, e.g., as discussed in block 540 of FIG. 5, and modifying the reference parameters using the primary parameter bias, e.g., as discussed in block 550 of FIG. 5.

[0131] The above description is intended to be illustrative and not limiting. For example, the above examples (or one or more aspects thereof) could be used in combination with each other. Other implementations could be used, for example, by one of ordinary skill in the art, upon reviewing the above description. Also, various features could be grouped together, and fewer than all features of a particular disclosed implementation could be used. Thus, the following aspects are hereby incorporated into the above description as examples or implementations, and it is contemplated that each aspect stands on its own as a separate implementation, and that such implementations can be combined with each other in various combinations or permutations. Therefore, the spirit and scope of the appended claims should not be limited to the foregoing description.

Claims

1. 1. A method for characterizing a device on a sample, comprising: obtaining measured metrology data from the device; determining at least one parameter of the device based on the measured metrology data using a machine learning model that uses composite metrology data, each of which includes an integration of measured metrology data for a reference device and first synthetic metrology data for a first model of the reference device.

2. The method of claim 1 , wherein the first model is generated by fitting measured metrology data from the reference device to synthetic metrology data for the first model.

3. 2. The method of claim 1, wherein each composite metrology data includes the integration of the metrology data and the first synthetic metrology data and further includes second synthetic metrology data for a second model of a modified reference device that is altered relative to the first model.

4. The method of claim 3 , wherein the second model of the modified reference device has at least one parameter that varies relative to the first model.

5. 4. The method of claim 3, wherein the machine learning model further uses composite reference parameters for the modified reference device based on a combination of reference parameters for the reference device, a first set of primary parameter values ​​generated for the first model, and a second set of primary parameter values ​​generated for the second model.

6. The method of claim 3 , wherein the second model of the modified reference device is generated by modifying at least one parameter of the first model.

7. 4. The method of claim 3, wherein second synthetic metrology data for the second model of the modified reference device is generated based on a variation between metrology data measured from the reference device and first synthetic metrology data for the first model.

8. 4. The method of claim 3, wherein the composite metrology data is generated by correcting the measured metrology data from the reference device using a determined difference between first composite metrology data for the first model and second composite metrology data for the second model of the corrected reference device.

9. The method of claim 3 , wherein the composite metrology data comprises a plurality of sets of composite metrology data for a corresponding plurality of modified reference devices.

10. The method of claim 1 , wherein the measured metrology data comprises a measured spectrum and the composite metrology data comprises a composite spectrum.

11. 1. A metrology system configured to assist in characterization of a device on a specimen, comprising: a source configured to generate radiation incident on the device on the sample; at least one detector configured to detect radiation from the device generated in response to the radiation incident on the device; at least one processor coupled to the at least one detector, the at least one processor comprising: Obtaining measured metrology data from the device; 1. A metrology system configured to determine at least one parameter of the device based on measured metrology data using a machine learning model that uses composite metrology data, the composite metrology data including an integration of measured metrology data for a reference device and first composite metrology data for a first model of the reference device, each of the metrology systems being configured to:

12. The metrology system of claim 11 , wherein the first model is generated by fitting measured metrology data from the reference device to synthetic metrology data for the first model.

13. 12. The measurement system of claim 11, wherein each composite measurement data includes the integration of the measurement data and the first composite measurement data and further includes second composite measurement data for a second model of a modified reference device that is altered relative to the first model.

14. The metrology system of claim 13 , wherein the second model of the modified reference device has at least one parameter that varies relative to the first model.

15. 14. The measurement system of claim 13, wherein the machine learning model further uses composite reference parameters for the modified reference device based on a combination of reference parameters for the reference device, a first set of primary parameter values ​​generated for the first model, and a second set of primary parameter values ​​generated for the second model.

16. The metrology system of claim 13 , wherein the second model of the modified reference device is generated by varying at least one parameter of the first model.

17. 14. The measurement system of claim 13, wherein second synthetic metrology data for the second model of the modified reference device is generated based on a variation between the measurement data measured from the reference device and the first synthetic metrology data for the first model.

18. 14. The measurement system of claim 13, wherein the composite measurement data is generated by correcting measurement data from the reference device using a determined difference between first composite measurement data for the first model and second composite measurement data for the second model for the corrected reference device.

19. The metrology system of claim 13 , wherein the composite metrology data includes a plurality of sets of composite metrology data for a corresponding plurality of corrected reference devices.

20. The metrology system of claim 11 , wherein the measured metrology data comprises a measured spectrum and the composite metrology data comprises a composite spectrum.

21. 1. A metrology system configured to assist in characterization of a device on a specimen, comprising: means for obtaining measured metrology data from said device; and means for determining at least one parameter of the device based on the measured metrology data using a machine learning model that uses composite metrology data, the composite metrology data including an integration of the measured metrology data for a reference device and first composite metrology data for a first model of the reference device.

22. 22. The measurement system of claim 21, wherein each composite measurement data includes the integration of the measurement data and the first composite measurement data and further includes second composite measurement data for a second model of a modified reference device that is altered relative to the first model.

23. 1. A method for characterizing a device on a sample, comprising: obtaining measured optical metrology data for a reference device of said device; generating composite optical metrology data, each generated by integrating measured optical metrology data for the reference device and first composite optical metrology data for a first model of the reference device; and training a machine learning model using a training dataset including at least the composite optical metrology data to characterize the device using the measured optical metrology data from the device.

24. 24. The method of claim 23, further comprising generating the first model by fitting the measured optical metrology data to a first set of synthetic optical metrology data for the first model.

25. 24. The method of claim 23, wherein each composite metrology data comprises an integration of the metrology data and the first synthetic metrology data and further comprises second synthetic metrology data for a second model of a modified reference device that is altered relative to the first model.

26. 26. The method of claim 25, further comprising generating the second model for the modified reference device by modifying at least one parameter of the first model.

27. generating the composite optical metrology data for the modified reference device includes: determining a variation between the measured optical metrology data and the first set of composite optical metrology data; and correcting a second set of synthetic optical metrology data using the variation between the measured optical metrology data and the first set of synthetic optical metrology data.

28. 28. The method of claim 27, wherein the variation between the measured optical metrology data and the first set of composite optical metrology data comprises a mismatch between the measured optical metrology data and the first set of composite optical metrology data or a spectral transformation between the measured optical metrology data and the first set of composite optical metrology data.

29. generating the composite optical metrology data for the modified reference device includes: determining a difference between the first set of combined optical metrology data and a second set of combined optical metrology data; and correcting the measured optical metrology data using the difference between the first set of composite optical metrology data and the second set of composite optical metrology data.

30. generating a plurality of sets of composite optical metrology data for a corresponding plurality of modified reference devices, the training data set comprising the plurality of sets of composite optical metrology data; 24. The method of claim 23.

31. 24. The method of claim 23, wherein the measured optical metrology data includes a measured spectrum and the first set of synthetic optical metrology data.

32. The reference device has reference parameters, and the method includes: generating a first set of primary parameter values ​​for the first model of the reference device; generating a second set of primary parameter values ​​for the second model; 26. The method of claim 25, further comprising: generating composite reference parameters for the modified reference device by integrating the reference parameters, the first set of primary parameter values, and the second set of primary parameter values, wherein the training data set includes the composite reference parameters along with the composite optical metrology data.

33. Generating the composite reference parameters for the modified reference device includes: determining a primary parameter bias between the first set of primary parameter values ​​and the second set of primary parameter values; and modifying the reference parameters using the primary parameter bias.

34. 1. A computer system configured to assist in characterizing a device on a specimen, comprising: at least one memory configured to store the measured optical metrology data and the composite optical metrology data; at least one processor coupled to the at least one memory, the at least one processor comprising: obtaining measured optical metrology data for a reference device of said device; generating composite optical metrology data, each generated by integrating measured optical metrology data for the reference device and first composite optical metrology data for a first model of the reference device; a computer system configured to train a machine learning model using a training dataset including at least the composite optical metrology data to characterize the device using measured optical metrology data from the device.

35. 35. The computer system of claim 34, wherein the at least one processor is further configured to generate the first model by fitting the measured optical metrology data to a first set of synthetic optical metrology data for the first model.

36. 35. The measurement system of claim 34, wherein each composite measurement data includes an integration of the measurement data and the first composite measurement data and further includes second composite measurement data for a second model of a modified reference device that is altered relative to the first model.

37. 37. The computer system of claim 36, wherein the at least one processor is further configured to generate the second model of the modified reference device by altering at least one parameter of the first model.

38. The at least one processor is configured to generate the composite optical metrology data for the modified reference device, the generating comprising: determining a variation between the measured optical metrology data and the first set of composite optical metrology data; 37. The computer system of claim 36, wherein the computer system is configured to use the variation between the measured optical metrology data and the first set of composite optical metrology data to correct a second set of composite optical metrology data.

39. 39. The computer system of claim 38, wherein the variation between the measured optical metrology data and the first set of composite optical metrology data comprises a mismatch between the measured optical metrology data and the first set of composite optical metrology data or a spectral transformation between the measured optical metrology data and the first set of composite optical metrology data.

40. The at least one processor is configured to generate the composite optical metrology data for the modified reference device, the generating comprising: determining a difference between the first set of combined optical metrology data and a second set of combined optical metrology data; 37. The computer system of claim 36, wherein the computer system is configured to correct the measured optical metrology data using the difference between the first set of composite optical metrology data and the second set of composite optical metrology data.

41. The at least one processor 35. The computer system of claim 34, further configured to generate a plurality of sets of composite optical metrology data for a corresponding plurality of modified reference devices, wherein the training data set comprises the plurality of sets of composite optical metrology data.

42. 35. The computer system of claim 34, wherein the measured optical metrology data includes a measured spectrum and the first set of composite optical metrology data.

43. The reference device has reference parameters, and the at least one processor: generating a first set of key parameter values ​​for the first model of the reference device; generating a second set of key parameter values ​​for the second model; 37. The computer system of claim 36, further configured to generate composite reference parameters for the modified reference device by integrating the reference parameters, the first set of primary parameter values, and the second set of primary parameter values, wherein the training data set includes the composite reference parameters together with the composite optical metrology data.

44. The at least one processor is configured to generate the composite reference parameters for the modified reference device, the generating comprising: determining a primary parameter bias between the first set of primary parameter values ​​and the second set of primary parameter values; 44. The computer system of claim 43, wherein the computer system is configured to modify the reference parameters with the primary parameter bias.

45. 1. A computer system for characterizing a device on a specimen, comprising: means for obtaining measured optical metrology data for a reference device of said device; means for generating composite optical metrology data, each generated by integrating measured optical metrology data for the reference device and first composite optical metrology data for a first model of the reference device; means for training a machine learning model using a training dataset including at least the composite optical metrology data to characterize the device using measured optical metrology data from the device.

46. 46. ​​The computer system of claim 45, wherein each composite metrology data includes an integration of the metrology data and the first composite metrology data and further includes second composite metrology data for a second model of a modified reference device that is altered relative to the first model.

Images