Transfer learning for measurement data analysis

Abstract

Non-contact measurements of structures, such as optical or X-ray measurements, are supported using transfer learning to train a machine learning (ML) model to predict key parameters. A first set of measurement data is acquired for a first set of structures and used to train a first ML model. A second set of measurement data is acquired for one or more second structures. Transfer learning is performed from the first ML model to the second set of measurement data to create a second ML model to predict key parameters of one or more second structures. Domain adaptation may be used, where the measurement data is selected from the first and second sets of measurement data using a feature extractor and used to train an ML model to predict key parameters of one or more second structures.

Description

Technical Field

[0001] (Cross - Reference to Related Applications) This application claims priority to U.S. Provisional Application No. 63 / 491,700, titled "TRANSFER LEARNING FOR METROLOGY DATA ANALYSIS," filed on March 22, 2023, and U.S. Non - Provisional Application No. 18 / 611,195, titled "TRANSFER LEARNING FOR METROLOGY DATA ANALYSIS," filed on March 20, 2024, both of which have been assigned to the present assignee and are hereby incorporated by reference in their entirety.

[0002] Implementations of the subject matter described herein generally relate to metrology, and more specifically, to modeling metrology data.

Background Art

[0003] In semiconductor and other similar industries, measurement equipment such as optical measurement equipment is often used to provide non - contact evaluation of samples during processing. In optical measurement, the sample under test is irradiated with light of, for example, a single wavelength or multiple wavelengths. After interacting with the sample, the resulting light is detected and analyzed to determine one or more characteristics of the sample.

[0004] The analysis typically involves a model of the structure under test. The model may be generated based on the structure's material and nominal parameters, such as film thickness, line width, and spacing width. One or more parameters of the model may vary, and predicted data may be calculated for each parameter variation based on the model, for example, using Rigorous Coupled Wave Analysis (RCWA) or other similar techniques. Measured data may be compared with predicted data for each parameter variation, for example, in a nonlinear regression process, until a good fit is achieved between the predicted data and the measured data, at which point the fitted parameters are determined to be an accurate representation of the structure's parameters under test.

[0005] Measurement techniques, including data analysis and recipes, function well under certain assumptions, such as the data used to optimize the recipe (spectrums and references) and test data from inline measurements being drawn from the same distribution. When the distribution changes, for example, when the substrate processing changes, the accuracy of the data analysis and / or recipe typically decreases. Therefore, what is needed is an improved process that can be used to enhance the robustness of the measurement techniques. [Overview of the project]

[0006] Non-contact measurements of structures, such as optical or X-ray measurements, are supported using transfer learning to train a machine learning (ML) model to predict key parameters. A first set of measurement data for one or more structures is acquired and used to train a first ML model. A second set of measurement data for a second or more structures is acquired. Transfer learning is performed from the first ML model to the second set of measurement data to create a second ML model to predict key parameters of a second or more structures. Transfer learning through domain adaptation can be used in another implementation, in which measurement data is selected from the first and second sets of measurement data using a feature extractor and used to train an ML model to predict key parameters of a second or more structures.

[0007] In one implementation, a method for supporting the measurement of structures includes: obtaining a first set of measurement data for one or more first structures; and using the first set of measurement data to train a first machine learning model for the one or more first structures. A second set of measurement data is obtained for one or more second structures. Transfer learning is performed from the first machine learning model to the second set of measurement data to create a second machine learning model for predicting key parameters for one or more second structures.

[0008] In one implementation, a computer system configured to support the measurement of structures includes at least one processor configured to acquire a first set of measurement data for one or more first structures and to train a first machine learning model for the one or more first structures using the first set of measurement data. The at least one processor is further configured to acquire a second set of measurement data for one or more second structures. The at least one processor is further configured to perform transfer learning from the first machine learning model to the second set of measurement data to produce a second machine learning model for predicting key parameters for one or more second structures.

[0009] In one implementation, a method for supporting the measurement of a structure includes obtaining a first set of measurement data for one or more first structures, and obtaining a second set of measurement data for one or more second structures. Measurement data from the first and second sets of measurement data are selected using a feature extractor. Using the selected measurement data, a machine learning model is trained to predict key parameters for one or more second structures.

[0010] In one implementation, a computer system configured to support the measurement of structures includes at least one processor configured to acquire a first set of measurement data for one or more first structures and a second set of measurement data for one or more second structures. The at least one processor is further configured to use a feature extractor to select measurement data from the first set of measurement data and the second set of measurement data. The at least one processor is further configured to train a machine learning model using the selected measurement data to predict key parameters for one or more second structures. [Brief explanation of the drawing]

[0011] [Figure 1]A schematic diagram illustrating a measuring device that may be configured to support optical measurements of one or more structures as discussed herein is illustrated. [Figure 2A] This section illustrates a typical workflow used in conventional modeling and library generation. [Figure 2B] This specification illustrates a workflow that illustrates one implementation form for modeling and predicting key parameters. [Figure 2C] This specification illustrates a workflow that demonstrates a generalized implementation for modeling and predicting key parameters. [Figure 3] This section illustrates an example of a transfer learning process for measurement data. [Figure 4] Examples of source and target domains are illustrated using spectral morphology. [Figure 5] This paper illustrates an example of transfer learning using a domain adaptation process for measurement data. [Figure 6] This document presents illustrative flowcharts illustrating exemplary behavior for supporting the measurement of structures across several implementations. [Figure 7] This document presents illustrative flowcharts illustrating exemplary behavior for supporting the measurement of structures across several implementations. [Modes for carrying out the invention]

[0012] During the fabrication of semiconductor devices and similar devices, it is very often necessary to monitor the fabrication process by non-destructively measuring the device. Optical and X-ray measurements are examples of non-contact measurement techniques that can be used for non-contact evaluation of samples during processing. For example, optical measurement techniques such as thin-film measurement and optical critical dimension (OCD) measurement can use structural modeling to generate predicted data that is compared with measured data from the sample. For example, variable parameters in the model, such as layer thickness, line width, spacing width, sidewall angle, and material properties, may change, and predicted data is generated for each variation. Measured data from the sample under test can be compared with predicted data for each parameter variation, for example, in a nonlinear regression process, until a good fit is achieved, at which point the fitted parameter values ​​are determined to be an accurate representation of the sample's parameters.

[0013] Modeling typically involves using physics-based techniques such as Rigorous Coupled Wave Analysis (RCWA), Finite-Difference Time-Domain (FDTD), or Finite Element Method (FEM), which require detailed knowledge of the structure. For example, modeling requires that preliminary structural and material information be known about the sample in order to generate an accurate representative model of the sample, which may include one or more variable parameters. Preliminary structural and material information about the sample may include the type of structure and a physical description of the sample with nominal values ​​for various parameters, such as layer thickness, line width, spacing width, sidewall angle, etc., along with the range over which these parameters may vary. The sample may further include one or more sample parameters that are not variable, i.e., are not expected to change in significant amounts during manufacturing.

[0014] Conventional modeling for measurement offers a high level of fidelity to the physical constraints of the sample (as well as the parameters of the measurement device). Libraries can be pre-generated to increase measurement throughput. However, constructing accurate libraries is slow in time-to-solution (TTS) and computationally expensive, especially for complex structures such as 3D-NAND devices. For example, a typical TTS for library generation can range from several days to several weeks and may require dozens of computing nodes (blades). Furthermore, the accuracy of pre-generated libraries can be compromised for complex structures with large misfits, i.e., achieving good correlation with the reference sample is difficult. Moreover, the robustness of the recipe is affected by model assumptions such as fixed and coupled parameters. Therefore, generating accurate libraries may require spectra for numerous model deformations, e.g., hundreds to thousands of model deformations, and furthermore, post-library optimization processes may be used, which further increases the TTS and computational power requirements.

[0015] Other measurement techniques, such as machine learning (ML), are available, offering significantly shorter TTS with reduced computational power requirements. However, ML typically requires large amounts of reference data that are costly to acquire, otherwise the robustness of the recipe is compromised.

[0016] The data analysis techniques for measurement assume that the reference data used for library construction or ML training, such as reference spectra and reference parameters, are from the same distribution as the measurement data from the sample under test. If the distribution changes, for example, if the process used to generate the sample from which the measurement data is obtained differs from the process assumptions used for the reference data, the accuracy of the recipe, and therefore the data analysis, will decrease. As a result, the lifespan of the library or ML training may be limited due to process variations, and new library generation and / or ML training may be required.

[0017] As discussed herein, transfer learning approaches are used to improve recipe generation for non-contact measurements such as optical measurements. Various aspects of this disclosure are described with reference to “optical measurements” and “optical measurement data,” but this disclosure may apply to other types of non-contact measurements, including X-ray measurements and other measurement techniques that produce measurement data using one or another form of radiation, and therefore, it should be understood that this disclosure is not limited to optical measurements unless otherwise specified.

[0018] For example, a first set of measurement data may be obtained for one or more structures, and a first machine learning model may be trained on one or more structures using this first set of measurement data. The first set of measurement data may be synthetic data generated from a physical model, experimental data measured from one or more reference structures, or a combination of both. The first set of measurement data may cover a wider range of process variations using more floating parameters than those used in typical measurement modeling. A second set of measurement data is obtained for a second one or more structures on a sample. The second set of measurement data may be experimental data measured from a second one or more structures, synthetic data generated from a physical model of a second one or more structures, or a combination of both. Transfer learning from the first machine learning model to the second set of measurement data is used to create a second machine learning model for predicting key parameters for the second one or more structures. For example, key parameters may include geometric dimensions of the structures on the sample, as well as parameters such as physical properties and material properties. For example, a model trained using synthetic data may be transferred to experimental data for inline measurements.

[0019] Transfer learning approaches are used to improve the robustness of recipes, reduce the number of reference samples required, and reduce the need to re-collect reference data and create reprocessed recipes when changes occur in the sample manufacturing process. Transfer learning approaches further have the advantages of reduced recipe creation steps, improved usability, and shorter TTS, especially for complex devices such as 3D NAND. The transfer learning approaches considered herein may require significantly fewer (2 to 100 times, depending on the device and hardware) samples in the spectrum than required for conventional library generation, which leads to significantly shorter TTS and reduced use of computational resources.

[0020] FIG. 1 illustrates, by way of example, a schematic diagram of a non-contact measurement device 100 that may be configured to support the measurement of structures, as described herein. For example, the measurement device 100 can be used to generate measurement data from test samples and / or reference samples and process the measurement data, as described herein. Although the measurement device 100 is illustrated as an optical measurement device, it should be understood that other types of non-contact measurement devices, including X-ray measurement devices, may be used. As illustrated, the measurement device 100 can be configured to perform, for example, spectroscopic reflectance measurement, spectroscopic polarization analysis (including Mueller matrix polarization analysis), spectroscopic scattering measurement, overlay scattering measurement, interferometric measurement, or FTIR measurement on a sample 101 that includes one or more structures to be measured. It should be understood that the measurement device 100 is illustrated as an example of a configuration for a measurement device, and other measurement device configurations, including normal incidence devices, non-polarization devices, etc., may be used if desired.

[0021] The measurement device 100 includes a light source 110 that generates light 102. For example, the light 102 is UV-visible light having a wavelength of, for example, 200 nm to 1000 nm. The light 102 generated by the light source 110 can include a range of wavelengths, that is, a continuous range or a plurality of discrete wavelengths, or can be a single wavelength. The measurement device 100 includes condenser optical components 120 and 130 that collect and receive light and direct the light to be obliquely incident on the uppermost surface of the sample 101. The optical components 120 and 130 can be refractive, reflective, or a combination thereof, and can be objective lenses.

[0022] The reflected light can be collected by the lens 114 and received by the detector 150. The detector 150 can be a conventional charge coupled device (CCD), a photodiode array, a CMOS, or a detector of a similar type. The detector 150 can be, for example, a spectrometer when broadband light is used, and the detector 150 can generate a spectral signal as a function of wavelength. The spectrometer can be used to disperse the entire spectrum of polarized light into spectral components across an array of detector pixels. One or more polarization elements can be present in the beam path of the measurement device 100. For example, the measurement device 100 can include one or both (or none) of one or more polarization elements 104 in the beam path before the sample 101 and a polarization element (analyzer) 112 in the beam path after the sample 101, and can include one or more additional optical elements 105 such as a wave plate, a compensator, or a photoelastic modulator, which can be before, after, or both before and after the sample 101.

[0023] The measurement device 100 further includes one or more computing systems 160 configured to acquire measurement data, which is used to train a machine learning model to predict key parameters (such as geometric dimensions, as well as physical and material properties) of a structure on a sample using methods described herein. For example, the measurement data may be measurement data acquired from a detector 150, or synthetic measurement data generated based on one or more models. As illustrated, one or more computing systems 160 may be coupled to the detector 150 to receive measured measurement data acquired by the detector 150 during the measurement of a structure on a sample 101. 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. One or more computing systems 160 may acquire multiple sets of measurement data about a structure on a sample and be configured to train one or more machine learning models with the measurement data to predict key parameters about the structure, including, for example, using transfer learning as described herein. One or more computing systems 160 may be further configured to measure the structure under test based on the predicted key parameters.

[0024] One or more computing systems 160 may be a single computer system or multiple separate or linked computer systems, and a computer system may include one or more processors that can be coupled to one or more computing nodes (blades). It should be understood that, as herein, computing system 160, at least one computing system 160, and one or more computing systems 160 may be interchangeable. In some implementations, the computing system 160 may be separate from the measurement device 100, while in some implementations, the computing system 160 may be included in the measurement device 100, or connected to or otherwise associated with the measurement device 100, or may be separate from the measurement device 100. Each different subsystem of the measurement device 100 may include a computing system configured to perform steps associated with the relevant subsystem. For example, the computing system 160 may control the positioning of the sample 101, for example, by controlling the movement of a step 109 coupled to a chuck. For example, the stage 109 may be capable of horizontal movement in Cartesian (i.e., X and Y) coordinates, polar (i.e., R and θ) coordinates, or any combination of the two. The stage may also be capable of vertical movement along the Z coordinate. The computing system 160 may further control the operation of the chuck 108 to hold or release the sample 101. The computing system 160 may further control or monitor, for example, one or more polarizing elements 104, 112, or monitor optical elements 105, etc.

[0025] The computing system 160 can be communicatively coupled to the detector 150 in any manner known in the art. For example, one or more computing systems 160 can be coupled to separate computing systems associated with the detector 150. The computing system 160 can be configured to receive and / or acquire measurement data or information from one or more subsystems of the measurement device 100, such as the detector 150, and control polarizing elements 104, 112, or optical elements 105, via a transmission medium which may include wired and / or wireless portions. Thus, the transmission medium can function as a data link between the computing system 160 and other subsystems of the measurement device 100.

[0026] The computing system 160 includes at least one processor 162 with memory 164, and a user interface (UI) 168, which are communicably coupled via a bus 161. The memory 164 or other non-temporary computer-readable storage medium includes its embodied computer-readable program code 166 and may be used by the computing system 160 to perform functions including controlling a measurement device 100 and / or predicting key parameters for a structure described herein. Data structures and software code for automatically implementing one or more actions described in this detailed description may be implemented by those skilled in the art in light of this disclosure and may be stored in a computer-readable storage medium, for example, memory 164, which may be any device or medium capable of storing code and / or data for use by a computer system such as the computing system 160. The computer-readable storage medium may be, but is not limited to, read-only memory, random-access memory, disk drives, magnetic tapes, and other magnetic and optical storage devices. Additionally, the functions described herein may be embodied in whole or in part within the circuitry of an application-specific integrated circuit (ASIC) or programmable logic device (PLD), and the functions may be embodied in a computer-readable descriptor language that can be used to create an ASIC or PLD that operates as described herein.

[0027] The computing system 160 may be configured, for example, to support the measurement of structures on a sample using transfer learning. At least one processor 162 may acquire measurement data for a first set of structures. The measurement data may be simulated (synthetic) data generated using a model of the first set of structures, or experimental data measured from the first set of structures using a measurement device 100, for example. At least one processor 162 may train a first machine learning model using the measurement data for the first set of structures. At least one processor 162 may also acquire measurement data for a second set of structures similar to the test structures being measured. The measurement data for the second set of structures may be experimental data measured using a measurement device 100, for example. At least one processor 162 performs transfer learning from the first machine learning model to the measurement data of the second set of structures to optimize a second machine learning model for predicting key parameters for one or more second structures.

[0028] In some implementations, the computing system 160 may be configured to support the measurement of structures on a sample using transfer learning with domain adaptation. At least one processor 162 may acquire, for example, a first set of measurement data for a first set of structures as a source and a second set of measurement data for a second set of structures as a target. The first set of measurement data may be simulation (synthetic) data generated using a model of the first set of structures, or experimental data measured from the first set of structures using, for example, a measurement device 100, or a combination thereof. The second set of measurement data for the second set of structures may be, for example, experimental data measured using the measurement device 100, or simulation (synthetic) data generated using a model of the second set of structures, or a combination thereof. At least one processor 162 may, for example, use a feature extractor to select measurement data from the first and second sets of measurement data. For example, measurement data may be selected that can minimize the regression model error of the main training task in order to predict key parameters and minimize the difference in measurement data between the first set of measurement data (source) and the second set of measurement data (target). In another example, computing system 160 may be configured to support the measurement of structures on a sample using other types of transfer learning, such as instance-based learning. Instance-based transfer learning can help improve the ML performance of a target domain by reweighting samples in the source or target domain to compensate for distributional differences between the two domains.

[0029] As a result, for example, the trained machine learning model may be stored in memory 164 and / or provided to other devices for measurement of the structure. For example, during measurement, measurement data is acquired from the target structure using a measurement device 100. The measurement data is provided to the trained machine learning model to acquire characteristic parameters of the structure. The results are reported and feedforward or feed back to process equipment to adjust appropriate fabrication steps and correct any variations detected in the fabrication process. The computing system 160 may include a communication port 169 which may be any type of communication connection, such as to the Internet or any other computer network. The communication 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 signals with measurement results and / or instructions to another system, such as an external process tool, in a feedforward or feedback process to adjust process parameters associated with the fabrication steps of the sample based on the measurement results.

[0030] Figure 2A illustrates a typical workflow 200 used in conventional modeling and library generation for predicting key parameters, for example. As illustrated, initial model setup and real-time fitting 202 are performed, where a physical model of the structure is generated based on preliminary structural and material information known about the structure, such as the type of structure, the physical description of the structure, and the material. The model includes nominal values ​​for various parameters, such as layer thickness, line width, space width, and sidewall angle, along with a range of variable parameters, such as a range where floating parameters can vary. The model is generated and validated using real-time fitting. Real-time fitting compares, for example, reference data, such as data measured from a reference structure with known parameter values ​​(e.g., measured using a transmission electron microscope (TEM) or critical dimension scanning electron microscope (CD-SEM)), with calculated data of a model having the same parameter values. If the comparison of the measured and calculated data for the model fits closely, for example within a few nanometers, the model can be assumed to be an accurate representation of the structure. The time required for initial model setup depends on the complexity of the structure, such as the number of layers, features, and parameters, as well as the number of reference structures used for fitting.

[0031] Once the model is set up, library generation 204 may then be performed. During library generation, the variable parameters change, and data is computed for each change, generating a set of data associated with the parameter values ​​for the model. The time required for library generation also depends on the complexity of the structure, as well as the size of the variable parameter range and the desired resolution of the library. For example, a sample with many variable parameters will have many permutations, each of which changes individually over its entire range, while a sample with few variable parameters will have significantly fewer permutations. Similarly, if the library is generated with a large variable parameter range or high resolution, i.e., with many possible values ​​within each variable parameter range, there will be more permutations for a small variable parameter range or low resolution. Even with a reasonable number of variable parameters with a limited range and resolution, the number of separate models (each with different permutations of parameters) for which data must be computed can be in the hundreds or thousands, requiring a lot of time or days to generate a useful library. Therefore, the resolution and range of variable parameters of the library may be limited in practical considerations, which limits the robustness of the library.

[0032] Following library generation 204, post-library recipe optimization 206 is performed, in which recipe details such as floating and fixed parameters are further optimized using the library, so that the reported measurement results for key parameters from the recipe may match the reference results within the desired tolerance, usually sub-nanometer. The time required for post-library optimization depends on the size of the reference sample and the complexity of the model.

[0033] As can be inferred, the TTS for conventional library generation is generally very long, for example, several days or weeks for complex structures. Furthermore, the accuracy of the model can be compromised for complex structures where significant misfits can exist during model setup and real-time fitting, and even after post-library optimization, it is difficult to achieve a good correlation between computed data and reference data for complex structures. Moreover, robustness can be affected by model assumptions such as fixed and coupled parameters, as well as practical considerations to reduce the number of model deformations.

[0034] Other measurement techniques may be used to avoid the need to generate a complete library. For example, pure machine learning (ML) can be used to provide a shorter TTS. However, pure ML requires a large amount of reference data, which is expensive to acquire. Without a large amount of reference data, robustness and accuracy may be compromised.

[0035] In addition, as discussed above, the process used to fabricate structures after conventional library generation may change over time. For example, if a limited range of variable parameters is used during initial model setup and real-time fitting, process variations for fabricating structures may reduce the accuracy and usefulness of the library. As a result, it may be necessary to generate additional libraries, and due to process variations, the recipe may need to be re-optimized over time using the newly generated libraries.

[0036] Figure 2B illustrates workflow 250, which illustrates one implementation form that models and predicts key parameters as described herein. Workflow 250 bypasses library generation and offers the advantages of fewer recipe creation steps, a shorter TTS, and ease of use.

[0037] The modeling and data analysis described herein, illustrated by Workflow 250 as one implementation, applies transfer learning to train a model using source domain data that can be transferred to a target domain, facilitating model training in the target domain due to a shorter time to solution, improved recipe robustness and generalizability, and a reduced number of reference data samples compared to the process illustrated in Figure 2A. For example, where a conventional process may require several days or weeks for TTS, this modeling and data analysis can be performed in less than a day, with significant improvements in robustness and generalizability, ease of use, and reference data requirements.

[0038] As illustrated in Figure 2B, an initial model setup and real-time fitting 252 is performed, which may be the same as or similar to the initial model setup and real-time fitting 202 in workflow 200 illustrated in Figure 2A. During the initial model setup and real-time fitting 252, the physical model of the structure may be generated based on, for example, the range over which the parameters can change, along with preliminary structural and material information known about the structure, such as a physical description of the structure with nominal values ​​for various parameters such as the type of structure, layer thickness, line width, space width, and sidewall angle, and one or more variable parameters. The model is generated and validated using real-time fitting. Real-time fitting compares, for example, data measured from reference data, such as data from a reference structure with known parameter values, with calculated data for a model with the same parameter values. If the comparison between the measured and calculated data of the model fits closely, the model can be assumed to be an accurate representation of the structure.

[0039] In some implementations, the initial model setup and real-time fitting 252 may include one or more models that can cover large process variations for the model generated in the initial model setup and real-time fitting 202. For example, the model may cover process variations that cannot be covered by the provided experimental data (reference data). The process window may be estimated by the user from prior experience or from a process window simulated using process simulation software. The model may be for 3D-NAND or other logic devices, a model for complex structures such as dynamic random access memory (DRAM), or a model for a combination of signals from any type of measurement tool or multiple tools.

[0040] Automated synthetic spectrum generation 254 is used to generate synthetic spectra from one or more models to cover large process variations. The synthetic data may be information that is artificially created using one or more models, for example, from the initial model setup and real-time fitting 252, rather than being measured from a reference sample. Although the synthetic spectra can cover larger process variations, they may be significantly sparser than those used in conventional library generation, and as a result, the automated synthetic spectrum generation 254 process may be substantially faster than library generation 204, for example, up to 40 times faster.

[0041] While workflow 250 illustrates the generation of a synthetic spectrum, it should be understood that any desired type of synthetic measurement data can be used. The synthetic measurement data may be labeled, unlabeled, or a combination of both. Furthermore, workflow 250 is not limited to synthetic data. For example, instead of generating synthetic measurement data from one or more models generated from the initial model setup and real-time fitting 252, measurement data, such as experimental measurement data, may be generated using measurements from one or more reference structures. Experimental measurement data measured from one or more models may be labeled, unlabeled, or a combination of both. Using experimental measurement data may eliminate the need for model setup and real-time fitting 252 and automated synthetic spectrum generation 254. In some implementations, both labeled synthetic and experimental measurement data may be generated. Furthermore, synthetic and experimental measurement data may be interchangeable with respect to the target or source domain. For example, the source domain may be synthetic measurement data, experimental measurement data, or a combination of synthetic and experimental measurement data, while the target domain may be synthetic measurement data, experimental measurement data, or a combination of synthetic and experimental measurement data.

[0042] Automated machine learning (ML) optimization 256 is used to generate key parameter predictions based on synthesized spectra (i.e., synthesized measurement data) from automated synthesized spectrum generation 254 or experimental measurement data. Automated ML optimization involves a transfer learning approach that develops recipes for one or more first targets and applies the recipes to one or more second targets of similar structures with small differences such as pitch, material properties, or minor geometric changes, and has the advantage of a shorter TTS by reducing the required references for the second set of target recipe creations, which are facilitated by the knowledge learned from the first set of target recipe creations.

[0043] Modeling and predicting key parameters, as illustrated in Workflow 250 and described herein, effectively reduces the TTS of measurement solutions for complex structures such as 3D-NAND from 1-2 weeks in conventional workflows, as illustrated in Figure 2A, to less than 1 day. Furthermore, the resulting modeled and predicted key parameters have improved robustness, generalizability, and ease of use (e.g., fewer adjustment knobs and floating / fixed parameters through automation) compared to those obtained from conventional library generation approaches. With respect to a pure ML approach, Workflow 250 exhibits lower dependence on experimental reference data with improved recipe robustness. Moreover, the required recipe creation steps are reduced, and less user experience is needed to develop high-quality recipes. For example, practical considerations that limit the number of floating parameters for modeling are relaxed so that large process variations can be covered. Furthermore, one or more physical models may be used, for example, to generate synthetic measurement data, but the required model fidelity and fit quality may be lower compared to conventional modeling methods. In addition, the robustness of the recipe against process variations (which result in changes in measured data over time) is maintained by retraining the model with unlabeled measurement data or by adding synthetic measurement data without the need for new reference data.

[0044] ML Optimization 256 may use transfer learning on measurement data, where a first ML model is trained using synthetic (and / or experimental) measurement data with or without reference data from, for example, TEM or CDSEM. Transfer learning is then performed from the first ML model to a second set of measurement data, which may be experimental, synthetic, or a combination thereof, to optimize a second ML model that can be used to predict measurement results for experimental measurement data. In some implementations, the first and second ML models may be simultaneously optimized or co-trained based on the first and second sets of measurement data and merged into a single ML model. The transfer learning process for measurement data significantly reduces reliance on reference data. Furthermore, in some implementations, one set of measurement data may be completely unlabeled (no references are used), for example, using domain adaptation. In some implementations, other types of transfer learning, such as instance-based learning, may be used.

[0045] Figure 2C illustrates workflow 270, which shows a generalized implementation that models and predicts key parameters as described herein. Workflow 270 may be similar to, for example, workflow 250, and bypasses the library generation shown in workflow 200, offering the advantages of fewer recipe creation steps, a shorter TTS, and ease of use.

[0046] As illustrated in Figure 2C, workflow 270 obtains a first set 272 of measurement data from one or more first structures. The first set of measurement data may be spectral data or any other desired data, and may be synthetic measurement data, experimental measurement data, or a combination thereof. For example, the first set of measurement data may be source data or target data used in a machine learning process. The first one or more structures may be a single structure or a set of structures, may be structures of different types, or may be structures of the same type produced using the same or different processes. Obtaining the first set of measurement data 272 may be similar to the model setup and real-time fitting 252 and automated synthetic spectrum generation 254 shown in Figure 2B for producing synthetic measurement data, for example, by using initial model setup and real-time fitting on one or more models of one or more first structures and generating synthetic measurement data from one or more models to cover relatively large variations in structural parameters, layer property parameters, material property parameters, or a combination thereof. Additionally or alternatively, as described above, obtaining the first set 272 of measurement data can be done, for example, by measuring one or more first structures to generate experimental measurement data.

[0047] As illustrated, workflow 270 obtains a second set 274 of measurement data from one or more second structures. The second set of measurement data may be spectral data or any other desired data, and may be synthetic measurement data, experimental measurement data, or a combination thereof. For example, the second set of measurement data may be target data or source data used in a machine learning process. The one or more second structures may be a single structure or a set of structures, may be structures of different types, or may be structures of the same type produced using the same or different processes. Obtaining the second set of measurement data 274 may be done, for example, by measuring one or more second structures to generate experimental measurement data. Additionally or alternatively, as described above, obtaining the second set of measurement data 274 may be done, for example, by producing synthetic measurement data, for example, by generating synthetic measurement data from one or more models using initial model setup and real-time fitting on one or more models of one or more second structures, similar to the model setup and real-time fitting 252 and automated synthetic spectrum generation 254 shown in Figure 2B for producing synthetic measurement data. The variations in structural parameters, layer property parameters, material property parameters, or combinations thereof covered by the second set of measurement data may be smaller than those in the first set of measurement data.

[0048] Workflow 270 trains a machine learning model using transfer learning 276 based on a first set of measurement data and a second set of measurement data. For example, the first model may be trained on measurement data in a source domain, e.g., a first set of measurement data, and the first model may be transferred using a parameter-based transfer learning process to measurement data in a target domain, e.g., a second set of measurement data, to produce a second machine learning model for predicting key parameters for one or more second structures. In the parameter-based transfer learning process, both the source and target domains contain labeled data, e.g., both the source and target domains contain at least partially labeled data. In another implementation, the first set of measurement data and the second set of measurement data may be used to train a machine learning model using transfer learning with domain adaptation to produce a machine learning model for predicting key parameters for one or more second structures. In domain adaptation, for example, the source domain and target domain may have different feature spaces and distributions, and the process generally attempts to improve the performance of training the target ML model by modifying both the source and target domains to bring their distributions closer together. In transfer learning using domain adaptation, only one of the source or target domains contains at least partially labeled data, while the remaining domain contains labeled data, unlabeled data, or a combination of labeled and unlabeled data.

[0049] Figure 3 illustrates an example of training a machine learning model using a transfer learning process 300 for measurement data, which may be used for ML optimization 256 shown in Figure 2B, and transfer learning 276 shown in Figure 2C. As illustrated, transfer learning is performed starting with a first ML model 350 (with readily available inputs and labels) for target A, which is used to optimize one or more second ML models 360 and 370 (with fewer inputs and labels) to predict measurement results for target B. In some implementations, the transfer learning process 300 may be adapted to be performed with respect to three or more targets. For example, there may be one or more targets used to train the first ML model 350, and one or more targets for one or more second ML models 360 and 370. The transfer learning process 300, shown in Figure 3, is sometimes referred to as parameter-based transfer learning and includes at least some labeled data for all domains (source domain and target domain). If there are more than two domains and some domains do not have labeled data, the data from these unlabeled domains can be combined with labeled data from other domains to train one of the ML models.

[0050] As illustrated, the measurement data 302 of target A is provided as the source domain to the input layer of the feature extractor 320. The measurement data 302 for target A may be synthetic measurement data generated using one or more models on target A, or experimental measurement data measured from target A, or a combination of synthetic and experimental measurement data. The measurement data 302 of target A is at least partially labeled. Target A may be one or more structures, e.g., one or more models of one or more structures generated from the initial model setup and real-time fitting 252 in Figure 2B used to produce synthetic measurement data, or one or more reference structures measured to generate experimental measurement data, as described in the acquisition of the first set 272 of measurement data in Figure 2C. The feature extractor 320 may be any linear or nonlinear feature extraction architecture, including but not limited to convolutional neural networks, shallow neural networks, deep learning, any other machine learning model that can co-optimize on both source and target data, or any combination thereof. The feature extractor 320 is coupled to the regression predictor 330, which provides 1 to n parameters in the output layer 340, exemplified as parameters P1 and P2 for target A of the first ML model 350. The output parameters P1 and P2 may be structural or layer parameters such as critical dimension (CD), height, thickness, depth, line width, space width, sidewall angle, etching recess, slope, overlay, surface roughness, and line edge roughness, or material properties such as doping concentration, composition, crystallinity, and conductivity.

[0051] As further illustrated, the measurement data 304 for target B is provided to the input layer 310 to the feature extractor 320 as a target domain. The measurement data 304 for target B may be experimental measurement data measured from target B, or synthetic measurement data generated using one or more models of target B, or a combination of experimental and synthetic measurement data. The experimental and synthetic measurement data are at least partially labeled. Similar to target A, target B may be one or more structures, for example, one or more models of one or more structures generated from modeling to produce synthetic measurement data, or one or more physical structures measured to generate experimental measurement data, for example, as described in the acquisition of the second set 274 of measurement data in Figure 2C. For example, each of target A and target B may be the same type of structure produced using the same or different processes. For example, target A and target B may be a single type of structure from the same process step but have different process conditions such as chemical concentration, etching time, and gas pressure. In another example, each of target A and target B could be multiple structure types produced using the same or different processes. In yet another example, target A could be a first set of structures, target B could be a second set of structures, and the first and second sets of structures could contain the same or different structures.

[0052] Two approaches can be used for transfer learning. The first model may be static except for one or more of the last layers, for example, by fixing parameters that have been pre-trained on measurement data 302 for target A. In the other approach, one or more of the first few layers of the model may have greater flexibility and may be trained at a lower learning rate using measurement data 304 for target B. The knowledge learned from target A is transferred to target B through shared parameters in the learner model, i.e., one or more of the first few layers are kept static or retrained at a lower learning rate, so this type of transfer learning is called parameter-based transfer learning.

[0053] As illustrated by scissors 332, 334, and 322, transfer learning can be applied in different layers within the regression predictor 330 or feature extractor 320 to produce an output layer 340 for a second ML model 360 with a small domain divergence, or for a third ML model 370 with a large domain divergence, for target B, illustrated as parameters P1 and P2. The parameters P1 and P2 for models 360 and 370 may be the same output parameters for model 350, for example. As domain divergence increases, generally more layers need to be retrained from the first model.

[0054] In some implementations, the transfer learning process 300 may use more than two targets. For example, multiple targets may be used to train the first ML model 350. In another example, an additional target (target C (not shown)) may be used to train an additional ML model, for example, training the first ML model 350 using measurement data 302 for target A, and training a third ML model using an additional set of measurement data for target C. The knowledge of the first ML model 350 and the knowledge of the third ML model can be transferred to a second ML model (e.g., model 360 / 370) trained using measurement data 304 from target B. In some implementations, different segments from the machine learning model may be transferred to the second machine learning model. For example, one or more of the first few layers from the first ML model 350 may be statically maintained or trained at a smaller learning rate using the measurement data 304 from target B of the second ML model 360 / 370, and one or more intermediate layers or one or more of the last few layers from the third ML model (for target C) may be statically maintained or trained at a smaller learning rate using the measurement data 304 from target B of the second ML model 360 / 370. The training of the first ML model 350 using the measurement data 302 for target A and the training of the third ML model using an additional set of measurement data for target C may be completely independent, or transfer learning may be performed between them. In addition, in some implementations, the first ML model 350 and the second ML model (360 / 370) may be used to train the third ML model, for example, by transferring the learning to an additional set of measurement data for target C. In some implementations, different segments from a machine learning model can be transferred to a third machine learning model.For example, one or more of the first few layers from the first ML model 350 may be statically maintained or trained at a smaller learning rate using measurement data from target C of the third ML model, and one or more intermediate layers or one or more of the last few layers from the second ML model (of target B) may be statically maintained or trained at a smaller learning rate using measurement data from target C of the third ML model.

[0055] In some implementations, the first ML model and the second ML model can be optimized simultaneously and merged into a single ML model.

[0056] In another implementation, transfer learning with domain adaptation may be used for measurement data, as discussed with reference to training machine learning models using ML optimization 256 shown in Figure 2B and transfer learning 276 in Figure 2C.

[0057] Domain adaptation is a transfer learning scenario in which the source and target domains have different feature spaces and distributions. Using domain adaptation, the process adapts one or more source domains to transfer information and improve the performance of training a target ML model. Therefore, this process generally involves changing the semantic representations of both the source and target domains to bring the source and target distributions closer together.

[0058] Figure 4 illustrates an example of a source domain 402 and a target domain 404 in spectral form. Although two domains are illustrated in Figure 4, it should be understood that there may be more than two domains. The source domain 402 may be, for example, a simulated (synthetic) spectrum, an experimental spectrum, or a combination thereof, and these spectra may be labeled, unlabeled, or a combination thereof. The target domain 404 may be an experimental spectrum, a simulated (synthetic) spectrum, or a combination thereof, and these spectra may be labeled, unlabeled, or a combination thereof, as described in relation to obtaining the first set of measurement data 272 and the second set of data 274 in Figure 2C. For example, in one implementation, the source domain may have labeled synthetic data or experimental data, and the target domain may have unlabeled experimental data. In this example, a regressor can be trained using target domain data, the purpose of the unlabeled data from the source domain is to facilitate the invariance of common features between the source and target domains, and the source domain covers larger process variations from the unlabeled data. In another implementation, the source domain may have labeled synthetic or experimental data, and the target domain may have unlabeled experimental data. The target domain may also contain a limited number of labeled data. In this example, a regressor can be trained using a combination of source domain data and labeled data from the target domain, if available, the purpose of the unlabeled data from the target domain is to facilitate the invariance of common features between the source and target domains, the target domain covers larger process variations from the unlabeled data, and has the advantage of reduced requirements compared to the reference data. It should be understood that source domain 402 and target domain 404 are not limited to spectra, but other measurement data may be used as desired.The source domain 402 and the target domain 404 may be similar, but may have different distributions due to misfits during, for example, the initial model setup and real-time fitting 252. In addition, if both the source and target contain experimental data, differences in distribution may be due to different fabrication processes. Therefore, if the source and target domains are collected from different structures with common key parameters, domain differences may be due to structural differences.

[0059] In addition to using regression with supervised machine learning where the key parameter is a continuous real value, classification may be used with supervised machine learning, in which case the key parameter indicates a categorization for diagnosing good (likely to pass electrical tests) / bad (likely to fail electrical tests) dies.

[0060] Figure 5 illustrates an example of transfer learning with a domain adaptation process 500 for usable measurement data, as discussed with reference to training a machine learning model using ML optimization 256 shown in Figure 2B and transfer learning 276 in Figure 2C. The transfer learning process 500 with domain adaptation shown in Figure 5 does not necessarily require that all domains have labeled data; for example, only one domain may have at least some labels. As illustrated, as described with reference to obtaining a first set of measurement data 272 and obtaining a second set of measurement data 274 in Figure 2C, measurement data 502, which may be labeled and / or unlabeled simulated and / or experimental measurement data, and measurement data 504, which may be labeled and / or unlabeled simulated and / or experimental measurement data, are provided to the feature extractor 510. Figure 5 illustrates the measurement data 502 and 504 as spectra, as discussed with reference to Figure 4, but the measurement data 502 and 504 are not limited to spectra and may be other measurement data. The feature extractor 510 may be any linear or nonlinear feature extraction architecture, including but not limited to convolutional neural networks, shallow neural networks, deep learning, or any combination thereof. The feature extractor 510 is coupled to a regression predictor 520, which provides parameters in the output layer 550, such as structural or layer parameters like limiting dimension (CD), height, thickness, depth, line width, space width, sidewall angle, etching recess, slope, overlay, surface roughness, and line edge roughness, or material properties like doping concentration, composition, crystallinity, and conductivity. The feature extractor 510 selects measurement data from measurement data 502 and measurement data 504 used by the regression predictor 520. In implementations where the domain classifier 530 is not used, minimizing the deviation between the source domain and the target domain is a result of cotraining (i.e., minimizing the regressor loss).

[0061] In some implementations, the feature extractor 510 is coupled to the domain classifier 530 via a gradient inversion layer 540, i.e., to carry out a feature-based transfer learning process. The domain classifier 530 and gradient inversion layer 540 facilitate the selection of "good" features from the measurement data, which can, for example, (i) minimize the regression model error for the main training task to predict key parameters, and (ii) minimize the difference in measurement data between measurement data 502 (source), which may be simulated data or experimental measurement data, and measurement data 504 (target), which may be simulated data or experimental measurement data.

[0062] In some implementations, transfer learning using the domain adaptation process 500 may use more than two targets, for example, by using three or more sets of measurement data. For example, in the case of domain adaptation using three or more sets of measurement data, the domain classifier 530 performs the task of multi-class classification instead of binary classification in the case of two sets of measurement data (as illustrated in Figure 5). In implementations using co-training, for example, when the domain classifier 530 is bypassed, using three or more sets of measurement data is equivalent to considering one or more first structures to contain multiple structures having multiple sets of measurement data.

[0063] In some implementations, the transfer learning process for measured data may include multiple types of transfer learning through hybridization. For example, the parameter-based transfer learning process 300 shown in Figure 3 may be combined with the domain adaptation-based transfer learning process 500 shown in Figure 5. For example, the first ML model 350 and / or the second ML models 360 / 370 may include a domain classifier referenced in the feature-based transfer learning using the domain adaptation process illustrated in Figure 5. In an example of a transfer learning process performed for three targets, one or more layers from the feature extractor 320 may be transferred from the first ML model 350, and one or more layers from the domain classifier or regression predictor 330 may be transferred from the third ML model (for target C) to the second ML model (for target B). Reweighted labeled samples from all domains may be used as input to a model using domain adaptation with instance-based transfer learning. As a further example, a transfer learning process for measurement data may include two or more of the following, as discussed herein: parameter-based transfer learning, feature-based transfer learning, and instance-based transfer learning.

[0064] For example, a transfer learning process for measurement data, as illustrated in Figures 3 and 5, can be used to transfer data between a source domain and a target domain in at least the following scenarios: from synthetic data to synthetic and / or experimental data (labeled and / or unlabeled) for a single structure from the same or different processes; from labeled experimental data to unlabeled synthetic and / or experimental data for a single structure from the same or different processes; from synthetic and labeled experimental data (co-trained) for a single structure from the same or different processes to synthetic and / or experimental data (labeled and / or unlabeled); or from synthetic and / or experimental data (labeled and / or unlabeled) for one or more structures from the same or different processes to synthetic and / or experimental data (labeled and / or unlabeled) for one or more similar but not identical structures.

[0065] Figure 6 shows an illustrative flowchart illustrating exemplary operation 600 for supporting non-contact measurement of a structure in several implementations. In some implementations, exemplary operation 600 may be performed by one or more processors, such as at least one processor 162 in at least one computing system 160 in Figure 1. Non-contact measurement may be, but is not necessarily limited to, optical measurement using measurement data, for example, optical measurement data. For example, in some implementations, non-contact measurement may be X-ray measurement using X-ray measurement data, or any other desired non-contact measurement in which radiation is used, for example.

[0066] One or more processors may obtain a first set of measurement data for one or more first structures (602). For example, the first set of measurement data may be synthetic measurement data generated based on modeling one or more reference structures as described with respect to the automated synthetic spectrum generation 254 in the workflow 250 shown in Figure 2B and obtaining the first set of measurement data shown in Figure 2C 272, or it may be measurement data 302 as considered with reference to Figure 3. In another example, the first set of measurement data may be experimental measurement data produced by a measurement device such as the measurement device 100 shown in Figure 1, for example, as referenced in the workflow 250 shown in Figure 2B and obtaining the first set of measurement data shown in Figure 2C 272, or it may be measurement data 302 as considered with reference to Figure 3. In some implementations, the first set of measurement data may be a combination of synthetic measurement data and experimental measurement data. The first set of measurement data may be labeled, unlabeled, or a combination thereof.

[0067] One or more processors may use a first set of measurement data to train a first machine learning model on one or more first structures (604). For example, a first machine learning model 350, as illustrated in Figure 3, may be trained using measurement data 302.

[0068] One or more processors may acquire a second set of measurement data for one or more second structures (606). In some implementations, one of the first set of one or more structures or one of the second set of one or more structures may have greater variability in structural parameters, or layer property parameters, or material property parameters, or combinations thereof, than the other structures of the first set of one or more structures or the second set of one or more structures. For example, the first set of one or more structures may be the result of greater manufacturing process variability than the structures used for the second set of one or more structures. For example, the first set of one or more structures may have greater variability in structural parameters, or layer property parameters, or material property parameters, or combinations thereof, than the second set of one or more structures. In another example, the second set of one or more structures may have greater variability in structural parameters, or layer property parameters, or material property parameters, or combinations thereof, than the first set of one or more structures. The second set of measurement data may be experimental measurement data produced by a measurement device, such as the measurement device 100 shown in Figure 1, as referenced in ML optimization 256 in workflow 250 shown in Figure 2B, to obtain the second set of measurement data shown in Figure 2C 274, or it may be measurement data 304 as considered with reference to Figure 3. The second set of measurement data may be synthetic measurement data generated based on modeling the second one or more second structures, as referenced in obtaining the second set of measurement data shown in Figure 2C 274, or it may be measurement data 304 as considered with reference to Figure 3. In some implementations, the second set of measurement data may be a combination of synthetic measurement data and experimental measurement data. The second set of measurement data may be labeled, unlabeled, or a combination thereof.

[0069] One or more processors perform transfer learning from the first machine learning model to a second set of measurement data to construct a second machine learning model for predicting key parameters for one or more second structures (608). Transfer learning from the first machine learning model to a second set of measurement data for constructing a second machine learning model for predicting key parameters for one or more second structures is considered, for example, in ML optimization 256 in the workflow 250 shown in Figure 2B, training of the machine learning model using transfer learning 276 shown in Figure 2C, and the transfer learning process 300 shown in Figure 3.

[0070] In some implementations, the first set of measurement data may be synthetic measurement data generated from one or more models of one or more first structures, experimental measurement data generated from one or more first structures, or a combination thereof.

[0071] In some implementations, the second set of measurement data may be synthetic measurement data generated from one or more models of one or more second structures, experimental measurement data generated from one or more second structures, or a combination thereof.

[0072] In some implementations, at least a portion of the first set of measurement data is labeled, and at least a portion of the second set of measurement data is labeled.

[0073] In some implementations, the first one or more structures may be a single structure or a set of structures, and the second one or more structures may be a single structure or a set of structures.

[0074] In some implementations, the first one or more structures and the second one or more structures are either different types of structures, or they are the same type of structures created using the same or different processes.

[0075] In some implementations, the first one or more structures may be a first single structure, and the second one or more structures may be a second single structure. In some implementations, the first and second single structures may be structures of the same type, produced using the same or different processes. In some implementations, the first and second single structures may be structures of different types.

[0076] In some implementations, the first set of one or more structures may be a first set of structures, and the second set of one or more structures may be a second set of structures. In some implementations, the first set of structures and the second set of structures may be structures of the same type produced using the same or different processes. In some implementations, the first set of structures and the second set of structures may be structures of different types. In some implementations, a set of structures may contain a single structure, while in other implementations, a set of structures may contain multiple structures; for example, the first set of one or more structures may be a single structure and the second set of one or more structures may be multiple structures, or conversely, the first set of one or more structures may be multiple structures and the second set of one or more structures may be a single structure.

[0077] In one implementation, one or more processors may further acquire a third set of measurement data for one or more third structures and use the third set of measurement data to train a third machine learning model for one or more third structures. One or more processors may further perform transfer learning (simultaneously or sequentially) from the third machine learning model to the second set of measurement data, together with the first machine learning model, to produce a second machine learning model for predicting key parameters for one or more second structures. For example, the first and third machine learning models may be transferred to the second machine learning model by transferring segments (layers or combinations of layers). In some implementations, different segments from the machine learning models may be transferred to the second machine learning model; for example, one or more layers from a feature extractor may be transferred from the first machine learning model, and one or more layers from a domain classifier may be transferred from the third machine learning model.

[0078] In one implementation, one or more processors may acquire a third set of measurement data for one or more third structures, perform transfer learning from the first and second machine learning models to the third set of measurement data, and construct a third machine learning model for predicting key parameters for one or more third structures. In some implementations, different segments from the machine learning models may be transferred to the third machine learning model; for example, one or more layers from the feature extractor may be transferred from the first machine learning model, and one or more layers from the domain classifier may be transferred from the second machine learning model.

[0079] Figure 7 shows an illustrative flowchart illustrating exemplary operations 700 for supporting non-contact measurement of structures in several implementations. In some implementations, exemplary operations 700 may be performed by one or more processors, such as at least one processor 162 in at least one computing system 160 in Figure 1. Non-contact measurement may be, but is not necessarily limited to, optical measurement using measurement data, for example, optical measurement data. For example, in some implementations, non-contact measurement may be X-ray measurement using X-ray measurement data, or any other desired non-contact measurement in which radiation is used, for example.

[0080] One or more processors may obtain a first set of measurement data for one or more first structures as a source domain (702). For example, the first set of measurement data may be synthetic measurement data generated based on modeling one or more reference structures as described in the automated synthetic spectrum generation 254 in the workflow 250 shown in Figure 2B and obtaining the first set of measurement data shown in Figure 2C 272, or measurement data of source domain 402, or measurement data 502 (source) as considered with reference to Figures 4 and 5. In another example, the first set of measurement data may be experimental measurement data produced by a measurement device such as the measurement device 100 shown in Figure 1, which measures one or more reference structures and obtains the first set of measurement data 272 shown in Figure 2C, as referenced in the workflow 250 shown in Figure 2B, or measurement data of source domain 402, or measurement data 502 (source) as considered with reference to Figures 4 and 5. In some implementations, the first set of measurement data may be a combination of synthetic measurement data and experimental measurement data. The first set of measurement data may be labeled, unlabeled, or a combination of both.

[0081] One or more processors may obtain a second set of measurement data for one or more second structures (704). The second set of measurement data may be experimental measurement data produced by a measurement device, such as the measurement device 100 shown in Figure 1, as referenced in ML optimization 256 in workflow 250 shown in Figure 2B, to obtain the second set of measurement data shown in Figure 2C 274, or it may be measurement data of the target domain 404, or measurement data 504 (target) as considered with reference to Figures 4 and 5. The second set of measurement data may be synthetic measurement data generated based on modeling one or more second structures, as referenced when obtaining the second set of measurement data 274 shown in Figure 2C, and may be measurement data of the target domain 404, or measurement data 504 (target) as considered with reference to Figures 4 and 5, respectively. In some implementations, the second set of measurement data may be a combination of synthetic measurement data and experimental measurement data. The second set of measurement data may be labeled, unlabeled, or a combination of both.

[0082] One or more processors use a feature extractor to select measurement data from a first set of measurement data and a second set of measurement data (706). The selection of measurement data may be performed using a feature extractor 510, as discussed with reference to Figure 5.

[0083] One or more processors train a machine learning model using selected measurement data to predict key parameters for one or more second structures, for example, as referenced in training a machine learning model using transfer learning 276 shown in Figure 2C (708). For example, a machine learning model such as the one illustrated in Figure 5 may be trained using selected measurement data.

[0084] In some implementations, one or more processors minimize the domain difference between a first set of measurement data and a second set of measurement data, as considered with reference to Figure 5. For example, one or more processors may minimize the domain difference using a domain classifier 530 via a gradient inversion layer 540, as shown in Figure 5.

[0085] In some implementations, one or more processors minimize domain differences by co-training on a first set of measurement data and a second set of measurement data, as discussed with reference to Figure 5.

[0086] In some implementations, the first set of measurement data may be synthetic measurement data generated from one or more models of one or more first structures, experimental measurement data generated from one or more first structures, or a combination thereof.

[0087] In some implementations, the second set of measurement data may be synthetic measurement data generated from one or more models of one or more second structures, experimental measurement data generated from one or more second structures, or a combination thereof.

[0088] In some implementations, the first set of measurement data is at least partially labeled, and the second set of measurement data is labeled, unlabeled, or a combination of both. In some implementations, the second set of measurement data is at least partially labeled, and the first set of measurement data is labeled, unlabeled, or a combination of both.

[0089] In some implementations, the first one or more structures may be a single structure or a set of structures, and the second one or more structures may be a single structure or a set of structures.

[0090] In some implementations, the first one or more structures and the second one or more structures are either different types of structures, or they are the same type of structures created using the same or different processes.

[0091] In some implementations, the first one or more structures may be a first single structure, and the second one or more structures may be a second single structure. In some implementations, the first and second single structures may be structures of the same type, produced using the same or different processes. In some implementations, the first and second single structures may be structures of different types.

[0092] In some implementations, the first set of one or more structures may be a first set of structures, and the second set of one or more structures may be a second set of structures. In some implementations, the first set of structures and the second set of structures may be structures of the same type produced using the same or different processes. In some implementations, the first set of structures and the second set of structures may be structures of different types. In some implementations, a set of structures may contain a single structure, while in other implementations, a set of structures may contain multiple structures; for example, the first set of one or more structures may be a single structure and the second set of one or more structures may be multiple structures, or conversely, the first set of one or more structures may be multiple structures and the second set of one or more structures may be a single structure.

[0093] In some implementations, one or more processors acquire a third set of measurement data for one or more third structures, and use a feature extractor to select measurement data from the third set of measurement data, along with the first and second sets of measurement data.

[0094] The above description is intended to be illustrative and not restrictive. For example, the above examples (or one or more of them) may be used in combination with each other. Other implementations may be used, for example, by those skilled in the art when considering the above description. Also, various features may be grouped together, and fewer features may be used than all features of a particular disclosed implementation. Accordingly, the following embodiments are incorporated herein as examples or implementations into the above description, and each embodiment stands on its own as a separate implementation, and such implementations are intended to be combined with each other in various combinations or rearrangements. Accordingly, the intent and scope of the appended claims should not be limited to the above description.

Claims

1. A method for supporting non-contact measurement of structures, Obtaining a first set of measurement data for one or more first structures, Using the first set of measurement data, a first machine learning model is trained on one or more of the first structures. Obtain a second set of measurement data for one or more second structures, A method comprising performing transfer learning from the first machine learning model to a second set of measurement data to produce a second machine learning model for predicting key parameters for one or more second structures.

2. The method according to claim 1, wherein one of the first one or more structures or the second one or more structures includes greater variation in structural parameters, layer property parameters, material property parameters, or combinations thereof than the other of the first one or more structures or the second one or more structures.

3. The method according to claim 1, wherein the first set of measurement data includes synthetic measurement data generated from one or more models of the first one or more structures, experimental measurement data generated from the first one or more structures, or a combination thereof, and the second set of measurement data includes synthetic measurement data generated from one or more models of the second one or more structures, experimental measurement data generated from the second one or more structures, or a combination thereof.

4. The method according to claim 1, wherein at least a portion of the first set of measurement data is labeled, and at least a portion of the second set of measurement data is labeled.

5. The method according to claim 1, wherein the first one or more structures and the second one or more structures are different types of structures, or are the same type of structures produced using the same or different processes.

6. Obtain a third set of measurement data for one or more third structures, Training a third machine learning model for one or more of the third structures using the third set of measurement data, Transfer learning is performed from the third machine learning model to the second set of measurement data together with the first machine learning model to produce the second machine learning model for predicting key parameters for one or more second structures. The method according to claim 1, further comprising:

7. The method according to claim 6, wherein different layers from the first machine learning model and the third machine learning model are transferred to the second machine learning model.

8. Obtain a third set of measurement data for one or more third structures, Transfer learning is performed from the first machine learning model and the second machine learning model to a third set of measurement data to create a third machine learning model for predicting key parameters for one or more of the third structures. The method according to claim 1, further comprising:

9. The method according to claim 8, wherein different layers from the first machine learning model and the second machine learning model are transferred to the third machine learning model.

10. A computer system configured to support non-contact measurement of a structure, It comprises at least one processor, and the at least one processor is Obtain a first set of measurement data for one or more first structures, A first machine learning model is trained on one or more of the first structures using the first set of measurement data. Obtain a second set of measurement data for one or more second structures. A computer system configured to perform transfer learning from the first machine learning model to a second set of measurement data to produce a second machine learning model for predicting key parameters for one or more second structures.

11. The computer system according to claim 10, wherein one of the first one or more structures or the second one or more structures includes greater variation in structural parameters, layer property parameters, material property parameters, or combinations thereof than the other of the first one or more structures or the second one or more structures.

12. The computer system according to claim 10, wherein the first set of measurement data includes synthetic measurement data generated from one or more models of the first one or more structures, experimental measurement data generated from the first one or more structures, or a combination thereof, and the second set of measurement data includes synthetic measurement data generated from one or more models of the second one or more structures, experimental measurement data generated from the second one or more structures, or a combination thereof.

13. The computer system according to claim 10, wherein at least a portion of the first set of measurement data is labeled, and at least a portion of the second set of measurement data is labeled.

14. The computer system according to claim 10, wherein the first one or more structures and the second one or more structures are different types of structures, or are the same type of structures produced using the same or different processes.

15. The aforementioned at least one processor, A third set of measurement data is obtained for one or more third structures. A third machine learning model is trained on one or more of the third structures using the third set of measurement data. The computer system according to claim 10, further configured to perform transfer learning from the third machine learning model to a second set of measurement data together with the first machine learning model to produce the second machine learning model for predicting key parameters for one or more second structures.

16. The computer system according to claim 15, wherein different layers from the first machine learning model and the third machine learning model are transferred to the second machine learning model.

17. The aforementioned at least one processor, A third set of measurement data is obtained for one or more third structures. The computer system according to claim 10, further configured to perform transfer learning from the first machine learning model and the second machine learning model to a third set of measurement data to produce a third machine learning model for predicting key parameters for one or more third structures.

18. The computer system according to claim 17, wherein different layers from the first machine learning model and the second machine learning model are transferred to the third machine learning model.

19. A method for supporting the measurement of a structure, wherein the method is Obtaining a first set of measurement data for one or more first structures, Obtain a second set of measurement data for one or more second structures, Using a feature extractor, select measurement data from the first set of measurement data and the second set of measurement data, A method comprising: training a machine learning model using selected measurement data to predict key parameters for one or more of the second structures.

20. The method according to claim 19, further comprising minimizing the domain difference between the first set of measurement data and the second set of measurement data.

21. The method according to claim 20, wherein the domain difference is minimized using a domain classifier via a gradient inversion layer.

22. The method according to claim 19, further comprising minimizing domain differences by co-training based on a first set of measurement data and a second set of measurement data.

23. The method according to claim 19, wherein the first set of measurement data includes synthetic measurement data generated from one or more models of the first one or more structures, experimental measurement data generated from the first one or more structures, or a combination thereof, and the second set of measurement data includes synthetic measurement data generated from one or more models of the second one or more structures, experimental measurement data generated from the second one or more structures, or a combination thereof.

24. The method according to claim 19, wherein one set of measurement data, which includes either the first set of measurement data or the second set of measurement data, is at least partially labeled, and the remaining set of measurement data is labeled, unlabeled, or a combination thereof.

25. The method according to claim 19, wherein the first one or more structures and the second one or more structures are different types of structures, or are the same type of structures produced using the same or different processes.

26. Obtain a third set of measurement data for one or more third structures, Using the feature extractor, select measurement data from the third set of measurement data, along with the first set and the second set of measurement data. The method according to claim 19, further comprising:

27. A computer system configured to support non-contact measurement of a sample, It comprises at least one processor, and the at least one processor is Obtain a first set of measurement data for one or more first structures, Obtain a second set of measurement data for one or more second structures. Using a feature extractor, select measurement data from the first set of measurement data and the second set of measurement data. A computer system configured to train a machine learning model using selected measurement data in order to predict key parameters for one or more of the second structures.

28. The computer system according to claim 27, wherein the at least one processor is further configured to minimize the domain difference between the first set of measurement data and the second set of measurement data.

29. The computer system according to claim 28, wherein the domain difference is minimized using a domain classifier via a gradient inversion layer.

30. The computer system according to claim 27, wherein the at least one processor is further configured to minimize domain differences by co-training on a first set of measurement data and a second set of measurement data.

31. The computer system according to claim 27, wherein the first set of measurement data includes synthetic measurement data generated from one or more models of the first one or more structures, experimental measurement data generated from the first one or more structures, or a combination thereof, and the second set of measurement data includes synthetic measurement data generated from one or more models of the second one or more structures, experimental measurement data generated from the second one or more structures, or a combination thereof.

32. The computer system according to claim 27, wherein one set of measurement data, including either the first set of measurement data or the second set of measurement data, is at least partially labeled, and the remaining set of measurement data is labeled, unlabeled, or a combination thereof.

33. The computer system according to claim 27, wherein the first one or more structures and the second one or more structures are different types of structures, or are the same type of structures produced using the same or different processes.

34. The aforementioned at least one processor, A third set of measurement data is obtained for one or more third structures. The computer system according to claim 27, further configured to use the feature extractor to select measurement data from a third set of measurement data, together with a first set of measurement data and a second set of measurement data.

Images