In-die overlay estimation using tool-induced shift correction

Abstract

The metrology module includes an estimation model configured to provide an estimate of independent overlay due to tool-induced shift on a received wafer based on only one azimuth spectrum. The estimation model can use at least one machine learning algorithm. The estimation model can be derived by a machine learning algorithm applied to training data calculated based on a first training sample set from initial metrology measurements and additional tool-induced shift training samples.

Description

[Technical Field]

[0001] The present disclosure relates to semiconductor metrology. [Background technology]

[0002] Evolution in the semiconductor manufacturing industry is placing greater demands on yield management, particularly on metrology and inspection systems. As critical dimensions continue to shrink, the industry must shorten the time to achieve high-yield, high-value production. Minimizing the total time from detecting a yield problem to fixing it maximizes the return on investment for semiconductor manufacturers.

[0003] Fabricating semiconductor devices, such as logic and memory devices, typically involves processing semiconductor wafers using multiple manufacturing processes to form various features and levels of the semiconductor devices. For example, lithography is a semiconductor manufacturing process that involves transferring a pattern from a reticle to a photoresist disposed on a semiconductor wafer. Further examples of semiconductor manufacturing processes include, but are not limited to, chemical mechanical polishing (CMP), etching, deposition, and ion implantation. An array of multiple semiconductor devices fabricated on a single semiconductor wafer can be separated into individual semiconductor devices.

[0004] Metrology processes are used at various steps during semiconductor manufacturing to monitor and control the process. Metrology processes differ from inspection processes, in that defects are detected on wafers and metrology processes are used to measure one or more characteristics of wafers that cannot be determined using existing inspection tools. Metrology processes can be used to measure one or more characteristics of a wafer so that the performance of the process can be determined from the one or more characteristics. For example, a metrology process can measure the dimensions (e.g., linewidth, thickness, etc.) of features formed on a wafer during the process. Furthermore, if one or more characteristics of a wafer are unacceptable (e.g., outside a predetermined range of characteristics), the measurements of the one or more characteristics of the wafer can be used to modify one or more parameters of the process so that additional wafers produced by the process have acceptable characteristics.

[0005] Tool-induced shift (TIS) is a measurement error resulting from tool asymmetry issues. TIS is commonly used to measure metrology tool accuracy in the semiconductor industry. Overlay (OVL) measurement inaccuracies are typically caused by lens aberrations, lens alignment, illumination alignment, or asymmetries on the measured target. TIS affects total measurement uncertainty (TMU) and tool-to-tool matching, and TIS variation across a wafer can explain the inaccuracy because it depends on incoming process conditions if not fully corrected. In addition, both lot-to-lot process variation and wafer-to-wafer process variation are affected by TIS in terms of overlay performance, which also includes metrology tool-to-tool efficiency in terms of throughput.

[0006] Typical metrology measurements of targets on wafer sites estimate the overlay between target structures in images of the target, typically for two images per target (0° and 180° rotated images), to achieve good accuracy. Previously, overlay measurements relied on models or specially designed targets, which used information extracted from the measured images, including noise levels across the target and overlay accuracy metrics, to find the ground truth for TIS. Specially designed targets were typically required. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] U.S. Patent Application Publication No. 2019 / 0178630 Summary of the Invention [Problem to be solved by the invention]

[0008] Improved techniques are needed to improve throughput and provide better performance. [Means for solving the problem]

[0009] In a first embodiment, a method is provided. The method includes determining a first training sample set from initial metrology measurements. The initial metrology measurements include spectra at two azimuth angles and, optionally, skew. Additional tool-induced shift training samples are prepared by adding a negative value of a reference to the opposite azimuth angle data from the first training sample set. The original azimuth angle data from the first training sample set and the opposite angle data in a second training sample set are stacked. At least one machine learning algorithm is applied to the first training sample set and the second training sample set to derive an estimation model. An independent overlay is obtained using a tool-induced shift correction embedded in an in-line wafer collected for a spectrum at one azimuth angle. The estimation model is used to provide an estimate of the independent overlay due to tool-induced shift.

[0010] Additional tool-induced shift training samples can serve as a benchmark for approaches based on external measurements.

[0011] An additional tool-induced shift training sample can be a self-calibrating overlay for a reference-free approach.

[0012] The machine learning algorithm can be one of a linear model, a neural network, or a convolutional network.

[0013] The machine learning algorithm can be a model-assisted approach, a model-less approach, a reference-based approach, or a self-calibrating recipe. For example, tool-induced shift signals can be used to calibrate system parameters.

[0014] The spectra at two azimuthal angles can be collected from a spectroscopic ellipsometer, soft x-ray reflectometer, small-angle x-ray scatterometer, or imaging system.

[0015] The initial metrology measurements may include tilt in the X direction and in the Y direction perpendicular to the X direction.

[0016] Using the estimation model can include determining a non-zero overlay and / or edge placement error.

[0017] The machine learning algorithm can be further trained using data with critical dimensions, heights, sidewall angles, or film variances.

[0018] Independent overlay with tool-induced shift can be determined for multiple tools.

[0019] The two azimuth angles can be separated by 180°.

[0020] At least one of determining, preparing, stacking, applying, or using can be obtained by at least one processor.

[0021] The computer-readable medium may store a program configured to instruct the processor to perform the method of the first embodiment. The instrumentation module may include a non-transitory computer-readable medium.

[0022] In a second embodiment, a system is provided. The system includes a metrology module including a processor. The metrology module includes an estimation model configured to provide an estimate of independent overlay due to tool-induced shift on a received wafer. The estimation model uses at least one machine learning algorithm. The estimation is based on only one azimuthal angle spectrum, and the estimation model is based on spectra at two azimuthal angles.

[0023] In one example, the estimation model is derived by a machine learning algorithm applied to training data calculated based on a first training sample set from initial metrology measurements and additional tool-induced shift training samples. The initial metrology measurements include spectra at two azimuth angles and, optionally, skew. The additional tool-induced shift training samples are at azimuth angles opposite to the first training sample set having a negative value of the reference.

[0024] In an example, the processor is further configured to stack the original azimuth angle data from the first training sample set with the opposite azimuth angle data in the second training sample set to obtain an independent overlay with tool-induced shift correction embedded in the in-line wafer collected for the spectrum at only one azimuth angle.

[0025] In some examples, the machine learning algorithm is further trained using critical dimensions, height, sidewall angle, or film variance.

[0026] In one example, the two azimuth angles are 180° apart.

[0027] The machine learning algorithm can be one of a linear model, a neural network, or a convolutional network.

[0028] The machine learning algorithm can be a model-assisted approach, a model-less approach, a criteria-based approach, or a self-calibrating recipe.

[0029] The metrology module may further be configured to determine non-zero overlay and / or edge placement error. [Brief explanation of the drawings]

[0030] For a fuller understanding of the nature and objects of the present disclosure, reference should be made to the following detailed description taken in conjunction with the accompanying drawings. [Figure 1] FIG. 1 is a block diagram of a system having a metrology module according to one embodiment of the present disclosure. [Figure 2] 1 is a flowchart of a method according to one embodiment of the present disclosure. [Figure 3] 10 shows an example overlay map representing real data comparing 2AZ technology with 1AZ machine learning recipe trained using TIS correction and TIS error. [Figure 4] FIG. 10 shows an example TIS error map representing actual data on two different wafers and different recipes. [Figure 5] FIG. 10 shows an example overlay map comparison representing actual data between before (top) and after (bottom) TIS correction outside the training process range using Recipe A and Recipe B when TIS is significant. [Figure 6] An exemplary overlay map comparison is shown between before TIS correction (top) and after TIS correction (bottom) representing actual data on a blind process of a recording wafer using two different layers (layer 1 and layer 2) when TIS is not significant. DETAILED DESCRIPTION OF THE INVENTION

[0031] Although the claimed subject matter is described in terms of certain embodiments, other embodiments, including embodiments that do not provide all of the benefits and features described herein, are also within the scope of this disclosure. Various structural, logical, process step, and electronic changes may be made without departing from the scope of this disclosure. Accordingly, the scope of this disclosure is defined solely by reference to the appended claims.

[0032] The embodiments disclosed herein provide a method for measuring independent overlay with TIS correction. This can be performed using standard metrology targets instead of specially designed metrology targets. These embodiments provide on-device measurements, which can avoid the need to find ground truth for TIS. The TIS-corrected overlay estimation can be performed quickly, which can avoid the need to retrain a library to correct for TIS. A single azimuth spectrum can be used for the TIS-corrected overlay estimation. Using a single azimuth spectrum can improve throughput by avoiding a second azimuth spectrum measurement.

[0033] The amount of TIS error can vary depending on the sample location within a wafer, across wafers, across lots, across tools, or using different machine learning recipes. Figure 4 shows an example of a TIS error map for different sample locations within a wafer, different wafers, and different recipes. As shown in Figure 4, it may not be feasible to estimate ground truth TIS for all samples across different wafers from a target, which requires a TIS training model when calculating overlay without TIS error or any other potential errors. The method disclosed herein can accurately measure in-die overlay using TIS error correction after a recipe has been trained.

[0034] When TIS error is negligible, the independent overlay from the base azimuth angle and the opposite angle has the same magnitude of the overlay value but opposite sign. When TIS error is introduced into the overlay measurement, the relationship is defined as follows: AZ180 and AZ0 represent azimuth angle measurements that are 180° opposite each other.

[0035]

number

[0036] By rearranging the above relationships, the TIS error and TIS correction overlay can be calculated using the two azimuth signals.

[0037]

number

[0038] This TIS correction can be applied to the inference stage of a machine learning recipe that has been trained without a TIS correction model.

[0039] To demonstrate this idea, we applied the TIS-corrected overlay to two different machine learning recipes that were trained without considering the TIS-corrected model. We selected a validated wafer outside the training process range to investigate whether the TIS-corrected overlay can alter the overlay map.

[0040] As shown in Figure 5, overlay measurements from two different recipes before TIS correction show different signatures outside the expected training process range due to different training of the machine learning module. Overlay measurements from two different recipes after TIS correction provide identical or nearly identical wafer maps and statistics, which were verified using separate external metrology. TIS correction allows the trained machine learning recipe to be more robust outside the training range. Another example, shown in Figure 6, demonstrates that when TIS error is negligible, overlay maps from both before and after TIS correction provide identical overlay maps and related statistics.

[0041] Collecting two azimuthal spectra on incoming in-line wafers can increase throughput by a factor of two compared to using a machine learning recipe without TIS correction with one azimuthal spectrum. To improve throughput, machine learning recipes with TIS correction can be trained as illustrated in FIG. 2. As shown in FIG. 2, two azimuthal spectra can be used in the training set to train the TIS model. First, two azimuthal spectra are collected on a skewed wafer (if any) and all training nominal wafers (e.g., process of record (FOR)). The skew is intentional and not due to the nature of the process. Training samples with TIS correction are then created with data augmentation. For example, in the OVLY (X) direction, the AZ180 (AZ90) spectrum is labeled with the original reference, and the AZ0 (AZ270) spectrum is labeled with the inverted OVL reference (i.e., -reference). In these examples, AZ is the azimuthal angle, and X and Y are vertical directions. By concatenating both the original (AZ180) and enhanced (AZ0) spectra, the number of training samples is doubled. In this way, both azimuth angle data are used as different samples to train the TIS model during recipe training. Therefore, the final recipe is created on the solution for one azimuth angle spectrum.

[0042] Method 200 describes this embodiment in more detail. Some or all of the steps of method 200 may be performed using a processor.

[0043] At 201, a first training sample set is determined from an initial metrology measurement. The initial metrology measurement includes spectra at two azimuthal angles and, optionally, a skew. For example, training spectra can be collected from a design of experiment (DOE) at 1AZ (e.g., AZ180), a skew DOE from 2AZ (e.g., AZ180+0), and a training nominal (FOR) wafer from 2AZ (AZ180-0). Thus, the two azimuthal angles can be separated by 180°.

[0044] The initial metrology measurements can include tilt in the X direction and in the Y direction, which is perpendicular to the X direction. By selecting a specific wavelength region that is only strongly sensitive to overlay, the tilt effect can be separated from the overlay measurement.

[0045] At 202, additional TIS training samples are prepared by adding the negative value of the reference for the opposite azimuth angle data from the first training sample set. For example, due to recording wafer skew and / or the training process, TIS-corrected samples can be generated for the opposite azimuth angle from the original azimuth angle (e.g., AZ180 reference - AZ0 reference) by multiplying the reference by -1. Creating additional TIS training samples (AZ0 with the inverted reference) may not affect the underlying measurement difference.

[0046] The additional TIS training samples can be references for external metrology-based approaches or self-calibrated overlays for reference-less approaches. Reference overlays or self-calibrated overlays can be negative values. External metrology-based approaches use dense reference data for each training nominal / skew wafer and can be trained against image-based approaches such as references including scanning electron microscopes (SEMs). Reference-less approaches do not use reference data to train the recipe. Self-calibrated overlays can be derived purely from the signal.

[0047] The original azimuth angle data from the first training sample set and the opposite angle data in the second training sample set may be concatenated or otherwise stacked at 203. This can double the number of training samples, as samples from both azimuth angles are stacked.

[0048] At 204, at least one machine learning algorithm is applied to the first training sample set and the second training sample set to derive an estimation model. The estimation model can operate on a recipe having one azimuth angle. The machine learning algorithm can be one of a linear model, a neural network, or a convolutional network, although other architectures are possible. In particular, the machine learning algorithm can be a neural network or a linear model.

[0049] In some examples, the machine learning algorithm is a model-assisted approach, a model-less approach, a reference-based approach, or a self-calibrating recipe. The TIS signal can be used to calibrate the system parameters.

[0050] In-line embedded TIS correction or independent overlay with other production wafers collected for spectra can be obtained using only one azimuthal angle at 205.

[0051] At 206, the estimation model is used to provide an estimate of independent overlay with the TIS. For example, the independent overlay can be determined using a recipe trained for a process on recorded wafers collected at only one azimuth angle (e.g., AZ180).

[0052] Using the estimation model may also include determining non-zero overlay and / or edge placement error.

[0053] The spectra at the two azimuthal angles can be collected from a spectroscopic ellipsometer, soft x-ray reflectometer, small angle x-ray scatterometer, imaging system, or other system.

[0054] In some examples, the machine learning algorithm is further trained using data involving critical dimensions, height, sidewall angle, or film variance. Estimation of parameters other than OVL can be used to detect process variations.

[0055] Independent overlays with TIS can be determined for multiple tools, which avoids the need to develop different machine learning strategies for multiple tools to train TIS errors.

[0056] To prove whether a single azimuth angle (1AZ) recipe training TIS model is sufficient to correct the TIS error, the single azimuth angle recipe solution was compared with a two azimuth angle (2AZ) solution on a blind wafer, as shown in Figure 3. As shown in Figure 3, the 1AZ recipe training TIS model performs similarly to the 2AZ model, but with improved throughput.

[0057] The TIS correction approach disclosed herein can be used to self-calibrate overlay machine learning recipes: instead of labeling the reference with a negative sign, the AZ0 spectrum on a skewed wafer can be labeled with a self-calibrating overlay derived from the asymmetric signal.

[0058] The TIS signal can be obtained using the sum of the first principal components of selected harmonics or Mueller element components in selected wavelength regions for AZ180 and AZ0. In some instances, the TIS signal is close to zero. Both the TIS signal and the TIS overlay error, TIS error = (OVL from AZ180 + OVL from AZ0) / 2, can be used to calibrate different system parameters, including the area of ​​interest (AOI), azimuth angle, wafer loading angle, or other parameters, by minimizing the TIS signal / TISOVL error.

[0059]

[0006] Embodiments of the present invention provide an efficient method and mechanism for performing metrology measurements in semiconductor manufacturing processes, which can improve semiconductor metrology. A metrology method, module, and system for using machine learning algorithms to improve metrology accuracy and overall process throughput are provided. The method disclosed herein includes calculating training data for metrology metrics from initial metrology measurements, applying a machine learning algorithm to the calculated training data to derive an estimation model for the metrology metrics, deriving metrology data from received images of sites on the wafer, and using the estimation model to provide an estimate of the metrology metric for the metrology data. While the training data may use two images per site, in operation, only one image per site may be required. This can reduce measurement time to less than half of current measurement times.

[0060] 1 is a block diagram of a system 100 having a metrology module 101 according to one embodiment of the present invention. The system 100 may include the metrology module 101 associated with and / or comprising at least one computer processor 102. Although shown as separate, the computer processor 102 may be part of the metrology module 101.

[0061] The metrology module 101 may comprise or be associated with one or more estimation models 120, which may be implemented on the computer processor 102 as part of the metrology module 101 or may be associated therewith (e.g., as a training or simulation module used to derive the estimation models 120). The estimation models 120 may be configured to provide estimates of at least one metrology metric for measurement data of sites on a received wafer. The estimation models 120 are derived by at least one machine learning algorithm 110, which is applied to the calculated training data 105 in a training phase 115 that associates at least one metrology metric with initial metrology measurements. The training phase 115 may include some of the steps in the method 200 of FIG. 2.

[0062] The metrology module 101 may be configured to calculate training data 105 for metrology metrics from initial metrology measurements. These metrology measurements may relate to multiple sites 85 within multiple fields 80 of one or more training wafers 70. The metrology module 101 may also apply machine learning algorithms 110 to the calculated training data 105 to derive estimation models 120 for metrology metrics, derive measurement data 122 from images of sites on the received wafers, and use the estimation models 120 to provide estimates of metrology metrics 124 for the measurement data 122.

[0063] For example, calculation of training data 105 may include using pairs of images 95 and 180° rotated images 95 derived from a site 85 on at least one training wafer 70. The training data 105 may include at least one processed feature 96 (as derived data 96) derived from each pair of images 95 for each site 85. For example, the processed feature may include a one-dimensional (ID) kernel representing an average along a specified direction in each image 95, and / or image transform data including a discrete cosine transform (DCT) or Fourier transform of each image 95.

[0064] In one embodiment, the training data 105 may include, at least in part, raw images 95 derived from multiple sites 85 on one or more training wafers 70 as initial metrology measurements.

[0065] The methods disclosed herein may be performed by the above-described system 100 and / or metrology module 101. For example, method 200 may be implemented using system 100 and / or metrology module 101. Method 200 may be implemented, at least in part, by at least one computer processor (e.g., 102), such as metrology module 101. Certain embodiments include a computer program product including a computer-readable storage medium having embodied thereon a computer-readable program configured to perform the relevant steps of method 200.

[0066] In one example, the metrology module 101 includes an estimation model 120 configured to provide an estimate of independent overlay with the TIS on the received wafer. The estimation model uses at least one machine learning algorithm 110. The estimation is based on only one azimuthal angle spectrum. The estimation model 120 can be based on spectra at two azimuthal angles relative to the surface of a training wafer 70. While a training wafer 70 is shown, a production wafer can also be used.

[0067] The estimation model 120 may be derived by the machine learning algorithm 110 applied to training data calculated based on a first training sample set from initial metrology measurements and additional TIS training samples. The initial metrology measurements include spectra at two azimuth angles and, optionally, skew. The additional TIS training samples are at azimuth angles opposite to the first training sample set having a negative value of the reference. The two azimuth angles may be separated by 180°. The at least one computer processor 102 may be configured to derive the estimation model 120.

[0068] The metrology module 101 may be further configured to stack the original azimuthal angle data from the first training sample set with the opposite angle data in the second training sample set to obtain an independent overlay with TIS corrections embedded in the in-line wafer collected for spectra at only one azimuthal angle.

[0069] The machine learning algorithm 110 can be further trained using critical dimensions, height, sidewall angle, or film variance.

[0070] The machine learning algorithm 110 can be one of a linear model, a neural network, or a convolutional network. The machine learning algorithm can use a model-assisted approach, a model-less approach, a reference-based approach, or a self-calibrating recipe.

[0071] The metrology module 101 may further be configured to determine non-zero overlay and / or edge placement error.

[0072] These computer program instructions may also be stored on a computer-readable medium that can direct a computer, other programmable data processing apparatus, or other device to function in a particular manner such that the instructions stored on the computer-readable medium can perform the steps of the methods disclosed herein.

[0073] Computer program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus, or other device, creating a computer-implemented process, such that the instructions executing on the computer or other programmable apparatus provide a process for implementing the functions / operations disclosed herein.

[0074] The system may further include an illumination system that illuminates a target (such as training wafer 70 or other wafers), a collection system that captures relevant information provided by the illumination system's interaction (or lack thereof) with the target, device, or feature, and a processing system (such as metrology module 101) that analyzes the collected information using one or more algorithms. System 100 can be used to measure structural and material properties related to various semiconductor manufacturing processes (e.g., material composition, dimensional properties of structures and films, e.g., film thickness and / or critical dimensions of structures, overlay, etc.). These measurements are used to facilitate process control and / or yield efficiency in the manufacture of semiconductor dies.

[0075] System 100 may include one or more hardware configurations that can be used in conjunction with certain embodiments of the present invention to, for example, measure various aforementioned semiconductor structure and material properties. Examples of such hardware configurations include, but are not limited to, a spectroscopic ellipsometer (SE), an SE with multiple illumination angles, an SE measurement Mueller matrix element (e.g., using a rotational compensator), a single-wavelength ellipsometer, a beam profile ellipsometer (angle-resolved ellipsometer), a beam profile reflectometer (angle-resolved reflectometer), a broadband reflectance spectrometer (spectroscopic reflectometer), a single-wavelength reflectometer, an angle-resolved reflectometer, an imaging system, or a scatterometer (e.g., a speckle analyzer).

[0076] Hardware configurations can be separated into separate operating systems. One or more hardware configurations can be combined into a single tool. U.S. Patent 7,933,026, incorporated herein by reference in its entirety, provides an example. Typically, numerous optical elements are present in such systems, including certain lenses, collimators, mirrors, λ / 4 wave plates, polarizers, detectors, cameras, apertures, and / or light sources. The wavelength of the optical system can vary from approximately 120 nm to 3 microns. For non-ellipsometer systems, the collected signal can be polarization-resolved or unpolarized. Multiple metrology tools can also be used for measurements on single or multiple metrology targets, such as those described in U.S. Patent 7,478,019, incorporated herein by reference in its entirety.

[0077] An illumination system of a particular hardware configuration can include one or more light sources. The light source may generate light having only one wavelength (i.e., monochromatic light), light having several discrete wavelengths (i.e., polychromatic light), light having multiple wavelengths (i.e., broadband light), and / or emit light that sweeps through wavelengths, either continuously or by hopping between wavelengths (e.g., using a tunable or swept light source). Examples of suitable light sources include white light sources, ultraviolet (UV) lasers, arc lamps or electrodeless lamps, laser-sustained plasma (LSP) sources, supercontinuum sources (such as broadband laser sources), or shorter wavelength sources such as x-ray sources, extreme ultraviolet sources, or some combination thereof. The light source may also be configured to provide light with sufficient brightness, which in some cases may be on the order of 1 W / (nm cm). 2 The brightness may exceed 1000 kJ / s (Sr). System 100 may also include high-speed feedback to the light source to stabilize its power and wavelength. The light source output may be delivered via free-space propagation, or in some cases, via an optical fiber or light guide.

[0078] System 100 can be designed to perform many different types of measurements related to semiconductor manufacturing. For example, system 100 can measure characteristics of one or more targets, such as critical dimensions, overlay, sidewall angle, film thickness, and process-related parameters (e.g., focus and / or dose). A target can include a particular region of interest that is periodic in nature, such as a grating in a memory die. A target can include multiple layers (or films) whose thicknesses can be measured by a metrology tool. A target can include target designs placed (or already present) on a semiconductor wafer for use in alignment and / or overlay registration operations, etc. A particular target can be located in various locations on a semiconductor wafer. For example, a target can be placed within a scribe line (e.g., between dies) and / or on the die itself. In some embodiments, multiple targets are measured (simultaneously or at different times) by the same or multiple metrology tools. Data from such measurements can be combined. Data from system 100 can be used in semiconductor manufacturing processes, for example, for feedforward corrections, feedback corrections, and / or feedside corrections to processes (e.g., lithography or etch).

[0079] As semiconductor device pattern dimensions continue to shrink, smaller metrology targets are often required. Furthermore, measurement accuracy and matching to actual device characteristics can increase the need for device-like targets as well as in-die and even on-device measurements. For example, focused-beam ellipsometry, primarily based on reflective optics, can be used. Apodizers can be used to mitigate the effects of optical diffraction, which causes the illumination spot to broaden beyond the size defined by geometric optics. High-numerical-aperture tools with simultaneous multiple-angle-of-incidence illumination can be used to achieve small-target throughput.

[0080] Other example measurements may include measuring the composition of one or more layers of a semiconductor stack, measuring specific defects on (or within) a wafer, or measuring the amount of photolithography radiation to which a wafer is exposed. In some cases, the system 100 and algorithms may be configured to measure non-periodic targets.

[0081] Measurement of a parameter of interest typically involves multiple algorithms. For example, the optical interaction of the incident beam with the sample is modeled using an electromagnetic (EM) solver, using algorithms such as rigorous coupled wave analysis (RCWA), finite element modeling (FEM), method of moments, surface integration, volume integration, and finite difference time domain (FDTD). The target of interest is typically modeled (parameterized) using a geometric engine, a process modeling engine, or a combination of both. Geometric engines are implemented, for example, in the AcuShape software product from KLA Corporation.

[0082] The collected data can be analyzed by several data fitting and optimization techniques and technologies, including libraries, fast reduced-order models, regression, machine learning algorithms, principal component analysis (PCA), independent component analysis (ICA), locally linear embedding (LEE), sparse representations such as Fourier transforms or wavelet transforms, Kalman filters, algorithms to facilitate matching from the same or different tool types, or others. The collected data can also be analyzed by algorithms that do not involve modeling, optimization, and / or fitting.

[0083] Computational algorithms are typically optimized for metrology applications using one or more approaches such as computational hardware design and implementation, parallelization, computation distribution, load balancing, multi-service support, or dynamic load optimization. Different implementations of the algorithms can be done in firmware, software, FPGAs, programmable optical components, etc.

[0084] The data analysis and fitting steps can have one or more objectives. Critical dimensions, sidewall angles, shapes, stresses, compositions, films, bandgaps, electrical properties, focus / dose, overlay, production process parameters (e.g., resist conditions, partial pressures, temperatures, focus models), and / or any combination thereof can be measured or otherwise determined. Metrology systems can be modeled or designed. Metrology targets can also be modeled, designed, and / or optimized.

[0085] The embodiments disclosed herein can also be used for mask metrology.

[0086] Each of the steps of the method may be performed as described herein. The method may also include any other steps that may be performed by a processor and / or computer subsystem or system described herein. The steps may be performed by one or more computer systems that may be configured according to any of the embodiments described herein. In addition, the above-described method may be implemented by any of the system embodiments described herein.

[0087] Although the present disclosure has been described with respect to one or more particular embodiments, it will be understood that other embodiments of the present disclosure may be made without departing from the scope of the present disclosure. Accordingly, the present disclosure is deemed to be limited only by the appended claims and their reasonable interpretation.

Claims

1. 1. A method comprising: determining a first training sample set from initial metrology measurements, said initial metrology measurements including spectra at two azimuth angles and optionally skew; preparing additional tool-induced shift training samples by adding a negative value reference to the opposite azimuth angle data from the first training sample set; stacking original azimuth data from the first training sample set with opposite azimuth data in a second training sample set; applying at least one machine learning algorithm to the first training sample set and the second training sample set to derive an estimation model; Obtaining an independent overlay with embedded tool-induced shift corrections on an in-line wafer collected for spectra at only one azimuthal angle; using the estimation model to provide an estimate of the independent overlay due to tool-induced shift; A method having the following.

2. The method of claim 1 , wherein the additional tool-induced shift training samples are a reference for an external metrology-based approach.

3. The method of claim 1 , wherein the additional tool-induced shift training samples are self-calibrating overlays for a reference-free approach.

4. 2. The method of claim 1, wherein the machine learning algorithm is one of a linear model, a neural network, and a convolutional network.

5. The method of claim 1 , wherein the machine learning algorithm is a model-assisted approach, a model-less approach, a reference-based approach, or a self-calibrating recipe.

6. 6. The method of claim 5, wherein the tool-induced shift signal is used to calibrate a system parameter.

7. 10. The method of claim 1, wherein the spectra at the two azimuthal angles are collected from a spectroscopic ellipsometer, a soft x-ray reflectometer, a small angle x-ray scatterometer, or an imaging system.

8. 2. The method of claim 1, wherein the initial metrology measurements include tilt in an X direction and a Y direction perpendicular to the X direction.

9. 10. The method of claim 1, wherein using the estimation model includes determining a non-zero overlay and / or edge placement error.

10. 10. The method of claim 1, wherein the machine learning algorithm is further trained using data having critical dimensions, heights, sidewall angles, or film variances.

11. The method of claim 1 , wherein the independent overlay with the tool-induced shift is determined for multiple tools.

12. 2. The method of claim 1, wherein the two azimuth angles are 180 degrees apart.

13. 10. The method of claim 1, wherein at least one of the determining, preparing, stacking, applying, or using steps is performed by at least one processor.

14. A computer readable medium storing a program configured to instruct a processor to perform the method of claim 1.

15. A metrology module comprising the computer readable medium of claim 14.

16. 1. A system comprising: a metrology module including a processor, the metrology module comprising an estimation model configured to provide an estimate of independent overlay due to tool-induced shift on a received wafer, the estimation model using at least one machine learning algorithm, the estimation being based on only one azimuthal angle spectrum, and the estimation model being based on spectra at two azimuthal angles; A system comprising:

17. 17. The system of claim 16, wherein the estimation model is derived by a machine learning algorithm applied to training data calculated based on a first training sample set from initial metrology measurements and additional tool-induced shift training samples, the initial metrology measurements including spectra at the two azimuth angles and optionally skew, and the additional tool-induced shift training samples being at azimuth angle data opposite to the first training sample set having negative values ​​of a reference.

18. The processor: stacking original azimuth data from the first training sample set with opposite azimuth data in a second training sample set; obtaining an independent overlay with embedded tool-induced shift corrections for in-line wafer collected spectra at one azimuthal angle; 20. The system of claim 17, further configured to:

19. 20. The system of claim 17, wherein the machine learning algorithm is further trained using critical dimensions, height, sidewall angle, or film variance.

20. 18. The system of claim 17, wherein the two azimuth angles are 180 degrees apart.

21. 17. The system of claim 16, wherein the machine learning algorithm is one of a linear model, a neural network, and a convolutional network.

22. 17. The system of claim 16, wherein the machine learning algorithm is a model-assisted approach, a model-less approach, a criteria-based approach, or a self-calibrating recipe.

23. 17. The system of claim 16, wherein the metrology module is further configured to determine non-zero overlay and / or edge placement error.

Images