Model related to autoencoder model or like for manufacturing process parameter estimation

By decomposing and optimizing potential elements in latent spatial representations through the modular autoencoder model, the problems of poor interpretability and long training time of existing autoencoder models are solved, and more accurate parameter estimation and model flexibility are achieved.

CN118974740APending Publication Date: 2024-11-15ASML NETHERLANDS BV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202380031637.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2022-03-29
Filing Date
2023-03-02
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

Existing automatic encoder models are difficult to interpret information flow during semiconductor manufacturing, and choosing the right network complexity requires intuition and trial-and-error, resulting in long training time and prone to overfitting.

Method used

Using a modular autoencoder model, by decomposing low-dimensional data in latent spatial representations, training and optimizing the second model component of each latent element, optimizing the output to sort potential elements, selecting the appropriate latent elements to build low-dimensional representations .

Benefits of technology

Improve the interpretability and flexibility of the model, reduce training time, avoid overfitting, and achieve more accurate estimation of manufacturing process parameters.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118974740A_ABST
    Figure CN118974740A_ABST
Patent Text Reader

Abstract

A method for ordering and / or selecting potential elements to model low-dimensional data in a potential spatial representation, the low-dimensional data being a dimensionality-reduced representation of input data determined by a first model component of a model, the method comprising the steps of training said model and selecting one of said potential element selections based on said training, the training includes: reducing a dimension of input data to generate the low-dimensional data in the potential spatial representation; training a second model component of the model for each of the one or more potential element selections; and for each of the potential element selection, optimizing an approximation of the input data output by the second model component to sort at least one of the plurality of potential elements in the potential spatial representation based on a contribution of each potential element to the input data.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to EP application 22164893.4 filed on March 29, 2022, and the entire contents of which are incorporated herein by reference. Technical Field

[0003] The present description relates to methods and systems for estimating manufacturing process parameters using modular autoencoder (or similar) models. Background Art

[0004] A lithographic apparatus is a machine used to apply a desired pattern onto a substrate. A lithographic apparatus can be used to manufacture integrated circuits (ICs). A lithographic apparatus can, for example, project a pattern (also often referred to as a "design layout" or "design") at a patterning device (e.g., a mask) onto a layer of radiation-sensitive material (resist) on a substrate (e.g., a wafer).

[0005] To project a pattern onto a substrate, a lithographic apparatus may use electromagnetic radiation. The wavelength of this radiation determines the minimum size of a feature that can be formed on a substrate. Typical wavelengths currently used are 365 nm (i-line), 248 nm, 193 nm, and 13.5 nm. A lithographic apparatus using extreme ultraviolet (EUV) radiation with a wavelength in the range of 4-20 nm (e.g., 6.7 nm or 13.5 nm) can be used to form smaller features on a substrate than a lithographic apparatus using, for example, radiation with a wavelength of 193 nm.

[0006] Low-k1 lithography can be used to process features with dimensions smaller than the traditional resolution limit of the lithographic equipment. In this process, the resolution formula can be expressed as CD = k1 × λ / NA, where λ is the wavelength of the radiation used, NA is the numerical aperture of the projection optics in the lithographic equipment, CD is the "critical dimension" (usually the minimum feature size printed, but in this case it is half the pitch), and k1 is an empirical resolution factor. In general, the smaller the k1, the more difficult it is to reproduce on the substrate a pattern similar in shape and size to that planned by the circuit designer to achieve specific electrical functions and performance.

[0007] To overcome these difficulties, complex fine-tuning steps can be applied to the lithographic projection equipment and / or the design layout. These include, for example, but not limited to, optimization of NA, customized illumination schemes, use of phase-shifted patterning devices, various optimizations of the design layout, such as optical proximity correction (OPC, sometimes also called "optical and process correction") in the design layout, or other methods generally defined as "resolution enhancement technology" (RET). Alternatively, a tight control loop for controlling the stability of the lithographic equipment can be used to improve pattern reproduction at low k1.

[0008] The autoencoder may be configured for measurement and / or other solutions, for parameter inference and / or other purposes. This deep learning model architecture is general and scalable to arbitrary size and complexity. The autoencoder is configured to compress a high-dimensional signal (e.g., pupil images in a semiconductor manufacturing process) into an efficient low-dimensional representation of the same signal. Next, or in parallel (parameter inference may be part of the autoencoder training), parameter inference (i.e., regression) from the low-dimensional representation to a set of known labels is performed. By first compressing the signal, the inference problem is greatly simplified compared to performing regression directly on the high-dimensional signal.

[0009] However, understanding the information flow inside a typical autoencoder is often difficult. One can infer the information at the input, compressed low-dimensional representation level, and output. One cannot easily interpret the information between these points.

[0010] Compared to traditional monolithic autoencoder models, the present modular autoencoder model is less rigid. The present modular autoencoder model has more trainable and / or adjustable components. The modularity of the present model makes it easier to interpret, define, and extend. The complexity of the present model is easy to adjust and is high enough to model the process that generates the data provided to the model, but low enough to avoid modeling noise or other unwanted features (e.g., the present model is configured to avoid overfitting the provided data). Because the process that generates the data (or at least some aspects of the process) is usually unknown, choosing the appropriate network complexity usually requires some intuition and trial and error. For this reason, it is very desirable to provide a model architecture that is modular, easy to understand, and easy to increase and decrease in complexity.

[0011] Note that the term "autoencoder" in relation to this modular autoencoder model generally refers to one or more autoencoders and / or other autoencoders configured for partially supervised learning using a latent space for parameter estimation. This can also include a single autoencoder, for example, trained using semi-supervised learning. Summary of the invention

[0012] The present invention discloses a method for sorting and / or selecting potential elements for modeling low-dimensional data within a latent space representation, wherein the low-dimensional data is a dimensionality-reduced representation of input data determined by a first model component of a model; the method comprises the following steps: obtaining the input data; training the model, wherein the training comprises reducing the dimension of the input data to generate the low-dimensional data in the latent space representation, wherein the latent space representation comprises a plurality of latent elements; training a second model component of the model for each of one or more latent element selections, each of the latent element selections comprising one or more of the plurality of latent elements; and optimizing an approximation of the input data output by the second model component for each of the latent element selections, thereby sorting at least one of the plurality of latent elements in the latent space representation based on the contribution of each latent element to the input data; and selecting one of the latent element selections based on the training.

[0013] In other aspects of the present invention, a computer program is provided, comprising program instructions that can execute the method of the first or second aspect when run on a suitable device, and a lithographic device configured to provide a product structure to a substrate in a lithographic process, the lithographic device comprising a processor and the above-mentioned computer program. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The accompanying drawings are incorporated into and constitute a part of this specification to illustrate one or more embodiments and together with the specification explain these embodiments. Embodiments of the present invention will now be described by way of example only with reference to the accompanying drawings, in which corresponding reference characters indicate corresponding parts, and in which:

[0015] Figure 1 Shown is a schematic overview of a lithographic apparatus according to an embodiment.

[0016] Figure 2 A schematic diagram of a lithography cell according to one embodiment is shown.

[0017] Figure 3 A schematic diagram of overall lithography is shown according to one embodiment, representing the cooperation between three technologies to optimize semiconductor manufacturing.

[0018] Figure 4 An example metrology device, such as a scatterometer, is shown according to an embodiment.

[0019] Figure 5 A first encoder-decoder architecture that may be adapted according to one embodiment is shown.

[0020] Figure 6 A second encoder-decoder architecture is shown that may be adapted according to one embodiment.

[0021] Figure 7 An embodiment of a modular autoencoder model that is adaptable according to one embodiment is shown.

[0022] Figure 8 A schematic diagram of an interconnected structure for generating labels to impose priors on a modular autoencoder model is shown according to one embodiment.

[0023] Fig. 9 An encoder-decoder architecture according to one embodiment is shown.

[0024] Fig.10 is a block diagram of an example computer system according to one embodiment. DETAILED DESCRIPTION

[0025] As described above, the autoencoder can be configured for measurement and / or other solutions for parameter inference and / or other purposes. This deep learning model architecture is general and can be extended to any size and complexity. The autoencoder is configured to compress a high-dimensional signal (such as a pupil image in a semiconductor measurement platform) into an effective low-dimensional representation of the same signal. Next, parameter inference (i.e., regression) from the low-dimensional representation to a set of known labels is performed. By first compressing the signal, the inference problem is significantly simplified compared to performing regression directly on the high-dimensional signal.

[0026] However, understanding the information flow inside a typical autoencoder is often difficult. One can see information at the input, at the compressed low-dimensional representation level, and at the output. One cannot easily interpret the information between these points.

[0027] Data-driven inference methods have also been proposed for semiconductor metrology operations and used for parameter estimation tasks. They rely on a large collection of measurements and models that map the measured features to parameters of interest and obtain labels for these parameters through carefully designed targets on the wafer or third-party measurements. Current methods are able to measure a large number of channels (multiple wavelengths, observations under multiple wafer rotations, four light polarization schemes, etc.). However, due to practical time constraints, the number of channels needs to be limited to a subset of the channels available for production measurements. In order to select the best channel, a brute force approach is usually used to test all possible channel combinations. This is time consuming, resulting in long measurement and / or process configuration scheme creation times. In addition, brute force methods are prone to overfitting, with each channel introducing different biases and / or other disadvantages.

[0028] In semiconductor manufacturing, optical metrology can be used to directly measure key stacking parameters on product (e.g., patterned wafer) structures. Machine learning methods are often applied on optical scatterometry data acquired using a metrology platform. These machine learning methods are conceptually equivalent to supervised learning methods, i.e., learning from labeled datasets. The success of such methods depends heavily on the quality of the labels. Typically, labeled datasets are created by measuring and labeling known targets in the wafer. One of the main challenges of using targets in this way is that they only provide very accurate relative labels. This means that within a cluster of targets, there is some unknown cluster deviation, based on which accurate labels are known. Determining this unknown cluster deviation, and thus obtaining absolute labels, is very important for the accuracy of target-based configuration solutions. The process of estimating cluster deviations is often called offset inference.

[0029] The modular autoencoder model provides greater flexibility than a traditional monolithic autoencoder model. The modular autoencoder model has more trainable and / or tunable components. The modularity of the model makes it easier to interpret, define, and extend. The complexity of the model is high enough to model the process that generates the data provided to the model, but low enough to avoid modeling noise or other undesirable characteristics (e.g., the model is configured to avoid overfitting the provided data). Since the process that generates the data (or at least some aspects of the process) is usually unknown, choosing the appropriate network complexity often requires some intuition and trial and error. Therefore, it is highly desirable to provide a model architecture that is modular, easy to understand, and easy to increase and decrease in complexity.

[0030] In addition, the modular autoencoder model is configured to estimate the amount of information content that can be retrieved based on the available channel combinations of the measurement data from the optical metrology platform by using a subset of the multiple input models based on the available channels. The model is configured to be trained by randomly or otherwise iteratively changing (e.g., sub-selecting) the multiple channels used to approximate the input during an iterative training step. This iterative change / sub-selection ensures that the model remains predictive / consistent for any combination of input channels. In addition, because the information content present in the input is representative of all channels (e.g., because each channel is part of the selected channel subset for at least one training iteration), the resulting model will not include biases that are specific to a particular channel.

[0031] The modular autoencoder model is also configured to embed known attributes of the input (e.g., domain knowledge) into the model during the training phase, thereby reducing or eliminating bias (e.g., clustering) in subsequent inferences of the model. In other words, the modular autoencoder is configured to embed known (e.g., symmetry) attributes of the input into the decoding portion of the model, and these embedded known attributes allow the model to perform unbiased inference.

[0032] Note that the term "autoencoder" used in conjunction with the present modular autoencoder model may generally refer to one or more autoencoders, or one or more portions of an autoencoder, configured for partially supervised learning using a latent space for parameter estimation and / or other operations. Furthermore, the various disadvantages (e.g., disadvantages of previous systems) and advantages (disadvantages of the present modular autoencoder model) described above are examples of many other possible disadvantages and advantages and should not be considered limiting.

[0033] Finally, although the fabrication of integrated circuits may be specifically mentioned herein, the description herein has many other possible applications. For example, it may be used to fabricate integrated optical systems, guidance and detection patterns for magnetic domain memories, liquid crystal-display panels, thin film magnetic heads, and the like. In these alternative applications, those skilled in the art will understand that any use of the terms "reticle," "wafer," or "die" herein in the context of these alternative applications should be considered interchangeable with the more general terms "mask," "substrate," and "target portion," respectively. Furthermore, it should be noted that the methods described herein may have many other possible applications in different fields such as language processing systems, self-driving cars, medical imaging and diagnosis, semantic segmentation, denoising, chip design, electronic design automation, and the like. The present method may be applied to any field where it is advantageous to quantify uncertainty in the predictions of machine learning models.

[0034] In this document, the terms "radiation" and "beam" are used to cover all types of electromagnetic radiation, including UV radiation (e.g., having a wavelength of 365nm, 248nm, 193nm, 157nm or 126nm) and EUV (extreme ultraviolet radiation, e.g., having a wavelength range of about 5-100nm).

[0035] The pattern forming device may include or may form one or more design layouts. The design layout may be generated using a CAD (computer-aided design) program. This process is generally referred to as EDA (electronic design automation). Most CAD programs follow a set of predetermined design rules to create a functional design layout / pattern forming device. These rules are set based on process and design constraints. For example, the design rules define the spatial tolerances between devices (such as gates, capacitors, etc.) or interconnects to ensure that the devices or lines do not interact in an undesirable manner. One or more design rule restrictions may be referred to as "critical dimensions" (CD). The critical dimension of a device may be defined as the minimum width of a line or hole, or the minimum space between two lines or two holes. Therefore, CD regulates the overall size and density of the design device. One of the goals of device manufacturing is to faithfully reproduce the original design intent on the substrate (via the pattern forming device).

[0036] The terms "reticle", "mask" or "patterning device" as used herein may be broadly interpreted as referring to a generic patterning device that can be used to impart a patterned cross-section to an incident radiation beam that corresponds to the pattern to be created in a target portion of the substrate. The term "light valve" may also be used in this context. In addition to classical masks (transmissive or reflective; binary, phase-shifting, hybrid, etc.), other examples of such patterning devices include programmable mirror arrays.

[0037] Let me give you a brief introduction. Figure 1 A lithographic apparatus LA is schematically shown. The lithographic apparatus LA comprises an illumination system (also referred to as an illuminator) IL configured to condition a radiation beam B (e.g., UV radiation, DUV radiation, or EUV radiation), a mask support (e.g., a mask stage) T constructed to support a patterning device (e.g., a mask) MA and connected to a first positioner PM configured to precisely position the patterning device MA according to certain parameters, a substrate support (e.g., a wafer stage) WT configured to hold a substrate (e.g., a resist-coated wafer) W and connected to a second positioner PW configured to precisely position the substrate support according to certain parameters, and a projection system (e.g., a refractive projection lens system) PS configured to project a pattern imparted by the patterning device MA to the radiation beam B onto a target portion C (e.g., comprising one or more chips) of the substrate W.

[0038] In operation, the illumination system IL receives a radiation beam from a radiation source SO, for example, via a beam delivery system BD. The illumination system IL may include various types of optical components, such as refractive, reflective, magnetic, electromagnetic, electrostatic and / or other types of optical components, or any combination thereof, for directing, shaping and / or controlling the radiation. The illuminator IL may be used to condition the radiation beam B so that it has a desired spatial and angular intensity distribution in its cross-section in the plane of the patterning device MA.

[0039] The term "projection system" PS as used herein should be interpreted broadly to cover various types of projection systems, including refractive, reflective, catadioptric, anamorphic, magnetic, electromagnetic and / or electrostatic optical systems, or any combination thereof, as appropriate for the exposure radiation used and / or other factors, such as the use of immersion liquid or the use of a vacuum. Any use of the term "projection lens" as used herein may be considered synonymous with the more general term "projection system" PS.

[0040] The lithographic apparatus LA may be of a type in which at least a portion of the substrate may be covered by a liquid having a relatively high refractive index, such as water, so as to fill the space between the projection system PS and the substrate W - this is also known as immersion lithography. For more information on immersion technology, see US6952253, which is incorporated herein by reference.

[0041] The lithographic apparatus LA may also be of a type having two or more substrate supports WT (also referred to as a "dual stage type"). In such a "multi-stage type" machine, the substrate supports WT may be used in parallel, and / or a subsequent exposure preparation step of a substrate W may be performed on a substrate W located on one substrate support WT while another substrate W on another substrate support WT is used to expose a pattern on another substrate W.

[0042] In addition to the substrate support WT, the lithographic apparatus LA may comprise a measuring table. The measuring table is arranged to accommodate sensors and / or cleaning means. The sensors may be arranged to measure properties of the projection system PS or properties of the radiation beam B. The measuring table may hold a plurality of sensors. The cleaning means may be arranged to clean a part of the lithographic apparatus, e.g. a part of the projection system PS or a part of the system for providing immersion liquid. The measuring table may be moved under the projection system PS when the substrate support WT is away from the projection system PS.

[0043] In operation, a radiation beam B is incident on a patterning device (e.g. a mask MA) which is held on a mask support MT and is patterned by a pattern (design layout) on the patterning device MA. After passing through the mask MA, the radiation beam B passes through a projection system PS which focuses the beam onto a target portion C of the substrate W. With the aid of a second positioner PW and a position measurement system IF, the substrate support WT can be precisely moved, for example, in order to position different target portions C in the path of the radiation beam B in a focused and aligned position. Similarly, a first positioner PM and possibly another position sensor (not shown) are provided. Figure 1The patterning device MA (shown explicitly in FIG. 1 ) can be used to accurately position the patterning device MA relative to the path of the radiation beam B. The patterning device MA and the substrate W can be aligned using the mask alignment marks M1, M2 and the substrate alignment marks P1, P2. Although the substrate alignment marks P1, P2 are shown occupying dedicated target portions, they can be located in the space between target portions. When the substrate alignment marks P1, P2 are located between target portions C, they are referred to as scribe line alignment marks.

[0044] Figure 2 Schematic diagram of the lithography unit LC is shown. Figure 2 As shown, the lithography apparatus LA may form part of a lithography cell LC, sometimes also referred to as a lithography cell or (lithography) cluster, which typically also includes equipment for performing pre-exposure and post-exposure processes on a substrate W. Conventionally, these include a spin coater SC configured to deposit a resist layer, a developer DE for developing the exposed resist, a cooling plate CH and a baking plate BK, for example for regulating the temperature of the substrate W, for example for regulating a solvent in the resist layer. A substrate handler or robot RO picks up substrates W from input / output ports I / O1, I / O2, moves them between different process equipment, and transfers the substrates W to a loading station LB of the lithography apparatus LA. The equipment in the lithography cell (also often collectively referred to as a track) is typically controlled by a track control unit TCU, which itself may be controlled by a monitoring system SCS, which may also control the lithography apparatus LA (e.g., via a lithography control unit LACU).

[0045] In order to expose the substrate W ( Figure 1 ) can be correctly and consistently exposed, the substrate needs to be inspected to measure properties of the patterned structure, such as overlay errors between subsequent layers, line thickness, critical dimensions (CD), etc. For this purpose, an inspection tool (not shown) may be included in the lithography cell LC. If errors are detected, adjustments may be made to the exposure of subsequent substrates or to other processing steps to be performed on the substrate W, particularly if the inspection is performed before other substrates W of the same batch or batch are still to be exposed or processed.

[0046] The inspection device (also called a metrology device) is used to determine the substrate W ( Figure 1), in particular to determine how properties of different substrates W vary, or how related properties of different layers of the same substrate W vary from layer to layer. The inspection device can also be configured to identify defects on the substrate W, for example, it can be part of the lithography unit LC, it can also be integrated into the lithography equipment LA, or it can even be a stand-alone device. The inspection device can measure the properties of the latent image (the image in the resist layer after exposure), it can also measure the properties of the semi-latent image (the image in the resist layer after the post-exposure bake step PEB), it can also measure the properties of the developed resist image (where the exposed or unexposed parts of the resist have been removed), and it can even measure the properties of the etched image (after a pattern transfer step such as etching).

[0047] Figure 3 FIG. 1 shows a schematic diagram of overall lithography, representing the collaboration between the three technologies to optimize semiconductor manufacturing. Generally, the patterning process in the lithography apparatus LA is one of the most critical steps in the process, which requires the substrate W( Figure 1 ) to dimension and place structures with high accuracy. To ensure this high accuracy, the three systems (in this example) can be combined in a so-called “holistic” control environment, such as Figure 3 As shown. One of these systems is a lithography apparatus LA, which is (virtually) connected to a metrology device (e.g., a metrology tool) MT (a second system) and a computer system CL (a third system). The "holistic" environment can be configured to optimize the cooperation between these three systems to enhance the overall process window and provide a tight control loop to ensure that the patterning performed by the lithography apparatus LA remains within the process window. The process window defines the range of process parameters (e.g., dose, focus, overlay) within which a specific manufacturing process can produce a defined result (e.g., a functional semiconductor device), and generally within which process parameter variations in the lithography process or patterning process are allowed.

[0048] The computer system CL may use the (portion of) the design layout to be patterned to predict which resolution enhancement technique should be used and perform computational lithography simulations and calculations to determine which mask layout and lithography equipment settings will achieve the maximum overall process window ( Figure 3 Typically, the resolution enhancement techniques are arranged to match the patterning possibilities of the lithographic apparatus LA. The computer system CL may also be used to detect where within the process window the lithographic apparatus LA is currently operating (e.g. using input from a metrology tool MT) to predict whether defects may exist due to, for example, suboptimal processing ( Figure 3 (as indicated by the arrow pointing to “0” in the second scale SC2).

[0049] The metrology device (tool) MT may provide input to the computer system CL to enable accurate simulations and predictions, and may provide feedback to the lithographic apparatus LA to identify possible drift, for example, in the calibration state of the lithographic apparatus LA (in Figure 3 (indicated by multiple arrows in the third scale SC3).

[0050] In the lithography process, it is necessary to frequently measure the created structures, for example for process control and verification. The tool for performing such measurements comprises a measurement tool (device) MT. Different types of measurement tools MT are known for performing such measurements, including scanning electron microscopes or various forms of scatterometer measurement tools MT. A scatterometer is a versatile device that allows the measurement of parameters of the lithography process by mounting a sensor in the pupil of the scatterometer objective or in a plane conjugated to the pupil, such measurements are generally referred to as pupil-based measurements, or by mounting a sensor in an image plane or in a plane conjugated to the image plane, in which case the measurements are generally referred to as image-based or field-based measurements. Such scatterometers and related measurement techniques are further described in patent applications US20100328655, US2011102753A1, US20120044470A, US20110249244, US20110026032 or EP1,628,164A, which are incorporated herein by reference in their entirety. For example, the scatterometer described above can use light from soft x-rays and visible to near infrared wavelengths to measure features of a substrate such as a grating.

[0051] In some embodiments, the scatterometer MT is an angle-resolved scatterometer. In these embodiments, scatterometer reconstruction methods may be applied to the measured signals to reconstruct or calculate properties of the grating and / or other features in the substrate. Such a reconstruction may be obtained, for example, by simulating the interaction of the scattered radiation with a mathematical model of the target structure and comparing the simulation results with the measurement results. The parameters of the mathematical model are adjusted until the simulated interaction produces a diffraction pattern similar to the diffraction pattern observed from a real target.

[0052] In some embodiments, the scatterometer MT is a spectroscopic scatterometer MT. In these embodiments, the spectroscopic scatterometer MT can be configured so that radiation emitted by the radiation source is directed onto a target feature of the substrate, while reflected or scattered radiation from the target is directed to a spectrometer detector, which measures the spectrum of the specularly reflected radiation (i.e., a measure of intensity as a function of wavelength). From this data, the structure or profile of the target that caused the detected spectrum can be reconstructed, for example by rigorous coupled wave analysis and nonlinear regression or by comparison with a library of simulated spectra.

[0053] In some embodiments, the scatterometer MT is an ellipsometric scatterometer. An ellipsometric scatterometer allows the parameters of a lithographic process to be determined by measuring the scattered radiation of each polarization state. Such a measurement device (MT) emits polarized light (e.g., linear, circular, or elliptical) by using, for example, an appropriate polarization filter in the illumination portion of the measurement device. A source suitable for the measurement device can also provide polarized radiation. Various embodiments of existing ellipsometric scatterometers are described in U.S. patent applications 11 / 451,599, 11 / 708,678, 12 / 256,780, 12 / 486,449, 12 / 920,968, 12 / 922,587, 13 / 000,229, 13 / 033,135, 13 / 533,110, and 13 / 891,410, which are incorporated herein by reference in their entirety.

[0054] In some embodiments, the scatterometer MT is suitable for measuring the overlay of two misaligned gratings or periodic structures (and / or other target features of the substrate) by measuring the reflection spectrum and / or an asymmetry in the detection configuration, which asymmetry is related to the degree of overlay. The two (usually overlapping) grating structures can be applied to two different layers (not necessarily continuous layers) and can be formed at substantially the same position on the wafer. The scatterometer can have a symmetrical detection configuration, such as described in patent application EP1,628,164A, so that any asymmetry can be clearly distinguished. This provides a method of measuring grating misalignment. Further examples of measuring overlay can be found in PCT patent application publication number WO 2011 / 012624 or in US patent application US20160161863, the entire contents of which are incorporated herein by reference.

[0055] Other parameters of interest may be focus and dose. Focus and dose may be determined simultaneously by scatterometry (or by scanning electron microscopy), as described in U.S. Patent Application US2011-0249244, which is incorporated herein by reference in its entirety. A single structure (e.g., a feature in a substrate) may be used that has a unique combination of critical dimension and sidewall angle measurements for each point in a focus energy matrix (FEM—also called a focus exposure matrix). If these unique combinations of critical dimensions and sidewall angles are available, focus and dose values ​​may be uniquely determined from these measurements.

[0056] The metrology target can be a collection of composite gratings and / or other features in a substrate, formed by a lithography process, typically in a resist, but can also be formed after an etching process. In some embodiments, one or more target groups can be gathered at different locations around the wafer. Typically, the pitch and line width of the structures in the grating are determined by the measurement optics (particularly the NA of the optics) so that the diffraction orders from the metrology target can be captured. The diffraction signal can be used to determine the offset between two layers (also called "overlay"), or can be used to reconstruct at least a portion of the original grating produced by the lithography process. This reconstruction can be used to provide guidance on the quality of the lithography process and can be used to control at least a portion of the lithography process. The target can have a smaller sub-division that is configured to simulate the size of a functional portion of the design layout in the target. Due to this sub-division, the behavior of the target will be more similar to the functional portion of the design layout, so that the overall process parameter measurement is similar to the functional portion of the design layout. The target can be measured in an underfill mode or an overfill mode. In the underfill mode, the spot produced by the measurement beam is smaller than the overall target. In the overfill mode, the spot produced by the measurement beam is larger than the overall target. In this overfill mode, different targets can also be measured simultaneously, thereby determining different process parameters simultaneously.

[0057] The overall measurement quality of the lithography parameters using a specific target is at least partially determined by the measurement scheme used to measure the lithography parameters. The term "substrate measurement scheme" can include one or more parameters of the measurement itself, one or more parameters of the one or more patterns measured, or both. For example, if the measurement used in the substrate measurement scheme is an optical measurement based on diffraction, the one or more parameters measured can include the wavelength of the radiation, the polarization of the radiation, the angle of incidence of the radiation relative to the substrate, the orientation of the radiation relative to the pattern on the substrate, etc. One of the criteria for selecting a measurement scheme can be the sensitivity of one of the measurement parameters to process variations. More examples are described in U.S. patent application US2016-0161863 and published U.S. patent application US2016 / 0370717A1, the entire contents of which are incorporated herein by reference.

[0058] Figure 4 An exemplary metrology device (tool or platform) MT is shown, such as a scatterometer. MT comprises a broadband (white light) radiation projector 40, which projects radiation onto a substrate 42. The reflected or scattered radiation is passed to a spectrometer detector 44, which measures a spectrum 46 (i.e. a measure of the intensity as a function of wavelength) of the specularly reflected radiation. From this data, a processing unit PU can reconstruct 48 the structure or profile that led to the detected spectrum, for example by rigorous coupled wave analysis and nonlinear regression or by comparison with a library of simulated spectra, such as Figure 3Typically, for the reconstruction, the general form of the structure is known and some parameters are assumed from knowledge of the structure fabrication process, leaving only some structural parameters to be determined from the scatterometry data. For example, such a scatterometer can be configured as a normal-incidence scatterometer or an oblique-incidence scatterometer.

[0059] Generally, it is desirable to be able to determine computationally how a patterning process produces a desired pattern on a substrate. For example, computational determination may include simulation and / or modeling. Models and / or simulations may be provided for one or more portions of a manufacturing process. For example, it is desirable to be able to simulate a lithography process that transfers a pattern forming device pattern to a resist layer of a substrate and a pattern produced in the resist layer after the resist is developed, simulate metrology operations (e.g., determine overlay), and / or perform other simulations. The purpose of the simulation may be to accurately predict metrology metrics (e.g., overlay, critical dimensions, reconstruction of a three-dimensional profile of a substrate feature, dose or focus of a lithography device when a substrate feature is printed using a lithography device, etc.), manufacturing process parameters (e.g., edge position, spatial image intensity slope, sub-resolution assist features (SRAF), etc.), and / or other information, which may then be used to determine whether an expected or target design has been achieved. An expected design is typically defined as a pre-optical proximity correction design layout, which may be provided in a standardized digital file format (such as GDSII, OASIS, or other file formats).

[0060] Simulation and / or modeling may be used to determine one or more metrology metrics (e.g., perform overlay and / or other metrology measurements), configure one or more features of a pattern forming device pattern (e.g., perform optical proximity correction), configure one or more features of the illumination (e.g., change one or more characteristics of the spatial / angular intensity distribution of the illumination, such as changing the shape), configure one or more features of the projection optics (e.g., numerical aperture, etc.), and / or for other purposes. Such determination and / or configuration may generally be referred to as mask optimization, source optimization, and / or projection optimization. Such optimizations may be performed individually or combined in different combinations. One such example is source mask optimization (SMO), which involves configuration of one or more features of the pattern forming device pattern and one or more features of the illumination. Optimization may use the parameterized models described herein to predict values ​​of various parameters (including images, etc.).

[0061] In some embodiments, the optimization process of the system can be expressed as a cost function. The optimization process can include finding a set of system parameters (design variables, process variables, inspection operation variables, etc.) that minimize the cost function. The cost function can have any suitable form, depending on the goal of the optimization. For example, the cost function can be the weighted root mean square (RMS) of the deviation of certain characteristics (evaluation points) of the system relative to the expected values ​​(e.g., ideal values) of these characteristics. The cost function can also be the maximum value (i.e., the worst deviation) of these deviations. The term "evaluation point" should be widely interpreted as including any characteristic of the system or manufacturing method. Due to the practicality of the implementation of the system and / or method, the design and / or process variables of the system can be limited to a limited range and / or interdependent. For lithographic projection and / or detection equipment, constraints are generally related to the physical properties and characteristics of the hardware (e.g., adjustable range) and / or patterning equipment manufacturability design rules. The evaluation point can include a physical point on the resist image on the substrate, as well as non-physical characteristics such as dose and focus.

[0062] In some embodiments, the present systems and methods may include an empirical model that performs one or more of the operations described herein. The empirical model may predict outputs based on correlations between various inputs (e.g., one or more characteristics of a pupil image, one or more characteristics of a complex electric field image, one or more characteristics of a design layout, one or more characteristics of a patterning device, one or more characteristics of an illumination used in a lithography process (e.g., wavelength, etc.).

[0063] For example, the empirical model can be a parameterized model and / or other model. The parameterized model can be a machine learning model and / or any other parameterized model. In some embodiments, the machine learning model (for example) can be and / or include mathematical equations, algorithms, graphs, charts, networks (e.g., neural networks), and / or other tools and machine learning model components. For example, the machine learning model can be and / or include one or more neural networks (e.g., neural network blocks) having an input layer, an output layer, and one or more intermediate layers or hidden layers. In some embodiments, one or more neural networks can be and / or include a deep neural network (e.g., a neural network having one or more intermediate layers or hidden layers between the input layer and the output layer).

[0064] For example, one or more neural networks can be based on a large number of neural units (or artificial neurons). One or more neural networks can loosely imitate the way biological brains work (for example, through a large number of biological neuron clusters connected by axons). Each neural unit of a neural network can be connected to many other neural units of the neural network. This connection can strengthen or inhibit the activation state of the connected neural units. In some embodiments, each individual neural unit can have a summation function that combines the values ​​of all its inputs. In some embodiments, each connection (or neural unit itself) can have a threshold function so that the signal must exceed the threshold to propagate to other neural units. These neural network systems can be self-learning and training, rather than explicitly programmed, and can perform better in certain problem-solving areas compared to traditional computer programs. In some embodiments, one or more neural networks can include multiple layers (for example, the signal path traverses from the front layer to the back layer). In some embodiments, the neural network can use backpropagation technology, in which forward stimulation is used to reset the weights of the "front end" neural unit. In some embodiments, the stimulation and inhibition of one or more neural networks can flow more freely, and the connections interact in a more chaotic and complex way. In some embodiments, the intermediate layer of one or more neural networks includes one or more convolutional layers, one or more recurrent layers, and / or other layers.

[0065] A set of training data (e.g., ground truth) can be used to train one or more neural networks (i.e., determine their parameters). The training data can include a set of training samples. Each sample can be a pair, including an input object (usually an image, measurement value, tensor, or vector, which can be called a feature tensor or vector) and a desired output value (also called a supervisory signal). The training algorithm analyzes the training data and adjusts the behavior of the neural network by adjusting the parameters of the neural network (e.g., the weights of one or more layers) based on the training data. For example, given a set of N training samples of the form {(x1, y1), (x2, y2), ..., (x N ,y N )},x i is the feature tensor / vector of the ith example, y i is its supervisory signal, and the training algorithm looks for a neural network g:X→Y, where X is the input space and Y is the output space. A feature tensor / vector is an n-dimensional tensor / vector that represents the numerical features of an object (e.g., a complex image). The tensor / vector space associated with these vectors is often called a feature or latent space. After training, the neural network can be used to make predictions using new samples.

[0066] As described herein, the modular autoencoder model includes one or more parameterized models (e.g., machine learning models, such as neural networks) and / or other models using an encoder-decoder architecture. In the middle (e.g., an intermediate layer) of the model (e.g., a neural network), the model formulates a low-dimensional encoding (e.g., in a latent space) that encapsulates information in the input (e.g., pupil images and / or other inputs related to patterns or other features of semiconductor manufacturing and / or measurement (and / or other sensing) processes) into the model. The modular autoencoder model exploits the low dimensionality and compactness of the latent space for parameter estimation and / or prediction.

[0067] Figure 5 A general encoder-decoder architecture 50 is shown by way of non-limiting example. The encoder-decoder architecture 50 has an encoding portion 52 (encoder) and a decoding portion 54 (decoder). Figure 5 In the example shown, the encoder-decoder architecture 50 may output a predicted pupil image 56 and / or other outputs.

[0068] In some embodiments, the low-dimensional encoding z represents one or more features of the input (e.g., a pupil image). One or more features of the input may be considered key or key features of the input. For example, features may be considered key or key features of the input because they are relatively more predictive and / or have other characteristics than other features of the desired output. The one or more features (dimensions) represented in the low-dimensional encoding may be predetermined (e.g., by a programmer when creating the present modular autoencoder model), determined by a previous layer of the neural network, adjusted by a user through a user interface associated with the system described herein, and / or may be determined by other methods. In some embodiments, the number of features (dimensions) represented by the low-dimensional encoding may be predetermined (e.g., by a programmer when creating the present modular autoencoder model), determined based on the output of a previous layer of the neural network, adjusted by a user through a user interface associated with the system described herein, and / or determined by other methods.

[0069] It should be noted that although machine learning models, neural networks and / or encoder-decoder architectures are mentioned in this specification, the machine learning models, neural networks and encoder-decoder architectures are merely examples, and the operations described herein can be applied to different parameterized models.

[0070] As described above, process information (e.g., images, measurements, process parameters, metrology indicators, etc.) can be used to guide various manufacturing operations. Exploiting the relatively low dimensionality of the latent space to predict and / or otherwise determine process information may be faster, more efficient, require less computing resources, and / or have other advantages over previous methods of determining process information.

[0071] Figure 6 An embodiment of an autoencoder model arrangement 600 to which the concepts disclosed herein can be applied is shown. In general, the autoencoder model can be applied to metrology and / or other solutions for parameter inference and / or other purposes. Inference can include estimating the parameters of interest from data and / or other operations. For example, this can include finding the potential representation in a forward manner by evaluating the encoder, or finding the potential representation in a reverse manner by solving the inverse problem using the decoder (as described herein). After finding the potential representation, the parameters of interest can be found by evaluating the prediction / estimation model (also as described herein). In addition, the potential representation provides a set of outputs (because the decoder can be evaluated given the potential representation), for example, which can be compared with the data. Essentially, in the current context, inference and estimation (parameters of interest) can be used interchangeably. The autoencoder model architecture is general and can be extended to any size and complexity. The autoencoder model is configured to compress a high-dimensional signal (input) into an effective low-dimensional representation of the same signal. From the low-dimensional representation, one or more outputs and / or other information, parameter inference is performed to a set of known labels (for example, which can include regression and / or other operations). The label may be a "reference" used in supervised learning. In this case, this may mean an external reference that one wants to replicate, or the design of a carefully crafted metrology target. Measuring a carefully crafted metrology target may include measuring a known target with known (absolute / relative) properties (e.g., overlay and / or other properties). By first compressing the (input) signal, the inference problem is significantly simplified compared to performing regression and / or other operations directly on the high-dimensional signal.

[0072] Figure 6 The autoencoder model 604 in FIG. 600 can be used to compress input information 611 (e.g., 611a, 611b, ..., 611n) to a bottleneck (e.g., a compression point or latent space 607 in model 600) and expand the information again to a level suitable for segmentation into multiple outputs. A prediction model 608 can be used to estimate parameters of interest from the information through the latent space.

[0073] In some embodiments, the autoencoder model 604 is configured to determine a latent space representation of a given input 611 in a latent space 607 (where the number of degrees of freedom to be analyzed is less than the number of degrees of freedom of the original input data from different sensors and / or tools). Process information can be estimated and / or predicted based on the latent space representation of the given input 611, and / or other operations can be performed.

[0074] In some embodiments, the autoencoder model 604 includes an encoder portion 605, a latent space 607, a decoder portion 609, and / or other components. In some embodiments, the autoencoder model includes a feedforward and / or residual layer and / or other components, but these example structures should not be considered limiting. The encoder portion 605 of the autoencoder model 604 can be configured to combine inputs 611 and reduce the dimensionality of the combined processed inputs to generate low-dimensional data in the latent space 607.

[0075] The autoencoder model 604 can be configured to expand the low-dimensional data in the potential space into one or more extended versions of one or more inputs 611. The low-dimensional data in the potential space 607 can be expanded into one or more extended versions of one or more inputs 611, for example, decoding, generating decoder signals and / or other operations. In general, one or more extended versions of one or more inputs can include outputs from the autoencoder model 604 (e.g., the last layer). However, one or more extended versions of one or more inputs 611 can include any extended version of any layer from the decoder part 609. One or more extended versions of one or more inputs 611 may have increased dimensions compared to the low-dimensional data in the potential space 607. One or more extended versions of one or more inputs 611 can be configured to be suitable for generating one or more different outputs 613 (e.g., 613a, 613b, ... 613n). Note that the input of the autoencoder model 604 is not necessarily restored to its output. This is intended to simply describe the interface. However, the recovery can be kept globally, input 611 to output 613.

[0076] The prediction model 608 may be configured to estimate one or more parameters (parameters of interest) 615 based on the low-dimensional data in the latent space 607, one or more different outputs 613, and / or other information. In some embodiments, for example, the one or more parameters may be semiconductor manufacturing process parameters (as described herein). In some embodiments, the prediction model 608 includes a feedforward, residual layer, and / or other components, but these example structures are not intended to be limiting. As a non-limiting example, the input 611 sensor signal may include a pupil image, and an encoded representation of the pupil image may be configured to be used by the prediction model 608 to estimate overlay and / or other parameters.

[0077] By way of non-limiting example, the input 611 may include a sensor signal associated with a sensing or measurement operation in a semiconductor manufacturing process (e.g., a measurement pupil or a processed representation thereof), the low-dimensional representation of the input 611 may be a compressed representation of the sensor signal, and the corresponding output 613 may be an approximation of the input sensor signal.

[0078] In some embodiments, input 611 is associated with a pupil, a target, and / or other components of a semiconductor manufacturing process and is received from one or more of a plurality of characterization devices configured to generate input 611. The characterization device may include various sensors and / or tools configured to generate data about the target. In some embodiments, the characterization device may include an optical metrology platform, such as Figure 4 The platform shown in . The data may include images, values ​​of various metrics, and / or other information. In some embodiments, input 611 includes one or more of an input image, an input process measurement, and / or a series of process measurements and / or other information. In some embodiments, input 611 can be a signal associated with a measurement data channel from one or more sensing (e.g., optical metrology and / or other sensing) platforms. A channel can be a mode of observing a stack, for example, a machine / physical configuration used when making measurements. As a non-limiting example, input 611 can include an image (e.g., any image related to or generated during semiconductor manufacturing).

[0079] In some embodiments, the autoencoder model 600 can be trained by comparing one or more different outputs 613 with corresponding inputs 611, and adjusting the parameterization of the autoencoder model 604 and / or the prediction model 608 to reduce or minimize the difference between the output 613 and the corresponding input 611. In some embodiments, the training can include applying changes to the low-dimensional data in the latent space 607 so that the public model 604 decodes a relatively more continuous latent space to generate a decoder signal (e.g., an output from the autoencoder model 604); recursively providing the decoder signal to the encoder portion 605 (autoencoder model 604) to generate new low-dimensional data; comparing the new low-dimensional data with previous low-dimensional data; and adjusting (e.g., changing weights, changing constants, changing architecture, etc.) one or more components (autoencoder model 604 and / or prediction model 608) based on the comparison to reduce or minimize the difference between the new low-dimensional data and the previous low-dimensional data.

[0080] Figure 7 An embodiment of an autoencoder model 700 is shown; wherein the autoencoder model 700 comprises a modular autoencoder model. In contrast to a typical autoencoder model, the modular autoencoder model 600 can be configured with several different inputs and several different outputs.

[0081] Understanding the flow of information inside a typical autoencoder can be difficult. Its architecture is often opaque and / or non-transparent, and typically only information at model inputs, model outputs, and compression points (i.e., latent space) can be inferred. Information between these points is not easily interpretable. In practice, there may be auxiliary information available (in addition to the inputs), such as the physical properties of targets on a wafer and corresponding sensors in the semiconductor manufacturing process. This auxiliary information can be used as prior knowledge (e.g., "priors") to ensure that model predictions match physical reality, improve the performance of the autoencoder model, or expand the applicability of the autoencoder model. However, in a typical autoencoder model with a rigid architecture consisting of inputs, compression points, and outputs, it is unclear how to incorporate any such information (e.g., it is unclear where or how such information is inserted into the model or used by the model).

[0082] This particular example of the autoencoder model 700 has a modular structure. This allows the construction of an intermediate level of abstraction that can be used to exploit auxiliary information. The autoencoder model can be embodied as a computer program (e.g., instructions stored on a non-transitory computer-readable medium) that causes a computer (e.g., one or more processors) to execute (e.g., train and / or evaluate) the model 700 for parameter estimation and / or prediction. In some embodiments, the model 700 (and / or any individual component of the model 700 described below) can be configured a priori before seeing training data. In some embodiments, the estimated and / or predicted parameters include one or more of an image (e.g., pupil image, electric field image, etc.), a process measurement (e.g., a metric value), and / or other information. In some embodiments, the process measurement includes one or more of a metrology index, intensity, xyz position, size, electric field, wavelength, illumination and / or detection pupil, bandwidth, illumination and / or detection polarization angle, illumination and / or detection delay angle, and / or other process measurements. The modular autoencoder model 700 is configured for partially supervised learning (as further described below) for parameter estimation using a latent space.

[0083] like Figure 7 As shown, the modular autoencoder model 700 is composed of four types of sub-models: input model 702 (e.g., one or more input models 702 (a, b, ..., n)), common model 704, output model 706 (e.g., one or more output models 706 (a, b, ..., n)), and prediction model 708 (although the number, type and / or arrangement of sub-models is arbitrary). Input model 702 is configured to process input data into a higher level of abstraction suitable for combination with other inputs. Common model 704 may include Figure 6The autoencoder model 604 of FIG. 700 is used to connect the input (the output of the input model 702), compress the information to a bottleneck (e.g., a compression point or latent space in the model 700), and expand the information again to a level suitable for segmentation into multiple outputs (for the output model 706). The output model 706 can process the information from this common level of abstraction into multiple outputs that approximate the corresponding inputs. The prediction model 708 can be used to estimate the parameters of interest based on the information through the latent space. Finally, note that the modular autoencoder model 700 can be configured for several different inputs and several different outputs, which is contrary to the typical autoencoder model.

[0084] In general, the modular autoencoder model 700 can be more complex (in terms of the number of free parameters) than the typical monolithic models discussed above. However, in exchange, such more complex models are easier to interpret, define, and extend. The complexity of the network is chosen for any neural network. This complexity should be high enough to simulate the process behind the data, but low enough not to simulate noisy realizations (which is often interpreted as a form of overfitting). For example, the model can be configured to simulate the way a sensor sees the results of a manufacturing process on a wafer. Because the process that generates the data is often unknown (or has unknown aspects), choosing the appropriate network complexity often requires some intuition and trial and error. For this reason, it is desirable to provide an easily understood model architecture with the modular autoencoder model 700, and in which it is clear how to increase and decrease model complexity.

[0085] In some embodiments, an input model 702 may be provided to perform at least part of the encoding. For example, encoding may include processing one or more inputs 711 to a first level of dimensionality (e.g., by input model 702), and reducing the dimensionality of the combined processed inputs (e.g., by encoder portion 705). This may include reducing the dimensionality of inputs 711 to form low-dimensional data in latent space 707, and / or performing any degree of dimensionality reduction before actually reaching a low-dimensional level in latent space 707 (e.g., by one or more layers of encoder portion 705). Note that this dimensionality reduction is not necessarily monotonic. For example, the combination of inputs (via concatenation) may be viewed as an increase in dimensionality.

[0086] Here, one or more input models 702, public models 704, one or more output models 706, and / or prediction models 708 can be separated from each other and can be configured to correspond to process physics differences in different parts of the manufacturing process and / or sensing operation. The model 700 can be configured in this way so that each of the one or more input models 702, public models 704, one or more output models 706, and / or prediction models 708 can be trained together and / or separately, but are configured separately based on the process physics of the corresponding part of the manufacturing process and / or sensing operation, and are separated from other models in the modular autoencoder model 700. As a non-limiting example, physically, the target and sensor contributions in an optical measurement device (tool, platform, etc.) are separable. In other words, different targets can be measured using the same sensor. Therefore, the target and sensor contributions can be modeled separately. In other words, one or more input models 702, public models 704, one or more output models 706, and / or prediction models 708 can be associated with the physics of light propagating through a sensor or stack.

[0087] One or more input models 702 are configured to process one or more inputs 711 (e.g., 711a, 711b, ..., 711n) into a first-level dimension suitable for combination with other inputs. Processing may include filtering and / or otherwise converting the input to a model-friendly format, compressing the input, projecting the data into a subspace of lower dimension to accelerate the training process, data normalization, processing signal contributions from sensors (e.g., source fluctuations, sensor dose configuration (amount of light generated), etc.), and / or other processing operations. Processing may be considered preprocessing, for example, ensuring that the input or data related to the input is suitable for the model 700, suitable for combination with other inputs, etc. The first-level dimensionality may be the same as or less than the dimensionality of a given input 711. In some embodiments, one or more input models 702 include dense (e.g., linear layers and / or dense layers with different activations) feed-forward layers, convolutional layers, and / or residual network architectures of a modular autoencoder model 700. These structures are examples only and should not be considered limiting.

[0088] In some embodiments, input 711 may include images associated with a pupil, target, and / or other components of a semiconductor manufacturing process and is received from one or more of a plurality of characterization devices configured to generate input 711. Each input image may be pre-processed by input model 702 and encoded by encoder portion 705 (described below) of general model 704 into low-dimensional data representing the image in latent space 707 (described below). Note that in some embodiments, input model 702 may be or be considered to be part of encoder portion 705. The low-dimensional data may then be decoded for estimating and / or predicting process information, and / or for other purposes.

[0089] The general model 704 may include an encoder-decoder architecture, a variational encoder-decoder architecture, and / or other architectures, and may include the above-described autoencoder model 604. The low-dimensional data in the latent space 707 may have a reduced dimensionality obtained at the second level, the degree of reduction obtained at the second level being less than the first level (e.g., the dimensionality level of the processed input). In other words, the resulting dimensionality after reduction is less than the dimensionality before reduction. In some embodiments, the low-dimensional data in the latent space may have one or more different forms, such as tensors, vectors, and / or other latent space representations (e.g., something with fewer dimensions than the dimensions associated with a given input 711).

[0090] Predictive model 708 may operate in a similar or identical manner to predictive model 608 already described.

[0091] The one or more output models 706 may be configured to generate one or more different outputs 713 using one or more extended versions of the one or more inputs 711. The one or more different outputs 713 may include approximations of the one or more inputs 711, and the one or more different outputs 713 may have the same or increased dimensions compared to the extended versions of the one or more inputs 711 (e.g., outputs from the common model 704). In some embodiments, the one or more output models 706 may include dense feed-forward layers, convolutional layers, and / or residual network architectures of a modular autoencoder model, but these example structures are not intended to be limiting.

[0092] In some embodiments, the modular autoencoder model 700 can be trained on all sub-models 702-708 in an overall manner (although each model can also be trained separately). In other words, changing the data in the latent space 707 affects other components of the modular autoencoder model 700. In some embodiments, the adjustment includes adjusting at least one weight, constant and / or architecture (e.g., number of layers, etc.) associated with one or more input models 702, common models 704, one or more output models 706, prediction models 708, and / or layers of other components of the model 700. These and other aspects of training the modular autoencoder model 700 will be described in more detail with respect to other figures.

[0093] The modular autoencoder model 700 may have an extended scope of applicability for estimating parameters of interest for manufacturing and / or sensing (e.g., optical metrology) operations because it may be configured to enforce known properties of the input 711 in a decoder portion 709, which may include one or more output models 706. In some embodiments, the decoder portion 709 may be configured to generate an output 713 corresponding to the input 711 by decoding a low-dimensional representation of the input 711, while enforcing the known properties of the encoded input 711 during decoding (a result of the enforcement during training) to generate the output 713. In practice, the enforcement initially occurs during training. After training, the enforcement becomes a property of the model, although, strictly speaking, decoding is also performed during training. The known properties may be associated with a known physical relationship between the low-dimensional representation of the input 711 in the latent space 707 and the output 713. In some embodiments, the known properties may be known symmetry properties, known asymmetry properties, and / or other known properties. In some embodiments, the decoder portion 709 can be configured to exploit the modularity of the model 700 to enforce known properties at certain intermediate decoding levels (e.g., at the interface between the common model 704 and the output model 706). The parameters of interest can be estimated based on the output 713 and / or a low-dimensional representation of the input 711 in the latent space 707 (as described herein). For example, in some embodiments, for the prediction model, with respect to the use of symmetry, the prediction model can be a selection mask (e.g., selecting parameters associated with the parameters of interest from the latent space). This can still be represented as a neural network layer. However, it may remain fixed during training (e.g., it becomes a fixed linear layer σ(W x+b), where each row in W contains only a value of 1, the other elements are set to 0, b contains only elements equal to 0 and σ(·) is the identity).

[0094] In some embodiments, the decoder portion 709 (which, in some embodiments, may include one or more output models 706) may be configured to enforce known symmetry properties and / or other properties of the encoded input during a training phase such that the modular autoencoder model 700 obeys the enforced known symmetry properties (and / or other properties) to generate outputs during an inference phase. Enforcement includes using a penalty term in a cost function associated with the decoder portion 709 (which may include one or more output models 706) to penalize differences between the output 713 and the output that should be generated according to the known properties. The penalty term may include differences between decoded versions of low-dimensional representations of the inputs that are related to each other by physical priors. In some embodiments, the known property is a known symmetry, and the penalty term may include differences between decoded versions of low-dimensional representations of the inputs 711 reflected on or rotated around a symmetry point relative to each other. In some embodiments, one or more of the input model 702, the encoder portion 705, the decoder portion 709, the output model 706, the prediction model 708, and / or other components of the model 700 (see Figure 7 ) are configured to adjust (e.g., train or further train) based on any differences between decoded versions of the low-dimensional representations.

[0095] As a non-limiting example, an optical metrology platform (e.g., apparatus, tool, etc.) is configured to measure key semiconductor stack parameters directly on a product structure. To this end, machine learning methods are often applied to optical scattering measurement data acquired using the optical metrology platform. These machine learning methods are conceptually equivalent to supervised learning methods, i.e., learning from labeled data sets. The success of this method depends on the quality of the labels.

[0096] There are several common approaches to acquiring labels. One approach is to use a self-reference target, which is a target designed specifically for acquiring label data. A second approach relies on recording tools in semiconductor wafer fabs (usually scanning electron microscopes). The self-reference target approach is often preferred due to the competitive advantage of freedom in design and independence from competing metrology solutions.

[0097] One of the main challenges of using self-reference targets is that they only provide very accurate relative labels. This means that within a target cluster, there is some unknown cluster bias (target or cluster offset) based on which the accurate labels are known. Determining this unknown cluster bias, and thus obtaining absolute labels, is very important for the accuracy of manufacturing and / or inspection parameter configuration schemes based on self-reference targets. The process of estimating cluster bias can be called offset inference or label correction.

[0098] This offset inference problem is not solvable for linear signals (e.g., inputs 711 such as pupil images, etc.) as a function of the parameter of interest. Therefore, methods are being explored that exploit nonlinearities in the signals (e.g., pupil images and / or other inputs 711). Currently, we are not aware of methods that exploit physical assumptions about the nonlinearity of the signal and / or the spatial orientation of the signal.

[0099] When all asymmetry parameters are negated simultaneously, a signal of interest (e.g., input 711) (e.g., from a metrology platform), such as an antisymmetric cross-polarization pupil signal caused by overlay, is antisymmetric with respect to the stack parameterization (odd symmetry function). More specifically, when all other asymmetry parameters are zero, the signal may be antisymmetric about 0 overlay (odd symmetry function). This domain knowledge can be embedded into the model 700 during the training phase, which adds physical interpretability to the model 700. In addition, the symmetry point is important because it defines the origin (zero point) of the model parameterization, which can be used to calibrate the absolute accuracy in order to find appropriate correction labels. The model 700 is configured to exploit this and other physical understandings and embed them into the model 700. In this example, the general pupil properties exploited are as follows:

[0100]

[0101] Among them I a DE represents the antisymmetric standard pupil, c a is a collection of asymmetry parameters.

[0102] Input 711P may be a pupil image (for ease of notation, P = I a DE ), (P) may encode this pupil image (e.g., via one or more input models 702 and / or public models 704) into a compressed representation Finally by Decoding to generate an approximate pupil The model is trained as follows: Approximate real engraving ov, that is One of the elements in represents the overlay. For the self-referential objective, this model can be trained using the following objective (e.g., cost function):

[0103]

[0104] Here, the true overlay is set to ov = L + B, where the label L is known and the clustering bias B is unknown. In practice, this approach may not be sufficient because there is some freedom in choosing the clustering bias B. This is actually equivalent to moving the parameterization This can be problematic because an absolute overlay estimate is required. To reduce this ambiguity, another term can be added to the objective (cost function) that embeds the symmetry property of the signal (e.g., input 711) into the decoding model. (e.g., the public model 704 and / or one or more output models 706):

[0105]

[0106] For any In fact, there is no guarantee that any This cost function is minimized, but points can be sampled from the process window to ensure that the third term is small for arbitrarily large samples.

[0107] The principles described herein (e.g., exploiting the relatively low dimensionality of the latent space in a trained parameterized model to predict and / or otherwise determine process information) can have a variety of additional applications (e.g., in addition to and / or in lieu of the applications described above). For example, the present systems and methods can be used to coordinate data from different process sensors and / or tools, which may be different even for the same measurement or imaging target. As another example (among many other possible examples), the autoencoder models disclosed herein can be configured to perform tilt inference (and / or estimation, prediction, etc.) using wafer-level priors and / or other information.

[0108] One problem is etcher-induced tilt, which is caused by the bending of the electric field relative to the normal wafer direction. In some cases, for example, the maximum absolute tilt area occurs at or near the edge of the wafer. Typically, a fully unsupervised principal component analysis (PCA) method is used for tilt inference (e.g., estimating or predicting the tilt at the edge of the wafer). The raw pupil measurements are projected onto multiple linear primitives, and one of them is manually selected as representative of the tilt signal based on the expected tilt behavior. The coefficients obtained by projecting the signal onto the selected primitives are then fit to an exponential model (e.g., exponential in radial coordinates - exponential in polar coordinates) to extract signal components expected to be related to tilt and reject other possible components. Sometimes, full profile measurements that rely on inverse problems (such as CD reconstruction) can also be used for tilt inference. Using this approach, a physical model can be built and a parameterized stack signal can be estimated using an electromagnetic solver. The tilt estimate is produced by solving an optimization problem to find the parameterization that ensures the best fit.

[0109] Advantageously, the present autoencoder model can be configured to use wafer priors to ensure that an informed decomposition is performed, rather than in conjunction with or in addition to the uninformed approach used by PCA-based methods. The modular autoencoder model can be configured to encode the behavior of the plasma in the etch chamber, for example, which causes a (modeled) radial behavior on the wafer. This is due to the electric field bending at the edge of the wafer and / or other factors. This radial effect is projected onto the stacked features, whose behavior depends on the specific structure. For example, for an infinite grating, a sinusoidal variation in the direction of the electric field bending is expected based on the normal to the wafer and based on the grating direction. This can be interpreted as a projection onto the grating normal vector, which is the normal vector ("grating tilt"); it is maximum if it is orthogonal to the grating and minimum if it is parallel to the grating.

[0110] Figure 8 A schematic diagram of applying 801 priors (via model 850) to a modular autoencoder model 800 is shown. More specifically, Figure 8 A schematic diagram of an interconnected structure for generating labels is shown to apply priors to a modular autoencoder model 800. Priors can be and / or include, for example, known, target, and / or other predetermined values ​​for specific wafer and / or patterning process variables. Applying priors can include ensuring that the model operates according to certain rules and / or expectations (e.g., based on prior knowledge and / or physical understanding). Such knowledge cannot typically be learned from data, so applying priors can effectively add additional knowledge to the model.

[0111] Please note that Figure 8 In FIG. 8 , model 808 is a given example embodiment of prediction model 608, 708 (described above). In general, model 808 includes a block that connects potential representation 807 to the output (e.g., slope in this example) (e.g., Figure 8 The output of model 808 is shown, but model 808 can be any general prediction model). The output is constrained to belong to the signal classes that can be encoded by the prior. Note that the output of model 850 can only belong to the allowed signal classes, while the output of 808 is free at this stage.

[0112] During training, the present systems and methods are configured to ensure that the output of model 808 belongs to the appropriate class by training the output of model 808 to approximate the output of model 850. In this case, model 850 can be trained to simulate any acceptable signal in a class of possible signals. By ensuring that the output of model 808 approximates the output of model 850, the present systems and methods ensure that the output of model 808 belongs to the class of signals of interest, while still allowing the information (provided to 850) to be used to determine the exact information encoded. This is possible because the output of model 850 can also be changed to simulate specific data, as long as the change is within the class of possible signals.

[0113] In some embodiments, the modular autoencoder model 850 includes one or more auxiliary models 852 (including models 852a...852n, each model receiving a corresponding input 811a...811n of the input data 811) configured to generate labels 854 for at least some of the low-dimensional data in the latent space 807. The labels 854 are configured to be used by the prediction model 806 to estimate (e.g., predict, infer, etc.) parameters 815 (e.g., tilt and / or other parameters). In some embodiments, the labels 854 are configured to be used by the modular autoencoder model 850 to impose a behavior (e.g., a behavior based on one or more independent variables) on the latent space 807 and / or the output of the prediction model 808 (e.g., an estimate of the parameter 815). The behavior is associated with a class of possible signals (e.g., a tilt signal in this example, but any number of other possible signals can be considered). If the prediction model is a simple mask, such as Figure 8 As shown in 806, portions of the latent space can be subselected and the behavior imposed directly on the latent space. If a different model is used for the prediction model (e.g., different model 808), the imposed behavior is added to the output of the prediction model (e.g., different model 808), and the link to the latent space is less direct because it moves backward through the prediction model.

[0114] In some embodiments, one or more auxiliary models 852 include one or more wafer models. The wafer model represents a trainable model that imposes a desired behavior on the latent space 807. This helps to incorporate physical knowledge about the etching process (in this example) and its interaction with the stack during training of one or more models (e.g., 802 (e.g., 802a-802n), 804, 805, 809, 806 (e.g., 806a-806n), 808, and / or 852) of the modular autoencoder model 850. As described herein, these models can be neural networks, graphical models, and / or other models that are constrained to simulate expected physical behaviors (in this example, radial and sinusoidal tilt behaviors).

[0115] In some embodiments, one or more wafer models (e.g., auxiliary models 852) may be configured to separate pattern tilt from other asymmetries in the stack and / or pattern features. In this example, one or more wafer models are associated with pattern tilt, and the generated labels 854 are coupled with the dimension data predefined in the latent space 807 to correspond to the tilt, so that the modular autoencoder model 850 performs an informed decomposition based on the wafer prior.

[0116] In some embodiments, the input to one or more wafer models (e.g., one or more auxiliary models 852) includes data associated with a wafer pattern shape and / or wafer coordinates, which is configured to generate, encode and / or constrain a class of signals (e.g., a tilt signal in this example). The input to one or more wafer models (e.g., auxiliary models 852) may include a wafer radius 810 (r) and / or a (grating to wafer) angle, which includes a position in polar coordinates associated with a pattern on the wafer, and / or other information. A second angle 812 (φ) associated with a pattern on the wafer may also be used, as well as a wafer identification and / or other information. The angle includes a polar coordinate angle and a constant phase associated with a pattern direction on the wafer.

[0117] One or more appropriate auxiliary models can be selected (e.g., by a processor) 820 and used based on the inputs so that the labels 854 match the underlying parameters of the entire wafer. In this example, a sine function is used because sine-like behavior is expected. In this example, the tilt prior model has two inputs, namely the radius r and the angle phi. Note that this angle (in this example) is the sum of a constant angle determined by the alignment of the grating to the wafer and an angle related to the position on the wafer. This model can be considered a model of radial behavior, which produces the maximum tilt value (i.e., 820) for the case when the tilt of the plasma is completely orthogonal to the grating direction in the XY plane of the wafer. This value represents the radial component of the tilt a priori. This component can be compared to the one that depends on the alignment of the plasma with the grating (via )'s stack tilt because the alignment varies depending on the location on the wafer. Once the tilt model is built, it can be coupled to an autoencoder (at 854).

[0118] Figure 8The equation sin(φ)l shown in , starting from the selection of the sin projection, is derived from, for example, a model of etching-induced tilt. The alignment of ions in the etching plasma relative to the grating affects the tilt of the grating and can be modeled by sin(φ)l, given an appropriate definition of φ. Due to this projection, the tilt becomes zero, for example, when sin(φ)=0. In this case, the plasma is still bending, it just does not cause the grating to tilt. These example inputs for tilt inference are not intended to be limiting. There may be other inputs. For example, another tilt-inducing factor may be wafer stress. In some embodiments, pattern feature density may be used to inspire a location-based tilt parameterized wafer map model. The same type of construction applies, but different auxiliary models are generated. Other possible example behaviors that can be implemented are related to the location on the wafer where the tilt occurs, i.e., the edge of the wafer. The auxiliary model 852n can be configured (e.g., trained) to ensure that the tilt signal value inside the wafer is small to zero. Knowledge of the usage of the etching chamber can be another type of example information that can be associated with the tilt behavior and / or amplitude (and can be trained into the auxiliary model 852n). Using this information, the age of the controlling electric field (e.g., RF hours) or etcher settings (e.g., ring height, DC voltage, etc.) can be correlated to monotonic changes in the induced etch tilt.

[0119] It should be noted that the above Figure 8 The description is not intended to be limiting. For example, there are different inputs for different applications. As described above, tilt-related inputs may be associated with etch chamber use, grating orientation, radial variation, circular (sinusoidal) variation, pattern feature density, and / or other stack information. However, inputs (or priors) (for tilt and / or any other application) can generally be considered to be any data that can be used to infer, estimate, predict, or otherwise determine the shape, geometric information, and / or other information (e.g., any information to be extracted) associated with one or more parameters of interest 815. Examples of other types of inputs to one or more auxiliary models 852 include pupil data, data related to slit shape, and the like.

[0120] As another example, the modular autoencoder model 850 may include more or fewer auxiliary models 852 than described above, and / or the auxiliary models 852 may be arranged in a manner different from that described above. Figure 8. For example, one or more auxiliary models 852 can be embedded in one or more other models (e.g., encoder portion 805) in the modular autoencoder model 850. As a third example, the prediction model 808 can be formed by multiple separate models. In some embodiments, the prediction model 808 includes one or more prediction models, and the one or more prediction models are configured to estimate one or more parameters 815 based on labels 854 and / or one or more different outputs from the one or more auxiliary models 852. As a fourth example, in some embodiments, one or more auxiliary models 852 are configured to be nested with one or more other auxiliary models 852 and / or one or more other models (e.g., 802, 804, 806, 808) of the modular autoencoder model 850.

[0121] It is worth noting that, for example, students can serve as input to the auxiliary model, which may come from some special / dedicated goals and / or other sources.

[0122] In some embodiments, one or more auxiliary models 852 are configured to be trained using a cost function to minimize the difference between the generated labels 854 and the outputs (e.g., parameters 815) of one or more prediction models 808. One or more prediction models 808 are configured to select appropriate latent variables (e.g., depending on the parameters 815 of interest). One or more auxiliary models 852 are configured to be trained simultaneously with one or more input models 852, general model 804, one or more output models 806, and / or prediction models 808.

[0123] For example, for in-device measurement (IDM) technology, a typical target product stack usually contains multiple geometric parameters, some of which are target parameters for measurement and control. Usually, the measurement signal includes information with different intensities / sensitivities related to all geometric parameters. Since only some of the parameters are target parameters, the signal contribution of the remaining parameters should be considered and separated. In fact, all parameters except one or more target parameters act as interference parameters, thus hindering the estimation of the target parameters.

[0124] To give a specific example, eliminating the signal contribution of interference parameters is important for self-referenced offset inference performed by current linear offset inference techniques as well as self-referenced offset inference and / or tilt inference using autoencoder models. For example, current self-referenced linear offset inference schemes involve computing multiple interference wafer maps (spatial distributions of the parameters of interest) by appropriately decomposing the signal into constant signal patterns in each self-referenced target cluster using a model related to the target mask deviations. As mentioned earlier, stack parameterization can be divided into symmetric parameters and asymmetric parameters, depending on the geometric symmetry present in the structural units. Using the autoencoder model, all parameters that affect the signal can be modeled. Therefore, the task of parameter (interference) selection becomes the task of deciding how to distribute the signal energy between the parameters of interest and the interference parameters.

[0125] The (e.g. asymmetric) interference selection for linear shift inference involves several manual inspection steps to determine the right number of parameters to consider. This task remains error-prone and entirely manual. Moreover, due to the simplified nature of the used linear models, the accuracy of this decomposition only remains good within the linear range of the parameter / signal mapping; this cannot be controlled, especially for interference signals.

[0126] A more complete modeling technique uses a machine learning model to define a latent representation, such as the one described above and in Figures 5 to 8 In the autoencoder model described in , the number of interfering (extra) parameters needs to be selected in advance. In such a model, the allowed size of the parameterization c is crucial for the model's ability to capture all signal information. However, the energy split between the parameters is still arbitrary and the choice of signal split will degenerate if the number of parameters chosen is larger than necessary. The cause of this degradation can also be due to correlations between the parameters or reparameterization of the stack. Using the tilt inference example, this can lead to simple solutions such as predicting zero tilt and thus modeling the tilted signal energy with additional parameters. For self-reference offset inference, this can lead to incorrect signal segmentation and thus incorrect selection of the self-reference target cluster offset. Currently, manual inspection or comparison with third-party data is required to determine the appropriate number of parameters. Even if the appropriate number of interfering / extra asymmetry parameters is determined, the results often still vary greatly.

[0127] As mentioned above, physically inspired priors can be imposed; however, these can only guide the solution and the problem may still be partially ill-posed. Therefore, even with such priors, it is still possible to obtain naive solutions and poor energy splitting decisions, resulting in solutions that vary widely between training instances.

[0128] Therefore, a method is proposed for imposing a consistent ordering or ranking of latent elements in a latent representation (e.g., obtained using an autoencoder architecture or otherwise) based on each latent element's contribution to the quality of the signal approximation (e.g., a quantification of the contribution). This method can find and order a subset of latent elements to maximize the modeled signal energy.

[0129] Therefore, a method for sorting and / or selecting potential elements to model low-dimensional data in a latent space representation is disclosed herein, the low-dimensional data being a dimensionality-reduced representation of input data determined by a first model component of a model. The method comprises the following steps: obtaining the input data; training the model and selecting one of the potential element selections based on the training. The training comprises: reducing the dimensionality of the input data to generate the low-dimensional data in the latent space representation, the latent space representation comprising a plurality of latent elements; training a second model component of the model for each of one or more latent element selections, each of the latent element selections comprising one or more of the plurality of latent elements; and optimizing an approximation of the input data output by the second model component for each of the latent element selections, thereby sorting at least one of the plurality of latent elements in the latent space representation based on the contribution of each latent element to the input data.

[0130] It is suggested to rank the latent variables in the latent representation according to their importance / contribution to the measured signal. This approach can include building a model with latent variables c1, ...c n (where n is the total number of latent variables), which are ordered based on their contribution to the accuracy of the signal approximation (e.g., the pupil measured by an optical metrology tool) of a model using an ordered latent representation (e.g., an autoencoder). In such a model, the most important signal content can be modeled by the first latent element c1 (i.e., the parameter c1 explains the most energy in the signal that can be modeled using just one parameter). The remaining signal content will be modeled by the other elements, also ordered, so that by increasing the number of elements in c, up to a certain point, better and better models of the signal are obtained. The model accuracy stops improving when a certain number k of latent elements or variables are included, such that these first k latent elements c1, ..., c k Approximate all signal contents. This method may include including all first k potential elements c1, ..., c k to maximize the achievable approximation accuracy, where for any k, k≤n and for any . This idea extends the concept of (linear) low-rank approximation to nonlinear autoencoder models.

[0131] Fig. 9Schematic diagram of an autoencoder architecture according to such an embodiment. The input arrangement of the autoencoder model (input data 911; for example, pupils P1, P2, ..., P n ) and the encoder portion 905 can be connected with Figures 6 to 8 The autoencoder architecture of any of the above is substantially similar (and thus may optionally include a modular architecture with an input model, not shown in this figure). However, in this embodiment, the corresponding latent parameterization, latent element selection or latent masking 9281, 9282, 9283 based on the latent representation 907 is performed. k To train the decoder part 909 。 For each potential choice 9281, 9282, 928 k , the decoder portion 909 outputs corresponding output data 9131 (e.g., approximate pupil ), 9132 (e.g., approximate pupil )...913 k (For example, approximating the pupil ).

[0132] This architecture allows for monolithic training of the encoder 905 and decoder 909 while ranking the potential parameters according to the accuracy of the signal approximation (e.g., by ranking the output data 9131, 9132, ..., 913 k , compared with the input data 911). For example, the attributes can be encoded during the training of an autoencoder model with the goal of constructing an increasingly accurate pupil P using appropriate latent variable selections. i approximation For c = f(P1, P2, ..., P n ), the k-th order approximation applied by decoder 909 can be described as a function of the output data:

[0133] The latent representation c constructed k may include the same elements as the first position k of the complete latent representation c. In an embodiment, for each latent representation c k , the remaining elements k+1, ... may be set to 0 (i.e., the first latent representation may include a single latent variable c1 and zeros for the remaining elements). Alternatively, the remaining elements may be set to another fixed value (e.g., for some parameters of interest, it may be desirable to set these elements to values ​​other than zero, such as for symmetry parameters). Alternatively, these elements may vary between wafers / targets; for example, the values ​​of these elements may depend on the input data.

[0134] Therefore, each parameter choice may include a different truncation of the latent representation (i.e., all elements are zeroed after the truncation point). Each of these truncations can be used to train the decoder while optimizing the approximation of the input data output by the encoder for each of the latent elements. In this way, after training, the parameters will be ranked in descending order of importance.

[0135] In general, the optimization can be such that model training consists of minimizing the approximation error e k :

[0136]

[0137] For all order approximations k. In this way, the first element of the potential function should simulate the most important part of the signal, and the other elements simulate the remaining parts, and their contribution to the signal power gradually decreases. The approximation error e1>e2>…>e k It decreases with the increase of the approximation order k.

[0138] By way of more specific example, for overlay self-reference target offset inference and using conventional techniques, the cost function to be minimized can be expressed as:

[0139]

[0140] where o describes the programming deviation, is the model estimation bias, δ is the unknown process overlay offset (applicable to all targets of the reference target cluster), and g(·) is an optional prior / regularization term. This is an ill-posed problem.

[0141] Using the proposed method, the cost function can be defined as:

[0142]

[0143] where the prior / regularization term g(·) is again optional. k is a diagonal identity matrix of size kxk, which can include 1s on the diagonal at the first k positions. This cost function masks Applied to the full latent representation. In such an example, minimization may result in the overlay parameter being the most important parameter in the latent representation and associated with the first position.

[0144] The same approach can be used for other parameters of interest. Replaced by (etch) tilt term where t is the unknown tilt (the parameter of interest), is the tilt estimated by the model. These are just a few examples, and one skilled in the art will recognize that model training and latent representation optimization can be performed for many other parameters of interest derived from the measurement signal (e.g., the measurement pupil).

[0145] During training, this method can be applied to any sequence of masks to select latent elements. Described above is one possible option for applying the natural order of the signal energies; however, a different mask / order may be more suitable for other applications.

[0146] Different groups of latent elements that have been separated in the latent space (e.g., by another regularization term in the objective) may be treated separately. For example, in an embodiment, sorting may be applied only to asymmetric parameters (or only to symmetric parameters). Thus, symmetric latent parameters and asymmetric latent parameters may be treated separately. For example, if it is known that the parameters of interest are related to asymmetric signals and therefore to asymmetric latent parameters (such as overlay), then only these parameters may be included in the latent representation c. k In , all symmetric latent parameters are zeroed (i.e., assumed to be noise parameters before training). The opposite operation can be performed if the parameters of interest are known to be correlated only with the symmetric parameters.

[0147] Each latent element choice or mask can be determined a priori so that the model can be trained for a fixed latent sampling option. In this way, multiple such choices can be tested by training multiple models and determining which choice is preferred.

[0148] Once training is complete, the optimized (ordered) latent representation c k and a corresponding (auto)encoder model can be used to infer the parameter of interest (e.g. the parameter of interest for which the optimized latent representation is optimized) from the measurement data (measurement signal or pupil). Thus, the prediction model can include an optimization mask for the parameter of interest. However, more complex prediction models can also be used.

[0149] For parameter inference of interest, the data can be passed through the encoder and observed in the optimized / masked latent data (possibly using an additional prediction model as described above), or the related inverse problem can be solved as an optimization task (e.g., find latent data that, if passed through the decoder, would produce a prediction signal that matches the measured data).

[0150] The proposed approach improves autoencoder modeling by promoting consistent behavior across different instances of model creation / training, as it imposes a strict order in the latent representation. When the maximum number of latent variables chosen is large, it removes the ability of the model to reparameterize the signal, which again ensures consistency across models. The accuracy of tilt and offset inference models can be improved by requiring the model to consider all signals associated with overlay or tilt in a single latent parameter. Without this prior, the latent parameterization may be over-completed, such that the signal associated with the parameter of interest is actually modeled by multiple latent elements. Furthermore, this approach increases the ability of the model to learn a minimal latent representation if the configured latent representation has more elements than required. In this case, the unwanted elements are associated with signal energies that are arbitrarily close to zero. This allows the signal to be modeled using a (minimal) appropriate number of latent elements. As a result, this also simplifies the choice of the number of elements in the full latent representation, as this initial choice can now be larger than the number required for the inferred parameter of interest.

[0151] While the concepts described above use examples where the parameters of interest are overlay (particularly IDM overlay / offset inference) and / or etch tilt, it should be understood that these are merely examples and that the concepts are equally applicable to other parameters of interest, such as layer thickness, grating imbalance, edge position error, critical dimensions, any one or more non-overlay structural asymmetries (e.g., sidewall angle differences or bottom grating asymmetry)

[0152] Fig.10 1 is a block diagram illustrating a computer system 100 that can perform and / or assist in implementing the methods, processes, systems, or devices disclosed herein. The computer system 100 includes a bus 102 or other communication mechanism for communicating information, and a processor 104 (or multiple processors 104 and 105) coupled to the bus 102 to process information. The computer system 100 also includes a main memory 106, such as a random access memory (RAM) or other dynamic storage device, which is coupled to the bus 102 and is used to store information and instructions to be executed by the processor 104. The main memory 106 can also be used to store temporary variables or other intermediate information during the execution of instructions to be executed by the processor 104. The computer system 100 also includes a read-only memory (ROM) 108 or other static storage device, which is coupled to the bus 102 and is used to store static information and instructions for the processor 104. A storage device 110, such as a magnetic disk or optical disk, is provided and coupled to the bus 102 for storing information and instructions.

[0153] The computer system 100 may be coupled via bus 102 to a display 112, such as a cathode ray tube (CRT) or a flat panel or touch panel display, for displaying information to a computer user. An input device 114, including alphanumeric and other keys, is coupled to bus 102 for communicating information and command selections to processor 104. Another type of user input device is a cursor control 116, such as a mouse, trackball, or cursor direction keys, for communicating direction information and command selections to processor 104 and for controlling cursor movement on display 112. The input device typically has two degrees of freedom in two axes, a first axis (e.g., x) and a second axis (e.g., y), allowing the device to specify a position in a plane. A touch panel (screen) display may also be used as an input device.

[0154] According to one embodiment, portions of one or more methods described herein may be performed by computer system 100 in response to processor 104 executing one or more sequences of one or more instructions contained in main memory 106. Such instructions may be read into main memory 106 from another computer-readable medium (e.g., storage device 110). Execution of the sequence of instructions contained in main memory 106 causes processor 104 to perform the processing steps described herein. One or more processors in a multi-processing arrangement may also be used to execute the sequence of instructions contained in main memory 106. In alternative embodiments, hard-wired circuits may be used in place of or in combination with software instructions. Therefore, the description herein is not limited to any specific combination of hardware circuitry and software.

[0155] As used herein, the term "computer-readable medium" or "machine-readable" refers to any medium that participates in providing instructions to processor 104 for execution. Such media can take a variety of forms, including but not limited to non-volatile media, volatile media, and transmission media. Non-volatile media include, for example, optical or magnetic disks, such as storage device 110. Volatile media include dynamic memory, such as main memory 106. Transmission media include coaxial cables, copper wires, and optical fibers, including the wires that make up bus 102. Transmission media can also take the form of sound waves or light waves, such as those generated during radio frequency (RF) and infrared (IR) data communications. Common forms of computer-readable media include, for example, floppy disks, flexible disks, hard disks, magnetic tapes, any other magnetic media, CD-ROMs, DVDs, any other optical media, punch cards, paper tapes, any other physical media with hole patterns, RAM, PROMs and EPROMs, FLASH-EPROMs, any other memory chips or cassettes, carrier waves as described below, or any other medium that can be read by a computer.

[0156] Various forms of computer readable media may be used to carry one or more sequences of one or more instructions to processor 104 for execution. For example, the instructions may initially be carried on a disk of a remote computer. The remote computer may load the instructions into its dynamic memory and send the instructions over a telephone line using a modem. A modem local to computer system 100 may receive the data on the telephone line and use an infrared transmitter to convert the data to an infrared signal. An infrared detector coupled to bus 102 may receive the data carried in the infrared signal and place the data on bus 102. Bus 102 carries the data to main memory 106, from which processor 104 retrieves and executes the instructions. The instructions received by main memory 106 may optionally be stored on storage device 110 before or after execution by processor 104.

[0157] The computer system 100 may also include a communication interface 118 coupled to the bus 102. The communication interface 118 provides a two-way data communication coupled to a network link 120 connected to a local network 122. For example, the communication interface 118 may be an integrated services digital network (ISDN) card or a modem to provide a data communication connection to a corresponding type of telephone line. As another example, the communication interface 118 may be a local area network (LAN) card to provide a data communication connection to a compatible LAN. A wireless link may also be implemented. In any such implementation, the communication interface 118 sends and receives electrical, electromagnetic, or optical signals that carry digital data streams representing various types of information.

[0158] The network link 120 typically provides data communication through one or more networks to other data devices. For example, the network link 120 may provide a connection through a local network 122 to a host computer 124 or to data equipment operated by an Internet Service Provider (ISP) 126. The ISP 126, in turn, provides data communication services through a global packet data communication network, now commonly referred to as the "Internet" 128. Both the local network 122 and the Internet 128 use electrical, electromagnetic, or optical signals that carry digital data streams. The signals through the various networks and the signals on the network link 120 and communication interface 118, which carry the digital data to and from the computer system 100, are exemplary forms of carrier waves that transport the information.

[0159] The computer system 100 can send messages and receive data (including program code) through the network, network link 120, and communication interface 118. In the Internet example, the server 130 can transmit the requested code of the application program through the Internet 128, ISP 126, local network 122, and communication interface 118. For example, one such downloaded application program can provide all or part of the methods described herein. The received code can be executed when received by the processor 104, and / or stored in the storage device 110 or other non-volatile memory for later execution. In this way, the computer system 100 can obtain the application program code in the form of a carrier wave.

[0160] Further embodiments are disclosed in the following list of numbered clauses:

[0161] 1. A method for ranking and / or selecting latent elements to model low-dimensional data within a latent space representation, the low-dimensional data being a dimensionality-reduced representation of input data determined by a first model component of the model; the method comprising the steps of:

[0162] obtaining the input data;

[0163] Training the model, the training comprising:

[0164] reducing the dimensionality of the input data to generate the low-dimensional data in the latent space representation, the latent space representation comprising a plurality of latent elements; training a second model component of the model for each of one or more latent element selections, each of the latent element selections comprising one or more latent elements of the plurality of latent elements; and

[0165] optimizing, for each of the latent element selections, an approximation of the input data output by the second model component to rank at least one of the plurality of latent elements in the latent space representation based on a contribution of each latent element to the input data; and

[0166] One of the potential element selections is selected based on the training.

[0167] 2. A method according to clause 1, wherein the optimizing step ranks the one or more potential elements according to their contribution to the accuracy of the signal approximation of the second model component.

[0168] 3. A method according to clause 1 or 2, wherein the first potential element within the ordering explains the maximum energy of the input data that can be modeled with only one parameter.

[0169] 4. A method according to any of the preceding clauses, wherein the selection step comprises selecting one of the potential element selections such that the signal energy of the input data is maximized by the low-dimensional data when modeling is performed according to the selected potential element selection.

[0170] 5. A method as defined in any one of the preceding clauses, wherein the selecting step comprises selecting the potential element selection that uses the fewest potential elements to maximize the signal energy of the input data.

[0171] 6. A method as defined in any one of the preceding clauses, wherein the training step comprises, for each potential element selection, comparing the output of the second model component with the input data.

[0172] 7. A method according to any of the preceding clauses, wherein the training step comprises, for each potential element selection, minimizing the difference between the output of the second model component and the input data.

[0173] 8. A method as defined in any one of the preceding clauses, wherein each potential element is selected by passing a The mask is applied to the latent space representation to define, where I k is a diagonal identity matrix of size kxk including ones on the diagonal at the first k positions, where k is the potential element index from 1 to M, where M is less than or equal to the total number of the plurality of potential elements.

[0174] 9. A method as defined in any one of the preceding clauses, wherein the training and selection steps are performed to optimize a model for inferring one or more parameters of interest using the latent element selection.

[0175] 10. A method as defined in clause 9 above, wherein the one or more parameters of interest include one or more of the following: overlay, etch slope, layer thickness, grating imbalance, edge placement error, critical dimension, any one or more non-overlay structural asymmetries.

[0176] 11. A method as defined in any one of the preceding clauses, wherein each of the potential element selections includes the highest ranked potential elements, the number of the highest ranked potential elements included in each potential element selection increasing successively over the potential element selections.

[0177] 12. A method as defined in any one of the preceding clauses, wherein all other latent elements except the latent element sampled in each latent element selection are set to fixed values.

[0178] 13. A method as defined in any one of the preceding clauses, wherein all other potential elements except the sampled potential elements in each potential element selection are set to values ​​that are dependent on the input data.

[0179] 14. A method as defined in any of the preceding clauses, wherein the model is an autoencoder model, the first model component comprises an encoding component, and the second model component comprises a decoding component.

[0180] 15. A method as defined in clause 14, wherein the autoencoder model comprises a modular autoencoder model, further comprising one or both of the following:

[0181] one or more input models configured to process the input data into an intermediate level of dimensionality suitable for combining the input data, the intermediate level of dimensionality being intermediate between the dimensionality of the input data and the low-dimensional data; and

[0182] One or more output models configured to use the output data from the second model component to generate different output data having the same or increased dimensionality as the output data from the second model component.

[0183] 16. A method as defined in any one of the preceding clauses, wherein during the training step, different groups of the latent elements are processed separately.

[0184] 17. A method as defined in clause 16, wherein the separated groups are separated in the latent representation by regularization.

[0185] 18. A method as defined in clause 16 or 17, wherein the separated groups include a first group and a second group, the first group including symmetric potential elements of the potential elements and the second group including asymmetric potential elements of the potential elements.

[0186] 19. A method as defined in any one of the preceding clauses, comprising using the resulting trained model and the selected potential element selection to infer one or more values ​​of one or more parameters of interest based on input data.

[0187] 20. A method as defined in any of the preceding clauses, wherein the input data comprises or is derived from one or more pupil images, each of the pupil images comprising an angularly resolved distribution of radiation scattered from a target, each of the angularly resolved distributions comprising one or more of: an angularly resolved intensity distribution, an angularly resolved amplitude distribution and an angularly resolved phase distribution.

[0188] 21. A computer program comprising program instructions operable to perform the method of any one of items 1 to 20 when run on a suitable device.

[0189] 22. A non-transitory computer program carrier comprising the computer program of clause 21.

[0190] 23. A processing system comprising:

[0191] A non-transitory computer program carrier according to clause 22; and

[0192] A processor operable to execute the program instructions.

[0193] 24. A metrology apparatus configured to measure a product structure formed on a substrate during a lithography process, the metrology apparatus comprising a processing system according to clause 23.

[0194] The concepts disclosed herein can simulate or mathematically model any general imaging system for imaging sub-wavelength features, and are particularly useful for emerging imaging technologies that can produce shorter and shorter wavelengths. Emerging technologies already in use include EUV (extreme ultraviolet), DUV lithography that can produce 193nm wavelengths using ArF lasers, and even 157nm wavelengths using fluorine lasers. In addition, EUV lithography can produce wavelengths in the range of 20-5nm by using synchrotrons or by bombarding materials (solid or plasma) with high-energy electrons to produce photons in this range.

[0195] Although the concepts disclosed herein can be used for imaging on substrates such as silicon wafers, it should be understood that the disclosed concepts can be used with any type of lithography imaging system, such as systems and / or metrology systems for imaging on substrates other than silicon wafers. In addition, combinations and subcombinations of the disclosed elements may include separate embodiments. For example, predicting complex electric field images and determining metrology metrics such as overlay may be performed by the same parameterized model and / or different parameterized models. These features may include separate embodiments, and / or these features may be used together in the same embodiment.

[0196] Although embodiments of the invention may be specifically mentioned herein in the context of a metrology apparatus, embodiments of the invention may be used in other apparatus. Embodiments of the invention may form part of a mask inspection apparatus, a lithography apparatus, or any apparatus that measures or processes an object such as a wafer (or other substrate) or a mask (or other pattern forming device). These apparatus may generally be referred to as lithography tools. Such lithography tools may use vacuum conditions or ambient (non-vacuum) conditions.

[0197] Although specific reference may have been made above to the use of embodiments of the present invention in the context of optical lithography, it will be appreciated that the present invention is not limited to optical lithography and may be used in other applications, such as imprint lithography, where the context permits. Although specific embodiments of the present invention have been described above, it will be appreciated that the present invention may be implemented in a manner other than that described. The foregoing description is intended to be illustrative, not limiting. Therefore, those skilled in the art will appreciate that modifications may be made to the present invention as described without departing from the scope of the claims set forth below.

Claims

1. A method for ranking and / or selecting latent elements to model low-dimensional data within a latent space representation, the low-dimensional data being a dimensionality-reduced representation of input data determined by a first model component of the model; The method comprises the following steps: obtaining the input data; Training the model, the training comprising: reducing the dimensionality of the input data to generate the low-dimensional data in the latent space representation, the latent space representation comprising a plurality of latent elements; training a second model component of the model for each of one or more latent element selections, each of the latent element selections comprising one or more latent elements of the plurality of latent elements; and optimizing, for each of the latent element selections, an approximation of the input data output by the second model component to rank at least one of the plurality of latent elements in the latent space representation based on a contribution of each latent element to the input data; and One of the potential element selections is selected based on the training.

2. The method of claim 1, wherein the optimizing step ranks the one or more potential elements according to their contribution to the accuracy of the signal approximation of the second model component.

3. A method according to claim 1 or 2, wherein a first potential element within the ordering explains the maximum energy of the input data that can be modeled with only one parameter.

4. A method according to any one of the preceding claims, wherein the selecting step comprises selecting one of the latent element selections such that the signal energy of the input data is maximized by the low-dimensional data when modeling is performed according to the selected latent element selection.

5. A method according to any one of the preceding claims, wherein the training step comprises, for each potential element selection, comparing the output of the second model component with the input data.

6. A method according to any one of the preceding claims, wherein the training step comprises minimizing, for each potential element selection, the difference between the output of the second model component and the input data.

7. A method according to any one of the preceding claims, wherein the training and selecting steps are performed to optimize the model for inferring one or more parameters of interest using the potential element selections.

8. The method of claim 7, wherein the one or more parameters of interest include one or more of: overlay, etch slope, layer thickness, grating imbalance, edge placement error, critical dimension, any one or more non-overlay structural asymmetries.

9. A method according to any one of the preceding claims, wherein each of the potential element selections includes the highest ranked potential elements, the number of the highest ranked potential elements included in each potential element selection increasing successively across the potential element selections.

10. A method according to any one of the preceding claims, wherein the model is an autoencoder model, the first model component comprises an encoding component, and the second model component comprises a decoding component, and Wherein the autoencoder model comprises a modular autoencoder model, and further comprises one or both of the following: one or more input models configured to process the input data into an intermediate level of dimensionality suitable for combining the input data, the intermediate level of dimensionality being intermediate between the dimensionality of the input data and the low-dimensional data; and One or more output models configured to use the output data from the second model component to generate different output data having the same or increased dimensionality as the output data from the second model component.

11. A method according to any one of the preceding claims, wherein: During the training step, different groups of the latent elements are processed separately.

12. The method of claim 11, wherein the separated groups include a first group and a second group, the first group including symmetric potential elements of the potential elements and the second group including asymmetric potential elements of the potential elements.

13. A method according to any one of the preceding claims, comprising using the resulting trained model and the selected potential element selection to infer one or more values ​​of one or more parameters of interest based on input data.

14. A method according to any of the preceding claims, wherein the input data comprises or is derived from one or more pupil images, each of the pupil images comprising an angularly resolved distribution of radiation scattered from the target, each of the angularly resolved distributions comprising one or more of: an angularly resolved intensity distribution, an angularly resolved amplitude distribution and an angularly resolved phase distribution.

15. A metrology apparatus configured to measure a product structure formed on a substrate in a photolithography process, the metrology apparatus comprising a processing system operable to perform the method of any one of claims 1 to 14.