Deep Learning for Semantic Segmentation of Patterns

Abstract

The present invention describes a method for training a deep learning model of a patterning process. The method includes obtaining (i) training data, the training data including an input image and a real image of at least a portion of a substrate having multiple features; (ii) a set of categories, each category corresponding to a feature of multiple features of the substrate in the input image; and (iii) a deep learning model configured to receive the training data and the set of categories; generate a predicted image by modeling and / or simulating the deep learning model using the input image; assign categories in the set of categories to features in the predicted image based on matching features in the predicted image with corresponding features in the real image; and generate a trained deep learning model by iteratively assigning weights using a loss function and by modeling and / or simulating.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to U.S. Application No. 62 / 634,540, filed on February 23, 2018, the entire contents of which are incorporated herein by reference. Technical Field

[0003] The present disclosure relates to a technique for improving the performance of a device manufacturing process. The technique can be used in conjunction with a lithography apparatus or a metrology apparatus. Background Art

[0004] A lithographic apparatus is a machine that applies a desired pattern to a target portion of a substrate. A lithographic apparatus can be used, for example, to manufacture an integrated circuit (IC). In this case, a pattern forming device, alternatively referred to as a mask or reticle, can be used to generate a circuit pattern corresponding to a single layer of the integrated circuit, and the pattern can be imaged onto a target portion (e.g., a portion of a die, one or more dies) on a substrate (e.g., a silicon wafer) having a layer of radiation-sensitive material (resist). Typically, a single substrate will contain a network of adjacent target portions that are exposed in succession. Known lithographic apparatus include so-called steppers, in which each target portion is irradiated by exposing the entire pattern to the target portion at once, and so-called scanners, in which each target portion is irradiated by using a beam scanning pattern in a given direction (the "scanning" direction), while synchronously scanning a substrate parallel or antiparallel to that direction.

[0005] Before the circuit pattern is transferred from the pattern forming device to the substrate, the substrate may undergo various processes, such as primer coating, resist coating and soft baking. After exposure, other processes may be performed on the substrate, such as post-exposure baking (PEB), development, hard baking and measurement / inspection of the transferred circuit pattern. This array of processes is used as the basis for making a single layer of a device (e.g., IC). The substrate may then undergo various processes, such as etching, ion implantation (doping), metallization, oxidation, chemical mechanical polishing, etc., all of which are intended to complete a single layer of the device. If multiple layers are required in the device, the entire process or a variation thereof is repeated for each layer. Ultimately, the device will be present in each target portion on the substrate. These devices are then separated from each other by techniques such as cutting or sawing, so that a single device can be mounted on a carrier, connected to pins, etc.

[0006] Thus, manufacturing devices such as semiconductor devices typically involves using multiple fabrication processes to process a substrate (e.g., a semiconductor wafer) to form various features and multiple layers of the device. Such layers and features are typically manufactured and processed using, for example, deposition, photolithography, etching, chemical mechanical polishing, and ion implantation. Multiple devices can be made on multiple dies on a substrate and then separated into individual devices. The device manufacturing process can be considered a patterning process. The patterning process involves a patterning step, such as optical and / or nanoimprint lithography using a pattern forming device in a lithographic apparatus to transfer the pattern on the pattern forming device to the substrate, and typically but optionally involves one or more related pattern processing steps, such as resist development by a developing apparatus, substrate baking using a baking tool, etching patterns using an etching apparatus, and the like. Summary of the invention

[0007] According to one embodiment, a method for training a deep learning model of a patterning process is provided. The method includes obtaining (i) training data, the training data including an input image of at least a portion of a substrate having multiple features and a real image corresponding to the input image; (ii) a set of categories, each category corresponding to a feature of multiple features of the substrate in the input image; and (iii) a deep learning model, the deep learning model being configured to receive the training data and the set of categories; generating a predicted image by modeling and / or simulating the deep learning model using the input image, wherein the deep learning model iteratively performs a convolution operation using at least one dilation kernel including multiple weights; assigning a category in the set of categories to the features in the predicted image based on a match between the features in the predicted image and the corresponding features in the real image; and generating a trained deep learning model by modeling and / or simulating by iteratively assigning weights to at least one dilation kernel based on back propagation through the deep learning model using a loss function so that the predicted image reproduces multiple features of the input image.

[0008] In one embodiment, the convolution operation includes: overlapping the input image with the at least one dilation kernel; determining the product between each overlapping element, wherein the overlapping element includes the weight of the at least one dilation kernel and the pixel value of the input image; and summing the results of the product to determine the value of the convolution operation.

[0009] In an embodiment, at least one dilation kernel comprises a plurality of weights, a dilation rate, and a depth, wherein the depth is a multiple of the number of categories in the set of categories.

[0010] In one embodiment, the loss function is a function of the true image and the probability assigned to each class in the set of classes of the image being predicted.

[0011] In one embodiment, iteratively assigning weights includes modifying a plurality of weights, a dilation rate, and / or a depth of the dilation kernel to characterize at least a portion of the patterning process based on a local derivative of a loss function.

[0012] In one embodiment, at least one dilated kernel includes: a first dilated kernel in at least one dilated kernel, comprising a first weight set, a first dilation rate, and a depth equal to the number of categories in the category set; a second dilated kernel in at least one dilated kernel, comprising a second weight set, a second dilation rate, and a depth equal to the number of categories in the category set, wherein the second weight set is different from the first weight set and / or the second dilation rate is different from the first dilation rate.

[0013] In an embodiment, a first expansion core characterizes a first portion of the patterning process and a second core characterizes a second portion of the patterning process, wherein the second portion is a projection system of the patterning process and the second portion is a parameter related to a feature of the substrate.

[0014] In one embodiment, the expansion kernel is symmetric to characterize the projection system of the device for the patterning process.

[0015] In one embodiment, iteratively assigning weights includes assigning a first set of weights to a plurality of weights of the dilated kernel to characterize a projection system of an apparatus of the patterning process.

[0016] In one embodiment, the expansion ratio is an integer value between 1 and 10.

[0017] In addition, according to an embodiment, a method for applying a trained deep learning model to a patterning process is provided. The method includes obtaining (i) a trained deep learning model; (ii) a set of categories, each category corresponding to a feature of a training substrate for training the trained deep learning model; and (iii) an input image of at least a portion of a substrate undergoing a patterning process having a plurality of features; generating a segmented image by modeling and / or simulating the trained deep learning model using the input image and the set of categories, wherein a plurality of features of the input image are reproduced in the segmented image; and determining parameters of the patterning process based on the plurality of features of the segmented image.

[0018] In one embodiment, generating a segmented image includes assigning a category from the category set to each pixel of the segmented image, wherein the label corresponds to a feature from the feature set having a maximum probability, and the segmented image is pixelated.

[0019] In one embodiment, each pixel of the segmented image is associated with information including: the class set, a probability set corresponding to each class in the class set, and a location on the segmented image.

[0020] In one embodiment, determining parameters of the patterning process includes performing measurements on a plurality of features within the segmented image, the measurements including feature sizes and / or distances between one or more of the plurality of features; and determining the parameters based on the measurements related to the plurality of features.

[0021] In an embodiment, the parameters are edge positioning error, overlay, line edge roughness, necking and / or CD.

[0022] In an embodiment, the method further comprises determining an adjustment to the patterning process based on a parameter associated with the feature, and adjusting the patterning process in response to the adjustment.

[0023] In an embodiment, the adjustment includes adjustment of process variables including dose and / or focus. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A lithographic apparatus according to an embodiment is schematically depicted.

[0025] Figure 2 An embodiment of a lithographic cell or cluster is schematically depicted;

[0026] Figure 3 Example inspection equipment and metrology techniques are schematically depicted.

[0027] Figure 4 An example inspection device is schematically depicted.

[0028] Figure 5 The diagram shows the relationship between the illumination spot of the inspection equipment and the measurement target.

[0029] Figure 6 The process of deriving multiple variables of interest based on measured data is schematically depicted.

[0030] Figure 7 Example categories of treatment variables are shown.

[0031] Figure 8 The flowchart of a method according to an embodiment is schematically shown.

[0032] Fig. 9 The flowchart of a method according to an embodiment is schematically shown.

[0033] Fig.10 The flowchart of a method for training a deep learning model according to an embodiment is schematically shown.

[0034] Fig.11 FIG. 1 illustrates a training method according to an embodiment of the present invention. Fig.10 An example of a deep learning model.

[0035] Fig.12 An example of a pixelated input image for training a deep learning model is illustrated according to one embodiment.

[0036] Fig.13A An example of a dilated kernel for training a deep learning model according to one embodiment is illustrated.

[0037] Fig. 13B Illustrated are example weights for a dilated kernel used to train a deep learning model, according to an embodiment.

[0038] Fig.14A Illustrated is an example convolution operation with a first dilated kernel for training a deep learning model, according to an embodiment.

[0039] Fig. 14B Illustrated is an example convolution operation with a second dilated kernel for training a deep learning model, according to an embodiment.

[0040] Fig. 14C Illustrated is an example convolution operation with a third dilated kernel for training a deep learning model according to an embodiment.

[0041] Fig.15 is an example graph of local derivatives of a loss function according to an embodiment.

[0042] Fig.16A An example input image to a deep learning model according to an embodiment is illustrated.

[0043] Fig. 16B The diagram shows a method according to an embodiment of the present invention. Fig.16A An example of an input image that is predicted.

[0044] Fig. 16C The diagram shows a method according to an embodiment of the present invention. Fig. 16B Example uncertainty map for a predicted image.

[0045] Fig.16D The diagram shows a method according to an embodiment of the present invention. Fig. 16B Example error map for the predicted image.

[0046] Fig.17A The flowchart of a method for semantic segmentation based on a trained deep learning model according to one embodiment is schematically shown.

[0047] Fig. 17B According to one embodiment, Fig.17A Example segmented images of the trained deep learning model.

[0048] Fig.18A is an example input image with a charging effect according to an embodiment.

[0049] Fig.18B According to an embodiment Fig.18A Example segmented images of the input image.

[0050] Fig.19A According to an embodiment Fig.18A Cropping of the input image.

[0051] Fig.19B According to an embodiment Fig.18B Cropping of the segmented image.

[0052] Fig.19C According to an embodiment Fig.19A A cropped, enlarged portion of the image.

[0053] Fig.19D According to an embodiment Fig.19B A cropped, enlarged portion of the image.

[0054] Fig.19E The diagram shows a method for performing a Fig.19D Example probability distribution of features within a zoomed-in portion of .

[0055] Fig. 20A and 20B are example input images and segmented images, respectively, for training a second deep learning model according to an embodiment.

[0056] Fig.21A and 21B are example input images and segmented images, respectively, for training a third deep learning model according to an embodiment.

[0057] Fig. 22 is a block diagram of an example computer system. DETAILED DESCRIPTION

[0058] Before describing the embodiments in detail, it is helpful to present an exemplary environment in which the embodiments may be implemented.

[0059] Figure 1 A lithographic apparatus LA is schematically depicted. The lithographic apparatus comprises:

[0060] - an illumination system (illuminator) IL configured to condition a radiation beam B (e.g. UV radiation, EUV or DUV radiation);

[0061] a support structure (e.g. mask table) MT configured to support a patterning device (e.g. mask) MA and connected to a first positioner PM configured to accurately position the patterning device according to certain parameters;

[0062] a substrate table (e.g., wafer stage) WT (e.g., WTa, WTb, or both) configured to hold a substrate (e.g., a resist-coated wafer) W and connected to a second positioner PW configured to accurately position the substrate according to certain parameters;

[0063] - a projection system (e.g. a refractive projection lens system) PS, configured to project the pattern imparted to the radiation beam B by the pattern forming device MA onto a target portion C (e.g. comprising one or more dies and commonly referred to as a field) of the substrate W, the projection system being supported on a reference frame (RF).

[0064] As depicted here, the device is transmissive (eg, employing a transmissive mask). Alternatively, the device may be reflective (eg, employing a programmable mirror array of the type described above, or employing a reflective mask).

[0065] The illuminator IL receives a radiation beam from a radiation source SO. For example, when the source is an excimer laser, the source and the lithographic apparatus may be separate entities. In this case, the source is not considered to constitute part of the lithographic apparatus, the radiation beam being transferred from the source SO to the illuminator IL with the help of a beam delivery system BD, which comprises, for example, suitable directional mirrors and / or a beam expander. In other cases, for example, when the source is a mercury lamp, the source may be an integral part of the apparatus. The source SO and the illuminator IL and, if necessary, the beam delivery system BD may be referred to as a radiation system.

[0066] The illuminator IL may vary the intensity distribution of the beam. The illuminator may be arranged to limit the radial extent of the radiation beam so that the intensity distribution is non-zero within an annular region in a pupil plane of the illuminator IL. Additionally or alternatively, the illuminator IL may be operable to limit the distribution of the beam in the pupil plane so that the intensity distribution is non-zero in a plurality of equally spaced sectors in the pupil plane. The intensity distribution of the radiation beam in the pupil plane of the illuminator IL may be referred to as an illumination pattern.

[0067] Thus, the illuminator IL may comprise an adjuster AM configured to adjust the (angular / spatial) intensity distribution of the beam. Typically, at least the outer and / or inner radial extent (generally referred to as σ-outer and σ-inner, respectively) of the intensity distribution in a pupil plane of the illuminator may be adjusted. The illuminator IL may be operable to change the angular distribution of the beam. For example, the illuminator may be operable to change the number and angular extent of sectors in the pupil plane where the intensity distribution is non-zero. By adjusting the intensity distribution of the beam in the pupil plane of the illuminator, different illumination modes may be achieved. For example, by limiting the radial extent and the angular extent of the intensity distribution in the pupil plane of the illuminator IL, the intensity distribution may have a multipolar distribution, such as, for example, a dipole, quadrupole or hexapole distribution. The desired illumination mode may be obtained, for example, by inserting an optical device providing the illumination mode into the illuminator IL or by using a spatial light modulator.

[0068] The illuminator IL may be operable to change the polarization of the beam, and may be operable to adjust the polarization using an adjuster AM. The polarization state of the radiation beam across the pupil plane of the illuminator IL may be referred to as a polarization mode. Using different polarization modes may allow greater contrast to be obtained in the image formed on the substrate W. The radiation beam may be unpolarized. Alternatively, the illuminator may be arranged to linearly polarize the radiation beam. The polarization direction of the radiation beam may vary across the pupil plane of the illuminator IL. The polarization direction of the radiation may be different in different zones in the pupil plane of the illuminator IL. The polarization state of the radiation may be selected depending on the illumination mode. For a multipolar illumination mode, the polarization of each pole of the radiation beam may generally be perpendicular to the position vector of the pole in the pupil plane of the illuminator IL. For example, for a dipole illumination mode, the radiation may be linearly polarized in a direction substantially perpendicular to a line bisecting two opposite sectors of the dipole. The radiation beam may be polarized in one of two different orthogonal directions, which may be referred to as an X polarization state and a Y polarization state. For a quadrupole illumination mode, the radiation in the sectors of each pole may be linearly polarized in a direction substantially perpendicular to a line bisecting the sector. This polarization mode may be referred to as XY polarization. Similarly, for a hexapole illumination mode, the radiation in the sectors of each pole may be linearly polarized in a direction substantially perpendicular to a line bisecting the sector. This polarization mode may be referred to as TE polarization.

[0069] In addition, the illuminator IL typically includes various other components, such as an integrator IN and a condenser CO. The illumination system may include various types of optical components, such as refractive, reflective, magnetic, electromagnetic, electrostatic or other types of optical components, or any combination thereof, to direct, shape, or control the radiation.

[0070] The illuminator thus provides a conditioned radiation beam B having a desired uniformity and intensity distribution in its cross-section.

[0071] The support structure MT holds the pattern forming device in a manner that depends on the orientation of the pattern forming device, the design of the lithographic equipment, and other conditions such as, for example, whether the pattern forming device is maintained in a vacuum environment. The support structure can use mechanical, vacuum, electrostatic or other clamping techniques to hold the pattern forming device. The support structure can be a frame or a table, for example, the frame or table can be fixed or movable as required. The support structure can ensure that the pattern forming device is located in a desired position (for example relative to the projection system). Any term "reticle" or "mask" used in the present invention can be considered to be synonymous with the more general term "pattern forming device".

[0072] The term "patterning device" as used herein should be broadly interpreted as referring to any device that can be used to impart a pattern in a target portion of a substrate. In an embodiment, the patterning device is any device that can be used to impart a pattern to a radiation beam in its cross-section so as to produce a pattern in the target portion of the substrate. It should be noted that, for example, if the pattern imparted to the radiation beam includes phase-shifting features or so-called assist features, the pattern may not exactly correspond to the desired pattern in the target portion of the substrate. Typically, the pattern imparted to the radiation beam will correspond to a specific functional layer in a device (such as an integrated circuit) produced in the target portion.

[0073] The pattern forming device can be transmissive or reflective. Examples of pattern forming devices include masks, programmable mirror arrays, and programmable liquid crystal display (LCD) panels. Masks are well known in photolithography and include mask types such as binary mask types, alternating phase shift mask types, attenuated phase shift mask types, and various hybrid mask types. An example of a programmable mirror array uses a matrix arrangement of small mirrors, each of which can be individually tilted so as to reflect an incident radiation beam in different directions. The tilted mirrors impart a pattern in the radiation beam reflected by the mirror matrix.

[0074] The term "projection system" as used herein should be interpreted broadly to cover any type of projection system suitable for the exposure radiation used or for other factors such as the use of an immersion liquid or the use of a vacuum, including refractive, reflective, catadioptric, magnetic, electromagnetic and electrostatic optical systems, or any combination thereof. Any use of the term "projection lens" herein may be considered synonymous with the more general term "projection system".

[0075] The projection system PS has an optical transfer function that may be non-uniform and may affect the pattern imaged on the substrate W. For unpolarized radiation, these effects can be described fairly well by two scalar maps that describe the transmission (apodization) and relative phase (aberration) of the radiation as a function of the position in the pupil plane of the radiation exiting the projection system PS. These scalar maps, which may be referred to as transmission maps and relative phase maps, may be expressed as linear combinations of a full set of basis functions. A particularly convenient set is the Zernike polynomials, which form a set of orthogonal polynomials defined on the unit circle. The determination of each scalar map may involve determining the coefficients in such an expansion. Since the Zernike polynomials are orthogonal on the unit circle, the Zernike coefficients may be determined by calculating the inner product of the measured scalar map with each Zernike polynomial in turn and dividing such inner product by the square of the norm of the Zernike polynomial.

[0076] The transmission map and the relative phase map are field and system dependent. That is, typically, each projection system PS will have a different Zernike expansion for each field point (i.e., for each spatial location in the image plane of the projection system PS). The relative phase of the projection system PS in its pupil plane can be determined by projecting radiation (e.g.) from a point source in the object plane of the projection system PS (i.e., the plane of the patterning device MA) through the projection system PS and using a shearing interferometer to measure the wavefront (i.e., the trajectory of points with the same phase). The shearing interferometer is a common path interferometer and therefore, advantageously, no secondary reference beam is required to measure the wavefront. The shearing interferometer may comprise: a diffraction grating, e.g., a two-dimensional grating in the image plane of the projection system (i.e., the substrate table WT); and a detector arranged to detect an interference pattern in a plane conjugate to the pupil plane of the projection system PS. The interference pattern is related to the derivative of the phase of the radiation with respect to the coordinates in the pupil plane in the shear direction. The detector may comprise an array of sensing elements such as, for example, a charge coupled device (CCD).

[0077] The projection system PS of the lithographic apparatus may not produce visible fringes, and therefore, a phase stepping technique (such as, for example, moving a diffraction grating) may be used to enhance the accuracy of the wavefront determination. Stepping may be performed in the plane of the diffraction grating and in a direction perpendicular to the scanning direction of the measurement. The stepping range may be one grating period, and at least three (uniformly distributed) phase steps may be used. Thus, for example, three scanning measurements may be performed in the y direction, each for a different position in the x direction. Such stepping of the diffraction grating effectively converts phase changes into intensity changes, allowing the phase information to be determined. The grating may be stepped in a direction perpendicular to the diffraction grating (z direction) to calibrate the detector.

[0078] The diffraction grating may be scanned sequentially in two perpendicular directions, which may coincide with the axes (x and y) of the coordinate system of the projection system PS, or may be at an angle, such as 45 degrees, to these axes. The scan may be performed over an integer number of grating periods, such as one grating period. The scan averages the phase changes in one direction, thereby allowing the phase changes in the other direction to be reconstructed. This allows the wavefront to be determined as a function of the two directions.

[0079] The transmission (apodization) of the projection system PS in its pupil plane may be determined by projecting radiation, for example, from a point source in the object plane of the projection system PS (i.e. the plane of the patterning device MA) through the projection system PS and using a detector to measure the intensity of the radiation in a plane conjugate to the pupil plane of the projection system PS. The same detector may be used as that used to measure the wavefront to determine the aberrations.

[0080] The projection system PS may include a plurality of optical (e.g., lens) elements and may also include an adjustment mechanism AM configured to adjust one or more of the optical elements to correct aberrations (phase variations in the pupil plane across the entire field). To achieve such correction, the adjustment mechanism may be operable to manipulate one or more optical (e.g., lens) elements within the projection system PS in one or more different ways. The projection system may have a coordinate system in which the optical axis of the projection system extends in the z direction. The adjustment mechanism may be operable to perform any combination of: shifting one or more optical elements; tilting one or more optical elements; and / or deforming one or more optical elements. The displacement of the optical element may be in any direction (x, y, z, or a combination thereof). The tilting of the optical element is typically by rotating around an axis in the x and / or y direction so as to be out of a plane perpendicular to the optical axis, but rotation around the z axis may be used for non-rotationally symmetric aspheric optical elements. The deformation of the optical element may include low-frequency shapes (e.g., astigmatism) and / or high-frequency shapes (e.g., free-form aspheric surfaces). For example, deformation of the optical element may be performed by using one or more actuators to apply forces to one or more sides of the optical element and / or by heating one or more selected regions of the optical element using one or more heating elements. Typically, it is not possible to adjust the projection system PS to correct for apodization (variations in transmission across the pupil plane). A transmission map of the projection system PS may be used when designing a patterning device (e.g., mask) MA for the lithographic apparatus LA. Using computational lithography techniques, the patterning device MA may be designed to at least partially correct for apodization.

[0081] The lithographic apparatus may be of a type having two (dual stage) or more tables (e.g. two or more substrate tables WTa, WTb, two or more patterning device tables, substrate table WTa and table WTb below the projection system in the absence of a dedicated substrate for, e.g., facilitating measurement and / or cleaning, etc.). In these "multi-stage" machines, the additional tables may be used in parallel, or preparatory steps may be performed on one or more tables while one or more other tables are being used for exposure. For example, alignment measurements using an alignment sensor AS and / or level (height, inclination, etc.) measurements using a level sensor LS may be performed.

[0082] The lithographic apparatus may also 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 (e.g., water) so as to fill the space between the projection system and the substrate. Immersion liquid may also be applied to other spaces in the lithographic apparatus, for example, the space between the patterning device and the projection system. Immersion techniques are known in the art for increasing the numerical aperture of the projection system. The term "immersion" as used herein does not mean that structures such as the substrate must be immersed in the liquid, but only means that the liquid is located between the projection system and the substrate during exposure.

[0083] Thus, in operation of the lithographic apparatus, a radiation beam is conditioned and provided by the illumination system IL. The radiation beam B is incident on a pattern forming device (e.g. a mask) MA held on a support structure (e.g. a mask table) MT and is patterned by the pattern forming device. After passing through the pattern forming device MA, the radiation beam B passes through a projection system PS, which focuses the radiation beam onto a target portion C of the substrate W. With the aid of a second positioner PW and a position sensor IF (e.g. an interferometer device, a linear encoder, a 2D encoder or a capacitive sensor), the substrate table WT can be accurately moved, for example in order to position different target portions C in the path of the radiation beam B. Similarly, for example, after mechanical retrieval from a mask library, or during scanning, the first positioner PM and a further position sensor (which is not in the Figure 1The support structure MT is used to accurately position the pattern forming device MA relative to the path of the radiation beam B (as explicitly depicted in Figure 1). Typically, the movement of the support structure MT can be achieved with the help of a long-stroke module (coarse positioning) and a short-stroke module (fine positioning), which form part of the first positioner PM. Similarly, the movement of the substrate table WT can be achieved using a long-stroke module and a short-stroke module, which form part of the second positioner PW. In the case of a stepper (as opposed to a scanner), the support structure MT may be connected only to the short-stroke actuator, or may be fixed. The pattern forming device MA and the substrate W can be aligned using pattern forming device alignment marks M1, M2 and substrate alignment marks P1, P2. Although the substrate alignment marks occupy dedicated target portions as shown, the substrate alignment marks may be located in the space between the target portions (these substrate alignment marks are called scribing alignment marks). Similarly, in the case where more than one die is provided on the pattern forming device MA, the pattern forming device alignment marks may be located between the dies.

[0084] The depicted device can be used in at least one of the following modes:

[0085] 1. In step mode, the support structure MT and substrate table WT are held substantially stationary while an entire pattern imparted to the radiation beam is projected at once onto a target portion C (i.e. a single static exposure). The substrate table WT is then shifted in the X and / or Y direction so that a different target portion C can be exposed. In step mode, the maximum size of the exposure field limits the size of the target portion C imaged in a single static exposure.

[0086] 2. In scan mode, the support structure MT and substrate table WT are scanned synchronously while a pattern imparted to the radiation beam is projected onto a target portion C (i.e. a single dynamic exposure). The speed and direction of the substrate table WT relative to the support structure MT may be determined by the (de-)magnification and image reversal characteristics of the projection system PS. In scan mode, the maximum size of the exposure field limits the width of the target portion in a single dynamic exposure (in the non-scanning direction), while the length of the scanning movement determines the height of the target portion (in the scanning direction).

[0087] 3. In another mode, the support structure MT holding the programmable patterning device is held substantially stationary and the substrate table WT is moved or scanned while a pattern imparted to the radiation beam is projected onto a target portion C. In this mode, a pulsed radiation source is typically employed and the programmable patterning device is updated as required after each movement of the substrate table WT or between successive radiation pulses during a scan. This mode of operation can be readily applied to maskless lithography using a programmable patterning device, such as a programmable mirror array of the type described above.

[0088] Combinations and / or variations on the above described modes of use, or entirely different modes of use, may also be employed.

[0089] Although specific reference may be made herein to the use of lithographic apparatus in the manufacture of ICs, it should be understood that the lithographic apparatus described herein may have other applications, such as the manufacture of integrated optical systems, guidance and detection patterns of magnetic domain memories, liquid crystal displays (LCDs), thin film magnetic heads, etc. Those skilled in the art will appreciate that in the context of such alternative applications, any term "wafer" or "die" used herein may be considered synonymous with the more general term "substrate" or "target portion", respectively. The substrate referred to herein may be processed before or after exposure in, for example, a coating and developing system (a tool that typically applies a resist layer to a substrate and develops the exposed resist), a metrology tool, or an inspection tool. Where applicable, the disclosure herein may be applied to such and other substrate processing tools. In addition, a substrate may be processed more than once, for example to produce a multi-layer IC, so that the term substrate used herein may also refer to a substrate that already contains multiple processed layers.

[0090] As used herein, the terms "radiation" and "beam" include all types of electromagnetic radiation, including ultraviolet (UV) radiation (e.g., having a wavelength of 365 nm, 248 nm, 193 nm, 157 nm or 126 nm) and extreme ultraviolet (EUV) radiation (e.g., having a wavelength in the range of 5 nm to 20 nm), as well as particle beams such as ion beams or electron beams.

[0091] Various patterns on or provided by a pattern forming device may have different process windows, i.e., the space of process variables under which patterns are produced within specifications. Examples of pattern specifications related to potential system defects include inspections of necking, line pullback, line thinning, CD, edge positioning, overlap, resist top loss, resist undercut and / or bridging. The process windows of all patterns on a pattern forming device or a region thereof may be obtained by merging (e.g., overlapping) the process windows of each individual pattern. The boundaries of the process windows of all patterns contain the boundaries of the process windows of some of the individual patterns. In other words, these individual patterns limit the process windows of all patterns. These patterns may be referred to as "hot spots" or "process window limit patterns (PWLP)", which are used interchangeably herein. When controlling portions of the patterning process, it is possible and economical to focus on hot spots. When there are no defects in the hot spots, it is likely that all patterns are defect-free.

[0092] like Figure 2 As shown, the lithographic equipment LA can constitute part of the lithographic unit LC, and is sometimes referred to as a lithographic cell or cluster, which can also include equipment for performing pre-exposure and post-exposure processing on the substrate. Conventionally, these equipment include: one or more spin coaters SC for depositing one or more resist layers, one or more developers DE for developing the exposed resist, one or more chill plates CH and one or more baking plates BK. The substrate handler or robot RO picks up one or more substrates from the input / output ports I / O1, I / O2, moves the substrate between different process equipment, and then transfers the substrate to the feed station LB of the lithographic equipment. These equipment are generally collectively referred to as coating and developing systems, and are controlled by a coating and developing system control unit TCU, which is itself controlled by a management control system SCS, which also controls the lithographic equipment via a lithography control unit LACU. Therefore, different equipment can be operated to maximize production volume and processing efficiency.

[0093] In order to correctly and consistently expose a substrate exposed by a lithographic apparatus, and / or in order to monitor a portion of a patterning process (e.g., a device manufacturing process) comprising at least one pattern transfer step (e.g., a lithography step), it is desirable to inspect a substrate or other object to measure or determine one or more properties, such as alignment, overlay (which may be, for example, between structures in superimposed layers, or between structures in the same layer that have been provided to the same layer by, for example, a double patterning process), line thickness, critical dimension (CD), focus offset, material properties, etc. Accordingly, the manufacturing facility in which the lithography cell LC is located also typically includes a metrology system MET that measures some or all of the substrates W that have been processed in the lithography cell or measures other objects in the lithography cell. The metrology system MET may be part of the lithography cell LC, for example, the metrology system may be part of the lithography apparatus LA (such as an alignment sensor AS).

[0094] One or more measured parameters may include, for example, overlap between successive layers formed in or on a patterned substrate, critical dimensions (CD) of features formed in or on a patterned substrate (e.g., critical line width), focus or focus error of an optical lithography step, dose or dose error of an optical lithography step, optical aberration of an optical lithography step, etc. This measurement may be performed on a target on the product substrate itself and / or on a dedicated metrology target disposed on the substrate. The measurement may be performed after resist development but before etching, or the measurement may be performed after etching.

[0095] There are a variety of techniques for measuring the structure formed in the patterning process, including the use of scanning electron microscopes, image-based measurement tools and / or various special tools. As described above, a fast and non-invasive form of a special measurement tool is a tool in which a radiation beam is directed to a target on the substrate surface and the properties of the scattered (diffraction / reflection) beam are measured. By evaluating one or more properties of the radiation scattered by the substrate, one or more properties of the substrate can be determined. This can be called diffraction-based measurement. One such application of diffraction-based measurement is in the measurement of feature asymmetry within the target. For example, this can be used as a measure of overlap, but other applications are also known. For example, asymmetry can be measured by comparing relative parts of the diffraction spectrum (for example, comparing the -1 order and +1 order in the diffraction spectrum of a periodic grating). This can be done as described above and as described, for example, in U.S. Patent Application Publication US2006-066855, the full text of which is incorporated herein by reference. Another application of diffraction-based metrology is the measurement of feature width (CD) within the target. This technology can use the equipment and methods described below.

[0096] Thus, in a device fabrication process (e.g., a patterning process or a lithography process), various types of measurements may be made of a substrate or other object during or after the process. The measurements may determine whether a particular substrate is defective, may make adjustments to the process and equipment used in the process (e.g., aligning two layers on a substrate or aligning a patterning device to a substrate), may measure the performance of the process and equipment, or may be used for other purposes. Examples of measurements include optical imaging (e.g., optical microscopy), non-imaging optical measurements (e.g., diffraction-based measurements such as ASML YieldStar metrology tools, ASML SMASH metrology systems), mechanical measurements (e.g., profiling using a stylus, atomic force microscopy (AFM)), and / or non-optical imaging (e.g., scanning electron microscopy (SEM)). The SMASH (Smart Alignment Sensor Hybrid) system, as described in U.S. Patent 6,961,116, (the entire contents of which are incorporated herein by reference), employs a self-referencing interferometer that produces two overlapping and relatively rotated images of an alignment mark, detects the intensity in the pupil plane at which the Fourier transforms of the images interfere, and extracts position information from the phase difference between the diffraction orders of the two images, which manifests itself as intensity variations of the interference orders.

[0097] The metrology results may be provided directly or indirectly to the supervisory control system SCS. If an error is detected, the exposure of subsequent substrates may be adjusted (especially if the inspection can be performed quickly and rapidly enough that one or more other substrates of the same batch are still to be exposed) and / or the subsequent exposure of the exposed substrate may be adjusted. Furthermore, already exposed substrates may be stripped and reworked to improve yield or scrapped, thereby avoiding further processing of substrates known to be defective. In the case that only some target portions of the substrate are defective, further exposures may be performed only on those target portions that are good.

[0098] Within the metrology system MET, a metrology device is used to determine one or more properties of a substrate, and in particular to determine how one or more properties of different substrates vary or how different layers of the same substrate vary between layers. As described above, the metrology device may be integrated into the lithographic apparatus LA or the lithographic cell LC, or may be a stand-alone device.

[0099] In order to achieve the measurement, one or more targets may be provided on the substrate. In one embodiment, the target is specifically designed and may include a periodic structure. In one embodiment, the target is a part of a device pattern, such as a periodic structure of the device pattern. In one embodiment, the device pattern is a periodic structure of a memory device (e.g., a bipolar transistor (BPT), a shared bit bus (BLC), etc.).

[0100] In an embodiment, the target on the substrate may include one or more one-dimensional periodic structures (e.g., gratings) that are printed such that after development, the periodic structural features are formed by solid resist lines. In an embodiment, the target may include one or more two-dimensional periodic structures (e.g., gratings) that are printed such that after development, one or more periodic structures are formed by solid resist pillars or vias in the resist. The bars, pillars, or vias may alternatively be etched into the substrate (e.g., into one or more layers on the substrate).

[0101] In one embodiment, one of the parameters of interest in the patterning process is overlap. Overlap can be measured using a dark field scattering method, in which the zero-order diffraction (corresponding to specular reflection) is blocked and only higher-order diffraction is processed. Examples of dark field measurements can be found in PCT patent application publications WO2009 / 078708 and WO2009 / 106279, the entire contents of which are incorporated herein by reference. Further developments of the technology have been described in U.S. patent application publications US2011-0027704, US2011-0043791, and US2012-0242970, the entire contents of which are incorporated herein by reference. Overlap based on diffraction detection using dark field detection of diffraction orders enables overlap measurements on smaller targets. These targets can be smaller than the irradiation spot and can be surrounded by device product structures on the substrate. In one embodiment, multiple targets can be measured in one radiation acquisition.

[0102] Figure 3 An example inspection device (e.g., a scatterometer) is depicted. The inspection device comprises a broadband (white light) radiation projector 2 which projects radiation onto a substrate W. The redirected radiation is passed to a spectrometer detector 4 which measures a spectrum 10 (intensity as a function of wavelength) of the specularly reflected radiation, as shown, for example, in the lower left figure. From this data, the specularly reflected radiation may be analyzed by a processing unit PU (e.g., by rigorous coupled wave analysis and nonlinear regression, or by comparing with, for example, Figure 3 By comparing the measured spectra to a library of simulated spectra shown in the lower right), the structure or profile that gave rise to the detected spectra can be reconstructed. Typically, for the reconstruction, the general form of the structure is known, and some variables are assumed based on knowledge of the process by which the structure was made, leaving only a few variables of the structure to be determined from the measured data. Such inspection devices can be configured as normal incidence inspection devices or oblique incidence inspection devices.

[0103] Another inspection device that can be used is Figure 4In this arrangement, radiation emitted by a radiation source 2 is collimated using a lens system 12, transmitted through an interference filter 13 and a polarizer 17, reflected by a partially reflective surface 16 and focused via an objective lens 15 into a spot S on a substrate W, the objective lens 15 having a high numerical aperture (NA), desirably at least 0.9 or at least 0.95. Immersion inspection equipment (using a relatively high refractive index fluid, such as water) may even have a numerical aperture greater than 1.

[0104] As in the lithographic apparatus LA, one or more substrate tables may be provided to hold a substrate W during measurement operations. The substrate table may be similar in form to Figure 1 The substrate table WT of the inspection apparatus is similar or identical to that of the lithographic apparatus. In an example where the inspection apparatus is integrated with the lithographic apparatus, the inspection apparatus and the lithographic apparatus may even be the same substrate table. The coarse positioner and the fine positioner may be provided on a second positioner PW, which is configured to accurately position the substrate relative to the measuring optical system. For example, various sensors and actuators are provided to acquire the position of an object of interest and bring the position of the object of interest to an appropriate position below the objective lens 15. Typically, many measurements will be made on targets at different locations across the substrate W. The substrate support can be moved in the X and Y directions to acquire different targets and in the Z direction to acquire the desired position of the target relative to the focus of the optical system. When (for example) in practice the optical system can remain essentially stationary (typically in the X and Y directions, but perhaps also in the Z direction) and only the substrate moves, it is convenient to consider and describe the operation as if the objective lens is being brought to different locations relative to the substrate. If the relative position of substrate and optical system is correct, then in principle it does not matter which of the substrate and the optical system moves in the real world, or if both are moving, or a combination of one part of the optical system moving (e.g. in Z and / or tilt direction), the rest of the optical system being stationary and the substrate moving (e.g. in X and Y direction, but optionally also in Z and / or tilt direction).

[0105] The radiation redirected by the substrate W is then passed through the partially reflective surface 16 to the detector 18 for detection of the spectrum. The detector 18 may be located at the back-projected focal plane 11 (i.e., at the focal length of the lens system 15), or the plane 11 may be re-imaged onto the detector 18 using auxiliary optics (not shown). The detector may be a two-dimensional detector, so that a two-dimensional angular scatter spectrum of the substrate target 30 may be measured. The detector 18 may be, for example, an array of CCD or CMOS sensors, and an integration time of, for example, 40 milliseconds per frame may be used.

[0106] The reference beam can be used, for example, to measure the intensity of the incident radiation. To this end, when the radiation beam is incident on the partially reflecting surface 16, a portion of the radiation beam is transmitted through the partially reflecting surface 16 as a reference beam towards the reference mirror 14. The reference beam is then projected onto a different portion of the same detector 18, or alternatively onto a different detector (not shown).

[0107] One or more interference filters 13 may be used to select wavelengths of interest in the range of, for example, 405 nm to 790 nm or even lower, such as 200 nm to 300 nm. The interference filters may be tunable, rather than comprising different sets of filters. Gratings may be used instead of interference filters. An aperture stop or spatial light modulator (not shown) may be provided in the illumination path to control the range of angles of incidence of the radiation on the target.

[0108] The detector 18 can measure the intensity of the redirected radiation at a single wavelength (or a narrow wavelength range), the intensity at multiple wavelengths, or the intensity integrated over the wavelength range. In addition, the detector can measure the intensity of the transverse magnetic polarization radiation and the transverse electric polarization radiation and / or the phase difference between the transverse magnetic polarization radiation and the transverse electric polarization radiation, respectively.

[0109] The target 30 on the substrate W may be a one-dimensional grating that is printed so that after development, the bars are formed by solid resist lines. The target 30 may be a two-dimensional grating that is printed so that after development, the grating is formed by solid resist columns or through holes in the resist. The bars, columns or through holes may be etched into or onto the substrate (e.g., etched into one or more layers on the substrate). The pattern (e.g., the pattern of bars, columns or through holes) is sensitive to changes in processing during the patterning process (e.g., optical aberrations, focus changes, dose changes, etc. in the lithographic projection equipment (particularly the projection system PS)) and will show changes in the printed grating. Accordingly, the measurement data of the printed grating is used to reconstruct the grating. One or more parameters of the one-dimensional grating (such as line width and / or shape) or one or more parameters of the two-dimensional grating (such as column or through hole width or length or shape) may be input into the reconstruction process performed by the processor PU based on knowledge of the printing step and / or other inspection processes.

[0110] In addition to measuring parameters by reconstruction, angle-resolved scatterometry can also be used to measure the asymmetry of features in a product and / or resist pattern. One particular application of asymmetry measurement is for overlay measurements, where the target 30 includes a set of periodic features that are superimposed on each other. For example, the use of Figure 3 or Figure 4The concept of asymmetry measurement by an instrument of is incorporated herein in its entirety. In short, while the positions of the diffraction orders in the diffraction spectrum of a target are determined solely by the periodicity of the target, the asymmetry in the diffraction spectrum represents the asymmetry in the individual features that make up the target. Figure 4 In an instrument of this type (where the detector 18 may be an image sensor), this asymmetry of the diffraction orders appears directly as an asymmetry in the pupil image recorded by the detector 18. The asymmetry may be measured by digital image processing in the unit PU and may be calibrated against known values ​​of the overlay.

[0111] Figure 5 A plan view of a typical target 30 is shown, and Figure 4 In order to obtain a diffraction spectrum that is not disturbed by surrounding structures, in one embodiment, the target 30 is a periodic structure (e.g., a grating) that is larger than the width (e.g., diameter) of the illumination spot S. The width of the spot S may be smaller than the width and length of the target. In other words, the target is "not filled" with illumination, and the diffraction signal is substantially unaffected by any signal from product features, etc., external to the target itself. The illumination arrangement 2, 12, 13, 17 may be configured to provide illumination of uniform intensity across the back focal plane of the objective lens 15. Alternatively, the illumination may be limited to an on-axis direction or an off-axis direction, for example by including holes in the illumination path.

[0112] Figure 6 An example process of determining values ​​of one or more variables of interest for a target pattern 30' based on measurement data obtained using metrology is schematically depicted.The radiation detected by the detector 18 provides a measured radiation distribution 108 for the target 30'.

[0113] For a given target 30', a radiation distribution 208 may be calculated / simulated from a parameterized model 206 using, for example, a numerical Maxwell solver 210. The parameterized model 206 shows example layers of various materials that make up and are associated with the target. The parameterized model 206 may include one or more variables of features and layers of a portion of the target under consideration that may be varied and derived. Figure 6As shown, the one or more variables may include the thickness t of one or more layers, the width w (e.g., CD) of one or more features, the height h of one or more features, and / or the sidewall angle α of one or more features. Although not shown, the one or more variables may also include, but are not limited to, the refractive index (e.g., real or complex refractive index, refractive index tensor, etc.) of one or more layers, the extinction coefficient of one or more layers, the absorption of one or more layers, the resist loss during development, the footing of one or more features, and / or the line edge roughness of one or more features. The initial value of the variable may be the expected value of the object being measured. The measured radiation distribution 108 is then compared to the calculated radiation distribution 208 at 212 to determine the difference between the two. If there is a difference, the value of one or more variables of the parameterized model 206 may be changed, a new calculated radiation distribution 208 is calculated, and the new calculated radiation distribution 208 is compared to the measured radiation distribution 108 until there is a sufficient match between the measured radiation distribution 108 and the calculated radiation distribution 208. At this point, the values ​​of the variables of the parameterized model 206 provide a good or optimal match to the geometry of the actual target 30'. In one embodiment, a sufficient match exists when the difference between the measured radiation distribution 108 and the calculated radiation distribution 208 is within a tolerance threshold.

[0114] The variations of the patterning process are referred to as “process variations.” The patterning process may include processes both upstream and downstream of the actual transfer of the pattern in the lithographic apparatus. Figure 7Example categories of process variables 370 are shown. The first category may be variables 310 of the lithographic apparatus or any other equipment used in the lithographic process. Examples of this category include variables of the illumination, projection system, substrate stage, etc. of the lithographic apparatus. The second category may be variables 320 of one or more processes performed during the patterning process. Examples of this category include focus control or focus measurement, dose control or dose measurement, bandwidth, exposure duration, development temperature, chemical composition used in development, etc. The third category may be variables 330 of the design layout and its implementation in or when using the patterning device. Examples of this category may include the shape and / or location of assist features, adjustments applied by resolution enhancement technology (RET), CD of mask features, etc. The fourth category may be variables 340 of the substrate. Examples include characteristics of structures below the resist layer, chemical composition and / or physical dimensions of the resist layer, etc. The fifth category may be characteristics of time variations of one or more variables of the patterning process 350. Examples of this category include characteristics of high-frequency stage movement (e.g., frequency, amplitude, etc.), high-frequency laser bandwidth changes (e.g., frequency, amplitude, etc.), and / or high-frequency laser wavelength changes. These high frequency changes or movements are those that are above the response time of the mechanism that adjusts the underlying variables (e.g., stage position, laser intensity). A sixth category may be characteristics 360 of processes upstream or downstream of pattern transfer in the lithographic apparatus, such as spin coating, post-exposure bake (PEB), development, etching, deposition, doping, and / or packaging.

[0115] It will be appreciated that many, if not all, of the variables will have an impact on the parameters of the patterning process and the parameters of general interest. Non-limiting examples of parameters of the patterning process may include critical dimension (CD), critical dimension uniformity (CDU), focus, overlay, edge position or placement, sidewall angle, pattern shift, etc. Often, these parameters represent errors from nominal values ​​(e.g., design values, averages, etc.). Parameter values ​​may be values ​​of a characteristic of a single pattern or statistical values ​​of a characteristic of a group of patterns (e.g., averages, variances, etc.).

[0116] The values ​​of some or all of the process variables or parameters associated therewith may be determined by appropriate methods. For example, the values ​​may be determined from data obtained using various metrology tools (e.g., substrate metrology tools). The values ​​may be obtained from various sensors or systems of equipment in the patterning process (e.g., sensors of the lithographic equipment, such as level sensors or alignment sensors, control systems of the lithographic equipment (e.g., substrate or pattern forming device stage control systems), sensors in coating and developing system tools, etc.). The values ​​may come from an operator of the patterning process.

[0117] exist Figure 8An exemplary flow chart for modeling and / or simulating portions of a patterning process is illustrated in . It will be appreciated that the models may represent different patterning processes and need not include all of the models described below. Source model 1200 represents the optical properties of the illumination of the patterning device (including the radiation intensity distribution, bandwidth, and / or phase distribution). Source model 1200 may represent the optical properties of the illumination, including but not limited to numerical aperture settings, illumination sigma (σ) settings, and any particular illumination shape (e.g., off-axis radiation shapes such as toroidal, quadrupole, dipole, etc.), where (or sigma) is the outer radial extent of the illuminator.

[0118] The projection optics model 1210 represents the optical properties of the projection optics (including changes in the radiation intensity distribution and / or phase distribution caused by the projection optics). The projection optics model 1210 may represent the optical properties of the projection optics, including aberrations, deformations, one or more refractive indices, one or more physical sizes, one or more physical dimensions, etc.

[0119] The patterning device / design layout model module 1220 captures how the design features are arranged in the pattern of the patterning device, and the patterning device / design layout model module 1220 may include a representation of the detailed physical properties of the patterning device, for example, as described in U.S. Pat. No. 7,587,704, the entire contents of which are incorporated herein by reference. In one embodiment, the patterning device / design layout model module 1220 represents the optical properties (including the changes in radiation intensity distribution and / or phase distribution caused by a given design layout) of a design layout (e.g., a device design layout corresponding to features of an integrated circuit, memory, electronic device, etc.), which is a representation of the arrangement of features on or formed by the patterning device. Since the patterning device used in the lithographic projection apparatus can be changed, it is desirable to separate the optical properties of the patterning device from the optical properties of the rest of the lithographic projection apparatus, including at least the illumination and projection optics. The purpose of the simulation is often to accurately predict (for example) edge placement or positioning and CD, which can then be compared to the device design. The device design is typically defined as a pre-OPC patterning device layout, and the device design will be provided in a standardized digital file format such as GDSII or OASIS.

[0120] An aerial image 1230 may be simulated based on the source model 1200, the projection optics model 1210, and the patterning device / design layout model 1220. The aerial image (AI) is the radiation intensity distribution at the substrate level. The optical properties of the lithographic projection apparatus (e.g., the properties of the illumination, patterning device, and projection optics) determine the aerial image.

[0121] The resist layer on the substrate is exposed to the aerial image, and the aerial image is transferred to the resist layer as a latent "resist image" (RI) therein. The resist image (RI) can be defined as the spatial distribution of the solubility of the resist in the resist layer. The resist model 1240 can be used to simulate the resist image 1250 according to the aerial image 1230. The resist model can be used to calculate the resist image from the aerial image, and an example of calculating the resist image from the aerial image can be found in U.S. Patent Application Publication US2009-0157360, the entire content of which is incorporated herein by reference. The resist model typically describes the effects of chemical processes occurring during resist exposure, post-exposure baking (PEB), and development in order to predict (for example) the profile of the resist features formed on the substrate, so the resist model is typically only related to such properties of the resist layer (for example, the effects of chemical processes occurring during exposure, post-exposure baking, and development). In one embodiment, the optical properties of the resist layer, such as refractive index, film thickness, propagation and polarization effects, may be captured as part of the projection optics model 1210 .

[0122] Therefore, in general, the connection between the optical model and the resist model is the simulated spatial image intensity within the resist layer, which is caused by the projection of radiation onto the substrate, the refraction at the resist interface, and the multiple reflections in the resist film stack. By absorbing the incident energy, the radiation intensity distribution (spatial image intensity) becomes a latent "resist image", which is further modified by diffusion processes and various loading effects. For full-chip applications, efficient simulation methods that are fast enough can approximate the actual 3D intensity distribution in the resist stack through a 2D spatial (and resist) image.

[0123] In one embodiment, the resist image may be used as input to the post-pattern transfer process model module 1260. The post-pattern transfer process model 1260 defines the performance of one or more post-resist development processes (eg, etching, developing, etc.).

[0124] The simulation of the patterning process can, for example, predict the profile, CD, edge positioning (e.g., edge positioning error), etc. in the resist and / or etched image. Thus, the purpose of the simulation is to accurately predict, for example, edge positioning and / or aerial image intensity slope and / or CD, etc. of the printed pattern. These values ​​can be compared to the expected design to, for example, correct the patterning process, identify locations where defects are expected to occur, etc. The expected design is typically defined as a pre-OPC design layout, which can be provided in a standardized digital file format (such as GDSII or OASIS) or other file format.

[0125] Thus, the model formula describes most, if not all, of the known physical and chemical properties of the overall process, and each of the model parameters desirably corresponds to a different physical or chemical effect. Thus, the model formula sets an upper limit on how well the model can be used to simulate the overall manufacturing process.

[0126] exist Fig. 9 An exemplary flow chart for modeling and / or simulating a measurement process is illustrated in FIG. It should be understood that the following models may represent different measurement processes and need not include all of the models described below (e.g., certain models may be combined). The source model 1300 represents the optical characteristics of the illumination of the measurement target (including radiation intensity distribution, radiation wavelength, polarization, etc.). The source model 1300 may represent the optical characteristics of the illumination, including but not limited to wavelength, polarization, illumination sigma (σ) settings (where (or sigma) is the radial range of the illumination in the illuminator), any specific illumination shape (e.g., off-axis radiation shapes such as annular, quadrupole, dipole, etc.), etc.

[0127] The metrology optics model 1310 represents the optical properties of the metrology optics (including changes in the radiation intensity distribution and / or phase distribution caused by the metrology optics). The metrology optics 1310 may represent the optical properties of the illumination of the metrology target by the metrology optics and the optical properties of the transmission of the redirected radiation from the metrology target to the detector of the metrology device. The metrology optics model may represent various properties related to the illumination of the target and the transfer of the redirected radiation from the metrology target to the detector, including aberrations, deformations, one or more refractive indices, one or more physical sizes, one or more physical dimensions, etc.

[0128] The measurement target model 1320 can represent the optical properties of the illumination redirected by the measurement target (including changes in the intensity distribution and / or phase distribution of the illumination radiation caused by the measurement target). Therefore, the measurement target model 1320 can model the conversion of the illumination radiation to the redirected radiation performed by the measurement target. Therefore, the measurement target model can simulate the resulting illumination distribution of the redirected radiation from the measurement target. The measurement target model can represent various properties related to the target illumination and the redirected radiation generated from the measurement, including one or more refractive indices, one or more physical sizes of the measurement, the physical layout of the measurement target, etc. Since the measurement target used can be changed, it is desirable to separate the optical properties of the measurement target from the rest of the measurement equipment including at least the illumination and projection optics and the detector. The purpose of the simulation is usually to accurately predict, for example, intensity, phase, etc., which can then be used to derive parameters of interest of the patterning process, such as overlap, CD, focus, etc.

[0129] A pupil or aerial image 1330 can be simulated from the source model 1300, the metrology optics model 1310 and the metrology target model 1320. The pupil or aerial image is the radiation intensity distribution at the detector level. The optical properties of the metrology optics and the metrology target (e.g., the properties of the illumination, the metrology target and the metrology optics) determine the pupil or aerial image.

[0130] The detector of the measurement device is exposed to the pupil or the aerial image and detects one or more optical properties of the pupil or the aerial image (e.g., intensity, phase, etc.). The detection model module 1320 represents how radiation from the measurement optics is detected by the detector of the measurement device. The detection model can describe how the detector detects the pupil or the aerial image and can include signal-to-noise ratio, sensitivity to incident radiation on the detector, etc. Therefore, in general, the connection between the measurement optics model and the detector model is the simulated pupil or aerial image, which is caused by the illumination of the measurement target by the optics, the redirection of the radiation by the target, and the transmission of the redirected radiation to the detector. By absorbing the incident energy on the detector, the radiation distribution (pupil or aerial image) becomes a detection signal.

[0131] The simulation of the metrology process can, for example, predict the spatial intensity signal at the detector, the spatial phase signal, etc. or other calculated values ​​from the detection system, such as overlay, CD, etc., based on the detection of the detector through the pupil or the spatial image. Therefore, the purpose of the simulation is, for example, to accurately predict the detector signal corresponding to the metrology target or derived values ​​such as overlay, CD, etc. These values ​​can be compared with the expected design values ​​to, for example, correct the patterning process, identify the location where defects are predicted to occur, etc.

[0132] Thus, the model formula describes most, if not all, of the known physical and chemical properties of the entire measurement process, and each of the model parameters desirably corresponds to a different physical or chemical effect in the measurement process.

[0133] As described above, simulation of the patterning process can be used to predict parameters in the resist and / or etched image, such as predicting contour, CD, edge location (e.g., edge location error), etc. However, since the contour is considered to be a good representation of the relevant object boundary, such predictions as edge location error (EPE) are based on high-accuracy contour extraction. Typically, contour techniques rely on careful tuning of the parameters of the contour algorithm, where the optimal parameters are usually image-dependent, resulting in image-to-image variance of the generated contour. In other words, for each image acquired during the patterning process, even if the features may be similar to previously acquired images, the contour algorithm needs to be tuned. Contour techniques also rely on denoising of the image to smooth the image gradient, and although there are techniques that do not blur the image spatially (e.g., Chan-Vese level sets), these methods also rely on carefully selected parameters of the denoising algorithm. Therefore, contour techniques rely on manual and temporary signal processing, which may not be robust and / or cannot be reused without tuning the newly acquired image.

[0134] The present disclosure includes methods using machine learning that can overcome several problems of contour techniques, such as those described above. In one embodiment, a method based on semantically segmenting an image of at least a portion of a patterned substrate rather than finding contours (which can be derived from the segmentation if necessary) is discussed. In addition, the present disclosure includes a method for training a machine learning model (specifically, a convolutional neural network) to identify a deep learning model that performs semantic segmentation based on an original input image (i.e., without performing denoising). Although in one embodiment, the input image provided can be preprocessed to remove noise data.

[0135] Methods employing machine learning techniques can be used to improve the prediction of process parameters based on the identification of patterns and features present in the acquired data (e.g., SEM images, measurement results, and contextual data). In addition, machine learning techniques can be used to guide users in selecting the patterning process data (e.g., process window, dose, focus, CD, etc.) that are most useful for controlling the patterning process to improve the yield and accuracy of the patterning process.

[0136] Both unsupervised machine learning models and supervised machine learning models can be used to identify features. Without limiting the scope of the present invention, the application of supervised machine learning algorithms is described below.

[0137] Supervised learning is the machine learning task of inferring a function from labeled training data. The training data consists of a set of training examples. In supervised learning, each example is a pair with an input object (e.g., a pixelated input image typically represented in the form of a tensor or vector) and a desired output value (also called a supervisory signal). A supervised learning algorithm analyzes the training data and generates an inferred model that can be used to map new examples. The best solution will allow the algorithm to correctly determine the output for unseen instances. This requires the learning algorithm to generalize the training data to unseen situations in a "reasonable" way.

[0138] Exemplary models for supervised learning include decision trees, ensembles (bagging, boosting, random forests), k-NN, linear regression, naive Bayes, (deep) neural networks, logistic regression, support vector machines (SVM), and relevance vector machines (RVM).

[0139] Deep learning refers to training a neural network consisting of multiple layers and can be supervised or unsupervised. In one embodiment, the deep learning model can be based on architectures such as residual nets, convolutional neural networks, deep belief networks, and recursive neural networks. In a convolution-based architecture, each layer performs certain convolution operations based on one or more kernels (typically defined by weights associated with the kernels). During the training process, the deep learning model can be modified by adjusting the kernels (i.e., changing the kernel weight values).

[0140] A convolutional neural network typically consists of several layers linked together in a sequential manner so that information flows from input to output. In practice, each layer takes in a tensor T in And output the new tensor T out . The input tensor is convolved with the kernel tensor W, and the resulting convolution may be augmented with a bias vector and passed through an activation function such as a rectifier linear unit (ReLU). In the present disclosure, one or more dilated kernels may be used to increase the receptive field without performing max pooling (as this may reduce spatial resolution). A dilated kernel is a kernel used with a dilation operation. Typically, max pooling is an operation that uses the maximum value of each cluster of neurons at the previous layer.

[0141] The deep learning model needs to be trained, and the object (or feature within the image) labels may be domain specific. Thus, in one embodiment, a training data set is generated or obtained based on some ground truth. Such ground truth may include a set of original input images and output images, where each pixel is assigned an object label. By using data augmentation (e.g., taking random crops from the ground truth and performing translation, rotation, scaling, etc.), a training set with sufficient variation and volume may be generated.

[0142] The following disclosure describes a method for training a deep learning model to accurately perform semantic segmentation of an image (e.g., a grayscale SEM image). For example, the purpose of the present method is to assign each image pixel x to a label L(x) from a predetermined set of labels L via a convolutional neural network (deep learning), specifically a dilated residual neural network. In the present disclosure, the term label is used interchangeably with the type or feature of a substrate to generally refer to a specific feature of a substrate.

[0143] Fig.10 A flow chart of a method for training a deep learning model is illustrated. As previously described, the deep learning model includes a neural network having several layers linked together in a continuous manner. Each layer performs a convolution operation using a (dilated) kernel. The dilated kernel expands the receptive field of the input image and retains the spatial resolution of the input image, and therefore retains the features in the input image. This enables accurate annotation of even the smallest features of the substrate. In this way, the trained deep learning model is able to detect and accurately annotate features of small sizes (such as less than 20nm). In other words, in one embodiment, the model can segment images at a pixel level, which can be significantly less than 20nm.

[0144] The trained deep learning model can be used to perform, for example, semantic segmentation of SEM images obtained during a patterning process, and determine various characteristics and parameters of the patterning process by measuring features identified in segmented images of the SEM images.

[0145] The deep learning model (e.g., 1006) trains the neural network set using the training data (e.g., 1002) and the set of categories (e.g., 1004) so ​​that the trained deep learning model can predict different features of any input image (e.g., a portion of a patterned substrate or a target image) and annotate each feature in the input image to generate a segmented image that accurately mimics the input image (e.g., the match between the features of the target image and the segmented image (e.g., the average segmentation score) is greater than 99%). Training involves forward propagation of the input through different layers of the neural network to reach the output, where each layer performs a convolution operation with a dilated kernel. However, especially in the first forward calculation or forward pass, the output of the last layer may deviate from the ground truth. The deviation is characterized by a loss function (or cost function or objective function) to quantify the amount of deviation between the target image and the output image. For example, the features of the ground truth image may not match the features of the output image. To minimize the deviation, the gradient of the loss with respect to the weights is propagated backward through the different layers of the network (i.e., from the last layer to the first layer of the neural network). In this way, the local derivative of the loss function with respect to the weights of a particular layer (e.g., the output of the 15th layer of the 25th layer of the neural network) can be determined. Based on the derivative of the local function, the weights of the particular layer (e.g., the 15th layer) can be adjusted to minimize the loss function. In other words, the deviation in the output is minimized. During the training process, the deep learning model 1006 can generate several intermediate images at each hidden layer of the neural network, which are generally referred to as output tensors 1010 for discussion purposes. The training of the deep learning model 1006 can start with the initialization of the weights of each layer, and then gradually modify the weights based on the back propagation of the error signal with respect to the loss function.

[0146] In one embodiment, the training process starts at process P10. In process P10, a deep learning model 1006, training data 1002, and a category set 1004 (ie, labels) may be acquired and / or generated.

[0147] The training data 1002 includes an input image 1000 of at least a portion of a substrate having a plurality of features (e.g., Fig.11 ) and a real image 1001 corresponding to the input image 1000 (e.g., Fig.111130 in the real image). Typically, the input image set and the corresponding real image set with different types of features can be used as training data to train the deep learning model 1006. Since the deep learning model 1006 needs to be trained and the labels are domain-specific, a training dataset can be generated based on a certain basic fact (i.e., the real image 1001) (e.g., as described above) using data augmentation (e.g., randomly taking out crops from such basic facts and performing translation, rotation, scaling, etc.) to generate a training set with sufficient variation.

[0148] The input image 1000 may be a pixelated image, which may be represented in the form of a tensor or vector to implement various mathematical operations (e.g., convolution, addition, multiplication, rotation, scaling, etc.) and statistical analysis for each pixel of the input image 1000. In addition, based on the analysis, each pixel may be labeled (e.g., with one or more categories). In one embodiment, the input image 1000 may be a SEM image (i.e., a grayscale image) of at least a portion of a patterned substrate obtained during or after a patterning process. Each pixel of the SEM image may have a grayscale value, such as Fig.12 . In one embodiment, the input image 1000 can be represented as an input tensor with dimensions (b, w, h, c), where b represents the batch size (e.g., the total number of images used for training), w represents the width of the input image (or tensor), h represents the height of the input image (or tensor), and c represents the channel (i.e., the number of categories / labels / features). In one embodiment, the n-th order tensor X can be defined as: X=R^a1*a2*…*an, where ai is the size of the i-th order.

[0149] The ground truth image 1001 (or ground truth image set) is a segmented image of the input image 1000 used to establish the ground truth for the input image 1000 (or input image set), so that these images are used to jointly train the deep learning model 1006 (e.g., a convolutional neural network with an expansion kernel). In one embodiment, such semantic segmentation can be performed manually, wherein each pixel of the real image 1001 is manually assigned a label (i.e., a category or feature of the patterned substrate). In one embodiment, the label (i.e., category) can be, for example, such as a contact hole, a line, a grid bar, a grating, an oxide layer, a CD, etc. Similar to the input image 1000, the real image 1001 can be represented in the form of a tensor or a vector.

[0150] In addition, process P10 can obtain a set of categories 1004, each category corresponding to a feature of a plurality of features of a substrate within an input image 1000 of the substrate. Each category can be a label, which can be a name of a feature, a color, or other appropriate identifier. For example, the set of categories 1004 can be contact holes, lines, bars, gratings, oxide layers, CDs, or other features that may appear on a patterned substrate. During the training process, one or more labels from a label set can be assigned to each pixel of the predicted image. In addition, each label can be associated with a probability distribution that provides a measure of confidence in assigning the label to a feature of, for example, a substrate. In addition, each label can be associated with a color to generate a segmented image with a label represented by a color (e.g., purple for gratings, black for oxide layers, yellow for power rails, green for bars, etc.). Such color coding enables easy identification of different features of the substrate, performing measurements such as CD, area, overlap, EPE, line edge roughness, etc.

[0151] The deep learning model 1006 can be a neural network configured to receive training data and perform convolution using a dilated kernel (e.g., through modeling and / or simulation). The neural network includes multiple layers linked together in a continuous manner. A layer in the multiple layers can include multiple interconnected nodes, and the interconnected nodes contain an "activation function". The activation function maps elements of an input tensor (e.g., values ​​associated with pixels of an input image) to an output tensor nonlinearly. For example, the activation function can be a nonlinear function or a linear function, such as a sigmoid (see Equation 1 below) or a rectified linear unit (ReLU).

[0152]

[0153] In the above equation (1), (i) a is the activation function, which may be a sigmoid function, and (ii) w i is the weight of the ith element of the input tensor, (iii) x iis the i-th element of the input tensor, (iv) θ is the bias. The set of all weights and biases can also be referred to as model parameters. Such an activation function is typically a nonlinear function that enables the deep learning model 1006 to generalize or adapt to various data (e.g., different pixels of the input tensor). Nonlinear activation functions allow the definition of an enriched model category that contains models that are effective for tasks based on, for example, computer vision. During the training process, the weights can be initialized to random values ​​(e.g., based on a Gaussian distribution between 0 and 1), and during the training process, these weights can be optimized via back propagation and a specific update rule (e.g., Adam update rule). Optimization using loss functions and back propagation will be discussed later in this disclosure. The weights can be updated to accurately identify specific features of the input image 1000, so that the trained deep learning model can reproduce the features of the input image 1000. The present disclosure is not limited to the type of activation function. For example, alternatively or additionally, a Softmax function can be used. In one embodiment, softmax is used to map activations to probability distributions such as activations, and is therefore used in the final layer that can interpret these values.

[0154] Each of the multiple layers of the neural network of the deep learning model 1006 can receive an input image 1000 in the form of a tensor (or vector) and output a new tensor (also referred to as a predicted image obtained at the final layer). The input tensor can be convolved with an expanded kernel tensor W, and the result of the convolution can be increased (or decreased) with a bias and passed through an activation function such as a sigmoid function or a rectifier linear unit (ReLU).

[0155] In addition, in process P12, the training process involves generating a predicted image by applying the deep learning model 1006 using the input image 1000 of the training data. In one example, the output tensor 1010 or the predicted image refers to the output tensor of the last layer of the deep learning model 1006. In another example, the output tensor 1010 can be the output tensor of the intermediate layer of the deep learning model 1006. The generation of the output tensor 1010 involves iteratively performing a convolution operation using at least one (expanded) kernel on each layer of the deep learning model 1006. The convolution operation includes overlapping a portion of the input image 1000 with the expanded kernel, thereby obtaining the product between each overlapping element including the weight of the expanded kernel and the pixel value of the input image 1000, and summing the results of the product. For example, in Fig.14A, a 3×3 first dilated kernel 1410 overlaps with a pixelated input tensor 1400, and each element of the dilated kernel overlaps with a pixel of the input tensor. The product of the value of each overlapping element (e.g., the value of the pixel and the weight of the kernel) can then be calculated, and the product can be further summed to determine an element of the output tensor. In addition, the first dilated kernel can then be shifted by a stride (e.g., one column of one pixel to the right), and a similar convolution operation is performed on the new overlapping elements of the input tensor and the kernel elements (i.e., weights) to determine the second element of the output tensor.

[0156] In one embodiment, iteratively performing convolutions may involve performing multiple convolution operations using one or more dilated kernels within a particular layer on the input tensor of the layer. For example, a first layer may receive an input tensor (e.g., a representation of an input image, such as Fig.12 1200 in ). On the input tensor, the first layer performs a convolution operation using a first dilated kernel. The first dilated kernel has a first set of weights (e.g., Fig. 13B The result of the convolution is the first output tensor (e.g., Fig.11 1115a in the figure), the first output tensor is passed to the second layer through the first activation function of the first layer. Similar to the first layer, the second layer also performs a convolution operation using a second dilated kernel. The second dilated kernel has a second set of weights, which may be the same as or different from the first set of weights of the first kernel. The output of the second layer may be a second output tensor (e.g., Fig.11 1115e in FIG. 1115b). In one embodiment, the first core and the second core themselves may be different, which will result in different output tensors.

[0157] The present disclosure is not limited to a single convolution operation with one dilated kernel per layer. The training process may involve convolving the input tensor with multiple dilated kernels per layer of the deep learning model 1006 to obtain multiple output tensors. For example, Fig.11 In the example, the first layer can output four different output tensors 1115a, 1115b, 1115c, and 1115d, each of which is obtained by convolving with a different expansion kernel (such as a first expansion kernel with an expansion rate of 1 and a depth of 2c and a second expansion kernel with an expansion rate of 2 and a depth of 2c). Similarly, the second layer can output four different output tensors 1115e, 1115f, 1115g, and 1115h, each of which is obtained by convolving with a different corresponding expansion kernel (such as a third expansion kernel with an expansion rate of 2 and a depth of 4c and a fourth expansion kernel with an expansion rate of 4 and a depth of 4c).

[0158] In one embodiment, the dilation kernel may be characterized by a plurality of weights, a depth (i.e., classes), a dilation rate, or a combination thereof. In one embodiment, the dilation rate is a model / architecture-dependent parameter that one may consider to be a hyperparameter. In one embodiment, an "optimal" value for the dilation rate is determined empirically; i.e., different architectures are considered and the value that performs best with respect to the loss function model / architecture is selected. In a first pass, the weights may be randomly assigned (e.g., based on a Gaussian distribution between 0 and 1) and subsequently modified during the training process. The dilation rate may be an integer value, e.g., between 1 and 10, that is selected at the beginning of the training process and remains fixed during the training process. For example, Fig.13A The dilated kernel 1300 is illustrated in the form of a 3×3 grid, with each element of the grid having a specific weight (e.g., w1 to w9). The dilated kernel 1300 also has a depth D (e.g., the total number of features (e.g., the depth D is equal to the number of categories in the last layer)). Any dilated kernel (e.g., 1300) can also be represented in the form of a tensor, also called a weight tensor (or weight kernel). In another example, Figures 14A to 14C Examples of different dilation kernels are shown, for example, a first 3x3 dilation kernel 1410 with a dilation rate of 1 (in Fig.14A ), the second 3x3 expansion kernel 1420 with expansion rate 2 (in Fig. 14B ), and a third 3×3 dilation kernel 1430 with a dilation rate of 3 (in Fig. 14C middle).

[0159] In an embodiment, a dilation kernel may be configured or constrained for a portion of a patterning process. For example, the dilation kernel may be symmetric to characterize the projection system of a device for the patterning process. For example, in a 3×3 dilation kernel, the center element (i.e., (2,2)) may be considered the origin, and the origin may overlap with a particular pixel of the pixelated input image. In an embodiment, the kernel is constrained such that the kernel mimics effects such as aberrations. In an embodiment, as non-limiting examples, the constraints may be spherical kernel constraints or point symmetries, similar to Zernike.

[0160] In one embodiment, multiple dilation cores may be used, such that one or more dilation cores may be configured as part of a characterization patterning process. For example, a first dilation core in the multiple dilation cores may include a first set of weights, a first dilation rate, and a depth equal to the number of categories in the category set 1004. A second dilation core in the multiple dilation cores may include a second set of weights, a second dilation rate, and a depth equal to the number of categories in the category set 1004. In one example, the second set of weights is different from the first set of weights and / or the second dilation rate is different from the first dilation rate.

[0161] Furthermore, in one embodiment, a first dilation kernel may characterize a first portion of a patterning process (eg, a projection system) and a second kernel may characterize a second portion of the patterning process (eg, parameters related to characteristics of the substrate).

[0162] During the training process, weights of one or more dilated kernels may be iteratively modified and / or assigned. Iterative assignment of weights may be performed to improve prediction of one or more features of the output tensor 1010. Such assignment of weights is done during the training process, such as using back propagation based on local derivatives of a loss function (discussed later in this disclosure). In one embodiment, iteratively assigning weights may include assigning / adjusting a set of weights of the dilated kernel to account for, for example, aberrations of a projection system of a device of a patterning process. In one embodiment, iteratively assigning weights may include assigning / adjusting a set of weights of the dilated kernel to characterize an aspect of the patterning process.

[0163] In one embodiment, process P12 can output an output tensor 1010 based on a convolution operation using one or more dilated kernels, which can be represented in the form of one or more output tensors 1010. In one embodiment, output tensor 1010 can be the output of the last layer of deep learning model 1006, where the depth of the dilated kernel is selected so that the depth of the output tensor is equal to the total number of categories in the category set 1004. In other words, the output tensor will have an element corresponding to each pixel of the input image, where each element of the output tensor also has a depth (i.e., the total number of categories). For example, if there are 10 categories, each element of the output tensor will have 10 categories associated with it. In addition, a probability distribution can be assigned to each category, which indicates the probability that the pixel belongs to that particular category.

[0164] The predicted image 1010 or output tensor 1010 includes elements (i.e., pixels) corresponding to pixels of the input image, each element of the output tensor 1010 has a value associated with a category set 1004 (i.e., one or more features of the substrate), and a unique label (such as a contact hole, line, grid bar, etc.) can be assigned to each element in process P14. In one embodiment, the categories of the category set 1004 can be assigned to the elements of the output tensor based on the comparison of the output tensor with the basic fact image 1001 or the real tensor. In one embodiment, the elements of the output tensor 1010 include a probability set associated with each category of the category set 1004 (or a probability distribution of each category for each pixel). The probability value indicates the probability that a particular element of the output tensor belongs to a particular category. Based on the probability, a label can be assigned to each pixel, where the label corresponding to the category has the maximum probability value for the element. For example, the first element associated with 10 categories (e.g., line category, gate category, contact hole category, etc.) may have the following probability value represented by a vector, [0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.50.95]. In one embodiment, the vector may have normalized probability values ​​so that their sum is 1 to make it a distribution. The values ​​shown here may be the logarithmic values ​​of the last layer, and then softmax will produce a probability distribution. For example, for 3 categories, logits = [9, 12, 10] probs = [0.04, 0.84, 0.11] = softmax(logits), then the first element may be assigned a label associated with the tenth category (e.g., contact hole).

[0165] Typically, in the first pass, the prediction of the label may be inaccurate, and therefore, the output tensor 1010 may not be accurately segmented. In order to generate an accurately segmented image that reproduces the features of the corresponding input image 1000, several iterations (e.g., more than 1000 iterations, 50,000 iterations, 100,000 iterations, 1 million iterations, etc., depending on how quickly the model performance converges) may be performed based on the following process.

[0166] In process P16, it can be determined whether the segmented image matches or accurately reproduces the input image 1000. In response to a determination that the segmented image is accurately reproduced, the training process can end and a trained deep learning model 1050 can be obtained. The trained deep learning model 1050 can have multiple layers, where nodes have multiple weights assigned as described above to obtain the segmented image.

[0167] However, if the segmented image does not reproduce the input image 1000 and / or the training does not converge, in process P18, the training process may further involve training the deep learning model based on back propagation through the deep learning model using a loss function, iteratively assigning weights to at least one dilated kernel until the output tensor 1010 reproduces multiple features of the input image 1000. In one embodiment, the loss function is a function of the probabilities of the real image 1001 and the output tensor 1010. For example, the loss function can be based on the probability of the category assigned to the output tensor 1010. The loss function can be based on a cross entropy function, a mean square error function, or other appropriate error function. In an example, the following cross entropy function (Equation 2) can be used as the loss function.

[0168]

[0169] In the above formula, L represents the loss function, i is the i-th pixel in the pixelated image I, c is the category of the category set C, and y c (i) is a tensor of the i-th pixel of the ground-truth image 1001, where the elements of the tensor are one-hot encoded to indicate that the pixel belongs to a specific category c, p c (i) is the probability that the i-th pixel may belong to each class of the predicted image or output tensor.

[0170] In the example, category C can be represented as a feature vector such as [line, gate bar, contact hole, grating, oxide]. A pixel in the real image Y may belong to one of the categories C. For example, if the i-th pixel y c (i) If it belongs to the contact hole category, the element of the real image tensor corresponding to the pixel may be one-hot encoded as a contact hole (i.e., y c (i) = [0, 0, 1, 0, 0]). In addition, using the probability of the predicted image, the above equation (2) can be used to determine the loss function. As the training process continues, the value of the loss function is gradually reduced, preferably minimizing the value of the loss function. In one embodiment, the ground truth can also be probabilistic.

[0171] Backpropagation through a deep learning model can involve inverse mapping the output tensor of a layer (e.g., layer 25) to the input tensor of that particular layer or multiple previous layers. For example, the output tensor of a layer (e.g., layer 25) can be inverse mapped through multiple layers to, for example, the input tensor of layer 5. In another example, backpropagation can involve inverse mapping the output tensor of layer 25 to the input tensor of layer 25 (i.e., the output of layer 24), the output of layer 24 to the input tensor of layer 24 (i.e., the output layer of layer 23), and so on. At each step during backpropagation (i.e., at the output of a particular layer), a local derivative of the loss function can be calculated. The local derivative of the loss function can be further plotted as a graph or map (e.g., see Fig.15 ), graphically determine the direction to change the layer weights. For example, for illustration purposes, up to 2 dimensions may be drawn. Gradient descent changes the weights so that the loss is reduced. In one embodiment, one of different optimization methods, such as gradient descent, may be used to determine the optimal weights so that the output tensor 1010 accurately depicts the input image 1000.

[0172] Fig.15 A graph of local derivatives at a specific layer is illustrated, where local derivatives are taken with respect to the weights w1 and w2 of the dilated kernel. The weights can then be updated based on an optimization method, such as a gradient descent method, which results in more optimized weights for that specific layer, thereby minimizing the loss for the given specific layer.

[0173] In one embodiment, a loss function may be defined such that the error in identifying a particular feature is reduced, preferably minimized. For example, the loss function may be to weight a particular feature more than other features in an image (e.g., a contact hole), resulting in a greater emphasis on improving the segmentation of that portion. Note that different image features are represented by different classes, and therefore different class weights allow features to contribute to the loss function in a feature-dependent manner.

[0174] After modifying the weights of one or more layers, the training method can continue from process P12 as previously discussed with a different training example, in which the output tensor 1010 is re-determined, i.e., a different output tensor 1010 can be generated based on the modified weights. Then, process P14 as previously discussed can be performed to reassign the categories and generate the segmented image 1020. In addition, determinations can be made as previously discussed in process P16, and process P18 can then be performed. After several iterations, the training process may converge, i.e., the loss does not decrease significantly on a valid set of images, at which point the training can be stopped and the deep learning model can be considered a trained deep learning model 1050.

[0175] Fig.11 An example of training a deep learning model such as a convolutional neural network using a training dataset including an input image 1100 and an output image (i.e., a segmented image) 1130 is illustrated. Fig.11 In the example, as in Fig.10 The training method discussed in can be applied to generate a trained deep learning model that can accurately identify each feature of the input image 1100. The training can involve determining the weights of the dilated kernel of each layer of the neural network so that the features of the input image 1100 can be identified.

[0176] Input image 1100 is a SEM image of a portion of a substrate having features, such as gratings 1104 (placed horizontally), oxide layer 1106 (placed horizontally between gratings), power rails 1102, a first set of vertical lines such as 1110 (the longest lines in the pattern), a second set of vertical lines such as 1114 (relatively shorter than 1110, but longer than the minimum line length), etc. In an embodiment, the first set of lines may be characterized according to various geometric measurements (such as size and orientation), for example, longitudinal length, thickness or width, number of crossed gratings, relative position relative to adjacent other features, or other suitable geometric properties of the features.

[0177] The deep learning model 1006 having multiple layers receives the input image 1100 and the segmented image 1130 (the segmented image 1130 is an example of the training data 1002). During the training process, forward and backward propagation (or propagation) can be performed as described above to match the input image 1100 and the segmented image 1130, so that the trained deep learning model can predict the segmented image from any input image, preferably with similar features or feature subsets used during the training process.

[0178] In a forward calculation (e.g., as discussed in process P12), the deep learning model can generate multiple predicted images represented as output tensors (collectively referred to as output tensors 1115). For example, the first layer can generate a first output tensor 1115a based on a first dilation kernel with a dilation rate of 1 and a depth of 2c, and generate a second output tensor 1115b based on a second dilation kernel with a dilation rate of 2 and a depth of 2c. Similarly, a third output tensor 1115c and a fourth output tensor 1115d are generated using dilation kernels based on a third dilation kernel and a fourth dilation kernel with a dilation rate of 2 and a depth of 2c, respectively. In one embodiment, output tensors 1115b to 1115d can be generated by different layers. Similarly, multiple output tensors 1115e, 1115f, 1115g, and 1115h can be output by the second layer or other layers of the deep learning model.

[0179] In the back propagation, the aforementioned dilated kernels may be modified based on the loss function, for example, as previously discussed in processes P12 and P18. In one embodiment, different weights may be assigned to each dilated kernel based on the features to be identified. For example, the first kernel may be characterized as identification feature 1102, the second kernel may be characterized as identification feature 1104, and so on.

[0180] Ultimately, after several iterations of forward and backward propagation, an output tensor with a specific class assigned to each element of the output tensor can be generated to ultimately produce a segmented image with labels, such as a first label 1132 corresponding to the class power rail (i.e., 1102 of the input image), a second label 1134 corresponding to the grating (i.e., 1104 of the input image), a third label 1136 corresponding to the oxide (i.e., 1106 of the input image), a fourth label 1140 corresponding to the line (i.e., 1110 of the input image), a fifth label 1142 corresponding to the line (i.e., 1112 of the input image), and a sixth label 1144 corresponding to the line (i.e., 1114 of the input image), and so on.

[0181] In one example, reference Figures 16A to 16D , the trained deep learning model can receive an input image 1600 and generate a segmented image 1630 corresponding to the input image 1600. Although the segmented image 1630 illustrates a unique label, each pixel of the segmented image 1630 also includes additional information including a probability distribution of a particular class. Therefore, based on the probability distribution, several probability-related metrics can be determined, such as uncertainty and / or error in identifying a particular label. For example, using the probability distribution of each pixel, an uncertainty map or graph 1640 can be generated. In another example, the probability can be used to generate an error map or graph 1650. Based on the amount of uncertainty in the map 1640, sub-pixel accuracy can be derived, for example, via interpolation.

[0182] In one embodiment, the segmented images can also be used to determine parameters related to features generated during the patterning process (e.g., CD, EPE, LER, necking, overlap, etc.) Based on these parameters, defects in the patterning process (e.g., EPE) can be determined.

[0183] Fig.17A A method of applying a trained deep learning model to determine parameters of features of a patterning process is illustrated. An input image of at least a portion of a substrate can be passed through a trained deep learning model to generate a segmented image. The segmented image clearly depicts features of the substrate that can be measured to determine parameters (e.g., CD) of the substrate. In an embodiment, the parameters can be used to determine adjustments to the patterning process. A patterning process based on a segmented image can produce more accurate results than a contour-based approach.

[0184] For example, segmented images generated from a trained deep learning model can identify or distinguish features, such as charge effects of light spots surrounding features, compared to unsegmented images or grayscale images (e.g., SEM images). Fig.18A), features with similar materials, etc. On the other hand, using an SEM image such as input image 1100 to determine parameters related to features without being segmented may produce inaccurate results. For example, in input image 1100, for example, when overlapping features are made of similar materials, it may not be clear where a particular feature ends and another begins. In this case, features such as feature 1114 and oxide layer 1106 may be indistinguishable based on the SEM image. Feature 1114 extends above grating 1104 and overlaps with oxide layer 1106 made of similar oxide-type material at the tip (top) of feature 1114. Measurements based on such images will obviously lead to inaccurate measurement results, and therefore defects may be introduced during the patterning process. In this way, compared with semantic segmentation based on grayscale images, semantic segmentation according to the present method can more accurately improve the recognition of various features. In general, the term semantic segmentation refers to the process of assigning labels or categories to each feature of any image, and the image resulting from the process is called a segmented image.

[0185] In process P170, a trained deep learning model 1050 (e.g., as shown in FIG. 1 ) having a set of categories 1704 (an example or subset of the set of categories 1004) used to train the model 1050 can be obtained. Fig.10 ). Each category corresponds to a feature of a training substrate (e.g., input image 1100 and segmented image 1130) used to train the trained deep learning model. In addition, an input image 1700 (e.g., a SEM image) of at least a portion of a substrate undergoing a patterning process having a plurality of features may be obtained. Input image 1700 is an image that can be segmented using trained deep learning model 1050. Input image 1700 may be any image different from that from which trained model 1050 is obtained.

[0186] In process P172, the method may involve generating a segmented image 1730 by modeling and / or simulating the trained deep learning model 1050 using the input image 1700 and the category set 1704, wherein a plurality of features of the input image are reproduced in the segmented image 1700. In process P172, because the weights of the dilation kernel have been assigned in the trained model 1050, the trained model 1050 may not be modified, and the segmented image 1730 may be generated in a single pass. Executing such a trained model P172 may produce accurate results. In addition, the probability associated with the label (i.e., category set) that can be further used to determine the confidence of the measurement result of the feature in the segmented image 1730 can be used to derive uncertainty and error data (e.g., error data that can be obtained when the training data includes ground truth). Fig. 17BAn exemplary segmented image 1730 is illustrated in FIG.

[0187] In one embodiment, generating the segmented image 1730 may include assigning a label (i.e., a category of the category set 1704) to each pixel of the segmented image 1730. Each pixel of the segmented image 1730 is associated with information including: the category set, a probability set corresponding to each category in the category set, and a location on the segmented image. The label may correspond to a feature in a feature set that has, for example, a maximum probability of the probability associated with the category set. In one embodiment, as Fig. 17B As shown, segmented image 1730 may include features such as lines 1710, 1714, 1716, etc. In an embodiment, a label may represent each feature, and each label may be color coded to clearly distinguish one feature from another.

[0188] In process P174, based on the segmented image 1730, parameters of the patterning process can be determined. The determination of the parameters of the patterning process can include performing measurements on multiple features within the segmented image 1730. The measurements can include measurements of feature sizes, measurements of distances between one or more features in the multiple features, or measurements of other geometric shapes that can be performed. The measurements can be automatically performed based on the location of each pixel. For example, the process can be programmed to identify the first pixel of a feature and the last pixel of the same feature along a horizontal line, and the distance between pixels can be derived based on the location information associated with each pixel. In addition, the uncertainty in the measurement of the distance can be calculated based on the probability associated with each pixel. Therefore, a measure of the confidence in a particular measurement is provided.

[0189] In addition, more complex measurements can be determined, such as curvature at the ends of a feature (e.g., point P11 of feature 1710), dimensional variations along the length of a feature, the area of ​​irregular features (e.g., dust particles), etc., which can be further supplemented by uncertainties and / or errors in measurements based on probabilities associated with individual pixels.

[0190] In addition, based on the measurement results, parameter values ​​of multiple features can be determined. The parameter can be edge positioning error, overlap, line edge roughness, necking and / or CD related to multiple features of the segmented image 1730. In an example, the parameter can be EPE between different features, minimum / maximum CD values ​​along the length (or width) of the feature, line edge roughness (LER) along the length of the feature, the total area of ​​the feature, or other suitable parameters that can be derived to determine the quality or defects in the substrate of the patterning process.

[0191] In the example, Fig. 17BAs shown, the EPE of feature 1710 can be determined. In one embodiment, the EPE can be an offset or distance between the edge of a feature relative to another feature. For example, the EPE of feature 1710 can be defined as the distance d1 measured from the edge of the grating (e.g., at P10) to the end of feature 1710. If the distance d1 is less than a predetermined threshold, there is an EPE between feature 1710 and grating 1704. Such a predetermined threshold can be based on experience or past patterning data that can be stored in a database. Distance d1 can indicate how much feature 1710 should extend to prevent defects such as short circuits. For example, feature 1710 can be designed to block a portion of the grating, and conductive material can be deposited in a portion of the grating during the patterning process. If the distance d1 is less than a predetermined threshold, or the edge at P1 does not completely block an area in the grating (e.g., leaving an open area), the conductive material located on the left and right sides of the feature 1710 can be bridged to create an electrical short circuit path. This short circuit is undesirable and may make the chip produced by this substrate defective, which may need to be discarded. Accordingly, it may be desirable to monitor such EPE to discard the substrate earlier in the patterning process to save time and processing costs, thereby effectively increasing the yield of the patterning process. In one embodiment, because the patterning process is a highly repetitive process in which similar operations (e.g., pattern transfer, etching, and deposition) are repeated on the same substrate or different substrates, the patterning process can be adjusted based on feature-related parameters (e.g., process variables such as dose or focus) to prevent similar defects from recurring on other substrates.

[0192] Similarly, in another example, a measurement may be performed on feature 1716, such as the width of a line along a length. In the case where the width is less than a certain threshold, then necking may be detected. For example, since width d2 (e.g., 10 nm) may be less than 50% of a selected width of feature 1716 (e.g., 20 nm), necking may be detected at the location of feature 1716.

[0193] In one embodiment, in process P176, based on the parameters of the patterning process, adjustments to the patterning process can be determined. The adjustments can be directed to process variables (e.g., dose, focus, scanner level, etc.) of the patterning process to prevent errors or defects in subsequent processing of the same substrate or a different substrate. The adjustments can be determined using simulation and / or modeling of the patterning process, such as the process models previously discussed in the present disclosure. In addition, in process P178, the adjustments can be applied to adjust the patterning process. In other words, patterning on the substrate can be performed by applying the adjustments to one or more devices of the patterning process.

[0194] The semantic segmentation process is useful for identifying features that suffer from charging effects that are often observed in HMI or SEM images. The charging effect can be seen as light spots around the features. This charging effect hinders the analysis of features on the substrate, and efforts are made to model this charging effect to account for this effect. Semantic segmentation can be used to analyze such images.

[0195] Fig.18A An image including a charge effect (i.e., light spot) around a feature (i.e., a contact hole) is illustrated. The input image 1800 includes three features (or categories): a contact hole 1802, a light spot 1804, and a background 1806. The contact hole 1802 is surrounded by a light spot 1804 (depicted as a white ring around a dark area (i.e., the contact hole)). A deep learning model can be trained to perform semantic segmentation of such a light spot (e.g., 1814), the contact hole 1812, and the background 1816. In one embodiment, a deep learning model can be trained to generate a segmented image 1810, in which the light spot 1814 may be lumped together with the background 1816 (corresponding to the background 1806) and interpreted as noise rather than part of the contact hole 1812. Therefore, the trained deep learning model may include weights to interpret the light spot as background, so that the contact hole can be clearly segmented. In one embodiment, a label can be assigned to the light spot to further explore the geometric or optical properties of the light spot relative to the feature size, location, etc.

[0196] Figures 19A to 19E The advantage of the semantic segmentation method over the charge effect is further illustrated. The cropped image 1900 is a crop of a portion of the input image 1800 discussed previously. The cropped image 1900 illustrates a feature 1902 (i.e., a contact hole), a light spot 1904 surrounding the contact hole, and a background 1906. After segmentation, a segmented image 1910 (in Fig.19B ), where a first label (or category / feature) 1912 corresponds to 1902, a second label (or category / feature) 1914 corresponds to the light spot 1902, and a third label 1916 (or category / feature) corresponds to the background 1906. In addition, a portion of the cropped image 1900 is Fig.19C , and the corresponding portion of the segmented image 1910 is enlarged in Fig.19D The cropped image 1920 does not clearly distinguish the three features 1902, 1904, and 1906. On the other hand, the segmented image 1930 more clearly distinguishes such features. In addition, each pixel within the segmented image 1930 is associated with a probability that the particular pixel belongs to a particular class. For example, in Fig.19E The probability distribution of pixels from A to B along line 1932 is illustrated in FIG. Fig.19E, at point A, the probability that the pixel belongs to the first label 1912 is the highest compared to the other labels 1914 and 1916. As the pixel moves away from point A toward point B, the probability that a particular pixel should be assigned the first label 1912 decreases. Similarly, since the first label is smaller than the second label 1914, the segmented image 1932 shows the second label 1914 instead of the first label 1912. In addition, as the pixel approaches point B, the probability of the third label 1916 is relatively higher than the other labels. Therefore, features on a substrate that are subject to charge effects can not only be better distinguished, but also can be measured using measurement results with a certain range of accuracy or confidence values.

[0197] In one embodiment, one or more deep learning models may be developed depending on the features or pattern layout of the substrate. For example, a first deep learning model may be trained to perform semantic segmentation of features, such as input image 1100 ( Fig.11 The training of the first deep learning model was discussed earlier in this disclosure.

[0198] In another example, a second deep learning model can be trained to perform tasks such as Fig. 20A and 20B Semantic segmentation of the features shown. Fig. 20A An input image 2000A having features 2002 and 2004 is illustrated. Fig. 20B 2000B in the segmented image 1100 illustrates features 2012 and 2014 corresponding to features 2002 and 2004. In an embodiment, features 2002 and 2004 may be different from the features in image 1100, and therefore, the first training model may not accurately label or identify the features. In other words, the amount of error and / or uncertainty in the label of unknown features may be relatively high compared to known features (i.e., features on which the first model may be trained).

[0199] Similarly, in another example, a third semantic model can be trained to recognize words such as Fig.21A and 21B A pattern or feature in a Fig.21A An input image 2100A is shown. Fig.21B Illustrated is a segmented image 2100B corresponding to input image 2100A. Input image 2100A includes several features having substantially different shapes and sizes than the features in images 1100 and 2000A.

[0200] Accordingly, in one embodiment, different models may be trained and applied, and then based on the uncertainty and error mapping of each model, a particular model may be used to identify certain types of features. For example, a first trained model may exhibit the lowest error when identifying features of image 1100, while the same first trained model may exhibit the highest error when identifying features of image 2100A. In this case, the first trained model may be selected to identify features similar to image 1100.

[0201] The first model, the second model, and the third model are discussed only as examples and do not limit the present disclosure. In one embodiment, a single model may be trained for all features of 1100, 2000A, and 2100A.

[0202] Semantic segmentation based on a trained deep learning model has multiple advantages. The trained deep learning model can identify features that the training model recognizes across different images without retraining or readjusting the parameters (e.g., weights, biases, or constants) of the model. In addition, when the trained model is applied to different images of a substrate or different substrates of a patterning process, no readjustment is required to identify or track the contour (i.e., shape) of the feature. In addition, the size of the feature on the substrate can be determined to derive parameters (e.g., CD, overlap, EPE, etc.) of the patterning process (as described above). Based on the parameters, one or more defects of the substrate can be identified, and the patterning process can be adjusted to improve the yield of the patterning process.

[0203] A trained deep learning model may not be re-adjusted (or retrained) to recognize similar feature types on different images; unlike contour determination techniques. Contour determination techniques typically require image-dependent re-adjustment of parameters (i.e., functions, variables, or constants) used in the contour determination algorithm, and may result in image-to-image variations. For example, a first image may include a particular feature (e.g., a contact hole), then a conventional contour determination algorithm may need to be re-adjusted for the same feature (e.g., a contact hole) on different images (e.g., SEM images acquired for different portions of the same substrate or different substrates). On the other hand, semantic segmentation based on a trained deep learning model (where the model is trained for a feature type (e.g., a contact hole)) may not be re-trained or re-adjusted to recognize the same feature type on different images. Since no re-adjustment is required, when a trained deep learning model is used during the patterning process compared to using a contour determination algorithm, the patterning process may be improved with less computational resources and time.

[0204] Fig. 221 is a block diagram illustrating a computer system 100 that may help implement the methods and processes 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 for processing 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 for storing information and instructions to be executed by the processor 104. The main memory 106 may 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 coupled to the bus 102 for storing 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.

[0205] The computer system 100 may be coupled to a display 112, such as a cathode ray tube (CRT) or a flat panel or touch panel display, via the bus 102 for displaying information to a computer user. An input device 114, including alphanumeric and other keys, is coupled to the bus 102 for communicating information and command selections to the 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 the processor 104 and controlling movement of a cursor on the display 112. The input device typically has two degrees of freedom in two axes, namely a first axis (e.g., x) and a second axis (e.g., y), which allows the device to specify a position in a plane. A touch panel (screen) display may also be used as an input device.

[0206] According to one embodiment, portions of the process 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, such as storage device 110. Execution of the sequences of instructions contained in main memory 106 causes processor 104 to perform the process steps described herein. One or more processors in a multi-processing arrangement may also be employed to execute the sequences of instructions contained in main memory 106. In an alternative embodiment, hard-wired circuitry may be used in place of or in combination with software instructions. Thus, the description herein is not limited to any specific combination of hardware circuitry and software.

[0207] As used herein, the term "computer-readable medium" refers to any medium that participates in providing instructions to processor 104 for execution. Such media can take many 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 wires that include 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, floppy 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 cartridges, and carrier waves described below or any other medium that can be read by a computer.

[0208] Various forms of computer readable media may be involved in carrying 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.

[0209] The computer system 100 also desirably includes 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 can be an integrated services digital network (ISDN) card or a modem to provide a data communication connection to a telephone line of a corresponding type. As another example, the communication interface 118 can be a local area network (LAN) card to provide a data communication connection to a compatible LAN. A wireless link can also be implemented. In any such implementation, the communication interface 118 sends and receives electrical signals, electromagnetic signals, or optical signals that carry digital data streams representing various types of information.

[0210] 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 to a host computer 124 through a local network 122 or to a data device 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 through the 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.

[0211] The computer system 100 can send messages and receive data (including program code) through one (or more) networks, network link 120, and communication interface 118. In the example of the Internet, server 130 can transmit the requested code of the application program through Internet 128, ISP 126, local network 122, and communication interface 118. For example, such a downloaded application can provide the illumination optimization of the embodiment. The received code can be executed by processor 104 when the code is received, and / or stored in storage device 110 or other non-volatile storage for later execution. In this way, computer system 100 can obtain application code in the form of a carrier wave.

[0212] The embodiments may be further described using the following aspects:

[0213] 1. A method for training a deep learning model for a patterning process, the method comprising:

[0214] Obtaining (i) training data, the training data comprising an input image of at least a portion of a substrate having a plurality of features and a real image corresponding to the input image; (ii) a set of categories, each category corresponding to a feature among the plurality of features of the substrate within the input image; and (iii) a deep learning model, the deep learning model being configured to receive the training data and the set of categories;

[0215] Generate a predicted image by modeling and / or simulating the deep learning model using the input image, wherein the deep learning model iteratively performs a convolution operation using at least one dilated kernel including a plurality of weights;

[0216] Assigning a category from the set of categories to a feature in the predicted image based on a match between the feature in the predicted image and a corresponding feature in the real image; and

[0217] A trained deep learning model is generated by modeling and / or simulation by iteratively assigning weights to at least one dilated kernel based on back propagation through the deep learning model using a loss function so that the predicted image reproduces multiple features of the input image.

[0218] 2. The method according to claim 1, wherein the convolution operation comprises

[0219] overlapping the input image with the at least one dilation kernel,

[0220] determining a product between each overlapping element comprising a weight of the at least one dilation kernel and a pixel value of the input image, and

[0221] The results of the multiplications are summed to determine the value of the convolution operation.

[0222] 3. A method according to any one of aspects 1 to 2, wherein the at least one dilation kernel comprises a plurality of weights, a dilation rate, and a depth, wherein the depth is a multiple of the number of categories in the category set.

[0223] 4. A method according to any one of aspects 1 to 3, wherein the loss function is a function of the probability of each class in the set of classes assigned to the true image and the predicted image.

[0224] 5. A method according to aspect 4, wherein the iteratively assigning weights includes: modifying multiple weights of the dilation kernel, the dilation rate and / or the depth to characterize at least a portion of the patterning process based on a local derivative of the loss function.

[0225] 6. The method according to any one of aspects 1 to 5, wherein the at least one dilated core comprises:

[0226] A first dilated kernel of the at least one dilated kernel comprises a first set of weights, a first dilation rate, and a depth equal to the number of categories in the set of categories;

[0227] A second dilated kernel of the at least one dilated kernel comprises a second set of weights, a second dilation rate, and a depth equal to the number of categories in the set of categories, wherein the second set of weights is different from the first set of weights and / or the second dilation rate is different from the first dilation rate.

[0228] 7. A method according to aspect 6, wherein the first expansion core characterizes a first part of the patterning process, and the second core characterizes a second part of the patterning process, wherein the second part is a projection system of the patterning process, and the second part is a parameter related to the characteristics of the substrate.

[0229] 8. A method according to any one of aspects 1 to 7, wherein the dilation kernel is symmetric to characterize a projection system of an apparatus for the patterning process.

[0230] 9. The method of clause 8, wherein the iteratively assigning weights comprises assigning a first set of weights to a plurality of weights of the dilated kernel to characterize a projection system of an apparatus of the patterning process.

[0231] 10. The method according to any one of aspects 1 to 9, wherein the expansion ratio is an integer value between 1 and 10.

[0232] 11. A method for applying a trained deep learning model to a patterning process, the method comprising:

[0233] Obtaining (i) a trained deep learning model; (ii) a set of categories, each category corresponding to a feature of a training substrate used to train the trained deep learning model; and (iii) an input image of at least a portion of a substrate undergoing a patterning process having a plurality of features;

[0234] generating a segmented image by modeling and / or simulating the trained deep learning model using the input image and the set of categories, wherein a plurality of features of the input image are reproduced in the segmented image; and

[0235] Parameters of the patterning process are determined based on a plurality of features of the segmented image.

[0236] 12. A method according to aspect 11, wherein generating a segmented image includes: assigning a category from the category set to each pixel of the segmented image, wherein the label corresponds to a feature from the feature set with the maximum probability, and the segmented image is pixelated.

[0237] 13. A method according to aspect 12, wherein each pixel of the segmented image is associated with information comprising: the category set, a probability set corresponding to each category of the category set, and a location on the segmented image.

[0238] 14. The method according to any one of aspects 11 to 13, wherein determining the parameters of the patterning process comprises

[0239] performing measurements on a plurality of features within the segmented image, the measurements comprising feature sizes and / or distances between one or more of the plurality of features; and

[0240] Parameters are determined based on measurements related to the plurality of features.

[0241] 15. A method according to any one of clauses 11 to 14, wherein the parameter is edge positioning error, overlay, line edge roughness, necking and / or CD.

[0242] 16. The method according to any one of aspects 11 to 15, further comprising:

[0243] determining adjustments to the patterning process based on parameters associated with the features; and

[0244] In response to the adjustment, the patterning process is adjusted.

[0245] 17. A method according to clause 16, wherein the adjustment comprises adjustment of a process variable including dose and / or focus.

[0246] The embodiments of the present disclosure may be implemented in hardware, firmware, software, or any combination thereof. The embodiments of the present disclosure may also be implemented as instructions stored on a machine-readable medium, which may be read and executed by one or more processors. A machine-readable medium may include any mechanism for storing or transmitting information in a form readable by a machine (e.g., a computing device). For example, a machine-readable medium may include a read-only memory (ROM); a random access memory (RAM); a disk storage medium; an optical storage medium; a flash memory device; an electrical, optical, acoustic, or other form of propagation signal (e.g., a carrier wave, an infrared signal, a digital signal, etc.), etc. In addition, firmware, software, routines, instructions may be described herein as performing certain actions. However, it should be understood that such descriptions are merely for convenience, and such actions are actually caused by a computing device, processor, controller, or other device that executes firmware, software, routines, instructions, etc.

[0247] In the block diagrams, the illustrated components are depicted as discrete functional blocks, but the embodiments are not limited to organizing the functions described herein into a system as shown. The functions provided by each component may be provided by software modules or hardware modules that are organized differently than currently depicted, for example, such software or hardware may be mixed, combined, replicated, decomposed, distributed (e.g., in a data center or geographic location), or otherwise organized differently. The functions described herein may be provided by one or more processors of one or more computers executing code stored on a tangible, non-transitory machine-readable medium. In some cases, a third-party content delivery network may host some or all of the information transmitted over the network, in which case, in the case of information (e.g., content) allegedly being supplied or otherwise provided, the information may be provided by sending instructions to obtain the information from the content delivery network.

[0248] Unless otherwise expressly stated, it will be clear from the discussion that throughout the specification, discussions utilizing terms such as "processing," "computing," "calculating," "determining," or the like refer to actions or processes of a specific apparatus, such as a special purpose computer or similar special purpose electronic processing / computing device, should be understood.

[0249] The reader should understand that this application describes multiple inventions. Because the related subject matter of these applications can make the application process more economical, the applicant did not separate these inventions into multiple separate patent applications, but instead grouped these inventions into one document. However, the unique advantages and aspects of this invention should not be lumped together. In some cases, multiple embodiments solve all the defects pointed out herein, but it should be understood that the present invention is independently useful, and some embodiments only solve a subset of these problems, or provide other unmentioned benefits, which is clear to those skilled in the art who review the technology disclosed herein. Due to cost constraints, some of the inventions disclosed herein may not be claimed for protection at present, but may be claimed in later applications, such as continuation applications or by amending the present claims. Similarly, due to space limitations, neither the abstract nor the content of the invention section of this document should be considered to contain a comprehensive list of all such inventions or all aspects of such inventions.

[0250] It should be understood that the specification and drawings are not intended to limit the invention to the particular forms disclosed, but on the contrary, the intention is to cover all modifications, equivalents, and alternatives falling within the spirit and scope of the invention as defined by the appended claims.

[0251] In view of this description, modifications and alternative embodiments of various aspects of the present invention will be clear to those skilled in the art. Accordingly, this description and the accompanying drawings should be interpreted as illustrative only, and the purpose is to teach those skilled in the art the general way of implementing the present invention. It should be understood that the form of the present invention shown and described herein will be regarded as an example of an embodiment. It is clear to those skilled in the art who have benefited from the description of the present invention that elements and materials can replace those illustrated and described herein, parts and processes can be reversed, changed in order or omitted, certain features can be utilized independently, and the features of embodiments or embodiments can be combined. Without departing from the spirit and scope of the present invention as described in the appended claims, the elements described herein may be changed. The titles used herein are only for organizational purposes and are not meant to limit the scope of this specification.

[0252] As used throughout this application, the word "may" is used in a permissive sense (i.e., meaning that it is possible) rather than a mandatory sense (i.e., must). The words "include", "including", and "includes" and the like mean, including, but not limited to. As used throughout this application, the singular forms "a", "an", and "the" include multiple references unless the context clearly indicates otherwise. Thus, for example, although other terms and phrases such as "one or more" are used for one or more elements, reference to "an element / aelement" includes a combination of two or more elements. Unless otherwise indicated, the term "or" is non-exclusive, i.e., includes both "and" and "or". Terms describing conditional relationships, such as "in response to X, and Y", "after X, Y", "if X, then Y", "when X, Y", etc., cover causal relationships where the antecedent is a necessary causal condition, the antecedent is a sufficient causal condition, or the antecedent is a contributing causal condition to the result, for example, "after condition Y is obtained, state X occurs" is common to "only after Y occurs, X occurs" and "after Y and Z, X occurs". These conditional relationships are not limited to results obtained by immediately following the antecedent, because some results may be delayed, and in the conditional statement, the antecedent is associated with its result, for example, the antecedent is related to the probability of the result occurring. Unless otherwise indicated, a statement that multiple attributes or functions are mapped to multiple objects (for example, one or more processors that perform steps A, B, C, and D) covers both all of these attributes or functions being mapped to all of these objects and subsets of attributes or functions being mapped to subsets of attributes or functions (for example, all processors each perform steps A to D, and where processor 1 performs step A, processor 2 performs step B and part of step C, and processor 3 performs part of step C and step D). Furthermore, unless otherwise indicated, a statement that a value or action is "based on" another condition or value encompasses both the case where the condition or value is the only factor and the case where the condition or value is one of multiple factors. Unless otherwise indicated, a statement that "each" instance of a set has certain properties should not be understood to exclude the case where some other identical or similar components of the larger set do not have that property, i.e., each does not necessarily mean every or every.

[0253] To the extent that certain U.S. patents, U.S. patent applications, or other materials (e.g., papers) are incorporated by reference, the text of such U.S. patents, U.S. patent applications, and other materials is incorporated by reference only to the extent that there is no conflict between such materials and the statements and drawings discussed herein. In the event of such a conflict, any such conflicting text in the incorporated by reference U.S. patents, U.S. patent applications, and other materials is not expressly incorporated by reference herein.

[0254] While specific embodiments of the present disclosure have been described above, it will be appreciated that the embodiments may be practiced otherwise than as described.

Claims

1. A method for training a deep learning model for a patterning process, the method comprising: obtaining (i) training data, the training data comprising an input image of at least a portion of a substrate having a plurality of features and a real image corresponding to the input image; (ii) a set of categories, each category corresponding to a feature among the plurality of features of the substrate within the input image; and (iii) a deep learning model configured to receive the training data and the set of categories; Generate a predicted image by modeling and / or simulating the deep learning model using the input image, wherein the deep learning model iteratively performs a convolution operation using at least one dilated kernel including a plurality of weights; Assigning a category from a category set to a feature in the predicted image based on a match between the feature in the predicted image and a corresponding feature in the real image; and A trained deep learning model is generated by modeling and / or simulation by iteratively assigning weights to at least one dilated kernel based on back propagation through the deep learning model using a loss function so that the predicted image reproduces the multiple features of the input image.

2. The method of claim 1, wherein the convolution operation comprises: overlapping the input image with the at least one dilation kernel, determining a product between each overlapping element comprising a weight of the at least one dilated kernel and a pixel value of the input image, and The results of the multiplications are summed to determine the value of the convolution operation.

3. The method of claim 1, wherein the at least one dilation kernel comprises a plurality of weights, a dilation rate, and a depth, wherein the depth is a multiple of the number of categories of the category set.

4. The method of claim 1, wherein the loss function is a function of the true image and the probability assigned to each class in a set of classes of the predicted image.

5. The method of claim 4, wherein: The iteratively assigning weights includes modifying a plurality of weights, a dilation rate, and / or a depth of the dilation kernel to characterize at least a portion of the patterning process based on a local derivative of the loss function.

6. The method of claim 1, wherein the at least one dilated kernel comprises: a first dilated kernel of the at least one dilated kernel comprising a first set of weights, a first dilation rate, and a depth equal to the number of categories in the set of categories; A second dilated kernel among the at least one dilated kernel comprises a second weight set, a second dilation rate, and a depth equal to the number of categories in the category set, wherein the second weight set is different from the first weight set, and / or the second dilation rate is different from the first dilation rate.

7. The method of claim 6, wherein: The first expansion core characterizes a first portion of the patterning process and the second core characterizes a second portion of the patterning process, wherein the second portion is a projection system of the patterning process and the second portion is a parameter related to a feature of the substrate.

8. The method of claim 1, wherein: The expansion kernel is symmetric to characterize the projection system of the device for the patterning process.

9. The method of claim 8, wherein: The iteratively assigning weights includes assigning a first set of weights to a plurality of weights of the dilated kernel to characterize a projection system of an apparatus of the patterning process.

10. The method of claim 3, wherein the expansion ratio is an integer value between 1 and 10.

11. The method of claim 7, wherein the parameters are edge positioning error, overlay, line edge roughness, necking and / or CD.

12. The method of claim 11, further comprising: determining an adjustment to the patterning process based on a parameter associated with the feature; and In response to the adjustment, the patterning process is adjusted.

13. The method of claim 12, wherein: The adjustments include adjustments to process variables including dose and / or focus.

14. A computer program product, comprising a non-transitory computer-readable medium, wherein the non-transitory computer-readable medium has recorded therein instructions, wherein when the instructions are executed by a computer, the following method is implemented: obtaining (i) training data, the training data comprising an input image of at least a portion of a substrate having a plurality of features and a real image corresponding to the input image; (ii) a set of categories, each category corresponding to a feature among the plurality of features of the substrate within the input image; and (iii) a deep learning model configured to receive the training data and the set of categories; Generate a predicted image by modeling and / or simulating the deep learning model using the input image, wherein the deep learning model iteratively performs a convolution operation using at least one dilated kernel including a plurality of weights; Assigning a category from a category set to a feature in the predicted image based on a match between the feature in the predicted image and a corresponding feature in the real image; and A trained deep learning model is generated by modeling and / or simulation by iteratively assigning weights to at least one dilated kernel based on back propagation through the deep learning model using a loss function so that the predicted image reproduces the multiple features of the input image.

Images