Method and device for determining topography contrast and / or material contrast of sample

Abstract

To provide a method (2800) and devices (800, 2800) for determining topographical contrast (585) and / or material contrast (550, 560) of samples (300, 890, 1000).SOLUTION: A method (2800) includes the steps of: (a) providing (2820) at least two image representations (720, 730) of samples (300, 890, 1000) recorded at least partly at different solid angles relative to the samples (300, 890, 1000); and (b) determining (2830) topographical contrast (585) and / or material contrast (550, 560) of the samples (300, 890, 1000) based at least partly on the at least two image representations (720, 730) of the samples (300, 890, 1000).SELECTED DRAWING: Figure 7

Description

[Technical Field]

[0001] This patent application claims priority from German Patent No. 102024103589.7 entitled "Method and Device for Determining the Topographical and / or Material Contrast of a Sample" (Method and Device for Determining the Topographical and / or Material Contrast of a Sample), filed with the German Patent and Trademark Office on February 8, 2024. In this connection, reference is made to German Patent No. 102024103589.7, the content of which is incorporated herein by reference.

[0002] The present invention relates to a method and device for determining the topographical and / or material contrast of an image representation of a specimen. [Background technology]

[0003] As a result of the ever-increasing integration density in microelectronics, lithography masks are required to image ever-smaller structural features into the photoresist layer of the wafer. To meet these requirements, exposure wavelengths are shifting to shorter and shorter wavelengths. Currently, argon fluoride (ArF) excimer lasers are often used for exposure; these lasers emit at a wavelength of 193 nm. However, light sources emitting in the extreme ultraviolet (EUV) wavelength range (10 nm to 15 nm) and corresponding EUV masks are also in use. The resolution capabilities of the wafer exposure process are being improved by the simultaneous further development of multiple variants of traditional binary lithography masks. Examples are phase masks or phase-shift masks and masks for multiple exposures.

[0004] Due to the continuous miniaturization of structural elements, lithography masks, specifically photolithography masks, or masks in general, cannot always be produced without defects that can be printed or seen on a wafer. Because photomasks are expensive to manufacture, defective masks are repaired whenever possible. This also applies to tiny samples or components, such as stamps in nanoimprint lithography.

[0005] Two important groups of defects in lithographic or photolithographic masks are black defects. These defects are areas where absorber and / or phase-shifting material is present but should not be. These defects are repaired by removing excess material, preferably using a localized etching process.

[0006] Second, there are so-called white defects. These are defects on photomasks that lack absorber and / or phase-shifting material and therefore, during exposure in a wafer stepper or wafer scanner, exhibit a greater optical transmittance than an identical defect-free reference position. In a mask repair process, these defects can be corrected by depositing a material with appropriate optical properties. Ideally, the optical properties of the material used for repair should correspond to those of the absorber or phase-shifting material. The layer thickness of the repaired location can then be adapted to the dimensions of the surrounding absorber or phase-shifting material layer.

[0007] Both types of defects are usually localized defects, whose dimensions are typically in the submicrometer range. These defects are often repaired by particle-beam-induced localized chemical processes. To repair localized defects on a specimen, at least one process gas is typically supplied to the processing site on the specimen, where a focused particle beam induces a localized chemical reaction. In the case of localized removal of excess material from the specimen, the process gas includes at least one etching gas. In the case of deposition of missing material, the process gas includes at least one deposition gas.

[0008] In the following description, various types of components such as masks, stamps for nanoimprint lithography, wafers to be repaired and MEMS (Micro-Electro-Mechanical Systems), NEMS (Nano-Electro-Mechanical Systems) or PICs (Photonic Integrated Circuits) are collectively grouped under the term specimen.

[0009] Unlike mechanical processes, such as sputtering processes, chemical processes are often slow processes. This means that a local chemical repair process of a sample can proceed on a time scale ranging from minutes to tens of minutes. To obtain reproducible results during sample repair, the local chemical process must be stopped at the right time, preferably automatically. The right time to deposit a phase-shift layer on a photomask is, for example, when the repair location has the same absorption and phase-shift properties as an equivalent defect-free area of ​​the mask. During a local etching process, the right time is reached when the layer to be removed has been etched, but preferably before the etching process begins to etch underlying layers of the sample. Therefore, to identify the time to stop a repair process performed using a local chemical reaction, an image representation of the sample that provides specific information about the local material composition of the sample can be used.

[0010] To image the sample to be repaired, a particle beam (e.g., an electron beam) (e.g., the same one used to initiate the local chemical repair process) is often used. This particle beam causes a local emission of, for example, secondary electrons (SEs) and backscattered electrons (BSEs) from the sample. Depending on the type of detector used to detect the SEs and BSEs, their placement relative to the solid angle distribution of the SEs and BSEs, and possibly the presence of an energy filter, the detector will record SE and BSE portions with different magnitudes.

[0011] SEs emitted from the sample are typically several times more frequent than BSEs, but more importantly, SEs are primarily responsible for the topographic contrast of the sample image, while BSEs, especially those that leave the sample in a narrow solid angle antiparallel to the primary particle beam, convey most of the information about the sample's local material composition.

[0012] Therefore, BSEs emitted substantially antiparallel to the beam direction of the primary particle beam are particularly suitable for determining the stop signal of a local chemical repair process. Nevertheless, detectors arranged around the primary particle beam, e.g., an electron beam, always detect SEs and BSEs simultaneously. This detector arrangement generally does not allow for complete separation of SEs and BSEs. Rather, such ring-shaped detectors typically receive the BSE signal together with a superimposed SE signal background. Introducing opposing electric fields makes it possible to at least partially separate SEs and BSEs in the sample image representation. However, image representations of samples imaged using detectors with built-in energy filters, which still actually contain portions of topographical and material contrast, are only to a limited extent suitable as a stop signal for a local particle-beam-induced repair process. Furthermore, incorporating detectors with built-in energy filters often poses significant technical challenges.

[0013] The present invention is therefore directed to the problem of defining a method and device for determining the topographical and / or material contrast of an image representation of a specimen. Summary of the Invention

[0014] According to one exemplary embodiment, this problem is solved by a method according to claim 1. In one embodiment, a method for determining topographical contrast and / or material contrast of a sample comprises (a) providing at least two image representations of the sample recorded at at least partially different solid angles relative to the sample, and (b) determining the topographical contrast and / or material contrast of the sample based at least in part on the at least two image representations of the sample.

[0015] The inventors have recognized that secondary particles emitted by a sample (either backscattered particles or particles emitted from the sample) have a non-uniform solid angle distribution. As a result, two sample image representations recorded at different solid angles and based on the detection of secondary particles, specifically secondary electrons (SE) and backscattered electrons (BSE), a priori contain different information regarding their topographical and material contrast. Furthermore, the energy distribution of secondary particles varies as a function of the solid angle from which the secondary particles are emitted. By linking the information contained in at least two image representations, it becomes possible to identify the topographical and material contrast portions contained in each individual image representation. In contrast, a single detector recording secondary particles (e.g., SE and BSE) from the same solid angle range as two or more detectors averages both the solid angle distribution and the energy distribution of the secondary particles used to generate the image representation. As a result, the information present in these distributions of secondary particles is lost.

[0016] One advantage of the method according to the present invention is that none of the detectors used to record the image representation require an energy filter for secondary particles. This, firstly, reduces the installation space required for the detectors. Secondly, the absence of an energy filter, for example in the form of a screening grid, can prevent interference with the primary charged particle beam, for example, an electron beam. For example, if a ring-shaped detector arrangement arranged around the primary charged particle beam includes a screening grid as an energy filter, the voltage of the screening grid can affect the incident energy of the primary particle beam on the sample, for example, the primary particles, i.e., the particles of the primary particle beam on the sample.

[0017] Furthermore, at the edge of the sample, for example at the edge of an absorbing pattern element of a photomask in the energy-filtered image representation or in the energy-filtered image, topographical effects arise as a result of screening effects. Therefore, a pure material contrast image cannot be measured in the region of the sample edge. The method according to the invention avoids this limitation by eliminating the energy filter for secondary particles.

[0018] During the simultaneous recording of at least two image representations using two detectors, it is advantageous if the solid angle ranges of the two detectors overlap as little as possible, and in the best case, not at all, in order to minimize the redundancy of information present in the two image representations. This is equally true if the two image representations are recorded successively by a single detector placed at different positions, or if the separate detectors are successively focused at different locations.

[0019] Providing at least two image representations recorded at different solid angles relative to the sample may include at least one element of the group consisting of: loading from a memory at least two image representations recorded at different solid angles; transmitting over a data connection at least two image representations recorded at different solid angles; or recording at least two image representations of the sample at different solid angles.

[0020] A representation or image of a sample can be recorded by scanning a focused primary electron beam that excites the sample over the sample or a partial area of ​​the sample and simultaneously detecting secondary electrons (SE) and backscattered electrons (BSE) emitted from the sample.

[0021] The focused primary particle beam may comprise a charged particle beam. The charged particle beam may comprise an electron beam and / or an ion beam. A primary particle beam in the form of an electron beam is advantageous.

[0022] The secondary particles can be detected by one or more detectors, at least one of which can use different detection principles, such as a scintillator counter in the form of a scintillator photomultiplier detector, a semiconductor detector, for example a diode structure with one or more segments, or an yttrium aluminum garnet (YAG) detector.

[0023] At least one detector can have a different geometry. Detectors placed near the sample, such as an Eberhart-Thornley detector, cover a solid angle range that is asymmetric with respect to both the polar angle (angle relative to the beam axis of the primary particle beam) and the azimuthal angle (relative to a line in the sample plane). So-called in-lens detectors, such as a Robinson detector, are positioned rotationally symmetric with respect to the beam axis of the primary particle beam and have an aperture for the passage of the primary particle beam. Therefore, due to their geometry and arrangement, in-lens detectors are polar angle independent. When an in-lens detector is placed near the sample, it records secondary particles from the sample over a wide solid angle range. This is advantageous because it improves the signal-to-noise ratio of the detector signal. On the other hand, the in-lens detector averages out a wide range of energy and solid angle distributions of secondary particles.

[0024] The number of secondary particles that leave the sample per particle of the primary or electron beam, i.e., the yield, depends on the kinetic energy or incident energy of the primary particle or electron. For very low incident energies, the secondary particle yield is <1, and is in the intermediate energy range >1, while for high incident energies the yield often drops again to the range <1.

[0025] The SE portion of secondary particles represents electrons emitted from the sample with kinetic energies up to approximately 50 eV, with most SEs having kinetic energies on the order of several eV (electron volts). When the primary particle beam includes an electron beam, electrons backscattered from the sample (BSE electrons) are more energetic than SEs, typically having kinetic energies of several keV (kiloelectron volts). The intensity of the BSE signal depends primarily on the atomic number, or, if the sample has a compound material composition, on the average atomic number. The intensity of the BSE signal increases with the (average) atomic number. This means that heavy elements produce strong backscattering, and regions of the sample with heavy elements appear bright as a result of a strong BSE signal. Therefore, in this application, the dependence of the intensity of the BSE signal on the local material composition can be used to detect changes in the local material composition of the sample in the z-direction, i.e., perpendicular to the sample plane.

[0026] Determining the topographical and / or material contrast can include applying a separation model to the at least two image representations. Specifically, determining can include applying a parameterized or trained separation model to the at least two image representations. The separation model includes a mathematical model applied to the at least two image representations recorded at different solid angles to determine the topographical and / or material contrast.

[0027] The presence of a signal indicative of the local material composition of a mask is important for monitoring purposes, e.g., with respect to mask repair processes. This signal should be as independent as possible of the local topography of the mask, or generally of the specimen, while at the same time reacting as sensitively as possible to local changes in the mask's material composition. The local topography takes into account the topography of the mask in its immediate vicinity. For example, determining the material contrast at the bottom of an edge is made more difficult by edge shading effects.

[0028] For example, in the case of a sample in the form of a binary photomask, it is only necessary to distinguish between the material of the absorbing pattern element and the material of the mask substrate. For this purpose, at least two, preferably multiple, detector signals are recorded from one point on the sample or mask. At least two signals from the points recorded at different solid angles are processed to form a function to generate a material signal. In the simplest form, a linear combination of different detector signals from one point on the mask can be determined. However, any other function of the detector signals can also be determined. The coefficients of this function can be calibrated, for example, so that the function takes the value 0 for the material composition of the pattern element and the value 1 for the material composition of the mask substrate, or vice versa. This makes it possible to determine an image representation of the sample that essentially contains only information about the local material composition of the sample (referred to as compositional contrast).

[0029] This image representation can be used to automatically stop the particle beam-induced local chemical repair process of the sample. Sample defects can be reproducibly repaired using the method according to the invention, while reliably avoiding sample damage due to unwanted deposition or removal of material onto or from the sample during the repair process.

[0030] At least two image representations can capture a range of solid angles of the sample that are at a small angle relative to the primary particle beam (the image representations can result, for example, from detectors that can be positioned around the primary particle or electron beam, so-called in-lens detectors). For example, if the detected secondary particles include secondary electrons (SEs) and backscattered electrons (BSEs), this detector arrangement will capture a large portion of the high-energy BSEs, which primarily provide information about the local material composition of the sample.

[0031] The method according to the present invention is also applicable to configurations including three or more detectors that record three or more image representations at partially different solid angles. The use of two or more detectors receiving secondary particles, or SEs and BSEs, from different solid angle ranges allows the secondary particle distribution to be resolved, its inhomogeneity to be identified, and used to determine topographic and / or material contrast. In contrast, a single detector receiving secondary particles from different solid angle ranges of two or more detectors integrates or averages the inhomogeneous solid angle distribution of secondary particles. Therefore, the accuracy of determining the local material composition of the sample increases with the number of image representations recorded at different solid angles and available as input data for the method according to the present invention.

[0032] When a flat sample surface is imaged and the focused particle or electron beam is incident on the sample surface at normal incidence, different solid angles involve different angles or different polar angles relative to the primary focused particle or electron beam.

[0033] In the case of a configuration with two in-lens detectors, where the first detector is positioned closer to the sample than the second detector, a first of the at least two image representations recorded by the first detector may encompass a solid angle in the range of 0.6 sr to 1.0 sr, preferably 0.26 sr to 1.2 sr, and most preferably 0.15 sr to 0.4 sr for the SE, and a solid angle in the range of 0.3 sr to 1.3 sr, preferably 0.5 sr to 1.6 sr, and most preferably 0.8 sr to 2.0 sr for the BSE. A second of the at least two image representations recorded by the second detector may have a solid angle for the SE in the range of 0.2 sr to 0.6 sr, preferably 0.15 sr to 1.3 sr, and most preferably 0.05 sr to 0.15 sr, and a solid angle for the BSE in the range of 0.15 to 0.6 sr, preferably 0.1 sr to 0.26 sr, and most preferably 0.05 sr to 0.15 sr. The abbreviation "sr" stands for steradian.

[0034] In the example of an in-lens arrangement of one or more detectors in the column of a scanning electron microscope (SEM), the solid angle range from which SEs and BSEs are incident on the detector depends on the kinetic energy of the SEs and BSEs. Secondary particles are focused by the SEM's objective lens at the back focal plane of the SEM. The lower the kinetic energy of the secondary particles, i.e., SEs and BSEs, the closer this focal point is to the sample. In the best possible arrangement, the first detector is positioned at the focal plane of the BSEs so that the maximum percentage of BSEs can reach the second detector and, at the same time, the first detector can direct the maximum possible portion of SEs away from the second detector. The primary role of the objective lens is to focus the primary charged particle beam onto the sample; i.e., the front and back focal planes of the objective lens depend on the selected settings of the primary electron beam or the distance between the sample and the objective lens.

[0035] At least two image representations can be recorded simultaneously by two detectors observing the sample from two partially different directions, where one of the detectors can partially conceal or shade the solid angle range seen by the second detector.

[0036] However, two image representations can also be recorded consecutively by one detector. In this case, the detector must be moved from the first imaging position to the second imaging position between the first and second imaging or image recording. When determining the topography contrast and material contrast portions of at least two image representations, the shading effect of simultaneous imaging using two detectors can also be taken into account during progressive image recording using one detector. It is also possible to cover a first portion of an individual detector during the first imaging and a second portion of the individual detector during the second imaging. In this case, it is advantageous if the individual detectors have the largest possible detection area.

[0037] A detector collects the primary particles or secondary particles that the electron beam detaches from a point on the sample and that reach the detector's entrance aperture. By scanning the primary electron beam two-dimensionally over the sample, a two-dimensional intensity distribution of the secondary particles, i.e., an image representation of the sample, is simultaneously generated on a monitor.

[0038] The sample is then imaged by scanning with a primary focused particle beam. Scanning the sample with a primary particle beam, preferably an electron beam, causes secondary particles (electrons) to leave the sample at the location of the primary particle's (electron) incidence. These secondary particles are recorded by one or more detectors, which then generate a two-dimensional intensity pattern of the sample. The number of secondary particles leaving the sample per incident primary particle depends not only on the sample's material but also on the sample's surface topography. From localized prominences, such as along the edges of elevated areas (e.g., along the edges of pattern elements), the local angular range of the sample surface from which secondary particles can be emitted is greater than 180 degrees or greater than π. Therefore, localized elevated areas and / or edges generate more secondary particles than flat sample surfaces. As a result, they appear brighter compared to flat sample surfaces. In contrast, in locally depressed areas, the angular range from which the sample can emit secondary particles is less than 180 degrees, and therefore they also appear darker in the image representation of the sample compared to flat sample surfaces.

[0039] For a flat sample surface and a predefined incident energy of the primary particles, the number and angular distribution of secondary particles depends substantially on the material or material composition of the sample. If the secondary particles are electrons scaled for incident energies >600 eV, the number of backscattered electrons (BSEs) generated is approximately proportional to the mass number of the sample material. This means that for a predefined particle flow, the stronger the local intensity of the sample image representation, the greater the mass number of the local sample material.

[0040] Contributions due to local material composition and local surface irregularities or surface topography are typically superimposed in the intensity distribution with corresponding contributions, and an important goal of this application is to separate these contributions.

[0041] As described below, a predefined sample image representation typically has both topographical and material contrast portions inherent to that image representation. Thus, in a coordinate system spanned by a purely topographical contrast portion (T axis) and a purely material contrast portion (M axis), the sample image representation typically does not lie on one of these axes. The parameterized empirical model can first rotate the sample image representation so that the sample image representation lies on the M axis or the T axis of the coordinate system (and thus has only the material contrast portion of the sample image representation or only the topographical contrast portion of the sample image representation). The respective rotation angles indicate the magnitude of the topographical or material contrast portion of the analyzed sample image representation.

[0042] The parameterized empirical model can more generally vary the position of the image representation of the sample within the plane spanned by the topography contrast and the material contrast.

[0043] For example, a parameterized empirical model can separate the image representation of the inspected sample into a material contrast portion and a topography contrast portion, i.e., the sample image representation can be expressed as a linear combination (e.g., of two unit vectors in the M and T axes). This is equivalent to projecting the sample image representation being analyzed or determined onto the T and M axes of an M / T coordinate system. For example, the material contrast portions of both image representations can then be added together to form the overall image.

[0044] In both cases, the material contrast of the specimen can be identified. Changes in the specimen image representation with only the material contrast clearly and reliably reveal changes in the specimen's local material composition. Such changes occur, for example, when an etching process reaches a layer boundary. Changes in the material contrast portion isolated from the specimen image representation can be used to control particle-beam-induced local chemical repair processes.

[0045] The material contrast of a sample can be determined by solving an optimization problem. In this case, it is advantageous to use all available data of the sample to solve the optimization problem. For example, a sample in the form of a lithography mask significantly limits the material composition of the mask's components (in the simplest case of a binary mask, these are the mask substrate and the absorbing pattern elements). In this regard, the pure M and T parts of the mask image representation can be determined from the design data. The design data can also be used to create an empirical model for a particular type of mask. The detector configuration used to record the at least two image representations can be taken into account when creating the empirical model, i.e., the distance between the individual detectors and the sample and the distance between the detectors. The empirical model can also include one or more parameters that define the operating point of the SEM. The operating point includes at least the parameters, namely, the kinetic energy of the particles of the primary charged particle beam, the distance between the sample and the objective lens, and the aperture angle of the primary particle beam.

[0046] The separation model may include at least one member of the group consisting of an empirical model or a transformation model.

[0047] Separation models can pursue various approaches to analyzing image representations with respect to the topographical and material contrast components they contain. Empirical models can establish analytical relationships between two or more image representations and their different topographical and material contrast components. In contrast, transformation models can dispense with establishing a functional relationship between the solid angle distributions contained in two or more image representations of detected secondary particles and the local topography and material composition of the sample. Instead, this relationship can be implicitly established or learned by training the transformation model.

[0048] A separation model can be designed to link at least two image representations together. Electron-optical simulations can be performed to determine the solid angle distribution or acceptance function of one or more detectors. In combination with the simulated angular distributions of the SE and BSE, a separation model can be established for the M and T portions of the at least two image representations.

[0049] Linking the at least two image representations can include linearly modifying the at least two image representations and combining the at least two modified image representations. Combining the at least two modified image representations can include at least one element of the group consisting of adding the at least two modified image representations, subtracting the at least two modified image representations, multiplying the at least two modified image representations, or dividing the at least two modified image representations. Linking the at least two modified image representations can also include determining a material contrast portion of at least one of the at least two image representations from the at least two linked image representations.

[0050] As explained above, a sample image representation produced by scanning a sample with a primary particle beam contains intensity components that originate from the surface topography and material composition of the sample. These components cannot be simply separated when considering a single sample image representation. However, in addition to the incident energy of the primary particles, these components also vary as a function of the solid angle through which the sample, or a portion or region of the sample, is observed.

[0051] The empirical model can link at least two image representations that generate secondary particles that leave the sample at at least partially different solid angles, the solid angles from which the secondary particles of the two image representations emanate being known from the detector arrangement, making it possible to identify topography contrast and material contrast portions of the sample using the at least two image representations.

[0052] The method according to the present invention may further comprise a step of adapting the empirical model to the sample. The complexity of the empirical model may be adapted to the complexity of the sample to be imaged. The complexity of the sample depends on the contour or structure of the sample or the material composition of the sample. For example, a sample with a structure having only small differences in atomic mass is more difficult to separate than a sample with a structure having large atomic mass differences. A sample with a structure having small atomic mass differences requires a complex separation model.

[0053] Adapting the empirical model can include adapting parameters of the empirical model to the number of different material compositions of the specimen. A specimen in the form of a binary photomask is characterized by two components with different material compositions. In contrast, a phase-shift mask has at least three components with different material compositions. An empirical model with two parameters is sufficient to determine a material contrast image representation and / or a topography contrast image representation of a binary mask. In contrast, an empirical model for a phase-shift mask requires at least three parameters.

[0054] The method according to the present invention may further comprise determining parameters of the empirical model.

[0055] Determining the parameters of the empirical model may include at least one element of the group consisting of: recording at least two image representations of the at least one calibrated test structure at at least two different solid angles; simulating at least two image representations of the at least one calibrated test structure at different solid angles; or recording at least two image representations of the at least one calibrated test structure at different solid angles, wherein at least one detector has an activated screening grating.

[0056] The recording of the at least two image representations can be performed at at least one first position and at least one second position of the calibrated test structure, where the at least one first position and the at least one second position have different material compositions. In the case of a binary mask calibrated test structure, this means that at least two image representations are recorded at different solid angles of the mask structure and at least two image representations are recorded at different solid angles of the absorbing pattern elements. In the case of a phase-shift mask, at least two image representations are recorded at different solid angles of the mask substrate, where at least two image representations of the phase-shift material are recorded and at least two image representations of the absorbing material of the pattern elements are recorded. The material contrast and topography contrast of the calibrated test structure are known from the calibration performed.

[0057] Simulating the at least two image representations may include performing a Monte Carlo simulation of the interaction of the primary particle beam with the sample and an electron optical simulation of the paths of secondary particles from the sample to the at least two detectors.

[0058] Determining the parameters of the empirical model may include modifying the parameters of the empirical model to minimize differences between the measured image representation of the sample and the measured or simulated image representation of the calibrated test structure.

[0059] To determine a local material signal or point-like image representation having only the signal portion resulting from the material contrast, a local or point-like sample image representation can be recorded by one or more detectors at a point on the sample (pixel-by-pixel determination) and processed with a function to determine the material signal at that point on the sample surface. In the simplest form, a linear combination of two or more detector signals can be taken for this purpose. It goes without saying that detector signals can be combined with other functions. It goes without saying that the coefficients of the function can be determined by comparison with a calibrated test structure for which the material signal (i.e., the material contrast portion of the image representation) is known.

[0060] For a sample consisting of two components, e.g., a binary photomask, the coefficients of the function can be calibrated to take the value 1 for the first component, e.g., the mask substrate, and the value 0 for the second component, e.g., the absorbing pattern element.

[0061] The last described embodiment for determining the parameters of the empirical model does not require the primary particle beam to be scanned over the sample, since only the sample signal emanating from one point on the sample is processed. The more signals from different detectors that can be processed, the more accurately the local material signal of the sample can be identified. This represents an advantage over the individual energy-filtered signals that are currently often used to determine material contrast. The use of two or more detectors allows more secondary particles emitted locally from the sample to be detected and thereby made usable. As a result, a better signal-to-noise ratio can be achieved and / or data recording can be performed more quickly.

[0062] Since the measured topography data of the sample is not completely local and, moreover, detector-dependent, it cannot be fully compensated for when determining the parameters of the empirical model mentioned above locally. However, the compensation part increases with the number of detectors used. Furthermore, the parameters of the empirical model need to be optimized for each operating point (incident energy of primary particles or electrons, distance between mask and objective lens, objective lens settings, etc.) and for each type of sample or mask.

[0063] Determining an image representation having only a material contrast contribution based on an empirical model can include scanning a primary particle beam over the sample and recording at least two detector signals at a scanning point of the primary particle beam or the electron beam.

[0064] The pixel-by-pixel determination of the parameters of the empirical model as described above ignores non-local effects due to sample topography. Using a primary particle beam scanned over a portion of the sample, the non-local signal portions of at least two detector signals can be taken into account when determining the parameters of the empirical model.

[0065] Rather than recording a signal at only one point, multiple image signals are recorded for each of at least two detectors using a scan of the primary beam over a portion of the sample. New signals can be generated from the images of the individual detectors by convolution with one or more convolution kernels. Parameters of the empirical model can be determined by adapting the convolved measurement data to measurements of a calibrated test structure.

[0066] At least one image filter from the group consisting of a Sobel filter, a Prewitt filter, a Laplacian filter, a Marr-Hildreth filter, a Gaussian filter, or a Sharpen filter can be used as the convolution kernel. The size of the convolution kernel can be adapted to the size of the scan area. The scan area of ​​the primary particle beam needs to be selected to be larger than the image area of ​​the convolution kernel. Furthermore, to maintain the accuracy of the convolution operation, it is advantageous to select the size of the scan area so that procedures in edge areas (e.g., dilation, lap convolution, or cropping) remain below a predefined error threshold. The scan area may include a size in the range of 2 nm to 50 nm in one dimension, preferably 5 nm to 20 nm.

[0067] Determining the parameters of the empirical model from a scanned area instead of a pixel-by-pixel determination allows for a significant reduction of residual topographical effects in the material contrast image representation of the sample or sample section. The advantage of significantly improved quality is countered by the disadvantage of significantly higher complexity. Scanning over a sample area takes longer than recording an image or image representation of an individual sample point. The size of the convolution kernel can be determined by performing Monte Carlo simulations. The optimization process for determining the parameters of the empirical model needs to be performed for each sample type and operating point.

[0068] After the parameters of the empirical model are defined, the empirical model can calculate at least two image representations of the sample, the image representations having different topography contrast portions and material contrast portions such that the topography contrast portions and material contrast portions of the at least two presented sample image representations can be extracted from the sample image representation.

[0069] The transformation model may include at least one transformation model, preferably a machine learning model and / or a generative model, having at least two transformation blocks each including at least one generic learnable function.

[0070] The trained transformation model can transform image representations from at least two image representations measured at least in part at different solid angles into an image representation having a predefined portion of topography contrast and / or material contrast. Specifically, the transformation model can be trained to transform the measured image representation so that the transformed image representation shows only the material contrast portion of the presented image representation being analyzed. Alternatively and / or additionally, the transformation model can be trained to output the topography contrast portion and / or material contrast portion of one of the two measured image representations as a numerical value. A change in the material contrast portion that exceeds a predefined threshold can be used to detect a change in local material composition.

[0071] However, the transformation model can be trained to transform both or all of the at least two measured image representations into image representations each having a predefined ratio of topographic contrast to material contrast, and the transformation model can be trained so that the first transformed image representation shows only topographic contrast and the second transformed image representation shows only material contrast.

[0072] A transformation model need not include a sequence of encoders, feature projections, and decoders. Instead, each of the N layers (e.g., of a neural network) can be provided with a generic learnable function that transforms an input to an output without requiring one of the intermediate steps to generate an appropriate, and therefore transferable, representation (feature) of the input.

[0073] The generalized trainable function of a transformation block in a transformation model takes the output data of the preceding transformation block as input data. Generally speaking, the Nth transformation block takes the output data of the (N-1)th transformation block (O N-1 ) is obtained by dividing the output data of the (N-1)th transformation block by the input data of the Nth transformation block (IN ), i.e. O N-1 =I N The output data of the (N-1)th transform block is N-1 may contain the original input data I1 of the first transform block in unchanged form, and the output data of the (N-1)th transform block may contain the input data (I1,...,I N-1 ) can be included. Also, the input data I N can contain the input data transformed by the preceding transform block, i.e., T N-1 (O N-2 ),P N-1 )=T N-1 (I N-1 ),P N-1 ) In this case, T N-1 is the input data I in the (N-1)th transformation block. N-1 This shows the transformation performed on P N denotes the model parameters of the transformation model in the Nth transformation block.

[0074] In the Nth transformation block, the transformation T N describes the convolution operator, this transformation block uses the transformed data T N-1 (I N-1 ,P N-1 ) is the only input variable or input data I N and the input data I1,...,I N-1 is ignored. The model parameters P N corresponds to the convolution weights, and the Nth transform block performs the function of the convolution layer or convolution block.

[0075] The generic learnable function of the transform block may include at least one element of the group consisting of a convolution block, a deconvolution block, a pooling block, an unpooling block, a Dense block, a Res block, an Inception block, an encoder, or a decoder.

[0076] The transformation model can transform at least one image representation of the tuple of measured image representations into a transformed image representation that appears like an image representation having a predefined ratio of topography contrast to material contrast, but the transformation model can also be trained to transform at least one simulated image representation of the tuple of simulated image representations into a transformed image representation having a predefined topography contrast ratio and / or material contrast ratio.

[0077] A machine learning model can include an encoder-decoder structure. In an encoder-decoder architecture, input data is mapped (encoded) on the encoder side to information-carrying features, or features, by a set of learnable functions. On the decoder side, from these features, target data, in this case at least one transformed image representation with predefined topographic contrast portions and / or material contrast portions, is then extracted (decoded) also using learnable functions. These individual functions on both the encoder and decoder sides are usually called layers. In an encoder-decoder architecture, the input available for a layer is typically the output of the previous layer. However, it is also possible for corresponding layers on the encoder side and the decoder side to be connected to each other.

[0078] The machine learning model may include at least one element from the group consisting of a parameter mapping, a neural network (NN), an artificial neural network (ANN), a deep neural network (DNN), a time-delay neural network, a convolutional neural network (CNN), a recurrent neural network (RNN), or a long short-term memory (LSTM) network.

[0079] The particle beam-induced repair process of the specimen can be considered as a time series of local image representations of the specimen, and thus can be interpreted as a video recording. This perspective and the corresponding choice of network architecture allow for a significant improvement in the accuracy of the transformations performed.

[0080] Machine learning models (ML models) can involve symbolic systems. Knowledge, i.e., training data and induced rules, is explicitly expressed in the case of symbolic systems. However, in the case of symbolic systems, the model learns computable behavior without detailed visibility into the learned solution path.

[0081] Additionally, the machine learning model (ML model) may include at least one element from the group consisting of kernel density estimators, statistical models, decision trees, linear models, time-varying models, nearest neighbor classification, and k-nearest neighbor algorithms, as well as nonlinear extensions thereof that use nonlinear feature transformations.

[0082] Kernel density estimators (KDEs) allow for continuous estimation of unknown probability distributions based on random samples. Kernel density estimators can include, for example, Gaussian, Cauchy, Pickard, or Epanechnikov kernels, and kernel parameters, such as bandwidth, in ML models can be assigned or estimated individually or collectively for all input parameters.

[0083] The statistical model can include at least one mixture distribution. The mixture distribution can include a member of the group consisting of a Gaussian mixture distribution (GMM (Gaussian mixture model), a multivariate normal distribution, and a categorical mixture distribution. The appropriate number of mixture distributions depends on the data available and can be optimized using a validation data set.

[0084] A decision tree (DT) can contain at least one element from the group consisting of a traditional decision tree (DT), a random decision tree (RDT) or a decision forest (DF) and a randomized version of the latter (RDF). In RDTs and RDFs, the degree or "level" of randomization can vary. For each node, all or only a random selection of the possible decisions can be present in training. For each leaf of the decision tree, all or only a subset of the training examples present up to that point can be used.

[0085] The linear model may include at least one element from the group consisting of latent Dirichlet allocation (LDA), support vector machine (SVM), logistic regression, least squares estimation, lasso regression, ridge regression, or perceptron. Advantageous applications of linear models require normalization of the input and training data.

[0086] ML models may include nonlinear extensions of SVMs in the form of kernel support vector machines, and ML models may include nonlinear extensions of Gaussian mixture distributions in the form of Gaussian process regression.

[0087] The time-varying model may include at least one element from the group consisting of a recurrent neural network or a hidden Markov model. The time-varying model may be simulated by a time-invariant model with parameters of earlier measurements provided as input data by the time-invariant model.

[0088] The ML model may also include two or more different types of machine learning models from the groups specified above. A machine learning model that uses a collection or group of multiple different types of models or multiple learning algorithms can generally achieve better results than an ML model based on a single type of model or learning algorithm. Computing the results of multiple different types of models typically takes longer than evaluating a single type of ML model. However, the tradeoff is that results corresponding to a single type of ML model or an ML model with a single learning algorithm can be obtained more quickly with a shallower computation depth.

[0089] The predictions of different components of the combination can contribute to the prediction of the machine learning model in an equally weighted manner. The predictions of different types of ML models can contribute to the prediction of the machine learning model in a weighted manner.

[0090] A machine learning model comprising a group of different types of ML models can be built incrementally in training stages, with each new type of model being added to the group being provided with training data that the previous types of models in the group were unable to predict or were able to predict poorly.

[0091] Two or more different ML model types of machine learning models can be selected using automated machine learning (or AutoML).

[0092] The transformation model can include a machine learning model, specifically a deep learning model. The machine learning model can use an artificial neural network. The deep learning model can use a deep neural network. A deep neural network includes multiple or many intermediate layers (called hidden layers). In addition to classifying the data presented to it, deep learning architectures can also extract features from the data presented to them.

[0093] The deep neural network (DNN) can include a U-net architecture or a ResNet architecture.

[0094] The U-Net architecture has symmetry between the encoder and decoder branches, resulting in the U-shaped structure of the architecture. In addition to relaying data between different layers of the encoder and decoder, the U-Net architecture can have additional connections between corresponding layers of the encoder and decoder, through which output data from the encoder layer is directly forwarded to the corresponding layer of the decoder as additional input data.

[0095] Residual network (ResNet) architectures are also encoder-decoder structures in that data is not simply relayed between adjacent layers in the encoder and decoder branches. ResNets have additional connections through which output data is forwarded across and / or through multiple layers as input data to every other layer. These additional data forwarding typically occurs on both the encoder and decoder sides of the architecture.

[0096] The machine learning model may include at least one additional parameter that is provided to the machine learning model (ML model) in its input.

[0097] As an alternative to the above procedure, one or more additional parameters are transferred to the ML model in addition to the measurement image representation tuple. The one or more additional parameters are available to the transformation model or ML model as inputs both in the training phase and for determining the topography contrast portion and / or material contrast portion of the measurement image representation tuple. This allows for the use of a generalized model for different types of masks for different parameter settings of the repair device for prediction. Only the generalized transformation model or ML model needs to be trained, which can then be used for different types of masks and different settings of the system parameters of the repair device.

[0098] The at least one additional parameter may include a system parameter of the repair device.

[0099] The at least one additional parameter may include at least one parameter of the lithography mask and / or at least one system parameter of the repair device or its imaging system. The at least one parameter of the mask may include the type of mask, the dimensions and material composition of the mask substrate and / or pattern elements, and / or the at least one system parameter may include the incident energy of the primary particle beam, the incident angle of the primary particle beam on the sample, the exposure setting of the objective lens, the aperture angle of the primary particle beam, the scan speed of the primary particle beam, the integration time (dwell time) of the primary particle beam, the scanning scheme (e.g., line-by-line, interlaced, or diagonal scan), the asymmetric positioning of the at least one detector with respect to the beam axis of the primary particle beam, the potential setting of the energy filter, and / or the potential setting of the screening grating at the particle optical column output.

[0100] The ML model can include at least one hyperparameter that characterizes the sample. The deep learning model can include at least one hyperparameter that characterizes the sample. The sample can include a photomask, and the hyperparameter can define the type of mask. Hyperparameters of machine learning and / or deep learning models are model parameters that are defined before the start of the training phase of the ML or deep learning (DL) model.

[0101] A conversion model, ML model, or deep learning model can include a common encoder branch for the input data and at least one additional parameter, and can have a dedicated decoder branch for each additional parameter. In this regard, a generic photomask can be converted into a binary mask, for example, by hyperparameters. This is done by enabling selected decoder branches while disabling other decoder branches, i.e., by multiplying them by 0, for example. Unlike hyperparameters, the model parameters of a conversion model, ML model, or DL ​​model are determined during a learning or training process.

[0102] Even when the at least one additional parameter is considered, the problem solved for the conversion model, ML model, or deep learning model is similar, and therefore the conversion model can "reuse" some of the model parameters already determined, so the at least one additional parameter increases the training complexity of the conversion model or deep learning model only sublinearly.

[0103] The method according to the invention may further comprise the step of training the transformation model with the training data set.

[0104] The key to predicting the material and / or topography contrast parts of tuples in the measured image representation is the training of an ML or DL ​​model, typically a transformation model, with a sufficient amount of training data. Especially for deep learning architectures with many hidden layers and therefore a large number of parameters, the quality and quantity of the training data or training dataset plays a key role.

[0105] The training dataset for the transformation model may include at least one element of a group consisting of a multiple tuple of at least two recorded image representations of at least one sample used for training, a multiple tuple of at least two recorded image representations of at least one test structure used for training, a multiple tuple of at least two simulated image representations of at least one sample used for training, or a multiple tuple of at least two simulated image representations of at least one test structure used for training, wherein the at least two tuples of image representations are recorded or simulated at at least partially different solid angles relative to the at least one sample and / or test structure used for training.

[0106] The simulation of the training data may include a Monte Carlo simulation of the interaction of a primary particle beam, e.g., an electron beam, with the sample and an electron-optical simulation of the paths of secondary particles, e.g., SEs and BSEs, from the sample to each of the at least two detectors. Random rotation and / or scaling of the sample may be performed to expand the training data generated by the simulation. Noise contributions from the simulation input data may be added. Various defects of the sample may be superimposed on the simulation input data in a defined manner. The measurement sample may also have all known types of defects. The test structure may have predefined defects of predefined sizes and / or shapes. Material contrast and / or topography contrast data may be known from a calibrated test structure.

[0107] The number of tuples in the measured and simulated image representations of the training dataset is 10 2 ~10 6 , preferably 5·10 2 ~3·10 5 , more preferably 10 3 ~10 5 , most preferably 3·10 3 ~3·104 may include a range of

[0108] The trained transformation model receives a tuple of measured or recorded and / or simulated image representations of the sample as input to a first transformation block. The tuple image representations include different portions of topography contrast and material contrast. The second transformation block of the trained transformation model outputs a transformed image representation of the sample having a predefined ratio of topography contrast to material contrast. Specifically, the transformed image representation output by the trained transformation model can have only topography contrast or only material contrast. The trained transformation model can also generate two transformed sample image representations from the provided tuple of sample image representations, where the first transformed image representation has only the material contrast portion and the second transformed image representation has only the topography contrast portion.

[0109] The transformation model or deep learning model can be trained to output tuples of transformed image representations, which can correspond to or be smaller than the tuples of measured image representations fed to the transformation model.

[0110] As specified above, a training dataset for training a machine learning or deep learning model can include a multiple tuple of at least two measured image representations of a sample used for training and a multiple tuple of at least two simulated image representations of the sample used for training, where the tuples of measured and simulated image representations are recorded at at least partially different solid angles relative to the sample used for training. The tuples of the training dataset can include two image representations representing a sample region of the sample used for training from two different solid angles. The size of the tuples of the training dataset can correspond to the number of detectors used to simultaneously record the image representations of the sample.

[0111] The method according to the invention may further comprise the step of recording a training data set for the transformation model.

[0112] The paradigm of machine learning (ML) is that a sufficient number of representative training data must be available to train the transformation model. This means that ML or DL ​​(deep learning) methods can typically ensure a mapping of input to output only for input data for which there exists similar training data that was taken as the basis for training the transformation model.

[0113] As a result, only the image representation of the sample contained in the training data can be simulated. For example, if the sample is a photomask, the latter can only be trained with the image representation of the photomask. Therefore, for a generally valid transformation model of a photomask, as many different photomasks as possible must be provided in the training data, whose structures or pattern elements correspond to the actual application. These include, for example, photomasks with different surface structures, such as different roughness, and masks with different material compositions due to contamination.

[0114] Since the trained transformation model is intended to be used in the context of a mask repair process, its training data should include all photomask defects that actually occur. It is also advantageous if the training data includes different intermediate states of defect repair. Photomasks with defined defects can be generated for training purposes based on simulations, which may result in long training periods.

[0115] Therefore, it may be advantageous for separate transformation models to describe and train distinct classes of photomasks (e.g., binary masks, phase-shift masks, and masks for multiple exposures). This means that properly training ML models, especially deep neural networks (DNNs), typically requires consistent training data that represents a one-to-one correspondence of input data to output data. In the case of photomasks, this means that dedicated ML models are needed for each distinct type of mask (e.g., OMOG (Opaque MoSi On Glass), COG (Chrome on Glass), PSM (Phase Shift Mask), APSM (Alternating Phase Shift Mask), etc.). As a result, first, the individual training phases are shorter and, second, the achievable accuracy of the transformed image representation can be improved. Alternatively, ML or DL ​​models can be pre-trained for masks, with distinct types of masks being defined by one or more hyperparameters.

[0116] Part of the required training data can be recorded during a test run of the repair device. This can be done using an image representation of the mask obtained during a calibration phase of the repair device. The training of the ML model can also be partially performed during the calibration phase of the repair device. To avoid the test run taking too long, the parameters of the ML model can be continuously optimized in the repair operation mode. This procedure results in incremental learning of the ML or DL ​​model. In this case, it can be advantageous to keep the original training data available for incremental learning to avoid overfitting the ML or DL ​​model to the new data.

[0117] The topography contrast and / or material contrast portions of the measurement image representation tuple can be identified sufficiently accurately at least ex-situ. For example, energy-dispersive x-ray spectroscopy (EDX) and / or energy-selective backscattered electron (EsB) detectors can be used for this purpose. In the case of programmed defects, design data can be used as a means.

[0118] Part of the training data can be implemented using tuples of simulated image representations of the specimen used for training, e.g., a photomask. Unlike measured image representations, the topography and material contrast of simulated image representations are known. As a result, they are particularly suitable as training data, since it is possible to simply verify whether the training transformation model can realistically transform the presented image representations. Also, image representations of less frequently used photomasks (e.g., tritone phase masks) can be reproducibly simulated for training. Furthermore, simulations can be used to realistically generate image representations of any type of defect at any stage of the repair process. As a result, measuring thousands of image representation tuples can be avoided. However, training an ML or DL ​​model requires a corresponding amount of training data, which can mean performing a large number of time-consuming simulations. However, these simulations can be performed cost-effectively at a central site using computer systems specifically optimized for this purpose.

[0119] Transformation models, ML models, or DL ​​models generate knowledge from experience. The model learns from examples provided to the model as training data during a learning or training phase. Therefore, appropriate values ​​can be assigned to the model's internal variables, such as parameters of a parametric mapping, so that relationships can be described in the training data. As a result, transformation models, or ML models in general, identify patterns and / or regularities in the training data rather than simply memorizing the training data during the training phase. To evaluate the generalizability of the trained model to new data, i.e., data unknown at the time of training, the quality of the learned relationships is typically evaluated based on validation data. The trained ML / DL model can be applied to image representation tuples to determine the portions of topographical and / or material contrast in the image representation tuples unknown to the ML / DL model. Thus, after the training phase is complete, a successfully trained ML / DL model, i.e., a model trained with good generalizability, can evaluate data unknown to the model, i.e., unknown image representation tuples of a sample, in terms of topographical and material contrast.

[0120] The transform model may include a generative model. The generative model may include a deep generative model. Hereinafter, a deep generative model shall be understood to be a model whose encoder and / or decoder include more than two sequential layers. A deep generative model typically includes 3 to 25 sequentially arranged encoder and / or decoder layers. However, it is also possible for the encoder and / or decoder of a generative model to include more than 100 sequential layers.

[0121] A discriminative model can generate output data from input data, and a generative model can generate output data from input data and can also reproduce the input data in the model output.

[0122] Generative models can include neural convolutional and deconvolutional networks. Neural convolutional and deconvolutional networks are commonly referred to as CNNs (Convolutional Neural Networks). When the input data to a generative model is a spatially structured image representation or image representation tuple, convolution is the appropriate operation for each individual layer in the encoder-decoder architecture. In this case, the learnable parameters are, for example, the weights of the filter masks in each individual convolutional layer. To increase the complexity of the model, the convolutional results of the layers are typically transformed into nonlinear expressions. For this purpose, the input of each neuron, determined using discrete convolutions, is converted to an output in a convolutional layer using an activation function, e.g., a sigmoid function (sig(t) = 0.5 · (1 + tanh(t / 2)) or a rectified linear unit (ReLU, f(x) = max(0,x)). The concatenation of multiple convolutional layers, each containing an activation function, enables the learning of complex patterns from the provided data for recognition tasks (known as cognition).

[0123] At least two layers of the encoder can include two or more convolutional layers and pooling layers, and / or at least two layers of the decoder can include two or more deconvolutional layers and depooling layers. In English usage, a pooling layer is alternatively referred to as a "sub-sampling layer." In English literature, an depooling layer is alternatively referred to as an "up-sampling layer." As a result of the pooling effect, the number of pixels used to represent object features in the layer is reduced, while at the same time, the feature depth or dimensionality in the encoder is increased. Feature depth is also referred to as the number of features per layer or per channel. As a result of depooling or an increase in the sampling rate (upsampling) when object data passes through the decoder, the number of pixels used to represent object features in the layer is increased.

[0124] At least two layers of the encoder are capable of determining information carrying features through a reduction in the number of pixels for representing the sample image representation.At least two layers of the encoder are capable of determining information carrying features through a reduction in the spatial dimensionality of the sample image representation.

[0125] The sample image representation may be an image representation of a photomask. The photomask image representation may include a measured and / or simulated mask image representation. The measured and / or simulated image representation may be in the form of a two-dimensional pixel matrix having grayscale values, for example.

[0126] To generate a simulated image representation of the sample, the sample is bombarded with primary particles and the interaction of the primary particles with the sample material is calculated statistically (Monte Carlo simulation). In the example of a photomask and an electron beam as the primary particle beam, this allows the number and angular distribution of SEs and BSEs leaving the sample to be determined probabilistically. The interaction of the primary particles as a function of their incident energy is well-described through modeling, and as a result, the intensity and angular distribution of secondary particles, e.g., SEs and BSEs, leaving the sample can be reproducibly calculated.

[0127] The input data input to the input layer of a DL model, or generally to a transformation model, are images or image representations from at least two of the detectors present, or from each detector. As with a typical color image, each detector can be considered a color channel. The size of the two-dimensional pixel matrix primarily determines the complexity of recording the sample image representation, or generating the training data, and transforming the at least two image representations. The size of the image representation or image must be chosen so that the DL model can at least recognize and correct the nonlocal topographic effects mentioned above. Nonlocal effects have dimensions between 5 nm and 20 nm. Based on a 200 nm x 200 nm scan area, image representations with pixel spacing in the range of 1 nm to 10 nm have been found to be advantageous. That is, sample image representations typically have matrix sizes in the range of 200 x 200 pixels to 20 x 20 pixels.

[0128] The at least one additional parameter may include image representation tuples recorded under different imaging conditions.

[0129] The portion of topographic contrast and material contrast varies depending on the conditions used to image the sample, and the imaging conditions can be adapted to the sample being analyzed.

[0130] The imaging conditions can include at least one parameter from the group consisting of the incident energy of the primary particles on the sample, the incident angle of the primary particle beam on the sample, the particle flow of the primary particles on the sample, the operating point of at least one detector for detecting secondary particles emitted from the sample, the objective setting of the particle scanning microscope (e.g., scanning electron microscope), the potential of the liner tube, the pressure to which the at least two detectors are subjected, or the temperature to which the at least two detectors are exposed. The imaging conditions can also include one or more of the system parameters listed above. The primary particle beam is typically incident on the sample at normal incidence. However, it is also possible to select a different angle of incidence, for example, adapted to the topography of the sample, for example, the topography along the edges of pattern elements of a photomask.

[0131] The incident energy of the primary particles may include an energy range of 2 eV to 50 keV, preferably 5 eV to 10 keV, more preferably 10 eV to 3 keV, and most preferably 20 eV to 1 keV. The primary particle flux may include an energy range of 1 pA to 10 nA, preferably 5 pA to 2 nA, more preferably 10 pA to 500 pA, and most preferably 20 pA to 100 pA. At least two detectors may include a detector having a current of <10 -3 mbar, preferably <3 10 -5 mbar, most preferably <10 -7 The pressure in the high vacuum chamber of a scanning electron microscope (SEM) can temporarily increase due to process gases required for localized chemical repair. As a result, the selection of available types of detectors can be limited.

[0132] At least one of the at least two detectors can include an energy filter, which can enable application of an electric field upstream of an input of the at least one detector, thereby accelerating or decelerating charged particles toward the detector.

[0133] The at least one first image representation may include an image representation of the sample area recorded at a first solid angle by at least one first detector, and the at least one second image representation of the sample may include at least one second image representation of the sample area recorded at a second solid angle by at least one second detector, the first solid angle being at least partially different from the second solid angle.

[0134] The sample area comprises the area of ​​the sample surface that is scanned by the primary particle beam.

[0135] The at least two image representations can be simultaneously recorded by at least two detectors capturing at least partially different solid angles.

[0136] The at least two detectors can be disposed in a particle optical column of a particle beam microscope. The particle beam microscope can include a scanning electron microscope (SEM), and the particle optical column can include an electron optical column. The at least one first detector can be disposed at an output of the electron optical column, and the at least one second detector can be disposed in the electron optical column, i.e., as an in-lens detector.

[0137] The at least two recorded image representations of the sample can be recorded using at least one first detector and at least one second detector that detect secondary electrons (SE) and backscattered electrons (BSE), and the SE / BSE ratio of the at least one first detector and the SE / BSE ratio of the at least one second detector can be different.

[0138] The at least one first detector and the at least one second detector can detect different BSE distributions or different portions of the BSE distribution. The BSE portions of the at least one first detector and the at least one second detector differ in at least one parameter from the group consisting of the BSE energy distribution and the BSE solid angle distribution. The local material composition of the sample can be determined from the at least two different BSE distributions. Using the method of the present invention, it is possible to determine relative changes or differences in the material composition of the sample perpendicular to the sample surface, i.e., in the z-direction, and / or when the sample is etched as the primary particle beam is scanned over the sample. Thus, the method of the present invention is intended to determine the material composition or local changes in material composition of the scanned sample area. However, the described method is not element-specific, i.e., the elements of the scanned sample area are not obtained.

[0139] The primary particle beam can include at least one particle type from the group consisting of electrons, ions, x-ray quanta, gamma quanta, or photons in the extreme ultraviolet wavelength range.

[0140] The particles emitted from the sample can include at least one member of the group consisting of electrons, ions, x-ray quanta, and photons in the ultraviolet wavelength range, the deep ultraviolet wavelength range, or the extreme ultraviolet wavelength range.

[0141] The specimen may include at least one element from the group consisting of a photomask, a stamp for nanoimprint lithography, a wafer, a Micro-Electro-Mechanical System (MEMS), a Nano-Electro-Mechanical System (NEMS), or a Photonic Integrated Circuit (PIC).

[0142] Recording the at least two image representations of the sample can include irradiating the sample region with a primary focused particle beam to separate secondary particles leaving the sample at different solid angles. Preferably, recording the at least two image representations of the sample can include irradiating the sample region with a primary focused electron beam to separate SEs and BSEs leaving the sample at different solid angles.

[0143] During, or before and after, recording of the at least two image representations by irradiating the sample with the primary focused particle beam, at least one precursor gas can be supplied to the sample area scanned by the primary focused particle beam, where the at least one precursor gas can comprise at least one element of the group consisting of an etching gas, a deposition gas, or an additive gas.

[0144] The method according to the invention may further comprise repairing at least one defect in the specimen using a local chemical reaction induced by the primary focused particle beam, which in the case of a focused electron beam as the focused particle beam comprises EBIE (electron beam induced etching) or EBID (electron beam induced deposition), depending on the precursor gas used.

[0145] Determining the topographical contrast and / or material contrast of the image representation may include determining an image representation that is substantially free of topographical contrast portions.

[0146] A sample image representation that does not include topography contrast and includes only material contrast is maximally sensitive in detecting changes in the local material composition of the sample, and therefore is optimal as a stop signal for the local chemical repair process.

[0147] The term "substantially" here, as elsewhere in this application, means representation of a measured variable within customary error limits of the prior art metrology used to measure that variable.

[0148] A computer program may include instructions for performing the method steps of any of the above aspects.

[0149] According to a further exemplary embodiment, the problem addressed by the present invention is solved by a device as set forth in claim 17. In one embodiment, a device for determining topographical and / or material contrast of a sample comprises: (a) means for providing at least two image representations of the sample recorded at least partially at different solid angles relative to the sample; and (b) means for determining the topographical and / or material contrast of the sample based on the at least two image representations.

[0150] The means for determining may be configured to apply a separation model to the at least two image representations to determine the topographical contrast and / or material contrast of the sample.

[0151] The device may further include at least one first detector and at least one second detector to provide at least two image representations, each detecting secondary electrons (SEs) and backscattered electrons (BSEs), and the SE / BSE ratio of the at least one first detector and the SE / BSE ratio of the at least one second detector are different from each other. The trajectories or paths of the SEs and / or BSEs may be influenced by an objective lens that focuses the primary particle beam. In this case, the trajectories of the SEs and / or BSEs are highly dependent on the kinetic energy of the secondary particles.

[0152] The first detector can be located in the electron optical column of the device, and / or the second detector can be located in the electron optical column of the device, and the first detector can be located closer to the sample than the second detector, meaning that both detectors can be implemented as in-lens detectors.

[0153] The solid angles of at least two detectors recording the two image representations may overlap or coincide. Thus, one detector may partially obscure the other. It is also possible for components of the repair device to limit the viewing or solid angle of one or both detectors relative to the sample. Partial obscuration may occur, particularly in the case of detectors integrated into the particle-optical column of the repair device. The solid angle through which each detector "sees" or focuses may also depend on the system settings of the repair device, for example, the magnitude of the electric and magnetic fields generated by the repair device's objective lens. It is also possible to record more than one image representation using one detector, with each image representation covering a different portion of the detector's detection area.

[0154] The second detector can include an energy filter, which generates an adjustable electric field upstream of the detector input, thereby accelerating or decelerating charged secondary particles in a defined manner toward the detector. The energy filter of the second detector allows for the substantial separation of SEs and BSEs.

[0155] The energy filter can be designed to generate a potential in the range of ±0.05 kV, preferably ±0.2 kV, more preferably ±0.5 kV, and most preferably ±2.0 kV. The energy filter can have a filter width of <400 eV, preferably <200 eV, and most preferably <100 eV, assuming a pass energy of 50 eV, preferably 200 eV, more preferably 500 eV, and most preferably 2000 eV.

[0156] The device may be configured to perform the method steps according to any of the above aspects.

[0157] The means for applying the separation model may include a dedicated hardware component for analyzing the at least two image representations of the sample, the dedicated hardware component may include at least one element from the group consisting of an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a programmable logic circuit (PLD, programmable logic device), and a graphics processor unit (GPU).

[0158] Many machine learning methods can be optimized for the use of specific computing units to significantly accelerate implementation. In this case, graphics processor units (GPUs) have proven particularly advantageous for DNNs. The size of the image representation that can be computed is typically limited by the available main memory of the GPU. However, the typical image area (FOV, field of view) of the image representation of a photomask may be significantly larger than can be handled by DNNs using current GPUs. This problem can be solved by dividing the image representation area to be computed into sub-areas. In this case, sub-areas of the image representation are computed separately. This applies both to training and to applying a transformation model to the image representation to be analyzed or determined. The computed sub-image representations are then combined to form the overall image or overall image representation.

[0159] The device may further include means for training the transformation model with a training dataset.

[0160] The device may further include a liner tube through which the potential of the electrons in the column may be altered by a predefined value.

[0161] Finally, the device may include a gas supply system and at least three gas storage containers. At least one etching gas, at least one deposition gas, and at least one additive gas may be stored in the gas storage containers. The gas supply system is designed to supply an adjustable amount of precursor gas stored in the gas storage containers to a sample region scanned by the primary particle beam.

[0162] In the following detailed description, presently preferred exemplary embodiments of the invention are described with reference to the following drawings: [Brief explanation of the drawings]

[0163] [Figure 1] FIG. 2 is a diagram showing a schematic energy spectrum of electrons generated by an electron beam incident on a sample. [Figure 2] FIG. 1 shows the solid angle distribution of backscattered electrons (BSE) of five different elements with different atomic numbers for two kinetic energies of the primary electron beam. [Figure 3] FIG. 1 shows a schematic representation of a mask section having areas of high topographic contrast and areas of low material contrast. [Figure 4] 4 shows the mask section of FIG. 3 in an image representation characterized by areas of low topographic contrast and areas of high material contrast. [Figure 5] FIG. 1 illustrates a coordinate system spanned by the contributions of topographic contrast and material contrast of an image representation. [Figure 6] FIG. 6 shows the coordinate system of FIG. 5 on which the image representations of FIGS. 3 and 4 are plotted. [Figure 7] FIG. 10 shows a schematic diagram of the reconstruction of material contrast and topography contrast from at least two specimen images using a separation model. [Figure 8] FIG. 1 shows a schematic cross section of a column of a scanning electron microscope, in which two detectors are placed at different distances from the sample. [Figure 9]9 is a schematic diagram showing a cross section of the solid angle distribution of secondary particles emitted from the sample and reaching the two detectors of FIG. 8. FIG. [Figure 10] 9 shows in the upper partial image a schematic cross-section of a mask having a ruthenium substrate and a tantalum pattern element, and in the lower partial image the signals of the two detectors of FIG. 8 of a simulated line scan along the edge of the pattern element, and a signal reconstructed from the signals of the two detectors and containing substantially only material signal information. [Figure 11] 11 shows the signal of FIG. 10 reconstructed from the signals of the two detectors of FIG. 8 as a vector addition of the two signals. [Figure 12] 9 is a flow diagram illustrating pixel-by-pixel reconstruction of material and / or topography contrast signal portions from at least two determined intensities of the detector of FIG. 8; [Figure 13] 1 is a flow diagram illustrating the optimization of parameters of an empirical model using a calibrated test sample. [Figure 14] FIG. 1 illustrates schematically a general-purpose deep learning model in the form of a U-Net architecture. [Figure 15] FIG. 15 illustrates the U-Net architecture of FIG. 14 with input data for solving the problem of the present application and output data shown in the form of material contrast images and topography contrast images. [Figure 16] FIG. 1 is a schematic diagram of a generative adversarial network (GAN) for generating SEM images for training. [Figure 17] FIG. 10 is a schematic diagram of a GAN adapting parameters of a simulation tool to perform Monte Carlo simulations to generate SEM images for training a generative deep learning model. [Figure 18] FIG. 1 illustrates a photomask test structure with defined defects for generating training data. [Figure 19]FIG. 1 shows LS (line width and line spacing) with predefined defects for generating training data. [Figure 20] FIG. 10 is a diagram showing a contact hole structure for generating training data. [Figure 21] FIG. 21 shows the contact hole structure of FIG. 20 with a defined defect. [Figure 22] FIG. 22 is a diagram showing a table summarizing the variation ranges of the defined defects in FIGS. 18 to 21. [Figure 23] FIG. 2 is a diagram illustrating the operation of a first embodiment of a machine learning model. [Figure 24] FIG. 10 illustrates a schematic diagram of the operation of a second exemplary embodiment of a machine learning model (ML model). [Figure 25] FIG. 10 illustrates schematically the operation of a third exemplary embodiment of an ML model. [Figure 26] FIG. 24 illustrates a schematic diagram of the training of the ML model of FIG. 23. [Figure 27] 1 shows a schematic cross-section of a device capable of carrying out the method and localized chemical sample remediation process according to the invention; [Figure 28] FIG. 1 shows a flow diagram of a method for determining topography and / or material contrast portions of an image representation of a sample. DETAILED DESCRIPTION OF THE INVENTION

[0164] In the following, a currently preferred embodiment of a method and a device according to the present invention for determining topographical and / or material contrast features of an image representation of a sample will be described in more detail. However, the method according to the present invention is not limited to photomasks, such as the sample examples described below. Rather, the method and the device according to the present invention can be used to determine topographical and material contrast features of any microstructured sample. The device according to the present invention will be described below based on the example of an improved scanning electron microscope. However, the device according to the present invention is not limited to the described exemplary embodiment. In addition to electrons, other charged particles and / or high-energy photons can also be used to determine topographical and material contrast features of a sample. Furthermore, the method according to the present invention will be described below based on the example of two detectors detecting secondary particles from different solid angles. However, the described method is not limited to the use of two detectors. Rather, the method can be used with one detector, and three or more detectors are advantageous. In addition to the signals emitted by the multiple detectors, the method according to the invention can also, in the case of conductive samples, make use of the detection current flowing through the sample to process the charges present on the sample at the same time so as to obtain more or more precise information about the properties of the sample.

[0165] In the following, the interaction of particle radiation with a sample is illustrated based on the example of the action of an electron beam on a sample. The incidence of a focused electron beam that scans an area of ​​the sample is described in detail. This area of ​​the sample is called the field of view (FOV). During the irradiation of the sample by electrons from the electron beam, the electrons interact with the sample. The process of interaction of the incident electron beam with the atoms of the sample generates free electrons in the sample. Some of the electrons generated in the interaction process can leave the sample surface and can be detected by one or more detectors and used to generate an SEM (Scanning Electron Microscope) image of the sample surface.

[0166] Diagram 100 of Figure 1 shows the energy spectrum of electrons generated by an electron beam at a sample. This figure can be taken from the book "Scanning Electron Microscopy" by L. Reimer. The energy spectrum of electrons emitted by a sample is divided into two main groups: low-energy electrons with a kinetic energy of up to 50 eV (electron volts) are called secondary electrons (SE); all other generated electrons with a spectral energy distribution ranging from 50 eV to substantially the kinetic energy of the electrons of the incident electron beam (E = e·U, where E is the incident energy of the electrons, e is the elementary charge, and U is the potential difference across which the electrons of the electron beam travel due to their acceleration) are called backscattered electrons (BSE).

[0167] If the sample surface has no surface charge, the secondary electron energy spectrum exhibits distinct material- and / or topography-specific peaks 110 (SE peaks 110) in the range of several volts. In the energy range from about 50 eV to about 2 keV, material-specific peaks can also occur in the spectrum of backscattered electrons caused by Auger electrons (AE). At the upper end of the backscattered electron energy spectrum, there is an elastic peak 120 (BSE peak 120) caused by electrons reflected from the sample surface at essentially the kinetic energy of the incident electrons. Below this peak 120, there is an adjacent so-called LLE (Low Loss Electron) range, which includes backscattered electrons, whose energies are typically 10 eV to 100 eV lower than the kinetic energy of the incident electrons. The LLE range also includes the range of plasma excitation (plasmon loss), and as a result, a relatively small number of backscattered electrons leave the sample surface in this spectral range.

[0168] Figure 100 reveals that a detector without an energy filter always detects both SEs and BSEs. It is also clear from Figure 1 that low-energy SEs are much more numerous than BSEs, which exhibit greater kinetic energy. SEs are primarily emitted at locally elevated portions of the sample, for example along edges. Therefore, they appear brighter than the sample surface in a sample image representation. SEs therefore primarily contribute to the topographic contrast of a sample image representation recorded using a focused electron beam.

[0169] The generation of an electric field of corresponding polarity allows the SEs to be decelerated so much that they cannot reach the detector. As a result, the intensity of the SEM image filtered in this way is significantly reduced, and in fact many SEs do not contribute to the image formation. However, the resulting image representation still contains a broad spectrum of BSEs and a substantial elastic scattering peak 120.

[0170] Diagram 200 in Figure 2 shows the different backscattering coefficients of BSE that various elements generate for two different kinetic energies of electrons in the incident electron beam. The incident primary electrons (PE) are elastically scattered in the nuclear fields. The greater the deflection of the primary electrons at the nuclei, the higher their positive charge. Figure 2 shows the distribution of beryllium (Be), aluminum (Al), copper (Cu), silver (Ag), and gold (Au) elements for two primary energies. The proportion of substantially elastic backscattered electrons, i.e., the height of the BSE peak 120, depends on the material of the sample irradiated by the primary focused electron beam.

[0171] As shown in Figure 2, the angular distribution of the radiation intensity of the BSE follows approximately a cosine distribution. The present application exploits this regularity to generate image representations of samples with a predefined ratio of topographic contrast to material contrast. Specifically, this regularity is used to generate image representations of samples with substantially only material contrast. The angular distribution of the BSE changes little as a function of the kinetic energy of the PE.

[0172] In the following, an empirical model is described as a first exemplary embodiment for determining the topographical and material contrast contributions of a sample image representation.

[0173] Image representation 395 in FIG. 3 schematically illustrates a section of a sample 300. Sample 300 may be a photolithography mask 300. Three pattern elements 320, 330, and 340 are disposed on a mask substrate 310 of the exemplary mask 300. Pattern element 320 includes two interconnected rectangular structures 322 and 324 interconnected on one side. Second pattern element 330 has a circular face, and third pattern element 340 has a triangular structure. Portion 322 of first pattern element 320 and triangular pattern element 340 include a first material 350. Additionally, round pattern element 330 and portion 324 of first pattern element 320 include a second material composition 360. 3, first material 350 causes a lower intensity in image representation 395, which is depicted in a lighter shade of gray, compared to second material 360, which is depicted in a slightly darker shade of gray in the image representation of mask 300. A conclusion that can be drawn from this is that material 360 has a higher atomic number than material 350 of mask 300. First portion 322 of first pattern element 320 and second pattern element 330 can be deposited on mask substrate 310 using, for example, a particle beam induced deposition process to create missing pattern element 330 and / or partially missing pattern element 324.

[0174] The wide dark edges 380 of the pattern elements 320, 330, 340 of the mask 300 make it clear that the image representation 395 has a large topography contrast contribution 380 in addition to the material contrast 350, 360. The relatively large topography contrast contribution 380 of the image representation 395 indicates that the detector that detected the SEs and BSEs, which are considered the basis for generating the image representation 395, must have a larger angle with respect to the beam axis of the primary electron beam.

[0175] It should be noted that in Figure 3, as in subsequent Figures 4-6, the image representation does not represent measurement data of the photomask, but merely serves to illustrate the principles of the present application. Also, it should be noted that in the exemplary Figures 3-6, light and dark correspond to dark and light, respectively, in subsequent figures. This exchange is made herein merely because it is easier to illustrate on a white background.

[0176] Image representation 495 of Figure 4 is a repeat of image representation 395 of Figure 3, but differs in the material contrast 450, 460 and the ratio of topography contrast 480. Edge 480 or edges 480 are significantly less visible in image representation 495 compared to image representation 395 of Figure 3. Conversely, the difference in gray levels 450, 460 is again significantly more pronounced compared to image representation 395 of Figure 3. The SE / BSE ratios of image representations 395 and 495 are clearly different. This means that the detector angle relative to the beam axis of the primary electron beam is relatively smaller when image representation 495 is recorded compared to when image representation 395 is detected.

[0177] A procedure in which the two image representations 395 and 495 recorded at different angles or solid angles relative to the beam axis of the primary beam are considered in combination, i.e. linked together in an appropriate manner, makes it possible to eliminate the contribution of topographic contrast at the edges 380, 480 of the pattern elements 320, 330, 340 of the mask 300, i.e. at the image representations 395, 495. As a result, it is possible to generate image representations that have substantially only material contrast.

[0178] Diagram 595 in FIG. 5 shows a coordinate system 500 spanned by the topography contrast component and material contrast component of the sample image representation. This means that the axes of coordinate system 500 completely separate the topography contrast component or topography contrast intensity from the material contrast component or material contrast intensity. In the example coordinate system 500 of FIG. 5, topography contrast is plotted on the abscissa axis, and material contrast is plotted on the ordinate axis. Subimage 520 shows a section of mask 300 where only the topography contrast intensity 585 of pattern elements 320, 330, and 340 of mask 300 contributes to generating the image representation. Contributions from different material compositions 350, 360 of pattern elements 320, 330, and 340 of mask 300, indicated by different gray levels in image representation 395 of FIG. 3, have disappeared.

[0179] At this point, attention is again directed to the general characteristics of image representations 395, 495, 510, and 520. In an actual image representation of pattern elements 320, 330, 340 present on mask substrate 310, the edges of the pattern elements will appear lighter compared to the surface of mask substrate 310 and the surfaces of pattern elements 320, 330, 340.

[0180] Partial image 510 shows a section of mask 300 where only the material contrast intensities 550, 560 of mask 300 contribute to generating the image representation. Edges 580 of pattern elements 320, 330, and 340 are no longer highlighted by intensity variations in partial image 510. The portions of mask 300 where the topography contributes to the intensity distribution of the image representation in partial image 510 are no longer visible in the partial image. This means that partial image 510 has no topographic contrast contribution. Conversely, the different material compositions 550, 560 of pattern elements 320, 330, and 340 of mask 300 are optimally visualized in partial image 510.

[0181] Partial images 510 and 520 show sections of mask 300 where only the topography of mask 300 (partial image 520) or the material composition of mask 300 (partial image 510) contributes to the image representation of that mask. Measuring partial images 510 and 520 is difficult. Because the contribution of high-energy BSEs to image generation cannot be eliminated, it is not possible to select the SEs emitted from mask 300 that primarily convey topographic information. In partial view 520 of Fig. 595, the topography of the sample, i.e., mask 300, is shown as an image representation of edges 380, and therefore the edges of pattern elements 320, 330, and 340. However, edges 380 of pattern elements 320, 330, and 340 are the same as the height profile of mask 300, and therefore partial image 520 shows the topography or height profile or profile of height variations of mask 300.

[0182] The material contrast 450, 460, 550, 560 of the mask 300, or more generally of the sample 300, during the creation of an image representation can be enhanced using a detector for detecting secondary particles positioned around the primary particle beam. Furthermore, by creating an electric field with a 50 eV potential difference of corresponding polarity upstream of the detector, SEs with kinetic energies less than 50 eV can be prevented from passing into the detector. A stronger electric field also prevents some of the low-energy BSEs from entering the detector. However, it is technically impossible to split the spectrum of secondary particles so that only peak 120 BSEs reach the detector. The broad background of partially inelastically scattered BSEs prevents the recording of a sample image representation that reproduces only the material contrast intensity. Furthermore, BSEs with decreasing polar angles experience increasingly increasing shading effects that result in the contribution of topographic contrast in the corresponding image representation.

[0183] However, by performing corresponding simulations, it is possible to generate particle images 510 and 520 that represent pure topographic contrast and pure material contrast. In Monte Carlo simulations, the interaction of primary particles (PEs) with the material and material composition of the mask substrate 310, as well as the topography of the pattern elements 320, 330, and 340, are also statistically simulated. From the results of these simulations, it is possible to generate images with an intensity distribution contributed only by SEs that convey topographic contrast information. Furthermore, from the simulation data, it is possible to generate an image representation in which only the BSEs of peak 120 contribute to the generation of the image. In simulations, the angular distribution of the BSEs that should contribute to the image generation can be selected in a simple manner. These complex simulations can be performed by a computing unit specifically designed for this purpose.

[0184] Diagram 695 in Figure 6 again illustrates the coordinate system of Figure 5. Image representations 395 and 495 of sections of mask 300 are now plotted in this coordinate system 500. As explained in the context of Figures 3 and 4, image representation 395 has a large contribution of mask topography contrast 380 to the intensity distribution of image representation 395. Therefore, image representation 395 lies near an axis representing the contribution of topography contrast to the intensity distribution of image representation 395. In contrast, in image representation 495, material contrast 450, 460 primarily contributes to the intensity distribution of image representation 495. Therefore, image representation 495 lies near an axis of coordinate system 500 representing the contribution of material contrast to imaging.

[0185] Sample topography contrast and sample material contributions can be extracted from two image representations recorded at different solid angles representing a sample 300, e.g., a photomask 300. Photomask 300 has pattern elements 320 and 350 and therefore has a non-planar surface. For a sample with a planar surface, different solid angles will involve different angles relative to the primary beam.

[0186] In coordinate system 500, image representations A, 395 and B, 495 are represented by a vector e on the topography contrast part axis (T axis) as follows: T and the vector e on the material contrast axis (M axis) M It can be expressed as a linear combination of A=a1·e T +a2·e M and B=b1·e T +b2·e M Or, in matrix notation,

[0187]

number

[0188] After the parameters or coefficients a1, a2, b1, and b2 have been determined, the contributions of the topography contrast and material contrast of the image representations A or 395 and B or 495 can be identified. The two image representations A or 395 and B or 495 provide four variables A that allow for defining the four parameters or coefficients a1, a2, b1, and b2 of the transfer matrix T. T , A M , B T and B M is obtained.

[0189] Alternatively, image representations A or 395 and B or 495 are angled on the T axis or M axis of the coordinate system.

number

number

[0190]

number

[0191] Angle of rotation

number

number

[0192] As explained above, one of the objectives of the present application is to use the change in the material contrast contribution of the sample image representation 395, 495 to derive a signal for stopping the local chemical repair process of the sample defect. For this purpose, it is advantageous to use an image representation 495 with a larger material contrast contribution. The signal change upon detection of the transition of the primary particle beam from the first sample layer to the second sample layer with a different material composition is larger than in the second image representation 395. This allows for more accurate identification of the time point for stopping the local chemical reaction. An image representation 495 that satisfies this prerequisite is an image representation recorded by a detector with a large portion of the BSE reflected substantially antiparallel to the primary particle beam. If the difference in the material contrast portion of the image representation 495 is not sufficient for this purpose, the method described herein allows for the generation of a sample image representation that essentially shows only the material contrast portion.

[0193] The above outline of the separation into material and topography contrast regions using linear algebra is idealized. It would be preferable to use more complex empirical or transformation models for this purpose.

[0194] The theoretical principles of this application are briefly outlined below: The basic premise is that for an uncharged sample on each detector, the image representation of the sample or its image signal is contributed by a combination of (a) topographic contrast (called a z-map or "height map") and (b) material contrast (element or material composition, crystalline structure, etc.).

[0195] Diagram 795 in Figure 7 illustrates the process of separating material contrast and topography contrast portions in an image representation using a separation model 700. Separation model 700 may include empirical models and / or transformation models. Images 720, 730, and optionally 740 are provided as input data 710 to separation model 700, which converts these images into a material contrast image 760 and a topography contrast or height profile image 770, which it provides at its output 750.

[0196] To record images 720, 730, 740, a primary electron beam can be scanned over the sample, and secondary particles emitted from the sample can be recorded by two or more detectors. When a focused electron beam, as an example of a charged focused particle beam, is scanned over the sample, electrons, as an example of secondary particles, with different energy and angular distributions are generated with a certain probability at each scanning point. These distributions Ψ, or emission distribution Ψ, depend on the local topography and material composition of the sample and are energy dependent, as follows: Ψ(E,φ)

[0197] Diagram 895 in FIG. 8 shows a schematic cross section of a column of a scanning electron microscope (SEM) 800. The SEM 800 includes an electron source 810 and a connection 820 for creating a reduced pressure or vacuum within the column of the SEM 800. The SEM 800 also includes a gas supply system 830 for supplying a precursor gas onto a sample 890 or photomask 890. An objective 840 or objective lens 840 focuses an electron beam or primary electron beam (not shown in FIG. 8 ) onto the sample 890. A portion of secondary particles generated by the sample 890 are recorded by a first detector 850 and a second detector 870. Mainly secondary electrons (SEs) 860 reach the first detector 850 via a trajectory 865. Mainly backscattered electrons (BSEs) 880 reach the second detector 870 via a path 875.

[0198] As shown in FIGS. 8 and 9, each detector 850, 870 has an acceptance function D1·2 Different sections of these distributions according to (E, φ) are used to detect secondary particles, namely SE860 and BSE880, from dedicated angular ranges as a function of the kinetic energy of the secondary particles. In this case, D 1.2 (E, φ) can only take the value 0 or 1, i.e., the value 0 if no secondary particle is incident on the corresponding detector 850, 870, and the value 1 if a secondary particle is incident on the detector 850, 870. In this case, the signals I 1.2 accepts function D 1.2 It is the integral of (E,φ) multiplied by the emission distribution Ψ(E,φ).

[0199]

number

[0200] Diagram 995 in Figure 9 shows the detector acceptance ranges of SE 860 and BSE 880 for the arrangement of two detectors 850 and 870 of Figure 8. Circle 910 corresponds to the angular distribution of SE 860, and highlighted areas 920 and 930 show the angular acceptance ranges of detectors 850 and 870. Circle 950 shows the angular distribution of BSE 880, and section 960 (970) shows the angular acceptance range of detector 850 (870). The detector acceptance function D 1.2(E,φ), i.e., segments 920, 930, 960, and 970, depend on the incident energy of the primary electrons on the sample, the working distance (i.e., the distance between the sample 890 and the objective lens 840), the convergence angle of the primary electron beam, and of course the positions of the detectors 850 and 870 within the column of the SEM 800. The first detector 850, which is at a smaller distance from the sample 890 than the second detector 870, "sees" secondary particles emitted by the sample 890 at a larger polar angle compared to the second detector 870. The detector acceptance function D 1.2 (E, φ) can be obtained using electron optical simulation.

[0201] 9 shows that detectors 850, 870 record different portions of the BSE 880 and SE 860, as well as different polar angle ranges 920, 930, 960, 970. Thus, with two or more detectors 850, 870, it is possible to approximate the local emission distribution Ψ(E,φ), thereby obtaining information about the topography and material composition of the sample 890.

[0202] In the examples described below, the emission distribution is not determined, but instead, an empirical model is adapted to the sample being analyzed and then parameterized or a trained transformation model is used to separate the material contrast and topography contrast portions of one or more image representations.

[0203] We first describe the process of determining the local material signal, i.e., a local image representation of a sample 890 that has substantial material contrast. For a binary sample 890, such as a binary photomask, one only needs to distinguish between the material of the absorbing pattern elements and the mask substrate.

[0204] In a first embodiment, in order to identify the local material signal using the detectors 850 and 870 at a point, the secondary particles emitted by the sample 890, or in the described embodiment the mask 890, are recorded, i.e. the detector signals are recorded pixel by pixel to determine the local material signal of the mask 890. The signals from the two detectors 850, 870, in the general case k detectors (k≧2), recorded at a point (i,j) of the sample 890 are processed with a function to generate the material signal S(i,j) of the sample 890 at that location (i,j). In the simplest form, the different detector signals I k However, the k detector signals I k Other functions based on k may also be used. For example, a new signal feature of the k detectors may be generated as a product of the individual detector signals.

[0205] The empirical model estimates the signals I from detectors 850 and 870. k , then the material signal, material contrast signal or material contrast signal function S(i,j) of the sample 890 at location (i,j) has the following form (k=2): S(i,j)=Σ k a k I k (i,j)+a0 where a0 is a normalization constant.

[0206] The coefficients of this function or parameters of the empirical model can be calibrated so that the function or empirical model takes on a value of 0 for the absorbing material of the pattern element and a value of 1 for the material of the mask substrate, or vice versa.

[0207] An empirical model for a particular material combination and particular operating point of sample 890 can be determined by the following procedure.

[0208] In the first step, the absorption test structure T A and mask substrate material T SThe recordings are made by each of the k detectors at different locations of a calibrated test sample that includes at least one calibrated test structure T A may include pattern elements with different dimensions. A may also include, for example, line structures with different spacing, holes or contact holes and islands of different sizes, and programmed defects, examples of which are shown in Figures 18-22.

[0209] At least one calibrated test structure T A For a calibrated test specimen having a material information M(i,j), the material information M(i,j) is known based on the calibration process that was performed. This is because M(i,j) is the material information of the test structure T A and the mask substrate T S This means that the material has a value of 1 in the region of

[0210] However, it is also possible to generate an image representation of the calibrated test structure by carrying out a simulation based on the calibrated test structure. As already explained above, the simulation first involves the combination of the primary beam with at least one test structure T of the calibrated test specimen. A and secondly, performing an electron optical simulation of the trajectories of the secondary particles up to the two detectors 850, 870 of the SEM 800, in the general case up to each of the k detectors of the detector configuration.

[0211] Diagram 1095 in FIG. 10 shows a schematic cross-section of sample 1000 in upper image 1005. Sample 1000 includes a ruthenium (Ru) substrate 1010 and a pattern element 1020 containing tantalum (Ta). In the simulation shown in lower image 1055, a line scan is performed on the Ru substrate 1010 along the Ta edge 1030 with a primary electron incident energy of 400 eV. In FIG. 10, the signal or intensity of the first (second) detector 850 (870) is shown by solid line 1050 and dashed line 1070, respectively. Curve 1080 represents a material contrast signal 1080 reconstructed from the signals of the first detector 850 and the second detector 870. The reconstructed material contrast signal 1080 follows the tantalum-to-ruthenium material transition due to the reduced influence of the topography of sample 1000. The horizontal dotted lines 1075 and 1085 are the normalization of the material contrast of the Ru substrate 1010 and the Ta pattern element 1020. As shown by the vertical dotted line 1035, a determination can be made very accurately from the reconstructed signal 1080 which corresponds to a change in topography, but which in the example of Figure 10 coincides with the location of the change in material composition Ta -> Ru.

[0212] Diagram 1195 in FIG. 11 shows a reconstruction of the material contrast signal 1080 of FIG. 10 using the signals I of detectors 850 and 870. k 10 and 11, the reconstructed material contrast signal 1080 is the result of subtracting 55% of the signal from the first detector 850 from the signal from the second detector 870. This is represented by point 110 in diagram 1195.

[0213] Alternatively, one of the k detectors for recording the test image can be equipped with a screening grid. Ideally, this is the detector located at the farthest distance from the sample 890. In the embodiment shown in FIG. 8, this is the second detector 870. This allows an additional EsB (Energy-selective Backscattered) image to be recorded at each point with the activated screening grid. In this case, this image can be considered a good approximation of the material signal M(i,j). However, once the parameters of the empirical model are defined, there is no need to use a screening grid. As a result, it is possible to avoid instabilities in the primary electron beam that may be caused by the high voltage required for the screening grid.

[0214] The second step involves minimizing the absolute value of the difference between the material signal S(i,j) of the sample 890 and the material contrast signal or material information M(i,j) of the calibrated test structure by changing the parameters of the material contrast signal or material signal function S(i,j) of the empirical model. Those skilled in the art can do this by trial and error in simple cases. In arbitrary or general cases, the parameters of the empirical model can be found by using known optimization methods, as follows: min|S(i,j)-M(i,j)|

[0215] By minimizing the absolute value of the difference between the topography signal H(i,j) of the sample 890 and the topography information T(i,j) of the calibrated test structure, min|H(i,j)-T(i,j)| A height profile or height profile map of the sample 890, 1000 can be determined that is substantially free of material contrast effects.

[0216] 12 shows a flow chart 1200 summarizing the pixel-by-pixel reconstruction of topographic and material contrast contributions from at least two image representations. The method starts at 1210. In a first step 1220, an intensity I is measured at a point (i,j) on the sample 890, 1000 by each of at least two detectors observing the sample at at least partially different solid angles. k is determined. The determination can be made by measurement and / or simulation. Next, step 1230 calculates the determined intensity I k Step 1240 then involves reconstructing the local material contrast portion M(i,j) or the local material information and / or the local topography contrast portion T(i,j) at point (i,j) of the sample 890, 1000 based on the established material signal function S(i,j) and / or topography signal function H(i,j). For the reconstruction, the separation model 700 of FIG. 7 in the form of an empirical model can be used. The process of defining the parameters of the empirical model or the material signal function S(i,j) and / or the topography signal function H(i,j) is described in FIG. 13 below. The method ends at 1250.

[0217] 13 illustrates a method for determining parameters of a separation model 700 in the form of an empirical model. The method begins at 1310. At step 1320, the intensity I of at least one test structure of a calibrated test specimen is calculated. k are recorded by at least two detectors 850, 870 observing the test structure of the calibrated test specimen at least partially at different solid angles. The calibrated test specimen is distinguished by its material contrast information M(i,j) and / or topography contrast information T(i,j), or material contrast distribution and / or topography contrast distribution, in the case of a calibrated test specimen, being known.

[0218] In step 1330, a reconstructed material contrast signal function S(i,j) is calculated based on the set material contrast signal function S(i,j) and / or the set topography contrast signal function H(i,j) of the calibrated test structure and the known material contrast information M(i,j) and / or the topography contrast information T(i,j). R (i,j) and / or the reconstructed topographic contrast signal function H R (i,j) are determined or calculated. Step 1340 then determines the parameters of the empirical model, i.e., the reconstructed material contrast signal function S R (i,j) and / or the reconstructed topographic contrast signal function H R By changing (i,j), we reconstruct the material contrast signal function S R This includes minimizing the absolute value of the difference between (i,j) and the known material contrast information M(i,j) and / or the reconstructed topography contrast signal function HR(i,j) and the known topography contrast information T(i,j).

[0219] Decision block 1350 involves checking whether the residual difference is less than a predefined threshold. If the residual difference is greater than the predefined threshold, the method returns to block 1330 and determines the reconstructed material contrast signal function S R (i,j) and / or the reconstructed topographic contrast signal function H R Using the current parameters of (i,j), we reconstruct a new material contrast signal function S R (i,j) and / or the new topographic contrast signal function H R (i,j). The method then proceeds to block 1340. If the condition of decision block 1350 is met, the method 1300 proceeds to block 1360, where the provisional parameters of the empirical model are selected as the best possible parameters for analyzing the sample 890, 1000, i.e., the reconstructed material contrast signal function S R (i,j) as and / or the reconstructed topographic contrast signal function H R(i,j). The method ends at block 1370.

[0220] The signal I from the two or more detectors 850, 870 k In the aforementioned particular case of the parameters of the empirical model based on pixel-wise determination of , it is not possible to take into account non-local effects in the signal. As a result, the topography part H(i,j) is calculated as the signal I of the detectors 850, 870. k It is not possible to completely separate the material contrast portion S(i,j) of the sample 890, 1000 from the point I(i,j). k The detector signals at (i-dx, j-dy) must be considered together. The maximum range (dx, dy) that must be considered here depends on the interaction area between the electrons of the primary electron beam and the sample 890, 1000, the topography of the sample 890, 1000, and the electron optics of the detection path (by the acceptance function Dk(E, φ)). The sample area considered, where non-local effects are present, typically has linear dimensions in the range of 5 nm to 20 nm.

[0221] This means that the signal I at each point of the sample 890, 1000 k This means that it is often not sufficient to simply record the image signals I of the detectors 850, 870 using a scan of the primary electron beam over the area around each point (i,j). k For each of at least two detectors 850, 870, in the general case k detectors, an image I k This relationship also applies to recording training data for training a transformation model or an ML model or a DL model.

[0222] Recorded image data I k one or more appropriate convolution kernels w using l By performing a convolution operation with k,l (i,j) is generated as follows:

[0223]

number

[0224] For example, known image filters such as the Sobel filter, Prewitt filter, Laplacian filter, Marr-Hildreth filter, Gaussian filter, or Sharpen filter can be calculated using the convolution kernel w l Further types of filters are possible as well.

[0225] In order for the convolution operation to have a corresponding effect on the separation of material contrast portions and topography contrast portions or material contrast signals and topography contrast signals, i.e., the effect of separating non-local effects, the scan area around each point must be selected sufficiently large. For this purpose, the scan area must be larger than the selected convolution kernel. Furthermore, for achievable accuracy, it is advantageous to select a size of the scan area such that procedures at the edge regions of the scan area, such as dilation, lap convolution, or cropping, do not substantially affect the results of the convolution operation. These considerations must be taken into account when selecting the size of the convolution kernel.

[0226] Further steps are then performed as described above. The material signal or material contrast signal function at point (i,j) of the sample 890, 1000 is given by: S(i,j)=Σ k Σ l b k,l G k,l (i,j)+a k l k (i,j)+a0

[0227] The absolute value of the difference between the material signal S(i,j) and the material information M(i,j) is the parameter a k , b k,l is minimized by modifying the reconstructed material contrast signal function S R(i,j) is generated. Data recording within the area of ​​the sample 890, 1000 rather than at individual points (i,j) allows minimizing the influence of topography effects on the local material signal of the sample 890, 1000. The optimization process described above must also be performed for each type of sample, e.g., each type of mask, and each operating point of the SEM 800 in FIG. 8.

[0228] As mentioned above, it is of course also possible to determine a topographic contrast image or a height profile or height profile map of the sample 890, 1000.

[0229] The following describes the use of a transformation model to represent a sample image or separate material contrast and topography contrast portions of a sample image. In the examples described below, the transformation model comprises a machine learning model (ML model), more precisely a deep learning model (DL model). A DL model can be viewed as a network of filters whose filter parameters are defined or optimized using deep learning methods.

[0230] Data recording is performed as described above by the primary electron beam being scanned around each individual point (i,j) on the sample 890, 1000. The requirements regarding the size of each individual scan area on the sample 890, 1000 and the number of scan points contained therein have already been discussed above.

[0231] A known architecture suitable for denoising and segmentation tasks is the U-Net architecture 1400 shown in Figure 14. At the encoder side 1410, the U-Net architecture 1400 includes alternating blocks that perform downconversion and pooling. At the decoder side 1420, corresponding blocks perform upconversion and upsampling instead of the inverse pooling operation. The output of the downconversion block at the encoder side 1410 is provided to the subsequent pooling block and is also supplied as input data to the corresponding upconversion block at the decoder side 1420. The U-Net architecture 1400 includes an input layer 1440 at the encoder side 1410 and an output layer 1450 at the decoder side 1420.

[0232] In the example U-Net architecture 1400 of diagram 1495 in FIG. 4, three images with a resolution of 256×256 pixels are fed as input data 1460 via input layer 1440. For the separation task described in this application, the input data 1460 to U-Net architecture 1400 is the signal I from at least two detectors 850 and 870, or in the general case k detectors. k The scanning area around each individual point (i,j) of the sample 890, 1000 may include a scanning area of ​​20 x 20 to 400 x 400 scanning points. As mentioned above, the accuracy with which non-local effects can be corrected increases with the size of the scanning area, specifically the number of scanning points contained within the scanning area. Meanwhile, the complexity of data recording increases as the square of the length of the scanning area. In the output layer 1460 of the decoder side 1420, the trained U-Net provides predicted output data 1470, which is the image data in the example of Figure 14.

[0233] Diagram 1595 in Figure 15 illustrates a U-Net architecture 1500 adapted for the purposes of this application. In the example of Figure 15, signals 1560 recorded by three detectors are fed to the U-Net architecture 1500 via its input layer 1540. From these data, the trained U-Net architecture 1500 predicts output data 1570 and outputs them at its output layer 1550. In the example shown in Figure 15, these output data are an image or image representation 1590 representing the material contrast of the sample 890, 1000, and an image 1580 having topographic contrast. In the example shown in Figure 15, the input data 1560 has dimensions (k, w, h), which is converted to output data 1570 having dimensions (2, w, h), where k indicates the number of image representations and w and h indicate the number of scan points or pixels in width and height.

[0234] The U-Net architecture 1500 can also be trained to output only one of the two images 1580 or 1590 at its output 1550. The U-Net architecture 1500 can also be trained to output only the material contrast or topography contrast content of one or both of the image representations 1580 and 1590 as a numerical value at its output 1550. The U-Net architecture 1500 can also be trained to provide only changes in material contrast and / or topography contrast at its output layer 1550. Thus, the U-Net architecture 1500 can specifically be trained to directly output a stop signal for a particle beam-induced repair process of a specimen 890, 1000, e.g., a photomask, when the predicted material contrast of one of the presented image representations recorded in a time series changes beyond a predefined threshold.

[0235] Image sizes (w, h) of 128x128, 64x64 or 32x32 with pixel dimensions in the range of 0.5nm to 1.5nm have proven advantageous, resulting in image sizes of 16nm x 16nm to 192nm x 192nm. Thus, non-local effects can be corrected to the maximum extent possible for separation models 700, both empirical and transformation model forms 1500, with reasonable complexity during image recording.

[0236] As an alternative to the U-Net architectures 1400 and 1500 shown in Figures 14 and 15 as examples of DL models, a Residual Network (ResNet) architecture can be used to separate material contrast signals and topography contrast signals in a sample image representation. In a ResNet architecture, the output of a layer is provided as input data to the subsequent layer as well as to the next or multiple subsequent layers on the encoder and decoder sides.

[0237] The quality and quantity of data on which the transformation model or ML or DL ​​model 1400, 1500 can be trained is crucial to the accuracy with which these models 1400, 1500 can predict data. To train the transformation model or U-Net architecture 1400, 1500 of FIGS. 14 and 15 , a training or test sample is recorded by at least two detectors 850, 870, or generally k detectors, that observe the test sample at least partially at different solid angles. However, the test sample or calibrated test sample approach may not record sufficiently diverse data to train the U-Net network 1400, 1500, and / or the complexity of generating a sufficient amount of training data by recording a corresponding number of image representations for training may be too high. In such cases, as described above, a first portion of the training data can be generated by measurement and a second portion by simulation.

[0238] As already explained above, Monte Carlo simulations can statistically determine the interactions of electrons of the primary beam with the nuclei of the sample 890, 1000 or test sample. Simulation tools capable of performing Monte Carlo simulations are known, such as Nebula (nebula-simulator.github.io). This makes it possible to generate additional images or image representations of test structures of one or more test samples that complement the measured image representation of the test sample. The simulation parameters of the tool are adapted so that the simulated image of the test sample can optimally reproduce the experimentally determined image.

[0239] After defining the parameters of the simulation tool for replicating the test specimen, the test specimen can be subjected to random and / or regular variations in the simulation so that the spectra of the simulated test image or training data generated by the simulation have random and / or regular variations. Specifically, the training data generated based on the simulation can include all possible defects of the specimen or test specimen.

[0240] Simulation-generated training data has an additional advantage. The simulated test sample or samples are typically based on the design file of the test sample and / or the specifications of the sample, e.g., a photomask. Therefore, the material and topography contrast portions or information of the simulated image of the test sample are generally known, and the simulation training data can be directly used to train a transformation model, e.g., a U-Net architecture 1400 or 1500, according to the specifications. In the case of a photomask, the specifications may include, for example, the sidewall angle and height of the absorbing pattern elements (see FIGS. 18-22). Compliance with the specifications can be determined, for example, by scanning the sample or test sample using the measuring tip of an atomic force microscope (AFM).

[0241] For experimentally determined test images, the material and topography contrast of the test images must be determined largely ex situ before they can be used as training data for training U-Net models 1400, 1500.

[0242] Once trained, it can be trained with a small amount of additional training data for relatively small variations in the sample, e.g., slightly different material compositions of the absorption pattern elements in the case of photomasks. For this purpose, additional layers can be used, e.g., in the encoder branch 1410, 1510 of the DL model 1400, 1500, so that the entire network 1400, 1500 does not need to be maintained. In an alternative embodiment, the DL model 1400, 1500 can be provided with hyperparameters in its input layer 1440, 1540 that select a trained model for a particular type of mask from the trained generic photomask models.

[0243] By providing additional parameters in the input layer 1440, 1540 of the DL model 1400, 1500, it becomes possible to adapt the trained DL model 1400, 1500 to a particular restoration device in order to provide the trained DL model 1400, 1500 with relatively small variations in the image representation of the detector 850, 870 and its specific characteristics.

[0244] Generative AI (artificial intelligence) methods can also be used to generate SEM images from design files (see Making digital twins using the Deep Learning Kit (DLK), spiedigitallibrary.org, or mask_defect_detection_with_hybrid_deep_learning_network_041205_1.pdf, zeiss.com). Diagram 1695 in FIG. 16 illustrates a schematic diagram of a generative adversarial network (GAN) 1600. A GAN typically includes two artificial neural networks (ANNs) that play a zero-sum game. The first ANN 1610, called the generator, creates candidates. In the example of FIG. 16 , from photomask design data 1620, the generator 1610 generates a simulated image 1630 that looks like an image recorded using an SEM. The goal of the generator 1610 is to generate a simulated SEM image 1630 that is indistinguishable from images generated using a different type of simulation, for example, by performing a Monte Carlo simulation 1640. A comparison of the image 1630 generated by generator 1610 and the simulated image 1640 is illustrated by box 1660 in FIG.

[0245] The second ANN 1650 of the GAN 1600, called the discriminator 1650, evaluates candidates. In the example of FIG. 16 , these candidates are the simulated image 1640 and the image 1630 generated by the generator 1610. To this end, the discriminator 1650 also has the image 1640 generated by performing a Monte Carlo simulation and the image 1670 measured by the SEM 800. The ANN of the discriminator 1650 is trained to distinguish the results provided by the generator 1610, i.e., the generated SEM image 1630, from the authentic measured SEM image 1670. At output 1680, the GAN 1600 outputs a decision whether to consider the presented SEM image 1630 generated by the generator 1610 as authentic in light of the measured SEM image 1670.

[0246] This means that by using an AI method, for example GAN 1600, training data can be generated directly from the design data 1620, in each case generating an SEM image 1670 for each detector 850, 870 for each test structure of the test sample of the design file 1620. However, a prerequisite for this method of generating training data is that the original training data, i.e. the experimental training data 1670 and / or the simulated training data 1640, contain sufficient information about the interaction process between the primary electron beam and the atomic nuclei of the sample to enable GAN 1600 to recognize the underlying structure.

[0247] With Monte Carlo simulation, the problem often arises of not being able to define parameters that allow the simulated image 1640 to perfectly match the image measured by the SEM 800. In this case, a hybrid approach can be used to achieve better results. Diagram 1795 in FIG. 17 again shows a schematic representation of the GAN 1600 of FIG. 16 . An image 1720 of a test structure generated using Monte Carlo simulation is provided as input data to the generator 1610 of the GAN 1600. From this input data, the generator 1610 synthesizes an image 1730 that appears to have been recorded by the SEM 800. The discriminator 1650 of the GAN 1600 compares or evaluates the image 1730 generated by the generator 1610 with the measured SEM image 1770 and determines, at output 1780, whether to consider the synthesized image authentic. As a result, the simulated image 1720 can be optimized so that the image 1730 synthesized by the generator 1610 is indistinguishable from the measured image 1770. 16, the synthetic image 1730 can then be used as input data to train the GAN 1600. Because the Monte Carlo simulation includes the physics of the interaction process between the primary electron beam and the sample 850, 870, the GAN 1600 can learn something new about the underlying physics.

[0248] The test structure of the test sample, which is considered the basis for generating the image or image representation used to train the DL model 1400, 1500, should contain the maximum possible number of different relevant features. If the sample includes a photomask, examples of relevant features include various edges of the transition between the absorbing pattern element and the mask substrate, LS (line-space) structures with various line widths and pitches, protrusions and intrusions, contact holes, etc. The image representation or image used to train the DL model 1400, 1500, or in general, the transformation model, does not necessarily need to include a full or large-area SEM image. Rather, an image segment is sufficient for this purpose. As mentioned above, the image segment should include linear dimensions of the sample 890, 1000 on the order of 200 nm so that non-local effects in the image or image segment can be corrected with high precision.

[0249] 18-21 show, by way of example, schematic representations of some structural elements or features whose image representations can be used for training the ML model 1500. The possible ranges of variation of the structural elements of FIGS. 18-21 are summarized in the table of FIG.

[0250] FIG. 18 shows a schematic cross-section of a photomask 1800 including a mask structure 1810 made of a first material M1 in each of upper and lower partial images 1805 and 1895. The upper partial image 1805 shows an absorbing and / or phase-shifting pattern element 1820, while the lower partial image 1895 shows an absorbing and / or phase-shifting pattern element 1830 made of a second material M2. The two pattern elements 1820 and 1830 have heights h and h1, respectively. Furthermore, both pattern elements 1820 and 1830 have a sidewall angle Θ 1840 that is significantly smaller than the sidewall angle Θ = 90° predefined, for example, by the specifications. The pattern element 1830 in the lower partial image 1895 also has an intermediate step 1850 having a width w1 and a height h2. The sidewall angle 1840 of the exemplary mask 1800 has the same value after the intermediate step 1850 as before the intermediate step 1850. However, it is possible that these two angles are not equal in magnitude.

[0251] FIG. 19 shows a line-spacing (LS) structure 1900 having a substrate 1910 made of a first material M1 and a strip 1920 made of a second material M2 disposed thereon. The two strips 1920 have a width w. The left strip 1920 also has a penetration 1930 of missing material M2 in the form of a trapezoid 1930. In the example of FIG. 19, the trapezoid has side lengths a and b and a height c. The right strip 1920 of FIG. 19 has a trapezoid-shaped protrusion of excess material M2. In FIG. 19, the trapezoids 1930 and 1940 have the same shape and size. Of course, the penetration 1930 and the protrusion 1940 can have different geometric shapes and areas.

[0252] 20 has test structures for Monte Carlo simulations in the form of contact hole 2030 in left partial image 2005 and rectangular pattern element 2050 in right partial image 2095. Rectangular contact hole 2030 can be etched in pattern element 2050 made of material M2 down to substrate 2010 made of material M1. Contact hole 2030 has width w, height h, and radius of curvature r at the corners of contact hole 2030. The same dimensions and same radius of curvature apply to pattern element 2050 in right partial image 2095.

[0253] FIG. 21 shows the test structure in the form of a contact hole of left partial image 2005 of FIG. 20 in both left partial image 2105 and right partial image 2195. However, in left partial image 2105, contact hole 2130 has a penetration 2140 of excess material M2. Penetration 2140 is again trapezoidal in shape and begins a distance d from the upper edge of contact hole 2130. Contact hole 2150 in right partial image 2195 has a bulge 2160 in the form of a defined "defect" 2160 of missing material M2. This defect has the same dimensions as defect 2140 in left partial image 2105 in the form of penetration 2140 and is positioned in mirror image with respect to defined penetration 2140.

[0254] Table 2200 in Figure 22 shows the range of variation of the parameters of Figures 18-21 that can intentionally create defects 1930, 1940, 2140, 2160 so that the training data generated based on these structural elements 1820, 1830, 1930, 1940, 2130, 2160 is sufficiently diverse. The generated structural elements, including their variations, are decomposed into polygons and used as input or input data for Monte Carlo simulations.

[0255] A second exemplary embodiment for determining the topographic and material contrast contributions of two image representations recorded simultaneously at different solid angles is described below. The second exemplary embodiment uses, for example, machine learning model 1500 as separation model 700.

[0256] Diagram 2395 in Figure 23 schematically illustrates an implementation of an ML model 2300 that converts the first measured image representation 720 and the second measured image representation 730 into a transformed image representation 2370 having a predefined ratio of topography contrast to material contrast. The ML model 2300 may include a deep learning model, for example, a U-Net architecture 1500.

[0257] For the specific purposes of this application, the transformed image representation 2370 is based on an intensity distribution that arises solely from the material composition of the sample 890, 1000, in this example, the photomask 890, 1000. In addition to the image representations 720, 730, at least one parameter 2350 and / or at least one hyperparameter 2360 are further provided to the trained machine learning model 2300 via the input layer 2310 of the trained machine learning model 2330. Additional parameters 2350 and hyperparameters 2360 are described in the "Summary of the Invention" section of this specification. The hyperparameter 2360 can be used, for example, to classify the sample 890, 1000, here in the form of a photomask 890, 1000, as a binary mask.

[0258] The image representations 720, 730 are two-dimensional pixel matrices representing grayscale values. 0 2 0 ~2 16 2 16 A pixel can have a size ranging from 2 4 ~2 10 The image representations 720, 730 may be encoded with a bit depth. The matrix size and pixel depth may be the same for both image representations 720, 730. However, the ML model 2300 may be trained to process image representations having different matrix sizes and / or pixel depths. The ML model 2300 may also be trained to process three or more image representations 720, 730 presented to the input layer 2310 to form a single transformed image representation 2370 (not shown in FIG. 23 ).

[0259] The trained model 2300 provides a transformed image representation 2370 at its output layer 2320. The transformed image representation 2370 may have the same matrix size and the same pixel depth as one of the two image representations 720, 730 presented at the input 2310. However, it is also possible that the transformed image representation 2370 has a different matrix size and / or pixel depth compared to one of the two image representations 720, 730.

[0260] The ML model 2300 may include one of the models described in the "Summary of the Invention." It is advantageous to select a model adapted to the problem to be solved from among the many available general-purpose ML models. It is also advantageous to adapt the selected general-purpose ML model 2300 to the problem to be solved and the required prediction accuracy. The U-Net architecture 1500 and / or the ResNet architecture have proven advantageous for the problem to be solved. The ML model 2300 can be adapted, for example, by adapting the complexity of the kernel function of the ML model 2300. In the case of an ML model 2300 having an encoder-decoder architecture, this can be achieved by a corresponding selection of the number of layers of the ML or DL ​​model 2300. In the case of an ML model 2300 implemented, for example, in the hybrid form described above, the number of leaves in the RDT or the number of trees in the RDF can be adapted to the problem to be solved.

[0261] Diagram 2495 in Figure 24 shows ML model 2400, which represents a modified version of ML model 2300 of Figure 23. Unlike ML model 2300, ML model 2400 is trained to generate two transformed image representations 2470, 2480 from two image representations 720, 730 presented in input 2310. The two transformed image representations 2470, 2480 may have the same matrix size and pixel depth, or these variables for the two transformed image representations 2470, 2480 may be different. The two transformed image representations 2470, 2480 may represent different ratios of topography contrast and material contrast of the image representations 720, 730. Specifically, ML model 2400 can be trained to the effect that, for example, transformed image representation 2470 represents only topography contrast and transformed image representation 2480 represents only material contrast. From the transformed image representation 2480, a signal to stop the local chemical feed repair process can be derived with high accuracy.

[0262] Diagram 2595 in Figure 25 illustrates a further possible variation of ML model 2300. Trained machine learning model 2500 has been trained to output the topography contrast and material contrast contributions of one or both of image representations 720, 730 at its output layer 2520. Of course, ML model 2500 could also be trained to output only the material contrast portion of image representation 730 at output layer 2520.

[0263] The ML model 2500 can also be trained to directly generate a signal to stop the local chemical repair process. To this end, the ML model 2500 is trained to recognize changes in the material contrast contribution between a first set of image representations 720, 730 and a second set of image representations 720, 730 recorded at a later time. If this change exceeds a predefined threshold, the trained ML model 2500 provides a corresponding signal in its output layer 2520. Conversely, as long as the temporal profile of the material contrast change in the set of image representations 720, 730 remains below the predefined threshold, no data is output by the correspondingly trained ML model 2500.

[0264] Before one of the machine learning models 1500, 2300, 2400, 2500 can be used to accomplish its intended task, it must be trained for its intended use. Diagram 2695 in FIG. 26 illustrates the training of the machine learning models 2300, 2400, 2500 or ML models 2300, 2400, 2500. Before the ML model 2300 can predict the transformed image representation 2370 from the presented image representation 720, 730, it must be trained with an extensive dataset or training dataset for the task. Of course, this also applies to the ML models 2400 and 2500.

[0265] To generate training data, a measurement device, e.g., SEM 800, performs a long series of measurements of the same type, including image representation tuples from the sample used for training. In the example of the present description, the sample 890, 1000 is a section of a photomask 300. Other samples can be wafers, nanoimprint lithography templates, MEMS, NEMS, or PICS. In the example of the present application, the image representation tuples include two pairs of image representations recorded from different representatives of a sample class at at least partially different solid angles. In the example of the present description, N different binary photomasks with different absorber patterns that additionally cover the entire range of known defects are measured in the same manner by a measurement device, e.g., SEM 800, in a detector configuration with two detectors 850, 870, where N should be chosen large enough so that the relevant characterization parameters of the image representation tuples, i.e., the topography contrast contribution and material contrast contribution of the image representation tuples, change significantly over the course of the measurements of the training data set. Furthermore, the measurement environment and therefore the characterization parameters can be deliberately varied when recording the training data in order to generate a database that is as representative as possible for training purposes. For this purpose, the sample set, i.e. the set of binary masks, can include, in addition to defect-free binary masks, defect binary masks and repaired binary masks 300, 890, 1000, in particular masks 300, 890, 1000 at any stage of the repair process.

[0266] Under a second value of the hyperparameter 2360, a generic model can be trained for a second type of mask, for example, a phase-shift mask.

[0267] The training dataset includes characterization image representation tuples 2630 and 2640 used for training along with at least one additional parameter characterizing the measurement device for measuring the image representation tuples 2630, 2640 and / or at least one hyperparameter 2650. The training data is provided to the training ML model 2300 at the input layer 2610. The hyperparameter 2660 indicates the classification of the characterization image representation tuples 2630 and 2640 used for training. During the training phase, the training or learning ML model 2600 generates a transformed image representation 2670 from the training characterization pair of image representations 2630 and 2640 and its associated additional parameters 2650, 2660. The predicted transformed image representation 2670 is compared to the image representation 2630 or 2640 of the assigned image representation tuple 2630, 2640. This is indicated by the double-headed arrow 2680 in FIG. 26 . The trained ML model 2600 outputs a predicted transformed image representation 2670 at its output layer 2620. Of course, it is also possible to train the ML model 2600 to provide two transformed image representations at its output layer 2620, a first representing material contrast and a second representing topography contrast of the specimen (not shown in FIG. 26 ).

[0268] Depending on the selected ML model 2300, 2400, 2500, there are various ways to adapt the parameters of the ML model 2300, 2400, 2500 during the training phase. For example, for deep neural networks (DNNs), which typically have a large number of parameters, the iterative technique "stochastic gradient descent" is established. In this case, training data is repeatedly "presented" to the learning ML model 2600, i.e., the ML model 2600 calculates predictions for the transformed image representation 2670 from the characterization image representation tuples 2630, 2640 used for training with the current parameter set. The above-mentioned comparison is then performed. If a deviation occurs between the transformed image representation 2670 and one of the image representations 2630 or 2640 of the image representation tuples 2630, 2640 and the actual values ​​of the topography contrast and / or material contrast of the selected image representation 2630 or 2640, the parameters of the learning ML model 2600 are adapted. The training phase ends when a local optimum is reached, i.e., when the deviation between the predicted topographic contrast and / or material contrast and the actual topographic contrast and / or material contrast of the image representation 2630 or 2640 no longer changes, or when the predefined time allocation for the training cycle of the learning or training ML model 2600 is exhausted.

[0269] The characterization image representation tuples 2630, 2640 used for training originate from a particle beam-based metrology device, such as the SEM 800 of FIG. 8 and / or the repair device 2700 described in the context of FIG. 27. However, the methods described herein can also be used with any metrology device that generally uses a particle beam to image elements of a photolithography process. In particular, the methods described herein can be used with scanning electron microscopes and / or metrology devices that use an ion beam to image a photomask or wafer.

[0270] FIG. 27 shows a schematic cross-section of some key components of a device 2700 designed to simultaneously record two image representations of a sample using two detectors with overlapping solid angles. In the example device 2700 of FIG. 27, the two image representations simultaneously recorded by the two detectors differ in polar angle. The device 2700 can also apply a separation model 700 to determine topography and / or material contrast in at least one of the at least two image representations 720, 730. The device 2700 can also be configured to train a transformation model and / or machine learning model 1500, 2300, 2400, or 2500, as examples of the separation model 700 in the form of a transformation model. The device 2700 can also be configured to repair sample defects by performing a particle-beam-induced local chemical process. The example device 2700 of FIG. 27 includes an improved scanning particle microscope (SEM) 2710 along with a gas supply system 2770.

[0271] The device 2700 includes a particle beam source 2705 in the form of an electron beam source 2705 that generates an electron beam 2715 as particle beam 2715. Compared to an ion beam, the electron beam 2715 has the advantage that the electrons incident on the sample 2725 or the lithography mask 300, 890, 1000 cannot substantially damage the sample 2725 or the mask 300, 890, 1000. However, it is also possible to use an ion beam, an atomic beam, a molecular beam or a high-energy photon beam, for example electrons in the extreme ultraviolet (EUV) wavelength range (not shown in FIG. 27 ), in the device 2700 for the purpose of processing the sample 2725.

[0272] The scanning particle microscope 2710 comprises an electron beam source 2705 and an electron optical column 2720 in which is arranged a beam optical unit 2713, e.g., in the form of an electron optical unit of the SEM 2710. In the SEM 2710 of Figure 27, the electron beam source 2705 generates an electron beam 2715, which is directed as a focused electron beam 2715 by imaging elements arranged in the column 2720, not shown in Figure 27, to a location 2722 on a sample 2725, which may comprise, e.g., a photolithography mask 300, 890, 1000. The beam optical unit 2713 thus forms the imaging system 2713 of the electron beam source 2705 of the SEM 2710.

[0273] The imaging elements of column 2720 of SEM 2710 can further scan electron beam 2715 over sample 2725. Sample 2725 can be inspected, i.e., analyzed and processed, using electron beam 2715 of SEM 2710. An aperture or aperture system comprising multiple apertures (not shown in FIG. 27 ) can be installed in electron optical column 2720 of SEM 2710, preferably downstream of the collecting lens of SEM 2710. The aperture or aperture system can be configured by configuration unit 2790 of computer system 2780 of device 2700.

[0274] Secondary particles, i.e., backscattered electrons (BSEs) and secondary electrons (SEs), generated by electron beam 2715 as a primary particle or electron beam 2715 in the interaction region of sample 2725 are recorded by a combination of two detectors 2717 and 2719. In the exemplary configuration of FIG. 27 , the two detectors 2717 and 2719 are referred to as “in-lens detectors.” Detector 2717 is mounted in a ring shape around primary particle beam 2715 in the vicinity of sample 2725, so that detector 2717 collects secondary particles over a wide solid angle range. The solid angle range is symmetric with respect to the polar angle relative to the beam axis of electron beam 2715. Detector 2717 also has only a small aperture for the passage of primary electron beam 2715, and therefore detects secondary particles, specifically BSEs, that are reflected from sample 2725 at small angles (polar angles) relative to the beam axis of primary particle beam 2715.

[0275] Detector 2719, located above detector 2717 in column 2720, primarily images BSEs emitted by sample 2725 at very small angles, i.e., very small polar angles, relative to the axis of primary particle beam 2715 or electron beam 2715. Detector 2717 partially obscures detector 2719, so the solid angles through which the two detectors 2717 and 2719 observe sample 2725 are at least partially different. In the exemplary detector configuration shown in Figure 27, the solid angle portions of the two detectors 2717 and 2719 are at different polar angles. The two detectors 2717 and 2719 can be located at different positions within column 2720 in different embodiments. The two detectors 2717 and 2719 convert the SEs generated by the electron beam 2715 at the measurement point 2722 and / or the BSEs backscattered from the sample 2725 into electronic measurement signals and transfer the electronic measurement signals to an evaluation unit 2785 of a computer system 2780 of the device 2700. The detector 2719 may include a filter or a filter system (not shown in FIG. 27 ) to discriminate the SEs and / or BSEs by energy. The detectors 2717 and 2719 are controlled by a configuration unit 2790 of the device 2700.

[0276] The exemplary device 2700 may also include a third detector 2721. The third detector 2721 may be designed specifically to detect electromagnetic radiation in the X-ray range. As a result, the detector 2721 makes it possible to analyze the material composition 450, 460 of the radiation emitted by the sample 2725 when inspecting the sample 2725. The detector 2721 is likewise controlled by the configuration unit 2790.

[0277] Device 2700 may also include a fourth detector (not shown in FIG. 27), often implemented in the form of an Everhart-Thornley detector, typically located external to column 2720. This is generally used to detect SEs.

[0278] The device 2700 includes a flood gun 2703, which can provide ions with low kinetic energy to a region of the sample 2725. The flood gun 2703 can be further configured to provide electrons 2707 with a configurable incident energy E0 to a region of the sample 2725 to be processed and / or analyzed. The ions with low kinetic energy and / or the electrons 2707 with the configurable incident energy E0 can compensate for charging of the sample 2725.

[0279] The device 2700 may also include a mesh or screening grid at the output of the column 2720 of the improved SEM 2710 (not shown in FIG. 27). By applying a voltage between the mesh or screening grid and a metal tube (e.g., a liner tube, also not shown in FIG. 27) attached to the region of the objective lens of the column 2720, it is possible to create a configurable potential of the electrons 2707 of the electron beam 2715 so that the incident energy E can be varied by a desired value. Furthermore, a mesh can also be used to compensate for the electrical resistance of the sample 2725. The screening grid mesh may also be grounded.

[0280] In addition to the electron beam source 2705, the device 2700 may include a second radiation source (not shown in FIG. 27), which may be a second electron beam source or a source of a different type of particle, such as ions, atoms, molecules, or high-energy photons.

[0281] The specimen 2725 is placed on a specimen stage 2730 or specimen holder 2730 for inspection. The specimen stage 2730 is also referred to in the art as a "stage." As represented by the arrows in Figure 27, the specimen stage 2730 can be moved in three spatial directions relative to the column 2720 of the SEM 2710, for example, using a micromanipulator not shown in Figure 27.

[0282] In addition to translation, the sample stage 2730 can rotate at least about an axis oriented parallel to the beam direction of the particle beam source 2705. The sample stage 2730 can also be implemented to be rotatable about one or two further axes, which or these axes are arranged in the plane of the sample stage 2730. The two or three rotation axes preferably form a rectangular coordinate system. As can be seen in Figure 27, rotation of the sample stage 2730 about rotation axes arranged in the plane of the sample stage 2730 is often only possible to a limited extent due to the short distance between the end of the column and the sample 2725.

[0283] The sample 2725 to be inspected can be any microstructured component or component part that requires analysis, i.e., sample imaging and possibly subsequent processing, e.g., repair of local defects in a lithography mask 300, 890, 1000. In this context, the sample 2725 can include, for example, a transmission or reflection photomask 300, 890, 1000 and / or a template for nanoimprint technology or nanoimprint lithography. The transmission or reflection photomask 300, 890, 1000 can include any kind of photomask, such as, for example, a binary mask, a phase-shift mask, an OMOG mask or a mask for double or multiple exposure.

[0284] The device 2700 of FIG. 27 may further include one or more scanning probe microscopes, for example in the form of an atomic force microscope (AFM) (not shown in FIG. 27), that can be used to analyze and / or process the sample 2725.

[0285] The exemplary scanning electron microscope 2710 shown in Figure 27 is operated within a vacuum chamber 2701. To create and maintain the necessary reduced pressure within the vacuum chamber 2701, the SEM 2710 of Figure 27 includes a pumping system 2709.

[0286] The device 2700 also includes a computer system 2780. The computer system includes a setting unit 2790 configured to set the incident energy E of the electrons 2707 of the electron beam 2715 to a predefined value. To this end, the setting unit 2790 can also set the acceleration and deceleration voltages of the electrons 2707 of the electron beam 2715. The setting unit 2790 can also set the potential of an energy filter of the detector 2719.

[0287] The computer system 2780 may also have an interface 2777 through which the computer system 2780 receives information about the specimen 2725, such as its material composition and / or surface topography. Additionally, the computer system 2780 may also obtain information about defects in the specimen 2725. Additionally, the computer system 2780 may obtain image representations 720, 730 of the specimen 2725 through the interface 2777 and / or transmit image representations 720, 730 of the specimen 2725 through the interface 2777.

[0288] The computer system 2780 may also include a scanning unit 2782 that scans the electron beam 2715 over the sample 2725. The setting unit 2790 may be further configured to set various parameters of the improved scanning particle microscope 2710 of the device 2700. The setting unit 2790 may further control the rotation of the micromanipulator and the sample stage 2730.

[0289] Additionally, the evaluation unit 2785 of the computer system 2780 can analyze the measurement signals from the detectors 2717 and 2719 and generate therefrom images or image representations 720, 730 of the specimen 2725 that can be displayed on the display 2795. In particular, the evaluation unit 2785 can be designed to determine the location and contours of missing material defects and / or excess material defects in the specimen 2725, e.g., the lithography mask 300, 890, 1000, from the measurement data of the detectors 2717 and 2719.

[0290] The computer system 2780 can also be configured to apply the separation model 700 to the image representations 720, 730 generated by the detectors 2717 and 2719 in order to determine the contributions of topographic contrast and material contrast to the image representations. To this end, the computer system 2780 can include one or more algorithms that allow for identifying parameters of the empirical model from the two image representations 720, 730 of the sample 2725. The algorithms of the computer system 2780 can be implemented using hardware, software, or a combination thereof. In particular, the one or more algorithms can be realized in the form of an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), and / or an FPGA (Field Programmable Gate Array).

[0291] The computer system 2780 and / or the evaluation unit 2785 may include a memory (not shown in FIG. 27 ), preferably a non-volatile memory, that stores the transformation models and / or machine learning models 1500, 2300, 2400, 2500 in generic and / or trained form. The non-volatile memory of the computer system 2780 may also store training data sets for the transformation models and / or ML models 2300, 2400, 2500. The evaluation unit 2785 may be configured to determine topography contrast and / or material contrast of the image representations 720, 730 of the sample 2725 from the image representations 720, 730. The computer system 2780 may also include an interface 2777 for exchanging data with the Internet, an intranet, and / or some other device. The interface 2777 may include a wireless or wired interface. The evaluation unit 2785 can provide the configuration unit 2790 with data that enables the configuration unit 2790 to stop the local chemical remediation process of the sample 2725.

[0292] The evaluation unit 2785 and / or the configuration unit 2790 may be incorporated into the computer system 2780 as shown in Figure 27. However, the evaluation unit 2785 and / or the configuration unit 2790 may also be implemented as a standalone unit inside or outside the device 2700. In particular, the evaluation unit 2785 and / or the configuration unit 2790 may be designed to perform some of its tasks using dedicated hardware implementations.

[0293] Additionally, computer system 2780 may be incorporated into device 2700 or may be designed as a stand-alone device (not shown in FIG. 27). Computer system 2780 may be implemented using hardware, software, firmware, or a combination.

[0294] The gas supply system 2770 implemented by device 2700 is described below. As already explained above, a sample 2725 is placed on a sample stage 2730. An imaging element 2713 of a column 2720 of the SEM 2710 can focus an electron beam 2715 and scan it over the sample 2725. The electron beam 2715 of the SEM 2710 can be used to induce particle beam induced deposition processes (EBID, electron beam induced deposition) and / or particle beam induced etching processes (EBIE, electron beam induced etching). To perform these processes, the example device 2700 of FIG. 27 includes three different supply vessels 2740, 2750, and 2760 for storing various precursor gases.

[0295] The first supply container 2740 stores a precursor gas, for example, a metal carbonyl such as chromium hexacarbonyl (Cr(CO)6) or molybdenum hexacarbonyl (Mo(CO)6). The precursor gas stored in the first supply container 2740 can be used to deposit missing material on the lithography mask 300, 890, 1000, for example, in a localized chemical deposition reaction. Also, a protective or sacrificial layer can be deposited on the mask 300, 890, 1000. Furthermore, the precursor gas stored in the first supply container 2740 can be used to deposit drift markers on the mask 300, 890, 1000 or sacrificial layer.

[0296] The electron beam 2715 of the SEM 2710 serves as an energy supplier for decomposing a precursor gas stored in a first supply container 2740 at the location where material is intended to be deposited on the sample 2725. This means that as a result of the combined supply of the electron beam 2715 and the precursor gas, an EBID process is performed for the local deposition of missing material, e.g., material missing in the mask 300, 890, 1000.

[0297] The electron beam 2715 can be focused to a spot diameter in the range of a few nanometers. The region of interaction, or scattering cone, at which the electron beam 2715 produces SEs depends first on the energy of the electron beam 2715 and second on the material composition on which the electron beam 2715 is incident. The diameter of the region of interaction is in the low single-digit nanometer range. Therefore, the diameter of the scattering cone of the electron beam 2715 defines the achievable resolution limit when performing localized particle beam-induced reactions. This resolution limit is currently in the single-digit nanometer range.

[0298] 27, the second supply container 2750 stores an etching gas that enables a localized electron beam induced etching (EBIE) process to be performed. The electron beam induced etching process can be used to remove excess material from the sample 2725, such as excess material in the right strip 1920 and / or the penetration 1940 of the contact hole 2130. For example, the etching gas can include xenon difluoride (XeF), a halogen, or nitrosyl chloride (NOCl).

[0299] Third supply container 2760 can store an additive or additional gas that can be added, as needed, to the etching gas kept available in second supply container 2750 or to the precursor gas stored in first supply container 2740. Alternatively, third supply container 2760 can store a second precursor gas or a second etching gas.

[0300] 27, each of the supply vessels 2740, 2750 and 2760 has its own control valve 2742, 2752 and 2762 in order to manage or control the amount of the corresponding gas supplied per unit time, i.e. the gas volumetric flow rate at the location 2722 of incidence of the electron beam 2715 on the sample 2725. The control valves 2742, 2752 and 2762 can be controlled or managed by a setting unit 2790 of the computer system 2780. By this means it is possible to set the partial pressure conditions of one or more gases supplied at the processing location in order to perform a wide range of EBID and / or EBIE processes.

[0301] Also in the example device 2700 of FIG. 27, each supply vessel 2740 , 2750 , and 2760 has its own gas supply line system 2745 , 2755 , and 2765 that terminates in a nozzle 2747 , 2757 , and 2767 near the point of incidence 2722 of the electron beam 2715 on the specimen 2725 .

[0302] Supply vessels 2740, 2750, and 2760 may have their own temperature setting and / or control elements that allow both cooling and heating of the corresponding supply vessels 2740, 2750, and 2760. This allows storing and particularly supplying precursor gases at optimal temperatures, respectively (not shown in FIG. 27 ). Setting unit 2790 may control the temperature setting and temperature control elements of supply vessels 2740, 2750, and 2760. The temperature setting elements of supply vessels 2740, 2750, and 2760 may further be used to set the vapor pressure of the precursor gases stored therein by selection of an appropriate temperature during the EBID and EBIE treatment processes.

[0303] The device 2700 can include multiple supply vessels 2740 for storing two or more precursor gases. The device 2700 can further include multiple supply vessels 2750 for storing two or more etching gases (not shown in FIG. 27).

[0304] 28 illustrates the key steps of a method for determining topography and / or material contrast features of an image representation 720, 730 of a sample 890, 1000, 2725. The method begins at step 2810.

[0305] Step 2820 includes providing at least two image representations 720, 730 of the sample 890, 1000, 2725 at at least partially different solid angles relative to the sample. Providing may include loading the at least two image representations 720, 730 from non-volatile memory, transmitting over a network, and / or recording the at least two image representations 720, 730 using, for example, detectors 2717 and 2719.

[0306] Step 2830 includes determining the topographical contrast and / or material contrast of the sample based at least in part on the at least two image representations 720, 730 of the sample 890, 1000, 2725. A separation model 700 can be used for this purpose. The separation model 700 can include a parameterized empirical model and / or a trained transformation model, e.g., a deep learning model 1500, 2300, 2400, 2500, applied to the at least two image representations 720, 730 of the sample 890, 1000, 2725. A computer system 2780 configured for this use, e.g., using a particular graphics processor unit and / or one of the hardware components identified above, can implement the separation model 700 by application to the at least two image representations 720 and 730.

[0307] The method ends at step 2840. [Explanation of symbols]

[0308] 300 samples 310 Mask substrate 320 pattern elements 322 parts 324 Missing Pattern Elements 330 Pattern Elements 340 pattern elements 350 First Ingredient 360 Second Material 380 Topographic Contrast Contribution 395 Image expression 450 Material contrast, material contrast part 460 Material contrast, material contrast part 480 Edge 495 Image expression 500 coordinate system 510 partial image 520 partial image 550 Material contrast, material contrast area, material contrast intensity 560 Material contrast, material contrast area, material contrast intensity 580 Edge 585 Topographic Contrast 595 Figure 695 Figure 700 Separated Model 710 Input Data 720 statue 730 statue 740 statue 750 output 760 Material contrast image 770 Topographic Contrast Image 795 Figures 800 Scanning Electron Microscope 810 Electron source 820 Connection 830 Gas Supply System 840 objective lens 850 detector 860 Secondary electron 870 detector 875 routes 880 Backscattered electrons 890 samples 895 Figures 910 yen 920 classification 930 classification 950 yen 960 classification 995 Figure 1000 samples 1005 Figure 1010 board 1020 Pattern Elements 1030 Ta edge 1035 vertical dotted line 1055 Lower part image 1070 dashed line 1075 horizontal dotted line 1080 reconstructed signal 1085 horizontal dotted line 1095 Figure 1195 Figure 1400 U-Net Architecture 1410 Encoder side 1420 decoder side 1440 Input Layer 1450 output layer 1460 output layer 1470 decoder side 1495 Figure 1500 U-Net Architecture 1550 output 1560 signal, input data 1570 output data 1580 Image expression 1590 Image expression 1595 Figure 1600 Generative Adversarial Networks 1610 Generator 1620 design data 1630 Simulation image 1640 statue 1650 Discriminator 1660 Compare Box 1670 SEM images, training data 1680 output 1695 Figure 1720 statue 1730 statue 1770 SEM image 1780 output 1795 Figure 1800 photomasks 1805 Upper part image 1810 Mask Structure 1820 pattern elements 1830 pattern elements 1840 side wall angle 1850 Middle Section 1895 Lower part image 1900LS construction 1910 PCB 1920 Strip 1930 Penetration, defect 1940 Protrusions, Defects 2005 Left partial image 2010 PCB 2030 Contact Hole 2050 pattern elements 2095 Right partial image 2105 Left partial image 2130 Contact Hole 2140 Penetrations, defects 2150 Contact Hole 2160 Defects, bulges 2195 Right partial image 2300 machine learning models 2310 Input Layer 2320 Output Layer 2350 Parameters 2360 Hyperparameters 2370 Image expression 2395 Figure 2400 machine learning models 2470 Image expression 2480 Image expression 2495 Figure 2500 machine learning models 2520 output layer 2595 Figures 2600 machine learning models 2610 Input Layer 2620 output layer 2630 Image Representation Tuple 2640 Image Representation Tuple 2650 Parameters 2660 Hyperparameters 2680 double arrow 2695 Figures 2700 devices 2701 Vacuum Chamber 2703 Flood Gun 2705 ​​Electron beam source 2707 Electronic 2709 Pump System 2710 Improved Scanning Particle Microscope 2713 Beam optical units, imaging elements, imaging systems 2715 Particle beams, electron beams 2717 Detector 2719 detector 2720 ​​Column 2721 Detector 2722 measurement points 2725 Samples 2730 Sample Stage 2740 First Supply Container 2742 Control valve 2745 Gas supply line system 2747 Nozzle 2750 Second Supply Container 2752 Control valve 2755 Gas supply line system 2757 Nozzle 2760 Third Supply Container 2762 Control valve 2765 Gas supply line system 2767 Nozzle 2770 Gas Supply System 2777 Interface 2780 Computer Systems 2782 Scanning Unit 2785 evaluation unit 2790 Setting Unit 2795 Display

Claims

1. 1. A method (2800) for determining topographic contrast (585) and / or material contrast (550, 560) of a sample (300, 890, 1000), comprising: providing (2820) at least two image representations (720, 730) of said sample (300, 890, 1000) recorded at least partially at different solid angles relative to said sample; and b. determining (2830) the topography contrast (585) and / or the material contrast (550, 560) of the sample (300, 890, 1000) based at least in part on the at least two image representations (720, 730) of the sample (300, 890, 1000).

2. The method (2800) of claim 1, wherein said determining comprises applying a separation model (700) to said at least two image representations (720, 730).

3. 3. The method (2800) of claim 2, wherein the separation model (700) comprises at least one member of the group consisting of an empirical model and a transformation model (1500, 2300, 2400, 2500).

4. 4. The method (2800) of claim 3, further comprising adapting the empirical model to the sample (300, 890, 1000).

5. The method (2800) of claim 4, further comprising determining parameters of the empirical model.

6. 6. The method (2800) of claim 5, wherein determining the parameters of the empirical model includes at least one element of the group consisting of: recording at least two image representations (720, 730) of at least one calibrated test structure at at least partially different solid angles; simulating at least two image representations (720, 730) of the at least one calibrated test structure at at least partially different solid angles; and recording at least two image representations (720, 730) of the at least one calibrated test structure at the at least partially different solid angles, and wherein at least one detector (870, 2719) has an activated screening grating.

7. 4. The method (2800) of claim 3, wherein the transformation model (1500, 2300, 2400, 2500) comprises at least one transformation model, preferably a machine learning model and / or a generative model (1500, 2300, 2400, 2500), having at least two transformation blocks each including at least one generic learnable function.

8. 8. The method (2800) of claim 3 or 7, wherein the transformation model (1500, 2300, 2400, 2500) comprises a machine learning model (2300, 2400, 2500), in particular a deep learning model (1500).

9. 9. The method (2800) of claim 7 or 8, wherein the machine learning model (2300, 2400, 2500) includes at least one additional parameter (2350) provided to the machine learning model (2300, 2400, 2500) at an input (1540, 2310) of the machine learning model (2300, 2400, 2500).

10. 10. The method (2800) of claim 9, wherein the at least one additional parameter (2350) comprises a system parameter of the repair device (2700).

11. The method (2800) of any one of claims 7 to 10, wherein the machine learning model (2300, 2400, 2500) comprises hyperparameters (2360) that characterize the sample (300, 890, 1000).

12. The method (2800) of claim 3 or any one of claims 7 to 11, further comprising the step of training said transformation model (1500, 2300, 2400, 2500) with a training dataset.

13. 13. The method of claim 12, wherein the training dataset for the transformation model comprises at least one element of a group consisting of a multiple tuple of at least two recorded image representations of at least one sample used for training, a multiple tuple of at least two image representations of at least one test structure used for training, a multiple tuple of at least two simulated image representations of at least one sample used for training, and a multiple tuple of at least two recorded image representations of at least one test structure used for training, wherein the tuples each comprise at least two image representations recorded or simulated at at least partially different solid angles relative to the at least one sample and / or the test structure used for training.

14. 14. The method (2800) of claim 12 or 13, further comprising recording the training data set for the transformation model (1500, 2300, 2400, 2500).

15. The method (2800) of any one of claims 1 to 14, wherein determining the topographic contrast (585) and / or the material contrast (550, 560) comprises determining an image representation (510) that is substantially free of topographic contrast portions.

16. A computer program comprising instructions for performing the method steps of any of claims 1 to 15 when said computer program is executed.

17. A device (800, 2700) for determining the topographic contrast (585) and / or material contrast (550, 560) of a sample (300, 890, 1000), comprising: a. means (850, 870, 2717, 2719) for providing at least two image representations (720, 730) of said sample (300, 890, 1000) recorded at least partially at different solid angles relative to said sample; and b. means for determining the topographic contrast and / or material contrast of the sample based at least in part on the at least two image representations of the sample.

18. 18. The device (800, 2700) of claim 17, wherein the determining means (2780) is configured to apply a separation model (700) to the at least two image representations (720, 730) to determine the topography contrast (585) and / or the material contrast (550, 560) of the sample (300, 890, 1000).

19. The device (800, 2700) of claim 17 or 18, comprising at least one first detector (850, 2717) and at least one second detector (870, 2719) for providing the at least two image representations (720, 730) of the sample (300, 890, 1000), wherein the at least one first detector (850, 2717) and the at least one second detector (870, 2719) preferably detect secondary electrons (SE) and backscattered electrons (BSE), respectively, and the SE / BSE ratio of the at least one first detector (850, 2717) and the SE / BSE ratio of the at least one second detector (870, 2719) are preferably different from each other.

20. The device (800, 2700) of claim 19, wherein the first detector (850, 2717) is disposed within an electron optical column of the device (800, 2700) and / or the second detector (870, 2719) is disposed within the electron optical column of the device (800, 2700).

Images