Method for determining a distortion-corrected position of a feature in an image formed with a multi-beam charged particle microscope, corresponding computer program product, and multi-beam charged particle microscope - Patents.com

Abstract

The present invention relates to a method for detecting a charged particle beamlet having a plurality of J primary charged particle beamlets (3) over a plurality of J image subfields (31.mn), a detection unit (200) including a detector for detecting a plurality of J secondary electron beamlets (9) each corresponding to one of the J image subfields (31.mn), a scan control unit (930) connected to the first raster scanner (110) and configured, in use, to control the raster scanning operation of the plurality of J primary charged particle beamlets (3) by the first raster scanner (110), a kernel generation unit (812) configured, in use, to generate a spatially varying filter kernel (910) for correcting spatially varying distortion of the image subfields (31.mn), and an image processing unit (812) synchronized in operation with the detector, the scan control unit (930) and the kernel generation unit. A multi-beam charged particle microscope (1) comprising: an image data acquisition unit (110); and a controller (800, 820) comprising: a data acquisition unit (810); for each of the J image subfields, an analog-to-digital converter (811) for converting, in use, an analog data stream received from the detector into a digital data stream representing the image subfield (31.mn); a hardware filter unit (813) configured to receive the digital data stream and configured, in use, to generate a distortion-corrected data stream by performing a convolution with a spatially varying filter kernel (910) of a segment (32) of the image subfield (31.mn); and an image memory (814) configured to store the distortion-corrected data stream as a 2D representation of the image subfield (31.mn).

Description

[Technical field]

[0001] The present invention relates to the field of multi-beam charged particle microscopes and related inspection tasks. More specifically, the present invention relates to a method for determining a distortion-corrected position of a feature in an image composed of one or more image patches, each composed of a plurality of image subfields imaged with associated beamlets of a multi-beam charged particle microscope. Furthermore, the present invention relates to a corresponding computer program product and a corresponding multi-beam charged particle microscope. [Background technology]

[0002] With the continuing development of miniaturization and high functionality of microstructures such as semiconductor devices, there is a demand for further development and optimization of planar processing techniques and inspection systems for processing and inspecting the critical dimensions of the microstructures. Development and processing of semiconductor devices requires, for example, design verification of test wafers, and planar processing techniques involve process optimization for reliable high-throughput processing. In addition, in recent years, analysis of semiconductor wafers has become necessary for reverse engineering and individual customized configuration of semiconductor devices. Therefore, there is a demand for high-throughput inspection tools that can examine the microstructures on the wafer with high accuracy.

[0003] Typical silicon wafers used in the manufacture of semiconductor devices are up to 12 inches (300 mm) in diameter. Each wafer is divided into 30-60 repeating areas ("dies") up to approximately 800 square mm in size. Semiconductor devices include multiple semiconductor structures fabricated in layers on the surface of the wafer by planar integration techniques. Semiconductor wafers typically have flat surfaces due to the fabrication processes associated therewith. Feature sizes of integrated semiconductor structures have shrunk from a few μm to critical dimensions (CDs) of 5 nm, and in the near future will be further reduced, for example, to feature sizes or critical dimensions (CDs) of less than 3 nm (e.g., 2 nm) or even 1 nm. The aforementioned small feature sizes require the identification of defects of the size of the critical dimensions over a very large area (relative to the feature size) in a short time. For some applications, the specification requirements for the accuracy of the measurements provided by the inspection equipment are even higher, for example, by a factor of two or an order of magnitude. For example, the width of semiconductor features must be measured with an accuracy of less than 1 nm (eg, 0.3 nm or less) and the relative positions of semiconductor structures must be determined with an overlay accuracy of less than 1 nm (eg, 0.3 nm or less).

[0004] It is therefore an object of the present invention to provide a charged particle system and a method of operating a charged particle system at high throughput that enables high precision measurements of semiconductor features with an accuracy of less than 1 nm, 0.3 nm or even less than 0.1 nm.

[0005] A recent development in the field of charged particle microscopy (CPM) is the multi-beam charged particle microscope (MSEM). Multi-beam scanning electron microscopes are disclosed, for example, in US Pat. No. 7,244,949 and US Patent Application Publication No. 2019 / 0355544. In a multi-beam electron microscope, a sample is illuminated by an array of electron beamlets, which may contain, for example, 4-10,000 electron beams as primary radiation, such that each electron beam is separated from the adjacent electron beam by a distance of 1-200 micrometers. For example, in a multi-beam charged particle microscope, approximately 100 separated electron beams or beamlets are arranged in a hexagonal array and separated by a distance of approximately 10 μm. A common objective lens focuses the multiple primary charged particle beamlets onto the surface of the sample under investigation (for example, a semiconductor wafer fixed on a wafer chuck mounted on a movable stage). During irradiation of the wafer surface by the primary charged particle beamlets, interaction products (e.g., secondary electrons) arise from multiple crossing points formed by the focal points of the primary charged particle beamlets, while the amount and energy of the interaction products depend on the material composition and topography of the wafer surface. The interaction products form multiple secondary charged particle beamlets, which are collected by a common objective lens and guided by a projection imaging system of the multi-beam inspection system onto a detector arranged at a detector plane. The detector includes multiple detection areas, each with multiple detection pixels, to obtain an image patch of, for example, 100 μm×100 μm for detecting the intensity distribution of each of the multiple secondary charged particle beamlets. The prior art multi-beam charged particle microscope comprises a series of electrostatic and magnetic elements. By adjusting at least a part of the electrostatic and magnetic elements, the focal position and stigma of the multiple secondary charged particle beams can be adjusted. The prior art multi-beam charged particle microscope comprises at least one crossover surface of the primary or secondary charged particles. The prior art multi-beam charged particle microscope comprises a detection system that facilitates the adjustment.Prior art multi-beam charged particle microscopes include at least one deflection scanner for collectively scanning a plurality of primary charged particle beamlets across an area of ​​the sample surface to obtain an image patch of the sample surface. More details of the multi-beam charged particle microscope and a method of operating the multi-beam charged particle microscope are described in PCT / EP2021 / 061216, filed April 29, 2021, which is incorporated herein by reference.

[0006] However, in charged particle microscopes for wafer inspection, it is desirable to keep the imaging conditions stable so that reliable and reproducible imaging can be performed. The throughput depends on several parameters (e.g., the speed of repositioning the stage and a new measurement site) as well as the measurement area per acquisition time itself. The latter is determined by the dwell time, resolution and number of beamlets. Also, in the case of multi-beam charged particle microscopes, time-consuming image post-processing is required. For example, the signals generated by the detection system of the multi-beam charged particle microscope need to be digitally corrected before stitching together image patches from multiple image sub-fields.

[0007] The multiple primary charged particle beamlets may deviate from a normal raster position in a raster configuration (e.g., a hexagonal raster configuration). Also, the multiple primary charged particle beamlets may deviate from a normal raster position of a raster scanning operation in a planar area segment, and the resolution of the multi-beam charged particle inspection system may be different for each individual scanning position of each beamlet of the multiple primary charged particle beamlets and may depend on the individual scanning positions. For the multiple primary charged particle beamlets, each beamlet enters the intersection volume of the common scanning deflector at a different angle, is deflected to a different exit angle, and traverses the intersection volume of the common scanning deflector by a different path. Therefore, each beamlet will have a different distortion pattern in the scanning operation. Prior art single-beam dynamic correctors are not suitable for mitigating any scanning-induced distortion of the multiple primary beamlets. US Patent Application Publication No. 2009 / 0001267 shows the calibration of a primary beam layout or static raster pattern configuration of a multi-beam charged particle system including five primary charged particle beamlets. Three causes of raster pattern anomalies are shown: rotation of the primary beam layout, scale-up or scale-down of the primary beam layout, and shift of the entire primary beam layout. Thus, US2009 / 0001267 considers basic first-order distortions (rotation, magnification, global shift, or displacement) of the static primary beam raster pattern formed by the static focus of multiple primary beamlets. US2009 / 0001267 also includes calibration of the first-order characteristics of the collective raster scanner (deflection width and deflection direction for collective raster scanning of multiple primary beamlets). And, although a means for compensating for these basic errors of the primary beam layout is discussed, no solution is provided for higher-order distortions of the static raster pattern (e.g., third-order distortion). Even after calibration of the primary beam layout and, optionally, the secondary electron beam path, scanning distortions are introduced during the scanning of each individual primary beamlet, which cannot be addressed by calibration of the static raster pattern of multiple primary beamlets.

[0008] Usually, in today's state-of-the-art multi-beam charged particle microscopes, the fundamental first order image distortions (rotation, magnification, and global shift or displacement) are corrected. However, with the increasing demand for higher accuracy in MSEM measurements in metrology, higher order distortions resulting from the scanning process become more important and need to be properly considered.

[0009] Undisclosed international patent application PCT / EP2021 / 066255, filed on June 16, 2021, deals with minimizing scan-induced distortion differences between multiple primary charged particle beamlets, the entire disclosure content of which is incorporated herein by reference. Said international patent application employs an approach to minimize scan-induced distortion by improving the raster scanner configuration itself. However, such improvements in the raster scanner configuration are usually only realized in newly built multi-beam charged particle microscopes. In contrast, there is also a demand for higher precision when working with existing microscopes, especially when dealing with quantitative metrology inspection tasks (e.g., determining feature sizes of integrated semiconductor structures).

[0010] International Patent Application Publication No. WO2021 / 239380 (corresponding to the aforementioned PCT / EP2021 / 061216) discloses a multi-beam charged particle inspection system and a method of operating the multi-beam charged particle inspection system for high throughput, high resolution, and high reliability wafer inspection. The method and the multi-beam charged particle inspection system are configured to extract a set of control signals from a plurality of sensor data to control the multi-beam charged particle inspection system to maintain imaging specifications including the movement of a wafer stage during a wafer inspection task. However, International Patent Application Publication No. WO2021 / 139380 does not solve the problem of time-consuming image post-processing. Furthermore, International Patent Application Publication No. WO2021 / 139380 does not address scanning-induced distortions or the specific problems caused by scanning-induced distortions. Summary of the Invention

[0011] It is therefore an object of the present invention to provide an alternative solution for correcting scanning-induced distortions in images acquired with a multi-beam charged particle microscope, which solution is particularly suitable for the accurate determination of feature sizes of integrated semiconductor structures.

[0012] The above object is solved by the independent claims. The dependent claims are directed to advantageous embodiments.

[0013] This patent application claims priority to German Patent Application No. 102022102548.9, filed on February 3, 2022, the entire disclosure of which is incorporated herein by reference.

[0014] In contrast to the hardware / physical approach adopted in PCT / EP2021 / 066255, the present invention employs an algorithmic approach. According to a first embodiment of the present invention, the scanning-induced distortion is corrected during image post-processing. The distortion correction is performed, for example by a PC, on the basis of an existing scanned distortion image. Said correction still requires less time and energy, while providing an elegant solution to the specific inspection task. According to a second embodiment of the present invention, the distortion correction is performed during image pre-processing. This is performed by specially configured or programmed hardware components of the MSEM. This MSEM is therefore an MSEM with integrated distortion correction. Furthermore, the first and second embodiments can be combined with each other.

[0015] According to a first aspect, the invention relates to a method for determining a distortion-corrected position of a feature in an image composed of one or more image patches, each composed of a plurality of image subfields, each imaged with associated beamlets of a multi-beam charged particle microscope, comprising: a) providing, for each image subfield, a number of vector distortion maps each characterizing a position-dependent distortion for each pixel of the associated image subfield; b) identifying a feature of interest in the image; c) extracting geometric properties of the features; d) determining a corresponding image subfield that includes the extracted geometric characteristics of the feature; e) determining the location of one or more of the extracted geometrical properties of the feature in the determined corresponding image sub-field; f) generating distortion-corrected image data by correcting one or more positions of the extracted geometric features in the image based on a vector distortion map of the corresponding image subfield; The present invention relates to a method, comprising:

[0016] The image typically comprises multiple image patches, although the method also works if the image comprises only one image "patch". In either case, the image patch comprises multiple image subfields, each of which is imaged (or has been imaged) by an associated beamlet of the multi-beam particle microscope.

[0017] The method is particularly suitable for the correction of scanning-induced distortions, which is a highly accurate correction. As an important aspect of the invention, a vector distortion map is prepared separately for each image sub-field, since the scanning-induced distortions are usually different for each sub-field, which is also the reason for the fact that a typical collective raster scanner cannot compensate for the scanning-induced distortions simultaneously for all beamlets (see above). The vector distortion map is not necessarily prepared as a "map". The term "map" is intended only to indicate that the distortions are vectors, and that this vector depends on the position. As a result, the vector distortion map is in principle a vector field.

[0018] To express the position of the distortion vector in an image subfield, the internal coordinates of the image subfield are used (usually called p, q in this patent application). Moreover, the internal coordinates must be connected to a global coordinate system (usually called x, y in this patent application). The position of each subfield, labeled with the subscript nm, relative to the global coordinate system can be, for example, the position of the center point (p0, q0) of each subfield in the global coordinate system (xnm,ynm).

[0019] A vector distortion map can be pre-determined for each subfield and therefore for each beamlet. Its determination is explained in more detail below. Typically, the vector distortion map will remain valid for multiple imaging procedures. In contrast to International Patent Application Publication No. WO2021 / 239380, the present invention is therefore particularly suitable for the correction of regularly or regularly occurring distortions, in particular regularly occurring scanning-induced distortions. However, the vector distortion map according to the present invention can also be regularly updated. This allows the correction of more unexpected or irregular distortions during image post-processing.

[0020] The method steps b) identifying features of interest in the image and c) extracting geometrical properties of the features can be performed separately or can be combined with each other. In principle, the features of interest can be features of any kind and shape. When investigating semiconductor structures, examples of features of interest are features such as HAR structures (high aspect ratio structures, also called pillars or holes or contact channels).

[0021] The geometrical property of a feature can for example be the contour of the feature. Or it can be only a part of said contour (for example an edge or a corner). In principle, the pixel itself can also represent the feature. According to one embodiment, the geometrical property of a feature is at least one of the following: contour, edge, corner, point, line, circle, ellipse, center, diameter, radius, distance.

[0022] The image data is typically data of interest to be measured (e.g., the position, size, area, or volume of the center or edge of an object of interest, or the distance or gap between multiple objects of interest). Additionally, the image data may also include characteristics such as line edge roughness, the angle between two lines, radius, etc.

[0023] Feature extraction itself is well known in image processing. An example of contour extraction is given in Li Huanliang, 4 th This can be seen in Computer-Based Image Contour Extraction Methods in National Conference on Electrical, Electronics and Computer Engineering (NCEECE 2015), 1185 - 1189 (2016).

[0024] According to one embodiment, the step of extracting the geometrical characteristics comprises the generation of a binary image. Images obtained with a multi-beam particle microscope are typically grayscale images showing the intensity of the detected secondary particles. The data size of such images is very large. In contrast, the data size of a binary image, which for example shows only the contour, is relatively small.

[0025] According to the present invention, the distortion correction is performed only on a part of the whole image, more precisely on the extracted geometric characteristics of the features (e.g. the extracted contours), which makes the distortion correction much faster compared to the conventional distortion correction according to the state of the art, which is performed on every pixel of the grayscale image. Moreover, the distortion correction according to the present invention requires less resources in terms of energy.

[0026] The distortion correction itself includes the steps of d) determining a corresponding image sub-field containing the extracted geometric properties of the feature, e) determining one or more positions of the extracted geometric properties of the feature in the determined corresponding image sub-field, and f) generating distortion-corrected image data by correcting one or more positions of the extracted geometric properties in the image based on the vector distortion map of the corresponding image sub-field.

[0027] Determining the corresponding image sub-fields is necessary to correct the extracted geometrical characteristics with an associated image distortion map. The corresponding image sub-fields can be indicated, for example, in the image metadata or can be determined based on the location of the data in memory or in an image data file.

[0028] The step of determining the position(s) of the extracted geometrical property(s) of the feature in the determined corresponding image sub-field is necessary since the distortion correction depends on said position(s).

[0029] According to one embodiment, the step of correcting one or more positions of the extracted geometrical feature in the image based on the vector distortion map of the corresponding image sub-field comprises determining a distortion vector for at least one position of the extracted geometrical feature. For example, if the center of a feature (the location of the feature) is a geometrical feature of said feature, it may be sufficient to determine only one distortion vector for this center location. For example, if the geometrical feature is an edge or a line, this edge or line is represented by multiple locations, so multiple distortion vectors need to be determined for each of the multiple locations, respectively. Similar considerations apply for geometrical features of other shapes.

[0030] According to one embodiment, each of the plurality of vector distortion maps is represented by a polynomial expansion in a vector polynomial. Thus, in principle, for any position or pixel in the image subfield, an associated distortion vector can be calculated. Alternatively, each of the plurality of vector distortion maps can be represented by a two-dimensional look-up table. In principle, other representations of the vector distortion "map" are possible.

[0031] The vector polynomial may be calculated, for example, as follows:

number

[0032] According to one embodiment, method steps b) to f) are performed repeatedly for a number of features, although method step a) is not necessarily repeated.

[0033] According to one embodiment, other areas in the image that do not contain features of interest are not dewarped, which significantly reduces the amount of computation and saves resources.

[0034] According to one embodiment, the step of extracting geometric properties of features of interest is performed on the entire image. In one example, the feature extraction results in a binary image with a relatively small data size. According to another example, the feature extraction results in the determination of the location of at least a geometric property (e.g., a center, a point, an edge, a contour, or a line).

[0035] According to one embodiment, the step of correcting one or more positions of the extracted geometric features in the image based on the vector distortion map of the corresponding image sub-field comprises transforming a pixel of the image into at least one pixel of the distortion-corrected image based on the distortion vector. This is due to the fact that the distortion correction does not necessarily shift the position of the whole pixel. In contrast, for example, one pixel can be shifted to two, three or four pixels (interpolation).

[0036] According to one embodiment, correcting one or more positions of the extracted geometric features in the image based on vector distortion of the corresponding image sub-field comprises transforming the image positions to distortion-corrected positions based on a distortion vector polynomial, the vector distortion polynomial being expressed by a vector polynomial expansion of a vector distortion map of the sub-field in sub-field coordinates (p,q), in global coordinates (x,y), or in both sets of coordinates.

[0037] According to one embodiment, the extracted geometrical property of the feature is divided into a number of parts, since it extends over a number of image sub-fields. In this case, one or more positions of each part of the extracted geometrical property are individually corrected based on the associated respective vector distortion map of the corresponding image sub-field of the respective part. Again, in principle, each part of the geometrical property is distortion corrected with respect to the vector distortion map of the image sub-field to which it belongs. In this way, the division of the feature into parts and the distortion correction for each part enables more accurate metrology applications.

[0038] According to one embodiment, the method comprises: determining dimensions of features of the semiconductor device in the distortion corrected image data; determining areas of semiconductor device features in the distortion corrected image data; determining locations of a plurality of regular objects, in particular HAR structures, in the semiconductor device in the distortion corrected image data; determining line edge roughness in the distortion corrected image data; and / or determining overlay errors between different features in the semiconductor device in the distortion corrected image data; The present invention further includes at least one of the following:

[0039] In each case, the determination / measurement is performed based on distortion-corrected image data, which may be represented, for example, as a set of position data or a binary image, thereby improving the accuracy of the determination or measurement.

[0040] According to one embodiment, the method comprises: Providing a test sample having a precisely known pattern, in particular a repeating pattern, that defines a target grid; imaging the test sample with a multi-beam charged particle microscope, analyzing the resulting image, and determining an actual lattice based on said analysis; determining a position deviation between the actual grating and the target grating; obtaining a vector distortion map for each image subfield based on the position deviation; The determination of the above-mentioned vector distortion map or vector distortion field is in principle known in the art from the imaging of a calibrated test sample. The accuracy of the obtained vector distortion map strongly depends on the manufacturing accuracy of the pattern on the test sample and on the measurement accuracy when analyzing the test sample.

[0041] According to an embodiment, the method further comprises the steps of shifting the test sample from a first position to a second position relative to the multi-beam charged particle microscope and imaging the test sample at the first and second positions. The stage is preferably moved for a shift of, for example, about half the image subfield. This method step contributes in particular to an increase in accuracy when high frequency structures / patterns that are statistically distributed across the sample are imaged.

[0042] According to one embodiment, the step of determining the position deviation between the actual grating and the target grating comprises a two-stage determination, in a first stage the shift of each image subfield, the rotation of each image subfield and the magnification of each image subfield are compensated, and in a second stage other, in particular higher order distortions are determined, the latter of which may be scanning-induced distortions. A clear distinction can therefore be made between scanning-induced distortions and other distortions.

[0043] According to one embodiment, the method further comprises updating the vector distortion map, for example at regular time intervals, on request by a user or whenever a configuration or operating parameters of the multi-beam charged particle microscope is changed.

[0044] According to a second aspect of the invention, the invention relates to a method for correcting distortions in an image composed of one or more image patches, each composed of a plurality of image subfields imaged respectively with associated beamlets of a multibeam charged particle microscope, comprising: g) providing, for each image subfield, a number of vector distortion maps each characterizing a position-dependent distortion for each pixel of the associated image subfield; h) for each pixel in the image, determining a corresponding image subfield that contains the pixel; i) for each pixel in the image, transforming the pixel in the image to at least one pixel in a distortion-corrected image based on a vector distortion map of a corresponding image subfield; The present invention relates to a method, comprising:

[0045] The definitions of the terms used above are the same as those explained or defined with respect to the first aspect of the invention. According to the second aspect of the invention, the distortion correction is performed on the whole distorted image, not on the extracted features. This can be performed by a PC, for example after combination by a multi-beam particle microscope.

[0046] According to a third aspect of the invention, the invention is directed to a computer program product comprising a program code for carrying out a method as described in any one of the embodiments described above with respect to the first and second aspects of the invention. The program code may be subdivided into one or more partial codes. For example, in one program part, it is appropriate to separately provide a code for controlling a multi-beam particle microscope, while in another program part, a routine for distortion correction is included. The distortion correction itself may be carried out, for example, on a PC.

[0047] According to a fourth aspect, the invention is directed to a multi-beam charged particle microscope comprising a controller configured to carry out the method as described above in the various embodiments.

[0048] According to a fifth aspect of the invention, the correction of scanning-induced distortions is performed during image pre-processing. This means that the correction is performed before the digitized image data is written into an image memory, which may be realized as a parallel access memory. For example, an FPGA ("field programmable gate array") is configured or programmed such that a spatially dependent distortion correction is performed for pixels representing an image sub-field. To realize each distortion correction, a filtering operation is realized by suitable hardware design / programming using a spatially varying filter kernel that takes into account the spatially varying distortions in the image sub-field, for example by referring to a vector distortion map determined for all image sub-fields as described above. To take into account the spatial variation of the filter kernel, a kernel generation unit is applied that calculates each filter kernel for each segment of the image sub-field separately and preferably "on the fly". While the distortion correction needs to be performed in parallel for the data streams of all beamlets, it needs to be numerically and individually adapted for the image sub-field / beamlet (imaging channel) of interest.

[0049] More specifically, the present invention relates to at least a first collectively raster scanner for collectively scanning the plurality of J primary charged particle beamlets over a plurality of J image subfields; a detection unit comprising a detector for detecting a plurality of J secondary electron beamlets, each corresponding to one of the J image subfields; a scan control unit connected to the first collective raster scanner and configured, in use, to control a raster scanning movement of the plurality of J primary charged particle beamlets by the first collective raster scanner; a kernel generation unit configured, in use, to generate a spatially varying filter kernel for correcting spatially varying distortion of an image subfield; an image data acquisition unit synchronized in operation with the detector, the scan control unit, and the kernel generation unit, for each of the J image subfields: an analog-to-digital converter for converting, in use, an analog data stream received from the detector into a digital data stream representative of an image sub-field; a hardware filter unit configured to receive the digital data stream and configured, in use, to generate a distortion-corrected data stream by performing a convolution of a segment of the image subfield with a spatially varying filter kernel; an image memory configured to store the distortion-corrected data stream as a 2D representation of an image subfield; an image data acquisition unit comprising: A controller (800, 820) comprising: The present invention relates to a multi-beam charged particle microscope comprising:

[0050] Characterization features according to a fifth aspect of the invention are a hardware filter unit and a kernel generation unit. A hardware filter unit configured to receive a digital data stream and further configured in use to generate a distortion-corrected data stream by performing a convolution with a spatially varying filter kernel of a segment of an image subfield is now realized for the first time in a multi-beam charged particle microscope. Since the distortion correction in an image subfield is not constant but varies within the image subfield, the filter kernel used likewise needs to be spatially varying. To take this spatial dependency into account, the application of the kernel generation unit allows the calculation / determination of a spatially varying filter kernel for each segment of the image subfield currently being filtered in the hardware filter unit.

[0051] Furthermore, in the case of multiple beamlets, the presence of multiple imaging channels in each case must also be taken into account. Therefore, the distortion correction needs to be performed for each imaging channel independently, in other words for each J image subfield. Therefore, the image data acquisition unit comprises an analog-to-digital converter, a hardware filter unit and an image memory for each imaging channel and thus for each J image subfield.

[0052] As mentioned above, distortion correction performed in image post-processing is usually only realized at the cost of huge computation time. However, if the image distortion correction for each image sub-field is performed by hardware filtering, the computation cost and the required energy can be significantly reduced. The effect of the hardware filtering itself is a reduction in the delay time during data generation before the data stream is stored in the image memory. The kernel generation unit can calculate the spatially-varying filter kernel for the spatially-varying distortion correction of each image sub-field "on the fly", but the computation cost of this filter kernel generation is rather moderate.

[0053] Of course, the operation of the various parts of the multi-beam charged particle microscope needs to be synchronized, for example by application of a clock signal and a counting unit; a feasibility recognized by those skilled in the art.

[0054] According to one embodiment of the present invention, the hardware filter unit comprises: a grid array of filter elements each comprising a first register for temporarily storing pixel values ​​and a second register for temporarily storing coefficients generated by the kernel generation unit, the pixel values ​​stored in the first register representing a segment of an image subfield; a plurality of multiplication blocks configured to multiply pixel values ​​stored in a first register by corresponding coefficients stored in a second register; a plurality of summation blocks configured to sum the results of the multiplications; As mentioned above, the hardware filter unit is configured, in use, to perform a convolution of a segment of the image sub-field with a spatially varying filter kernel. Mathematically, the convolution between two matrices can be expressed as a sum of products calculated from the entries in the matrices. In the context of the present invention, a first register stores the entries of the first matrix (pixel values ​​of a segment of the image sub-field) and the entries of the second matrix correspond to the coefficients generated by the kernel generation unit. A number of multiplication blocks are provided to perform the necessary multiplications of the entries in the two matrices with each other. Similarly, a number of summation blocks are provided for the necessary summation of the products.

[0055] The term grid arrangement is intended to indicate the internal relationship / situation of pixel values ​​and coefficients. A grid arrangement corresponds logically to a matrix representation.

[0056] Typically, the filtering is a neighbourhood operation. This means that the filter unit does not act on the whole image subfield, but only on a segment of the image subfield. Thus, according to an embodiment, the hardware filter unit comprises a number of shift registers configured to realise a grid arrangement of filter elements and to maintain the order of the data in the data stream as it passes through said hardware filter unit. By these means, the grid arrangement is a realisation of a segment of the image subfield and thus of pixels in the image subfield that are located in the neighbourhood of the image pixel to be distortion corrected. The shift registers typically have a predefined size (for example 512 bits, 1024 bits, 2048 bits or 4096 bits). The shift registers can therefore store a corresponding number of pixels. However, the size of the grid array of filter elements is typically much smaller. Typically, an image segment may comprise, for example, 11x11 filter elements, 21x21 filter elements or 31x31 filter elements. If the grid array of filter elements has a typical size AxA, a number of A shift registers may be applied. The first A entries in the shift register belong to a representation of a segment of the image subfield, and the other entries in the shift register are filled with the other pixels of the row (or column) of the image subfield, so essentially the size of the shift register limits the number of pixels in a row (or column) of the image subfield.

[0057] According to an embodiment of the present invention, the size of the grid array of filter elements is configured to correct distortions of at least 10 times the pixel size of the image subfield. This means that the size of the grid array of filter elements is at least 20x20, more precisely 21x21 entries. It should be noted that the number of filter elements in a row or column is usually chosen to be an odd number, since this allows the filter kernel to be represented in a symmetrical manner with a unique center. However, mathematically, it is also possible for the size of the grid array of the filter kernel to be an even number. Furthermore, the pixel size can be the same in different scanning directions, but can also be different in different scanning directions.

[0058] As an example, the pixel size of the image subfield can be 2 nm, and by applying a 20x20 or 21x21 filter kernel, approximately 20 nm of distortion can be corrected.

[0059] Generally, the size of the grid array of filter elements determines the maximum distortion that can be corrected, which is approximately half the size / dimension of the grid array multiplied by the pixel size in each dimension or direction.

[0060] According to one embodiment, the size of the lattice array corresponds to the size of the filter kernel. The number of multiplications that need to be performed is therefore the number of filter elements. However, the number of multiplications required grows quadratically with the number of pixels in a row or column. Thus, the amount of calculations increases, as does the number of logic units. This is because the hardware filter unit is realized in hardware. Therefore, it is preferable to reduce the number of logic units. According to one embodiment of the present invention, the size of the given kernel window is equal to or smaller than the size of the lattice array of filter elements. Here, it is necessary to take into account that the filtering according to the invention is performed for the purpose of distortion correction. Distortion correction can be understood as a shifting of pixels. This means that even if a full convolution of a full-size kernel filter with the pixel value stored in the first register of the filter element is performed, there are many multiplications that do not affect the result. In other words, the shifting of a pixel usually results in the "distribution" of that pixel to, for example, four other pixels. Thus, the kernel window reflects that part of the filter kernel, and the entries of the filter kernel affect the result. Other multiplications, which in theory could also be performed in a full convolution, can be omitted since they do not have any effect. This saves on logic units, more precisely on multiplication and summation blocks. Of course, one must take into account the position in the overall filter kernel where the kernel window needs to be placed. Consequently, according to one embodiment of the invention, the kernel generation unit is configured to determine, when in use, the position of the kernel window relative to the grid arrangement of the filter elements.

[0061] According to one embodiment, the hardware filter unit further comprises a number of switching means configured to logically combine the entries and filter elements in the multiplication blocks based on the position of the kernel window when in use. Thus, to reduce the number of multiplication blocks and summation blocks, the number of switching means (e.g. multiplexers) needs to be increased, which is still easier to implement.

[0062] According to an embodiment of the invention, the kernel generation unit is configured to determine the spatially varying filter kernel based on a vector distortion map characterizing the spatially varying distortion in the image sub-fields, for details describing the vector distortion map reference is made to the definitions and explanations given with respect to the first to fourth aspects of the invention.

[0063] According to one embodiment, the vector distortion map is represented by a polynomial expansion in a vector polynomial, or alternatively, the vector distortion map is represented by a multi-dimensional look-up table.

[0064] According to an embodiment, the kernel generation unit is configured to determine the filter kernel based on a function f that represents a pixel representatively. In other words, apart from the distortion subject itself, the filter kernel also takes into account the "shape" of the pixel. A possible function for representing a pixel could be, for example, the Rect2D function, which represents a rectangular pixel. This corresponds to a linear or bilinear filter. As the pixel may be blurred in the scanning direction, a possible function f could also be the function Rect(p,q), which blurs differently in different scanning directions p and q.

[0065] Alternatively, the function f representing the pixel may have the shape of the pixel's beam focus (e.g., Gaussian, anisotropic, cubic, sinusoidal, Airy pattern, etc.), and the filter is rounded at some low level value. Furthermore, according to one example, the filter is assumed to be energy-conserving, and therefore the high-order rounding filter kernel is assumed to be normalized so that the weights sum to 1. Alternatively, this normalization may be performed at a later stage rather than directly in the filter, although those skilled in the art will recognize the advantages and disadvantages of one particular implementation.

[0066] It should be noted that pixels at the boundaries of the image subfields become unusable. However, this effect is well known due to filtering processes in image post-processing. To handle this effect, cut-offs are required depending on the size of the filter kernel. This still does not cause any problems, since in multi-beam charged particle microscopes an overlap is usually realized between adjacent image subfields.

[0067] According to an embodiment, the image data acquisition unit further comprises a counter which, in use, is arranged to indicate the local coordinates p, q of a pixel in the image sub-field being filtered, which is relevant on the one hand for synchronization purposes and on the other hand for determining the individual spatially dependent scan-induced distortions in the image sub-field.

[0068] According to one embodiment, the image data acquisition unit further comprises an averaging unit implemented in the direction of the data stream after the analog-to-digital converter and before the hardware filter unit. The application of the averaging unit can increase the signal-to-noise ratio. Possible implementations are described in International Patent Application Publication No. WO2021 / 156198, the entire contents of which are incorporated herein by reference.

[0069] According to an embodiment, the image data acquisition unit further comprises another hardware filter unit configured in use to perform another filtering operation, in particular low-pass filtering, morphological operations and / or deconvolution with a point spread function. Naturally, the image data acquisition unit can comprise a number of other hardware filter units as well. Here, the principle applies that filtering operations are also realized by specially configured hardware and that it is not necessary to be forced to perform filtering operations in image post-processing.

[0070] According to one embodiment, the hardware filter unit comprises a field programmable gate array (FPGA) or an application specific integrated circuit (ASIC).

[0071] According to one embodiment, the hardware filter unit comprises a series of FIFOs, which can implement a grid array of filter elements by implementing a shift register as described above.

[0072] In one embodiment, the FIFO is implemented as a BlockRAM. In another embodiment, the FIFO can be implemented as a LUT (Look Up Table) or an externally connected SRAM / DRAM (Static or Dynamic Random Access Memory). Note that there are usually off-the-shelf IP blocks from the corresponding chip manufacturers to instantiate the hardware.

[0073] The embodiments of the fifth aspect of the present invention may be combined with each other in part or in whole, unless a technical contradiction arises.

[0074] Naturally, other realizations and configurations of the hardware filter unit are possible. According to a sixth aspect of the invention, the invention is directed to a system comprising a multi-beam charged particle microscope as described above in the many embodiments, and an image post-processing unit configured to perform distortion correction of image data. The image post-processing unit may be additionally provided to the multi-beam charged particle microscope. This may include, for example, an additional PC. However, as an alternative, the image post-processing unit may also be included in the multi-beam charged particle microscope. The image post-processing unit may be configured to perform distortion correction by image post-processing as described above with respect to the first aspect of the invention. Importantly, according to this embodiment of the invention, two different types of distortion correction may be combined with each other. A first distortion correction may be performed in image pre-processing (realized as data stream processing) and a second distortion correction may then be performed in image post-processing, preferably only on the extracted geometrical properties of the feature of interest. In this respect, explicit reference is made to the description of the first aspect of the invention.

[0075] Likewise, different aspects of the invention may be combined with each other, in whole or in part, unless a technical contradiction arises. Definitions given with respect to one aspect of the invention are also valid for other aspects of the invention.

[0076] According to one example, in a first step, regularly occurring scanning-induced distortions are corrected according to the second embodiment of the invention (distortion correction is performed during image pre-processing), and then in a second step, further residual distortions can be corrected according to the first embodiment of the invention (scanning-induced distortions are corrected during image post-processing).

[0077] The invention may be better understood with reference to the accompanying drawings. [Brief description of the drawings]

[0078] [Figure 1] FIG. 1 illustrates a multi-beam charged particle microscope system according to one embodiment. [Diagram 2] FIG. 2 illustrates the coordinates of a first test site and a second test site including first and second image patches. [Diagram 3] FIG. 1 illustrates the static distortion offsets of multiple primary charged particle beamlets (3). [Figure 4a] FIG. 13 illustrates the scan deflection in the scan deflector for axial beamlets. [Figure 4b] FIG. 13 shows the scan deflection at the scan deflector versus the scan induced distortion for an off-axis beamlet with a propagation angle β. [Diagram 5] FIG. 13 illustrates scanning-induced decentering aberrations for off-axis beamlets with propagation angle β. [Figure 6] FIG. 13 illustrates typical scan-induced distortion of a single beamlet when scanned across an image subfield with image subfield coordinates (p,q). [Figure 7] FIG. 1 is a diagram illustrating distortion correction in general image processing. [Figure 8] FIG. 1 illustrates distortion correction and subsequent feature extraction in a grayscale image. [Figure 9] FIG. 2 illustrates feature extraction and subsequent distortion correction according to the present invention. [Figure 10] 4 is a flow chart of a method for determining a distortion-corrected position of a feature in accordance with the present invention. [Figure 11] FIG. 13 illustrates the determination of a vector distortion map based on a target grid. [Figure 12] FIG. 1 illustrates the determination of distortion vectors. [Figure 13] FIG. 13 is a diagram illustrating the determination of grid points. [Figure 14] FIG. 13 shows dimension measurements based on distortion-corrected image data. [Figure 15] FIG. 13 illustrates a statistical evaluation of regular object positions based on distortion-corrected image data. [Figure 16] FIG. 2 shows an image data acquisition unit and associated units or modules. [Figure 17] FIG. 2 illustrates a hardware filter unit. [Figure 18] FIG. 13 shows the convolution of a segment of an image subfield with a filter kernel. [Figure 19] FIG. 2 illustrates a filter element and some associated elements. [Figure 20] FIG. 2 illustrates a hardware filter unit with a 3×3 filter kernel window. [Figure 21] FIG. 2 illustrates a hardware filter unit with a 2×2 filter kernel window. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0079] In the exemplary embodiments described below, functionally and structurally similar components are denoted by similar or identical reference numerals whenever possible.

[0080] The schematic representation of Fig. 1 illustrates basic features and functions of a multi-beam charged particle microscope system 1 according to some embodiments of the present invention. It should be noted that the symbols used in the figure are chosen to symbolically represent their respective functions, rather than to represent the physical configuration of the components shown. The system as shown is that of a multi-beam scanning electron microscope (MSEM or multi-SEM) using multiple primary electron beamlets 3 to generate multiple primary charged particle beam spots 5 on the surface of an object 7, such as a wafer, whose upper surface 25 is located in the object plane 101 of an objective lens 102. For simplicity, only five primary charged particle beamlets 3 and five primary charged particle beam spots 5 are shown. The features and functions of the multi-beamlet charged particle microscope system 1 can be implemented using other types of primary charged particles, such as electrons or ions, in particular helium ions.

[0081] The microscope system 1 comprises an object illumination unit 100, a detection unit 200 and a beam splitter unit 400 for separating the secondary charged particle beam path 11 from the primary charged particle beam path 13. The object illumination unit 100 comprises a charged particle multi-beam generator 300 for generating a plurality of primary charged particle beamlets 3 and is configured to focus the plurality of primary charged particle beamlets 3 by means of a sample stage 500 onto an object plane 101 where a surface 25 of a wafer 7 is located.

[0082] The primary beam generator 300 generates a number of primary charged particle beamlet spots 311 in an intermediate image plane 321, which is usually a spherically curved surface to compensate for the field curvature of the object illumination unit 100. The primary beamlet generator 300 comprises a primary charged particle (e.g. electron) source 301. The primary charged particle source 301 emits a divergent primary charged particle beam 309, which is collimated by at least one collimator lens 303 to form a parallel beam. The collimator lens 303 typically consists of one or more electrostatic or magnetic lenses, or a combination of electrostatic and magnetic lenses. The collimated primary charged particle beam is incident on a primary multi-beam construction unit 305. The multi-beam construction unit 305 essentially comprises a first multi-aperture plate 306.1, which is illuminated by the primary charged particle beam 309. The first multi-aperture plate 306.1 comprises a plurality of apertures in a raster arrangement for generating a plurality of primary charged particle beamlets 3 by transmission of a collimated primary charged particle beam 309. The multi-beamlet arrangement unit 305 comprises at least other multi-aperture plates 306.2 and 306.3 arranged downstream of the first multi-aperture plate 306.1 with respect to the direction of movement of the electrons of the beam 309. For example, the second multi-aperture plate 306.2 has the function of a microlens array and is preferably set to a defined potential so that the focal positions of the plurality of primary beamlets 3 in the intermediate image plane 321 are adjusted. The third multi-aperture active plate arrangement 306.3 (not shown) comprises an individual electrostatic element for each of the plurality of apertures, thereby exerting an individual influence on each of the plurality of beamlets. The multi-aperture active plate arrangement 306.3 consists of one or more multi-aperture plates with electrostatic elements such as circular electrodes, multi-pole electrodes or a series of multi-pole electrodes for microlenses constituting a static deflector array, a microlens array or an astigmatism corrector array. The multi-beamlet arrangement unit 305 comprises an adjacent first electrostatic field lens 307 which, together with a second field lens 308 and the second multi-aperture plate 306.2, focuses the multiple primary charged particle beamlets 3 onto or near the intermediate image plane 321.

[0083] Located at or near the intermediate image plane 321 is a static beam steering multi-aperture plate 390 including a number of apertures with electrostatic elements (e.g. deflectors) for individually steering each of the multiple charged particle beamlets 3. The apertures of the beam steering multi-aperture plate 390 have a large diameter to allow the passage of the multiple primary charged particle beamlets 3 even if their focal spots are offset from the intermediate image plane or their respective lateral design positions. In one example, the beam steering multi-aperture plate 390 can also be configured as a single multi-aperture element.

[0084] The multiple focal points of the primary charged particle beamlets 3 passing through the intermediate image plane 321 are imaged by the field lens group 103 and the objective lens 102 in the object plane 101, where the investigation surface 25 of the object 7 is located. The object illumination system 100 further comprises a multi-beam raster scanner 110 capable of deflecting the multiple charged particle beamlets 3 in a direction perpendicular to the beam propagation direction or the optical axis 105 of the objective lens 102, close to the first beam crossover 108. In the example of FIG. 1, the optical axis 105 is parallel to the z-direction. The objective lens 102 and the multi-beam raster scanner 110 are centered on the optical axis 105 of the multi-beamlet charged particle microscope system 1, which is perpendicular to the wafer surface 25. The wafer surface 25, which is located in the image plane 101, is then raster scanned by the multi-beam raster scanner 110. This causes a plurality of primary charged particle beamlets 3 constituting a plurality of beam spots 5 arranged in a raster configuration to be synchronously scanned over the wafer surface 101. In one example, the raster configuration of focal spots 5 of the plurality of primary charged particle beamlets 3 is a hexagonal raster of approximately 100 or more primary charged particle beamlets 3. The primary beam spots 5 are approximately 6 μm to 15 μm apart and have a diameter of less than 5 nm (e.g., 3 nm, 2 nm, or less). In one example, the beam spot size is approximately 2 nm and the distance between two adjacent beam spots is 8 μm. At each scanning position of each of the plurality of primary beam spots 5, a plurality of secondary electrons are generated to constitute a plurality of secondary electron beamlets 9 in the same raster configuration as the primary beam spots 5. The intensity of the secondary charged particle beamlets generated at each beam spot 5 depends on the intensity of the impinging primary charged particle beamlets 3 illuminating the corresponding spot, the material composition and the topography of the object 7 under the beam spot 5. The secondary charged particle beamlets 9 are accelerated by an electrostatic field generated by the sample charging unit 503, collected by the objective lens 102 and guided by the beam splitter 400 to the detection unit 200. The detection unit 200 images the secondary electron beamlets 9 onto an image sensor 207 to form a number of secondary charged particle image spots 15. The detector may include a number of detector pixels or individual detectors.The intensity of each of the multiple secondary charged particle beam spots 15 is detected separately to detect the material composition of the wafer surface 25 with high resolution, achieving a large image patch with high throughput. For example, a raster of 10x10 beamlets with 8 μm pitch generates an image patch of about 88 μm x 88 μm in one image scan by the collective multi-beam raster scanner 110, with an image resolution of, for example, 2 nm or less. For example, the image patch is sampled at half the beam spot size, resulting in 8000 pixels per image line for each beamlet, and a digital data set representing an image patch generated by 100 beamlets contains 6.4 billion pixels. The image data is collected by the control unit 800. Details of the collection and processing of image data, for example using parallel processing, are described in German Patent Application No. 102019000470.1 and US Pat. No. 9,536,702, which are incorporated herein by reference.

[0085] The secondary electron beamlets 9 are deflected when passing through the first collective multi-beam raster scanner 110 and guided by the beam splitter unit 400 to follow the secondary beam path 11 of the detection unit 200. The secondary electron beamlets 9 travel in the opposite direction to the primary charged particle beamlets 3, and the beam splitter unit 400 is configured to separate the secondary beam path 11 from the primary beam path 13, typically by a magnetic field or a combination of a magnetic field and an electrostatic field. Optionally, additional magnetic correction elements 420 are also present in the primary or secondary beam path. The projection system 205 further comprises at least a second collective raster scanner 222 connected to a projection system control unit 820, more generally an imaging control module 820. The control unit 800 is configured to compensate for residual errors in the positions of the multiple foci 15 of the multiple secondary electron beamlets 9, such that the positions of the multiple secondary electron focal spots 15 remain constant on the image sensor 207.

[0086] The projection system 205 of the detection unit 200 includes other electrostatic or magnetic lenses 208, 209, 210 and a second crossover 212 of the multiple secondary electron beamlets 9, where an aperture 214 is arranged. In one example, the aperture 214 further comprises a detector (not shown) connected to a projection system control unit 820. The projection system control unit 820 is further connected to at least one electrostatic lens 206 and a third deflection unit 218. The projection system 205 further comprises at least a first multi-aperture corrector 220 with an aperture and an electrode having an individual influence on each of the multiple secondary electron beamlets 9, and optionally another active element 216 (e.g. a multipole element connected to the control unit 800).

[0087] The image sensor 207 is configured with an array of sensing areas in a pattern that matches the raster arrangement of the secondary electron beamlets 9 focused on the image sensor 207 by the projection lens 205. This allows individual secondary electron beamlets 9 to be detected independently of other secondary electron beamlets 9 incident on the image sensor 207. A plurality of electrical signals are generated, converted into digital image data and processed in the control unit 800. During an image scan, the control unit 800 is configured to trigger the image sensor 207 to detect a plurality of time resolved intensity signals from the plurality of secondary electron beamlets 9 in a predefined time interval, and digital images of image patches from all scanning positions of the plurality of primary charged particle beamlets 3 are accumulated and stitched together.

[0088] The image sensor 207 shown in FIG. 1 can be an electron-sensitive detector array, such as a CMOS or CCD sensor. Such an electron-sensitive detector array can include an electron-photon conversion unit, such as a scintillator element or an array of scintillator elements. In one example, the image sensor 207 can be configured as an electron-photon conversion unit or scintillator plate located at the focal plane of the multiple secondary electron particle image spots 15. In this example, the image sensor 207 can further include relay optics for imaging and guiding the photons generated by the electron-photon conversion unit at the secondary charged particle image spots 15 on dedicated photon detection elements, such as multiple photomultiplier tubes or avalanche photodiodes (not shown). Such an image sensor is disclosed in the above-cited U.S. Pat. No. 9,536,702. In one example, the relay optics further includes a beam splitter for splitting and guiding the light to a first slow photodetector and a second fast photodetector. The second high speed photodetector is formed by a photodiode array, e.g. avalanche photodiodes, and is fast enough to resolve the image signals of the multiple secondary electron beamlets 9 according to the scanning speed of the multiple primary charged particle beamlets 3. The first slow photodetector is preferably a CMOS or CCD sensor, which monitors the multiple secondary electron beamlets 9 or focal spots 15 and provides high resolution sensor data signals for controlling the operation of the multi-beam charged particle microscope.

[0089] In this example, the primary charged particle source is implemented in the form of an electron source 301 featuring an emitter tip and extraction electrodes. If primary charged particles other than electrons are used, such as helium ions, the configuration of the primary charged particle source 301 can differ from that shown. The primary charged particle source 301, the active multi-aperture active plate configurations 306.1...306.3 and the beam steering multi-aperture plate 390 are controlled by a primary beamlet control module 830 which is connected to the control unit 800.

[0090] Preferably, the stage 500 does not move during acquisition of an image patch by scanning the multiple primary charged particle beamlets 3, but moves after acquisition of the image patch to the next image patch to be acquired. In an alternative embodiment, the stage 500 moves continuously in a second direction while an image is acquired by scanning the multiple primary charged particle beamlets 3 in a first direction by the collective multi-beam raster scanner 110. The stage movements and stage positions are monitored and controlled by sensors known in the art, such as laser interferometers, grating interferometers, confocal microlens arrays, etc.

[0091] The method of wafer inspection by acquiring image patches is explained in more detail in Fig. 2. The wafer is positioned with its wafer surface 25 in the focal plane of a number of primary charged particle beamlets 3 at the centre 21.1 of a first image patch 17.1. Predefined positions of the image patches 17.1...k correspond to inspection sites of the wafer for inspection of semiconductor features. The application is not limited to the wafer surface 25 but is also applicable, for example, to lithography masks used in semiconductor processing. The word "wafer" is therefore not limited to semiconductor wafers but is intended to include general objects used or processed during semiconductor processing.

[0092] The predefined positions of the first inspection site 33 and the second inspection site 35 are loaded from an inspection file in a standard file format. The predefined first inspection site 33 is divided into a number of image patches (e.g., a first image patch 17.1 and a second image patch 17.2), and a first center position 21.1 of the first image patch 17.1 is aligned under the optical axis 105 of the multi-beam charged particle microscope system 1 for a first image acquisition step of the inspection task. The first center 21.1 of the first image patch is selected as the origin of a first local wafer coordinate system for the acquisition of the first image patch 17.1. It is well known in the art how to align the wafer 7 such that the wafer surface 25 is positioned and a local coordinate system of wafer coordinates is generated.

[0093] The plurality of primary beamlets 3 are distributed in a substantially regular raster configuration in each image patch 17.1···k, and a digital image of the image patch is generated by scanning with a raster scanning mechanism. In this example, the plurality of primary charged particle beamlets 3 are arranged in a rectangular raster configuration with N primary beam spots 5.11, 5.12-5.1N in a first line as N beam spots, and M lines including beam spots 5.11-5.MN. For simplicity, M=5×N=5 beam spots are illustrated, but the number of beam spots J=M×N can be larger (e.g. J=61 beamlets or about 100 beamlets or more) and the plurality of beam spots 5.11-5.MN can have various raster configurations, such as hexagonal or circular.

[0094] Each primary charged particle beamlet is scanned on the wafer surface 25, and as shown, in the example of the primary charged particle beamlet including beam spots 5.11 and 5.MN, there are scan paths 27.11 and 27.MN. The scanning of each of the multiple primary charged particles is performed, for example, by moving back and forth along scan paths 27.11...27.MN, while each focal point 5.11...5.MN of each primary charged particle beamlet is moved by scanning the collective multi-beam scanning deflection system 110 in the x-direction from the start position of the image subfield line, which is, for example, the leftmost image point of the image subfield 31.mn in this example. Then, after each focal point 5.11...5.MN is scanned collectively by the collective scanning of the primary charged particle beamlet 3 to the right position, the collective multi-beam raster scanner 110 moves each of the multiple charged particle beamlets in parallel to the line start position of the next line in each subfield 31.11...31.MN. The movement back to the line start position of the subsequent scan line is called a flyback. The primary charged particle beamlets 3 follow the scan paths 27.11-27.MN in substantially parallel fashion, so that multiple scan images of each of the sub-fields 31.11-31.MN are obtained in parallel. As for the image acquisition, as described above, multiple secondary electrons are emitted at the focal points 5.11-5.MN, generating multiple secondary electron beamlets 9. The multiple secondary electron beamlets 9 are collected by the objective lens 102, pass through the first collective multi-beam raster scanner 110, guided to the detection unit 200, and detected by the image sensor 207. The continuous stream of data of each of the multiple secondary electron beamlets 9 is converted into multiple two-dimensional data sets in synchronization with the scan paths 27.11-27.MN, constituting the digital image data of each image sub-field. The multiple digital images of the multiple image sub-fields are finally stitched together by the image stitching unit to constitute the digital image of the first image patch 17.1. Each image subfield has a small overlap area with the adjacent image subfield, as shown by the overlap area 39 of subfield 31.mn and subfield 31.m(n+1).

[0095] Next, the requirements or specifications of the wafer inspection task are described. For high throughput wafer inspection, the image acquisition time of each image patch 17.1···k must be fast, including the time required for image post-processing. Meanwhile, strict specifications on image quality such as image resolution, image accuracy, and repeatability must be maintained. For example, the image resolution requirement is typically 2 nm or less with high repeatability. Image accuracy is also referred to as image fidelity. For example, the edge position of a feature, typically the absolute position accuracy of a feature, is determined with high absolute accuracy. Typically, the position accuracy requirement is approximately 50% or less than the resolution requirement. For example, a measurement task requires absolute accuracy of the dimensions of a semiconductor feature with an accuracy of less than 1 nm, 0.3 nm, or even 0.1 nm. Therefore, the lateral position accuracy of each of the focal spots 5 of the multiple primary charged particle beamlets 3 must be less than 1 nm (for example, less than 0.3 nm or even less than 0.1 nm). Under high image repeatability, it is understood that repeated image acquisition of the same area produces first and second repeated digital images, and the difference between the first and second repeated digital images is below a predetermined threshold. For example, the difference in image distortion between the first and second repeated digital images should be less than 1 nm (for example, less than 0.3 nm or preferably less than 0.1 nm), and the difference in image contrast should be less than 10%. In this way, repeated imaging operations can produce similar image results. This is important, for example, for image acquisition and comparison of similar semiconductor structures on different wafer dies, or for comparison of acquired images against representative or reference images obtained from CAD-based image simulations or databases.

[0096] One of the requirements or specifications of a wafer inspection task is the throughput. The measurement area per acquisition time depends on the dwell time, pixel size and number of beamlets. Typical examples of dwell times are 2ns to 800ns. Thus, the pixel rate in the high-speed image sensor 207 ranges from 1.25Mhz to 500MHz, and it is also possible to acquire about 15 to 20 image patches or frames per minute. For 100 beamlets, a typical example of throughput in high resolution mode with a pixel size of 0.5nm is about 0.045mm2 / min (square millimeters per minute), and for an increased number of beamlets (e.g., 10000 beamlets and a dwell time of 25ns), a throughput of more than 7mm2 / min is possible. However, in prior art systems, the requirement for digital image processing significantly limits the throughput. For example, the digital compensation of scanning distortions in the prior art is very time-consuming and therefore undesirable.

[0097] The imaging performance of the charged particle microscope 1 is limited by the design and higher order aberrations of the electrostatic or magnetic elements of the object projection unit 100, as well as the manufacturing tolerances of the primary multi-beamlet configuration unit 305, for example. The imaging performance is limited by aberrations such as distortion, focus aberration, decentering, and astigmatism of the multiple charged particle beamlets. FIG. 3 shows as an example the usual static distortion aberrations of multiple primary charged particle beamlets 3 in the image plane 101. The multiple primary charged particle beamlets 3 are focused in the image plane to form multiple primary charged particle beam spots 5 (three shown) in a raster configuration (hexagonal raster in this example). In an ideal system, the beam spots 5 are formed at the center positions 29.mn (see FIG. 2) of the corresponding image subfields 31.mn (subscript m is the line number and n is the column number) with the collective multi-beam raster scanner 110 switched off. However, in real systems, the beam spot 5 is formed at a position that deviates slightly from the ideal position on the ideal raster, as shown by the static distortion vector in FIG. 3. In the illustrated example of the primary beam spot 141, the deviation from the ideal position on the hexagonal raster is represented by the distortion vector 143. The distortion vector gives the lateral difference [dx, dy] from the ideal position, and the maximum absolute value of the distortion vector can be in the range of several nanometers (e.g., more than 1 nm, 2 nm or even more than 5 nm). Typically, the static distortion vector of a real system is measured and compensated for by an array of static deflection elements, such as one of the multi-aperture active plate configurations 306.2. Drift or dynamic changes in the static distortion are also taken into account and compensated for as described in German Patent Application No. 102020206739.2, filed May 28, 2020, which is incorporated herein by reference. Control and compensation of the aberrations is achieved by a monitoring or detection system and a control loop which may drive a compensator, for example multiple times during an image scan, thereby compensating for the aberrations of the multi-beam charged particle microscope 1 .

[0098] However, the imaging performance of a charged particle microscope is limited not only by design and drift aberrations of electrostatic or magnetic elements of the object projection unit 100, but also in particular by the first collective multi-beam raster scanner 110. For single-beam microscopes, deflection scanning systems and their respective characteristics have been thoroughly investigated. However, for multi-beam microscopes, conventional deflection scanning systems for scanning deflection of multiple charged particle beamlets exhibit specific characteristics, which are shown in more detail in the beam path through the deflection scanner in FIG. 4.

[0099] FIG. 4a shows the beam path of a single primary charged particle beam through a prior art scanning deflector 110 with deflection electrodes 153.1 and 153.2 and a voltage source. For simplicity, only the deflection scanner electrodes for raster scanning deflection in the first direction are shown. In use, a scanning deflection voltage difference VSp(t) is applied and an electrostatic field is formed with equipotential lines 155 between electrodes 153.1 and 153.2. An axial charged particle beamlet 150a corresponding to image patch 31.c with image patch center 29.c coincides with the optical axis 105 and is deflected by the electrostatic field to pass through an intersection volume 189 between deflection electrodes 153.1 and 153.2 along a real beam path 151f. The beam trajectory can be approximated by primary beam paths 150a and 150f with a single virtual deflection at pivot point 159. Charged particle beamlets traveling along path 150z are focused by objective lens 102 into an object plane 101 shown in the lower part of Figure 4a. The subfield coordinates are given in (p,q) relative to the centre point 29.c of the subfield 31.c.

[0100] coordinate p f For maximum deflection to the maximum subfield point, the maximum voltage difference VSp max is applied, and the distance p zIn the case of deflection of the incident beamlet 150a to a subfield point of p, a corresponding voltage VSp is applied and the incident beamlet 150a is deflected by a deflection angle α in the direction of the beam path 150z. The nonlinearity of the deflector is compensated by determining the functional dependence of the deflection angle α and the deflector voltage difference VSp. By calibrating the functional dependence VSp(sin(α)), an almost ideal scanner for a single charged particle beamlet is realized with a single common pivot point 159 for the deflection scanning of a single primary charged particle beamlet. It should be noted that the lateral displacement (p,q) of the beam spot position in the image plane is proportional to the focal length f of the objective lens 102 multiplied by sin(α). In the example of a strip-shaped field point, p z =fsin(α z ) for small angles α, the function sin(α) is usually approximated by α. As explained in more detail below, despite the fact that scanning-induced distortions can be minimized in the case of single-beam microscopes, other scanning-induced aberrations such as astigmatism, defocus, coma, or spherical aberration can degrade the resolution of charged particle microscopes with large field sizes. Also, as the field size increases, deviations from the virtual pivot point 159 become increasingly significant.

[0101] In a multi-beam system, multiple charged particle beamlets are scanned in parallel by the same deflection scanner with the same voltage difference according to the functional dependency VSp(sin(α)). In FIG. 4b, the crossover 108 of multiple primary charged particle beamlets coincides with the virtual pivot point 159 of the axial primary beamlet 150a, and each charged particle beamlet passes through the electrostatic field at a different angle. A charged particle beamlet 157a is shown with an incidence angle β together with the corresponding subfield 31.o with the image subfield center 29.o. The angle β is related to the distance X from the center coordinate 29.o to the optical axis 105 by sin(β)=X / f using the focal length f of the objective lens 102. With the deflection scanner 110 switched off (VSp(t)=0V), the beamlet travels along the path 157a and is focused by the objective lens 102 to the center point 29.o of the subfield 31.o. However, despite the fact that the deflection scanner is nearly ideal for axial beamlets when a voltage difference is applied, as shown in FIG. 4a, it is not ideal for field beamlets with incidence angles β. Due to the finite thickness of the deflection field, the path length through the electrostatic field is different for incident beamlets with different incidence angles β, and the real beam paths 157z and 157f deviate from the ideal primary beam paths 163z and 163f. This is called the coordinate p z and p f The beam paths for the two subfield points are shown as real beam paths 157z and 157f. The angles of the real beam paths 157z and 157f deviate from the angles of the ideal beam paths 163z and 163f, and each beam is virtually deflected at a different virtual pivot point 161z and 161f that deviates from the beam crossover 108. For example, when a voltage Vsp(sin(α0)) is applied, the primary charged particle beamlet 157a is deflected by an angle α1 instead of an angle α0 and follows the beam path 157z that includes the virtual deflection point 161z. Thus, the charged particle beam spot is distorted by a local distortion vector dpz.

[0102] The deviation in deflection angle increases with increasing incidence angle β, resulting in large scanning-induced distortions caused by the collective multi-beam raster scanner 110 .

[0103] The difference in the deflection angle α causes the scanning-induced distortion, and the difference in the location of the virtual pivot point causes the scanning-induced decentering aberration. FIG. 5 shows a simplified system 171 in front of the scanning collective multi-beam raster scanner 110, from which a plurality of primary charged particles enter the first collective multi-beam raster scanner 110. The plurality of charged particle beamlets are shown by two beamlets, including an axial charged particle beamlet 3.0 and an off-axis beamlet 3.1, which pass through an intersection volume 189 of the raster scanner 110 and are focused by the objective lens 102 to form a plurality of foci shown as foci 5.0 and 5.1 on the surface 25 of the wafer 7. When the raster scanner 110 is in the off state and no voltage difference VSp is applied to the electrode 153, the beam spots 5.0 and 5.1 are at the center points 29.0 and 29.1 of each image subfield. When a voltage difference VSp(sin(α0)) is applied, the beamlet 3.0 is deflected along an ideal path 150 to a strip field point Z0. In the linear representation of FIG. 5, the beamlet 3.0 is seen to be deflected at the beam crossover 108, which corresponds to the virtual pivot point 159 of FIG. 4a. Thus, the beamlet 3.0 illuminates the wafer surface 25 with the same angle of incidence as the central position 29.0. The off-axis beamlet 3.1 is seen to be deflected to a corresponding strip field point Z1 in the corresponding image subfield. The off-axis beamlet 3.1 is seen to be deflected along a representative beam path 157 at a virtual deflection point 161 that deviates from the beam crossover 108. Thus, the decentering angle of the beamlet 3.1 at the scan position relative to the strip field point Z1 deviates from the decentering angle at the central field 29.1, which corresponds to the above-mentioned distortions as well as the scan-induced decentering aberrations of the beamlet 3.1. In a third embodiment of the present invention, a second multi-beam scanning correction system 602 reduces scan-induced decenter aberrations.

[0104] The deviation of the focal position at the scan position of each of the multiple charged particle beamlets 3 is represented by a scan distortion vector field (also called vector distortion map) for each of the image subfields 31.11-31.MN. FIG. 6 shows the scan distortion for the example of image subfield 31.15 (see FIG. 7). Throughout this disclosure, image subfield coordinates (p,q) for each center of each image subfield 31.mn are used and the scan distortion is represented by a vector [dp,dq] as a function of the image subfield coordinates (p,q) for each image subfield 31.mn. The center position (p,q)=(0,0) of each image subfield is represented by its (x,y) coordinates relative to the optical axis 105. Each image center coordinate can be distorted from a predetermined ideal raster configuration as a function of the (x,y) coordinates by a static offset (dx,dy) as shown in FIG. 3. Static distortion is usually compensated for by the static multi-aperture plate 306.2 and is not taken into account in the scanning distortion [dp,dq]. Since the scanning distortion is different in each image subfield 31.11...31.MN, the scanning distortion vector [dp,dq]=[dp,dq](p,q;x ij ,y ij The four coordinates are the local coordinates (p, q) of the image subfield and the discrete center coordinates (x ij ,y ij )

[0105] FIG. 6 shows the scanning distortion vector [dp,dq] over the entire image subfield 31.15. In this example, the maximum scanning distortion is at the maximum image subfield coordinate p=q=6 μm, with a scanning distortion vector [dp,dq]=[2.7 nm,-1.6 nm]. The length of the maximum scanning distortion vector in this image subfield is 3.5 nm. Typical maximum scanning distortion aberrations in the image subfield are in the range of 1 nm to 4 nm, but can exceed 5 nm.

[0106] FIG. 7 illustrates a diagram of distortion correction in general image processing. Image distortion correction itself is well known in the art. Image distortion correction is then performed in image post-processing. The distortion correction can be represented as a displacement of pixels by a position-dependent displacement vector, since the distortion is different for each pixel. The position-dependent displacement vector can be mathematically represented by the result of a matrix-vector multiplication. Furthermore, it must be taken into account that the distortion is usually not given over the entire pixel. In other words, in addition to the simple displacement, it is necessary to perform an interpolation of the pixel values. These facts are illustrated diagrammatically in FIG. 7. A pixel 700 is displaced by a distortion, and the resulting pixel position is indicated by reference symbol 700'. The value of pixel 700 is set to 1. Due to the displacement, the value or intensity 1 should be distributed over four pixels in the distortion-corrected image. Each pixel has an intensity / value l1, l2, l3, and l4.

[0107] For the whole image dewarping by image processing, this is computationally expensive: for each original pixel in the distorted image, a multiplication with an n x m matrix needs to be performed, and also an interpolation needs to be performed. As an example, an image of a multi-beam charged particle microscope contains 10 billion pixels. Thus, dewarping requires at least 40 billion operations, which is a huge amount, since four operations are required per pixel, plus the interpolation.

[0108] However, what really matters in metrology is the exact position of the image detail. According to the invention, the position of the image detail is determined in the original distorted image and then the distortion is corrected. For example, if the aim is to determine the position of HAR structures (high aspect ratio structures) in a semiconductor sample, the computational effort can be reduced by a factor of about 100,000 (100×100 μm 2 (Assume the image field of view contains 10 billion pixels and the HAR structures have a diameter of about 100 nanometers and a pitch of about 300 nanometers.)

[0109] According to the invention, the distortion is determined for each image subfield 31.mn with respect to a vector distortion map 730, since the distortion is different for each image subfield 31.mn and varies in each image subfield 31.mn. The generation of the vector distortion map is known per se. The distortion in each image subfield 31.mn can be represented, for example, by a polynomial expansion in a vector polynomial. This is known in principle, for example, by measuring a calibrated object. The object or test sample can also be displaced between a first and a second measurement, and the distortion can be determined based on the difference between the two measurements. These measurements can also be performed repeatedly. Thus, the distortion can be determined. The distortion, more precisely the representation as a vector distortion map 730 and / or a polynomial expansion in a vector polynomial, can be stored in a memory for each image subfield. It can also be updated at predefined time intervals.

[0110] 8 and 9 show, on the one hand (FIG. 8) conventional image processing and, on the other hand (FIG. 9), distortion correction according to the invention. More precisely, FIG. 8A shows a grayscale image 702. The grayscale image 702 can in principle be the whole image, only an image patch or only an image subfield. There is no difference in the explanation of the principle. The grayscale image 702 contains three features of interest 701a, 701b and 701c. In principle, these features 701a, 701b and 701c can be distorted. Here, the distortion is illustrated in the curved feature 701c. The original grayscale image 702 has been distortion corrected according to the state of the art, this distortion correction being performed for all pixels of the grayscale image. The result is shown in FIG. 8B. The feature 801c is no longer distorted (no longer curved). In a next step, the contours of features 701a, 701b, and 701c are extracted from the grayscale image 702 to generate a binary image 710 as shown in Fig. 8C. Based on the contours in the binary image 710, precision measurement or metrology applications can be performed. Note that for purposes of illustration and distinction, the grayscale image 702 includes a dotted background, and the binary image 710 includes a white background.

[0111] Now referring to Fig. 9, this figure shows the correction process according to the invention, but with the same original situation as shown in Fig. 9A. However, first all features of interest are then identified and extracted. Fig. 9B shows a binary image 710 containing only the contours of features 701a, 701b and 701c. These contours are still distorted. However, the amount of data in this binary image is significantly reduced compared to the state of the art grayscale image. Then, in a next step, the contours of features 701a, 701b and 701c are corrected. Now, due to the nature of the distortions, which are scanning-induced distortions, the correction is performed separately for each image subfield, and the correction of each pixel in each image subfield 31.mn is position-dependent.

[0112] For illustrative reasons, Fig. 9 shows a simplified approach for improved correction of scanning-induced distortion. According to another example of a method for correcting scanning-induced distortion, the positions of at least the features of interest 701a, 701b, 701c are extracted from the uncorrected digital image, and distortion correction is applied only to the positions of the features of interest 701a, 701b, 701c, for example by polynomial expansion of the vector distortion map. Thus, distortion correction is not limited to the pixel raster of the digital image.

[0113] Thus, more generally, the illustration shown in Figure 9B may alternatively be interpreted as a visualization of connected line segments consisting of a set of non-integer positions or coordinates of features of interest 701a, 701b, 701c obtained by feature extraction from grayscale image 702. Similarly, Figure 9C may be interpreted as a visualization of connected line segments consisting of non-integer positions or coordinates of distortion-corrected features of interest 701a, 701b, 701c.

[0114] Fig. 10 shows a flow chart of a method for determining a distortion-corrected position of a feature 701 in an image composed of one or more image patches, each composed of a plurality of image subfields 31.mn, each imaged with an associated beamlet of a multi-beam charged particle microscope. In a first method step S1, for each image subfield 31.mn, a plurality of vector distortion maps 730 are provided. Each vector distortion map 730 characterizes the position-dependent distortion per pixel of the associated image subfield 31.mn. Furthermore, as already explained in the overview of the present application, the term "map" should be interpreted broadly. It is intended to indicate that for each image subfield 31.mn, a vector field is provided, which includes distortion vectors. For example, each of the plurality of vector distortion maps 730 can be represented by a polynomial expansion in vector polynomials. From the polynomial expansion, then, the specific distortions for positions p, q in the image subfield 31.mn can be calculated. Alternatively, each of the plurality of vector distortion maps 730 can be represented by a two-dimensional look-up table. In principle, other representations are also possible.

[0115] In method step S2, a feature of interest 701 in the image is identified. In method step S3, geometric properties of the feature 701 are extracted. Method steps S2 and S3 can be performed separately or can be combined with each other. In principle, the geometric properties of the feature of interest 701 can be of any kind and shape. The geometric properties of the feature 701 can be, for example, the contour of the feature 701. Alternatively, only a part of said contour (for example an edge or a corner). Also the center of the feature of interest 701 is possible. Examples of geometric properties of the feature 701 can be at least one of the following: contour, edge, corner, point, line, circle, ellipse, center, diameter, radius, distance. Irregular shapes as well as other geometric properties are possible. The geometric properties can also include properties such as line edge roughness, angle between two lines, or area or volume.

[0116] In a next step S4, a corresponding image sub-field 31.mn is determined which contains the extracted geometrical property of the feature 701. In a step S5, one or more positions of the extracted geometrical property of the feature 701 in the determined corresponding image sub-field 31.mn are determined. Whether one or more positions are determined depends on the nature of the extracted geometrical property. By determining the corresponding image sub-fields 31.mn and by determining one or more positions of pixels in each image sub-field 31.mn, it is possible to unambiguously assign the distortion vector 715 (or a number of distortion vectors 715) to the correction performed in method step S6. According to method step S6, distortion-corrected image data is generated by correcting one or more positions of the extracted geometrical property in the image based on the vector distortion map 730 of the corresponding image sub-field 31.mn. Method steps S2 to S6 can be performed repeatedly for a number of features 701.

[0117] Thereafter, in method S7, the procedure can end or one or more metrology applications or measurements can be performed. Examples include determining the dimensions of features of the semiconductor device in the distortion-corrected image, determining the area of ​​features of the semiconductor device in the distortion-corrected image, determining the positions of a plurality of regular objects in the semiconductor device, particularly the HAR structure, in the distortion-corrected image, determining the line edge roughness in the distortion-corrected image, and / or determining the overlay error between different features in the semiconductor device in the distortion-corrected image. These exemplary applications are described in more detail below.

[0118] The extracted geometrical characteristic 701 of a feature may be divided into a number of parts, respectively, since it extends over a number of image sub-fields 31.mn. In such a case, one or more positions of each part of the extracted geometrical characteristic are individually de-distorted based on the associated individual vector distortion map 730 of the corresponding image sub-field 31.mn of each part. This significantly improves the accuracy of the measurement process, since the scanning-induced distortion is not necessarily a smooth function on the sub-field boundary 725.

[0119] FIG. 11 illustrates the determination of a vector distortion map 730 based on a target grid 711. FIG. 11A shows a test sample with a precisely known pattern of structures 712 that define a target grid, in this example a repeating pattern. In this case, the target grid 711 includes a number of circles. However, other target grids 711 (e.g., target grids including squares or a combination of squares and circles) can also be selected. The target grid is ideally a perfect grid with a nominal pitch between a number of structures 712 arranged in a regular pattern. The test sample is then imaged with the multi-beam charged particle microscope 1, the resulting image is analyzed, and an actual grid 720 is determined based on said analysis. The target grid 711 and the actual grid 720 differ from each other. This difference is expressed with respect to the center 713 of the structures 712 and is illustrated by the distortion vectors 715 in FIG. 11B. The field of the distortion vectors 715 is an example of a vector distortion map 730 used for distortion correction.

[0120] 12 shows a diagram of the determination of the distortion vector 715. A vector 717 defined in the internal coordinate system of coordinates p, q points to the center 713 of the structures 712 of the ideal target grating 711. However, when determining the real grating, this center 713 is imaged at a position 714, which can be represented by a vector 716 with respect to the internal coordinates p, q of the image subfield. Subtracting the vector 717 from the vector 716 gives the distortion vector 715. It should be noted that the distortion vector 715 is defined as a vector that points from the ideal grating center 713 to the actually measured grating center 714. However, in principle it is also possible to define the distortion vector 715 as the inverse vector of the vector shown here. Depending on this definition, it is the distortion vector 715 itself or its inverse that is used for the correction of one or more positions of the extracted geometrical features in the image subfield 31.mn.

[0121] FIG. 13 shows the determination of grid points in an actual grid 720. The target grid 710 includes a number of regular and precisely known structures 712. These structures 712 have ideal contours. In the illustrated example, the structures 712 are circles. In the case of imaging of a test sample, a number of positions 721 of a single contour are determined. In this case, due to the geometrical properties of the structures 712, which are point-symmetric, a connecting line 722 can be defined that connects two edge positions on either side of the structure 712. Reference symbol 723 indicates the area of ​​the center point 724 of the line that includes the structure center 713. The structure center 713 is used to define the grid position. The average position of these center points 724 can be used as the actual structure center, i.e. the structure center with respect to the actual grid 720. The standard deviation of the center point positions 724 is a measure of the accuracy or confidence with which the feature center 713 can be determined. If this deviation is too large, this structure can be excluded from further processing.

[0122] FIG. 14 illustrates dimensional measurements based on distortion-corrected shape data. FIG. 14A illustrates two image subfields 31.mn and 31.m(n+1) whose corresponding vector distortion map 730 includes a field of distortion vectors 715. In a conventional single-beam charged particle microscope with a single image field, the distortion is a slowly varying continuous function over the single image field and has negligible effect on the measurement of the dimension. However, in a multi-beam charged particle microscope with multiple image subfields 31.mn, e.g. subfields 31.mn and 31.m(n+1), the overall distortion is a discontinuous function at the subfield boundary 725. Therefore, the dimensional measurement of a feature 701 extending across two image subfields 31.mn and 31.m(n+1) may be degraded by a large difference in the discontinuous distortion function. According to the invention, the two parts 726 and 727 of the feature 701 are separately dewarped according to the vector distortion map 730 of the image subfields 31.mn and 31.m(n+1), respectively. More precisely, the geometrical characteristic of the feature extracted from the image is the distance dv, or more precisely, the two positions (p1;q) and (p2;q). The value of q is identical and will not be described further. However, the coordinate (p1;q) is determined with respect to the image subfield 31.mn, whereas the coordinate (p2;q) is determined with respect to the image subfield 31.m(n+1). The position of (p1;q) is corrected based on the vector distortion map 730 of the image subfield 31.mn, and the position of (p2;q) is corrected based on the vector distortion map 730 of the image subfield 31.m(n+1). The respective distortion vectors vp1 and vp2 are also shown in FIG. 14B. As a result, the distance dv is dewarped to the distance d.

[0123] Fig. 14 shows the situation when the static distortion of multiple primary beamlets is compensated. Thus, the vector distortion map 730 at the center position of each image subfield 31.mn, 31.m(n+1) does not show any distortion or distortion vector offset. However, the vector distortion map 730 due to the scanning-induced distortion of each image subfield 31.mn, 31.m(n+1) may each include an additional distortion vector offset due to static distortion of the multi-beam charged particle system 1. For example, as shown in Fig. 3, each distortion vector offset of each image subfield may be different.

[0124] FIG. 15 illustrates a statistical evaluation of the position of regular objects based on distortion-corrected image data. FIG. 16A illustrates a number of HAR features, with reference symbols 80.1 and 80.2 respectively labeling a first HAR structure and a second HAR structure. These HAR features 80.1, 80.2 can be identified, for example, by pattern recognition, the principles of which are well known in the art. The pattern recognition can be assisted, for example, by machine learning. The geometrical properties of the HAR features 80.1 and 80.2 are the center positions of the HAR features 80.1 and 80.2, respectively. The center positions of each HAR structure 80 are extracted and their positions are determined. Furthermore, the image subfield 31.mn to which the center positions of the HAR structures 80 belong is determined. In this case, the center of the HAR structure 80.1 belongs to the image subfield 31.mn, and the center of the HAR structure 80.2 belongs to the image subfield 31.m(n+1). The positions of the centers of the HAR structures 80.1 and 80.2 are then corrected based on the corresponding vector distortion map 703 of each of the corresponding image subfields 31.mn and 31.m(n+1). The corrected center positions are then analyzed (e.g. compared with the design center positions 96 of the HAR structures) and the deviations 97 from the design center positions 96 are analyzed. Also, in the example shown in Fig. 15, it is important to first perform feature extraction and localization on the binary image that remains distorted. Then, a distortion correction is performed position-dependently for the associated image subfields 31.mn, 31.m(n+1).

[0125] Besides the specific applications shown in figures 14 and 15, several other applications of the invention are possible, one of which is the LER determination (determination of line edge roughness) across a subfield boundary 725. Due to distortion discontinuities across the subfield boundary 725, the line itself may also become discontinuous. A possible solution according to the invention is basically to first extract the line, split it into parts belonging to different image subfields, apply distortion correction to each part of the line and then determine the line edge roughness.

[0126] The deviation in the position of the second feature 701' of the second layer relative to the first feature 701 of the first layer is referred to as the overlay error. The overlay error may be determined for features 701, 701' produced in different lithography steps or in different layers. Again, according to the invention, the features 701, 701' are first extracted. Then a distortion correction is applied to the features 701, 701'. The invention is particularly important when the first feature 701 and the second feature 701' are in different image subfields 31.mn.

[0127] A general object of the present invention is to suppress or avoid distortion compensation during image post-processing of 2D image data. As described above, distortion compensation during post-processing of 2D image data requires storage of original image data and computation of distortion-corrected target image data. According to the above-described improved distortion correction method, distortion correction is performed on a suppressed set of extracted parameters, such as edge or center positions, instead of the entire 2D image data. This reduces the amount of computation and power consumption by at least one order of magnitude or even up to five orders of magnitude. According to another embodiment of the present invention, the amount of computation and power consumption required for post-processing is further reduced. In this embodiment, the digital image data stream received from the image sensor 207 is written directly to the memory 814 so that distortion aberrations are suppressed or compensated during processing of the data stream. This allows at least a large part of the distortion of each sub-field of view 31.mn to be compensated during stream processing.

[0128] Fig. 16 shows an image data acquisition unit 810 and related units or modules. For ease of explanation, only one image channel is shown in Fig. 16, and the other image channels are not shown. In this case, the number of image channels corresponds to the number of J beamlets applied for imaging by the multi-beam charged particle microscope 1.

[0129] In one example, the image sensor 207 comprises a plurality of J photodiodes corresponding to a plurality of J secondary electron beamlets. Each of the J photodiodes (e.g., avalanche photodiodes (APDs)) is connected to a respective analog-to-digital converter. The image sensor may further comprise an electron-to-photon converter, for example as described in DE 102018007455 B4, the entire contents of which are incorporated herein by reference.

[0130] The analog-to-digital converter 811 converts the analog data stream into a plurality of J digital data streams. After conversion to a digital data stream, the data is provided to an averaging unit 815. However, the averaging unit 815 may also be omitted. In principle, pixel averaging or line averaging may be performed, but for further details reference is made to International Patent Application Publication No. WO2021 / 156198, the entire contents of which are incorporated herein by reference.

[0131] The image data acquisition unit comprises, for each of the J image subfields, a hardware filter unit 813 arranged to receive the digital data stream and, when using the multi-beam charged particle microscope 1, to generate a distortion-corrected data stream by performing a convolution of a segment of the image subfields 32.mn with a spatially varying filter kernel 910. Details of this distortion correction are explained in more detail below.

[0132] The image data acquisition unit 810 further comprises an image memory 814 configured to store the distortion corrected data stream as a 2D representation of the image sub-field of view 31.mn.

[0133] In the illustrated example, the image data acquisition unit 810 is part of an imaging control module 820, which also includes a scan control unit 930. In this example, the scan control unit 930 is configured to control the first collective raster scanner 110 as well as the second collective raster scanner 220. Other control mechanisms of the scan control unit 930 (not shown in FIG. 16 ) may also be implemented in the multi-beam charged particle microscope 1.

[0134] In principle, the overall control of the multi-beam charged particle microscope 1 includes various units or modules. However, it should be noted that the illustrated representation of the various modules belonging to this control can also be differently selected and realized. The structure shown in FIG. 16 is therefore only an example. Besides the imaging control module 820, a control unit 800 is provided. An image memory 814 is connected to the control unit 800 for parallel readout and is configured to read out a number of J digital images corresponding to J image subfields 31.11 to 31.mn. An image stitching unit 817 of the control unit 800 is configured to stitch the J digital image subfields into a digital image file corresponding to an image patch (e.g. image patch 17.k). The image stitching unit 817 is configured to extract information from the digital image file and is connected to an image data processor output 818 configured to write the digital image file into a memory or provide information from the digital image file to a display.

[0135] It should be noted that the modules and processes shown in Fig. 16 are strictly synchronized, which can be achieved by providing suitable clock signals (not shown in detail in Fig. 16). Furthermore, since the hardware filter unit 813 is arranged to perform a convolution of a segment of the image subfield with a spatially-varying filter kernel 910, a counting unit 816 is realised in the control unit 800 providing an input to a kernel generation unit 812 which provides data for the filter kernel to the hardware filter unit 813. It should be emphasised again that the filter kernel 910 is calculated for each imaging channel, which however is not shown in detail in Fig. 16 for ease of illustration.

[0136] The imaging control module 820 of the multi-beam charged particle microscope 1 may comprise a plurality of L image data acquisition units 810.n including at least a first image data acquisition unit 810.1 and a second image data acquisition unit 810.2 arranged in parallel. Each image data acquisition unit 810.n may be configured to receive sensor data of the image sensor 207 corresponding to a subset of S beamlets of the plurality of J primary charged particle beamlets and generate a subset of S streams of digital image data values ​​of the plurality of J streams of digital image data values. The number of S beamlets attributed to each of the L image data acquisition units 810.n may be the same, S*L=J. The number of S may be, for example, 6 to 10 (for example, S=8). The number L of parallel image data acquisition units 810.n may be, for example, 10 to 100 or more depending on the number J of primary charged particle beamlets. Due to the modular concept of the imaging control module 820, the number J of charged particle beams in the multi-beam charged particle microscope 1 can be increased by adding parallel image data acquisition units 810.n.

[0137] FIG. 17 illustrates the hardware filter unit 813. The arrows in FIG. 17 indicate the data input to the hardware filter unit 813. In the illustrated embodiment, the hardware filter unit 813 comprises a grid array 900 including 5×5 filter elements 901. The grid array 900 of filter elements 901 is intended to reflect or correspond to a representation of a segment of the image subfield 31.mn. The order and arrangement of the data within the grid array 900 is therefore important to ensure this relationship or equivalence. In an exemplary embodiment, the hardware filter unit 813 is realized by a series of FIFOs 906. The series of FIFOs 906 ensures that the order of the data entering the hardware filter unit 813 is maintained. Furthermore, the FIFOs 906 ensure the correct jump from a first row or line of the image subfield 31.mn to a second row or line of the image subfield, etc. Thus, when the filter elements 901 are filled stepwise with pixel values ​​and a sequence of pixel values ​​is passed through the filter unit 813, the pixel value entries in the grid array 900 may correspond to segments of the image subfield 31.mn that are to be distortion corrected.

[0138] As mentioned above, the hardware filter unit 813 is configured to perform a convolution with a spatially-varying filter kernel 910 of a segment 32 of the image sub-field 31.mn. In other words, the values ​​of the coefficients of the filter kernel 910 must be calculated individually for the filtering process of the particular segment 32 to be filtered. Each filter element 901 in the illustrated grid array 900 contains two types of entries: the pixel value itself and the coefficients generated by the kernel generation unit. To perform the convolution, it is necessary to perform a multiplication of the entries in the filter element 901. Then, a summation of the results of this multiplication is required, which is indicated by the line in FIG. 17 connecting the filter element 901 with the box 905. The performed filtering operations (multiplication and summation) result in a time delay, which remains constant throughout the filtering process of the entire image sub-field 31.mn. The distorted data stream (data input) is converted into a distortion-corrected data stream (data output).

[0139] Fig. 18 shows the convolution of a segment 32 of an image subfield 31.mn with a filter kernel 910. Both the segment 32 of the image subfield 32.mn and the filter kernel 910 are shown as a grid arrangement of filter elements 901, where the size of the filter kernel 910 is the same. A 5x5 realization is shown here. On the left side of Fig. 18A, the uncorrected pixel value or intensity I is shown in a first register 902. In the filter kernel 910, a number of coefficients 903 generated by the kernel generation unit 812 are stored in a second register 903.

[0140] FIG. 18B is mathematically equivalent to the situation shown in FIG. 18A, showing two matrices that need to be convolved. The result is the sum of two of the specific products of the matrix entries with each other. It should be noted that it is usually necessary to multiply different entries of a matrix with each other. For example, it is usually necessary to multiply the entry I 11 and entry K 11 It is not necessary to multiply the different entries by each other. This is only the case for symmetric filter kernels. However, there is still a fixed scheme in which different entries need to be multiplied. This scheme is already realizable by the respective hardware representation of the filter kernel 910 (the flipping process of both the rows and columns of the kernel).

[0141] FIG. 19 is a diagram illustrating some of the filter elements 901 and related elements. More specifically, according to the illustrated embodiment, each filter element 901 comprises a first register 902 for temporarily storing a pixel value and a second register 903 for temporarily storing a coefficient generated by the kernel generation unit 812. Furthermore, the filter element 901 comprises a multiplication block 904 configured to multiply the pixel value stored in the first register 902 by a corresponding coefficient stored in the second register 903. It is noted that the multiplication block 904 is not necessarily part of the filter element 901 itself, but can be realized separately. After the multiplication is performed in the multiplication block 904, the respective results are presented to a summation block 905. It is noted that although FIG. 19 only shows two filter elements 901 and one summation block 905, more filter elements 901 and multiple summation blocks 905 are usually provided to successfully realize the distortion correction. The arrows in FIG. 19 indicate data flow. Furthermore, the entries of the second register 903 are provided by the kernel generation unit 812 (not shown in FIG. 19).

[0142] According to a more general embodiment, the hardware filter unit 813 may comprise a grid arrangement 900 of filter elements 901 each comprising a first register 902 for temporarily storing a pixel value and a second register 903 for temporarily storing a coefficient generated by the kernel generation unit 812, the pixel value temporarily stored in the first register 902 representing a segment of the image subfield 31.mn. The hardware filter unit 813 may further comprise a number of multiplication blocks 904 configured to multiply the pixel value stored in the first register 902 by a corresponding coefficient stored in the second register 903. The hardware filter unit 813 may further comprise a number of summation blocks 905 configured to sum the results of the multiplication. According to this more general configuration, the number of multiplication blocks is not necessarily the same as the number of filter elements 901 and can be reduced.

[0143] The latter situation is illustrated in FIG. 20, which shows a hardware filter unit 813 with a filter kernel window of 3×3. Thus, the filter kernel window (3×3) is smaller than the lattice arrangement 900 (5×5). It is important to note that the filtering process according to the present invention is performed for a specific purpose, namely distortion correction. The distortion correction can be interpreted as a shift of pixels. This means that even if a full convolution of the full size kernel filter 901 with the pixel value stored in the first register 902 of the filter element 901 is performed, there are many multiplications that do not affect the result of the distortion correction, or more precisely, the sum generated. For this reason, even if all filter elements 901 are taken into account in the convolution, the result (sum) does not differ. Instead, it is important that the relevant filter elements 901 are selected for the calculation process. This can be done by selecting an appropriate kernel window 907. Naturally, the exact placement position of the filter window 907 in the lattice 900 is not arbitrary. The position of the kernel window 907 can be determined by the kernel generation unit 812, in particular "on the fly". If this embodiment variant is selected, there is no need to provide a multiplication block for each filter element 901. The number of logic units in the hardware filter unit 813 can therefore be reduced. However, it must be ensured that the position of the kernel window 907 is not fixed for each segment 32 of the image subfield 31.mn, so that different multiplications can be performed. For this reason, it must be provided that a number of switching means are arranged to logically combine the entries and filter elements 901 in the multiplication block 904, in use, depending on the position of the kernel window 907.

[0144] According to an embodiment, the kernel generation unit 812 is configured to determine the spatially varying filter kernel 910 based on a vector distortion map 730 characterizing the spatially varying distortion in the image subfield 31.mn. According to an embodiment, the vector distortion map 730 is represented by a polynomial expansion in a vector polynomial. Alternatively, the vector distortion map 730 is represented by a multidimensional lookup table. Furthermore, the kernel generation unit 812 may be configured to determine the filter kernel 910 based on a function f that represents a pixel representatively. A possible function f for representing a pixel can be, for example, a Rect2D function representing a rectangular pixel. Alternatively, the function f can be the shape of the pixel's beam focus (for example, a Gaussian function, an anisotropic function, a cubic function, a sinusoidal function, an Airy pattern, etc.), and the filter is truncated at a certain low level value. Furthermore, the filter is assumed to be energy-conserving, and therefore the high-order rounding filter kernel 910 is assumed to be normalized so that the weights sum to one.

[0145] 7 of this patent application, pixel 700 is "distributed" over four pixels 700' in the distortion-corrected image. Therefore, a kernel window 907 of size 2x2 may be applied.

[0146] Figure 21 illustrates a hardware filter unit 813 with a 2x2 filter kernel window 907. The example shown in Figure 21 corresponds to the shift shown in Figure 7 of this patent application.

[0147] According to an embodiment of the invention, distortion compensation during image post-processing of 2D image data is minimized or avoided. Thus, there is no need to perform distortion compensation for every pixel of a huge 2D image, which contains billions of pixels and requires a large amount of image memory. Instead, distortion compensation is performed for a suppressed set of extracted parameters, such as edge or center positions, rather than for the entire 2D image data. According to another example, distortions of each subfield of view 31.mn are compensated during stream processing of the data stream from the image sensor 207. Stream processing of analog data from the image sensor 207 is necessary anyway, and additional distortion compensation during stream processing requires only a small additional computing power and reduces the amount of additional memory. This allows the present invention to reduce the amount of computation and power consumption by at least one order of magnitude or even up to five orders of magnitude. The two methods and configurations can also be combined. In one example, it is advantageous to compensate a first part of the vector distortion polynomial for each image subfield of view 31.mn by stream processing and to compensate a second part of the vector distortion polynomial by distortion compensation with a suppressed set of extracted parameters or geometrical properties. For example, the linear part of the distortion polynomial is compensated during stream processing, and higher order distortions are compensated for by distortion correction with a suppressed set of extracted parameters. This reduces the additional computational effort of computing higher order vector polynomials during stream processing. In general, the present invention enables distortion correction of a multi-beam charged particle inspection system 1 with reduced computational effort and energy consumption. This allows inspection or metrology tasks during semiconductor fabrication processes to be performed with high efficiency and with low computational effort and energy consumption.

[0148] It should be noted that the embodiments of the present invention described with reference to the drawings are not intended to limit the present invention in any way, and the drawings merely show possible implementations of the present invention.

[0149] Other examples of the present invention are described below, which can be combined with the other embodiments and examples described above.

[0150] Example 1. A method for determining a distortion-corrected position of a feature in an image composed of one or more image patches, each composed of a plurality of image subfields, each imaged with an associated beamlet of a multi-beam charged particle microscope, comprising: a) providing, for each image subfield, a number of vector distortion maps each characterizing a position-dependent distortion for each pixel of the associated image subfield; b) identifying a feature of interest in the image; c) extracting geometric properties of the features; d) determining a corresponding image subfield that includes the extracted geometric characteristics of the feature; e) determining the location of one or more of the extracted geometrical properties of the feature in the determined corresponding image sub-field; f) generating distortion-corrected image data by correcting one or more positions of the extracted geometric features in the image based on a vector distortion map of the corresponding image subfield; A method comprising:

[0151] Example 2. The method of example 1, wherein method steps b) to f) are performed repeatedly for multiple features.

[0152] Example 3. The method of example 1 or 2, wherein other areas in the image that do not contain features of interest are not dewarped.

[0153] Example 4. The method of any one of Examples 1-3, wherein the geometric characteristic of the feature is at least one of a contour, an edge, a corner, a point, a line, a circle, an ellipse, a center, a diameter, a radius, and a distance.

[0154] Example 5. The method of any one of Examples 1 to 4, wherein the step of extracting geometric characteristics includes generating a binary image.

[0155] Example 6. The extracted geometrical characteristics of a feature are divided into a number of parts, each of which extends over a number of image subfields, The method according to any one of Examples 1 to 5, wherein one or more positions of each portion of the extracted geometric characteristics are individually corrected based on an associated individual vector distortion map of the corresponding image sub-field of each portion.

[0156] Example 7. The method of any one of Examples 1 to 6, wherein the step of extracting geometric properties of the feature of interest is performed on the entire image.

[0157] Example 8. A method according to any one of Examples 1 to 7, wherein the step of correcting one or more positions of the extracted geometric characteristic in the image based on a vector distortion map of a corresponding image subfield includes determining a distortion vector for at least one position of the extracted geometric characteristic.

[0158] Example 9. The method according to any one of Examples 1 to 8, wherein the step of correcting one or more positions of the extracted geometric features in the image based on a vector distortion map of a corresponding image subfield comprises transforming a pixel of the image into at least one pixel of a distortion-corrected image based on the distortion vector.

[0159] Example 10. The method of any one of Examples 1-9, wherein each of the plurality of vector distortion maps is represented by a polynomial expansion in a vector polynomial.

[0160] Example 11. The method of any one of Examples 1-9, wherein each of the plurality of vector distortion maps is represented by a two-dimensional lookup table.

[0161] Example 12. Determining dimensions of features of a semiconductor device in distortion-corrected image data; determining areas of semiconductor device features in the distortion corrected image data; determining locations of a plurality of regular objects, in particular HAR structures, in the semiconductor device in the distortion corrected image data; determining line edge roughness in the distortion corrected image data; and / or determining overlay errors between different features in the semiconductor device in the distortion corrected image data; The method of any one of Examples 1 to 11, further comprising at least one of:

[0162] Example 13. Providing a test sample having a precisely known pattern, particularly a repeating pattern, defining a target grid; imaging the test sample with a multi-beam charged particle microscope, analyzing the resulting image, and determining an actual lattice based on said analysis; determining a position deviation between the actual grating and the target grating; obtaining a vector distortion map for each image subfield based on the position deviation; The method according to any one of Examples 1 to 12, further comprising:

[0163] Example 14. The method of example 13, further comprising shifting the test sample from a first position to a second position relative to the multi-beam charged particle microscope and imaging the test sample at the first position and the second position.

[0164] Example 15. The method according to example 13 or 14, wherein the step of determining the positional deviation comprises a two-stage determination, in a first stage, the shift of each image subfield, the rotation of each image subfield and the magnification of each image subfield are compensated for, and in a second stage, other higher order distortions are determined.

[0165] Example 16. The method of any one of Examples 1 to 15, further comprising updating the vector distortion map.

[0166] Example 17. The method according to any one of Examples 1 to 16, further comprising correcting distortions in the image by stream processing of data during image pre-processing.

[0167] Example 18. A method for correcting distortions in an image composed of one or more image patches, each composed of a plurality of image subfields imaged by associated beamlets of a multi-beam charged particle microscope, comprising: g) providing, for each image subfield, a number of vector distortion maps each characterizing a position-dependent distortion for each pixel of the associated image subfield; h) for each pixel in the image, determining a corresponding image subfield that contains the pixel; i) for each pixel in the image, transforming the pixel in the image to at least one pixel in a distortion-corrected image based on a vector distortion map of a corresponding image subfield; A method comprising:

[0168] Example 19. A computer program product comprising program code for carrying out the method according to any one of Examples 1 to 18.

[0169] Example 20. A multi-beam charged particle microscope comprising a controller configured to execute the method described in any one of Examples 1-18. [Explanation of symbols]

[0170] 1. Multi-beamlet charged particle microscope system 3 Primary charged particle beamlets (comprising multiple primary charged particle beamlets) 5 Primary charged particle beam spot 7 Object 9 Secondary electron beamlet (Consists of multiple secondary electron beamlets) 11 Secondary electron beam path 13 Primary beam path 15 Secondary charged particle image spot 17 Elephant Patch 19 Image patch overlap area 21 Image patch center position 25 Wafer surface 27 Scan path of primary beamlet 29 Center of image subfield 31 Image subfield 32 Image subfield segments 33 First inspection site 35 Second Test Site 39 Superimposed area of ​​secondary field of view 31 80.1 HAR Structure 80.2 HAR Structure 96 Design center position of HAR structure 97 Deviation from the design center position of the HAR structure 100 Object Irradiation Unit 101 Object or image plane 102 Objective Lens 103 Field of view lens group 105 Optical axis of multi-beam charged particle microscope system 108 First Beam Crossover 110 First multi-beam raster scanner 112 Correction element for multi-beam raster scanner 120 Scanning correction control module 141 Example of primary beam spot position 143 Static displacement vector of the primary beam spot 150 central beamlets 151 Real beamlet trajectories 153 Deflection electrode 155 Equipotential lines of electrostatic potential 157 Off-Axis or Field of View Beamlets 159 Virtual common pivot point 161 Virtual pivot point 163 1st beam path 171 Scanner 110 Front-stage System 189 Intersection Volume of Transverse Beams 200 detection units 205 Projection system 206 Electrostatic Lens 207 Image Sensor 208 Imaging Lens 209 Imaging Lens 210 Imaging Lens 212 The Second Crossover 214 Aperture Filter 216 Active Elements 218 Third deflection system 220 Multi-aperture corrector 222 Second deflection system 300 Charged Particle Multi-Beamlet Generator 301 Charged Particle Source 303 Collimator Lens 305 Primary Multi-Beamlet Construction Unit 306 Active Multi-Aperture Plate 307 First Field Lens 308 Second Field Lens 309 Electron Beam 311 Primary electron beamlet spot 321 Intermediate image plane 390 Beam Steering Multi-Aperture Plate 400 Beam splitter unit 420 Magnetic Elements 500 Sample Stage 503 Sample Voltage Source 700 pixels 701 Features 702 Grayscale Image 710 Binary Image 711 Target Grid 712 Structure 713 Center of structure 714 Actual Grid Points 715 Distortion Vector 716 Vector 717 Vector 720 Actual lattice 721 Single Contour Position 722 A connecting line connecting two edge positions on either side of the structure 723 Area of ​​the center point of the line including the center of the structure 724 Line Center Point 725 Subfield boundary 726 First Part of Feature Second part of 727 feature 730 Vector Distortion Maps 800 Control Unit 810 Image Data Acquisition Unit 811 Analog-to-Digital Converter 812 Kernel Generation Unit 813 Hardware Filter Unit 814 Image Memory 815 Averaging Unit 816 Counting Unit 817 Image Stitching Unit 818 Image processing and output 820 Projection system control module, imaging control module 830 Primary Beam Path Control Module 900 Grid Array 901 Filter Element 902 A first register for storing pixel values 903 Second register for storing coefficients 904 Multiplication Block 905 Sum Block 906 Shift Register 907 Kernel Window 910 Filter Kernel 930 Scanning Control Unit Prepare multiple vector distortion maps for each S1 image subfield Identifying features of interest in S2 images S3 Extract geometric characteristics of features S4 Determine the corresponding image subfield containing the extracted geometric characteristics of the feature. S5. Determine the location or locations of the extracted geometrical characteristics of the feature in the corresponding image sub-fields. S6. Generate distortion-corrected image data by correcting one or more locations of the extracted geometric features in the image based on the vector distortion map of the corresponding image subfield. S7 Exit or other method steps dv Distance in distorted image d Distance in the undistorted image vp1 First part of the distortion vector vp2 Second part of the distortion vector p Internal coordinates of the image subfield q Internal coordinates of the image subfield x global coordinate y global coordinate

Claims

1. at least a first collective raster scanner (110) for collective scanning a plurality of J primary charged particle beamlets (3) over a plurality of J image subfields (31.mn); a detection unit (200) comprising a detector for detecting a plurality of J secondary electron beamlets (9), each of which corresponds to one of the J image subfields (31.mn); a scan control unit (930) connected to the first collective raster scanner (110) and configured, in use, to control a raster scanning movement of the plurality of J primary charged particle beamlets (3) by the first collective raster scanner (110); a kernel generation unit (812) configured, in use, to generate a spatially varying filter kernel (910) for the correction of spatially varying distortions of said image subfields (31.mn); an image data acquisition unit (810) synchronized in operation with the detector, the scan control unit (930), and the kernel generation unit, for each of the J image subfields: an analog-to-digital converter (811) for converting, in use, an analog data stream received from said detector into a digital data stream representative of said image sub-fields (31.mn); a hardware filter unit (813) configured to receive the digital data stream and configured, in use, to generate a distortion-corrected data stream by performing a convolution of a segment (32) of the image subfield (31.mn) with the spatially varying filter kernel (910); an image memory (814) configured to store the distortion-corrected data stream as a 2D representation of the image subfields (31.mn); an image data acquisition unit (810) comprising: A controller (800, 820) comprising: A multi-beam charged particle microscope (1) comprising:

2. The hardware filter unit (813) a grid array (900) of filter elements (901) each comprising a first register (902) for temporarily storing pixel values ​​and a second register (903) for temporarily storing coefficients generated by the kernel generation unit (812), the pixel values ​​stored in the first register (902) representing a segment of the image subfield (31.mn); a plurality of multiplication blocks (904) configured to multiply pixel values ​​stored in the first register (902) by the corresponding coefficients stored in the second register (903); a plurality of summation blocks (905) configured to sum the results of the multiplications; 2. A multi-beam charged particle microscope (1) according to claim 1, comprising:

3. 3. The multi-beam charged particle microscope (1) of claim 1 or 2, wherein the hardware filter unit (813) comprises a plurality of shift registers (906) configured to realize a grid arrangement (900) of the filter elements (901) and to maintain the order of data in the data stream as it passes through the hardware filter unit (813).

4. A multi-beam charged particle microscope (1) according to any one of claims 1 to 3, wherein the image data acquisition unit (810) further comprises a counter (816) configured, in use, to indicate the local coordinates (p, q) of a pixel within an image sub-field (31.mn) being filtered.

5. The multi-beam charged particle microscope (1) according to any one of claims 2 to 4, wherein the size of the lattice arrangement (900) of the filter elements (901) is configured to correct a distortion of at least 10 times the pixel size of the image subfield (31.mn).

6. A multi-beam charged particle microscope (1) according to any one of claims 2 to 5, wherein the size of the grid array (900) of filter elements (901) is at least 21x21 filter elements (901).

7. The multi-beam charged particle microscope (1) according to any one of claims 2 to 6, wherein the size of a predefined kernel window (907) is equal to or smaller than the size of the grid arrangement (900) of filter elements (901).

8. 8. The multi-beam charged particle microscope (1) of claim 7, wherein the kernel generation unit (812) is configured, in use, to determine a position of the kernel window (907) relative to a grid arrangement (900) of the filter elements (901).

9. 9. The multi-beam charged particle microscope (1) of claim 8, wherein the hardware filter unit (813) further comprises a plurality of switching means configured, in use, to logically combine entries of filter elements (901) in a multiplication block (904) based on the position of the kernel window (907).

10. The multi-beam charged particle microscope (1) according to any one of claims 1 to 9, wherein the kernel generation unit (812) is configured to determine the spatially varying filter kernel based on a vector distortion map (730) characterizing the spatially varying distortion in an image subfield (31.mn).

11. The multi-beam charged particle microscope (1) of claim 10, wherein the vector distortion map (730) is represented by a polynomial expansion in vector polynomials.

12. The multi-beam charged particle microscope (1) according to any one of the preceding claims, wherein the vector distortion map (730) is represented by a multi-dimensional look-up table.

13. The multi-beam charged particle microscope (1) according to any one of claims 1 to 12, wherein the kernel generation unit (812) is configured to determine the filter kernel (910) based on a function f that represents a pixel representatively.

14. 14. The multi-beam charged particle microscope (1) according to claim 13, wherein the function f is the same for different scanning directions or different for different scanning directions.

15. The multi-beam charged particle microscope (1) according to any one of claims 1 to 14, wherein the image data acquisition unit (810) further comprises an averaging unit (815) implemented in the direction of the data stream after the analog-to-digital converter (811) and before the hardware filter unit (813).

16. The multi-beam charged particle microscope (1) according to any one of claims 1 to 15, wherein the image data acquisition unit (810) further comprises a further hardware filter unit configured, in use, to perform further filtering operations, in particular low-pass filtering, morphological operations and / or deconvolution with a point spread function.

17. The multi-beam charged particle microscope (1) according to any one of the preceding claims, wherein the hardware filter unit (813) comprises a Field Programmable Gate Array (FPGA) or an Application Specific Integrated Circuit (ASIC).

18. The multi-beam charged particle microscope (1) according to any one of the preceding claims, wherein the hardware filter unit (813) comprises a series of FIFOs (906).

19. 20. The multi-beam charged particle microscope (1) of claim 18, wherein the FIFO (906) is realized as a BlockRAM, a LUT, or an externally connected SRAM or DRAM.

20. A multi-beam charged particle microscope (1) according to any one of claims 1 to 19, an image post-processing unit configured to perform distortion correction of the image data; A system comprising:

Images