KR20240112881A - Method and apparatus for measuring semiconductor features with increased throughput - Google Patents

Abstract

Systems and methods are provided for measurement of parameter values of semiconductor objects in a wafer with increased throughput. The measurement method uses a modified machine learning algorithm to extract measurement results from instances of semiconductor objects. A training method is provided to train a modified machine learning algorithm where user interaction is reduced to a minimum. The method is more flexible and robust and requires less user interaction than conventional methods. The system and method can be used for quantitative metrology of integrated circuits within a semiconductor wafer.

Description

Method and apparatus for measuring semiconductor features with increased throughput

Related applications

This application is filed in connection with U.S. Provisional Application No. 63/291,569, filed December 20, 2021, each of which is hereby incorporated by reference in its entirety, pursuant to 35 U.S.C. 35 U.S.C. in U.S. Application Serial No. 17/701,054, filed March 22, 2022, which claims priority under § 119(e). Claiming benefit under § 120.

Field

The present invention relates to a method for measuring three-dimensional circuit patterns of semiconductor objects within a semiconductor wafer, and more specifically to methods, computer program products and corresponding semiconductor inspection devices for measuring parameters of semiconductor objects such as HAR structures with increased throughput. It's about. Parameter values of a repetitive semiconductor object are measured by a metrology method using machine learning by a semiconductor inspection device configured to perform the method. The methods, computer program products, and semiconductor inspection devices can be used for quantitative metrology, defect detection, process monitoring, or defect review of integrated circuits within a semiconductor wafer.

Semiconductor structures are one of the most precise artificial structures. Semiconductor manufacturing involves precise manipulation, for example lithography or etching, of materials such as silicon or oxides at very fine scales in the nm range. Wafers made from thin slices of silicon serve as a substrate for microelectronic devices containing semiconductor structures formed in and on the wafer. Semiconductor structures are constructed layer by layer using repetitive processing steps involving repetitive chemical, mechanical, thermal and optical processes. The dimensions, shape, and arrangement of semiconductor structures and patterns are affected by many factors. For example, during the fabrication of 3D-memory devices, important processes are currently etching and deposition. Other related process steps, such as lithographic exposure or implantation, may also affect the properties of the integrated circuit elements. Therefore, the manufactured semiconductor structures suffer from rare and different imperfections. Devices for quantitative metrology, defect-detection or defect review are looking for these imperfections. These devices are not only required during wafer fabrication. Because this manufacturing process is complex and highly non-linear, optimization of production process parameters is difficult. As a solution, an iterative approach called process window qualification (PWQ) can be applied. In each iteration, test wafers are manufactured based on current best process parameters, and different dies on the wafer are exposed to different manufacturing conditions. By detecting and analyzing test structures with devices for quantitative metrology and defect-detection, the best manufacturing process parameters can be selected. In this way, production process parameters can be tweaked towards optimum. Thereafter, highly accurate quality control processes and devices for metrology semiconductor structures on wafers are required.

The manufactured semiconductor structure is based on prior knowledge. Semiconductor structures are manufactured from a series of layers parallel to a substrate. For example, in a logic type sample, metal lines run parallel in the metal layer or HAR (high aspect ratio) structure and metal vias extend perpendicular to the metal layer. The angle between the metal lines of different layers is 0° or 90°. On the other hand, in the case of VNAND type structures, the cross-section is known to be circular on average. Moreover, the semiconductor wafer has a diameter of 300 mm and consists of a plurality of different parts, so-called dies, each containing at least one integrated circuit pattern, for example for a memory chip or for a processor chip. During manufacturing, a semiconductor wafer undergoes approximately 1000 process steps to form approximately 100 or more parallel layers, including transistor layers, mid-line layers, and interconnect layers, and within memory devices, memory cells. A plurality of 3D arrays are formed.

The aspect ratio and number of layers of integrated circuits continue to increase and structures are growing into a third (vertical) dimension. The current height of the memory stack exceeds tens of microns. In contrast, feature sizes are becoming smaller. The minimum feature size or critical dimension is less than 10 nm, for example 7 nm or 5 nm, and will approach feature sizes of less than 3 nm in the near future. Although the complexity and dimensions of semiconductor structures are growing in a third dimension, the lateral dimensions of integrated semiconductor structures are becoming increasingly smaller. Therefore, measuring the shape, dimensions and orientation of features and patterns and their overlays in 3D with high precision becomes a challenge. The lateral measurement resolution of charged particle systems is typically limited by the sampling raster of individual image points or dwell times per pixel on the sample and the charged particle beam diameter. The sampling raster resolution can be set within the imaging system and adapted to the charged particle beam diameter on the sample. Typical raster resolution is 2 nm or less, but the raster resolution limit can be reduced without physical limitations. The charged particle beam diameter has limited dimensions that depend on the charged particle beam operating conditions and the lens. Beam resolution is limited by approximately half the beam diameter. The lateral resolution may be less than 2 nm, for example even less than 1 nm.

A common way to generate 3D tomography data from nm-scale semiconductor samples is the so-called slice and image approach, acquired by, for example, a dual beam device. The slice and image approach is described in WO 2020/244795 A1. According to the method of WO 2020/244795 A1, a 3D volume inspection is obtained on an inspection sample extracted from a semiconductor wafer. In another example, as described in WO 2021/180600 A1, the slice and image method is applied within the surface of a semiconductor wafer under an oblique angle. According to this method, a 3D volume image of the inspection volume is obtained by slicing and imaging a plurality of cross-sectional surfaces within the inspection volume. For precise measurements, multiple (N) cross-sectional surfaces are created within the inspection volume, with the number N exceeding 100 or even more image slices. For example, in a volume with a lateral dimension of 5 μm and a slicing distance of 5 nm, 1000 slices are milled and imaged. For example, in a typical sample of multiple HAR structures with a pitch of 70 nm, approximately 5000 HAR structures are in one field of view, resulting in a total sum of over 5 million cross-sections of HAR structures. Several improvements have been proposed to reduce the enormous computational effort of extracting the required measurement results. WO 2021/180600 A1 illustrates several methods using a reduced number of image slices. In one example, the method applies a priori information.

One important task in semiconductor inspection is to determine a set of specific parameters of a semiconductor object, such as a high aspect ratio (HAR) structure inside the inspection volume. These parameters are, for example, dimensions, areas, shapes or other measurement parameters. Typically, prior art measurement tasks involve several computational steps, such as object detection, feature extraction, and any kind of metrology operation, for example calculation of distance, radius or area from the extracted features. Of these multiple steps, each requires high computational effort.

Generally, semiconductors contain multiple repetitive three-dimensional structures. During a manufacturing process or process development, several selected physical or geometrical parameters of a plurality of representative three-dimensional structures must be measured with high accuracy and high throughput. To monitor manufacturing, an inspection volume is defined containing a plurality of representative three-dimensional structures. This examination volume is then analyzed, for example by a slice and image approach, resulting in a 3D volume image of the examination volume with high resolution.

The multiple repetitive three-dimensional structures within the inspection volume may exceed hundreds or even thousands of individual structures. By this, a huge number of cross-sectional images are generated, for example at least 100 three-dimensional structures are examined by 100 cross-sectional image slices, so the number of measurements to be performed may easily reach 10000 or more.

Machine learning is a field of artificial intelligence. Machine learning algorithms typically build machine learning models based on training data consisting of multiple training samples. After training, the algorithm generalizes the knowledge gained from the training data to new, previously unencountered samples, thereby making it possible to make predictions on new data. There are a number of machine learning algorithms, for example linear regression k-means or neural networks. For example, deep learning is a class of machine learning that uses artificial neural networks with numerous hidden layers between input and output layers. This extensive internal structure allows the network to incrementally extract higher-level features from raw input data. Each level learns to transform its input data into a slightly more abstract and complex representation, thus deriving lower-level and higher-level knowledge from the training data. Hidden layers can have different sizes and operations, such as convolutional layers or pooling layers. To date, it is known that machine learning is frequently applied to defect detection or classification during semiconductor inspection. For example, during defect detection, machine learning algorithms are trained to flag defects and classify defects into discrete defect classes. The required training data typically requires images of pre-identified defects with several classification labels. The steps of object detection or feature extraction typically involve classic pattern recognition algorithms or machine learning algorithms.

US 6,054,710 discloses an example of the application of machine learning techniques to measurement tasks. Here, the cross section of the semiconductor line will be estimated. At this time, the semiconductor line has a topography that reduces the resolution of the top-down image produced by the electron microscope. The effect of edge slope on the top-down image is taken into account by a machine learning network trained by edge slope data acquired with AFM. However, in practical applications and current requirements, milled or polished 2D cross-sections of semiconductor wafers without any edge tilt have to be examined, and electron microscopy images do not suffer from topography effects. US 2020/0258212 A1 proposes a new example of the application of machine learning to the measurement of semiconductor objects of interest, with a specific focus on edge roughness. US 2020/0258212 A1 discloses the general concept of the application of machine learning to measurement tasks, but requires further improvements for practical implementation with less training effort.

Typical machine learning algorithms require intensive training, including intensive operator or user interaction. Users need to annotate large sets of images with annotation tags to successfully train machine learning algorithms. This is rarely feasible due to large-scale annotation efforts. To manage the labeling effort for annotations in large datasets, active learning has been proposed. This active learning system for classification of anomalies is disclosed in US 11,138,507 B2, where a plurality of defects in a sample are associated with a set of predefined classes by a trained classifier. To obtain new training data, low-likelihood samples are selected from each class and presented together with high-likelihood samples of the same class to obtain a binary decision from the user. However, again sophisticated user interaction is required.

One object of the present invention is to provide an efficient method for performing measurement operations on semiconductor objects of interest. The aim of the present invention is to reduce the number of steps in the measurement task and reduce the high computational effort. Another object of the present invention is to improve the method provided in US 2020/0258212 A1. Another object of the present invention is to improve prior art methods for measuring HAR channels. Another objective of the present invention is to reduce the amount of user interaction for implementing measurement tasks. In general, the object of the present invention is to provide a wafer inspection method for measurement of semiconductor structures within an inspection volume with high throughput and high accuracy. The object of the present invention is to provide a generalized wafer inspection method for measurement of semiconductor structures in an inspection volume, which can be quickly adapted to changes in measurement tasks, measurement systems or changes in the semiconductor object of interest. Another object of the present invention is to provide a method for fast, robust and reliable measurement of a set of parameters describing a semiconductor structure within an inspection volume with high precision and reduced measurement artifacts.

The above objects are solved by the present invention. The invention is illustrated by the claims, and details are provided by examples and examples.

The present disclosure provides an efficient method for performing measurement tasks on semiconductor objects of interest. The present disclosure reduces the number of steps in the measurement task and reduces the high computational effort. This disclosure further improves the method provided in US 2020/0258212 A1. Additionally, the present disclosure improves certain known methods for measuring HAR channels. The present disclosure also reduces the amount of user interaction to implement measurement tasks. In general, the present disclosure provides a wafer inspection method for measurement of semiconductor structures within an inspection volume with high throughput and high accuracy. The present disclosure also provides a generalized wafer inspection method for measurement of semiconductor structures within an inspection volume that can be rapidly adapted to changes in measurement tasks, measurement systems, or changes in semiconductor objects of interest. Moreover, the present disclosure provides a method for fast, robust and reliable measurement of at least one parameter value of a parameter set describing a semiconductor structure within an inspection volume with high precision and reduced measurement artifacts.

According to aspects of the present disclosure, continuous machine learning algorithms are applied directly to the measurement task. However, to perform the measurements in one step involving successive machine learning algorithms, large amounts of annotated training image data are required. Annotated training image data is required to cover the desired range of measurements of the selected physical or geometric parameters.

According to a first embodiment, a segmentation and annotation method is provided to generate training image data for measurement tasks with reduced used interaction. A method according to a first embodiment includes a computer-assisted method for generating training data for a measurement algorithm using machine learning. By means of annotated training image data, training of machine learning algorithms for measurement tasks can be achieved or modified and measurement results that are robust against defects or deviations can be obtained. Additionally, the training data may further include identifiers for fault classes depending on the specific fault type. The method for generating training data according to the first embodiment relies on prior information. By using a priori information, effort and especially user interaction during training are reduced to a minimum.

The first part of the prior information is given by the selection of an appropriate parameterized description of the semiconductor object of interest. The parameterized description of the semiconductor object of interest may include geometric shapes such as circles, rings, ellipses, polygons, or expansion coefficients of series expansions. For example, a parameterized description of a HAR channel cross section includes a set of concentric circles. The parameterized description can be displayed as an overlay on a cross-sectional image, acquired by a charged particle beam system. User interaction is then limited to changing the values of a few parameters of the parameterized technique, for example, changing the center position, the radius of the ring or changing the positions of the points of the polygon instead of pixel-wise annotation. Parameterized techniques can be selected according to CAD data. Parameterized techniques can be selected from libraries containing common user templates of parameterized techniques. In another example, the parameterized description may be derived entirely from CAD data of the semiconductor object of interest.

The second part of the prior information is given by the parameters of the charged particle imaging system (such as residence time, pixel size, landing energy, material contrast, topography contrast, charge compensation measurements).

During the segmentation step according to the prior art, pixels or voxels within the 2D or 3D image are assigned to selected structures. Training machine learning segmentation requires providing pixel/voxel annotations for multiple pixels or voxels in multiple images. This standard method for segmentation by user interaction and pixel-by-pixel segmentation is very time consuming - especially with precision requirements such as in the semiconductor industry where structures are sometimes limited to only a few 3 or 4 pixels.

In a first example according to the first embodiment the generation of training image data is achieved by using a parameterized description of the semiconductor object of interest and the training image data is generated to cover the range of expected parameter values to be measured. Examples of elements of parameterized geometric descriptions are rings or shells of high aspect ratio (HAR) structures or HAR channels. The parameter is given by the radius of the circle or ring, for example. The parameter value range is then the expected deviation of the radius from the design value, for example with a radius value of 5 nm to 6 nm. For example, the edges or boundaries of the HAR channel are used as input to a contour engine that derives measurement results as continuous parameter values. The measurement results are used as annotation labels for the identified semiconductor objects of interest, with which the training images are annotated. According to a first example of the first embodiment, the user provides prior information by selecting a parameterized description, for example the number of circles or rings, the radius of the circles, the distance between the centers of the HAR structure, etc. The parameterized technique is then displayed on the training cross-sectional image segments, either by manual placement of elements of the parameterized technique or by automated methods such as image processing. For example, the user selects the center, number of circles, approximate radius of the circles, or other parameters of the selected parameterized technique. The processing algorithm supports the user for semi-automatic placement and adjustment of the elements of the parameterized technique, whereby the center, radius, etc. can be readjusted. Processing algorithms are also supported by automatically determining parameter values, for example by image processing, whereby elements can be readjusted to represent measured parameter values. In an example, the automated determination of the initial annotation value includes the step of applying a physically-influenced forward simulation model to a parametric description of the semiconductor object of interest. By means of a physical simulation model, an expected image of the semiconductor object of interest acquired with a charged particle beam imaging device is generated by simulation. In the optimization phase, unknown parameter values of the parameter description of the semiconductor object of interest are obtained.

At each step, elements of the parametric description, such as circles or rings, can be superimposed on the training cross-sectional image segments, and the user can modify the parameter values by known computer-graphics methods for geometric elements. The final step of segmentation and annotation involves annotating each training cross-section image segment with parameter values of a parameterized description of detected instances of semiconductor objects of interest within each training cross-section image segment. The segmentation results can also be used for mapping of pixels of 2D training cross-sectional image segments and pixelated segmentation is achieved. Pixelated segmentation can also be used, for example, to combine measurement machine learning algorithms with defect detection and classification algorithms.

From a plurality of representative training cross-sectional image segments, typically a limited range of measurement results of parameter values is determined. Training By means of a set of cross-sectional image segments, training data for a limited range of parameter values is thus generated. Accordingly, a machine learning algorithm or regressor according to the second embodiment, trained with the training data, is capable of detecting and determining continuous quantitative measurements of parameters of the semiconductor structure of interest in new images. The machine learning algorithm's regressor generates measurements of multiple HAR channels directly from new 2D images. For metrology use cases, the quality measure of a machine learning algorithm may be defined, for example, by the deviation of the actually measured values from the training parameter values. Therefore, quality measures can be improved by using dense coverage of predetermined parameter value ranges.

According to a second example of the method of the first embodiment, the coverage of a predetermined range of parameter values is automatically improved by using the first segmented and annotated training cross-sectional image segment obtained by user interaction. Additional annotated training cross-sectional image segments with additional parameter values within a predetermined range of parameter values may be generated, for example, by image processing. Examples of suitable image processing include at least one of variation of scale, change of shape, interpolation, morphological operations, or pattern replacement within the training cross-sectional image segments. Additional training cross-sectional image segments can be generated by physical simulation of the process of acquiring the cross-sectional image segments based on the selected parameterized description of the semiconductor object of interest and the imaging parameters of the charged particle beam system.

In a third example of the method according to the first embodiment, user interaction is reduced even further. Methods for generating training data for training machine learning algorithms to perform quantitative measurements can generally be improved by a priori knowledge. During measurement operations of semiconductor structures, two essential steps are generally known in advance.

First, the imaging characteristics of the imaging process by the imaging device are known. Typically, charged particle beam systems are used that utilize electrons or ions, such as helium ions, as primary charged particles. Typically, imaging parameters such as dwell time, pixel size, landing energy, material contrast or charge compensation measurements are known. Imaging contrast depends on the contrast method selected and the known material contrast. Resolution generally depends on the pixel resolution and point spread function of the charged particle beam system. Imaging noise is determined by imaging conditions, such as scanning speed or dwell time during imaging.

Second, the target shape of the semiconductor structure is generally known from CAD information and information about the manufacturing process steps involved during manufacturing. Additionally, the list of relevant materials is limited and the material contrast is known or can be determined with high precision.

According to a third example of the first embodiment, the method for generating training data applies a physical simulation of the imaging process. From the CAD information, a parameterized description of the repetitive semiconductor object of interest is selected and a plurality of semiconductor objects of interest are created according to the parameterized description. In a plurality of semiconductor objects of interest, parameter values vary within a predetermined range of parameter values. Thereby, large quantities of training cross-sectional image segments covering a predetermined range of parameter values can be automatically generated with minimal user interaction. Physical simulation employs simulation of the imaging process for a given charged particle imaging system, for which imaging parameters may be predetermined and stored in the memory of the charged particle imaging system.

A general advantage of the segmentation and annotation method according to the first embodiment is a significant reduction in annotation effort by reducing user interaction to a minimum. Accordingly, a system configured to perform the method according to the first embodiment can quickly adapt to new measurement tasks of new semiconductor objects of interest, for example after reduction of an integrated semiconductor structure, or of semiconductor objects of interest of different sizes or scales. possible. In an example, the set of training data is based on existing annotated training images and covers, for example, a narrow range of measurements or a different range or measurement. Additional annotated training images are generated by image processing of the annotated training images to cover the required measurement range. This is useful, for example, when reduction is applied to a semiconductor structure, or when the magnification of the imaging process is changed.

The segmentation and annotation method according to the first embodiment provides the additional advantage of rapidly adapting the parameterized technique to new semiconductor objects or variations in the shape of the semiconductor object of interest by introducing new additional parameters. For example, if the HAR channel exhibits frequent deviations from a circular shape, additional parameters such as ellipticity may be introduced into the existing training data set. The introduction can be carried out, for example, in an improvement step or by automatic generation of additional training cross-sectional image segments covering the parameter value range of the newly introduced parameter, for example the ellipticity. For example, if the material or material composition of the semiconductor object of interest changes, the training data can be automatically adapted to the new material contrast corresponding to the new material or material composition.

In the example, generation of training image data is not complete after initial training. Typically, not all impacts or defects during manufacturing are predicted in initial training, and machine learning algorithms are frequently modified during monitoring of the manufacturing process. According to aspects of the first embodiment, a method is provided to further modify training image data for a measurement task with reduced interactivity used.

In a first example, modified annotated training image data is provided to cover unexpected effects during manufacturing. During manufacturing, certain error types may additionally be detected, for example defects due to regular and repetitive errors or defects, for example from contamination of the lithographic mask used during manufacturing. Other defects may be random, for example, sporadic contamination during a manufacturing process step. For example, additional modified annotated training image data is provided by image processing, including, for example, additional noise levels, addition of particle artifacts, or image contrast deviations.

In a second example, modified annotated training image data may be provided to cover unexpected effects during the imaging process. The initial annotated training image data initially covers the effects of the image generation process, such as magnification, noise level, material contrast, and convolution kernels such as point spread functions. Imaging effects may be, for example, curtain effects during the slicing and imaging process, resulting in unexpected topography contrast in addition to the material contrast in the SEM image. This contrast from the curtain can be taken into account in the modified annotated training image data. Another effect is the charging effect. Local sample charging can affect the local secondary electron yield and thus change the material contrast and primary particle beam focus position and thus produce a locally distorted image or a locally altered image contrast. These images with locally altered image contrast and local distortion can be considered in the modified annotated training image data. According to the method, initial or modified annotated training images can be automatically generated depending on the expected range of imaging parameter values of the measurement task. With this additional modified annotated training data, inspection or measurement tasks using machine learning algorithms are more robust.

According to another aspect, an improved method for annotating training data is provided. The method includes receiving a set of training cross-sectional image segments of a semiconductor object of interest and selecting a parameterized description of the semiconductor object of interest. According to the method, the parameter values of the parameters of the parameterized technique are adjusted for each detected instance of a semiconductor object of interest. Each detected instance of a semiconductor object of interest within the set of training cross-sectional image segments is annotated with adjusted parameter values. Parameterized techniques with adjusted parameter values can be displayed to a user via a user graphical user interface or display.

According to another aspect, a method is provided for generating training data for a machine learning algorithm for image segmentation. The method includes receiving a set of training cross-sectional images of a semiconductor object of interest and selecting a parameterized description of the semiconductor object of interest. The method includes adjusting the parameter values of the parameters of the parameterized technique for each detected instance of a semiconductor object of interest in a set of training cross-sectional image segments and annotating pixels according to the adjusted parameter values of the parameters of the parameterized technique. It involves automatically annotating each image pixel with a value.

According to a second embodiment, a method is provided for measuring parameter values of at least a plurality of semiconductor objects of interest by one single step or a continuous machine learning algorithm. As a result, the method for measuring at least parameter values of a plurality of semiconductor objects of interest extracts a list of parameter values of a parameter description of the plurality of semiconductor objects in a 2D image of a region of interest within a wafer. In an example, a set of measurement results is generated for each cross-sectional image of each HAR structure within the inspection volume. The set of measurement results according to the parametric description may include diameter, centroid position offset, area, distance, ellipticity or other general geometric properties. The set of parameters may further include an estimated material composition or other physical properties of the fabricated semiconductor device, or the number of instances of a repetitive three-dimensional structure within the region of interest.

According to a second embodiment, machine learning methods (e.g. deep neural networks) are used to accomplish metrology or measurement tasks, providing continuous measurement results. These instrumentation or measurement tasks by machine learning require large training data sets that cover the range of expected measurements. However, instead of annotating with discrete, distinct classes, each training image is annotated with a list of corresponding measurement results of at least one semiconductor structure of interest within the training image. Accordingly, the measurement method according to the second embodiment is configured for training by a set of annotated training images. By the method for generating or completing training images according to the first embodiment, successive machine learning algorithms can be trained with less user interaction and lower effort, and thus can be modified or adapted to new or changing measurement tasks. there is. Accordingly, the second embodiment provides a fast, robust and adaptive measure extractor.

According to an aspect of the method according to the second embodiment, the method is further configured for generating modified annotated training data during a monitoring task and for further training with the modified annotated training images. The initial annotated training images initially cover the expected range of parameter values. Modified annotated training images can cover an extended range of parameter values. In an example, a method is provided for automatically generating initial or modified annotated training images within a range of parameter values to be measured by a measurement task.

The method of the second embodiment is configured for minimal user interaction during monitoring tasks. The measurement method according to the second embodiment uses a continuous machine learning algorithm to measure parameter values according to the parameter description. Measurement methods using parametric techniques offer the advantage of being very robust against imaging variations. Measurement methods using machine learning algorithms can operate at high noise levels where classical measurement methods fail. The measurement method according to the second embodiment can be implemented to be performed at very high speed. Moreover, because the complex physical simulation models of classical metrology methods do not need to be implemented, the method is easily adaptable to a variety of semiconductor objects of interest, including changes in the semiconductor object of interest or changes in the imaging conditions of a charged particle imaging system. It can be.

According to an example, the measurement method includes acquiring a series of J cross-sectional image slices of a semiconductor wafer at an inspection location. The series of J cross-sectional image slices includes at least a first cross-sectional image slice at a first angle and a second cross-sectional image slice at a second angle through the examination volume. The first and second angles may be the same or different. Cross-sectional image slices are created and acquired by the slice and imaging method, using a dual beam device containing a FIB column for milling and a charged particle imaging system for imaging.

According to an example, the method is applied to a HAR channel of a memory device, and a plurality of HAR channel cross-sectional images are generated by the slicing and imaging method. The measurement method is trained according to the parametric description of the HAR channel, including a set of elements containing circles or rings. The measurement method provides as a measurement result a parameter value, for example a numerical value of a radius, a diameter, an ellipticity, or a deviation from the average central position of a set of circles or rings.

In an example, the measurement method according to the second method further includes determining an inspection position of the inspection volume within the wafer and adjusting the wafer to the inspection position in the cross section of the dual beam device. Inspection locations may be obtained from additional inspection tools or for inspection control files or lists that are created and provided from a list of locations in a process control monitor.

A system configured to implement the measurement method according to the second embodiment is described in the third embodiment. The system according to a third embodiment comprises a user interface configured to receive user information about the measurement task, for example CAD information of the inspection volume or expected ranges of desired measurement value ranges. The system is configured to combine user information with process information of an image creation process, for example selected imaging parameters of a dual beam system. The user interface is coupled to a processing engine configured to combine user and process information and generate annotated training images with reduced user interaction. Process information of the image creation process may include, for example, a library of effects during image creation. According to an example, the processing engine is configured to generate annotated training images by physical simulation or image processing.

In an example, the processing engine is configured to generate training images by physical simulation, based on CAD-data of the three-dimensional structure. Accordingly, the user interface is configured to receive, display, select and store CAD-data in memory. The processing engine is configured to take into account information from an imaging process, for example with a dual beam device. Accordingly, the user interface is configured to receive, display, and select imaging parameters of the dual beam device. Imaging parameters may be selected by the user depending, for example, on the required speed or accuracy of the measurement task. The processing engine is also configured to take material contrast into account, for example by means of predetermined and stored library data of secondary electron yield for a particular material composition of the semiconductor object of interest. The processing engine is also configured to take into account imaging noise depending on secondary electron collection efficiency and residence time depending on the material composition and imaging parameters of the semiconductor object of interest.

In an example of the third embodiment, a system for performing a measurement task is configured to measure parameter values of a semiconductor object in an inspection volume with high throughput. The system includes a FIB column arranged and configured to mill a series of cross-section surfaces into the wafer surface at the inspection site and a charged particle imaging microscope arranged and configured to acquire digital images of the series of cross-section surfaces. The system includes a stage configured to hold and position an inspection area of the wafer and a control unit configured to control operations for milling and imaging a series of cross-sectional surfaces. The system further includes a calculation or processing unit configured to measure a plurality of parameter values inside the inspection volume of the semiconductor wafer according to the method of the second embodiment. The computing or processing unit includes a memory with installed software and a processing unit configured to compute and process digital images of a series of cross-sectional surfaces according to installed software code. The computational or processing unit communicates with an interface for receiving commands and a control unit for receiving digital image data. The calculation or processing unit communicates with the interface or control unit of the dual beam device to exchange and provide control commands such as milling angle (GF) and y-position of the cross section through the inspection volume.

A system for performing measurement operations of semiconductor objects has the following features: an imaging device configured to provide an imaging dataset of a wafer, a graphical user interface configured to present data to a user and obtain input data from the user, one or more processing devices, and one or more machine-readable hardware storage devices containing instructions executable by one or more processing devices to perform operations comprising one of the methods disclosed herein. The invention also relates to one or more machine-readable hardware storage devices containing instructions executable by one or more processing devices to perform operations including one of the methods according to the first or second embodiment.

Accordingly, according to the present disclosure, a wafer inspection method for measurement of a semiconductor object of interest can be adapted to quickly respond to changing conditions, such as changes in the measurement task, changes in the charged particle beam imaging system, or changes in the semiconductor object of interest itself. It is possible to adapt. Accordingly, a generalized wafer inspection method with high flexibility is provided.

In accordance with the present disclosure, systems and methods are provided for volume inspection of semiconductor wafers with increased throughput. The system and method are configured to mill and image a suitable cross-sectional surface in an inspection volume and determine measurement parameters of a 3D object from the cross-sectional surface image. The present invention provides a device and method for 3D inspection of an inspection volume within a wafer and for measuring parameter values of semiconductor objects within the inspection volume with high throughput, high accuracy, and reduced wafer damage. The method and device are used for quantitative metrology, but may also be used for defect detection, process monitoring, defect review, and inspection of integrated circuits within semiconductor wafers.

Although the examples and embodiments are described in the example of a semiconductor wafer, it is understood that the present invention is not limited to semiconductor wafers, but can also be applied to reticles or masks for semiconductor manufacturing, for example.

The present invention described by examples and examples is not limited to the examples and examples, and can be implemented by those skilled in the art through various combinations or modifications. The invention will be understood much more completely with reference to the following drawings:
1 shows a diagram of a wafer inspection or metrology system for 3D volume inspection by a dual beam device.
Figure 2 is a diagram of a slice and image method of volume inspection within a wafer.
Figure 3 shows an example of a cross-sectional image obtained by the slice and image method.
Figure 4 is a simplified diagram of the method steps according to the invention.
Figure 5 shows an example of a training method according to the first embodiment.
Figure 6 shows a first example of a user interface display configured for a training method according to the first embodiment.
Figure 7 shows a second example of a user interface display configured for a training method according to the first embodiment.
Figure 8 shows a third example of a user interface display configured for a training method according to the first embodiment.
Figure 9 shows an example of a physical simulation of a training image according to the first embodiment.
Figure 10 shows an example of improvement of training data according to the first embodiment.
Figure 11 is a diagram of process steps of an annotation process according to the first embodiment.
Figure 12 shows a method for measuring parameter values according to the second embodiment.
Figure 13 shows a machine learning algorithm for the measurement task.
Figure 14 shows two examples of analysis and display of measurement results.
Figure 15 shows a system for performing a measurement task according to a third embodiment.
Figure 16 shows an example of a parameter description describing an abstract relationship.
Figure 17 shows an example of measurement of parameter values of an abstract description.

Throughout the drawings and detailed description, like reference numbers are used to describe like features or elements. The coordinate system is chosen so that the wafer surface 55 coincides with the XY plane.

Recently, for the investigation of the 3D inspection volume of semiconductor wafers, a slicing and imaging method applicable to the inspection volume inside the wafer was proposed. Thereby, a 3D volume image is created in the inspection volume inside the wafer with a so-called “wedge cutting” approach or wedge cutting geometry, without the need for removal of the sample from the wafer. The slice and image method is applied to inspection volumes of dimensions of a few μm, with lateral extensions of 5 μm to 10 μm, for example on wafers with a diameter of 200 mm or 300 mm. The lateral extensions can also be larger and reach up to tenths of a micrometer. V-shaped grooves or edges are milled from the top surface of the integrated semiconductor wafer to provide access to the cross-sectional surface at an angle relative to the top surface. A 3D volume image of the inspection volume is acquired from a limited number of measurement sites, for example from representative areas of the die, from areas identified, for example, by a process control monitor (PCM), or by another inspection tool. The slice and image method will only locally destroy the wafer, and other dies may still be used, or the wafer may still be used for further processing. The method and inspection system for generating 3D volume images are described in WO 2021/180600 A1, which is incorporated herein by reference in its entirety. An example of a wafer inspection system 1000 for 3D volume inspection is shown in Figure 1. The wafer inspection system 1000 is configured for slicing and imaging methods under a wedge cut geometry with a dual beam device 1. For the wafer 8, several measurement sites, including measurement sites 6.1 and 6.2, are defined in a location map or inspection list generated from an inspection tool or from design information. The wafer 8 is placed on the wafer support table 15. The wafer support table 15 is mounted on a stage 155 with actuators and position control. Actuators and means for precision control for wafer stages, such as laser interferometers, are known in the art. The control unit 16 is configured to control the wafer stage 155 and adjust the measurement area 6.1 of the wafer 8 at the intersection 43 of the dual beam device 1. The dual beam device 1 includes a FIB column 50 with a FIB optical axis 48 and a charged particle beam (CPB) imaging system 40 with an optical axis 42 . At the intersection 43 of both optical axes of the FIB and CPB imaging systems, the wafer surface is arranged at a tilt angle GF with respect to the FIB axis 48. FIB axis 48 and CPB imaging system axis 42 comprise an angle (GFE), and the CPB imaging system axis forms an angle (GE) perpendicular to the wafer surface 55. In the coordinate system of Figure 1, the normal to the wafer surface 55 is given by the z-axis. A focused ion beam (FIB) 51 is generated by a FIB-column 50 and impinges on the surface 55 of the wafer 8 under an angle GF. The inclined cross-sectional surface is milled into the wafer by ion beam milling at the inspection area 6.1 at approximately the tilt angle GF. In the example of Figure 1, the tilt angle GF is approximately 30°. The actual tilt angle of the inclined cross-sectional surface may deviate from the tilt angle GF by up to 1° to 4°, for example due to beam divergence of the focused ion beam, such as a gallium-ion beam. With the charged particle beam imaging system 40 tilted at an angle GE relative to the wafer normal, an image of the milled surface is acquired. In the example of Figure 1, angle GE is approximately 15°. However, for example, when GE = GF, the CPB imaging system axis 42 is perpendicular to the FIB axis 48, or when GE = 0°, the CPB imaging system axis 42 is perpendicular to the wafer surface 55. Other vertical arrangements are likewise possible.

During imaging, the charged particle beam 44 is scanned along a scan path over the cross-sectional surface of the wafer at the measurement site 6.1 by the scanning unit of the charged particle beam imaging system 40, and secondary particles as well as scattering particles are detected. is created. Particle detector 17 collects at least some of the secondary and scattered particles and communicates the particle count with control unit 19. Other detectors for other types of interaction products may likewise exist. The control unit 19 controls the charged particle beam imaging column 40 and the FIB column 50 and is connected to the control unit 16 to control the position of the wafer mounted on the wafer support table through the wafer stage 155. do. The control unit 19 communicates with the motion control unit 2, which triggers, for example, the positioning and alignment of the measuring area 6.1 of the wafer 8 at the intersection 43 via wafer stage movement, FIB milling, Trigger the operations of image acquisition and stage movement repeatedly.

Each new intersecting surface is milled by the FIB beam 51 and imaged by a charged particle imaging beam 44, for example a scanning electron beam or a helium-ion-beam in a helium ion microscope (HIM). In an example, the dual beam system includes a first focused ion beam system 50 arranged at a first angle GF1 and a second focused ion column 50 arranged at a second angle GF2, and the wafer is aligned at the first angle GF2. is rotated between milling at a second angle GF1) and GF2, while imaging is performed, for example, by an imaging charged particle beam column 40 arranged perpendicularly to the wafer surface.

Figure 2 shows a wedge cut geometry in an example of a 3D-memory stack. Figure 2 shows the situation when surface 52 is a new cross-sectional surface that was last milled by FIB 51. The cross-sectional surface 52 is scanned, for example, by a SEM beam 44 arranged to be of normal incidence on the wafer surface 55 in the example of Figure 2, and a high-resolution cross-sectional image slice is produced. The cross-sectional surfaces 53.1...53.N are subsequently milled with the FIB beam 51 at an angle GF of approximately 30° to the wafer surface 9, but for example GF = 20° and GF = 60. Other angles (GF) between ° are likewise possible. The cross-sectional image slice is a first cross-sectional image feature formed by the intersection with a high aspect ratio (HAR) structure or via (e.g., the first cross-sectional image feature of the HAR structure 4.1, 4.2, 4.3) and, for example, SiO2, SiN - or a second cross-sectional image feature formed by the intersection with a layer comprising tungsten lines (L.1...L.M). Some lines are also referred to as “word lines.” The maximum number of layers (M) is typically greater than 50, for example greater than 100 or even greater than 200. The HAR structures and layers extend throughout most of the inspection volume of the wafer but may also include gaps. HAR structures typically have a diameter of less than 100 nm, for example about 80 nm or for example 40 nm. Accordingly, a cross-sectional image slice includes a first cross-sectional image feature as an intersection or cross-section of the HAR structure footprint at a different depth (Z) at each XY-position. For vertical memory HAR structures of cylindrical shape, the first cross-sectional image features obtained are circular or elliptical structures of varying depth determined by the location of the structures on the inclined cross-sectional surface 52. The memory stack extends in the Z-direction perpendicular to the wafer surface 55. The thickness (d) or minimum distance (d) between two adjacent cross-sectional image slices is typically adjusted to values of the order of a few nm, e.g. 30 nm, 20 nm, 10 nm, 5 nm, 4 nm or even less. do. Once a layer of material of predetermined thickness d has been removed with the FIB, the next cross-sectional surface 53.i...53.J is exposed and accessible for imaging with the charged particle imaging beam 44.

Figure 3 shows an example of a cross-sectional image slice 311.1 produced by an imaging charged particle beam and corresponding to a cross-sectional surface 52. Cross-sectional image slice 311.1 includes an edge line 315 between the inclined cross-section and the surface 55 of the wafer at edge coordinate y1. To the right of the edge, image slice 311.1 shows several sections 307.1...307.S through the HAR structure intersected by section surface 52. Additionally, image slice 311.1 includes cross-sections of several word lines 313.1 to 313.3 at different depths or z-positions. With these word lines 313.1 to 313.3, a depth map Z ₁ (x, y) of the inclined cross-sectional surface 52 can be created.

According to aspects of the invention, a fast and robust method for performing measurement tasks is provided. Additional details of the method for performing the measurement task are described below in the second embodiment, and a brief description of the method is illustrated in the example in FIG. 4 .

From high-resolution 3D volumetric image data (i.e., having a plurality of more than 100, preferably more than 1000 image slices), characteristics of the included 3D structures, such as average tilt angle, minimum diameter of the repetitive semiconductor structures, Distance, curvature, and overlay errors can be determined, and multiple image slices or virtual image splices can be extracted. With a plurality of image slices or virtual image slices, a first machine learning algorithm can be trained and a minimal set of cross-sectional image slices for measurement of properties of repetitive 3D structures can be determined. With the second machine learning algorithm, properties of repetitive 3D structures can be determined from a set of minimal cross-sectional image slices with high accuracy and high throughput. The method is explained by the following steps.

In step M1, multiple high-resolution 3D volume images of a representative examination volume are generated. Multiple high-resolution 3D volume images can be generated by slicing and imaging methods applied to a representative test wafer, or by simulation, for example, by altering the 3D volume images obtained by measurements.

In step ML2, the properties of interest of the repetitive semiconductor structure are determined from a plurality of high-resolution 3D volume images. Accordingly, each high-resolution 3D volume image expresses certain characteristics described by at least one parameter. The plurality of high-resolution 3D volume images are labeled with parameter values of at least one parameter.

In step ML3, a plurality of labeled or annotated cross-sectional image slices are extracted or created from the plurality of labeled high-resolution 3D volume images. Slices can be measured as image slices calculated from high-resolution 3D volume images or as virtual image slices.

In step ML4, a machine learning model is trained with multiple annotated cross-sectional image slices and multiple labeled high-resolution 3D volume images. Training can be accomplished iteratively to determine the required minimum set of cross-sectional image slices and determine parameter values of interest with a given accuracy and confidence.

In step ML5, a set of measured cross-sectional image slices are determined, for example from measurements at a new inspection site on the wafer.

In step ML6, the trained model according to step ML4 is applied to the set of measured cross-sectional image slices.

In step ML7, outputs of parameter values and confidence values according to the trained model are generated for a new inspection site on the wafer.

The method and the trained model further iterate the trained model on new inspection results of a new wafer for measurement, including generation of a new 3D volume image by simulation triggered, for example, by a low confidence value according to step ML7. and can be improved by adaptation.

The preferred method for performing the measurement task involves one single continuous machine learning algorithm. Successive machine learning algorithms are applied directly to the image and multiple measurement results are obtained. These continuous machine learning algorithms are sometimes also referred to as deep learning algorithms. Methods for performing measurement tasks by machine learning are generally described by way of example in US 2020/0258212 A1, which is fully incorporated herein by reference. According to the improvement provided by the present invention, the regressor of the machine learning algorithm generates measurements of the plurality of HAR channels directly from the new 2D image. The sequential machine learning algorithm is described below in a second embodiment. To perform the measurement task in a single step consisting of a single sequential machine learning algorithm, a large amount of annotated training image data is required. Annotated training image data is required to cover the desired range of measurements of the selected physical or geometric parameters. During the segmentation step according to the prior art, pixels or voxels within the 2D or 3D image are assigned to selected structures. Training machine learning segmentation requires providing pixel/voxel annotations for multiple pixels or voxels in multiple images. This standard method for segmentation by user interaction and pixel-by-pixel segmentation is very time consuming - especially with precision requirements such as in the semiconductor industry where structures are sometimes limited to only a few 3 or 4 pixels.

According to a first embodiment, an improved segmentation and annotation method is provided to generate training image data for measurement tasks with reduced used interaction. The method for generating training data according to the first embodiment relies on prior information. By using a priori information, the effort and especially user interaction during generation of training data is reduced to a minimum.

In a first example according to the first embodiment, the method includes a computer-assisted method for generating training data for a measurement algorithm. In a first example the generation of training image data is achieved by using a parameterized description of the semiconductor object of interest and the training image data is generated to cover the range of expected parameter values to be measured. Examples of elements of parameterized geometric descriptions are rings or shells of high aspect ratio (HAR) structures or HAR channels. The method is illustrated in Figure 5.

In the first step MSA1, a first set of training images of the examination area is acquired, for example the cross-sectional image 311.1 in Figure 3 acquired by the dual beam charged particle microscope 1 described above. Each training image includes at least one cross-section of a semiconductor object of interest, for example a cross-section of a HAR structure 307.1 to 307.S. At least a first image of the first set of training images is displayed to the user via a user interface.

In the second step MSA2, a user selection of a parameterized description of the semiconductor object of interest is received. The parameterized description of the semiconductor object of interest may include a geometric shape such as a circle, ring, ellipse, polygon, or a series expansion, such as the expansion coefficients of a Fourier series expansion. Parameterized descriptions can be expressed in different coordinate systems, for example Cartesian or polar coordinates. Parameterized techniques can be selected from libraries containing common user templates of parameterized techniques. A public list of parameterized techniques may be displayed via the user interface. In the example shown in Figure 6, the user may select the number of four circular rings as the parameterized description 409.j, with the center coordinates of the circular rings and the four radii (r1 to r4). The user interface 400 shown in FIG. 6 includes an image display area 403, a toolbar 405 and the display of at least some graphical representations of predefined parameterized techniques 409.i...409.l. It includes a selection box 407 with. The user interface is configured for graphical interaction with the user and may receive instructions for selection of parameter descriptions. By means of the elements of the toolbar 405 , conventional tools of a graphical interface can be selected, for example tools for changing the magnification of the cross-sectional image 311.1 displayed in the image display area 403 .

In the third step MSA3, a graphical representation according to the selected parameter description is displayed as an overlay over the training image on the display of the user interface.

The user interaction is then limited to changing a few parameters of the parameterized technique, for example changing the center position, the radius of the circle or changing the positions of the points of the polygon instead of pixel-wise annotation. The user interface is configured, for example, for manual placement of elements of a graphical representation of a parameterized technique and for manual selection of parameter values.

As an annotation value, the selected parameter value according to the selected parameter description is stored and serves as an annotation for each training image. Step MSA 3 can be repeated for as many cross-sections of the HAR structure (307.1 to 307.S) as desired.

An example is shown in Figure 7. The user interface 400 provides an image display of the 2D image 311.1 to be annotated, a toolbar 405 and a selection box 407 along with the display of a graphical representation of the parameterized technique 409.j selected according to step MSA 2. Display again in area 403. The user interface is configured for graphical interaction with the user and can receive commands for the selection of parameter values of the selected parameter description, in this example the two coordinates of the lateral positions x and y and the four radii of the circle. . The user interface can be configured for graphical interaction as well as numerical input of parameter values. The user is assisted by the user interface display 400 and can easily detect the edges or boundaries of the HAR channel and adjust the circle of parameter descriptions accordingly. The measurement results are used as annotation labels for the identified semiconductor objects of interest. An example of a graphical representation of a selected parameter description (409.j3) is shown; The user interface is configured for the location of the graphical representation 409.j3 and graphical adjustment of the parameter values of the selected parameter description. The parameter values of the selected parameter description are displayed as tuples in the list of annotated parameter values 411.

In general, the determination of the parameter values to be measured relies on the known imaging scale or magnification of the cross-sectional image 311.1, which can be predetermined, for example, in the calibration phase of the charged particle beam imaging system 40. However, an exact determination of the imaging scale is not required, and the parameter values can also be given for a constant scale of the training image.

Step MSA3 can be repeated for the first set of plurality of training images and a plurality of annotated training images are generated. The annotation contains a tuple of parameter values according to the selected parameter description. An example of the tuple [x,y,r1,r2,r3,r4] is illustrated in the annotation parameter value list 411 for display of tuples annotated on actual training images. The annotation parameter value list 411 is the annotation [x,y,5,7.3,11,13] ₁ , [x,y,6,7,11,13.5] ₂ , and the graphical representation 409 achieved in this example. contains one undecided tuple [x,y,r1,r2,r3,r4] ₃ according to the actual annotation with j3).

According to an example, user interaction may be further reduced by computer-assisted detection and positioning of instances of semiconductor objects of interest. In this example, the method according to the first example further comprises a computer automated segmentation and annotation step (CASA) (see Figure 5). During step CASA, instances of e.g. HAR cross-sections are automatically detected and analyzed. Detection and analysis are based on selected parameterized techniques and, for example, filtering algorithms. The processing algorithm supports the user for semi-automatic or automatic placement and adjustment of the elements of the parameterized technique, thereby allowing readjustment of the center, radius, etc.

For example, in step CASA 1, a repeating instance of a semiconductor object of interest is determined by a Fourier method, and the repeating instance may be predicted from a raster position derived from a filtered Fourier spectrum of the image.

In step CASA 2, initial annotation parameter values for each detected instance according to step CASA 1 are automatically determined. The initial parameter values of the selected parameter description can be obtained by known image processing methods, such as correlation with a matching filter or application of circular Hough transformation to directly detect the radius of a circle. In general, the Hough transform is a robust method for detecting a general parameterized description of a line or circle or geometric shape. Matching filters can be automatically generated according to the selected parameter description. For example, other methods include, for example, contour detection methods. Thereby, at least some edges or boundaries of the HAR channel can be detected and annotation parameter values can be determined. Annotation parameter values representing measurement results are used as annotation labels for identified instances of semiconductor objects of interest. Another method for recognition of instances of semiconductor objects of interest and measurement of parameter values may employ, for example, the RANSAC algorithm.

In another example for determination of initial annotation parameter values for each detected instance of a semiconductor object of interest, the best fitting cross section of the semiconductor object of interest is derived from information from the parameter description and CAD data. Simulated cross-section images of the semiconductor object of interest with different specific parameter values are generated by physically influenced forward simulation and compared to the measured cross-sections. The parameter values of the parameter description most likely to be present in the measured image are determined by optimization. Details of the physically affected forward simulation will be described below.

In another example for determination of initial annotation parameter values for each detected instance of a semiconductor object of interest according to step CASA 2, a matching step is applied. The matching step includes determination of matching terms through a modified intersection-over-union (IoU) ratio. IoU is known to determine the degree of similarity between two objects A and B, for example, object A in an image and object class B used for training data. According to an example of the present invention, the modified IoU adopts the area of maximum intersection or overlap between the bounded object and the selected parameter description. In an example, accurate determination of IoU may be applied using parametric techniques rather than a surrounding rectangular bounding box. Thereby, object detection is further improved.

In an example, the matching step includes calculating a first intersection/union ratio based on the outermost boundary ring of the parameter description selected according to the example of the HAR channel shown in FIGS. 6 and 7. The second and other intersection/union ratios may be calculated iteratively based on the additional parameters of the parametric description. However, it is also possible to match a bounded object to multiple parameter values of a parameter description in a single intersection/union step.

In an example, the matching step may further include calculating a detection loss. Various algorithms of loss functions are known and applicable to the present invention. However, instead of general modeling, the joint box parameter loss according to the second embodiment is adapted to include predicted or “measured” parameter values and their corresponding L1 differences with respect to the training data. In general, loss modeling for predicted tuples versus annotated tuples included in the training data is

(1) Classification loss;

(2) bounding box parameter loss or confidence value, and/or

(3) It can be constructed by bounding box overlap loss.

The parameter loss is described by the absolute difference between the value of at least one parameter of the predicted object and the corresponding annotated parameter value for the training data, e.g., an L1 metric. In the example, the overlap loss is described by a generalized intersection/union ratio based on a geometric model according to the parametric description of the predicted object and a corresponding geometric model of the training data. The bounding box overlap loss may be limited to only the outer bounding box of the largest ring, for example.

The detection loss may include at least one or a weighted sum of two or more detection losses. Other formulations of machine learning algorithms that detect objects by predicting their parameter representations are likewise suitable. For example, elements of a parametric description, for example all circles of a selected parametric description according to FIG. 7 , may be subjected to individual iterative object detection. By the matching step, a forward model from the parameter description to a plausible image is provided and parameter values for the best match of the image, for example of a HAR channel cross-section, are derived.

At step CASA 3, a graphical representation of the selected parameter description is displayed graphically via user interface display 400 at the detected instance of the semiconductor object of interest as an overlay on the actual training image 311.1. Each graphical representation may be displayed according to the initial annotation parameter values derived in step CASA 2. Initial annotation parameter values may also be displayed in the annotation parameter list 411. During step CASA 3, the user interface is configured to receive confirmation or improvement of the initial annotation value, via user input. The user interface display 400 for confirmation or improvement may be similar to the user interface display 400 shown in FIG. 7 . Completion or confirmation is for example illustrated by a checkbox in the annotation parameter value list 411, where the first two tuples are displayed as confirmed. With the computer-automated segmentation and annotation step CASA 3, user interaction is reduced even further.

Steps MSA 3 and CASA can be performed in parallel. Users can perform segmentation and annotation in parallel according to step MSA 3 and check or improve the results of computer-assisted segmentation and annotation in step CASA 3.

In the third step MSA4, the parameter value ranges of the annotated sections are evaluated and displayed. The user is hereby informed whether the desired parameter value range has already been achieved within the first set of training data. An example is shown in Figure 8. The drawing may include a parameter list 413 with minimum and maximum annotation values. In the example of Figure 8, the parameter value of parameter r2 is selected for further display in histogram 414. Thereby, the user is informed about the parameter value ranges reached during the annotation method.

The training image set is obtained from actual cross-sectional images from the inspection volume and includes additional semiconductor structures or objects, such as word lines (reference numeral 313 in Figure 3). Parametric techniques help segment and annotate training images even in the presence of these additional semiconductor structures. Therefore, machine learning algorithms trained with annotated training images will be robust even in the presence of additional semiconductor structures.

At this point, a first set of training data (TD.1) can be created and stored in memory for further use. A training data set (TD) typically contains a plurality of annotated training images and a parameter description where tuples containing annotated parameter values are interpreted accordingly. The first set of training data (TD.1) typically has a limited range of measurement results for parameter values. With the set of training cross-sectional image segments, a first set of training data (TD.1) of a limited range of parameter values is generated. Therefore, it is possible for a machine learning algorithm or regressor trained with limited training data (TD.1) to detect and determine continuous quantitative measurements of the semiconductor structure of interest in new images only within a limited range of parameter values. Moreover, for metrology applications, a quality measure of a machine learning algorithm may be defined, for example, by the deviation of actually measured parameter values from training or annotation parameter values. Therefore, quality measures can be improved by using dense coverage of predetermined parameter value ranges. Therefore, multiple training data containing multiple annotations must be generated. According to a second example of the method of the first embodiment, coverage of the predetermined parameter value range is automatically achieved by using the first set of training images (TD.1) obtained by user interaction during steps MSA 1 to MSA 4. It improves. For example, during the MSA4 step, the user interface may be configured to receive user input containing, for example, a range of parameter values required for a measurement task or a specific range of values within a range of parameter values, wherein an annotated training image Better sampling with is required. During step MSA 5, additional annotated training cross-sectional image segments with different or additional parameter values within the specified parameter value range are generated from the first set of training images. Thereby, a second set of training images with different or additional parameter values can be generated, for example by image processing. Examples of suitable image processing include at least one of: variation of scale, change of shape, interpolation, morphological operations, or pattern replacement within the training image.

The first and second sets of training images are combined in step MSA 6 to build a complete set of training images or training data (TD.2). The segmentation and annotation results of the set of training images can be further used for mapping the pixels of the 2D training images and pixelated segmentation is achieved. Pixelated segmentation can also be used, for example, to combine measurement machine learning algorithms with defect detection and classification algorithms.

The completed training image or set of secondary training data (TD.2) is finally stored in memory for further use.

However, the generation of additional training images according to step MSA 5 is not limited to additional training images with different parameter values of the parameter description of the semiconductor object of interest. Additional training images may also include variations in image acquisition parameters, such as noise level or image contrast. Thereby, the machine learning algorithm trained with the second training data is more robust and provides measurement results with higher accuracy.

Accordingly, a system configured to perform the method according to the second example of the first embodiment is also capable of quickly adapting to new measurement tasks of semiconductor objects of interest of different size or scale, for example after reduction of the integrated semiconductor structure. According to a second example, the set of training data is based on existing annotated training images and covers, for example, a narrow range of measurements or a different range or measurements. Additional annotated training images are generated by image processing of the annotated training images to cover the required measurement range. This is useful, for example, when reduction is applied to a semiconductor structure, or when the magnification of the imaging process is changed.

3 to 8, the cross-sectional images and training images are simplified for illustration. As described above, cross-sectional images typically include hundreds to thousands of cross-sections of the HAR structure. Training images are typically much smaller than the actual measured cross-sectional images and contain, for example, only a small number of HAS structures, for example less than 50 or even fewer. Moreover, the training images should have a smaller pixel count, for example containing only 256, 512 or 1024 pixels in one dimension. Accordingly, the training image may be formed by, for example, 256×1024 pixels or 512×512 pixels. Therefore, training images are typically formed by segments of a cross-sectional image. The reduced number of pixels allows for fast and effective training of machine learning algorithms. However, with advances in processors, larger pixel counts are also possible, for example 2048 or 4096 pixels in one dimension.

In a third example of the method according to the first embodiment, user interaction is reduced even further. Methods for generating training data for training machine learning algorithms to perform quantitative measurements can generally be improved by a priori knowledge. According to a third example of the first embodiment, the method for generating training data applies a physical simulation of an imaging process of a cross-section through an inspection volume of a semiconductor wafer. The method is illustrated in Figure 9.

In step PSA1, user input is received for the expected cross-sectional area through the inspection volume of the semiconductor wafer. The expected cross-sectional area can be obtained from CAD information, for example. CAD-data can be displayed graphically, for example through 3D projection, with the difference that the cross-section through CAD data is displayed, for example, rather than the actual cross-section image, as shown in FIG. 6 It can be presented similarly. The interface is also configured for selection of a parameterized description of the semiconductor object of interest. Parameterized techniques can be selected according to CAD data. In the example, the parameterized techniques used are applied directly to CAD data of the semiconductor object of interest. Multiple instances of a semiconductor object of interest according to a parameterized description are created with material compositions available from CAD information. This creates a raw image, which is later used in physical simulations of image formation. For a set of training images, the parameter values of the parameterized description of the semiconductor object of interest are varied within a predetermined range of parameter values, and a plurality of raw images are generated. The parameter value range may be specified by user input, and the user interface is further configured to receive the parameter value range according to the parameters of the selected parameterized technology. Additional objects of the semiconductor, such as word lines, can be extracted from the CAD-data and considered in the raw image for the generation of annotated training images.

In step PSA 2, the imaging parameters of the charged particle beam imaging system 40 are determined. During this step, the user interface is configured to receive user instructions on how measurement images are to be acquired at the actual examination site. Imaging parameters can be described within the typical settings in which the charged particle imaging system operates. Typically, charged particle imaging systems provide a set of predefined imaging settings from which the user can select task-specific imaging settings. For example, imaging parameters required for physical simulations include electron energy, pixel size, scanning method, residence time, settings of the charged particle beam imaging system, e.g., selected contrast method including selected detector. These imaging parameters may be stored with each imaging setting in a list of predefined imaging settings.

The imaging parameters further include a list of predetermined material contrasts. The imaging parameters are not limited to just imaging, and may further include, for example, parameters of a milling operation by the FIB beam 51. Accordingly, the imaging parameters may further include the expected roughness or slope of the cross-sectional surface. The imaging parameters may further include a predetermined model of the curtain effect during the milling operation.

A number of imaging parameters may be predetermined and stored in look-up tables in the memory of a control unit of a charged particle inspection and metrology system, such as a system for performing automated measurements of semiconductor objects. Several imaging parameters may be specified for the charged particle beam imaging system 40 and may be predetermined during the calibration step.

In step PSA 3, a physically-influenced forward simulation (simply: physical simulation) of the imaging process is performed for a given charged particle beam imaging system according to step PSA 2 and the raw image according to step PSA 1. The physical simulation can be performed in the processing engine of the charged particle inspection and metrology system or in any other processing engine that has access to the information generated in steps PSA 1 and PSA 2.

The physical simulation typically includes scaled or raw image data for the image to be measured according to the pixel raster of the selected scanning method. The spatially resolved image contrast is determined by the material contrast of the materials present within the raw image data of the training image. After scaling and application of the material contrast values, convolution of the raw image data is performed with a convolution kernel according to the point spread function of the imaging system. The point spread function can be determined according to the expected interaction volume created by the primary charged particle imaging beam in a cross section through the wafer. The interaction volume typically depends on the electron energy. Noise levels can be added depending on the dwell time at each raster location. Thereby also the limited detection count of the selected detector geometry is taken into account.

Imaging parameters may also depend on the material composition within the inspection volume of the wafer and may include the curtain effect of the milling operation depending on the material composition of the cross-section to be milled. The curtain effect is accessible in a simple model of the milling operation and can therefore be taken into account as well. The milling effect creates, for example, additional topographic contrast superimposed on the material contrast.

The physical simulation can also consider additional structures, such as word lines, within the inspection area. By this, the measurement effect of the real image is taken into account in the training image.

In step PSA 4, training images obtained by simulation can optionally be presented for the user via a user interface, and the user can reconsider choosing between steps PSA 1 or PSA 2, for example. For example, the user may select different magnifications during imaging, or different milling angles through the examination volume. If the user requests any changes, the process will repeat with step PSA 3. Once the user confirms the simulation image, the method is complete. The annotated training images obtained by simulation form a set of training data (TD.3) and are finally stored in memory for further use. In a third example of the first embodiment, a large number of training cross-sectional image segments covering the range of required parameter values are automatically generated with minimal user interaction.

A general advantage of the segmentation and annotation method according to the first embodiment is a significant reduction in annotation effort by reducing user interaction to a minimum. According to a first embodiment, training data may be generated by one of the examples described above, or by a combination of examples. This generates a large amount of training data, including manually annotated training images, training images acquired by computer-assisted segmentation and annotation, training images acquired by image processing, or training images acquired by physical simulation. It can be.

In the example, generation of training image data is not complete after initial training. Typically, not all impacts or defects during manufacturing are predicted in initial training, and the training data of machine learning algorithms are frequently modified during monitoring of the manufacturing process. According to a fourth example of the first embodiment, a method is provided to further improve training data for a measurement task with reduced used interactions. For example, modified annotated training images can be provided to cover unexpected effects during the imaging process. This improves machine learning algorithms for measurement tasks to extract additional measurement results with increased robustness to deviations or aberrations.

The initial annotated training data (TD) may initially cover only a few effects of the image generation process, such as magnification, noise level, material contrast, and convolutional kernels such as interaction volumes. Imaging effects may be, for example, unexpected curtain effects during the slicing and imaging process, resulting in unexpected topography contrast in addition to the material contrast in the SEM image. This contrast from the curtain can be taken into account in the modified annotated training image data. Another effect is the charging effect. Local sample charging can affect the local secondary electron yield and thus change the material contrast and primary particle beam focus position and thus produce a locally distorted image or a locally altered image contrast. These images with locally altered image contrast and local distortion can be considered in the modified annotated training image data. According to a fourth example, modified annotated training images are generated, for example according to the expected range of imaging parameter values of the measurement task. A fourth example of the first embodiment for generating annotated training images is shown in Figure 10.

During an inspection operation, a plurality of cross-sectional images (MI) within at least one inspection volume of the wafer are generated. Enhanced segmentation and annotation methods can be automatically triggered in step ESA1. For example, during measurement methods that use machine learning algorithms, the reliability value or quality measure of the measurement result is monitored. The machine learning algorithm is trained with initial training data (TD.i) generated, for example, by any one of the first to third examples described above. If the confidence value or quality measure of the measurement result frequently exceeds a predefined threshold, automated generation of additional or modified annotated training images can be triggered. If image quality measurements, such as noise level or contrast level of measured images, frequently deviate from predefined image quality measurements, automated creation of additional or modified annotated training images may be triggered. The creation of additional or modified annotated training images can also be triggered by user interaction. During a measurement operation, at least one of the cross-sectional images is displayed to a user via user interface display 400 and the user interface is configured to receive user input to trigger advanced segmentation and annotation steps.

In step ESA 2, a method according to any one of the first to third examples of the first embodiment, or a combination, may be applied for generating an added or modified annotated training image. According to a first example, an additional set of annotated training images is generated by computer-assisted annotation of actual measured cross-sectional images. According to a second example, training images with added or modified annotations are generated by image processing. According to a third example, the added or modified annotated training images are generated by physical simulation, taking into account new or modified effects of cross-sectional image generation, or taking into account the modified raw images of step PSA 1 described above. . In the example, modified annotated training image data is provided to cover unexpected effects during manufacturing. During the measurement operation, certain error types can additionally be detected, for example defects due to regular and repetitive errors or defects, for example from lithography mask contamination. Other defects may be random, for example, sporadic contamination during a manufacturing process step. As a result of step ESA 2, improved training data (TD.e) is generated and stored in memory for further use.

The segmentation and annotation method according to the first embodiment provides the additional advantage of rapidly adapting the parameterized technique to new semiconductor objects or deviations in the geometry of the semiconductor object of interest by introducing new additional parameters to the selected parametric technique. . For example, if the HAR channel exhibits a deviation from a circular shape, additional parameters such as ellipticity or eccentricity may be introduced into the existing training data set. The introduction can be carried out, for example, in an improvement step or by automatic generation of additional training cross-sectional image segments covering the parameter value range of the newly introduced parameter, for example eccentricity. For example, if the material or material composition of the semiconductor object of interest changes, the training data can be automatically adapted to the new material contrast corresponding to the new material or material composition.

Figure 11 shows a fifth example of a method for generating training data according to the first embodiment. According to a fifth example, the initial parameter description is selected according to step MSA 2 and the cross-sectional image 311 is acquired as described in step MSA 1. Figure 11 shows a cross-section 307.i of an instance of a semiconductor object of interest in a cross-sectional image 311 acquired by the dual beam device 1. The semiconductor object of interest in Figure 11 is again a HAR structure, but other semiconductor objects of interest are likewise possible. The illustration of Figure 11 is limited to one instance of cross section 307.i of the HAR channel.

In a first step, all instances of a cross-section of the HAR-channel within the cross-sectional image are identified, for example with the method described in step CASA 1 above. Each cross section 307.i includes one central circle 317.1 with five circles 317.2 to 317.6 with different image contrasts (FIG. 11A).

In the second step, initial parameter values, including the radius according to the selected parameter description, are automatically determined according to the CASA 2 step. As shown in Figure 11b, not all geometric shapes of the parametric description can be detected, only the circles 319.1, 319.2, 319.5, 319.6 are automatically detected, while the additional shapes of the circles 317.3, 317.4 It may happen that external circles are not detected.

In the third step according to CASA 3, incorrect or missing annotations can be corrected by the user. The user interface is configured to receive commands to modify, delete or add annotations, similar to step CASA 3 described above. The results are shown in Figure 11c with all circles (319.1 to 319.6) depending on the selected parameter description. The user interface can also be configured to completely ignore comments.

In a fourth step, the user interface is configured to receive commands to modify the parameter description. The result of the graphic overlay is shown in Figure 11D.

In the example, the actual cross-sectional image contains three significant deviations (325.1 to 325.3) from the original of the initial parameter description. For example, the user interface is configured to receive instructions to modify one element of the parameter description to another element. In the example of Figure 11D, the initial parametric description of circles 319.1, 319.2 is modified to a modified parametric description of spline curves or closed polygonal lines 323.1, 323.2. Thereby, the deviations (325.1, 325.2) from the ideal circle are minimized. Accordingly, the user interface is configured to receive instructions for modifying the selected initial parameter description 319 to the selected modified parameter description 325 and to receive instructions for adjusting parameter values of the modified parameter description 323. .

In step 5, parametric techniques, including modified parametric techniques 323, are used to automatically generate pixel-wise labels of cross-sectional images. The results are shown in Figure 11E. Each pixel is automatically labeled according to ring segment 327.1 to 327.6. The processing engine is configured to perform labeling according to the location of the pixel relative to the modified parameter description 323. The processing engine may also be configured to perform labeling according to the dominant or average gray value within each region defined by the modified parameter description 323 shown in FIG. 11D. By this, it is possible to improve the labeling to cover further deviations, for example deviation 325.3.

In step 6, the post-processing routine algorithmically optimizes the pixel-wise annotation, for example by snapping to strong edges or by smoothing the labeled regions. The results are shown in Figure 11f. Additional modification of the parameter description and fine tuning of parameter values according to the modified parameter description are induced. In the example shown in Figure 11, circle 319.5 does not cover deviation 325.3. This missing deviation is automatically corrected in step 6 by modifying the parameter description “circle” to an ellipse with one additional parameter value, for example, describing the numerical eccentricity. Thereby, the final parameter description is automatically determined and the training images are annotated by the improved annotation parameter values with high precision. In optional step 7, the results are presented to the user via a user interface. The interface is configured to receive confirmation or further modifications, such as step 4 or step CASA 3 described above.

In a further example of automated determination of initial parameter values according to Step 2 or Step CASA 2, a physically-influenced forward simulation model is applied for initial proposal of parameter values. As explained in the third example of the first embodiment above, the physically-influenced forward simulation model is a series of models that model a physical or optical measurement process, for example with a charged particle beam imaging device 40 (see Figure 1). It is given as a numerical method. The physically-influenced forward simulation model (F) is applied to the raw image generated according to the parametric description (P) of the semiconductor object with selected imaging settings (S) of the charged particle beam imaging device (40). In the example, the parameter description (P) is a parameter description (P[x,y,r1,r2,r3,r4,r5,r6, e1,e2,e3,e4,e5, e6,e7]) and the materials (ei) for different rings corresponding to different material contrasts of the SEM imaging process. Materials are typically known in advance and can be set to fixed values. However, it is also possible to work with unknown materials. The physically-influenced forward simulation model (F) is typically independent of the input model (P).

The measured cross-sectional image 311.1 is given by Y. Ideally, the simulation model (F) represents the measured image Y = F(S; P), including noise. Therefore, the unknown object described by the parameter values of P may be obtained by optimization of the merit function (M(P)):

For example, the input parameters are given by random initial parameter values of the parameter description (P) and the selected imaging settings (S). The initial parameter values may also incorporate existing information about the expected radius of the HAR channel cross-section, which may be available from different gold standard measurements or prior information. Initial parameter values for further optimization can also be determined according to the method described above in step CASA 2. In this example, the initial parameter values are determined, for example, by image processing, and the optimized parameter values are determined following optimization using a physically influenced forward simulation model (F).

As a result of the optimization, a tuple [x, y, r1, r2, r3, r4, r5, r6] of the actual parameter values of the HAR channel cross section in the measured image (Y) is obtained. Since the parameters of the parametric technique (P) are often limited to include only a few parameters, classical optimization methods such as simulated annealing may also be applied. However, in other examples, the parametric description (P) of the HAR channel cross-section may include several additional parameters to describe any deviations from the circular shape. An example is shown in Figure 11. In order to efficiently compute a solution to this optimization problem with a large number of parameters to be determined, methods of differentiable programming can be applied. These methods are well known in the art.

As described above, optimization can be performed in an iterative manner, starting with a simple parametric description of the semiconductor object of interest and increasing the complexity of the parametric description using, for example, the deviation of the ring. This method is more robust compared to the example of using multiple parameters in advance according to a comprehensive parameter description.

Since the initial proposal of parameters can be verified and corrected according to Step CASA 3 or Steps 3 and 4 described in the fifth example, more consistent training data can be generated. Because annotation suggestions can be easily verified and corrected, correction of training data can not only be performed by very-skilled application experts.

It is understood that the examples and method steps of the first embodiment may be combined with each other. Thereby, user interaction is reduced to a minimum and a large set of annotated training images is generated by a small number of user interactions and the application of a priori information. User interaction is reduced to selection of parameter description, required parameter value range, and imaging mode of the imaging device. Additional user interaction may include a few initial annotations, a few modifications or corrections to the automated annotations, a few modifications to the parameterized description, and confirmation of selected annotation results.

Although the previous example of the first embodiment has been described for annotation of training data (TD) with parameter values of a semiconductor object of interest, the method according to the first embodiment can also be applied to a segmentation method. During segmentation, the cross-sectional image of the semiconductor object of interest is typically annotated pixel-by-pixel. Thereby, different elements of the semiconductor object of interest are labeled pixel-by-pixel within the image with corresponding element identifiers. For example, the rings in the HAR structure are labeled or annotated with their respective ring numbers. These training images can then be used to train an object detection machine learning model. Pixel-by-pixel annotation by users is a very time-consuming and error-prone method. Similar to the first embodiment, the parameterized description of the semiconductor object of interest enables the segmentation workflow with less effort. The pixels of the training images, segmented pixel-wise, can be separated by the method steps described above, for example by user selection of parameterized techniques and reduced user interaction, or by image processing or physically influenced simulations and optimizations. It is annotated according to the parameterized technique with appropriately adjusted parameter values. After annotating the initial training images during the initial segmentation workflow, additional pixel-by-pixel annotated training images can be generated by image processing or physically influenced simulations. Thereby, by the method according to the first embodiment, full image segmentation is possible with less effort and training image data for a machine learning model for object detection can be generated with less user interaction.

The calculation steps according to examples of the first embodiment may be implemented in a system comprising a processing engine configured to generate initial proposals for parameter values of a parameterized description of a semiconductor object of interest. Algorithms, such as physically-influenced forward simulation models, can be precompiled and stored as executable files in memory connected to the processing engine, ready for automatic execution or execution on demand.

According to the above-described method, training data (TD) can be generated on which a robust machine learning method can be trained. A machine learning algorithm trained by training data containing parameter values as annotation values is capable of performing measurements even if new images are subject to high noise levels. An example of an image containing high noise levels is shown in Figure 11g.

In previous solutions, collecting and annotating sufficiently large and diverse training data is extremely expensive, time-consuming, and dependent on user interaction. For example, efforts to manually annotate images of HAR channel cross-sections on a pixel-by-pixel basis scale very poorly, especially in the presence of acquisition noise. However, with the method according to the first embodiment, training also becomes possible for images containing high noise levels, since pixel-wise annotation is not required.

By means of annotated training image data, training of machine learning algorithms for measurement tasks can be achieved or modified and measurement results that are robust against defects or deviations can be obtained. Additionally, the training data may further include identifiers for fault classes depending on the specific fault type. According to a second embodiment, a method is provided for performing a measurement task of parameter values of at least a plurality of semiconductor objects of interest by a sequential machine learning algorithm. As a result, the method of performing a measurement operation extracts a list of parameter values of a plurality of semiconductor objects of interest within an inspection volume within the wafer. The parameter values represent the measurement results of the parameters of the parametric description, for example four circles with the center position and four radii described above. For each instance of a semiconductor object of interest, a tuple of parameter values, e.g. [x,y,r1,r2,r3,r4], is extracted as the measurement result. In an example, measurement results are generated for each cross-sectional image of each HAR structure within the inspection volume. The set of measurement results may include diameter, centroid position offset, area, distance, ellipticity or other general geometric properties. The set of measurement results may also include relative deviations from expected parameter values, for example according to the design specifications of the semiconductor object. The set of parameters may further include an estimated material composition or other physical properties of the fabricated semiconductor device, or the number of instances of a repetitive three-dimensional structure within the region of interest. The measurement method according to the second embodiment is not a classical measurement, for example by comparison with a scale or reference, by counting, or by using a classical measuring tool. An example would be a laser interferometer on stage. In contrast to classical measurements, the measurement method according to the second embodiment relies on appropriately trained and modified machine learning algorithms described below. The method according to the second embodiment is also referred to as a measurement extractor, and the measurement results are limited to the parameters of the parameter description used in the training data. The modified machine learning algorithm can be simplified as regression, whereby parameter values present in the training data can be reproduced with high confidence, and parameter values not present in the training data are measured with lower confidence. For example, measurement of the channel cross-section characteristics of a HAR channel is thus achieved by a trainable machine learning algorithm and no further empirical post-processing is required. The method of the measurement value extractor according to the second embodiment is illustrated in Figure 12.

In step ME1, a measurement operation is configured. A machine learning algorithm is selected for the measurement task. The selection may be received by user interaction or may be selected automatically depending on the parameter description used during generation of the training data.

If training data is not available for the measurement task, step TDG for training data generation is triggered. Stage TDG may be configured according to the first embodiment. As a result, training data (TD) is generated and stored in memory.

During step ME 2, training data (TD) from memory is used to train the selected machine learning algorithm. The trained algorithm (MA) is finally stored, for example, in the memory of the processing engine of the control unit of the measurement system. Once the machine learning algorithm is trained to associate a tuple to each detected instance of a semiconductor object of interest, it is possible to run the trained machine learning algorithm on any new cross-sectional image. The trained algorithm (MA) is then capable of detecting instances of semiconductor objects of interest and generating tuples of parameter values according to the selected parameter description used in the training images. Interpretation of parameter values is linked to the selected parameter description.

In step ME 3, at least one new cross-sectional image slice is created, for example at a new inspection site on the wafer. Inspection sites are typically listed on an inspection list and may include the typical locations of repeat inspection sites on a die-by-die or wafer-by-wafer basis. An inspection area of the wafer is placed by a wafer table under a charged particle beam microscope and at least one cross-sectional image is acquired. During the measurement operation, a plurality of cross-sectional image slices, for example through the examination volume, are generated, for example by means of a slice and imaging method. The method of image acquisition according to the slicing and imaging method and the device for image acquisition are explained in the context of FIGS. 1 to 3 of the present application. As long as new cross-sectional image slices are scaled to the same or very similar magnification scale as the training image during the measurement operation, a tuple of representative measurement results according to the selected parameter description of the semiconductor object of interest can be generated as output and the measurement operation can be performed. there is. In this way, the geometric properties of the cross section can therefore be extracted in a single step approach.

In step ME 4, a trained machine learning algorithm (MA) is applied to at least one new cross-sectional image slice and measurement results are obtained. The measurement result includes a list of tuples of extracted parameter values of detected instances of the semiconductor object of interest and a reliability value for each of the extracted parameter values. Measurement results (MR) are stored in memory for further use. It is understood that steps ME 3 and ME 4 can be repeated multiple times for multiple cross-sectional images, for example at different depths inside the examination volume, resulting from slice and image methods. Thereby, a plurality of measurement results (MR) are stored in memory for further use.

Machine learning algorithms typically include a series of layers, including an input layer, an output layer, and a so-called hidden layer. The hidden layer contains a series of abstract mathematical operations, which are available from established libraries for programming machine learning algorithms, for example. Machine learning algorithms as applied in the second embodiment are known as object detection. Object detection is typically implemented as a convolutional neural network (CNN). A CNN according to the invention has for example 10 to 200 convolutional layers followed by 1 to 3 fully connected layers. Convolutional layers are not limited to convolutions, and may also include non-linear operations such as normalization, pooling, or skip operations. Typically, the output of an object detection task is a list of tuples, each tuple containing five values corresponding to the bounding box parameters position (x, y), width and height, class label and confidence value.

An example of a modified machine learning algorithm according to the second embodiment includes the YOLO (“you only look once”) method as an object detection method known in the art. According to the YOLO-method, a machine learning algorithm “looks only once” at an image to predict where an object exists, along with its classification label and confidence value. By means of a single object detection network, these machine learning algorithms simultaneously predict bounding boxes and class probabilities. Another example of object detection according to the second embodiment is the “detection-transformer” or DETR-method, which is well known in the art. Within the DETR method, sets of predictions (or measurement results) are generated directly in parallel. DETR uses conventional CNNs in combination with converter-encoders and converter-decoders, all known in the art. Transformer-based architectures are originally used in the field of natural language processing. They can efficiently handle non-local correlations, for example, to connect the correlations of words in a sentence to extract their exact meaning. In image processing, this leads to the possibility of correlating non-local information. By this, the “order” or semantic contest of the pixels is determined, e.g. detection of a first object of the parametric description increases the probability of detecting further objects of the parametric description, e.g. within the image. A detected ring increases the probability of detecting additional rings in neighboring HAR-channels.

A schematic diagram of the modified machine learning algorithm (MA) is shown in Figure 13. In step MA 1 a new cross-sectional image (NI) is received. The new cross-sectional image (NI) may contain imperfections such as defects or noise and deviations from parameter values in the training data. In an example, a received cross-sectional image (NI) is improved by applying an image processing algorithm such as, for example, any of convolution, normalization, filtering, edge handling, etc. Accordingly, the purpose of step MA 1 may be to pre-process the new cross-sectional image (NI) so that the trained machine learning algorithm can operate more efficiently.

In step MA 2, one or more instances of the semiconductor object of interest are detected in the new cross-sectional image (NI). Detection is configured such that each instance of a detected object contains image features corresponding to a selected parametric description of the training image data. The output of step MA 2 may include one or more bounding boxes, at least some of which, preferably most of them and most preferably all of them contain instances of the semiconductor object of interest. One or more of the bounding boxes, most preferably each bounding box, may include a class label and/or confidence level associated with the bounded object. Association of bounded objects with class labels and/or confidence levels can be implemented through known algorithms.

In optional filtering step MA 3, the bounding box output by step MA 2 may be filtered based on its confidence level. This allows selecting bounding boxes with a higher confidence level than a predetermined threshold for subsequent steps. Accordingly, steps MA 2 and MA 3 allow the identification and selection of bounding boxes containing instances of the semiconductor object of interest with high probability. Therefore, each bounding box contains a new real image of the cross-section of the semiconductor object of interest, for example the cross-section through the HAR channel.

In step MA 4, a predicted tuple of parameter values is generated for at least one semiconductor object of interest within the bounding box. Each predicted tuple for each detected instance of a semiconductor object of interest corresponds to a tuple used during training of the training data, e.g. the center coordinates, e.g. the measurement results and confidence of the four radii of the HAR channel. Contains value. Other tuples corresponding to the selected parameter description are likewise possible.

Steps MA 2 to MA 4 are illustrated here as consecutive steps for simplified illustration. It is understood that the machine learning algorithm will perform the functions of steps MA 2 to MA 4 in an entangled manner.

Accordingly, the tuple generated by the modified machine learning algorithm (MA) includes a plurality of parameter values of the measured semiconductor object. Instead of trying to measure geometric properties from a new cross-sectional image (NI) of a bounded object, the method of performing the measurement task according to the second embodiment depends on the parametric technique used during training of the modified machine learning algorithm (MA). Accordingly, a modified machine learning algorithm (MA) is employed to match predicted parameter values with detected instances of semiconductor objects of interest.

According to a second embodiment, the output of the object detection operation is modified to include a tuple of parameter values defined in the selected parameter description. Accordingly, the output of the object detection operation is a direct prediction of at least a tuple of measurements for each detected instance of a semiconductor of interest within the cross-sectional image. Every predicted tuple represents the measurement or parameter value of one instance of the detected semiconductor object of interest. Because the parametric description corresponds to a “compressed” representation of the semiconductor object of interest, noise in individual pixels or deviation of the semiconductor object of interest from the parametric description does not significantly affect the predictive quality of object detection. The tuple may also further contain an object classifier “object/no object” or a classifier for distinguishing multiple objects with different parameterized descriptions. As an example, each tuple contains the center coordinates of an instance of the semiconductor object of interest and the four radii shown in Figures 6 and 7. The tuples correspond to annotation tuples used during generation of training data, for example according to the first embodiment.

In optional step ME 5 (Figure 12), the measurement results (MR) stored in memory are also analyzed, for example statistical analysis is performed or a graphical display is generated. An example is shown in Figure 14. In Figure 14A, the typical z-dependence of the average center position of a HAR structure or channel is known from a 3D volume inspection at a representative inspection location or a reference location. The average HAR channel trajectory 363 (dashed line) is derived from the distribution of centroid positions (x, y) at different depth levels (z). Figure 14b shows the measured parameter values, in the example of r1 for the number (#) of instances of the semiconductor object of interest here. As shown in Figure 14, each parameter value can be displayed along with its reliability value.

According to a second embodiment, machine learning methods are used to accomplish a metrology or measurement task, providing continuous measurement results. Accordingly, the second embodiment provides a fast, robust and adaptive measure extractor.

According to an aspect of the method according to the second embodiment, the method is further configured for generating modified annotated training image data during a monitoring task and configured for further training with the modified annotated training image data. As shown in Figure 12, generating or enhancing training data during step TDG and training or retraining machine learning algorithms to achieve the measurement task may be performed in parallel with the measurement task during step ME 4. You can.

The method of the second embodiment is configured for minimal user interaction during measurement or inspection operations. Measurement methods that utilize a single continuous machine learning algorithm offer the advantage of being very robust against imaging variations and can operate at high noise levels where classical measurement methods fail. The method can be implemented to perform very quickly. Moreover, since the complex physical simulation models of dedicated classical metrology methods do not need to be implemented, the method can easily be adapted to a variety of semiconductor objects of interest, including changes in the semiconductor object of interest, or changes in the imaging conditions of a charged particle imaging system. It can be adapted.

A system configured to implement the measurement method according to the second embodiment is described in the third embodiment. A system configured to perform measurement tasks by machine learning is shown in Figure 15. A system 1000 for performing a measurement task includes a dual beam system 1. The dual beam system is shown in more detail in Figure 1 and reference is made to the description of Figure 1. The essential features of the dual beam system 1 are a first charged particle or FIB column 50 for milling and a second charged particle beam imaging system 40 for high-resolution imaging of the cross-sectional surface. The dual beam system 1 comprises at least one detector 17 for detecting secondary particles, which can be electrons or photons. The dual beam system 1 further includes a wafer support table 15 configured to hold the wafer 8 during use. The wafer support table 15 is position controlled by a stage control unit 16 connected to the control unit 19 of the dual beam system 1. The control unit 19 consists of memory and logic to control the operation of the dual beam system 1.

The system 1000 for performing a measurement operation further comprises an operation control unit 2. The operation control unit 2 includes at least one processing engine 201 that can be formed by a GPU processor and multiple parallel processors with a common integrated memory. The motion control unit 2 further includes an SSD storage 203 memory, for example 8TB or more, for storing the training data generated during the training method, the trained machine learning algorithm and the plurality of cross-sectional images. The motion control unit 2 further includes a user interface 205 comprising a user interface display 400 and a user command device 401 configured to receive input from a user. The operation control unit 2 further includes a memory or storage 219 for storing process information of the image creation process of the dual beam device 1. The process information of the image creation process by the dual beam device 1 may for example include a library of effects and a list of predetermined material contrasts during image creation.

The operation control unit 2 is also connected to an interface unit 231 configured to receive additional commands or data, for example CAD data, from an external device or network. The interface unit 231 is also configured to exchange information, for example measurement results (MR), with an external device, and to store a set of training data or a trained machine learning algorithm or a plurality of cross-sectional images in an external storage.

The processing engine 201 is configured to take into account process information of an image generation process, for example with a dual beam device 1 , including for example selected imaging parameters of the dual beam system. Imaging parameters may be selected by the user depending, for example, on the required speed or accuracy of the measurement task.

The system 1000 according to a third embodiment is configured to receive user information about the measurement task, including, for example, CAD information of the inspection volume or expected ranges of desired measurement value ranges. System 1000 is configured to combine user information with process information of the image creation process. Processing engine 201 is configured to combine user and process information and generate training data (TD) with reduced user interaction. For example, processing engine 201 is configured to generate annotated training images by physical simulation or image processing. In an example, the processing engine 201 is configured to generate training image data by physical simulation, based on a CAD-model of a three-dimensional semiconductor structure. Accordingly, user interface 205 is configured to receive, display, and select CAD data. The processing engine 201 is also configured to take material contrast into account, for example by means of predetermined and stored library data of secondary electron yield for a particular material composition of the semiconductor object of interest. Processing engine 201 is also configured to consider imaging noise depending on secondary electron collection efficiency and residence time depending on the material composition and imaging parameters of the semiconductor object of interest.

Processing engine 201 is also configured to train a selected machine learning algorithm with training data stored in storage 203 and store the trained machine learning algorithm (MA) in storage 203 for later use.

According to a third embodiment, a system 1000 is provided for measuring parameter values of semiconductor objects within an inspection volume at high throughput. Accordingly, the processing unit 201 is configured to measure a plurality of parameter values inside the inspection volume of the semiconductor wafer according to the method of the second embodiment. Computation or processing unit 201 computes and processes digital images of a series of cross-sectional surfaces according to a trained machine learning algorithm (MA) stored in storage 203 and via user interface 205 or through interface unit 231 It is configured to provide measurement results (MR) extracted through ).

During the above-described examples, the method according to the first or second embodiment was applied to the measurement of parameter values of the parameter description of the HAR channel. The method can of course also be applied to other repetitive semiconductor objects of interest. The method can also be applied to rasters of repetitive semiconductor objects of interest, for example. An example is shown in Figure 16. Figure 16 shows a cross-sectional image slice with an ideal raster configuration 345 of the HAR channel cross-section. Each of the M HAR channel cross-sections can be described by a parametric description comprising, for example, four circles with radii (r1 to r4) and center positions (321.1...M). Moreover, the parametric description of this example includes at least one raster unit cell, such as raster unit cells 345.1 and 345.2 shown in FIG. 16B. By raster unit cell, the distance of regularity of the HAR channel cross-section within a cross-sectional image slice can be described. The unit cell may be included in a parametric technology similar to other parameterized technologies, with the unit cell distance (D) as a parameter. During training according to the first embodiment, training images are annotated with different parameter values for the unit cell distance D, such that the annotation tuple of the training data further includes a plurality of different parameter values for the parameter D. do. Figure 17 shows an example of a cross-sectional image slice with a deviation from the ideal raster configuration 345 of the HAR channel cross-section. As shown in Figure 17B, unit cell 345.3 in this example has a larger parameter value for the unit cell parameter (D). During the method according to the second embodiment, the parameter value D is included in a tuple of measurement results extracted from the new cross-sectional image. Therefore, the method according to the invention using a parameterized description is not limited to a parameterized description of the shape or form of the semiconductor objects of interest, but also includes relationships between the semiconductor objects of interest, for example distances or regularities. It can also be applied to abstract parameterized technologies. Thereby, computational effort in numerical analysis of measurement results can be reduced.

The present invention provides a device and method for 3D inspection of an inspection volume within a wafer and for measuring parameter values of semiconductor objects within the inspection volume with high throughput, high accuracy, and reduced wafer damage. The method and device can be used for quantitative metrology, but can also be used for defect detection, process monitoring, defect review, and inspection of integrated circuits within semiconductor wafers. The present invention is based on the concept that measurement results are obtained by the application of a single appropriately trained machine learning algorithm, instead of applying a generally two-stage or more approach. The prior art two-stage or higher approach is to first apply segmentation by conventional machine learning, and then - based on the output of the segmentation - calculate geometric properties, such as parameter values, by classical methods of image processing and metrology. . By the present invention, the second step is avoided at the expense of generating more sophisticated training data. Accordingly, the present invention provides a method and device for generating training data with reduced user interaction. Methods and devices for generating training data rely on prior knowledge of the object to be measured and training can be achieved, for example, without repeating the pixel-wise segmentation of the prior art. According to the present invention, geometric properties can be obtained directly from input images in a single-step approach without classical methods of image processing and metrology involving empirically motivated calculations. Instead, embodiments of the present invention allow measurement of parameter values of a semiconductor object of interest in a single step process.

According to the present invention, a system or device and method are provided for measurement of parameters of semiconductor wafers with increased throughput. The system and method provide higher flexibility and robustness during measurement operations and require less effort during implementation compared to classical methods of the prior art. Systems and methods for measurement of parameters of semiconductor wafers can be easily adapted or modified and can be accomplished with minimal user interaction. One feature of the systems and methods is the application of prior knowledge given, for example, by CAD information or by given image characteristics of a charged particle beam imaging system. The systems and methods rely on a parametric description of the semiconductor object of interest. Thereby, user interaction is reduced to a minimum and the parameter values are obtained directly by the modified machine learning algorithm according to the invention. Although strictly classical methods for metrology are rather utilized, the application of machine learning for measurement tasks can further benefit from future improvements in the rapidly growing libraries and tools for machine learning. Systems and methods for measurement of parameters of semiconductor wafers according to the present invention may further benefit from further improvements in machine learning in the classical domain of machine learning.

The invention, illustrated by examples, can be explained by the following clauses:

Clause 1: A method of generating training data for training a machine learning algorithm to provide quantitative measurement results of parameters of a semiconductor object of interest from a cross-sectional image generated by a charged particle beam system,

- generating a set of training cross-section image segments of a semiconductor object of interest, wherein the training cross-section image segments comprise variations in parameter values of the semiconductor object of interest,

- The variation of the parameter value is within the selected parameter value range.

Clause 2: The method of clause 1, further comprising selecting a parameterized description of the semiconductor object of interest, wherein the parameterized description includes parameters.

Clause 3: The method of clause 2, wherein the parameters of the parameterized description of the semiconductor object of interest are length, diameter, distance, area, angle, radius, ellipticity, aspect ratio, curvature, periodicity, and polygon parameter.

Clause 4: The method of clause 2 or 3, wherein selecting a parameterized technology includes presenting a plurality of parameterized technologies for selection through a user interface and configuring the selected parameterized technology through user input. How to.

Clause 5: The method of any one of clauses 1-4, further comprising receiving imaging parameters of the charged particle beam system.

Clause 6: The method of clause 5, wherein the imaging parameters include at least one of resolution, contrast, and noise level.

Clause 7: The method of clause 6, wherein the imaging parameters further comprise at least one of point spread function, residence time, contrast method, material contrast, and topography contrast.

Clause 8: The method of any one of clauses 1-7, further comprising receiving a selected parameter value range via user input.

Clause 9: The method of any one of clauses 1-8, wherein generating the set of training cross-sectional image segments further comprises annotating the plurality of cross-sectional images containing the semiconductor object of interest with at least one annotation value. and each annotation value represents a measurement result of a parameter in each detected instance of a semiconductor object of interest.

Clause 10: The method of any one of clauses 1-9, wherein generating the set of training cross-sectional image segments comprises receiving at least a first set of training cross-sectional image segments of the semiconductor object of interest from the charged particle beam system. and wherein the first set of training cross-sectional image segments covers a first parameter value range.

Clause 11: The method of clause 10, further comprising automatically detecting instances of semiconductor objects of interest within the first set of training cross-sectional image segments, based on the selected parameterized technique.

Clause 12: The method of clause 11, further comprising automatically determining an initial annotation value for each detected instance of a semiconductor object of interest within the first set of training cross-sectional image segments.

Clause 13: The method of clause 12, wherein automatically determining initial annotation values comprises applying a physical simulation model to a parametric description of the semiconductor object of interest and determining parameter values of the parametric description by optimization.

Clause 14: The method of clause 12 or 13, wherein in at least one detected instance of a semiconductor object of interest within the first set of training cross-sectional image segments, graphically presenting the parameterized description with initial annotation values via a user interface. A method further comprising the steps of:

Clause 15: The method of clause 14, further comprising receiving confirmation or improvement of the initial annotation value via user input.

Clause 16: The method of any of clauses 10-15, wherein the method comprises comparing the first parameter value range to a selected parameter value range and, depending on the results of the comparison, determining whether additional training cross-sectional image segments are required. How to.

Clause 17: The method of clause 16, wherein generating the set of training cross-section image segments comprises: generating a second set of training cross-section image segments of the semiconductor object of interest within a second parameter range from the first set of training cross-section image segments by image processing. A method comprising generating a set.

Clause 18: The method of clause 17, wherein the image processing includes at least one of scale variation, shape change, interpolation, morphological operation, and pattern replacement.

Clause 19: The method of clause 16, wherein generating the set of training cross-sectional image segments comprises a second physical simulation of the cross-sectional image segments based on the selected parameterized description of the semiconductor object of interest and the imaging parameters of the charged particle beam system. A method comprising generating a second set of training cross-sectional image segments of a semiconductor object of interest within a parameter range.

Clause 20: The method of any one of clauses 1 to 19, wherein generating the set of training cross-sectional image segments comprises a physical simulation of the cross-sectional image segments based on a priori information of the semiconductor object of interest and imaging parameters of the charged particle beam system. Including, method.

Clause 21: The method of clause 20, wherein the physical simulation comprises receiving CAD data of a semiconductor object of interest and selecting a parameterized description of the semiconductor object of interest according to the CAD data; and

varying the CAD data according to a selected parameter value range of the parameterized technology; and

A method comprising performing a physical simulation of imaging with a charged particle beam system to acquire a set of training cross-sectional image segments.

Clause 22: The method of any of clauses 1-21, wherein the training data is configured for training a machine learning algorithm to provide continuous quantitative measurements of parameters of the semiconductor object of interest.

Clause 23: A method of performing measurements of semiconductor objects within a wafer,

- acquiring a digital 2D cross-sectional image slice comprising at least one cross-section of the semiconductor object of interest,

- determining at least one quantitative measurement result of at least one predefined parameter of the semiconductor object of interest by one continuous machine learning algorithm applied directly to the digital 2D cross-sectional image slice.

Clause 24: The method of clause 23, wherein the at least one predefined parameter is a parameter of a parameterized geometric description of the semiconductor object of interest.

Clause 25: The method of clause 24, wherein the at least one predefined parameter is one of dimension, length, diameter, distance, area, angle, radius, ellipticity, aspect ratio, curvature, periodicity, and polygon.

Clause 26: The method of any of clauses 23-25, wherein determining a quantitative measurement result comprises determining a continuous quantitative measurement of at least one predefined parameter.

Clause 27: The method of clause 26, wherein determining a quantitative measurement result further comprises measuring a quantity or instance of a semiconductor object of interest.

Clause 28: The method of clauses 26 or 27, wherein determining quantitative measurement results further comprises determining successive quantitative measurement results of predefined parameters for each of the quantities or instances of the semiconductor object of interest.

Clause 29: The method of any of clauses 23-28, wherein the 2D cross-sectional image slice is acquired by a charged particle beam system comprising at least one charged particle beam column.

Clause 30: The method of any one of clauses 23-29, wherein the 2D cross-sectional image slices are obtained by a slice and image method comprising repeatedly milling and imaging a plurality of 2D cross-sectional image slices through an inspection volume within the wafer. One of the 2D cross-sectional image slices of the method.

Clause 31: The method of any of clauses 23-30, further comprising training a machine learning algorithm with training data.

Clause 32: The method of any of clauses 23-31, further comprising generating training data according to the method steps of any of clauses 1-22.

Clause 33: The method of any of clauses 23-32, wherein the semiconductor object of interest is at least one HAR channel in a wafer and the parameterized description comprises a set of rings or circles.

Clause 34: A system for performing automated measurements of semiconductor objects within a wafer,

- a charged particle beam system comprising at least one charged particle beam column, configured to acquire at least one 2D cross-sectional image slice;

- a control unit in communication with the charged particle beam system, wherein the control unit:

- a user interface configured for displaying information and receiving user input;

- further comprising a processing engine configured to determine at least one quantitative measurement result of at least one predefined parameter of the semiconductor object of interest by one continuous machine learning algorithm applied directly to the digital 2D cross-sectional image slice during use. A system that does.

Clause 35: The system of clause 34, wherein the processing engine is further configured to train one continuous machine learning algorithm with training data comprising a set of annotated training cross-sectional image segments of semiconductor objects of interest.

Clause 36: The system of clauses 34 or 35, wherein the processing engine is further configured to generate training data according to any one of the method steps of any of clauses 1-22.

Clause 37: The system of any of clauses 34-36, wherein the processing engine is configured to perform a measurement of a semiconductor object of interest according to any one of the method steps of any of clauses 23-32.

Clause 38: The method of any one of clauses 34 to 37, wherein the charged particle beam system is a dual beam comprising a focused ion beam system and a charged particle imaging system configured to acquire a plurality of 2D cross-sectional image slices by a slice and image method. A system, wherein the control unit is configured to perform a slice and image method by repeatedly milling and imaging a plurality of 2D cross-sectional image slices through an inspection volume within the wafer.

Clause 39: The system of any of clauses 34-38, further comprising a wafer table for receiving and holding a wafer, the wafer table comprising actuators and sensors configured to position and move the wafer table.

Clause 40: The method of any of clauses 1-22, wherein the semiconductor object of interest is a HAR channel and the parameterized description comprises a set of rings or circles.

Clause 41: A method of generating training data for training a machine learning algorithm to provide quantitative measurement results of parameters of a plurality of repetitive HAR channels from cross-sectional images generated by a charged particle beam system,

- selecting a parameterized description comprising a set of circles or rings, the parameterized description comprising at least a radius of the circle or ring as a parameter with a parameter value,

- generating a set of training cross-section image segments of the HAR channel, wherein the training cross-section image segments contain variations of parameter values of the parameterized technique, each parameter value representing a measurement result of the parameter of an instance of the HAR channel cross-section Including expressing and generating steps,

- The variation of the parameter value is within the selected parameter value range.

Item 42: The method of item 41, further comprising any one of method steps 2-22.

Clause 43: A method of generating first training data for training a machine learning algorithm to provide measurement results of parameters of a semiconductor object of interest with a charged particle beam system,

- Receiving CAD data of a semiconductor object of interest; and

- receiving imaging parameters of the charged particle beam system; and

- generating first training data through a physical simulation of imaging with a charged particle beam system based on CAD data of the semiconductor object of interest and imaging parameters of the charged particle beam system.

Clause 44: The method of clause 43, further comprising selecting a parameterized description of the CAD data of the semiconductor object of interest, wherein the parameterized description includes parameters.

Clause 45: The method of clause 43, wherein the parameters of the parameterized description of CAD data are members selected from the group consisting of length, diameter, distance, area, angle, radius, ellipticity, aspect ratio, curvature, periodicity, and polygon parameter. .

Clause 46: The method of clauses 44 or 45, wherein generating the first training data includes varying parameter values of the CAD data according to a selected parameter value range of the parameter.

Clause 47: The method of clause 46, wherein generating the first training data further comprises annotating each training data with parameter values.

Clause 48: The method of any of clauses 43-47, wherein the imaging parameters comprise at least one member of the group consisting of resolution, contrast and noise level achieved by a charged particle beam system.

Clause 49: The method of clause 48, wherein the imaging parameters of the charged particle beam system further comprise at least a member of the group consisting of point spread function, residence time, contrast method, material contrast, and topography contrast of the semiconductor object of interest.

Clause 50: The method of any of clauses 43-49, comprising generating second training data by image processing the first training data.

Clause 51: The method of clause 50, wherein the image processing includes at least a member of the group consisting of variation of scale, variation of shape, interpolation, morphological operations, pattern substitution, addition of noise, and addition of particle defects.

Clause 52: A method of annotating training data for training a machine learning algorithm to provide measurement results of parameters of a semiconductor object of interest,

- receiving a set of training cross-sectional image segments of a semiconductor object of interest, wherein the training cross-sectional image segments comprise variations in parameter values of the semiconductor object of interest;

- selecting a parameterized description of the semiconductor object of interest, the parameterized description comprising parameters;

- adjusting the parameter values of the parameters of the parameterized technique for each detected instance of a semiconductor object of interest in the set of training cross-sectional image segments;

- annotating each detected instance of a semiconductor object of interest within the set of training cross-sectional image segments with adjusted parameter values.

Clause 53: The method of clause 52, wherein the parameters of the parameterized description of the semiconductor object of interest include members of the group consisting of length, diameter, distance, area, angle, radius, ellipticity, aspect ratio, curvature, periodicity, and polygon parameters. How to.

Clause 54: The method of clause 52 or 53, wherein selecting a parameterized technique comprises presenting a plurality of parameterized techniques for selection via a user interface and configuring the selected parameterized technique via user input. How to.

Clause 55: The method of any of clauses 52-54, further comprising automatically detecting instances of semiconductor objects of interest within the set of training cross-sectional image segments based on the selected parameterized technique.

Clause 56: The method of any of clauses 52-55, further comprising automatically adjusting parameter values for each detected instance of a semiconductor object of interest by image processing.

Clause 57: The method of any of clauses 52-55, comprising applying a physical simulation model to a description of the parameters of the semiconductor object of interest with parameter values and adjusting the parameter values by physically influenced forward optimization. , method.

Clause 58: The physical simulation model of clause 57, wherein the physical simulation model comprises imaging parameters of the charged particle device, wherein the imaging parameters consist of resolution, contrast and noise level, point spread function, residence time, contrast method, material contrast and topography contrast. A method containing at least a member of a group.

Clause 59: The method of any one of clauses 52-58, comprising graphically presenting the parameterized description with adjusted parameter values in at least one detected instance of a semiconductor object of interest via a user interface. method.

Clause 60: The method of clause 59, comprising receiving confirmation or improvement of parameter values in at least one detected instance of a semiconductor object of interest via user input.

Clause 61: The method of any of clauses 52-60, wherein selecting a parameterized technique is based on CAD information of the semiconductor object of interest.

Clause 62: A method for generating training data for a machine learning algorithm for image segmentation of a cross-sectional image of a semiconductor object of interest,

- receiving a set of training cross-sectional images of a semiconductor object of interest, wherein the training cross-sectional images contain variations in parameter values of the semiconductor object of interest;

- selecting a parameterized description of the semiconductor object of interest;

- adjusting the parameter values of the parameters of the parameterized technique for each detected instance of a semiconductor object of interest in the set of training cross-sectional images;

- Automatically annotating each image pixel with a pixel annotation value according to adjusted parameter values of the parameters of the parameterized technique.

Clause 63: The method of clause 62, wherein the parameters of the parameterized description of the semiconductor object of interest include members of the group consisting of length, diameter, distance, area, angle, radius, ellipticity, aspect ratio, curvature, periodicity, and polygon parameters. How to.

Clause 64: The method of clause 62 or 63, wherein selecting a parameterized technique comprises presenting a plurality of parameterized techniques for selection via a user interface and configuring the selected parameterized technique via user input. How to.

Clause 65: The method of any of clauses 62-64, further comprising automatically detecting instances of semiconductor objects of interest within the set of training cross-sectional images based on the selected parameterized technique.

Clause 66: The method of any of clauses 62-65, further comprising automatically adjusting parameter values for each detected instance of a semiconductor object of interest by image processing.

Clause 67: The method of any one of clauses 62-65, comprising applying a physical simulation model to a description of the parameters of the semiconductor object of interest with parameter values and adjusting the parameter values by physically influenced forward optimization. , method.

Clause 68: The physical simulation model of clause 67, wherein the physical simulation model comprises imaging parameters of the charged particle device, the imaging parameters consisting of resolution, contrast and noise level, point spread function, residence time, contrast method, material contrast and topography contrast. A method containing at least a member of a group.

Clause 69: The method of any one of clauses 62-68, comprising graphically presenting the parameterized description with adjusted parameter values in at least one detected instance of a semiconductor object of interest via a user interface. method.

Clause 70: The method of clause 69, comprising receiving confirmation or improvement of parameter values in at least one detected instance of a semiconductor object of interest via user input.

Clause 71: The method of any of clauses 62-70, wherein selecting a parameterized technique is based on CAD information of the semiconductor object of interest.

However, the present invention described by examples and embodiments is not limited to the claims and can be implemented by those skilled in the art by various combinations or modifications.

A list of reference numbers is provided.

1: Dual beam system
2: Motion control unit
4: First cross-sectional image feature
6: Measurement area
8: wafer
15: wafer support table
16: Stage control unit
17: Secondary electron detector
19: Control unit
40: Charged particle beam (CPB) imaging system
42: Optical axis of the imaging system
43: intersection
44: Imaging charged particle beam
48: Fib optical axis
50: FIB column
51: Focused ion beam
52: cross-sectional surface
53: cross-sectional surface
55: wafer top surface
155: wafer stage
160: Inspection volume
201: Processing Engine
203: memory
205: User interface
219: memory
231: Interface unit
307: Measured cross-sectional image of HAR structure
311: Cross-sectional image slice
313: word line
315: edge with surface
317: Ring section of HAR structure
319: Initial parameter description
321: Center position
323: Modified parameter description
325: Defect or deviation
327: Ring annotated per pixel
329: Final parameter description
345: Raster
363: Average HAR channel trajectory
400: User interface display
401: User command device
403: Image display area
405: Toolbar
407: Selection box
409: Graphical example of a parameterized technique
411: List of annotation parameter values
413: List of parameter value ranges
414: Histogram
421: Graphic pointer
1000: System for performing measurements of semiconductor objects

Claims (28)

A method of generating training data for training a machine learning algorithm to provide quantitative measurement results of parameters of a semiconductor object of interest from cross-sectional images generated by a charged particle beam system,
- selecting a parameterized description of the semiconductor object of interest, the parameterized description comprising parameters; and
- generating a set of training cross-section image segments of a semiconductor object of interest, wherein the training cross-section image segments comprise variations in parameter values of the semiconductor object of interest,
- The variation of the parameter value is within the selected parameter value range. The method of claim 1 further comprising receiving a selected parameter value range via user input. 3. The method of claim 1 or 2, wherein the parameters of the parameterized description of the semiconductor object of interest are selected from the group consisting of dimensions, length, diameter, distance, area, angle, radius, ellipticity, aspect ratio, curvature, periodicity, and polygonal parameters. A method containing selected members. 4. The method of any one of claims 1 to 3, wherein selecting a parameterized technique comprises:
using a user interface to present a plurality of parameterized techniques for selection; and
A method comprising using user input to configure the selected parameterized technique. 5. The method of any one of claims 1 to 4, further comprising receiving imaging parameters of the charged particle beam system, wherein the imaging parameters include resolution, contrast, noise level, point spread function, residence time, contrast method, A method comprising at least one member selected from the group consisting of material contrast and topography contrast. According to any one of claims 1 to 5,
Generating a set of training cross-sectional image segments includes annotating a plurality of cross-sectional images containing a semiconductor object of interest with at least one annotation value,
The annotation value represents a measurement result of a parameter value of an instance of a semiconductor object of interest. 7. The method of claim 6, further comprising automatically detecting instances of semiconductor objects of interest within the set of training cross-sectional image segments, based on the selected parameterized technique. 8. The method of claim 7, further comprising automatically determining an initial annotation value for each detected instance of a semiconductor object of interest within the first set of training cross-sectional image segments. 9. The method of claim 8, wherein automatically determining initial annotation values comprises applying a physical simulation model to a parametric description of a semiconductor object of interest and determining parameter values of the parametric description by optimization. According to clause 8,
- Graphically presenting the parameterized description with initial annotation values via a user interface at at least one detected instance of a semiconductor object of interest within the set of training cross-sectional image segments; and
- The method further comprising receiving confirmation or improvement of the initial annotation value through user input. 11. The method of any preceding claim, wherein generating a set of training cross-sectional image segments comprises:
- receiving from the charged particle beam system at least a first set of training cross-sectional image segments of a semiconductor object of interest, wherein the first set of training cross-sectional image segments cover a first parameter value range, and
- using image processing to generate a second set of training cross-section image segments of a semiconductor object of interest within a second parameter range from the first set of training cross-section image segments;
The method of claim 1, wherein image processing includes at least a member selected from the group consisting of scale variation, shape variation, interpolation, morphological operations, and pattern substitution. 12. The method of any one of claims 1 to 11, wherein generating the set of training cross-sectional image segments comprises physically simulating the cross-sectional image segments based on a priori information of the semiconductor object of interest and imaging parameters of the charged particle beam system. A method comprising the steps of: 13. The method of claim 12, wherein the physical simulation:
Receiving CAD data of a semiconductor object of interest;
selecting a parameterized description of the semiconductor object of interest according to the CAD data;
varying the CAD data according to a selected parameter value range of the parameterized technology; and
A method comprising physically simulating imaging with a charged particle beam system to acquire a set of training cross-sectional image segments. The method of claim 1 , wherein the training data is configured to train a machine learning algorithm to provide continuous quantitative measurements of parameters of a semiconductor object of interest. 15. One or more machine-readable hardware storage devices comprising instructions executable by one or more processing devices to perform operations comprising the method of any one of claims 1-14. It is a system,
one or more processing devices; and
15. A system comprising one or more machine-readable hardware storage devices containing instructions executable by one or more processing devices to perform operations comprising the method of any one of claims 1-14. A method of performing measurements of semiconductor objects within a wafer,
- acquiring at least one digital 2D cross-sectional image slice comprising at least one cross-section of the semiconductor object of interest,
- determining at least one quantitative measurement result of at least one predefined parameter of the semiconductor object of interest by a continuous machine learning algorithm applied directly to the digital 2D cross-sectional image slice,
The method of claim 1, wherein the at least one predefined parameter comprises a parameter of a parameterized geometric description of the semiconductor object of interest. 18. The method of claim 17, wherein the at least one predefined parameter of the parameterized geometric description is a member selected from the group consisting of length, diameter, distance, area, angle, radius, ellipticity, aspect ratio, curvature, periodicity, and polygonal parameters. Including, method. 19. The method of claim 17 or 18, wherein determining a quantitative measurement result comprises determining a continuous quantitative measurement of at least one predefined parameter. 20. The method of any one of claims 17-19, wherein determining a quantitative measurement result further comprises measuring a quantity of instances of a semiconductor object of interest. 21. The method of claim 20, wherein determining quantitative measurement results further comprises determining successive quantitative measurement results of predefined parameters for each quantity of instances of the semiconductor object of interest. 22. The method of any one of claims 17 to 21, wherein the at least one 2D cross-sectional image slice is acquired by a charged particle beam system comprising at least one charged particle beam column, and the 2D cross-sectional image slice is located in an inspection volume within the wafer. A method, one of a plurality of 2D cross-sectional image slices acquired by a slice and image method comprising repeatedly milling and imaging a plurality of 2D cross-sectional image slices. 23. The method of any one of claims 17 to 22, further comprising generating training data according to the method steps of any of claims 1 to 14. One or more machine-readable hardware storage devices comprising instructions executable by one or more processing devices to perform operations comprising the method of any one of claims 17-23. It is a system,
- a charged particle beam system comprising at least one charged particle beam column, configured to acquire at least one 2D cross-sectional image slice;
- one or more processing devices; and
- A system comprising one or more machine-readable hardware storage devices containing instructions executable by one or more processing devices to perform operations comprising the method of any one of claims 17 to 23. It is a system for performing automated measurement of semiconductor objects within a wafer,
- a charged particle beam system comprising at least one charged particle beam column, configured to acquire at least one 2D cross-sectional image slice;
- a control unit in communication with the charged particle beam system, wherein the control unit:
- a user interface configured for displaying information and receiving user input;
- a processing engine configured to determine, in use, at least one quantitative measurement result of at least one predefined parameter of the semiconductor object of interest by one continuous machine learning algorithm applied directly to the digital 2D cross-sectional image slice, at least The system further comprising a processing engine, wherein one predefined parameter is a parameter of a parameterized geometric description of a semiconductor object of interest. 27. The system of claim 26, wherein the processing engine is configured to perform measurements of a semiconductor object of interest according to the method of any one of claims 17-23. 27. The method of claim 26, wherein the charged particle beam system is a dual beam system comprising a focused ion beam system and a charged particle imaging system configured to acquire a plurality of 2D cross-sectional image slices by a slice and image method, and the control unit is configured to A system configured to perform a slice and image method by iteratively milling and imaging a plurality of 2D cross-sectional image slices through an examination volume.

Images