Pattern inspection apparatus

Abstract

Based on the reference waveform (112) and the BSE signal waveform (211) representing the intensity of the backscattered electron signal from the pattern along the first direction extracted from the backscattered electron image, a difference waveform is generated, representing the relationship between the difference between the coordinates of the BSE signal waveform with the same backscattered electron signal intensity and the coordinates of the reference waveform, and the backscattered electron signal intensity. Based on the difference waveform, it is determined whether there is a shielding area (203) on the sidewall of the pattern that was not irradiated by the primary electron beam. Here, the reference waveform represents the intensity of the backscattered electron signal from the reference pattern along the first direction when a primary electron beam is scanned on the reference pattern, wherein the reference pattern is a pattern with sidewalls formed perpendicularly to the surface and bottom surface of the pattern.

Description

Technical Field

[0001] The present invention relates to a pattern inspection apparatus for inspecting the cross-sectional shape of patterns formed on semiconductor wafers or the like. Background Technology

[0002] To date, semiconductor devices have undergone significant advancements in precision and integration due to the increasing capacity and decreasing cost of memory. In recent years, the demand for high integration has intensified, leading to the development and manufacturing of three-dimensional devices as an alternative to precision.

[0003] By making the planar structure three-dimensional, the entire device is in a thickening direction. Therefore, in structures such as 3D-NAND and DRAM, the number of laminated films increases, and in the process of forming holes or trenches, there is a tendency for the ratio of the planar size to the depth (aspect ratio) of the holes or trenches to also increase. While it is possible to obtain accurate cross-sectional shapes of holes or trenches with high aspect ratios by cutting the wafer for measurement, it is time-consuming and costly to obtain the cross-sectional shapes of multiple holes or trenches when investigating the uniformity within the wafer surface during semiconductor manufacturing processes. Patent Document 1 discloses a backscattered electron (reflected electron) image obtained by observing a pattern from top to bottom, comparing the intensity of backscattered electron signals from the top surface, bottom surface, and sidewalls of the pattern, and measuring the cross-sectional shape of the pattern sidewalls based on the height difference between the top and bottom surfaces.

[0004] Patent document 2 discloses a method for generating electron microscope images through electron beam simulation, and recording simulated images of various shapes into a library. Electron beam simulation refers to simulating the process of generating electron microscope images in a scanning electron microscope, and calculating the relationship between the shape of the measured object and the waveform of the electron microscope image.

[0005] Existing technical documents

[0006] Patent documents

[0007] Patent Document 1: Japanese Patent Application Publication No. 2019-87518

[0008] Patent Document 2: Japanese Patent Application Publication No. 2007-218711 Summary of the Invention

[0009] The problem that the invention aims to solve

[0010] Patterns with high aspect ratios are difficult to control in terms of the shape of the sidewalls and bottom, sometimes resulting in conical, bowing, or twisting shapes. For example, as in Patent Document 1, when determining the cross-sectional or three-dimensional shape of the pattern by observing it from top to bottom using a charged particle beam device, depending on the cross-sectional shape, sometimes a partially or completely un-irradiated area of ​​the sidewalls may occur. In this case, the accurate cross-sectional shape cannot be estimated for the un-irradiated area. However, by irradiating the pattern with the charged particle beam at an angle, the un-irradiated area can be eliminated.

[0011] Therefore, in this invention, when inspecting the cross-sectional shape of a pattern by top-down observation, the presence or absence of obstructed areas is determined based on the intensity of the backscattered electron signal. Only when an obstructed area is determined to exist is the beam tilted differently to obtain additional images, thereby improving the throughput of cross-sectional shape inspection of patterns performed by top-down observation.

[0012] Furthermore, in order to perform cross-sectional shape checks on patterns at high speed, a learned neural network model is used to estimate the cross-sectional shape.

[0013] Methods for solving problems

[0014] A pattern inspection apparatus according to one embodiment of the present invention inspects the cross-sectional shape of a pattern formed on a sample. It comprises: an electron optical system that irradiates the sample with a single electron beam under the control of an apparatus control unit; an electron detector that detects backscattered electrons emitted when the electron beam is scanned across the pattern; an image processing unit that forms an image based on the detection signal from the electron detector; and a calculation unit that generates a difference waveform based on a reference waveform and a BSE signal waveform representing the intensity of the backscattered electron signal from the pattern along a first direction extracted from the BSE image formed by the image processing unit based on the detection signal from the electron detector. This difference waveform represents the relationship between the difference between the coordinates of the BSE signal waveforms with the same backscattered electron signal intensity and the coordinates of the reference waveform, and the backscattered electron signal intensity. Based on the difference waveform, it determines whether there is an obstruction area on the sidewall of the pattern that was not irradiated by the single electron beam. Here, the reference waveform represents the intensity of the backscattered electron signal from the reference pattern along the first direction when the electron beam is scanned across the reference pattern. The reference pattern is a pattern in which the sidewalls are formed perpendicularly to the surface and bottom surface of the pattern.

[0015] Invention Effects

[0016] It can perform cross-sectional shape inspections of deep holes, deep grooves, etc. with high precision and high throughput. Attached Figure Description

[0017] Figure 1 This is used to explain the method for determining occluded areas.

[0018] Figure 2 This is used to explain the method for determining occluded areas.

[0019] Figure 3 This is used to explain the method for determining occluded areas.

[0020] Figure 4 This refers to a pattern inspection device.

[0021] Figure 5 This is a flowchart for checking the cross-sectional shape of a pattern.

[0022] Figure 6 This is an example of the hardware architecture of a computing server.

[0023] Figure 7 This is an example of a neural network model.

[0024] Figure 8 This indicates the relationship between the parameters and the neural network model.

[0025] Figure 9A This is a flowchart for updating a neural network model.

[0026] Figure 9B Represents a labeled pattern in the feature space.

[0027] Figure 10A This is another example of a neural network model.

[0028] Figure 10B This is a flowchart for updating a neural network model.

[0029] Figure 11 This indicates a computing server connected to multiple SEMs.

[0030] Figure 12 This is an example of a screen displaying the inspection results.

[0031] Figure 13 This is an example of a screen displaying the inspection results. Detailed Implementation

[0032] The following describes a pattern inspection apparatus particularly suitable for inspecting the cross-sectional shape of hole patterns and trench patterns with high aspect ratios in semiconductor manufacturing processes, and a method for estimating the cross-sectional shape of patterns using this pattern inspection apparatus. As a sample to be inspected, a patterned semiconductor wafer is shown, but it is not limited to semiconductor patterns; any sample that can be observed using a charged particle beam device is acceptable.

[0033] First, use Figures 1-3The method for determining the occlusion area using the pattern inspection device in this embodiment will be described. The upper section represents the cross-sectional shape of the pattern to be inspected along the X direction, the middle section represents the BSE (backscattered electron) signal waveform of the pattern to be inspected along the X direction, and the lower section represents the difference between the BSE signal waveform of the middle section and the reference waveform.

[0034] Figure 1 This illustrates the case where a charged particle beam is perpendicularly irradiated onto the sample surface for a pattern without obstruction. In the upper section, the cross-sectional shape 101 of the pattern under inspection is represented by a solid line, and the reference cross-sectional shape 102, serving as the reference pattern, is represented by a dashed line. The sidewalls of the reference pattern are formed perpendicularly to the surface and bottom of the pattern. The perpendicularly irradiated beam 100 has a predetermined aperture angle, therefore the spot diameter on the bottom surface of the sample is larger than the spot diameter on the surface. Regarding... Figure 1 The pattern of the object being inspected, when viewed from above, has its bottom surface contained within the surface, and its diameter gradually increases from the bottom to the surface. Therefore, the vertically illuminating beam 100 illuminates the entire area of ​​the sidewall without any obstruction. This cross-sectional shape is called a conical shape.

[0035] In the middle section, the BSE signal waveform 111 along the X direction of the pattern with cross-sectional shape 101 is represented by a solid line, and the BSE signal waveform (reference waveform 112) along the X direction of the reference pattern with reference cross-sectional shape 102 is represented by a dashed line. The BSE signal waveform of the reference pattern is called the reference waveform. The vertical axis represents the BSE signal intensity, and the horizontal axis represents the X coordinate. Here, assuming the aperture angle of the vertically irradiated beam 100 is 0, the BSE signal quantity generated by irradiating the aperture pattern or slot pattern with a charged particle beam is expressed as η·e －H (Where H is the depth from the sample surface, and η is a coefficient). When the aperture angle is not 0, although affected by the spot expansion corresponding to the aperture angle, the BSE signal intensity becomes a function of the depth from the sample surface of the position irradiated by the vertically irradiated beam 100.

[0036] The lower section shows the difference between the BSE signal waveform 111 on the right sidewall of the pattern and the reference waveform 112. The vertical axis represents the BSE signal intensity, and the horizontal axis represents the difference between the X-coordinate of the BSE signal waveform 111 with the same BSE signal intensity and the X-coordinate of the reference waveform 112. In the case of a positive cone shape, the difference in the X-coordinate represents a waveform that monotonically increases as the BSE signal intensity decreases, with the difference on the sample surface (BSE signal intensity = Max) set to 0 and the difference on the bottom surface (BSE signal intensity = Min) set to the maximum (d12).

[0037] Figure 2This illustrates the case where a charged particle beam is perpendicularly irradiated onto the sample surface for a pattern with a shielded area. The cross-sectional shape 201 (solid line) of the pattern being inspected, shown in the upper section, represents a bulging, curved shape. Therefore, there exists a shielded area 203 that is not reached by the perpendicularly irradiated beam 100.

[0038] In the middle section, the BSE signal waveform 211 along the X direction of the pattern of cross-sectional shape 201 is represented by a solid line. Thus, in the blocked region 203, since the vertical illumination beam 100 does not illuminate it, the depth information of the sidewall is not included in the BSE signal intensity. Only in the sidewall portion illuminated by the vertical illumination beam 100 does the BSE signal intensity vary according to the depth of the position of the vertical illumination beam 100 from the sample surface.

[0039] Therefore, the difference between the BSE signal waveform 211 and the reference waveform 112 on the right sidewall of the pattern shown in the lower section becomes 0 in the region corresponding to the occlusion region 203. On the other hand, the sidewall outside the occlusion region 203 in the cross-sectional shape 201 becomes a positive cone shape, so in the region corresponding to the region outside the occlusion region 203, the difference in the X coordinate increases monotonically as the BSE signal strength decreases.

[0040] Thus, it exhibits the following characteristic: when there is an obstruction area that the charged particle beam cannot irradiate, there is a region in the difference waveform where the difference in the X-coordinate becomes 0, independent of the BSE signal intensity. Therefore, for example, by determining whether there is a region below a predetermined tilt θth in the tilt of the difference waveform relative to the BSE signal intensity, it is possible to determine whether there is an obstruction area in the pattern being inspected. As a tilt-based determination method, the differential value of the difference waveform can also be calculated, and the presence or absence of an obstruction area can be determined based on the magnitude of the differential value. Furthermore, this determination method is not limited to the illustrated bowing shape, and can generally be used for cross-sectional shapes such as twisting shapes that contain obstruction areas in the sidewalls.

[0041] Figure 3 Indicates that for those having the same Figure 2 The pattern of the same shielding area is used to irradiate the charged particle beam at an angle in the X direction. In this example, by irradiating the angled beam 300, the shielding area 203 that cannot be irradiated by the vertical irradiation beam 100 can be irradiated with the charged particle beam, and the shielding area on the right sidewall disappears.

[0042] In the middle section, the BSE signal waveform 311 along the X direction of the pattern with cross-sectional shape 201 is represented by a solid line, and the BSE signal waveform (reference waveform 312) along the X direction of the reference pattern with reference cross-sectional shape 102 is represented by a dashed line. Since there is no obstruction area on the right sidewall, the region corresponding to the right sidewall in the BSE signal waveform 311 contains the depth information of the sidewall. Therefore, the cross-sectional shape of the right sidewall of the pattern can be estimated from the BSE signal waveform 311. Because the charged particle beam is tilted, the reference waveform 312 is different from the reference waveform 112 when the beam 100 is vertically irradiated.

[0043] The lower section shows the difference between the BSE signal waveform 311 on the right sidewall of the pattern and the reference waveform 312. In this example, the difference in the X-coordinate represents a waveform where the difference on the sample surface is set to 0, initially decreasing monotonically as the BSE signal intensity decreases, and then monotonically increasing as the BSE signal intensity decreases after exceeding the inflection point. Even with the presence of an obstruction region 203, an area where the difference in the X-coordinate is 0 will appear in the difference waveform regardless of the BSE signal intensity when irradiated with the tilted beam 300. In this case, the presence or absence of an obstruction region can also be determined based on the tilt of the difference waveform.

[0044] Figure 4 This describes the pattern inspection apparatus of this embodiment. An example using a scanning electron microscope (SEM) is shown as the pattern inspection apparatus. The scanning electron microscope body consists of an electron optical column 1 and a sample chamber 2. Inside the column 1, as the main structure of the electron optical system, there are: an electron gun 3, a source for releasing a primary electron beam that generates electrons and energizes them with a predetermined accelerating voltage; a focusing lens 4 for focusing the electron beam; a deflector 6 for scanning the primary electron beam on the wafer (sample) 10; and an objective lens 7 for focusing the primary electron beam to irradiate the sample. Additionally, a deflector 5 is provided that deflects the primary electron beam away from the ideal optical axis 3a and deflects the deviated beam in a direction tilted relative to the ideal optical axis 3a, thereby making the primary electron beam a tilted beam. These optical elements constituting the electron optical system are each controlled by an electron optical system control unit 14. The wafer 10, serving as the sample, is placed on the XY stage 11 provided in the sample chamber 2, and the wafer 10 is moved according to control signals given from the stage control unit 15. The device control unit 20 of the control unit 16 scans the electron beam once on the observation area of ​​the wafer 10 by controlling the electron optical system control unit 14 and the stage control unit 15.

[0045] In this embodiment, to inspect the cross-sectional shape of deep holes or trenches with high aspect ratios, a high-energy (high accelerating voltage) primary electron beam is irradiated onto the wafer 10 to reach the deep portion of the pattern. Electrons generated by scanning the primary electron beam on the wafer 10 are detected by a first electron detector 8 and a second electron detector 9. The detection signals output from each detector are converted by amplifiers 12 and 13, respectively, and then input to the image processing unit 17 of the control unit 16.

[0046] The first electron detector 8 primarily detects secondary electrons generated by irradiating the sample with a primary electron beam. Secondary electrons are electrons excited from the atoms constituting the sample through inelastic scattering of primary electrons within the sample, and their energy is below 50 eV. The release rate of secondary electrons is sensitive to the surface shape of the sample; therefore, the detection signal from the first electron detector 8 mainly represents the pattern information of the wafer surface (top surface). On the other hand, the second electron detector 9 detects backscattered electrons generated by irradiating the sample with a primary electron beam. Backscattered electrons (BSE) are electrons that are backscattered during the scattering process of primary electrons irradiating the sample. The BSE reflection electron release rate mainly reflects the material information of the irradiated area of ​​the primary electron beam.

[0047] The control unit 16 has an input unit and a display unit (not shown), which receive instructions from the user regarding the scanning electron microscope and display the examination results. The storage unit 19 stores the processing program executed by the control unit 16. In addition, images output from the image processing unit 17 are also stored in the storage unit 19.

[0048] The computing unit 18 uses images captured by SEM (backscattered electron (BSE) images, secondary electron images) to check the cross-sectional shape of the pattern, details of which will be described later.

[0049] In the pattern inspection apparatus of this embodiment, it is assumed that a learned neural network model is used to quickly estimate the cross-sectional shape of the pattern. Since learning the neural network model requires high computer processing power, a computing server 22 connected to the control unit 16 via network 21 can be provided. The computing server 22 is not limited to the purpose of learning the neural network model. For example, as described later, multiple SEMs can be connected to network 21, and pattern inspection of multiple SEMs can be performed on the computing server 22 instead of the control unit 16. Regarding the task allocation between the computing unit 18 and the computing server 22, it is acceptable as long as the allocation, including network resources, is optimal.

[0050] Figure 5 A flowchart showing the cross-sectional shape of the inspection pattern performed by the control unit 16.

[0051] S11: The device control unit 20 controls the electro-optical system to irradiate the sample 10 with a primary electron beam at a predetermined tilt angle. The image processing unit 17 forms a top-down BSE image of the pattern based on the detection signal from the second electron detector 9. The formed top-down BSE image is stored in the storage unit 19. The calculation unit 18 obtains the BSE signal waveform from the acquired top-down BSE image of the pattern based on the brightness values ​​of the pixels along the X direction of the pattern. The X direction is preset to the cross-sectional direction of the pattern that the user wants to observe. At this time, the tilt angle of the primary electron beam is set to, for example, 0° (perpendicular to the sample surface).

[0052] S12: The arithmetic unit 18 generates the difference waveform between the BSE signal waveform obtained in S11 and the reference waveform. The reference waveform is calculated in advance through electron beam simulation and stored in the storage unit 19. The reference waveform varies depending on the optical conditions of the primary electron beam and the material of the sample. The optical conditions that affect the reference waveform are the accelerating voltage, aperture angle, and tilt angle of the primary electron beam. Therefore, it is preferable to calculate the reference waveform in advance for each combination of these conditions. Alternatively, the reference waveform may not be calculated in advance, but rather based on the shooting conditions each time a BSE image is acquired. In this case, since the calculation of the reference waveform requires more computer resources, it is preferable to send the shooting conditions to the computing server 22 to calculate the reference waveform.

[0053] S13: Determine whether there is an obstruction area based on whether there is a region below a predetermined tilt θth in the tilt of the difference waveform obtained in step S12. The difference waveform represents the relationship between the difference in X-coordinate between the BSE signal waveform with the same BSE signal strength and the reference waveform, and the BSE signal strength. Regions where the tilt of the difference waveform is below the predetermined tilt θth are considered obstruction areas.

[0054] S14: In the absence of an obstruction area (S13: No), the cross-sectional shape is estimated based on the acquired BSE signal waveform. To estimate the cross-sectional shape at high speed, a learned neural network model is used in this embodiment. The computation unit 18 can either pre-store the learned neural network model in the storage unit 19 and retrieve it for estimation, or send the acquired BSE signal waveform to the computing server 22 and use the learned neural network model to estimate the cross-sectional shape. The computing server 22 and the control unit 16 have the same learned neural network model.

[0055] S15: The calculation unit 18 displays the cross-sectional shape estimated in step S14 on the display unit.

[0056] In contrast, when there is an obstruction area (S13: Yes), the cross-sectional shape cannot be determined from the BSE signal waveform. Therefore, the tilt angle of the electron beam is changed once so that the obstruction area is irradiated with an electron beam once.

[0057] S21: Determine the imaging conditions for irradiating the obstructed area with an electron beam once. For example, if it is Figure 3 In the example, the obstruction area on the right side wall disappears due to the tilted beam 300, while the obstruction area remains on the left side wall. To eliminate the obstruction area on the left side wall, a top-down BSE image needs to be obtained using a primary electron beam with a different tilt angle than the tilted beam 300 (hereinafter, when we want to specifically indicate the image obtained by a primary electron beam with a positive or negative tilt angle, it is referred to as a tilted BSE image). Thus, in this step, the calculation unit 18 determines one or more shooting conditions (tilt angle) such that the obstruction area on the pattern side wall disappears.

[0058] Based on the top-down BSE image, the relative depth position of the obstructed area within the cross-sectional shape of the pattern is determined. Therefore, knowing the size of the pattern opening and the absolute depth of the bottom surface (distance from the sample surface to the bottom surface) allows for the calculation of the tilt angle to eliminate the obstructed area. The size of the pattern opening is determined from a top-down secondary electron image of the pattern formed based on the detection signal from the first electron detector 8. Furthermore, the absolute depth of the bottom surface can be determined based on the magnitude of the bottom surface offset in the top-down BSE image taken with a changed tilt angle. Alternatively, instead of calculating the desired tilt angle separately, a predetermined tilt angle can be pre-determined as the tilt angle when the obstructed area exists.

[0059] S22: The arithmetic unit 18 confirms whether the tilt angle calculated in step S21 is within the range for tilting the scanning electron microscope. If it is within the allowable range, the processing after step S11 is performed according to the imaging conditions (tilt angle) determined in step S21.

[0060] S23: In step S22, if the tilt angle calculated in step S21 exceeds the range allowed by the scanning electron microscope, the obstruction area cannot be eliminated. In this case, the calculation unit 18 displays on the display unit that there is an obstruction area that cannot be photographed. In this case, the cross-sectional shape cannot be determined.

[0061] When multiple patterns are present on a sample and are to be inspected, the process of estimating the cross-sectional shape based on the BSE image is separated from the subsequent process, and the images are taken at the same tilt angle (including 0°), thereby improving processing efficiency.

[0062] In this embodiment, a learned neural network is used to rapidly estimate the cross-sectional shape based on a top-down BSE image. In semiconductor manufacturing process management, it is not always necessary to measure the exact cross-sectional shape; as long as the occurrence of abnormal cross-sectional shapes can be determined quickly and accurately based on the top-down BSE image, the throughput of inspections can be improved.

[0063] Figure 6 This illustrates an example of the hardware structure of a computing server 22. The computing server 22 includes a processor 601, a GPU (Graphics Processing Unit) 602, a memory 603, a storage device 604, a network interface 605, and an input / output interface 606, which are connected via a bus 607. The input / output interface 606 connects to input devices such as a keyboard and operation buttons (not shown), and display devices such as a monitor, to implement a GUI (Graphical User Interface). The network interface 605 is used for connecting to a network 21.

[0064] Storage device 604 is typically composed of non-volatile memory such as HDD (Hard Disk Drive), SSD (Solid State Drive), ROM (Read Only Memory), or flash memory, storing programs executed by the computing server 22, image data that is the object of processing by the program, etc. Memory 603 is composed of RAM (Random Access Memory), temporarily storing programs and data required for program execution according to commands from processor 601. Processor 601 executes the program loaded from storage device 604 into memory 603. In the figure, the functions of the processor are shown as an inference unit 601a that performs inference through a neural network model and a learning unit 601b that learns the neural network model; these functions are implemented by executing the program stored in storage device 604. Since it is assumed that the computing server 22 performs neural network model learning, a GPU 602 is also included in addition to the processor 601 (Central Processing Unit). This is because calculations need to be performed repeatedly during the learning process.

[0065] Figure 7 Here is an example of a neural network model. Neural network model 703 outputs the following probability data (p1, p2, ..., p...). M The model, wherein the probability data represents the BSE signal waveform data (s1, s2, ..., s) as input data 701. NThis corresponds to the probability of each label among M pre-determined labels (patterns 1 to M with different cross-sectional shapes). Patterns 1 to M are defined as patterns with normal or abnormal cross-sectional shapes, set by the user. The neural network model 703, for example, contains K hidden layers 704. Input data 701 is fed into the first hidden layer 704-1, and the output of the final hidden layer 704-K is weighted in the fully connected layer 705, ultimately averaging the probability data (p1, p2, ..., p...). M This is output as data 702. In this example, in the above... Figure 5 In step S14 of the flowchart, the arithmetic unit 18 (or the computation server 22) estimates the cross-sectional shape of the labeled pattern with the highest probability output by the learned neural network model as the cross-sectional shape of the obtained BSE signal waveform pattern. The BSE signal waveform data as input data 701 may be, for example, data representing the relative signal intensity when the signal intensity at the bottom surface is set to 0 and the signal intensity at the sample surface is set to 1.

[0066] The neural network model used in this embodiment will be explained. As described with respect to the reference waveform, even with the same pattern, the BSE signal waveform can vary depending on the optical and sample conditions. The optical conditions affecting the BSE signal waveform are the accelerating voltage of the primary electron beam, the aperture angle, and the tilt angle; the sample conditions affecting the BSE signal waveform are the sample material. Therefore, a neural network model is learned for each combination of these three optical conditions and sample material, and inference is performed using a conditionally consistent neural network model. Figure 8 Examples of the relationship between these parameters and neural network models are shown. When the accelerating voltage is of type a (α1 to αa), the aperture angle is of type b (β1 to βb), the tilt angle is of type c (γ1 to γc), and the sample material is of type d (δ1 to δd) (a, b, c, and d are all 1 or more), J (=a×b×c×d) neural network models corresponding to each combination are learned, and the cross-sectional shape is inferred using the corresponding neural network models. The J neural network models are stored in the storage device 604 (storage unit 19). Regarding the materials used as sample conditions, as long as the BSE signal generated when irradiated with a primary electron beam with equal optical conditions is approximately the same, even if the materials have different chemical compositions, they can be treated as the same material.

[0067] The learning data used in the neural network model is a combination of BSE signal waveform data, which serves as the input to the neural network model, and labeled patterns (pattern i (i = 1 to M) of the cross-sectional shape shown by the BSE signal waveform. As the BSE signal waveform data, multiple derived cross-sectional shapes approximating the cross-sectional shape of pattern i can be generated. The BSE signal waveform data for these derived cross-sectional shapes is obtained through electron beam simulation and used as learning data. Therefore, the learning of the neural network model can be effectively advanced even in the early stages when there is limited measured data. On the other hand, through pattern inspection in this embodiment, measured top-down BSE images are collected. Therefore, it is preferable to measure the pattern cross-sectional shape based on the top-down BSE images and update the neural network model based on the pattern cross-sectional shape. Figure 9A A flowchart illustrating the update of the neural network model. Storage device 604 stores the neural network model update program, and processor 601 executes the update program (processor 601 executing the update program is equivalent to learning unit 601b).

[0068] S31: The processor 601 acquires the top-down BSE image stored in the storage unit 19 of the control unit 16. At this time, it is set to simultaneously acquire the data required for calculating the cross-sectional shape. Specifically, this includes a top-down secondary electron image of the same pattern and absolute depth information. Alternatively, it may not be the absolute depth information itself, but a top-down BSE image acquired by changing the tilt angle. The top-down BSE image acquired at this time is determined to be an image where there is no occlusion area ("No" in step S13).

[0069] S32: The processor 601 calculates the cross-sectional shape about a predetermined direction based on the top-down BSE image. In calculating the cross-sectional shape, for example, the pattern measurement method disclosed in Patent Document 1 can be used.

[0070] S33: Processor 601 determines whether the cross-sectional shape calculated in step S32 corresponds to one of the existing labeled patterns. For example... Figure 9B As shown, labeled patterns i (i = 1 to M) are positioned in an m-dimensional feature space. Patterns contained within a predetermined distance centered on a typical pattern are determined to be equivalent to the labeled pattern. For example, in the feature space, patterns contained within a predetermined distance relative to pattern i (901, 902) (referred to as approximate regions 910, 920) are determined to be equivalent to pattern i. Furthermore, a pattern being approximate to pattern i means that the pattern is contained within the approximate region of pattern i.

[0071] That is, for the cross-sectional shape, the characteristic quantity is calculated. If the position of the calculated cross-sectional shape in the characteristic quantity space is contained in the approximate region of pattern i, it is determined that the pattern of the cross-sectional shape is equivalent to pattern i. If it is not contained in any approximate region, it is determined that the pattern of the cross-sectional shape is not equivalent to any of the existing marking patterns.

[0072] S34: When the processor 601 determines that the calculated cross-sectional shape is equivalent to pattern i, it sets the combination of the BSE signal waveform data and pattern i (label) as new learning data.

[0073] On the other hand, if the calculated cross-sectional shape is determined not to be equivalent to any of the existing labeled patterns, it can be excluded from the learning data, and a new labeled pattern can be defined. The definition of a new labeled pattern will be explained here.

[0074] S41: Based on the calculated cross-sectional shape, define a new labeled pattern (M+1) 903. The cross-sectional shape of the pattern (M+1) can be the calculated cross-sectional shape itself, or it can be set by the user based on the calculated cross-sectional shape.

[0075] S42: Processor 601 generates multiple derived cross-sectional shapes that approximate the cross-sectional shape of pattern (M+1). The derived cross-sectional shape refers to... Figure 9B The approximate region 930 of the pattern (M+1) shown contains the cross-sectional shape of pattern 904. In order to learn using derived cross-sectional shapes, it is necessary to generate the required number of derived cross-sectional shapes for learning.

[0076] S43: Processor 601 performs electron beam simulation on each derived cross-sectional shape to obtain BSE signal waveform data.

[0077] S44: The combination of the BSE signal waveform data obtained in step S43 and the pattern (M+1) (label) is used as new learning data. The number of new learning data is the number of derived cross-sectional shapes generated in step S42.

[0078] S35: The processor 601 uses the new learning data generated in step S34 or step S44 to perform training of the neural network model.

[0079] S36: The processor 601 stores the updated neural network model in the storage device 604 via step S35.

[0080] Here, by pre-defining the conditions for defining new labeled patterns (e.g., the case where a certain number of cross-sectional shapes not included in existing labeled patterns are generated at relatively close positions in the feature space), extended learning can be automatically implemented.

[0081] The neural network model used in this embodiment is not limited to the neural network model described above. Figure 10A This represents another example of a neural network model. Neural network model 1003 contains K hidden layers 1004 and fully connected layers 1005, and is similar to... Figure 7 The model shown has the same structure, but the output data used as the estimation object is different. It is the BSE signal waveform data (s1, s2...s1) used as input data 1001. N (Here, the BSE signal waveform data is set to relative values), outputting relative depth data (h1, h2, ..., h) at the same X coordinate. N This is used as the model for output data 1002. The relative depth of each X coordinate represents the depth of the pattern section at that location relative to the depth of the bottom surface of the pattern. In this case, in Figure 5 In step S14 of the flowchart, the arithmetic unit 18 (or the computing server 22) can estimate the accurate cross-sectional shape by multiplying the relative depth of the pattern in the X direction output by the neural network model by the separately calculated absolute depth of the bottom surface.

[0082] exist Figure 10A In the case of a neural network model, it also learns and Figure 8 The J (=a×b×c×d) neural network models corresponding to each combination shown are used to infer the cross-sectional shape.

[0083] for Figure 10A The neural network model is preferably updated based on the measured top-down BSE images collected through pattern inspection in this embodiment. Figure 10B This indicates the process.

[0084] S51: Use a neural network model to estimate the cross-sectional shape. This process is equivalent to... Figure 5 The process of step S14 in the flowchart.

[0085] S52: The processor 601 compares the estimated cross-sectional shape with the cross-sectional shapes in the existing learning data. Specifically, it calculates feature quantities for each cross-sectional shape, based on... Figure 9B The degree of similarity is determined by the distance in the m-dimensional feature space shown.

[0086] When the estimated cross-sectional shape differs significantly from the cross-sectional shape in the existing learning data, the estimation accuracy may decrease. Therefore, for patterns with significantly different cross-sectional shapes from the learned ones, it is preferable to regenerate learning data containing patterns similar to the original pattern and update the neural network model. Thus, a baseline value for the distance in the m-dimensional feature space is predetermined, and it is determined whether the shortest distance to the existing learning data exceeds the baseline value (step S53). If it exceeds the baseline value (step S53: Yes), proceed to step S61; if it is below the baseline value (step S53: No), proceed to step S54.

[0087] S61: This process is equivalent to Figure 9A The flowchart describes the processing steps S31 to S32.

[0088] S62: Processor 601 generates multiple derived cross-sectional shapes that approximate the cross-sectional shape of the pattern calculated in step S61. The derived cross-sectional shapes are... Figure 9B The approximate region of the pattern shown contains the cross-sectional shape of the pattern. To learn using derived cross-sectional shapes, the required number of derived cross-sectional shapes need to be generated.

[0089] S63: This process is equivalent to Figure 9A The process of step S43 in the flowchart.

[0090] S64: The processor 601 uses the combination of the BSE signal waveform data obtained in step S63 and the relative depth data of the derived cross-sectional shape generated in step S62 as new learning data. The number of new learning data is the number of derived cross-sectional shapes generated in step S62.

[0091] On the other hand, if the estimated cross-sectional shape is sufficiently similar to the cross-sectional shape in the existing learning data (no in step S53), the combination of the measured BSE signal waveform data and the relative depth data of the corresponding cross-sectional shape is used as new learning data.

[0092] S54: This process is equivalent to Figure 9A The flowchart steps S31 to S32 are processed. However, it is assumed that the cross-sectional shape estimation accuracy based on the neural network model is high enough at this time, so the processing of step S54 can be omitted, and the estimated cross-sectional shape can be used directly in step S55.

[0093] S55: The processor 601 uses the combination of the measured BSE signal waveform data and the relative depth data of the cross-sectional shape estimated by the neural network model or the relative depth data of the cross-sectional shape calculated in step S54 as new learning data.

[0094] S56~S57: This process is equivalent to Figure 9AThe flowchart steps S35 to S36 are the processing steps.

[0095] like Figure 11 As shown, multiple SEMs can be connected to the computing server 22. For example, in the same semiconductor manufacturing process, if SEMs 1101 to 1103 are used for inspecting the cross-sectional shape of patterns, a shared computing server 22 can be set up for SEMs 1101 to 1103. The computing server 22 trains and updates the neural network model 1110 based on the top-down BSE images obtained from SEMs 1101 to 1103, and stores the updated neural network model 1110 in the storage unit 19 of SEMs 1101 to 1103. In each SEM, the learned neural network model 1110 stored in the storage unit 19 is used to inspect the cross-sectional shape of the patterns.

[0096] Figure 12 An example of displaying the inspection results for cross-sectional shape. A rectangle 1202 dividing the inspection area is displayed within a circle 1201 representing the wafer being inspected. In this example, for rectangular areas where patterns containing bent cross-sections are found, color or patterns are applied to display them in a recognizable manner. In this case, it is possible to determine whether a shape is bent based on the presence of occlusion areas on the sidewalls of the pattern. Therefore, if it is determined to be only a bent shape, it is possible to... Figure 5 Steps S14 to S15 are omitted in the process. The twisting shape can also be determined based on the presence of the occluded area.

[0097] Figure 13 Another example of the inspection results showing the cross-sectional shape is shown. In display screen 1300, rectangles 1302 dividing the inspection area are displayed within a circle 1301 representing the wafer being inspected. In this example, the color or pattern is changed for each discovered cross-sectional shape to make it easily identifiable. The color or pattern displayed in rectangle 1302 in the legend bar 1303 represents the cross-sectional shape of the pattern. The divisions in the legend correspond to... Figure 7 The labeled pattern shown. Through inference based on a neural network model, the color or pattern displayed in rectangle 1302 is determined according to which labeled pattern it corresponds to. Alternatively, in using... Figure 10A In the case of the neural network model shown, the inference results are classified based on the legend's benchmark.

[0098] The embodiments of the present invention have been described above using the accompanying drawings. However, the present invention is not limited to the description of the embodiments shown above. Changes can be made to the specific structure without departing from the spirit or essence of the invention.

[0099] Furthermore, regarding the position, size, shape, and extent of the various structures shown in the accompanying drawings, etc., for ease of understanding of the invention, the actual position, size, shape, and extent are sometimes not indicated. Therefore, the present invention is not limited to the position, size, shape, and extent disclosed in the accompanying drawings, etc.

[0100] Furthermore, in the embodiments, control lines and information lines refer to the control lines and information lines required for the description, but may not necessarily represent all control lines and information lines on the product. For example, all structures may be interconnected.

[0101] Furthermore, the present invention is not limited to the illustrated embodiments, but includes various modifications. The illustrated embodiments are examples of structures described in detail for the purpose of readily understanding the present invention, and are not limited to having all the illustrated structures. In addition, it is possible to delete some structures in each embodiment, or to add or replace other structures, without causing contradictions.

[0102] Furthermore, regarding the aforementioned structures, functions, and processing units, some or all of them can be implemented in hardware using integrated circuits, for example. Additionally, the present invention can also be implemented using software program code for implementing the functions of the embodiments. In this case, a storage medium storing the program code is provided to a computer, and a processor of the computer reads the program code stored in the storage medium. In this case, the program code read from the storage medium itself implements the functions of the aforementioned embodiments.

[0103] Explanation of reference numerals in the attached figures

[0104] 1: Electron optical column, 2: Sample chamber, 3: Electron gun, 3a: Ideal optical axis, 4: Focusing lens, 5, 6: Deflector, 7: Objective lens, 8, 9: Electron detector, 10: Wafer, 11: XY stage, 12, 13: Amplifier, 14: Electron optical system control unit, 15: Stage control unit, 16: Control unit, 17: Image processing unit, 18: Computation unit, 19: Storage unit, 20: Device control unit, 21: Network, 22: Computing server, 100: Vertical illumination beam, 101, 201: Cross-sectional shape, 102: Reference Cross-sectional shape, 111, 211, 311: BSE signal waveform, 112, 312: reference waveform, 203: occlusion area, 300: tilted beam, 601: processor, 602: GPU, 603: memory, 604: storage device, 605: network interface, 606: input / output interface, 607: bus, 701, 1001: input data, 702, 1002: output data, 703, 1003, 1110: neural network model, 704, 1004: hidden layer, 705, 1005: fully connected layer.

Claims

1. A pattern inspection apparatus for inspecting the cross-sectional shape of a pattern formed on a sample, characterized in that, having: a storage device that stores a plurality of learned neural network models; an inference unit that estimates a cross-sectional shape of the pattern using one of the plurality of learned neural network models stored in the storage device; a learning unit that updates the learned neural network model by training the learned neural network model using new learning data, the learned neural network model takes, as input data, BSE signal waveform data representing backscattered electron signal intensity from the pattern in a first direction extracted from a top-down backscattered electron image taken for the pattern using a charged particle beam device, the inference unit selects, as a learned neural network model for estimating the cross-sectional shape of the pattern, a learned neural network model in which the optical conditions of an electron beam and the conditions of the sample at the time when the charged particle beam device took the top-down backscattered electron image are consistent, from among the plurality of learned neural network models, the learned neural network model takes, as output data, a probability that the input data corresponds to a predetermined plurality of labeled patterns, the learning unit takes, as the new learning data, a combination of the BSE signal waveform data and an approximate labeled pattern when the cross-sectional shape of the pattern measured based on the top-down backscattered electron image is approximate to the cross-sectional shape of the one of the plurality of labeled patterns.

2. The pattern inspection device according to claim 1, characterized in that the learning unit, when the cross-sectional shape of the pattern measured based on the top-down backscattered electron image is not approximate to any one of the cross-sectional shapes of the plurality of labeled patterns, generates a plurality of derived cross-sectional shapes approximate to the cross-sectional shape of a new labeled pattern defined when the cross-sectional shape of the new labeled pattern is not approximate to any one of the cross-sectional shapes of the plurality of labeled patterns, generates, for each of the plurality of derived cross-sectional shapes, BSE signal waveform data representing backscattered electron signal intensity from a pattern having the derived cross-sectional shape in the first direction by electron beam simulation, and takes, as the new learning data, a combination of the BSE signal waveform data for the derived cross-sectional shape and the new labeled pattern.

3. A pattern inspection apparatus which inspects a cross-sectional shape of a pattern formed on a sample, characterized by comprising: having: a storage device that stores a plurality of learned neural network models; an inference unit that estimates a cross-sectional shape of the pattern using one of the plurality of learned neural network models stored in the storage device; a learning unit that updates the learned neural network model by training the learned neural network model using new learning data, the learned neural network model takes, as input data, BSE signal waveform data representing backscattered electron signal intensity from the pattern in a first direction extracted from a top-down backscattered electron image taken for the pattern using a charged particle beam device, The inference unit selects a learned neural network model in which the conditions of the electron beam and the conditions of the sample at the time when the charged particle beam device acquired the top-down BSE image are consistent among the plurality of learned neural network models, as a learned neural network model for estimating the cross-sectional shape of the pattern, The learned neural network model outputs relative depth data representing the depth of the cross section of the pattern in the first direction with respect to the depth of the bottom surface of the pattern, When the difference between the cross-sectional shape of the pattern estimated by the inference unit and the cross-sectional shape of the existing learned data is equal to or less than a predetermined reference, the learning unit generates a combination of the BSE signal waveform data and the relative depth data of the cross-sectional shape of the pattern estimated by the inference unit or the relative depth data of the cross-sectional shape of the pattern measured based on the top-down BSE image as the new learned data.

4. The pattern inspection apparatus according to claim 3, wherein When the difference between the cross-sectional shape of the pattern estimated by the inference unit and the cross-sectional shape of the existing learned data exceeds the predetermined reference, the learning unit generates a plurality of derived cross-sectional shapes approximating the cross-sectional shape of the pattern measured based on the top-down BSE image, generates, for each of the plurality of derived cross-sectional shapes, BSE signal waveform data representing the backscattered electron signal intensity from a pattern having the derived cross-sectional shape in the first direction by electron beam simulation, and generates a combination of the BSE signal waveform data for the derived cross-sectional shape and the relative depth data of the derived cross-sectional shape as the new learned data.

5. The pattern inspection apparatus according to any one of claims 1 to 4, wherein the conditions of the primary electron beam include an acceleration voltage, an aperture angle, and a tilt angle of the primary electron beam, and the conditions of the sample include a material of the sample. ​

Images