Learning model suitable region detection device, learning model suitable region detection method, and learning model operation method

Abstract

Provided are a learning model suitable region detection device, a learning model suitable region detection method, and a learning model operation method, which are suitable for verification and correction of a suitable region of a learning model in machine learning. A learning model suitable region detection device is characterized by comprising: an image quality improvement processing unit that uses a learning model to convert a low-image-quality image into an image quality improved image that is a high-image-quality image; a model-suited region storage unit that stores, as a model-suited region, a region in which the learned information has been applied in a learning stage of learning the model; a model suitable region accuracy verification unit that verifies the accuracy of a measurement value obtained from the image quality-improved image on the basis of the measurement value of the high-quality image; and a model suited region correction processing unit that corrects the model suited region in the model suited region storage unit using the accuracy of the accuracy improvement information detected by the model suited region accuracy verification unit.

Description

Technical Field

[0001] The present invention relates to a learning model suitable area detection device, a learning model suitable area detection method, and a learning model application method that detects the suitable area of ​​a learning model and is suitable for verification and correction. Background Art

[0002] When using machine learning models to construct various systems and processes, it is necessary to appropriately determine the appropriate range of the learning model. To optimize the appropriate range, the validity of the appropriate range must be verified, and if it is inappropriate, it must be corrected to make it appropriate.

[0003] As a specific example of this requirement, there is a device used to measure the shape of a circuit pattern formed on a wafer using an electron microscope, etc. Here, for example, an image quality improvement system that converts a low-quality image into a high-quality image, and an example of a length measurement system using an image with improved image quality are envisioned.

[0004] When using an electron microscope to photograph circuit patterns formed on a wafer, an electron beam is irradiated onto the wafer, and a detector detects the rebounding electrons to produce an image. When the wafer is irradiated with an electron beam, the electron beam can scrape away the circuit pattern, damaging the circuit pattern formed on the wafer. Therefore, it is desirable to obtain an image with less electron beam exposure.

[0005] However, images obtained with a low electron beam irradiation rate contain a lot of particle-like noise and are unsuitable for shape measurement (hereinafter referred to as low-quality images). Image-based shape measurement detects the edge of a circuit pattern through image processing and measures the distance to the target measurement location based on the number of pixels in the image. The presence of particle-like noise can obscure the edge and lead to it being mistaken for the edge of a circuit pattern.

[0006] Therefore, in order to reduce such granular noise, the same shooting position is photographed multiple times, and the cumulative average of the images is taken to reduce the line noise (hereinafter, this image is referred to as a high-quality image).

[0007] This high-quality image is equivalent to an image of a true circuit pattern (a noise-free, high-quality image). Using this image enables accurate image measurement. The measured values ​​from this image are then set as true values.

[0008] However, to obtain high-quality images, multiple low-quality images must be collected. This means exposing the wafer to a large number of electron beams, which can increase wafer damage. Therefore, a different method than capturing multiple low-quality images to obtain a high-quality image is required to obtain high-quality images and perform accurate image measurements.

[0009] In contrast, methods using AI (machine learning) to improve image quality are examples of methods for obtaining (predicting) high-quality images from low-quality images. While these AI-based image quality improvement methods effectively improve image quality, they can sometimes produce adverse side effects during image measurement, thus requiring improvements to address these side effects.

[0010] As an example of a system that uses machine learning to transform low-quality images into high-quality images, in the system described in patent document 1, a learning model for machine learning is generated using low-quality images and high-quality images, and the learning model is used to transform the low-quality input image into a high-quality image through machine learning.

[0011] Prior art literature

[0012] Patent Literature

[0013] Patent Document 1: WJ 21 / 095256 Summary of the Invention

[0014] Problems to be solved by the invention

[0015] The system described in Patent Document 1 uses AI to improve image quality. Learning models are prepared in advance based on the purpose of image quality improvement. For example, if the goal is noise removal, a learning model is prepared that corresponds to the noise level; if the goal is aberration improvement, a learning model is prepared that corresponds to the aberration level. Based on this, the user selects a learning model based on the purpose of image quality improvement and the quality status of the low-quality image (noise and aberration conditions) and performs image quality improvement processing.

[0016] However, in the method of Patent Document 1, since the user visually judges and selects a learning model, there are cases where the selected learning model is ineffective in improving the image quality of the processing target image (hereinafter referred to as "learning model unsuitability"). However, there is no description of how to deal with such situations. Furthermore, there is no description of how to deal with the situation where separate learning models for noise removal and aberration improvement are used and then combined (for example, to improve both noise removal and aberration).

[0017] Furthermore, the system described in Patent Document 1 is designed for image quality improvement. Therefore, when using AI for image quality improvement, even when noise removal is visually excellent, the learned model can sometimes cause a side effect such as shifting of shapes and patterns within the image. When using the improved image for image measurement, this inappropriate shifting of shapes and patterns reduces measurement accuracy, a process that must be avoided. The system does not describe the side effects of this AI-based image quality improvement process or how to address them.

[0018] Based on the above, the present invention aims to provide a learning model suitable area detection device, a learning model suitable area detection method, and a learning model application method suitable for verifying and correcting the suitable area of ​​a learning model in machine learning.

[0019] Means for solving problems

[0020] According to the above content, in the present invention, it is set as "a learning model suitable area detection device, characterized in that it comprises: an image quality improvement processing unit, which uses a learning model to transform a low-quality image into an image quality improved image which is a high-quality image; a model suitable area storage unit, which stores the area where the learning information is implemented as a model suitable area during the learning stage of the learning model; a model suitable area accuracy verification unit, which verifies the accuracy of the measurement value obtained based on the image quality improved image with the measurement value of the high-quality image as a reference; and a model suitable area correction processing unit, which uses the accuracy of the accuracy improvement information detected by the model suitable area accuracy verification unit to correct the model suitable area in the model suitable area storage unit."

[0021] In addition, the present invention is set as "a learning model suitable area detection method, characterized in that a low-quality image is transformed into a quality-improved image as a high-quality image using a learning model, and in the learning stage of the learning model, the area where the learned information is implemented is stored as a model suitable area, and the accuracy of the measurement value obtained based on the quality-improved image is verified based on the actual measurement value, and the accuracy of the accuracy improvement information is used to correct the model suitable area."

[0022] In addition, the present invention is set as "a learning model application method, characterized in that it includes: a learning stage, obtaining a learning model for inferring a quality-improved image from a low-quality image using AI, and storing an area where the information of the learning is implemented as a model-suitable area; and an application stage, based on the information of the model-suitable area, determining whether the accuracy improvement status when using the learning model to infer a quality-improved image from a low-quality image and using the quality-improved image to perform image measurement is information of the model-suitable area of ​​the learning model, and verifying the accuracy improvement status based on high-precision information, and correcting the model-suitable area using the accuracy of the accuracy improvement status."

[0023] In addition, in the present invention, it is set as "a learning model suitable area detection device, characterized in that it comprises: an improvement processing unit, which uses a learning model to transform low-precision information into improved information as high-precision information; a model suitable area storage unit, which stores the area where the learned information is implemented as a model suitable area during the learning stage of the learning model; a model suitable area accuracy verification unit, which verifies the accuracy of the measurement value obtained based on the improvement information with the measurement value of the high-precision information as a reference; and a model suitable area correction processing unit, which uses the accuracy of the accuracy improvement information detected by the model suitable area accuracy verification unit to correct the model suitable area in the model suitable area storage unit."

[0024] Effects of the Invention

[0025] According to the present invention, it is possible to provide a suitable region detection device and method for a machine learning model that are suitable for verification and correction of the suitable region of a learning model in machine learning.

[0026] Specifically, according to the embodiment of the present invention, it is possible to understand the decrease in measurement accuracy of image measurement by notifying that the learning model is not suitable, thereby preventing the decrease in measurement accuracy caused by using inappropriately transformed images.

[0027] Furthermore, according to the embodiment of the present invention, even if a previously used learning model has not been registered as a measurement area of ​​a model adaptation area, the model adaptation area of ​​the learning model can be added or deleted by verifying the measurement accuracy during operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a diagram of a suitable area detection device and method for a machine learning model according to Example 1 of the present invention, and particularly shows a structural example for determining a suitable area of ​​a model.

[0029] Figure 2 It is a diagram showing an example of upper and lower layers of a multilayer semiconductor pattern.

[0030] Figure 3 This is a diagram showing the relationship between a low-quality image, an improved-quality image, and a high-quality image.

[0031] Figure 4 This is a diagram showing a configuration example of a processing block for OVL measurement.

[0032] Figure 5 This is a diagram showing the occurrence of measurement errors inside and outside the learning area.

[0033] Figure 6 This is a diagram showing the occurrence of errors in the learning area.

[0034] Figure 7 This is a diagram showing how the model adaptation area changes within the learning area.

[0035] Figure 8 This is a diagram showing processing blocks during actual operation.

[0036] Figure 9 This is a diagram showing an example of a case where measurement accuracy improves even outside the learning area.

[0037] Figure 10 This diagram shows how the model adaptation area outside the learning area changes.

[0038] Figure 11 This is a diagram showing an example of point shape measurement. DETAILED DESCRIPTION

[0039] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0040] Example 1

[0041] In the first embodiment of the present invention, the learning model fit region detection device and method are described, applying the device and method to an image quality improvement system and a length measurement system using improved image quality images. Specifically, the first embodiment describes the accuracy verification of the model fit region and the reduction of the achieved region.

[0042] Figure 1 The following describes an example of a configuration for detecting an appropriate region of a learning model and a method thereof, particularly for determining an appropriate region of a model, in accordance with the first embodiment of the present invention.

[0043] exist Figure 1In the figure, PL is the low-quality image to be improved, Ph is the high-quality image capable of high-precision image measurement, and PhA is the high-quality image (hereinafter referred to as the image quality improved image) estimated from the low-quality image PL through AI processing. Furthermore, DB1 is the learning model, 3 is the image quality improvement processing unit, 5 is the accuracy verification unit for the model-fitting region, 6 is the correction processing unit for the model-fitting region, DB2 is the storage unit for the model-fitting region, 8 is the output unit for measurement and judgment results, and 4L and 4h are the OVL value measurement processing units.

[0044] Here, a low-quality image PL refers to an electron microscope image captured using a low electron beam irradiation rate. If the image is noisy and image measurement processing is performed, measurement becomes impossible or measurement accuracy is reduced, resulting in a low-quality image PL. In contrast, a high-quality image Ph reflects the actual posture of the observed part and is considered an accurate value. A high-quality image Ph is an image obtained by capturing multiple low-quality cumulative images PL of the same part and averaging these images to remove noise.

[0045] In the learning model suitable area detection device and method of the present invention, the deviation of the posture, shape and position represented by the image quality improved image obtained by using AI image quality improvement processing to estimate the image obtained by shooting the low-quality image PL from the actual posture, shape and position of the observed object part represented by the high-quality image Ph is detected.

[0046] First, use Figure 2 and Figure 3 An example of low-quality image PL will be described. In the present description, an example of wafer chip image measurement is described, in which the deviation (OVL value) of the upper and lower layers of a multilayer circuit structure is measured (overlay measurement).

[0047] Figure 2 2a is an example of an upper circuit pattern, and 2c is an upper circuit wiring. In addition, 2b is a lower circuit pattern, and 2d is a lower circuit wiring. In the semiconductor manufacturing process, circuit pattern 2a is formed on circuit pattern 2b.

[0048] Figure 3 It is a diagram showing the relationship between a low-quality image and an improved-quality image. Figure 3 The measured low-quality image PL shown on the left side of FIG is transformed into the image quality improved image PhA on the upper right side by improving the image quality through machine learning. Figure 3 The high-quality image Ph shown in the lower part is for comparison.

[0049] Among them, for the low-quality image PL, by changing the acceleration voltage of the electron beam and the position of the electron beam detector, it is possible to capture not only the upper circuit pattern 2c but also the lower circuit pattern 2d that is blocked by the upper layer. Figure 3 The low-quality image PL on the upper left schematically depicts a state where the shooting time is short and there is much granular noise.

[0050] The image obtained by converting the low-quality image PL into an improved image equivalent to a high-quality image using the image quality improvement processing unit 3 is Figure 3 The image quality improved image PhA on the upper right. The image quality improvement processing unit 3 can be implemented using machine learning such as CNN (Convolutional Neural Network). CNN learns a high-quality image Ph obtained by photographing the same position and improved in quality by cumulatively averaging low-quality images PL, and learns a means for converting a low-quality image into a high-quality image. Hereinafter, a low-quality image and a high-quality image obtained by photographing the same part are referred to as an image pair. The learning model learned here is learning model DB1. Furthermore, using learning model DB1, the low-quality image PL is converted into a quality equivalent to a high-quality image obtained by taking a cumulative average of multiple low-quality images.

[0051] The image quality improvement processing unit 3 removes the noise from the low-quality image PL, which contains a lot of granular noise, and transforms it into an improved-quality image PhA with good image quality. This improved-quality image PhA clearly shows not only the upper-layer circuit wiring 2c but also the lower-layer circuit wiring 2d. Furthermore, the edge positions 3d' and 3e' of each wiring, as well as the distance 3c' between the upper and lower wirings, can be clearly determined.

[0052] In contrast, for example, if the information of the high-quality image Ph obtained by cumulative averaging of low-quality images is compared with the measurement information of the transformed image PhA with improved image quality, the deviation between the two can be grasped. The difference between the edge position 3d', 3e', and distance 3c of each wiring in the transformed image PhA with improved image quality relative to the edge position 3d, 3e, and distance 3c of each wiring in the high-quality image Ph can be grasped as a deviation. The deviation in this case is a deviation caused by the learning model. In order to grasp this deviation, Figure 1 The OVL value measurement processing units 4L and 4h perform a process of measuring (superimposing measurement) the amount of deviation (OVL value) between upper and lower layers of a multilayer circuit structure.

[0053] Figure 4The structure of the processing block 4L for measuring the OVL value is shown. The transformed image quality improved image PhA is input to the edge position detection processing block 4b of the upper pattern and the edge position detection processing block 4c of the lower pattern, and the edge positions 3d' and 3e' are detected respectively. Figure 3 The edge position 3d' of the upper layer circuit pattern 2c is detected by the lower layer pattern edge position detection block 4c. The edge position 3e' of the lower layer circuit pattern 2d is detected by the lower layer pattern edge position detection block 4c. The upper and lower layer distance calculation block 4d calculates the upper and lower layer distance 3c' based on the difference in the detected positions of the upper and lower layer edges and outputs it as the OVL value.

[0054] Although not shown in the figure, similar processing is also performed in the OVL value measurement processing block 4h, and the edge position 3d of the upper layer pattern, the edge position 3e of the upper layer pattern, and the distance 3c between the upper and lower layers are measured.

[0055] exist Figure 1 The accuracy verification unit 5 of the model fit region uses the distances 3c and 3c' from the upper and lower layers of the processing blocks 4L and 4h, which measure the OVL values, to perform accuracy verification of the model fit region. While the following embodiments describe an example where the model fit region is defined using measurement accuracy information, the model fit region can be defined using various methods besides "measurement accuracy information," such as "image brightness information," "noise information," "edge feature information," and objective image evaluation metrics such as SNR (Signal to Noise Ratio), PSNR (Peak Signal to Noise Ratio), and SSIM (Structural Similarity). In this specification, all of these are referred to as measurement accuracy information. Furthermore, measurement accuracy information can be used together with information on its reliability. In other words, it is best to manage the accuracy of a process in conjunction with the degree of reliability it supports.

[0056] Figure 5 An example showing the relationship between the error in the learning area and other areas. Figure 5 In the figure, the horizontal axis represents the OVL value. A certain reference value (design value, etc.) is corrected to 0. The OVL of the learning data is indicated by a triangle mark (5b). The measurement of the OVL value is recorded using the high-quality image Ph as a reference.

[0057] Even if OVL = 0 is used during design, the OVL value will not be zero due to variations in manufacturing equipment, but will vary. In this example, the OVL value ranges from -5 to +5. The vertical axis represents the difference (error) between the OVL value 3c obtained from the high-quality image Ph and the OVL value 3c' obtained from the improved image PhA.

[0058] Region (OVL value: -5 to +5) 5a represents the range of OVL values ​​in the learning data (hereinafter referred to as the learning region). Within this region, the error is centered around the error 0. However, as shown in 5c, as the OVL value becomes increasingly negative, the center of the error deviation becomes proportionally larger. This indicates that as the OVL value becomes increasingly negative, the OVL value obtained from the image quality-improved image PhA becomes smaller than the OVL value obtained from the high-quality image Ph.

[0059] That is, while removing the particle noise, the learning model also learns the deviation between the upper and lower layers remembered during learning. When a low-quality image PL with a large negative OVL value is input, the upper and lower layers are moved in order to return to the state where the distance between the upper and lower layers remembered during learning is small. Therefore, it is believed that such a result is obtained.

[0060] Due to this phenomenon, the accuracy cannot be guaranteed in the area of ​​the OVL value that has not been learned during learning. Therefore, regarding the initial state of the model suitable area stored in the model suitable area storage unit DB2, the range of the OVL value learned when generating the learning model is associated with the learning model as the model suitable area and registered. Figure 5 In the example, the initial state of the model's suitable area is the area with OVL values ​​of -5 to +5.

[0061] However, depending on the learning model, accuracy may not be guaranteed within the range of OVL values ​​learned during training. Figure 6 Indicate one example.

[0062] Figure 6 This diagram shows the occurrence of errors within the learning area. The OVL values ​​of the learning data are indicated by the triangular markers 6a and fall within the range 6c (-28 to 14). The deviation of the errors within this range varies, and as shown in 6b, there are areas where large errors occur. Therefore, the model fit area accuracy verification unit 5 performs accuracy verification to identify such areas where the errors are locally large.

[0063] This accuracy verification is performed in the model fit region accuracy verification unit 5 as follows. First, the OVL value to be determined is set to x, and within a window with a width of ±d centered on x, an allowable value is set for the difference between the OVL value 3c obtained from the high-quality image Ph and the OVL value 3c′ obtained from the improved image PhA, which is referred to as ±p (allowable error).

[0064] If the number of data points in the window that meet the allowable error is set to N(kp(x)), the number of data points that do not meet the allowable error is set to N(kn(x)), and the proportion that meets the allowable error is set to the allowable error achievement rate Ae(x), then formula (1) holds.

[0065] [Formula 1]

[0066]

[0067] In equation (1), the permissible error rate Ae(x) represents the error generation status within the window. Therefore, an acceptable permissible error rate Aeth is determined. If Ae(x) exceeds Aeth, the model is considered to be within the model fit region. Thus, the model fit region determination function Mf(x) becomes as shown in equation (2).

[0068] [Formula 2]

[0069]

[0070] Figure 7 A schematic diagram showing a state of change of the model adaptation region performed by the model adaptation region accuracy verification unit 5 is shown.

[0071] exist Figure 7 In the example, 7a is the initial state. In the initial state, when generating the learning model, the range of OVL values ​​of the learning image (learning area) a to b is used as the model fitting area (7d).

[0072] exist Figure 7 In the example, 7b is the accuracy verification result of the model fit region in the initial state, obtained using the calculation method shown in equation (1). The allowable error achievement rate Ae(x) for regions a to e (7e) is greater than the allowable error achievement rate Aeth, while the allowable error achievement rate Aeth for regions e to b (7f) is less than the allowable error achievement rate Aeth. Therefore, according to equation (2), region 7e is determined to be a model fit region, and region 7f is determined to be outside the model fit region.

[0073] Based on this result, the model fit region after accuracy verification is shown in 7c. Region 7h, which was determined to be outside the model fit region, is excluded from the initial model fit region (7d), resulting in regions a through e (7g). Furthermore, as described above, along with the reduction of the model region, the permissible error achievement rate Ae(x) for each value x within the model region may be registered, or the function of equation (1) may be stored in the model fit region storage unit DB2. This allows the model reliability Ae(x) to be displayed (prompted) along with the determination of whether the model is inside or outside the model fit region.

[0074] Based on this result, the model adaptation region correction processing unit 6 corrects the model adaptation region in the model adaptation region storage unit DB2 associated with the learning model used in the current image quality improvement processing to region 7g.

[0075] As described above, to verify and correct the adaptive range of the learning model, image pairs consisting of low-quality images PL and their paired high-quality images Ph are required. This process can be performed when the learning model DB1 is generated, or during testing during production line modifications, low-quality images PL and their paired high-quality images Ph are collected and the model adaptive range storage unit DB2 is updated.

[0076] In Example 1, the case where the learning model-adapted region detection device and method are applied to an image quality improvement system and a length measurement system using an image quality-improved image is described. However, the application example is not limited to the image quality improvement system.

[0077] More generally, by "having: an accuracy conversion unit that uses a learning model to convert low-precision information into high-precision accuracy improvement information; a model suitable area storage unit that stores the area where the learned information is implemented as a model suitable area during the learning phase of the learning model; a model suitable area accuracy verification unit that verifies the accuracy of the accuracy improvement information based on the high-precision information; and a model suitable area correction processing unit that uses the accuracy of the accuracy improvement information detected by the model suitable area accuracy verification unit to correct the model suitable area in the model suitable area storage unit", it can be applied to a wider range of fields.

[0078] Example 2

[0079] In the first embodiment, accuracy verification is performed within the model fit region to correct the information in the model fit region storage unit DB2. In contrast, in the second embodiment, measurement and judgment of a measured low-quality image are performed using the corrected model fit region.

[0080] Figure 8This section shows the processing blocks used for actual operation using information from the model fit region storage unit DB2, which has undergone accuracy verification within the model fit region. During actual operation, image measurement processing is performed using only the low-quality image PL. By performing image quality improvement processing on the low-quality image PL using the learning model DB1 in the image quality improvement processing unit 3, the low-quality image PL is transformed into an improved image PhA of comparable high quality. The OVL value of the transformed improved image PhA is calculated using the OVL value measurement processing unit 4L.

[0081] The model fit region determination unit 80a uses this OVL value to perform a determination. The determination is made using information from the model fit region storage unit DB2. If the OVL value is outside the model fit region, a warning message such as "Measurement is not possible due to the learning model not fitting" can be displayed using the measurement and determination result output unit 80a. Alternatively, a display such as "OVL value is xx.x [nm]. Tolerance error achievement rate xx [%], accuracy reliability is low" can be displayed using the recorded tolerance error achievement rate Ae(x) information. (The "xx" value is a combination of the measured value and the reliability value.)

[0082] According to this embodiment, measurement data satisfying measurement accuracy can be obtained in the learning area of ​​the selected learning model.

[0083] Example 3

[0084] In Examples 1 and 2, we described methods related to accuracy verification within the learning area and addressed the countermeasures for sometimes poor accuracy even within the learning area. In Example 3, we described accuracy verification outside the learning area and explained how to modify the model's fit region. Modification, in this context, refers to expanding or changing the model's fit region.

[0085] Figure 9 This shows the deviation of the error outside the learning area using the learning model. 9a is the OVL value of the learning data, and 9b is the learning area that represents the range of the OVL value of the learning data. The area to the right of 9b is outside the learning area, so the error becomes larger. However, the error in area 9c to the left of 9b does not increase. Figure 5 Different results. This means that depending on the learning model used, the error may not increase even outside the learning area. Therefore, if accuracy verification outside the learning area can be performed and confirmed, the model fit range associated with the learning model used can be expanded.

[0086] The extension of the model adaptation range outside the learning area can be done in the same way as described in Example 1. Figure 1The same processing structure as the processing block shown in the figure is used for processing. However, in Example 1, the accuracy verification was performed using the low-quality image PL and the high-quality image Ph in the learning area, but in this embodiment, it is necessary to obtain an image pair outside the learning area. For example, such data can also be collected during test operations when the manufacturing process is changed. In addition, in the accuracy verification unit 5 of the model suitable area, as shown in formula (1), image data corresponding to a predetermined window width is required, so a plurality of image data are required to meet the window width. Hereinafter, this will be referred to as an image group.

[0087] Figure 10 A schematic diagram showing the change in the model fit region outside the learning region is shown. Figure 10 is Figure 7 This is a scenario where the accuracy verification of the model fit area within the learning area is followed by the accuracy verification outside the learning area.

[0088] exist Figure 10 In , 10d represents an example of a case where an image group with a new OVL value appears outside the learning area. Figure 7 In the example, a and b are the learning areas, and the image groups with new OVL values ​​are captured on the left side of them. These image groups are subjected to accuracy verification processing by the accuracy verification unit of the five model fitting areas to obtain the allowable error achievement rate Ae(x).

[0089] The result is 10e. The tolerance rate of the newly appeared interval 9a is 10b, which exceeds the accuracy threshold and is within the model fit region. 10f indicates that Figure 7 The model described in is suitable for the case where the results are added in the region. Figure 7 As shown in Figure 1. This shows the case where the model fit intervals for the allowable error achievement rate Ae(x) (7e) for data within the learning region are combined, and the additional model fit intervals are c to e (10c). These results are registered in the model fit region storage unit of DB2 in a state associated with the used learning model.

[0090] The actual operation is the same as in Example 2. Figure 8 The processing blocks shown perform processing and output measurement results and accuracy determination results.

[0091] According to this embodiment, even outside the learning area of ​​the selected learning model, measurement data satisfying the measurement accuracy can be obtained.

[0092] Furthermore, by using this embodiment in combination with the above-described first embodiment, length measurement data satisfying measurement accuracy can be obtained within the model fitting period where measurement accuracy can be verified, both inside and outside the learning period.

[0093] Moreover, as a method different from this method, it is also considered to generate a new learning model within a wide range of OVL values ​​to expand the model suitability range. In this case, in order to reconstruct (relearn) the learning model, a large amount of calculation time and accuracy confirmation work and confirmation work accompanying the switching of the learning model are generated. This method does not change the learning model. It is a method that maximizes the model suitability range of the existing learning model. The formulas (1) and (2) used in this method are simple calculations. Compared with the reconstruction of the learning model, the required amount of calculation is greatly reduced, and the change of the suitability range can be completed in a short time.

[0094] in addition, Figure 3 An example of measuring the deviation between upper and lower layers (OVL measurement) is shown. However, as an example of the side effect of AI-based image quality improvement, unnecessary line width correction is performed to make the line width of the circuit pattern close to the line width of the learned image, or Figure 11 In the case of a circular pattern shape such as that shown, unnecessary corrections are sometimes performed to make it close to the size of the learned circular shape. Even in such a case, the method described in the present invention can be applied. The model application area can be set to the size of the line width, inner diameter, and outer diameter of the circular shape. If the line width, outer diameter, and inner diameter are the same as the model suitable area, the reliability of the measurement value obtained based on the image quality improved image can be guaranteed. However, if not, as described in Examples 1-3 of the present invention, a warning message such as "Accuracy cannot be guaranteed" can be prompted, and the reduction or expansion correction of the model suitable area can also be performed.

[0095] According to the present invention, it is possible to provide a suitable region detection device and method for a machine learning model that are suitable for verification and correction of the suitable region of a learning model in machine learning.

[0096] Specifically, according to the embodiment of the present invention, it is possible to understand the decrease in measurement accuracy of image measurement by notifying that the learning model is not suitable, thereby preventing the decrease in measurement accuracy caused by using inappropriately transformed images.

[0097] Furthermore, according to the embodiment of the present invention, even if a previously used learning model has not been registered as a measurement area of ​​a model adaptation area, the model adaptation area of ​​the learning model can be added or deleted by verifying the measurement accuracy during operation.

[0098] Example 4

[0099] While the above embodiments describe a learning model-based region detection device and method, in Embodiment 4, a learning model application method utilizing these devices and methods will be described.

[0100] As can be understood from Examples 1-3, this learning model application method is described in two stages: the learning model learning phase and the learning model application phase. In summary, relearning is typically used to address the situation where the appropriateness of a learning model generated during initial learning decreases during subsequent application. However, this is not a burdensome relearning, but rather a method that changes the appropriate range of the initial learning model based on the detection accuracy during normal operation.

[0101] Therefore, in the learning model application method of Example 4, it is set to "comprise: a learning stage, obtaining a learning model for inferring a quality-improved image from a low-quality image using AI, and storing an area where the information of the learning is implemented as a model-suitable area; and an application stage, based on the information of the model-suitable area, determining whether the accuracy improvement status when inferring a quality-improved image from a low-quality image using the learning model and performing image measurement using the quality-improved image is the information of the model-suitable area of ​​the learning model, and verifying the accuracy improvement status based on the high-precision information, and correcting the model-suitable area using the accuracy of the accuracy improvement status."

[0102] Example 5

[0103] In the above embodiment, an example of obtaining a high-quality image by using a learning model for a measured low-quality image is described. However, the application examples of the present invention are not limited to image quality.

[0104] The present invention is applicable to any device that uses a learning model to improve the accuracy of low-precision information. This is because, when using a learning model to improve accuracy, the present invention addresses the side effects of unwanted bias caused by the learning model, a crucial issue to address in scenarios where learning models are used to improve accuracy.

[0105] Therefore, the present invention can also be applied to sounds measured as low-precision information instead of low-quality images, or pressure, temperature, vibration, etc. in various factories. In this case, the present invention can be configured as "a learning model suitable area detection device, characterized by comprising: an improvement processing unit that uses a learning model to convert low-precision information into improved information as high-precision information; a model suitable area storage unit that stores the area where the learned information is implemented as a model suitable area during the learning stage of the learning model; a model suitable area accuracy verification unit that verifies the accuracy of the measurement value obtained based on the improved information with the measurement value of the high-precision information as a reference; and a model suitable area correction processing unit that uses the accuracy of the accuracy improvement information detected by the model suitable area accuracy verification unit to correct the model suitable area in the model suitable area storage unit."

[0106] Example 6

[0107] In Example 6, the matters described in Example 1 to Example 4 are summarized.

[0108] exist Figure 1 In the learning model-suitable region detection device of the present invention shown above, since three types of images are processed, the methods for obtaining the characteristics of each of these images and the characteristics of the images are collectively described as follows.

[0109] First, the low-quality image PL is an SEM image obtained by electron beam irradiation of a small portion of the observation target, and contains particle-like noise. Therefore, when using this image for image measurement, measurement errors may occur.

[0110] In contrast, the high-quality image Ph reflects the correct, original posture of the observed part and should be considered the correct value. The high-quality image Ph is obtained by capturing multiple low-quality images PL and calculating the cumulative average. The granular noise contained in the low-quality images PL is randomly generated and can therefore be removed through this cumulative averaging process. However, due to the damage to the wafer caused by increased electron beam irradiation, it is not possible to frequently obtain these high-quality images Ph obtained through multiple irradiations. Therefore, they are only acquired to prepare for damage when generating the learning model.

[0111] Furthermore, the improved image PhA is a high-quality image estimated from the low-quality image PL through AI processing using a learning model. To clearly distinguish it from the high-quality image Ph calculated by cumulatively averaging the low-quality images PL, it is labeled "improved image PhA" here. The improved image PhA is obtained by AI inferring a noise-free image from a single low-quality image using a learning model. Using an undesirable learning model in this case can result in the side effect of arbitrarily correcting the offset between the upper and lower layers. However, this side effect does not occur with the high-quality image calculated by cumulatively averaging the low-quality images.

[0112] In the learning model suitable area detection device and method according to the present invention, an attempt is made to detect how much the actual posture, shape and position of the part of the observation object represented by the low-quality image PL representing the observation part of the actual manufactured product deviates from the original correct posture, shape and position of the part of the observation object represented by the high-quality image Ph.

[0113] However, since sufficient detection accuracy cannot be achieved in the state of the low-quality image PL, the low-quality image PL is converted into a high-quality improved image PhA through machine learning and then compared with the high-quality image Ph.

[0114] This comparison is performed in consideration of the characteristics and features of the three types of images described above, and is correlated with the measurement values ​​(OVL values ​​4L and 4h) obtained by image measurement.

[0115] Regarding the measured values ​​obtained through image measurement, first, in the case of the low-quality image PL, the edges of the circuit pattern are unclear due to noise, making measurement impossible. The measured values ​​obtained from the high-quality image Ph include "errors due to manufacturing process factors" in the "designed value of the upper and lower layer deviation amount." Furthermore, in contrast, the measured values ​​obtained from the improved image PhA also include "unnecessary deviations (side effects) caused by AI-based image quality improvement" in the "designed value of the upper and lower layer deviation amount" and "errors due to manufacturing process factors."

[0116] In view of these characteristics and features, what the present invention should achieve is to grasp the area (model suitable area) where the "unnecessary deviation (side effect) caused by AI-based image quality improvement" is small. If the "unnecessary deviation (side effect) caused by AI-based image quality improvement" is small, the size of the "error of manufacturing process factors" can be accurately measured.

[0117] When the "design value of the upper and lower layer deviation amount in the high-quality image Ph based on the design information of the observation target site" is used as a reference, the "measured value obtained from the image quality improved image" minus the "design value of the upper and lower layer deviation amount in the high-quality image Ph based on the design information of the observation target site" is a combination of "errors due to manufacturing process factors" and "unnecessary deviations (side effects) caused by AI-based image quality improvement." In other words, the error obtained is a mixture of "errors due to manufacturing process factors" and "unnecessary deviations (side effects) caused by AI-based image quality improvement," making it impossible to understand the "unnecessary deviations caused by AI-based image quality improvement."

[0118] At this time, in order to calculate the "unnecessary deviation caused by AI-based image quality improvement", the "measurement value obtained based on the image with improved image quality" - "measurement value obtained based on the image with improved image quality" becomes the "unnecessary deviation (side effect) caused by AI-based image quality improvement", and the "unnecessary deviation caused by AI-based image quality improvement" can be calculated.

[0119] In the present invention, attention is paid to the above-mentioned features shown in each image, and high accuracy is achieved by utilizing the relationship between them.

[0120] Description of Reference Numerals

[0121] PL: low-quality image, Ph: high-quality image, DB1: learning model, 3: image quality improvement processing unit, 4 (4L, 4h): OVL value measurement processing unit, 5: model suitable area accuracy verification unit, 6: model suitable area correction processing unit, DB2: model suitable area storage unit.

Claims

1. A learning model suitable region detection device, characterized in that The learning model is suitable for the region detection device having: an image quality improvement processing unit that converts the low-quality image into an image quality improved image that is a high-quality image using a learning model; a model-fitting region storage unit that stores, during a learning phase of learning a model, a region where information on learning has been performed as a model-fitting region; The model fit area accuracy verification unit verifies the accuracy of the measurement values ​​obtained from the image quality improved image, using the measurement values ​​of the high-quality image as a benchmark; as well as The model fit region correction processing unit corrects the model fit region in the model fit region storage unit using the accuracy of the accuracy improvement information detected by the model fit region accuracy verification unit.

2. The learning model suitable area detection device according to claim 1, characterized in that The learning model applicable region detection device includes a model region determination unit that determines whether the accuracy improvement information is information on the model applicable region of the learning model based on information on the model applicable region in the model applicable region storage unit.

3. The learning model suitable area detection device according to claim 1 or 2, characterized in that The low-quality image information is information about a low-quality image obtained by photographing a target product, and the high-quality image information is information about a high-quality image of the target product.

4. The learning model suitable area detection device according to claim 1, characterized in that The correction of the model adaptation region by the model adaptation region storage unit in the model adaptation region correction processing unit is correction to change a part of the model adaptation region or to expand the model adaptation region.

5. The learning model suitable area detection device according to claim 4, characterized in that The model adaptation region storage unit in the model adaptation region correction processing unit corrects the model adaptation region by adding measurement accuracy information and reliability information of the model adaptation region.

6. The learning model suitable area detection device according to claim 5, characterized in that The measurement accuracy information is any one of an image measurement value, image brightness information, noise amount information, an edge feature amount, and an objective evaluation scale of an image.

7. The learning model suitable area detection device according to claim 5, characterized in that The model region determination unit determines whether the accuracy improvement information is performed within the model region based on the measurement result of the accuracy improvement information. If not, a warning is output and the measurement accuracy information or reliability information is added to the measurement value and outputted.

8. The learning model suitable area detection device according to claim 3, characterized in that The target product is a semiconductor having a multi-layer structure, and the information of the low-quality image is the distance between the upper layer and the lower layer of the semiconductor pattern.

9. The learning model suitable area detection device according to claim 7, characterized in that The learning model is suitable for an area detection device to detect the line width of a semiconductor having a multilayer structure.

10. A learning model suitable for region detection method, characterized in that A learning model is used to transform a low-quality image into a quality-improved image, which is a high-quality image. During the learning phase of the learning model, the area where the learned information is implemented is stored as a model-fitting area. The accuracy of the measurement values ​​obtained from the quality-improved image is verified using the actual measurement values ​​as a reference, and the accuracy of the accuracy improvement information is used to correct the model-fitting area.

11. The learning model suitable area detection method according to claim 10, characterized in that: Whether the accuracy improvement information is information on the model fit region of the learning model is determined based on information on the model fit region.

12. A method for applying a learning model, characterized in that: include: In the learning phase, a learning model is obtained for estimating an improved image from a low-quality image using AI, and the area where the learned information is applied is stored as a model-fitting area. as well as During the application phase, based on the information about the model's fit area, a determination is made as to whether the improvement in accuracy when using the learned model to infer an improved image from a low-quality image and performing image measurement using the improved image corresponds to the information about the model's fit area for the learned model. Furthermore, the improvement in accuracy is verified using the high-precision information as a benchmark, and the model's fit area is corrected using the accuracy of the improvement in accuracy.

13. A learning model suitable area detection device, characterized in that The learning model is suitable for the region detection device having: an improvement processing unit that converts low-precision information into improved information that is high-precision information using a learning model; a model-fitting region storage unit that stores, during a learning phase of learning a model, a region where information on learning has been performed as a model-fitting region; The model fit area accuracy verification department verifies the accuracy of the measurement values ​​obtained based on the improved information, using the measurement values ​​of the high-precision information as a benchmark; as well as The model fit region correction processing unit corrects the model fit region in the model fit region storage unit using the accuracy of the accuracy improvement information detected by the model fit region accuracy verification unit.

Images