Surface defect detection method, surface defect detection device, and surface defect detection program

Abstract

To detect the presence or absence of a surface defect and position information of a semiconductor wafer by utilizing a machine learning even if teacher data cannot be sufficiently secured.SOLUTION: A surface defect detection method for detecting a surface defect from thickness measurement data of an entire surface of a semiconductor wafer measured using a wafer thickness measurement apparatus includes steps of: extracting and quantifying a feature of the surface defect from the thickness measurement data, adding a pseudo surface defect to the thickness measurement data of the semiconductor wafer having no surface defect using a quantified surface defect data, machine learning a detection model of the surface defect by including the thickness measurement data of the semiconductor wafer to which the pseudo surface defect is added in teacher data, and detecting the surface defect from the thickness measurement data of the semiconductor wafer to be inspected using the detection model of the surface defect.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a surface defect detection method, a surface defect detection device, and a surface defect detection program. [Background technology]

[0002] In the manufacturing of semiconductor wafers, including silicon wafers, wafers after grinding or etching processes have a rougher surface than the mirror-polished final wafer. Wafers with such rough surfaces can have dimples, large-size pits (LSPs; holes created by abrasive grains), and protrusions in the center of the wafer caused by the grinder (collectively referred to as surface defects). In this case, the wafer may be rejected even after subsequent manufacturing processes. In particular, during the grinding process, if the next wafer is held by suction on the chuck table's holding surface while grinding debris is still attached, the area of ​​the wafer where the grinding debris is present can be excessively ground, resulting in dimples on the surface of the wafer after grinding. These dimples are called dimples.

[0003] Dimples cause problems such as uneven wafer thickness and reduced wafer quality, so they must be removed during the process. Conventionally, dimples that occur on the wafer surface are detected by converting the state of the wafer surface into an image using a magic mirror (see, for example, Non-Patent Document 1). This method not only requires the human eye to judge each converted image, but also makes it difficult to judge surface defects if the shape of the surface defect is distorted or the surface of the defect is rough. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Tokura, Seitaro et al. “Characterization of mirror-polished silicon wafers by Makyoh method.” Journal of Crystal Growth 103 (1990): 437-442. [Patent documents]

[0005] [Patent Document 1] Japanese Patent Application Laid-Open No. 2011-232267 [Patent Document 2] Japanese Patent Application Laid-Open No. 2024-32486 Summary of the Invention [Problem to be solved by the invention]

[0006] When surface defects such as dimples occur, it is desirable to detect them early in the manufacturing process and remove the wafers. However, surface defects that occur on wafers with rough surfaces after the grinding or etching process but before the mirror polishing process, particularly dimples that occur on wafers that have been ground and have scratches (also called grinding scratches), are shallow surface defects relative to the wafer surface (rough surface), so if the shape of the surface defect is distorted, it may be overlooked by the human eye.

[0007] Furthermore, a method for inspecting wafer surface defects by utilizing scattered light has been proposed (see, for example, Patent Document 1), but if the wafer surface is rough, it has not been possible to clearly detect surface defects.

[0008] Furthermore, while there are hopes for automatic defect detection methods using machine learning, surface defects such as dimples occur very rarely to begin with, and it can be difficult to secure sufficient training data. In this case, there is a problem that the detection accuracy of machine learning is low with limited training data, and surface defects are not detected sufficiently, leading to wafer sampling being overlooked.

[0009] There is also a need for a method to identify the location of surface defects that occur on wafers. Knowing defect location information will help to identify the cause of detected surface defects, and when a defect is determined to be present, it will also make it easier to visually check the actual location of the surface defect on the wafer. In particular, to achieve zero oversight of surface defects, it is hoped that a slightly excessive number of wafers suspected of having surface defects will be detected as errors, thereby improving the workability of the final visual inspection by the operator.

[0010] In view of the above-mentioned problems, the object of the present invention is to provide a surface defect detection method, a surface defect detection device, and a surface defect detection program that utilize machine learning to detect the presence or absence of surface defects on a wafer and their position information even if sufficient training data cannot be secured. [Means for solving the problem]

[0011] In order to solve the above problems, the surface defect detection method of the present invention is a surface defect detection method that detects surface defects from thickness measurement data of the entire surface of a semiconductor wafer measured using a wafer thickness measurement device, and is characterized in that it extracts and quantifies characteristics of the surface defects from the thickness measurement data, uses the quantified surface defect data to assign pseudo surface defects to thickness measurement data of a semiconductor wafer that has no surface defects, includes the thickness measurement data of the semiconductor wafer with the pseudo surface defects assigned as training data, machine-learns a surface defect detection model, and uses the surface defect detection model to detect surface defects from the thickness measurement data of the semiconductor wafer to be inspected.

[0012] The surface defect detection method of the present invention adds pseudo surface defects to thickness measurement data of a semiconductor wafer that has no surface defects, includes this as training data, and machine-learns a surface defect detection model, thereby making it possible to utilize machine learning even if it is not possible to secure sufficient training data.

[0013] In the above-described surface defect detection method, the semiconductor wafer may be a semiconductor wafer after a grinding process or an etching process but before a mirror polishing process, or the semiconductor wafer may be a semiconductor wafer having scratches. The surface defects may be dimples.

[0014] The training data preferably includes position information on the semiconductor wafer, in order to enable detection of the presence or absence of surface defects on the semiconductor wafer and their position information.

[0015] It is preferable to randomly generate data on pseudo surface defects within a range of specified size, thickness, shape, and position on the wafer surface in order to obtain a sufficient amount of training data.

[0016] The surface defect detection model is preferably configured using a convolutional neural network (CNN), as this allows for suitable processing of thickness measurement data on the semiconductor wafer surface.

[0017] The wafer thickness measuring device preferably uses the principle of optical interference as its detection principle, which is preferable from the viewpoint of suppressing deviations due to axial misalignment and displacement caused by vibration.

[0018] The surface defect detection device of the present invention is a surface defect detection device that detects surface defects from thickness measurement data of the entire surface of a semiconductor wafer measured using a wafer thickness measurement device, and is characterized in that it extracts and quantifies characteristics of the surface defects from the thickness measurement data, uses the quantified surface defect data to assign pseudo surface defects to thickness measurement data of a semiconductor wafer that has no surface defects, includes the thickness measurement data of the semiconductor wafer to which the pseudo surface defects have been assigned as training data, machine-learns a surface defect detection model, and uses the surface defect detection model to detect surface defects from the thickness measurement data of the semiconductor wafer to be inspected.

[0019] Furthermore, the present invention provides a surface defect detection program for detecting surface defects from thickness measurement data of an entire surface of a semiconductor wafer measured using a wafer thickness measurement device, the program comprising: extracting and quantifying surface defect characteristics from the thickness measurement data; using the quantified surface defect data to assign pseudo surface defects to thickness measurement data of a semiconductor wafer having no surface defects; including the thickness measurement data of the semiconductor wafer to which the pseudo surface defects have been assigned as training data; and causing an information processing device to execute a process of detecting surface defects from the thickness measurement data of the semiconductor wafer to be inspected using the surface defect detection model. This program can be recorded on a computer-readable storage medium. The storage medium can be a non-transitory medium such as a semiconductor memory, a hard disk, a magnetic recording medium, or an optical recording medium. The present invention can also be embodied as a computer program product. [Effects of the Invention]

[0020] According to each aspect of the present invention, it is possible to provide a surface defect detection method, a surface defect detection device, and a surface defect detection program that utilize machine learning to detect the presence or absence of surface defects on semiconductor wafers and their position information, even if sufficient training data cannot be secured. [Brief explanation of the drawings]

[0021] [Figure 1] FIG. 1 is a flowchart showing the steps of machine learning in the surface defect detection method according to this embodiment. [Figure 2] FIG. 2 shows an example of an image visualizing coordinate data before and after the surface defect is created, and an example of an image visualizing position data of the surface defect. [Figure 3] FIG. 3 is a flowchart showing an example of the procedure of the surface defect detection method. [Figure 4]FIG. 4 shows an example of a visualized image based on thickness data of a silicon wafer to which the created pseudo surface defects have been added, and a visualized image of position data of the surface defects. [Figure 5] FIG. 5 is a diagram showing the evaluation results of 515 silicon wafers used in the example. [Figure 6] FIG. 6 is a diagram illustrating an example of the hardware configuration of a surface defect detection device. DETAILED DESCRIPTION OF THE INVENTION

[0022] Hereinafter, embodiments of the present invention will be described with reference to the drawings. However, the present invention is not limited to the embodiments described below. In addition, the same or corresponding elements in each drawing are appropriately designated by the same reference numerals. Furthermore, it should be noted that the drawings are schematic, and the dimensional relationships and ratios of each element may differ from those in reality. There may also be parts in which the dimensional relationships and ratios differ between the drawings.

[0023] The surface defect detection method according to the embodiment of the present invention is a surface defect detection method for detecting surface defects from thickness measurement data of the entire surface of a semiconductor wafer measured using a wafer thickness measurement device. Therefore, the method for acquiring the thickness measurement data of the entire surface of a semiconductor wafer will be described first.

[0024] Dimples are shallow surface defects on the surface (rough surface) of a semiconductor wafer that has scratches (also called grinding marks or grinding scratches) left after the grinding process. The spacing between the scratches is very narrow, and depending on the size of the surface defect, conventional detection methods may overlook the surface defect. On the other hand, thickness measurement data can accurately observe the target semiconductor wafer and obtain information on the entire surface of the semiconductor wafer, regardless of the spacing between scratches or the size of the surface defect.

[0025] The surface defect detection method according to this embodiment can employ a wafer thickness measurement device that uses a conventionally known thickness measurement method. For example, a displacement sensor or a spectral interference sensor can be used to measure the thickness of the entire surface of a semiconductor wafer. One method using a displacement sensor involves installing a displacement sensor on each of the front and back sides of the semiconductor wafer. This method allows the thickness to be calculated by subtracting the distance to the front and back sides of the semiconductor wafer from the distance between the two displacement sensors. A method using a spectral interference sensor involves measuring the reflected light from the front and back sides of the semiconductor wafer using a coaxial laser to calculate the thickness (see, for example, Patent Document 2). From the perspective of suppressing axial misalignment and displacement deviations due to vibration, the surface defect detection method according to this embodiment preferably employs a wafer thickness measurement device that uses a spectral interference sensor.

[0026] Next, the creation of training data and machine learning of a detection model will be described. Fig. 1 is a flowchart showing the steps of machine learning in the surface defect detection method according to this embodiment. As shown in Fig. 1, the surface defect detection method according to this embodiment includes a process of acquiring thickness measurement data (step S1), a process of quantifying the characteristics of surface defects (step S2), a process of adding pseudo surface defects (step S3), and a process of machine learning a detection model (step S4).

[0027] In the process of acquiring thickness measurement data (step S1), as described above, thickness measurement data of the entire surface of a semiconductor wafer is acquired using a wafer thickness measurement device. Note that the process of acquiring thickness measurement data here does not necessarily include a process of measuring the thickness of a semiconductor wafer anew, but also includes a process of acquiring thickness measurement data measured in the past.

[0028] In the process of quantifying the characteristics of the surface defects (step S2), the characteristics of the surface defects are quantified from the thickness data of the semiconductor wafer having the surface defects. The quantification of the characteristics refers to quantifying the characteristics related to the size, thickness, shape, and position on the wafer surface of the detected actual surface defects.

[0029] In the process of adding pseudo surface defects (step S3), the quantified surface defect data is used to add pseudo surface defects to the thickness measurement data of a semiconductor wafer without surface defects. From the characteristics of the quantified surface defects, ranges are defined for the size, thickness, shape, and wafer surface position of the surface defects, and thickness data of the pseudo surface defects randomly generated from the defined ranges is incorporated into the thickness data of the semiconductor wafer without surface defects. The thickness data (coordinate data) in which the pseudo surface defects are incorporated into the thickness measurement data of the semiconductor wafer without surface defects is used as training data.

[0030] An example of an image visualizing the coordinate data before and after the surface defects are created is shown in Figure 2. The coordinate data used here is two-dimensional data, and for example, data showing the thickness at each point as a number on the x-coordinate and y-coordinate plane can be used.

[0031] Furthermore, data indicating the positions of surface defects relative to the created coordinate data (position data) is also obtained and used as training data. Like coordinate data, position data is also used as two-dimensional data, but in the case of position data, it is possible to use numerical data in which, for example, positions on the x- and y-coordinate plane where pseudo defects have been incorporated are output as 1 and other points are output as 0. Figure 2 shows an example of an image visualizing the thickness data (coordinate data) of a semiconductor wafer with the created pseudo surface defects added, and an image visualizing the corresponding position data of the surface defects.

[0032] In the process of machine learning a detection model (step S4), the thickness measurement data of the semiconductor wafer to which the pseudo surface defects have been added is included in the training data, and a surface defect detection model is machine learned. The surface defect detection model is a model that uses the thickness measurement data of the semiconductor wafer as input and outputs the position of the surface defect. In the process of machine learning a detection model (step S4), the thickness measurement data of the semiconductor wafer to which the pseudo surface defects have been added is included in the training data, and the surface defect detection model is machine learned, so that the position of the surface defect can be output even for thickness measurement data of the semiconductor wafer that is not included in the training data.

[0033] In this case, it is preferable to perform machine learning using as training data the position data of the surface defects of the semiconductor wafer to which the pseudo surface defects are added, which is paired with the thickness data (coordinate data) of the semiconductor wafer to which the pseudo surface defects are added. In the present invention, by using the position data as training data, the positions of the surface defects can be clearly determined in the determination step.

[0034] The surface defect detection model can be processed using a known model structure, such as ViT (Vision Transformer) or CNN (Convolutional Neural Network). ViT extracts features from the entire image in parallel, and the receptive field performs global recognition across the entire image. On the other hand, CNN extracts features within the range of the kernel, performing local recognition. In the present invention, since the purpose is to detect the presence or absence of surface defects and their position information over the entire surface of a semiconductor wafer, it is preferable to use CNN for model learning. This enables accurate automatic detection of surface defects using semiconductor wafer thickness data.

[0035] Fig. 3 is a flowchart showing an example of the procedure for the surface defect detection method using the detection model for surface defects learned by machine learning as described above. As shown in Fig. 3, the procedure for the surface defect detection method includes a process for acquiring thickness measurement data (step S5), a process for detecting surface defects using the detection model (step S6), a process for confirming the results (step S7), and a process for extracting the semiconductor wafer or proceeding to the next processing (step S8).

[0036] In the process of acquiring thickness measurement data (step S5), as described above, thickness measurement data for the entire surface of the semiconductor wafer being inspected is acquired using a wafer thickness measurement device. In the process of detecting surface defects using a detection model (step S6), surface defects are detected from the thickness measurement data of the semiconductor wafer being inspected using a surface defect detection model that has been machine-learned as described above. In the process of confirming the results (step S7), an operator confirms whether or not the semiconductor wafer that was determined to have a surface defect in the process of detecting surface defects using the detection model (step S6) actually has a surface defect. In the process of sampling the semiconductor wafer or proceeding to the next processing (step S8), the semiconductor wafer that was determined to have a surface defect is sampled, or if it was determined that no surface defect was present, the wafer proceeds to the next processing.

[0037] [Example] Hereinafter, an example will be described in which the effect of the surface defect detection method according to this embodiment is verified.

[0038] (wafer thickness measurement) The semiconductor wafers used were silicon wafers extracted immediately after the grinding process and had scratches caused by the grinding process. Wafer thickness data was obtained by measuring the entire surface of the wafer at 1.5 mm intervals using a wafer thickness measurement device (Keyence SI-F microhead type spectral interference laser displacement meter) that uses a laser clamping spectral interferometry method.

[0039] (Provisions for surface defects) A dimple is defined as a portion where the thickness has been reduced within a specific range of size and depth on a silicon wafer that has undergone the same process, and the ranges of the size, thickness, shape, and wafer surface position of the target dimples from silicon wafers having dimples are specified as follows. Size: 4~8mm 2 Thickness: 2~10μm Shape: Normal distribution shape in the depth direction Location: Wafer surface outside the central circle with a radius of 6 mm

[0040] (Creating training data) For the thickness data of silicon wafers without dimples, dimple information was randomly generated within a range of specified dimple sizes, thicknesses, shapes, and wafer surface positions, and this was incorporated into the silicon wafer thickness data, and thickness data (two-dimensional coordinate data) and corresponding two-dimensional position data were created for 1,000 silicon wafers with artificial surface defects added. Figure 4 shows an example of a visualized image based on the thickness data of silicon wafers with artificial surface defects added.

[0041] (machine learning detection model) Using the thickness data of the 1,000 silicon wafers with the pseudo-surface defects created, data showing the location of each surface defect, and the thickness data of four silicon wafers without dimples, machine learning of a surface defect detection model was performed using a convolutional neural network (CNN).

[0042] (Surface defect determination) Dimples were detected using a surface defect detection model based on the silicon wafer thickness data measured for the sample to be inspected. For silicon wafers determined to have dimples, the locations of dimples on the actual sample to be inspected were visually confirmed based on the position information of the detected dimples. Samples in which dimples were confirmed were removed from the process, while silicon wafers determined to have no dimples were returned to the process and wafers were manufactured. All of the manufactured silicon wafers were determined to be pass-grade products with no dimples. Figure 5 shows the results of the inspection of 515 silicon wafers used in the example. At this time, 8 of the silicon wafers determined to have defects by implementing the present invention were determined to have no defects by visual inspection, resulting in a 0% oversight rate of silicon wafers with dimples.

[0043] As shown in Figure 5, in the evaluation of 515 silicon wafers in this example, 66 were determined to have dimples. Of the 66 wafers determined to have dimples, 8 were visually determined to have no dimples and were subsequently determined to be acceptable products on the production line. On the other hand, of the 449 wafers determined to have no dimples, none actually had dimples. In other words, the oversight rate was 0%. [Hardware configuration example] Fig. 6 is a diagram showing an example of the hardware configuration of a surface defect detection device. That is, the surface defect detection device makes it possible to realize each function of the surface defect detection device by executing the above-described surface defect detection method as a program on an information processing device (computer) 10 employing the hardware configuration shown in Fig. 6. However, the example of the hardware configuration shown in Fig. 6 is an example of a hardware configuration that realizes each function of the surface defect detection device, and is not intended to limit the hardware configuration of the surface defect detection device. The surface defect detection device may include hardware not shown in Fig. 6.

[0044] As shown in FIG. 6, the hardware configuration that can be adopted by the surface defect detection device includes a CPU (Central Processing Unit) 11, a main memory device 12, an auxiliary memory device 13, and an IF (Interface) unit 14, which are interconnected by, for example, an internal bus.

[0045] The CPU 11 executes each command included in the surface defect detection program executed by the information processing device (computer) 10. The main storage device 12 is, for example, a RAM (Random Access Memory), and temporarily stores various programs, such as the surface defect detection program executed by the information processing device (computer) 10, for processing by the CPU 11.

[0046] The auxiliary storage device 13 is, for example, an HDD (Hard Disk Drive), and is capable of storing, for the medium to long term, various programs such as the surface defect detection program executed by the information processing device (computer) 10. Various programs such as the surface defect detection program can be provided as a program product recorded on a non-transitory computer-readable storage medium.

[0047] The IF unit 14 provides an interface relating to input and output of, for example, a surface defect detection device.

[0048] The information processing device (computer) 10 employing the above-described hardware configuration executes the above-described surface defect detection method as a program, thereby realizing each function of the surface defect detection device.

[0049] Although the present invention has been described above based on the embodiments, the present invention is not limited to the above-described embodiments. Modifications and adjustments of the embodiments and examples are possible based on the basic technical concept of the present invention. Furthermore, various combinations and selections (including partial deletions) of various disclosed elements (including elements of each claim, each element of each embodiment or example, each element of each drawing, etc.) are possible within the overall disclosure of the present invention. [Explanation of symbols]

[0050] 10. Information processing equipment 11 CPU 12 Main storage 13 Auxiliary storage device 14 IF Section

Claims

1. A surface defect detection method for detecting surface defects from thickness measurement data of an entire surface of a semiconductor wafer measured using a wafer thickness measurement device, comprising: extracting and quantifying surface defect characteristics from the thickness measurement data; adding pseudo surface defects to thickness measurement data of a semiconductor wafer having no surface defects using data of the quantified surface defects; a detection model for surface defects is machine-learned by including thickness measurement data of the semiconductor wafer to which the pseudo surface defects have been added as training data; A surface defect detection method, characterized in that surface defects are detected from thickness measurement data of a semiconductor wafer to be inspected using the surface defect detection model.

2. 2. The surface defect detection method according to claim 1, wherein the semiconductor wafer is a semiconductor wafer that has been subjected to a grinding process or an etching process but has not yet been subjected to a mirror polishing process.

3. 2. The surface defect detection method according to claim 1, wherein the semiconductor wafer is a semiconductor wafer having scratches.

4. 2. The surface defect detection method according to claim 1, wherein the surface defect is a dimple.

5. 2. The surface defect detection method according to claim 1, wherein the teaching data includes position information on the semiconductor wafer.

6. 2. The surface defect detection method according to claim 1, wherein the data of the pseudo surface defects is randomly generated within a range of a specified size, thickness, shape, and position on the wafer surface.

7. 2. The surface defect detection method according to claim 1, wherein the surface defect detection model is configured using a convolutional neural network (CNN).

8. 2. The surface defect detection method according to claim 1, wherein the wafer thickness measuring device uses the principle of optical interference as a detection principle.

9. A surface defect detection device that detects surface defects from thickness measurement data of an entire surface of a semiconductor wafer measured using a wafer thickness measurement device, extracting and quantifying surface defect characteristics from the thickness measurement data; adding pseudo surface defects to thickness measurement data of a semiconductor wafer having no surface defects using data of the quantified surface defects; a detection model for surface defects is machine-learned by including thickness measurement data of the semiconductor wafer to which the pseudo surface defects have been added as training data; A surface defect detection device that detects surface defects from thickness measurement data of a semiconductor wafer to be inspected using the surface defect detection model.

10. A surface defect detection program for detecting surface defects from thickness measurement data of an entire surface of a semiconductor wafer measured using a wafer thickness measurement device, extracting and quantifying surface defect characteristics from the thickness measurement data; adding pseudo surface defects to thickness measurement data of a semiconductor wafer having no surface defects using data of the quantified surface defects; a detection model for surface defects is machine-learned by including thickness measurement data of the semiconductor wafer to which the pseudo surface defects have been added as training data; A surface defect detection program that causes an information processing device to execute a process of detecting surface defects from thickness measurement data of a semiconductor wafer to be inspected using the surface defect detection model.

Images